BREAST TUMOUR GRADING

CROSS REFERENCE TO RELATED APPLICATIONS

This application claims priority from Singapore Patent Application No 200607354- 8. Reference is also made to US provisional application serial number 60/862,519 filed October 23 ^rd , 2007.

The foregoing applications, and each document cited or referenced in each of the present and foregoing applications, including during the prosecution of each of the foregoing applications ("application and article cited documents"), and any manufacturer's instructions or catalogues for any products cited or mentioned in each of the foregoing applications and articles and in any of the application and article cited documents, are hereby incorporated herein by reference. Furthermore, all documents cited in this text, and all documents cited or reference in documents cited in this text, and any manufacturer's instructions or catalogues for any products cited or mentioned in this text or in any document hereby incorporated into this text, are hereby incorporated herein by reference. Documents incorporated by reference into this text or any teachings therein may be used in the practice of this invention. Documents incorporated by reference into this text are not admitted to be prior art.

FIELD

The present invention relates to the fields of medicine, cell biology, molecular biology and genetics. More particularly, the invention relates to a method of assigning a grade to a breast tumour which reflects its aggressiveness.

BACKGROUND

The effective treatment of cancer depends, to a large extent, on the accuracy with which malignant tissue can be subtyped according to clinicopathological features that reflect disease aggressiveness.

Some clinical subtypes, despite phenotypic homogeneity, are associated with substantial clinical heterogeneity (e.g., refractory response to treatment) confounding their clinical meaning. Recent studies using DNA microarray technology suggest that such clinical heterogeneity may be resolvable at the molecular level (1-4). Indeed, some have demonstrated that gene expression signatures underlying specific biological properties of cancer cells may be superior indicators of clinical subtypes with robust prognostic value (1, 2). Thus, global analysis of gene expression has the potential to uncover molecular

determinants of clinical heterogeneity providing a more objective and biologically-rational approach to cancer subtyping.

Accordingly, there is a need in the art for gene markers which are diagnostic or reflective of tumourigenicity.

In breast cancer, histologic grade is an important parameter for classifying tumours into morphological subtypes informative of patient risk. Grading seeks to integrate measurements of cellular differentiation and replicative potential into a composite score that quantifies the aggressive behaviour of the tumour.

The most studied and widely used method of breast tumour grading is the Elston- Ellis modified Scarff, Bloom, Richardson grading system, also known as the Nottingham grading system (NGS) (5, 6, Haybittle et al, 1982). The NGS is based on a phenotypic scoring procedure that involves the microscopic evaluation of morphologic and cytologic features of tumour cells including degree of tubule formation, nuclear pleomorphism and mitotic count (6). The sum of these scores stratifies breast tumours into Grade I (Gl) (well-differentiated, slow-growing), Grade II (G2) (moderately differentiated), and Grade III (G3) (poorly-differentiated, highly-proliferative) malignancies.

Multivariate analyses in large patient cohorts have consistently demonstrated that the histologic grade of invasive breast cancer is a powerful prognostic indicator of disease recurrence and patient death independent of lymph node status and tumour size (6-9). Untreated patients with Gl disease have a -95% five-year survival rate, whereas those with G2 and G3 malignancies have survival rates at 5 years of ~75% and ~50%, respectively.

However, the value of histologic grade in patient prognosis has been questioned by reports of substantial inter-observer variability among pathologists (10-13) leading to debate over the role that grade should play in therapeutic planning (14, 15). Furthermore, where the prognostic significance of Gl and G3 disease is of more obvious clinical relevance, it is less clear what the prognostic value is of the more heterogeneous, moderately differentiated Grade II tumours, which comprise approximately 50% of all breast cancer cases (9, 15, 16).

There is therefore a need for methods which are capable of discriminating between heterogenous tumour grades, particularly Grade II breast tumours.

SUMMARY

We have now demonstrated that a gene expression signature comprising one or more of a set of 232 genes, represented by 264 probesets (e.g., Affymetrix probesets), is capable of discriminating between high and low grade tumours. Such a gene expression signature may be used to provide an objective and clinically valuable measure of tumour grade.

We further describe a novel strategy of clinical class discovery that combines gene discovery and class prediction algorithms with patient survival analysis, and between- group statistical analyses of conventional clinical markers and gene ontologies represented by differentially expressed genes.

Our findings show that the genetic reclassification of histologic grade reveals new clinical subtypes of invasive breast cancer and can improve therapeutic planning for patients with moderately differentiated tumours.

Furthermore, our results support the view that tumours of low and high grade, as defined genetically, may reflect independent pathobiological entities rather than a continuum of cancer progression.

According to a 1 ^st aspect of the present invention, we provide a method of assigning a grade to a breast tumour, which grade is indicative of the aggressiveness of the tumour, the method comprising detecting the expression of a gene selected from the genes set out in Table Dl (SWS Classifier 0).

There is provided, according to a 2 ^nd aspect of the present invention, a method of classifying a histological Grade 2 tumour into a low aggressiveness tumour or a high aggressiveness tumour, the method comprising assigning a grade to the histological Grade 2 tumour according to the 1 ^st aspect of the invention.

We provide, according to a 3 ^rd aspect of the present invention, a method of predicting a survival rate for an individual with a histological Grade 2 breast tumour, the method comprising assigning a grade to the breast tumour by a method according to any preceding aspect of the invention.

As a 4 ^th aspect of the present invention, there is provided a method of prognosis of an individual with a breast tumour, the method comprising assigning a grade to the breast tumour by a method as described,

We provide, according to a 5 ^th aspect of the present invention, a method of diagnosis of aggressive breast cancer in an individual, the method comprising assigning a

grade indicative of high aggressiveness to a breast tumour of the individual by a method as described.

The present invention, in a 6 ^th aspect, provides a method of choosing a therapy for an individual with breast cancer, the method comprising assigning a grade to the breast tumour by a method as described, and choosing an appropriate therapy based on the aggressiveness of the breast tumour.

In a 7 ^{l 1} aspect of the present invention, there is provided a method of treatment of an individual with breast cancer, the method comprising assigning a grade to the breast tumour by a method as described, and administering an appropriate therapy to the individual based on the aggressiveness of the breast tumour.

According to an 8* aspect of the present invention, we provide a method of determining the likelihood of success of a particular therapy on an individual with a breast tumour, the method comprising comparing the therapy with the therapy determined by a such a method.

We provide, according to a 9* aspect of the invention, a method of assigning a breast tumour patient into a prognostic group, the method comprising applying the Nottingham Prognostic Index to a breast tumour, in which the histologic grade score of the breast tumour is replaced by a grade obtained by a method as described.

There is provided, in accordance with a 10 ^th aspect of the present invention, a method of assigning a breast tumour patient into a prognostic group, the method comprising deriving a score which is the sum of the following: (a) (0.2 x tumour size in cm); (b) tumour grade in which the tumour grade is assigned by a method as described; and (c) lymph node stage; in which the tumour size and the lymph node stage are determined according to the Nottingham Prognostic Index, in which a patient with a score of 2.4 or less is categorised to a EPG (excellent prognostic group), a patient with a score of less than 3.4 is categorised to a GPG (good prognostic group), a patient with a score of between 3.4 and 5.4 is categorised to a MPG (moderate prognostic group), a patient with a score of greater than 5.4 is categorised to a PPG (poor prognostic group).

As an 11 ^th aspect of the invention, we, provide a method of determining whether a breast tumour is a metastatic breast tumour, the method comprising assigning a grade to the breast tumour by a method as described.

We provide, according to a 12 ^th aspect of the invention, a method of identifying a molecule capable of treating or preventing breast cancer, the method comprising: (a) grading a breast tumour; (b) exposing the breast tumour to a candidate molecule; and (c)

detecting a change in tumour grade; in which the grade or change thereof, or both, is assigned by a method as described.

According to a 13 ^th aspect of the present invention, we provide a molecule identified by such a method.

There is provided, according to a 14 ^th aspect of the present invention, use of such a molecule in a method of treatment or prevention of cancer in an individual.

We provide, according to a 15 ^th aspect of the present invention, a method of treatment of an individual suffering from breast cancer, the method comprising modulating the expression of a gene set out in Table Dl (SWS Classifier 0).

According to a 16 ^th aspect of the present invention, we provide a method of determining the proliferative state of a cell, the method comprising detecting the expression of a gene selected from the genes set out in Table Dl (SWS Classifier 0), in which: (a) a high level of expression of a gene which is annotated "3" in Column 7 ("Grade with Higher Expression") indicates a highly proliferative cell; (b) a high level of expression of a gene which is annotated "1" in Column 7 ("Grade with Higher Expression") indicates a non-proliferating cell or a slow-growing cell; (c) a low level of expression of a gene which is annotated "3" in Column 8 ("Grade with Lower Expression") indicates a highly proliferative cell; and (d) a low level of expression of a gene which is annotated "1" in Column 8 ("Grade with Lower Expression") indicates a non-proliferating cell or a slow-growing cell.

According to a 17 ^th aspect of the present invention, we provide a combination comprising the genes set out in Table Dl (SWS Classifier 0). We provide, according to an 18 ^th aspect of the present invention, a combination comprising the probesets set out in Table Dl (SWS Classifier 0). According to a 19 ^th aspect of the present invention, we provide a combination comprising the genes set out in the above aspects of the invention. As an 20 ^th aspect of the invention, we provide a combination comprising the probesets set out in the above aspects of the invention. According to a 21 ^st aspect of the present invention, we provide a combination according to any of the above aspects of the invention in the form of an array. According to a 21 ^st aspect of the present invention, we provide a combination according to the above aspects of the invention in the form of a microarray.

There is provided, according to a 22 ^nd aspect of the present invention, a kit comprising such a combination, array or microarray, together with instructions for use in a method as described. We provide, according to a 23 ^rd aspect of the present invention, use of such a combination, array or a microarray or kit in a method as described.

The method may comprise a method of assigning a grade to a breast tumour as described.

As a 24 ^{f l} aspect of the present invention, there is provided a computer implemented method of assigning a grade to a breast tumour, the method comprising processing expression data for one or more genes set out in Table Dl (SWS Classifier 0) and obtaining a grade indicative of aggressiveness of the breast tumour.

We provide, according to a 25 ⁿ aspect of the present invention, a program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine to perform method of assigning a grade to a breast tumour, the method comprising: processing expression data for one or more genes set out in Table Dl (SWS Classifier 0); and obtaining a grade indicative of aggressiveness of the breast tumour.

The practice of the present invention will employ, unless otherwise indicated, conventional techniques of chemistry, molecular biology, microbiology, recombinant DNA and immunology, which are within the capabilities of a person of ordinary skill in the art. Such techniques are explained in the literature. See, for example, J. Sambrook, E. F. Fritsch, and T. Maniatis, 1989, Molecular Cloning: A Laboratory Manual, Second Edition, Books 1-3, Cold Spring Harbor Laboratory Press; Ausubel, F. M. et al. (1995 and periodic supplements; Current Protocols in Molecular Biology, ch. 9, 13, and 16, John Wiley & Sons, New York, N. Y.); B. Roe, J. Crabtree, and A. Kahn, 1996, DNA Isolation and Sequencing: Essential Techniques, John Wiley & Sons; J. M. Polak and James O 'D. McGee, 1990, In Situ Hybridization: Principles and Practice; Oxford University Press; M. J. Gait (Editor), 1984, Oligonucleotide Synthesis: A Practical Approach, IrI Press; D. M. J. Lilley and J. E. Dahlberg, 1992, Methods of Enzymology: DNA Structure Part A: Synthesis and Physical Analysis of DNA Methods in Enzymology, Academic Press; Using Antibodies : A Laboratory Manual : Portable Protocol NO. I by Edward Harlow, David Lane, Ed Harlow (1999, Cold Spring Harbor Laboratory Press, ISBN 0-87969-544-7); Antibodies : A Laboratory Manual by Ed Harlow (Editor), David Lane (Editor) (1988, Cold Spring Harbor Laboratory Press, ISBN 0-87969-314-2), 1855, Lars-Inge Larsson "Immunocytochemistry: Theory and Practice", CRC Press inc., Baca Raton, Florida, 1988, ISBN 0-8493-6078-1, John D. Pound (ed); "Immunochemical Protocols, vol 80", in the series: "Methods in Molecular Biology", Humana Press, Totowa, New Jersey, 1998, ISBN 0-89603-493-3, Handbook of Drug Screening, edited by Ramakrishna Seethala, Prabhavathi B. Fernandes (2001, New York, NY, Marcel Dekker, ISBN 0-8247-0562-9); Lab Ref: A Handbook of Recipes, Reagents, and Other Reference Tools for Use at the Bench, Edited Jane Roskams and Linda Rodgers, 2002, Cold Spring Harbor Laboratory, ISBN 0-87969-630-3; and The Merck Manual of Diagnosis and Therapy (17th Edition,

Beers, M. H., and Berkow, R, Eds, ISBN: 0911910107, John Wiley & Sons). Each of these general texts is herein incorporated by reference.

BRIEF DESCRIPTION OF THE FIGURES

Figure 1. Schema of discovery and validation of the genetic G2a and G2b breast cancer groups. SWS: Statistically Weighted Syndromes method; PAM: Prediction Analysis for Microarray method; CER: Class Error Rate Function; p.s. probe setGl : Grade 1; G3: Grade 2;G3: Grade 3; G2a: Grade 2a; G2b: Grade 2b; GO: gene ontology.

Figure 2. Probability (Pr) scores from the SWS classifier. Pr scores (0-1) generated by the class prediction algorithm are shown on the y-axes. Number of tumours per classification exercise is shown on the x-axis. Green indicates Grade 1 tumours; red denotes Grade 3 tumours.

Figure 3. Survival differences between G2a and G2b genetic grade subtypes. Kaplan-Meier survival curves for G2a and G2b subtypes are shown superimposed on survival curves of histologic grades 1, 2, and 3 (see key). Uppsala cohort survival curves are shown for all patients (A), patients who did not receive systemic therapy (B), patients treated with systemic therapy (C), and patients with ER+ disease who received anti- estrogen therapy only (D). Stockholm cohort survival curves are shown for patients treated with systemic therapy (E) and those with ER+ cancer treated with anit-estrogen therapy only (F). The p-value (likelihood ratio test) reflects the significance of the hazard ratio between the G2a and G2b curves.

Figure 4. Expression profiles of the top 264 grade (G1-G3) associated gene probesets. Gene probesets (rows) and tumours (columns) were hierarchically clustered by average linkage (Pearson correlation), then tumours were grouped according to grade while maintaining original cluster order within groups. Red reflects above mean expression, green denotes below mean expression, and black indicates mean expression. The degree of color saturation reflects the magnitude of expression relative to the mean.

Figure 5. Statistical analysis of clinicopathological markers. Measurements (or percentages of binary measurements) of clinicopathological variables assessed at the time of surgery were compared between different tumour subgroups: Gl vs. G2a, G2a vs. G2b, and G2b vs. G3. P-values are noted below subgroup designations. Average scores (or percentages) within each subgroup are shown as vertical bars with standard deviations.

Figure 6. Stratification of patient risk by classic NPI and ggNPI. (A) Kaplan-Meier survival curves are shown for the classic NPI categories: Good Prognostic Group (GPG); Moderate Prognostic Group (MPG); Poor Prognostic Group (PPG). (B) Kaplan-Meier

survival curves are shown for risk groups determined by the classic NPI (black curves) and the NPI calculated with genetic grade assignments (ggNPI; colored curves). (C) Kaplan-Meier survival curves are shown for patients reclassified by ggNPI (colored curves indicate that reclassified patients have survival curves similar to the good, moderate and poor prognostic groups of the classic NPI (black curves)). (D) The disease-specific survival curves of node negative, untreated patients classified into the Excellent Prognostic Group (EPG) by classic NPI (black curve) or ggNPI (green curve) are compared.

Figure 7. Stratification of patient risk by classic NPI and ggNPI. Kaplan-Meier survival curves are shown for risk groups determined by the classic NPI (A) and the NPI calculated with genetic grade predictions (ggNPI) (B). Survival curves of patients reassigned to new risk groups by the ggNPI are, shown (C). The disease-specific survival curve of the EPG patients (by classic NPI) is compared to that of patients identified as EPG exclusively by the ggNPI (D). Classic NPI curves from (A) are shown superimposed on (B-D).

DETAILED DESCRIPTION

BREAST TUMOUR GRADING

We have identified a number of genes whose expression is indicative of breast tumour aggressiveness. Accordingly, we provide for methods of grading breast tumours, and therefore assigning a measure of their aggressiveness, by detecting the level of expression of one or more of these genes. The genes are provided in a number of gene sets, or classifiers.

In a general aspect, we provide for the detection of any one or more of a small set of 264 gene probesets, which we term the "SWS Classifier 0". This classifier represents 232 genes. In some embodiments, the expression of all of the 264 gene probesets are detected. For example, the expression of all the 232 genes represented by such probesets may be detected.

The genes comprised in this classifier are set out in Table Dl in the section "SWS Classifier 0" below, in Table Sl in Example 20, as well as in Appendix Al. This and the other tables D2, D3, D4 and D5 (see below) contain the GenBank ID and the Gene Symbol of the gene, as well as the "Affi ID", or the "Affymetrix ID" number of a probe. Affymetrix probe set IDs and their corresponding oligonucleotide sequences, as well as the GenBank mRNA sequences they are designed from, can be accessed on the world wide web at the ADAPT website

(http://bioinfoi-matics.picr.man.ac.uk/adapt/ProbeToGene. adapt) hosted by the Paterson Institute for Cancer Research.

In such an embodiment, therefore, our method comprises determining the expression level of at least one of the genes of the 264 gene probesets (for example, at least one of the 232 genes) in the classifier which we term the "SWS Classifier 0". More than one, for example, a plurality of the genes of such a set may also be detected. The 264 gene probesets of the SWS Classifier 0 gene set are set out in Table Dl below.

In some embodiments, the expression level of more than one gene is detected. For example, the expression level of 5 or more genes may be detected. The expression level of a plurality of genes may therefore be determined. In some embodiments, the expression level of all 264 gene probesets (for example, the expression level of all 232 genes) may be detected, though it will be clear that this does not need to be so, and a smaller subset may be detected.

We therefore provide for the detection of one or more, a plurality or all, of subsets comprising 17 genes and several subsets of 5-17 genes from the 264 gene probesets.

Thus, alternatively, or in addition, our method may comprise determining the expression level of at least one, a plurality, or all of the genes of an 5 gene set which we term the "SWS Classifier 1". The 5 genes of the SWS Classifier 1 gene set are set out in Table D2 below.

In other embodiments, our method may comprise determining the expression level of at least one, a plurality, or all of the genes of an 17 gene set which we term the "SWS Classifier 2". The 17 genes of the SWS Classifier 2 gene set are set out in Table D3 below.

In other embodiments, our method may comprise determining the expression level of at least one, a plurality, or all of the genes of an 7 gene set which we term the "SWS Classifier 3". The 7 genes of the SWS Classifier 3 gene set are set out in Table D4 below.

In other embodiments, our method may comprise determining the expression level of at least one, a plurality, or all of the genes of an 7 gene set which we term the "SWS Classifier 4". The 7 genes of the SWS Classifier 4 gene set are set out in Table D5 below.

In specific embodiments, the methods comprise detection of the expression level of all of the genes in the gene set of interest. For example, all 5 genes in the "SWS Classifier 1" are detected, all 17 genes in the "SWS Classifier 2" are detected, all 7 genes in the "SWS Classifier 3" are detected and all 7 genes of the SWS Classifier 4 gene set are detected in these embodiments.

Where the SWS Classifier I ₃ the SWS Classifier 2, the SWS Classifier 3 or the SWS Classifier 4 are used, each of Tables D2, D3, D4 and D5 provide indications of the grades to be assigned to the tumour depending on the level of expression of the relevant gene which is detected (in Columns 7 and 8 respectively).

Thus, the tables also contain columns showing the grades associated with high and low levels of expression of a particular gene, in Columns 7 and 8 of Table Dl for example. Thus, for example, the gene Barren homolog (Drosophilά) is annotated to the effect that the "Grade with Higher Expression" is 3, while the "Grade with Lower Expression" is 1. Accordingly, our method provides that the tumour has a grade of 3 if a high level of expression of Barren homolog (Drosophilά) is detected in or from the tumour. If a low level of this gene is detected in or from the tumour, then a grade of 1 may be assigned to that tumour.

Detection of gene expression, for example for tumour grading, may suitably be done by any means as known in the art, and as described in further detail below.

The methods described here for gene expression analysis and tumour grading may be automated, or partially or completely controlled by a controller such as a microcomputer. Thus, any of the methods described here may comprise computer implemented methods of assigning a grade to a breast tumour. For example, such a method may comprise processing expression data for one or more genes set out in Table Dl (SWS 0 Classifier) and obtaining a grade indicative of aggressiveness of the breast tumour.

The methods described here are suitably capable of classifying a breast tumour to an accuracy of at least 85%, at least 90% accuracy, or at least 95% accuracy, with reference to the grade obtained by conventional means, such as for example grading of the breast tumour by histological grading. For example, the methods may be capable of classifying tumours with grades corresponding to histological Grade 1 and histological Grade 3 tumours with an accuracy of 70% or above, 80% or above, or 90% or above.

DETECTION OF HIGHER AND LOWER EXPRESSION

In a refinement of our methods, we provide for a "cut-off level of expression, by which the expression of a gene in or from a tumour may be judged in order to establish whether the expression is at a "high" level, or at a "low" level. The cut-off level is set out in Column 9 of Tables Dl, D2, D3, D4 and D5.

Accordingly, in some embodiments, our methods include assigning a grade based on whether the level of expression falls below or exceeds the cut-off. In some embodiments, the cut-off values are determined as the natural log transform normalised

signal intensity measurement for Affymetrix arrays. In such embodiments, the cut-off values may be determined as a global mean normalisation with a scaling factor of 500.

For example, referring back to Table 1, the cut-off level of expression for the gene Barren homolog (Drosophilά) is 5.9167 units (see above and formula (1), Microarray Method). Where a given tumour contains a level of expression of this gene that exceeds this level, then it is determined to be a "high" level of expression. A grade of 3 may then be assigned to that tumour. On the other hand, if the expression of the Barren homologue falls below this cut-off level, then the expression is judged to be a "low" level of expression. A grade of 1 may be assigned to the tumour in this event.

Thus, we provide for a method which comprises detecting a high level of expression of a gene in SWS Classifier 0 and assigning the grade set out in Column 7 of Table Dl to the breast tumour. The method may comprise, or optionally further comprise detecting a low level of expression of the gene and assigning the grade set out in Column 8 of Table Dl to the breast tumour. A high level of expression may be detected if the expression level of the gene is above the expression level set out in Column 9 of Table Dl, and a low level of expression is detected if the expression level of the gene is below that level.

DETECTION OF GENE EXPRESSION

There are various methods by which expression levels of a gene may be detected, and these are known in the art. Examples include RT-PCR, RNAse protection, Northern blotting, Western blotting etc. The gene expression level may be determined at the transcript level, or at the protein level, or both. The detection may be manual, or it may be automated. It is envisaged that any one or a combination of these methods may be employed in the methods and compositions described here.

The detection of expression of a plurality of genes is suitably detected in the form of an expression profile of the plurality of genes, by conventional means known in the art. In some embodiments, the detection is by means of microarray hybridisation.

For example, a sample of a tumour may be taken from a patient and processed for detection of gene expression levels. Gene expression levels may be detected in the form of nucleic acid or protein levels or both, for example. Analysis of nucleic acid expression levels may be suitably performed by amplification techniques, such as polymerase chain reaction (PCR), rolling circle amplification, etc. Detection of expression levels is suitably performed by detecting RNA levels. This can be performed by means known in the art, for example, real time polymerase chain reaction (RT-PCR) or RNAse protection, etc. For this purpose, we provide for sets of one or more primers or primer pairs which are capable of

amplifying any one or more of the genes in the classifiers disclosed herein. Specifically, we provide for a set of primer pairs capable of amplifying all of the genes in the SWS Classifier 0, SWS Classifier 1, SWS Classifier 2, SWS Classifier 3 or SWS Classifier 4 sets.

Suitably, RNA expression levels may be detected by hybridisation to a microchip or array, for example, a microchip or array comprising the genes or probesets corresponding to the specific classifier of interest, as described in the Examples. In some embodiments, the gene expression data or profile is derived from microarray hybridisation to for example an Affymetrix microarray.

Detection of protein levels may be performed by for example, immunoassays including ELISA or sandwich immunoassays using antibodies against the protein or proteins of interest (for example as described in US6664114. The detection may be performed by use of a "dip stick" which comprises impregnated antibodies against polypeptides of interest, such as described in US2004014094.

We provide therefore for sets of one or more antibodies which are capable of binding specifically to any one or more of the proteins encoded by the genes in the classifiers disclosed herein. Specifically, we provide for a set of antibodies capable of amplifying all of the genes in the SWS Classifier O ₅ SWS Classifier 1, SWS Classifier 2, SWS Classifier 3 or SWS Classifier 4 sets.

The grade may be assigned by any suitable method. For example, it may be assigned applying a class prediction algorithm comprising a nearest shrunken centroid method (Tibshirani, et al, 2002, Proc Natl Acad Sci USA. 99(10): 6567-6572) to the expression data of the plurality of genes. The class prediction algorithm may suitably comprise Statistically Weighted Syndromes (SWS) or Prediction Analysis of Microarrays (PAM).

In some embodiments, the grade of the tumour may be assigned by applying a class prediction algorithm comprising one or more of the steps set out here. First, a set of predictor parameters (i.e., probesets) may be obtained, based on predictors which discriminate the histologic tumours Gl and G3. Next, the potentially predictive parameters (i.e. signal intensity values of micro-array) may be recoded to obtain cut-off values for robust discrete-valued variables. The recoding may be done in such a way as to maximize an informativity measure of discrimination ability of the parameter and minimize its instability to the discrimination object (i.e. patients) belonging to distinct classes (i.e. Gl and G3). Then, statistically robust discrete-valued variables and combinations thereof may be selected for further construction of class prediction algorithm. A sum of the statistically weighted discrete-valued variables and combinations thereof may be obtained based on the

Weighted Voting Procedure procedure described in SWS method section. Finally, a predictive outcome (classification) scores of breast cancer subtypes based on the sum for sub-typing (re-classification) histologic G2 tumours may be obtained.

APPLICATION TO GRADE 2 TUMOURS

In suitable embodiments, the method is applied to grade breast tumours which are traditionally graded as Grade 2 by conventional means, such as by histological grading as known in the art. Our method is capable of distinguishing the aggressiveness of tumours within the group of tumours in Grade 2 (which were hitherto thought to be homogenous) into Grade 1 like tumours (i.e., more aggressive) and Grade 3 like tumours (i.e., less aggressive). This is described in detail in the Examples.

Accordingly, we provide for a method of classifying a histological Grade 2 tumour into a low aggressiveness tumour or a high aggressiveness tumour. In other words, we provide a method for reassigning a more precise grading to a tumour which has been graded histologically as a Grade 2 tumour.

Such a method comprises assigning a grade to the histological Grade 2 tumour according to any of the methods described above. For example, the expression of any one or more genes, for example, all the genes, in any of the SWS Classifiers described here may be detected and a grade of 1 or 3 assigned using Columns 7, 8 or 9 individually or in combination, as described above.

Such a tumour which has been reassigned will suitably have one or more characteristics or features of the reassigned grade. The characteristics or features may include one or more histological or morphological features, susceptibility to treatment, rate of growth or proliferation, degree of differentiation, aggressiveness, etc. As an example, the characteristic or feature may comprise aggressiveness.

For example, a histological Grade 2 breast tumour which has been assigned a low aggressiveness grade by the gene expression detection methods described here may suitably have at least one feature of a histological Grade 1 breast tumour. Similarly, a breast tumour assigned a high aggressiveness grade may have at least one feature of a histological Grade 3 breast tumour.

Such a feature may comprise degree of differentiation (e.g., well-differentiated, moderately differentiated or poorly-differentiated). The feature may comprise rate of growth (e.g., slow-growing, fast-growing). The feature may comprise rate of proliferation (e.g., slow-proliferation, highly-proliferative). The feature may comprise likelihood of tumour recurrence post-surgery. The feature may comprise survival rate. The feature may

comprise likelihood of tumour recurrence post-surgery and survival rate. The feature may comprise a disease free survival rate. The feature may comprise susceptibility to treatment.

Accordingly, application of the grading methods described here enables the classification of the histological Grade 2 tumour into a Grade 1 tumour or a Grade 3 tumour, so as to allow the clinician to treat the tumour accordingly in view of its aggressiveness, prognosis, etc.

Such regrading using our methods is suitably capable of classifying histological Grade 2 tumours into Grade 1 like and/ or Grade 3 like tumours with an accuracy of 70% or above, 80% or above, or 90% or above.

The histological grading may be performed by any means known in the art. For example, the breast tissue or tumour may be graded by the Nottingham Grading System (NGS) or the Elston-Ellis Modified Scarff, Bloom, Richardson Grading System, both methods being well known in the art.

The information obtained from the regrading may be used to predict any of the parameters which may be useful to the clinician. The parameter may include, for example, likelihood of tumour metastasis, prognosis of the patient, survival rate, possibility of recovery and recurrence, etc, depending on the grade of the tumour which has been reassigned to the histological Grade 2 tumour. We therefore describe a method of determining whether a breast tumour is a metastatic breast tumour, the method comprising assigning a grade to the breast tumour as described using gene expression data.

We describe a method of predicting a survival rate for an individual with a histological Grade 2 breast tumour, the method comprising assigning a grade to the breast tumour using gene expression data as described. A low aggressiveness grade may suitably indicate a high probability of survival and a high aggressiveness grade may suitably indicate a low probability of survival. We also provide for a method of prognosis of an individual with a breast tumour, the method comprising assigning a grade to the breast tumour by a method as described, and a method of diagnosis of aggressive breast cancer in an individual, the method comprising assigning a grade indicative of high aggressiveness to a breast tumour of the individual by a method as described.

The methods of gene expression analysis may be employed for determining the proliferative state of a cell. For example, such a method may comprise detecting the expression of a gene selected from the genes set out in Table Dl (SWS Classifier 0). Where a high level of expression of a gene which is annotated "3" in Column 7 is detected, this may indicate a highly proliferative cell. Similarly, where a high level of expression of a gene which is annotated "1" in Column 7 is detected, a non-proliferating

cell or a slow-growing cell may be indicated. If a low level of expression of a gene which is annotated "3" in Column 8 is detected, this may indicate a highly proliferative cell and where a low level of expression of a gene which is annotated "1" in Column 8 is detected, this indicates a non-proliferating cell or a slow-growing cell.

The classifiers are described herein as combinations of probesets, and the skilled person will be aware that more than one probeset can correspond to one gene. Accordingly, the SWS Classifier 0 contains 264 probesets which represent 232 genes. It will be clear therefore that the invention encompasses detection of expression level of one or more genes, and/or one or more probesets within the relevant classifiers, or any combination of this.

DIAGNOSIS AND TREATMENT

Suitably, the information obtained by the regarding may also be used by the clinician to recommend a suitable treatment, in line with the grade of the tumour which has been reassigned.

Thus, a tumour which has been reassigned to Grade 1 may require less aggressive treatment than a tumour which has been reassigned to Grade 3, for example. We therefore describe a method of choosing a therapy for an individual with breast cancer, the method comprising assigning a grade to the breast tumour by a method as described herein, and choosing an appropriate therapy based on the aggressiveness of the breast tumour. In general, the method may be employed for the treatment of an individual with breast cancer, by assigning a grade to the breast tumour and administering an appropriate therapy to the individual based on the aggressiveness of the breast tumour.

In general, we disclose a method of treatment of an individual suffering from breast cancer, the method comprising modulating the expression of a gene set out in Table Dl (SWS Classifier 0), Table D2 (SWS 1 Classifier), Table D3 (SWS Classifier 2), Table D4 (SWS Classifier 3) and/or Table D5 (SWS Classifier 4).

It will be evident that any of the diagnosis and treatment methods may suitably be combined with other methods of assessing the aggressiveness of the tumour, the patient's health and susceptibility to treatment, etc. For example, the diagnosis or choice of therapy may be determined by further assessing the size of the tumour, or the lymph node stage or both, optionally together or in combination with other risk factors

Specifically, the choice of therapy may be determined by assessing the Nottingham Prognostic Index (NPI). The NPI is described in detail in Haybittle, et al., 1982. In combination with the grading methods described here, the method is suitable for assigning a breast tumour patient into a prognostic group. Such a combined method comprises

deriving a score which is the sum of the following: (a) (0.2 x tumour size in cm); (b) tumour grade in which the tumour grade is assigned by a method according to any of the gene expression detection methods described herein; and (c) lymph node stage; in which the tumour size and the lymph node stage are determined according to the Nottingham Prognostic Index, in which a patient with a score of 2.4 or less is categorised to a EPG (excellent prognostic group), a patient with a score of less than 3.4 is categorised to a GPG (good prognostic group), a patient with a score of between 3.4 and 5.4 is categorised to a MPG (moderate prognostic group), a patient with a score of greater than 5.4 is categorised to a PPG (poor prognostic group).

Alternatively, or in addition, a method of assigning a breast tumour patient into a prognostic group may comprise applying the Nottingham Prognostic Index to a breast tumour, but modified such that the histologic grade score of the breast tumour is replaced by a grade obtained by a gene expression detection method as described in this document.

Other factors which may of course be assessed for determining the choice of therapy may include receptor status, such as oestrogen receptor (ER) or progesterone receptor (PR) status, as known in the art. For example, the choice of therapy may be determined by further assessing the oestrogen receptor (ER) status of the breast tumour.

GENE COMBINATIONS

We further provide for combinations of genes according to the various classifiers disclosed in this document. Such combinations may comprise mixtures of genes or corresponding probes, such as in a form which is suitable for detection of expression. For example, the combination may be provided in the form of DNA in solution.

In other embodiments, a microarray or chip is provided which comprises any combination of genes or probes, in the form of cDNA, genomic DNA, or RNA, within the classifiers. In some embodiments, the microarray or chip comprises all the genes or probes in SWS Classifier 0, SWS Classifier 1, SWS Classifier 2, SWS Classifier 3 or SWS Classifier 4. The genes may be synthesised or obtained by means known in the art, and attached on the microarray or chip by conventional means, as known in the art. Such microarrays or chips are useful in monitoring gene expression of any one or more of the genes comprised therein, and may be used for tumour grading or detection as described here.

We further describe a probe set consisting of a probe or probes having Affymetrix ID numbers as set out in Column 6 of Table Dl, Table D2, Table D3, Table D4 or Table D5. Specifically, we describe an array, such as a microarray, comprising the probesets set out in Table Dl (SWS Classifier 0). We also describe an array such as a microarray

comprising the genes or probesets set out in Table D2 (SWS 1 Classifier), an array such as a microarray comprising the genes or probesets set out in Table D3 (SWS Classifier 2), an array such as a microarray comprising the genes or probesets set out in Table D4 (SWS3 Classifier), and an array such as a microarray comprising the genes or probesets set out in Table D5 (SWS Classifier 4).

The probes or probe sets are suitably synthesised or made by means known in the art, for example by oligonucleotide synthesis, and may be attached to a microarray for easier carriage and storage. They may be used in a method of assigning a grade to a breast tumour as described herein.

We describe the use of Statistically Weighted Syndromes (SWS) on gene expression data which may comprise microarray gene expression data. We describe the use of SWS for gene discovery. We further describe such use in combination with Prediction Analysis of Microarrays (PAM). We describe the use of SWS to identify gene sets diagnostic of cancer status, such as breast cancer status or proliferative status.

SCREENING

The methods and compositions described here may be used for identifying molecules capable of treating or preventing breast cancer, which may be used as drugs for cancer treatment. Such a method comprises: (a) grading a breast tumour as described using gene expression data; (b) exposing the breast tumour to a candidate molecule; and (c) detecting a change in tumour grade. The change in tumour grade is suitably determined by grading a breast tumour as described using gene expression data before and after exposure of the breast tumour to a candidate molecule. We provide molecule identified by such a method, for example for use in breast cancer treatment.

Particular screening applications relate to the testing of pharmaceutical compounds in drug research. The reader is referred generally to the standard textbook "In vitro Methods in Pharmaceutical Research", Academic Press, 1997, and U.S. Pat. No. 5,030,015). Assessment of the activity of candidate pharmaceutical compounds generally involves combining the breast cancer cells with the candidate compound, determining any change in the tumour grade, as determined by the gene expression detection methods described herein of the cells that is attributable to the compound (compared with untreated cells or cells treated with an inert compound), and then correlating the effect of the compound with the observed change.

The screening may be done, for example, either because the compound is designed to have a pharmacological effect on certain cell types such as tumour cells, or because a compound designed to have effects elsewhere may have unintended side effects. Two or

more drugs can be tested in combination (by .combining with the cells either simultaneously or sequentially), to detect possible drug-drug interaction effects. In some applications, compounds are screened initially for potential toxicity (Castell et al, pp. 375- 410 in "In vitro Methods in Pharmaceutical Research," Academic Press, 1997). Cytotoxicity can be determined in the first instance by the effect on cell viability, survival, morphology, and expression or release of certain markers, receptors or enzymes. Effects of a drug on chromosomal DNA can be determined by measuring DNA synthesis or repair. [ Hjthymidine or BrdU incorporation, especially at unscheduled times in the cell cycle, or above the level required for cell replication, is consistent with a drug effect. The reader is referred to A. Vickers (PP 375-410 in "In vitro Methods in Pharmaceutical Research," Academic Press, 1997) for further elaboration.

Candidate molecules subjected to the assay and which are found to be of interest may be isolated and further studied. Methods of isolation of molecules of interest will depend on the type of molecule employed, whether it is in the form of a library, how many candidate molecules are being tested at any one time, whether a batch procedure is being followed, etc.

The candidate molecules may be provided in the form of a library. In an embodiment, more than one candidate molecule is screened simultaneously. A library of candidate molecules may be generated, for example, a small molecule library, a polypeptide library, a nucleic acid library, a library of compounds (such as a combinatorial library), a library of antisense molecules such as antisense DNA or antisense RNA, an antibody library etc, by means known in the art. Such libraries are suitable for high- throughput screening. Tumour cells may be exposed to individual members of the library, and the effect on tumour grade, if any, cell determined. Array technology may be employed for this purpose. The cells may be spatially separated, for example, in wells of a microtitre plate.

In an embodiment, a small molecule library is employed. By a "small molecule", we refer to a molecule whose molecular weight may be less than about 50 kDa. In particular embodiments, a small molecule has a molecular weight may be less than about 30 kDa, such as less than about 15 kDa, or less than 10 kDa or so. Libraries of such small molecules, here referred to as "small molecule libraries" may contain polypeptides, small peptides, for example, peptides of 20 amino acids or fewer, for example, 15, 10 or 5 amino acids, simple compounds, etc.

Alternatively or in addition, a combinatorial library, as described in further detail below, may be screened for candidate modulators of tumour function.

COMBINATORIAL LIBRARIES

Libraries, in particular, libraries of candidate molecules, may suitably be in the form of combinatorial libraries (also known as combinatorial chemical libraries).

A "combinatorial library", as the term is used in this document, is a collection of multiple species of chemical compounds that consist of randomly selected subunits. Combinatorial libraries may be screened for molecules which are capable of changing the choice by a stem cell between the pathways of self-renewal and differentiation.

Various combinatorial libraries of chemical compounds are currently available, including libraries active against proteolytic and non-proteolytic enzymes, libraries of agonists and antagonists of G-protein coupled receptors (GPCRs), libraries active against non-GPCR targets (e.g., integrins, ion channels, domain interactions, nuclear receptors, and transcription factors) and libraries of whole-cell oncology and anti-infective targets, among others. A comprehensive review of combinatorial libraries, in particular their construction and uses is provided in Dolle arid Nelson (1999), Journal of Combinatorial Chemistry, VoI 1 No 4, 235-282. Reference is also made to Combinatorial peptide library protocols (edited by Shmuel Cabilly, Totowa, NJ. : Humana Press, cl998. Methods in Molecular Biology ; v. 87). Specific combinatorial libraries and methods for their construction are disclosed in United States Patent 6,168,914 (Campbell, et al), as well as in Baldwin et al. (1995), "Synthesis of a Small Molecule Library Encoded with Molecular Tags," J. Am. Chem. Soc. 117:5588-5589, and in the references mentioned in those documents.

Further references describing chemical combinatorial libraries, their production and use include those available from the URL http://www.netsci.org/Science/Combichem/, including The Chemical Generation of Molecular Diversity. Michael R. Pavia, Sphinx Pharmaceuticals, A Division of Eli Lilly (Published July, 1995); Combinatorial Chemistry: A Strategy for the Future - MDL Information Systems discusses the role its Project Library plays in managing diversity libraries (Published July, 1995); Solid Support Combinatorial Chemistry in Lead Discovery and SAR Optimization, Adnan M. M. Mj alii and Barry E. Toyonaga, Ontogen Corporation (Published July, 1995); Non-Peptidic Bradykinin Receptor Antagonists From a Structurally Directed Non-Peptide Library. Sarvajit Chakravarty, Babu J. Mavunkel, Robin Andy, Donald J. Kyle*, Scios Nova Inc. (Published July, 1995); Combinatorial Chemistry Library Design using Pharmacophore Diversity Keith Davies and Clive Briant, Chemical Design Ltd. (Published July, 1995); A Database System for Combinatorial Synthesis Experiments - Craig James and David Weininger, Daylight Chemical Information Systems, Inc. (Published July, 1995); An Information Management Architecture for Combinatorial Chemistry, Keith Davies and Catherine White, Chemical Design Ltd. (Published July, 1995); Novel Software Tools for

Addressing Chemical Diversity, R. S. Pearlman, Laboratory for Molecular Graphics and Theoretical Modeling, College of Pharmacy, University of Texas (Published June/July, 1996); Opportunities for Computational Chemists Afforded by the New Strategies in Drug Discovery: An Opinion, Yvonne Connolly Martin, Computer Assisted Molecular Design Project, Abbott Laboratories (Published June/July, 1996); Combinatorial Chemistry and Molecular Diversity Course at the University of Louisville: A Description, Arno F. Spatola, Department of Chemistry, University of Louisville (Published June/July, 1996); Chemically Generated Screening Libraries: Present and Future. Michael R. Pavia, Sphinx Pharmaceuticals, A Division of Eli Lilly (Published June/July, 1996); Chemical Strategies For Introducing Carbohydrate Molecular Diversity Into The Drug Discovery Process.. Michael J. Sofia, Transcell Technologies Inc. (Published June/July, 1996); Data Management for Combinatorial Chemistry. Maryjo Zaborowski, Chiron Corporation and Sheila H. DeWitt, Parke-Davis Pharmaceutical Research, Division of Warner-Lambert Company (Published November, 1995); and The Impact of High Throughput Organic Synthesis on R&D in Bio-Based Industries, John P. Devlin (Published March, 1996).

Techniques in combinatorial chemistry are gaming wide acceptance among modern methods for the generation of new pharmaceutical leads (Gallop, M. A. et al., 1994, J. Med. Chem. 37:1233-1251; Gordon^ E. M. et al., 1994, J. Med. Chem. 37:1385- 1401.). One combinatorial approach in use is based on a strategy involving the synthesis of libraries containing a different structure on each particle of the solid phase support, interaction of the library with a soluble receptor, identification of the 'bead" which interacts with the macromolecular target, and determination of the structure carried by the identified "bead ^λ (Lam, K. S. et al., 1991, Nature 354:82-84). An alternative to this approach is the sequential release of defined aliquots of the compounds from the solid support, with subsequent determination of activity in solution, identification of the particle from which the active compound was released, and elucidation of its structure by direct sequencing (Salmon, S. E. et al., 1993, Proc.Natl.Acad.Sci.USA 90:11708-11712), or by reading its code (Kerr, J. M. et al., 1993, J.Am.Chem.Soc. 115:2529-2531; Nikolaiev, V. et al., 1993, Pept. Res. 6:161-170; Ohlmeyer, M. H. J. et al., 1993, Proc.Natl.Acad.Sci.USA 90:10922-10926). ^'

Soluble random combinatorial libraries may be synthesized using a simple principle for the generation of equimolar mixtures of peptides which was first described by Furka (Furka, A. et al., 1988, Xth International Symposium on Medicinal Chemistry, Budapest 1988; Furka, A. et al., 1988, 14th International Congress of Biochemistry, Prague 1988; Furka, A. et al., 1991, Int. J. Peptide Protein Res. 37:487-493). The construction of soluble libraries for iterative screening has also been described (Houghten, R. A. et al.1991, Nature 354:84-86). K. S. Lam disclosed the novel and unexpectedly powerful technique of using insoluble random combinatorial libraries. Lam synthesized

random combinatorial libraries on solid phase supports, so that each support had a test compound of uniform molecular structure, and screened the libraries without prior removal of the test compounds from the support by solid phase binding protocols (Lam, K. S. et al., 1991, Nature 354:82-84).

Thus, a library of candidate molecules may be a synthetic combinatorial library (e.g., a combinatorial chemical library), a cellular extract, a bodily fluid (e.g., urine, blood, tears, sweat, or saliva), or other mixture of synthetic or natural products (e.g., a library of small molecules or a fermentation mixture). '

A library of molecules may include, for example, amino acids, oligopeptides, polypeptides, proteins, or fragments of peptides or proteins; nucleic acids (e.g., antisense; DNA; RNA; or peptide nucleic acids, PNA); aptamers; or carbohydrates or polysaccharides. Each member of the library can be singular or can be a part of a mixture (e.g., a compressed library). The library may .contain purified compounds or can be "dirty" (i.e., containing a significant quantity of impurities).

Commercially available libraries (e.g., from Affymetrix, ArQuIe, Neose Technologies, Sarco, Ciddco, Oxford Asymmetry, Maybridge, Aldrich, Panlabs, Pharmacopoeia, Sigma, or Tripose) may also be used with the methods described here.

In addition to libraries as described above, special libraries called diversity files can be used to assess the specificity, reliability, or reproducibility of the new methods. Diversity files contain a large number of compounds (e.g., 1000 or more small molecules) representative of many classes of compounds that could potentially result in nonspecific detection in an assay. Diversity files are commercially available or can also be assembled from individual compounds commercially available from the vendors listed above.

ANALYSIS METHOD - RNA PURIFICATION

The breast tumour is surgically resected, processed, and snap frozen. A frozen portion of the tumour is processed for total RNA extraction using the Qiagen RNeasy kit (Qiagen, Valencia, CA). Briefly, frozen tumours are cut into minute pieces, and pieces totalling -50-100 milligrams (mg) are homogenized for 40 seconds in RNeasy Lysis Buffer (RLT). Proteinase K is added, and the ^' samples are incubated for 10 minutes at 55 degrees C, followed by centrifugation and the addition of ethanol. After transferring the supernatant into RNeasy columns, DNase is added. Collected RNA is then assessed for quality using an Agilent 2100 bioanalyzer (Agilent Technologies, Rockville, MD) or by agarose gel. The RNA is stored at minus -70 degrees C.

MlCROARRAY ANALYSIS

Labeled cRNA target is generated for microarray hybridization essentially according to the Affymetrix protocol (Affymetrix, Santa Clara, CA). Briefly, approximately 5 micrograms (μg) of total RNA are reversed transcribed into first-strand cDNA using a T7-linked oligo-dT primer, followed by second strand synthesis. A T7 RNA polymerase is then used to linearly amplify antisense RNA. This "cRNA" is biotinylated and chemically fragmented at 95°C. Ten μg of the fragmented, biotinylated cRNA is hybridized at 45°C for 16 hours to an Affymetrix high-density oligonucleotide GenChip array. The array is then washed and stained with streptavidin-phycoerythrin (10 μg/ml). Signal amplification is achieved using a biotinylated anti-streptavidin antibody. The scanned images are inspected for the presence of artifacts. In case of defects, the hybridization procedure is repeated. Expression values and detection calls are computed from raw data following the procedures outlined for the Affymetrix MAS 5.0 analysis software. Global mean normalization of the gene expression by hybridization signals across all arrays is used to control for differences in chip hybridization signal intensity values. To do that for a given array j(j = 1,2,...M) , we calculated normalization coefficients k _/ (j = 1,2,..., ή) , by the following formula:

^ = « * ln(500)/£ln(a _y ) , (1) ι=\

where n is the number of observed probe sets, a _tJ is the signal intensity value of the i-th

Affymetrix probesets representing a gene expression. Then the natural logarithm of the signal intensity value of the given array/ was multiplied by this normalization coefficient. A normalisation coefficient of 500 is used in determining the cut-offs shown in the Tables in this document.

SWS ANALYSIS

The microarray-derived normalized numerical expression values corresponding to the genetic grade signature genes are used as input for the SWS algorithm.

OTHER METHODS

The RNA purification and microarray analysis methodologies above reflect only our "preffered methods", and that other variants exist that could be used in conjunction with our Process for Predicting Patient Outcome...For example, the starting material could be formalin-fixed paraffin-embedded tumour material instead of fresh frozen material, or the RNA might be extracted using a Cesium Chloride Gradient method, or the RNA could be analyzed by NimbleGen Microarrays that include DNA probes corresponding to our

genes of interest. And it should also be noted that a microarray may not be necessary at all to determine the expression levels of our signature genes, but rather their expression could be quantitatively measured by PCR-based techniques such as real time-PCR.

CLASSIFIERS, GENE SETS AND PROBE SETS

SWS CLASSIFIER O (TABLE Dl)

4>

Ni

to

NS

OO

LO o

I ¹ O

Table Dl: SWS Classifier 0: 264 Probesets.

For any particular gene, where the value of the cell in column 7 "Grade with Higher Expression" is 3, the value of the cell in column 8 "Grade with Lower Expression" is 1, and where the value of cell "Grade with Higher Expression" is I ₅ the value of column "Grade with Lower Expression" is 3. Colum 9 shows the cut off (Optimal Variance Cut-off) expressed as the natural log transform normalised signal intensity measurement for Affymetrix arrays (global mean normalisation with a scaling factor of 500).

SWS CLASSIFIER 1 (TABLE D2)

Table D2. SWS Classifier 1: 6 Probe Sets (5 Genes)

For any particular gene, where the value of the cell in column 7 "Grade with Higher Expression" is 3, the value of the cell in column 8 "Grade with Lower Expression" is 1, and where the value of cell "Grade with Higher Expression" is 1, the value of column "Grade with Lower Expression" is 3. Colum 9 shows the cut off (Optimal Variance Cut-off) expressed as the natural log transform normalised signal intensity measurement for Affymetrix arrays (global mean normalisation with a scaling factor of 500).

SWS CLASSIFIER 2 (TABLE D3)

U) -P-

Table D3. SWS Classifier 2: 18 Probe Sets (17 Genes)

For any particular gene, where the value of the cell in column 7 "Grade with Higher Expression" is 3, the value of the cell in column 8 "Grade with Lower Expression" is 1, and where the value of cell "Grade with Higher Expression" is 1, the value of column "Grade with Lower Expression" is 3. Colum 9 shows the cut off (Optimal Variance Cut-off) expressed as the natural log transform normalised signal intensity measurement for Affymetrix arrays (global mean normalisation with a scaling factor of 500).

Ln

SWS CLASSIFIER 3 (TABLE D4)

Table D4. SWS Classifier 3 : 7 Probe Sets (7 Genes)

For any particular gene, where the value of the cell in column 7 "Grade with Higher Expression" is 3, the value of the cell in column 8 "Grade with Lower Expression" is 1, and where the value of cell "Grade with Higher Expression" is 1, the value of column "Grade with Lower Expression" is 3. Colum 9 shows the cut off (Optimal Variance Cut-off) expressed as the natural log transform normalised signal intensity measurement for Affymetrix arrays (global mean normalisation with a scaling factor of 500).

SWS CLASSIFIER 4 (TABLE D5)

Table D5. SWS Classifier 4: 7 Probe Sets (7 Genes)

For any particular gene, where the value of the cell in column 7 "Grade with Higher Expression" is 3, the value of the cell in column 8 "Grade with Lower Expression" is 1, and where the value of cell "Grade with Higher Expression" is 1, the value of column "Grade with Lower Expression" is 3. Colum 9 shows the cut off (Optimal Variance Cut-off) expressed as the natural log transform normalised signal intensity _s measurement for Affymetrix arrays (global mean normalisation with a scaling factor of 500).

EXAMPLES

Example 1. Materials and Methods: Patients and Tumour Specimens

Clinical characteristics of patient and tumour samples of the Uppsala, Stockholm and Singapore cohorts are summarized in Table El.

Table El. Distribution of patients and tumour characteristics

All cohorts are of unselected populations, and in each case, the original tumour material was collected at the time of surgery and freshly frozen on dry ice or in liquid nitrogen and stored under liquid nitrogen or at -70 ⁰ C.

Example 2. Methods: Details of Uppsala, Singapore and Stockholm

Cohorts

Uppsala Cohort

The Uppsala cohort originally comprised of 315 women representing 65% of all breast cancers resected in Uppsala County, Sweden from January I ₅ 1987 to December 31, 1989. Information pertaining to patient therapies, clinical follow up, and sample processing are described elsewhere (41).

Histological Grading

For histological grading, new tumour sections are prepared from the original paraffin blocks, stained with eosin, and graded in a blinded fashion by H.N. according to the Nottingham grading system (6, Haybittle et al., 1982) as follows:

Tubule Formation: 3=poor, if < 10% of the tumour showed definite tubule formation, 2=moderate, if >10% but <75%, and l=well, if >75%.

Mitotic Index: l=low, if <10 mitoses, 2=medium, if 10-18 mitoses, and 3=high, if >18 mitoses (per 10 high-power fields). The field diameter was 0.57 mm.

Nuclear Grade: I=Io w, if there was little variation in the size and shape of the nuclei, 2=medium for moderate variation, an,d 3=high for marked variation and large size.

Scores are then summed, and tumour samples with scores ranging from 3-5 are classified as Grade I; 6-7 as Grade II; and 8-9 as Grade III. Protein Assays

Protein levels of Estrogen Receptor (ER) and Progesterone Receptor (PgR) are assessed by immunoassay (monoclonal 6Fl 1 anti-ER and monoclonal NCL-PGR, respectively, Novocastra Laboratories Ltd, Newcastle upon Tyne, UK) and deemed positive if >0.1 fmol/ug DNA. VEGF was measured in tumour cytosol by a quantitative immunoassay kit (Quantikine-human VEGF; R&D Systems, Minneapolis, MN, USA) as described (42). Protein levels of Ki-67 are analyzed using anti-Ki67 antibody (MIB-I) by the grid-graticula method with cut-offs: low=2, medium >2 and <6, high=6. Cyclin E was measured using the antibody HE 12 (Santa Cruz Inc., USA) with cutoffs: low=0-4%, medium=5-49%, and higb=50-100% stained tumour cells (43). S -phase fraction was determined by flow cytometry and defined as high if >7% in diploid tumours, or >12% in aneuploid tumours. TP53 mutational status was determined by cDNA sequencing as previously described (41). The Uppsala tumour samples ar

approved for microarray profiling by the ethical committee at the Karolinska

Institute, Stockholm, Sweden.

Stockholm Cohort

The Stockholm samples are derived from breast cancer patients that were operated on at the Karolinska Hospital from January 1, 1994 through December 31, 1996 and identified in the Stockholm-Gotland breast cancer registry.

Information on patient age, tumour size, number of metastatic axillary lymph nodes, hormonal receptor status, distant metastases, site and date of relapse, initial therapy, and date and cause of death are obtained from patient records filed with the Stockholm- Gotland registry.

Tumour sections are classified using the Nottingham grading system (Haybittle et al., 1982). The Stockholm tumour samples are approved for microarray profiling by the ethical committee at the Karolinska Hospital, Stockholm, Sweden.

Singapore Cohort The Singapore samples are derived from patients that were operated on at the

National University Hospital (Singapore) from February 1, 2000 through January 31, 2002.

Information on patient age, tumour size, number of metastatic lymph nodes and hormonal receptor status are obtained from hospital records. Tumour sections are graded in a blinded fashion according to the Nottingham grading system (Haybittle et al., 1982) as applied to the Uppsala and Stockholm cohorts, with the following exception: Mitotic Index: l=low, if <8 mitoses, 2=medium, if 9-16 mitoses, and 3=high, if >16 mitoses (per 10 high-power fields). The field diameter is 0.55 mm. The Singapore tumour samples are approved for microarray profiling by the Singapore National University Hospital ethics board.

After exclusions based on tissue availability, RNA integrity, clinical annotation and microarray quality control, expression profiles of 249, 147, and 98 tumours from the Uppsala, Stockholm and Singapore cohorts, respectively, were deemed suitable for further analysis. Example 3. Materials and Methods: Microarray Expression Profiling and Processing

All tumour samples are profiled on the Affymetrix Ul 33 A and B genechips. Microarray analysis of the Uppsala and Singapore samples was carried out at the Genome

Institute of Singapore (44). The Stockholm samples are analyzed by microarray at

Bristol-Myers Squibb, Princeton, New Jersey, USA. RNA processing and microarray hybridizations are carried out essentially as described (44).

Microarray data processing: all microarray data are processed as previously described (44).

Example 4. Materials and Methods: Statistical Analysis of Gene Ontology (GO) Terms

GO analysis is facilitated by PANTHER software

(https://panther.appliedbiosystems.com/) (46). Selected gene lists are statistically compared (Mann- Whitney) with a reference list (ie, NCBI Build 35) comprised of all genes represented on the microarray to identify significantly over- and under-represented GO terms.

Example 5. Materials and Methods: Survival Analysis

The Kaplan Meier estimate is used to compute survival curves, and the p-value of the likelihood-ratio test is used to assess the statistical significance of the resultant hazard ratios. For standardization, events occurring beyond 10 years are censored. All cases of contralateral disease are censored. Disease-free survival (DFS) is defined as the time interval from surgery until the first recurrence or last follow-up.

Multivariate analysis by Cox proportional hazard regression, including a stepwise model selection procedure based on the Akaike information criterion, and all survival statistics are performed in the R survival package. Remaining predictors in the Cox models are assessed by Likelihood-ratio test p-values.

Example 6. Methods: Scoring by the Nottingham Prognostic Index (NPI)

NPI scores (Haybittle et al., 1982) are calculated according to the following formula:

NPI score= (0.2 x tumour size (cm)) + grade (1,2 or 3) + LN stage (1,2 or 3)

Tumour size is defined as the longest diameter of the resected tumour. LN stage is 1, if lymph node negative, 2, if 3 or fewer nodes involved, and 3, if >3 nodes involved (47). As the number of cancerous lymph nodes are not available for the Uppsala cohort, a LN stage score of 2 is assigned if 1 or more nodes are involved, and a score of 3 is assigned if nodal involvement showed evidence of periglandular growth. For ggNPI calculations, grade scores (1,2 or 3) are replaced by genetic grade predictions (1 or 3).

NPI scores <3.4 = GPG (good prognostic group); scores of 3.4 to 5.4 =

MPG (moderate prognostic group); scores >5.4 = PPG (poor prognostic group). Scores of 2.4 or less = EPG (excellent prognostic group).

Example 7. Methods: Descriptive Statistics For inter-group comparisons using the clinicopathological measurements, non- parametric Mann- Whitney U-test statistics are used for continuous variables and onesided Fisher's exact test used for categorical variables. This work is facilitated by the Statistica-6 and StatXact-6 software packages.

Example 8. Materials and Methods: Details of Genetic Reclassification Algorithm of Grade 2 Tumours Based on SWS Approach

In simplified terms, the algorithm of genetic re-classification of Grade 2 tumour, based on SWS approach can be described as follows.

A training set consisting of samples of known classes (eg, histologic Grade I (Gl) and histologic Grade III (G3) tumours) is used to select the variables (ie, gene expression measurements; probesets or predictors), that' allow the most accurate discrimination (or prediction) of the samples in the training set. Once the SWS algorithm is trained on the optimal set of variables, it is then applied to an independent exam set (ie, a new set of samples not used in training) to validate it's prediction accuracy. More details are given below. Briefly, for constructing the class prediction function, the SWS method uses the training set ^ ₀ (comprised of Gl and G3 tumour samples) to evaluate statistically the weight of the graduated "informative" variables (predictors), and all possible pairs of these predictors. The predictors are automatically selected by SWS from n («=44,500) probe sets (which represents the gene expression measurements) on Ul 33 A and U133B Affymetrix Genechips.

The description of each patient includes n (potential) prognostic variables X ₁ ,..., X _n (signals from probe sets of the Ul 33 A and U133B chips) and information about class to which a patient belongs. In particular, the predictors might be able to discriminate Gl and G3 tumours with minimum "a posteriori probability". Reliability of the SWS class prediction function is based on the standard "leave-one-out procedure" and on an additional exam of the class prediction- ability on one or more independent sample populations (ie, patient cohorts). In this application of SWS, the G2 tumour samples of the

Uppsala cohort and two other cohorts (NUH and Stockholm cohorts; see

Methods) have been used as exam datasets to test the SWS class prediction function.

Let us consider the available ^-dimension domain of the variables (the probesets) X _x ,..., X _n as prognostic variable space. The SWS algorithm is based on calculating the a posteriori probabilities of the tumours belonging to one of two classes using a weighted voting scheme involving the sets of so called "syndromes". A syndrome is the sub-region of prognostic variable space. For a syndrome to be useful in the algorithm, within the syndrome, one class of samples (for instance, G3 tumours) must be significantly highly represented than another class (for instance, GIs), and in other sub-region(s) the inverse relationship should be observed. In the present version of the SWS method, one- dimensional and two-dimensional sub-regions (syndromes) are used.

Let δ' ₍ and b'\ denote the boundaries of the sub-region for the variable X ₁ (the i-th probe set); b\ > X ₁ > b" . One-dimensional syndrome for the variable X ₁ is defined as the set of points in variable space for which inequalities V ₁ ≥ X ₁ > b" are satisfied. Two- dimensional syndrome for variables X., and X^, is defined as a set of points in variable space for which inequalities b', ≥ X ₁ , > b" and b'.,, ≥ X _n > b". _n are satisfied. The syndromes are constructed at the initial stage of training using the optimal partitioning (OP) algorithm described below.

SWS Training Algorithm SWS training algorithm is based on three major steps:

1) optimal recoding (partitioning) of the given covariates (signal intensity values) to obtain discrete-valued variables with low and high gradation;

2) selection of the most informative and robust of these discrete-valued variables and their paired combinations (termed syndromes) that together best characterize the classes of interest;

3) tallying the statistically weighted votes of these syndromes to allow us to compute the value of the outcome prediction- function.

Optimal Partitioning (OP)

The OP method is used for constructing the optimal syndromes for each class (Gl and G3) using the training set S ₀ . The OP is based on the optimal partitioning of some

potential prognostic variable X ₁ range that allows the best separation of the samples belonging to different classes. To evaluate the separating ability of partition R (see below) in the training set SQ the chi-2 functional is used (Kuznetsov et al, 1998). The optimal partitions are searched inside observed variable domain that contain partitions with cut-off values not greater than a fixed threshold (defined below). The partition with the maximal value of the chi-2 functional is considered optimal for the given variable.

Stability of Partitioning

Another important characteristic that allows evaluation the prognostic ability of partitioning model for specific variables is the index of boundary instability. Let R ₀ , R ₁ , ... , R _m be optimal partitions of variable X ₁ ranges that is calculated by training set

S ₀ , S _{ ,... , S _1n , where S _k is the training set without description of the k ^{t l} sample. Let K _j denote the different classes (J=I ₁ I). Let &* , ...,δ*_, be boundary points of optimal partition R _k found by training set S _k ; D ₁ is the variance of variable X ₁ . The boundary instability index rc(S ₀ , K^r) for partitioning with r elements is calculated as the ratio (Kuznetsov et al, 1996) :

Selecting of Optimal Variables Set

The OP can be used at the initial stage of training for reducing the dimension of the prognostic variables set. Selection of the optimal set of prognostic variables depends on a sufficiently high partition value determined by the Chi-2 function. The additional criterion of selection of prognostic variables is the instability index κ(S ₀ , K _j , r) . The variable is used if value κ(S ₀ ,K _; ,r) is less than threshold K ₀ , defined a priori by the user. When the partition of the given variable is instable ( κ(S ₀ ,K _j ,r) <κ _ϋ ), the variable is removed from the final optimal set of prognostic variables. Finally, the optimal set of prognostic variables is defined if both selection criteria are fulfilled.

The Weighted Voting Procedure

Let Q° denote the set of constructed syndromes for class K ₁ . Let x ^* denote the point of parametric space. The SWS estimates a posteriori probability P, ^JV (x ^* ) of the class

K ₁ at the point x ^* that belongs to the intersection of syndromes q _λ , ...,q _r from

Q° as follows:

where v/ is the fraction of class K _} among objects with prognostic variables vectors belonging to syndrome q _} , W ₁ is the so-called "weight" of syndrome q _t . The weight W ₁ is calculated by the formula,

w ,— - ^m i

¹ m^l d. ^'

where d _f = (l -V- )v? + — (l~ Vn ) ^v n (Kuznetsov, 1996). The estimate of fraction v/

1 ^{■ ■} --- variance has the second term (1 — V L )vl , which is used to avoid a value d- equal to

zero in cases when the given syndrome is associated only with objects of one class from the training set.

The results of testing applied and simulated tasks have demonstrated that formula (1) gives too low of estimates of conditional probabilities for classes that are of smaller fraction in the training set. So the additional correction of estimates in (1) has been implemented. The final estimates of conditional probability at point x ^* are calculated as and X, is the vector of prognostic variables for the k-th samples from the training set.

Example 9. Derivation of a Classifier Comprising 264 Probe Sets (SWS Classifier 0)

Schema of the SWS-Based Discovery Method of Novel Classes of Tumours Our methodology is based on the schema presented in Figure 1.

Beginning with the Uppsala dataset comprised of 68 Gl and 55 G3 tumours, we used SWS optimal partitioning (OP) at the initial stage of training to reduce the dimension

of the prognostic variables set. SWS rank orders the set of probes according to specific algorithmic criteria for assessing differential expression between classes.

Based on this two-criteria (chi-2 and instability index) selection algorithm, we used SWS chi-2 values bigger than 24.38 (at p-value less then 0.00001); in combination with low boundary instability index criteria ( K ₀ <0.1 for 90% of the selected informative variables and K ₀ <0.4 for 10% of the other informative variables). Visual presentation on scatchard plot (log K ₀ , chi-2) distribution of probesets, these two cut-off values discriminated the relatively small and compact group of probesets. We observed that this group of probesets provide a local minima on the Class Error Rate (CER) function and provide an optimal selection of 264 probesets classifier of Gl and G3. Using these 264 probe sets, the both SWS and PAM methods ^' provide a small misclassification error (4.5% for Gl, and 5.5% for G3, respectively) when the leave-one-out cross-validation procedure is used. We also used the U-test with critical value p=0.05 (with Bonferroni correction) and all 264 probesets follow this cut-off value. Based on our selection criteria, we selected a classifier comprising 264 probe sets, which we term the "SWS Classifier 0". See Table Dl in section "SWS Classifier 0" of the Description as well as Appendix 1.

Details are shown in Appendix IA, Appendix 2, Appendix 3 and Appendix 4.

Example 10. A Posteriori Probability for SWS Classifier 0 (264 Probe Sets) Gl and G3 Estimated by SWS Classifier 2

A posterior probability for Gl and G3 was also estimated by SWS Classifier 2 for each tumour sample by the classical leave-one-out cross-validation procedure.

We estimated the class error rate based on the misclassification error rate plot (Tibshirani et al, 2002) and found that for the 264 selected probe sets, CER consists of 5% for Gl, and 6% for G3, respectively. Similar discrimination was obtained with SWS methods (see above).

Based on consistency between SWS and U-tests and PAM CER validation of the selection procedure, we further considered the classification results using the 264 variables. In two-group comparisons, high CER were observed in the G1-G2 and G2-G3 predictions (data not shown), while G1-G3 classification accuracy was high (<5% errors). Complementary to SWS classification method, the PAM method confirms that G2 tumours are not molecularly distinct from either low or high grade tumours, possibly owing to substantial molecular heterogeneity within the G2 class.

Example 11. Derivation of Classifiers of 6 Genes (SWS Classifier 1)

To extract the smallest possible classifier from the 264 variables, we varied the initial parameters of the SWS algorithm to minimize the number of predictors in training set providing the maximum correlation coefficient between posteriori probabilities and true class indicators (specifically, 1 was the indicator of Gl tumours, and 3 was the indicator of G3 tumours in the G1-G3 comparison). The predictive power of the predictor set was estimated using standard leave-one-out procedure and counting the numbers of errors of class predictions.

We derived a classifier comprising 6 gene probe sets (5 genes) which we term the "SWS Classifier 1". 4.4% for class Gl; and 5.5% for class G3 CERs were obtained with the SWS Classifier 1. See Figure 1 and Table D2 in section "SWS Classifier 1" of the Description.

Appendix 5 A, Appendix 5B and Appendix 5 C show detailed information about selected gene probe sets, optimal partition boundaries, true classes, posterior probabilities and clinical significance of the SWS Classifier 1 predictor (estimated by patient survival analysis).

Example 12. Derivation of Classifiers of 18 Genes (SWS Classifier 2)

By SWS, for the G1-G3 comparisons, maximal prediction accuracies are obtained with 18 probe sets (17 genes). We refer to this 18 probe set as the "SWS Classifier 2". See Table D3 in section "SWS Classifier 2" of the Description. This classifier includes all five genes represented by SWS Classifier 1.

Appendix 6A, Appendix 6B and Appendix 6C show detailed information about selected gene probe sets, optimal partition boundaries, true classes, posterior probabilities and clinical significance of the SWS Classifier 2 (estimated by patient survival analysis). With the 18 probe sets, both the SWS Classifier 2 and PAM correctly classify

-96% (65/68) of the GIs and -95% (52/55) of the G3s (by leave one-out method).

The smaller number of probes sets required by SWS Classifier 1 (6 probe sets) compared to PAM (18 probe sets, data not presented) may reflect the ability of SWS to use more diverse interaction and/or co-expression patterns during variable selection. The posterior probability (Pr) is an estimate of the likelihood that a sample from the exam group of tumours belongs to one class (termed "Gl -like") or the other (ie, "G3-

like"). Both 18 probesets SWS and PAM classifiers scored the vast majority of Gl and G3 tumours with high probabilities of class membership.

Example 13. The SWS Classifier 0 (264 Gene Probe Set) Contains Many Small Subsets Which Can Provide Equally High Discrimination Ability of the Genetic G2a and G2b Tumours

Due to the highly informative and stable nature of each gene (represented by Affymetrix probe-sets) of the 264 predictor set we hypothesized that there are many small alternative gene sub-sets that could be used to classify tumours with high accuracy (and therefore classify patients according to outcome with high prognostic significance). For example, high Pr scores for the class assignments of Gl and G3 by SWS classifier 1 (6 probesets, as shown in Table D2 in section "SWS classifier 1" of the Description and Appendix 5A) and SWS class assignments of Gl -like and G3-like classes within G2 class were observed.

Notably, 95% of the tumours of the Uppsala cohort showed >75% probability of belonging to either the Gl-like or G3-like class, indicating a highly discriminant statistical basis for the class prediction function of the SWS classifier 1 for the G2 class.

Example 14. SWS Classifier 3 and SWS Classifier 4

To find other classifiers, we excluded the best 6 probe sets (SWS classifier 1) from the 264 probe sets, and randomly selected two non-overlapping subsets (each of 40 probe sets) from the remaining 258 probe sets and applied the SWS algorithm to each subset.

In this way, we selected two additional classifiers: SWS classifier 3 (6-probe sets; Table D4 in section "SWS Classifier 3" of the Description and Appendix 7A) and SWS classifier 4 (7 -probe sets; Table D5 in section "SWS Classifier 4" of the Description and Appendix 8A). Tables D4 and D5 are organized as Table D3. For Uppsala , Stockholm and

Singapore cohorts, each of three SWS classifiers provide similar high accuracy of classification in G1-G3 comparisons (Tables D3-D5). SWS also provided high and reproducible levels of separation of G2a and G2b sub-groups for different cohorts and highly significant differences in G2a-G2b comparison based on survival analysis (Tables D3-D5).

These tables show the values of parameters of SWS algorithm for selected classifies, predicted individual probabilities of belonging to the given class, and gene annotation, clinical significance etc.

Thus, we could consider the 264 probe sets as a general genetic classifier of the G2a (Gl -like) and G2b (G3-like) tumour types.

Example 15. Dichotomy of G2 Tumours by 264 Probe Sets Gene Grade Classifier

We next applied our grade classifiers directly to the 126 G2 tumours of the Uppsala cohort to ask if these genetic determinants of low and high grade might resolve moderately differentiated G2 tumours into separable classes. Using SWS for the 264 predictor set, we observed that the G2 tumours could be separated into Gl -like (n=83) and G3-like (n=43) classes with few tumours exhibiting intermediate Pr scores (Appendix 2).

The probabilities of the SWS class assignments are shown in Figure 2B (Figure 2, Panel B) and more detailed information in Appendix 2.

We found 96% of the G2 tumours were assigned by the SWS classifier (and 94% by the PAM classifier, data not shown) to either the Gl -like or G3-like classes with >75% probability, indicating that almost all G2 tumours can be molecularly well separated into distinct low- and high-grade-like classes (henceforth referred to as "G2a" and "G2b" genetic grades) (Appendix 2).

We validated the separation ability of GIa and G2b based on individual predictors and showed that all of them are statistically significant by U-test and t-test (Appendix 3).

Clinical validation (survival analysis) of G2a and G2b tumour subtypes based on the predictor set (or genetic classifier), showed a highly significant difference between survival curves of the GIa and G2b patients (Appendix 4).

Example 16. Genetic Grade is Prognostic of Tumour Recurrence

To determine if the genetic grade classification correlates with patient outcome, we compared the disease-free survival (DFS) of patients with histologic G2 tumours classified as G2a or G2b by the SWS algorithm. (Due to space limitations and high concordance between the SWS and PAM classifiers, only data for the SWS classifier are presented hereafter.)

The Kaplan-Meier survival curves for these patients are shown in Figure 3 (green and red curves) superimposed on the survival curves of histologic Gl, G2 and G3 patients (black curves) for comparison. Patients with G2a tumours showed a significantly better disease-free survival than those with G2b disease, regardless of therapeutic background (p=0.001; Figure 3A).

This finding is consistent in specific therapeutic contexts including untreated patients (Figure 3B), systemic therapy (Figure 3C), and hormone therapy only (Figure 3D) with survival differences significant at p=0.019, p=0.10 and p=0.022, respectively. These findings demonstrate a robust prognostic power of the genetic grade classifier in moderately differentiated tumours independent of therapeutic effects.

Example 17. External Validation of the Genetic Grade Signature on the Stockholm and Singapore Cohorts

For external validation, we directly applied the SWS classifier to two large independent cohorts of primary breast cancer cases that are also graded according to the NGS guidelines and profiled on the Affymetrix platform (albeit at different times and in different laboratories). The results of the grade classifications are shown in Figure 2.

In both the Stockholm and Singapore cohorts, the Gl tumours are correctly classified with high accuracies similar to that observed in the training set: 96% (27/28) for Stockholm and 91% (10/11) for Singapore (Figure 2C and Figure 2E). However, both cohorts showed less accuracy in classifying the G3 tumours: 75% (46/61) for Stockholm and 72% (34/47) for Singapore. Despite this, the classifier remained capable of dividing the vast majority of the tumour samples into Gl-like and G3-like classes with high Pr scores, and this remained true for the G2 tumours of both the Stockholm and Singapore cohorts (Figure 2D and Figure 2F). As clinical histories are available on the Stockholm patients, we tested the prognostic performance of the classifier on this new G2 population of which 79% (46/58) of tumours are classified as G2a and 21% (12/58) are classified as G2b. Though this set is considerably smaller than the Uppsala G2 set, similar survival associations are observed.

As Figure 3E and Figure 3F show, patients with the G2a subtype are significantly less likely to relapse than those with tumours of the G2b subtype, indicating that the prognostic performance of the genetic grade classifier is reproducible in a second, independent population of G2 patients.

Example 18. The Prognostic Power of Genetic Grade is Independent of Other Risk Factors To assess the prognostic novelty of the classifier, we used multivariate Cox regression models to compare its performance to that of other conventional prognostic indicators assessed in the Uppsala cohort including lymph node status, tumour size, patient age, and estrogen (ER) and progesterone (PgR) receptor status. See Table E3 below.

Table E3. The genetic grade signature is a strong independent indicator of disease-free survival in a multivariate analysis with conventional risk factors.

As Table E3 shows, the genetic grade signature remained significantly associated with outcome in the different therapeutic contexts independent of the classical predictors, and is superior to both LN status and tumour size in all four treatment subgroups with the exception of systemic therapy where only tumour size is more significant.

This finding is further substantiated by a robust model selection approach (the Akaike Information Criterion) whereby the genetic grade classifier remained more significant than LN status and tumour size in all therapeutic subgroups (data not shown). These results demonstrate a powerful and additive contribution of the genetic grade classifier to patient prognosis.

Example 19. G2a and G2b Subtypes Are Molecularly and Pathologically Distinct

The prognostic performance of the classifier suggests that G2a and G2b genetic grades may in fact represent distinct pathological entities previously unrecognized. We investigated this possibility by several approaches.

First we examined the histopatho logical composition of the G2a and G2b tumours and found that the predominant histologic subtypes - ductal, lobular and tubular - are equally distributed within the two classes and therefore not correlated with genetic grade (data not shown). Next, we analyzed the expression levels of the selected 264 probesets (i.e., representing -232 genes) as the maximum number of probesets capable of recapitulating a high G1/G3 classification accuracy (see Methods). These genes represent

the top most significantly differentially expressed genes between Gl and G3 tumours after correcting for false discovery (see Table Dl above).

As shown in Figure 4, hierarchical cluster analysis using this set of genes shows a striking separation of the G2 population into two primary tumour profiles highly resembling the Gl and G3 profiles and that separate well into the G2a and G2b classes. Indeed, all but 11 of these 264 gene probesets are also differentially expressed (at p<0.05, Wilcoxon rank-sum test) between the G2a and G2b tumours.

This finding shows that extensive molecular heterogeneity exists within the G2 tumour population, and this heterogeneity is robustly defined by the major determinants of Gl and G3 cancer. It also demonstrates that a much larger and pervasive transcriptional program underlies the genetic grade predictions of the SWS signature — despite its composition of a mere 5 genes. Furthermore, statistical analysis of the gene ontology (GO) terms associated with the G2a-G2b differentially expressed genes revealed the significant enrichment of numerous biological processes and molecular functions.

Table E4 displays a selected set of significantly enriched GO categories which includes cell cycle, inhibition of apoptosis, cell motility and stress response, suggesting an imbalance of these cellular processes between the G2a- and G2b-type tumour cells.

Table E4. Gene ontology analysis of differentially expressed genes. Selected terms are shown with corresponding p-values that reflect significance of term enrichment

Table S2 below shows the complete list of GO categories and their p values.

Table S2. Comprehensive table of significant gene ontology terms identified in the different tumour group comparisons.

To extend our analysis beyond the transcript level, we investigated the differences between G2a and G2b tumours using conventional clinicopathological markers.

Of the three histologic grading criteria, both mitotic count and nuclear pleomorphism are found to significantly vary between the G2a and G2b tumours (p=0.007 and p=0.05; Figure 5A and Figure 51). Protein levels of the proliferation marker Ki67 are also found to be significantly different between the G2a and G2b tumours (p<0.0001; Figure 5B).

These findings, together with those of the gene ontology analysis, suggest that the genetic grade classifier may largely mirror cell proliferation and thus reflect the replicative potential of the breast tumour cells. However, proliferation is not the only oncogenic factor found to be associated with genetic grade. In the G2b tumours, protein levels of VEGF (Figure 5C), a major inducer of angiogenesis, and the degree of vascular growth (Figure 5D) are both found to be significantly higher compared to the G2a samples

(p=0.015 and p=0.002, respectively) suggesting that a difference in angiogenic potential also distinguishes the two genetic grade classes.

Further analysis of bio-markers revealed yet more oncogenic differences. P53 mutations are found in only 6% of the G2a tumours, whereas 44% of the G2b tumours are p53 mutants (p<0.0001 ; Figure 5E) consistent with their higher replicative potential, and likely conferring a further survival advantage to these tumours via decreased apoptotic potential. We also observed higher levels of cyclin El protein (p=0.04; Figure 5F) in the G2b tumours which, in addition to contributing to enhanced proliferation (20), may also confer greater genomic instability (21, 22). Finally, we observed a significant difference in hormonal status between the G2a and G2b tumours, with an increasing fraction of ER negative (7% versus 19%; p=0.06) and PgR negative (8.5% versus 23%; p=0.02) tumours in the G2b class, indicating differences in hormome sensitivity and dependence.

Taken together, these results show that multiple tumourigenic properties measured at the RNA, DNA, protein, and cellular levels can subdivide the G2a and G2b tumour subtypes — a finding that may explain, in part, the different patient survival outcomes observed between these two genetic classes.

Example 20. The Grade Signature is More Than a Proliferative Marker

The genetic and clinicopathological evidence suggests that the genetic grade signature reflects, among other properties, the proliferative capacity of tumour cells. That proliferation rate is positively correlated with poor outcome in breast cancer (23) could explain the prognostic capacity of the genetic grade signature.

To further investigate this possibility, we analyzed the major proliferation markers, Ki67, S-phase fraction and mitotic index, together with the genetic grade signature, for survival correlations in Cox regression models (Table S3).

Table S3. Multivariate analysis of proliferation markers and the genetic grade signature for disease-free survival correlations among patients with Grade II tumours.

Multivariate analysis showed that the genetic grade signature remained a significant independent predictor of recurrence (p=0.0075) in the presence of these proliferation markers, suggesting that the prognostic power of the grade signature derives from more than just and association with cell proliferation.

Example 21. G2a and G2b Tumours Are Not Identical to Histologic Gl and G3 Cancers In the survival analysis (Figure 3), we observed no significant survival differences between patients with Gl and G2a tumours, nor those with G3 and G2b tumours. This observation, together with the transcriptional analysis in Figure 4, suggests that the G2a and G2b classes may be clinically and molecularly indistinguishable from histologic Gl and G3 tumours, respectively. To address this, we further analyzed the expression patterns of the 264 grade- associated probesets described in Figure 4. We discovered 14 genes and 57 genes significantly differentially expressed (p<0.01, Mann- Whitney U-test) between the Gl and G2a tumours and G3 and G2b tumours, respectively.

Notably, FOS and FOSB, central components of the AP-I transcription factor complex, are expressed at higher levels in the Gl tumours, while genes involved in cell cycle progression such as CCNE2, MAD2L1, ASK and ECT2 are expressed at higher levels in the G2a tumours. In a similar fashion, the G3 tumours showed higher expression of cell cycle genes such as CDC20, BRRNl and TTK as well as proliferative genes with oncogenic potential including MYBL2, ECT2 and CCNEl when compared to the G2b tumours, while the anti-apoptotic gene, BCL2, is expressed at higher levels in the G2b tumours.

GO analysis of these differentially expressed genes indicated larger biological differences. In the Gl-G2a comparison, the differentially expressed genes pointed to differences primarily in cell cycle-related processes and oncogenesis, while differences between the G2a and G3 tumours included cell cycle-related processes, inhibition of apoptosis, oncogenesis and cell motility (Table E4, Table S2).

Statistical analysis of conventional clinicopathological markers revealed further distinctions in the Gl-G2a and the G2b-G3 tumour comparisons. As shown in Figure 5, G2a tumours showed significant increases in tumour size (K), lymph node positivity (L),

cellular mitoses (A), tubule formation (J) and Ki67 levels (B) compared to histologic Gl tumours, and the G3 population showed significant increases in tumour size (K), vascular growth (D), mitoses (A), tubule formation (J), cyclin El (F) and ER negative status (G) when compared to the G2b tumours. Taken together, these data indicate that the G2a and G2b populations, though highly similar to Gl and G3 tumours in terms of survival and transcriptional configuration, remain separable at multiple molecular and clinicopathological levels.

Example 22. Prognostic Potential of the Genetic Grade Signature in G3 Tumours

The prognostic performance of the genetic grade signature in the G2 population suggests that the molecular "misclassifications" in the G1-G3 comparisons might con-elate with survival differences. Of the 68 Uppsala and 28 Stockholm Gl tumours, too few are classified as G3-like (ie, 4 in total) for a reliable Kaplan-Meier estimate.

However, among the 55 Uppsala and 61 Stockholm G3 tumours, a total of 18 are classified as Gl -like. Kaplan-Meier analysis could not confirm a significant disease-free- survival advantage for these patients, though a trend is observed (Figure 7A).

Interestingly, scaling of the SWS probability (Pr) score to a threshold of Pr>0.8 (for Gl- like) resulted in the selection of 12 Gl -like G3 tumours associated with only two relapse events (one being a local recurrence only), thus having a survival curve moderately different from that of the remaining G3 population (p=0.077; Figure 7A). This finding suggests that the prognostic significance of the classifier may extend also to the poorly differentiated G3 tumours, and that scaling based on the classifier Pr score may allow the fine tuning of prognostic sensitivity and/or specificity, depending on the clinical application.

Example 23. Genetic Grade Improves Prognosis by the Nottingham Prognostic Index The Nottingham Prognostic Index (NPl) is a widely accepted method of stratifying patients into prognostic groups (good (GPG), moderate (MPG) and poor (PPG)) based on lymph node stage, tumour size, and histologic grade (24). It is described in detail in Haybittle et al., 1982. We investigated whether incorporating genetic grade into the NPI could improve patient stratification. A simplified substitution method was explored. For all tumours of the Uppsala and Stockholm cohorts for which NPI scores and survival information could be obtained (n=382), histologic grade (1, 2 or 3) is replaced by the genetic grade prediction (1 or 3) and new NPI (ie, ggNPI) scores are computed (see

Methods). The survival of patients stratified into risk groups is then compared between classic NPI and ggNPI.

Though the survival curves of the NPI and ggNPI prognostic groups are comparable (Figure 6A and Figure 6B) ₃ the ggNPI reclassifϊed 96 patients into different prognostic groups (ie, 46 into GPG, 36 into MPG, and 13 into PPG). The survival curves of these reclassified patients are highly similar to the GPG, MPG and PPG of the classic NPI (Figure 6C) indicating that reclassification by genetic grade improves prognosis of patient risk.

Practical guidelines that use the NPI in therapeutic decision making often recognize an excellent prognostic group (EPG) comprised of patients with NPI scores </= 2.4 (25, 26). Untreated patients in this group with lymph node negative disease have a 95% 10-year survival probability - equivalent to that of an age-matched female population without breast cancer (26). Thus, patients in this group are routinely not recommended for post-operative adjuvant therapy (25-27). We compared the NPI and ggNPI stratifications on a subset of 161 lymph-node- negative patients who received no adjuvant systemic therapy. Forty-three and 87 patients are classified into the EPG by the classic NPI and ggNPI, respectively. Of the 43 patients classified into the EPG by the classic NPI, only one was considered different by the ggNPI; whereas, of those classified as needing adjuvant therapy by the classic NPI (ie, scores >2.4), 45 are reclassified by the ggNPI into the EPG.

When examined for outcome, the survival curves of the 43 and 87 EPG patients by NPI and ggNPI, respectively, are statistically indistinguishable, both showing -94% survival at 10 years (Figure 6D).

Thus, twice as many patients could be accurately classified into the EPG by the ggNPI, suggesting that the use of genetic grade can improve prediction of which patients should be spared systemic adjuvant therapy.

Example 24. Discussion

The clinical subtyping of cancer directly impacts disease management. Subtypes indicative of tumour recurrence or drug resistance indicate the need for more aggressive or specific therapeutic strategies, while those that suggest less aggressive disease may specify milder therapeutic options. While clinical subtyping has historically been based primarily on the phenotypic properties of cancer, comprehensive genomic and transcriptomic analyses are beginning to reveal robust genotypic determinants of tumour subtype. In this context, we have studied the transcriptomes of primary invasive breast cancers using

expression microarray technology to elucidate the genetic underpinnings of histologic grade, and to use this information to resolve the clinical heterogeneity associated with histologic grade.

Using two different supervised learning algorithms, SWS and PAM, we identified small gene subsets capable of classifying histologic Grade I and Grade III tumours with high accuracy. The smallest gene signature (SWS), comprised of a mere 5 genes (6 probesets), partitioned the large majority of G2 tumours into two highly distinguishable subclasses with Gl -like and G3-like properties (G2a and G2b, respectively). Not only are the G2a and G2b tumours molecularly similar to those of histologic Gl and G3, respectively, but the disease-free survival curves of G2a and G2b patients are also highly resemblant of those of Gl and G3 patients. Moreover, these observations are confirmed in a large independent breast cancer cohort. Further analysis revealed that extensive genetic differences between the G2a and G2b classes are accompanied by a host of biological and tumourigenic differences know to separate low and high grade cancer (28) including proliferation rate (mitotic index, Ki67), angiogenic potential (VEGF, vascular growth), p53 mutational status, and estrogen and progesterone dependence, to name a few. Together, these findings demonstrate that the genetic grade signature recognizes and delineates two novel grade-related clinical subtypes among moderately differentiated G2 tumours. Ma et. al. (2003) were the first to report a histologic grade signature capable of distinguishing low and high grade breast tumours. Using 12K cDNA microarrays to analyse material from 10 Gl, 11G2 and 10 G3 microdissected tumours, they identified from a list of 1,940 variably expressed, well-measured genes (the top 200 differentially expressed between Gl and G3 tumours (p<0.01 after false discovery correction) (29). Using these genes to cluster their graded tumours, they observed that the majority of G2 tumours possessed a hybrid signature intermediate to that of Gl and G3 with few exceptions (see Figure 3 in Ma et. al., PNAS, 2003). Notably, this finding is in contrast with our discovery that the majority of G2 tumours do not display hybrid signatures (Figure 4; profiles of the top 264 gene probesets), but rather possess clear Gl-like or G3- like gene features. According to our SWS classifiers, only a small percentage (6%) of the Grade 2 tumours has intermediate grade measurements (i.e. Pr score <0.75 for Gl-like and G3-like).

To address this discrepancy, we cross-compared their list of 200 grade-associated genes to our list of 232 and observed a significant overlap of 35 genes (p<l.0x10-7; Monte Carlo simulation) including 2 of our 5 SWS signature genes, MELK and STK6. However, this overlap, despite its significance, represents only a small percentage of either gene list. That the two lists are mostly dissimilar in composition, and that the Ma et. al. study

included both invasive (IDC) and noninvasive (DCIS) tumours could explain, to some degree, the variable results observed. Nevertheless, our finding that G2 tumours are predominantly Gl -like or G3-like is clinically substantiated by the significant and reproducible survival differences observed between the G2a and G2b classes. It is also possible that differences in sample size (we have much larger number of patients than in Ma etc work), sample preparation, sample size, RNA purification, data normalization could have contributed to the variable results.

To better understand the prognostic value of the genetic grade signature, we compared its performance to other major indicators of outcome in multivariate Cox regression models. In G2 tumours, not only did the classifier remain an independent predictor of disease recurrence, but it is consistently a more powerful predictor than lymph node status and tumour size, underscoring its value as a new prognostic indicator. When incorporated into the Nottingham Prognostic Index (Haybittle et al., 1982), the genetic grade signature improved risk stratification for 25% of patients (compared to the classic NPI) and more than doubled the fraction of lymph node negative patients that should be classified into the excellent prognostic group and thus spared adjuvant treatment.

Breast cancer is thought to progress from a hyperplastic state, to a noninvasive malignant form (carcinoma in situ), to invasive carcinoma and, ultimately, to metastatic disease (30-32). Both the noninvasive and invasive forms can be stratified according to histologic grade. Whether grade is a continuum through which breast cancer progresses, or whether it is merely the endpoint of distinct genetic pathways has been debated (33-38). Studies comparing primary tumours to their subsequent metastases have supported the grade progression model, particularly when multiple metachronous recurrences are analyzed (38, 39). However, comparative genomic studies have identified reproducible chromosomal alterations that distinguish low and high grade disease including a 16q deletion unique to Gl carcinomas (36, 37, 40). These studies argue against the progression model and point to genetic origins of histologic grade. In our study of 494 invasive primary tumours, 94% could be molecularly classified with high probability of being Gl- like or G3-like, while only 6% showed intermediate Pr scores (ie, <0.75 for Gl-like or G3- like). Notably, we observed these same percentages in the G2 population of 224 tumours. These findings support the genetic pathways 'model of grade origin, as they suggest that the large majority of breast cancers fundamentally exist in one of two predominant forms marked by the molecular and clinical essence of low or high grade. Whether these forms correlate with the grade-specific genomic alterations previously reported (36, 37, 40) remains to be elucidated.

It should also be noted that although a small percentage (~6%) of the tumours in our study had intermediate genetic grade measurements (ie, analogous to the hybrid

signature observed in Ma et. al. (2003)), too few were discovered to determine the clinical relevance of this intermediate genotype. Furthermore, it is unclear whether these intermediates arise as homogeneous cells that truly borderline low and high grade, or rather represent heterogeneous tumours comprised of distinct low and high grade cell types, such as that observed in tubular mixed carcinoma (38). Alternatively, that we observed the same percentage of intermediacy in tumour classification of all grades and across cohorts, suggests that this class represent a baseline level of uncertainty owing the technical noise.

In conclusion, our results show that the genetic essence of histologic grade can be distilled down to the expression patterns of a mere 5 genes with powerful prognostic implications, particularly in the Grade II setting and in the context of the NPI. The results indicate that G2 invasive breast cancer, at least in genetic terms, does not exist as a significant clinical entity. Indeed, our genetic grade signature dichotomized G2 tumours into two biologically and clinically distinct subtypes that could further be distinguished from Gl and G3 populations. Thus histologic grading, together with measurements of genetic grade, provide a rational basis for the refinement of the G2 subtype into subgrades "2a" and "2b" with immediate clinical ramifications.

Furthermore, our finding that the genetic grade signature could further resolve outcome prediction in G3 tumours, and in a manner dependent on Pr score thresholding, suggests that the genetic grade classifier, viewed as a scalable continuous variable, may have robust prognostic benefit in the diagnosis of all breast tumours. How to optimally weight the genetic grade measurement in combination with other risk factors for greatest prognostic return is a clinical challenge that must next be addressed.

REFERENCES 1. Alizadeh, A.A., Eisen, M.B., Davis, R.E., Ma, C, Lossos, I.S., Rosenwald,

A., Boldrick, J.C., Sabet, H., Tran, T., Yu, X., et al. 2000. Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature 403:503-511.

2. Sorlie, T., Perou, CM., Tibshirani, R., Aas, T., Geisler, S., Johnsen, H., Hastie, T., Eisen, M.B., van de Rijn, M., Jeffrey, S. S., et al. 2001. Gene expression patterns of breast carcinomas distinguish tumour subclasses with clinical implications. Proc Natl Acad Sci U S A 98:10869-10874. ,

3. van 't Veer, LJ., Dai, H., van de Vijver, M.J., He, Y.D., Hart, A.A., Mao, M., Peterse, H.L., van der Kooy, K., Marton, M.J., Witteveen, A.T., et al. 2002. Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530-536.

4. Bullinger, L., Dohner, K., Bair, E., Frohling, S., Schlenk, R.F.,

Tibshirani, R., Dohner, H., and Pollack, J.R. 2004. Use of gene-expression profiling to identify prognostic subclasses in adult acute myeloid leukemia. N Engl J Med 350:1605- 1616. 5. Bloom, H. J., and Richardson, W.W. 1957. Histological grading and prognosis in breast cancer; a study of 1409 cases of which 359 have been followed for 15 years. Br J Cancer 11 :359-377.

6. Elston, C.W., and Ellis, 1.0. 1991. Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer: experience from a large study with, long-term follow-up. Histopathology 19 :403-410.

7. Schumacher, M., Schmoor, C, Sauerbrei, W., Schauer, A., Ummenhofer, L., Gatzemeier, W., and Rauschecker, H. 1993. The prognostic effect of histological tumour grade in node-negative breast cancer patients. Breast Cancer Res Treat 25:235- 245. 8. Roberti, N.E. 1997. The role of histologic grading in the prognosis of patients with carcinoma of the breast: is this a neglected opportunity? Cancer 80:1708- 1716.

9. Lundin, J., Lundin, M., Holli, K., Kataja, V., Elomaa, L., Pylkkanen, L., Turpeenniemi-Huj anen, T., and Joensuu, H. 2001. Omission of histologic grading from clinical decision making may result in overuse of adjuvant therapies in breast cancer: results from a nationwide study. J Clin Oncol 19:28-36.

10. Harvey, J.M., de Klerk, N.H., ^' and Sterrett, G.F. 1992. Histological grading in breast cancer: interobserver agreement, and relation to other prognostic factors including ploidy. Pathology 24:63-68. 11. Frierson, H.F., Jr., Wolber, R. A., Berean, K.W., Franquemont, D.W.,

Gaffey, M.J., Boyd, J.C., and Wilbur, D.C. 1995. Interobserver reproducibility of the Nottingham modification of the Bloom and Richardson histologic grading scheme for infiltrating ductal carcinoma. Am J Clin Pathol 103:195-198.

12. Robbins, P., Pinder, S., de Klerk, N., Dawkins, H., Harvey, J., Sterrett, G., Ellis, L, and Elston, C. 1995. Histological grading of breast carcinomas: a study of interobserver agreement. Hum Pathol 26:873-879.

13. Dalton, L.W., Pinder, S.E., Elston, C.E., Ellis, I.O., Page, D.L., Dupont, W.D., and Blarney, R.W. 2000. Histologic grading of breast cancer: linkage of patient outcome with level of pathologist agreement. Mod Pathol 13:730-735.

14. Younes, M., and Laucirica, R. 1997. Lack of prognostic significance of histological grade in node-negative invasive breast carcinoma. Clin Cancer Res 3:601-604.

15. Hayes, D.F., Isaacs, C, and Stearns, V. 2001. Prognostic factors in breast cancer: current and new predictors of metastasis. J Mammary Gland Biol Neoplasia 6:375- 392.

16. Trudeau, M.E., Pritchard, K.I., Chapman, J.A., Hanna, W.M., Kalin, H.J., Murray, D., Sawka, C.A., Mobbs, B.G., Andrulis, L, McCready, D.R., et al. 2005.

Prognostic factors affecting the natural histoiy of node-negative breast cancer. Breast Cancer Res Treat 89:35-45.

17. Tibshirani, R., Hastie, T., Narasimhan, B., and Chu, G. 2002. Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc Natl Acad Sci U S A 99:6567-6572.

18. Kuznetsov, V.A., Ivshina, A.V., Sen'ko, O.V., Kuznetsova, A.V. 1996. Syndrome approach for computer recognition of fuzzy systems and its application to immunological diagnostics and prognosis of human cancer. Math. Comput. Modeling 23:92-112. 19. Jackson, A.M., Ivshina, A.V., Senko, O., Kuznetsova, A., Sundan, A.,

O'Donnell, M.A., Clinton, S., Alexandroff, A.B., Selby, P.J., James, K., et al. 1998. Prognosis of intravesical bacillus Calmette-Guerin therapy for superficial bladder cancer by immunological urinary measurements: statistically weighted syndrome analysis. J Urol 159:1054-1063. 20. Lukas, J., Herzinger, T., Hansen, K., Moroni, M.C., Resnitzky, D., Helin,

K., Reed, S.I., and Bartek, J. 1997. Cyclin E-induced S phase without activation of the pRb/E2F pathway. Genes Dev 11 :1479-1492.

21. Spruck, C.H., Won, K.A., and Reed, S.I. 1999. Deregulated cyclin E induces chromosome instability. Nature 401 :297-300. 22. Minella, A.C., Swanger, J., Bryant, E., Welcker, M., Hwang, H., and

Clurman, B.E. 2002. p53 and p21 form an inducible barrier that protects cells against cyclin E-cdk2 deregulation. Curr Biol 12:1817-1827.

23. van Diest, P.J., van der ^' Wall, E., and Baak, J.P. 2004. Prognostic value of proliferation in invasive breast cancer: a review. J Clin Pathol 57:675-681.

24. Haybittle, J.L., Blarney, R.W., Elston, C.W., Johnson, J., Doyle, P.J., Campbell, F. C, Nicholson, R.I., and Griffiths, K. 1982. A prognostic index in primary breast cancer. Br J Cancer 45:361-366.

25. Blarney, R. W. 1996. The design and clinical use of the Nottingham Prognostic Index in breast cancer. Breast 5:156-157.

26. Stotter, A. 1999. A prognostic table to guide practitioners advising patients on adjuvant systemic therapy in early breast cancer. Eur J Surg Oncol 25:341-343. 27. Feldman, M., Stanford, R., Catcheside, A., and Stotter, A. 2002. The use of a prognostic table to aid decision making on adjuvant therapy for women with early breast cancer. Eur J Surg Oncol 28:615-619.

28. Lacroix, M., Toillon, R. A., and Leclercq, G. 2004. Stable 'portrait' of breast tumours during progression: data from biology, pathology and genetics. Endocr Relat Cancer 11:497-522.

29. Ma, X.J., Salunga, R., Tuggle, J.T., Gaudet, J., Enright, E., McQuary, P., Payette, T., Pistone, M., Stecker, K., Zhang, B.M., et al. 2003. Gene expression profiles of human breast cancer progression. Proc Natl Acad Sci U S A 100:5974-5979.

30. Lakhani, S.R. 1999. The transition from hyperplasia to invasive carcinoma of the breast. J Pathol 187:272-278.

31. Shackney, S.E., and Silverman, J.F. 2003. Molecular evolutionary patterns in breast cancer. Adv Anat Pathol 10:278-290.

32. Simpson, P.T., Reis-Filho, J.S., Gale, T., and Lakhani, S.R. 2005. Molecular evolution of breast cancer. J Pathol 205:248-254. 33. Tubiana, M., and Koscielny, S. 1991. Natural history of human breast cancer: recent data and clinical implications. Breast Cancer Res Treat 18:125-140.

34. Tabar, L., Fagerberg, G., Chen, H.H., Duffy, S.W., and Gad, A. 1996. Tumour development, histology and grade of breast cancers: prognosis and progression. Int J Cancer 66:413-419.

35. Millis, R.R., Barnes, D.M., Lampejo, O.T., Egan, M.K., and Smith, P. 1998. Tumour grade does not change between primary and recurrent mammary carcinoma. Eur J Cancer 34:548-553.

36. Roylance, R., Gorman, P., Harris, W., Liebmann, R., Barnes, D., Hanby, A., and Sheer, D. 1999. Comparative genomic hybridization of breast tumours stratified by histological grade reveals new insights into the biological progression of breast cancer. Cancer Res 59:1433-1436.

37. Buerger, H., Otterbach, F., Simon, R., Schafer, K.L., Poremba, C, Diallo, R., Brinkschmidt, C, Dockhom-Dworniczak, B., and Boecker, W. 1999. Different genetic pathways in the evolution of invasive breast cancer are associated with distinct morphological subtypes. J Pathol 189:521-526.

38. Cserni, G. 2002. Tumour histological grade may progress between primary and recurrent invasive mammary carcinoma. J Clin Pathol 55:293-297.

39. Hitchcock, A., Ellis, I.O., Robertson, J.F., Gilmour, A., Bell, J., Elston, CW. , and Blarney, R.W. 1989. An observation of DNA ploidy, histological grade, and imrmmoreactivity for tumour-related antigens in primary and metastatic breast carcinoma. J Pathol 159:129-134.

40. Buerger, H., Mommers, E. C, Littmann, R., Simon, R., Diallo, R., Poremba, C, Dockhorn-Dworniczak, B., van Diest, PJ., and Boecker, W. 2001. Ductal invasive G2 and G3 carcinomas of the breast are the end stages of at least two different lines of genetic evolution. J Pathol 194:165-170.

41. Bergh, J., Norberg, T., Sjogren, S., Lindgren, A., and Holmberg, L. 1995. Complete sequencing of the p53 gene provides prognostic information in breast cancer patients, particularly in relation to adjuvant systemic therapy and radiotherapy. Nat Med 1 :1029-1034.

42. Linderholm, B.K., Lindahl, T., Holmberg, L., Klaar, S., Lennerstrand, J., Henriksson, R., and Bergh, J. 2001. The expression of vascular endothelial growth factor correlates with mutant p53 and poor prognosis in human breast cancer. Cancer Res 61 :2256-2260. 43. Lindahl, T., Landberg, G., Ahlgren, J., Nordgren, H., Norberg, T., Klaar,

S. ₅ Holmberg, L., and Bergh, J. 2004. Overexpression of cyclin E protein is associated with specific mutation types in the p53 gene and poor survival in human breast cancer. Carcinogenesis 25:375-380.

44. Miller, L.D., Smeds, J., George, J., Vega, V.B., Vergara, L.,

Ploner, A., Pawitan, Y., Hall, P., Klaar, S., Liu, E.T., et al. 2005. An expression signature for p53 status in human breast cancer predicts mutation status, transcriptional effects, and patient survival. Proc Natl Acad Sci U S A. 45. Kuznetsov, V. A., Knott, G.D., Ivshina, A. V. 1998. Artificial immune system based on syndromes-response approach: Theory and their application to recognition of the patterns of immune response and prognosis of therapy outcome. In Proc. of IEEE Intern. Conf. on Systems, Man, and Cybernetics. San Diego, CA, USA. 3804- 3809. 46. Mi, H., Lazareva-Ulitsky, B., Loo, R., Kejariwal, A., Vandergriff, J.,

Rabkin, S., Guo, N., Muruganujan, A., Doremieux, O., Campbell, M.J., et al. 2005. The PANTHER database of protein families, subfamilies, functions and pathways. Nucleic Acids Res 33 -.D284-288.

47. 1996. Randomized trial of two versus five years of adjuvant tamoxifen for postmenopausal early stage breast cancer. Swedish Breast Cancer Cooperative Group. J Natl Cancer Inst 88:1543-1549.

Sotiriou , T., Perou, CM., Tibshirani, R., Aas, T., Geisler, S., Johnsen, H., Hastie, T., Eisen, M.B., van de Rijn, M., Jeffrey, S. S., et al. 2001. Gene expression patterns of breast carcinomas distinguish tumour subclasses with clinical implications. Proc Natl Acad Sci USA 98:10869-10874. van 't Veer, L.J., Dai, H., van de Vijver, M.J., He, Y.D., Hart, A.A., Mao, M., Peterse, H.L., van der Kooy, K., Marion, M.J., Witteveen, A.T., et al. 2002. Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530-536.

Ma, XJ., Salunga, R., Tuggle, J.T., Gaudet, J., Enright, E., McQuary, P., Payette, T., Pistone, M., Stecker, K., Zhang, B.M., et al. 2003. Gene expression profiles of human breast cancer progression. Proc Natl Acad Sd USA 100:5974-5979.

Miller, L.D., Smeds, J., George, J., Vega, V.B., Vergara, L., Ploner, A., Pawitan, Y., Hall, P., Klaar, S., Liu, E.T., et al. 2005. An expression signature for p53 status in human breast cancer predicts mutation status, transcriptional effects, and patient survival. Proc Natl Acad Sci U S A.

Kuznetsov, V.A., Ivshina, A.V., Sen'ko, O.V., Kuznetsova, A.V. 1996. Syndrome approach for computer recognition of fuzzy systems and its application to immunological diagnostics and prognosis of human cancer. Math. Comput. Modeling 23:92-112.

Kuznetsov, V. A., Knott, G.D., Ivshina, A.V. 1998. Artificial immune system based on syndromes-response approach: Theory and their application to recognition of the patterns of immune response and prognosis of therapy outcome. In Proc. of IEEE Intern. Conf. on Systems, Man, and Cybernetics. San Diego, CA, USA. 3804-3809.

Jackson, A.M., Ivshina, A.V., Senko, O., Kuznetsova, A., Sundan, A., O'Donnell, M.A., Clinton, S., Alexandroff, A.B., Selby, P.J., James, K., Kuznetsov,V.A. 1998. Prognosis of intravesical bacillus Cahnette-Guerin therapy for superficial bladder cancer by immunological urinary measurements: statistically weighted syndrome analysis. J Urol 159:1054-1063.

Mueller, B.U. , Zeichner, SX. , Kuznetsov, V.A., Heath-Chiozzi,M., Pizzo P.A., and Dimitrov, D. S. Individual prognoses of long-term responses to antiretroviral treatment based on virological, immunological and pharmacological parameters measured during the first week under therapy. AIDS, 13, 1998, pp. fl91-fl96. Tibshirani, R., Hastie, T., Narasimhan, B., and Chu, G. 2002. Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc Natl Acad Set USA 99:6567-6572.

Haybittle JL, Blarney RW ₅ Elston CW, Johnson J, Doyle PJ, Campbell FC, Nicholson RI, Griffiths K. et al. A prognostic index in primary breast cancer. Br J Cancer 1982; 45 (3):361-6

Each of the applications and patents mentioned in this document, and each document cited or referenced in each of the above applications and patents, including during the prosecution of each of the applications and patents ("application cited documents") and any manufacturer's instructions or catalogues for any products cited or mentioned in each of the applications and patents and in any of the application cited documents, are hereby incorporated herein by reference. Furthermore, all documents cited in this text, and all documents cited or referenced in documents cited in this text, and any manufacturer's instructions or catalogues for any products cited or mentioned in this text, are hereby incorporated herein by reference. Various modifications and variations of the described methods and system of the invention will be apparent to those skilled in the art without departing from the scope and spirit of the invention. Although the invention has been described in connection with specific preferred embodiments, it should be understood that the invention as claimed should not be unduly limited to such specific embodiments and that many modifications

and additions thereto may be made within the scope of the invention. Indeed, various modifications of the described modes for carrying out the invention which are obvious to those skilled in molecular biology or related fields are intended to be within the scope of the claims. Furthermore, various combinations of the features of the following dependent claims can be made with the features of the independent claims without departing from the scope of the present invention.

APPENDIX 1

SWS Classifier 0

N)

~-4

OO

APPENDIX IA

SWS Classifier 0 Accuracy Gl vs G3

APPENDIX 2

SWS Classifier 0: Prediction of genetic G2a and G2b tumour sub-types based on 264 gene classifier

APPENDIX 3

SWS Classifier O: Tests of differences G2a and G2b by 264 gene classifier

OO

Oo

O

vO

N3

OJ

⁽

,

<

.

APPENDIX 4

SWS Classifier 0: Clinical validation (survival analysis) of G2a and G2b tumour subtypes (264 classifier).

APPENDIX 5A

SWS Classifier 1

Order UGID(build GeneSym Genbank Affi ID Cut-off

#183) Unigen boi Ace eName

1 Hs.528654 Hypothe FLJ 11029 BG16501 B.228273. .a 7.706303 tical 1 t protein

FLJ110

29

2 ace NM 0031 Serine/t STK6 NM 0031 A.208079 _s 6.652593

58.1 hreonin 58 _at e kinase

6. ¹ transcri

Pt 1

3 Hs.35962 CDNA clone BG49235 B.226936 _a 7.561905

IMAGE:4452583, 9 t partial cds

4 Hs.308045 Barren BRRN1 D38553 A.212949 _a 5.916703 homolo t g

(Drosop hila)

5 Hs.184339 Materna MELK NM 0147 A.204825 _a 7.107259

I 91 t embryo nic leucine zipper kinase

6 Hs.250822 Serine/t STK6 NM 0036 A.204092 _s 6.726571 hreonin 00 ^' _at e kinase

6, transcri pt 2

APPENDIX 5B

SWS Classifier 1 : Classifier Accuracy

APPENDIX 5C

SWS Classifier 1 : Prediction validation

* DFS Event defined as any type of recurrence or death because of breast cancer, whichever comes first

00357

103

APPENDIX 6A

SWS Classifier 2

APPENDIX 6B

SWS Classifier 2: Accuracy

APPENDIX 6C

SWS Classifier 2: G2a-G2b Prediction and Survival

* DFS Event defined as any type of recurrence or death because of breast cancer, whichever comes first

2007/000357

109

APPENDIX 7 A

SWS Classifier 3

APPENDIX 7B

SWS Classifier 3: Classifier Accuracy

|"V/IIW"- ^» fc« W w ^« / u v/ U _* J o i

113

APPENDIX 7C

SWS Classifier 3: G2a-G2b Prediction Validation

* DFS Event defined as any type of recurrence or death because of breast cancer, whichever comes first

G2007/000357

117

APPENDIX 8A

SWS Classifier 4

APPENDIX 8B

SWS Classifier 4: Classifier Accuracy

APPENDIX 8C

SWS Classifier 4: G2a-G2b Prediction Validation

* DFS Event defined as any type of recurrence or death because of breast cancer, whichever comes first