METHODS OF DESIGNING SHORT HAIRPIN RNAS (shRNAs) FOR GENE

SILENCING

This application claims benefit under 35 U. S. C. § 119(e) of U.S. Provisional Patent Application No. 60/995,156, filed on September 24, 2007, which is incorporated by reference herein in its entirety.

1. FIELD OF THE INVENTION

The invention relates to methods for designing shRNAs with higher overall silencing efficacy. The invention also relates to methods for designing DNA constructs that can express shRNAs with high transcription efficiencies.

2. BACKGROUND OF THE INVENTION

RNA interference (RNAi) suppresses gene expression in cells from various organisms including mammalian cells, and has generated much excitement in the scientific community (Couzin, 2002, Science 298:2296-2297; McManus et al. , 2002, Nat. Rev. Genet. 3:737-747; Hannon, G. J., 2002, Nature 418:244-251; Paddison et al., 2002, Cancer Cell 2:17-23). RNAi is conserved throughout evolution, from C. elegans to humans, and is believed to function in protecting cells from invasion by RNA viruses. When a cell is infected by a double stranded RNA (dsRNA) virus, the dsRNA is recognized and targeted for cleavage by an RNaselll-type enzyme termed Dicer. The Dicer enzyme "dices" the RNA into short duplexes of 21 to 23 nucleotides (nt), termed short-interfering RNAs (siRNAs), which are composed of 19 to 21 nt of perfectly paired ribonucleotides with two unpaired nucleotides on the 3' end of each strand. These short duplexes associate with a multiprotein complex termed RNA-induced silencing complex (RISC), and direct this complex to mRNA transcripts with sequence similarities to the siRNAs. As a result, nucleases present in the RISC complex cleave the mRNA transcripts, thereby abolishing expression of the gene product. In the case of viral infection, this mechanism would result in destruction of viral transcripts, thus preventing viral synthesis. Since the siRNAs are double-stranded, either strand has the potential to associate with RISC and direct silencing of transcripts with sequence similarity.

Specific gene silencing promises the potential to harness human genome data to elucidate gene function, identify drug targets, and develop more specific therapeutics.

Many of these applications assume high degree of specificities of siRNAs for their intended targets. Cross-hybridization with transcripts containing partial identity to the siRNA sequence may elicit phenotypes reflecting silencing of unintended transcripts in addition to the target gene. This could confound the identification of the gene implicated in the phenotype. Numerous reports in the literature purport the exquisite specificity of siRNAs, suggesting a requirement for near-perfect identity with the siRNA sequence (Elbashir et al, 2001 EMBOJ. 20:6877-6888; Tuschl et al, 1999, Genes Dev. 13:3191- 3197; and Hutvagner et al, 2002, Sciencexpress 297:2056-2060). One recent report suggests that perfect sequence complementarity is required for siRNA-targeted transcript cleavage, while partial complementarity will lead to translational repression without transcript degradation, in the manner of microRNAs (Hutvagner et al, Sciencexpress 297:2056-2060).

Besides siRNAs, microRNAs (miRNAs) also function to regulate gene expression. miRNAs are regulatory RNAs expressed from the genome, and are processed from precursor stem-loop structures to produce single-stranded nucleic acids that bind to sequences in the 3' UTR of the target mRNA (Lee et al, 1993, Cell 75:843-854; Reinhart et al, 2000, Nature 403:901-906; Lee et al, 2001, Science 294:862-864; and Lau et al, 2001, Science 294:858-862; Hutvagner et al, 2001, Science 293:834-838). miRNAs bind to transcript sequences with only partial complementarity (Zeng et al , 2002, Molec. Cell 9:1327-1333) and repress translation without affecting steady-state RNA levels (Lee et al, 1993, Cell 75:843-854; and Wightman et al, 1993, Cell 75:855-862). Both miRNAs and siRNAs are produced by Dicer and associate with components of the RISC complex (Hutvagner et al, 2001, Science 293:834-838; Grishok et al, 2001, Cell 106: 23-34; Ketting et al, 2001, Genes Dev. 15:2654-2659; Williams et al, 2002, Proc. Natl Acad. Sci. USA 99:6889-6894; Hammond et al, 2001, Science 293:1146-1150; and Mourlatos et al, 2002, Genes Dev. 16:720-728). A recent report (Hutvagner et al, 2002, Sciencexpress 297:2056-2060) hypothesizes that gene regulation through the miRNA pathway versus the siRNA pathway is determined solely by the degree of complementarity to the target transcript. It is speculated that siRNAs with only partial complementarity to the mRNA target will function in translational repression, similar to an miRNA, rather than triggering RNA degradation. Synthetic siRNAs have been shown to be able to induce RNAi in live cells, for example, plant cells and mammalian cells.

The ability of utilizing siRNAs for gene silencing in vivo enables possible selection and development of siRNAs for therapeutic use. A recent report highlights the potential

therapeutic application of siRNAs. Fas-mediated apoptosis is implicated in a broad spectrum of liver diseases, where lives could be saved by inhibiting apoptotic death of hepatocytes. When mice were injected intravenously with siRNAs targeting the Fas receptor, the Fas gene was silenced in mouse hepatocytes at the mRNA and protein levels, thereby preventing apoptosis and protecting the mice from hepatitis-induced liver damage (Song et ah, 2003, Nat. Medicine 9:347-351). Thus, silencing Fas expression holds therapeutic promise to prevent liver injury by protecting hepatocytes from cytotoxicity. In another example, mice were injected intraperitoneally with siRNAs targeting TNF-a. Lipopolysaccharide-induced TNF-a gene expression was inhibited, and these mice were protected from sepsis. Collectively, these results suggest that siRNAs can function in vivo, and may hold potential as therapeutic drugs (Sorensen et al, 2003, J. MoI. Biol. 327: 761-766).

United States Publication No. 6,506,559 discloses an RNA interference process for inhibiting expression of a target gene in a living cell. The patent discloses introducing partially or fully doubled-stranded RNAs having a sequence in the duplex region that is identical to a sequence in the target gene into the living cell or into the extracellular environment. RNA sequences with insertions, deletions, and single point mutations relative to the target sequence are also disclosed as effective for expression inhibition. United States Patent Publication No. US 2002/0086356 discloses RNA interference in a Drosophila in vitro system using RNA segments of 21 -23 nt in length. The patent publication teaches that when these 21-23 nt fragments are purified and added back to Drosophila extracts, they mediate sequence-specific RNA interference in the absence of long double-stranded RNAs (dsRNAs). The patent publication also teaches that chemically synthesized oligonucleotides (oligos) of the same or similar nature can be used to target specific mRNAs for degradation in mammalian cells.

PCT publication WO 02/44321 discloses that double-stranded RNAs (dsRNAs) of 19-23 nt in length induce sequence-specific post-transcriptional gene silencing in a Drosophila in vitro system. The PCT publication teaches that siRNAs duplexes can be generated by an RNase Ill-like processing reaction from long dsRNAs or by chemically synthesized siRNA duplexes with overhanging 3' ends mediating efficient target RNA cleavage in the lysate where the cleavage site is located near the center of the region spanned by the guiding siRNA. The PCT publication also provides evidence that the direction of dsRNA processing determines whether sense or antisense-identical target RNA can be cleaved by the produced siRNA complex. Systematic analyses of the effects

of length, secondary structure, sugar backbone and sequence specificity of siRNAs on RNA interference have been disclosed to aid siRNA design. In addition, silencing efficacy has been shown to correlate with the GC content of the 5' and 3' regions of the 19 base pair target sequence. It was found that siRNAs targeting sequences with a GC rich 5' and GC poor 3' perform the best. More detailed discussion may be found in Elbashir et al, 2001, EMBO J. 20:6877-6888 and Aza-Blanc et al, 2003, MoI. Cell 12:627-637; each of which is hereby incorporated by reference herein in its entirety.

In addition, siRNA design algorithms are disclosed in PCT publications WO 2005/018534 A2 and WO 2005/042708 A2; each of which is hereby incorporated by reference herein in its entirety. Specifically, WO 2005/018534 A2 discloses methods and compositions for gene silencing using siRNA having partial sequence homology to its target gene. The application provides methods for identifying common and / or differential responses to different siRNAs targeting a gene. The application also provides methods for evaluating the relative activity of the two strands of an siRNA. The application further provides methods of using siRNAs as therapeutics for treatment of diseases. WO 2005/042708 A2 provides a method for identifying siRNA target motifs in a transcript using a position-specific score matrix approach. It also provides a method for identifying off-target genes of an siRNA using a position-specific score matrix approach. The application further provides a method for designing siRNAs with improved silencing efficacy and specificity as well as a library of exemplary siRNAs.

Despite the developments in oligo design resulting in improved silencing efficacies, the oligo-based siRNA technologies have limitations. The delivery options for the RNA molecules are limited. For example, in cell lines that are difficult to transfect, siRNA duplexes may not work at all. In addition, siRNAs can be expensive because oligos must be purchased or synthesized for every experiment. It is also difficult to trace or quantify oligo delivery, making normalization difficult.

DNA expression vector based short hairpin RNA (shRNA) technologies have been introduced to overcome some of the limitations of the siRNA based gene silencing technologies. An shRNA molecule contains a sense strand and an antisense strand, and a loop sequence between the sense and antisense strands. Due to the sequence complementarity of the sense and antisense strands, such RNA molecules form hairpin- shaped double-stranded RNAs (dsRNAs). Usually, sequence encoding an shRNA is cloned into a vector and the vector is introduced into a cell and transcribed by the cell's transcription machinery (Chen et al, 2003, Biochem Biophys Res Commun 311 :398-404).

The shRNAs can then be transcribed, for example, by RNA polymerase III (Pol III) in response to a Pol Ill-type promoter in the vector (Yuan et al, 2006, MoI Biol Rep 33:33- 41 and Scherer et al, 2004, MoI Ther 10:597-603). The expressed shRNAs are then exported into the cytoplasm where they are processed by proteins such as Dicer into siRNAs, which then trigger RNAi (Amarzguioui et al. , 2005, FEBS Letter 579:5974- 5981). It has been reported that purines are required at the 5' end of a newly initiated RNA for optimal RNA polymerase III transcription. More detailed discussion can be found in Zecherle et a/. , 1996, Mo/. Cell. Biol. 16:5801-5810; Fruscoloni et al, 1995, Nucleic Acids Res, 23:2914-2918; and Mattaj et al., 1988, Cell, 55:435-442. It has also been reported that cleavage of primary microRNA hairpins by Drosha requires flanking nonstructured RNA sequences (Zeng et al, 2005, J. Biol. Chem. 280:27595-27603; and Han et al, 2006, Cell 125:887-901). In addition, for example, for shRNAs expressed from a DNA construct by RNA polymerase III (Pol III) transcription, the transcription efficiency depends on whether the transcription start site contains purines or pyrimidines (see, e.g., Zecherle et al, 1996, Molecular and Cellular Biology, 16:5801-5810). The vector based expression of shRNA core sequences is advantageous over the oligo-based siRNAs for several reasons. It is more versatile than oligo- based siRNAs. The shRNAs core sequences can be expressed stably in cells, allowing long-term gene silencing in cells both in vitro and in vivo, e.g., in animals (see, McCaffrey et al, 2002, Nature 418:38-39; Xia et al, 2002, Nat. Biotech. 20:1006-1010; Lewis et al, 2002, Nat. Genetics 32:107- 108; Rubinson et al, 2003, Nat. Genetics 33:401-406; and Tiscornia et al, 2003, Proc. Natl. Acad. Sci. USA 100:1844-1848). The vector based expression of shRNA core sequences allows both long-term and transient gene expression for better controlled gene silencing than the oligo-based siRNA technology. Because DNA handling and delivery methods are well-established, long term shRNA-based RNAi experiments are more cost- effective. In addition, transfection efficiency can be normalized to accurately quantify reduction of gene expression (gene knockdown). Sophisticated DNA handling tools allow fluorescently tagged fusion proteins to be tagged to the corresponding siRNA duplex to easily trace sequence delivery. It is also possible to use flow cytometry to enrich transfected populations.

Despite the advantages, the silencing efficacies of shRNA core sequences are not as potent as siRNA duplexes harboring the same nucleotide sequence obtained by oligo synthesis. Multiple factors may account for the attenuated silencing efficacies of shRNAs when compared with their corresponding siRNAs. For example, the silencing efficacy of

an shRNA core sequence may be compromised if the DNA construct encoding the shRNA core sequence is not properly or efficiently transcribed. What is needed in the art are systems and methods for improving the silencing efficacies of shRNAs.

Discussion or citation of a reference herein shall not be construed as an admission that such reference is prior art to the present invention.

3. SUMMARY OF THE INVENTION

The present invention provides methods for classifying siRNA functional sequence motifs for use in an shRNA or shRNA core sequence designed to produce an siRNA for silencing a target gene in a cell. The shRNA core sequence comprises an antisense strand sequence and a sense strand sequence. In one aspect, the invention provides a method for selecting (a) selecting a first plurality of siRNA functional sequence motifs, wherein each siRNA functional sequence motif of said first plurality of siRNA functional sequence motifs has a predicted silencing efficacy against a targeted sequence motif within a transcript of a target gene and said predicted silencing efficacy is equal or greater than a chosen threshold of silencing efficacy; and (b) classifying each siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs as belonging to a first class of siRNA functional sequence motifs when said siRNA functional sequence motif has a 3' terminal nucleotide that is a cytosine, or as belonging to a second class of siRNA functional sequence motifs when said siRNA functional sequence motif does not have a 3 ' terminal nucleotide that is a cytosine. In some embodiments, the chosen threshold is 50%, 75%, 80% or 90% at an siRNA concentration of 100 nM in culture medium in which a cell is expressing said target gene is incubated.

In some embodiments, the method further comprises (aθ) determining, prior to said selecting step (a), said predicted silencing efficacy of an siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs based on a PSSM matrix. In other embodiments, the method further comprises (aθ) determining, prior to said selecting step (a), said predicted silencing efficacy of an siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs based on a predicted binding affinity between said siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs and a susceptible sequence motif of said target gene. In still other embodiments, the method further comprises (aθ) determining, prior to said selecting step (a), said predicted silencing efficacy of an siRNA functional sequence motif in said first

plurality of siRNA functional sequence motifs based on an observed silencing efficacy of said siRNA functional sequence motif.

In some embodiments in accordance with the present invention, an siRNA functional sequence motif of said first plurality of siRNA functional sequence motifs has at least 19 nucleotides, at least 21 nucleotides, or least 27 nucleotides.

In some embodiments, the further comprises (bl) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; the classification of one or more or all of said siRNA functional sequence motifs in said first plurality of siRNA functional sequence motifs as belonging to said first class or said second class resulting from said classifying step. In some embodiments, the method further comprises (iiil) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said sense strand sequence and /or said antisense strand sequence.

The present invention also provides methods of designing an shRNA or shRNA core sequence or sense strand sequence thereof or antisense strand sequence thereof, wherein said shRNA core sequence is designed to produce an siRNA for silencing a target gene in a cell of an organism. The shRNA core sequence comprises an antisense strand sequence and a sense strand sequence. In another aspect, the method comprises: (i) carrying out any method comprising any one of the selecting (a), classifying (b) or determining (aθ) steps; (ii) selecting an siRNA functional sequence strand sequence in an shRNA core sequence; and (iii) designing a sense strand sequence based on said antisense strand sequence in said shRNA core sequence, wherein said sense strand sequence is an inverted repeat to said antisense strand sequence. In some embodiments, the method further comprises (iv) designing a nucleotide sequence (hereinafter "loop sequence") in between said antisense strand sequence and said sense strand sequence, thereby completing said shRNA core sequence. In some embodiments, the loop sequence consists of at least 3 nucleotides, at least 7 nucleotides, at least 9 nucleotides, at least 15 nucleotides or at least 25 nucleotides.

In some embodiments, the method further comprises (v) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; one or more of said loop sequence, said sense strand sequence, said antisense strand sequence, and said shRNA core sequence.

The present invention also provides methods of designing an shRNA or shRNA core sequence or sense strand sequence thereof or antisense strand sequence thereof,

wherein said shRNA core sequence is designed to produce an siRNA for silencing a target gene in a cell. The shRNA core sequence comprises an antisense strand sequence and a sense strand sequence. In another aspect, the method comprises: (i) carrying out any method comprising any one of the selecting (a), classifying (b) or determining (aθ) steps; (ii) selecting an siRNA functional sequence motif from said second class of siRNA functional sequence motifs; (iii) replacing the nucleotide at the 3' terminus of said siRNA functional motif with a cytosine, thereby creating a new siRNA functional sequence motif to be an antisense strand sequence in an shRNA core sequence; and (iv) designing a sense strand sequence based on said antisense strand sequence in said shRNA core sequence, wherein said sense strand sequence is an inverted repeat to said antisense strand sequence. In some embodiments, the method further comprises (v) designing a nucleotide sequence (hereinafter "loop sequence") in between said antisense strand sequence and said sense strand sequence, thereby completing said shRNA core sequence. In some embodiments, the loop sequence consists of at least 3 nucleotides, at least 7 nucleotides, at least 9 nucleotides, at least 15 nucleotides or at least 25 nucleotides.

In some embodiments, the method further comprises (el) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said sense strand sequence and / or said antisense strand sequence. In some embodiments, the method further comprises (fl) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; one or more of said loop sequence, said sense strand sequence, said antisense strand sequence, and said shRNA core sequence.

The present invention also provides methods of designing an shRNA or shRNA core sequence or sense strand sequence thereof or antisense strand sequence thereof, wherein said shRNA core sequence is designed to produce an siRNA for silencing a target gene in a cell. The shRNA core sequence comprises an antisense strand sequence and a sense strand sequence. In still another aspect, the method comprises (a) selecting a first plurality of siRNA functional sequence motifs, wherein each siRNA functional sequence motif of said first plurality of siRNA functional sequence motifs has a predicted silencing efficacy against a targeted sequence motif within a transcript of a target gene and said predicted silencing efficacy is equal or greater than a chosen threshold of silencing efficacy; (b) selecting an siRNA functional sequence motif from said first plurality of siRNA functional sequence motifs; (c) adding a cytosine to the 3' terminus of said siRNA

functional sequence motif, thereby creating a new siRNA functional sequence motif to be an antisense strand sequence in an shRNA core sequence; and (d) designing a sense strand sequence based on said antisense strand sequence in said shRNA core sequence, wherein said sense strand sequence is an inverted repeat to said antisense strand sequence. In some embodiments, the method further comprise (e) designing a nucleotide sequence (hereinafter "loop sequence") in between said antisense strand sequence and said sense strand sequence, thereby completing said shRNA core sequence. In some embodiments, the loop sequence consists of at least 3 nucleotides, at least 7 nucleotides, at least 9 nucleotides, at least 15 nucleotides or at least 25 nucleotides. In some embodiments, the method further comprises (dl) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said sense strand sequence and / or said antisense strand sequence.

In other embodiments, the method further comprises (el) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; one or more of said loop sequence, said sense strand sequence, said antisense strand sequence, and said shRNA core sequence.

In some embodiments, the method further comprises (dl) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said sense strand sequence and / or said antisense strand sequence. In some embodiments, the method further comprises (el) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; one or more of said loop sequence, said sense strand sequence, said antisense strand sequence, and said shRNA core sequence.

The present invention provides methods of designing an shRNA or an shRNA core sequence or sense strand sequence thereof or antisense strand sequence thereof or designing siRNA functional sequence motifs for use in an shRNA core sequence where the shRNA core sequence is designed to produce an siRNA for silencing a target gene in a cell of an organism. The shRNA core sequence comprises an antisense strand sequence and a sense strand sequence. In still another aspect, the method comprises (a) selecting a first plurality of siRNA functional sequence motifs, wherein each siRNA functional sequence motif of said first plurality of siRNA functional sequence motifs has a predicted silencing efficacy against a targeted sequence motif within a transcript of a target gene and said predicted silencing efficacy is equal or greater than a chosen threshold of silencing efficacy; and (b) deleting from said first plurality of siRNA functional sequence motifs

each siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs that does not have a 3 ' terminal nucleotide that is a cytosine, thereby forming a first class of siRNA functional sequence motifs with a 3' terminal cytosine. In some embodiments, the chosen threshold is 50%, 75%, 80% or 90% at an siRNA concentration of 100 nM in culture medium in which a cell is expressing said target gene is incubated.

In some embodiments, the method further comprises (aθ) determining, prior to said selecting step (a), said predicted silencing efficacy of an siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs based on a PSSM matrix.

In some embodiments, the method further comprises (aθ) determining, prior to said selecting step (a), said predicted silencing efficacy of an siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs based on a predicted binding affinity between said siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs and a susceptible sequence motif of said target gene.

In some embodiments, the method further comprises (aθ) determining, prior to said selecting step (a), said predicted silencing efficacy of an siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs based on an observed silencing efficacy of said siRNA functional sequence motif.

In some embodiments, the loop sequence consists of at least 3 nucleotides, at least 7 nucleotides, at least 9 nucleotides, at least 15 nucleotides or at least 25 nucleotides. In some embodiments, the method further comprises (c) selecting an siRNA functional sequence motif from said first class of siRNA functional sequence motifs to be an antisense strand sequence in an shRNA core sequence; and (d) designing a sense strand sequence based on said antisense strand sequence in said shRNA core sequence, wherein said sense strand sequence is an inverted repeat to said antisense strand sequence. In other embodiments, the method further comprises (e) designing a nucleotide sequence

(hereinafter "loop sequence") in between said antisense strand sequence and said sense strand sequence, thereby completing said shRNA core sequence. In some embodiments, the loop sequence consists of at least 3 nucleotides, at least 7 nucleotides, at least 9 nucleotides, at least 15 nucleotides or at least 25 nucleotides. In some embodiments, the method further comprises (bl) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said first class of siRNA functional sequence motifs or one or more siRNA functional sequence motifs in said first class of siRNA functional sequence motifs.

In some embodiments, the method further comprises (dl) outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said sense strand sequence and / or said antisense strand sequence.

The present invention also provides methods of designing an shRNA or shRNA core sequence or sense strand sequence thereof or antisense strand sequence thereof or designing siRNA functional sequence motifs for use in an shRNA core sequence, where said shRNA core sequence is designed to produce an siRNA for silencing a target gene in a cell of an organism. The shRNA core sequence comprises an antisense strand sequence and a sense strand sequence. In still another aspect, the method comprises (a) selecting a first plurality of siRNA functional sequence motifs, wherein each siRNA functional sequence motif of said first plurality of siRNA functional sequence motifs has a predicted silencing efficacy against a targeted sequence motif within a transcript of a target gene and said predicted silencing efficacy is equal or greater than a chosen threshold of silencing efficacy; and (b) deleting from said first plurality of siRNA functional sequence motifs each siRNA functional sequence motif in said first plurality of siRNA functional sequence motifs that has a 3 ' terminal nucleotide that is a cytosine, thereby forming a second class of siRNA functional sequence motifs that do not have a 3' terminal cytosine. In some embodiments, the chosen threshold is 50%, 75%, 80% or 90% at an siRNA concentration of 100 nM in culture medium in which a cell is expressing said target gene is incubated. In some embodiments, the method further comprises (c) selecting an siRNA functional sequence motif from said second class of siRNA functional sequence motifs; (d) replacing the nucleotide at the 3' terminus of said siRNA functional motif with a cytosine, thereby creating a new siRNA functional sequence motif to be an antisense strand sequence in an shRNA core sequence; and (e) designing a sense strand sequence based on said antisense strand sequence in said shRNA core sequence, wherein said sense strand sequence is an inverted repeat to said antisense strand sequence.

In some embodiments, the method further comprises (f) designing a nucleotide sequence (hereinafter "loop sequence") in between said antisense strand sequence and said sense strand sequence, thereby completing said shRNA core sequence. In some embodiments, the loop sequence consists of at least 3 nucleotides, at least 7 nucleotides, at least 9 nucleotides, at least 15 nucleotides or at least 25 nucleotides.

In some embodiments, the method further comprises (el) outputting to a user interface device, a computer readable storage medium, or a local or remote computer

system; or displaying; one or more of said loop sequence, said sense strand sequence, said antisense strand sequence, and said shRNA core sequence.

In some embodiments, the method further comprises adding a flanking region at the 5' terminus of the shRNA core sequence; adding a flanking region at the 3' terminus of the shRNA core sequence; adding a first flanking region at the 3' terminus of said shRNA core sequence, and a second flanking region at the 5' terminus of said shRNA core sequence; or adding a promoter sequence 5' to said shRNA core sequence. In some embodiments, the promoter sequence is selected from the group consisting of a human, bovine or mouse Hl promoter, a human, bovine or mouse U6 promoter, and a human, bovine or mouse U2 promoter.

In some embodiments, the method further comprises outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said flanking region(s) at the 5' terminus and / or the 3' terminus of said shRNA core sequence. In some embodiments, the method further comprises outputting to a user interface device, a computer readable storage medium, or a local or remote computer system; or displaying; said promoter sequence and / or said shRNA core sequence. In some embodiments, the method further comprises (aθ) selecting, prior to said selecting step (a), said first plurality of siRNA functional sequence motifs from a database of siRNA and miRNA sequences, or (aθ) selecting, prior to said selecting step (a), said first plurality of siRNA functional sequence motifs from sequences of 21U-RNAs.

The present invention also provides an apparatus comprising a processor, and a memory coupled to said processor and encoding one or more programs, wherein said one or more programs cause the processor to carry out a method of any embodiments in accordance with the present invention.

The present invention also provides a computer program product for use in conjunction with a computer having a processor and a memory connected to the processor, said computer program product comprising a computer readable storage medium having a computer program mechanism encoded thereon, wherein said computer program mechanism may be loaded into the memory of said computer and cause said computer to carry out the method of any embodiments in accordance with the present invention.

In one aspect, the invention provides a method for selecting from a plurality of different siRNAs one or more siRNAs for silencing a target gene in an organism, each of the plurality of different siRNAs targeting a different target sequence in a transcript of the target gene, the method comprising (a) ranking the plurality of different siRNAs according to positional base compositions of a corresponding targeted sequence motifs in the transcript, wherein each targeted sequence motif comprises at least a portion of the target sequence of the corresponding siRNA and/or a second sequence in a sequence region flanking the target sequence; and (b) selecting one or more siRNAs from the ranked siRNAs. In a preferred embodiment, each sequence motif comprises the target sequence of the targeting siRNA. In another embodiment, the ranking step is carried out by (al) determining a score for each different siRNA, wherein the score is calculated using a position-specific score matrix; and (a2) ranking the plurality of different siRNAs according to the score.

In one embodiment, each sequence motif is a nucleotide sequence of L nucleotides, L being an integer, and the position-specific score matrix is {\og(e , _j /p, _j )}, where e _υ is the weight of nucleotide i at position/, p _v is the weight of nucleotide i at position/ in a random sequence, and / = G, C, A, U(T), j = 1, ..., L. In another embodiment, each sequence motif is a nucleotide sequence of Z nucleotides, L being an integer, and the position-specific score matrix is {log(e , _j /p, _j )}, where e _υ is the weight of nucleotide i at position/, p _υ is the weight of nucleotide i at position/ in a random sequence, and i = G or C, A, U(T), j = 1, .... L.

In one embodiment, the score for each siRNA is calculated according to equation

L

Score = ∑\n(e _t l p _t )

wherein e _t and/?, are respectively weights of the nucleotide at position t in the sequence motif as determined according to the position-specific score matrix and in a random sequence.

In another embodiment, each sequence motif comprises the target sequence of the targeting siRNA and at least one flanking sequence. Preferably, each sequence motif comprises the target sequence of the targeting siRNA and a 5' flanking sequence and a 3' flanking sequence. In one embodiment, the 5' flanking sequence and the 3' flanking sequence are each a sequence of D nucleotides, D being an integer. In a specific embodiment, each target sequence is a sequence of 19 nucleotides, and each 5' flanking sequence and 3' flanking sequence are a sequence of 10 nucleotides. In another specific

embodiment, each target sequence is a sequence of 19 nucleotides, and each 5' flanking sequence and 3' flanking sequence are a sequence of 50 nucleotides.

Preferably, the one or more siRNAs consist of at least 3 siRNAs. In another embodiment, the method further comprises a step of de-overlapping, comprising selecting a plurality of siRNAs among the at least 3 siRNAs such that siRNAs in the plurality are sufficiently different in a sequence diversity measure. In one embodiment, the diversity measure is a quantifiable measure, and the selecting in the de-overlapping step comprises selecting siRNAs having a difference in the sequence diversity measure between different selected siRNAs above a given threshold. In one embodiment, the sequence diversity measure is the overall GC content of the siRNAs. In one embodiment, the given threshold is 5%. In another embodiment, the sequence diversity measure is the distance between siRNAs along the length of the transcript sequence. In one embodiment, the threshold is 100 nucleotides. In still another embodiment, the sequence diversity measure is the identity of the leading dimer of the siRNAs, wherein each of the 16 possible leading dimers is assigned a score of 1-16, respectively. In one embodiment, the threshold is 0.5.

In another embodiment, the method further comprises a step of selecting one or more siRNAs based on silencing specificity, the step of selecting based on silencing specificity comprising, (i) for each of the plurality of siRNAs, predicting off-target genes of the siRNA from among a plurality of genes, wherein the off-target genes are genes other than the target gene and are directly silenced by the siRNA; (ii) ranking the plurality of siRNAs according to their respective numbers of off-target genes; and (iii) selecting one or more siRNAs for which the number of off-target genes is below a given threshold.

In one embodiment, the predicting comprises (il) evaluating the sequence of each of the plurality of genes based on a predetermined siRNA sequence match pattern; and (i2) predicting the gene as an off-target gene if the gene comprise a sequence that matches the siRNA based on the sequence match pattern. In one embodiment, the step of evaluating comprises identifying an alignment of the siRNA to a sequence in a gene by a low stringency FASTA alignment.

In one embodiment, each siRNA has L nucleotides in its duplex region, and the match pattern is represented by a position match position-specific score matrix

(pmPSSM), the position match position-specific score matrix consisting of weights of different positions in an siRNA to match transcript sequence positions in an off-target transcript [P _j ), where 7 = 1, ..., L, P ₁ is the weight of a match at position/

In another embodiment, the step (il) comprises calculating a position match score pmScore according to equation

L pmScore = ∑ In(E ₁ /0.25) ι=l where E ₁ = P ₁ if position / is a match and E ₁ = (l-P,)/3 if position i is a mismatch; and the step (i2) comprises predicting the gene as an off-target gene if the position match score is greater than a given threshold.

In a preferred embodiment, L is 19, and the pmPSSM is given by Table I. Preferably, the plurality of genes comprises all known unique genes of the organism other than the target gene. In one embodiment, the position-specific score matrix (PSSM) is determined by a method comprising (aa) identifying a plurality of iV siRNAs consisting of siRNAs having

19-nucleotide duplex region and having a silencing efficacy above a chosen threshold;

(bb) identifying for each siRNA a functional sequence motif, the functional sequence motif comprising a 19-nucleotide target sequence of the siRNA and a 10-nucleotide 5' flanking sequence and a 10-nucleotide 3' flanking sequence; (cc) calculating a frequency matrix {f _υ }, where / = G, C, A, U(T); j = 1, 2, ..., L, and where f _υ is the frequency of the

/th nucleotide at they ^' th position, based on the siRNAs functional sequence motifs according to equation

, and (d) determining the PSSM by calculating e,, according to 0,if k ≠ i equation

In another embodiment, the position-specific score matrix (PSSM) is obtained by a method comprising (aa) initializing the PSSM with random weights; (bb) selecting randomly a weight w _υ obtained in (aa); (cc) changing the value of the selected weight to generate a test psPSSM comprising the selected weight having the changed value; (dd) calculating a score for each of a plurality of siRNAs functional sequence motifs using the test PSSM according to equation

L

Score = ∑\n(w _k lp _k )

wherein the w* and p _k are respectively weights of a nucleotide at position k in the functional sequence motif and in a random sequence; (ee) calculating correlation of the score and a metric of a characteristic of an siRNA among the plurality of siRNAs functional sequence motifs; (ff) repeating steps (cc)-(ee) for a plurality of different values of the selected weight in a given range and retain the value that corresponds to the best correlation for the selected weight; and (gg) repeating steps (bb)-(ff) for a chosen number of times; thereby determining the PSSM.

In one embodiment, the method further comprises selecting the plurality of siRNA functional sequence motifs by a method comprising (i) identifying a plurality of siRNAs consisting of siRNAs having different values in the metric; (ii) identifying a plurality of siRNA functional sequence motifs each corresponding to an siRNA in the plurality of siRNAs. In a preferred embodiment, the characteristic is silencing efficacy.

In one embodiment, the plurality of N siRNAs targets a plurality of different genes having different transcript abundances in a cell. In one embodiment, step (b) is carried out by selecting one or more siRNAs having the highest scores. In another embodiment, step (b) is carried out by selecting one or more siRNAs having a score closest to a predetermined value, wherein the predetermined value is the score value corresponding to the maximum median silencing efficacy of a plurality of siRNA sequence motifs. In a preferred embodiment, the plurality of siRNA sequence motifs are sequence motifs in transcript having abundance level of less than about 3-5 copies per cell.

In another embodiment, step (b) is carried out by selecting one or more siRNAs having a score within a predetermined range, wherein the predetermined range is a score range corresponding to a plurality of siRNAs sequence motifs having a given level of silencing efficacy. In one embodiment, the silencing efficacy is above 50%, 75%, or 90% at an siRNA concentration of about 10 μM, l μM, 100 nM, 50 nM, 10 nM or 1 nM , preferably at 100 nM.

In a preferred embodiment, the plurality of siRNA sequence motifs are sequence motifs in transcript having abundance level of less than about 3-5 copies per cell. In another preferred embodiment, the plurality of N siRNAs comprises at least 10,

50, 100, 200, or 500 different siRNAs.

In another embodiment, the position-specific score matrix (PSSM) comprises W _k , k =1, ..., L, W _k being a difference in probability of finding nucleotide G or C at sequence

position k between a first type of siRNA and a second type of siRNA, and the score for each strand is calculated according to equation

Score = ^T w i _k t=i

In one embodiment, the first type of siRNA consists of one or more siRNAs having silencing efficacy no less than a first threshold and the second type of siRNA consists of one or more siRNAs having silencing efficacy less than a second threshold.

In one embodiment, the difference in probability is described by a sum of Gaussian curves, each of the Gaussian curves representing the difference in probability of finding a G or C at a different sequence position. In one embodiment, the first and second thresholds are both 75% at an siRNA concentration of 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM, preferably at 100 nM.

In another aspect, the invention provides a method for selecting from a plurality of different siRNAs one or more siRNAs for silencing a target gene in an organism, each of the plurality of different siRNAs targeting a different target sequence in a transcript of the target gene, the method comprising (a) ranking the plurality of different siRNAs according to positional base composition of reverse complement sequences of sense strands of the siRNAs; and (b) selecting one or more siRNAs from the ranked siRNAs.

In one embodiment, the ranking step is carried out by (al) determining a score for each different siRNA, wherein the score is calculated using a position-specific score matrix; and (a2) ranking the plurality of different siRNAs according to the score.

In one embodiment, the siRNA has a nucleotide sequence of L nucleotides in its duplex region, L being an integer, wherein the position-specific score matrix comprises Wi ₁ , k =1, ..., L, W _k being a difference in probability of finding nucleotide G or C at sequence position k between reverse complement of sense strand of a first type of siRNA and reverse complement of sense strand of a second type of siRNA, and the score for each reverse complement is calculated according to equation

Score = ^ W i ₁ ,

In one embodiment, the first type of siRNA consists of one or more siRNAs having silencing efficacy no less than a first threshold and the second type of siRNA consists of one or more siRNAs having silencing efficacy less than a second threshold.

In another embodiment, the difference in probability is described by a sum of Gaussian curves, each of the Gaussian curves representing the difference in probability of finding a G or C at a different sequence position.

In one embodiment, the first and second thresholds are both 75% at an siRNA concentration of 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM.

In still another aspect, the invention provides a method for selecting from a plurality of different siRNAs one or more siRNAs for silencing a target gene in an organism, each of the plurality of different siRNAs targeting a different target sequence in a transcript of the target gene, the method comprising, (i) for each of the plurality of different siRNAs, predicting off-target genes of the siRNA from among a plurality of genes, wherein the off-target genes are genes other than the target gene and are directly silenced by the siRNA; (ii) ranking the plurality of different siRNAs according to the number of off-target genes; and (iii) selecting one or more siRNAs for which the number of off-target genes is below a given threshold. In one embodiment, the predicting comprises (il) evaluating the sequence of each of the plurality of genes based on a predetermined siRNA sequence match pattern; and (i2) predicting a gene as an off-target gene if the gene comprise a sequence that matches the siRNA based on the sequence match pattern.

In one embodiment, each siRNA has L nucleotides in its duplex region, and the sequence match pattern is represented by a position match position-specific score matrix (pmPSSM), the position match position-specific score matrix consisting of weights of different positions in an siRNA to match transcript sequence positions in an off-target transcript {P,}, where j = 1, .... L, P ₁ is the weight of a match at position^.

In another embodiment, the step (il) comprises calculating a position match score pmScore according to equation

L pmScore = ∑ln(£, /0.25)

where E ₁ = P ₁ if position / is a match and E ₁ = (l-P,)/3 if position i is a mismatch; and the step (i2) comprises predicting the gene as an off-target gene if the position match score is greater than a given threshold. In a preferred embodiment, L is 19, and the pmPSSM is given by Table I.

In one embodiment, the plurality of genes comprises all known unique genes of the organism other than the target gene.

In still another aspect, the invention provides a library of siRNAs, comprising a plurality of siRNAs for each of a plurality of different genes of an organism, wherein each siRNA achieves at least 75%, at least 80%, or at least 90% silencing of its target gene. In one embodiment, the plurality of siRNAs consists of at least 3, at least 5, or at least 10 siRNAs. In another embodiment, the plurality of different genes consists of at least 10, at least 100, at least 500, at least 1,000, at least 10,000, or at least 30,000 different genes. In still another aspect, the invention provides a method for determining a base composition position-specific score matrix (bsPSSM) {\og{e, Jp ₁ J)) for representing base composition patterns of siRNA functional sequence motifs of L nucleotides in transcripts, wherein i = G, C, A, U(T) and/ = 1, 2, ..., L, and each siRNA functional sequence motif comprises at least a portion of the target sequence of the corresponding targeting siRNA and/or a sequence in a sequence region flanking the target sequence, the method comprising (a) identifying a plurality of N different siRNAs consisting of siRNAs having a silencing efficacy above a chosen threshold; (b) identifying a plurality of N corresponding siRNA functional sequence motifs, one for each different siRNA; (c) calculating a frequency matrix {f _υ }, where i = G, C, A, U(T); j = 1, 2, ..., L, and where f _y is the frequency of the /th nucleotide at the/th position, based on the plurality of N siRNAs functional sequence motifs according to equation

, and (d) determining the psPSSM by calculating e,, according to 0,if k ≠ i equation

. • -4 N. ^'

In one embodiment, each siRNA functional sequence motif comprises the target sequence of the corresponding targeting siRNA and one or both flanking sequences of the target sequence.

In one embodiment, each siRNA has M nucleotides in its duplex region, and each siRNA functional sequence motif consists of an siRNA target sequence of M nucleotides, a 5' flanking sequence of Z) _/ nucleotides and a 3' flanking sequence of D ₂ nucleotides.

In a specific embodiment, each siRNA has 19 nucleotides in its duplex region, and each siRNA functional sequence motif consists of an siRNA target sequence of 19 nucleotides, a 5' flanking sequence of 10 nucleotides and a 3' flanking sequence of 10

nucleotides. In another specific embodiment, each siRNA has 19 nucleotides in its duplex region, and each siRNA functional sequence motif consists of an siRNA target sequence of 19 nucleotides, a 5' flanking sequence of 50 nucleotides and a 3' flanking sequence of 50 nucleotides. In one embodiment, the plurality of N siRNAs each targets a gene whose transcript abundance is within a given range. In one embodiment, the range is at least about 5, 10, or 100 transcripts per cell. In another embodiment, the range is less than about 3-5 transcripts per cell.

In another embodiment, the silencing threshold is 50%, 75%, or 90% at an siRNA concentration of about 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM. In still another embodiment, the plurality of N siRNAs comprises 10, 50, 100, 200, or 500 different siRNAs.

In still another aspect, the invention provides a method for determining a base composition position-specific score matrix (bsPSSM) {w _υ } for representing a base composition pattern representing a plurality of different siRNA functional sequence motifs of L nucleotides, wherein / = G, C, A, U(T) andj = 1, 2, ..., L, and each siRNA functional sequence motif comprises at least a portion of the target sequence of the corresponding targeting siRNA and/or a sequence in a sequence region flanking the siRNA target sequence, the method comprising (a) initializing the bsPSSM with random weights; (b) selecting randomly a weight w _y obtained in (a); (c) changing the value of the selected weight to generate a test psPSSM comprising the selected weight having the changed value; (d) calculating a score for each of the plurality of siRNAs functional sequence motifs using the test psPSSM according to equation

L

Score = ∑ ln(w _k I p _k )

wherein the W _k and/ty are respectively weights of a nucleotide at position k in the functional sequence motif and in a random sequence; (e) calculating correlation of the score and a metric characterizing an siRNA among the plurality of siRNAs functional sequence motifs; (f) repeating steps (c)-(e) for a plurality of different values of the selected weight in a given range and retain the value that corresponds to the best correlation for the selected weight; and (g) repeating steps (b)-(f) for a chosen number of times; thereby determining the psPSSM.

The invention also provides a method for determining a base composition position- specific score matrix (bsPSSM) {w _υ } for representing a base composition pattern

representing a plurality of different siRNA functional sequence motifs of L nucleotides, wherein / = G/C, A, U(T) and 7 = 1, 2, .... L, and each siRNA functional sequence motif comprises a least a portion of the target sequence of the corresponding siRNA and/or a sequence in a sequence region flanking the siRNA target sequence, the method comprising (a) initializing the bsPSSM with random weights; (b) randomly selecting a weight w _tJ obtained in (a); (c) changing the value of the selected weight to generate a test psPSSM comprising the selected weight having the changed value; (d) calculating a score for each of the plurality of siRNA functional sequence motifs using the test psPSSM according to equation

L Score = ∑ln(w _k /p _k )

wherein the Wj ₁ andp _k are respectively weights of a nucleotide at position k in the functional sequence motif and in a random sequence; (e) calculating a correlation of the score and a metric of a characteristic of an siRNA among the plurality of siRNAs functional sequence motifs; (f) repeating steps (c)-(e) for a plurality of different values of the selected weight in a given range and retain the value that corresponds to the best correlation for the selected weight; and (g) repeating steps (b)-(f) for a chosen number of times; thereby determining the psPSSM.

In one embodiment, each siRNA functional sequence motif comprises the target sequence of the corresponding targeting siRNA and one or both flanking sequences of the target sequence.

In another embodiment, the method further comprises selecting the plurality of siRNA functional sequence motifs by a method comprising (i) identifying a plurality of siRNAs consisting of siRNAs having different values in the metric; (ii) identifying a plurality of siRNA functional sequence motifs each corresponding to an siRNA in the plurality of siRNAs.

In one embodiment, each siRNA has M nucleotides in its duplex region, and each siRNA functional sequence motif consists of an siRNA target sequence of M nucleotides, a 5' flanking sequence of Dj nucleotides and a 3' flanking sequence of Z) _? nucleotides.

In a specific embodiment, each siRNA has 19 nucleotides in its duplex region, and each siRNA functional sequence motif consists of an siRNA target sequence of 19 nucleotides, a 5' flanking sequence of 10 nucleotides and a 3' flanking sequence of 10 nucleotides. In another specific embodiment, each siRNA has 19 nucleotides in its duplex region, and each siRNA functional sequence motif consists of an siRNA target sequence

of 19 nucleotides, a 5' flanking sequence of 50 nucleotides and a 3' flanking sequence of 50 nucleotides.

In one embodiment, the metric is silencing efficacy.

In one embodiment, the plurality of N siRNAs each targets a gene whose transcript abundance is within a given range. In one embodiment, the range is at least about 5, 10, or 100 transcripts per cell. In another embodiment, the range is less than about 3-5 transcripts per cell. In another embodiment, the threshold is 50%, 75%, or 90% at an siRNA concentration of about 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 10O nM. In another embodiment, the method further comprises evaluating the psPSSM using an ROC (receiver operating characteristic) curve of the sensitivity of the psPSSM vs. the non-specificity of the psPSSM curve, the sensitivity of the PSSM being the proportion of true positives detected using the psPSSM as a fraction of total true positives, and the non-specificity of the PSSM being the proportion of false positives detected using the psPSSM as a fraction of total false positives.

In one embodiment, the plurality of siRNA functional sequence motifs consists of at least 50, at least 100, or at least 200 different siRNAs functional sequence motifs.

In still another embodiment, the method further comprises testing the psPSSM using another plurality of siRNA functional sequence motifs. The invention also provides a method for determining a position match position- specific score matrix (pmPSSM) {E,} for representing position match pattern of an siRNA of L nucleotides with its target sequence in a transcript, wherein E ₁ is a score of a match at position i, i = 1, 2, ..., L, the method comprising (a) identifying a plurality of N siRNA off-target sequences, wherein each off-target sequence is a sequence on which the siRNA exhibits silencing activity; (b) calculating a position match weight matrix {P,}, where / = 1, 2, .... L, based on the plurality of N siRNAs off-target sequences according to equation where δk(j) is 1 if k is a match, and is 0 if k is a mismatch; and (c) determining the psPSSM by calculating E ₁ such that E ₁ = P ₁ if position / is a match and E ₁ = (l-P,)/3 if position i is a mismatch.

In a preferred embodiment, L = 19. In another preferred embodiment, the position match weight matrix is given by Table I.

The invention also provides a method for evaluating the relative activity of the two strands of an siRNA in off-target gene silencing, comprising comparing position specific base composition of the sense strand of the siRNA and position specific base composition of the antisense strand of the siRNA or reverse complement strand of the sense strand of the siRNA, wherein the antisense strand is the guiding strand for targeting the intended target sequence.

In one embodiment, the comparing is carried out by a method comprising (a) determining a score for the sense strand of the siRNA, wherein the score is calculated using a position-specific score matrix; (b) determining a score for the antisense strand of the siRNA or the reverse complement strand of the sense strand of the siRNA using the position-specific score matrix; and (c) comparing the score for the sense strand and the score for the antisense strand or the reverse complement strand of the sense strand, thereby evaluating strand preference of the siRNA.

In one embodiment, the siRNA has a nucleotide sequence of L nucleotides in its duplex region, L being an integer, wherein the position-specific score matrix is {w _y }, where w _y is the weight of nucleotide i at position/, / = G, C, A, U(T), j = 1, ..., L.

In another embodiment, the siRNA has a nucleotide sequence of L nucleotides in its duplex region, L being an integer, and the position-specific score matrix is {w _y }, where w, _j is the weight of nucleotide / at position/, i = G or C, A, U(T), j = 1, ..., L. In another embodiment, the position-specific score matrix is obtained by a method comprising (a) initializing the position-specific score matrix with random weights; (b) selecting randomly a weight w _y obtained in (a); (c) changing the value of the selected weight to generate a test position-specific score matrix comprising the selected weight having the changed value; (d) calculating a score for each of a plurality of siRNAs using the test position-specific score matrix according to equation

L

Score = γ \n{W _j l P _j )

wherein the w, and /7, are respectively weights of a nucleotide at position/ in the siRNA and in a random sequence; (e) calculating correlation of the score with a metric of a characteristic of an siRNA among the plurality of siRNAs; (f) repeating steps (c)-(e) for a plurality of different values of the selected weight in a given range and retain the value that corresponds to the best correlation for the selected weight; and (g) repeating steps Qo)- (f) for a chosen number of times; thereby determining the position-specific score matrix. In one embodiment, the metric is siRNA silencing efficacy.

In one embodiment, the siRNA has 19 nucleotides in its duplex region.

In another embodiment, the siRNA has a nucleotide sequence of Z nucleotides in its duplex region, L being an integer, wherein the position-specific score matrix comprises Wk, k =1, ..., L, W _k being a difference in probability of finding nucleotide G or C at sequence position k between a first type of siRNA and a second type of siRNA, and the score for each strand is calculated according to equation

L

Score = ^ w _k .

In one embodiment, the first type of siRNA consists of one or more siRNAs having silencing efficacy no less than a first threshold and the second type of siRNA consists of one or more siRNAs having silencing efficacy less than a second threshold, and the siRNA is determined as having antisense preference if the score determined in step (a) is greater than the score determined in step (b), or as having sense preference if the score determined in step (b) is greater than the score determined in step (a).

In another embodiment, the difference in probability is described by a sum of Gaussian curves, each of the Gaussian curves representing the difference in probability of finding a G or C at a different sequence position.

In one embodiment, the first and second thresholds are both 75% at an siRNA concentration of about 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM. In still another aspect, the invention provides a computer system comprising a processor, and a memory coupled to the processor and encoding one or more programs, wherein the one or more programs cause the processor to carry out any one of the method of the invention.

In still another aspect, the invention provides a computer program product for use in conjunction with a computer having a processor and a memory connected to the processor, the computer program product comprising a computer readable storage medium having a computer program mechanism encoded thereon, wherein the computer program mechanism may be loaded into the memory of the computer and cause the computer to carry out any one of the method of the invention.

4. BRIEF DESCRIPTION OF FIGURES FIGS. IA-C show that base composition in and around an siRNA target sequence affects the silencing efficacy of the siRNA. A total of 377 siRNAs were tested by quantitative reverse transcription PCR (qRT-PCR) Taqman analysis for their ability to silence their target sequences 24hr following transfection into HeLa cells. Median target

silencing was -75%. This dataset was divided into two subsets, one having less than median and one having equal to or greater than median silencing ability (referred to as "bad" and "good" siRNAs, respectively). Shown here are the mean difference within a window of 5 (i.e., averaged over all 5 bases) in GC content (FIG. IA), A content (FIG. IB), and U content (FIG. 1C) between good and bad siRNAs at different relative positions on a target sequence.

FIGS. 2A-C show nucleotide contents of various siRNAs. (A) GC content of good and bad siRNAs; (B) A content of good and bad siRNAs; (C) U content of good and bad siRNAs. The figures show average compositions of each base. For example, 0.5 on the y- axis corresponds to an average base content of 50%.

FIG. 3 shows the performance of an actual siRNA base composition model used in the siRNA design method of the invention. siRNA efficacy data were subdivided into two pairs of training and test sets. Different PSSMs were optimized on each of the training sets and verified on the test sets. The performance of each PSSM was evaluated by its ability to distinguish good siRNAs (true positives) and bad siRNAs (false positives) as an increasing number of siRNAs were selected from a list ranked by PSSM score. Shown are Receiver Operating Characteristics (ROC) curves demonstrating the performance of two different PSSMs on their respective training and test sets (heavy black and dotted gray lines, respectively). The expected performance of the PSSMs on randomized data is shown for comparison (i.e., no improvement in selection ability, 45 line).

FIG. 4 demonstrates the predictive ability of PSSMs on an independent experimental data set. New siRNAs were designed for five genes by the standard method as described in Elbashir et al, 2001, Nature 411 :494-498, with the addition of the specificity prediction method disclosed in this application, and by the PSSM based efficacy and specificity prediction method of the invention. The top three ranked siRNAs per gene were selected for each method and purchased from Dharmacon. All six siRNAs for each of the five genes were then tested for their ability to silence their target sequences. Shown is a histogram of the number of siRNAs that silence their respective target genes by a specified amount. The Solid curve represents silencing by siRNAs designed by the present method; the dashed curve represents silencing by siRNAs designed by the standard method; and the dotted gray curve represents silencing by the data set of 377 siRNAs.

FIGS. 5A-C show mean weights of GC, A or U from the two ensembles of base composition PSSM trained and tested with siRNAs in set 1 and set 2, respectively. FIG. 5A shows mean weights for GC; FIG. 5B shows mean weights for A; and FIG. 5C shows mean weights for U. siRNAs in set 1 and set 2 are shown in Table II.

FIG. 6 shows an example of alignments of transcripts of off-target genes to the core 19mer of an siRNA oligo sequence. Off-target genes were selected from the Human 25k v2.2.1 microarray by selecting for kinetic patterns of transcript abundance consistent with direct effects of siRNA oligos. The left hand column lists transcript sequence identifiers. Alignments were generated with FASTA and edited by hand. The black boxes and grey area demonstrate the higher level of sequence similarity in the 3' half of the alignment.

FIG. 7 shows a position match position-specific scoring matrix for predicting off- target effects. The chart shows the weight associated with each position in a matrix representing the alignment between an siRNA oligo and off-target transcripts. The weight represents the probability that a match will be observed at each position / along an alignment between an siRNA oligo and an observed off-target transcript.

FIG. 8 shows optimization of the threshold score for predicting off-target effects of siRNAs. The R values result from the correlation of number of alignments scoring above the threshold with number of observed off-target effects.

FIG. 9 shows a flow chart of an exemplary embodiment of the method for selecting siRNAs for use in silencing a gene.

FIG. 10 illustrates sequence regions that can be used for distinguishing good and bad siRNAs. PSSMs were trained on chunks of sequence 10+ bases in length, from 50 bases upstream to 50 bases downstream of the siRNA 19mer, and tested on independent test sets. The performance of models trained on chunks of interest was compared with models trained on random sequences. Position 1 corresponds to the first 5' base in the duplex region of a 21 nt siRNA.

FIGS. 1 IA-B show curve models for PSSM. 1 IA: an exemplary set of curve models for PSSM. 1 IB: the performance of the models on training and test sets.

FIG. 12 illustrates an exemplary embodiment of a computer system useful for implementing the methods of the present invention.

FIG. 13 shows a comparison of the distribution of silencing efficacies of the siRNAs among the 30 siRNAs designed using the method of the invention (solid circles) and siRNAs designed using the standard method (open circles). X-axis: 1, KIF14; 2, PLK; 3, IGFlR; 4, MAPK14; 5, KIFl 1. Y-axis: RNA level. The siRNAs designed using the standard method exhibited a broad distribution of silencing abilities, while those designed with the method of the invention show more consistent silencing within each gene, as well as across genes. A narrow distribution is very important for functional genomics with siRNAs.

FIGS. 14A-B show the results of a comparison of the GC content of siRNAs and their reverse complements with the GC content of bad siRNAs. The results indicate that bad siRNAs have sense strands similar to good siRNAs, while good siRNAs have sense strands similar to bad siRNAs. RC: reverse complement of the siRNA target sequence.

FIG. 15 shows that less effective siRNAs have active sense strands. Strand bias of 61 siRNAs was predicted from expression profiles by the 3 '-biased method, and from comparison of the GC PSSM scores of the siRNAs and their reverse complements. Strand bias predictions were binned by siRNA silencing efficacy.

FIG. 16 shows that silencing efficacy relates to transcript expression level. A total of 222 siRNAs (3 siRNAs per gene for 74 genes) were tested by branched DNA (bDN A) or Taqman analysis for their ability to silence their target sequences 24hr following transfection into HeLa cells. Percent silencing (y-axis) was plotted as a function of transcript abundance (x-axis) measured as intensity on microarray. Shown is the median target silencing observed for 3 siRNAs per gene selected by the previous siRNA design algorithm. The dependence of silencing on gene expression level, as the average of intensities from 2 array types, is shown for 74 genes. TaqMan assays were used for 8 genes. b-DNA data is shown for the remaining 66 genes.

FIG. 17 shows that the silencing efficacy of an siRNA relates to its base composition. siRNAs to poorly-expressed genes were tested by bDNA analysis for their ability to silence their target sequences. Data were divided into subsets having less than 75% silencing and equal to or greater than 75% silencing (bad and good siRNAs, respectively). Shown here is the difference in GC content between good and bad siRNAs (y-axis) at each position in the siRNA sense strand (x-axis.) The dataset includes both poorly-expressed and highly-expressed genes from 570 siRNAs selected to 33 poorly- and 41 highly-expressed genes by Tuschl rules or randomized selection. The siRNA sequences are listed in Table IV. The GC profile for good siRNAs to poorly-expressed genes (gray dotted curve) shows some similar composition preferences to good siRNAs for well-expressed genes (black curve), but also some differences.

FIG. 18 shows the efficacies of newly design siRNAs. siRNAs were designed for 18 poorly-expressed genes by the standard method and by the new algorithm. Standard pipeline: selection for maximum pssm score; minimax filter for long off-target matches. Improved pipeline: selection for 1-3 G+C in sense 19mer bases 2-7, base 1 & 19 asymmetry, -300 < pssm score < +200, and blast matches less than 16, 200 bases on either side of the 19mer are not repeat or low-complexity sequences. The top three ranked siRNAs per gene were selected for each method. All six siRNAs for each of the five genes were then tested for their ability to silence their target sequences. Shown is a histogram of the number of siRNAs silencing their target genes by a specified amount. Dotted curve, silencing by siRNAs designed by the new algorithm; solid curve, silencing by siRNAs designed by the standard method. Median silencing improved from 60% (standard algorithm) to 80% (new algorithm).

FIG. 19. Design features of efficacious siRNAs. Studies of design criteria that correlate with siRNA silencing efficacy have revealed a number of features that predict efficacy. These include a base asymmetry at the two termini to direct the antisense (guide) strand into RISC, a U at position 10 for effective cleavage of the transcript, a low GC stretch encompassing the center and 3' end of the guide strand for enhanced cleavage, and the "seed" region at the 5' end of the antisense strand implicated in transcript binding. Gray lines above the duplex indicate sequence preferences, light gray lines below the duplex indicate functional attributes.

FIG. 20 shows expression vs. median silencing in 371 siRNAs. These are siRNAs from the original training set of 377 siRNAs. 6 siRNAs were not included in the analysis, as the expression level of their target gene was not available.

FIG. 21 shows that shRNAs and siRNAs have similar requirements for GC-content asymmetry. The fractional differences in GC content preference (Y axis) of 218 shRNAs (dashed line) and 377 siRNAs (solid line) are shown for each position in the shRNA sense strand (X axis). shRNAs were divided into those with greater ("good") or less ("bad") than 66% silencing, which is the median silencing for shRNAs designed by the second generation algorithm. siRNAs were divided into those with greater ("good") or less ("bad") than 75% silencing, which is the median silencing measured for siRNAs designed by a pseudorandom approach. Base composition profiles are smoothed over a window of 5 bases.

FIG. 22 shows the preferences of siRNAs and shRNAs for individual bases at each position of the core sequence. As above, shRNAs were divided into those with greater or less than 66% silencing, the median silencing for shRNAs designed using the second generation design, and siRNAs were divided into those with greater or less than 75% silencing, the median silencing of siRNAs designed in a pseudo-random manner. The fractional difference in content of each base is shown between good and bad shRNAs (dotted line) and between good and bad siRNAs (solid line) at each position in the sense strand. The strongest preference is for a G at position 1 of the sense strand.

FIG. 23 shows improved efficacies of shRNAs designed using the third generation algorithm. shRNAs were selected by a pseudorandom design (dotted line), the second generation algorithm (dashed line), or the third generation algorithm with a requirement for G at position 1 (solid line). Shown is a histogram of the number of shRNAs that silence their target genes by a specified amount. Median silencing values were ~54% for shRNAs designed using the pseudorandom algorithm, -66% for shRNAs designed using the second generation algorithm (RSTA shRNA V2), and -75% for shRNAs designed using the third generation algorithm (RSTA shRNA V3) with the additional requirement that there is a G at position 1.

FIG. 24 shows some exemplary genes used in studies of pseudorandom shRNAs in accordance with the present invention. A dozen genes were selected based on a distribution of their overall transcript GC content (expressed as a percentage in parentheses) and of the mean log expression level as determined by microarray analysis (also in parentheses).

5. DETAILED DESCRIPTION OF THE INVENTION

The present invention provides systems and methods for designing shRNA sequences for gene silencing in a cell through RNA interference (RNAi). In embodiments in accordance with the present invention, the cell can be a cell from an organism, a cell of an organism, a cell from a cell line, a primary cell, in vivo or ex vivo (e.g., in cell culture). Each of the designed shRNA "sequences (the "shRNA designs") comprises an shRNA core sequence that comprises nucleotide sequence motifs joined by a loop sequence. The shRNA design, when transcribed and post-transcriptionally processed (e.g. , by Dicer), is predicted to yield a corresponding siRNA. In most embodiments, nucleotide sequence motifs in the shRNA designs are inverted respeats that correspond to the sense and antisense sequences of the corresponding siRNA. The silencing efficacy of the corresponding siRNA is embedded in the sequence motifs of the sense or antisense strands. Accordingly, the sequence motif of one of the strands is therefore referred to as an "siRNA functional sequence motif or simply a "functional sequence motif in the present application. An siRNA with a functional sequence motif causes gene silencing of a target gene by binding to an siRNA susceptible sequence motif within the target gene. A sequence motif in a transcript that may be targeted by an siRNA for degradation of the transcript, e.g., a sequence motif that is likely to be a highly effective siRNA targeting site is also referred to as an siRNA susceptible motif. A sequence motif in a transcript of the target gene that may be less desirable for targeting by an siRNA, e.g., a sequence motif that is likely to be a less effective siRNA targeting site is also referred to as an siRNA resistant motif.

In embodiments in accordance with the present invention, transcription efficiency of an shRNA core sequence may be manipulated by adding flanking sequences on one or both sides of the shRNA core sequence.

The present invention provides systems and methods for shRNA design by separately optimizing the shRNA core sequence for silencing efficacy and the transcription sequence motif for transcription efficiency within a specific shRNA design.

Systems and methods for predicting silencing efficacy of a core sequence have been disclosed previously, for example, in PCT publication WO 2005/042708 A2 (hereinafter "WO 05/042708"), published on May 12, 2005. As disclosed above, an shRNA core sequence comprises the sense and antisense sequences (i.e., the functional sequence motifs) of a corresponding siRNA. Sections 5.1.1 through 5.1.4 herein disclose methods and systems for designing a shRNA core sequence based on a predicted or observed silencing efficacy of a functional sequence motif that correspond to the core sequence. An shRNA core sequence further comprises a loop sequence. Section 5.1.5 discloses methods and system for designing sense strand and antisense strand sequences (e.g., the inverted repeats) in the shRNA core sequences using the functional sequence motifs identified in sections 5.1.1 through 5.1.4. Section 5.1.5.2 further discloses exemplary methods for enhancing transcription efficiency of an shRNA sequence design and provides examples of loop sequence.

The invention provides a method of designing shRNA core sequences, in particular the sequences in the sense and antisense strands, by identifying the siRNA susceptible motifs for a target gene. Gene silencing occurs when a strand of an siRNA binds through complementarity to an siRNA susceptible sequence motif in a target gene. In most embodiments, defining the sequence of an siRNA susceptible sequence motif sufficiently defines the functional sequence motif of an siRNA, which in turn sufficiently defines the sense strand and antisense strand sequences of an shRNA core sequence.

Specifically, the present invention provides a method for identifying siRNA susceptible motifs in a transcript using a position-specific score matrix approach, for example, as disclosed in PCT publication No. WO 05/042708 ("hereinafter WO 05/042708"). WO 05/042708 provides a method for identifying off-target genes of an siRNA and for predicting specificity of an siRNA using a position-specific score matrix approach. WO 05/042708 further provides a method for designing siRNAs with high silencing efficacy and specificity as well as a library of siRNAs comprising siRNAs with high silencing efficacy and specificity. Based on WO 05/042708, the present application invention further provides a method for designing shRNAs based on not only the silencing efficiencies of siRNAs produced from the shRNAs but also the transcription efficiencies of the shRNAs.

In this application, an shRNA core sequence is often said to target a gene through an siRNA that is produced from the shRNA core sequence. It will be understood that when such a statement is made, it means that the shRNA that corresponds to the shRNA is

designed to target and cause degradation of a transcript of the gene. Such a gene is also referred to as a target gene of the shRNA, and the sequence in the transcript that is acted upon by the siRNA is referred to as the target sequence. For example, a 19-nucleotide sequence in a transcript which is identical to the sequence of the 19-nucleotide sequence in the sense strand of the duplex region of the corresponding siRNA is the target sequence of the shRNA. The antisense strand of the corresponding siRNA, a strand that acts upon the target sequence, is also referred to as the guiding strand. In the above example, the antisense strand of the 19-nucleotide duplex region of the corresponding siRNA is the guiding strand. In this application, features of an siRNA are often referred to with reference to its sequence, e.g., positional base composition. It will be understood that, unless specifically pointed out otherwise, such a reference is made to the sequence of the sense strand of the corresponding siRNA. It will be understood that the target sequence of an shRNA. In this application, a nucleotide or a sequence of nucleotides in an siRNA is often described with reference to the 5 ' or 3' end of the siRNA. It will be understood that when such a description is employed, it refers to the 5' or 3' end of the sense strand of the siRNA. It will also be understood that, when a reference to the 3' end of the siRNA is made, it refers to the 3' duplex region of the siRNA, i.e., the two nucleotides of the 3' overhang are not included in the numbering of the nucleotides.

In this disclosure, design of shRNA is discussed in reference to silencing a sense strand target, i.e., transcript target sequence binding to the antisense strand of the shRNA. It will be understood by one skilled in the art that the methods of the invention are also applicable to the design of siRNA for silencing an antisense target (see, e.g., Martinez et al, 2002, Cell 1 10:563-574).

5.1. METHODS OF IDENTIFYING SEQUENCE MOTIFS IN A GENE FOR

TARGETING BY A SMALL INTERFERING RNA

As noted above, an siRNA functional sequence motif (or simply functional sequence motif) binds to an siRNA susceptible sequence motif in a target gene prior to RNA silencing. In some embodiments, the silencing efficacy of a functional sequence motif to an siRNA susceptible sequence motif is determined by experimental methods (e.g., methods for measuring silencing efficacies or binding coefficients) or alternatively, by methods of predicting silencing efficacies (e.g., computer algorithms based on previous experimental data, as disclosed in WO 05/042708). This invention provides a method for identifying siRNA functional sequence motifs and subsequently builds a profile of siRNA

functional sequence motifs using, e.g., a library of siRNAs for which silencing efficacies have been determined. In one embodiment, the sequence region of interest is scanned to identify sequences that match the profile of the functional sequence motif.

5.1.1. SEQUENCE PROFILE AND TARGET SILENCING EFFICACY

In a preferred embodiment, the profile of a functional sequence motif is represented using a position-specific score matrix (PSSM). A general discussion of PSSM can be found in, e.g., "Biological Sequence Analysis" by R. Durbin, S. Eddy, A. Krogh, and G. Mitchison, Cambridge Univ. Press, 1998; and Henikoff et al. , 1994, J MoI Biol. 243:574-8; each of which is hereby incorporated by reference herein in its entirety. A PSSM is a sequence motif descriptor which captures the characteristics of a functional sequence motif. In this disclosure, a PSSM is used to describe sequence motifs of the invention, e.g., an siRNA susceptible or siRNA resistant motif. A PSSM of an siRNA susceptible motif is also referred to as a susceptible PSSM. Accordingly, a PSSM of an siRNA resistant motif is also referred to as a resistant PSSM. A skilled person in the art will know that a position-specific score matrix is also termed a position specific scoring matrix, a position weight matrix (PWM), or a Profile, and there terms are used interchangeably herein.

In the present invention, a functional sequence motif comprises one or more sequences in an siRNA target sequence. For example, the one or more sequences in an siRNA target sequence may be a sequence at 5' end of the target sequence, a sequence at 3' end of the target sequence. The one or more sequences in an siRNA target sequence may also be two stretches of sequences, one at 5' end of the target sequence and one at 3' end of the target sequence. A functional sequence motif can also comprise one or more sequences in a sequence region that flanks the siRNA target sequence. Such one or more sequences can be directly adjacent to the siRNA target sequence. Such one or more sequences can also be separated from the siRNA target sequence by an intervening sequence. FIG. 10 illustrates some examples of functional sequence motifs.

In one embodiment, a functional sequence motif, e.g., an siRNA susceptible or siRNA resistant sequence motif, comprises at least a portion of a sequence targeted by an siRNA. In one embodiment, the functional sequence motif comprises a contiguous sequence stretch of at least 7 nucleotides of the target sequence. In a preferred embodiment, the contiguous sequence stretch is in a 3' region of the target sequence, e.g., beginning within 3 bases at the 3' end. In another embodiment, the contiguous sequence

stretch is in a 5' region of the target sequence. In another embodiment, the functional sequence motif comprises a contiguous sequence stretch of at least 3, 4, 5, 6, or 7 nucleotides in a 3' region of the target sequence and comprises a contiguous sequence stretch of at least 3, 4, 5, 6, or 7 nucleotides in a 5' region of the target sequence. In still another embodiment, the functional sequence motif comprises a contiguous sequence stretch of at least 11 nucleotides in a central region of the target sequence. Sequence motifs comprise less than the full length of the target sequence can be used for evaluating siRNA target transcripts that exhibit only partial sequence identify to an siRNA (International Application No. PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is incorporated herein by reference in its entirety). In a preferred embodiment, the functional sequence motif comprises the full length siRNA target sequence.

The functional sequence motif may also comprise a flanking sequence. It has been discovered that the sequence of such flanking region plays a role in determining the efficacy of silencing. In one embodiment, a functional sequence motif, e.g. , an siRNA susceptible or siRNA resistant sequence motif, comprises at least a portion of a sequence targeted by an siRNA and one or more sequences in one or both flanking regions. Thus, a sequence motif can include an M nucleotides siRNA target sequence, a flanking sequence of Di nucleotides at one side of the siRNA target sequence and a flanking sequence of D^ nucleotides at the other side of the siRNA target sequence where M, Di and D ₂ are appropriate integers. In one embodiment, Di = D ₂ = D. In one embodiment, M= 19. In some preferred embodiments, Di, D ₂ , or D is at least 5, 10, 20, 30, 50 nucleotides in length. In a specific embodiment, an siRNA susceptible or siRNA resistant sequence motif consists of an siRNA target sequence of 19 nucleotides and a flanking sequence of 10 nucleotides at either side of the siRNA target sequence. In another specific embodiment, an siRNA susceptible or siRNA resistant sequence motif consists of a 19 nucleotides siRNA target sequence and a 50 nucleotides flanking sequence at either side of the siRNA target sequence.

In another embodiment, a sequence motif can include an M nucleotides siRNA target sequence, and one or more of the following: a contiguous sequence stretch of Di nucleotides flanking the 5' end of the target sequence, a contiguous sequence stretch of D ₂ nucleotides flanking the 3' end of the target sequence, a contiguous sequence stretch of D ₃ nucleotides which starts about 35 nucleotides upstream of the 5' end of the target sequence, a contiguous sequence stretch of D ₄ nucleotides which starts about 25

nucleotides downstream of the 3' end of the target sequence, and a contiguous sequence stretch of D ₅ nucleotides which starts about 60 nucleotides downstream of the 3' end of the target sequence, where Dj, D ₂ , Di, D ₄ , and Ds are appropriate integers. In one embodiment, D _/ = D ₂ = D. In some preferred embodiments, each of Di, D 2, D$ _t D 4, and D ₅ is at least 5, 10, or 20 nucleotides in length. The length of the functional sequence motif is L = M+ Di + D ₂ + D ₃ + D ₄ + D ₅ . In a specific embodiment, the sequence motif include 19 nucleotides siRNA target sequence, a contiguous sequence stretch of about 10 nucleotides flanking the 5' end of the target sequence, a contiguous sequence stretch of about 10 nucleotides flanking the 3' end of the target sequence, a contiguous sequence stretch of about 10 nucleotides which starts about 35 nucleotides upstream of the 5' end of the target sequence, a contiguous sequence stretch of about 10 nucleotides which starts about 25 nucleotides downstream of the 3' end of the target sequence, and a contiguous sequence stretch of about 10 nucleotides which starts about 60 nucleotides downstream of the 3' end of the target sequence (see FIG. 10). In other embodiments, a functional sequence motif, e.g., an siRNA susceptible or siRNA resistant sequence motif, comprises one or more sequences in one or both flanking regions of an siRNA target sequence but does not comprise any siRNA target sequence. In one embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides flanking the 5' end of the target sequence. In another embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides flanking the 3' end of the target sequence. In a preferred embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides flanking the 5' end of the target sequence and a contiguous sequence stretch of about 10 nucleotides flanking the 3' end of the target sequence. In one embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides which starts about 35 nucleotides upstream of the 5' end of the target sequence. In another embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides which starts about 25 nucleotides downstream of the 3' end of the target sequence. In still another embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides which starts about 60 nucleotides downstream of the 3' end of the target sequence. In a preferred embodiment, the functional motif comprises a contiguous sequence stretch of about 10 nucleotides flanking the 5' end of the target sequence, a contiguous sequence stretch of about 10 nucleotides flanking the 3' end of the target sequence, a contiguous sequence stretch of about 10 nucleotides which starts about 35 nucleotides upstream of the

5' end of the target sequence, a contiguous sequence stretch of about 10 nucleotides which starts about 25 nucleotides downstream of the 3' end of the target sequence, and a contiguous sequence stretch of about 10 nucleotides which starts about 60 nucleotides downstream of the 3' end of the target sequence. Thus, a functional sequence motif can include a contiguous sequence stretch of Dy nucleotides flanking the 5' end of the target sequence, a contiguous sequence stretch of D ₂ nucleotides flanking the 3' end of the target sequence, a contiguous sequence stretch of D ₃ nucleotides which starts about 35 nucleotides upstream of the 5' end of the target sequence, a contiguous sequence stretch of D ₄ nucleotides which starts about 25 nucleotides downstream of the 3' end of the target sequence, and a contiguous sequence stretch of D ₅ nucleotides which starts about 60 nucleotides downstream of the 3' end of the target sequence, where Di, D ₂ , D ₃ D ₄ , and D ₅ are appropriate integers. In some preferred embodiments, each of Di, D ₂ , D _3< D ₄ , and Ds is at least 5, 10, or 20 nucleotides in length. The length of the functional functional motif is L = Di + D ₂ + D ₃ + D ₄ + D ₅ . In one embodiment, the characteristics of a functional sequence motif are characterized using the frequency of each of G, C, A, U(or T) observed at each position along the functional sequence motif. In the disclosure, U(or T), or sometimes simply U(T), is used to indicate nucleotide U or T. The set of frequencies forms a frequency matrix, in which each element indicates the number of times that a given nucleotide has been observed at a given position. A frequency matrix representing a sequence motif of length L is a 4 ^• L matrix {f _tJ }, where / = G, C, A, U(T); j = 1, 2, ..., L; where f _y is the frequency of the ith nucleotide at the 7th position. A frequency matrix of a functional sequence motif can be derived or built from a set of iVsiRNA target sequences that exhibit a desired quality, e.g., a chosen level of susceptibility or resistance to siRNA silencing:

where 4, CT) (2)

In embodiments in which a functional sequence motif consists of M nucleotides siRNA target sequence, a flanking sequence of Z) _/ nucleotides at one side of the siRNA target sequence and a flanking sequence of D ₂ nucleotides at the other side of the siRNA target sequence, L - M+ Di + D ₂ . In embodiments in which the functional sequence motif consists of M nucleotides siRNA target sequence, a contiguous sequence stretch of D _/

nucleotides flanking the 5' end of the target sequence, a contiguous sequence stretch of D2 nucleotides flanking the 3' end of the target sequence, a contiguous sequence stretch of D ₃ nucleotides which starts about 35 nucleotides upstream of the 5' end of the target sequence, a contiguous sequence stretch of D ₄ nucleotides which starts about 25 nucleotides downstream of the 3' end of the target sequence, and a contiguous sequence stretch of D 5 nucleotides which starts about 60 nucleotides downstream of the 3' end of the target sequence, L = D _/ + D ₂ + D ₃ + D ₄ + D ₅ .

In another embodiment, the characteristics of a functional sequence motif are characterized using a set of weights, one for each nucleotide occurring at a position in the motif. In such an embodiment, a weight matrix {e _y }, where i = G, C, A, U(T); j = 1, 2, ..., L, can be used for representing a functional sequence motif of length L, where e _XJ is the weight of finding the /th nucleotide at the/th position. In one embodiment, the weight e _l} is the probability of finding the /th nucleotide at the 7th position in the functional sequence motif. When a probability is used for the weight, the matrix is also called a probability matrix. A probability matrix of a sequence motif can be derived from a frequency matrix according to equation

_ λ ^J V e _H = — (3)

N

In a preferred embodiment, a position-specific score matrix is used to characterize a functional sequence motif. The PSSM can be constructed using log likelihood values \og(e _υ /p, _j ), where e _υ is the weight of finding nucleotide i at position/, and p _y is the weight of finding nucleotide i at position/ in a random sequence. In some embodiments, the probability of finding the /th nucleotide at the/th position in the functional sequence motif is used as e _y , the probability of finding nucleotide / at position/ in a random sequence is used asp, _j . The weight or probability p _υ is an "a priori" weight or probability. In some embodiments, p, _} = 0.25 for each possible nucleotide / e {G, C, A, U(T)) at each position /. Thus, for a given sequence of length L, the sum of log likelihood ratios at all positions can be used as a score for evaluating if the given sequence is more or less likely to match the functional sequence motif than to match a random sequence:

Score = X ln(e, I _P] ) (4)

wherein e _} and/?, are respectively weights of a nucleotide at position/ in the functional sequence motif and in a random sequence. For example, if such a score is zero, the sequence has the same probability to match the functional sequence motif as to that to

match a random sequence. A sequence is more likely to match the functional sequence motif if the ratio is greater than zero.

In another embodiment, when two or more different nucleotides are not to be distinguished, a PSSM with a reduced dimension can be used. For example, if the relative base compositions of G and C in a sequence motif are not to be distinguished, a PSSM can be a 3 L matrix {\og(E, _j lp, _j )}, where i = G/C, A, U(T); j = 1, 2, ..., L; where E _y is the weight, e.g., probability, of finding nucleotide / at position^ ^' , andp, _j is the weight, e.g., probability, of finding nucleotide i at position y in a random sequence. Thus, in such cases, a PSSM has 3 sets of weights: GC-specific, A-specific and U-specific, e.g., if the base at a position is a G or a C, the natural logarithm of the ratio of the GC weight and the unbiased probability of finding a G or C at that position is used as the GC-specific weight for the position; and the natural logarithms of the position-specific A and T weights divided by the unbiased probability of respective base are used as the A- and T-specific weights for the position, respectively. The log likelihood ratio score is represented by Eq. (5):

Score = £ln(E ₇ / _λ ) (5)

7=1 where E ₁ is the weight assigned to a base — A, U or G/C — at position j, and/?, = 0.25 for A or U and 0.5 for G/C.

In still another embodiment, when the relative base compositions of G and C in a sequence motif are not to be distinguished and the relative base compositions of A and T in the sequence motif are also not to be distinguished, a PSSM can be a 1 ^• L matrix {\og{E _υ lp, _j )}, where / = G/C; j = 1, 2, ..., L; where E _y is the weight, e.g., probability, of finding nucleotide / at position j, andp _υ is the weight, e.g., probability, of finding nucleotide / at position y in a random sequence. Thus, in such cases, a PSSM has 1 set of GC-specific weights: if the base at a position is a G or a C, the natural logarithm of the ratio of the GC weight and the unbiased probability of finding a G or C at that position is used as the GC-specific weight for the position. The log likelihood ratio score is represented by Eq. (5), except that E, is the weight assigned to a base — G/C — at position j, and p _} = 0.50. Silencing efficacies are used to validate methods for RNA interference, i.e., to assess the effect of siRNA on its target gene. In some embodiments, the silencing efficacy of a functional sequence motif in an siRNA is determined by percentage reduction of the expression level of its target gene in the presence of a given concentration of the

functional sequence motif, in comparison to the expression level of the same target gene in the absence of the functional sequence motif. The reduction of target gene expression level may be quantified, for example, as a reduction in transcription level, translation level or protein activity level of the target gene or a product thereof. Any suitable methods may be applied to quantify the reduction in target expression level in order to determine a silencing efficacy for the functional sequence motif. Changes in transcription level (e.g., endogenous mRNA level) may be quantified by, for example, Northern blot or quantitative Reverse Transcription Polymerase Chain Reaction (qRT-PCR). In some embodiments, the silencing efficacy of an siRNA functional sequence motif is determined by a branched DNA assay. In this type of assay, labeled branched DNA is mixed with a sample to be tested. If the branched DNA binds to viral RNA, then the luminescent compound will react, and the result can be measured via luminometer. See, for example, Murphy et ah, 1999, "Reproducibility and performance of the second-generation branched-DNA assay in routine quantification of human immunodeficiency virus type 1 RNA in plasma," J CHn Microbiol. 37:812-814.

Changes in translation level (e.g., endogenous protein level or exogenous fusion protein level) may be quantified by, for example, Western blot or other antibody-related assays, immunofluorescence (IF), enzyme-linked immunosorbent assay (ELISA), or fluorescence-activated cell sorting (FACS). Furthermore, changes of translation level or protein expression level may be quantified by analyzing the activity of the protein product of the target gene, by using, for example, any enzymatic assay that detects changes in the activities of the protein product. Exemplary methods for determining silencing efficacies can be found, for example, in Sandy et al, 2005, "Mammalian RNAi: a practical guide," BioTechniques 39:215-224, which is incorporated by reference herein in its entirety. Silencing efficacies are often expressed as a percentage in reduction of expression levels at a given siRNA concentration. For example, for a given functional sequence motif, a silencing efficacy of 90% at 100 nM siRNA means that expression of the target gene of the functional sequence motif is reduced by 90% when the siRNAs comprising the functional sequence motif are present at a concentration of 100 nM. It is to be understood that any method for determining siRNA concentration may be used to quantify the concentration of the siRNA functional sequence motif corresponding to a particular shRNA design, for example, those as disclosed in WO 05/042708 and in Section 5.5 of the instant application. It is to be understood that the concentration of the corresponding siRNA functional sequence motif of a particular

shRNA design will be affected by the transcription efficiency of the shRNA design. A more detailed description of transcription efficiency is found in Section 5.1.5.3. hi some embodiments, the intracellular concentration of the siRNA functional sequence motif corresponding to a particular shRNA design may be quantified experimentally by chemical probes or tags that are associated with the transcription system (e.g., expression vector for transcribing the shRNA design). For example, stably transfected cells with an Enhanced Green Fluorescence Protein (EGFP) expression vector for live cell visualization and imaging may be used. Understanding the silencing efficiency of EGFP shRNA in soft agar will help to identify the optimal shRNA concentration necessary for suppression of genes with varying levels of expression. See, United States Patent Publication No. 20060166247, United States Patent Publication No. 20070149470 and in Wang et al., 2005, "Small hairpin RNAs efficiently inhibit hepatitis C IRES-mediated gene expression in human tissue culture cells and a mouse model," MoI Ther. 12:562-568, each of which is hereby incorporated by reference herein in its entirety. In other embodiments, the concentration of an shRNA may be determined based on a predicted concentration of the corresponding siRNA functional sequence motif determined using any method known in the art, for example, those as disclosed in United States Patent Publication No. 20060166247, United States Patent Publication No. 20070149470 and in Wang et al., 2005, "Small hairpin RNAs efficiently inhibit hepatitis C IRES-mediated gene expression in human tissue culture cells and a mouse model," MoI Ther. 12:562-568.

5.1.2. METHODS OF DETERMINING A PROFILE

Methods of determining a PSSM of a functional sequence motif based on a plurality of siRNAs for which some quantity or quantities characterizing the siRNAs have been determined. For example, a plurality of siRNAs whose silencing efficacy has been determined can be used for determination of a PSSM of an siRNA susceptible or siRNA resistant sequence motif as disclosed WO 05/042708. In the current disclosure, for simplicity reasons, silencing efficacy is often used as a measure for classifying siRNAs and accordingly shRNAs with the siRNA susceptible or siRNA resistant sequence motifs embedded in the sequences of their inverted repeats. The silencing efficacies of the shRNA designs directly correlate with the sequence motif of their corresponding siRNAs. Therefore, for all intents and purposes, discussion of silencing efficacies of the corresponding siRNAs should also apply to shRNA design.

In some embodiments, silencing efficacy of an siRNA is measured in the absence of other siRNAs designed to silence the target gene. It will be apparent to one skilled person in the art that the methods of the invention are equally applicable in cases where siRNAs are classified based on another measure. Such a plurality of siRNAs is also referred to as a library of siRNAs. In cases where the functional sequence motif of interest comprises one or more sequences in one or both flanking regions, a plurality of siRNA functional motifs, i.e., a sequence comprising the siRNA target sequence and the sequences in the flanking region(s) in a transcript, can be used to determine the PSSM of the functional motif. In a preferred embodiment, the siRNA functional sequence motif consists of an siRNA target sequence of 19 nucleotides and a flanking sequence of 10 nucleotides at either side of the siRNA target sequence. For simplicity reasons, in this disclosure, unless specified, the term "a library of siRNAs" is often used to refer to both a library of siRNAs and a library of siRNA functional sequence motifs. It will be understood that in the latter cases, when the efficacy of an siRNA is referred to, it refers to the efficacy of the siRNA that targets the motif. Preferably, the plurality of siRNAs or siRNA target motifs comprises at least 10, 50, 100, 200, 500, 1000, or 10,000 different siRNAs or siRNA target motifs.

Each different siRNA in the plurality or library of siRNAs or siRNA functional sequence motifs can have a different level of efficacy. In one embodiment, the plurality or library of siRNAs consists of siRNAs having a chosen level of efficacy. In another embodiment, the plurality or library of siRNAs comprises siRNAs having different levels of efficacy. In such an embodiment, siRNAs may be grouped into subsets, each consisting of siRNAs that have a chosen level of efficacy.

In one embodiment, a PSSM of an siRNA functional motif is determining using a plurality of siRNAs having a given efficacy. In one embodiment, a plurality of N siRNAs consisting of siRNAs having a silencing efficacy above a chosen threshold is used to determine a PSSM of an siRNA susceptible motif. The PSSM is determined based on the frequency of a nucleotide appeared at a position (see Section 5.1.1). The chosen threshold can be 50%, 75%, 80%, or 90%. In another embodiment, a plurality of N siRNAs consisting of siRNAs having a silencing efficacy below a chosen threshold is used to determine a PSSM of an siRNA susceptible motif. The chosen threshold can be 5%, 10%, 20%, 50%, 75% or 90%. In a preferred embodiment, the PSSM has a reduced dimension with a weight for G/C.

In preferred embodiments, a PSSM of an siRNA susceptible or siRNA resistant motif is derived or built using a classifier approach with a set of N sequences. In such embodiments, a library of siRNAs comprising siRNAs with different levels of silencing efficacies is used. In one embodiment, siRNAs in the library may be randomly grouped into subsets, each consisting of siRNAs that have different levels of silencing efficacy, one subset is used as a training set for determining a PSSM and the other is used as a testing set for validating the PSSM. Different criteria can be used to divide the existing siRNA library into training and test sets. For an siRNA library in which a majority of siRNA oligos are designed with the standard method, which requires an AA dimer immediately before the 19mer oligo sequence, several partitions were used and more than one trained PSSMs (rather than single PSSMs) were combined to assign scores to the test oligos. An exemplary siRNA library and divisions of the library into training and test sets are shown in Table II.

In a preferred embodiment, the sequence motif consists of 39 bases in the transcript sequence, beginning 10 bases upstream of the 19mer siRNA target sequence and ending 10 bases downstream of the 19mer. The PSSM characterizing such a sequence motif is described in Section 5.1.1.

In a preferred embodiment, the PSSM is determined by an iterative process. A PSSM is initialized with random weights {e _y } or {E,y} within a given search range for all bases at all positions. In another preferred embodiment, PSSM is initialized to the smoothed mean base composition difference between good and bad siRNAs in the training set. As an example, a PSSM describing a 39 nucleotide sequence motif can have 117 elements. In another embodiment, the weights are optimized by comparing the correlation of scores generated to a quantity of interest, e.g., silencing efficacy, and selecting the PSSM whose score best correspond to that quantity. Improvement in PSSM performance is scored by comparing correlation values before and after a change in weights at any one position. In one embodiment, there is no minimum requirement for a change in correlation. Aggregate improvement is calculated as the difference between the final correlation and the initial correlation, hi one embodiment, for a PSSM characterizing a 39mer sequence motif, the aggregate improvement threshold after 117 cycles for termination of optimization is a difference of 0.01.

In one embodiment, the weights are optimized to reflect base composition differences between good siRNAs, i.e., siRNAs having at least median silencing efficacy, and bad siRNAs, i.e., siRNAs having below median silencing efficacy, in the range of

allowed values for weights. If the PSSM is initialized with a frequency matrix, the range of allowed values corresponds to the frequency matrix elements +/- 0.05. If an unbiased search is used, the ranges of the allowed values for weights are 0.45-.55 for G/C and 0.2- 0.3 for A or U. In one embodiment, weights are allowed to vary from initial values by +/- 0.05. If an unbiased search is used, the PSSM weights can be set to random initial values within the unbiased search range described above.

In one embodiment, the PSSM is determined by a random hill-climbing mutation optimization procedure. In each step of the process, one base at one position is randomly selected for optimization. For example, for a PSSM describing a 39 nucleotide sequence motif, the 39 bases become a vector of 117 weights: 39 G/C weights, 39 A weights and 39 U weights. One of these 117 weights is selected for optimization in each step, and is run through all values in the search range at that step. For each value in the search range, scores for a training set of siRNAs are calculated. The correlation of these scores with the silencing efficacy of the siRNAs is then calculated. The weight for the position which generates the best correlation between the scores and silencing efficacy is retained as the new weight at that position.

In one embodiment, the metric used to measure the effectiveness of the training and testing is the aggregate false detection rate (FDR) based on the ROC curve, and is computed as the average of the FDR scores of the top 33% oligos sorted by the scores given by the trained PSSM. In computing the FDR scores, those oligos with silencing levels less than the median are considered false, and those with silencing level higher than the median level are considered true. The "false detection rate" is the number of false positives selected divided by the total number of true positives, measured at each ranked position in a list. The false detection rate can be a function of the fraction of all siRNAs selected. In one embodiment, the area under the curve at 33% of the list selected as a single number representing performance. In one embodiment, all at-least-median siRNAs are called as "positives" and all worse-than-median siRNAs are called "negatives." Thus, half the data are positives and the other half are "false positives." In an ideal ranking, the area under the curve at 33% or even at 50% of the list selected should be 0. In contrast, a random ranking would cause equal numbers of true positives and false positives to be selected. This corresponds to an area under the curve of 0.17 at 33% of the list selected, or 0.25 at 50% of the list selected.

Correlations between percet of silencing (e.g., percent reduction of expression level) and PSSM score are calculated according to methods known in the art (see, e.g.,

Applied Multivariate Statistical Analysis, 4th ed., R.A. Johnson & E. W. Wichern, Prentice-hall, 1998).

The process is continued until the aggregate improvement over a plurality of iterations fell below a threshold. In a preferred embodiment, a plurality of PSSMs is obtained for a functional sequence motif using an siRNA training set. In this disclosure, a plurality of PSSMs is also referred to as an "ensemble" of PSSMs. Each round of optimization may stop at a local optimum distinct from the global optimum. The particular local optimum reached is dependent on the history of random positions selected for optimization. A higher improvement threshold may not bring a PSSM optimized to a local optimum closer to the global optimum. Thus it is more effective to run multiple optimizations than one long optimization. Additional runs (e.g., up to 200) were found to enhance performance. Running more than 200 optimizations was not seen to provide further enhancements in performance. Empirically, scoring siRNAs via the average of multiple runs is less effective than scoring candidate siRNAs on the PSSMs generated by each run and then summing the scores. Thus, in one embodiment, the PSSMs are used individually or summed to generate a composite score for each sequence match. The plurality of matrices can be tested individually or as a composite on an independent set of siRNA target motifs with known silencing efficacy to evaluate the utility for identifying sequence motifs and in siRNA design. In a preferred embodiment, the plurality of PSSM consists of at least 2, 10, 50, 100, 200, or 500 PSSMs.

In a preferred embodiment, one or more different siRNA training sets are used to obtain one or more ensemble of PSSMs. These different ensembles of PSSMs may be used together in determining the predicted silencing efficacy of a sequence motif. Sequence weighting methods have been used in the art to reduce redundancy and emphasize diversity in multiple sequence alignment and searching applications. Each of these methods is based on a notion of distance between a sequence and an ancestral or generalized sequence. Here a different approach is presented, in which base weights on the diversity observed at each position in the alignment and the correlation between the base composition and the observed efficacy of the siRNAs, rather than on a sequence distance measure.

In still another embodiment, PSSMs are generated by a method which hypothesized dependency of the base composition of any one position on its neighboring positions, referred to as "curve models."

In one embodiment, curve models are generated as a sum of normal curves (i.e. , Gaussian). It will be apparent to one skilled person in the art that other suitable curve functions, e.g. , polynomials, can also be used. Each curve represents the probability of finding a particular base in a particular region. The value at each position in the summed normal curves is the weight given to that position for the base represented by the curve. The weights for each base present at each position in each siRNA and its flanking sequences are then summed to generate an siRNA's score, i.e., the score is σ Wj. The score calculation can also be described as the dot product of the base content in the sequence with the weights in the curve model. As such, it is one way of representing the correlation of the sequence of interest with the model.

Curve models can be initialized to correspond to the major peaks and valleys present in the smoothed base composition difference between good and bad siRNAs, e.g., as described in FIGS. IA-C and 5A-C. In one embodiment, curve models for G/C, A and U are obtained. In one embodiment, the initial model can be set up for the 3 -peak G/C curve model as follows: Peak 1 mean: 1.5 standard deviation: 2 amplitude: 0.0455 Peak 1 mean, standard deviation and amplitude are set to correspond to the peak in the mean difference in GC content between good and bad siRNAs occurring within bases -2 - 5 of the siRNA target site in Set 1 training and test sets. Peak 2 mean: 11 standard deviation: 0.5 amplitude: 0.0337

Peak 2 mean, standard deviation and amplitude are set to correspond to the peak in the mean difference in GC content between good and bad siRNAs occurring within bases 10- 12 of the siRNA target site in Set 1 training and test sets. Peak 3 mean: 18.5 standard deviation: 4 amplitude: -0.0548

Peak 3 mean, standard deviation and amplitude are set to correspond to the peak in the mean difference in GC content between good and bad siRNAs occurring within bases 12- 25 of the siRNA target site in Set 1 training and test sets.

Peak height (amplitude), center position in the sequence (mean) and width (standard deviation) of a peak in a curve model can be adjusted. Curve models are optimized by adjusting the amplitude, mean and standard deviation of each peak over a preset grid of values. In one embodiment, curve models are optimized on several training sets and tested on several test sets, e.g., training sets and test sets as described in Table II. Each base - G/C, A and U (or T) - is optimized separately, and then combinations of optimized models are screened for best performance.

Preferably, optimization criteria for curve models are: (1) the fraction of good oligos in the top 10%, 15%, 20% and 33% of the scores, (2) the false detection rate at 33% and 50% of the siRNAs selected, and (3) the correlation coefficient of siRNA silencing vs. siRNA scores used as a tiebreaker. When the model is trained, a grid of possible values for amplitude, mean and standard deviation of each peak is explored. The models with the top value or within the top range of values for any of the above criteria were selected and examined further.

In a preferred embodiment, G/C models are optimized with 3 or 4 peaks, A models are optimized with 3 peaks, and U models are optimized with 5 peaks. Exemplary ranges of parameters optimized for curve models are shown in Example 3, infra.

Preferably, the performance of the obtained PSSM is evaluated. In one embodiment, the PSSM is evaluated using an ROC (receiver operating characteristic) curve. An ROC curve is a plot of the sensitivity of a diagnostic test as a function of non- specificity. An ROC curve indicates the intrinsic properties of a test's diagnostic performance and can be used to compare relative merits of competing procedures. In one embodiment, the sensitivity of a PSSM is calculated as the proportion of true positives detected as a fraction of total true positives, whereas the non-specificity of the PSSM is calculated as the proportion of false positives detected as a fraction of total false positives (see, e.g., Campbell, 1994, Statistics in Medicine 13:499-508; Metz, 1986, Investigative Radiology 21 :720-733; Gribskov et al. , 1996, Computers Chem. 20:25-33). FIG. 3 shows ROC curves of the two PSSMs selected for the current best practice of the invention.

In another embodiment, the performance of a PSSM is evaluated by comparing a plurality of sequence motifs identified using the PSSM with a plurality of reference sequence motifs. The PSSM is used to obtain the plurality of sequence motifs by, e.g.,

scanning one or more transcripts and identifying sequence motifs that match the PSSM, e.g., with a score above a threshold. Preferably, the plurality comprises at least 3, 5, 10, 20 or 50 different sequence motifs. The reference sequence motifs can be from any suitable source. In one embodiment, a plurality of reference sequence motifs is obtained using a standard method (e.g., Elbashir et al., 2001, Nature. 411 :494-8). The two pluralities are then compared using any standard method known in the art to determine if they are identical.

In a preferred embodiment, the two pluralities are compared using a Wilcoxon rank sum test. A Wilcoxon rank sum test tests if two pluralities of measurements are identical (see, e.g., Snedecor and Cochran, Statistical Methods, Eighth Edition, 1989, Iowa State University Press, pp. 142-144; McClave and Sincich, 2002, Statistics, Ninth Edition, Prentice Hall, Chapter 14). The Wilcoxon rank sum test can be considered a non- parametric equivalent of the unpaired t-test. It is used to test the hypothesis that two independent samples have come from the same population. Because it is non-parametric, it makes only limited assumptions about the distribution of the data. It assumes that the shape of the distribution is similar in the two groups. This is of particular relevance if the test is to be used as evidence that the median is significantly different between the groups. The test ranks all the data from both groups. The smallest value is given a rank of 1 ; the second smallest is given a rank of 2, and so on. Where values are tied, they are given an average rank. The ranks for each group are added together (hence the term rank sum test). The sums of the ranks are compared with tabulated critical values to generate a p value. In a Wilcoxon rank sum test, p, a function of X, Y, and α, is the probability of observing a result equal or more extreme than the one using the data (X and Y) if the null hypothesis is true. The value of p indicates the significance for testing the null hypothesis that the populations generating the two independent samples, X and Y, are identical. X and Y are vectors but can have different lengths, i.e., the samples can have different number of elements. The alternative hypothesis is that the median of the X population is shifted from the median of the Y population by a non-zero amount, α is a given level of significance and is a scalar between zero and one. hi some embodiment, the default value of α is set to 0.05. If p is near zero, the null hypothesis may be rejected.

In one embodiment, the PSSM approach of the present invention was compared to the standard method (e.g., Elbashir et al., 2001, Nature 411 :494-8) for its performance in identifying siRNAs having high efficacy. The results obtained with three siRNAs selected

by each method are shown in Figure 3. siRNAs selected by the method using the PSSM showed better median efficacy (88% as compared to 78% for the standard method siRNA) and were more uniform in their performance. The minimum efficacy was greatly improved (75% as compared to 12% for the standard method). The distribution of silencing efficacies of siRNAs designed using the algorithm based on PSSM was significantly better than that of the siRNAs designed using the standard method for the same genes (p=0.004, Wilcoxon rank sum test).

5.1.3. ALTERNATIVE METHOD FOR EVALUATING SILENCING EFFICACY OF siRNAS

Position-specific scoring matrix approaches are the preferred method of representing siRNA functional motifs, e.g., siRNA susceptible and siRNA resistant motifs. However, the information represented by PSSMs can also be represented by other methods which also provide weights for base-composition at particular positions. This section provides such methods for evaluating siRNA functional motifs.

5.1.3.1. METHODS BASED ON SEQUENCE WINDOWS A common method of weighting base-composition at positions in a sequence is to tally the number of a particular base or set of bases in a "window" of sequence positions. Alternatively, the tally is represented as a percentage. The number of values of such a score, referred to as a window score, depends on the size of the window. For example, scoring a window of size 5 for G/C content may give values of 0, 1, 2, 3, 4 or 5; or 0%, 20%, 40%, 60%, 80% or 100%.

An alternative method of scoring a window is to calculate the duplex melting temperature or δG for the bases in that window. These thermodynamic quantities reflect the composition of all bases in the window as well as their particular order. It is readily apparent to one of skill in the art that these thermodynamic quantities directly depend on the base composition of each window, and are dominated by the G/C content of the window while showing some variation with the order of the bases. In one embodiment, the information represented by the base-composition differences, e.g., in Figures IA, IB and 1C, is represented by windows of base- composition corresponding to the positions to the peaks of increased or decreased composition of a particular base(s). These windows can be scored for content of the

particular base(s), with increased or decreased base composition corresponding to sequences which are more or less functional or resistant for siRNA targeting. For example, a 5 -base window of increased G/C content from base -1 to base 3 relative to the siRNA 19mer duplex, and a 16-base window of decreased G/C content from base 14 to base 29 relative to the siRNA 19mer duplex, can be used to represent some of the siRNA functional motifs reflected in Figure IA.

The scores may be used directly as a classifier: in the example of a 5-base window, a 5-part classifier is automatically available. Scores can also be compared to a calculated or empirically derived threshold to use the window as a 2-part classifier. Windows can also be used in combination. The scores of each sequence over multiple windows can be summed with or without normalization or weighting. In one embodiment, scores for each window are normalized by subtracting the mean score in a set of scores and then dividing by the standard deviation in the set of scores. In another embodiment, scores are weighted by the Pearson correlation coefficient obtained by comparing that window's score with the measured efficacy of a set of siRNAs. In another embodiment, scores are normalized, and then weighted before summation.

As an example of the use of windows to represent siRNA functional motifs, the following list of parameters was considered for prediction of siRNA efficacy:

1. Straight-forward parameters. ATG Dist - distance to the start codon.

STOP_Dist - distance to the end of the coding region

Coding Percent - ATG_Dist as percentage of the length of coding region

End_Dist - distance to the end of the transcript

Total Percent - start position as a percentage of the length of the transcript sequence.

2. Window-based parameters.

119 bases on the transcript sequence were considered (19mer plus 50 bases downstream and 50 bases upstream). Windows of sizes 3-10 were examined for each position from the beginning to the end of the 119-base chunk. The following items were counted for each window position: a. Numbers of bases: A, C, G, or U. b. Numbers of pairs of bases: M (A or C), R (A or G), W (A or U), S (C or G), Y (C or U) ₃ and K (G or U). c. Numbers of various ordered dimers: AC, AT, AG, MM, RY ₅ KM, SW, etc.

d. The longest stretches of the above one base or two-base units.

3. Motif-based parameters.

These parameters are also based on the 119-base chunks. The letters include the bases (A, C, G, and U) and pairs of bases (M, R, W, S, Y, and K). (1). Position-Specific one-mer, dimers, or trimers.

(2). Numbers of lmers to 7mers in four large regions: 50 bases upstream, 19mer proper, 50 bases downstream, and the whole 1 19mer region.

4. Structural parameters.

The structural parameters are based on the following regions. the 19mer oligo proper (prefix: proper) the 20mer immediate upstream the oligo (prefix: up20) the 40mer immediate upstream the oligo the 60mer immediate upstream the oligo the 20mer immediate downstream the oligo (prefix: down20) the 40mer immediate downstream the oligo the 60mer immediate downstream the oligo

Base-pairing predicted by RNAStructure was examined and the following parameters were calculated: the count of bulge loops (parameter: bulge) the total bases in the bulge loops (bulge_b) the count of internal loops (internal) the total bases in the internal loops (internal b) the count of hairpins (hairpin) the total bases in the hairpins (hairpin b) the count of other motif regions (other) the total bases in the other motif regions (other_b) the total paired bases (total_pairs_b) the total non-paired bases (total nonpairs b) the longest stretch of paired bases (longest_pairs_b) the longest stretch of non-paired bases (longest_nonpairs_b)

Thus, a total of 12*7=84 parameters were computed about the secondary structure motifs for each siRNA.

5. Parameters on off-target predictions.

10 different parameters were computed using the weighted FASTA score discussed in Section 5.2., the minimax score and the predicted duplex δG discussed in Section 5.4, using different conditions.

Parameters were normalized and weighted by the Pearson correlation coefficient of the scores with the silencing efficacy of the siRNAs examined. Various methods were used to select the parameters with the greatest predictive power for siRNA efficacy; the various methods agreed on the selection 1750 parameters. 1190 of these are window- based base composition parameters, 559 are motif-based base composition parameters, and only 1 structural parameter was selected. No other parameters were selected.

5.1.3.2. SEQUENCE FAMILY SCORING METHODS

Sequence consensus patterns, hidden Markov models and neural networks can also be used to represent siRNA functional motifs, e.g., siRNA susceptible or siRNA resistant motifs as an alternative to PSSMs. First, an siRNA functional motif, e.g. , siRNA susceptible or siRNA resistant motif can be understood as a loose consensus sequence for a family of distantly related sequences - e.g. the family of functional siRNA target sites. Scoring sequences for similarity to a family consensus is well known in the art. More detailed discussion may be found in Gribskov, et al, 1987, "Profile analysis: detection of distantly related proteins," Proc. Natl. Acad. Sci. USA 84:4355-4358 and Gribskov et al, 1990, "Profile Analysis," Meth. Enzymol. 183:146-159, each of which is hereby incorporated by reference herein in its entirety. Such scoring methods are most commonly referred to as "profiles", but may also be referred to as "templates" or "flexible patterns" or similar terms. Such methods are more or less statistical descriptions of the consensus of a multiple sequence alignment, using position-specific scores for particular bases or amino acids as well as for insertions or deletions in the sequence. Weights can be derived from the degree of conservation at each position. A difference between consensus profiles and PSSMs as the term is used in this text is that spacing can be flexible in consensus profiles: discontinuous portions of an siRNA functional motif, e.g., siRNA susceptible or siRNA resistant motif can be found at varying distances to each other, with insertions or deletions permitted and scored as bases are.

Profile hidden Markov models are statistical models which also represent the consensus of a family of sequences. Krogh and colleagues (Krogh et al, 1994, Hidden

Markov models in computational biology: Applications to protein modeling. J. MoI Biol. 235: 1501-1531) applied HMM techniques to modeling sequence profiles, adopting techniques from speech recognition studies (Rabiner, 1989, A tutorial on hidden Markov models and selected applications to speech recognition. Proc. IEEE 77:257-286). The use of hidden Markov models for analysis of biological sequences is now well known in the art and applications for hidden Markov model calculation are readily available; for example, the program HMMER from the HHMI Janelia Farm Research Campus in Ashburn, Virginia.

Profile hidden Markov models differ from consensus profiles as described above in that profile hidden Markov models have a formal probabilistic basis for setting the weights for each base, insertion or deletion at each position. Hidden Markov models can also perform the alignment of unknown sequences for discovery of motifs as well as determining position-specific weights for said motifs, while consensus profiles are generally derived from previously aligned sequences. Consensus profiles and profile hidden Markov models can assume that the base composition at a particular position is independent of the base composition of all other positions. This is similar to the random-hill-climbing PSSMs of this invention but distinct from the windows and curve model PSSMs.

To capture dependency of base composition at a particular position on the composition of neighboring positions, Markov models can be used as fixed-order Markov chains and interpolated Markov models. Salzberg and colleagues applied interpolated Markov models to finding genes in microbial genomes as an improvement over fixed- order Markov chains (Salzberg et al., 1998, Nucl. Acids Res. 26: 544-548). A fixed-order Markov chain predicts each base of a sequence as a function of a fixed number of bases preceding that position. The number of preceding bases used to predict the next is known as the order of the Markov chain. Interpolated Markov models use a flexible number of preceding bases to predict the base composition at a particular position. This permits training on smaller sequence sets. Sufficient predictive data may be available for n-mers of various lengths in a training set such that some predictions of succeeding bases can be made, while insufficient data may be available for all oligomers at any fixed length.

Interpolated Markov models thus have more freedom to use preferable longer oligomers for prediction than fixed-order Markov chains, when said long oligomers are sufficiently frequent in the training set. Interpolated Markov models employ a weighted combination of probabilities from a plurality of oligomer lengths for classification of each base.

Fixed-order Markov chains and interpolated Markov models can represent siRNA functional motifs, e.g., siRNA susceptible or siRNA resistant motifs in terms of the dependency of the base-composition at a particular position on the composition of the preceding positions. An interpolated Markov model building process will discover the oligomers most predictive of siRNA functional or nonfunctional motifs.

Neural networks are also employed to score sequences for similarity to a family of sequences. A neural network is a statistical analysis tool used to build a model through an iterative learning process. The trained network will then perform a classification task, dependent upon the desired output and the training input initially associated with that output. Typically a neural network program or computational device is supplied with a training set of sequences and sets up a state representing those sequences. The neural network is then tested for performance on a test set of sequences. Neural networks can be used to predict and model siRNA functional motifs, e.g., siRNA susceptible and siRNA resistant motifs. A disadvantage of neural networks is that the actual sequence features of a motif can be difficult or impossible to determine from examination of the state of the trained network.

5.1.4. METHODS OF IDENTIFYING SEQUENCE MOTIFS IN A GENE FOR

TARGETING BY AN SIRNA The invention provides a method for identifying one or more sequence motifs in a transcript which are siRNA-susceptible or siRNA-resistant motifs. The corresponding functional or unfunctional siRNAs are thereby also provided by the method. In one embodiment, the sequence region of interest is scanned to identify sequences that match the profile of a functional motif. In one embodiment, a plurality of possible siRNA sequence motifs comprises siRNA sequence motifs tiled across the region at steps of a predetermined base intervals are evaluated to identify sequences that matched the profile. In a preferred embodiment, steps of 1, 5, 10, 15, or 19 base intervals are used. In a preferred embodiment, the entire transcript sequence is scanned. A score is calculated for each different sequence motif using a PSSM as described in Sections 5.1.1 through 5.1.3. The sequences are then ranked according to the score. One or more sequences are then selected from the rank list. In one embodiment, siRNA sequence motifs having the highest scores are selected as siRNA-susceptible motifs. In another embodiment, siRNA sequence motifs having the lowest scores are selected as siRNA resistant motifs.

It has been discovered that the correlation between silencing efficacy and the base composition profiles of siRNA functional motifs may depend on one or more factors, e.g., the abundance of the target transcript. For example, it has been found that for silencing poorly-expressed genes, e.g., genes whose transcript levels are less than about 5 copies per cell, siRNA functional motifs having high GC content asymmetry at the two ends of the target sequence and having high GC content in the sequence regions flanking the target sequence have lower silencing efficacy than siRNA functional motifs having moderate GC content asymmetry at the two ends of the target sequence and low GC content in the flanking regions. The effect of target transcript abundance on silencing efficacy is illustrated in Example 6.

While not to be confined by any theory, is has been reasoned that the silencing efficacy of a particular siRNA functional motif is a result of the interplay of a number of processes, including RISC formation and siRNA duplex unwinding, diffusion of the RISC and target mRNA, reaction of the RISC/target complex, which may include diffusion of the RISC along the target mRNA, cleavage reaction, and products dissociation, etc. Thus, the abundance of the transcript, the base composition profile of the siRNA, the base composition profile of the target sequence and flanking sequences, and the concentration of the siRNA and RISC in a cell may all affect silencing efficacy. Different processes may involve different sequence regions of an siRNA or siRNA sequence motif, i.e., different sequence regions of an siRNA or siRNA sequence motif may have different functions in transcript recognition, cleavage, and product release, siRNAs may be designed based on criteria that take one or more of such features into account. For example, bases near the 5' end of the guide strand are implicated in transcript binding (both on- and off-target transcripts), and have been shown to be sufficient for target RNA- binding energy. Weaker base pairing at the 5' end of the antisense strand (3' end of the duplex) encourages preferential interaction of the antisense strand with RISC, e.g., by facilitating unwinding of the siRNA duplex by a 5 '-3' helicase component of RISC. A preference for U at position 10 of the sense strand of an siRNA has been associated with improved cleavage efficiency by RISC as it is in most endonucleases. Low GC content sequence flanking the cleavage site may enhance accessibility of the RISC/nuclease complex for cleavage, or release of the cleaved transcript, consistent with recent studies demonstrating that base pairs formed by the central and 3' regions of the siRNA guide strand provide a helical geometry required for catalysis. Thus, the invention provides a method of identifying siRNA sequence motifs (and thus siRNAs) by obtaining siRNAs

that have an optimal sequence composition in one or more sequence regions such that these siRNAs are optimal in one or more the siRNA functional processes. In one embodiment, the method comprises identifying siRNA sequence motifs whose overall sequence and/or different sequence regions have desired composition profiles. The method can be used to identify siRNAs motifs that have desired sequence composition in a particular region, thus are optimized for one functional process. The method can also be used to identify siRNAs that have desired sequence composition in a number of regions, thus are optimized for a number of functional processes.

In a preferred embodiment, a single siRNA functional profile, e.g., a profile as represented by a set of PSSMs, is obtained, e.g. , by training with silencing efficacy data of a plurality of siRNAs that target genes having different transcript abundances using a method described in Section 5.1.2 or Section 5.1.3., and is used to evaluate siRNA sequence motifs in gene transcripts having abundances in all ranges. In one embodiment, the siRNA sequence motifs in gene transcripts having abundances in any range are evaluated based on the degree of similarity of their sequence base composition profiles to the profile or profiles represented by the set of PSSMs. In one embodiment, the PSSM scores of siRNA functional motifs for a gene of interest are obtained by a method described in Section 5.1.1. A predetermined reference value or reference range of values of the PSSM score is determined based on siRNAs that target genes having expression levels in different ranges. Methods for determining the reference value or range of reference value is described below. siRNA functional motifs in a particular gene are then ranked based on the closeness of their scores to the predetermined reference value or within the reference range. One or more siRNAs having scores closest to the predetermined value or within the reference range are then selected. In another embodiment, a predetermined reference value of the PSSM score or a reference range of the PSSM scores is used for genes having expression levels in a given range. The reference value or the reference range is determined based on siRNAs that target genes having expression levels in the range. siRNA functional motifs in a particular gene are then ranked based on the closeness of their scores to the predetermined reference value or within the reference range. One or more siRNAs having scores closest to the predetermined value or within the reference range are then selected.

The reference value or the reference range can be determined in various ways. In a preferred embodiment, correlation of PSSM scores of a plurality of siRNAs having one or more features, e.g., having particular efficiency in one or more siRNA functional

processes, with silencing efficacy is evaluated. In a preferred embodiment, the feature is that the plurality of siRNAs targets poorly-expressed genes. The value of the score corresponding to maximum median silencing is used as the reference value. In a specific embodiment, the reference value is 0. One or more siRNAs having PSSM scores the closest to the reference score are selected.

In another embodiment, the range of scores corresponding to siRNAs having a given level of silencing efficacy, e.g., efficacy above 75%, is used as the range for the reference values. In one embodiment, effective siRNAs are found to have scores between -300 and +200 as long as the GC content in bases 2-7 is controlled. In a specific embodiment, a reference value of between -300 and +200 is used. One or more siRNAs having PSSM scores within the range are selected.

In another preferred embodiment, a particular score range within the range of PSSM scores of the plurality of siRNAs having one or more features, e.g., having particular efficiency in one or more siRNA functional processes, is used as the range of the reference value. In a preferred embodiment, the feature is that the plurality of siRNAs targets poorly-expressed genes. In one embodiment, a certain percentile in the range of PSSM scores is used as the range of the reference value, e.g., 90%, 80%, 70%, or 60%. In a specific embodiment, the combined PSSM score range in the training set has a maximum of 200, with 97% of the scores being 0 or below and 60% of the scores are below -300. In still another preferred embodiment, a sum of scores from a plurality of sets of

PSSMs (see Section 5.1.2) is used as the reference score. In a specific embodiment, the plurality of sets consists of the two sets of PSSMs described previously. The two sets of PSSMs differ in the base composition preferred for siRNAs, in particular with respect to the GC content of the 19mer and flanking sequences. With a combined score of 0, the PSSM sets are in balance in their preference for the siRNA.

In another preferred embodiment, in addition to the PSSM scores, the siRNA sequence motifs are also ranked according to GC content at positions corresponding to positions 2-7 of the corresponding siRNAs, and one or more siRNA sequence motifs that have a GC content approximately 0.15 to 0.5 (corresponding to 1-3 G or C) in the region are selected.

In still another preferred embodiment, siRNA sequence motifs having a G or C at the position corresponding to position 1 of the corresponding 19mer siRNA and a A or T at the position corresponding to position 19 of the corresponding 19mer siRNA are selected. In still another preferred embodiment, siRNAs motifs in which 200 bases on

either side of the 19mer target region are not repeat or low-complexity sequences are selected.

In a specific embodiment, the siRNA sequence motifs selected in the following manner: (1) they are first ranked according to GC content at positions corresponding to positions 2-7 of the corresponding siRNAs, and one or more siRNA sequence motifs that have a GC content approximately 0.15 to 0.5 (corresponding to 1-3 G or C) in the region are selected; (2) next, siRNA sequence motifs having a G or C at the position corresponding to position 1 of the corresponding 19mer siRNA and a A or T at the position corresponding to position 19 of the corresponding 19mer siRNA are selected; (3) siRNAs having PSSM scores in the range of -300 to 200 or most close to 0 are then selected; (4) number of off-target BLAST match less than 16 are then selected; and (5) siRNAs motifs in which 200 bases on either side of the 19mer target region are not repeat or low-complexity sequences are selected.

In another embodiment, a reference value or reference range for each of a plurality of different abundance ranges is determined. Selection of siRNA functional motifs in a gene of interest is achieved by using the appropriate reference value or reference range for the abundance range in which the gene of interest falls. In one embodiment, the plurality of different abundance ranges consists of two ranges: below about 3-5 copies per cell, corresponding to poorly-expressed genes, and above 5 copies per cell, corresponding to highly-expressed genes. The reference value or reference range can be determined for each abundance range using any one of the methods described above.

In another embodiment, a plurality of siRNA functional motif profiles are determined for a plurality of different transcript abundance ranges. Each such profile is determined based on silencing efficacy data of siRNAs that target genes having expression levels in a given range, /. e. , genes whose transcript abundances fall within a given range, using a method described in Sections 5.1.2 and 5.1.3., supra. In one embodiment, a set of one or more PSSMs for genes having expression levels in a given range are trained as described in Section 5.1.2., using siRNAs that target genes having expression levels in the range. The PSSMs are then used for identifying siRNA functional motifs in a target gene whose expression level falls in the range, e.g., by ranking according to the PSSM scores obtained using a method described in Section 5.1.1. hi a preferred embodiment, the transcript abundance ranges are divided into two ranges: below about 3-5 copies per cell, corresponding to poorly-expressed genes, and above 5 copies per cell, corresponding to highly-expressed genes. Two sets of PSSMs are obtained, one for each abundance range.

siRNA functional motifs in a gene of interest can be identified using the set of PSSMs that is appropriate for the abundance of the gene of interest.

Methods for evaluating the silencing efficacies of siRNA sequence motifs under different siRNA concentrations are known in the art, see for example, WO 05/042708. The methods described above for evaluating silencing efficacy of siRNA sequence motifs in transcripts having different abundances can be used for such purposes by replacing the abundance parameter with the concentration parameter. In one embodiment, a plurality of siRNA functional motif profiles are determined for a plurality of different siRNA concentration ranges. Each such profile can be determined based on silencing efficacy data of different concentration of siRNAs targeting genes having a different expression level or having an expression level in a different range. In one embodiment, such profiles are determined for transcripts having a given abundance or having a abundance within a range of abundances. Each such profile can be determined based on silencing efficacy data of different concentration of siRNAs targeting genes having the expression level or having an expression level in the range. In one embodiment, one or more PSSMs for a given siRNA concentration range are trained based on silencing efficacy data of siRNAs having a concentration in the range. The PSSMs can then be used for selecting siRNAs that have high efficiency at a concentration that falls in the concentration range. In a preferred embodiment, the transcript abundance ranges is selected to be below 5 copies per cell. In another embodiment, the transcript abundance ranges is selected to be above 5 copies per cell. The invention thus provides a method for selecting one or more siRNA functional motifs for targeting by siRNAs of a given concentration.

The methods can be used for identifying one or more siRNA functional motifs that can be targeted by siRNAs of a given concentration with desired silencing efficacy. The given concentration is preferably in the low nanomolar to sub-nanomolar range, more preferably in the picomolar range. In specific embodiments, the given concentration is 50 nmol, 20 nmol, 10 nmol, 5 nmol, 1 nmol, 0.5 nmol, 0.1 nmol, 0.05 nmol, or 0.01 nmol. The desired silencing efficacy is at least 50%, 75%, 90%, or 99% under a given concentration. Such methods are particularly useful for designing therapeutic siRNAs. For therapeutic uses, it is often desirable to identify siRNAs that can silence a target gene with high efficacy at sub-nanomolar to picomolar concentrations. The invention thus also provides a method for design of therapeutic siRNAs.

A method for determining if a gene is suitable for targeting by a therapeutic siRNA is known; see for example, WO 05/042708. In one embodiment, the desired siRNA

concentration and the desired silencing efficacy are first determined. A plurality of possible siRNA sequence motifs in the transcript of the gene is evaluated using a method of this invention. One or more siRNA sequence motifs that exhibit the highest efficacy, e.g., having PSSM scores satisfying the above described criterion or criteria, are identified. The gene is determined as suitable for targeting by a therapeutic siRNA if the one or more siRNA sequence motifs can be targeted by the corresponding siRNAs with silencing efficacy above or equal to the desired efficacy. In one embodiment, the plurality of possible siRNA sequence motifs comprises siRNA sequence motifs that span or are tiled across a part of or the entire transcript at steps of a predetermined base intervals, e.g. at steps of 1, 5, 10, 15, or 19 base intervals. In a preferred embodiment, successive overlapping siRNA sequence motifs are tiled across the entire transcript sequence. In another preferred embodiment, successive overlapping siRNA sequence motifs tiled across a region of or the entire transcript sequence at steps of 1 base intervals.

5.1.5 METHODS FOR DESIGNING shRNAS CORE SEQUENCES

Gene silencing can also be achieved using a cell-expressed short hairpin RNA (shRNA), e.g., by introducing into a cell a DNA construct which can be transcribed to express an shRNA (see, e.g., Paddison et al, 2002, Genes Dev. 16:948-958; Brummelkamp et al, 2002, Science 296:550-553; Sui et al, 2002, Proc. Natl. Acad. ScL USA 99:5515-5520; Ventura et al, 2004, Proc. Natl. Acad. ScL USA 101 :10381-10385; Chen et al, 2003, Cancer Research 63:4801-4804, each of which is hereby incorporated by reference herein in its entirety). In contrast to synthetic siRNA sequences which can be produced by in vitro oligo-based chemical synthesis methods, when an siRNA is to be produced intracellularly from an shRNA, by way of example, a DNA fragment encoding an shRNA core sequence can be inserted into an expression vector, delivered into a cell where it is subsequently transported into the nucleus of the cell, for example, a human liver cell. Once inside the nucleus, a promoter sequence embedded in the expression vector is recognized by an RNA polymerase, for example, human RNA polymerase II (Pol II) or human RNA polymerase III. In the present invention, an shRNA core sequence comprises two inverted repeats that are separated by an intervening loop region. Upon transcription, the shRNA core sequence forms a hairpin like structure, hence the name short hairpin RNA, with the inverted repeats as the duplex stem connected at one end by the loop sequence. When the properly transcribed shRNAs are transported out of the cell nucleus, processing enzymes,

for example Dicer, cleave the loop ends to convert shRNA into functional siRNA with gene silencing capacity. The inverted repeats of the shRNA core sequence form the sense and antisense strands of the functional siRNA. For each shRNA core sequence, a functional siRNA may be produced. The functional siRNA is accordingly termed a corresponding siRNA sequence to the original shRNA core sequence.

It has been discovered that the overall silencing efficacy of an shRNA core sequence, i.e., the silencing efficacy of a cell-expressed shRNA, is determined by an interplay between the silencing efficacy of the corresponding siRNA (i.e., the siRNA produced from the shRNA core sequence) and the transcription efficiency of the shRNA. For example, it has been shown that a G at the first or second 5' nucleotide of the sense strand provides the best transcription level under an RNA polymerase III promoter.

The methods, computer systems, and computer program products of the present invention take advantage of the discovery by selecting and using siRNA sequence motifs with high silencing efficacy as the antisense strand sequences of shRNA core sequences. In addition, the transcription efficiency of the shRNA core sequences may be optimized to further enhancing the silencing efficacy of the shRNA core sequences.

5.1.5.1 INVERTED REPEATS IN shRNA CORE SEQUENCES In some embodiments, the invention provides that the inverted repeat sequences in an shRNA core sequence may be designed based on a predicted silencing efficacy for the corresponding siRNA of the original shRNA core sequence. In these embodiments, sequences for the inverted repeats are designed in accordance with the methods described in sections 5.1.1 through 5.1.4. .

The present invention provides a method of identifying a sequence motif in a target transcript which may be targeted by a cell expressed shRNA, e.g., a sequence motif that is likely to be an effective shRNA targeting site. Such a sequence motif is also referred to as an shRNA susceptible motif. Similarly, the method can also be used for identifying a sequence motif in a transcript which may be less desirable for targeting by an shRNA, e.g., a sequence motif that is likely to be a less effective shRNA targeting site. Such a sequence motif is also referred to as an shRNA resistant motif.

An shRNA core sequence which produces a corresponding siRNA that binds to an shRNA susceptible motif likely exhibits high silencing efficacy against the target transcript. The silencing efficacy of an shRNA is embedded in the inverted repeats, which are designed in accordance with the methods for identifying an siRNA susceptible motif,

as disclosed in sections 5.1.1 through 5.1.4. Essentially, the silencing efficacy of an shRNA core sequence may be predicted based on a predicted or observed silencing efficacy of its corresponding siRNA, for example, using methods as disclosed in WO 05/042708. In some embodiments, the silencing efficacies of the corresponding siRNAs may be based on known experimental data, for example, by selecting the siRNA sequences with previously determined silencing efficacies from an siRNA database. Exemplary siRNA or miRNA databases or libraries include siRNAdb hosted by Center for Genomics and Bioinformatics, Karolinska Institutet in Stockholm, Sweden (see Chalk et al, 2005, Nucleic Acids Research 33 (Database Issue):D131-D134); HuSiDa (Human siRNA Database) hosted by Department of Pediatrics, Laboratory of Molecular Biology at Humboldt-University in Berlin, Germany; and the miRBase hosted by the Sanger Institute in Cambridge United Kingdom (see, for example, Griffiths- Jones et al., 2006, "miRBase: microRNA sequences, targets and gene nomenclature," Nucleic Acids Research, 34 (Database Issue):D140-D144; Griffiths- Jones, 2004, "The microRNA Registry," Nucleic Acids Research, 32 (Database Issue) :D109-Dl 11; and Ambros et al., 2003, "A uniform system for microRNA annotation," RNA 9: 277-279).

In embodiments in accordance with the present invention, siRNA functional sequence motifs are selected based on a predicted silencing efficacy value that is greater than a pre-determined threshold value. In some embodiments, all siRNA functional sequence motifs with predicted silencing efficacy values above the pre-determined value will be separated into two sub-groups based on the composition of the last nucleotide residue in the sequence motif. The first sub-group comprises all siRNA functional sequence motifs that end with a Cytosine (C) at the 3 '-terminus. In some embodiments, an siRNA functional sequence motif will be selected from the first sub-group as the antisense strand sequence of the shRNA core sequence. The sense sequence will be created as the inverted repeat to the antisense strand, thereby ensuring that the sense strand will start with a Guanine (G) at the 5 '-terminus. The second sub-group comprises those siRNA functional sequence motifs that do not end with a Cytosine (C) at the 3 '-terminus. In some embodiments, an siRNA functional sequence motif will be selected from the second sub- group. The last nucleotide at the 3 '-terminus of the selected siRNA functional sequence motif will be replaced with a Cytosine (C) before it is used as the antisense strand of an shRNA core sequence. The sense sequence will be created as the inverted repeat to the antisense strand, thereby ensuring that the sense strand will start with a Guanine (G) at the 5 '-terminus.

In some embodiments, siRNA functional sequence motif will be selected from any of the siRNA functional sequence motifs that have silencing efficacies above a predetermined value. A Cytosine (C) will be added to the 3' terminus of the selected siRNA functional sequence motif before it is used as the antisense strand of an shRNA core sequence. The sense sequence will be created as the inverted repeat to the antisense strand, thereby ensuring that the sense strand will start with a Guanine (G) at the 5'- terminus.

It is to be appreciated that the methods of the present invention are not limited to shRNA core sequences that are based on siRNA functional sequence motifs or miRNA sequence motifs. Additional small RNA sequence motifs which exhibit RNA silencing may also be used in constructing the shRNA core sequences. For example, the new classes of small RNAs other than siRNAs and miRNAs (e.g., 21U-RNAs) may also be used to construct shRNA core sequences. Examples of 21U-RNAs in C. elegans can be found in Ruby et al, 2006, "Large-scale Sequencing Reveals 21U-RNAs and Additional microRNAs and Endogenous siRNAs in C. elegans," Cell 127: 1193-1207.

The silencing efficacy of an shRNA core sequence can be actually measured or predicted. In some embodiments, the silencing efficacy of an siRNA produced from an shRNA core sequence can be predicted using any of the methods described in Sections 5.1.1 to 5.1.4. For example, the silencing efficacy of a corresponding siRNA can be evaluated based on positional base composition of a targeted sequence motif in the transcript, e.g., a targeted sequence motif that comprises at least a portion of the target sequence of the siRNA and/or a second sequence in a sequence region within about 200 nucleotides from said target sequence in the RNA transcript. When predicting the silencing efficacy of an siRNA functional sequence motif against a targeted sequence motif within a transcript of a target gene, the targeted sequence motif comprises at least a portion of the target sequence of the siRNA and/or a second sequence in a sequence region within about 200 nucleotides from said target sequence in the RNA transcript. Preferably, the targeted sequence motif comprises the target sequence of the siRNA in the transcript. In some embodiments, the length of the duplex stems of the corresponding siRNAs varies from 19 nt to 21 nt. Because a functional siRNA duplex has a two-nucleotide overhang at one end of each of the sense and antisense strands, each of the inverted repeats form the corresponding siRNAs accordingly contain from 21 nt to 23 nt. In some embodiments, longer inverted repeats may be use to construct the shRNAs core sequences, for example, Kim et αl. , have used as long as 27 nt to construct double stranded RNAs (dsRNAs) for

gene silencing purpose. See, Kim, et al, 2005, "Synthetic dsRNA Dicer substrates enhance RNAi potency and efficacy," Nature Biotechnology 23:222-226, which is hereby incorporated by reference herein by reference in its entirety.

5.1.5.2 LOOP SEQUENCES IN shRNA CORE SEQUENCES

In some embodiments, the loop region that connects the two inverted repeats in an shRNA core sequence varies from three nucleotides to over 20 nucleotides, for example over 25 nucleotides or even longer. In a preferred embodiments, a loop consists of nine nucleotides. Loop sequences that facilitate hairpin formation have been reported in the literatures. Examples of these loop sequences include but are not limited ATG, CCC, TTCG, CCACC, CTCGAG, AAGCTT, CCACACC, TTCAAGAGA. More details on loop sequence selection and design can be found in Sui, et al , 2002, "A DNA vector- based RNAi technology to suppress gene expression in mammalian cells," Proc. Natl. Acad. ScL USA 99(8):5515-5520; Lee et al, 2002, "Expression of small interfering RNAs targeted against HIV-I rev transcripts in human cells," Nature Biotechnology 20:500-505; Yu et al. , 2002, "RNA interference by expression of short-interfering RNAs and hairpin RNAs in mammalian cells," Proc. Natl. Acad. Sci. USA 99(9):6047-6052; and Paul, et al., 2002, "Effective expression of small interfering RNA in human cells," Nature Biotechnology 20:505-508.

5.1.5.3 TRANSCRIPTION EFFICIENCY OF shRNA CORE SEQUENCES The transcription efficiency of the shRNA can be predicted based on the sequence characteristics preferred by the cellular mechanism employed for the transcription of the shRNA from its encoding DNA construct (e.g., an shRNA design). The transcription efficiency of an shRNA is affected by the identity of one or more nucleotides at the 5' end of said shRNA or by the context into which the shRNA is embedded. For example, an shRNA can be expressed as a simple hairpin stem loop structure or an shRNA can be inserted into a microRNA-like context such that flanking sequence derived from natural microRNAs is expressed upstream and downstream of the shRNA itself. For example, the flanking sequence can be derived from human microRNA 30 (mir30). Examples of mir30 may be found in Stegmeier et al., 2005, Proc. Natl. Acad. Sci. USA 102:13212-13217; Miyagishi et al, 2004, The Journal of Gene Medicine 6:715-723; and Shin et al, 2006, Proc. Natl. Acad. Sci. USA 103: 13759-13764; each of which is hereby incorporated by

reference herein in its entirety. The present invention also provides that a G at the first or second 5' nucleotide of the sense strand enhances silencing efficacy of a corresponding siRNA, possibly through enhancing the transcriptional efficiency of the shRNA core sequence from which the corresponding siRNA is produced. In one embodiment, the invention provides a method for identifying an shRNA susceptible sequence motif in a transcript of a target gene. The method involves identifying one or more sequence motifs that have a high silencing efficacy, e.g., using any of the methods described in Sections 5.1.1 to 5.1.4. For example, the method can involves identifying one or more sequence motifs for the inverted repeats that have a silencing efficacy above a predetermined threshold. The threshold can be a rank based threshold. Thus, in one embodiment, a plurality of candidate sequence motifs shRNA inverted repeats is ranked according to their predicted silencing efficacies. The candidate target sequence motifs having a silencing efficacy above a threshold percentile, e.g., 5%, 10% or 20%, in the rank list are selected. From the identified one or more candidate sequence motifs, a sequence motif corresponding to an shRNA having a GY (Y is any nucleotide) or XG (X is A or T) at the first and second nucleotide positions is identified as an shRNA susceptible sequence motif. The overall silencing efficacy of such shRNAs from highest to lowest is as follows: for Y, G>T>A>C; for X, T>A. Thus, shRNA susceptible sequence motifs and the corresponding shRNAs can be selected according to these orders for desired overall silencing efficacy. In a specific embodiment, a sequence motif corresponding to an shRNA having G at the first nucleotide position at the 5' end is identified as an shRNA susceptible sequence motif. In another specific embodiment, a sequence motif corresponding to an shRNA having a G at the second nucleotide position at the 5' end and not having a C at the first nucleotide position is identified as an shRNA susceptible sequence motif. In still another specific embodiment, a sequence motif corresponding to an shRNA having a G at the first nucleotide position at the 5' end of said shRNA, and a G at the second nucleotide position at the 5' end is identified as an shRNA susceptible sequence motif. In still another specific embodiment, a sequence motif corresponding to an shRNA having a G at the first nucleotide position at the 5' end, and a T at the second nucleotide position at the 5' end is identified as an shRNA susceptible sequence motif. The method can also be used for identifying shRNAs for silencing a gene. For example, the shRNAs can be selected according to the overall silencing efficacies of their target shRNA functional sequence motifs by the above described methods.

In another embodiment, the invention provides a method for designing an shRNA for silencing a target gene. The method involves identifying one or more shRNAs whose corresponding siRNAs have silencing efficiencies above a selected threshold value, e.g., using any of the methods described in Sections 5.1.1 to 5.1.4. The threshold can be a rank based threshold. Thus, in one embodiment, a plurality of candidate shRNA target sequence motifs are ranked according to their silencing efficacy. The shRNA target sequence motifs having a silencing efficacy above a threshold percentile, e.g., 5%, 10% or 20%, in the rank list are selected. One or more nucleotides at the 5' end of these shRNAs are examined and replaced with different nucleotides such that the modified shRNAs have enhanced overall silencing efficacy. In preferred embodiments, for example, when the RNA polymerase III promoter is used, the first and second nucleotide positions at the 5' end are modified to be GY (Y is any nucleotide) or XG (X is A or T). The overall silencing efficacy of such shRNAs from highest to lowest is as follows: for Y, G>T>A>C; for X, T>A. Thus, shRNAs having desired silencing efficiencies can be designed according to these orders. In a preferred embodiment, for example, when the RNA polymerase III promoter is used, the nucleotide at the first nucleotide position at the 5' end in the identified one or more shRNAs is replaced with a G, if it is not a G. In another preferred embodiment, for example, when the RNA polymerase III promoter is used, the nucleotide at the second position of the 5' end in the identified one or more shRNAs is replaced with a G, if it is not a G, and the nucleotide at the first position of the 5' end is replaced such that it is not a C. The replacement of the one or more nucleotides creates mismatches between the produced siRNA and its target sequence in the transcript. However, it has been found that such mismatches do not reduce the silencing efficacy of the siRNA. In a preferred embodiment, the shRNA design containing the nucleotide sequence encoding the shRNA core sequence is under the control of a mammalian promoter (e.g., a human or mouse Hl promoter, a human or mouse U6 promoter, or a human U2 promoter) for transcription by a RNA polymerase III. In another preferred embodiment, the DNA construct containing the nucleotide sequence encoding the shRNA is under the control of a RNA polymerase II (Pol II) promoter, e.g., a promoter containing a TATA upstream of the shRNA encoding sequence, for transcription by a RNA polymerase II. An example of shRNA expression utilizing Pol II can be found in Stegmeier et αl, 2005, Proc. Nαtl. Acαd. Sci. USA 102: 13212-13217, which is hereby incorporated by reference herein in its entirety.

In still another embodiment, the invention provides a method for designing a DNA construct that can be introduced into a cell of an organism to produce a short hairpin RNA (shRNA) in the cell. The method involves evaluating the silencing efficacies of a plurality of different siRNAs, each of which targets a different target sequence in a transcript of the target gene, and selecting an siRNA that has a desired silencing efficacy, e.g., an siRNA that has a predicted silencing efficacy above a predetermined threshold. The selected siRNA sequence is then filtered to eliminate the siRNA sequence motifs that do not contain a Guanine (G) at the 5' terminal position. Any of the methods described in this Section and Sections 5.1.1 to 5.1.4 can be used for this purpose. siRNA sequence motifs selected after the filtering step can be used as construct a shRNA core sequence, for example, as the antisense strand. The shRNA core sequence can be further designed and constructed by creating a sense strand based on the antisense strand and linking the two strands of the siRNA with a loop sequence. Upon transcription, the shRNA core sequence can produce an siRNA with a functional sequence motif corresponding to the selected siRNA.

In some embodiments, an shRNA design can be designed and constructed by fusing the shRNA core sequence with a promoter and at least one flanking nucleotide sequences. In a preferred embodiment, the shRNA design comprises in 5' to 3' order (i) a promoter, (ii) the optional first flanking nucleotide sequence, (iii) the nucleotide sequence encoding the shRNA, and (iv) the optional second flanking nucleotide sequence.

Preferably, the optional first flanking nucleotide sequence in the shRNA design encodes a 5' single-stranded flanking RNA sequence, and the optional second flanking nucleotide sequence in the shRNA design encodes a 3' single-stranded flanking RNA sequence. Thus, when transcribed under the control of the promoter, the shRNA design produces an RNA hairpin (hereinafter termed "the primary RNA hairpin") consisting of the shRNA hairpin and at least one of the 5' single-stranded flanking RNA sequence and the 3' single- stranded flanking RNA sequence.

In one embodiment, the shRNA design comprises the first flanking nucleotide sequence but not the second flanking nucleotide sequence. In such an embodiment, the primary RNA hairpin consists of in 5' to 3' order the 5' single-stranded flanking RNA sequence and the shRNA hairpin. In another embodiment, the shRNA design comprises the second flanking nucleotide sequence but not the first flanking nucleotide sequence. In such an embodiment, the primary RNA hairpin consists of in 5' to 3' order the shRNA hairpin and 3' single-stranded flanking RNA sequence. In a preferred embodiment, the

shRNA design comprises both the first flanking nucleotide sequence and the second flanking nucleotide sequence. In such an embodiment, the primary RNA hairpin consists of in 5' to 3' order the 5' single-stranded flanking RNA sequence, the shRNA hairpin, and the 3' single-stranded flanking RNA sequence. The first flanking nucleotide sequence preferably consists of at least 3 nucleotides.

In more preferred embodiments, the first flanking nucleotide sequence consists of at least 5, 8, 10, 15, 20, 30, 40 or 50 nucleotides. Similarly, the second flanking nucleotide sequence preferably consists of at least 3 nucleotides. In more preferred embodiments, the second flanking nucleotide sequence consists of at least 5, 10, 15, 20, 30, 40 or 50 nucleotides. Preferably, the first and the second flanking nucleotide sequences are such that the encoded 5' single-stranded flanking RNA sequence and the 3' single-stranded flanking RNA sequence do not form any secondary structures under physiological conditions (e.g., at approximately 37 ⁰ C and a pH level around 7.0). In a preferred embodiment, secondary structures may be avoided by selecting flanking nucleotide sequences that have been previously shown to discourage secondary structures formation during in vitro analysis where any secondary structures of the RNA molecules are monitored by spectroscopic methods such as UV absorbance. In some embodiments, the flanking nucleotide sequences are selected based on a computer-based prediction algorithm, e.g., original compiled based on experimental data. Preferably, the first and/or the second flanking nucleotide sequences is such that the encoded 5' single-stranded flanking RNA sequence and/or the 3' single-stranded flanking RNA sequence facilitates cleavage of the primary RNA hairpin that are produced from the shRNA design by Drosha.

In another embodiment, the shRNA design comprises a Pol III promoter, e.g., a human or mouse Hl promoter or a human or mouse U6 promoter. In such an embodiment, when the shRNA design is introduced into a cell expressing or containing Pol III, the primary RNA hairpin is produced by transcription by Pol III. Preferably, the first flanking nucleotide sequence starts with a guanine (G) at the 5' end. Thus, the invention provides a method utilizing Pol III for producing a primary RNA hairpin. In still another embodiment, the DNA construct comprises a Pol II promoter. In such an embodiment, when the shRNA design is introduced into a cell expressing or containing Pol H, the primary RNA hairpin is produced by transcription by Pol II.

In any of the methods described above, the promoter can be an inducible promoter such that the expression of the shRNA and hence the silencing of its target gene can be

turned on or off as desired. Inducible expression of an siRNA is particularly useful for targeting essential genes. In one embodiment, the expression of the shRNA is under the control of a regulated promoter that allows tuning of the silencing level of the target gene. In one embodiment, a tetracycline regulated gene expression system is used (see, e.g., Gossen et al, 1995, Science 268:1766-1769; U.S. Patent No. 6,004,941; each of which is hereby incorporated herein by reference in its entirety). In other embodiments, an ecdysone-regulated gene expression system (see, e.g., Saez et al, 2000, Proc. Natl. Acad. ScL USA 97:14512-14517, which is hereby incorporated herein by reference in its entirety), or an MMTV glucocorticoid response element regulated gene expression system (see, e.g., Lucas et al., 1992, Annu. Rev. Biochem. 61:1131, which is hereby incorporated herein by reference in its entirety) may be used to regulate the expression of the shRNA.

In any of the methods described above, the shRNA can be expressed from a plasmid (or virus). In another embodiment, the shRNA is expressed from recombinant vectors introduced either transiently or stably integrated into the genome of the target cells (see, e.g., Paddison et al, 2002, Genes Dev 16:948-958; Sui et al, 2002, Proc Natl Acad Sci USA 99:5515-5520; Yu et al, 2002, Proc Natl Acad Sci USA 99:6047-6052; Miyagishi et al, 2002, Nat Biotechnol 20:497-500; Paul et al, 2002, Nat Biotechnol 20:505-508; Kwak et al, 2003, J Pharmacol Sci 93:214-217; Brummelkamp et al, 2002, Science 296: 550-553; Boden et al , 2003, Nucleic Acids Res 31 :5033-5038; Kawasaki et al, 2003, Nucleic Acids Res 31 :700-707, each of which is hereby incorporated by reference herein in its entirety).

5.2. METHODS OF IDENTIFYING OFF-TARGET GENES OF AN siRNA

The invention also provides a method of identifying off-target genes of an siRNA. As used herein, an "off-target" gene is a gene which is directly silenced by an siRNA that is designed to target another gene (see, International application No. PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is incorporated herein by reference in its entirety). An off-target gene can be silenced by either the sense strand or the antisense strand of the siRNA.

5.2.1. SEQUENCE MATCH PROFILE AND OFF-TARGET SILENCING

Microarray experiments suggest that most siRNA oligos result in downregulation of off-target genes through direct interactions between an siRNA and the off-target transcripts. While sequence similarity between dsRNA and transcripts appears to play a

role in determining which off-target genes are affected, sequence similarity searches, even combined with thermodynamic models of hybridization, are insufficient to predict off-target effects accurately. However, alignment of off-target transcripts with offending siRNA sequences reveals that some base pairing interactions between the two appear to be more important than others (Fig. 6).

The invention provides a method of identifying potential off-target genes of an siRNA using a PSSM that describes the sequence match pattern between an siRNA and a sequence of an off-target gene (pmPSSM). In one embodiment, the sequence match pattern is represented by weights of different positions in an siRNA to match the corresponding target positions in off-target transcripts {P,}, where P ₁ is the weight of a match at position /, i = 1, 2, .... L, where L is the length of the siRNA. Such a match pattern can be determined based on the frequency with which each position in an siRNA is found to match affected off-target transcripts identified as direct targets of the siRNA by simultaneous downregulation with the intended target through kinetic analysis of expression profiles (see, International application No. PCT/US2004/015439 by Jackson et al., filed on May 17, 2004). A pmPSSM can be {E,}, where E ₁ = P ₁ if position i in the alignment is a match and E ₁ = (l-P,)/3 if position i is a mismatch. An exemplary {P,} for a 19mer siRNA sequence is plotted in FIG. 7 and listed in Table I. Table I Weights of an exemplary pmPSSM for 21nt siRNAs having a 19 nt duplex region

1 0.25

2 0.32

3 0.32

4 0.46

5 0.39

6 0.38

7 0.36

8 0.45

9 0.61

10 0.47

11 0.76

12 0.96

13 0.94

14 0.81

15 0.92

16 0.94

17 0.89

18 0.78

19 0.58

In one embodiment, sequence match pattern of off-target transcripts are used to obtain a pmPSSM. Off-target genes of an siRNA can be identified using a method disclosed in International application No. PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is hereby incorporated by reference herein in its entirety. For example, off-target genes of an siRNA are identified based on silencing kinetics (see, e.g., International application No. PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is hereby incorporated by reference herein in its entirety). A pmPSSM can then be generated using the frequency of matches found for each position. In one embodiment, the alignment shown in Fig. 6 and similar data for other siRNAs were combined to generate the exemplary position-specific scoring matrix for use in predicting off-target effects.

The degree of match between an siRNA and a sequence in a transcript can be evaluated with the pmPSSM using a score (also referred to as a position match score, pmScore) according to the following equation

L Score = ∑ In(E, /0.25)

(6) where L is the length of the alignment, e.g., 19. A pmScore above a given threshold identifies the sequence as a potential off-target sequence.

It has been discovered that for a given siRNA the number of alignments with a score above a threshold is predictive of the number of observed off-target effects. The score threshold can be optimized by maximizing the correlation between predicted and observed numbers of off-target effects (Fig. 8). The optimized threshold can be used to favor selection of siRNAs with relatively small numbers of predicted off-target effects.

5.2.2. METHOD OF IDENTIFYING OFF-TARGET GENES OF AN siRNA

Off-target genes of a given siRNA can be identified by first identifying off-target transcript sequences that align with the siRNA. Any suitable method for pair-wise alignment, such as but not limited to BLAST and FASTA, can be used. The position-

specific scoring matrix is then used to calculate position match scores for these alignments. In a preferred embodiment, alignments are established with a low-stringency FASTA search and the score for each alignment is calculated according to Eq. 6. A score above a given threshold identifies the transcript comprising the sequence as a potential off-target gene.

The invention thus also provides a method of evaluating the silencing specificity of an siRNA. In one embodiment, potential off-target genes of the siRNA are identified. The total number of such off-target genes in the genome or a portion of the genome is then used as a measure of the silencing specificity of the siRNA.

5.3. METHOD FOR PREDICTION OF STRAND PREFERENCE OF siRNAS

The invention provides a method for predicting strand preference and/or the efficacy and specificity of siRNAs based on position specific base composition of the siRNAs. It has been discovered that an siRNA whose base composition PSSM score (see Section 5.1.) is greater than the base composition PSSM (G/C PSSM) score of its reverse complement is predicted to have an antisense strand that is more active than its sense strand. In contrast, an siRNA whose base composition PSSM score is less than the base composition PSSM score of its reverse complement is predicted to have a sense strand that is more active than its antisense strand.

It has been shown that increased efficacy of an siRNA in silencing a sense- identical target gene corresponds to greater antisense strand activity and lesser sense strand activity. It has been discovered that base composition PSSMs can be used to distinguish siRNAs with strong sense strands as bad siRNAs from siRNAs with weak sense strands as good siRNAs. The reverse complements of bad siRNAs were seen to be even more different from the bad siRNAs themselves than are good siRNAs. On the average, the reverse complements of bad siRNAs had even stronger G/C content at the 5' end than the good siRNAs did and were similar in G/C content to good siRNAs at the 3' end. In contrast, the reverse complements of good siRNAs were seen to be substantially more similar to bad siRNAs than the good siRNAs were. On the average, the reverse complements of good siRNAs hardly differed from bad siRNAs in G/C content at the 5' end and were only slightly less G/C rich than bad siRNAs at the 3' end. These results indicate that the G/C PSSMs distinguish siRNAs with strong sense strands as bad siRNAs from siRNAs with weak sense strands as good siRNAs.

FIG. 14A shows the difference between the mean G/C content of the reverse complements of bad siRNAs with the mean G/C content of the bad siRNAs themselves, within the 19mer siRNA duplex region. The difference between the mean G/C content of good and bad siRNAs is shown for comparison. The curves are smoothed over a window of 5 (or portion of a window of 5, at the edges of the sequence).

FIG. 14B shows the difference between the mean G/C content of the reverse complements of good siRNAs with the mean G/C content of bad siRNAs, within the 19mer siRNA duplex region. The difference between the mean G/C content of good and bad siRNAs is shown for comparison. The curves are smoothed over a window of 5 (or portion of a window of 5, at the edges of the sequence).

In FIG. 15, siRNAs were binned by measured silencing efficacy, and the frequency of sense-active calls by the 3 '-biased method and G/C PSSM method was compared. Although these techniques are based on different analyses, the agreement is quite good. Both show that a higher proportion of low-silencing siRNAs vs. high-silencing siRNAs are predicted to be sense active. The correlation coefficient for (siRNA G/C PSSM score - reverse complement G/C PSSM score) vs. logio(sense-identity score/antisense-identity score) is 0.59 for the set of 61 siRNAs binned in FIG. 15.

Thus, in one embodiment, the invention provides a method for predicting strand preference, i.e., which of the two strands is move active, of siRNAs based on position specific base composition of the siRNAs. In one embodiment, the method comprises evaluating the strand preference of an siRNA in gene silencing by comparing the base compositions of the sense and the antisense strands of the siRNA. In another embodiment, the method comprises evaluating the strand preference of an siRNA in gene silencing by comparing the base compositions of the sense and the reverse complement of the target sequence of the siRNA.

In one embodiment, the sequence of the antisense strand of an siRNA or the reverse complement of the target sequence of the siRNA in a transcript is compared with the target sequence using a PSSM approach (see Section 5.1.). An siRNA and its reverse complement are scored using a PSSM based on a smoothed G/C content difference between good and bad siRNAs within the duplex region as the weight matrix. In one embodiment, a base composition weight matrix as described by FIG. 14A is used as the weight matrix. In a preferred embodiment, the PSSM score of each strand can be calculated as the dot product of the siRNA strand G/C content with the G/C content difference matrix (as the score calculation method of curve model PSSMs). In one

embodiment, an siRNA is identified as sense-active if its reverse complement PSSM score exceeded its own PSSM score.

In another embodiment, the 3 '-biased method as described in International application Number PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is incorporated by reference herein in its entirety, is used in conjunction with the PSSM score to determining the strand preference of an siRNA. In such an embodiment, an siRNA is identified as sense-active by the 3 '-biased method of strand preference determination if the antisense-identical score exceeded the sense-identical score.

The method based on comparison of G/C PSSMs of siRNAs and their reverse complements for prediction of strand bias was tested by comparison with estimation of strand bias from siRNA expression profiles by the 3 '-biased method.

The invention also provides a method for identifying siRNAs having good silencing efficacy. The method comprises identifying siRNAs having dominant antisense strand activity ("antisense-active" siRNAs) as siRNAs having good silencing efficacy and specificity (for silencing sense-identical target). In one embodiment, the method described in Section 5.1 is used to identify siRNAs having good sense strand (i.e., identifying siRNAs having good silencing efficacy towards an antisense-identical target). Such siRNAs are then eliminated from uses in silencing sense-identical targets. The method can also be used to eliminate siRNAs with dominant sense strand activity ("sense-active" siRNAs) as siRNAs having less efficacy and specificity for silencing sense-identical targets. In one embodiment, the method described in International application Number PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is incorporated by reference herein in its entirety, is used to determine strand preference of an siRNA. The reverse complements of bad siRNAs, on the average, appear to have a GC content profile which differs from that of bad siRNAs in the same manner as the GC content profile of good siRNAs differs from that of bad siRNAs. However, the reverse complements of bad siRNAs show even more extreme differences from bad siRNAs than do the good siRNAs. This observation is in accord with the evidence in siRNA expression profiles that many bad siRNAs have active sense strands.

The combination of data and analysis thus suggests that the reverse complements of bad siRNAs form an alternative, or perhaps even more advantageous, model for effective siRNAs than the good siRNAs do. Thus, the invention also provides a method

for selecting siRNAs based on the base composition of the sequence of a reverse complement of the sense strand of the siRNAs. In one embodiment, a plurality of different siRNAs designed for silencing a target gene in an organism at a different target sequence in a transcript of the target gene is ranked according to positional base composition of the reverse complement sequences of their sense strands. One or more siRNAs whose reverse complement sequences' positional base composition matches the positional base composition of desired siRNAs can then be selected. Preferably, the ranking of siRNAs is carried out by first determining a score for each different siRNA using a position-specific score matrix. The siRNAs are then ranked according to the score. Any method described in Section 5.1., supra, can be used to score reverse complement sequences. In one embodiment, for siRNAs that have a nucleotide sequence of Z nucleotides in the duplex region, L being an integer, the position-specific score matrix comprises a difference in probability of finding nucleotide G or C at sequence position k between reverse complement of a first type of siRNA and reverse complement of a second type of siRNA designated as W _k , k =1, ..., L. The score for each reverse complement is calculated according to equation

L

Score = ∑w _k (7)

The first type of siRNA can consist of one or more siRNAs having silencing efficacy no less than a first threshold, e.g., 75%, 80% or 90% at a suitable concentration, e.g., 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM, and the second type of siRNA can consist of one or more siRNAs having silencing efficacy less than a second threshold, e.g., 25%, 50%, or 75% at a suitable concentration, e.g., 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM . In a preferred embodiment, the difference in probability is described by a sum of Gaussian curves, each of said Gaussian curves representing the difference in probability of finding a G or C at a different sequence position.

The methods of this invention can also be applied to developing models, e.g., PSSMs, of siRNA functional motifs by training position-specific scoring matrices to distinguish between bad siRNAs and their reverse complements (see, e.g., Section 5.1.). A restriction in this analysis is that the reverse complements of bad siRNAs have no designated targets. Thus, in one embodiment, position-specific scoring matrices of 19mer siRNA duplex sequences are trained to distinguish between bad siRNAs and their reverse complements.

Flanking sequence training can be performed on off-target genes in the case of distinguishing between bad siRNAs and their reverse complements, as well as in the case of distinguishing between any two groups of siRNAs. In other words, the off-target activity of siRNAs can be hypothesized to have the same flanking sequence requirements as the on-target activity, as the same RNA-protein complexes are thought to be involved in both processes.

Thus, if the methods of the off-target application are used to identify genes directly down-regulated by an siRNA (i.e. through kinetic analysis of down-regulation to identify a group of genes down-regulated with the same half-life as the intended target), the regions flanking the alignment of the siRNA with the directly regulated off-target genes can be used to train and test models of flanking sequence requirements. These models can be developed by any of the methods of this invention: random hill-climbing PSSMs, curve-model PSSMs, good-bad difference frequency matrices, good-composition frequency matrices, and/or bad-composition frequency matrices, etc.

5.4. METHODS OF DESIGNING siRNAS FOR GENE SILENCING The invention provides a method for designing siRNAs for gene silencing. The method can be used to design siRNAs that have full sequence homology to their respective target sequences in a target gene. The method can also be used to design siRNAs that have only partial sequence homology to a target gene. Methods and compositions for silencing a target gene using an siRNA that has only partial sequence homology to its target sequence in a target gene is disclosed in International application No. PCT/US2004/015439 by Jackson et al., filed on May 17, 2004, which is incorporated by reference herein in its entirety. For example, an siRNA that comprises a sense strand contiguous nucleotide sequence of 11-18 nucleotides that is identical to a sequence of a transcript of the target gene but the siRNA does not have full length homology to any sequences in the transcript may be used to silence the transcript. Such contiguous nucleotide sequence is preferably in the central region of the siRNA molecules. A contiguous nucleotide sequence in the central region of an siRNA can be any continuous stretch of nucleotide sequence in the siRNA which does not begin at the 3' end. For example, a contiguous nucleotide sequence of 11 nucleotides can be the nucleotide sequence 2-12, 3-13, 4-14, 5-15, 6-16, 7-17, 8-18, or 9-19. In preferred embodiments, the contiguous nucleotide sequence is 11-16, 11-15, 14-15, 11, 12, or 13 nucleotides in length. Alternatively, an siRNA that comprises a 3' sense strand contiguous nucleotide sequence

of 9-18 nucleotides which is identical to a sequence of a transcript of the target gene but which siRNA does not have full length sequence identity to any contiguous sequences in the transcript may also be used to silence the transcript. A 3' 9-18 nucleotide sequence is a continuous stretch of nucleotides that begins at the first paired base, i.e., it does not comprise the two base 3' overhang. In preferred embodiments, the contiguous nucleotide sequence is 9-16, 9-15, 9-12, 11, 10, or 9 nucleotides in length.

In preferred embodiments, the method of Section 5.1 is used for identifying from among a plurality of siRNAs one or more siRNAs that have high silencing efficacy. In one embodiment, each siRNA in the plurality of siRNAs is evaluated for silencing efficacy by base composition PSSMs. In one embodiment, this step comprises calculating one or more PSSM scores for each siRNA. The plurality of siRNAs is then ranked based on the scores and one or more siRNAs are selected using a method described in Section 5.1.4.

In other preferred embodiments, the method of Section 5.2 is used for identifying from among a plurality of siRNAs one or more siRNAs that have high silencing specificity. In one embodiment, alignments of each siRNA with sequences in each of a plurality of non-target transcripts are identified and evaluated with the pmPSSM approach (see Section 5.2.). A pmScore is calculated for each of the alignments. A pmScore above a given threshold identifies a sequence as a potential off-target sequence. Such a pmScore is also termed an alignment score. For example, when FASTA is used for the alignment, a pmScore can be a weighted FASTA alignment score. The transcript that comprises the potential off-target sequence is identified as a potential off-target transcript. The total number of such off-target transcripts in the genome or a portion of the genome is used as a measure of the silencing specificity of the siRNA. One or more siRNAs having less off- target transcripts may then be selected.

The siRNAs having the desired levels of efficacy and specificity for a transcript can be further evaluated for sequence diversity. In this disclosure, sequence diversity is also referred to as "sequence variety" or simply "diversity" or "variety." Sequence diversity can be represented or measured based on some sequence characteristics. The siRNAs can be selected such that a plurality of siRNAs targeting a gene comprises siRNAs exhibiting sufficient difference in one or more of such diversity characteristics.

Preferably the sequence diversity characteristics used in the method of the invention are quantifiable. For example, sequence diversity can be measured based on GC content, the location of the siRNA target sequence along the length of the target transcript,

or the two bases upstream of the siRNA duplex (i.e., the leading dimer, with 16 different possible leading dimers). The difference of two siRNAs can be measured as the difference between values of a sequence diversity measure. The diversity or variety of a plurality of siRNAs can be quantitatively represented by the minimum difference or spacing in a sequence diversity measure between different siRNAs in the plurality.

In the siRNA design method of the invention, the step of selection of siRNAs for diversity or variety is also referred to as a "de-overlap" step. In a preferred embodiment, for a sequence diversity measure that is quantifiable, the de-overlapping selects siRNAs having differences of a sequence diversity measure between two siRNAs above a given threshold. For example, de-overlapping by position establishes a minimum distance between selected oligos along the length of the transcript sequence. In one embodiment, siRNAs positioned at least 100 bases apart in the transcript are selected. De-overlapping by GC content establishes a minimum difference in GC content. In one embodiment, the minimum difference in GC content is 1%, 2% or 5%. De-overlapping by leading dimers establishes the probability of all or a portion of the 16 possible leading dimers among the selected siRNAs. In one embodiment, each of the 16 possible dimers is assigned a score of 1-16, and a 0.5 is used to select all possible leading primer with equal probability.

In some embodiments, the candidates are preferably de-overlapped on GC content, with a minimum spacing of 5%, a maximum number of duplicates of each value of GC% of 100 and at least 200 candidates selected; more preferably they are de-overlapped on GC content with a minimum spacing of 5%, a maximum number of duplicates of each value of GC% of 80 and at least 200 candidates selected; and still more preferably they are de- overlapped on GC content with a minimum spacing of 5%, a maximum number of duplicates of each value of GC% of 60 and at least 200 candidates selected. siRNAs can be further selected based additional selection criteria.

In one embodiment, siRNAs targeting sequences not common to all documented splice forms are eliminated.

In another embodiment, siRNAs targeting sequences overlapping with simple or interspersed repeat elements are eliminated. In still another embodiment, siRNAs targeting sequences positioned at least 75 bases downstream of the translation start codon are selected.

In another embodiment, siRNAs targeting sequences overlapping or downstream of the stop codon are eliminated. This avoids targeting sequences absent in undocumented alternative polyadenylation forms.

In still another embodiment, siRNAs with GC content close to 50% are selected. In one embodiment, siRNAs with GC% < 20% and > 70% are eliminated. In another embodiment, 10% < GC% < 90%, 20% < GC% < 80%, 25% < GC% < 75%, 30% < GC% < 70% are retained. In still another embodiment, siRNAs targeting sequence containing 4 consecutive guanosine, cytosine, adenine or uracil residues are eliminated. In still another embodiment, siRNAs targeting a sequence with a guanine or cytosine residue at the first position in the 19mer duplex region at the 5' end are selected. Such siRNAs target sequences that are effectively transcribed by RNA polymerase III. In still another embodiment, siRNAs targeting a sequence containing recognition sites for one or more given restriction endonucleases, e.g., Xhol or EcoRI restriction endonucleases, are eliminated. This embodiment may be used to select siRNAs sequences for construction of the shRNA vectors.

In still another embodiment, the siRNAs are evaluated for binding energy. See WO 01/05935 for an exemplary method of determining binding energy. In a preferred embodiment, the binding energy is evaluated by calculating the nearest-neighbor 21mer δG.

In still another embodiment, the siRNAs are evaluated for binding specificity. See WO 01/05935 for an exemplary method of determining binding specificity of a 21mer. In a preferred embodiment, the binding specificity is evaluated by calculating a 21mer minimax score against the set of unique sequence representatives of genes of an organism, e.g., the set of unique sequences representatives for each cluster of Homo sapiens Unigene build 161 (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=unigene).

In still another embodiment, the method for predicting strand preference and/or the efficacy and specificity of siRNAs based on position specific base composition of the siRNAs as described in Section 5.3 can be used to evaluate the siRNA candidates.

A flow chart of an exemplary embodiment of the method used to select the siRNAs is shown in FIG. 9.

In step 101, siRNA sequences that target a transcript are selected. In one embodiment, all 19mer subsequences of the transcript are considered. The appropriate flanking sequences for each siRNA sequence are also obtained and considered. The siRNAs are evaluated against the following filters: (1) eliminating siRNAs targeting sequences not common to all documented splice forms; (2) eliminating siRNAs targeting

sequences overlapping with simple or interspersed repeat elements; (3) eliminating siRNAs targeting sequences positioned within 75 bases downstream of the translation start codon; and (4) eliminating siRNAs overlapping or downstream of the stop codon.

For shRNA selection, the following steps are also taken: (5) eliminating siRNAs targeting sequence containing 4 consecutive guanosine, cytosine, adenine or uracil residues; (6) retaining siRNAs targeting a sequence with a guanine residue at the first and/or the second position in the 19mer duplex region at the 5' end; and (7) eliminating siRNAs targeting a sequence containing recognition sites for one or more given restriction enzymes, e.g., Xhol or EcoRI restriction endonucleases, if siRNAs sequences used in construction of the shRNA vectors.

In step 102, the siRNA is evaluated for silencing efficacy by base composition PSSMs. In one embodiment, step 102 comprises calculating a first PSSM score, i.e., the PSSM-I score, and a second PSSM score, i.e., the PSSM-2 score, for an siRNA. The two scores are sum to calculate the combined PSSM- l+PSSM-2 score for the siRNA. In one embodiment, the PSSMs used are those whose performance is shown in Figure 2. The siRNA is retained if the combined score is above a given threshold.

The siRNA is then evaluated for its binding energy by calculating the nearest- neighbor 21mer δG. The siRNA is then evaluated for binding specificity by calculating a 21mer minimax score against the set of unique sequence representatives of genes of an organism, e.g., the set of unique sequences representatives for each cluster of Homo sapiens Unigene build 161. See WO 01/05935 for methods of calculating the δG and the minimax score. In one embodiment, the parameters for the BLAST alignments and nearest-neighbor delta-G calculations based on the BLAST alignments, which are used to compute minimax scores, are as follows: -p blastn -e 100 -F F -W 11 -b 200 -v 10000 -S 3; and delta-G: temperature 66°; salt IM; concentration 1 pM; type of nucleic acid, RNA. In one embodiment, the siRNA is eliminated if the (21mer δG - 21mer minimax) < 0.5. In step 103, siRNAs are screened for overall GC content. In one embodiment, siRNAs with GC content significantly deviated from 50%, e.g., GC% < 20% and > 70%, are eliminated. In step 104, siRNAs are screened for diversity or variety. Position simply refers to the position of the oligo in the transcript sequence and is automatically provided by identifying the oligo. Variety is enforced in one or more "de-overlap" steps in the method. Briefly, de-overlapping selects for above-threshold spacing between selected oligos in

some calculable parameter. To de-overlap, oligos are first ranked according to some parameter thought to distinguish better from poorer performers and then selected for spacing between oligos according to some other parameter. To begin, the top ranked oligo is selected. Then the ranked list is examined, and the next-best oligo with at least the minimum required spacing from the selected oligo is selected. Then the next-best oligo with at least the minimum spacing from the two selected oligos is also selected. The process continues until the desired number of oligos is selected. In one embodiment, multiple oligos may share the same value if a parameter is few- valued, and the number of oligos sharing the same value is limited by a set threshold. In one embodiment, if an insufficient number of oligos is selected in a first pass of de-overlapping, the spacing requirement can be relaxed until the desired number, or the set of all remaining available oligos, is selected.

For example, de-overlapping by position establishes a minimum distance between selected oligos along the length of the transcript sequence. In one embodiment, siRNAs are ranked by a PSSM score and the ranked siRNAs positioned at least 100 bases apart in the transcript are selected. De-overlapping by GC content establishes a minimum difference in GC content, hi one embodiment, the minimum difference in GC content is 1%, 2% or 5%. Duplicates are allowed for few- valued parameters such as the GC% of a 19mer. De-overlapping by leading dimers establishes the probability of all or a portion of the 16 possible leading dimers among the selected siRNAs. In one embodiment, each of the 16 possible dimers is assigned a score of 1-16, and a 0.5 is used to selected all possible leading primer with equal probability, i.e., to distribute candidate siRNAs over all possible leading dimer values.

De-overlapping with different parameters may be combined. In step 105, off-target activity of an siRNA is evaluated according to the method described in Section 5.2. Alignments of each siRNA with sequences in each of a plurality of non-target transcripts are identified and evaluated with a pmPSSM using a pmScore calculated according to equation (6). A pmScore above a given threshold identifies the sequence as a potential off-target sequence. The transcript that comprises the potential off-target sequence is identified as a potential off-target transcript. The total number of such off-target transcripts in the genome or a portion of the genome is used as a measure of the silencing specificity of the siRNA. One or more siRNAs having less off-target transcripts are selected.

In one embodiment, transcripts of genes are scanned using FASTA with the parameters: KTUP 6 -r 3/-7 -g -6 -f -6 -d 14000 -b 14000 -E 7000. A pmScore is determined for each alignment as described in Section 5.2. The FASTA weighted score is used to: (1) quantify the nearest sequence match to the candidate siRNA; and (2) count the total matches to the candidate siRNA with weighted scores over a threshold. The total number of such off-target genes in the genome or a portion of the genome is then used as a measure of the silencing specificity of the siRNA.

In a preferred embodiment, the selected siRNAs are subjected to a second round of selection for variety (step 106), and re-ranked by their base composition PSSM scores (step 107). The desired number of siRNAs is retained from the top of this final ranking (step 108).

The invention also provides a method for selecting a plurality of siRNAs for each of a plurality of different genes, each siRNA achieving at least 75%, at least 80%, or at least 90% silencing of its target gene. The method described above is used to select a plurality of siRNAs for each of a plurality of genes. Preferably, the plurality of siRNAs consists of at least 3, 5, or 10 siRNAs. Preferably, the plurality of different genes consists of at least 100, 500, 1,000, 5,000, 10,000 or 30,000 different genes.

The invention also provides a library of siRNAs which comprises a plurality of siRNAs for each of a plurality of different genes, each siRNA achieves at least 75%, at least 80%, or at least 90% silencing of its target gene. The standard conditions are 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, siRNA, preferably at 100 nM, silencing assayed by TaqMan 24 hours post-transfection. Preferably, the plurality of siRNAs consists of at least 3, at least 5, or at least 10 siRNAs. Preferably, the plurality of different genes consists of at least 10, 100, 500, 1,000, 5,000, 10,000 or 30,000 different genes.

5.5. METHODS AND COMPOSITIONS FOR RNA INTERFERENCE

AND CELL ASSAYS

Any standard method for gene silencing can be used in conjunction with the present invention, e.g., to carry our gene silencing using siRNAs designed by a method described in the present invention (see, e.g. , Guo et al , 1995, Cell 81 :611 -620; Fire et al. , 1998, Nature 391 :806-81 1; Grant, 1999, Cell 96:303-306; Tabara et al, 1999, Cell 99: 123-132; Zamore et al, 2000, Cell 101 :25-33; Bass, 2000, Cell 101 :235-238; Petcherski et al, 2000, Nature 405:364-368; and Elbashir et al, Nature 41 1 :494-498; Paddison et al, Proc. Natl. Acad. Sci. USA 99:1443-1448; each of which is incorporated

by reference herein in its entirety). In one embodiment, gene silencing is induced by presenting the cell with the siRNA, mimicking the product of Dicer cleavage (see, e.g., Elbashir et al, 2001, Nature 411:494-498; and Elbashir et al, 2001, Genes Dev. 15:188- 200, each of which is incorporated by reference herein in its entirety). Synthetic siRNA duplexes maintain the ability to associate with RISC and direct silencing of mRNA transcripts. siRNAs can be chemically synthesized, or derived from cleavage of double- stranded RNA by recombinant Dicer. Cells can be transfected with the siRNA using standard method known in the art.

In one embodiment, siRNA transfection is carried out as follows: one day prior to transfection, 100 microliters of chosen cells, e.g., cervical cancer HeLa cells (ATCC, Cat. No. CCL-2), grown in DMEM/10% fetal bovine serum (Invitrogen, Carlsbad, CA) to approximately 90% confluency are seeded in a 96- well tissue culture plate (Corning, Corning, NY ) at 1500 cells/well. For each transfection 85 microliters of OptiMEM (Invitrogen) is mixed with 5 microliter of serially diluted siRNA (Dharmacon, Denver) from a 20 micro molar stock. For each transfection 5 microliter OptiMEM is mixed with 5 microliter Oligofectamine reagent (Invitrogen) and incubated 5 minutes at room temperature. The 10 microliter OptiMEM/Oligofectamine mixture is dispensed into each tube with the OptiMEM/siRNA mixture, mixed and incubated 15-20 minutes at room temperature. 10 microliter of the transfection mixture is aliquoted into each well of the 96-well plate and incubated for 4 hours at 37°C and 5% CO ₂ .

In one embodiment, RNA interference is carried out using pool of siRNAs. In a preferred embodiment, an siRNA pool containing at least k (k = 2, 3, 4, 5, 6 or 10) different siRNAs targeting a target gene at different sequence regions is used to transfect the cells. In another preferred embodiment, an siRNA pool containing at least k (k = 2, 3, 4, 5, 6 or 10) different siRNAs targeting two or more different target genes is used to supertransfect the cells. In a preferred embodiment, the total siRNA concentration of the pool is about the same as the concentration of a single siRNA when used individually, e.g., 10 μM, lμM, 100 nM, 50 nM, 10 nM or 1 nM, preferably at 100 nM. Preferably, the total concentration of the pool of siRNAs is an optimal concentration for silencing the intended target gene. An optimal concentration is a concentration further increase of which does not increase the level of silencing substantially. In one embodiment, the optimal concentration is a concentration further increase of which does not increase the level of silencing by more than 5%, 10% or 20%. In a preferred embodiment, the composition of the pool, including the number of different siRNAs in the pool and the concentration of

each different siRNA, is chosen such that the pool of siRNAs causes less than 30%, 20%, 10% or 5%, 1%, 0.1% or 0.01% of silencing of any off-target genes. In another preferred embodiment, the concentration of each different siRNA in the pool of different siRNAs is about the same. In still another preferred embodiment, the respective concentrations of different siRNAs in the pool are different from each other by less than 5%, 10%, 20% or 50%. In still another preferred embodiment, at least one siRNA in the pool of different siRNAs constitutes more than 90%, 80%, 70%, 50%, or 20% of the total siRNA concentration in the pool. In still another preferred embodiment, none of the siRNAs in the pool of different siRNAs constitutes more than 90%, 80%, 70%, 50%, or 20% of the total siRNA concentration in the pool. In other embodiments, each siRNA in the pool has an concentration that is lower than the optimal concentration when used individually. In a preferred embodiment, each different siRNA in the pool has an concentration that is lower than the concentration of the siRNA that is effective to achieve at least 30%, 50%, 75%, 80%, 85%, 90% or 95 % silencing when used in the absence of other siRNAs or in the absence of other siRNAs designed to silence the gene. In another preferred embodiment, each different siRNA in the pool has a concentration that causes less than 30%, 20%, 10% or 5% of silencing of the gene when used in the absence of other siRNAs or in the absence of other siRNAs designed to silence the gene. In a preferred embodiment, each siRNA has a concentration that causes less than 30%, 20%, 10% or 5% of silencing of the target gene when used alone, while the plurality of siRNAs causes at least 80% or 90% of silencing of the target gene.

Another method for gene silencing is to introduce into a cell an shRNA, for short hairpin RNA (see, e.g., Paddison et al, 2002, Genes Dev. 16:948-958; Brummelkamp et al, 2002, Science 296:550-553; and Sui, G. et al, 2002, Proc. Natl. Acad. Sci. USA 99:5515-5520, each of which is incorporated by reference herein in its entirety), which can be processed in the cells into siRNA. hi this method, a desired siRNA sequence is expressed from a plasmid (or virus) as an inverted repeat with an intervening loop sequence to form a hairpin structure. The resulting RNA transcript containing the hairpin is subsequently processed by Dicer to produce siRNAs for silencing. Plasmid-based shRNAs can be expressed stably in cells, allowing long-term gene silencing in cells both in vitro and in vivo, e.g., in animals (see, McCaffrey et al., 2002, Nature 418:38-39; Xia et al., 2002, Nat. Biotech. 20:1006-1010; Lewis et al, 2002, Nat. Genetics 32:107-108; Rubinson et al, 2003, Nat: Genetics 33:401-406; Tiscornia et al, 2003, Proc. Natl. Acad.

Sci. USA 100:1844-1848, each of which is incorporated by reference herein in its entirety). Thus, in one embodiment, a plasmid-based shRNA is used.

In a preferred embodiment, shRNAs are expressed from recombinant vectors introduced either transiently or stably integrated into the genome (see, e.g., Paddison et al, 2002, Genes Dev 16:948-958; Sui et al, 2002, Proc Natl Acad Sci USA 99:5515-5520; Yu et al, 2002, Proc Natl Acad Sci USA 99:6047-6052; Miyagishi et al, 2002, Nat Biotechnol 20:497-500; Paul et al, 2002, Nat Biotechnol 20:505-508; Kwak et al, 2003, J Pharmacol Sci 93:214-217; Brummelkamp et al, 2002, Science 296:550-553; Boden et al , 2003, Nucleic Acids Res 31 :5033-5038; Kawasaki et al, 2003, Nucleic Acids Res 31 :700-707; each of which is incorporated by reference herein in its entirety). The siRNA that disrupts the target gene can be expressed (via an shRNA) by any suitable vector which encodes the shRNA. The vector can also encode a marker which can be used for selecting clones in which the vector or a sufficient portion thereof is integrated in the host genome such that the shRNA is expressed. Any standard method known in the art can be used to deliver the vector into the cells. In one embodiment, cells expressing the shRNA are generated by transfecting suitable cells with a plasmid containing the vector. Cells can then be selected by the appropriate marker. Clones are then picked, and tested for knockdown. In a preferred embodiment, a plurality of recombinant vectors is introduced into the genome such that the expression level of the siRNA can be above a given value. Such an embodiment is particular useful for silencing genes whose transcript level is low in the cell.

In a preferred embodiment, the expression of the shRNA is under the control of an inducible promoter such that the silencing of its target gene can be turned on when desired. Inducible expression of an siRNA is particularly useful for targeting essential genes. In one embodiment, the expression of the shRNA is under the control of a regulated promoter that allows tuning of the silencing level of the target gene. This allows screening against cells in which the target gene is partially knocked out. As used herein, a "regulated promoter" refers to a promoter that can be activated when an appropriate inducing agent is present. An "inducing agent" can be any molecule that can be used to activate transcription by activating the regulated promoter. An inducing agent can be, but is not limited to, a peptide or polypeptide, a hormone, or an organic small molecule. An analogue of an inducing agent, i.e., a molecule that activates the regulated promoter as the inducing agent does, can also be used. The level of activity of the regulated promoter induced by different analogues may be different, thus allowing more flexibility in tuning

the activity level of the regulated promoter. The regulated promoter in the vector can be any mammalian transcription regulation system known in the art (see, e.g., Gossen et al,

1995, Science 268:1766-1769; Lucas et al, 1992, Annu. Rev. Biochem. 61 :1131; Li et al,

1996, Cell 85:319-329; Saez et al, 2000, Proc. Natl. Acad. ScL USA 97: 14512-14517; and Pollock et al., 2000, Proc. Natl. Acad Sci. USA 97:13221-13226; each of which is incorporated by reference herein in its entirety). In preferred embodiments, the regulated promoter is regulated in a dosage and/or analogue dependent manner. In one embodiment, the level of activity of the regulated promoter is tuned to a desired level by a method comprising adjusting the concentration of the inducing agent to which the regulated promoter is responsive. The desired level of activity of the regulated promoter, as obtained by applying a particular concentration of the inducing agent, can be determined based on the desired silencing level of the target gene.

In one embodiment, a tetracycline regulated gene expression system is used (see, e.g., Gossen et al., 1995, Science 268:1766-1769; U.S. Patent No. 6,004,941 ; each of which is incorporated by reference herein in its entirety). A tet regulated system utilizes components of the tet repressor/operator/inducer system of prokaryotes to regulate gene expression in eukaryotic cells. Thus, the invention provides methods for using the tet regulatory system for regulating the expression of an shRNA linked to one or more tet operator sequences. The methods involve introducing into a cell a vector encoding a fusion protein that activates transcription. The fusion protein comprises a first polypeptide that binds to a tet operator sequence in the presence of tetracycline or a tetracycline analogue operatively linked to a second polypeptide that activates transcription in cells. By modulating the concentration of a tetracycline, or a tetracycline analogue, expression of the tet operator-linked shRNA is regulated. In other embodiments, an ecdyson regulated gene expression system (see, e.g.,

Saez et al., 2000, Proc. Natl. Acad. Sci. USA 97: 14512-14517, which is incorporated by reference herein in its entirety), or an MMTV glucocorticoid response element regulated gene expression system (see, e.g. , Lucas et al. , 1992, Annu. Rev. Biochem. 61 : 1 131, which is incorporated by reference herein in its entirety) may be used to regulate the expression of the shRNA.

In one embodiment, the pRETRO- SUPER (pRS) vector which encodes a puromycin-resistance marker and drives shRNA expression from an Hl (RNA Pol III) promoter is used. The pRS-shRNA plasmid can be generated by any standard method known in the art. In one embodiment, the pRS-shRNA is deconvoluted from the library

plasmid pool for a chosen gene by transforming bacteria with the pool and looking for clones containing only the plasmid of interest. Preferably, a 19mer siRNA sequence is used along with suitable forward and reverse primers for sequence specific PCR. Plasmids are identified by sequence specific PCR, and confirmed by sequencing. Cells expressing the shRNA are generated by transfecting suitable cells with the pRS-shRNA plasmid. Cells are selected by the appropriate marker, e.g., puromycin, and maintained until colonies are evident. Clones are then picked, and tested for knockdown. In another embodiment, an shRNA is expressed by a plasmid, e.g., a pRS-shRNA. The knockdown by the pRS-shRNA plasmid can be achieved by transfecting cells using Lipofectamine 2000 (Invitrogen).

In yet another method, siRNAs can be delivered to an organ or tissue in an animal, such a human, in vivo (see, e.g., Song et al, 2003, Nat. Medicine 9:347-351; Sorensen et al, 2003, J MoI Biol. 327:761-766; and Lewis et al, 2002, Nat. Genetics 32:107-108, each of which is incorporated by reference herein in their entirety). In this method, a solution of siRNA is injected intravenously into the animal. The siRNA can then reach an organ or tissue of interest and effectively reduce the expression of the target gene in the organ or tissue of the animal.

The siRNAs can also be delivered to an organ or tissue using a gene therapy approach. Any of the methods for gene therapy available in the art can be used to deliver the siRNA. For general reviews of the methods of gene therapy, see Goldspiel et al,

1993, Clinical Pharmacy 12:488-505; Wu and Wu, 1991, Biotherapy 3:87-95; Tolstoshev, 1993, Ann. Rev. Pharmacol. Toxicol. 32:573-596; Mulligan, 1993, Science 260:926-932; and Morgan and Anderson, 1993, Ann. Rev. Biochem. 62:191-217; Robinson, 1993, TIBTECH 11(5):155-215; each of which is incorporated by reference herein in its entirety). In a preferred embodiment, the therapeutic comprises a nucleic acid encoding the siRNA as a part of an expression vector. In particular, such a nucleic acid has a promoter operably linked to the siRNA coding region, in which the promoter being inducible or constitutive, and, optionally, tissue-specific. In another particular embodiment, a nucleic acid molecule in which the siRNA coding sequence is flanked by regions that promote homologous recombination at a desired site in the genome is used (see e.g., Koller and Smithies, 1989, Proc. Natl. Acad. ScL U.S.A. 86:8932-8935; Zijlstra et al, 1989, Nature 342:435-438; each of which is incorporated by reference herein in its entirety).

In a specific embodiment, the nucleic acid is directly administered in vivo. This can be accomplished by any of numerous methods known in the art, e.g., by constructing it as part of an appropriate nucleic acid expression vector and administering it so that it becomes intracellular, e.g., by infection using a defective or attenuated retroviral or other viral vector (see U.S. Patent No. 4,980,286, which is incorporated by reference herein in its entirety), or by direct injection of naked DNA, or by use of microparticle bombardment (e.g., a gene gun; Biolistic, Dupont), or coating with lipids or cell-surface receptors or transfecting agents, encapsulation in liposomes, microparticles, or microcapsules, or by administering it in linkage to a peptide which is known to enter the nucleus, by administering it in linkage to a ligand subject to receptor-mediated endocytosis (see e.g., Wu and Wu, 1987, J. Biol. Chem. 262:4429-4432, which is incorporated by reference herein in its entirety) (which can be used to target cell types specifically expressing the receptors), etc. In another embodiment, a nucleic acid-ligand complex can be formed in which the ligand comprises a fusogenic viral peptide to disrupt endosomes, allowing the nucleic acid to avoid lysosomal degradation. In yet another embodiment, the nucleic acid can be targeted in vivo for cell specific uptake and expression, by targeting a specific receptor (see, e.g., PCT Publications WO 92/06180 dated April 16, 1992 (Wu et al); WO 92/22635 dated December 23, 1992 (Wilson et al); WO92/20316 dated November 26, 1992 (Findeis et al); WO93/14188 dated July 22, 1993 (Clarke et al), WO 93/20221 dated October 14, 1993 (Young)). Alternatively, the nucleic acid can be introduced intracellularly and incorporated within host cell DNA for expression, by homologous recombination (Koller and Smithies, 1989, Proc. Natl Acad. Sci. U.S.A. 86:8932-8935; Zijlstra et al , 1989, Nature 342:435-438). Each of the aforementioned references is incorporated by reference herein in it entirety In a specific embodiment, a viral vector that contains the siRNA coding nucleic acid is used. For example, a retroviral vector can be used (see Miller et al, 1993, Meth. Enzymol. 217: 581-599, which is incorporated by reference herein in its entirety). These retroviral vectors have been modified to delete retroviral sequences that are not necessary for packaging of the viral genome and integration into host cell DNA. The siRNA coding nucleic acid to be used in gene therapy is cloned into the vector, which facilitates delivery of the gene into a patient. More detail about retroviral vectors can be found in Boesen et al., 1994, Biotherapy 6:291-302, which describes the use of a retroviral vector to deliver the mdrl gene to hematopoietic stem cells in order to make the stem cells more resistant to chemotherapy. Other references illustrating the use of retroviral vectors in gene therapy

are: Clowes et al, 1994, J. Clin. Invest. 93:644-651; Kiem et al, 1994, Blood 83:1467- 1473; Salmons and Gunzberg, 1993, Human Gene Therapy 4:129-141 ; and Grossman and Wilson, 1993, Curr. Opin. Genet. and Devel. 3:1 10-114; each of which is incorporated by reference herein in its entirety. Adenoviruses are other viral vectors that can be used in gene therapy.

Adenoviruses are especially attractive vehicles for delivering genes to respiratory epithelia. Adenoviruses naturally infect respiratory epithelia where they cause a mild disease. Other targets for adenovirus-based delivery systems are liver, the central nervous system, endothelial cells, and muscle. Adenoviruses have the advantage of being capable of infecting non-dividing cells. Kozarsky and Wilson (1993, Current Opinion in Genetics and Development 3:499-503, which is incorporated by reference herein in its entirety) present a review of adenovirus-based gene therapy. Bout et al. (1994, Human Gene Therapy 5:3-10) demonstrated the use of adenovirus vectors to transfer genes to the respiratory epithelia of rhesus monkeys. Other instances of the use of adenoviruses in gene therapy can be found in Rosenfeld et al, 1991, Science 252:431-434; Rosenfeld et al , 1992, Cell 68:143-155; and Mastrangeli et α/. , 1993, J. Clin. Invest. 91 :225-234, each of which is incorporated by reference herein in its entirety. Adeno-associated virus (AAV) may also been used in gene therapy (Walsh et al., 1993, Proc. Soc. Exp. Biol. Med. 204:289-300, which is incorporated by reference herein in its entirety). Degree of silencing can be determined using any standard RNA or protein quantification method known in the art. For example, RNA quantification can be performed using Real-time PCR, e.g., using AP Biosystems TaqMan pre-developed assay reagent (#4319442). Primer probe for the appropriate gene can be designed using any standard method known in the art, e.g. using Primer Express software. RNA values can be normalized to RNA for actin (#4326315). Protein levels can be quantified by flow cytometry following staining with appropriate antibody and labeled secondary antibody. Protein levels can also be quantified by western blot of cell lysates with appropriate monoclonal antibodies followed by Kodak image analysis of chemiluminescent immunoblot. Protein levels can also be normalized to actin levels. Effects of gene silencing on a cell can be evaluated by any known assay. For example, cell growth can be assayed using any suitable proliferation or growth inhibition assays known in the art. In a preferred embodiment, an MTT proliferation assay (see, e.g., van de Loosdrechet, et al., 1994, J. Immunol Methods 174:31 1-320; Ohno et al, 1991 , J. Immunol. Methods 145:199-203; Ferrari et al, 199O ₅ J. Immunol Methods 131 : 165-172;

Alley et al, 1988, Cancer Res. 48: 589-601; Carmichael et al, 1987, Cancer Res. 47:936- 942; Gerlier et al, 1986, J. Immunol. Methods 94:57-63; Mosmann, 1983, J. Immunological Methods 65:55-63; each of which is incorporated by reference herein in its entirety) is used to assay the effect of one or more agents in inhibiting the growth of cells. The cells are treated with chosen concentrations of one or more candidate agents for a chosen period of time, e.g. , for 4 to 72 hours. The cells are then incubated with a suitable amount of 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) for a chosen period of time, e.g., 1-8 hours, such that viable cells convert MTT into an intracellular deposit of insoluble formazan. After removing the excess MTT contained in the supernatant, a suitable MTT solvent, e.g., a DMSO solution, is added to dissolve the formazan. The concentration of MTT, which is proportional to the number of viable cells, is then measured by determining the optical density at e.g. , 570 nm. A plurality of different concentrations of the candidate agent can be assayed to allow the determination of the concentrations of the candidate agent or agents which causes 50% inhibition. In another preferred embodiment, an alamarBlue™ Assay for cell proliferation is used to screen for one or more candidate agents that can be used to inhibit the growth of cells (see, e.g., Page et al., 1993, Int. J. Oncol. 3:473-476, which is incorporated by reference herein in its entirety). An alamarBlue™ assay measures cellular respiration and uses it as a measure of the number of living cells. The internal environment of proliferating cells is more reduced than that of non-proliferating cells. For example, the ratios of NADPH/NADP, FADH/FAD, FMNH/FMN, and NADH/NAF increase during proliferation. AlamarBlue can be reduced by these metabolic intermediates and, therefore, can be used to monitor cell proliferation. The cell number of a treated sample as measured by alamarBlue can be expressed in percent relative to that of an untreated control sample. alamarBlue reduction can be measured by either absorption or fluorescence spectroscopy. In one embodiment, the alamarBlue reduction is determined by absorbance and calculated as percent reduced using the equation:

% Re duced = _{χ l QQ} (ε _red λ _λ )(A'λ ₂ )- (ε _red λ ₂ )(A'λ _λ ) where: λi = 570 nm X ₂ = 600 nm (ε _red λj) = 155,677 (Molar extinction coefficient of reduced alamarBlue at 570 nm)

(ε _re d λ ₂ ) = 14,652 (Molar extinction coefficient of reduced alamarBlue at 600 nm) (ε _ox λθ = 80,586 (Molar extinction coefficient of oxidized alamarBlue at 570 nm) (ε _ox λ ₂ ) = 117,216 (Molar extinction coefficient of oxidized alamarBlue at 600 nm) (A = Absorbance of test wells at 570 nm (A λ ₂ ) = Absorbance of test wells at 600 nm

(A ¹ X ₁ ) = Absorbance of negative control wells which contain medium plus alamar Blue but to which no cells have been added at 570 nm.

(A'λ ₂ ) = Absorbance of negative control wells which contain medium plus alamar Blue but to which no cells have been added at 600 nm. Preferably, the % Reduced of wells containing no cell was subtracted from the % Reduced of wells containing samples to determine the % Reduced above background.

Cell cycle analysis can be carried out using standard method known in the art. In one embodiment, the supernatant from each well is combined with the cells that have been harvested by trypsinization. The mixture is then centrifuged at a suitable speed. The cells are then fixed with, e.g., ice cold 70% ethanol for a suitable period of time, e.g., ~ 30 minutes. Fixed cells can be washed once with PBS and resuspended, e.g., in 0.5 ml of PBS containing Propidium Iodide (10 microgram/ml) and RNase A (lmg/ml), and incubated at a suitable temperature, e.g., 37°C, for a suitable period of time, e.g., 30 min. Flow cytometric analysis is then carried out using a flow cytometer. In one embodiment, the Sub-Gl cell population is used as a measure of cell death. For example, the cells are said to have been sensitized to an agent if the Sub-Gl population from the sample treated with the agent is larger than the Sub-Gl population of sample not treated with the agent.

5.6. IMPLEMENTATION SYSTEMS AND METHODS Any of the methods of the present invention can preferably be implemented using an apparatus such as a computer system (of one or more computers), such as the computer system described in this section, according to the following programs and methods. Such a computer system can also preferably store and manipulate measured signals obtained in various experiments that can be used by a computer system implemented with the analytical methods of this invention. Accordingly, such computer systems are also considered part of the present invention.

An exemplary computer system suitable from implementing the analytic methods of this invention is illustrated in FIG. 12. Computer system 1201 is illustrated here as

comprising internal components and as being linked to external components. The internal components of this computer system include one or more processor elements 1202 interconnected with a main memory 1203. For example, computer system 1201 can be an Intel Pentium IV®-based processor of 2 GHZ or greater clock rate and with 256 MB or more main memory. In a preferred embodiment, computer system 1201 is a cluster of a plurality of computers comprising a head "node" and eight sibling "nodes," with each node having a central processing unit ("CPU"). In addition, the cluster also comprises at least 128 MB of random access memory ("RAM") on the head node and at least 256 MB of RAM on each of the eight sibling nodes. Therefore, the computer systems of the present invention are not limited to those consisting of a single memory unit or a single processor unit.

The external components can include a mass storage 1204. This mass storage can be one or more hard disks that are typically packaged together with the processor and memory. Such hard disks are typically of 10 GB or greater storage capacity and more preferably have at least 40 GB of storage capacity. For example, in a preferred embodiment, described above, wherein a computer system of the invention comprises several nodes, each node can have its own hard drive. The head node preferably has a hard drive with at least 10 GB of storage capacity whereas each sibling node preferably has a hard drive with at least 40 GB of storage capacity. A computer system of the invention can further comprise other mass storage units including, for example, one or more floppy drives, one more CD-ROM drives, one or more DVD drives or one or more DAT drives.

Other external components typically include a user interface device 1205, which is most typically a monitor and a keyboard together with a graphical input device 1206 such as a "mouse." The computer system is also typically linked to a network link 1207 which can be, e.g., part of a local area network ("LAN") to other, local computer systems and/or part of a wide area network ("WAN"), such as the Internet, that is connected to other, remote computer systems. For example, in the preferred embodiment, discussed above, wherein the computer system comprises a plurality of nodes, each node is preferably connected to a network, preferably an NFS network, so that the nodes of the computer system communicate with each other and, optionally, with other computer systems by means of the network and can thereby share data and processing tasks with one another.

Loaded into memory during operation of such a computer system are several software components that are also shown schematically in FIG. 12. The software

components comprise both software components that are standard in the art and components that are special to the present invention. These software components are typically stored on mass storage such as the hard drive 1204, but can be stored on other computer readable media as well including, for example, one or more floppy disks, one or more CD-ROMs, one or more DVDs or one or more DATs. Software component 1210 represents an operating system which is responsible for managing the computer system and its network interconnections. The operating system can be, for example, of the Microsoft Windows™ family such as Windows 95, Window 98, Windows NT, Windows 2000 or Windows XP. Alternatively, the operating software can be a Macintosh operating system, a UNIX operating system or a LINUX operating system. Software components 1211 comprises common languages and functions that are preferably present in the system to assist programs implementing methods specific to the present invention. Languages that can be used to program the analytic methods of the invention include, for example, C and C++, FORTRAN, PERL, HTML, JAVA, and any of the UNIX or LINUX shell command languages such as C shell script language. The methods of the invention can also be programmed or modeled in mathematical software packages that allow symbolic entry of equations and high-level specification of processing, including specific algorithms to be used, thereby freeing a user of the need to procedurally program individual equations and algorithms. Such packages include, e.g., Matlab from Mathworks (Natick, MA), Mathematica from Wolfram Research (Champaign, IL) or S-Plus from MathSoft (Seattle, WA).

Software component 1212 comprises any analytic methods of the present invention described supra, preferably programmed in a procedural language or symbolic package. For example, software component 1212 preferably includes programs that cause the processor to implement steps of accepting a plurality of measured signals and storing the measured signals in the memory. For example, the computer system can accept measured signals that are manually entered by a user (e.g., by means of the user interface). More preferably, however, the programs cause the computer system to retrieve measured signals from a database. Such a database can be stored on a mass storage (e.g., a hard drive) or other computer readable medium and loaded into the memory of the computer, or the compendium can be accessed by the computer system by means of the network 1207.

In some embodiments, the present invention can be implemented as a computer program product that comprises a computer program mechanism embedded in a computer readable storage medium. Further, any of the methods of the present invention that don't

involve a measuring step can be implemented in one or more computers or computer systems. Further still, any of the methods of the present invention that don't involve a measuring step can be implemented in one or more computer program products. Some embodiments of the present invention provide a computer system or a computer program product that encodes or has instructions for performing any or all of the methods disclosed herein. Such methods/instructions can be stored on a CD-ROM, DVD, magnetic disk storage product, or any other computer readable data or program storage product. Such methods can also be embedded in permanent storage, such as ROM, one or more programmable chips, or one or more application specific integrated circuits (ASICs). Such permanent storage can be localized in a server, 802.11 access point, 802.11 wireless bridge/station, repeater, router, mobile phone, or other electronic devices. Such methods encoded in the computer program product can also be distributed electronically, via the Internet or otherwise, by transmission of a computer data signal (in which the software modules are embedded) either digitally or on a carrier wave. Some embodiments of the present invention provide a computer program product that contains any or all of the program modules shown in Fig. 12. These program modules can be stored on a CD-ROM, DVD, magnetic disk storage product, or any other computer readable data or program storage product. The program modules can also be embedded in permanent storage, such as ROM, one or more programmable chips, or one or more application specific integrated circuits (ASICs). Such permanent storage can be localized in a server, 802.11 access point, 802.11 wireless bridge/station, repeater, router, mobile phone, or other electronic devices. The software modules in the computer program product can also be distributed electronically, via the Internet or otherwise, by transmission of a computer data signal (in which the software modules are embedded) either digitally or on a carrier wave.

In some embodiments, the computer program products contain programs with instructions for carrying out all or part of any of the methods of the present invention.

In some embodiments, the systems and methods of the present invention further comprises displaying or outputting to a user, to an output medium such as paper or a computer screen, to a computer readable storage medium, or to a remote computer, a result or an indicia of the present invention.

In addition to the exemplary program structures and computer systems described herein, other, alternative program structures and computer systems will be readily apparent to the skilled artisan. Such alternative systems, which do not depart from the above

described computer system and programs structures either in spirit or in scope, are therefore intended to be comprehended within the accompanying claims.

6. EXAMPLES The following examples are presented by way of illustration of the present invention, and are not intended to limit the present invention in any way.

6.1. EXAMPLE 1: DESIGNING SIRNA FOR HIGH SILENCING EFFICACY

A library of siRNAs targeting more than 700 genes was constructed. The siRNAs in the library were designed by use of a "standard" approach, based on a combination of limited design principles available from the scientific literature (Elbashir et al., 2001, Nature 411 :494-8) and a method for predicting off target effects by sequence similarity scoring as described in Section 5.2. A set of 377 siRNAs was tested by Taqman analysis for their ability to silence their respective target genes. The set of 377 siRNAs are listed in Table II. Table II lists the following information for the 377 siRNAs: the ID number of the siRNA, the accession number of the target gene, start position of the target sequence, target sequence, % silencing, the set it belongs (i.e., training or test) in Set 1, the set it belongs in Set 2, and the SEQ ID NO. The results of this test showed that most siRNAs successfully silenced their target genes (median silencing, -75%), but individual siRNAs still showed a wide range of silencing performance. Good (or poor) silencing ability was not consistently associated with any particular base at any position, overall GC content, the position of the siRNA sequence within the target transcript, or with alternative splicing of target transcripts.

The potential relationship between target gene silencing and the base-composition, thermodynamics and secondary structure of the siRNA and target sequences was explored using a classifier approach. siRNAs were divided into groups containing those with less than median silencing ability ("bad" siRNAs) and those with median or better silencing ability ("good" siRNAs). A number of metrics were evaluated for their ability to distinguish good and bad siRNAs, including base composition in windows of the 19mer siRNA duplex sequence and the flanking target region, secondary structure predictions by various programs and thermodynamic properties. These tests revealed that siRNA efficacy correlated well with siRNA and target gene base composition, but poorly with secondary structure predictions and thermodynamic properties. In particular, the GC content of good siRNAs differed substantially from that of bad siRNAs in a position-

specific manner (FIGS. 1-3). For example, good siRNA duplexes were not observed to be associated with any particular sequence, but tended to be GC rich at the 5' end and GC poor at the 3' end. The data indicate that a good siRNA duplex encourages preferential interaction of the antisense strand by being GC poor at its 3' end and discourages interaction of the sense strand by being GC rich at its 5' end. The data further demonstrate that position-specific sequence preferences extend beyond the boundaries of the siRNA target sequence into the adjacent sequence(s). This suggests that steps during RNA silencing other than unwinding of the siRNA duplex are affected by position-specific base composition preferences. The GC-content difference between good and bad siRNAs shown in FIGS. 1 and 2 was used to develop methods for selecting good siRNAs. Best results were obtained with a position-specific scoring matrix (PSSM) approach. The PSSM provides weights for GC, A or U at every position on the sense strand of the target gene sequence from 10 bases upstream of the start to 10 bases downstream of the end of the siRNA duplex. The siRNA efficacy data were divided into two sets, one to be used for training and the other for an independent test. A random-mutation hill-climbing search algorithm was used to optimize the weights for each base at each position of the PSSM simultaneously. The optimization criterion was the correlation coefficient between the target silencing of the siRNA and its PSSM score. Multiple runs of optimization on the training data set were averaged to complete each PSSM. Each PSSM was then tested on the independent (test) set of siRNAs. The performance of two PSSMs on their training and test data sets is demonstrated in Figure 2.

An siRNA design method was developed based on a position-specific score matrix (PSSM). A scoring scheme is used to predict the efficacy of siRNA oligos. The score is a weighted sum of 39 bases (10 bases upstream of the 19mer, 19 bases on the siRNA proper, and 10 bases downstream) computed as follows:

Score = |]ln(E, Ip ₁ )

where P ₁ equals the random probability of any base, i.e., 0.25, and E, the weight assigned to the base A, U, G or C at position /. Therefore, a total of 117 weights (39 positions times 3 base types - G or C, A, U) need to be assigned and optimized.

A random-mutation hill climbing (RMHC) search algorithm was utilized to optimize the weights based on a training oligo set and the resulting profiles applied to a test set, with the optimizing criteria being the correlation coefficient between the knock-

down (KD) levels of the oligos and the computed PSSM scores. The metric to measure the effectiveness of the training and testing is the aggregate false detection rate (FDR) based on the ROC curve, and is computed as the average of the FDR scores of the top 33% oligos sorted by the scores given by the trained predictor. In computing the FDR scores, those oligos with silencing levels less than the median are considered false, and those more than the median silencing levels considered true.

Different criteria were used to divide the existing siRNA performance data into training and test sets. The greatest obstacle to an ideal partition is that the vast majority of siRNA oligos are designed with the standard method, which requires an AA dimer immediately before the 19mer oligo sequence. This limitation was found later to be detrimental rather than helpful to the design process and was abolished. To limit the influence of this on the training procedure, several partitions were used and more than one trained predictors, i.e., PSSMs, (rather than single predictors) were combined to assign scores to the test oligos. Finally, a state-of-the-art siRNA oligo design procedure (also referred to as

"pipeline") was constructed. It incorporates the off-target prediction procedure and two ensembles of siRNA oligo efficacy predictors trained and tested on different data sets. A total of 30 siRNA oligos (6 oligos for each of 5 genes) were selected and tested. The results were significantly better than any of the previously existing pipelines. The initial training and testing results showed that the PSSM is very effective in predicting the on-target efficacy of siRNA oligos. Typically the aggregate FDR scores for training are between 0.02 and 0.08, and those for testing between 0.05 and 0.10. As a reference, random predictions have a mean aggregate FDR of 0.17, with the standard deviation being 0.02 (data computed with 10,000 randomly generated predictions). FIG. 3 illustrates typical ROC curves, generated by an ensemble of about 200 randomly optimized predictors. It could be seen that the performance of the training is better than the test set, which is hardly surprising. Both curves are significantly better than random.

FIG. 5 illustrates the resulting sequence profiles from training and testing on several different oligo sets. This profile illustrates that G or C bases are strongly preferred at the beginning, i.e., the 5' end, and strongly disfavored at the end, i.e., the 3' end, of the 19mer sequence. To confirm this observation, the average knock-down levels for oligos starting and ending with G/C or AAJ are computed, and those oligos starting with G/C and ending with AAJ have the best performance, far superior to the other three categories. Simply by comparing the weights at different positions, a 19mer oligo having a sequence

of GCGTTAATGTGATAATATA (SEQ ID NO: 1), and the oligos that are most similar to this sequence are identified as an siRNA that may have high silencing efficacy.

The design method incorporated both PSSMs shown in FIG. 3 because the combination gave better performance as compared to using either one PSSM alone. The improved siRNA design method selected oligonucleotides based on 4 principles: base composition, off-target identity, position in the transcript, and sequence variety. Certain oligonucleotides containing sequence from features such as untranslated regions, repeats or homopolymeric runs were eliminated. Remaining oligonucleotides were ranked by their PSSM scores. Top-ranking oligonucleotides were selected for variety in GC content, in start position, and in the two bases upstream of the siRNA 19mer duplex. Selected oligonucleotides were then filtered for predicted off-target activity, which was calculated as a position-weighted FASTA alignment score. Remaining oligonucleotides were ranked by PSSM scores, subjected to a second round of selection for variety and finally re-ranked by their PSSM scores. The desired number of siRNAs was retained from the top of this final ranking.

The improved method was compared to the standard method by side-by-side testing of new siRNAs selected by each. The results obtained with three siRNAs selected by each method are shown in Figure 3. siRNAs designed by the improved algorithm showed better median efficacy (88%, compared to 78% for the standard method siRNA) and were more uniform in their performance. The distribution of silencing efficacies of the improved algorithm siRNAs was significantly better than that of the standard method siRNAs for the same genes (p=0.004, Wilcoxon rank sum test).

The test results of 30 experimental oligos using the new pipeline proved to be successful. Table III lists the 30 siRNAs. In the past, an siRNA design with the standard method had a median silencing level of 75%. Of the 30 experimental oligos, 28 had silencing levels equal to or better than 75%, 26 better than or equal to 80%, and 37% better than 90%, comparing with only 10% better than 90% using the standard method. Two target genes (KIF 14 and IGFlR) had been very difficult to silence by siRNAs, with previous oligos achieving only 40% to 70% and no more than 80% silencing levels in the past. The 12 new oligos targeting these genes all achieved silencing of at least 80% and 6 achieved 90% levels. The two oligos among the 30 oligos which had less than 75% silencing level turned out to be targeting an exon that is unique to one target transcript sequence, but absent in all other alternative splice forms of the same gene. Therefore, the failure of these two oligos was due to improper input sequence rather than the PSSM

method. Therefore, when given proper input sequences, the pipeline appears to be able to pick oligos that can knock down target genes by at least 75% for 100% of the target genes. Table II A library of 377 siRNAs accession start % SEQID

BioID number position 19mer sequence silencing Set 1 Set 2 NO.

31 NM_ _. 000075 437 TGTTGTCCGGCTGATGGAC 27.0 Training Training 2

36 NM_ _. 001813 1036 ACTCTTACTGCTCTCCAGT 86.1 Test Training 3

37 NM_ _. 001813 1278 CTTAACACGGATGCTGGTG 60.1 Test Training 4

38 NM_ _. 001813 3427 GGAGAGCTTTCTAGGACCT 88.0 Test Training 5

39 NM_ .004073 192 AGTCATCCCGCAGAGCCGC 55.0 Training Training 6

40 NM_ _. 004073 1745 ATCGTAGTGCTTGTACTTA 70.0 Training Training 7

41 NM_ _. 004073 717 GGAGACGTACCGCTGCATC 65.0 Training Training 8

42 AK092024 437 GCAGTGATTGCTCAGCAGC 93.0 Training Training 9

43 NM_ _. 030932 935 GAGTTTACCGACCACCAAG 81.0 Training Training 10

44 NM_ _. 030932 1186 TGCGGATGCCATTCAGTGG 35.0 Training Training 11

45 NM_ 030932 1620 CACGGTTGGCAGAGTCTAT 73.0 Training Training 12

49 U53530 169 GCAAGTTGAGCTCTACCGC 59.0 Training Training 13

50 U53530 190 TGGCCAGCGCTTACTGGAA 75.0 Training Training 14

64 NM _{. .} 006101 1623 GTTCAAAAGCTGGATGATC 79.0 Test Training 15

65 NM_ 006101 186 GGCCTCTATACCCCTCAAA 74.4 Test Training 16

66 NM _. _006101 968 AGAACCGAATCGTCTAGAG 80.3 Test Training 17

67 NM _. _000859 253 CACGATGCATAGCCATCCT 25.0 Training Training 18

68 NM _000859 1075 CAGAGACAGAATCTACACT 45.0 Training Training 19

69 NM _. _000859 1720 CAACAGAAGGTTGTCTTGT 50.0 Training Training 20

70 NM _. _000859 2572 TTGTGTGTGGGACCGTAAT 80.0 Training Training 21

71 NM _. 000875 276 GCTCACGGTCATTACCGAG 63.9 Training Training 22

72 NM. _000875 441 CCTGAGGAACATTACTCGG 0.0 Training Training 23

73 NM _. 000875 483 TGCTGACCTCTGTTACCTC 50.0 Training Training 24

74 NM _000875 777 CGACACGGCCTGTGTAGCT 58.0 Training Training 25

75 NM _. _000875 987 CGGCAGCCAGAGCATGTAC 63.0 Training Training 26

76 NM _. _000875 1320 CCAGAACTTGCAGCAACTG 70.0 Training Training 27

81 NM _. 000875 351 CCTCACGGTCATCCGCGGC 0.0 Training Training 28

83 NM _. _000875 387 CTACGCCCTGGTCATCTTC 32.0 Training Training 29

84 NM _. _000875 417 TCTCAAGGATATTGGGCTT 54.0 Training Training 30

85 NM _, 000875 423 GGATATTGGGCTTTACAAC 71.0 Training Training 31

86 NM _. _000875 450 CATTACTCGGGGGGCCATC 53.0 Training Training 32

87 NM _000875 481 AATGCTGACCTCTGTTACC 54.6 Training Training 33

117 NM _. _004523 1689 CTGGATCGTAAGAAGGCAG 74.7 Training Test 34

1 18 NM 004523 484 TGGAAGGTGAAAGGTCACC 16.0 Training Test 35

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

119 NM_004523 802 GGACAACTGCAGCTACTCT 84.1 Training Test 36

139 NM_002358 219 TACGGACTCACCTTGCTTG 83.0 Training Training 37

144 NM_001315 779 GTATATACATTCAGCTGAC 78.5 Training 38

145 NM_001315 1080 GGAACACCCCCCGCTTATC 27.2 Training 39

146 NM_001315 1317 GTGGCCGATCCTTATGATC 81.3 Training 40

152 NM_001315 607 ATGTGATTGGTCTGTTGGA 95.0 Training 41

153 NM_OO1315 1395 GTCATCAGCTTTGTGCCAC 92.0 Training 42

154 NM_001315 799 TAATTCACAGGGACCTAAA 82.0 Training 43

155 NM_001315 1277 TGCCTACTTTGCTCAGTAC 95.0 Training 44

193 NM_001315 565 CCTACAGAGAACTGCGGTT 90.0 Training 45

190 NM_001315 763 TTCTCCGAGGTCTAAAGTA 87.0 Training 46

192 NM_001315 1314 CCAGTGGCCGATCCTTATG 89.0 Training 47

194 NM_OO1315 1491 GGCCTTTTCACGGGAACTC 97.0 Training 48

201 NM_016195 2044 CTGAAGAAGCTACTGCTTG 80.3 Test Training 49

202 NM_016195 4053 GACATGCGAATGACACTAG 75.9 Test Training 50

203 NM_016195 3710 AGAGGAACTCTCTGCAAGC 84.7 Test Training 51

204 NM_014875 4478 AAACTGGGAGGCTACTTAC 93.0 Test Training 52

205 NM_014875 1297 ACTGACAACAAAGTGCAGC 37.0 Test Training 53

206 NM_014875 5130 CTCACATTGTCCACCAGGA 91.6 Test Training 54

210 NM_004523 4394 GACCTGTGCCTπTAGAGA 63.7 Training Test 55

211 NM_004523 21 17 GACTTCATTGACAGTGGCC 71.0 Training Test 56

212 NM_004523 799 AAAGGACAACTGCAGCTAC 49.0 Training Test 57

213 NM_000314 2753 TGGAGGGGAATGCTCAGAA 40.0 Training Training 58

214 NM_000314 2510 TAAAGATGGCACTTTCCCG 79.0 Training Training 59

215 NM_000314 2935 AAGGCAGCTAAAGGAAGTG 55.0 Training Training 60

234 NM_007054 963 TATTGGGCCAGCAGATTAC 76.9 Training Training 61

235 NM_007054 593 TTATGACGCTAGGCCACAA 74.4 Training Training 62

236 NM_007054 1926 GGAGAAAGATCCCTTTGAG 78.3 Training Training 63

237 NM_006845 324 ACAAAAACGGAGATCCGTC 72.2 Training Training 64

238 NM_006845 2206 ATAAGCAGCAAGAAACGGC 30.9 Training Training 65

239 NM_006845 766 GAATTTCGGGCTACTTTGG 65.8 Training Training 66

240 NM_005163 454 CGCACCTTCCATGTGGAGA 86.8 Training Training 67

241 NM_005163 1777 AGACG I I I I I GTGCTGTGG 76.0 Training Training 68

242 NM_005163 1026 GCTGGAGAACCTCATGCTG 87.8 Training Training 69

243 NM_005733 2139 CrCTACCACTGAAGAGTTG 90.7 Training Training 70

244 NM_005733 1 106 AAGTGGGTCGTAAGAACCA 82.5 Training Training 71

245 NM_005733 696 GAAGCTGTCCCTGCTAAAT 93.4 Training Training 72

246 NM 001813 3928 GAAGAGATCCCAGTGCTTC 86.8 Test Training 73

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

247 NM_001813 4456 TCTGAAAGTGACCAGCTCA 82.5 Test Training 74

248 NM 001813 2293 GAAAATGAAGCTTTGCGGG 78.4 Test Training 75

249 NM_005030 1135 AAGAAGAACCAGTGGTTCG 83.0 Test Test 76

250 NM_005030 572 CCGAGTTATTCATCGAGAC 93.6 Test Test 77

251 NM_005030 832 AAGAGACCTACCTCCGGAT 85.0 Test Test 78

255 NM_001315 3050 AATATCCTCAGGGGTGGAG 36.0 Training 79

256 NM_001315 1526 GTGCCTCTTGTTGCAGAGA 88.0 Training 80

257 NM_001315 521 GAAGCTCTCCAGACCATTT 96.0 Training 81

261 NM_006218 456 AGAAGCTGTGGATCTTAGG 65.3 Test Training 82

262 NM_006218 3144 TGATGCACATCATGGTGGC 68.9 Test Training 83

263 NM_006218 2293 CTAGGAAACCTCAGGCTTA 94.7 Test Training 84

264 NM_000075 1073 GCGAATCTCTGCCTTTCGA 79.0 Training Training 85

265 NM_000075 685 CAGTCAAGCTGGCTGACTT 78.0 Training Training 86

266 NM_000075 581 GGATCTGATGCGCCAGTTT 77.0 Training Training 87

288 NM_020242 1829 GCACAACTCCTGCAAATTC 87.4 Training Training 88

289 NM_020242 3566 GATGGAAGAGCCTCTAAGA 82.7 Training Training 89

290 NM_020242 2631 ACGAAAAGCTGCTTGAGAG 73.4 Training Training 90

291 NM_004073 570 GAAGACCATCTGTGGCACC 65.0 Training Training 91

292 NM_004073 1977 TCAGGGACCAGCTTTACTG 60.0 Training Training 92

293 NM_004073 958 GTTACCAAGAGCCTCTTTG 75.0 Training Training 93

294 NM_005026 3279 AACCAAAGTGAACTGGCTG 56.3 Training Training 94

295 NM_005026 2121 GATCGGCCACTTCCTTTTC 70.9 Training Training 95

296 NM_005026 4004 AGAGATCTGGGCCTCATGT 67.3 Training Training 96

303 NM_000051 5373 AGTTCGATCAGCAGCTGTT 60.9 Training Training 97

304 NM_000051 3471 TAGATTGTTCCAGGACACG 71.2 Training Training 98

305 NMJJ00051 7140 GAAGTTGGATGCCAGCTGT 56.3 Training Training 99

309 NM_004064 1755 TGGTGATCACTCCAGGTAG 25.3 Training Training 100

310 NM_004064 1505 TGTCCCTTTCAGAGACAGC 5.0 Training Training 101

311 NM_004064 1049 GACGTCAAACGTAAACAGC 50.2 Training Training 102

312 NM_006219 1049 AAGTTCATGTCAGGGCTGG 76.6 Test Training 103

313 NM_006219 2631 CAAAGATGCCCTTCTGAAC 88.9 Test Training 104

314 NM_006219 453 AATGCGCAAATTCAGCGAG 32.9 Test Training 105

339 NM_003600 437 GCACAAAAGCTTGTCTCCA 96.0 Test Training 106

340 NM_003600 1071 TTGCAGATπTGGGTGGTC 37.0 Test Training 107

341 NM_003600 1459 ACAGTCTTAGGAATCGTGC 61.1 Test Training 108

342 NM_004958 1476 AGGACTTCGCCCATAAGAG 61.8 Test Training 109

343 NM 004958 5773 CAACCTCCAGGATACACTC 80.9 Test Training 1 10

344 NM 004958 7886 CCAACTTTCTAGCTGCTGT 71.1 Test Training 1 1 1

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

348 NM_004856 1999 GAATGTGAGCGTAGAGTGG 92.2 Training Training 112

349 NM_004856 1516 CCATTGGTTACTGACGTGG 87.7 Training Training 1 13

350 NM_004856 845 AACCCAAACCTCCACAATC 71.8 Training Training 114

369 XM_294563 1 17 GAAAGAAGCAGTTGACCTC 59.9 Training Training 115

370 XM_294563 2006 CTAAAAGCTGGGTGGACTC 69.4 Training Training 116

371 XM_294563 389 GAAAGCACCTCTTTGTGTG 64.2 Training Training 117

399 NM_000546 1286 TGAGGCCTTGGAACTCAAG 17.8 118

400 NM_000546 2066 CCTCTTGGTCGACCTTAGT 74.5 119

401 NM_000546 1546 GCACCCAGGACTTCCATTT 93.2 120

417 NM_001184 3790 GAAACTGCAGCTATCTTCC 75.8 Training Training 121

418 NM_001184 7717 GTTACAATGAGGCTGATGC 73.0 Training Training 122

419 NMJ)Ol 184 5953 TCACGACTCGCTGAACTGT 78.8 Training Training 123

453 NMJ)05978 323 GACCGACCCTGAAGCAGAA 91.3 Test Test 124

454 NMJ)05978 254 TTCCAGGAGTATGCTGTTT 74.4 Test Test 125

455 NMJ)05978 145 GGAACTTCTGCACAAGGAG 96.5 Test Test 126

465 NM_000551 495 TGTTGACGGACAGCCTATT 75.5 Test Training 127

466 NMJ)00551 1056 GGCATTGGCATCTGCTTTT 89.7 Test Training 128

467 NMJ)00551 3147 GTGAATGAGACACTCCAGT 82.2 Test Training 129

468 NM_002658 1944 GAGCTGGTGTCTGATTGTT 82.8 Test Training 130

469 NMJ)02658 1765 GTGTAAGCAGCTGAGGTCT 44.4 Test Training 131

470 NMJ)02658 232 CTGCCCAAAGAAATTCGGA 47.8 Test Training 132

507 NMJ)03391 792 ATTTGCCCGCGCATTTGTG 27.2 Test Training 133

508 NMJ)03391 2171 AGAAGATGAATGGTCTGGC 69.4 Test Training 134

509 NMJ)03391 981 AACGGGCGATTATCTCTGG 43.3 Test Training 135

540 NMJ)02387 3490 GACTTAGAGCTGGGAATCT 83.7 Test Training 136

541 NM_002387 4098 AGTTGAGGAGGTTTCTGCA 86.1 Test Training 137

542 NMJ)02387 1930 GGATTATATCCAGCAGCTC 82.3 Test Training 138

585 NMJH4885 509 GTGGCTGGATTCATGTTCC 81.5 Training Training 139

586 NMJ) 14885 798 CAAGGCATCCGTTATATCT 84.7 Training Training 140

587 NMJ) 14885 270 ACCAGGATTTGGAGTGGAT 84.7 Training Training 141

639 NMJ)01274 250 CTGAAGAAGCAGTCGCAGT 77.7 142

640 NMJ)01274 858 ATCGATTCTGCTCCTCTAG 86.2 143

641 NMJ)01274 1332 TGCCTGAAAGAGACTTGTG 85.4 144

651 NMJ)01259 807 TCTTGGACGTGATTGGACT 89.8 Training Training 145

652 NMJ)01259 1036 AGAAAACCTGGATTCCCAC 88.9 Training Training 146

653 NMJ)01259 556 ACCACAGAACATTCTGGTG 89.3 Training Training 147

672 NMJ)03161 2211 GAAAGCCAGACAACTTCTG 87.1 Test Training 148

673 NM 003161 1223 CTCTCAGTGAAAGTGCCAA 91.2 Test Training 149

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

674 NM_003161 604 GACACTGCCTGCTTTTACT 98.1 Test Training 150

678 NM_004972 3526 AAGAACCTGGTGAAAGTCC 57.2 Training Training 151

679 NM_004972 4877 GAAGTGCAGCAGGTTAAGA 54.8 Training Training 152

680 NM_004972 1509 AGCCGAGTTGTAACTATCC 74.9 Training Training 153

684 NM_007194 1245 GATCACAGTGGCAATGGAA 80.9 154

685 NM 007194 1432 AAACTCTTGGAAGTGGTGC 39.2 155

686 NM_007194 2269 ATGAATCCACAGCTCTACC 44.6 156

687 NM_007313 3866 GAATGGAAGCCTGAACTGA 92.4 Test Training 157

688 NM_007313 2451 AGACATCATGGAGTCCAGC 5.0 Test Training 158

689 NM_007313 1296 CAAGTTCTCCATCAAGTCC 91.1 Test Training 159

711 NM_139049 129 GGAATAGTATGCGCAGCTT 92.5 Test Training 160

712 NM_139049 369 GTGATTCAGATGGAGCTAG 89.0 Test Training 161

713 NM_139049 969 CACCCGTACATCAATGTCT 77.0 Test Training 162

858 NMJ)01253 522 TCATTGGAAGAACAGCGGC 0.0 Test Training 163

859 NMJ)01253 2571 AAGAAGACGTTCAGCGACA 93.5 Test Training 164

860 NM_001253 911 AAAAAGCCTGCCCTTGGTT 88.1 Test Training 165

1110 NMJ)06101 1847 CTTGCAACGTCTGTTAGAG 72.3 Test Training 166

1 11 1 NMJ)06101 999 CTGAAGGCTTCCTTACAAG 82.9 Test Training 167

1 112 NM_006101 1278 CAGAAGTTGTGGAATGAGG 79.1 Test Training 168

1182 NMJ) 16231 1302 GCAATGAGGACAGCTTGTG 79.8 Test Training 169

1183 NMJ) 16231 1829 TGTAGCTTTCCACTGGAGT 79.3 Test Training 170

1184 NMJ) 16231 1019 TCTCCTTGTGAACAGCAAC 62.5 Test Training 171

1212 NMJ)01654 1072 AGTGAAGAACCTGGGGTAC 79.3 Test Training 172

1213 NMJ)01654 595 GTTCCACCAGCATTGTTCC 86.2 Test Training 173

1214 NMJ)01654 1258 GAATGAGATGCAGGTGCTC 86.9 Test Training 174

1287 NMJ)05417 2425 CAATTCGTCGGAGGCATCA 73.9 Test Training 175

1288 NMJJ05417 1077 GGGGAGTTTGCTGGACTTT 66.4 Test Training 176

1289 NMJJ05417 3338 GCAGTGCCTGCCTATGAAA 68.2 Test Training 177

1290 NMJ)01982 3223 CTAGACCTAGACCTAGACT 63.5 Test Training 178

1291 NMJ)01982 3658 GAGGATGTCAACGGTTATG 49.4 Test Training 179

1292 NM OO 1982 2289 CAAAGTCTTGGCCAGAATC 45.3 Test Training 180

1293 NMJJ05400 249 GATCGAGCTGGCTGTCTTT 85.4 Test Training 181

1294 NMJ)05400 1326 GGTCTTAAAGAAGGACGTC 63.4 Test Training 182

1295 NMJ)05400 1848 TGAGGACGACCTATTTGAG 0.0 Test Training 183

1317 NMJJ02086 465 TGAGCTGGTGGATTATCAC 85.5 Test Test 184

1318 NMJ ⁾ 02086 183 CTGGTACAAGGCAGAGCTT 95.5 Test Test 185

1319 NMJ ⁾ 02086 720 CCGGAACGTCTAAGAGTCA 92.3 Test Test 186

1332 NM 006219 2925 TACAGAAAAGTTTGGCCGG 20.1 Test Training 187

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

1333 NM_006219 2346 AATGAAGCCTTTGTGGCTG 22.4 Test Training 188

1334 NM_006219 2044 GTGCACATTCCTGCTGTCT 79.0 Test Training 189

1335 NM_003600 1618 CCTCCCTATTCAGAAAGCT 84.2 Test Training 190

1336 NM_003600 650 GACTTTGAAATTGGTCGCC 52.1 Test Training 191

1337 NM_003600 538 CACCCAAAAGAGCAAGCAG 96.3 Test Training 192

1338 XM_294563 2703 TAAGCCTGGTGGTGATCTT 78.1 Training Training 193

1339 XM_294563 1701 AAGGTCTTTACGCCAGTAC 29.5 Training Training 194

1340 XM_294563 789 GGAATGTATCCGAGCACTG 73.5 Training Training 195

1386 NM_033360 493 GGACTCTGAAGATGTACCT 91.0 Test Training 196

1387 NMJ333360 897 GGCATACTAGTACAAGTGG 84.8 Test Training 197

1388 NM_033360 704 GAAAAGACTCCTGGCTGTG 0.0 Test Training 198

1389 NM_024408 4735 CTTTGAATGCCAGGGGAAC 91.6 Test Training 199

1390 NM_024408 2674 CCAAGGAACCTGCTTTGAT 96.4 Test Training 200

1391 NM_024408 5159 GACTCAGACCACTGCTTCA 95.8 Test Training 201

1392 NM_000435 6045 GCTGCTGTTGGACCACTTT 0.0 Test Training 202

1393 NM_000435 5495 TGCCAACTGAAGAGGATGA 0.0 Test Training 203

1394 NM_000435 4869 TGATCACTGCTTCCCCGAT 0.0 Test Training 204

1410 AF308602 770 ATATCGACGATTGTCCAGG 36.7 Test Training 205

1411 AF308602 3939 AGGCAAGCCCTGCAAGAAT 81.3 Test Training 206

1412 AF308602 1644 CACTTACACCTGTGTGTGC 81.3 Test Training 207

1581 NM 005633 3593 TATCAGACCGGACCTCTAT 70.8 Test Training 208

1582 NM_005633 364 ATTGACCACCAGGTTTCTG 1.4 Test Training 209

1583 NM_005633 3926 CTTACAAAAGGGAGCACAC 66.9 Test Training 210

1620 NM_002388 1097 GTCTCAGCTTCTGCGGTAT 95.0 Test Training 21 1

1621 NM_002388 286 AGGATTTTGTGGCCTCCAT 94.6 Test Training 212

1622 NM_002388 2268 TCCAGGTTGAAGGCATTCA 92.5 Test Training 213

1629 NM_012193 3191 TTGGCAAAGGCTCCTTGTA 80.0 Test Test 214

1630 NM_012193 5335 CCATCTGCTTGAGCTACTT 85.0 Test Test 215

1631 NM_012193 2781 GTTGACTTACCTGACGGAC 43.1 Test Test 216

1632 NM_004380 3708 GACATCCCGAGTCTATAAG 85.3 Test Training 217

1633 NM_004380 339 TGGAGGAGAATTAGGCCTT 81.1 Test Training 218

1634 NM_004380 5079 GCACAAGGAGGTCTTCTTC 79.0 Test Training 219

1641 NMJ)17412 2331 CAGATCACTCCAGGCATAG 97.3 Test Training 220

1643 NM 017412 2783 ATGTGTGGTGACTGCTTTG 95.7 Test Training 221

1695 NM OO 1903 2137 TGACATCATTGTGCTGGCC 38.4 Test Training 222

1696 NM_001903 655 CGTTCCGATCCTCTATACT 97.9 Test Training 223

1697 NM OO 1903 31 17 TGACCAAAGATGACCTGTG 40.1 Test Training 224

1815 NM 020168 3064 GAGAAAGAATGGGGTCGGT 85.0 Training Training 225

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

1816 NM_020168 681 CGACATCCAGAAGTTGTCA 86.1 Training Training 226

1817 NM_020168 1917 TGAGGAGCAGATTGCCACT 72.1 Training Training 227

2502 NM_000271 237 GAGGTACAATTGCGAATAT 87.0 Training Training 228

2503 NM_000271 559 TACTACGTCGGACAGAGTT 76.0 Training Training 229

2504 NM_000271 1783 AACTACAATAACGCCACTG 39.0 Training Training 230

2505 NM_000271 2976 GCCACAGTCGTCTTGCTGT 84.0 Training Training 231

2512 NM_005030 245 GGGCGGCTTTGCCAAGTGC 88.6 Test Test 232

2513 NM_005030 1381 CACGCCTCATCCTCTACAA 90.5 Test Test 233

2514 NM_005030 834 GAGACCTACCTCCGGATCA 91.0 Test Test 234

2521 NM_000314 1316 CCCACCACAGCTAGAACTT 93.0 Training Training 235

2522 NM_000314 1534 CTATTCCCAGTCAGAGGCG 89.0 Training Training 236

2523 NM_000314 2083 CAGTAGAGGAGCCGTCAAA 90.0 Training Training 237

2524 NM_006622 1928 CAGTTCACTATTACGCAGA 65.0 Training Training 238

2525 NM_006622 586 TGTTACGAGATGACAGATT 73.0 Training Training 239

2526 NM_006622 1252 AACCCAGAGGATCGTCCCA 70.0 Training Training 240

2527 NM_139164 200 CTGTTTGGAGAAAACCCTC 79.0 Training Training 241

2528 NMJ39164 568 GACAACCCAAACCAGAGTC 71.0 Training Training 242

2529 NMJ39164 488 GTCTTGACTGGGATGAAAA 66.0 Training Training 243

2530 NM_139164 578 ACCAGAGTCTTTTGACAGG 82.0 Training Training 244

2546 NM_014875 1090 TAGACCACCCATTGCTTCC 63.5 Test Training 245

2547 NM_014875 1739 AGAGCCTTCGAAGGCTTCA 73.2 Test Training 246

2548 NM_014875 3563 GACCATAGCATCCGCCATG 87.1 Test Training 247

2602 NM_002387 2655 TAGCTCTGCTAGAGGAGGA 71.0 Test Training 248

2603 NM_002387 1418 ACAGAACGGCTGAATAGCC 43.5 Test Training 249

2604 NM 002387 941 GAGAATGAGAGCCTGACTG 81.0 Test Training 250

2605 NM_016231 1683 GGAAACAGAGTGCCTCTCT 55.3 Test Training 251

2606 NM_016231 915 CCACTCAGCTCAGATCATG 82.3 Test Training 252

2607 NM_016231 111 TCTGGTCTCTTGCAAAAGG 30.3 Test Training 253

261 1 NM_004380 4230 A I I I I I GCGGCGCCAGAAT 79.0 Test Training 254

2612 NM_004380 2197 GAAAAACGGAGGTCGCGTT 85.9 Test Training 255

2613 NM_004380 5701 GAAAACAAATGCCCCGTGC 55.4 Test Training 256

2614 NM_005978 276 TGGCACTCATCACTGTCAT 91.8 Test Test 257

2615 NM_005978 229 TGAGAACAGTGACCAGCAG 91.9 Test Test 258

2616 NM_005978 369 GGGCCCAGGACTGTTGATG 94.5 Test Test 259

2617 NM_017412 3128 AGAGATGGGCATTGTTTCC 94.3 Test Training 260

2618 NM_017412 814 GCTCATGGAGATGTTTGGT 88.7 Test Training 261

2619 NM_017412 1459 AGCATTGCTGTTTCACGCC 93.1 Test Training 262

2620 NM 001654 1902 TTGAGCTGCTGCAACGGTC 67.2 Test Training 263

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

2621 NM_001654 1006 GTCCCCACATTCCAAGTCA 90.0 Test Training 264

2622 NM_001654 2327 CCTCTCTGGAATTTGTGCC 85.7 Test Training 265

2623 NMJ)02658 202 CAAGTACTTCTCCAACATT 87.2 Test Training 266

2624 NMJ)02658 181 TGGAGGAACATGTGTGTCC 0.0 Test Training 267

2625 NMJ)02658 436 TTACTGCAGGAACCCAGAC 0.0 Test Training 268

2629 NMJ)06218 1334 TGGCTTTGAATCTTTGGCC 3.5 Test Training 269

2630 NMJ)06218 2613 AGGTGCACTGCAGTTCAAC 53.8 Test Training 270

2631 NMJ)06218 1910 TTCAGCTAGTACAGGTCCT 78.0 Test Training 271

2632 NMJ)03161 1834 TTGATTCCTCGCGACATCT 88.3 Test Training 272

2633 NMJ)03161 1555 GCTTTTCCCATGATCTCCA 90.7 Test Training 273

2634 NM_003161 217 CTTGGCATGGAACATTGTG 61.4 Test Training 274

2635 NMJ)03391 2072 GCCTCAGAAAGGGATTGCT 79.1 Test Training 275

2636 NMJ)03391 1318 GCTCTGGATGTGCACACAT 60.5 Test Training 276

2637 NM_003391 1734 GTGTCTCAAAGGAGCTTTC 87.1 Test Training 277

2641 AF308602 4260 ATTCAACGGGCTCTTGTGC 0.0 Test Training 278

2642 AF308602 1974 GATCGATGGCTACGAGTGT 84.0 Test Training 279

2643 AF308602 5142 CATCCCCTACAAGATCGAG 41.6 Test Training 280

2644 NM_024408 8232 GCAACTTTGGTCTCCTTTC 91.0 Test Training 281

2645 NM_024408 10503 GCAATTGGCTGTGATGCTC 86.6 Test Training 282

2646 NM_024408 8643 GAGACAAGTTAACTCGTGC 89.4 Test Training 283

2647 NM_007313 4222 TCCTGGCAAGAAAGCTTGA 65.6 Test Training 284

2648 NM_007313 3237 AAACCTCTACACGTTCTGC 53.5 Test Training 285

2649 NM_007313 302 CTAAAGGTGAAAAGCTCCG 67.8 Test Training 286

2650 NM_000551 631 GATCTGGAAGACCACCCAA 70.9 Test Training 287

2651 NM_000551 4678 CAGAACCCAAAAGGGTAAG 0.0 Test Training 288

2652 NM_000551 4382 AGGAAATAGGCAGGGTGTG 4.3 Test Training 289

2653 NMJ)01903 1888 AGCAGTGCTGATGATAAGG 89.1 Test Training 290

2654 NMJ)01903 2606 AAGCCATTGGTGAAGAGAG 91.9 Test Training 291

2655 NMJ)01903 1583 TGTGTCATTGCTCTCCAAG 90.3 Test Training 292

2656 NM_002388 842 GCAGATGAGCAAGGATGCT 86.8 Test Training 293

2657 NM_002388 1754 GTACATCCATGTGGCCAAA 94.6 Test Training 294

2658 NM 002388 2642 TGGGTCATGAAAGCTGCCA 93.1 Test Training 295

2662 NM_005633 3251 GAACACCGTTAACACCTCC 31.2 Test Training 296

2663 NM_005633 2899 ATAACAGGAGAGATCCAGC 21.7 Test Training 297

2664 NMJJ05633 2607 TGGTGTCCTTGAGGTTGTC 75.1 Test Training 298

2665 NM_033360 329 ACCTGTCTCTTGGATATTC 81.4 Test Training 299

2666 NMJ333360 529 TAAATGTGATTTGCCTTCT 47.8 Test Training 300

2667 NM 033360 585 GAAGTTATGGAATTCCTTT 94.2 Test Training 301

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

2668 NM_139049 745 CACCATGTCCTGAATTCAT 80.7 Test Training 302

2669 NM l 39049 433 TCAAGCACCTTCATTCTGC 42.6 Test Training 303

2670 NM l 39049 550 CGAGTπTATGATGACGCC 79.9 Test Training 304

2671 NM_002086 555 ATACGTCCAGGCCCTCTTT 87.9 Test Test 305

2672 NM_002086 392 TGCAGCACTTCAAGGTGCT 36.9 Test Test 306

2673 NM_002086 675 CGGGCAGACCGGCATGTTT 92.6 Test Test 307

2674 NM_004958 5024 GACATGAGAACCTGGCTCA 77.8 Test Training 308

2675 NM_004958 2155 CTTGCAGGCCTTGTTTGTG 83.2 Test Training 309

2676 NM_004958 6955 TAATACAGCTGGGGACGAC 52.3 Test Training 310

2677 NM_012193 467 AGAACCTCGGCTACAACGT 71.5 Test Test 311

2678 NM_012193 473 TCGGCTACAACGTGACCAA 51.3 Test Test 312

2679 NM_012193 449 TCCGCATCTCCATGTGCCA 37.5 Test Test 313

2680 NM 005400 665 TCACAAAGTGTGCTGGGTT 43.9 Test Training 314

2681 NM_005400 2178 CCAGGAGGAATTCAAAGGT 41.6 Test Training 315

2682 NM_005400 1022 GCTCACCATCTGAGGAAGA 64.2 Test Training 316

2686 NMJ)01982 948 TGACAGTGGAGCCTGTGTA 65.8 Test Training 317

2687 NMJ)01982 1800 CTTTCTGAATGGGGAGCCT 61.7 Test Training 318

2688 NMJ)01982 2860 TACACACACCAGAGTGATG 0.0 Test Training 319

2692 NMJ) 16195 5331 ATGAAGGAGAGTGATCACC 10.5 Test Training 320

2693 NMJ)16195 4829 AATGGCAGTGAAACACCCT 67.3 Test Training 321

2694 NMJ) 16195 1480 AAGTTTGTGTCCCAGACAC 80.5 Test Training 322

2695 NMJ)00435 2107 AATGGCTTCCGCTGCCTCT 0.0 Test Training 323

2696 NMJ)00435 5193 GAACATGGCCAAGGGTGAG 15.5 Test Training 324

2697 NMJ)00435 7273 GAGTCTGGGACCTCCTTCT 0.0 Test Training 325

2802 NMJ)04523 46 CCAGGGAGACTCCGGCCCC 6.7 Training Test 326

2803 NMJ)04523 132 GGGACCGTCATGGCGTCGC 8.2 Training Test 327

2804 NMJ)04523 221 ATTTAATTTGGCAGAGCGG 0.0 Training Test 328

2805 NM 004523 322 GCTCAAGGAAAACATACAC 76.2 Training Test 329

2806 NMJ)04523 365 TACTAAACAGATTGATGTT 77.9 Training Test 330

2807 NMJ)04523 581 TACTGATAATGGTACTGAA 93.8 Training Test 331

2808 NMJ)04523 716 AGGAGTGATAATTAAAGGT 84.8 Training Test 332

2809 NMJ)04523 852 GTTTTCTCTGTTACAATAC 85.4 Training Test 333

2810 NMJJ04523 995 TGGAAATATAAATCAATCC 0.0 Training Test 334

281 1 NMJ)04523 1085 ACTAACTAGAATCCTCCAG 0.0 Training Test 335

2812 NMJ)04523 1174 AAACTCTGAGTACATTGGA 81.9 Training Test 336

2813 NM 004523 1375 TAACTGTTCAAGAAGAGCA 14.1 Training Test 337

2814 NM_004523 1570 AAGAAGAATATATCACATC 0.0 Training Test 338

2815 NM 004523 1706 AGTTGACCAACACAATGCA 86.0 Training Test 339

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

2816 NM_ _. 004523 2197 TACATGAACTACAAGAAAA 90.0 Training Test 340

2817 NM_ _004523 2858 GACTAAGCTTAATTGCTTT 87.0 Training Test 341

2818 NM_ _004523 3089 GGGGCAGTATACTGAAGAA 64.5 Training Test 342

2819 NM_ _„ 004523 3878 TTCTTGTATATTATTAAGT 0.0 Training Test 343

2820 NM_ _004523 4455 TCTATAATTTATATTCTTT 9.3 Training Test 344

2821 NM_ _004523 4648 TACAAAGAATAAATTTTCT 23.5 Training Test 345

2823 NM_ _005030 45 CAGCGCAGCTTCGGGAGCA 72.1 Training Test 346

2824 NM_ _005030 131 CGGAGTTGCAGCTCCCGGA 85.7 Training Test 347

2825 NM_ _005030 303 GGCAAGATTGTGCCTAAGT 80.1 Training Test 348

2826 NM_ _005030 346 GGGAGAAGATGTCCATGGA 100.0 Training Test 349

2827 NM_ _005030 432 GACTTCGTGTTCGTGGTGT 89.3 Training Test 350

2828 NM _. 005030 519 GCCCGATACTACCTACGGC 86.2 Training Test 351

2829 NM_ _005030 648 GGACTGGCAACCAAAGTCG 86.7 Training Test 352

2830 NM_ 005030 777 TGTATCATGTATACCTTGT 84.3 Training Test 353

2831 NM_ _005030 821 TTCTTGCCTAAAAGAGACC 26.8 Training Test 354

2832 NM _. 005030 907 TCCAGAAGATGCTTCAGAC 90.8 Training Test 355

2833 NM _. _005030 952 ACGAGCTGCTTAATGACGA 87.7 Training Test 356

2834 NM _. _005030 1038 TCGATTGCTCCCAGCAGCC 31.4 Training Test 357

2835 NM _005030 1082 CACAGTCCTCAATAAAGGC 62.9 Training Test 358

2836 NM _. _OO5O3O 1214 CAATGCCTCCAAGCCCTCG 0.0 Training Test 359

2837 NM _. _005030 1300 AGTGGGTGGACTATTCGGA 84.9 Training Test 360

2838 NM _. _OO5O3O 1515 TACATGAGCGAGCACTTGC 20.3 Training Test 361

2839 NM _005030 1860 CTCAAGGCCTCCTAATAGC 74.2 Training Test 362

2840 NM. _005030 1946 CCGCGGTGCCATGTCTGCA 79.7 Training Test 363

2841 NM _. 005030 2075 CCCCTCCCCCTCAACCCCA 34.6 Training Test 364

3041 NM _. _014875 4629 ATTTTCTAGAAAACGGTAA 91.8 365

3042 NM _. 014875 77 GAGGGGCGAAGTTTCGGCA 71.2 366

3043 NM _. _014875 243 CTGGGACCGGGAAGCCGGA 0.0 367

3044 NM _. _014875 5094 CTTCTACTTCTGTTGGCAG 85.9 368

3045 NM _. _014875 4354 ACTTACTATTCAGACTGCA 85.7 369

3046 NM. _014875 524 GCCCTCACCCACAGTAGCC 68.1 370

3047 NM _. _014875 5349 CAGAGGAATGCACACCCAG 73.6 371

3048 NM _014875 4824 GATTGATTAGATCTCTTGA 91.3 372

3049 NM _014875 3014 GTGAGTATTATCCCAGTTG 41.5 373

3050 NM 014875 2959 ATCTGGGGTGCTGATTGCT 46.3 374

3051 NM _. _014875 1514 GTGACAGTGGCAGTACGCG 67.7 375

3052 NM _014875 11 14 TCAGACTGAAGTTGTTAGA 80.8 376

3053 NM 014875 2079 GTTGGCTAGAATTGGGAAA 91.8 377

accession start % SEQID

BioID number position 19mer sequence silencing Set l Set 2 NO.

3054 NM 014875 3560 GAAGACCATAGCATCCGCC 74.8 378

Table III 30 siRNAs designed using the method of this example BioID Accession Gene name Sequence (sense strand) lencinj ; SEQ ID NO

3844 NM_014875 KIF14 CAGGTAAAGTCAGAGACAT 87 379

3845 NM_014875 KIF 14 GGGATTGACGGCAGTAAGA 89 380

3846 NM 014875 KIF14 CACTGAATGTGGGAGGTGA 92 381

3847 NM_014875 KIF 14 GTCTGGGTGGAAATTCAAA 93 382

3848 NM_014875 KIF 14 CATCTTTGCTGAATCGAAA 86 383

3849 NM_014875 KIF 14 CAGGG ATGCTGTTTGG ATA 95 384

3850 NM_OO5O3O PLK CCCTGTGTGGGACTCCTAA 87 385

3851 NM OO5O3O PLK GGTGTTCGCGGGCAAGATT 86 386

3852 NM_005030 PLK CGCCTCATCCTCTACAATG 88 387

3853 NM_005030 PLK GTTCTTTACTTCTGGCTAT 97 388

3854 NM OO5O3O PLK CTCCTTAAATATTTCCGCA 92 389

3855 NM_OO5O3O PLK CTGAGCCTGAGGCCCGATA 75 390

3856 NM_000875 IGFlR CAAATTATGTGTTTCCGAA 90 391

3857 NM_000875 IGFlR CGCATGTGCTGGCAGTATA 84 392

3858 NM_000875 IGFlR CCGAAGATTTCACAGTCAA 79 393

3859 NM_000875 IGFlR ACCATTGATTCTGTTACTT 86 394

3860 NM_000875 IGFlR ACCGC AAAGTCTTTGAGAA 88 395

3861 NM_000875 IGFlR GTCCTGACATGCTGTTTGA 79 396

3862 NM_001315 MAPK 14 GGAATTCAATGATGTGTAT 85 397

3863 NM_001315 MAPK 14 GCTGTTGACTGGAAGAACA 84 398

3864 NM_001315 MAPK 14 CTCCTG AG ATCATGCTGAA 81 399

3865 NM_001315 MAPK 14 CCATTTCAGTCCATCATTC 88 400

3866 NM 001315 MAPK 14 CAGATTATGCGTCTGACAG 25 401

3867 NM_001315 MAPK 14 CGCTTATCTCATTAACAGG 14 402

3871 NM_004523 KIFI l GAGCCCAGATCAACCTTTA 87 403

3872 NM_004523 KIFI l CTGACAAGAGCTCAAGGAA 89 404

3873 NM_004523 KIFI l GGCATTAACACACTGGAGA 92 405

3874 NM 004523 KIFI l GATGGCAGCTCAAAGCAAA 93 406

3875 NM_004523 KIFI l CAGCAGAAATCTAAGGATA 86 407

3876 NM 004523 KIFI l CGTTCTGGAGCTGTTGATA 95 408

6.2. EXAMPLE 2: SELECTION OF SIRNAS FOR SILENCING SPECIFICITY

The importance of off-target effects of siRNA and shRNA sequences has been shown. Microarray experiments suggest that most siRNA oligos result in downregulation of off-target genes through direct interactions between dsRNA and the off-target transcripts. While sequence similarity between dsRNA and transcripts appears to play a role in determining which off-target genes will be affected, sequence similarity searches, even combined with thermodynamic models of hybridization, are insufficient to predict off-target effects accurately. However, alignment of off-target transcripts with offending siRNA sequences reveals that some base pairing interactions between the two appear to be more important than others (Fig. 6). Figure 6 shows an example of alignments of transcripts of off-target genes to the core 19mer of an siRNA oligo sequence. Off-target genes were selected from the Human 25k v2.2.1 microarray by selecting for kinetic patterns of transcript abundance consistent with direct effects of siRNA oligos. Alignments were generated with FASTA and edited by hand. The black boxes and grey area demonstrate the higher level of sequence similarity in the 3 ' half of the alignment.

The alignment shown in Fig. 6 and similar data for other siRNAs were combined to generate a position-specific scoring matrix for use in predicting off-target effects. The matrix, which reflects the frequency with which each position in the oligo is found to match affected off-target transcripts, is represented in Fig. 7. The position-specific scoring matrix is used to calculate scores for alignments between a candidate RNAi sequence and off-target transcript sequences. Alignments of interest are established with a low-stringency FASTA search and the score for each alignment is calculated with the Eq. 6

Score = ∑ In(E ₁ /0.25)

1=1 where: n is the length of the alignment (generally 19); E,— P, from Fig. 7 if position / in the alignment is a match and E, = (l-P,)/3 if position i is a mismatch. It was observed that the number of alignments for a given siRNA which score above a threshold is predictive of the number of observed off-target effects. The threshold of the score was optimized to maximize the correlation between predicted and observed numbers of effects (Fig. 8). The selection pipeline uses the optimized threshold to favor sequences with relatively small numbers of predicted off-target effects.

6.3. EXAMPLE 3: CURVE MODEL PSSMS

PSSMs were also generated by a method which hypothesized dependency of the base composition of any one position on its neighboring positions, referred to as "curve models".

The curve models were generated as a sum of normal curves. Each curve represents the probability of finding a particular base in a particular region. The value at each position in the summed normal curves is the weight given to that position for the base represented by the curve. The weights for each base present at each position in each siRNA and its flanking sequences were summed to generate an siRNA's score, i.e., the score is σ Wj. The score calculation can also be described as the dot product of the base content in the sequence with the weights in the curve model. As such, it is one way of representing the correlation of the sequence of interest with the model.

Curve models can be initialized to correspond to the major peaks and valleys present in the smoothed base composition difference between good and bad siRNAs, e.g., as described in FIGS. IA-C and 5A-C. The initial model can be set up for the 3-peak G/C curve model as follows: Peak l mean: 1.5 standard deviation: 2 amplitude: 0.0455 Peak 1 mean, standard deviation and amplitude are set to correspond to the peak in the mean difference in GC content between good and bad siRNAs occurring within bases -2 - 5 of the siRNA target site in Set 1 training and test sets. Peak 2 mean: 1 1 standard deviation: 0.5 amplitude: 0.0337

Peak 2 mean, standard deviation and amplitude are set to correspond to the peak in the mean difference in GC content between good and bad siRNAs occurring within bases 10- 12 of the siRNA target site in Set 1 training and test sets. Peak 3 mean: 18.5 standard deviation: 4 amplitude: -0.0548

Peak 3 mean, standard deviation and amplitude are set to correspond to the peak in the mean difference in GC content between good and bad siRNAs occurring within bases 12-25 of the siRNA target site in Set 1 training and test sets.

Peak height (amplitude), center position in the sequence (mean) and width (standard deviation) of a peak in a curve model can be adjusted. Curve models were optimized by adjusting the amplitude, mean and standard deviation of each peak over a preset grid of values. Curve models were optimized on several training sets and tested on several test sets, e.g., training sets and test sets as described in Table II. Each base - G/C, A or U - was optimized separately, and then combinations of optimized models were screened for best performance.

The optimization criteria for curve models were: (1) the fraction of good oligos in the top 10%, 15%, 20% and 33% of the scores, (2) the false detection rate at 33% and 50% of the siRNAs selected, and (3) the correlation coefficient of siRNA silencing vs. siRNA scores used as a tiebreaker. When the model is trained, a grid of possible values for amplitude, mean and standard deviation of each peak is explored. The models with the top value or within the top range of values for any of the above criteria were selected and examined further.

G/C models were optimized with 3 or 4 peaks. A models were optimized with 3 peaks. U models were optimized with 5 peaks. Exemplary optimization ranges for the models are listed below:

3 Peak G/C models: peak 1 : amplitudes: gel = 0 - 0.091 means: gel = -2.5 - 1.5 standard deviations: gel = 2.5 - 4 peak 2: amplitudes: gc2 = 0.0337 - 0.1011 means: gc2 = 11 - 11.5 standard deviations: gc2 = 0.5 - 0.9 peak 3: amplitudes: gc3 = -0.1644 - -0.0822 means: gc3 = 18.75 - 20.75 standard deviations: gc3 = 2.5 - 3.5

4 Peak G/C models: peak 0: amplitudes: gcO = O - 0.091 means: gcO = -5.5 - -3.5 standard deviations: gcO = 1-2.5 peak 1: amplitudes: gel = O - 0.091 means: gel =-2.5- 1.5 standard deviations: gel =2.5-4 peak 2: amplitudes: gc2 = 0.0337 - 0.1011 means: gc2 =11-11.5 standard deviations: gc2 = 0.5-0.9 peak 3: amplitudes: gc3 = -0.1644 - -0.0822 means: gc3 = 18.75 - 20.75 standard deviations: gc3 = 2.5-3.5

5 Peak U models: U peak 1 : amplitudes: ul = -0.2 - 0.0 means: ul = 1 - 2 standard deviations: ul = .75 — 1.5

U peak 2: amplitudes: u2 = 0.0 - 0.16 means: u2 = 5 - 6 standard deviations: u2 = .75 - 1.5 U peak 3: amplitudes: u3 = 0.0 - 0.1 means: u3 = 10- 11 standard deviations: u3 = 1 - 2

U peak 4:

amplitudes: u4 = 0.0 - 0.16 means: u4 = 13 - 14 standard deviations: u4 = .75 - 1.5 U peak 5: amplitudes: u5 = 0.0 - 0.16 means: u5 = 17 - 18 standard deviations: u5 = 1 — 3

3 Peak A model: A peak 1 : amplitudes: al = 0.0442 - 0.2210 means: al = 5.5 - 6.5 standard deviations: al = 1 - 2

A peak 2: amplitudes: a2 = -.05 - 0 means: a2 = 10 - 12.5 standard deviations: a2 = 2.5 - 4.5 A peak 3: amplitudes: a3 = 0.0442 - 0.2210 means: a3 = 18 - 20 standard deviations: a3 = 4 - 6

An exemplary set of curve models for PSSM is shown in FIG. 1 IA. FIG. 1 IB shows the performance of the models on training and test sets.

6.4. EXAMPLE 4: BASE COMPOSITION MODELS FOR PREDICTION OF

STRAND PREFERENCE OF siRNAS

The mean difference in G/C content between good and bad siRNAs provides a model for G/C PSSMs which can be used to classify siRNA functional and resistant motifs. As it is known that both strands of the siRNA can be active (see, e.g., Elbashir et al., 2001, Genes Dev. 15:188-200), it was of interest to discover how well the G/C contents of both sense and antisense strands of siRNAs fit the model of siRNA functional target motif G/C content derived from the mean difference in G/C content between good and bad siRNAs. To this end, the reverse complements of good and bad siRNAs were

examined. These reverse complements correspond to the hypothetical perfect match target sites for the sense strands of the siRNA duplexes. The reverse complements were compared to the actual good and bad siRNAs, represented by the actual perfect match target sites of the antisense strands of the siRNA duplexes. FIG. 14A shows the difference between the mean G/C content of the reverse complements of bad siRNAs with the mean G/C content of the bad siRNAs themselves, within the 19mer siRNA duplex region. The difference between the mean G/C content of good and bad siRNAs is shown for comparison. The curves were smoothed over a window of 5 (or portion of a window of 5, at the edges of the sequence). FIG. 14B shows the difference between the mean G/C content of the reverse complements of good siRNAs with the mean G/C content of bad siRNAs, within the 19mer siRNA duplex region. The difference between the mean G/C content of good and bad siRNAs is shown for comparison. The curves were smoothed over a window of 5 (or portion of a window of 5, at the edges of the sequence). The reverse complements of bad siRNAs were seen to be even more different from the bad siRNAs themselves than are good siRNAs. On the average, the reverse complements of bad siRNAs had even stronger G/C content at the 5' end than the good siRNAs did and were similar in G/C content to good siRNAs at the 3' end. In contrast, the reverse complements of good siRNAs were seen to be substantially more similar to bad siRNAs than the good siRNAs were. On average, the reverse complements of good siRNAs hardly differed from bad siRNAs in G/C content at the 5' end and were only slightly less G/C rich than bad siRNAs at the 3' end.

These results appear to imply that the G/C PSSMs are distinguishing siRNAs with strong sense strands as bad siRNAs from siRNAs with weak sense strands as good siRNAs. An siRNA whose G/C PSSM score is greater than the G/C PSSM score of its reverse complement is predicted to have an antisense strand that is more active than its sense strand. In contrast, an siRNA whose G/C PSSM score is less than the G/C PSSM score of its reverse complement is predicted to have a sense strand that is more active than its antisense strand. It has been shown that increased efficacy corresponds to greater antisense strand activity and lesser sense strand activity. Thus the G/C PSSMs of this invention would appear to distinguish good siRNAs with greater efficacy due to dominant antisense strand activity ("antisense-active" siRNAs) from siRNAs with dominant sense strand activity ("sense-active" siRNAs).

The relevance of comparison of G/C PSSMs of siRNAs and their reverse complements for prediction of strand bias was tested by comparison with estimation of strand bias from siRNA expression profiles by the 3 '-biased method. siRNAs and their reverse complements were scored using the smoothed G/C content difference between good and bad siRNAs within the 19mer, shown in FIG. 14 A, as the weight matrix. The G/C PSSM score of each strand was the dot product of the siRNA strand G/C content with the G/C content difference matrix, following the score calculation method of curve model PSSMs. siRNAs were called sense-active by the 3 '-biased method of expression profile analysis if the antisense-identical score exceeded the sense-identical score. siRNAs were called sense-active by the G/C PSSM method if their reverse complement G/C PSSM score exceeded their own G/C PSSM score.

In FIG. 15, siRNAs were binned by measured silencing efficacy, and the frequency of sense-active calls by the expression profile and G/C PSSM methods was compared. Although these techniques are based on distinct analyses, the agreement is quite good. Both show that a higher proportion of low-silencing siRNAs vs. high-silencing siRNAs are predicted to be sense active. The correlation coefficient for (siRNA G/C PSSM score - reverse complement G/C PSSM score) vs. log _! o(sense-identity score/antisense-identity score) is 0.59 for the set of 61 siRNAs binned in FIG. 15.

6.5. EXAMPLE 5: DESIGNING SIRNAS FOR SILENCING GENES HAVING

LOW TRANSCRIPT LEVELS

In the previous examples, an improved siRNA design algorithm that permits selection of siRNAs with greater and more uniform silencing ability was described. Despite this dramatic improvement, some genes remain difficult to silence with high efficacy. A general trend toward poorer silencing for poorly-expressed genes (less than - 0.5 intensity on microarray; <5 copies per cell; Figure 16) was observed. This example describes identification of parameters affecting silencing efficacy of siRNAs to poorly expressed genes. Twenty-four poorly-expressed genes were selected for detailed analysis of parameters affecting siRNA silencing efficacy. A number of criteria were evaluated for their ability to distinguish good and bad siRNAs, including base composition of the 19mer siRNA duplex sequence and the flanking target region. In addition, the contribution of the GC content of the target transcript was considered. These tests revealed that siRNA

efficacy correlated well with siRNA and target gene base composition. In particular, the GC content of good siRNAs differed substantially from that of bad siRNAs in a region- specific manner (Figure 17). The sequences of siRNAs used in generating Figure 17 are listed in Table IV. Good siRNA duplexes tended to be GC poor at positions 2-7 of the 5' end of the sense strand, and GC poor at the 3' end (positions 18-19). Furthermore, siRNA efficacy correlated with low GC content in the transcript sequence flanking the siRNA binding site. The requirement for low GC content as a determinant of siRNA efficacy may explain the difficulty in silencing the poorly-expressed transcripts, as these transcripts tend to be GC rich overall. Base composition of the siRNA duplex also affected silencing of poorly expressed genes. In particular, the GC content of good siRNAs differed substantially from that of bad siRNAs in a region-specific manner (Figure 17). Good siRNA duplexes tended to be GC rich at the first position, GC poor at positions 2-7 of the 5' end of the sense strand, and GC poor at the 3' end (positions 18-19.) Of the criteria examined, low GC content in positions 2-7 of the sense strand (Figure 17, dotted line) produced the greatest improvement in silencing efficacy. This is consistent with the region of the siRNA implicated in the catalysis step of transcript silencing. Low GC content in this region may provide accessibility or optimal helical geometry for enhanced cleavage. Requiring low GC content in this region of the siRNA may also select for target sites that contain low GC content flanking the binding site, which also correlated with silencing efficacy.

The base composition for good siRNAs to poorly-expressed genes diverges somewhat from our previously-derived base composition criteria for good siRNAs to well- expressed genes (Figure 17, solid line). Good siRNAs to both types of genes show a preference for high GC at position 1, and low GC at the 3' end. However, siRNAs for well-expressed genes show an extreme asymmetry in GC content between the two termini, while siRNAs for poorly-expressed genes prefer a more moderate asymmetry. Our previous design algorithm seeks to maximize asymmetry, in accordance with the features seen in good siRNAs to well-expressed genes. Our current results indicate that base composition of more than one region of the siRNA can influence efficacy. Different regions of the siRNA may be more critical for silencing of different targets, perhaps depending on target transcript features such as expression level or overall GC content. Consistent with this idea, different commercially available design algorithms work well on different subsets of genes (data not shown).

A new siRNA design algorithm (the third generation siRNA design algorithm, RSTA siRNA V3) was developed based on the GC composition derived for poorly- expressed genes. The new algorithm includes the following adjustments to the previous algorithm: (1) selection for 1-3 G+C in sense 19mer bases 2-7,

(2) sense 19mer base 1 & 19 asymmetry (position 1, G or C; position 19, A or T),

(3) -300<pssm score<+200,

(4) greatest off-target BLAST match no more than 16, and

(5) 200 bases on either side of the 19mer are not repeat or low-complexity sequences. The new algorithm was compared to the algorithm described in previous examples, by side-by-side testing of new siRNAs selected by each. The results obtained with three siRNAs selected by each method are shown in Figure 18. siRNAs designed by the new algorithm of the present example showed better median efficacy (80%, compared to 60% for the standard method siRNA) and were more uniform in their performance. The distribution of silencing efficacies of siRNAs obtained by the new algorithm was significantly better than that of the previous algorithm for the same genes (p = 10 ^"5 , Wilcoxon rank-sum). siRNAs designed using the new design algorithm also appear effective at silencing more highly-expressed transcripts, based on an examination of 12 highly-expressed genes. The new design criteria may capture features important to siRNA functionality in general (Figure 19), and emphasize that different regions of siRNAs have different functions in transcript recognition, cleavage, and product release. Bases near the 5' end of the guide strand are implicated in transcript binding (both on- and off-target transcripts), and have recently been shown to be sufficient for target RNA-binding energy. The design criteria are also consistent with available data on how siRNAs interact with RISC, the protein-RNA complex that mediates RNA silencing. These studies show that weaker base pairing at the 5' end of the antisense strand (3' end of the duplex) encourages preferential interaction of the antisense strand with RISC, perhaps by facilitating unwinding of the siRNA duplex by a 5 '-3' helicase component of RISC. As in the previous design, our new design maintains the base composition asymmetry that encourages preferential interaction of the antisense strand. This suggests that the previous inefficiency of silencing poorly- expressed transcripts is not due to inefficient association with RISC, but rather is likely due to inefficient targeting of the RISC complex to the target transcript, or inefficient cleavage and release of the target transcript. The designs described in these examples

include a preference for U at position 10 of the sense strand, which has been associated with improved cleavage efficiency by RISC as it is in most endonucleases. The observed preference for low GC content flanking the cleavage site may enhance accessibility of the RISC/nuclease complex for cleavage, or release of the cleaved transcript, consistent with recent studies demonstrating that base pairs formed by the central and 3' regions of the siRNA guide strand provide a helical geometry required for catalysis. The new design criteria may increase the efficiency of these and additional steps in the RNAi pathway, thereby providing efficient silencing of transcripts at different levels of expression. Table IV siRNAs for Figure 17

ACCESSION NUMBER GENE siRNλ sequence SEQ ID NO

AKQ92024_NM_030932 DIAPH3 GCAGTGATTGCTCAGCAGC 409

AK092024_NM_030932 DIAPH3 GAGTTTACCGACCACCAAG 410

AK092024_NM_030932 DIAPH3 CACGGTTGGCAGAGTCTAT 411

AK092024_NM_030932 DIAPH3 TGCGGATGCCATTCAGTGG 412

NM_01487 ₅ KIF 14 AAACTGGGAGGCTACTTAC 413

NM_01487 ₅ KIF 14 CTCACATTGTCCACCAGGA 414

NM_014875 KIF14 GACCATAGCATCCGCCATG 4I ₅

NM_01487 ₅ KIF 14 AGAGCCTTCGAAGGCTTCA 416

NM_01487 ₅ KIF14 TAGACCACCCATTGCTTCC 417

NM_01487 ₅ KIF14 ACTGACAACAAAGTGCAGC 418

U53 ₅ 3O DNCHl TGGCCAGCGCTTACTGGAA 419

U ₅ 3 ₅ 30 DNCHl GCAAGTTGAGCTCTACCGC 420

NM_0008 ₅ 9 HMGCR TTGTGTGTGGGACCGTAAT 421

NM_0008 ₅ 9 HMGCR CAACAGAAGGTTGTCTTGT 422

NM_000859 HMGCR CAGAGACAGAATCTACACT 423

NM_0008 ₅ 9 HMGCR CACGATGCATAGCCATCCT 424

NM_0Q0271 NPCl GAGGTACAATTGCGAATAT 42 ₅

NM_000271 NPCl GCCACAGTCGTCTTGCTGT 426

NM_000271 NPCl TACTACGTCGGACAGAGTT 427

NM_ ₀₀ 0271 NPCl AACTACAATAACGCCACTG 428

NM_0Q4 ₅ 23 KNSLl TACTGATAATGGTACTGAA 429

NM_004 ₅ 23 KNSLl TACATGAACTACAAGAAAA 430

NM_004523 KNSLl GACTAAGCTTAATTGCTTT 431

NM_004 ₅ 23 KNSLl AGTTGACCAACACAATGCA 432

NM_0Q4 ₅ 23 KNSLl GTTTTCTCTGTTACAATAC 433

NM_004 ₅ 23 KNSLl AGGAGTGATAATTAAAGGT 434

NM_004 ₅ 23 KNSLl AAACTCTGAGTACATTGGA 435

NM_004523 KNSLl TACTAAACAGATTGATGTT 436

NM_004523 KNSLl GCTCAAGGAAAACATACAC 437

NM_004523 KNSLl CTGGATCGTAAGAAGGCAG 438

NM_004 ₅ 23 KNSLl GACTTCATTGACAGTGGCC 439

NM_004523 KNSLl GGACAACTGCAGCTACTCT 440

NM_004 ₅ 23 KNSLl GGGGCAGTATACTGAAGAA 441

NM_004 ₅ 23 KNSLl GACCTGTGCCTTTTAGAGA 442

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_Q04523 KNSLl AAAGGACAACTGCAGCTAC 443

NM_004523 KNSLl TACAAAGAATAAATTTTCT 444

NM_004 ₅ 23 KNSLl TGGAAGGTGAAAGGTCACC 445

NM_Q04 ₅ 23 KNSLl TAACTGTTCAAGAAGAGCA 446

NM_004 ₅ 23 KNSLl TCTATAATTTATATTCTTT 447

NM_004 ₅ 23 KNSLl GGGACCGTCATGGCGTCGC 448

NM_Q04 ₅ 23 KNSLl CCAGGGAGACTCCGGCCCC 449

NMJ304523 KNSLl ATTTAATTTGGCAGAGCGG 4 ₅ 0

NM_0G4 ₅ 23 KNSLl TGGAAATATAAATCAATCC 451

NM_004523 KNSLl ACTAACTAGAATCCTCCAG 452

NM_0Q4 ₅ 23 KNSLl AAGAAGAATATATCACATC 453

NM_0G4 ₅ 23 KNSLl TTCTTGTATATTATTAAGT 454

NM_GG4G64 CDKNlB GACGTCAAACGTAAACAGC 455

NM_QQ4 ₀₆ 4 CDKNlB TGGTGATCACTCCAGGTAG 456

NM_GG4064 CDKNlB TGTCCCTTTCAGAGACAGC 457

NM_004073 CNK GTTACCAAGAGCCTCTTTG 458

NM_004073 CNK ATCGTAGTGCTTGTACTTA 459

NM_GG4G73 CNK GAAGACCATCTGTGGCACC 460

NM_004073 CNK GGAGACGTACCGCTGCATC 461

NM_GG4G73 CNK TCAGGGACCAGCTTTACTG 462

NM_0G4073 CNK AGTCATCCCGCAGAGCCGC 463

NM_GG1315 MAPK14 GGCCTπTCACGGGAACTC 464

NM_OO131 ₅ MAPK14 GAAGCTCTCCAGACCATTT 465

NM_0G131 ₅ MAPK14 TGCCTACTTTGCTCAGTAC 466

NM_G01315 MAPK14 ATGTGATTGGTCTGTTGGA 467

NM_OG131 ₅ MAPK 14 GTCATCAGCTTTGTGCCAC 468

NM_0G1315 MAPK14 CCTACAGAGAACTGCGGTT 469

NM_QQ131S MAPK14 CCAGTGGCCGATCCTTATG 470

NM_O01315 MAPK14 GTGCCTCTTGTTGCAGAGA 471

NMJ)0131 ₅ MAPK14 TTCTCCGAGGTCTAAAGTA 472

NMJ)01315 MAPK14 TAATTCACAGGGACCTAAA 473

NM_00131 ₅ MAPK14 GTGGCCGATCCTTATGATC 474

NMJ)0131 ₅ MAPK14 GTATATACATTCAGCTGAC 475

NM_001315 MAPK 14 AATATCCTCAGGGGTGGAG 476

NMJ)01315 MAPK14 GGAACACCCCCCGCTTATC 477

NMJ)06101 HEC CTGAAGGCTTCCTTACAAG 478

NM_006101 HEC AGAACCGAATCGTCTAGAG 479

NM_006101 HEC CAGAAGTTGTGGAATGAGG 480

NMJJ06101 HEC GTTCAAAAGCTGGATGATC 481

NM_006101 HEC GGCCTCTATACCCCTCAAA 482

NMJ)06101 HEC CTTGCAACGTCTGTTAGAG 483

NMJ)00314 PTEN CCCACCACAGCTAGAACTT 484

NM_000314 PTEN CAGTAGAGGAGCCGTCAAA 485

NM_000314 PTEN CTATTCCCAGTCAGAGGCG 486

NM_000314 PTEN TAAAGATGGCACTTTCCCG 487

NM_000314 PTEN AAGGCAGCTAAAGGAAGTG 488

NM_000314 PTEN TGGAGGGGAATGCTCAGAA 489

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_ ₀₀ 007 ₅ CDK4 GCGAATCTCTGCCTTTCGA 490

NM_000075 CDK4 CAGTCAAGCTGGCTGACTT 491

NM_00007 ₅ CDK4 GGATCTGATGCGCCAGTTT 492

NM_00Q075 CDK4 TGTTGTCCGGCTGATGGAC 493

NM_006622 SNK TGTTACGAGATGACAGATT 494

NM_006622 SNK AACCCAGAGGATCGTCCCA 49 ₅

NM_006622 SNK CAGTTCACTATTACGCAGA 496

NM_139164 STARD4 ACCAGAGTCTTTTGACAGG 497

NM_1391ό4 STARD4 CTGTTTGGAGAAAACCCTC 498

NMJ39164 STARD4 GACAACCCAAACCAGAGTC 499

NM_139164 STARD4 GTCTTGACTGGGATGAAAA ₅ 00

NM_0O5O3O PLK GGGAGAAGATGTCCATGGA 501

NM_005030 PLK CCGAGTTATTCATCGAGAC 502

NM_ ₀ 05030 PLK GAGACCTACCTCCGGATCA 503

NM_OO5O3O PLK TCCAGAAGATGCTTCAGAC 504

NM_005030 PLK CACGCCTCATCCTCTACAA 505

NM_ ₀₀₅₀ 30 PLK GACTTCGTGTTCGTGGTGT 506

NM_0OSO3O PLK GGGCGGCTTTGCCAAGTGC 507

NM_OO5O3O PLK ACGAGCTGCTTAATGACGA 508

NM_ ₀₀₅₀ 30 PLK GGACTGGCAACCAAAGTCG 509

NM_0O ₅ O3O PLK GCCCGATACTACCTACGGC 510

NM_ _OO5 O3O PLK CGGAGTTGCAGCTCCCGGA 511

NM_OO _5O 3O PLK AAGAGACCTACCTCCGGAT 512

NM_ ₀₀₅₀ 30 PLK AGTGGGTGGACTATTCGGA 513

NM_QO5O3O PLK TGTATCATGTATACCTTGT 514

NM_00 ₅ 030 PLK AAGAAGAACCAGTGGTTCG 515

NM_ ₀₀₅₀ 30 PLK GGCAAGATTGTGCCTAAGT 516

NM_0O ₅ O3O PLK CCGCGGTGCCATGTCTGCA 517

NM_OO ₅₀ 3O PLK CTCAAGGCCTCCTAATAGC 518

NM_0Q ₅ 03O PLK CAGCGCAGCTTCGGGAGCA 519

NM_0 ₀₅ 030 PLK CACAGTCCTCAATAAAGGC 520

NM_00 ₅ 030 PLK CCCCTCCCCCTCAACCCCA 521

NM_ _OO5 O3O PLK TCGATTGCTCCCAGCAGCC 522

NM_OO ₅ O3O PLK TTCTTGCCTAAAAGAGACC 523

NM_OO _5O 3O PLK TACATGAGCGAGCACTTGC 524

NM_0O ₅₀ 3O PLK CAATGCCTCCAAGCCCTCG 525

NM_00087 ₅ IGFlR GGATATTGGGCTTTACAAC 526

NM_00Q87 ₅ IGFlR CTTGCAGCAACTGTGGGAC 527

NM_0Q087 ₅ IGFlR GCTCACGGTCATTACCGAG 528

NM_00087 ₅ IGFlR GATGATTCAGATGGCCGGA 529

NM_OO087 ₅ IGFlR CGACACGGCCTGTGTAGCT 530

NM_00087 ₅ IGFlR AATGCTGACCTCTGTTACC 531

NM_000875 IGFlR TCTCAAGGATATTGGGCTT 532

NM_000875 IGFlR CATTACTCGGGGGGCCATC 533

NM_00087 ₅ IGFlR TGCTGACCTCTGTTACCTC 534

NM_00087 ₅ IGFlR CTACGCCCTGGTCATCTTC 535

NM_000875 IGFlR CCTCACGGTCATCCGCGGC 536

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_00087 ₅ IGFlR CCTGAGGAACATTACTCGG 537

NM_Q01813 CENPE GGAGAGCTTTCTAGGACCT 538

NM_001813 CENPE GAAGAGATCCCAGTGCTTC 539

NM_001813 CENPE ACTCTTACTGCTCTCCAGT 540

NMJ)01813 CENPE TCTGAAAGTGACCAGCTCA 541

NMJ)Ol 813 CENPE GAAAATGAAGCTTTGCGGG 542

NMJ)01813 CENPE CTTAACACGGATGCTGGTG 543

NM_004958 FRAPl CTTGCAGGCCTTGTTTGTG 544

NMJ)049 ₅ 8 FRAPl CAACCTCCAGGATACACTC 545

NM_0049 ₅ 8 FRAPl GACATGAGAACCTGGCTCA 546

NMJ)049 ₅ 8 FRAPl CCAACTTTCTAGCTGCTGT 547

NMJ ⁾ 049 ₅ 8 FRAPl AGGACTTCGCCCATAAGAG 548

NMJ)049 ₅ 8 FRAPl TAATACAGCTGGGGACGAC 549

NMJ)O ₅ 163 AKTl GCTGGAGAACCTCATGCTG 550

NMJ)O ₅ 163 AKTl CGCACCTTCCATGTGGAGA 551

NMJ)O ₅ 163 AKTl AGACGTTTTTGTGCTGTGG 552

NMJ)023S8 MAD2L1 TACGGACTCACCTTGCTTG 553

NMJ)OO ₅₅ I VHL GGCATTGGCATCTGCTTTT 554

NMJ)OO ₅₅ I VHL GTGAATGAGACACTCCAGT 555

NMJ)OO ₅₅ I VHL TGTTGACGGACAGCCTATT 556

NMJ)OO ₅₅ I VHL GATCTGGAAGACCACCCAA 557

NMJ)00 ₅ 51 VHL AGGAAATAGGCAGGGTGTG 558

NMJ)OO ₅₅ I VHL CAGAACCCAAAAGGGTAAG 559

NMJ)01654 ARAFl GTCCCCACATTCCAAGTCA 560

NMJ)01654 ARAFl GAATGAGATGCAGGTGCTC 561

NMJ)01654 ARAFl GTTCCACCAGCATTGTTCC 562

NMJ)01654 ARAFl CCTCTCTGGAATTTGTGCC 563

NMJ)01654 ARAFl AGTGAAGAACCTGGGGTAC 564

NMJ)01654 ARAFl TTGAGCTGCTGCAACGGTC 565

NMJ50O435 NOTCH3 GAACATGGCCAAGGGTGAG 566

NM_000435 NOTCH3 GAGTCTGGGACCTCCTTCT 567

NMJJ00435 NOTCH3 AATGGCTTCCGCTGCCTCT 568

NM_000435 NOTCH3 TGATCACTGCTTCCCCGAT 569

NM_000435 NOTCH3 TGCCAACTGAAGAGGATGA 570

NMJJ00435 NOTCH3 GCTGCTGTTGGACCACTTT 571

NMJ)24408 NOTCH2 CCAAGGAACCTGCTTTGAT 572

NMJ)24408 NOTCH2 GACTCAGACCACTGCTTCA 573

NMJ)24408 NOTCH2 CTTTGAATGCCAGGGGAAC 574

NM_024408 NOTCH2 GCAACTTTGGTCTCCTTTC 575

NMJJ24408 NOTCH2 GAGACAAGTTAACTCGTGC 576

NMJ)24408 NOTCH2 GCAATTGGCTGTGATGCTC 577

NMJ)12193 FZD4 CCATCTGCTTGAGCTACTT 578

NMJ ⁾ 12193 FZD4 TTGGCAAAGGCTCCTTGTA 579

NM_012193 FZD4 AGAACCTCGGCTACAACGT 580

NM_012193 FZD4 TCGGCTACAACGTGACCAA 581

NMJH2193 FZD4 GTTGACTTACCTGACGGAC 582

NMJH2193 FZD4 TCCGCATCTCCATGTGCCA 583

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_007313 ABLl GAATGGAAGCCTGAACTGA ₅ 84

NM_007313 ABLl CAAGTTCTCCATCAAGTCC ₅ 85

NM_0Q7313 ABLl CTAAAGGTGAAAAGCTCCG ₅ 86

NM_007313 ABLl TCCTGGCAAGAAAGCTTGA ₅ 87

NM_007313 ABLl AAACCTCTACACGTTCTGC 588

NM_007313 ABLl AGACATCATGGAGTCCAGC 589

NM_017412 FZD3 CAGATCACTCCAGGCATAG 590

NM_017412 FZD3 ATGTGTGGTGACTGCTTTG 591

NM_017412 FZD3 AGAGATGGGCATTGTTTCC 592

NMJ)17412 FZD3 AGCATTGCTGTTTCACGCC 593

NMJH7412 FZD3 GCTCATGGAGATGTTTGGT 594

NM_005633 SOSl TGGTGTCCTTGAGGTTGTC 595

NM_ ₀₀₅ 633 SOSl TATCAGACCGGACCTCTAT 596

NM_00 ₅ 633 SOSl CTTACAAAAGGGAGCACAC 597

NM_Q ₀₅ 633 SOSl GAACACCGTTAACACCTCC 598

NM_005633 SOSl ATAACAGGAGAGATCCAGC 599

NM_ ₀₀₅ 633 SOSl ATTGACCACCAGGTTTCTG 600

NM_ ₀₀₅ 417 SRC CAATTCGTCGGAGGCATCA 601

NM_00 ₅ 417 SRC GCAGTGCCTGCCTATGAAA 602

NM_005417 SRC GGGGAGTTTGCTGGACTTT 603

NM_00 ₅ 400 PRKCE GATCGAGCTGGCTGTCTTT 604

NM_QO ₅ 4G0 PRKCE GCTCACCATCTGAGGAAGA 605

NM_GG54GG PRKCE GGTCTTAAAGAAGGACGTC 606

NM_00S40G PRKCE TCACAAAGTGTGCTGGGTT 607

NM_ ₀₀₅ 4G0 PRKCE CCAGGAGGAATTCAAAGGT 608

NM_Q0 ₅ 40G PRKCE TGAGGACGACCTATTTGAG 609

NM_GG2388 MCM3 GTCTCAGCTTCTGCGGTAT 610

NM_G02388 MCM3 GTACATCCATGTGGCCAAA 611

NM_0G2388 MCM3 AGGATTTTGTGGCCTCCAT 612

NM_QQ2388 MCM3 TGGGTCATGAAAGCTGCCA 613

NM_G02388 MCM3 TCCAGGTTGAAGGCATTCA 614

NM_0G2388 MCM3 GCAGATGAGCAAGGATGCT 615

NM_GQ438G CREBBP GAAAAACGGAGGTCGCGTT 616

NM_GG438G CREBBP GACATCCCGAGTCTATAAG 617

NM_Q04380 CREBBP TGGAGGAGAATTAGGCCTT 618

NM_00438Q CREBBP ATTTTTGCGGCGCCAGAAT 619

NM_004380 CREBBP GCACAAGGAGGTCTTCTTC 620

NM_QG438Q CREBBP GAAAACAAATGCCCCGTGC 621

NM_GQ6219 PIK3CB CAAAGATGCCCTTCTGAAC 622

NM_006219 PIK3CB GTGCACATTCCTGCTGTCT 623

NM_006219 PIK3CB AAGTTCATGTCAGGGCTGG 624

NM_0Q6219 PIK3CB AATGCGCAAATTCAGCGAG 625

NM_006219 PIK3CB AATGAAGCCTTTGTGGCTG 626

NM_006219 PIK3CB TACAGAAAAGTTTGGCCGG 627

NM_006218 P1K3CA CTAGGAAACCTCAGGCTTA 628

NM_006218 PIK3CA TTCAGCTAGTACAGGTCCT 629

NM_006218 PIK3CA TGATGCACATCATGGTGGC 630

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_006218 PIK3CA AGAAGCTGTGGATCTTAGG 631

NM_006218 PIK3CA AGGTGCACTGCAGTTCAAC 632

NM_006218 PIK3CA TGGCTTTGAATCTTTGGCC 633

NM_002086 GRB2 CTGGTACAAGGCAGAGCTT 634

NM_002086 GRB2 CGGGCAGACCGGCATGTTT 635

NM_0Q2086 GRB2 CCGGAACGTCTAAGAGTCA 636

NM_002086 GRB2 ATACGTCCAGGCCCTCTTT 637

NM_002086 GRB2 TGAGCTGGTGGATTATCAC 638

NM_002086 GRB2 TGCAGCACTTCAAGGTGCT 639

NM_Q01982 ERBB3 TGACAGTGGAGCCTGTGTA 640

NM_001982 ERBB3 CTAGACCTAGACCTAGACT 641

NM_001982 ERBB3 CTTTCTGAATGGGGAGCCT 642

NM_001982 ERBB3 GAGGATGTCAACGGTTATG 643

NM_001982 ERBB3 CAAAGTCTTGGCCAGAATC 644

NMJ)01982 ERBB3 TACACACACCAGAGTGATG 645

NMJ)01903 CTNNAl CGTTCCGATCCTCTATACT 646

NMJ)01903 CTNNAl AAGCCATTGGTGAAGAGAG 647

NMJJ01903 CTNNAl TGTGTCATTGCTCTCCAAG 648

NMJ)01903 CTNNAl AGCAGTGCTGATGATAAGG 649

NMJ)01903 CTNNAl TGACCAAAGATGACCTGTG 650

NMJ)01903 CTNNAl TGACATCATTGTGCTGGCC 651

NMJJ03600 STK6 CACCCAAAAGAGCAAGCAG 652

NM_OO36OO STK6 GCACAAAAGCTTGTCTCCA 6S3

NM_003600 STK6 CCTCCCTATTCAGAAAGCT 654

NM_003600 STK6 ACAGTCTTAGGAATCGTGC 655

NM_003600 STK6 GACTTTGAAATTGGTCGCC 656

NMJJ03600 STK6 TTGCAGATTTTGGGTGGTC 657

NMJJ03161 RPS6KB1 GACACTGCCTGCTTTTACT 658

NM_003161 RPS6KB1 CTCTCAGTGAAAGTGCCAA 659

NMJJ03161 RPS6KB1 GCTTTTCCCATGATCTCCA 660

NMJJ03161 RPS6KB1 TTGATTCCTCGCGACATCT 661

NM_003161 RPS6KB1 GAAAGCCAGACAACTTCTG 662

NMJ)03161 RPS6KB1 CTTGGCATGGAACATTGTG 663

AF308602 NOTCHl GATCGATGGCTACGAGTGT 664

AF308602 NOTCHl CACTTACACCTGTGTGTGC 665

AF308602 NOTCHl AGGCAAGCCCTGCAAGAAT 666

AF308602 NOTCH 1 CATCCCCTACAAGATCGAG 667

AF308602 NOTCH 1 ATATCGACGATTGTCCAGG 668

AF308602 NOTCH 1 ATTCAACGGGCTCTTGTGC 669

NMJ) 16231 NLK CCACTCAGCTCAGATCATG 670

NMJ) 16231 NLK GCAATGAGGACAGCTTGTG 671

NMJ) 16231 NLK TGTAGCTTTCCACTGGAGT 672

NMJ) 16231 NLK TCTCCTTGTGAACAGCAAC 673

NMJ) 16231 NLK GGAAACAGAGTGCCTCTCT 674

NMJ) 16231 NLK TCTGGTCTCTTGCAAAAGG 675

NMJ)01253 CDC5L AAGAAGACGTTCAGCGACA 676

NMJ)01253 CDC5L AAAAAGCCTGCCCTTGGTT 677

ACCESSION NUMBER GENE siRiNA sequence SEQ ID NO

NM_0012 ₅ 3 CDC5L TCATTGGAAGAACAGCGGC 678

NM_Q03391 WNT2 GTGTCTCAAAGGAGCTTTC 679

NM_003391 WNT2 GCCTCAGAAAGGGATTGCT 680

NM_003391 WNT2 AGAAGATGAATGGTCTGGC 681

NM_003391 WNT2 GCTCTGGATGTGCACACAT 682

NM_003391 WNT2 AACGGGCGATTATCTCTGG 683

NM_003391 WNT2 ATTTGCCCGCGCATTTGTG 684

NM_002387 MCC AGTTGAGGAGGTTTCTGCA 685

NM_002387 MCC GACTTAGAGCTGGGAATCT 686

NM_002387 MCC GGATTATATCCAGCAGCTC 687

NM_002387 MCC GAGAATGAGAGCCTGACTG 688

NM_002387 MCC TAGCTCTGCTAGAGGAGGA 689

NM_0Q2387 MCC ACAGAACGGCTGAATAGCC 690

NM_0 ₀₅ 978 S100A2 GGAACTTCTGCACAAGGAG 691

NM_005978 S100A2 GGGCCCAGGACTGTTGATG 692

NM_0 ₀₅ 978 S100A2 TGAGAACAGTGACCAGCAG 693

NM_00 ₅ 978 S100A2 TGGCACTCATCACTGTCAT 694

NM_ ₀₀₅ 978 S100A2 GACCGACCCTGAAGCAGAA 695

NM_00 ₅ 978 S100A2 TTCCAGGAGTATGCTGTTT 696

NM_033360 KRAS2 GAAGTTATGGAATTCCTTT 697

NM_033360 KRAS2 GGACTCTGAAGATGTACCT 698

NM_ ₀ 3336Q KRAS2 GGCATACTAGTACAAGTGG 699

NM_033360 KRAS2 ACCTGTCTCTTGGATATTC 700

NM_033360 K.RAS2 TAAATGTGATTTGCCTTCT 701

NM_033360 KRAS2 GAAAAGACTCCTGGCTGTG 702

NM_139049 MAPK8 GGAATAGTATGCGCAGCTT 703

NM_139049 MAPK8 GTGATTCAGATGGAGCTAG 704

NM_139049 MAPK.8 CACCATGTCCTGAATTCAT 705

NM_139049 MAPK8 CGAGTTTTATGATGACGCC 706

NMJ 39049 MAPK8 CACCCGTACATCAATGTCT 707

NM_139049 MAPK8 TCAAGCACCTTCATTCTGC 708

NM_002 ₆₅ 8 PLAU CAAGTACTTCTCCAACATT 709

NM_0026 ₅ 8 PLAU GAGCTGGTGTCTGATTGTT 710

NM_0026 ₅ 8 PLAU CTGCCCAAAGAAATTCGGA 711

NM_0026 ₅ 8 PLAU GTGTAAGCAGCTGAGGTCT 712

NM_0026 ₅ 8 PLAU TGGAGGAACATGTGTGTCC 713

NM_0026S8 PLAU TTACTGCAGGAACCCAGAC 714

NM_016195 MPHOSPHl AGAGGAACTCTCTGCAAGC 715

NMJ)16195 MPHOSPHl AAGTTTGTGTCCCAGACAC 716

NMJ)16195 MPHOSPHl CTGAAGAAGCTACTGCTTG 717

NMJ)16195 MPHOSPHl GACATGCGAATGACACTAG 718

NMJM6195 MPHOSPHl AATGGCAGTGAAACACCCT 719

NMJ)16195 MPHOSPHl ATGAAGGAGAGTGATCACC 720

NM_020168 PAK6 CGACATCCAGAAGTTGTCA 721

NM_020168 PAK6 GAGAAAGAATGGGGTCGGT 722

NMJJ20168 PAK6 TGAGGAGCAGATTGCCACT 723

NM_000051 ATM TAGATTGTTCCAGGACACG 724

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NMJ)OOO ₅ I ATM AGTTCGATCAGCAGCTGTT 725

NM_0000 ₅ l ATM GAAGTTGGATGCCAGCTGT 726

NM_0012 ₅ 9 CDK6 TCTTGGACGTGATTGGACT 727

NM_0012 ₅ 9 CDK ₆ ACCACAGAACATTCTGGTG 728

NM_0012 ₅ 9 CDK6 AGAAAACCTGGATTCCCAC 729

NM_004856 KNSL5 GAATGTGAGCGTAGAGTGG 730

NM_0048 ₅ 6 KNSL5 CCATTGGTTACTGACGTGG 731

NM_0048S6 KNSL ₅ AACCCAAACCTCCACAATC 732

NM_006845 KNSL6 ACAAAAACGGAGATCCGTC 733

NM_00684 ₅ KNSL6 GAATTTCGGGCTACTTTGG 734

NM_00684 ₅ KNSL6 ATAAGCAGCAAGAAACGGC 73 ₅

NM_004972 JAK2 AGCCGAGTTGTAACTATCC 736

NM_004972 JAK2 AAGAACCTGGTGAAAGTCC 737

NM_004972 JAK2 GAAGTGCAGCAGGTTAAGA 738

NM_00 ₅ 026 PIK3CD GATCGGCCACTTCCTTTTC 739

NM_005026 PIK3CD AGAGATCTGGGCCTCATGT 740

NM_00 ₅ 026 PIK3CD AACCAAAGTGAACTGGCTG 741

NM_01488 ₅ APClO CAAGGCATCCGTTATATCT 742

NM_01488 ₅ APClO ACCAGGATTTGGAGTGGAT 743

NM_01488 ₅ APClO GTGGCTGGATTCATGTTCC 744

NM_00 ₅ 733 RAB6KIFL GAAGCTGTCCCTGCTAAAT 74 ₅

NM_005733 RAB6K1FL CTCTACCACTGAAGAGTTG 746

NM_00 ₅ 733 RAB ₆ KIFL AAGTGGGTCGTAAGAACCA 747

NM_0070 ₅ 4 KIF3A GGAGAAAGATCCCTTTGAG 748

NM_0070 ₅ 4 KIF3A TATTGGGCCAGCAGATTAC 749

NM_0070 ₅ 4 KIF3A TTATGACGCTAGGCCACAA 7 ₅ 0

NM_020242 KNSL7 GCACAACTCCTGCAAATTC 751

NM_020242 KNSL7 GATGGAAGAGCCTCTAAGA 752

NM_020242 KNSL7 ACGAAAAGCTGCTTGAGAG 753

NMJ)Ol 184 ATR TCACGACTCGCTGAACTGT 754

NMJ)Ol 184 ATR GAAACTGCAGCTATCTTCC 7 ₅₅

NMJ)01184 ATR GTTACAATGAGGCTGATGC 756

NMJ)1487 ₅ KIF 14 ATTTTCTAGAAAACGGTAA 757

NMJ)1487 ₅ KIF 14 GAGGGGCGAAGTTTCGGCA 758

NMJ11487S KIF 14 CTGGGACCGGGAAGCCGGA 759

NMJ)1487 ₅ KIF 14 CTTCTACTTCTGTTGGCAG 760

NMJ)1487 ₅ KIF 14 ACTTACTATTCAGACTGCA 761

NMJ)1487 ₅ KIF 14 GCCCTCACCCACAGTAGCC 762

NMJ) 14875 KIF14 CAGAGGAATGCACACCCAG 763

NMJ) 14875 KIF 14 GATTGATTAGATCTCTTGA 764

NMJ)1487 ₅ KIF 14 GTGAGTATTATCCCAGTTG 765

NMJ114875 KIF 14 ATCTGGGGTGCTGATTGCT 766

NMJ)14875 KIF 14 GTGACAGTGGCAGTACGCG 767

NMJ)1487 ₅ K1F14 TCAGACTGAAGTTGTTAGA 768

NMJ114875 KIF 14 GTTGGCTAGAATTGGGAAA 769

NMJ)14875 KIF14 GAAGACCATAGCATCCGCC 770

NMJ)01274 CHEKl TGCCTGAAAGAGACTTGTG 771

ACCESSION NUMBER GENE SiRiNA sequence SEQ ID NO

NMJ)01274 CHEKl ATCGATTCTGCTCCTCTAG 772

NMJ)01274 CHEKl CTGAAGAAGCAGTCGCAGT 773

NM_007194 CHEK2 GATCACAGTGGCAATGGAA 774

NMJJ07194 CHEK2 ATGAATCCACAGCTCTACC 775

NM_007194 CHEK2 AAACTCTTGGAAGTGGTGC 776

NMJ)00 ₅ 46 TP ₅ 3 GCACCCAGGACTTCCATTT 777

NM_00054 ₆ TP ₅ 3 CCTCTTGGTCGACCTTAGT 778

NMJ ⁾ 00 ₅ 46 TP53 TGAGGCCTTGGAACTCAAG 779

NM_00 ₅ 400 PRKCE AGCGCCTGGGCCTGGATGA 780

NM_005400 PRKCE ACCGGGCAGCATCGTCTCC 781

NMJ)05400 PRKCE CAGCGGCCAGAGAAGGAAA 782

NMJ)054 ₀₀ PRKCE CAGAAGGAAGAGTGTATGT 783

NM_00 ₅ 400 PRKCE TGCAGTGTAAAGTCTGCAA 784

NM_00 ₅ 400 PRKCE GCGCATCGGCCAAACGGCC 785

NMJ)0 ₅ 400 PRKCE ATTGCAGAGACTTCATCTG 786

NMJ) ₀₅ 400 PRKCE GAAGAGCCGGTACTCACCC 787

NM_005400 PRKCE AGTACTGGCCGACCTGGGC 788

NM_005400 PRKCE GGATGCAGAAGGTCACTGC 789

NMJ)0 ₅ 400 PRKCE CGTGAGCTTGAAGCCCACA 790

NM_00 ₅ 400 PRKCE CACAAAGTGTGCTGGGTTA 791

NM_005400 PRKCE GACGAAGCAATTGTAAAGC 792

NMJ ⁾ 0 ₅ 400 PRKCE CACCCTTCAAACCACGCAT 793

NMJ)05400 PRKCE GTCAGCATCTTGAAAGCTT 794

NMJ ⁾ ₀₅ 400 PRKCE CAACCGAGGAGAGGAGCAC 795

NMJ)0 ₅ 400 PRKCE TACATTGCCCTCAATGTGG 796

NMJ)05400 PRKCE GAGGAATCGCCAAAGTACT 797

NMJ)0 ₅ 400 PRKCE GGGATTTGAAACTGGACAA 798

NMJJ06218 PIK3CA TTACACGTTCATGTGCTGG 799

NMJ)06218 PIK3CA CACAATCCATGAACAGCAT 800

NMJ ⁾ 06218 PIK3CA CAATCAAACCTGAACAGGC 801

NMJ)06218 PIK3CA CAGTTCAACAGCCACACAC 802

NM_006218 PIK3CA GTGTTACAAGGCTTATCTA 803

NM_006218 PIK3CA GATCCTATGGTTCGAGGTT 804

NM_006218 PIK3CA CTCCAAATAATGACAAGCA 805

NMJ)06218 PIK3CA ACTTTGCCTTTCCATTTGC 806

NMJJ06218 PIK3CA AGAATATCAGGGCAAGTAC 807

NM_006218 PIK3CA TTGGATCTTCCACACAATT 808

NMJJ06218 PIK3CA AGTAGGCAACCGTGAAGAA 809

NMJJ06218 PIK3CA CAGGGCTTGCTGTCTCCTC 810

NMJ ⁾ 06218 PIK3CA GAGCCCAAGAATGCACAAA 811

NM_006218 P1K3CA GCCAGAACAAGTAATTGCT 812

NMJ ⁾ 06218 P1K3CA GGATGCCCTACAGGGCTTG 813

NMJJ06218 PIK3CA TCAAATTATTCGTATTATG 814

NMJJ06218 PIK3CA GAATTGGAGATCGTCACAA 815

NM_006218 PIK3CA TGAGGTGGTGCGAAATTCT 816

NMJJ06218 P1K3CA GATTTACGGCAAGATATGC 817

NM_006218 PIK3CA TGATGAATACTTCCTAGAA 818

ACCESSION NUMBER GENE si RNA sequence SEQ ID NO

NMJ)Ol 982 ERBB3 GCTGCTGGGACTATGCCCA 819

NMJ)01982 ERBB3 ATCTGCACAATTGATGTCT 820

NMJW1982 ERBB3 CTTTGAACTGGACCAAGGT 821

NMJ)01982 ERBB3 CATCATGCCCACTGCAGGC 822

NMJ)01982 ERBB3 AACTTTCCAGCTGGAACCC 823

NMJ)01982 ERBB3 TGAAGGAAATTAGTGCTGG 824

NMJ)01982 ERBB3 AATTCGCCAGCGGTTCAGG 82 ₅

NMJ)01982 ERBB3 ACCAGAGCTTCAAGACTGT 826

NMJ)01982 ERBB3 GAGGCTACAGACTCTGCCT 827

NMJKH982 ERBB3 TGGAGCCAGAACTAGACCT 828

NMJ)01982 ERBB3 ACACTGTACAAGCTCTACG 829

NMJ)01982 ERBB3 TAATGGTCACTGCTTTGGG 830

NMJ ⁾ 01982 ERBB3 ACAGGCACTCCTGGAGATA 831

NMJ)01982 ERBB3 GTTTAGGACAAACACTGGT 832

NMJ)01982 ERBB3 GATTACTGGCATAGCAGGC 833

NMJ)01982 ERBB3 ATGAATACATGAACCGGAG 834

NMJ)01982 ERBB3 CACTTAATCGGCCACGTGG 83 ₅

NMJJ01982 ERBB3 GGCCTGTCCTCCTGACAAG 836

NMJ)01982 ERBB3 TCTGCGGAGTCATGAGGGC 837

NMJJ01982 ERBB3 TAGACCTAGACTTGGAAGC 838

NM_004283 RAB3D GATTTCAGGTCTCCCTGTC 839

NMJJ04283 RAB3D GCCACAGTGGTTATCTCCA 840

NM_004283 RAB3D GCAATCCCTTCCCTCCTGT 841

NMJJ04283 RAB3D TCTCTGATCCTGAAGTGAA 842

NMJ ⁾ 04283 RAB3D CATCAATGTGAAGCAGGTC 843

NM_004283 RAB3D CATGAGCTTGCTGCTTTCC 844

NM_004283 RAB3D AACGTGTTGTGCCTGCTGA 84 ₅

NM_004283 RAB3D CTGCTTTCCAGGGTGTGTT 846

NM_004283 RAB3D GCGGCCAGGGCCAAGCCGC 847

NMJW4283 RAB3D CTTCTAGCTTAGAACCATT 848

NMJ ⁾ 04283 RAB3D CAGGGTGTGTTGAGGGTGG 849

NMJJ04283 RAB3D CTCTTTCTCAGGTCCTGCA 8 ₅ 0

NMJ)04283 RAB3D CTTGTGCCAAGATGGCATC 851

NMJ)04283 RAB3D GCACCATCACCACGGCCTA 852

NMJJ04283 RAB3D CGCGGACGACTCCTTCACT 853

NM_004283 RAB3D TCATCCAGGGAAGGCGGCG 854

NMJ ⁾ 04283 RAB3D GACACTGACGTGCATGAGC 855

NM_004283 RAB3D CCCTCCCAGGCCCTGTTTA 856

NMJJ04283 RAB3D AGGTCTTCGAGCGCCTGGT 857

NMJJ04283 RAB3D CCTCTTTCTCAGGTCCTGC 858

NMJ)03620 PPMlD TTGCCCGGGAGCACTTGTG 859

NMJ)03620 PPMlD CGTGTGCGACGGGCACGGC 860

NM_003620 PPMlD ATTAGGTCTTAAAGTAGTT 861

NMJX)3620 PPMlD AGCCCTGACTTTAAGGATA 862

NMJW3620 PPMlD TGTGGAGCCCGAACCGACG 863

NMJ ⁾ 03620 PPMlD GCGACGGGCACGGCGGGCG 864

NMJW3620 PPMlD GATTATATGGGTATATATT 865

ACCESSION NUMBER GENE si RNA sequence SEQ ID NO

NM_003620 PPMlD TTAGAAGGAGCACAGTTAT 86β

NM_003620 PPMlD CCGGCCAGCCGGCCATGGC 867

NM_00362Q PPMlD GAGCAGATAACACTAGTGC 868

NM_003620 PPMlD AGATGCCATCTCAATGTGC 869

NM_003620 PPMlD GCGGCACAGTTTGCCCGGG 870

NM_003620 PPMlD CGTAGCAATGCCTTCTCAG 871

NM_003620 PPMlD TATATGGGTATATATTCAT 872

NM_0Q3620 PPMlD GCTGCTAATTCCCAACATT 873

NM_003620 PPMlD ACAACTGCCAGTGTGGTCA 874

NM_003620 PPMlD TTGACCCTCAGAAGCACAA 87 ₅

NM_003620 PPMlD GTCTTAAAGTAGTTACTCC 876

NM_003620 PPMlD ATGCTCCGAGCAGATAACA 877

NM_003620 PPMlD GCGCCTAGTGTGTCTCCCG 878

NM_022048 CSNKlGl TAGCCATCCAGCTGCTTTC 879

NM_022048 CSNKlGl TTCTCATTGGAAGGGACTC 880

NM_022048 CSNKlGl CACGCATCTTGGCAAAGAG 881

NM_022048 CSNKlGl TAGCTTGGAGGACTTGTTT 882

NM_022048 CSNKlGl ACTCAATTGTACCTGCAGC 883

NM_022048 CSNKlGl CTAAGTGCTGCTGTTTCTT 884

NM_022048 CSNKlGl GCAAAGCCGGAGAGATGAT 88 ₅

NM_022048 CSNKlGl CCTCTTCACAGACCTCTTT 886

NM_022048 CSNKlGl GAAGGGACTCCTCTTTGGG 887

NM_022048 CSNKlGl GAGAGCTCAGATTAGGTAA 888

NM_022048 CSNKlGl CACGTAGATTCTGGTGCAT 889

NM_022048 CSNKlGl ATGAGTATTTACGGACCCT 890

NM_022048 CSNKlGl GGTGGGACCCAACTTCAGG 891

NM_022048 CSNKlGl AGAGCTGAATGTTGATGAT 892

NM_022048 CSNKlGl GATTCTGGTGCATCTGCAA 893

NM_022048 CSNKlGl AACTTCAGGGTTGGCAAGA 894

NM_022048 CSNKlGl TCTCGAATGGAATACGTGC 89 ₅

NM_022048 CSNKlGl CCGAGGAGAGTGGGAAATT 896

NM_022Q48 CSNKlGl GGGAGCCCACTCCAATGCA 897

NM_022048 CSNKlGl GTCAAGCCAGAGAACTTCC 898

NM_000082 CKNl TTAGCAGTTTCCTGGTCTC 899

NM_000082 CKNl ATGTGAGAAGAGCATCAGG 900

NM_000082 CKNl AGCAGTGTGTTCCATTGGC 901

NM_0QDQ82 CKNl GGATCCTGTTCTCACATTC 902

NM_000082 CKNl CAGCAGTGATGAAGAAGGA 903

NM_0Q0082 CKNl GATAACTATGCTTAAGGGA 904

NM_OQ0082 CKNl TGGACTTCACCTCCTCACT 9O ₅

NM_000082 CKNl TTGAAGTCTGGATCCTGTT 906

NM_Q00082 CKNl AGGAACTTTATAGTGGTAG 907

NM_ ₀ 00082 CKNl AAGTGATGGACTTCACCTC 908

NM_000082 CKNl TGTTTATACAGTTTACTCA 909

NM_000Q82 CKNl GAAGGGAGATACATGTTAT 910

NM_000Q82 CKNl GGGTTTGGAGGACCCTCTT 911

NM_000082 CKNl ATATGTCTCCAGTCTCCAC 912

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_Q00082 CKNl GATGGACTTCACCTCCTCA 913

NM_Q00082 CKNl TGAAAGTATGGGATACAAA 914

NM_Q ₀₀ 082 CKNl ATGTAAAGCAGTGTGTTCC 9I ₅

NM_ ₀₀ 0082 CKNl TCTACAGGGTCACAGACAA 916

NM_000082 CKNl GAGGCCATCAGTATTGACT 917

NM_000082 CKNl ACTGTTTGGTAGCAGTTGG 918

NM_002843 PTPRJ AGGAGGAGGCGAAGGAGAC 919

NM_002843 PTPRJ CTACGTCACCACCACGGAG 920

NM_002843 PTPRJ TCGCCTAATTCCAAAGGAA 921

NM_002843 PTPRJ CAAGTATGTAGTAAAGCAT 922

NM_002843 PTPRJ AAGCTGGTCACCCTTCTGC 923

NM_002843 PTPRJ CACAGAAGGTGGCTTGGAT 924

NM_002843 PTPRJ TGGAATCTAGCCGATGGAA 925

NM_002843 PTPRJ ATAAACAGAATGGAACTGG 926

NM_002843 PTPRJ CCTGGAGAGCTGCTCCTCT 927

NM_002843 PTPRJ AACTTTAAGTTGGCAGAAC 928

NM_002843 PTPRJ ACACAGTGGAGATCTTTGC 929

NM_002843 PTPRJ CAGTACACACGGCCCAGCA 930

NM_002843 PTPRJ TTGAACAGGGAAGAACCAA 931

NM_002843 PTPRJ ATTATGTTGACTAAATGTG 932

NM_002843 PTPRJ TGACTCAAGACTCAAGACT 933

NM_002843 PTPRJ AACTTTCGGTCCAGACCCA 934

NM_002843 PTPRJ GGCCAGACCACGGTGTTCC 935

NM_002843 PTPRJ TCACTGGAACCTGGCCGGA 936

NM_002843 PTPRJ ACACAGGAGGGAGCTGGCA 937

NM_002843 PTPRJ TGTTCTCATTTGATCAGGG 938

NM_004037 AMPD2 TCATCCGGGAGAAGTACAT 939

NM_004037 AMPD2 ACCCAACTATACCAAGGAA 940

NM_004037 AMPD2 CCTGCATGAACCAGAAGCA 941

NM_004037 AMPD2 CTGCGGGAGGTCTTTGAGA 942

NM_004037 AMPD2 GCCTCTTTGATGTGTACCG 943

NM_004037 AMPD2 GACAACATGAGAAATCGTG 944

NM_004037 AMPD2 GCCACCCAGTGAAAGCAAA 945

NM_004037 AMPD2 CAGGAACACTTTCCATCGC 946

NM_004037 AMPD2 TGTGGGAGAGGCAGCTGCC 947

NM_0Q4037 AMPD2 GCCGTGAACAGACGCTGCG 948

NM_004037 AMPD2 AAATATCCCTTTAAGAAGC 949

NM_0Q4037 AMPD2 GTAAAGAGCCACTGGCTGG 950

NM_ ₀ 04037 AMPD2 CGTCCTGCATGAACCAGAA 951

NM_0Q4037 AMPD2 GCTCAGCAACAACAGCCTC 952

NM_004037 AMPD2 CACATCATCAAGGAGGTGA 953

NMJ)04037 AMPD2 CTCATTGTTGTTTGGGCTC 954

NM_004037 AMPD2 AAGCTCAGCTCCTGCGATA 955

NM_004037 AMPD2 TGCGATATGTGTGAGCTGG 956

NM_Q04037 AMPD2 CTGGGCCCATCCACCACCT 957

NM_004037 AMPD2 GAAGGACCAGCTAGCCTGG 958

NM_016218 POLK TATTTCATTTCTTGTCAAT 959

ACCESSION NUMBER GENE siRNA sequence SEQ ID NO

NM_016218 POLK GACGAGGGATGGAGAGAGG 96Q

NM_Q16218 POLK AGTAGATTGTATAGCTTTA 961

NM_016218 POLK TATAGATAACTCATCTAAA 962

NM_016218 POLK AAGAACTTTGCAGTGAGCT 963

NMJ)16218 POLK GAATTAGAACAAAGCCGAA 964

NM_016218 POLK TGTGCTATCAATGAGTTCT 965

NM_016218 POLK ACACCTGACGAGGGATGGA 966

NM_016218 POLK TGCATCTACAGTTTCATCT 967

NM_016218 POLK ACACACCTGACGAGGGATG 968

NM_016218 POLK TGGATAGCACAAAGGAGAA 969

NMJ ⁾ 16218 POLK AGGGTGCATCAGTCTGGAA 970

NMJ ⁾ 16218 POLK TATAGCTTTAGTAGATACT 971

NMJ ⁾ 16218 POLK TGTTTCTACTGCAGAAGAA 972

NMJ ⁾ 16218 POLK GTTGTTTCTACTGCAGAAG 973

NMJ ⁾ 16218 POLK CTGACAAAGATAAGTTTGT 974

NMJ116218 POLK GCATCAGTCTGGAAGCCTT 975

NMJ ⁾ 16218 POLK CTCAGGATCTACAGAAAGA 976

NMJ ⁾ 16218 POLK AAGGAGATTTGGTGTTCGT 977

NMJ) 16218 POLK TAGTGCACATTGACATGGA 978

6.6. EXAMPLE 6: DESIGNING SHORT HAIRPIN RNAS FOR SILENCING

GENES

This example illustrates an exemplary embodiment of shRNA design. Although several effective approaches exist for the design of siRNAs (Aza-Blanc et al., 2003, MoI. Cell, 12:627-637; Reynolds et al., 2004, Nature Biotech. 22:326-330; Hsieh et al., 2004, Nucleic Acids Res. 32:893-901 ; Ui-Tei et al., 2004, Nucleic Acids Res. 32:936-948), optimal approaches for selecting short-hairpin RNAs (shRNAs) remain unknown. To remedy this deficiency, a comprehensive set of randomly selected shRNAs was designed and tested to systematically determine features that contribute to efficacy in shRNA gene silencing.

In this example, a simple hairpin structure having a single-stranded RNA encoding a 19-bp stem with a 9-nucleotide loop was used. This choice was based on a comparison of simple hairpins with hairpins designed to mimic a mir30 microRNA precursor, both selected by the siRNA design algorithm that only takes into account the silencing efficiency of the siRNA (the "second generation siRNA algorithm"). In general, despite the difference in shRNA core sequence context, the overall performance of these two sets of constructs was nearly identical, indicating that the actual sequence of the hairpin is an important factor in determining efficacy. shRNAs designed by the second generation siRNA design algorithms have a median silencing of roughly 66%. This performance

contrasts with the siRNAs designed by these algorithms whose median silencing is around 80%-90% for well expressed genes.

Figure 24 shows a table exhibiting the genes that were targeted in this example. In selecting these targets, a broad collection of genes representing a variety of expression levels in the cell as well as a range of overall GC contents was chosen. shRNAs that were distributed across coding regions and that varied in their own GC content were also selected. Because such parameters figured into the selection of shRNA core sequences, this approach was referred to as pseudorandom design.

The selected shRNA sequence was cloned into an expression vector that contained an RNA polymerase III promoter, the human Hl promoter. The vector was transiently transfected the cloned vectors into HeLa cells. After 24 hr, total RNA was harvested from the transfected cells and the silencing of target genes was measured by quantitative RT- PCR (qRT-PCR). Silencing in samples transfected with shRNA-containing constructs was compared with silencing in samples transfected with an empty vector control. In analyzing the data, the GC content of effective vs. ineffective shRNAs was first compared across the sense 19mer core sequence (Figure 21). In comparing these results to an analogous study of siRNAs, the preference for GC content asymmetry across the core sequence was found to be essentially the same. These results support the notion that processed shRNAs enter the RNA-induced silencing complex (RISC) via the same mechanisms as transfected siRNAs and, therefore, effective siRNA algorithms are an appropriate starting point for designing shRNAs.

To determine whether other features might contribute to shRNA efficacy, the distribution of each of the four individual nucleotides (G, C, A, and T) in effective vs. ineffective shRNAs across the 19mer core sequence was examined (Figure 22). The distribution was then compared to that of the siRNA preferences. For the most part, the base preferences across the 19mer were similar, except in the first position and/or the second position. As indicated in Figure 22 (top left), there was a very strong preference for a G at position 1 and/or 2 of the shRNA. Such a preference was not observed for siRNAs. This preference may reflect a start site preference of the Hl RNA polymerase III promoter. A preference for a purine at the start site of Pol III promoters has been previously reported (Fruscoloni et al, 1995, Nucleic Acids Res, 23(15): 2914-2918; Mattaj et al, 1988, Cell, 55(3): 435-442; Zecherle et al, 1996, MoI Cell Biol. 16: 5801-5810). The efficiency of overall silencing correlates to the first and/or second position in the following order: G>T>A>C.

To test if limiting the first position of the 19mer to a G nucleotide, an improvement in overall shRNA performance may be observed, shRNAs targeting a subset of genes from the pseudorandom study was designed and their performance was compared with that of shRNAs designed by second generation designs. As shown in the histograms in Figure 23, a set of shRNAs designed with such an siRNA design algorithm (version 3, RSTA siRNA V3, with the additional requirement of a G start nucleotide (Gl)) had a median percent remaining mRNA in the range of 22%-27% (i.e., median silencing of ~70%-80%), whereas shRNAs designed by the second generation algorithm (RSTA siRNA V2) had a median percent remaining of about 34% (i.e., median silencing of -66%). In contrast to these two approaches, the pseudorandom design gave a median percent remaining in the realm of 46%-50% percent. The differences in performance between these sets were all statistically significant. Because early siRNA designs was based on an essentially random approach outlined by Tuschl and colleagues (Mattaj et ah, 1988, Cell, 55(3):435-442), it is expected that libraries based on Tuschl "rules" (e.g., a library constructed by the Broad Institute) will perform about as well as our pseudorandom set.

To summarize, in an effort to improve short hairpin RNA (shRNA) design, a large set of randomly designed shRNAs was tested to uncover features that contribute to efficacy in gene silencing. Analysis of data generated from transient transfection of plasmids expressing these sequences showed that effective shRNAs have the same requirements for GC content asymmetry as small interfering RNAs (siRNAs). A very strong preference for a G nucleotide in position 1 and/or 2 of the shRNA sense strand was also observed. On the basis of these results, a method for designing shRNA was developed which incorporated the second generation algorithm for siRNA selection with the added constraint of a G in position 1 and/or 2. shRNAs designed by this improved approach silenced their target genes more effectively than shRNAs to the same targets generated randomly or by use of algorithms for siRNA design, e.g., the second generation algorithm, alone. The example shows that shRNAs can be effectively designed by use of an siRNA selection algorithm with some minor modifications. The example also shows that despite of the additional steps involved in incorporation of shRNAs into the RNA interference transcript-silencing machinery, GC content asymmetry remains a primary determinant of performance.

7. REFERENCES CITED

All references cited are incorporated herein by reference in their entireties and for all purposes to the same extent as if each individual publication or patent or patent application was specifically and individually indicated to be incorporated by reference in its entirety for all purposes.

Many modifications and variations of the present invention can be made without departing from its spirit and scope, as will be apparent to those skilled in the art. The specific embodiments described herein are offered by way of example only, and the invention is to be limited only by the terms of the appended claims along with the full scope of equivalents to which such claims are entitled.