(12) INTERNATIONAL APPLICATION PUBLISHED UNDER THE PATENT COOPERATION TREATY (PCT) (19) World Intellectual Property Organization International Bureau (10) International Publication Number (43) International Publication Date WO 2012/066451 Al 24 May 2012 (24.05.2012) P O P CT

(51) International Patent Classification: (74) Agent: BENSON, Gregg C ; Pfizer Inc. Eastern Point C12Q 1/68 (2006.01) Road, MS 9 114, Groton, Connecticut 06340 (US). (21) International Application Number: (81) Designated States (unless otherwise indicated, for every PCT/IB201 1/054962 kind of national protection available): AE, AG, AL, AM, AO, AT, AU, AZ, BA, BB, BG, BH, BR, BW, BY, BZ, (22) Date: International Filing CA, CH, CL, CN, CO, CR, CU, CZ, DE, DK, DM, DO, 7 November 201 1 (07.1 1.201 1) DZ, EC, EE, EG, ES, FI, GB, GD, GE, GH, GM, GT, HN, (25) Filing Language: English HR, HU, ID, IL, IN, IS, JP, KE, KG, KM, KN, KP, KR, KZ, LA, LC, LK, LR, LS, LT, LU, LY, MA, MD, ME, (26) Publication Language: English MG, MK, MN, MW, MX, MY, MZ, NA, NG, NI, NO, NZ, (30) Priority Data: OM, PE, PG, PH, PL, PT, QA, RO, RS, RU, RW, SC, SD, 61/413,806 15 November 2010 (15. 11.2010) US SE, SG, SK, SL, SM, ST, SV, SY, TH, TJ, TM, TN, TR, 61/470,381 31 March 201 1 (3 1.03.201 1) US TT, TZ, UA, UG, US, UZ, VC, VN, ZA, ZM, ZW. (71) Applicants (for all designated States except US): PFIZER (84) Designated States (unless otherwise indicated, for every INC. [US/US]; 235 East 42nd Street, New York, New kind of regional protection available): ARIPO (BW, GH, York 10017 (US). CENTRE HOSPITALIER UNI- GM, KE, LR, LS, MW, MZ, NA, RW, SD, SL, SZ, TZ, VERSITAIRE VAUDOIS [CH/CH]; Rue du Bugnon 21, UG, ZM, ZW), Eurasian (AM, AZ, BY, KG, KZ, MD, RU, CH-101 1 Lausanne (CH). SWISS INSTITUTE OF TJ, TM), European (AL, AT, BE, BG, CH, CY, CZ, DE, BIOINFORMATICS [CH/CH]; Quartier Unil Sorge-Bati- DK, EE, ES, FI, FR, GB, GR, HR, HU, IE, IS, IT, LT, LU, LV, MC, MK, MT, NL, NO, PL, PT, RO, RS, SE, SI, SK, ment Genopode, CH-1015 Lausanne (CH). SM, TR), OAPI (BF, BJ, CF, CG, CI, CM, GA, GN, GQ, (72) Inventors; and GW, ML, MR, NE, SN, TD, TG). (75) Inventors/Applicants (for US only): BUDINSKA, Eva [SK/FR]; 2 bis Rue du Jura, F-01 630 St. Genis Pouilly Declarations under Rule 4.17 : (FR). DELORENZI, Mauro Claudio [CH/CH]; Route du — as to the identity of the inventor (Rule 4.1 7(Ϊ)) jorat 118, CH-1000 Lausanne 26 (CH). PAVLICEK, — as to applicant's entitlement to apply for and be granted a Adam [CZ/US]; Pfizer Global Research & Development, patent (Rule 4.1 7(H)) 10777 Science Center Drive, San Diego, California 92121 (US). POPOVICI, Vlad Calin [RO/CH]; Rue du Bugnon — as to the applicant's entitlement to claim the priority of the 37, CH-1020 Renens, Vaud (CH). TEJPAR, Sabine earlier application (Rule 4.1 7(in)) [BE/BE]; Groenlaan 22, B-1560 Hoeilaart (BE). WEIN- — of inventorship (Rule 4.17(iv)) RICH, Scott Lawrence [US/US]; Pfizer Global Research & Development, 10777 Science Center Drive, San Diego, Published: California 92121 (US). — with international search report (Art. 21(3))

©

(54) Title: PROGNOSTIC AND PREDICTIVE GENE SIGNATURE FOR COLON CANCER (57) Abstract: The application provides methods of prognosing and classifying colon cancer patients into poor survival groups or good survival groups and for determining the benefit of adjuvant chemotherapy by way of a multigene signature. The application S also includes kits and computer products for use in the methods of the application. Prognostic and Predictive Gene Signature for Colon Cancer This application claims priority to U.S. Provisional Application No. 61/413,806 filed on November 15, 2010, and to U.S. Provisional Application No. 61/470,381 filed on

March 3 1, 201 1, both of which are incorporated herein by reference in their entireties.

Field The application relates to compositions and methods for prognosing and classifying colon cancer and for determining the benefit of adjuvant chemotherapy. Background As the third most common form of cancer, over 1 million new cases of colorectal cancer (CRC) are diagnosed worldwide each year. Despite significant advances in detection, surgery, and chemotherapeutic treatment, CRC is the fourth most common cause of cancer death worldwide, and second most common cause of cancer death in the United States (Tenesa & Dunlop, Nat Rev Genet 10:353-358 (2009); Jemal et al., Methods Mol. Biol. 471 :3-29 (2009)). CRC that is confined within the wall of the colon (TNM (tumor-node metastasis) stages I and II) are typically curable with surgery. However, if left untreated, such tumors may spread to regional lymph nodes (stage III), where up to 73% are curable by surgery and chemotherapy. Once CRC metastasizes to distant sites within the body (stage IV), the disease is typically not curable, although chemotherapy can extend the rate of survival. Clinical benefit in CRC patients has recently been observed with drugs that target vascular endothelial growth factor (VEGF) or epidermal growth factor receptor (EGFR). In particular, monoclonal antibodies that target EGFR (e.g. cetuximab and panitumumab) and VEGF (bevacizumab) are approved for clinical use to treat CRC. The constitutive activation of the mitogen-activated protein kinase (MAPK) pathway is a key driver of CRC tumorigenesis. The extracellular signal-regulated kinase (ERK) pathway plays a key role in cell proliferation and its aberrant activation is often due to oncogenic mutations in KRAS or BRAF genes (Fang and Richardson, Lancet Oncol. 6:322-327 (2005); Tejpar et al., Oncologist 15:390-404 (2010)). RAF is a serine-threonine-specific protein kinase that is activated downstream of the small G-protein RAS and which activates the MAP kinase (MEK) pathway, which in turn activates ERK. BRAF is one of the three highly conserved RAF genes in mammals (the other two being ARAF and CRAF) and its somatic mutations have been reported in approximately 7% of human cancers (Davies et al., Nature 417:949-954 (2002); Dhomen & Marais, Curr. Opin. Genet. Dev. 17:31-39 (2007)). In CRC, the BRAF mutations occur in 8-10% of sporadic cancers and generally are markers of poor prognosis. For example, the V600E mutation in BRAF, is believed to be associated with microsatellite instability (MSI), and may confer resistance to anti-EGFR therapy (Richman et al. J. Clin. Oncol. 27(35):5931-5937 (2009)). Furthermore, KRAS mutations are known to lead to EGFR-independent activation of the MAPK pathway, suggesting that therapies targeting EGFR will not be effective in patients with KRAS mutations (Benvenuti et al. Cancer Res 67: 2643-2648 (2007); Di Fiore et al., Br. J. Cancer 96:1 166-1 169 (2007)). Accordingly, there is an ongoing need to develop biomarkers that can effectively identify CRC patients that are best suited for certain therapeutic modalities. Summary As will be discussed in more detail herein, the present disclosure relates to the identification, from historical CRC patient data, several gene signatures that identify a subpopulation of patients that may be sensitive to novel targeted treatments. In particular, the present disclosure provides several gene signatures that are characteristic of BRAF mutated CRC tumors. The present disclosure provides methods and kits useful for obtaining and utilizing expression information for the genes identified herein, to obtain prognostic and diagnostic information for patients with CRC. The methods of the present disclosure generally involve obtaining relative expression data from a patient, at the DNA, messenger RNA (mRNA), or protein level, for each of the genes identified herein, processing the data and comparing the resulting information to one or more reference values. Relative expression levels are expression data normalized according to techniques known to those skilled in the art. Expression data may be normalized with respect to one or more genes with invariant expression, such as "housekeeping" genes. In some embodiments, expression data may be processed using standard techniques, such as transformation to a z-score, and/or software tools, such as RMAexpress v0.3.

In one aspect, a multi-gene signature is provided for prognosing or classifying patients with CRC. In some embodiments, a 39-gene pair signature is provided, comprising reference values for each of 39 pairs of different genes based on relative expression data for each gene from a historical data set with a known outcome, such as good or poor survival, and/or known treatment, such as adjuvant chemotherapy.

In one aspect, relative expression data from a patient are combined with the gene-specific reference values on a gene-by-gene basis for each of the genes identified herein, to generate a test value which allows prognosis or therapy recommendation. In some embodiments, relative expression data are subjected to an algorithm that yields a single test value, or combined score, which is then compared to a control value obtained from the historical expression data for a patient or pool of patients. In some embodiments, the control value is a numerical threshold for predicting outcomes, for example good and poor outcome, or making therapy recommendations, for example adjuvant therapy in addition to surgical resection or surgical resection alone. In some embodiments, a test value or combined score greater than the control value is predictive, for example, of high risk (poor outcome) or benefit from adjuvant therapy, whereas a combined score falling below the control value is predictive, for example, of low risk (good outcome) or lack of benefit from adjuvant therapy. The present disclosure provides gene signatures that are prognostic for survival as well as predictive for benefit from adjuvant chemotherapy. For example, the disclosure provides methods that can be used to select or identify subjects who might benefit from adjuvant chemotherapy as opposed to subjects who are not likely to benefit from such adjuvant chemotherapy. Accordingly, in one embodiment, the disclosure provides a method of prognosing or classifying a subject with CRC comprising: a) analyzing at least one of the gene pairs shown in Table 3.1 . 1 or Table 3.1 .2 according to the top scoring pair method; and b) classifying the subject into a BRAF mutant-like group or a wild-type group. For example, in one embodiment at least 10 of the gene pairs shown in Table 3.1 . 1 or 3.1 .2 are analyzed according to the top scoring pair method. In a further embodiment at least

30 of the gene pairs shown in Table 3.1 . 1 or Table 3.1 .2 are analyzed according to the top scoring pair method. In a further embodiment, the 39 gene pairs shown in Table

3.1 . 1 are analyzed according to the top scoring pair method. In a further embodiment, the 32 gene pairs shown in Table 3.1 .2 are analyzed according to the top scoring pair method.

In a further embodiment, the top scoring pair method is carried out by comparing the average value of the relative expression levels of all Genel genes used in the analysis with the average value of relative expression levels of all Gene2 genes used in the analysis, wherein if the average Genel value is less than the average Gene2 value, then the subject is classified as BRAF mutant-like. In a further embodiment, the top scoring pair method is carried out as described above, wherein if the average Genel value is greater than or equal to the average Gene2 value, then the subject is classified as wild-type. In some embodiments, the top scoring pair method uses the 39 pairs of genes shown in Table 3.1 . 1 . In some embodiments, the top scoring pair method uses the 32 pairs of genes shown in Table 3.1 .2.

In a further embodiment, the disclosure provides a method of prognosing or classifying a subject with CRC comprising: a) calculating a score using the AdaBoost method as described in Example 3, using the relative expression values of the genes shown in Table 3.2; and b) classifying the subject into a BRAF mutant-like group or a wild-type group. For example, in one embodiment the subject is classified as wild-type if the calculated score is less than 0.5, and the subject is classified as BRAF mutant-like if the score is 0.5 or greater.

In a further embodiment, the disclosure provides a method of prognosing or classifying a subject with CRC by using the CCP2 gene signature as described in Example 3. For example, the relative expression levels of the genes noted in Example 3.2.2 can be determined and the CCP2 method carried out as described in Example 3.1 .2. Using the CCP2 gene signature as described in Example 3 can be used to classify or prognose a subject with CRC as either BRAF mutant-like, or wild-type. In a further aspect, the present disclosure provides a method for selecting therapy comprising the steps of classifying or prognosing a subject with CRC using any of the methods described herein, and further comprising selecting adjuvant chemotherapy for a subject classified as wild-type, or selecting no adjuvant chemotherapy for a subject classified as BRAF mutant-like.

In one embodiment, the present disclosure provides a method for selecting therapy comprising the steps of classifying or prognosing a subject with CRC using any of the methods described herein, and further comprising selecting adjuvant chemotherapy for a subject classified as wild-type, or selecting a treatment regimen comprising a BRAF mutant-specific inhibitor for a subject classified as BRAF mutant like. In a further aspect, the present disclosure provides a method of treating a subject with CRC comprising administering a BRAF mutant-specific inhibitor to said subject, wherein said subject is classified as BRAF mutant-like according to any of the methods described herein. In one embodiment, the present disclosure provides any of the methods described herein, wherein said subject is a human.

In a further aspect, the present disclosure provides a CRC prognosticator comprising a mechanism for determining relative expression levels in a CRC tumor sample of the genes listed in Table 3.1 .1, Table 3.1 .2, Table 3.2, or those listed in Example 3.2.2. In one embodiment, the mechanism comprises a microarray. In a further embodiment, the mechanism comprises an assay of reverse transcription polymerase chain reaction. The application also provides for kits used to prognose or classify a subject with CRC into a good survival group or a poor survival group or for selecting therapy for a subject with CRC that includes detection agents that can detect the expression products of the biomarkers described herein, for example the gene pairs shown in Table 3.1 .1, Table 3.1 .2, or the genes listed in Table 3.2, or those listed in Example 3.2.2. Accordingly, in a further aspect the present disclosure provides a kit for classifying a subject with CRC comprising detection agents capable of detecting the expression products of at least one gene pair shown in Table 3.1 .1, or Table 3.1 .2, or of the genes shown in Table 3.1 . 1 or Table 3.1 .2, or in Example 3.2.2. In some embodiments, said agents are capable of detecting the expression products of at least 5, at least 10, at least 20, at least 30, at least 35, or the 39 gene pairs shown in Table

3.1 . 1 . In some embodiments, said agents are capable of detecting the expression products of at least 5, at least 10, at least 20, at least 30, or the 32 gene pairs shown in

Table 3.1 .2. In a further embodiment, any of the kits described above comprise an addressable array that comprises probes for the expression products of the at least one, at least 5, at least 10, at least 20, at least 30, at least 35, or the 39 gene pairs of Table

3.1 . 1 . In a further embodiment, any of the kits described above comprise an addressable array that comprises probes for the expression products of the at least one, at least 5, at least 10, at least 20, at least 30, or the 32 gene pairs of Table 3.1 .2. In a further embodiment, the detection agents comprise primers capable of hybridizing to the expression products of the gene pairs. In a further embodiment, the present disclosure provides any of the kits described herein, further comprising a computer implemented product for comparing: a) the relative expression level values for Genel genes in Table 3.1 . 1 or Table 3.1 .2 for a subject to b) the relative expression level values for Gene2 genes in Table 3.1 . 1 or

Table 3.1 .2 for said subject. In one embodiment, the average value of the relative expression levels of all Genel genes used in the analysis is compared with the average value of relative expression levels of all Gene2 genes used in the analysis. In one embodiment, the 39 gene pairs in Table 3.1 . 1 are used in the analysis. In one embodiment, the 32 gene pairs in Table 3.1 .2 are used in the analysis. The present disclosure provides probes for detecting the biomarkers described herein, for example the genes disclosed in Table 3.1 .1, Table 3.1 .2, Table 3.2, and those disclosed in Example 3.2.2. Exemplary probes include mRNA oligonucleotides, cDNA oligonucleotides, and PCR primers. The probes are capable of detecting or hybridizing to, each of the 39 pairs or 32 pairs of genes described in Example 3 .

In one aspect, the present disclosure provides kits useful for carrying out the diagnostic and prognostic tests described herein. The kits generally comprise reagents and compositions for obtaining relative expression data for the 39 gene pairs or 32 gene pairs, described in Table 3.1 . 1 or Table 3.1 .2, the genes shown in Table 3.2, or the genes noted in Example 3.2.2. The kits typically comprise probes for detecting the 39 gene pairs. The present disclosure also provides antibodies capable of specifically binding to the protein products of the biomarkers described herein. As will be recognized by skilled artisans, the contents of the kits will depend upon the means used to obtain the relative expression information. Kits may comprise a labeled compound or agent capable of detecting protein product(s) or nucleic acid sequence(s) in a sample and means for determining the amount of the protein or mRNA in the sample (e.g., an antibody which binds the protein or a fragment thereof, or an oligonucleotide probe which binds to DNA or mRNA encoding the protein). Kits can also include instructions for interpreting the results obtained using the kit.

In some embodiments, the kits are oligonucleotide-based kits, which may comprise, for example: ( 1 ) an oligonucleotide, e.g., a detectably labeled oligonucleotide, which hybridizes to a nucleic acid sequence encoding a marker protein or (2) a pair of primers useful for amplifying a marker nucleic acid molecule. Kits may also comprise, e.g., a buffering agent, a preservative, or a protein stabilizing agent. The kits can further comprise components necessary for detecting the detectable label (e.g., an enzyme or a substrate). The kits can also contain a control sample or a series of control samples which can be assayed and compared to the test sample. Each component of a kit can be enclosed within an individual container and all of the various containers can be within a single package, along with instructions for interpreting the results of the assays performed using the kit.

In some embodiments, the kits are antibody-based kits, which may comprise, for example: (1) a first antibody (e.g., attached to a solid support) which binds to a marker protein; and, optionally, (2) a second, different antibody which binds to either the protein or the first antibody and is conjugated to a detectable label. A further aspect provides computer implemented products, computer readable mediums and computer systems that are useful for the methods described herein. Other features and advantages of the present disclosure will become apparent from the following detailed description. It should be understood, however, that the detailed description and the specific examples while indicating preferred embodiments of the disclosure are given by way of illustration only, since various changes and modifications within the spirit and scope of the disclosure will become apparent to those skilled in the art from this detailed description.

Brief Description of the Drawings Figures 1 A and B: AUC and error rates when the model is built on the phase 1 data and validated on phase 2 data, for increasing model size.

Figures 2 A and B: AUC and error rates when the model is built on the phase 2 data and validated on phase 1 data, for increasing model size. Figure 3 : (AdaBoost) Distribution of BRAF scores: all scores above 0.5 (grey vertical line) indicate the "BRAF-like" samples. The small hash lines at the bottom right show the score of the BRAFmut samples and the small hash lines along the top are those of KRASmut samples. Figure 4 : (CCP2) Distribution of BRAF scores: all scores above the threshold (grey vertical line) indicate the "BRAF-like" samples. The small hash lines along the bottom show the score of the BRAFmut samples and the small hash along the top are those of KRASmut samples. Figure 5 : Classifiers agreement: The diagrams show the number of samples that are predicted to be either BRAF-like or WT2-like by the three classifiers. For some samples, the three classifiers agree on their predictions, while for others there is no agreement. Figure 6 : Kaplan-Meier plots for the BRAF-like group predicted by mTSP

(Figures 6 A , C, and E) and BRAFmut status (Figures 6 B, D, and F) and the OS, RFS and SAR endpoints, on the PETACC3 data set. Figure 7 : KRASmut samples stratified by the mTSP signature in BRAF-like samples (dashed-line - BRAF high) and non-BRAF-like samples (solid line - BRAF low).

Figure 8 : (PETACC3 data/mTSP) Overall survival: BRAF.hi (predicted) and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D only KRASmut; E only WT2). Figure 9: (PETACC3 data/mTSP) Relapse-free survival: BRAF.hi (predicted) and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D only KRASmut; E only WT2).

Figure 10: (PETACC3 data/mTSP) Survival after relapse: BRAF.hi (predicted) and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D only KRASmut; E only WT2). Figure 11: Kaplan-Meier plots for the BRAF-like group predicted by mTSP (A and C) and BRAFmut status (B and D) and the OS and PFS, on the CETUX data set. Figure 12: (Overall survival) KRASmut samples stratified by the mTSP signature in BRAF-like samples (BRAF high) and non-BRAF-like samples (BRAF low) in the CETUX data set. Figure 13: Kaplan-Meier plots for the BRAF-like group predicted by CCP2 and the OS, RFS and SAR endpoints, on the PETACC3 data set. Figure 14: (PETACC3 data/CCP2) Overall survival: BRAF.hi (predicted) and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D only KRASmut; E only WT2).

Figure 15: (PETACC3 data/CCP2) Relapse-free survival: BRAF.hi (predicted) and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D only KRASmut; E only WT2). Figure 16: (PETACC3 data/CCP2) Survival after relapse: BRAF.hi (predicted) and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D - only KRASmut; E - only WT2). Figure 17: Kaplan-Meier plots for the BRAF-like group predicted by CCP2 and the OS and PFS on the CETUX data set. Figure 18: Overall survival: Population stratification by binarized BRAF score. Figure 19: Overall survival: BRAFhi and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D - only KRASmut; E - only WT2). Figure 20: Relapse-free survival: Population stratification by binarized BRAF score.

Figure 2 1: Survival after relapse: Population stratification by binarized BRAF score. Figure 22: Survival after relapse: BRAFhi and MSI status interaction within different subpopulations (A - whole population; B - no BRAFmut; C - only BRAFmut and KRASmut; D - only KRASmut; E - only WT2).

Detailed Description The present disclosure provides several gene signatures that can be used to predict BRAFmut status, and provides methods, compositions, computer implemented products, detection agents and kits for prognosing or classifying a subject with CRC and for determining the benefit of adjuvant chemotherapy. The term "biomarker" as used herein refers to a gene that is differentially expressed in individuals with CRC according to prognosis and is predictive of different survival outcomes and of the benefit of adjuvant chemotherapy. In some embodiments, a 39-gene pair signature comprises 39 gene pairs listed in Table 3.1 . 1 . In some embodiments, a 32-gene pair signature comprises 32 gene pairs listed in Table 3.1 .2. As used herein the following terms have the following meanings. The term "reference expression profile" refers to the expression of the 39 gene pairs listed in

Table 3.1 . 1 associated with a clinical outcome in a CRC patient. The reference expression profile comprises 78 values, each value representing the expression level of a biomarker, wherein each biomarker corresponds to one gene in Table 3.1 . 1 . The reference expression profile is identified using one or more samples comprising tumor wherein the expression is similar between related samples defining an outcome class or group such as poor survival or good survival and is different to unrelated samples defining a different outcome class such that the reference expression profile is associated with a particular clinical outcome. The reference expression profile is accordingly a reference profile of the expression of the 78 genes in Table 3.1 . 1 , to which the subject expression levels of the corresponding genes in a patient sample are compared in methods for determining or predicting clinical outcome. Similarly, such a reference expression profile can also refer to the 32 gene pairs listed in Table 3.1 .2. As used herein, the term "control" refers to a specific value or dataset that can be used to prognose or classify the value, e.g., expression level or reference expression profile obtained from the test sample associated with an outcome class. In one embodiment, a dataset may be obtained from samples from a group of subjects known to have CRC and good survival outcome or known to have CRC and have poor survival outcome or known to have CRC and have benefited from adjuvant chemotherapy or known to have CRC and not have benefited from adjuvant chemotherapy. The expression data of the biomarkers in the dataset can be used to create a "control value" that is used in testing samples from new patients. A control value is obtained from the historical expression data for a patient or pool of patients with a known outcome. In some embodiments, the control value is a numerical threshold for predicting outcomes, for example good and poor outcome, or making therapy recommendations, for example adjuvant therapy in addition to surgical resection or surgical resection alone.

In some embodiments, the "control" is a predetermined value for the set of 78 biomarkers obtained from CRC patients whose biomarker expression values and survival times are known. Alternatively, the "control" is a predetermined reference profile for the set of 78 biomarkers obtained from CRC patients whose survival times are known. Using values from known samples allows one to develop an algorithm for classifying new patient samples into good and poor survival groups as described in the Examples. Accordingly, in one embodiment, the control is a sample from a subject known to have CRC and good survival outcome. In another embodiment, the control is a sample from a subject known to have CRC and poor survival outcome. A person skilled in the art will appreciate that the comparison between the expression of the biomarkers in the test sample and the expression of the biomarkers in the control will depend on the control used. For example, if the control is from a subject known to have CRC and poor survival, and there is a difference in expression of the biomarkers between the control and test sample, then the subject can be prognosed or classified in a good survival group. If the control is from a subject known to have CRC and good survival, and there is a difference in expression of the biomarkers between the control and test sample, then the subject can be prognosed or classified in a poor survival group. For example, if the control is from a subject known to have CRC and good survival, and there is a similarity in expression of the biomarkers between the control and test sample, then the subject can be prognosed or classified in a good survival group. For example, if the control is from a subject known to have CRC and poor survival, and there is a similarity in expression of the biomarkers between the control and test sample, then the subject can be prognosed or classified in a poor survival group. As used herein, a "reference value" refers to a gene-specific coefficient derived from historical expression data. The multi-gene signatures of the present disclosure comprise gene-specific reference values. In some embodiments, the multi-gene signature comprises one reference value for each gene in the signature. In some embodiments, the multi-gene signature comprises four reference values for each gene in the signature. In some embodiments, the reference values are the first four components derived from principal component analysis for each gene in the signature. The term "differentially expressed" or "differential expression" as used herein refers to a difference in the level of expression of the biomarkers that can be assayed by measuring the level of expression of the products of the biomarkers, such as the difference in level of messenger RNA transcript expressed or proteins expressed of the biomarkers. In a preferred embodiment, the difference is statistically significant. The term "difference in the level of expression" refers to an increase or decrease in the measurable expression level of a given biomarker as measured by the amount of messenger RNA transcript and/or the amount of protein in a sample as compared with the measurable expression level of a given biomarker in a control. In one embodiment, the differential expression can be compared using the ratio of the level of expression of a given biomarker or biomarkers as compared with the expression level of the given biomarker or biomarkers of a control, wherein the ratio is not equal to 1.0. For example, an RNA or protein is differentially expressed if the ratio of the level of expression in a first sample as compared with a second sample is greater than or less than 1.0. For example, a ratio of greater than 1, 1.2, 1.5, 1.7, 2, 3, 5, 10, 15, 20 or more, or a ratio less than 1, 0.8, 0.6, 0.4, 0.2, 0.1, 0.05, 0.001 or less. In another embodiment the differential expression is measured using p-value. For instance, when using p-value, a biomarker is identified as being differentially expressed as between a first sample and a second sample when the p-value is less than 0.1, preferably less than 0.05, more preferably less than 0.01 , even more preferably less than 0.005, the most preferably less than 0.001 . The term "similarity in expression" as used herein means that there is no or little difference in the level of expression of the biomarkers between the test sample and the control or reference profile. For example, similarity can refer to a fold difference compared to a control. In a preferred embodiment, there is no statistically significant difference in the level of expression of the biomarkers. The term "most similar" in the context of a reference profile refers to a reference profile that is associated with a clinical outcome that shows the greatest number of identities and/or degree of changes with the subject profile. The term "prognosis" as used herein refers to a clinical outcome group such as a poor survival group or a good survival group associated with a disease subtype which is reflected by a reference profile such as a biomarker reference expression profile or reflected by an expression level of the biomarkers disclosed herein. The prognosis provides an indication of disease progression and includes an indication of likelihood of death due to CRC. In one embodiment the clinical outcome class includes a good survival group and a poor survival group. The term "prognosing or classifying" as used herein means predicting or identifying the clinical outcome group that a subject belongs to according to the subject's similarity to a reference profile or biomarker expression level associated with the prognosis. For example, prognosing or classifying comprises a method or process of determining whether an individual with CRC has a good or poor survival outcome, or grouping an individual with CRC into a good survival group or a poor survival group. The term "good survival" as used herein refers to an increased chance of survival as compared to patients in the "poor survival" group. For example, the biomarkers of the application can prognose or classify patients into a "good survival group." These patients are at a lower risk of death after surgery. The term "poor survival" as used herein refers to an increased risk of death as compared to patients in the "good survival" group. For example, biomarkers or genes of the application can prognose or classify patients into a "poor survival group." These patients are at greater risk of death from surgery. Accordingly, in one embodiment, the biomarker reference expression profile comprises a poor survival group. In another embodiment, the biomarker reference expression profile comprises a good survival group. The term "subject" as used herein refers to any member of the animal kingdom, preferably a human being, that has CRC or that is suspected of having CRC. CRC patients are classified into stages, which are used to determine therapy. Staging classification testing may include any or all of history, physical examination, routine laboratory evaluations, x-rays, and computed tomography scans or positron emission tomography scans with infusion of contrast materials. As used herein, the term "BRAF mutant-specific inhibitor" refers to a substance that decreases the activity and/or expression of a BRAF mutant protein, but that does not substantially decrease the activity and/or expression of wild type BRAF. Such inhibitors include small molecules, antibodies, and antisense molecules. BRAF mutant proteins include those with mutations as compared with the wild type sequence. A DNA missense mutation leading to a valine to glutamic acid amino acid substitution (V600E) is the most frequent BRAF mutation observed, and functionally the most important involved in the aberrant activation of the MEK-ERK pathway and CRC carcinogenesis.

Other known BRAF mutations include R461 I, I462S, G463E, G463V, G465A, G465E, G465V, G468A, G468E, N580S, E585K, D593V, F594L, G595R, L596V, T598I, V599D, V599E, V599K, V599R, K600E, A727V. Most of such mutations are clustered in two regions: the glycine-rich P loop of the N lobe, and the activation segment and flanking regions. BRAF mutant-specific inhibitors currently in development include, without limitation, compounds such as PLX-4720 (Plexxikon), PLX-4032 (Plexxikon), XL-281 (Exelixis), GSK-21 18436 (Glaxo Smith Kline). As used herein, the term "BRAF mutant-like" refers to a classification of subjects with CRC as predicted by the gene signatures disclosed herein, where subjects with CRC that are classified as "BRAF mutant-like" are those expected to possess at least one BRAF mutation, and/or are expected to respond to adjuvant chemotherapy in a manner that is similar to subjects with CRC who have BRAF mutations and/or possess mutations that result in the aberrant activation of the MEK-ERK pathway and are thus expected to exhibit poor survival when treated with adjuvant chemotherapy. For example, subjects with CRC that have at least one BRAF mutation are generally expected to show a poor response to adjuvant chemotherapy. Furthermore, subjects with CRC that are BRAF mutant-like have a poor survival outcome. As used herein, the term "WT2", or "wild-type" refers to a classification of subjects with CRC as predicted by the gene signatures disclosed herein, where subjects with CRC that are classified as "WT2" or "wild-type" are those expected to be wild type for both BRAF and KRAS genes (i.e. have no mutations in either BRAF or KRAS genes), and/or are expected to respond to adjuvant chemotherapy in a manner that is similar to subjects with CRC who are wild type for both BRAF and KRAS genes. Subjects with CRC that are wild type for both BRAF and KRAS genes are generally expected to show a good response to adjuvant chemotherapy and have a good survival outcome.

In one aspect, a multi-gene signature is prognostic of patient outcome and/or response to adjuvant chemotherapy. The present disclosure provides prognostic signatures that are stage-independent classifiers. In some embodiments, a 39 gene pair or 32 gene pair signature is provided as described herein. In one embodiment, the signature comprises reference values for each of the 39 gene pairs listed in Table 3.1 .1, or the 32 gene pairs listed in Table 3.1 .2. In some embodiments, this gene signature is prognostic of patient outcome and/or response to adjuvant chemotherapy. In some embodiments, the gene pairs listed in Table 3.1 . 1 or Table 3.1 .2 are used in a "top scoring pair" algorithm/method to predict whether or not a patient is classified as "BRAF mutant-like". Table 3.1 . 1 and Table 3.1 .2 lists pairs of genes, where the first gene in the pair is the "Genel" gene, and the second gene in the pair is the "Gene2" gene. In one embodiment, a single gene pair can be analyzed according to the top scoring pair method by comparing the relative gene expression value of a Genel gene in Table

3.1 . 1 or Table 3.1 .2 with the relative gene expression value of the second gene in the pair (i.e. Gene2). If the Genel value of this gene pair is less than the Gene2 value, then the method predicts BRAF mutant-like status. If the Genel value of this gene pair is greater than or equal to the Gene2 value, then the method predicts wild-type ("WT2") status.

Several of the gene pairs shown in Table 3.1 . 1 or Table 3.1 .2 can be used in a similar way, where each of the gene pairs that predict BRAF mutant-like status count as a "vote" for BRAF mutant-like status, so that if there are more "votes" for BRAF mutant like status, then the method would predict BRAF mutant-like status overall. This could be applied, for example, using any number of the gene pairs from Table 3.1 .1, or Table 3.1 .2, for example, less than 39 pairs, less than 30 pairs, less than 25 pairs, less than 20 pairs, less than 15 pairs, less than 10 pairs, less than 5 pairs, or less than 4, 3, or 2 pairs of the genes in Table 3.1 . 1 . Further, for example, for a given number of gene pairs from Table 3.1 . 1 or Table 3.1 .2 that will be used in the prediction method, the average value of all the Genel values can be compared to the average value of all the Gene2 values. Accordingly, if the average Genel value is less than the average Gene2 value, then the method predicts BRAF mutant-like status. For example, as described in Example 3, when using all 39 gene pairs, the average relative expression value of all the Genel genes in Table 3.1 . 1 can be compared to the average relative expression value of all the Gene2 genes in Table 3.1 . 1 . If the average Genel value is less than the average Gene2 value, then the top scoring pair method predicts BRAF mutant-like. Those of skill in the art will recognize that in other embodiments, this method could be applied, for example, using relative expression levels of any number of the gene pairs from Table 3.1 . 1 , for example, less than 39 pairs, less than 30 pairs, less than 25 pairs, less than 20 pairs, less than 15 pairs, less than 10 pairs, less than 5 pairs, or less than 4 , less than 3, or less than 2 pairs. The term "test sample" as used herein refers to any cancer-affected fluid, cell or tissue sample from a subject which can be assayed for biomarker expression products and/or a reference expression profile, e.g., genes differentially expressed in subjects with CRC according to survival outcome. The phrase "determining the expression of biomarkers" as used herein refers to determining or quantifying RNA or proteins expressed by the biomarkers. The term "RNA" includes mRNA transcripts, and/or specific spliced variants of mRNA. The terms "RNA product of the biomarker," "biomarker RNA," or "target RNA" as used herein refers to RNA transcripts transcribed from the biomarkers and/or specific spliced variants. In the case of "protein," it refers to proteins translated from the RNA transcripts transcribed from the biomarkers. The term "protein product of the biomarker" or "biomarker protein" refers to proteins translated from RNA products of the biomarkers. A person skilled in the art will appreciate that a number of methods can be used to detect or quantify the level of RNA products of the biomarkers within a sample, including arrays, such as microarrays, RT-PCR (including quantitative PCR), nuclease protection assays and Northern blot analyses. Any analytical procedure capable of permitting specific and quantifiable (or semi-quantifiable) detection of the genes described here and, optionally, additional biomarkers may be used in the methods herein presented, such as the microarray methods set forth herein, and methods known to those skilled in the art. Accordingly, in one embodiment, the biomarker expression levels are determined using arrays, optionally microarrays, RT-PCR, optionally quantitative RT-PCR, nuclease protection assays or Northern blot analyses.

In some embodiments, the biomarker expression levels are determined by using an array. cDNA microarrays consist of multiple (usually thousands) of different cDNA probes spotted (usually using a robotic spotting device) onto known locations on a solid support, such as a glass microscope slide. Microarrays for use in the methods described herein comprise a solid substrate onto which the probes are covalently or non-covalently attached. The cDNAs are typically obtained by PCR amplification of plasmid library inserts using primers complementary to the vector backbone portion of the plasmid or to the gene itself for genes where sequence is known. PCR products suitable for production of microarrays are typically between 0.5 and 2.5 kB in length. In a typical microarray experiment, RNA (either total RNA or poly A RNA) is isolated from cells or tissues of interest and is reverse transcribed to yield cDNA. Labeling is usually performed during reverse transcription by incorporating a labeled nucleotide in the reaction mixture. A microarray is then hybridized with labeled RNA, and relative expression levels calculated based on the relative concentrations of cDNA molecules that hybridized to the cDNAs represented on the microarray. Microarray analysis can be performed by commercially available equipment, following manufacturer's protocols, such as by using Affymetrix GeneChip technology, Agilent Technologies cDNA microarrays, lllumina Whole-Genome DASL array assays, or any other comparable microarray technology.

In some embodiments, probes capable of hybridizing to one or more biomarker RNAs or cDNAs are attached to the substrate at a defined location ("addressable array"). Probes can be attached to the substrate in a wide variety of ways, as will be appreciated by those in the art. In some embodiments, the probes are synthesized first and subsequently attached to the substrate. In other embodiments, the probes are synthesized on the substrate. In some embodiments, probes are synthesized on the substrate surface using techniques such as photo-polymerization and photolithography.

In some embodiments, microarrays are utilized in a RNA-primed, Array-based Klenow Enzyme ("RAKE") assay. See Nelson, P. T. et al. (2004) Nature Methods 1(2):1-7; Nelson, P. T. et al. (2006) RNA 12(2):1-5, each of which is incorporated herein by reference in its entirety. In these embodiments, total RNA is isolated from a sample. Optionally, small RNAs can be further purified from the total RNA sample. The RNA sample is then hybridized to DNA probes immobilized at the 5'-end on an addressable array. The DNA probes comprise a base sequence that is complementary to a target RNA of interest, such as one or more biomarker RNAs capable of specifically hybridizing to a nucleic acid comprising a sequence that is identically present in one of the genes described in Example 3 under standard hybridization conditions.

In some embodiments, the addressable array comprises DNA probes for no more than the 78 genes listed in Table 3.1 . 1 , or the 64 genes listed in Table 3.1 .2, or the genes listed in Table 3.2, or those listed in Example 3.2.2. In some embodiments, the addressable array comprises DNA probes for each of the 78 genes listed in Table 3.1 .1, or each of the 64 genes listed in Table 3.1 .2, or each of the genes listed in Table 3.2, or each of the genes listed in Example 3.2.2.

In some embodiments, quantitation of biomarker RNA expression levels requires assumptions to be made about the total RNA per cell and the extent of sample loss during sample preparation. In some embodiments, the addressable array comprises

DNA probes for each of the 78 genes listed in Table 3.1 . 1 , or for each of the 64 genes listed in Table 3.1 .2, or the genes listed in Table 3.2, or those listed in Example 3.2.2.

In some embodiments, expression data are pre-processed to correct for variations in sample preparation or other non-experimental variables affecting expression measurements. For example, background adjustment, quantile adjustment, and summarization may be performed on microarray data, using standard software programs such as RMAexpress v0.3, followed by centering of the data to the mean and scaling to the standard deviation. After the sample is hybridized to the array, it is exposed to exonuclease I to digest any unhybridized probes. The Klenow fragment of DNA polymerase I is then applied along with biotinylated dATP, allowing the hybridized biomarker RNAs to act as primers for the enzyme with the DNA probe as template. The slide is then washed and a streptavidin-conjugated fluorophore is applied to detect and quantitate the spots on the array containing hybridized and Klenow-extended biomarker RNAs from the sample.

In some embodiments, the RNA sample is reverse transcribed using a biotin/poly-dA random octamer primer. The RNA template is digested and the biotin- containing cDNA is hybridized to an addressable microarray with bound probes that permit specific detection of biomarker RNAs. In typical embodiments, the microarray includes at least one probe comprising at least 8, at least 9, at least 10, at least 11, at least 12, at least 13, at least 14, at least 15, at least 16, at least 17, at least 18, at least

19, even at least 20, 2 1, 22, 23, or 24 contiguous nucleotides identically present in each of the genes listed in Table 3.1 . 1 or Table 3.1 .2, or each of the genes listed in Table 3.2, or each of the genes listed in Example 3.2.2. After hybridization of the cDNA to the microarray, the microarray is exposed to a streptavidin-bound detectable marker, such as a fluorescent dye, and the bound cDNA is detected. In one embodiment, the array is a U133A chip from Affymetrix. In another embodiment, a plurality of nucleic acid probes that are complementary or hybridizable to an expression product of the genes listed in Table 3.1 . 1 , or Table 3.1 .2, or the genes listed in Table 3.2, or the genes listed in Example 3.2.2, are used on the array. In some embodiments, a plurality of nucleic acid probes that are complementary or hybridizable to an expression product of some or all the genes described in Example 3 are used on the array. The term "nucleic acid" includes DNA and RNA and can be either double stranded or single stranded. The term "hybridize" or "hybridizable" refers to the sequence specific non- covalent binding interaction with a complementary nucleic acid. In a preferred embodiment, the hybridization is under high stringency conditions. Appropriate stringency conditions which promote hybridization are known to those skilled in the art, or can be found in Current Protocols in Molecular Biology, John Wiley & Sons, N.Y. (1989), 6.3.1 6.3.6. For example, 6.OX sodium chloride/sodium citrate (SSC) at about 45°C, followed by a wash of 2.0XSSC at 50°C may be employed. The term "probe" as used herein refers to a nucleic acid sequence that will hybridize to a nucleic acid target sequence. In one example, the probe hybridizes to an RNA product of the biomarker or a nucleic acid sequence complementary thereof. The length of probe depends on the hybridization conditions and the sequences of the probe and nucleic acid target sequence. In one embodiment, the probe is at least 8, 10, 15, 20, 25, 50, 75, 100, 150, 200, 250, 400, 500 or more nucleotides in length.

In some embodiments, compositions are provided that comprise at least one biomarker or target RNA-specific probe. The term "target RNA-specific probe" encompasses probes that have a region of contiguous nucleotides having a sequence that is either (i) identically present in one of the genes described in Example 3, or (ii) complementary to the sequence of a region of contiguous nucleotides found in one of the genes described in Example 3, where "region" can comprise the full length sequence of any one of the genes described in Example 3, a complementary sequence of the full length sequence of any one of the genes described in Example 3, or a subsequence thereof.

In some embodiments, target RNA-specific probes consist of deoxyribonucleotides. In other embodiments, target RNA-specific probes consist of both deoxyribonucleotides and nucleotide analogs. In some embodiments, biomarker RNA-specific probes comprise at least one nucleotide analog which increases the hybridization binding energy. In some embodiments, a target RNA-specific probe in the compositions described herein binds to one biomarker RNA in the sample.

In some embodiments, more than one probe specific for a single biomarker RNA is present in the compositions, the probes capable of binding to overlapping or spatially separated regions of the biomarker RNA. It will be understood that in some embodiments in which the compositions described herein are designed to hybridize to cDNAs reverse transcribed from biomarker RNAs, the composition comprises at least one target RNA-specific probe comprising a sequence that is identically present in a biomarker RNA (or a subsequence thereof).

In some embodiments, a biomarker RNA is capable of specifically hybridizing to at least one probe comprising a base sequence that is identically present in one of the genes described in Example 3. In some embodiments, a biomarker RNA is capable of specifically hybridizing to at least one nucleic acid probe comprising a sequence that is identically present in one of the genes described in Example 3.

In some embodiments, the composition comprises a plurality of target or biomarker RNA-specific probes each comprising a region of contiguous nucleotides comprising a base sequence that is identically present in one or more of the genes described in Example 3, or in a subsequence thereof. As used herein, the terms "complementary" or "partially complementary" to a biomarker or target RNA (or target region thereof), and the percentage of "complementarity" of the probe sequence to that of the biomarker RNA sequence is the percentage "identity" to the reverse complement of the sequence of the biomarker RNA. In determining the degree of "complementarity" between probes used in the compositions described herein (or regions thereof) and a biomarker RNA, such as those disclosed herein, the degree of "complementarity" is expressed as the percentage identity between the sequence of the probe (or region thereof) and the reverse complement of the sequence of the biomarker RNA that best aligns therewith. The percentage is calculated by counting the number of aligned bases that are identical as between the two sequences, dividing by the total number of contiguous nucleotides in the probe, and multiplying by 100.

In some embodiments, the microarray comprises probes comprising a region with a base sequence that is fully complementary to a target region of a biomarker RNA. In other embodiments, the microarray comprises probes comprising a region with a base sequence that comprises one or more base mismatches when compared to the sequence of the best-aligned target region of a biomarker RNA. As noted above, a "region" of a probe or biomarker RNA, as used herein, may comprise or consist of 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 2 1, 22, 23, 24, 25, 26, 27, 28, 29 or more contiguous nucleotides from a particular gene or a complementary sequence thereof. In some embodiments, the region is of the same length as the probe or the biomarker RNA. In other embodiments, the region is shorter than the length of the probe or the biomarker RNA.

In some embodiments, the microarray comprises 78 probes each comprising a region of at least 10 contiguous nucleotides, such as at least 11 contiguous nucleotides, such as at least 13 contiguous nucleotides, such as at least 14 contiguous nucleotides, such as at least 15 contiguous nucleotides, such as at least 16 contiguous nucleotides, such as at least 17 contiguous nucleotides, such as at least 18 contiguous nucleotides, such as at least 19 contiguous nucleotides, such as at least 20 contiguous nucleotides, such as at least 2 1 contiguous nucleotides, such as at least 22 contiguous nucleotides, such as at least 23 contiguous nucleotides, such as at least 24 contiguous nucleotides, such as at least 25 contiguous nucleotides with a base sequence that is identically present in one of the genes described in Table 3.1 . 1 , or Table 3.1 .2.

In another embodiment, the biomarker expression levels are determined by using quantitative RT-PCR. RT-PCR is one of the most sensitive, flexible, and quantitative methods for measuring expression levels. The first step is the isolation of mRNA from a target sample. The starting material is typically total RNA isolated from human tumors or tumor cell lines. General methods for mRNA extraction are well known in the art and are disclosed in standard textbooks of molecular biology, including Ausubel et al., Current Protocols of Molecular Biology, John Wiley and Sons (1997). Methods for RNA extraction from paraffin embedded tissues are disclosed, for example, in Rupp and Locker, Lab Invest. 56:A67 (1987), and De Andres et al., BioTechniques 18:42044

(1995). In particular, RNA isolation can be performed using a purification kit, buffer set and protease from commercial manufacturers, such as Qiagen, according to the manufacturer's instructions. For example, total RNA from cells in culture can be isolated using Qiagen RNeasy mini-columns. Numerous RNA isolation kits are commercially available.

In some embodiments, the primers used for quantitative RT-PCR comprise a forward and reverse primer for each gene listed in Table 3.1 .1, or Table 3.1 .2.

In some embodiments the analytical method used for detecting at least one biomarker RNA in the methods set forth herein includes real-time quantitative RT-PCR. Although PCR can use a variety of thermostable DNA-dependent DNA polymerases, it typically employs the Taq DNA polymerase, which has a 5'-3' nuclease activity but lacks a 3'-5' proofreading endonuclease activity. In some embodiments, RT-PCR is done using a TaqMan™ assay sold by Applied Biosystems, Inc. In a first step, total RNA is isolated from the sample. In some embodiments, the assay can be used to analyze about 10 ng of total RNA input sample, such as about 9 ng of input sample, such as about 8 ng of input sample, such as about 7 ng of input sample, such as about 6 ng of input sample, such as about 5 ng of input sample, such as about 4 ng of input sample, such as about 3 ng of input sample, such as about 2 ng of input sample, and even as little as about 1 ng of input sample containing RNA. The TaqMan™ assay utilizes a stem-loop primer that is specifically complementary to the 3'-end of a biomarker RNA. The step of hybridizing the stem-loop primer to the biomarker RNA is followed by reverse transcription of the biomarker RNA template, resulting in extension of the 3' end of the primer. The result of the reverse transcription step is a chimeric (DNA) amplicon with the step-loop primer sequence at the 5' end of the amplicon and the cDNA of the biomarker RNA at the 3' end. Quantitation of the biomarker RNA is achieved by RT-PCR using a universal reverse primer comprising a sequence that is complementary to a sequence at the 5' end of all stem-loop biomarker RNA primers, a biomarker RNA-specific forward primer, and a biomarker RNA sequence-specific TaqMan™ probe. The assay uses fluorescence resonance energy transfer ("FRET") to detect and quantitate the synthesized PCR product. Typically, the TaqMan™ probe comprises a fluorescent dye molecule coupled to the 5'-end and a quencher molecule coupled to the 3'-end, such that the dye and the quencher are in close proximity, allowing the quencher to suppress the fluorescence signal of the dye via FRET. When the polymerase replicates the chimeric amplicon template to which the TaqMan™ probe is bound, the 5'- nuclease of the polymerase cleaves the probe, decoupling the dye and the quencher so that FRET is abolished and a fluorescence signal is generated. Fluorescence increases with each RT-PCR cycle proportionally to the amount of probe that is cleaved.

In some embodiments, quantitation of the results of RT-PCR assays is done by constructing a standard curve from a nucleic acid of known concentration and then extrapolating quantitative information for biomarker RNAs of unknown concentration. In some embodiments, the nucleic acid used for generating a standard curve is an RNA of known concentration. In some embodiments, the nucleic acid used for generating a standard curve is a purified double-stranded plasmid DNA or a single-stranded DNA generated in vitro.

In some embodiments, where the amplification efficiencies of the biomarker nucleic acids and the endogenous reference are approximately equal, quantitation is accomplished by the comparative Ct (cycle threshold, e.g., the number of PCR cycles required for the fluorescence signal to rise above background) method. Ct values are inversely proportional to the amount of nucleic acid target in a sample. In some embodiments, Ct values of the target RNA of interest can be compared with a control or calibrator, such as RNA from normal tissue. In some embodiments, the Ct values of the calibrator and the target RNA samples of interest are normalized to an appropriate endogenous housekeeping gene (see above).

In addition to the TaqMan™ assays, other RT-PCR chemistries useful for detecting and quantitating PCR products in the methods presented herein include, but are not limited to, Molecular Beacons, Scorpion probes and SYBR Green detection.

In some embodiments, Molecular Beacons can be used to detect and quantitate PCR products. Like TaqMan™ probes, Molecular Beacons use FRET to detect and quantitate a PCR product via a probe comprising a fluorescent dye and a quencher attached at the ends of the probe. Unlike TaqMan™ probes, Molecular Beacons remain intact during the PCR cycles. Molecular Beacon probes form a stem-loop structure when free in solution, thereby allowing the dye and quencher to be in close enough proximity to cause fluorescence quenching. When the Molecular Beacon hybridizes to a target, the stem-loop structure is abolished so that the dye and the quencher become separated in space and the dye fluoresces. Molecular Beacons are available, e.g., from Gene Link™.

In some embodiments, Scorpion probes can be used as both sequence-specific primers and for PCR product detection and quantitation. Like Molecular Beacons, Scorpion probes form a stem-loop structure when not hybridized to a target nucleic acid. However, unlike Molecular Beacons, a Scorpion probe achieves both sequence-specific priming and PCR product detection. A fluorescent dye molecule is attached to the 5'- end of the Scorpion probe, and a quencher is attached to the 3'-end. The 3' portion of the probe is complementary to the extension product of the PCR primer, and this complementary portion is linked to the 5'-end of the probe by a non-amplifiable moiety. After the Scorpion primer is extended, the target-specific sequence of the probe binds to its complement within the extended amplicon, thus opening up the stem-loop structure and allowing the dye on the 5'-end to fluoresce and generate a signal. Scorpion probes are available from, e.g., Premier Biosoft International. In some embodiments, RT-PCR detection is performed specifically to detect and quantify the expression of a single biomarker RNA. The biomarker RNA, in typical embodiments, is selected from a biomarker RNA capable of specifically hybridizing to a nucleic acid comprising a sequence that is identically present in one of the genes described in Example 3. In some embodiments, the biomarker RNA specifically hybridizes to a nucleic acid comprising a sequence that is identically present in at least one of the genes in Table 3.1 .1, or Table 3.1 .2. In other embodiments, the biomarker RNA specifically hybridizes to a nucleic acid comprising a sequence that is identically present in at least one of the genes in Table 3.2 or in Example 3.2.2. In various other embodiments, RT-PCR detection is utilized to detect, in a single multiplex reaction, each of 78 biomarker RNAs. The biomarker RNAs, in some embodiments, are capable of specifically hybridizing to a nucleic acid comprising a sequence that is identically present in one of the 78 genes listed in Table 3.1 . 1 , or Table

3.1 .2. In various other embodiments, RT-PCR detection is utilized to detect, in a single multiplex reaction, RNAs corresponding to each of the biomarkers listed in Table 3.2, or in Example 3.2.2.

In some multiplex embodiments, a plurality of probes, such as TaqMan™ probes, each specific for a different RNA target, is used. In typical embodiments, each target

RNA-specific probe is spectrally distinguishable from the other probes used in the same multiplex reaction.

In some embodiments, quantitation of RT-PCR products is accomplished using a dye that binds to double-stranded DNA products, such as SYBR Green. In some embodiments, the assay is the QuantiTect SYBR Green PCR assay from Qiagen. In this assay, total RNA is first isolated from a sample. Total RNA is subsequently poly- adenylated at the 3'-end and reverse transcribed using a universal primer with poly-dT at the 5'-end. In some embodiments, a single reverse transcription reaction is sufficient to assay multiple biomarker RNAs. RT-PCR is then accomplished using biomarker RNA-specific primers and a miScript Universal Primer, which comprises a poly-dT sequence at the 5'-end. SYBR Green dye binds non-specifically to double-stranded DNA and upon excitation, emits light. In some embodiments, buffer conditions that promote highly-specific annealing of primers to the PCR template (e.g., available in the QuantiTect SYBR Green PCR Kit from Qiagen) can be used to avoid the formation of non-specific DNA duplexes and primer dimers that will bind SYBR Green and negatively affect quantitation. Thus, as PCR product accumulates, the signal from SYBR green increases, allowing quantitation of specific products. RT-PCR is performed using any RT-PCR instrumentation available in the art. Typically, instrumentation used in real-time RT-PCR data collection and analysis comprises a thermal cycler, optics for fluorescence excitation and emission collection, and optionally a computer and data acquisition and analysis software.

In some embodiments, the method of detectably quantifying one or more biomarker RNAs includes the steps of: (a) isolating total RNA; (b) reverse transcribing a biomarker RNA to produce a cDNA that is complementary to the biomarker RNA; (c) amplifying the cDNA from step (b); and (d) detecting the amount of a biomarker RNA with RT-PCR. As described above, in some embodiments, the RT-PCR detection is performed using a FRET probe, which includes, but is not limited to, a TaqMan™ probe, a Molecular beacon probe and a Scorpion probe. In some embodiments, the RT-PCR detection and quantification is performed with a TaqMan™ probe, i.e., a linear probe that typically has a fluorescent dye covalently bound at one end of the DNA and a quencher molecule covalently bound at the other end of the DNA. The FRET probe comprises a base sequence that is complementary to a region of the cDNA such that, when the FRET probe is hybridized to the cDNA, the dye fluorescence is quenched, and when the probe is digested during amplification of the cDNA, the dye is released from the probe and produces a fluorescence signal. In such embodiments, the amount of biomarker RNA in the sample is proportional to the amount of fluorescence measured during cDNA amplification. The TaqMan™ probe typically comprises a region of contiguous nucleotides comprising a base sequence that is complementary to a region of a biomarker RNA or its complementary cDNA that is reverse transcribed from the biomarker RNA template (i.e., the sequence of the probe region is complementary to or identically present in the biomarker RNA to be detected) such that the probe is specifically hybridizable to the resulting PCR amplicon. In some embodiments, the probe comprises a region of at least 6 contiguous nucleotides having a base sequence that is fully complementary to or identically present in a region of a cDNA that has been reverse transcribed from a biomarker RNA template, such as comprising a region of at least 8 contiguous nucleotides, or comprising a region of at least 10 contiguous nucleotides, or comprising a region of at least 12 contiguous nucleotides, or comprising a region of at least 14 contiguous nucleotides, or even comprising a region of at least 16 contiguous nucleotides having a base sequence that is complementary to or identically present in a region of a cDNA reverse transcribed from a biomarker RNA to be detected. Preferably, the region of the cDNA that has a sequence that is complementary to the TaqMan™ probe sequence is at or near the center of the cDNA molecule. In some embodiments, there are independently at least 2 nucleotides, such as at least 3 nucleotides, such as at least 4 nucleotides, such as at least 5 nucleotides of the cDNA at the 5'-end and at the 3'-end of the region of complementarity. In typical embodiments, all biomarker RNAs are detected in a single multiplex reaction. In these embodiments, each TaqMan™ probe that is targeted to a unique cDNA is spectrally distinguishable when released from the probe. Thus, each biomarker RNA is detected by a unique fluorescence signal.

In some embodiments, expression levels may be represented by gene transcript numbers per nanogram of cDNA. To control for variability in cDNA quantity, integrity and the overall transcriptional efficiency of individual primers, RT-PCR data can be subjected to standardization and normalization against one or more housekeeping genes as has been previously described. See, e.g., Rubie et al., Mol. Cell. Probes 19(2):101-9 (2005). Appropriate genes for normalization in the methods described herein include those as to which the quantity of the product does not vary between different cell types, cell lines or under different growth and sample preparation conditions. In some embodiments, endogenous housekeeping genes useful as normalization controls in the methods described herein include, but are not limited to, ACTB, BAT1, B2M, TBP, U6 snRNA, RNU44, RNU 48, and U47. In typical embodiments, the at least one endogenous housekeeping gene for use in normalizing the measured quantity of RNA is selected from ACTB, BAT1 , B2M, TBP, U6 snRNA, U6 snRNA, RNU44, RNU 48, and

U47. In some embodiments, normalization to the geometric mean of two, three, four or more housekeeping genes is performed. In some embodiments, one housekeeping gene is used for normalization. In some embodiments, two, three, four or more housekeeping genes are used for normalization.

In some embodiments, labels that can be used on the FRET probes include colorimetric and fluorescent labels such as Alexa Fluor dyes, BODIPY dyes, such as BODIPY FL; Cascade Blue; Cascade Yellow; coumarin and its derivatives, such as 7- amino-4-methylcoumarin, aminocoumarin and hydroxycoumarin; cyanine dyes, such as Cy3 and Cy5; eosins and erythrosins; fluorescein and its derivatives, such as fluorescein isothiocyanate; macrocyclic chelates of lanthanide ions, such as Quantum Dye™; Marina Blue; Oregon Green; rhodamine dyes, such as rhodamine red, tetramethylrhodamine and rhodamine 6G; Texas Red; fluorescent energy transfer dyes, such as thiazole orange-ethidium heterodimer; and, TOTAB. Specific examples of dyes include, but are not limited to, those identified above and the following: Alexa Fluor 350, Alexa Fluor 405, Alexa Fluor 430, Alexa Fluor 488, Alexa Fluor 500. Alexa Fluor 514, Alexa Fluor 532, Alexa Fluor 546, Alexa Fluor 555, Alexa Fluor 568, Alexa Fluor 594, Alexa Fluor 610, Alexa Fluor 633, Alexa Fluor 647, Alexa Fluor 660, Alexa Fluor 680, Alexa Fluor 700, and, Alexa Fluor 750; amine-reactive BODIPY dyes, such as BODIPY 493/503, BODIPY 530/550, BODIPY 558/568, BODIPY 564/570, BODIPY 576/589, BODIPY 581/591 , BODIPY 630/650, BODIPY 650/655, BODIPY FL, BODIPY R6G, BODIPY TMR, and, BODIPY-TR; Cy3, Cy5, 6-FAM, Fluorescein Isothiocyanate, HEX, 6-JOE, Oregon Green 488, Oregon Green 500, Oregon Green 514, Pacific Blue, REG, Rhodamine Green, Rhodamine Red, Renographin, ROX, SYPRO, TAMRA, 2',4',5',7'-Tetrabromosulfonefluorescein, and TET.

Specific examples of fluorescently labeled ribonucleotides useful in the preparation of RT-PCR probes for use in some embodiments of the methods described herein are available from Molecular Probes (Invitrogen), and these include, Alexa Fluor 488-5-UTP, Fluorescein-12-UTP, BODIPY FL-14-UTP, BODIPY TMR-14-UTP, Tetramethylrhodamine-6-UTP, Alexa Fluor 546-1 4-UTP, Texas Red-5-UTP, and BODIPY TR-1 4-UTP. Other fluorescent ribonucleotides are available from Amersham Biosciences (GE Healthcare), such as Cy3-UTP and Cy5-UTP. Examples of fluorescently labeled deoxyribonucleotides useful in the preparation of RT-PCR probes for use in the methods described herein include Dinitrophenyl (DNP)- r-dUTP, Cascade Blue-7-dUTP, Alexa Fluor 488-5-dUTP, Fluorescein-12-dUTP, Oregon Green 488-5-dUTP, BODIPY FL-14-dUTP, Rhodamine Green-5-dUTP, Alexa Fluor 532-5-dUTP, BODIPY TMR-14-dUTP, Tetramethylrhodamine-6-dUTP, Alexa Fluor 546-14-dUTP, Alexa Fluor 568-5-dUTP, Texas Red-12-dUTP, Texas Red-5-dUTP, BODIPY TR-14-dUTP, Alexa Fluor 594-5-dUTP, BODIPY 630/650-1 4-dUTP, BODIPY 650/665-1 4-dUTP; Alexa Fluor 488-7-OBEA-dCTP, Alexa Fluor 546-1 6-OBEA-dCTP, Alexa Fluor 594-7-OBEA-dCTP, Alexa Fluor 647-1 2-OBEA-dCTP. Fluorescently labeled nucleotides are commercially available and can be purchased from, e.g., Invitrogen.

In some embodiments, dyes and other moieties, such as quenchers, are introduced into nucleic acids used in the methods described herein, such as FRET probes, via modified nucleotides. A "modified nucleotide" refers to a nucleotide that has been chemically modified, but still functions as a nucleotide. In some embodiments, the modified nucleotide has a chemical moiety, such as a dye or quencher, covalently attached, and can be introduced into an oligonucleotide, for example, by way of solid phase synthesis of the oligonucleotide. In other embodiments, the modified nucleotide includes one or more reactive groups that can react with a dye or quencher before, during, or after incorporation of the modified nucleotide into the nucleic acid. In specific embodiments, the modified nucleotide is an amine-modified nucleotide, i.e., a nucleotide that has been modified to have a reactive amine group. In some embodiments, the modified nucleotide comprises a modified base moiety, such as uridine, adenosine, guanosine, and/or cytosine. In specific embodiments, the amine-modified nucleotide is selected from 5-(3-aminoallyl)-UTP; 8-[(4-amino)butyl]-amino-ATP and 8-[(6- amino)butyl]-amino-ATP; N6-(4-amino)butyl-ATP, N6-(6-amino)butyl-ATP, N4-[2,2-oxy- bis-(ethylamine)]-CTP; N6-(6-Amino)hexyl-ATP; 8-[(6-Amino)hexyl]-amino-ATP; 5- propargylamino-CTP, 5-propargylamino-UTP. In some embodiments, nucleotides with different nucleobase moieties are similarly modified, for example, 5-(3-aminoallyl)-GTP instead of 5-(3-aminoallyl)-UTP. Many amine modified nucleotides are commercially available from, e.g., Applied Biosystems, Sigma, Jena Bioscience and TriLink.

In some embodiments, the methods of detecting at least one biomarker RNA described herein employ one or more modified oligonucleotides, such as oligonucleotides comprising one or more affinity-enhancing nucleotides. Modified oligonucleotides useful in the methods described herein include primers for reverse transcription, PCR amplification primers, and probes. In some embodiments, the incorporation of affinity-enhancing nucleotides increases the binding affinity and specificity of an oligonucleotide for its target nucleic acid as compared to oligonucleotides that contain only deoxyribonucleotides, and allows for the use of shorter oligonucleotides or for shorter regions of complementarity between the oligonucleotide and the target nucleic acid.

In some embodiments, affinity-enhancing nucleotides include nucleotides comprising one or more base modifications, sugar modifications and/or backbone modifications.

In some embodiments, modified bases for use in affinity-enhancing nucleotides include 5-methylcytosine, isocytosine, pseudoisocytosine, 5-bromouracil, 5- propynyluracil, 6-aminopurine, 2-aminopurine, inosine, diaminopurine, 2-chloro-6- aminopurine, xanthine and hypoxanthine.

In some embodiments, affinity-enhancing modifications include nucleotides having modified sugars such as 2'-substituted sugars, such as 2'-0-alkyl-ribose sugars, 2'-amino-deoxyribose sugars, 2'-fluoro- deoxyribose sugars, 2'-fluoro-arabinose sugars, and 2'-0-methoxyethyl-ribose (2'MOE) sugars. In some embodiments, modified sugars are arabinose sugars, or d-arabino-hexitol sugars.

In some embodiments, affinity-enhancing modifications include backbone modifications such as the use of peptide nucleic acids (e.g., an oligomer including nucleobases linked together by an amino acid backbone). Other backbone modifications include phosphorothioate linkages, phosphodiester modified nucleic acids, combinations of phosphodiester and phosphorothioate nucleic acid, methylphosphonate, alkylphosphonates, phosphate esters, alkylphosphonothioates, phosphoramidates, carbamates, carbonates, phosphate triesters, acetamidates, carboxymethyl esters, methylphosphorothioate, phosphorodithioate, p-ethoxy, and combinations thereof.

In some embodiments, the oligomer includes at least one affinity-enhancing nucleotide that has a modified base, at least nucleotide (which may be the same nucleotide) that has a modified sugar and at least one internucleotide linkage that is non-naturally occurring.

In some embodiments, the affinity-enhancing nucleotide contains a locked nucleic acid ("LNA") sugar, which is a bicyclic sugar. In some embodiments, an oligonucleotide for use in the methods described herein comprises one or more nucleotides having an LNA sugar. In some embodiments, the oligonucleotide contains one or more regions consisting of nucleotides with LNA sugars. In other embodiments, the oligonucleotide contains nucleotides with LNA sugars interspersed with deoxyribonucleotides. The term "primer" as used herein refers to a nucleic acid sequence, whether occurring naturally as in a purified restriction digest or produced synthetically, which is capable of acting as a point of synthesis when placed under conditions in which synthesis of a primer extension product, which is complementary to a nucleic acid strand is induced (e.g., in the presence of nucleotides and an inducing agent such as DNA polymerase and at a suitable temperature and pH). The primer must be sufficiently long to prime the synthesis of the desired extension product in the presence of the inducing agent. The exact length of the primer will depend upon factors, including temperature, sequences of the primer and the methods used. A primer typically contains 15-25 or more nucleotides, although it can contain less. The factors involved in determining the appropriate length of primer are readily known to one of ordinary skill in the art.

In addition, a person skilled in the art will appreciate that a number of methods can be used to determine the amount of a protein product of the biomarker of the disclosure, including immunoassays such as Western blots, ELISA, and immunoprecipitation followed by SDS-PAGE and immunocytochemistry. Accordingly, in another embodiment, an antibody is used to detect the polypeptide products of the 78 biomarkers listed in Table 3.1 .1, or Table 3.1 .2. In another embodiment, the sample comprises a tissue sample. In a further embodiment, the tissue sample is suitable for immunohistochemistry. The term "antibody" as used herein is intended to include monoclonal antibodies, polyclonal antibodies, and chimeric antibodies. The antibody may be from recombinant sources and/or produced in transgenic animals. The term "antibody fragment"" as used herein is intended to include Fab, Fab', F(ab')2, scFv, dsFv, ds-scFv, dimers, minibodies, diabodies, and multimers thereof and bispecific antibody fragments. Antibodies can be fragmented using conventional techniques. For example, F(ab')2 fragments can be generated by treating the antibody with pepsin. The resulting F(ab')2 fragment can be treated to reduce disulfide bridges to produce Fab' fragments. Papain digestion can lead to the formation of Fab fragments. Fab, Fab' and F(ab')2, scFv, dsFv, ds-scFv, dimers, minibodies, diabodies, bispecific antibody fragments and other fragments can also be synthesized by recombinant techniques. Conventional techniques of molecular biology, microbiology and recombinant DNA techniques are within the skill of the art. Such techniques are explained fully in the literature. See, e.g., Sambrook, Fritsch & Maniatis, 1989, Molecular Cloning: A Laboratory Manual, Second Edition; Oligonucleotide Synthesis (M. J. Gait, ed., 1984); Nucleic Acid Hybridization (B. D. Harnes & S. J. Higgins, eds., 1984); A Practical Guide to Molecular Cloning (B. Perbal, 1984); and a series, Methods in Enzymology

(Academic Press, Inc.); Short Protocols In Molecular Biology, (Ausubel et al., ed., 1995). For example, antibodies having specificity for a specific protein, such as the protein product of a biomarker, may be prepared by conventional methods. A mammal, (e.g., a mouse, hamster, or rabbit) can be immunized with an immunogenic form of the peptide which elicits an antibody response in the mammal. Techniques for conferring immunogenicity on a peptide include conjugation to carriers or other techniques well known in the art. For example, the peptide can be administered in the presence of adjuvant. The progress of immunization can be monitored by detection of antibody titers in plasma or serum. Standard ELISA or other immunoassay procedures can be used with the immunogen as antigen to assess the levels of antibodies. Following immunization, antisera can be obtained and, if desired, polyclonal antibodies isolated from the sera. To produce monoclonal antibodies, antibody producing cells (lymphocytes) can be harvested from an immunized animal and fused with myeloma cells by standard somatic cell fusion procedures thus immortalizing these cells and yielding hybridoma cells. Such techniques are well known in the art, as well as other techniques such as the human B-cell hybridoma technique. Hybridoma cells can be screened immunochemically for production of antibodies specifically reactive with the peptide and the monoclonal antibodies can be isolated.

In some embodiments, recombinant antibodies are provided that specifically bind protein products of the genes described in Example 3 . Recombinant antibodies include, but are not limited to, chimeric and humanized monoclonal antibodies, comprising both human and non-human portions, single-chain antibodies and multi-specific antibodies. A chimeric antibody is a molecule in which different portions are derived from different animal species, such as those having a variable region derived from a murine monoclonal antibody (mAb) and a human immunoglobulin constant region. Single-chain antibodies have an antigen binding site and consist of single polypeptides. They can be produced by techniques known in the art. Multi-specific antibodies are antibody molecules having at least two antigen-binding sites that specifically bind different antigens. Such molecules can be produced by techniques known in the art. Monoclonal antibodies directed against any of the expression products of the genes described in Example 3 can be identified and isolated by screening a recombinant combinatorial immunoglobulin library (e.g., an antibody phage display library) with the polypeptide(s) of interest. Kits for generating and screening phage display libraries are commercially available (e.g., the Pharmacia Recombinant Phage Antibody System, Catalog No. 27-9400-01 ; and the Stratagene SurfZAP Phage Display Kit, Catalog No. 240612). Humanized antibodies are antibody molecules from non-human species having one or more complementarity determining regions (CDRs) from the non-human species and a framework region from a human immunoglobulin molecule. Humanized monoclonal antibodies can be produced by recombinant DNA techniques known in the art.

In some embodiments, humanized antibodies can be produced, for example, using transgenic mice which are incapable of expressing endogenous immunoglobulin heavy and light chains genes, but which can express human heavy and light chain genes. The transgenic mice are immunized in the normal fashion with a selected antigen, e.g., all or a portion of a polypeptide corresponding to a protein product. Monoclonal antibodies directed against the antigen can be obtained using conventional hybridoma technology. The human immunoglobulin transgenes harbored by the transgenic mice rearrange during B cell differentiation, and subsequently undergo class switching and somatic mutation. Thus, using such a technique, it is possible to produce therapeutically useful IgG, IgA and IgE antibodies. Antibodies may be isolated after production (e.g., from the blood or serum of the subject) or synthesis and further purified by well-known techniques. For example, IgG antibodies can be purified using protein A chromatography. Antibodies specific for a protein can be selected or (e.g., partially purified) or purified by, e.g., affinity chromatography. For example, a recombinantly expressed and purified (or partially purified) expression product may be produced, and covalently or non-covalently coupled to a solid support such as, for example, a chromatography column. The column can then be used to affinity purify antibodies specific for the protein products of the genes described in Example 3 from a sample containing antibodies directed against a large number of different epitopes, thereby generating a substantially purified antibody composition, i.e., one that is substantially free of contaminating antibodies. By a substantially purified antibody composition it is meant, in this context, that the antibody sample contains at most only 30% (by dry weight) of contaminating antibodies directed against epitopes other than those of the protein products of the genes described in Example 3, and preferably at most 20%, yet more preferably at most 10%, and most preferably at most 5% (by dry weight) of the sample is contaminating antibodies. A purified antibody composition means that at least 99% of the antibodies in the composition are directed against the desired protein.

In some embodiments, substantially purified antibodies may specifically bind to a signal peptide, a secreted sequence, an extracellular domain, a transmembrane or a cytoplasmic domain or cytoplasmic membrane of a protein product of one of the genes described in Example 3.

In some embodiments, antibodies directed against a protein product of one of the genes described in Example 3 can be used to detect the protein products or fragment thereof (e.g., in a cellular lysate or cell supernatant) in order to evaluate the level and pattern of expression of the protein. Detection can be facilitated by the use of an antibody derivative, which comprises an antibody coupled to a detectable substance. Examples of detectable substances include various enzymes, prosthetic groups, fluorescent materials, luminescent materials, bioluminescent materials, and radioactive materials. Examples of suitable enzymes include horseradish peroxidase, alkaline phosphatase, β-galactosidase, or acetylcholinesterase; examples of suitable prosthetic group complexes include streptavidin/biotin and avidin/biotin; examples of suitable fluorescent materials include umbelliferone, fluorescein, fluorescein isothiocyanate, rhodamine, dichlorotriazinylamine fluorescein, dansyl chloride or phycoerythrin; an example of a luminescent material includes luminol; examples of bioluminescent materials include luciferase, luciferin, and aequorin, and examples of suitable radioactive material include 1251, 131 1, 35S or 3H. A variety of techniques can be employed to measure expression levels of each of the products from the 78 genes shown in Table 3.1 . 1 or the 64 genes shown in Table

3.1 .2 given a sample that contains protein products that bind to a given antibody. Examples of such formats include, but are not limited to, enzyme immunoassay (EIA), radioimmunoassay (RIA), Western blot analysis and enzyme linked immunoabsorbant assay (ELISA). A skilled artisan can readily adapt known protein/antibody detection methods for use in determining protein expression levels of the products of the genes described in Example 3.

In one embodiment, antibodies, or antibody fragments or derivatives, can be used in methods such as Western blots or immunofluorescence techniques to detect the expressed proteins. In some embodiments, either the antibodies or proteins are immobilized on a solid support. Suitable solid phase supports or carriers include any support capable of binding an antigen or an antibody. Well-known supports or carriers include glass, polystyrene, polypropylene, polyethylene, dextran, nylon, amylases, natural and modified celluloses, polyacrylamides, gabbros, and magnetite.

One skilled in the art will know many other suitable carriers for binding antibody or antigen, and will be able to adapt such support for use with the present disclosure. The support can then be washed with suitable buffers followed by treatment with the detectably labeled antibody. The solid phase support can then be washed with the buffer a second time to remove unbound antibody. The amount of bound label on the solid support can then be detected by conventional means. Immunohistochemistry methods are also suitable for detecting the expression levels of the prognostic markers. In some embodiments, antibodies or antisera, including polyclonal antisera, and monoclonal antibodies specific for each marker may be used to detect expression. The antibodies can be detected by direct labeling of the antibodies themselves, for example, with radioactive labels, fluorescent labels, hapten labels such as, biotin, or an enzyme such as horse radish peroxidase or alkaline phosphatase. Alternatively, unlabeled primary antibody is used in conjunction with a labeled secondary antibody, comprising antisera, polyclonal antisera or a monoclonal antibody specific for the primary antibody. Immunohistochemistry protocols and kits are well known in the art and are commercially available. Immunological methods for detecting and measuring complex formation as a measure of protein expression using either specific polyclonal or monoclonal antibodies are known in the art. Examples of such techniques include enzyme-linked immunosorbent assays (ELISAs), radioimmunoassays (RIAs), fluorescence-activated cell sorting (FACS) and antibody arrays. Such immunoassays typically involve the measurement of complex formation between the protein and its specific antibody. Numerous labels are available which can be generally grouped into the following categories:

a. Radioisotopes, such as 36S, 14C, 1251, 3H, and 131 1. The antibody variant can be labeled with the radioisotope using the techniques described in Current

Protocols in Immunology, Vol. 1-2, Coligen et al., Ed., Wiley-lnterscience, New York, Pubs. (1991) for example and radioactivity can be measured using scintillation counting. b. Fluorescent labels such as rare earth chelates (europium chelates) or fluorescein and its derivatives, rhodamine and its derivatives, dansyl, Lissamine, phycoerythrin and Texas Red are available. The fluorescent labels can be conjugated to the antibody variant using techniques well known in the art. Fluorescence can be quantified using a fluorimeter; c. Various enzyme-substrate labels are available and well known to those skilled in the art. The enzyme generally catalyzes a chemical alteration of the chromogenic substrate which can be measured using various techniques. For example, the enzyme may catalyze a color change in a substrate, which can be measured spectrophotometrically. Alternatively, the enzyme may alter the fluorescence or chemiluminescence of the substrate. Techniques for quantifying a change in fluorescence are described above. The chemiluminescent substrate becomes electronically excited by a chemical reaction and may then emit light which can be measured (using a chemiluminometer, for example) or donates energy to a fluorescent acceptor. Examples of enzymatic labels include luciferases (e.g., firefly luciferase and bacterial luciferase, luciferin, 2,3-dihydrophthalazinediones, malate dehydrogenase, urease, peroxidase such as horseradish peroxidase (HRPO), alkaline phosphatase, .beta.-galactosidase, glucoamylase, lysozyme, saccharide oxidases (e.g., glucose oxidase, galactose oxidase, and glucose-6-phosphate dehydrogenase), heterocyclic oxidases (such as uricase and xanthine oxidase), lactoperoxidase, microperoxidase, and the like. Techniques for conjugating enzymes to antibodies are well known in the art.

In some embodiments, a detection label is indirectly conjugated with the antibody. The skilled artisan will be aware of various techniques for achieving this. For example, the antibody can be conjugated with biotin and any of the three broad categories of labels mentioned above can be conjugated with avidin, or vice versa. Biotin binds selectively to avidin and thus, the label can be conjugated with the antibody in this indirect manner. Alternatively, to achieve indirect conjugation of the label with the antibody, the antibody is conjugated with a small hapten (e.g., digoxin) and one of the different types of labels mentioned above is conjugated with an anti-hapten antibody

(e.g., anti-digoxin antibody). In some embodiments, the antibody need not be labeled, and the presence thereof can be detected using a labeled antibody, which binds to the antibody. The 39 gene pair signature described herein can be used to select treatment for CRC patients. As explained herein, the biomarkers can classify patients with CRC into a poor survival group or a good survival group and into groups that might benefit from adjuvant chemotherapy or not. The term "adjuvant chemotherapy" as used herein means treatment of cancer with standard chemotherapeutic agents after surgery where all detectable disease has been removed, but where there still remains a risk of small amounts of remaining cancer. Typical chemotherapeutic agents include cisplatin, carboplatin, vinorelbine, gemcitabine, doccetaxel, paclitaxel and navelbine. Chemotherapeutic agents that are typically used to treat CRC, such as 5-fluorouracil, leucovorin, bevacizumab, cetuximab, panitumumab, and oxaliplatin are known to those in the art.

In yet another aspect, the application also provides for kits used to prognose or classify a subject with CRC into a good survival group or a poor survival group or to select a therapy for a subject with CRC that includes detection agents that can detect the expression products of the biomarkers described herein.

In some embodiments, kits are provided containing antibodies to each of the protein products of the genes described in Example 3, conjugated to a detectable substance, and instructions for use. In some embodiments, the kits comprise antibodies to the protein products of the 78 genes (39 gene pairs) listed in Table 3.1 .1, or the 64 genes listed in Table 3.1 .2. Kits may comprise an antibody, an antibody derivative, or an antibody fragment, which binds specifically with a marker protein, or a fragment of the protein. Such kits may also comprise a plurality of antibodies, antibody derivatives, or antibody fragments wherein the plurality of such antibody agents binds specifically with a marker protein, or a fragment of the protein.

In some embodiments, kits may comprise antibodies such as a labeled or label- able antibody and a compound or agent for detecting protein in a biological sample; means for determining the amount of protein in the sample; means for comparing the amount of protein in the sample with a standard; and instructions for use. Such kits can be supplied to detect a single protein or epitope or can be configured to detect one of a multitude of epitopes, such as in an antibody detection array. Arrays are described in detail herein for nucleic acid arrays and similar methods have been developed for antibody arrays.

In some aspects, a multi-gene signature is provided for prognosis or classifying patients with CRC. In some embodiments, a 39-gene pair signature is provided as described in Example 3, comprising reference values for each of the 78 genes based on relative expression data from a historical data set with a known outcome, such as good or poor survival, and/or known treatment, such as adjuvant chemotherapy.

In one aspect, relative expression data from a patient are combined with the gene-specific reference values on a gene-by-gene basis for each of genes being assessed, to generate a test value which allows prognosis or therapy recommendation. In some embodiments, relative expression data are subjected to an algorithm that yields a single test value, or combined score, which is then compared to a control value obtained from the historical expression data for a patient or pool of patients.

In some embodiments, the control value is a numerical threshold for predicting outcomes, for example good and poor outcome, or making therapy recommendations for a subject, for example adjuvant chemotherapy in addition to surgical resection or surgical resection alone. In some embodiments, a test value or combined score greater than the control value is predictive, for example, of a good outcome or benefit from adjuvant chemotherapy, whereas a combined score falling below the control value is predictive, for example, of a poor outcome or lack of benefit from adjuvant chemotherapy for a subject. In another embodiment, the test value or combined score can be used to predict BRAFT mutant-like status, as described herein.

In a further aspect, the application provides computer programs and computer implemented products for carrying out the methods described herein. Accordingly, in one embodiment, the application provides a computer program product for use in conjunction with a computer having a processor and a memory connected to the processor, the computer program product comprising a computer readable storage medium having a computer mechanism encoded thereon, wherein the computer program mechanism may be loaded into the memory of the computer and cause the computer to carry out the methods described herein.

In another embodiment, the application provides a computer implemented product for predicting a prognosis or classifying a subject with CRC comprising: a. a means for receiving values corresponding to a subject expression profile in a subject sample; and b. a database comprising a reference expression profile associated with a prognosis, wherein the subject biomarker expression profile and the biomarker reference profile each has 78 values, each value representing the expression level of a biomarker, wherein each biomarker corresponds to one gene in Table 3.1 . 1 ; wherein the computer implemented product selects the biomarker reference expression profile most similar to the subject biomarker expression profile, to thereby predict a prognosis or classify the subject. In yet another embodiment, the application provides a computer implemented product for determining therapy for a subject with CRC comprising: a. a means for receiving values corresponding to a subject expression profile in a subject sample; and b. a database comprising a reference expression profile associated with a therapy, wherein the subject biomarker expression profile and the biomarker reference profile each has 78 values, each value representing the expression level of a biomarker, wherein each biomarker corresponds to one gene in Table 3.1 .1; wherein the computer implemented product selects the biomarker reference expression profile most similar to the subject biomarker expression profile, to thereby predict the therapy. Another aspect relates to computer readable mediums such as CD-ROMs. In one embodiment, the application provides computer readable medium having stored thereon a data structure for storing a computer implemented product described herein.

In one embodiment, the data structure is capable of configuring a computer to respond to queries based on records belonging to the data structure, each of the records comprising: a. a value that identifies a biomarker reference expression profile of the 78 genes in Table 3.1 .1; b. a value that identifies the probability of a prognosis associated with the biomarker reference expression profile.

In another aspect, the application provides a computer system comprising a. a database including records comprising a biomarker reference expression profile of the 78 genes in Table 3.1 . 1 associated with a prognosis or therapy; b. a user interface capable of receiving a selection of gene expression levels of the 78 genes in Table 3.1 . 1 for use in comparing to the biomarker reference expression profile in the database; and c. an output that displays a prediction of prognosis or therapy according to the biomarker reference expression profile most similar to the expression levels of the 78 genes.

In some embodiments, the application provides a computer implemented product comprising a. a means for receiving values corresponding to relative expression levels in a subject, of the 39 gene pairs in Table 3.1 .1; b. an algorithm for calculating the top scoring pair method as described herein using the relative expression levels of the 39 gene pairs as shown in Table 3.1 . 1 ; c. an output that displays the combined score; and, optionally, d. an output that displays a prognosis or therapy recommendation based on the combined score. A more complete understanding of the present disclosure can be obtained by reference to the following specific examples. These examples are described solely for the purpose of illustration and are not intended to limit the scope of the overall disclosure. Changes in form and substitution of equivalents are contemplated as circumstances might suggest or render expedient. Although specific terms have been employed herein, such terms are intended in a descriptive sense and not for purposes of limitation.

In the Examples below, and as used in other parts of the specification, the following terms have the following meanings: "BRAFhi" is an indicator variable (C {0,1}) that is obtained by binarizing the BRAF score at 0.5 level; "BRAFhi. t" is an indicator variable (C {0,1}) that is obtained by binarizing the BRAF score at t/100-level, for example, BRAFhi.80 = 1 if and only if BRAF score ≥ 0.80; "BRAF-like" refers to samples with a high BRAF score (≥ 0.5); "BRAFmut" refers to samples with mutation of BRAF, as determined by RT-PCR; it is also the indicator variable for the BRAF mutation status; "BRAF score" is the score produced by the classifier, which can be interpreted as a posteriori probability (C {0,1}); "HR" is a hazard ratio; "KM" means Kaplan-Meier; "KRASmut" means samples with mutation of KRAS, as determined by RT-PCR; it is also the indicator variable, for the KRAS mutation status; "MSI/MSS" means microsatellite instable/stable; "OS" means overall survival; "RFS" means relapse-free survival; "SAR" means survival after relapse; and "WT2" means double wild-type, i.e. no BRAFmut nor KRASmut. Furthermore, the names used to identify specific genes referenced herein (e.g. in the Examples below, and throughout the specification) follow the Hugo Gene nomenclature system (i.e. as stored in the Hugo Gene Nomenclature Committee database - see www.genenames.org), and will be well understood by those of skill in the art. In some instances, GenelD numbers are also provided, which is the gene identification method of the Entrez Gene database (operated by the National Center for Biotechnology Information - see www.ncbi.nlm.nih.gov), and is also well understood by those of skill in the art.

EXAMPLES Example 1: Data used to generate the models The BRAF signatures described herein were built by modeling a binary classification problem (BRAF mutants vs. non-BRAF and non-KRAS mutants, i.e. WT2) using three different classification algorithms: (multiple) top scoring pairs, compound covariate predictor and AdaBoost. While the signatures were derived from a dataset consisting solely of BRAF mutants and WT2 samples, they have been applied to the full population of patients, in contradiction somehow with the usual modeling paradigm which requires a representative data set for classifier training. Nevertheless, this exercise allows the identification of a larger subpopulation of patients with a consistent gene expression pattern, which is generically called a "BRAF-like" subpopulation. The modeling set consisted of gene expression data from tumor samples from phase 1 and 2 of the PETACC3 study, and were either BRAFmut (all V600E mutants) or WT2. The PETACC3 study was an international, randomized clinical study that involved comparison of infused irinotecan + 5-fluorouracil/folinic acid (5-FU/FA) versus 5-FU/FA in patients with stage II and stage III colon cancer. One important feature of the PETACC3 study was the coordinated collection of formalin fixed, paraffin embedded (FFPE) colon cancer tumor samples. RNA from 1378 FFPE colon cancer samples was extracted for expression profiling on the Affymetrix-based platform Colorectal Cancer Disease Specific Array (DSA™) developed by Almac Diagnostics. The KRASmut were discarded from the modeling phase. To reduce batch effects as much as possible, the data from the two phases was aligned using the 45 bridging samples. All models were assessed in two steps on the modeling set: in a first step one of the phases (either one or two, called training set) was used to estimate the performance parameters and their variance (by repeated 5-fold cross-validation) and the other phase (called validation set) was used as an independent validation set on which the performance parameters were simply measured. Then the two data sets were swapped and the processed repeated all over again. We label the two processes as Phase 1 vs 2 (phase 1 for training, phase 2 for independent validation) and Phase 2 vs 1 (phase 2 for training, phase 1 for independent validation), respectively. To avoid biasing the performance estimation, the WT2 and BRAFmut bridging samples are always considered in the training set, and not in the validation set (also for enriching the number of BRAFmut samples in the training set). Table 1. 1 shows the samples sizes for the training and validation sets used in the modeling phase.

Table 1.1 : Sample sizes for the training and validation sets, on the modeling data.

Different independent data sets were used for external independent validation of the BRAF signatures. However, not all models can be assessed on totally different platforms. For example, the AdaBoost model (see Example 3.1 .3) cannot be applied to data that originates from a different platform than the one used for creating the model, due to its sensitivity to exact numerical values. On the other hand, mTSP (see Example 3.1 .1) can be applied on any data set which is feasible to assume that the relative ordering of the genes expression do not change from the modeling set. The following data sets were used for external validation: 1. Cetuximab data set (CETUX): 2 . Kim data set (KIM - Kim et al., Carcinogenesis 27:392-404 (2006)), which consists of 20 tissue samples of CRC of which 9 are KRASmut and 11 are BRAFmut (9 V600E mutants, one D594G mutant, and one G464V mutant). We considered 2 versions of this data set: one containing all the samples (KIM (all)) and one from which the non-V600E BRAF mutants are discarded (KIM (V600E)), because those mutants were considered to be more 'KRAS-like', by the authors of the paper.

Performance parameters For a binary classification problem, the following parameters are used to describe the performance: sensitivity (Se), also called the true positive fraction, gives the proportion of "positive" samples (BRAFmut in the present case) that are correctly classified; specificity (Sp) gives the proportion of negative samples (WT2 in the present case) that are correctly classified; error rate (Err) gives the proportion of samples that are misclassified; area under the ROC curve (AUC) - this parameter is the most descriptive and summarizes the performance of the classifier across all possible values of the threshold and is independent of the prevalence of any of the classes in the population. However, a proper estimation needs a classifier that outputs a score and not simply a binary decision. Nevertheless, even for the binary decisions, the AUC can be approximated by AUC = 0.5 x (Se + Sp)

Example 2 : Differentially Expressed Genes The differential gene expression (BRAF mutants vs. WT2) was assessed using a multivariate linear model (and the limma R package (Gordon, Stat. App. Gen. and Mol.

Biol. 3 (2004); Smyth et al., Bioinform. 2 1:2067-2075 (2005)). The linear model used adjusted for the effects of KRAS mutation status and for the known interactions with MSI status, and had the form: gene expression ~ BRAF +KRAS+MSI+BRAF * MSI+KRAS * MSI

In this model, a gene was called differentially expressed if the adjusted p-value associated with the BRAF variable was below the preselected false discovery rate. The analysis was carried out on the full set of samples (pooled data), at probeset-level, not gene-symbol level. While there are roughly 1,000 probesets differentially expressed at

FDR=0.001, Table 2.1 . 1 provides only the top 100 differentially expressed genes (ordered by the corresponding adjusted p-value).

Table 2.1.1 Top 100 differentially expressed genes Log-fold Adjusted P- Gene symbol Probeset ID change value AQP5 ADXCRPD.7995.C1_x_at 4.07 2.27E-038 F5 ADXCRAG_M14335_s_at 1.47 4.95E-018 REG4 ADXCRIH.384.C1_s_at 3.21 5.32E-017 HSF5 ADXCRSS.Hs#S29881 80_at -2.34 1.68E-016 CTSE ADXCRAG_AJ25071 7_s_at 3.57 9.55E-016 GGH ADXCRIH.546.C1_at -2.25 2.01 E-015 TM4SF4 ADXCRAD_BM825250_s_at 1.26 5.89E-015 CDX2 ADXCRAG_BC0 1446 1_x_at - 1 .80 7.03E-015 LYZ ADXCRIH.1305.C1_s_at 2.36 9.13E-015 Log-fold Adjusted P- Gene symbol Probeset ID change value RNF43 ADXCRPDRC.4289.C1_at - 1 .63 8.31 E-014 TFCP2L1 ADXCRPDRC.8321 .C1_s_at - 1 .63 7.96E-013

MIRN142 ADXCRPD.15182.C1_at - 1 .28 1. 1 1E-012 VNN1 ADXCRAD BP299698 s at 1.26 2.65E-012 C13orf18 ADXCRAD_NM_0251 13_s_at -2.1 1 2.97E-012 RPUSD1 ADXCRPD.7300.C1_s_at 1.88 4.93E-012 ANXA10 ADXCRAD_CK8231 69_at 1.08 4.99E-012 SYT13 ADXCRPD.12823.C1_s_at 0.96 5.07E-012 SATB2 ADXCRPD.10016.C1_at -2.31 8.90E-012 PIWIL1 ADXCRAG_BC028581_x_at 1.97 9.60E-012 SOX8 ADXCRAG_AK024491_s_at 2.31 1.90E-01 1 EPDR1 ADXCRPD.4010.C1_s_at - 1.52 1.97E-01 1 VAV3 ADXCRAD_NM_0061 13_s_at - 1 .71 2.38E-01 1 AMACR ADXCRAG_AF047020_s_at - 1 .27 3.10E-01 1 PLAGL2 ADXCRPD.16547.C1_at - 1 .03 5.00E-01 1 ARID3A ADXCRAD_BP38951 1_at - 1 .48 6.12E-01 1 DPP4 ADXCRAD_BX1 10831_s_at 0.79 7.18E-01 1 TSPAN6 ADXCRIH.1064.C1_at - 1.42 7.20E-01 1 INPP5J ADXCRAG_U45975_s_at - 1 .30 1.80E-010 LOC100130716 ADXCRAD_XM_1 68585_s_at -2.32 2.21 E-010 SPINK1 ADXCRIH.4080.C1_s_at -2.54 2.34E-010 KLK10 ADXCRPD.7217.C1_at 1.44 2.73E-010 C1orf67 ADXCRAD_AI07681 0_s_at - 1 .44 2.91 E-010 SLC14A1 ADXCRAD_BU664688_s_at 1.21 3.58E-010 CCDC1 13 ADXCRAD_CN256031_s_at - 1 .51 3.61 E-010 GRM8 ADXCRAD_BG1 98589_at - 1 .76 4.25E-010 MUC12 ADXCRAD_XM_1 68585_at - 1 .86 4.64E-010 PPP1 R14C ADXCRAG_AF407165_at -2.53 9.85E-010 AGRN ADXCRAG_XM_3721 95_s_at 0.56 9.96E-010 C1orf225 ADXCRAD_BM71 821 6_s_at 1.52 1.15E-009 CEACAM5 ADXCRPD.1 1630.C1_at - 1 .33 1.30E-009 KIAA0802 ADXCRAG_XM_03 1357_s_at 1.02 1.73E-009 C10orf99 ADXCRIH.1562.C1_s_at - 1 .49 1.99E-009 PLLP ADXCRPD.6652.C1_s_at 0.91 2.26E-009 Log-fold Adjusted P- Gene symbol Probeset ID change value SPRED1 ADXCRPD.3092.C1_at 0.84 3.31 E-009 OSBPL1A ADXCRAD_CN36621 1_s_at 0.64 3.34E-009 AIFM3 ADXCRAG_BC032485_s_at -2.21 3.68E-009 LOC1 001 34361 ADXCRAD_XM_086879_at - 1 .34 4.05E-009 TRPM6 ADXCRAG_NM_01 7662_s_at - 1 .54 4.55E-009 TNNC2 ADXCRAG_M33772_s_at - 1 .87 6.25E-009 DPYSL2 ADXCRPD.15073.C1_s_at 0.98 6.32E-009 POFUT1 ADXCRPDRC.5238.C1_at -0.92 6.33E-009 AXIN2 ADXCRPDRC.1943.C1_at - 1 .49 6.34E-009 B3GALT5 ADXCRAG_NM_006057_s_at 0.94 7.26E-009 TP53RK ADXCRAD_BX460824_s_at -0.92 1.02E-008 ANTXR2 ADXCRPD.1 1175.C1_x_at 0.79 1.41 E-008

MUC3B ADXCRPD.9207.C1 s at - 1 .53 1.44E-008 PCLO ADXCRAD_XM_374484_at -2.08 1.47E-008 SLC26A2 ADXCRIH.2831 .C1_s_at - 1 .70 1.78E-008 C 11orf9 ADXCRAD_CK005805_s_at 1.40 2.42E-008 GDA ADXCRAG_NM_004293_s_at 1.18 2.59E-008 AK3L1 ADXCRPDRC.963.C1_at - 1 .21 3.14E-008 CELP ADXCRAD_NM_001 808_s_at - 1 .93 3.40E-008 MYRIP ADXCRAG_AF396687_s_at - 1 .39 3.49E-008 XKR9 ADXCRSS.Hs#S5955802_at 0.68 3.98E-008 ATP5EP2 ADXCRPD.7790.C1_at -0.68 4.20E-008 TSPAN8 1_RDCR049_C08_at 0.69 4.73E-008 DPEP1 ADXCRPD.6962.C1_x_at - 1 .27 6.94E-008 ACSF2 ADXCRPD.6142.C1_at - 1 .33 7.16E-008 MEGF6 ADXCRSS.Hs#S3733570_at 1.01 7.20E-008 TMEM56 ADXCRIH.578.C1_s_at -0.96 7.85E-008 MLPH ADXCRPD.1 115.C1_s_at 1.54 8.13E-008 TFAP2A ADXCRAD_BM719170_at 1.28 8.46E-008 WDR18 ADXCRPD.470.C1_x_at - 1 .43 9.22E-008 CRIP1 ADXCRPD.5224.C1_x_at 1.17 9.32E-008 DUSP4 ADXCRAD_BM852899_at 1.23 1.00E-007 TIMM8AP ADXCRAD_AI09251 1_at 0.58 1.18E-007 CABLES1 ADXCRAD_BX1 08451_s_at 0.80 1.29E-007 Log-fold Adjusted P- Gene symbol Probeset ID change value PABPC1 L ADXCRPD.4612.C1_s_at - 1 .25 1.43E-007 CBFA2T2 ADXCRAD_BE21 891 6_s_at -0.84 2.37E-007 FLJ32063 ADXCRAD_BX1 19160_s_at - 1 .03 2.44E-007 SLC19A3 ADXCRAG_AF28331 7_s_at -0.77 2.93E-007 ASAP2 ADXCRIH.1590.C1_s_at 0.59 3.43E-007 RASGRF2 ADXCRPD. 15964. C1_at 1.12 3.50E-007 ZNRF3 ADXCRAG_BC021 570_at -0.88 3.82E-007 GPR143 ADXCRPD.3734.C1_s_at -0.80 4.05E-007 EPHA4 ADXCRAG_BC016981_s_at 0.52 4.17E-007 C20orf1 12 ADXCRPDRC.3468.C1_s_at -0.86 4.22E-007 PPP1 R14D ADXCRPD.7272.C1_s_at -0.87 4.25E-007 SLC5A6 ADXCRAG_BC015631_s_at - 1.05 4.30E-007 BST2 ADXCRIH.286.C1_s_at 0.82 4.49E-007 TM9SF4 ADXCRAG_BC022850_s_at -0.80 4.53E-007 NT5C3L ADXCRPD.10192.C1_s_at -0.90 4.62E-007 TRERF1 ADXCRAD_CV572327_x_at 0.67 4.75E-007 GNG4 ADXCRPD. 1216.C1_s_at -0.91 4.75E-007 KLK7 ADXCRAG_NM_005046_s_at 1.64 4.80E-007 KLK1 1 ADXCRAD_NM_1 44947_s_at 0.91 4.92E-007 GCNT3 ADXCRPD. 15673. C1_at 1.67 4.93E-007 PIPOX ADXCRAD_BM690859_at - 1 .55 4.97E-007

PLCB4 ADXCRPDRC.13802.C1 at - 1 .17 5.03E-007 CLDN2 ADXCRAG NM 020384 at 1.78 5.38E-007

Further analyses, using some additional samples not included previously, resulted differentially expressed genes shown in Table 2.1 .2.

Table 2.1.2

Gene Gene name Log-Fold change Adjusted P-value AQP5 Aquaporin 5 5.5 9.7E-2 1 CTSE cathepsin E 4,8 5.43E-1 1 SRY (sex determining SOX8 2,6 0,00021 7 region Y)-box 8 Gene Gene name Log-Fold change Adjusted P-value Regenerating islet-derived REG4 3,7 1,68E-07 family, member 4 PIWIL1 piwi-like 1 2,5 9,43E-05 AXIN2 Axin 2 -2, 1 5,65E-06 CDX2 caudal type homeobox 2 -2,3 1.16E-1 0 heat shock transcription HSF5 -3, 1 5.79E-1 0 factor family member 5 transcription factor CP2- TFCP2L1 -2,3 4,5E-08 like 1 gamma-glutamyl GGH -3,0 8,54E-08 hydrolase RNF43 ring finger protein 43 -2,2 1,35E-07 protein tyrosine PTPRO phosphatase, receptor -3,2 5,65E-06 type, 0 serine peptidase inhibitor, SPINK1 -2,9 9,73E-06 Kazal type 1 SATB2 SATB homeobox 2 -2,5 1.17E-05 DPEP1 -2,4 2.76E-05 TNNC2 -2,2 9,21 E-05 PCLO -2,5 0,0001 7

Example 3: Braf Signatures

3.1 Introduction

3.1 .1 Top Scoring Pairs - and extensions

Let X be a data matrix with variables by columns (in the present case a gene expression matrix, with genes by columns, samples by rows). The top scoring pairs

(TSPs) method (Geman et al., Stat. Appl . Genet. Mol. Biol. 3, Article 19 (2004)) seeks a pair of variables i, j such that X k < X for all samples k labeled as (positive class) and k > k for all samples k labeled as "0". While in a real life situation, there is no pair of variables to provide a perfect classification, the method ranks the pairs according to the proportion of erroneous predictions they make. Those top ranking pairs are usually considered for making the predictions. However, the practice demonstrates that the top scoring pairs share a lot of common variables, so their predictions are correlated . To increase the prediction power of the classifier, we propose to combine several TSPs (or their decisions). To this end, we filtered the list of TSPs in such a way that each variable (gene) appears only once in the list on either side of the inequality and then we average the values of these variables, to produce a new prediction rule, as follows. From a list of TSPs of the form

X.,/?< X.,/ → class + 1 → X.,/ < X.,i 4 class + 1 we produce the rule ∑ → X.,ik < X.,// class + 1 /

This rule is referred to as e a- TSP (mTSP) herein. The main advantage of using (m)TSP consists in its applicability across platforms and technologies. Its drawbacks come from the fact that it does not produce a score and its somewhat reduced performance, depending on the particular problem.

3.1 .2 Compound Covariate Predictor - and extensions Compound covariate predictor (CCP - Radmacher et al., J. Comput. Biol. 9:505-

5 1 1 (2002)) is another simple classification rule that, in contrast with TSPs, builds a score which is used for making the final prediction. The score for the sample k has the form ∑ ck = tiXk,i / where t, is some coefficient. In its original form, CCP proposes to use the t-statistic to rank all variables (genes), and use the corresponding statistic as the coefficient in the sum above. Only the top m variables are used in the sum, with m to be tuned via some cross-validation process, for example. For making a prediction, a threshold Co must be chosen. The simplest choice is to take

where lo and l+ are the indexes of the samples in class "0" and "+1", respectively. While this version works fairly well in practice, several extensions can be imagined to either improve robustness (for cross-platform applicability) or to improve the variable (gene) ranking/selection process. By convention, we divide the variables (genes) into two groups: those positively associated with class (they have a positive sign for the ranking statistics), and those negatively associated with class Then, the following extensions are immediate: • instead of using a coefficient specific to each variable, just average all variables - either globally or within the groups of positively and negatively associated variables. The sign is, nevertheless, preserved. • instead of using t-statistics to rank the genes, use a linear model of the form X.,, ~ class indicator + other covariates. and rank the genes by the p-values corresponding to the class indicator. Again, for the coefficient one can choose any of the options above. For example, in the case of BRAFm prediction, the linear model used was

X., / ~ BRAFm + MSI + site, so the variable selection is done with adjustment for MSI status and tumor site. The version of CCP used reported herein, called CCP2, uses the linear model for gene ranking (with adjustment for MSI status and tumor site) and takes the averages of positively and negatively associated genes separately:

I positively -associated -genes 2 negatively -associated -genes

3.1 .3 AdaBoost method Boosting refers to a general class of methods that produce accurate decision rules by combining rough and slightly better than chance base rules (weak learners). Boosting proceeds by repeatedly training the weak learners on different distributions over the training set. For a given sample, the final prediction is obtained by combining the predictions of the individual weak learners. Different combination approaches can be attempted, but usually a simple weighted majority voting scheme is adopted. Even though the early versions of the boosting algorithm were provably converging to an improved classification rule (with respect to the performance of any of the weak learners), they suffered from serious practical drawbacks. The first practically usable version of boosting was AdaBoost, introduced in 1995 by Freund and Schapire. The version of AdaBoost used in developing the BRAF-gene signature fits a generalized linear model using the boosting algorithm based on univariate linear models as weak learners (Buhlmann and Yu, J. Amer. Stat. Assoc. 98:324-339 (2003)). This algorithm is implemented in the R package mboost available from http://stat.ethz.ch/CRAN/. There are a number of advantages in using AdaBoost, particularly the version mentioned above:

· the algorithm produces a sparse classifier - in the sense that the number of variables (genes in the present case) in the final model is small when compared with the initial dimensionality of the feature space; • the classification rule is robust - this is a consequence of the fact that AdaBoost algorithms converge to a large margin classifier (maximize the separation between classes); • by adopting univariate weak learners, AdaBoost will implicitly perform a variable selection as well (selecting those genes that contribute most to the discrimination between classes); • AdaBoost is resistant to overfitting, meaning that there is a high probability that the training performance will be reproduced on other independent data sets.

One thing must be stressed: the model produced is minimalistic, in the sense that not all genes that could be included in the model are considered. Rather, the minimal set of genes that lead to a good classifier is selected. This means that other genes that are correlated with those in the model could also be considered. However, this strategy would not lead to an improved classification performance and the model would become redundant.

3.2 Signatures 3.2.1 mTSP The individual TSP predicts "BRAFm" when Genel

Further analyses, using some additional samples not included previously, resulted in a 32 gene pair meta-TSP signature as shown below in Table 3.1 .2. Table 3.1.2

Gene 1 Gene2 1 C 13orf1 8 CTSE 2 DDC AQP5 3 PPP1 R 14D REG4 4 HSF5 RSBN 1L 5 SATB2 RASSF6 6 TNNC2 CRIP1 7 GGH PPPDE2 8 SPINK1 PLK2 9 PTPRO TM4SF4 10 ZSWIM 1 MLPH 1 1 RNF43 RBM8A 12 CELP SOX8 13 CBFA2T2 PIWIL1 14 PTPRD LOC3881 99 15 CDX2 S 100A1 6 16 TSPAN6 RBBP8 17 VAV3 OSBP2 18 CFTR KLK1 0 19 PHYH DUSP4 20 PLCB4 HOXD3 2 1 ZNF1 4 1 C 11orf9 22 PPP1 R 14C CD55 23 FLJ32063 TRNP1 Gene 1 Gene2 24 APCDD1 FSCN 1 25 ACOX1 KIAA0802 26 C 10orf99 PLLP 27 MIR1 42 IRX3 28 ARID3A SLC25A37 29 C20orf 111 PIK3AP1 30 AMACR TPK1 3 1 AIFM3 ZIC2 32 CTTNBP2 SERPINB5

3.2.2 CCP2 As mentioned before, CCP2 takes the difference between the average of positively associated genes and the average of the negatively associated genes with BRAFm, from a linear model (see Example 3.1 .2). One has to choose the number of genes to be included in the model. A sensitivity analysis on the modeling set was performed using phase 1 data as training and phase 2 as validation and then swapping the two data sets, and varying the number of selected genes from 10 to 300, in increments of 10. While the performance varied slightly with the number of selected genes, this variability remained limited. In Figures 1 A and B and 2 A and B, the AUC and error rates obtained are presented. The final model contains 100 genes, which are provided below: • positively associated genes: ABLIM3, ANXA1, ANXA10, AP1S3, AQP5, B3GALT2, B3GALT5, BST2, CABLES1, CD109, CRIP1, CTSE, DCBLD2, DNAH2, DPP4, DPYSL2, DUSP4, EPHA4, EPHB6, F5, GABRE, GDA, GPR126, HCRP1 ,

HOXB2, INSM1 , KIAA0802, KLK1 1, KLK6, KLK7, LYZ, MLPH, NT5E, PFKP, PIWIL1 , PKM2, PLLP, PMAIP1 , PON3, PRDM16, REG4, SERPINB5, SLC14A1, SLC1A1 , SMCHD1, SOX13, SOX2, SOX8, SPRR1A, SPRR1 B, STS, TFAP2A, TIMM8AP1 , TM4SF4, TRERF1, TRNP1, UBASH3B, VNN1 , XKR9 · negatively associated genes: AIFM3, AK3L1 , AMACR, APCDD1, ARID3A, AXIN2, CCDC1 13, CDHR1, CDX2, DDC, DPEP1, EPDR1 , FAM84A, GGH, GPR143, GPR160, GRM8, GUCY2C, H2AFY2, HSF5, INPP5D, INPP5J, LDLRAD3, MUC12, MUC3B, MYRIP, PARM1, PCLO, POFUT1 , PPP1 R14C, PPP1 R14D, PTPRO, RNF43, SATB2, SEMA5A, SPINK1, SUPT4H1, TM9SF4, TSPAN6, VAV3, ZNF518B

3.2.3 AdaBoost The AdaBoost signature contains 29 genes which are combined through a weighted mean. Table 3.2 lists these genes and the corresponding coefficients. Table 3.2 The AdaBoost model.

GenelD Symbol Coeff 1 6690 SPINK1 -0.1989 2 9221 1 CDHR1 -0.1559 3 83998 REG4 0.1 191 4 362 AQP5 0.1 131 5 6698 SPRR1A 0.0855 6 4885 NPTX2 -0.0852 7 30812 SOX8 0.0549 8 5473 PPBP -0.0538 9 2918 GRM8 -0.0528 10 412 STS 0.0489 11 145447 ABHD12B -0.0488 12 181 1 SLC26A3 -0.0462 13 7546 ZIC2 0.0459 14 146336 LOC146336 0.0415 15 6699 SPRR1 B 0.0413 16 4923 NTSR1 0.0340 17 55328 RNLS -0.031 1 18 54749 EPDR1 -0.0250 19 57415 C3orf14 -0.0219 20 100134361 LOC100134361 -0.0182 2 1 5650 KLK7 0.0170 22 51268 PIPOX -0.0157 23 3223 HOXC6 0.0094 24 84959 UBASH3B 0.0081 25 5366 PMAIP1 0.0074 26 5800 PTPRO -0.0058 GenelD Symbol Coeff 27 4291 MLF1 0.0039 28 1670 DEFA5 -0.0034 29 5968 REG1 B 0.0024

3.3 Internal validation of signatures The signature development process has been validated in two stages, using one data batch as a training/modeling set and the other one as an independent validation set. The 45 bridging samples were always considered in the training set (to keep the number of BRAFmut samples at a reasonable level), and their replicates have been removed from the validation set. On the training set, the performance of the classifier has been estimated by repeated (10 times) stratified 5-fold cross validation. The same performance parameters (area under the ROC curve - AUC, sensitivity, specificity and error rate) were measured on the validation sets. Table 3.3 lists these performance measures. The main criterion for judging the performance of the classifiers was the AUC as it is independent of the classifier threshold and of the prevalence of BRAF mutations. Note that this is only a subset of the full PETACC3 data set, which contains only BRAFmut and WT2, the KRASmut being discarded.

Table 3.3: Estimated and validation performance of the BRAF classifiers. For the estimated parameters, the standard deviation of the estimates are given between parentheses. T - train, V - validation; Ph. 1 - phase 1 data, Ph. 2 - phase 2 data. The pooled estimates correspond to the results of repeated cross-validation on the pooled data.

Train/ Valid AUC Sensitivity Specificity Error rate 88.40% 85.36% 14.12% T: Ph. 1 0.869 0.948 (14.07) (7.86) (6.78) V : Ph. 2 100.0% 89.67% 9.69 88.40% 89.69% 10.44% mTSP T: Ph. 2 0.890 0.892 (15.00) (4.76) (4.59) V : Ph. 1 92.86% 85.45% 13.71% 92.50% 87.77% 11.74% Pooled 0.901 (8.94) (4.12) (3.71) 3.4 External validation of signatures Only the CCP2 and mTSP signatures are susceptible to work on other platforms than those on which they were built. The AdaBoost signature is bound to the platform on which the model was produced. Although CCP2 requires a threshold to produce the labels (which is platform-dependent), it can still be applied on various other platforms, but its performance must be judged only by AUC (threshold-independent). To keep in mind when interpreting the results: • the models were built to discriminate between BRAFmut and WT2; there were no KRASmut in the modeling set; • KIM data set contains only KRASmut and BRAFmut; 2 out of 11 (18.18%)

BRAFmut are not V600E mutants (as were those in the modeling set) and they are always classified as non-BRAFmut; • CETUX data set originates from an Almac platform, as the one which generated the modeling set - that is why the AdaBoost classifier could be applied as well;

• CCP2 uses a threshold that is tuned on the modeling set; this threshold is not portable across platforms and that is why only AUC is given for this classifier • mTSP predictions were made with only 13 pairs (out of 39) due to missing genes on the KIM dataset Table 3.4: External validation of the different models. The results marked with a star are approximations of the real AUC.

3.5 BRAF score distributions The scores produced by the AdaBoost classifier can be interpreted as a posteriori probability that a sample belongs to the category "BRAF mutants", so a score of at least 0.5 can be considered as predicting the "BRAF mutants" class. Or, as it will be called later on, "BRAF-like samples". While the models have been constructed without taking into account the KRAS mutants, they were applied to the whole population, including the KRASes. Figure 3 shows the distribution of BRAF scores as well as the scores for KRASmut (small hashes along the top) and BRAFmut (small hashes along the bottom) samples. Note that all the BRAFmut samples have a high BRAF score (≥ 0.5). Also, there are 96 KRASmut samples out of 248, which have a high BRAF score (see Table 3.5 for details).

Table 3.5: AdaBoost Stratification of BRAF scores by mutation status.

BRAF score BRAFmut KRASmut WT2 BRAF high 4 1 96 42 BRAF low 0 152 298 On the other hand, CCP2 does not produce a posteriori probabilities, but a real value that is to be thresholded to produce the final label. This real value (the difference between the average expression level of positively and negatively associated genes, respectively) can be used as a surrogate for a score. The distribution of these values is shown in Figure 4 along with the scores of the BRAF mutants and KRAS mutants. There are two BRAF mutant samples that are misclassified by CCP2, one of them harboring a K601 E mutation and the other one the 'classical' V600E mutation. See also Table 3.6 for details.

Table 3.6: CCP2 stratification of BRAF scores by mutation status.

3.6 Comparison of predictions by the three classifiers While the performance parameters of the three classifiers are roughly equivalent, each of them has its own strengths and weaknesses. The Venn diagrams in Figure 5 show the overlap between the predictions (agreement of classifiers), for both the BRAF- like samples (those predicted to be BRAF mutants) and WT2-like samples (those predicted to be WT2). Note that the figures do not necessarily add up to those in the clinical table, because of the missing values (even if the BRAF/KRAS status is missing in the clinical table, the sample's status was predicted). We can further stratify the samples by their BRAF/KRAS status to see each classifier's preferences. Table 3.7 shows such a stratification for the common predictions and classifier-specific predictions. For example, intersection of all three classifier stands for the common predictions made by the three classifiers (the intersection of the three sets in Figure 5). Taking the row BRAF-like/intersection of all three as an example, one can see that out of the 126 samples that were predicted to be "BRAF-like" by all three classifiers (Figure 5), 25 are actually WT2, 36 are BRAF mutants and 56 are KRAS mutants, respectively (9 have missing values). Similarly, the row BRAF-like/mTSP shows that out of 14 samples that are predicted to be BRAF-like solely by mTSP, 4 are actually WT2 and 10 are KRAS mutants respectively, and so forth. Table 3.7: Stratification of the predictions made by the three classifiers: common and classifier-specific predictions.

Example 4 : Survival Analysis 4.1 Gene-wise univariate analyses In this section we simply assess the statistical significance of each of the genes selected in the signatures with respect to their capacity to model overall survival (OS), relapse-free survival (RFS) and survival after relapse (SAR). Cox proportional hazards models are employed and the p-values and hazard ratios are reported for each of the genes. The p-values are not adjusted for multiple testing. Each signature is analyzed separately and all its corresponding genes are listed, even though there are genes that are repeated in each signature.

4.1 .1 mTSP signature The univariate analyses for the 78 genes (39 pairs) in the mTSP signature are given in Table 4.1 .

Table 4.1 : Hazard rations (HR) and p-values for the 78 genes in the mTSP signature. OS RFS SAR Gene HR p-value HR p-value HR p-value AMACR 0.759074 0.000722 0.8891 80 0.077870 0.81 3820 0.002669 ANP32E 0.993829 0.953706 0.9291 35 0.4221 84 1.107437 0.248244 ANXA1 1.194339 0.0 19506 1.108505 0 .118909 1.187781 0.008244 APCDD 1 1.01 1976 0.81 43 17 1.046970 0.289266 0.946524 0.223520 AQP5 1.065522 0 .138348 0.979870 0.625296 1.13331 5 0.002084 ARID3A 0.97651 4 0.7 12791 1.032744 0.552308 0.847283 0.007934 ASPHD2 0.91 3907 0.457890 0.81 5255 0.051 926 1.044 158 0.701 053 CCDC1 13 0.759774 0.0000 12 0.8382 17 0.001 040 0.867007 0.008989 CCDC56 0.655098 0.000037 0.7367 18 0.000744 0.71 0 177 0.000748 CD55 1.138029 0.047873 1.044060 0.432369 1.14561 1 0.027358 CDX2 0.839908 0.000669 0.934687 0 .164696 0.778608 0.000001 CELP 0.934963 0 .137288 1.0 114 15 0.761 537 0.882329 0.001 031

CRIP1 1.081 349 0.258736 1.0 14 100 0.81 5023 1.175387 0.01 5704 CTSE 1.040962 0 .198035 0.9931 35 0.804060 1.074 130 0.01 1275 CYP4F2 0.970872 0.576600 1.0401 18 0.357507 0.867677 0.001 681 DPEP1 0.96341 4 0.5741 87 1.0781 97 0 .170482 0.894459 0.045753

DPP4 1.063 18 1 0 .138558 1.033895 0.346706 1.033349 0.3351 43 DUSP4 1.055373 0.383004 1.001 065 0.983946 1.09541 5 0.095239 EPDR1 0.958408 0.505291 1.023628 0.667766 0.886263 0.046942 FAM84A 0.873787 0.084929 0.956789 0.51 1838 0.884573 0.068570 FLJ23867 1.037426 0.634071 1.116663 0.085763 0.968377 0.59661 9 FSCN 1 1.067861 0.336202 0.996680 0.953727 1.220631 0.001 4 11 FUT8 0.91 1963 0.2736 13 0.850633 0.023629 1.220007 0.01 1093 GGH 0.833575 0.000466 0.897208 0.0 13533 0.887023 0.007099 GNG4 0.965531 0.683367 1.070530 0.3431 2 1 0.893757 0 .12 1575 GPR1 60 0.906843 0 .128825 0.979221 0.7026 16 0.90941 4 0 .10521 2 GRM8 0.849902 0.022081 0.9 14437 0 .123485 0.81 3905 0.002349 GUCY2C 0.948688 0.385435 0.9973 17 0.959904 0.889945 0.037220 HEPH 0.907534 0.206448 0.985275 0.8241 59 0.85381 0 0.030337 HSF5 0.880988 0.006229 0.971 788 0.490682 0.820950 0.000002 IFT52 0.869604 0 .154256 0.940783 0.459983 0.849 186 0.060800 KIAA0802 1.155620 0 .112572 1.021 792 0.784838 1.124320 0 .145340 LDLRAD3 0.888478 0 .123067 0.939502 0.338006 0.91 501 0 0.2251 20 LM04 1.189495 0.060034 1.055665 0.491 373 1.243096 0.01 0835 OS RFS SAR Gene HR p-value HR p-value HR p-value LSM7 0.983847 0.882237 0.935364 0.478460 1.054322 0.604924 LYPLA1 0.87061 4 0 .104542 0.940632 0.398505 0.91 3033 0 .199888 LYZ 1.037584 0.373932 1.000349 0.9921 42 1.122 158 0.00391 1

MALL 1.029591 0.7 10 195 1.036597 0.588273 1.01 16 14 0.878504 MAP3K5 1.003794 0.963268 0.936950 0.341 428 1.186521 0.025234 MEGF6 0.996246 0.943752 0.972325 0.537979 1.063554 0 .17721 6 MLPH 1.130788 0.0 17551 1.081 469 0.075207 1.134625 0.006565 MUC1 2 0.87701 7 0.009283 0.9341 27 0 .114455 0.882433 0.006 190

MYRIP 0.854320 0.0 13 1 11 0.965635 0.51 4853 0.809667 0.000303 PCTP 0.743664 0.021 626 0.881 145 0.238765 0.787532 0.031 600 PIN4 0.972783 0.652297 1.076756 0 .1600 11 0.863862 0.004064 PKM2 1.10371 9 0.480483 0.996628 0.9771 93 1.263841 0.065897

PLK2 1.104036 0 .198052 1.0561 88 0.3986 19 1.196036 0.01 119 1 PLLP 1.123741 0.245703 0.981 086 0.831 081 1.255523 0.006585 P0FUT1 0.879065 0.089546 0.968652 0.6 12847 0.885235 0.061 4 16

PPP1 R 14C 0.933526 0.045321 0.992239 0.797096 0.906457 0.002887 PPP1 R 14D 0.895082 0.2 18894 0.984250 0.836757 0.856302 0.033034

PTPRO 0.871 3 13 0.000546 0.9 16871 0.009726 0.920625 0.022089 QSOX1 1.100855 0.3083 11 1.071 244 0.390655 1.162947 0.06271 9 RAB26 1.03661 5 0.7231 35 1.027328 0.756324 1.036250 0.656771 RASSF6 1.204860 0.024594 1.122079 0 .102738 1.182388 0.01 4732

RBBP8 0.973 107 0.8051 85 0.827032 0.048632 1.21 9093 0.055071 REG4 1.048030 0 .18 112 1 0.993976 0.847632 1.0981 96 0.004958 RNF43 0.803797 0.0001 92 0.897388 0.042545 0.856799 0.001 527 S 100A1 6 1.01 1868 0.902990 0.999442 0.994562 1.09051 2 0.329960

SATB2 0.879355 0.001 637 0.9431 3 1 0 .107627 0.876323 0.000283 SEMA5A 0.885409 0.024470 0.955283 0.327092 0.881 056 0.00981 0 SMCHD1 0.88561 4 0.490233 0.855678 0.292335 1.31 051 1 0.089404 SOX 13 1.126294 0.3347 17 1.070240 0.51 6693 1.258041 0.05651 5 SPINK1 0.936081 0.078244 0.976927 0.473389 0.932000 0.023642 SPRED1 1.059666 0.596464 0.994338 0.951 386 1.237983 0.029843 TBC 1D8 1.139950 0.228368 1.039978 0.671 127 1.229 105 0.037039 TM4SF4 1.189333 0.000259 1.104730 0.023585 1.126832 0.007 163 TNNC2 0.955299 0.261 048 1.030227 0.396237 0.88431 7 0.000464 OS RFS SAR Gene HR p-value HR p-value HR p-value TP53RK 0.866625 0.132189 1.015147 0.854809 0.806361 0.005695 TPK1 0.816910 0.180659 0.821094 0.122846 0.865062 0.305323 TRNP1 1.181406 0.004930 1.098287 0.068505 1.144291 0.016540 TSPAN6 0.848330 0.012430 0.971464 0.608883 0.816825 0.000158 VAV3 0.823752 0.008269 0.967883 0.589143 0.760513 0.000076 XKR9 1.095569 0.313464 0.948245 0.527779 1.158396 0.076050 ZNF518B 0.880308 0.184318 1.058794 0.467954 0.808457 0.006073

The predictive power of each individual pair of genes was also estimated by constructing a new set of variables of the form d = Gene 2 - Genei. The results are shown in Table 4.2.

Table 4.2: Hazard ratios (HR) and p-values for the 39 pairs in the mTSP signature. From each pair, a new variable is constructed as the difference between the two genes.

OS RFS SAR Difference HR p-value HR p-value HR p-value LYPLA1-GGH 1.2534459 0.0010314 1.1590868 0.0106207 1.2252445 0.0027639 AQP5-PTPRO 1.0919755 0.0007171 1.0361712 0.121 1483 1.1402320 0.0000042 REG4- 1.0837008 0.0025540 1.0314297 0.1897915 1.1 139664 0.0001082 CCDC1 13 CTSE- 1.0414463 0.1333025 0.9962422 0.8758005 1.1026721 0.0003238 PPP1R14D LM04-HSF5 1.1 181697 0.0028084 1.0283001 0.3987142 1.2185076 0.0000002 DUSP4-GRM8 1.0842092 0.0470370 1.0332763 0.3408591 1.1507371 0.0013195 RBBP8- 1.1297166 0.0359132 0.9675698 0.5184169 1.3759971 0.0000003 TSPAN6 RASSF6- 1.1523124 0.0000895 1.0740050 0.0261665 1.1919828 0.0000018 SATB2 CRIP1-TNNC2 1.0465923 0.1582336 0.9843564 0.5704062 1.1659913 0.0000174 MLPH-MUC12 1.1202864 0.0007898 1.0680848 0.0230445 1.1387998 0.0003660 MALL-CDX2 1.1347794 0.0044204 1.0570062 0.1579294 1.2234772 0.0000058 TBC1D8-VAV3 1.1708764 0.0058146 1.031 1399 0.5293330 1.3360842 0.0000028 S100A16- 1.1482330 0.0037658 1.0678105 0.1 199634 1.1536007 0.0012956 RNF43 ANXA1-EPDR1 1.1 122661 0.0366551 1.0294241 0.51 18602 1.2039764 0.0005607 OS RFS SAR Difference HR p-value HR p-value HR p-value

LSM7-CELP 1.0560865 0 .1877028 0.9809036 0.5768485 1.1566 158 0.0007974 LYZ-PIN4 1.03322 11 0.3357549 0.9779082 0.43939 17 1.1341 4 11 0.0003624 SPRED 1- 1.1248048 0.0737450 1.0220826 0.69 14 117 1.229891 1 0.00 16650 POFUT1 KIAA0802- 1.22391 29 0.0024448 1.07553 11 0 .19771 04 1.2861 6 12 0.00021 8 1 ACOX1 PLK2-SPINK1 1.0709370 0.0367 193 1.0284068 0.3237235 1.113351 1 0.0008255 TRNP1 - 1.14 11827 0.0027688 1.0755800 0.05441 50 1.1574420 0.0022457 LDLRAD3 CD55- 1.0777298 0.008951 4 1.01 49708 0.5580232 1.133231 6 0.0000358 PPP1 R 14C SMCHD1 -PCTP 1.1605628 0 .1425769 1.0272059 0.7539837 1.3997527 0.0023022 TPK1 -AMACR 1.1577333 0.0254243 1.0431 116 0.4407091 1.1938752 0.0084244 AGR2-GUCY2C 1.0032258 0.9382656 0.97293 13 0.4459302 1.0694452 0 .1289237

PLLP-MYRIP 1.1160892 0.01 77857 1.01 5 1075 0.7093447 1.1937869 0.0001 954 FSCN 1- 1.01 5 1403 0.7085034 0.9697070 0.3692627 1.122001 2 0.0051 660 APCDD1 TM4SF4-GNG4 1.1257204 0.0031 938 1.0450995 0.2254462 1.1473762 0.0003651 RAB26-ARID3A 1.0232534 0.6438379 0.9871 875 0.7606257 1.0977248 0.0634740 FUT8-AK3L1 1.025551 4 0.6085967 0.96561 73 0.41 27725 1.1404701 0.0099008 XKR9-CYP4F2 1.0440546 0.3382601 0.9620422 0.2939290 1.175061 3 0.0006332 SOX1 3-IFT52 1.175451 3 0.0580465 1.0801 871 0.2847481 1.2781 237 0.0077750 ANP32E- 1.0831 204 0.2750790 0.95561 97 0.4700689 1.2596667 0.00 11774 TP53RK QSOX1 -DPEP1 1.0550599 0.31 43480 0.97 1951 5 0.5247499 1.1937 187 0.00 15984 FLJ23867- 1.0741 366 0 .1446725 1.0593436 0 .1687958 1.053281 9 0.2709038 GPR1 60 PKM2-CCDC56 1.37871 18 0.0002362 1.2224768 0.0086323 1.3727521 0.0002984 MEGF6- 1.0399873 0.3704800 0.99546 19 0.9031 479 1.10351 78 0.0266460 FAM84A ASPHD2- 1.0408367 0.5722 13 1 0.8993359 0.0824427 1.240841 9 0.0023261 ZNF51 8B MAP3K5- 1.1167793 0.0320731 1.01 33378 0.7654574 1.2751 488 0.00001 72 SEMA5A DPP4-HEPH 1.081 7938 0.0429931 1.0335324 0.3249376 1.0938932 0.01 74591

4.1 .2 CCP2 signature The univariate analyses for the 100 genes in the CCP2 signature are shown Table 4.3. Table 4.3 : Hazard ratios (H R) and p-values for the 100 genes in the CCP signature.

OS RFS SAR Gene HR p-value HR p-value HR p-value ABLIM3 1.497490 0.000047 1.303227 0.002872 1.3841 03 0.000431 AIFM3 0.904398 0.01 5891 0.937492 0.06581 6 0.896340 0.0051 67 AK3L1 0.898902 0 .107229 0.960577 0.4831 79 0.9227 14 0 .179480 AMACR 0.759074 0.000722 0.8891 80 0.077870 0.81 3820 0.002669 ANXA1 1.194339 0.01 9506 1.108505 0 .118909 1.187781 0.008244 ANXA1 0 1.134066 0 .134587 0.974327 0.771 200 1.465642 0.000033 AP1 S3 1.113324 0 .147471 1.0371 04 0.572433 1.124993 0.071 436

APCDD1 1.01 1976 0.81 43 17 1.046970 0.289266 0.946524 0.223520 AQP5 1.065522 0 .138348 0.979870 0.625296 1.1333 15 0.002084 ARID3A 0.97651 4 0.71 2791 1.032744 0.552308 0.847283 0.007934 AXIN2 0.942664 0.307607 1.037246 0.462772 0.836303 0.001 240 B3GALT2 1.01 1082 0.921 7 19 1.084372 0.37636 1 0.925781 0.420204 B3GALT5 1.30950 1 0.003539 1.072206 0.4 12556 1.330909 0.000872 BST2 1.049824 0.61 8907 1.00271 3 0.97433 1 1.128929 0 .152963 CABLES1 1.233657 0.0471 50 1.14 170 1 0 .145670 1.136693 0 .118884 CCDC1 13 0.759774 0.0000 12 0.83821 7 0.001 040 0.867007 0.008989

CD109 1.255953 0.0047 16 1.16 1989 0.032736 1.21 251 6 0.004030 CDHR1 0.882343 0.01 5251 0.94443 1 0 .172553 0.903573 0.0 15955 CDX2 0.839908 0.000669 0.934687 0 .164696 0.778608 0.000001 CRIP1 1.081 349 0.258736 1.0 14 100 0.81 5023 1.175387 0.0 15704 CTSE 1.040962 0 .198035 0.9931 35 0.804060 1.0741 30 0.0 11275

DCBLD2 1.377086 0.000231 1.1921 8 1 0.023270 1.2261 88 0.007232 DDC 0.873088 0.007056 0.9 14 186 0.040853 0.91 2274 0.0371 27 DNAH2 1.12 1460 0 .1351 74 1.059443 0.384370 1.137699 0.057982 DPEP1 0.96341 4 0.5741 87 1.0781 97 0 .170482 0.894459 0.045753

DPP4 1.0631 8 1 0 .138558 1.033895 0.346706 1.033349 0.3351 43 DPYSL2 1.305706 0.005342 1.206497 0.020454 1.156984 0.06081 7 DUSP4 1.055373 0.383004 1.001 065 0.983946 1.0954 15 0.095239 EPDR1 0.958408 0.505291 1.023628 0.667766 0.886263 0.046942 EPHA4 1.349624 0.051 575 1.282245 0.063399 1.160280 0 .184862 EPHB6 1.196603 0.094872 1.2 13457 0.032779 1.055901 0.565234 F5 1.2691 11 0.002292 1.160289 0.03451 0 1.244789 0.001 560 OS RFS SAR Gene HR p-value HR p-value HR p-value FAM84A 0.873787 0.084929 0.956789 0.51 1838 0.884573 0.068570 GAB RE 1.196660 0.041 797 1.139479 0.082907 1.132824 0 .126430 GDA 1.100078 0.204991 1.0 10734 0.869262 1.154380 0.031 994 GGH 0.833575 0.000466 0.897208 0.0 13533 0.887023 0.007099

GPR1 26 1.079345 0.4521 65 0.948483 0.564600 1.21 0875 0.064309 GPR1 43 0.830655 0.076539 0.96452 1 0.6741 3 1 0.739 189 0.001 583 GPR1 60 0.906843 0 .128825 0.97922 1 0.70261 6 0.90941 4 0 .1052 12 GRM8 0.849902 0.022081 0.9 14437 0 .123485 0.81 3905 0.002349 GUCY2C 0.948688 0.385435 0.99731 7 0.959904 0.889945 0.037220 H2AFY2 0.794496 0.028053 0.9 17900 0.331 243 0.939388 0.4707 10 HCRP1 1.444374 0.0071 68 1.145086 0.259057 1.279577 0.024342 HOXB2 1.276536 0.0 18571 1.11670 1 0.2 19808 1.234947 0.0 17446 HSF5 0.880988 0.006229 0.971 788 0.490682 0.820950 0.000002 INPP5D 0.898273 0.250249 0.974600 0.7521 02 0.894546 0 .158873 INPP5J 0.757369 0.000435 0.878427 0.050404 0.733601 0.000006 INSM 1 1.0900 11 0.395273 0.93833 1 0.5481 80 1.284272 0.007707 KIAA0802 1.155620 0 .112572 1.021 792 0.784838 1.124320 0 .145340

KLK1 1 1.1754 13 0.049403 1.0 1471 0 0.847447 1.277231 0.000735 KLK6 1.2451 11 0.000038 1.2 113 14 0.000045 1.063 122 0 .169059 KLK7 1.166322 0.000290 1.116986 0.003296 1.132638 0.001 365 LDLRAD3 0.888478 0 .123067 0.939502 0.338006 0.91 501 0 0.2251 20 LYZ 1.037584 0.373932 1.000349 0.9921 42 1.122 158 0.0039 11 MLPH 1.130788 0.0 17551 1.081 469 0.075207 1.134625 0.006565 MUC 12 0.8770 17 0.009283 0.9341 27 0 .114455 0.882433 0.0061 90 MUC3B 0.926527 0 .179849 0.997926 0.9671 74 0.828397 0.001 302

MYRIP 0.854320 0.0 13 1 11 0.965635 0.51 4853 0.809667 0.000303 NT5E 0.991 194 0.881 7 16 0.9601 45 0.4 10964 1.134883 0.026849 PARM1 0.932707 0.201 808 1.004359 0.928395 0.868928 0.0 15853 PCLO 0.940934 0 .12681 7 0.953603 0 .1641 05 0.950 175 0 .14351 5 PFKP 1.163223 0 .147553 1.0 18043 0.842848 1.21 6293 0.051 877 PIWIL1 1.051 252 0.337307 1.0401 66 0.376955 1.01 771 7 0.686420 PKM2 1.1037 19 0.480483 0.996628 0.9771 93 1.263841 0.065897 PLLP 1.123741 0.245703 0.981 086 0.831 081 1.255523 0.006585 PMAIP1 0.9971 73 0.979369 0.9 16793 0.3671 22 1.070 145 0.451 7 13 OS RFS SAR Gene HR p-value HR p-value HR p-value POFUT1 0.879065 0.089546 0.968652 0.6 12847 0.885235 0.061 4 16 PON3 1.134636 0.224728 1.117593 0.2 16849 1.166422 0.060832

PPP1 R 14C 0.933526 0.045321 0.992239 0.797096 0.906457 0.002887 PPP1 R 14D 0.895082 0.2 18894 0.984250 0.836757 0.856302 0.033034 PRDM 16 1.05851 0 0.746084 0.962358 0.798247 1.086322 0.577336 PTPRO 0.871 3 13 0.000546 0.9 1687 1 0.009726 0.920625 0.022089 REG4 1.048030 0 .18 112 1 0.993976 0.847632 1.0981 96 0.004958 RNF43 0.803797 0.0001 92 0.897388 0.042545 0.856799 0.001 527 SATB2 0.879355 0.001 637 0.9431 3 1 0 .107627 0.876323 0.000283 SEMA5A 0.885409 0.024470 0.955283 0.327092 0.881 056 0.00981 0 SERPINB5 1.028221 0.720547 0.981 598 0.780908 1.080777 0.324920 SLC1 4A1 1.0842 13 0.3 12425 1.05531 2 0.4441 03 1.051 996 0.495986 SLC 1A1 1.01 24 14 0.9056 13 0.999092 0.991 97 1 1.01 7 134 0.834070 SMCHD1 0.8856 14 0.490233 0.855678 0.292335 1.31 051 1 0.089404 SOX 13 1.126294 0.3347 17 1.070240 0.51 6693 1.258041 0.05651 5 SOX2 1.14 12 17 0.000045 1.079902 0.006649 1.112738 0.000222 SOX8 1.068053 0 .119300 1.02461 2 0.51 7469 1.094492 0.0 13953 SPINK1 0.936081 0.078244 0.976927 0.473389 0.932000 0.023642 SPRR1A 1.000828 0.992821 0.941 992 0.470270 1.175592 0.075792

SPRR1 B 1.076602 0.28051 3 1.00481 2 0.939228 1.18 1609 0.0 16949 STS 1.0651 45 0.656022 0.9 10328 0.460252 1.332961 0.0 153 19 SUPT4H 1 0.765986 0.024948 0.844335 0 .10581 7 0.806960 0.029800

TFAP2A 1.0721 8 1 0.229260 1.006607 0.89591 3 1.162769 0.0091 46

TIMM8AP1 1.31 66 10 0.037070 1.135399 0.278472 1.179795 0.2 18924 TM4SF4 1.189333 0.000259 1.104730 0.023585 1.126832 0.0071 63

TM9SF4 0.799885 0.022205 0.882396 0 .130886 0.841 469 0.0401 3 1 TRERF1 1.1521 20 0.223466 0.99347 1 0.948759 1.430096 0.000356

TRNP1 1.18 1406 0.004930 1.098287 0.068505 1.144291 0.0 16540 TSPAN6 0.848330 0.0 12430 0.971 464 0.608883 0.81 6825 0.0001 58 UBASH3B 1.076483 0.425824 0.98439 1 0.847303 1.290780 0.000973 VAV3 0.823752 0.008269 0.967883 0.5891 43 0.76051 3 0.000076 VNN 1 1.303692 0.000575 1.069727 0.347998 1.41 6 119 0.000001 XKR9 1.095569 0.3 13464 0.948245 0.527779 1.158396 0.076050 ZNF51 8B 0.880308 0 .1843 18 1.058794 0.467954 0.808457 0.006073 4.1 .3 AdaBoost Signature The univariate analyses for the 29 genes in the AdaBoost signature are provided in Table 4.4.

Table 4.4: Hazard ratios (HR) and p-values for the 29 genes in the AdaBoost signature.

OS RFS SAR Gene HR p-value HR p-value HR p-value ABHD12B 0.874514 0.072398 0.914256 0.139512 0.987078 0.836374 AQP5 1.065522 0.138348 0.979870 0.625296 1.133315 0.002084 C3orf14 0.934716 0.377755 1.009220 0.884522 0.939038 0.306504

CDHR1 0.882343 0.015251 0.944431 0.172553 0.903573 0.015955 DEFA5 0.961939 0.332530 0.949784 0.137190 1.010263 0.773808 EPDR1 0.958408 0.505291 1.023628 0.667766 0.886263 0.046942 GRM8 0.849902 0.022081 0.914437 0.123485 0.813905 0.002349 HOXC6 1.050021 0.245995 1.022963 0.528481 1.163585 0.0001 15

KLK7 1.166322 0.000290 1.1 16986 0.003296 1.132638 0.001365 LOC100134361 0.851882 0.018675 0.910384 0.102897 0.895754 0.066705 LOC 146336 1.008758 0.905163 0.910496 0.190946 1.197185 0.006389 MLF1 1.102477 0.715277 1.027213 0.910235 0.888758 0.610615 NPTX2 0.969961 0.362040 0.999669 0.990423 0.958906 0.184604 NTSR1 1.252817 0.051561 1.057497 0.632747 1.174101 0.092968 PIPOX 0.859552 0.007045 0.929001 0.106368 0.859023 0.001 155 PMAIP1 0.997173 0.979369 0.916793 0.367122 1.070145 0.451713 PPBP 0.914964 0.080723 0.905417 0.024926 0.948058 0.237915 PTPRO 0.871313 0.000546 0.916871 0.009726 0.920625 0.022089 REG1B 0.968052 0.324659 0.968596 0.252862 1.004730 0.866556 REG4 1.048030 0.181 121 0.993976 0.847632 1.098196 0.004958 RNLS 0.977795 0.766893 1.004652 0.942227 0.899600 0.101564 SLC26A3 0.979071 0.530018 1.016791 0.558492 0.939665 0.024468 SOX8 1.068053 0.1 19300 1.024612 0.517469 1.094492 0.013953 SPINK1 0.936081 0.078244 0.976927 0.473389 0.932000 0.023642 SPRR1A 1.000828 0.992821 0.941992 0.470270 1.175592 0.075792 OS RFS SAR Gene HR p-value HR p-value HR p-value

SPRR1 B 1.076602 0.28051 3 1.00481 2 0.939228 1.18 1609 0.0 16949 STS 1.0651 45 0.656022 0.91 0328 0.460252 1.332961 0.0 153 19 UBASH3B 1.076483 0.425824 0.984391 0.847303 1.290780 0.000973 ZIC2 0.958494 0 .123379 0.953836 0.0427 14 1.0321 16 0 .188028

4.2 mTSP: BRAFmut status vs. predicted BRAF-mut 4.2.1 PETACC3 samples On the PETACC3 data, the following endpoints were considered: overall survival (OS), relapse-free survival (RFS) and survival after relapse (SAR). For each of the endpoints the BRAFmut status as predicted by mTSP is compared with the BRAFmut status given by PCR.

In Table 4.5, the results of the Cox proportional models analysis are given for the predicted BRAFmut status and for the golden standard (BRAFmut by PCR).

Table 4.5: (PETACC3 data/mTSP) Hazard ratios (HR) and p-values for predicted and assessed BRAFmut status, produced by Cox proportional harzards model.

The Kaplan-Meier curves for the same three endpoints are shown in Figure 6 . Note that the p-values given in the figures correspond to the likelihood ratio test for the differences between the two groups. The effect of stratification induced by the mTSP signature was also studied within the group of KRASmut samples, without taking into account the MSI status. While there is no statistically significant difference in survival experience for OS and RFS endpoints, at 0.05 level, there is a significant difference for the SAR endpoint (p-value=0.04, HR=1 .58) - see Figure 7. The interactions between MSI status and predicted BRAF mutation status were also studied within different subpopulations of the PETACC3 data set for all the three endpoints. The results are given in Figures 8, 9, and 10. Within the MSS subpopulation there is a statistically significant difference in survival experience between BRAF.hi and BRAF.Io groups (predicted by mTSP), for OS and SAR, in all stratifications. For the RFS endpoint, the only stratifications with significant differences are the whole population and all but BRAF mutants (see Figures 9A and 9B).

Within the MSI subpopulation there is no statistically significant difference in survival experience, in any stratification and for all endpoints.

Table 4.6: CETUX data/mTSP - Hazard ratios (HR) and p-values for predicted and assessed BRAFmut status, produced by the Cox proportional hazards model.

In the PETACC3 data set, within WT2 samples, there is no statistically significant difference in survival experience between samples classified as BRAFIike and BRAFmut-like, by the mTSP, for any of the three endpoints.

4.2.2 CETUX samples All CETUX samples represent metastatic patients (stage IV) and two endpoints are considered: overall survival (OS) and progression-free survival (PFS). In Table 4.8, the results of Cox proportional models analyses are given for the predicted BRAFmut status and for the golden standard (BRAFmut by PCR). The Kaplan-Meier curves for the two endpoints (OS and PFS) are given in Figure 17. Note that the p-values given in the figures correspond to the likelihood ratio test for the differences between the two groups.

In this smaller data set, only two KRASmut samples are classified as BRAFIike by the mTSP, so any conclusion about the separation between BRAF-like and non- BRAF-like samples within the KRASmut group is speculative. Nevertheless, for the sake of completion, we mention that there is a significant difference between the two groups for both OS and PFS (see Figure 12). The test within WT2 samples cannot be performed in the CETUX data set because a single WT2 sample is misclassified, meaning that the BRAF-like group is too small.

4.3 CCP2: BRAFmut status vs. predicted BRAF-mut 4.3.1 PETACC3 samples On the PETACC3 data, the following endpoints were considered: overall survival (OS), relapse-free survival (RFS) and survival after relapse (SAR). For each of the endpoints the BRAFmut status as predicted by CCP2 is compared with the BRAFmut status given by PCR.

In Table 4.7, the results of Cox proportional models analyses are given for the predicted BRAFmut status and for the golden standard (BRAFmut by PCR). The Kaplan-Meier curves for the same three endpoints are given in Figure 13.

Table 4.7: (PETACC3 data/CCP2) Hazard ratios (HR) and p-values for predicted and assessed BRAFmut status, produced by Cox proportional harzards model.

Note that the p-values given in the figures correspond to the likelihood ratio test for the differences between the two groups. Within KRASmut samples, the effect of the stratification induced by the CCP2 signature within the group of KRASmut samples was also studied, without taking into account the MSI status. There is no statistically significant difference in survival experience between the two groups defined by the CCP2 classifier, for all three endpoints. The interactions between MSI status and predicted BRAF mutation status were also studied within different subpopulations of the PETACC3 data set, for all three endpoints. The results are given in Figures 14, 15, and 16. Within the MSS subpopulation there are statistically significant differences in survival experience between BRAF.hi and BRAF.Io groups (predicted by CCP2) for various combinations of endpoint and stratification: • OS endpoint: all but within KRAS mutants stratifications show a significant difference between BRAF.Io and BRAF.hi • RFS endpoint: the only significant differences are within the whole population and all but BRAF mutants • SAR endpoint: all but within KRAS mutant stratifications show a significant difference between BRAF.Io and BRAF.hi

Within the MSI subpopulation there is no statistically significant difference in survival experience, in any stratification and for all endpoints.

4.3.2 CETUX samples All CETUX samples represent metastatic patients (stage IV) and two endpoints are considered: overall survival (OS) and progression-free survival (PFS).

In Table 4.8, the results of Cox proportional models analyses are given for the predicted BRAFmut status and for the golden standard (BRAFmut by PCR). The Kaplan-Meier curves for the two endpoints (OS and PFS) are given in Figure 17. Note that the p-values given in the figures correspond to the likelihood ratio test for the differences between the two groups.

Table 4.8: (CETUX data/CCP2) Hazard ratios (HR) and p-values for predicted (by CCP2) and assessed BRAFmut status, produced by Cox proportional harzards model.

mTSP prediction BRAFmut status Endpoint HR p-value HR p-value OS 7.42 1.3e-05 9.43 1.1e-06 PFS 13.8 2.7e-07 5.80 2.0e-05 4.4 Ada Boost: BRAFmut status vs. predicted BRAFmut 4.4.1 Overall survival. Univariate analysis: BRAF score vs. BRAFhi vs. BRAFmut The continuous BRAF score is predictive for overall survival (p-value=0.0045,

HR=1 .91 ). Also, its binarized version, BRAFhi, is predictive for OS (p-value=0.0013, HR=1 .62). In contrast, the BRAFmut variable is only marginally significant for OS prediction (p-value=0.059, HR=1 .64). Figure 18 shows the KM curves for the subpopulations identified by BRAFhi and BRAFmut indicator variables, in the whole patient population.

In a bivariate model including either BRAF score or BRAFhi and BRAFmut, the BRAF score and BRAFhi were always significant, with BRAFmut being redundant. Another way of assessing the predictive power of a variable/score is to use the time- dependent ROC curves (Heagerty et al., Biometrics 56:337-344 (2000)). These are a generalization of the usual ROC curves and give an indication of the dichotomization power of the variable/score at a given time point. Nevertheless, the BRAF score and BRAFhi indicator are always better than the BRAFmut status - and they work also in WT2 and KRASmut subgroups.

Multivariate models Starting with a full model including all the variables (BRAFscore, age, grade, tstage, nstage, site, MSI, KRASm) and their pairwise interactions, and using automatic stepwise variable elimination (with AIC criterion) led to the following model: coxph(formula = Surv(os_time, os_event) ~ BRAF.score + age + grade + tstage + nstage + site + MSI + KRASm. any + BRAF.score:MSI + tstage:site + tstage:KRASm.any + BRAF.score:tstage + site:KRASm.any + age:tstage + nstage:site) To further reduce the model, this step was followed by manual variable selection: those variables (or interaction terms) with non-significant (at 0.05 level) p-values were removed and the models re-assessed. The final model was: coxph(formula = Surv(os_time, os_event) ~ BRAF.score + grade + tstage + KRASm. any) with the following HRs and p-values: coef HR p-value BRAF.score 0.544 1.72 0.0270 gradeG-34 0.444 1.56 0.0470 tstageT3 0.744 2.10 0.0760 tstageT4 1.506 4.51 0.0005

KRASm.any 0.270 1.31 0.0230

Interestingly, MSI status does not seem to be significant in a model with KRASmut and BRAF score, with or without interaction between BRAF score and MSI: coxph(formula = Surv(os_time, os_event) ~ BRAF.score + grade + tstage + MSI + KRASm.any + BRAF.score * MSI)

coef HR D-value BRAF.score 2.003 7.408 0.01 100 gradeG-34 0.626 1.870 0.00660 tstageT3 0.832 2.298 0.04700 tstageT4 1.593 4.917 0.00023 MSIMSI-H 0.629 1.875 0.49000 MSIMSS 1.126 3.083 0.03800 KRASm.any 0.318 1.374 0.02700 BRAF.score:MSIMSI-H -2.260 0.104 0.08900

BRAF.score:MSIMSS - 1 .276 0.279 0.12000 and coxph(formula = Surv(os_time, os_event) ~ BRAF.score + grade + tstage + MSI + KRASm.any)

coef HR D-value BRAF.score 0.813 2.254 0.00180 gradeG-34 0.632 1.882 0.00600 tstageT3 0.821 2.273 0.05000 tstageT4 1.592 4.914 0.00024 MSIMSI-H -0.690 0.502 0.12000 MSIMSS 0.491 1.633 0.09600 KRASm.any 0.306 1.357 0.03200 BRAFhi and MSI status within different subpopulations

• whole population: there is a clear difference in OS between BRAF-high and BRAF-low groups within MSS (p-value=0.00036), but not within MSI. Also, within BRAF-high there is a significant difference between MSI and MSS patients (p- value=0.016), but not within BRAF-low patients (see Figure 19A). • all but BRAFmut. within MSS, BRAF-high has a worse prognostic than BRAF- low (p-value=0.0066), however, there is no difference between BRAF-high and -low within MSI. On the other hand, there is no significant difference between MSI and MSS within BRAF-high or BRAF-low subpopulations (see Figure 19B). · only BRAFmut and KRASmut. the only significant difference is between BRAF- low and -high within MSI subpopulation (p-value=0.013) (see Figure 19C). • only KRASmut. the only marginally significant difference (p-value=0.07) is between MSS and MSI in the BRAF-high group of KRASmuts (see Figure 19D).

• only WT2: there is no significant difference between various subgroups (see Figure 19E).

4.4.2 Relapse-free survival. Univariate analysis: BRAF score vs. BRAFhi vs. BRAFmut The continuous BRAF score is not predictive for RFS (p-value=0.36). However, its binarized version, BRAFhi, is marginally predictive for RFS (p-value=0.07, HR=1 .27). On the other hand, the BRAFmut variable is not predictive for RFS (p-value=0.63), either. Figure 20 shows the KM curves for the subpopulations identified by BRAFhi and BRAFmut indicator variables, in the whole patient population.

4.4.3 Survival after relapse. Univariate analysis: BRAF score vs. BRAFhi vs. BRAFmut The continuous BRAF score is strongly predictive for SAR (p-value=0. 00000023, HR=3.56), as is its binarized version, BRAFhi (p-value=0. 000053, HR=1 .85). The

BRAFmut variable is also significant for SAR prediction (p-value=0. 00006, HR=2.91 ).

Figure 2 1 shows the KM curves for the subpopulations identified by BRAFhi and BRAFmut indicator variables, in the whole patient population. The AUCs at 3 years are better than in the case of OS, but they remain below 0.7. Multivariate models Starting with a full model including all the variables (BRAFscore, age, grade, tstage, nstage, site, MSI, KRASm) and their pairwise interactions, and using automatic stepwise variable elimination (with AIC criterion) led to the following model: coxph(formula = Surv(os_time - rfs_time, os_event) ~ BRAF. score + grade + tstage + site + MSI + tstage:site + BRAF.score:grade, data = D) coef HR p-value BRAF. score 1.001 1 2.721 0.00044 gradeG-34 -0.5918 0.553 0.29000 tstageT3 0.6093 1.839 0.30000 tstageT4 0.6837 1.981 0.26000 siteright 0.9637 2.621 0.24000

MSIMSI-H - 1.1436 0.319 0.00530 MSIMSS -0.0277 0.973 0.91000 tstageT3:siteright -0.8073 0.446 0.34000 tstageT4:siteright 0.4503 1.569 0.61000 BRAF.score:gradeG-34 2.0818 8.019 0.01400

To further reduce the model, this step was followed by manual variable selection: those variables (or interaction terms) with non-significant (at 0.05 level) p-values were removed and the models re-assessed. The final model was: coxph(formula = Surv(os_time - rfs_time, os_event) ~ BRAF. score + grade + tstage + site + MSI, data = D)

coef HR D-value BRAF. score 1.23e+00 3.419 4.6e-06 gradeG-34 4.80e-01 1.615 3.3e-02 tstageT3 3.32e-01 1.393 4.3e-01 tstageT4 8.48e-01 2.336 5.2e-02 siteright 3.95e-01 1.485 1.7e-02 MSIMSI-H -8.93e-01 0.409 2.7e-02 MSIMSS -9.20e-06 1.000 1.0e+00

BRAFhi and MSI status within different subpopulations • whole population: there is a clear difference in SAR between BRAF-high and

BRAF-low groups within MSS (p-value=0.000331 ), but not within MSI (however, there are not many MSIs) (see Figure 22A). • all but BRAFmut. there is a significant (p-value=0.015) difference between BRAF-high and BRAF-low patients. Also, within MSS, BRAF-high has a worse prognostic than BRAF-low (p-value=0.019), however, there is no difference between BRAF-high and -low within MSI. There is no significant difference between MSI and MSS within BRAF-high or BRAF-low subpopulations (see Figure 22B). • only BRAFmut and KRASmut there is no significant difference between subgroups (see Figure 22C). • only KRASmut: there is no significant difference between subgroups (see Figure 22D).

• only WT2: there are only a few MSIs, so most of the effect is due to MSSs. There is a marginally significant difference between BRAF-low and BRAF-high subgroups both within all WT2 and within MSS (p=0.04) (see Figure 22E). CLAIMS

1. A method of classifying a subject with CRC comprising:

a. analyzing at least one of the gene pairs shown in Table 3.1 . 1 according to the top scoring pair method; and

b. classifying the subject into a BRAF mutant-like group or a wild-type group.

2 . The method according to claim 1, wherein at least 10 of the gene pairs shown in

Table 3.1 . 1 are analyzed according to the top scoring pair method.

3 . The method according to claim 1, wherein at least 30 of the gene pairs shown in

Table 3.1 . 1 are analyzed according to the top scoring pair method.

4 . The method according to claim 1, wherein the 39 gene pairs shown in Table

3.1 . 1 are analyzed according to the top scoring pair method.

5 . The method of claim 1, wherein the top scoring pair method is carried out by comparing the average value of the relative expression levels of all Genel genes used in the analysis with the average value of relative expression levels of all Gene2 genes used in the analysis, wherein if the average Genel value is less than the average Gene2 value, then the subject is classified as BRAF mutant-like.

6 . The method of claim 6, wherein if the average Genel value is greater than or equal to the average Gene2 value, then the subject is classified as wild-type.

7 . The method of claim 5 or 6, wherein the analysis uses the 39 pairs of genes shown in Table 3.1 . 1 .

8 . A method for selecting therapy comprising the steps of any one of claims 1-7, and further comprising selecting adjuvant chemotherapy for a subject classified as wild- type, or selecting no adjuvant chemotherapy for a subject classified as BRAF mutant like. 9 . A method for selecting therapy comprising the steps of any one of claims 1-7, and further comprising selecting adjuvant chemotherapy for a subject classified as wild- type, or selecting a treatment regimen comprising a BRAF mutant-specific inhibitor for a subject classified as BRAF mutant-like.

10. A method of treating a subject with CRC comprising administering a BRAF mutant-specific inhibitor to said subject, wherein said subject is classified as BRAF mutant-like according to any of the methods of claims 1-7.

11. The method according to claim 8 or 9, wherein said subject is a human.

12. A CRC prognosticator comprising a mechanism for determining relative expression levels in a CRC tumor sample of the genes listed in Table 3.1 . 1 .

13. The CRC prognosticator of claim 12, wherein the mechanism comprises a microarray.

14. The CRC prognosticator of claim 12, wherein the mechanism comprises an assay of reverse transcription polymerase chain reaction.

15. A kit for classifying a subject with CRC comprising detection agents capable of detecting the expression products of at least one gene pair shown in Table 3.1 . 1 .

16. The kit of claim 15, further comprising an addressable array comprising probes for the expression products of the at least one gene pair.

17. The kit of claim 15, wherein the detection agents comprise primers capable of hybridizing to the expression products of the at least one gene pair.

18. The kit of claim 15, comprising detection agents capable of detecting the expression products of at least 10 gene pairs shown in Table 3.1 . 1 .

19. The kit of claim 15, comprising detection agents capable of detecting the expression products of at least 20 gene pairs shown in Table 3.1 . 1 . 20. The kit of claim 15, comprising detection agents capable of detecting the expression products of at least 30 gene pairs shown in Table 3.1 . 1 .

2 1. The kit of claim 15, comprising detection agents capable of detecting the expression products of the 39 gene pairs shown in Table 3.1 . 1 .

22. A kit according to claim 15, further comprising a computer implemented product for comparing a) the relative expression level values for Genel genes in Table 3.1 . 1 for a subject to b) the relative expression level values for Gene2 genes in Table 3.1 . 1 for said subject.

23. The kit according to claim 22, wherein the average value of the relative expression levels of all Genel genes used in the analysis is compared with the average value of relative expression levels of all Gene2 genes used in the analysis.

24. The kit according to claim 23, wherein the 39 gene pairs in Table 3.1 . 1 are used in the analysis.

International application No PCT/IB2011/054962

A . CLASSIFICATION O F SUBJECT MATTER INV. C12Q1/68 ADD.

According to International Patent Classification (IPC) or to both national classification and IPC

B. FIELDS SEARCHED Minimum documentation searched (classification system followed by classification symbols C12Q

Documentation searched other than minimum documentation to the extent that such documents are included in the fields searched

Electronic data base consulted during the international search (name of data base and, where practical, search terms used)

EPO-Internal , BIOSIS, CHEM ABS Data, Sequence Search, EMBASE, WPI Data

C . DOCUMENTS CONSIDERED TO BE RELEVANT

Category* Citation of document, with indication, where appropriate, of the relevant passages Relevant to claim No.

T . M. PITTS ET AL: "Devel opment of an 12-24 Integrated Genomi c Cl assi fier for a Novel Agent i n Colorectal Cancer: Approach t o Indivi dual zed Therapy i n Early Devel opment" , CLINICAL CANCER RESEARCH , vol . 16, no. 12, 15 June 2010 (2010-06-15) , pages 3193-3204, XP55019519 , ISSN : 1078-0432, D0I : 10. 1158/1078-0432 . CCR-09-3191 Abstract; page 3195, right-hand col umn; 1-9 page 3196, second and last paragraph ; page 3200

/ -

Further documents are listed in the continuation of Box C . See patent family annex.

* Special categories of cited documents : "T" later document published after the international filing date or priority date and not in conflict with the application but "A" document defining the general state of the art which is not cited to understand the principle or theory underlying the considered to be of particular relevance invention "E" earlier document but published on or after the international "X" document of particular relevance; the claimed invention filing date cannot be considered novel or cannot be considered to "L" document which may throw doubts on priority claim(s) or involve an inventive step when the document is taken alone which is cited to establish the publication date of another " document of particular relevance; the claimed invention citation or other special reason (as specified) cannot be considered to involve an inventive step when the "O" document referring to a n oral disclosure, use, exhibition or document is combined with one or more other such docu¬ other means ments, such combination being obvious to a person skilled in the art. "P" document published prior to the international filing date but later than the priority date claimed "&" document member of the same patent family

Date of the actual completion of the international search Date of mailing of the international search report

16 February 2012 28/02/2012

Name and mailing address of the ISA/ Authorized officer European Patent Office, P.B. 5818 Patentlaan 2 NL - 2280 HV Rijswijk Tel. (+31-70) 340-2040, Fax: (+31-70) 340-3016 Hennard, Chri stophe International application No PCT/IB2011/054962

C(Continuation). DOCUMENTS CONSIDERED TO BE RELEVANT

Category* Citation of document, with indication, where appropriate, of the relevant passages Relevant to claim No.

X J . J . TENTLER ET AL: " Identi f i cation of 12-24 Predi ctive Markers of Response t o the MEK1/2 Inhibi tor Sel umeti nib (AZD6244) i n K-ras-Mutated Col orectal Cancer" , MOLECULAR CANCER THERAPEUTICS, vol . 9 , no. 12, 5 October 2010 (2010-10-05) , pages 3351-3362 , XP55019521 , ISSN : 1535-7163, D0I : 10. 1158/1535-7163 . MCT-10-0376 Y pages 3352-2254; tabl e 1 1-9

X J . J . ARCAR0LI ET AL: "Gene Array and 12-24 Fl uorescence I n si t u Hybri di zation Bi omarkers of Acti v i t y of Saracati nib (AZD0530) , a Src Inhi bi tor, i n a Precl i ni cal Model of Colorectal Cancer" , CLINICAL CANCER RESEARCH , vol . 16, no. 16, 15 August 2 1 (2010-08-15) , pages 4165-4177 , XP55019523 , ISSN : 1078-0432, D0I : 10. 1158/1078-0432 . CCR-10-0066 Y pages 4166-4168, "Material s and Methods" ; 1-9 page 4171 , last paragraph

X AI K CH00N TAN ET AL: "Simple deci sion 12-24 rules for classi fying human cancers from gene expressi on profi l es" , BI0INF0RMATICS, OXFORD UNIVERSITY PRESS, SURREY, GB, vol . 21, no. 20, 1 January 2005 (2005-01-01) , pages 3896-3904, XP002545348, ISSN : 1367-4803, D0I : 10. 1093/BI0INF0RMATICS/BTI631 [retri eved on 2005-08-16] Y page 3896, " Introducti on" ; page 3897 , i tem 1-9 2 . 1 ; tabl es 1 and 3 ; page 3899, i tem 3

X 0 2007/084992 A2 (UNIV CHICAGO [US] ; 12-21 WEICHSELBAUM RALPH [US] ; ROIZMAN BERNARD [US] ; MINN) 26 July 2007 (2007-07-26) page 8 , l i ne 20 - page 9 , l ine 8 ; page 12 , l i nes 13-19; claims; examples

X W0 2007/100859 A2 (PFIZER PROD INC [US] ; 12-21 DEL RIO MARGUERITE [FR] ; MOLINA FRANCK [FR] ; PAU) 7 September 2007 (2007-09-07) exampl es ; claims ; pages 4-11

X W0 2010/006225 Al (N0VARTIS AG [CH] ; 10-12 GARCIA-ECHEVERRIA CARLOS [CH] ; MAI RA SAUVEUR-MICHEL) 14 January 2010 (2010-01-14) c l aims; page 12, l ines 5-8; page 15 International application No Information on patent family members PCT/IB2011/054962

Patent document Publication Patent family Publication cited in search report date member(s) date

WO 2007084992 A2 26-07-2007 2009011439 Al 08-01-2009 2011200998 Al 18-08-2011 2007084992 A2 26-07-2007

O 2007100859 A2 07--09--2007 CA 2643225 Al 07-09-2007 EP 1994177 A2 26-11-2008 P 2009528061 A 06-08-2009 US 2009221609 Al 03-09-2009 WO 2007100859 A2 07-09-2007

wo 2010006225 Al 14--01--2010 AU 2009268469 Al 14-01-2010 CA 2729914 Al 14-01-2010 CN 102089007 A 08-06-2011 EP 2310050 Al 20-04-2011 P 2011527703 A 04-11-2011 KR 20110028651 A 21-03-2011 US 2011105521 Al 05-05-2011 W0 2010006225 Al 14-01-2010