Field of the invention

The invention relates to the treatment, detection and classification of cancer, including the identification of heterogeneous tumours and in particular relates to breast cancer. It also relates to identifying patients who are likely to respond to cancer therapy with a PP2A activator.

The invention defines the use of biomarkers (ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 expression) and the need of the novel biomarkers AURKA and KIF2C to classify breast cancer patients as Basal or Luminal. In addition, a method is described to further classify Luminal cases into good (Luminal A-like) or bad (Luminal B-like) prognoses.

The invention also relates to methods useful in predicting if a sample comprises a gene amplification or gene reduction, or high or low gene expression.

Background

The provision of new treatments for cancer is of high importance, including for cancers resistant to known treatments. There is also a need for the identification of markers that allow for the detection, prognosis and classification of cancer, and which can predict responsiveness of a patient to a given therapy. For example, the deregulation of the protein phosphatase 2A (PP2A) complex is known to be a common event in cancer (Grech et al; Tumour Biol; D01 10.1007/s13277-016-5145-4 (2016)) and PP2A activators have recently been shown to be useful in treating breast cancers that overexpression AURKA and/or KIF2C (GB. 1704536.0, the disclosure of which is specifically incorporated herein by reference).

Due to the genetic instability of tumour cells, genomic rearrangements frequently result in gene amplification or gene loss. Thus the overexpression of some genes, for instance HER2 in some breast cancers, is not mediated at the transcriptional level per se but is instead the result of an increase in gene copy number. The identification of gene amplification or the loss of genes that are directly involved in tumorigenesis or tumour suppression can be clear indicators of the presence of a tumour, or of a particular tumour subtype. Additionally the detection of an amplification or loss event of a gene that is not itself involved in tumorigenesis but which occurs as the result of the type of genomic instability that is frequently found in tumours can also be a marker of the presence of tumour cells. For example, a particular indicator gene may be duplicated as part of the translocation process that results in the amplification of an oncogene, and increased expression from the indicator gene may be used as a surrogate to indicate duplication of the oncogene.

Molecular classification of breast cancer interrogates molecular markers to categorise patients into four molecular classes, namely luminal A-like or luminal A, luminal B-like or luminal B, the human epidermal growth factor receptor 2 (HER2)-enriched and the basal types. Luminal A subtype is positive for oestrogen receptor (ER) and/or progesterone receptor (PgR) expression with low expression of Ki-67, while luminal B, apart from having an ER/PgR positive expression, includes also HER2 positive and negative subgroups associated with high Ki-67 expression. The HER2-enriched subtype is well defined, with high expression of HER2 receptor, due to the ERBB2 gene amplification, combined with low or absent ER and PgR. Interestingly, HER2 expression was found to be present in microvesicles originating from tumour cells, that induced activation of mitogenic signals in recipient fibroblasts ¹⁰ . Hence, circulating microvesicles such as exosomes are potentially vehicles of early detection and early indicators of metastasis and relapse ¹¹ . Basal type tumours are in general negative for the 3 receptors significantly overlapping with the triple negative breast cancer (TNBC) diagnostic subtype ⁵ · ⁶ . Other markers are used to determine epithelial and mesenchymal characteristics. Fibronectin (FN1 ) is a main component of breast tissue mesenchymal compartment. Increased FN1 expression is accompanied by high Ki67 staining showing a signature for a more invasive tumour ^{7 8} and an increased expression is associated with metastasis ⁹ .

The luminal A-like or luminal A, and luminal B-like or luminal B subclasses are classified based immunohistochemical and clinicopathological criteria set out in the St Gallen 2013 conference (Harbeck, N., Thomssen, C., & Gnant, M. (2013). St. Gallen 2013: Brief Preliminary Summary of the Consensus Discussion. Breast Care, 8(2), 102-109. http://doi.Org/10.1 159/000351 193). Alternative classifications have been proposed (Maisonneuve et al 2014 Breast Cancer Res 16(3): R65; and Inic et al 2014 Clinical Medicine Insights. Oncology, 8, 107-1 1 1. http://doi.orq/10.4137/CMQ.S18006).

Each subtype of cancer may be associated with a different response to a particular treatment and so identification of the subtype of cancer is important to ensure that the subject receives the most appropriate treatment. In the clinical setting, gene amplification events are routinely assayed by either immunohistochemistry (IHC) and in-situ hybridization (ISH). IHC testing can show how much of the particular protein is present on the cancer cell surface which may in some situations be considered to be the most relevant parameter to assay, while ISH testing measures the number of copies of the particular gene inside each cell. There are two main types of ISH tests: fluorescence and bright-field ISH (www.cancer.net).

These are the only two methods currently approved by the FDA in the United States to test for the HER2 status of a tumour (Perez et al 2014 Cancer Treatment Reviews

Volume 40, Issue 2, March 2014, Pages 276-284).

However, both IHC and ISH suffer from many problems that make them less than ideal tests to be carried out on large numbers of samples. For example, both tests require a visual interpretation of the degree of hybridisation of a probe, either to the DNA or protein, and are therefore susceptible to variation between pathologists and pathology labs. Both methods require a high quality sample and are affected by degradation, fixation time and care in handling.

ISH is considered to be more accurate than IHC, though it is IHC that is the more widely used since it is more practical.

However, ISH is also susceptible to problems such as loss of the control CEP17 chromosomal region, or the concomitant amplification of CEP17 in ERBB2 amplified cells, both of which would present misleading results.

In addition, some breast cancer cell lines such as MDAMB453 and JIMT1 harbour the ERBB2 gene amplification but show a relatively low expression of HER2 that is undetectable by HER2 immunohistochemistry (Koninki et al., 2010. Kao et al., 2009). Despite the amplification of ERBB2, the JIMT1 ERBB2 gene amplified cell line is resistant to trastuzumab and has reduced levels of the ERBB2 transcript compared to SKBR3 and BT474 HER2 positive cell lines (Tanner et al., 2004, Koninki et al., 2010).

The Cancer Genome Atlas (TCGA) data accessed through the cBioPortal shows that the discrepancy between gene copy number and transcript is also seen in breast cancer patient samples. For example (15.31 %) of cases show ERBB2 transcript overexpression while only 1 19 (12.4%) cases show ERBB2 gene amplification. Hence 4.17% of cases show increased HER2 expression but no gene amplification (false negative cases by genetic studies) and 1.25% of cases are ERBB2 gene amplified but do not show overexpression (false positive cases by genetic studies) (Figure 1 ).

The use of whole sections from resected material indicate the presence of heterogenous samples that are missed when diagnosis is done on biopsies (Table 1 ). A local study assessed 258 positive or equivocal HER2 resection specimens by Immunohistochemistry to assess heterogeneity. Technical staining variation attributed to inadequate specimen fixation, represented by peripheral tissue staining accompanied by a negative central area, was identified in 16 out of 258 (6.20%) sections. 45 (17.44%) resection specimens exhibited HER2 staining heterogeneity. Hence the need of high throughput methodologies to test biopsies and resected material will provide additional information on patient status.

Table 1: A comparative audit of diagnostic HER2 results on breast biopsy specimens as opposed to the patient-matched breast resection specimen where data was available. Out of 2303 breast cancer cases diagnosed at Mater Dei Hospital , Malta between 2009-2016, only 111 breast cancer cases (4.8%) had a diagnostic result for HER2 for both the biopsy and the main specimen. A discrepancy in the HER2 status was observed in 29/111 (26.13%) of cases (unpublished results).

Thus both IHC and ISH have several drawbacks and are not ideal for the routine and widespread determination of tumour status. Gene amplification and gene expression diagnostic testing using archival material or material that requires transportation to servicing laboratories, needs a more robust and accurate test adapted to current clinical workflow.

Some attempts have been made to produce panels of biomarkers to aid in breast cancer diagnosis and classification. For instance, the Prosigna Breast Cancer Prognostic Gene Signature Assay (NanoString) based on PAM50 biomarker panel was US FDA Approved in September 2013 and CE mark in 2012. However, the Prosigna assay is designed as an indicator of prognosis of tumours and not classification, therapeutic selection or to detect response ¹² .

In addition a 12-gene signature (W02009055823) was developed for Basal versus Luminal classification only as compared to PAM50 classification ¹⁸ and to predict therapeutic response to polyamine type chemotherapy.

Summary of the invention

The present invention provides an accurate method or test to allow the classification of a cancer as either cancerous, or a sub-class of that cancer, and in particular relates to cancers or sub-cancers that are known to be associated with a gene amplification event, for example the ERBB2 gene in some breast cancers.

By using an algorithm trained on we 11 -an notated breast cancer samples, the transcription level is used to diagnose the status of the cell with respect to that gene or genes, for example whether a breast cancer sample is HER2 positive or negative. The present method is improved over the currently used ISH method since discrepancies between amplification at the level of chromosomes and levels of transcribed RNA are recorded both in cell lines and patient material; and is improved over the currently used IHC method for at least the reasons given above.

The invention also provides a further method which classifies breast cancer tumours into various sub-classes of breast cancer.

The tests described herein in some circumstances rely on the selection of appropriate normalisation genes. Accordingly, the invention also provides a set of normalisation genes useful for use with the tests described above when applied to breast cancer.

The inventors have previously shown (GB. 1704536.0, data reproduced herein) that cancer cells sensitive to treatment with a PP2A activator exhibit overexpression of the markers AURKA and KIF2C. These include cancer cells exemplified by Triple negative breast cancer (TNBC) cells that do not benefit from targeted therapy and have bad prognosis when current front-line anti-cancer therapies are administered. The markers AURKA and KIF2C further correlate with low PP2A enzymatic activity. The markers AURKA and KIF2C may thus be used to predict responsiveness of a patient to treatment with a PP2A activator. AURKA and KIF2C may also be targeted with antagonists to thereby treat cancer, based on their correlation with cancer. In addition, the inventors have also shown that the markers AURKA and KIF2C have utility in detecting, prognosing and classifying cancer. The markers may be used to classify cancers with a worse prognosis, in particular basal versus luminal breast cancer subtype and subsets of luminal cases with bad prognosis (Luminal B). The markers may thus be assayed using suitable reagents in kits for detection and classification of cancer and for predicting therapeutic benefit from activation of the PP2A complex.

The present invention aims to address the problems of the prior art discussed above and provides methods (exemplified by the role of HER2 in breast cancer and the classification of breast cancers as HER2 positive (HER2+), triple negative breast cancer (TNBC), Basal, Luminal A or Luminal B) but which can now be generalised and extrapolated to other genes involved in various diseases, for the determination of the status of a given tumour, tumour sample or tumour subsample with regards to a particular gene or set of genes that are quick, can be standardised, are much less susceptible to user interpretation and which advantageously can be used on degraded samples, which makes these methods suitable for use in situations where it isn’t possible to handle or store samples under optimum conditions.

Optimisation of RNA based assays using archival formalin-fixed paraffin-embedded (FFPE) material is challenging due to variability in surgical tissue processing and degradation of RNA caused by tissue integrity preservation using Formalin ¹ · ² . To overcome the limitation of performing accurate gene expression studies from archival material, the invention makes use of branched-chain DNA (bDNA) technology, for example the QuantiGene ^® technology. bDNA technology replaces enzymatic amplification of target template with hybridisation of specific probes and amplification of a reporter signal ³ . The short recognition sequences of the capture and detection probes are designed to hybridise to short fragments of target RNA ⁴ . In addition, the use of tissue homogenates directly as starting material of this assay, overcomes the inevitable loss of RNA occurring in assays requiring prior RNA extraction and purification and also allows multiple assays to be performed on the same sample. Multiplex technology such as the Luminex technology provides the possibility to multiplex the assay, measuring expression of a panel of targets from low material input.

The fact that unlike previous methods the present invention can be used in conjunction with archived, degraded, formalin-fixed and paraffin-embedded samples means that the present invention can be used to readily identify novel biomarkers by using historic samples.

The invention is further advantageous since the multiplex nature of one embodiment of the invention allows a single assay to be performed which can fully classify a cancer, for example a breast cancer.

In contrast to the biomarker panels and methods of the prior art the methods of the invention (which include in one embodiment the following markers ESR1 , HER2, PGR, AURKA, KIF2C, FOXC1 ) along with particular normalisation genes is able to classify breast cancer in the current diagnostic groups (as determined by IHC and FISH) with high accuracy (see for example Examples 2-7). The potential of this molecular-based assay with quantitative measurements has potential to digitalise the current methodologies of breast cancer diagnosis. Therefore, this assay can assist in the diagnosis of breast cancer as well as determine therapeutic decisions. In addition, based on the earlier finding (GB. 1704536.0) that AURKA and KIF2C are biomarkers of PP2A activity, classification of breast cancer subtypes using these markers set another potential therapeutic subtype, that is considered to benefit from PP2A activation therapy.

The prediction by an algorithm as used herein based on ESR1 , ERBB2, PGR, AURKA, KIF2C and FOXC1 expression (6 biomarker panel) allows the definition of Luminal and Basal subtypes and the same measurements can be used to discriminate between Luminal A and Luminal B when prediction is run using the Luminal cases only (see for example Examples 2-4). The use of the methods and panels of the invention reduces the cost of runs, digitalise the workflow, and allows minimal sample requirement through multiplexing of the RNA-based assay.

The compatibility of this assay and algorithm with H&E staining allows for the use of precisely laser microdissected material that allows isolation of specific cellular populations within heterogeneous samples which is considered to result in higher sensitivity due to increased percentage of tumour cell content per sample. In contrast, the prior art methods, for example those of Prosigna requires an unstained macro-dissection FFPE specimen for analysis.

Furthermore, PAM50 algorithms have been trained on microarray and RT-PCR data which require RNA extraction ¹⁹ . The methods described herein utilise algorithms that have been assessed on RNA Sequencing data as well as Quantigene 2.0 data. The latter methodology eliminates extraction bias, and is minimally affected by degraded RNA.

Methods of the invention can also be used to detect heterogeneity within a single tumour sample, for instance, can detect the presence of both luminal and basal subtypes within the same tumour, in some instances without the need of, for instance, microdissection and separate analysis of morphologically distinct tumours identified following staining and microscopy. This is important since whilst the majority of a tumour may be of one sub- type and may respond to a particular treatment, the presence of an additional tumour subtype which may not respond to that treatment indicates that an alternative or additional treatment may be needed.

Detailed description of the invention

In a first aspect, the present invention provides a method for classifying a cancer into one or more sub-classes wherein the method involves the use of bDNA technology.

As discussed above, it is well known that although on the face of it a cancer is described as, for example, a breast cancer, within that broad classification exists a heterogenous range of cancers, distinguished by unique phenotypic and genotypic differences and which can each require a different therapy. For example, breast cancer can be classified as HER2+, Basal, Luminal (which encompasses both the Luminal A and Luminal B sub classes), and each of which responds better to different therapies. For example, HER2+ breast cancers are known to be suitable for treatment with Herceptin (trastuzumab), Kadcyla (Herceptin and emtansine), Nerlynx (neratinib), Perieta (pertuzumab), and Tykerb (lapatinib), all of which take advantage of the overexpression of the HER2 protein to specifically target the cancer cell. Such treatments are considered to be less effective and/or not targeted to the cancer cell if the cell does not either harbour the HER2 gene amplification or the has some equivalent mutation/alteration that increases the expression of HER2. It is important to note that, as stated above, not all gene amplification events of the ERRB2 gene which encodes HER2 leads to an increase in HER2 mRNA and HER2 protein. Since the standard clinical tests generally involves assessment of the DNA gene copy number of ERRB2 such a sample would erroneously be classified as HER2+. In this case a patient may be subjected to treatment with an anti-HER2 therapy which is likely to be ineffective. As also stated above, although some clinical tests look at the actual level of the protein made, for example the HER2 protein, these tests are highly subjective and are subject to inconsistencies between laboratories, degree of fixation and user interpretation. Similarly, the ER and PR status of breast cancer samples is also assessed by IHC and susceptible to the same problems. The present invention addresses these issues.

Although the present invention is exemplified by the development and optimisation of a method to classify breast cancer samples, for example as HER2+ cancers, or basal cancers or Luminal A or Luminal B cancers, the methods and algorithms described herein can be used to similarly develop and optimise methods to classify other cancers, for example other cancers that have a sub-class associated with a gene amplification. The skilled person will understand for which cancers and sub-classes of cancers the present invention is appropriate, for example by understanding which cancers or sub-classes of cancers are associated with which gene amplifications.

Examples of other cancers and sub-classes of cancers for which the present invention is considered useful are:

RUNX1 gene amplification

Intrachromosomal amplifications of chromosome 21 involves amplification of the gene RUNX1 , defining a subgroup of B-cell precursor Acute Lymphoblastic Leukemia (ALL), predicting high relapse rare. Intensifying the therapy in these patients significantly reduce the replication rate [Harrison CJ: Blood spotlight on iAMP21 acute lymphoblastic leukemia (ALL), a high-risk pediatric disease. Blood 125: 1383-1386 (2015)].

MYCN gene amplification

MYCN gene amplification is found in various cancer types including colorectal cancer, neuroblastoma and others. MYCN gene amplification is an independent adverse prognostic factor in neuroblastoma predicting patients associated with rapid disease progression, involving all ages and stages [Thompson D, Vo KT, London WB, Fischer M, Ambros PF, et al: Identification of patient subgroups with markedly disparate rates of MYCN amplification in neuroblastoma: a report from the International Neuroblastoma Risk Group Project. Cancer 122: 935-945 (2016).]. hTERC gene amplification

hTERC gene amplification in liquid-based cervical samples was found in 37% of HPV genotype positive individuals, and 70% of hTERC amplified samples were diagnosed as CIN2+. hTERC amplification significantly improves the specificity and positive predictive value of HPV screening [Zappacosta R, lanieri M M, Buca D, Repetti E, Ricciardulli A, and Liberati M. Clinical Role of the Detection of Human Telomerase RNA Component Gene Amplification by Fluorescence in situ Hybridization on Liquid-Based Cervical Samples: Comparison with Human Papillomavirus-DNA Testing and Histopathology. Acta Cytologica 2015;59:345-354]

PD-L1 and PD-L2 expression level

Programmed death-ligand 1 (PD-L1 ) is a protein that in humans is encoded by the CD274 gene while PD-L2 is encoded by the PDCD1 LG2 gene. The binding of PD-L1 or PD-L2 with the PD-1 receptor on T cells induces a signal that inhibits TCR-mediated activation of IL-2 production and T cell proliferation [Sheppard KA, Fitz LJ, Lee JM, Benander C, George JA, Wooters J, Qiu Y, Jussif JM, Carter LL, Wood CR, Chaudhary D (September 2004). "PD-1 inhibits T-cell receptor induced phosphorylation of the ZAP70/CD3zeta signalosome and downstream signaling to PKCtheta". FEBS Letters. 574 (1 -3): 37-41 ; Karwacz K, Bricogne C, MacDonald D, Arce F, Bennett CL, Collins M, Escors D (October 201 1 ). "PD-L1 co-stimulation contributes to ligand-induced T cell receptor down- modulation on CD8+ T cells". EMBO Molecular Medicine. 3 (10): 581 -92.]. . This interaction is thought to be one of the causes of how tumour cells might evade detection and destruction by the body’s immune system. Gene amplification, gene translocation and gene overexpression have been implicated in the overexpression of these surface bound proteins [Georgiou K Chen L Berglund M Ren W de Miranda N et. al. Genetic basis of PD- L1 overexpression in diffuse large B-cell lymphomas. Blood 2016 vol: 127 (24) pp: 3026- 34] Many PD-L1 inhibitors are already in clinical use with many others in development and are showing good results in clinical trials ["Immune checkpoint inhibitors to treat cancer" www.cancer.org. Retrieved 2018-02-1 1 ]

MET amplification

MET is a proto-oncogene that encodes a receptor tyrosine kinase. The aberrant activation of MET signalling in a subgroup of cancers, is an example of how certain cancer become dependent on a single overactive oncogene for their proliferation and survival, a phenomenon that has become known as“oncogene addiction”. This activation is usually the result of gene amplification, polysomy, and gene mutations. MET deregulation can be identified in various human malignancies, including cancers of kidney, liver, stomach, breast, and brain [Kawakami H Okamoto I Okamoto W Tanizaki J Nakagawa K et. al. Targeting MET Amplification as a New Oncogenic Driver. Cancers 2014 vol: 6 (3) pp: 1540-52] AR-V7

Androgen deprivation therapy provides effective, though temporary tumour control in patients with metastatic prostate cancer [G. Attard, J.S. de Bono Translating scientific advancement into clinical benefit for castration-resistant prostate cancer patients Clin Cancer Res, 17 (201 1 ), pp. 3867-3875,]. Androgen deprivation therapy can be achieved through a number of chemotherapeutic protocols. But there is clinical evidence that not all patients benefit from such therapy [H.l. Scher, K. Fizazi, F. Saad, M.-E. Taplin, C.N. Sternberg, K. Miller, et al. Increased survival with enzalutamide in prostate cancer after chemotherapy N Engl J Med, 367 (2012), pp. 1 187-1197,] Overexpression of nuclear AR- V7 protein in primary prostate cancer has been identified as an independent negative prognostic marker in men with high-risk disease receiving adjuvant therapy [Chen X Bernemann C Tolkach Y Heller M Nientiedt C et. al. Overexpression of nuclear AR-V7 protein in primary prostate cancer is an independent negative prognostic marker in men with high-risk disease receiving adjuvant therapy Urologic Oncology: Seminars and Original Investigations 2017]

FGFR1 amplifications

The fibroblast growth factor receptor (FGFR) family is a novel potential therapeutic target in cancer. FGFR plays an impoartant role in stimulating cell proliferation and migration as well as in promoting survival of various types of cells [Beenken A, Mohammadi M. The FGF family: biology, pathophysiology and therapy. Nat Rev Drug Discov 2009; 8: 235- 253.] Though different FGFR aberrations, including receptor overexpression through gene amplification or post-transcriptional regulation, FGFR mutations, FGFR translocations, alternative splicing of FGFR, have been identified, all active aberrations constitutively activate downstream pathways and contribute to tumour development. Most of these abnormalities achieve this through the overexpression of FGFR [Dienstmann R, Rodon J, Prat A, et al Genomic aberrations in the FGFR pathway: opportunities for targeted therapies in solid tumors. Ann Oncol 2014; 25: 552-563]. Aberrant FGFR signalling has been observed in many cancers, including NSCLC, bladder, breast, prostate, ovarian, endometrial, gastric, colorectal cancer, head and neck and glioblastoma [Parker BC, Engels M, Annala M, et al Emergence of FGFR family gene fusions as therapeutic targets in a wide spectrum of solid tumours. J Pathol 2014; 232: 4-15]. Inhibition with several small molecules known to target the FGF/FGFR-pathway has shown in vitro anti-tumour activity within tumours with FGFR1 -3 aberrations [Greenman C, Stephens P, Smith R, et al Patterns of somatic mutation in human cancer genomes. Nature 2007; 446: 153-158; Wu YM, Su F, Kalyana-Sundaram S, et al Identification of targetable FGFR gene fusions in diverse cancers. Cancer Discov 2013; 3: 636-647]

Other amplifications include PIK3CA in Lung, ovarian and Head and Neck cancers [Meric- Bernstam F et al 2015. A Decision Support Framework for Genomically Informed Investigational Cancer Therapy JNCI J Natl Cancer Inst (2015) 107(7): djv098].

The method of the invention may be used to identify new sub-classes of cancers. For example, the method could be used to assess the RNA expression level of a multitude of samples and analysis of the results can reveal novel therapeutic and/or prognostic tumour sub-classes of cancers of distinct tissue origin, for example new sub-classes of lung cancer, or new sub-classes of colon or breast cancer, each with a unique and distinctive expression pattern. Methods according to this embodiment are discussed below in subsequent aspects of the invention.

Cancers are generally classified according to one or more various markers that the cancer displays, for example the cancer may have one or two or more genes are expressed to a higher level than the expression level of those genes in another sub-class of the same cancer. For example, although the HER2+ subclass of cancers is typically defined by the presence of an amplification of the ERBB2 gene, it is the important consequence of this amplification, i.e. an increased expression of the ERBB2 gene, i.e. an increased level of the HER2 mRNA and protein that truly makes the HER2+ sub-class of breast cancer a suitable target for certain therapies such as Herceptin. Similarly, an ER+ cancer is a cancer in which a certain number of the cells in the sample display the ER, as determined by IHC.

There is an extensive list of other classifiers, for example in breast cancer diagnosis E- Cadherin positivity defines cancers as of ductal origin as opposed to being of lobular origin (E-Cadherin negative) ; whilst a mutation in the c-kit gene is associated with c- kit gastrointestinal stromal tumours (GIST).

It will be understood then that the level of RNA associated with each of the key genes that are used to classify the different sub-classes of cancers is a more accurate and informative method of producing clinically relevant classifications than assessment of gene copy number. The method provided herein, which is quick, sensitive and not subject to pathologist error or interpretation is also improved over those methods such as IHC which directly assess the resultant protein.

The present invention makes use of the RNA expression levels of at least 1 , for example at least 2, optionally at least 3, or at least 4, or at least 5, or at least 6, or at least 7, or at least 8, or at least 9, or at least 10 genes that are known to be associated with various cancer sub-classes. Although the RNA expression level can be assessed by any means, for example by reverse transcription PCR (rtPCR), a preferred embodiment utilises branched-DNA (bDNA) technology. This is advantageous for a number of reasons.

In a first aspect, the present invention provides a method for classifying a cancer into one or more sub-classes wherein the method involves the use of bDNA technology. bDNA technology replaces enzymatic amplification of target template with hybridisation of specific probes and amplification of a reporter signal ³ . The short recognition sequences of the capture and detection probes are designed to hybridise to short fragments of target RNA ⁴ .

Since the use of bDNA removes the requirement for RNA extraction and allows the direct use of a tissue homogenate from, for example, fresh tissue, fresh-frozen tissue, FFPE tissue sections or a laser dissected, stained or unstained sample, or exosomes, the technical variation obtained by using this assay is much reduced.

The skilled person has the necessary skills and knowledge to be able to design suitable target and amplification probes for use in a bDNA assay, see for example US8426578 and Bushnell et al 1999 Bioinformatics 15: 348-55.

Examples of commercially available kits for performing bDNA assays include the QuantiGene Plex assay from Thermo Fisher; Chiron branched DNA signal amplification (bDNA) assay used for viral molecular diagnositics [VERSANT® HIV-1 RNA 3.0 Assay (bDNA) ]; Diacarta (http://www.diacarta.com/technology/bdna-signal-amplificatio n- technology/); and RNAscope (http://acdbio.com/science/technoloqv-overview).

Briefly, branched DNA uses sequential hybridization of oligonucleotides to a captured target RNA in order to amplify a signal for quantitative measurement. In the context of a multiplex assay, for example using Luminex beads, the sample is first added to a bead mix that consists of both the magnetic Luminex beads as well as a set of probes that are used to capture the target RNA. During the first incubation, the capture extenders hybridize to the capture probes conjugated to the beads while also hybridizing to the target RNA sequence. This captures the target RNA onto the desired beads through a process called ‘cooperative hybridization’. Each bead colour has its own target-specific set of probes, allowing multiple genes to be captured onto different beads.

Also hybridizing to the target RNA are label extenders, which provide the basis for the branched DNA signal amplification structures. These label extenders are always designed in pairs to enhance the specificity of the assay. The third type of probe, the blocking probe, hybridizes to any piece of the target sequence that is not already targeted by the capture extenders or label probes. The purpose of the blocking probes is to form a complete double-stranded piece of RNA, protecting it from RNases and helping to prevent secondary structure within the target region.

The capture extenders, label extenders, and blocking probes all comprise the target- specific probe set which can be designed and provided commercially, for example from Thermo Fisher.

Next, the branched DNA oligonucleotides form the signal amplification structure. First, a pre-amplifier hybridizes to the label extender pairs. In the next incubation, many amplifiers hybridize to each pre-amplifier, and in the following incubation, many label probe oligonucleotides hybridize to each amplifier. The label probe molecule is conjugated with biotin, so when streptavidin-phycoerythrin is added in the last step, a fluorescent signal is created and measured.

Although the invention is herein exemplified by use of the QuantiGene Plex Assay, it is considered that any bDNA assay can be used in the methods of the invention. In one advantageous embodiment, the method involves the detection of RNA levels of relevant genes via the QuantiGene Plex Assay.

Advantageously, in one embodiment the bDNA assay, for example a QuantiGene Plex Assay, is a multiplex assay, which allows the simultaneous detection the expression level of many RNA species. Accordingly, in one embodiment the bDNA assay is a multiplex assay, for example an assay that allows the simultaneous detection of the expression level of more than 1 RNA species, optionally more than 2 RNA species, optionally more than 3 RNA species, optionally more than 4 RNA species, optionally more than 5 RNA species, optionally more than 6 RNA species, optionally more than 7 RNA species, optionally more than 8 RNA species, optionally more than 9 RNA species, optionally more than 10 RNA species, optionally more than 1 1 RNA species, optionally more than 12 RNA species, optionally more than 14 RNA species, optionally more than 16 RNA species, optionally more than 18 RNA species, optionally more than 20 RNA species, optionally more than 30 RNA species, optionally more than 40 RNA species, optionally more than 50 RNA species.

Multiplex assays of any format are considered useful in the present invention. For example microbead technology, for example the Luminex platform, is considered useful and advantageous. Accordingly in one embodiment the detection of the level of RNA expression is by the use of a multiplex platform, optionally a microbead platform, optionally the Luminex platform.

As described above, in one advantageous and preferred embodiment the detection of the level of RNA expression is by the use of the QuantiGene Plex assay (Thermo Fisher).

In one embodiment the present invention allows the classification of the cancer into at least two sub-classes, optionally at least three sub-classes, optionally at least four-subclasses, optionally at least five or more sub-classes.

As discussed above, the sub-classes of cancer may be associated with a gene amplification or gene reduction, or may be associated with a change in RNA expression level that occurs independently of a gene amplification or reduction event. In one embodiment one or more of the sub-classes of cancer is associated with a gene amplification or a gene reduction event. By associated with a gene amplification or gene reduction event we include the meaning that that particular class of cancer typically encompasses cancers with a gene amplification or gene reduction event, but does not mean that all cancers that make up that sub-class necessarily have the gene amplification or gene reduction. For example, the well known HER2+ sub-class of breast cancers typically arises due to a gene amplification of the ERBB2 gene, and currently clinical diagnosis of the HER2+ phenotype often involves detection of a gene amplification event. However, the skilled person will appreciate that there are many factors that can affect the expression level of a particular gene, for example whether a particular gene is overexpressed or not, for example there may be some breast cancers that have a mutation in the ERBB2 gene which results in an increased level of mRNA, for example due to a decreased rate of ERBB2 mRNA turnover. In this case, such a sample may still be considered to be a HER2+ cancer, provided the expression level of ERBB2 reaches a particular threshold. Typically the threshold is set according to the level of RNA expression of samples that are known to harbour the gene amplification or gene reduction event, for example that are known to be HER2. The skilled person is well equipped to determine the relevant threshold levels of expression using standard lab techniques.

In another embodiment, one or more of the sub-classes of cancer is not associated with a gene amplification or gene reduction event, for example is associated with a change in the RNA expression level of a particular gene due to other mechanisms. For example, such mechanisms can involve a mutation in the gene that is used to classify the cancer, or a mutation in a gene that encodes for a regulatory protein that affects the expression level of the gene that is used to classify the cancer.

Other mechanisms that can result in altered gene expression that is used to classify the cancer include:

altered signal transduction pathway due to increased or decreased proliferation signals or lack of a negative feedback mechanism (exemplified by low PP2A activity); genetic translocations;

activating mutations;

reduced expression/activity of tumour suppressors due to gene deletion, inactivating mutations; and

epigenetic mechanisms such as hypermethylation.

It is considered that hypermethylation in particular has significant effects on gene expression and the resulting changes in RNA expression can be detected using the methods of the invention.

In one embodiment at least one sub-class of cancer is associated with a gene amplification event. In another embodiment, at least one sub-class of cancer is associated with an overexpression of a particular gene.

Accordingly in one embodiment one or more of the sub-classes of cancer is associated with an increase or decrease in copy number of a gene.

In a further embodiment the cancer can be classified into at least one sub-class of cancer associated with a gene amplification event and at least one sub-class of cancer that is not associated with a gene amplification event, optionally at least one sub-class of cancer that is associated with an overexpression or an underexpression of a particular gene. The expression level of any number of RNA species may be determined in order to classify the cancer into the various sub-classes. For example, in one embodiment the method comprises the determination of the level of RNA expression of optionally at least one RNA species, optionally more than 2 RNA species, optionally more than 3 RNA species, optionally more than 4 RNA species, optionally more than 5 RNA species, optionally more than 6 RNA species, optionally more than 7 RNA species, optionally more than 8 RNA species, optionally more than 9 RNA species, optionally more than 10 RNA species, optionally more than 1 1 RNA species, optionally more than 12 RNA species, optionally more than 14 RNA species, optionally more than 16 RNA species, optionally more than 18 RNA species, optionally more than 20 RNA species, optionally more than 30 RNA species, optionally more than 40 RNA species, optionally more than 50 RNA species. Some of the RNA species may be associated with the cancer or sub-class of the cancer, whilst others may not be associated with the cancer or sub-class of the cancer, and are used for normalising the expression levels of the genes that are associated with the cancer or sub cancer.

The skilled person will be very aware of the significance in choosing the correct normalisation genes. Methods of doing so are described herein, along with a specific and novel set of markers that can used with the method of the invention for the classification of breast cancer. Accordingly, the invention provides particular markers, ACTB; PPIB; HPRT 1 ; and/or TBP, which either alone, in particular combinations, or all together, can be used to normalise the data obtained from the method of the invention when applied to breast cancer. The skilled person will appreciate that the selection of appropriate normalisation genes is affected by cell or tissue type as well as the expression level of the “test” genes, i.e. the genes that are associated with the cancer or sub-cancer.

Accordingly, in one embodiment, the method involves the determination of the expression level:

of at least one or more genes associated with the cancer or at least one sub-class of cancer, optionally at least one gene associated with the cancer or at least one sub-class of the cancer, optionally at least two genes associated with the cancer or at least one sub class of the cancer, optionally at least three genes associated with the cancer or at least one sub-class of the cancer, optionally at least four genes associated with the cancer or at least one sub-class of the cancer, optionally at least five genes associated with the cancer or at least one sub-class of the cancer, optionally at least six genes associated with the cancer or at least one sub-class of the cancer, optionally more than six genes associated with the cancer or at least one sub-class of the cancer, optionally wherein the at least one or more genes associated with the cancer or at least one sub-class of cancer are selected from the group consisting of ERBB2, ESR1 , AURKA, KIF2C, PGR and FOXC1 and/or

of at least one gene not associated with the cancer or a sub-class of the cancer, of at least two genes not associated with the cancer or a sub-class of the cancer, of at least three genes not associated with the cancer or a sub-class of the cancer, of at least four genes not associated with the cancer or a sub-class of the cancer, optionally wherein the one or more genes has been determined to be suitable for use as a normalising gene, optionally determined by an algorithm trained on a dataset of known cancer classification, optionally wherein the at least one gene not associated with the cancer or sub-cancer is selected from the group consisting of ACTB; PPIB; FIPRT1 ; and TBP.

It will be appreciated that a test which requires the determination of fewer RNA expression levels is preferable over a test that requires the determination of more RNA expression levels, provided that the test meets the required level of accuracy. In one embodiment it is possible to classify the cancer into a sub-class of cancer by determining the expression level of just one gene, for example of the FIER2 (ERBB2) gene or the ER (ESR1 ) gene which can be used to classify breast cancer as FIER2+ or ER+ using the method of the invention. In other cases, to accurately distinguish between two sub-classes of cancer, for example Luminal A and Luminal B breast cancer, multiple markers are required, for example when assessed using the present invention, the expression levels of the ERBB2, ESR1 , PGR and FOXC1 genes can be used. Flowever, a greater accuracy of classification can be achieved by determining the expression levels of all of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes.

The skilled person will know what is meant by the genes ERBB2, ESR1 , FOXC1 , PGR, AURKA, KIF2C, ACTB, PPIB, HPRT1 and TBP. For the avoidance of doubt, the table below provides sequence information for the relevant genes and proteins. However, it will be understood that proteins or genes with minor variations in the sequences are also encompassed by the terms ERBB2, ESR1 , FOXC1 , PGR, AURKA, KIF2C, ACTB, PPIB, HPRT1 and TBP. For example, the genes or proteins may contain polymorphisms that are not represented in the sequences provided below. Accordingly, by the terms ERBB2, ESR1 , FOXC1 , PGR, AURKA, KIF2C, ACTB, PPIB, HPRT1 and TBP we include the meaning of the relevant gene or protein sequences as provided below, as well as genes or proteins with at least 75% homology to those sequences, for example at least 80% homology to those sequences, for example at least 85% homology to those sequences, for example at least 90% homology to those sequences, for example at least 95% homology to those sequences, for example at least 98% homology to those sequences, for example at least 99% homology to those sequences, for example 100% homology to those sequences.

Table 2 - protein and nucleic acid sequences of the exemplified genes associated with the cancer or at least one sub-class of cancer and normalisation genes

An accuracy of 99.5% is required in order for a test to have clinical significance and to be adopted in practice. By“accuracy” we include the meaning of concordance with samples of known sub-class as determined by a different method. For example, as described herein samples are assessed with the method of the invention and compared to samples of known sub-class, the status of which has in some embodiments been determined previously by IHC or FISH. An accuracy or concordance of 100% means that all samples that were previously identified as of a particular sub-class are also identified by the methods described herein as being of that same sub-class. In one embodiment accuracy is determined using well defined samples such as cell lines and patient samples that are classified positive or negative to the measured analyte, example HER2 IHC scores of 3+ and 0 or 1 + on FFPE whole sections from tumour resections showing no heterogeneity (only one tumour site). Discrepancies may occur wherein, for example, the status of a sample has been previously characterised using FISH, i.e. determining gene copy number, and has been determined as comprising a gene amplification, for example of the ERBB2 gene. If that amplification does not result in the expected increase in RNA level, then the two results will not correlate. However it is considered that the methods of the present invention provide a more accurate means of classification than at least FISH, and also IHC due to the inherent variability in a subjective test.

In one embodiment the cancer is breast cancer. In a further embodiment the sub-classes of cancer comprise at least one or more of HER2+, ER+, Basal, Triple Negative Breast Cancer (TNBC), and Luminal A, Luminal B and HER2 enriched. In another embodiment the sub-classes are TNBC, ER+ and HER2+. In yet another embodiment the sub-classes are Basal, Luminal and HER2-enriched. The Luminal sub-class can be further classified as Luminal A or Luminal B.

Accordingly, in one embodiment the method can distinguish, or classify, a breast cancer as HER2+ or not HER2+. In another embodiment the method can distinguish, or classify, a breast cancer as ER+ or not ER+. In another embodiment the method can distinguish or classify a breast cancer as TNBC or not TNBC. In yet another embodiment the method can distinguish or classify a breast cancer as Basal or not Basal. In another embodiment the method can distinguish or classify a breast cancer as Luminal or not Luminal. In another embodiment the method can distinguish or classify a breast cancer as Luminal A or not Luminal A. In another embodiment the method can distinguish or classify a breast cancer as Luminal B or not Luminal B. In another embodiment the method can distinguish or classify a breast cancer as HER2-enriched or not Her2-enriched.

In particular embodiments the method can distinguish or classify a breast cancer as TNBC, ER+ and/or HER2+ or not TNBC, ER+ and/or HER2+. In another embodiment the method can distinguish or classify a breast cancer as Basal, Luminal and HER2-enriched or not Basal, Luminal and HER2-enriched. The Luminal sub-class can be further classified as Luminal A or Luminal B.

In yet another embodiment the method can distinguish or classify a breast cancer as HER2+; ER+, HER2+ and ER+; or TNBC. In yet another embodiment the method can classify a breast cancer as Luminal or Basal breast cancer. Further, in another embodiment, the method can classify a breast cancer as Luminal A or Luminal B breast cancer. In a preferred embodiment the method can classify a breast cancer as HER2+; ER+; HER2+ and ER+; TNBC; Basal; Luminal A; and/ or Luminal B.

In one embodiment, the method comprises the determination of the level of RNA expression of ERBB2.

In one embodiment, the method comprises the determination of the level of RNA expression of ESR1 .

In one embodiment, the method comprises the determination of the level of RNA expression of AURKA.

In one embodiment, the method comprises the determination of the level of RNA expression of KIF2C.

In one embodiment, the method comprises the determination of the level of RNA expression of PGR.

In one embodiment, the method comprises the determination of the level of RNA expression of FOXC1.

As discussed above, any determination of the expression level of a gene associated with a cancer or sub-class of a cancer requires normalisation against appropriately selected normalising genes. The inventors have shown that normalisation of expression data to at least any one of ACTB, HPRT1 , TBP and PPIB; preferably all of ACTB, HPRT1 , TBP and PPIB allows a highly accurate classification of breast cancers. Accordingly, in any embodiment of the invention, the expression levels as determined by the bDNA are normalised to at least any one of ACTB, HPRT1 , TBP and PPIB; preferably all of ACTB, HPRT1 , TBP and PPIB.

Accordingly, in one embodiment the method comprises the determination of the level of RNA expression of any one or more of ACTB, PPIB, HPRT1 and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ERBB2, ACTB, PPIB, HPRT1 and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 , ACTB, PPIB, HPRT1 and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 , ERBB2, ACTB, PPIB, HPRT1 and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ERBB2; AURKA and/or KIF2C; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 ; AURKA and/or KIF2C; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ERBB2; ESR1 ; AURKA and/or KIF2C; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ERBB2; ESR1 ; AURKA; KIF2C; PGR; FOXC1 ; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ERBB2; ESR1 ; PGR; FOXC1 ; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of AURKA; KIF2C; PGR; FOXC1 ; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of PGR; FOXC1 ; ACTB; PPIB; HPRT1 ; and TBP. In one embodiment the method comprises the determination of the level of RNA expression of PGR; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 ; PGR; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of AURKA; KIF2C; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 ; ERBB2; PGR; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 ; ERBB2; AURKA; ACTB; PPIB; HPRT1 ; and TBP.

In one embodiment the method comprises the determination of the level of RNA expression of ESR1 ; ERBB2; KIF2C; ACTB; PPIB; HPRT1 ; and TBP.

It will be appreciated that whilst it is possible for the skilled person to compare the expression levels of various genes in cancers from different sub-classes to identify which genes are expressed to which levels in each sub-class, the use of machine learning greatly enhances both the speed and accuracy of such determinations.

Accordingly in one embodiment the method further comprises the use of machine learning to classify the cancer into the sub-classes. The skilled person will be aware of suitable programmes and algorithms that can analyse the data, for example in one embodiment the machine learning is an algorithm, optionally is an algorithm selected from the group consisting of the Neural Network algorithm, Decision Tree, Random Forest and Support Vector Machine. In a preferred embodiment the machine learning is the Neural Network algorithm. Accordingly in one embodiment the method comprises the use of the Neural Network algorithm, Decision Tree, Random Forest and/or Support Vector Machine. In a preferred embodiment the method comprises the use of the Neural Network algorithm. In another preferred embodiment the method comprises the use of RapidMiner Studio Community software.

Such algorithms require training on a defined data set so that the algorithm can learn which features are associated with each other, for example which level of HER2 expression is associated with a declaration of HER2+ status; which level of ER expression is associated with a declaration of ER+ status; which expression levels of certain genes are associated with the Luminal A and Luminal B sub-classes etc. Accordingly in one embodiment the algorithm, for example the Neural Network, has been trained on a suitable data set.

Suitable datasets are considered to include at least large databases of cancers samples, for example those found within the The Cancer Genome Atlas and Oncomine portals. An example of just one of the datasets that are available through such portals is the Breast Invasive Carcinoma (TCGA Cell, 2015) N=818 dataset. For example, such databases have details or the classification of samples, for examples samples may have been classed as various sub-classes of cancer based on FISH or IHC data. In a preferred embodiment the dataset contains details of samples that have been classified using a clinically relevant method, for example IHC and FISH, preferably without equivocal or heterogenous cases. The algorithm may be trained on all samples that samples that have been classified using a clinically relevant method, for example IHC and FISH, preferably without equivocal or heterogenous cases, or it may be necessary to remove certain samples prior to analysis, for example it may be necessary to first identify samples of a certain sub-class, for example HER2+ samples, and remove them from the dataset prior to training the algorithm for, for example Luminal vs Basal classification.

In one embodiment the algorithm used in the present invention is any one or more of:

Algorithm 1 - trained on normalised expression of ERBB2 in:

a sample set comprising HER2+ samples as defined by IHC and FISH;

and a sample set not comprising HER2+ samples as defined by IHC and FISH;

Algorithm 2 - trained on normalised expression of ESR1 in:

a sample set comprising ER+ samples as defined by IHC;

and a sample set not comprising ER+ samples as defined by IHC;

Algorithm 3 - trained on normalised expression of ERBB2 and ESR1 in:

a sample set comprising HER2+ and/or ER+ samples as defined by IHC and FISH; and a sample set not comprising HER2+ and ER+ samples as defined by IHC and FISH;

Algorithm 4 - trained on normalised expression of ESR1 , ERBB2, AURKA and/or KIF2C in: a sample set comprising HER2+ and/or ER+ samples as defined by IHC and FISH; and a sample set not comprising HER2+ and/or ER+ samples as defined by IHC and FISH;

Algorithm 5 - trained on normalised expression of ERBB2 samples in a dataset from The Cancer Genome Atlas (TGCA) comprising

HER2+ samples as defined by IHC and FISH; and

samples defined as not HER2+ by IHC and FISH;

Algorithm 6 - trained on normalised expression of ERBB2 and ESR1 in samples in a dataset from the TCGA, comprising

HER2+ and/or ER+ samples as defined by IHC and FISH; and

Samples defined as not HER2+ and not ER+ by IHC and FISH;

Algorithm 7 - trained on normalised expression of ERBB2 and ESR1 in samples in a dataset from the TCGA which had HER2-enriched samples as defined by PAM50 and HER2+ as defined by IHC/FISH removed.

Algorithm 8 - trained on normalised expression of ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 in samples in a dataset from the TCGA which had HER2-enriched samples as defined by PAM50 and HER2+ samples as defined by IHC/FISH removed.

Algorithm 9 - trained on normalised expression of ERBB2, ESR1 , FOXC1 , PGR normalised expression in:

samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8;

Algorithm 10 - trained on normalised expression of ERBB2, ESR1 , FOXC1 , PGR, AURKA and KIF2C in:

samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8; Algorithm 1 1 - trained on normalised expression of AURKA, KIF2C, FOXC1 and PGR in: samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8;

Algorithm 12 - Trained on normalised expression of FOXC1 and PGR in:

samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8; optionally wherein all expression levels were normalised to the expression levels of ACTB; PPIB; HPRT1 ; and/or TBP, preferably wherein all expression levels were normalised to the expression levels of ACTB; PPIB; HPRT1 ; and TBP.

Algorithm 1 identifies samples as HER2 positive or HER2 Negative.

Algorithm 2 identifies samples as ER positive or ER Negative.

Algorithm 3 identifies samples as HER2 positive and/or ER positive or TNBC.

Algorithm 4 identifies samples as HER2 positive and/or ER positive or TNBC.

Algorithm 5 identifies samples as HER2 positive or HER2 Negative.

Algorithm 6 identifies samples as HER2 positive and/or ER positive or TNBC.

Algorithm 7 identifies samples as Luminal or Basal.

Algorithm 8 identifies samples as Luminal or Basal.

Algorithm 9 identifies samples as Luminal A or Luminal B.

Algorithm 10 identifies samples as Luminal A or Luminal B.

Algorithm 1 1 identifies samples as Luminal A or Luminal B.

Algorithm 12 identifies samples as Luminal A or Luminal B.

In particular embodiments the method classifies the cancer as a HER2+ cancer or a HER2 negative cancer, optionally wherein the method comprises the determination of the level of expression of the ERBB2 gene, optionally further comprises normalisation of the level of expression of the ERBB2 gene to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of any one or more of Algorithm 1 , Algorithm 3, Algorithm 4 and Algorithm 5 as defined above.

In another embodiment, the method classifies the cancer as an ER+ cancer or an ER negative cancer, optionally wherein the method comprises the determination of the level of expression of the ESR1 gene, optionally further comprises normalisation of the level of expression of the ESR1 gene to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of any one or more of Algorithm 2, Algorithm 3, Algorithm 4 and Algorithm 6 as defined above.

In a further embodiment the method classifies the cancer as a HER2+, ER+ or TNBC, optionally wherein the method comprises the determination of the level of expression of the ERBB2, ESR1 , AURKA and/or KIF2C genes, optionally further comprises normalisation of the level of expression of the ERBB2, ESR1 , AURKA and/or KIF2C genes to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of any one or more of Algorithm 3, Algorithm 4 and Algorithm 6 as defined above.

In yet another embodiment the method classifies the cancer as a Basal cancer or a Luminal cancer, optionally wherein the method comprises the determination of the level of expression of the ERBB2 and ESR1 genes, optionally further comprises normalisation of the level of expression of the ERBB2 and ESR1 genes to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of Algorithm 7 as defined above.

In another embodiment the method classifies the cancer as a Basal cancer or a Luminal cancer, optionally wherein the method comprises the determination of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes, optionally further comprises normalisation of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of Algorithm 8 as defined above.

In a further embodiment the method classifies the cancer as a Luminal A or a Luminal B cancer, optionally wherein the method comprises the determination of the level of expression of the ERBB2, ESR1 , PGR and FOXC1 genes, optionally further comprises normalisation of the level of expression of the ERBB2, ESR1 , PGR and FOXC1 genes, to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of Algorithm 9 as defined above.

In yet another embodiment the method classifies the cancer as a Luminal A or a Luminal B cancer, optionally wherein the method comprises the determination of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes, optionally further comprises normalisation of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes, to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of Algorithm 10 as defined above.

In another embodiment the method classifies the cancer as a Luminal A or a Luminal B cancer, optionally wherein the method comprises the determination of the level of expression of the PGR, AURKA, KIF2C and FOXC1 genes, optionally further comprises normalisation of the level of expression of the PGR, AURKA, KIF2C and FOXC1 genes, to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of Algorithm 1 1 as defined above.

In another embodiment the method classifies the cancer as a Luminal A or a Luminal B cancer, optionally wherein the method comprises the determination of the level of expression of the PGR and FOXC1 genes, optionally further comprises normalisation of the level of expression of the PGR and FOXC1 genes, to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of Algorithm 12 as defined above.

In a further embodiment the method classifies the cancer as HER2+, ER+, TNBC, Luminal, Basal, Luminal A and Luminal B by a) Identifying the HER2+ class as defined diagnostically by IHC (accuracy compared to IHC/FISH positive cases); b) Define between Luminal and Basal (concordance with PAM50); c) Luminal subclassification into Luminal A or Luminal B (prognostic significance); wherein part a) provides information to further classify the clinically actionable classes: HER+, ER+ and TNBC.

The skilled person will be aware of the algorithm parameters that must be defined using appropriate training sets. An example of suitable settings for the Neural Network Algorithm computed using RapidMiner Studio Community software is as follows: 500 training cycles

0.3 learning rate

0.2 momentum

Shuffle

Normalise

Error Epsilon: 1 .0E-5.

See for example Figure 15.

Accordingly, through determination of the RNA expression level of the genes described herein, along with normalisation to the specific set of normalising genes as described herein, and the use of a trained algorithm as described herein, the method of the invention is able to classify a breast cancer as either HER2+, ER+, Basal, Luminal A, Luminal B or TNBC, optionally wherein the method comprises the determination of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KI F2C and FOXC1 genes, optionally further comprises normalisation of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes, to the expression levels of ACTB; PPIB; HPRT1 ; and TBP, optionally further comprise the use of the Algorithms as defined above.

Suitable combinations of algorithms are as follows: a) Algorithm 1 and Algorithm 2 together; or Algorithm3; or Algorithm 4; or Algorithm 6; to classify the ER+, HER2+ and TNBC diagnostic classification. A preferred algorithm is Algorithm 4 (see Table 4). b) the HER2+ diagnostic class (by IHC/FISH) can be defined by Algorithm 1 or 5 A preferred algorithm is Algorithm 1 . c) the ER+ diagnostic class (by IHC) can be defined by Algorithm 2. d) The ER+ and TNBC case set (for example as defined by any of algorithms 1 and 2; 3 or 4) or the HER2 negative case set (for example as defined by any of algorithms 1 or 5) can be re-classified into the molecular classes Luminal and Basal using Algorithms 7 or 8. A preferred algorithm is Algorithm 8 (see Table 10). e) The Luminal class (as defined in (c) above) can be re-classified into Luminal A and Luminal B using any of Algorithms 9-12. A preferred algorithm is Algorithm 10 (see Table 12). A preferred sequential application of algorithms to expression data obtained from a single sample is as follows (summarised in Figure 4):

1 . Algorithm 4 - Identify HER2+ (Table 4)

2. Algorithm 8 - identify Basal (Table 10)

3. Algorithm 10 - identify Luminal A and Luminal B (Table 12)

The method of the invention is considered to be an in vitro method, wherein the method is performed on a sample obtained from a subject. However, there may be some situations whereby the method is suitable for in vivo use and in vivo detection of the cancer and/or sub-class of cancer, for example where the means required for performing the method are incorporated into an implantable sensor. Such in vivo use is also encompassed by the present invention and the method of the invention is in one embodiment an in vivo method. However in a preferred embodiment the method of the invention in an in vitro method.

It is considered that any sample type in which RNA (intact or degraded) is present, will work with the present invention. Data provided herein demonstrate that the method of the invention can classify breast cancer with a high degree of accuracy when the sample is a Hematoxylin and Eosin stained sample and/or is a sample comprising totally degraded RNA, i.e. a sample with RNA with a RIN value of around 2.0. Data also provided herein demonstrates that the use of the method with a sample comprising exosomes is appropriate.

The sample may be from any organism, for example any mammal, for example a human, a dog, a primate, cattle or any virus and other microorganisms. Preferably the sample is from a human.

The sample may be a fresh sample, for example may be a fresh biopsy of a tumour, or may be a fresh blood or plasma sample. The sample may be a sample that comprises exosomes. The sample may also be not a fresh sample, for example may be an archived sample, for example a sample that has been stained and/or frozen and/or fixed for example fixed with formalin and/or embedded for example embedded in paraffin.

The sample may be a tissue sample obtained from a subject; a cell line; a liquid biopsy, for example a blood sample or a plasma sample. In a preferred embodiment the liquid biopsy comprises circulating tumour cells. In another embodiment the liquid biopsy comprises exosomes. In one embodiment the sample is a homogenate or lysate, for example a homogenate or lysate of a homogenous tumour. In another embodiment the sample is a homogenate or lysate of a heterogeneous tumour.

In one embodiment, the sample is an archived/historical sample, for example the sample is between 1 month-100 years old, optionally wherein the sample is at least 1 month old, optionally wherein the sample is at least 2 months old, or at least 3 months old, or at least

4 months old, or at least 5 months old, or at least 6 months old, or at least 7 months old, or at least 8 months old, or at least 9 months old, or at least 10 months old, or at least 1 1 months old, or at least 12 months old for example 1 year old, or at least 1 .5 years old, or at least 2 years old, or at least 3 years old, or at least 4 years old, or at least 5 years old, or at least 6 years old, or at least 7 years old, or at least 8 years old, or at least 9 years old, or at least 10 years old, or at least 1 1 years old, or at least 12 years old, or at least 13 years old, or at least 14 years old, or at least 15 years old, or at least 16 years old, or at least 17 years old, or at least 18 years old, or at least 19 years old, or at least 20 years old, or at least 25 years old, or at least 30 years old, or at least 40 years old, or at least 50 years old, or at least 60 years old, or at least 70 years old, or at least 80 years old, or at least 90 years old or at least 100 years old.

In another embodiment the sample is a fresh sample, for example the sample is less than 1 month old, optionally less than 4 weeks old, optionally less than 21 days old, optionally less than 14 days old, optionally less than 7 days old, optionally less than 6 days old, optionally less than 5 days old, optionally less than 4 days old, optionally less than 3 days old, optionally less than 2 days old optionally less than 1 day old, optionally less than 18 hours old, optionally less than 12 hours old, optionally less than 6 hours old, optionally less than 5 hours old, optionally less than 4 hours old, optionally less than 3 hours old, optionally less than 2 hours old, optionally less than 1 hour old, optionally less than 30 minutes old, optionally less than 15 minutes old, optionally less than 10 minutes old, optionally less than

5 minutes old.

As discussed above, the present inventors have found that the present invention, involving the use of bDNA technology, is able to accurately classify cancer samples even when the sample comprises degraded RNA. Accordingly in one embodiment the sample comprises RNA, including degraded RNA, for example wherein the sample has a RIN of 8 or less, for example less than 7.5, for example less than 7.0, for example less than 6.5, for example less than 6.0, for example less than 5.5, for example less than 5.0, for example less than 4.5, for example less than 4.0, for example less than 3.5, for example less than 3.2 , for example less than 3.0, for example less than 2.8, for example less than 2.6, for example less than 2.4, for example less than 2.2, for example less than 2.0, for example less than 1 .8, for example less than 1 .6, for example less than 1 .4, for example less than 1.2, for example 1 .0. The skilled person will be aware of the RIN parameter, for example it is discussed in https://www.aqilent.com/cs/library/applications/5989-1 165EN.pdf.

As demonstrated herein, the method of the present invention is not affected by formalin fixation of paraffin embedding. Accordingly, in one embodiment the sample has been formalin-fixed (FF). In the same or alternative embodiment the sample has been paraffin- embedded (PE). In a further embodiment the sample is a formalin-fixed paraffin- embedded sample (FFPE).

Also as demonstrated herein, the method of the present invention is not affected by staining, for example is not affected by Haematoxylin and Eosin staining. Accordingly in one embodiment the sample is a stained sample, for example is a Haematoxylin and Eosin stained sample.

It is considered useful if the sample is a selected part of the sample that is taken from the subject. As discussed above, tumours are often heterogeneous, often comprising numerous individual morphologically distinct tumours. For example, a breast cancer sample may contain several different tumours within it, for example may comprise both HER2+ and HER2 negative tumours. Whilst analysis of the entire tumour sample is considered to be useful, it is considered more useful if the individual tumours can be isolated and assessed separately. For example, analysing a whole tumour sample that actually contains both HER2+ and HER2- microtumours may give an overall HER2+ phenotype. This may direct the clinician to administer Herceptin therapy. Whilst Herceptin may be effective against the HER2+ microtumour, the HER2 negative microtumour is unlikely to be affected and may continue to grow unchecked.

Accordingly, an advantage of the fact that the present invention is not affected by staining is that it allows the staining of a sample, for example with Haematoxylin and Eosin, and the dissection of the individual morphologically distinct tumours followed by direct determination of RNA expression level, according to the present invention, classification of each individual microtumour that was isolated and analysed. The skilled person will be well aware of suitable techniques for staining the samples and subsequent identification and isolation of morphologically distinct tumours. For example, in one embodiment the sample is a laser dissected sample or a macro dissected sample.

In another embodiment the sample is a sample of a morphologically distinct tumour within a larger tumour.

It is clear then that the present invention has the advantage of being able to detect morphologically distinct tumours of different sub-classes.

Accordingly, the invention further provides a method for the detection of heterogeneity in a tumour. Preferences for features discussed in relation to the first aspect of the invention also apply here. For example in one embodiment the method for the detection of heterogeneity in a tumour involves the detection of the RNA expression level of particular genes, for instance the ERBB2 and ESR1 gene, via bDNA technology, for example via the QuantiGene Plex Assay, wherein the expression levels have been normalised to ACTB, PPIB, HPRT1 , and TBP.

In one embodiment the method comprises:

the identification of morphologically distinct tumours identified in a sample obtained from a patient;

dissection, optionally laser dissection or macro dissection of the morphologically distinct tumours into two or more samples; and

performing a method according to the first aspect of the invention separately on each sample.

In other embodiments, the method comprises performing a method according to the first aspect of the invention separately on more than one sample obtained from the tumour. For example, it may be preferable for samples to be taken from several different sites in the tumour. The advantages of the present invention which include a rapid and accurate assessment of cancer type or cancer sub-class permits the routine classification of multiple samples from the same tumour. The skilled person can then readily determine if a subject has a tumour that has heterogeneity, and what that heterogeneity is, for example if it is a tumour that has FIER2+ and FIER2 negative microtumours, or both Luminal and Basal microtumours. It is considered that such an approach would greatly inform the clinical practitioner’s therapeutic strategy. In addition, the present invention also provides the use of bDNA technology to predict the presence of a gene amplification.

It will be appreciated that the various methods described herein are amenable to being packaged as a kit or part of a kit. The invention therefore also provides kits that are suitable for use in the methods of the present invention.

In one embodiment the invention provides a kit for use in one or more methods of the invention, said kit comprising bDNA probes directed towards at least two of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 , ACTB, PPIB, HPRT1 , and TBP.

For example the kit may contain bDNA probes directed towards ERBB2 and ACTB, PPIB, HPRT1 , and TBP; or ESR1 and ACTB, PPIB, HPRT1 , and TBP; or ERBB2, ESR1 and ACTB, PPIB, HPRT1 , and TBP; or all of ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 , ACTB, PPIB, HPRT1 , and TBP.

The kit may additionally or alternatively contain bDNA probes directed towards any two or more of ACTB, PPIB, HPRT1 , and TBP; any three or more of ACTB, PPIB, HPRT1 , and TBP; or all four of ACTB, PPIB, HPRT1 , and TBP.

The kit may additionally or alternatively contain bDNA probes directed towards:

Any one or more of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 ;

Any two or more of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 ;

Any three or more of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 ;

Any four or more of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 ;

Any five or more of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 ; or

All six of: ERBB2, ESR1 , PGR, AURKA, KIF2C, FOXC1 .

The kit may additionally or alternatively contain bDNA probes to AURKA and/or KIF2C.

The kit may also additionally or alternatively contain means for staining the sample to identify morphologically distinct tumours within the sample, for example may contain Haematoxylin & Eosin.

Any part of the kits described above are also for use in the diagnosis or prognosis of cancer. The present invention also provides the use of any or part of the kits as described above in a method of manufacture of a composition for use in diagnosing or prognosing cancer.

As discussed above, the present methods are suitable for use on archived samples which means that a wealth of data can be mined using the present invention. The invention also therefore provides a method of validating a potential gene amplification or gene reduction as a biomarker, the method comprising the use of any one or more of the methods of the invention.

It will be apparent to the skilled person that the present invention can be used in a method of diagnosis or prognosis. For examples the methods described herein can be used to diagnose cancer, for example to diagnose breast cancer or can be used to diagnose a particular sub-class of cancer, for example to diagnose a subject has having a HER2+ breast cancer and/or an ER+ breast cancer and/or a TNBC and/or a Basal breast cancer and/or a Luminal breast cancer, for example a Luminal A or a Luminal B breast cancer.

Similarly, the methods described herein can be used in a method for aiding in determining the prognosis of a subject, for example where a subject has been diagnosed with breast cancer, the methods of the invention can be used to aid in determining the prognosis of the subject by classifying the sub-type or sub-types of cancer that the subject has. The skilled person will be aware that a diagnosis of Basal breast cancer is associated with a worse prognosis than HER2-enriched, for example, and that HER2-enriched is associated with a worse prognosis than Luminal B, for example, and that Luminal B is associated with a worse prognosis than Luminal A.

The skilled person will also understand that TNBC is associated with a worse prognosis than HER2+, and that HER2+ is associated with a worse prognosis than ER+.

It will be appreciated, particularly in view of the ability to use the present invention on archived samples that the present invention is suitable for research use. Accordingly in one embodiment the methods of the present invention are for research use.

It will be evident from the discussions above that knowledge of the cancer sub-class or sub-classes that a subject has can inform treatment strategy. Accordingly, the present invention provides a method of selecting a suitable treatment strategy wherein the method comprises any of the methods of the invention as described above. For instance, where the method is used to classify breast cancers, a diagnosis of a HER2+ breast cancer can be used to select the subject for treatment with Herceptin (trastuzumab), Kadcyla (Herceptin and emtansine), Nerlynx (neratinib), Perieta (pertuzumab), and/or Tykerb (lapatinib); a diagnosis of ER+ breast cancer can be used to select the subject for treatment with Tamoxifen, Aromatase Inhibitors, and/or SERMs; and a diagnosis of Basal breast cancer can be used to select the subject for treatment with other chemotherapeutic agents.

Accordingly, the invention also provides Herceptin (trastuzumab), Kadcyla (Herceptin and emtansine), Nerlynx (neratinib), Perieta (pertuzumab), and/or Tykerb (lapatinib) for use in the treatment of a subject with breast cancer wherein a sample from the subject has been identified as HER2+ by use of any of the preceding methods. Similarly the invention provides a method for treating HER2+ cancer wherein the subject has been diagnosed as having a HER2+ cancer or microtumour, wherein said method comprises administration of any one or more of Herceptin (trastuzumab), Kadcyla (Herceptin and emtansine), Nerlynx (neratinib), Perieta (pertuzumab), and/or Tykerb (lapatinib) to said subject.

Accordingly, the invention also provides Tamoxifen, Aromatase Inhibitors, and/or SERMs for use in the treatment of a subject with breast cancer wherein a sample from the subject has been identified as ER+ by use of any of the preceding methods.

The invention also provides various algorithms that have been trained on different datasets. Preferences for the algorithms as given above, for instance a preferred algorithm is the Neural Network algorithm. Accordingly the invention provides any one or more of Algorithms 1 -12 as defined above. In one embodiment the algorithm is the Neural Network algorithm.

Since it has been previously shown that an overexpression of AURKA and/or KIF2C makes cancers susceptible to treatment with a PP2A activator, the invention also provides a method of treating a patient having a cancer comprising an overexpression of AURKA and/or KIF2C, comprising administering to the patient a PP2A activator and thereby treating the cancer, wherein the overexpression of AURKA and/or KIF2C has been determined by use of any of the methods disclosed herein. The invention further provides a method of treating cancer in a patient, the method comprising administering an AURKA antagonist and/or a KIF2C antagonist to the patient and thereby treating the cancer wherein the patient has a cancer with an overexpression of AURKA and/or KIF2C, wherein the overexpression of AURKA and/or KIF2C has been determined by use of any of the methods disclosed herein. The invention additionally provides a method of treating cancer in a patient, or predicting sensitivity of a cancer to PP2A activators, the method comprising (a) measuring the amount of AURKA and/or KIF2C RNA in the cancer according to any of the methods described herein and optionally (b) if the cancer comprises an overexpression of AURKA and/or KIF2C, administering to the patient a PP2A activator and thereby treating the cancer.

The invention also provides a PP2A activator for use in treating cancer in a patient, wherein the cancer comprises an overexpression of AURKA and/or KIF2C, wherein the overexpression of AURKA and/or KIF2C has been determined by use of any of the methods disclosed herein, and use of a PP2A activator in the manufacture of a medicament for treating cancer in a patient, wherein the cancer comprises an overexpression of AURKA and/or KIF2C wherein the overexpression of AURKA and/or KIF2C has been determined by use of any of the methods disclosed herein. The invention further provides an AURKA antagonist and/or a KIF2C antagonist for use in treating cancer in a patient, and use of an AURKA antagonist and/or a KIF2C antagonist in the manufacture of a medicament for treating cancer in a patient wherein the cancer has an overexpression of AURKA and/or KIF2C, wherein the overexpression of AURKA and/or KIF2C has been determined by use of any of the methods disclosed herein.

The invention additionally provides a method for detecting cancer in a patient, the method comprising measuring expression of AURKA and KIF2C in the patient in accordance with any of the methods described herein, wherein overexpression of AURKA and KIF2C indicates that the patient comprises a cancer. The invention further provides a method for prognosing a cancer in a patient, the method comprising determining whether or not the cancer comprises an overexpression of AURKA and/or KIF2C in accordance with any of the methods described herein, wherein an overexpression of AURKA and/or KIF2C in the cancer indicates that the patient has a worse prognosis than in the situation of normal expression of AURKA and/or KIF2C. The invention also provides a method for determining whether or not a patient having or suspected of having or being at risk of developing cancer will respond to treatment with a PP2A activator, which method comprises measuring expression of AURKA and/or KIF2C in the individual in accordance with any of the methods described herein, and thereby predicting whether or not the patient will respond to treatment with a PP2A activator.

The invention further provides a method for classifying a cancer in a patient, the method comprising measuring expression of AURKA and/or KIF2C in the patient in accordance with any of the methods described herein, and classifying the cancer as of a particular subtype based on the expression. The invention also provides a kit for detecting a cancer comprising a deregulation of PP2A, comprising reagents suitable for detecting expression of AURKA and/or KIF2C. The invention also provides a system for detecting, classifying or prognosing cancer in a patient, or for predicting responsiveness of a cancer patient to treatment with a PP2A activator, the system comprising: (a) a measuring module for determining expression of AURKA and/or KIF2C in the patient, (b) a storage module configured to store control data and output data from the measuring module, (c) a computation module configured to provide a comparison between the value of the output data from the measuring module and the control data; and (d) an output module configured to display whether or not the patient has cancer based on the comparison, wherein an overexpression of AURKA and/or KIF2C in the patient indicates the presence of cancer, classifies the cancer, indicates a worse prognosis of cancer, or predicts that the patient will respond to treatment with a PP2A activator.

Figure Legends

Figure 1

A boxplot of ERBB2 mRNA transcript expression across cases grouped by the ERBB2 putative gene copy number. N = 960 where 147 (15.31 %) cases show ERBB2 transcript overexpression while only 1 19 (12.4%) cases show ERBB2 gene amplification. Flence 4.17% of cases show increased FIER2 expression but no gene amplification (false negative cases by genetic studies) and 1.25% of cases are ERBB2 gene amplified but do not show overexpression (false positive cases by genetic studies). Data retrieved from the cBIOportal from the TCGA Breast Cancer dataset (TCGA, Provisional) on 13 ^th June, 2017.

Figure 2

ERBB2 raw expression data against HER2 positive total RNA input derived from BT474 cell line. [A] LX200 - shows saturation at approximately 100ng of input RNA [B] Magpix - shows signal saturation at approximately 180ng of RNA.

Figure 3

Quantigene results of Breast Cancer receptor status and normalising genes across a range of RNA concentration. The x-axis represents the concentration of RNA (pg/ml) and the y-axis represent the Mean Fluorescence Intensity (MFI) measured using the Magpix Luminex instrument. Gene expression was measured using the breast cancer cell line BT474, positive for both oestrogen receptor and HER2 (ERBB2). [A] Shows the full range of MFIs measures; [B] zooms into the low MFIs to show the oestrogen receptor expression and normalising genes.

Figure 4

Algorithm flow chart from sample processing down to classification. [C] Shows the preferred algorithm 4 for defining HER2+ cases. [D] Includes the preferred algorithm 8 to define Basal and Luminal cases. [E] subclassifies Luminal cases into Luminal A and Luminal B using the preferred algorithm 10.

Figure 5

Distribution of expression levels of the TBP normalising gene of 98 breast cancer samples. [A] Raw MFI data of TBP [B] TBP MFI data expressed as Log10 and [C] TBP MFI Log10 data expressed as z-scores (MFI = Median Fluorescence Intensity)

Figure 6

Kaplan-Meier overall survival curve derived from the cBIOportal for the TCGA Breast cancer cohort for Luminal A and Luminal B as annotated with PAM50. Log Rank (Mantel- Cox) x ² = 6.310, p-value = 0.012

Figure 7

Kaplan-Meier overall survival curve derived from the cBIOportal for the TCGA Breast cancer cohort for Luminal A and Luminal B as predicted by the Neural Network based on the expression data of ESR1 , ERBB2, PGR and FOXC1 (Algorithm 9). Log Rank (Mantel- Cox) x ² = 1.394, p-value = 0.238

Figure 8

Kaplan-Meier overall survival curve derived from the cBIOportal for the TCGA Breast cancer cohort for Luminal A and Luminal B as predicted by the Neural Network based on the expression data of AURKA, KIF2C, ESR1 , ERBB2, PGR and FOXC1 (Algorithm 10). Log Rank (Mantel-Cox) c ² = 3.953, p-value = 0.047

Figure 9

Kaplan-Meier overall survival curve derived from the cBIOportal for the TCGA Breast cancer cohort for Luminal A and Luminal B as predicted by the Neural Network based on the expression data of PGR, FOXC1 , AURKA and KIF2C (Algorithm 1 1 ). Log Rank (Mantel-Cox) c ² = 3.906, p-value = 0.048 Figure 10

Kaplan-Meier overall survival curve derived from the cBIOportal for the TCGA Breast cancer cohort for Luminal A and Luminal B as predicted by the Neural Network based on the expression data of PGR and FOXC1 (Algorithm 12). Log Rank (Mantel-Cox) c ² = 0.95, p-value = 0.330.

Figure 1 1

Prediction of breast cancer molecular classification using known and novel luminal / basal classifiers and HER2 expression. The Principal component analysis (PCA) plot shows prediction of 12-gene annotated Basal (red), Luminal (blue) and HER2-enriched (green) samples using Quantigene 6-plex expression data from FFPE patient material. The table shows a 100% concordance of the HER2 enriched with IHC and FISH results, basal subtype representing the Triple negative breast cancer (TNBC) patients and a high representation of ER positivity in the Luminal group as expected.

Figure 12

RNA Integrity numbers and concentration for a ladder, MDAMB453 RNA and degraded RNA controls, a BT474 RNA degradation array and 3 RNA samples extracted from patient derived breast cancer from FFPE blocks.

Figure 13

Raw expression profiles across degraded BT474 RNA samples for [A] ERBB2 expression as MFI and [B] normalising gene expression as MFI.

Figure 14

Normalised ERBB2 expression profile across degraded BT474 RNA samples for Relative expression between different genes in the RNA profile is maintained across the RNA degradation gradient.

Figure 15

Structure of data processing module. Expression Data from FFPE patient data [Localdata] is filtered for outliers using the Filter Examples operator. ERBB2 expression data and HER2 diagnostic (IHC/FISH) status is retrieved by the Select Attributes operator. The Set Role operator defines the HER2 diagnostic status as the variable to predict. The streamlined data is used to train the Neural Network operator which is applied using the Apply model operator to the BT474 RNA degradation series data that is processed in an identical fashion. The output from the Apply model operator provides a label for the unlabelled data along with a confidence value for the prediction.

Figure 16

Correlation between an expression profile derived from an unstained tumour section as opposed to a stained tumour section. [Pearson Correlation p-value = 5.34E-26]

Figure 17

Laser Microdissection of FFPE tissues. [A]: H&E stained slide; [B]: Immunohistochemical staining for ER expression; [C]: HER2 immunohistochemical staining; [D]: Unstained 20pm section on laser microdissection membrane slides. The yellow arrow indicates an area of invasive tumour that is not clearly demarcated due to lack of staining. [E]: A 20pm section stained with FI&E for better delineation of areas of interest. White arrows indicate laser dissection trail while the Red arrow shows the laser focus during dissection. All illustrations were captured at 10x magnification.

Figure 18

Expression of [A]: receptor status and [B]: mesenchymal marker FN1 , in breast tumour compared to matched normal. [K-W: Kruskal Wallis Test Statistic]

Figure 19

Case Study: Tumour heterogeneity. Morphologically distinct tumours were microdissected and treated as distinct samples. [A]: shows the master scan of the FI&E section. [B] & [C]: show a 10x and 40x magnification respectively for each tumour morphology identified; [D]: Immunohistochemical staining for ER at 10x magnification; [E]: Immunohistochemical staining for Ki67 at 10x magnification showing a higher mitotic activity in tumour 1 ; [F]: Normalised expression levels for the ESR1 gene in each tumour showing relatively high and equal expression between tumours as expected from the immunohistochemical result; [G]: Normalised expression levels of FN1 , a mesenchymal marker, where increased FN1 expression is accompanied by high Ki67 staining showing a signature for a more invasive tumour. The inverse is observed in tumour 2 which appears to be a slower proliferating tumour with a lower malignant potential represented by reduced FN1 expression.

Figure 20

Table indicating sources of material and equipment used in the Examples. Figure 21

MTT assays for cell viability of various breast cancer cell lines at increasing concentrations of FTY720 ranging from 0 (untreated) to 5mM. Cells were left to adhere for 24 hours following seeding, then treated for 48 hours in duplicate experiments. The cell viability of the cell lines at each dose is expressed as a percentage of the untreated cells. Error bars are expressed as percentages of the untreated cell viability and represent the standard deviation from triplicate assay values for biological replicates. [A] Sensitive basal and triple negative breast cancer (TNBC) cell lines; [B] Basal and triple negative cell lines that are not sensitive to FTY720; [C] Normal-like cell lines: HB-2 and MCF10A; [D] MCF-7, MDAMB453 Luminal, ER positive cell lines; and [E] Luminal, HER2 positive SKBR-3 and BT-474.

Figure 22

Distribution of gene expression levels derived from Quantigene assay intensity data, normalised to housekeeping genes, comparing sensitive versus non-sensitive cell lines. RNA expression was measured by Quantigene and normalised to housekeeping gene expression of each cell line respectively. [A] Expression levels of PP2A complex subunits: PPP2CA (white), PPP2R2A (hatched) and the inhibitory subunit, CIP2A (black). The adjacent table shows the statistical significance of the difference in the median expression between the two categories of cell lines that are sensitive or not sensitive to FTY720 [ ^* shows significant P-values to 95% confidence interval]. [B] expression levels of AURKA (black) and KIF2C (white). The adjacent table shows a statistically significant difference in expression between sensitive and non-sensitive cell lines for AURKA and KIF2C gene expression. Statistical analysis was done using the Mann-Whitney U test and a p-value smaller than 0.05 was considered significant. The expression of CIP2A, AURKA and KIF2C are significantly higher in the FTY720-senstive cell lines (p values of <0.05, <0.02, <0.001 respectively). Sensitive cell lines thus include the TNBC cell lines MDAMB231 , BT- 20 and Hs578T, while non-sensitive cell lines include HCC1937, MDAMB436, MDAMB468, MDAMB453, BT-474, MCT-7 and SKBR-3.

Figure 23

AURKA (white) and KIF2C (black) expression levels derived from RNASeq dataset downloaded from TCGA data portal (n=72), [A] Distribution of AURKA and KIF2C expression across breast cancer subtypes as defined by the PAM50 annotation, compared to normal tissues. [Normal N = 82; Basal N = 97; HER2-Enriched N = 82; Luminal B N = 129; Luminal A N = 228] [B] Comparison of expression in patient tumour tissue and matched normal tissue. The statistical significance of the differential expression between tumours and matched normal tissue is shown in the adjacent table using the Related samples Wilcoxon signed rank test. The tumour versus normal column shows the direction of significant difference where applicable [n = 72; RNASeqV2 normalised expression as downloaded from TCGA data portal]. [C] Comparison of expression in patient tumour tissue and matched normal tissue across different PAM50 defined patient subtypes (N = 55).

Figure 24

Percent patients with amplifications or an expression level (z-score) greater than 2 (AMP; EXP>2) for AURKA (white bars) and KIF2C (black bars) in tumours of different origin. The analysis of data was done using the data portal (cBioPortal for Cancer Genomics (Cerami et al., 2012), available at http://www.cbioportal.org. The normalised RNASeqV2 data was used from the TCGA dataportal (https://tega-data.nci.nih.gov/tcga/).

Figure 25

Expression of the selected probes (genes) in the colorectal cancer cell lines COLO320. The x-axis represents the expression in exosomes derived from COLO320 and the y-axis the expression measured using RNA from originator cells.

Figure 26

Expression of AURKA (Log-io) in cell line derived exosomes can be used to normalise and characterise HER2 expression into positive or negative.

Further embodiments of the invention provide:

1. A method of treating a patient having a cancer comprising an overexpression of AURKA and/or KIF2C, comprising administering to the patient a PP2A activator and thereby treating the cancer.

2. The method according to embodiment 1, wherein said patient has an overexpression of

AURKA and KIF2C.

3. A method of treating cancer in a patient, the method comprising administering an AURKA antagonist and/or a KIF2C antagonist to the patient and thereby treating the cancer. 4. The method according to embodiment 3, further comprising administering a PP2A activator.

5. The method according to any one of the preceding embodiments, which is for treating

breast cancer.

6. The method according to embodiment 5, which is for treating basal breast cancer.

7. The method according to any one of the preceding embodiments, wherein the cancer has

an underexpression of one or more of HER2, ER and PR.

8. A method according to embodiment 7, wherein the cancer has an underexpression of

HER2, ER and PR.

9. A method according to any one of the preceding embodiments, wherein the patient is

selected for said treatment on the basis of having a cancer comprising an overexpression ofAURKA and/or KIF2C.

10. A method according to any one of the preceding embodiments, wherein the patient is

selected for said treatment on the basis of the cancer having an underexpression of one or more of HER2, ER and PR.

11. A method according to embodiment 9 or 10 wherein said selection for treatment is carried

out on diagnosis of the cancer, during treatment of the cancer, or following resistance to a cancer therapy.

12. A method according to any one of the preceding embodiments, wherein the PP2A activator

or the AURKA antagonist and/or KIF2C antagonist is a small molecule, a protein, an antibody, a polynucleotide, an oligonucleotide, an antisense RNA, a small interfering RNA (siRNA) or a small hairpin RNA (shRNA). 13. A method according to any one of the preceding embodiments, wherein the PP2A activator

and/or the AURKA antagonist and/or KIF2C antagonist is administered in combination with another cancer therapy.

14. A method according to any one of the preceding embodiments, wherein the patient is

human.

15. A method of treating cancer in a patient, the method comprising (a) measuring the amount of AURKA and/or KIF2C in the cancer and (b) if the cancer comprises an overexpression of AURKA and/or KIF2C, administering to the patient a PP2A activator and thereby treating the cancer.

16. A PP2A activator for use in treating cancer in a patient, wherein the cancer comprises an overexpression of AURKA and/or KIF2C.

17. An AURKA antagonist and/or a KIF2C antagonist for use in treating cancer in a patient.

18. Use of a PP2A activator in the manufacture of a medicament for treating cancer in a patient, wherein the cancer comprises an overexpression of AURKA and/or KIF2C.

19. Use of an AURKA antagonist and/or a KIF2C antagonist in the manufacture of a medicament for treating cancer in a patient.

20. A method for detecting cancer in a patient, the method comprising measuring expression of AURKA and KIF2C in the patient, wherein overexpression of AURKA and KIF2C indicates that the patient comprises a cancer.

21. A method for prognosing a cancer in a patient, the method comprising determining whether or not the cancer comprises an overexpression of AURKA and/or KIF2C, wherein an overexpression of AURKA and/or KIF2C in the cancer indicates that the patient has a worse prognosis than in the situation of normal expression of AURKA and/or KIF2C. 22. A method for determining whether or not a patient having or suspected of having or being at risk of developing cancer will respond to treatment with a PP2A activator, which method comprises measuring expression of AURKA and/or KIF2C in the individual, and thereby predicting whether or not the patient will respond to treatment with a PP2A activator.

23. A method for classifying a cancer in a patient, the method comprising measuring expression of AURKA and/or KIF2C in the patient, and classifying the cancer as of a particular subtype based on the expression.

24. The method according to any one of embodiments 20 to 23, which further comprises

measuring expression of one or more of HER2, ER and PR in the patient.

25. The method according to any one of embodiments 20 to 24, which comprises determining whether there is a deregulation of PP2A in the patient.

26. The method according to any one of embodiments 20 to 25, wherein the cancer is breast cancer.

27. The method according to embodiment 23 or any of embodiments 24 to 26 as dependent on embodiment

23, comprising classifying the cancer as a luminal or basal breast cancer, wherein overexpression of AURKA and/or KIF2C classifies the cancer as a basal breast cancer, and wherein normal expression of AURKA and/or KIF2C classifies the cancer as a luminal breast cancer.

28. The method according to embodiment 27, further comprising measuring expression of one

or more of HER2, ER and PR to further classify the cancer.

29. A kit for treating cancer comprising (a) one or more reagents suitable for measuring expression of AURKA and/or KIF2C and (b) a PP2A activator.

30. A kit for detecting a cancer comprising a deregulation of PP2A, comprising reagents suitable for detecting expression of AURKA and/or KIF2C. 31. The kit according to embodiment 30, which comprises reagents suitable for detecting

expression of AURKA and KIF2C.

32. The kit according to embodiment 30 or 31 further comprising reagents suitable for detecting expression of an endogenous inhibitor of PP2A or a PP2A subunit.

33. The kit according to any one of embodiments 30-32 wherein the reagents suitable for

detecting expression are selected from nucleic acid probes or primers and antibodies.

34. A system for detecting, classifying or prognosing cancer in a patient, or for predicting responsiveness of a cancer patient to treatment with a PP2A activator, the system comprising:

(a) a measuring module for determining expression of AURKA and/or KIF2C in the patient,

(b) a storage module configured to store control data and output data from the measuring module,

(c) a computation module configured to provide a comparison between the value of

the output data from the measuring module and the control data; and

(d) an output module configured to display whether or not the patient has cancer based on the comparison,

wherein an overexpression of AURKA and/or KIF2C in the patient indicates the presence of cancer, classifies the cancer, indicates a worse prognosis of cancer, or predicts that the patient will respond to treatment with a PP2A activator.

The listing or discussion of an apparently prior-published document in this specification should not necessarily be taken as an acknowledgement that the document is part of the state of the art or is common general knowledge.

Preferences and options for a given aspect, feature or parameter of the invention should, unless the context indicates otherwise, be regarded as having been disclosed in combination with any and all preferences and options for all other aspects, features and parameters of the invention. For example, the invention provides a method for classifying a cancer as a Luminal A or a Luminal B cancer, and involves the determination of the level of expression of the ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 genes in a formalin- fixed paraffin-embedded sample (FFPE) of a morphologically distinct tumours identified in a sample obtained from a patient wherein the sample is 5 months old.

Examples

Here we describe a technology that provides advantages to measure gene amplifications using RNA-based multiplex measurements that can be adapted to various actionable gene amplifications summarised below.

Example 1 - Selection of appropriate normalisation genes

Assay Linearity - HER2

This exercise aims to identify the dynamic range for HER2 on both the LX200 and Magpix. This establishes the ideal RNA concentration for the detection of these expressions (linear phase).

Selection of normalising genes using Cell line assays

Well annotated Breast cancer cell lines were cultured and RNA isolated. RNA was quantified using Agilent Bioanalyser and diluted to prepare a range of concentrations. A multi-plex Luminex based RNA assay was used to study the expression of HER2 (ERBB2) and ESR1 (oestrogen receptor) across the various concentrations to assess sensitivity of the assay. Normalising genes were included to select the best combination to ensure proper threshold settings for the simultaneous measurement of the lowly expressing ESR1 and the amplified HER2 gene.

As indicated in Figure 2, the normalising genes ACTB and GAPDH are expressed at high levels using the cell line BT474 and at a concentration of 8 pg/ml (400ng total RNA input) the signal reaches saturation and hence can result in false negative when normalising the HER2 (ERBB2) expression. To utilise these normalising genes, the concentration should not exceed 4pg/ml (200ng input RNA). GAPDH shows a significantly lower expression in patient material with some extreme outliers, becoming among the lowest expressing normalising gene within the analysed normalising geneset. GAPDH does not follow the same trends as the other normalising genes analysed and thus was excluded from further analysis. The normalising genes ACTB, PPIB, HPRT1 and TBP were selected to normalise the MFI. This allowed us to normalise the lowly expressed oestrogen receptor gene and the highly expressed amplified HER2 in one run. Example 2 - Algorithm description

See Figure 2.

Sample population and assay methodology

A local breast cancer dataset (N=98) was assembled from breast cancer diagnosed between 2007-2012. Equal numbers from each breast cancer class (ER, HER2, TNBC and ER and HER2 positive) were recruited. Haematoxylin & Eosin (H&E) slides were examined to identify the tumour area. The tumour area was isolated by microdissection or macrodissection depending on the tumour morphology, with the aim of having the highest tumour content within the sample. The dissected section was stained with H&E to confirm tumour cell content. Following sample lysis using a ratio of 2.5mm ³ per 300mI_ of Homogenizing mix, RNA expression is measured using a custom designed Invitrogen Quantigene 2.0 10-plex assay. fAl Data quality control

Quality control starts by assessing the bead count per region. If the bead count is less than 30, sample is regarded as inadequate and a repeat would be necessary. Low bead counts are generally caused by bead clumping which can be avoided by obtaining a cleaner sample (less paraffin and tissue fragments) or by diluting the sample.

The Limit of Detection (LOD) is determined as the average of the blank + 3 Standard Deviations of the same blanks for each gene. A minimum of 3 blank wells are required per assay. Sample Median Fluorescence Intensity (MFI) expression values lower than the LOD are regarded as undetected for each respective gene. fBl Detecting outliers

Normalising gene data is expressed as a log base 10, resulting in a normally distributed dataset.

The z-score of the log is computed based on the average and standard deviation derived from the population set (N=98; N=94 complete with HER2 and ER IHC data). A threshold to select outliers was set based on concordance for the classification of HER2 status defined by Immunohistochemistry (IHC) and Fluorescence in situ Hybridisation (FISH) (N=28/28 HER2 positive), using ERBB2 expression normalised to normalising gene sets. The same exercise was performed using the ESR1 expression to predict the Oestrogen receptor (ER) status (N=42/43 ER positive). The best classification with the minimal loss of samples as outliers was selected: Outlier threshold for saturated or too dilute samples: N of outliers = 12; N; N of adequate samples = 86, 82 are annotated by the ER and HER2 IHC.

-1.6 < TBP log z-score < 1.6 Classification of Breast Cancer Classes

The Breast cancer classes (TNBC, ER+ and HER2+) can be predicted using ESR1 and ERBB2 expression data normalised to 4 normalising genes (ACTB, PPIB, HPRT1 , and TBP) with up to 98.78% accuracy using the Neural network algorithm (1 out of 82 cases is misclassified).

Breast Cancer Classification by IHC (ER, HER2) true TNBC true ER true HER2 class precision:

100.00% i

96.67%

100.00%

class recall 95.83% 100.00% 100.00%

Table 3: Crosstab of Breast cancer classification defined by IHC and FISH compared to Neural Network prediction based on ESR1 and ERBB2 normalised expression. Classification accuracy of 98.78%, N=82

Adding AURKA and/or KIF2C gives a 100% accuracy for this classification when using Neural Net algorithm.

Breast Cancer Classification by IHC (ER, HER2) true TNBC true ER true HER2 class

precision

100,00%

100.00%

100 00

¹ class recall 100.00% 100.00% 100.00%

Table 4: Crosstab of Breast cancer classification defined by IHC and FISH compared to Neural Network prediction based on ESR1, ERBB2 and AURKA and/or KIF2C normalised expression.

Classification accuracy of 100%, N-82 TCGA - PAM50 and Diagnostic IHC breast cancer classification

RNASeq V2 data of N = 91 1 Breast cancer cases annotated with PAM50 (N=512)

annotated with complete IHC and FISH data (N=470)

annotated with complete IHC and FISH data and PAM50 (N=316) Data was retrieved on 16 ^th January, 2016 from The Cancer Genome Atlas Portal (https i//tcqa-data.nci.nih.qov/tcqa/) from the Breast Invasive Carcinoma (TCGA, Provisional) cohort. Samples derived from normal tissue was omitted for this analysis.

Concordance of PAM50 with HER2 classification and Breast cancer classes as defined by IHC (N = 317).

HER2 HER2 Class

HER2 Status by IHC and FISH

Negative Positive Precision

Table 5: Crosstab of HER2 Status defined by IHC and FISH compared to PAM50 HER2-Enriched or otherwise. Classification concordance of 88.64%, N=317

HER2 cases were selected using the Neural Network algorithm based on the ERBB2 gene expression. Other algorithms such as Decision Tree, Random Forest and Support Vector Machine can be applied here with as similar prediction (Respective concordance of 95.0%, 95.3% and 94.0%).

HER2 HER2 Class

HER2 status by IHC

Negative Positive precision

Neural Net prediction based 94.8%

94.6% on ERBB2 expression

Class recall 99.1% 74.3% Table 6: Crosstab of Breast cancer classification of HER2 status defined by IHC and FISH compared to Neural Net prediction based on the normalised expression of ERBB2 and shown as pred. HER2 negative and pred. HER2 positive. Classification concordance of 94.76%, N-401 Class

Breast Cancer Class by IHC FISH ER HER2 TNBC

precision

Class recall 93.3% 57.1% 90.2%

Table 7: Crosstab of Breast cancer classification defined by IHC and FISH compared to PAM50 Luminal, Basal and HER2-Enriched. Classification concordance of 86.39%, N=316

Class

Breast Cancer Class by IHC/FISH ER HER2 TNBC

precision

Table 8: Crosstab of Breast cancer classification defined by IHC and FISH compared to Neural Net prediction based on the normalised expression of ESRl and ERBB2 and shown as pred. ER, pred. HER2 and pred. TNBC. Classification concordance of 92.79%, N=402 iCl Algorithm to select HER2 positive patients:

HER2 status as determined by IHC and FISH can be predicted perfectly by Neural Network based on the expression of ERBB2 normalised expression. A Neural Network model trained on available dataset can be used to predict status of unknown cases.

Neural Network prediction of TCGA IHC/FISH data based on ERBB2 normalised expression (ACTB, PPIB, HPRT1 , TBP).

Prediction of N = 401 breast cancer for HER2 status (Table 65)

[D] Luminal Basal Classification

The HER2 cases N=55/401 were removed from the TCGA dataset as HER2 positive cases. The PAM50 HER2-Enriched were also omitted to obtain a reference dataset specific for the Luminal/Basal classification based on the PAM50 classification (N = 372). Cases that were regarded as HER2 positive by IHC/FISH but not with the Neural Network algorithm or the PAM50 (N=9) were also removed as potential confounders.

Final dataset = 363

PAMSO Luminal Basal class

precisian

Prediction by Neural Network based on pred. Luminal 98.9%

pred. Basal 91.1%

ESR1 and ERBB2 expression

class recall 97.6% 96.0%

Table 9: Crosstab of the molecular classification of breast cancer with PAM50 as reference. A Neural Net prediction based on the normalised expression of ESR1 and ERBB2 is shown as pred. Luminal and pred. Basal. Classification concordance of 97.25%

When including also PGR, AURKA, KIF2C and FOXC1 normalised expression as predicting variables, the algorithm provides a more accurate classification of Luminal and Basal patients (Table 10).

PAMSO Luminal Basal class

precision

Prediction by Neural Network based on pred. Luminal 286 98.96%

pred. Basal 2 97.30%

ESR1, ERBB2, PGR, AURKA, KSF2C and

class recall

FOXC1 expression 99.31% 96.00%

Table 10: Crosstab of the molecular classification of breast cancer with PAMSO as reference. A Neural Net prediction based on the normalised expression of ESRl, ERBB2, PGR, AURKA, KIF2C and F0XC1 is shown as pred. Luminal and pred. Basal. Classification concordance of 98.62%

{El Dissecting Luminal Breast Cancer into Luminal A and Luminal B

Breast cancer cases that were classified as Luminal by both PAM50 and the Neural Network Algorithm were selected to train the algorithm on but also assess the performance as the patient set subclasses are known. (N=286 of which 3 cases had no survival data).

PAMSO Luminal A vs Luminal B Survival Data Processing Summary

AU RKA and KIF2C are required for clinically significant Luminal A and Luminal B categorisation

Algorithm 9: Gene signature: ERBB2, ESR1 , FOXC1 , PGR (Normalised to ACTB, HPRT1 and TBP)

PAM50 true true class

Luminal A Luminal B precision

Table 11: Crosstab of the Luminal classification of breast cancer with PAM50 as reference. A Neural Net prediction based on the normalised expression of ESRl, ERBB2, PGR and FOXC1 is shown as pred. Luminal A and pred. Luminal B. Classification concordance of 76.14%, N-285

Survival Data Processing Summary

Algorithm 10: Neural Net Prediction using gene signature: ERBB2, ESR1, FOXCl, PGR, AURKA, KIF2C (Normalised to ACTB, HPRT1 and TBP)

PAMSO true Luminal true Luminal class

A B precisian

Prediction by Neural Network based pred. Luminal 171 91.44%

pred. Luminal B 22 77.55% on ESR1, ERBB2, PGR, AURKA, KIF2C

and FOXCl expression class recall 88.60% 82.61%

Table 12: Crosstab of the Luminal classification of breast cancer with PAM50 as reference. A Neural Net prediction based on the normalised expression of AURKA, KIF2C, ESR1, ERBB2, PGR and FOXCl is shown as pred. Luminal A and pred. Luminal B. Classification concordance of 86.67%, N=285

Case Processing Summary

Algorithm 11 : Gene signature: PGR, FOXC1 , AURKA and KIF2C (Normalised to ACTB, HPRT1 and TBP)

PAM50 true true class

Luminal A Luminal B precision

pred.

87.38%

Lumina

Prediction by Neural Network pred.

13 66 83.54% Luminal B

based on PGR, AURKA, KIF2C and

class

FOXC1 expression 93.26% 71.74%

recall

Table 13: Crosstab of the Luminal classification of breast cancer with PAM50 as reference. A Neural Net prediction based on the normalised expression of AURK A, KIF2C, PGR and F0XC1 is shown as pred. Luminal A and pred. Luminal B. Classification concordance of 8632%, N-285

Case Processing Summary

Algorithm 12: Gene signature: PGR and FOXC1 (Normalised to ACTB, HPRT1 and TBP)

PAM50 true true class

Luminal A Luminal B precision

Prediction by Neural Network based on pred. Luminal A 171 77.73 %

PGR and FOXC1 expression pred. luminal B 22 43 §6.15%

I class recall 88.6096 46.74 %

Table 14: Crosstab of the Luminal classification of breast cancer with PAM50 as reference. A Neural Net prediction based on the normalised expression of PGR and FOXC1 is shown as pred. Luminal A and pred. Luminal B. Classification concordance of 75.09%, N-285

Case Processing Summary

Example 3

APPENDIX A - Neural Network computation details

Neural Net details (computed using RapidMiner Studio Community software):

All dataset was used to train and as an unlabelled dataset.

500 training layers

0.3 learning rate

0.2 momentum

Shuffle

Normalise

Error Epsilon: 1.0E-5

Example 4 - Algorithms

Algorithm 1 - trained on normalised expression of ERBB2 in:

a sample set comprising HER2+ samples as defined by IHC and FISH;

and a sample set not comprising HER2+ samples as defined by IHC and FISH;

Algorithm 2 - trained on normalised expression of ESR1 in:

a sample set comprising ER+ samples as defined by IHC;

and a sample set not comprising ER+ samples as defined by IHC;

Algorithm 3 - trained on normalised expression of ERBB2 and ESR1 in:

a sample set comprising HER2+ and/or ER+ samples as defined by IHC and FISH; and a sample set not comprising HER2+ and ER+ samples as defined by IHC and FISH;

Algorithm 4 - trained on normalised expression of ESR1 , ERBB2, AURKA and/or KIF2C in: a sample set comprising HER2+ and/or ER+ samples as defined by IHC and FISH; and a sample set not comprising HER2+ and/or ER+ samples as defined by IHC and FISH;

Algorithm 5 - trained on normalised expression of ERBB2 samples in a dataset from The Cancer Genome Atlas (TGCA) comprising

HER2+ samples as defined by IHC and FISH; and

samples defined as not HER2+ by IHC and FISH;

Algorithm 6 - trained on normalised expression of ERBB2 and ESR1 in samples in a dataset from the TCGA, comprising

HER2+ and/or ER+ samples as defined by IHC and FISH; and

Samples defined as not HER2+ and not ER+ by IHC and FISH;

Algorithm 7 - trained on normalised expression of ERBB2 and ESR1 in samples in a dataset from the TCGA which had HER2-enriched samples as defined by PAM50 and HER2+ as defined by IHC/FISH removed.

Algorithm 8 - trained on normalised expression of ERBB2, ESR1 , PGR, AURKA, KIF2C and FOXC1 in samples in a dataset from the TCGA which had HER2-enriched samples as defined by PAM50 and HER2+ samples as defined by IHC/FISH removed.

Algorithm 9 - trained on normalised expression of ERBB2, ESR1 , FOXC1 , PGR normalised expression in:

samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8;

Algorithm 10 - trained on normalised expression of ERBB2, ESR1 , FOXC1 , PGR, AURKA and KIF2C in:

samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8; Algorithm 1 1 - trained on normalised expression of AURKA, KIF2C, FOXC1 and PGR in: samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8;

Algorithm 12 - Trained on normalised expression of FOXC1 and PGR in:

samples in a dataset that comprises breast cancer cases that were classified as Luminal by both PAM50 and the Algorithm 8; and

samples in a dataset that comprises breast cancer cases that were classified as not Luminal by both PAM50 and the Algorithm 8; optionally wherein all expression levels were normalised to the expression levels of ACTB; PPIB; HPRT1 ; and/or TBP, preferably wherein all expression levels were normalised to the expression levels of ACTB; PPIB; HPRT1 ; and TBP.

Algorithm 1 identifies samples as HER2 positive or HER2 Negative.

Algorithm 2 identifies samples as ER positive or ER Negative.

Algorithm 3 identifies samples as HER2 positive and/or ER positive or TNBC.

Algorithm 4 identifies samples as HER2 positive and/or ER positive or TNBC.

Algorithm 5 identifies samples as HER2 positive or HER2 Negative.

Algorithm 6 identifies samples as HER2 positive and/or ER positive or TNBC.

Algorithm 7 identifies samples as Luminal or Basal.

Algorithm 8 identifies samples as Luminal or Basal.

Algorithm 9 identifies samples as Luminal A or Luminal B.

Algorithm 10 identifies samples as Luminal A or Luminal B.

Algorithm 1 1 identifies samples as Luminal A or Luminal B.

Algorithm 12 identifies samples as Luminal A or Luminal B

Example 5 - Concordance with receptor status using FFPE material

A Neural Net algorithm (as described in Algorithm section Example 2) has been trained based on a subset of breast cancer patients that have been previously analysed and classified into Luminal, Basal and HER2-Enriched by PAM50 (94.81 % overall concordance). The reference classification was defined using a patented 12-gene signature that was validated on Quantigene 2.0 (W02009055823). Likewise, a Neural Net algorithm trained using 6-gene Quantigene data (ERBB2, ESR1 , AURKA, KIF2C, FOXC1 ) against the 12-gene signature (W02009055823) molecular classification as reference, can use predict Luminal, Basal and HER2-Enriched molecular classification with 100% concordance. The 6-gene signature presented here, can be applied to the TCGA breast cancer data to predict PAM50 with 95.03% concordance when using the Neural Net algorithm.

Normalisation using various normalising genes, shows that a combination of 4 normalising genes (ACTB, PPIB, HPRT1 and TBP) will allow a 100% concordance of HER2 overexpression when compared to IHC and FISH results.

Example 6 - Quantigene expression analysis using degraded RNA

Sample used:

• BT474 RNA - a HER2 positive, ER positive breast cancer cell line derived from a primary infiltrative ductal carcinoma of luminal type

• MDAMB453 RNA - derived from a Luminal breast cancer cell line with amplified ERBB2 gene that is not overexpressed.

• Breast cancer tumours in paraffin blocks (Formalin Fixed Paraffin Embedded - FFPE)

Methodology:

RNA degradation was performed by heating RNA at 95°C (as suggested by TATAA Biocenter - AIT 2016) over the following time points:

1 . Untreated,

2. 1 minute,

3. 5 minutes,

4. 10 minutes

5. 30minutes

6. 60 minutes

7. 120 minutes

RNA degradation was significant at 5 minutes until complete degradation at 30 minutes where it becomes comparable to degradation seen in tissue FFPE blocks, also in terms of RIN (Figure 12).

Table 15

Stable RNA Expression profiles in heat degraded RNA - see Figure 14.

RAW data is defined by MFI data with subtracted the blank average and any signals below (blank average + 3SDs) are indicated as 0 (undetected)

On analysis of the 40-gene plex a significant reduction in signal is observed when RNA was heat denatured for over 60 minutes.

When assessing normalised data this effect seems to be overcome. The housekeeping gene RNA and the target RNA seem to bear the same effect from degradation, hence data derived from degraded RNA can yield the same information when normalised.

The Neural Network algorithm was trained based on patient FFPE data (N=82, 29/82 are HER2 positive) without outliers using the HER2 IHC/FISH annotation. This algorithm is applied on the BT474 degraded RNA series data to predict HER2 status based on ERBB2 gene expression (Figure 15). The Neural network can accurately determine the HER2 status across the RNA degradation series using ERBB2 or ERBB2/ESR1 (Table 16).

BT474 RNA NEURAL NET ALGORITHM CONFIDENCE

DEGRADATION SERIES HER2 PREDICTION

Stable Positive 0.999

5min (5 ⁾ 95 ^D C Positive 0.999 lOmin Positive 0.999

30min @ 95°C Positive 0.999

120min @ 95°C Positive 0.999 Stable Positive 0.999

5min (5 ⁾ 95°C Positive 0.999 lOmin Positive 0.999

30min @ 95°C Positive 0.999 Positive 0.999

Table 16 Neural Network algorithm HER2 status prediction results for the BT474 heat degraded RNA.

Example 7 - Subtyping Breast Cancer Cell lines in FFPE

To provide a proof of principle for the performance of Quantigene on degraded RNA samples, cell lines were subjected to the same tissue processing that a tumour tissue would undergo during the diagnostic process. The paraffinized cell lines were cored, lysed and analysed using Quantigene.

Cell blocks were produced to represent well annotated cell lines that are formalin fixed, and embedded in paraffin (FFPE-cell line blocks). Using a 12-plex Quantigene assay, normalising genes and classifier genes are measured in the FFPE-cell blocks to provide quality control material for the patient classification runs.

Samples used:

Patient-derived breast cancer samples from FFPE n = 82 (12 more cases did not pass QC)

Breast Cancer Cell Lines (BCCL) FFPE = 11 (2 HER2 positive + 9 HER2 Negative)

Total N = 93

Neural Network Prediction of BCCL in FFPE

The Neural Network algorithm trained on the patient set (N=82) can predict the HER2 status of the BCCL FFPEs with 100% accuracy. A larger dataset will be required to validate and provide a more robust algorithm for class prediction.

BREAST HER 2

CANCER CELL HER2 STATUS ALGORITHM

PREDICTION

Positive Positive

Positive Positive

Negative Negative

MCF10A FFPE Negative Negative HB2 F Negative Negative

436 FF Negative Negative

231 FF Negative Negative

468 FF Negative Negative

HCC1 Negative Negative

HS578 Negative Negative

BT20 Negative Negative

Table 17: Neural Network Predicted HER2 status based on Breast Cancer cell line ( BCCL ) FFPE compared with the actual HER2 status confirmed by IHC.

Table 18: Neural Network Performance in determining the HER2 status on breast cancer FFPE (N=82) and cell line FFPE dataset (N = 11). The algorithm can detect the HER2 status of the available dataset with complete accuracy.

Example 8

Optimisation of RNA based assays using archival formalin-fixed paraffin-embedded (FFPE) material is challenging due to variability in surgical tissue processing and degradation of RNA caused by tissue integrity preservation using formalin ¹ · ² . To overcome the limitation of performing accurate gene expression studies from archival material, our group used the branched-chain DNA (bDNA) multiplex magnetic bead assay. bDNA technology replaces enzymatic amplification of target template with hybridisation of specific probes and amplification of a reporter signal ³ . The short recognition sequences of the capture and detection probes are designed to hybridise to short fragments of target RNA ⁴ . In addition, the use of tissue homogenates directly as starting material of this assay, overcomes the inevitable loss of RNA occurring in assays requiring prior RNA extraction and purification. Signal amplification, the use of short recognition sequences and the lack of a purification step, contribute to reduce technical variation of the assay. The technology provides the possibility to multiplex the assay (up to 80 RNA targets), measuring expression of a panel of targets from low material input. This protocol describes the preparation and staining of tissue samples for laser microdissection. Staining on membrane slides facilitate imaging of tumour and histological architecture providing accurate selection and profiling of (1 ) tumour and normal ducts in breast tissue and (2) malignant cell clones within heterogeneous tumours.

Molecular classification of Breast cancer interrogates molecular markers to categorise patients into molecular classes, namely luminal classes, the human epidermal growth factor receptor 2 (HER2)-enriched and the basal types. The HER2-enriched subtype is well defined, with high expression of HER2 receptor, due to the ERBB2 gene amplification, combined with low or absent ER and PgR. Luminal subtype is generally positive for oestrogen receptor (ER) and the Basal type tumours are in general negative for the 3 receptors significantly overlapping with the triple negative breast cancer (TNBC) diagnostic subtype ⁵ · ⁶ . Other markers are used to determine epithelial and mesenchymal characteristics. Fibronectin (FN1 ) is a main component of breast tissue mesenchymal compartment. Increased FN1 expression is accompanied by high Ki67 staining showing a signature for a more invasive tumour ^{7 8} and an increased expression is associated with metastasis ⁹ . Interestingly, FN1 was found to be present in microvesicles originating from tumour cells, that induced activation of mitogenic signals in recipient fibroblasts ¹⁰ . Hence, circulating microvesicles such as exosomes are potentially vehicles of early detection and early indicators of metastasis and relapse ¹¹ .

Tumour area selection for breast cancer transcriptional subtyping has recurrently been done by macrodissection ^{12 13} . To overcome tissue heterogeneity and increase sensitivity, we have reliably combined classical tissue staining with multiplex molecular profiling methods. As a proof of principle, two distinct breast cancer clones have been defined by their epithelial mesenchymal signature and metastatic potential. The workflow of the described protocol can be easily translated to the current clinical setup and used to selectively isolate and characterise tissue subtypes using targeted mRNA profiling.

The described method has been applied for the simultaneous measurement of 40 transcripts in H&E stained (Figure 16), microdissected (Figure 17) highly degraded FFPE material. Using this method, we show accurate characterisation of receptor status (Figure 18A), classification of tumours into luminal and basal molecular subtypes ¹⁷ and differential expression of the mesenchymal marker, FN1 when comparing tumour and matched control tissue (Figure 18B), in the various receptor positive and negative subtypes. A bead-based multiplex branched DNA (bDNA) assay has been optimised to quantify gene expression on degraded RNA derived from formalin fixed paraffin embedded (FFPE) breast cancer tissue and normal breast ducts. Optimising the assay, involved developing an algorithm to classify breast cancer patients in luminal and basal subtypes utilising a 8 well-known biomarkers and 5 potential normalising genes. Data normalisation was done using permutations of the normalising genes. The selection of the normalising genes was based on the best prediction of receptor status using the Luminal/Basal classifier genes. To classify Luminal/Basal subtypes from FFPE tissues, the normalising genes selected were Beta-actin (ACTB), Glyceraldehyde 3-phosphate dehydrogenase (GAPDFI) and Flypoxanthine Phosphoribosyltransferase 1 (FIPRT1 ).

The method can be adapted for use in other diagnostic and research areas following adequate selection of the normalising gene set. One important application of this method in the research sector is the measurement of biomarkers in archival material that is well annotated with clinical outcomes. This provides the possibility to validate potential predictive markers in retrospective studies, quickly and accurately, avoiding long-term prospective studies awaiting disease-free survival and overall survival data. Figures 25 and 26 show the correlation of RNA expression level of genes from colorectal cell lines with the expression level in exosomes. The data shows that AURKA can be used as a normalising gene to characterise transcript expression of diagnostic genes such as FIER2 within cancer-derived exosomes. FIER2 expression is in concordance with HER2 amplification. Similarly, this can be applied to other diagnostic gene amplifications and/or overexpression. The use of liquid biopsies and robust gene expression assays will allow high throughput multiplex assays adapted for patient management during treatment, providing the means to follow treatment efficacy, potential relapse due to resistance to therapy and the metastatic capacity of the tumour.

This method has also a wide range of possible applications in the diagnosis of tumours and is adapted to the current diagnostic workflow. The main advantages of the method in the diagnostic field include (1 ) implementation of high throughput assays, (2) excluding subjectivity and equivocal results originating from image-based measurements; (3) accurate detection of multiple targets simultaneously, enhancing accuracy and minimising use of precious patient samples and (4) there is no need for highly specialised facilities and human resources. The optimised sampling process, together with the low input of material required for the bead-based multiplex assay, allows further investigation of tumour heterogeneity by the use of laser microdissection to accurately separate multiple foci of malignant tissue from the same patient section and compare multiple gene expression between them and also with matched normal tissue (Figure 19). Low material input is vital for diagnostic application on tumour biopsies that provide limited tumour tissue. The capacity of the assay to measure gene expression from degraded RNA samples allows easy transportation of samples for analysis within an institution or to servicing laboratories. In addition, whole section analysis was also possible using H&E stained material (Figure 16).

For the success of this protocol, it is imperative to (1 ) ensure proper sampling of the tumour site/s that are lysed for the assay and (2) develop well optimised and validated data normalisation algorithms, for each gene expression panel and/or individual prognostic or predictive biomarkers. The former depends on the technical experience of the technician/scientist performing the sampling. It is recommended to take an additional core and prepare a tissue microarray (TMA) in the same format of the multiplex magnetic bead assay runs (96 well format). This will provide an archive of tumour sites as a replica of samples used for the RNA-based assay. TMAs can also be assessed with other techniques for follow-up research or validation of results. The development of normalisation algorithms is dependent on the material being investigated and the normalising genes selected for normalisation. Different panels of normalising genes are selected based on the level and variability of expression in the sample analysed and this varies between cancer tissues from different origins, exosomes from plasma or circulating tumour cells. Validation of the assay includes sample processing since various preparations will also result in different normalisation algorithms.

To summarise, the use of bDNA technology in combination with magnetic bead technology and the selection of the proper panel of target genes, will provide the added advantage of measuring gene expression directly in tissue lysates derived from small amounts of patient material, including microdissected material, exosomes and circulating tumour cells. In addition to detection of tumour heterogeneity, the proper use of panels has the potential to detect tumour derived exosomes for early diagnostics and early detection of relapses. Since there is no need of any nucleic acid amplification step, signal amplification using the bDNA technology, combined with the bead-based multiplex measures multiple gene expression in clinically-annotated archival material providing a resource for biomarker validation.

REFERENCES:

1 Abrahamsen, H. N., Steiniche, T., Nexo, E., Hamilton-Dutoit, S. J. & Sorensen, B. S. Towards quantitative mRNA analysis in paraffin-embedded tissues using real-time reverse transcriptase-polymerase chain reaction: a methodological study on lymph nodes from melanoma patients. J Mol Diagn. 5 (1 ), 34-41 , doi:S1525-1578(10)60449-7 [pii]

10.1016/S1525-1578(10)60449-7, (2003).

2 Macabeo-Ong, M. et al. Effect of duration of fixation on quantitative reverse transcription polymerase chain reaction analyses. Mod Pathol. 15 (9), 979-987 (2002).

3 Yang, W. et al. Direct quantification of gene expression in homogenates of formalin-fixed, paraffin-embedded tissues. Biotechniques. 40 (4), 481 -486 (2006).

4 Flagella, M. et al. A multiplex branched DNA assay for parallel quantitative gene expression profiling. Anal Biochem. 352 (1 ), 50-60, doi:S0003-2697(06)001 17-5 [pii] 10.1016/j.ab.2006.02.013, (2006).

5 Goldhirsch, A. et al. Strategies for subtypes-dealing with the diversity of breast cancer: highlights of the St. Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 201 1 . Ann Oncol. 22 (8), 1736-1747, doi:10.1093/annonc/mdr304 mdr304 [pii], (201 1 ).

6 Schnitt, S. J. Classification and prognosis of invasive breast cancer: from morphology to molecular taxonomy. Mod Pathol. 23 Suppl 2 S60-64, doi: 10.1038/mod pathol.2010.33

modpathol201033 [pii], (2010).

7 Bae, Y. K. et al. Fibronectin expression in carcinoma cells correlates with tumor aggressiveness and poor clinical outcome in patients with invasive breast cancer. Hum Pathol. 44 (10), 2028-2037, doi:10.1016/j.humpath.2013.03.006

S0046-8177(13)001 18-4 [pii], (2013).

8 Park, J. & Schwarzbauer, J. E. Mammary epithelial cell interactions with fibronectin stimulate epithelial-mesenchymal transition. Oncogene. 33 (13), 1649-1657, doi:10.1038/onc.2013.1 18

onc20131 18 [pii], (2014).

9 Zhou, Z. et al. MRI detection of breast cancer micrometastases with a fibronectin- targeting contrast agent. Nat Commun. 6 7984, doi:10.1038/ncomms8984

ncomms8984 [pii], (2015).

10 Antonyak, M. A. et al. Cancer cell-derived microvesicles induce transformation by transferring tissue transglutaminase and fibronectin to recipient cells. Proc Natl Acad Sci U S A. 108 (12), 4852-4857, doi: 10.1073/pnas.1017667108

1017667108 [pii], (201 1 ).

1 1 Moon, P. G. et al. Fibronectin on circulating extracellular vesicles as a liquid biopsy to detect breast cancer. Oncotarget. 7 (26), 40189-40199, doi:10.18632/oncotarget.9561 9561 [pii], (2016). 12 Parker, J. S. et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol. 27, doi:10.1200/jco.2008.18.1370, (2009).

13 Nielsen, T. et al. Analytical validation of the PAM50-based Prosigna Breast Cancer Prognostic Gene Signature Assay and nCounter Analysis System using formalin-fixed paraffin-embedded breast tumor specimens. BMC Cancer. 14 177-177, doi: 10.1186/1471 - 2407-14-177, (2014).

14 Fischer, A. H., Jacobson, K. A., Rose, J. & Zeller, R. Hematoxylin and Eosin Staining of Tissue and Cell Sections. Cold Spring Harbor Protocols. 2008 (5), pdb.prot4986, doi:10.1 101/pdb.prot4986, (2008).

15 Molecular Machines and Industries. H&E Staining Kit Plus for Laser Microdissection and more, Prod. No 70302.

16 Lin, F. & Prichard, J. Handbook of Practical Immunohistochemistry: Frequently asked questions. (Springer, 201 1 ).

17 Grech, G. et al. Molecular Classification of Breast Cancer Patients Using Formalin- fixed Paraffin-embedded Derived RNA Samples. Journal of Molecular Biomarkers & Diagnosis. 7 (S8), doi:doi: 10.4172/2155-9929.S8-016, (2016).

18 MACKAY A et al 201 1. Microarray-Based Class Discovery for Molecular Classification of Breast Cancer: Analysis of Interobserver Agreement. JNCI: Journal of the National Cancer Institute, 103, 662-673.

19 PEROU C et al 2010. Clinical implementation of the intrinsic subtypes of breast cancer. The Lancet Oncology, 1 1 , 718-719.

Example 9

Identification of markers of sensitivity to PP2A activators

Data analysis of well annotated, publicly available databases (TCGA) showed that PP2A deregulation is a common event in breast cancer, with 59.6% of basal breast cancer patients showing a low expression of one of the PP2A complex components or a high expression of the inhibitory subunits. The main aim of the study was to identify genes whose expression depends directly on PP2A activity. In the ideal scenario, these novel biomarkers would reflect PP2A activity regardless of the mechanism driving the activity regulation.

We first established cell line models to investigate possible markers of sensitivity to PP2A activators. Basal breast cancer cell line MDAMB231 was sensitive to the PP2A activator, 20 FTY720, and showed a suppressed PP2A activity while the MDAMB453 Luminal cell line was not sensitive to FTY720 and had relatively high PP2A activity. Drug sensitivity is the inverse of percent viability.

In more detail, the dose-dependent effect of FTY720 on breast cancer cell lines provided information on the sensitivity of the cell line to PP2A activation. Cytotoxicity 25 results for the 12 breast cancer cell lines tested are illustrated in Figure 21 where the % cell viability is expressed as a percentage of the parallel control culture (without drug) against FTY720 concentration. The average vehicle control (VC) % cell viability for each cell line was never under 85% viability compared to the untreated ruling out any interference by the vehicle. The 10mM and the 25mM doses were considered to be cytotoxic as cellular 30 morphology was generally altered, and hence are not shown.

The three sensitive cell lines, defined by % cell viability under 50% (IC50) at a dose lower than 5mM, were represented by a Triple Negative phenotype (Figure 21 A). BT-20 was the most sensitive at 0.5mM, followed by MDAMB231 and Hs578T showing and IC50 at 2.5mM FTY720 and 5mM FTY720, respectively. From this experiment, the effective dose of FTY720, that is not lethal, was set at 5mM for assessing the effect of 5 FTY720 on RNA expression. MCF10A is a cell-line derived from normal epithelium. This cell line was used as a baseline control in the study of the PP2A mechanism in breast cancer cell lines.

PP2A biomarker discovery and evaluation

The knowledge gained from the cellular models implied the need to characterise novel biomarkers for PP2A activity. The PP2A deregulation criteria was designed based on this information to probe datasets for gene candidates.

Commonly deregulated regulatory subunits of PP2A were identified. Preliminary 15 analysis of the PP2A regulators was done on TCGA data to assess possible pathways for deregulation within breast cancer. Initial analysis of transcript expression and protein expression of PP2A regulators in breast cancer cellular models did not identify clear markers that consistently reflect FTY720 sensitivity. Moreover, the assessment of PP2A regulators and phosphotargets of 20 PP2A, in local patients and a set of tissue arrays (TMAs), highlighted the complexity and heterogeneity of the feedback mechanism. Biomarkers that directly measure the activity of the complex or the specific function of the variety of complexes were required to predict sensitivity to drugs at a tissue level.

A number of candidate PP2A biomarker genes were identified. As outlined in 25 various studies, a biomarker needs to be well defined in its objectives, use and target population (Altar et al., 2007; Pepe, Feng, Janes, Bossuyt, & Potter, 2008). The approach for biomarker discovery involved the use of breast cancer data in the TCGA data portals in the context of PP2A deregulation. Genes were shortlisted using network analysis and known biological evidence connecting biomarkers to the PP2A pathway.

Following discovery, the viability of using these biomarkers in the same cohort based on RNASeqV2 expression was assessed. Validation analysis was performed using the TCGA RNASeqV2 data analysed manually providing a better insight into the data. The in silico analysis was then compared to observations within the cellular models and more importantly a cohort of local patients (N = 95 tumours, 3 Relapse tumours, 7 in situ, non- invasive tumours, 36 Normal breast tissue, of which 33 were patient-matched with 5 tumours).

The set of biomarkers was validated and evidence-based selection of biomarker candidates, identified AURKA and KIF2C as strong predictors of sensitivity to FTY720. These genes were used to classify patients and cell lines into those that are eligible to PP2A activation and those that are not predicted to have PP2A deregulation. The novel 10 therapeutic class was mainly represented by the Basal or Triple Negative groups and was associated with higher tumour aggressivity.

In more detail, the expression of the PP2A complex and regulator genes was measured in the 12 breast cancer cell lines when untreated, but also when treated with FTY720 (5mM). 15

Figure 22 shows that Cip2a (KIAA1524) RNA expression correlates with sensitivity of breast cancer cell lines to FTY720 (Figure 22). MDA-MB-231 exemplifies FTY720- sensitive breast cancer cell lines, showing a low PP2A enzymatic activity and a high Cip2a RNA expression relative to MCF10A (a cell line derived from normal epithelium). AURKA and KIF2C RNA expression were found to be positively correlated with 20 FTY720-sensitive breast cancer cell lines (Figure 22B). Of interest expression also correlated positively with PP2A endogenous inhibitors, cip2a and SET (Table 19).

Table 19: Spearman correlation analysis between the expression of the PP2A complex subunits and the 2 biomarkers. The normalised RNASeqV2 data as downloaded from the TCGA database was used. [N=873 tumour cases; Sig. (2-tailed) stands for two-tailed significance, significance at the 95% confidence interval is highlighted]

In addition, the expression of the genes within breast cancer tissue was also compared to the respective patient matched control tissue (Figure 23B). AURKA and KIF2C RNA expression was found to be significantly higher in tumour material relative to 5 matched normal tissues. This is also seen when patients are subdivided by the hormone status (Figure 23C).

AURKA and KIF2C are thus preferred as they are exclusive for aggressive tumours, providing evidence that they are directly involved in the malignant phenotype (Table 18).

Table 18: Occurrence of overexpression of the candidate biomarkers over the PAM50 breast cancer subtypes expressed as a percentage of the subtype population (N = 525). PAM50 annotation retrieved from data generated by TCGA Research Network: http://cancergenome.nih.gov/and data generated from the cBioPortal accessible through http://www.cBioPortal.org/ (Cerami, et ai, 2012; Gao, et al., 2013).

KIF2C and AURKA expression also correlated with percentage cell viability (Table 19) during FTY720 treatment in FTY720-sensitive cell lines. There was no correlation, when non sensitive cell lines were used. AURKA and KIF2C also correlates with CIP2A 5 across treatment.. This supports the use of KIF2C and AURKA as biomarkers to identify novel therapeutic groups within breast cancer patients eligible for PP2A activation therapy.

Table 19: Spearman correlation analysis between the expression of AURKA, KIF2C and CIP2A following treatment at 0.1, 1 and 5pM FTY720 in the FTY720-sensitive cell lines. 10 Correlation of expression with cellular viability is included. The normalised Quantigene expression levels were used.

[N=16 conditions (4 sensitive cell lines at 4 treatment conditions); Sig. (2-tailed) stands for two-tailed significance, significance at the 95% confidence interval is highlighted] PP2A activity biomarkers are correlated with protein expression of inhibitors and pS6K

AURKA expression was also found to correlate strongly with cytoplasmic cip2a staining (Table 20). AURKA significantly correlates with nuclear pS6K, establishing an association between AURKA expression and growth factor (PI3K/mTOR) signaling attenuation as part of the negative feedback driven by PP2A.

Table 20: Spearman correlations between PP2A activity biomarker expression and the protein expression of inhibitors of PP2A and downstream phospho-proteins with defined localisation in the breast cancer cell lines.

Expression of PP2A activity biomarkers is normalised to the housekeeping genes, while protein expression was scored using the H-score (0-300) for each cellular compartment and then expressed as positive or negative based on thresholds set on significance with survival and/or correlation with histopathological factors [ ^* significant at 95% confidence interval; + significant at 90% confidence interval]

Hence, in this study we identified a novel subgroup of basal-type patients, using KIF2C 5 and AURKA as biomarkers of sensitivity to the PP2A activator FTY720. The sensitivity of this class to FTY720 is supported following proper and expected classification of the sensitive cell lines within this subgroup. References for Example 9

KONINKI, K., BAROK, M., TANNER, M., STAFF, S., PITKANEN, J., HEMMILA, P., ILVESARO, J. & ISOLA, J. 2010. Multiple molecular mechanisms underlying

trastuzumab and lapatinib resistance in JIMT-1 breast cancer cells. Cancer Letters, 294, 21 1 -219.

KAO, J., S ALAR I, K., BOCANEGRA, M., CHOI, Y.-L, GIRARD, L, GANDHI, J„ KWEI, K. A., HERNANDEZ-BOUSSARD, T., WANG, P., GAZDAR, A. F., MINNA, J. D. & POLLACK, J. R. 2009. Molecular Profiling of Breast Cancer Cell Lines Defines Relevant Tumor Models and Provides a Resource for Cancer Gene Discovery. PLoS ONE, 4, e6146.

TANNER, M„ KAPANEN, A. L, JUNTTILA, T„ RAHEEM, 0., GRENMAN, S„ ELO, J., ELENIUS, K. & ISOLA, J. 2004. Characterization of a novel cell line established from a patient with Herceptin-resistant breast cancer. Molecular Cancer Therapeutics, 3, 1585- 1592.

Example 10

The expression of genes in RNA isolated from colorectal cancer cell lines was compared to the expression in exosomes derived from matched originator cells. Figure 1 shows that using the colorectal cancer cell line, COLO320 the expression of these genes correlated (r ² =0.9198), hence exosomes represent the expression in originator cells.

Cell lines were incubated for 48 hours at 0% FBS to produce exosomes. Measuring the proliferation marker AURKA indicates that originator cells are proliferating and provides an indication of the exosome quantity within the sample. Therefore, AURKA can be used as a normalising gene to characterise transcript expression of diagnostic genes such as HER2 within cancer-derived exosomes. HER2 expression is in concordance with HER2 amplification. Similarly, this can be applied to other diagnostic gene amplifications and/or overexpression.