WO2006093507A2

WO2006093507A2 - Methods and systems for predicting cancer outcome

Info

Publication number: WO2006093507A2
Application number: PCT/US2005/017988
Authority: WO
Inventors: Timothy J. Yeatman; Steven Eschrich; Gregory C. Bloom
Original assignee: H. Lee Moffitt Cancer Center And Research Institute, Inc.; University Of South Florida
Priority date: 2005-02-25
Filing date: 2005-05-19
Publication date: 2006-09-08
Also published as: EP1894131B1; WO2006093507A3; US20060195269A1; EP1894131A4; US10181009B2; US20060195266A1; EP1894131A2

Abstract

The invention provides a molecular marker set that can be used for prognosis of colorectal cancer in a colorectal cancer patient. The invention also provides methods and computer systems for evaluating prognosis of colorectal cancer in a colorectal cancer patient based on the molecular marker set. The invention also provides methods and computer systems for determining chemotherapy for a colorectal cancer patient and for enrolling patients in clinical trials.

Description

METHODS AND SYSTEMS FOR PREDICTING CANCER OUTCOME

1. FIELD QF THE INVENTION

The invention relates to molecular markers that can be used for prognosis of colorectal cancer. The invention also relates to methods and computer systems for determining a prognosis of colorectal cancer in a colorectal cancer patient based on the molecular markers. The invention also relates to methods and computer systems for determining chemotherapy for a colorectal cancer patient and for enrolling patients in clinical trials.

2. BACKGROUND OF THE INVENTION

Ranked as the third most commonly diagnosed cancer and the second leading cause of cancer deaths in the United States (American Cancer Society, "Cancer facts and figures," Washington, DC: American Cancer Society (2000)), colon cancer is a deadly disease afflicting nearly 130,000 new patients yearly in the United States. Colon cancer is the only cancer that occurs with approximately equal frequency in men and women. There are several potential risk factors for the development of colon and/or rectal cancer. Known factors for the disease include older age, excessive alcohol consumption, sedentary lifestyle (Reddy, Cancer Res., 41:3700-3705 (1981)), and genetic predisposition (Potter J Natl Cancer Institute, 91:916-932 (1999)).

Several molecular pathways have been linked to the development of colon cancer (see, for example, Leeman et ah, J Pathol., 201(4):528-34 (2003); Kanazawa et ah, TumorL, 89(4):408-ll (2003); and Notarnicola et ah, Oncol Rep., 10(6):1987-91 (2003)), and the expression of key genes in any of these pathways may be affected by inherited or acquired mutation or by hypermethylation. A great deal of research has been performed with regard to identifying genes for which changes in expression may provide an early indicator of colon cancer or a predisposition for the development of colon cancer. Unfortunately, no research has yet been conducted on identifying specific genes associated with colorectal cancer and specific outcomes to provide an accurate prediction of prognosis. Survival of patients with colon and/or rectal cancer depends to a large extent on the stage of the disease at diagnosis. Devised nearly seventy years ago (Dukes, 1932, J Pathol Bacterid 35:323), the modified Dukes' staging system for colon cancer, discriminates four stages (A, B, C, and D), primarily based on clinicopathologic features such as the presence or absence of lymph node or distant metastases. Specifically, colonic tumors are classified by four Dukes' stages: A, tumor within the intestinal mucosa; B, tumor into muscularis mucosa; C, metastasis to lymph nodes and D, metastasis to other tissues. Of the systems available, the Dukes' staging system, based on the pathological spread of disease through the bowel wall, to lymph nodes, and to distant organ sites such as the liver, has remained the most popular. Despite providing only a relative estimate for cure for any individual patient, the Dukes' staging system remains the standard for predicting colon cancer prognosis, and is the primary means for directing adjuvant therapy.

The Dukes' staging system, however, has only been found useful in predicting the behavior of a population of patients, rather than an individual. For this reason, any patient with a Dukes A, B, or C lesion would be predicted to be alive at 36 months while a patient staged as Dukes D would be predicted to be dead. Unfortunately, application of this staging system results in the potential over-treatment or under-treatment of a significant number of patients. Further, Dukes' staging can only be applied after complete surgical resection rather than after a pre-surgical biopsy.

DNA array technologies have made it possible to monitor the expression level of a large number of genetic transcripts at any one time (see, e.g., Schena et al, 1995, Science 270:461-410; Lockhart et al, 1996, Nature Biotechnology 74:1675-1680; Blanchard et al, 1996, Nature Biotechnology 14:1649; Ashby et al, U.S. Patent No. 5,569,588, issued October 29, 1996). Of the two main formats of DNA arrays, spotted cDNA arrays are prepared by depositing PCR products of cDNA fragments with sizes ranging from about 0.6 to 2.4kb, from full length cDNAs, ESTs, etc., onto a suitable surface (see, e.g., DeRisi et al, 1996, Nature Genetics 14:451-460; Shalon et al, 1996, Genome Res. (5:689-645; Schena et al, 1995, Proc. Natl Acad. Sci. U.S.A. 95:10539-11286; and Duggan et al., Nature Genetics Supplement 27:10-14). Alternatively, high-density oligonucleotide arrays containing thousands of oligonucleotides complementary to defined sequences, at defined locations on a surface are synthesized in situ on the surface by, for example, photolithographic techniques (see, e.g., Fodor et al, 1991, Science 251:161-113; Pease et al, 1994, Proc. Natl. Acad. Sci. U.S.A. 97:5022-5026; Lockhart et al, 1996, Nature Biotechnology 14:1615; McGaIl et al., 1996, Proc. Natl. Acad. ScL U.S.A. 93:13555-13560; U.S. Patent Nos. 5,578,832; 5,556,752; 5,510,270; and 6,040,138). Methods for generating arrays using inkjet technology for in situ oligonucleotide synthesis are also known in the art (see, e.g., Blanchard, International Patent Publication WO 98/41531, published September 24, 1998; Blanchard et al. , 1996,

Biosensors and Bioelectronics ii:687-690; Blanchard, 1998, in Synthetic DNA Arrays in Genetic Engineering, Vol. 20, J.K. Setlow, Ed., Plenum Press, New York at pages 111-123).

By simultaneously monitoring tens of thousands of genes, microarrays have permitted identification of biomarkers of cancer (Welsh et al, PNAS, 100(6):3410-3415 (March 2003)), creating gene expression-based classifications of cancers (Alzadeh et al., Nature, 403:513-11 (2000); and Garber et al., Proc Natl Acad Sci USA, 98:13784-9 (2001); development of gene based multi-organ cancer classifiers (Bloom et al, Am J Pathol 164:9-16, 2004; Giordano et al, Am J Pathol, 159:1231-8 (2001); Ramaswamy et al, Proc Natl Acad Sci USA, 98:15149- 54 (2001); and Su et al, Cancer Res, 61:7388-93 (2001)), identification of tumor subclasses (Dyrskjot et al., Nat Genet, 33:90-6 (2003); Bhattacharjee et al, Proc Natl Acad Sci USA, 98: 13790-5 (2001); Garber et al, Proc Natl Acad Sci USA, 98: 13784-9. (2001); and Sorlie et al, Proc Natl Acad Sci USA, 98:10869-74 (2001)), discovery of progression markers (Sanchez-Carbayo et al, Am J Pathol, 163:505-16 (2003); and Frederiksen et al, J Cancer Res Clin Oncol, 129:263-71 (2003)); and prediction of disease outcome (Henshall et al, Cancer Res, 63:4196-203 (2003); Shipp et al, Nat Med, 8:68-74 (2002); Beer et al, Nat

Med, 8:816-24 (2002); Pomeroy et al, Nature, 415:436-42 (2002); van 't Veer et al, Nature, 415:530-6 (2002); Vasselli et al, Proc Natl Acad Sci USA, 100:6958-63 (2003); Takahashi et al, Proc Natl Acad Sci USA, 98:9754-9 (2001); WO 2004/065545 A2; WO 02/103320 A2)); and in drug discovery (Marton et al, Nat Med, 4(11): 1293-301 (1998); and Gray et al, Science, 281:533-538 (1998)).

One tool that has been applied to microarrays to decipher and compare genome expression patterns in biological systems is Significance Analysis of Microarrays, or SAM (Tusher et al, 2001, Proc. Natl. Acad. Sci. 98:5116-5121). This statistical method was developed as a cluster tool for use in identifying genes with statistically significant changes in expression. SAM has been used for a variety of purposes, including identifying potential drugs that would be effective in treating various conditions associated with specific gene expressions (Bunney et al, Am J Psychiatry, 160(4):657-66 (Apr. 2003)). Sophisticated and powerful machine learning algorithms have been applied to transcriptional profiling analysis. For example, a modified "Fisher classification" approach has been applied to distinguish patients with good prognosis from those who do not have a good prognosis, based on their expression profiles (van 't Veer et al., 2002, Nature 415: 530-6). A similar study has been reported using an artificial neural network (Bloom et al, Am J Pathol 164:9-16, 2004; Khan et al., 2001, Nat Med 7: 673-9). Support Vector Machine (SVM) (see, e.g., Brown et al., Proc. Natl. Acad. Sci. 97(l):262-67 (2000); Zien et al, Bioinformatics, 16(9):799-807 (2000); Furey et al, Bioinformatics, 16(10):906-914 (2000)) is a correlation tool shown to perform well in multiple areas of biological analysis, including evaluating microarray expression data (Brown et al, Proc Natl Acad Sci USA, 97:262-267 (2000)), detecting remote protein homologies (Jaakkola et al., Proceedings of the 7^th International Conference on Intelligent Systems for Molecular Biology, AAAI Press, Menlo Park, CA (1999)), and classification of cancer tissues (Furey et ah, Bioinformatics, 16(10):906-914 (2000)). Furey describes using SVM to classify colon cancer tissues based on expression levels of a set of 2000 genes or a set of 1000 genes having the highest minimal intensity across 60 colon tissue samples (40 tumors and 22 normal tissues) on an Affymetrix^® oligonucleotide microarray.

Wang et al. (Wang et al., 2004, J. Clinical Oncology 22:1564-1571) reported identification of a 60-gene and a 23-gene signature for prediction of cancer recurrence in Dukes' B patients using an Affymetrix^® U133a GeneChip. This signature was validated in 36 independent patients. Two supervised class prediction approaches were used to identify gene markers that could best discriminate between patients who would experience relapse and patients who would remain disease-free. A multivariate Cox model was built to predict recurrence. The overall performance accuracy was reported as 78%.

Resnick et al. (Resnick et al., 2004, Clin. Can. Res. 10:3069-3075) reported a study of the prognostic value of epidermal growth factor receptor, c-MET, b-catenin, and p53 protein expression in TNM stage II colon cancer using tissue microarray technology.

Muro et al. (Muro et al., 2003, Genome Biology 4:R21) describes identification and analysis of the expression levels of 1,536 genes in colorectal cancer and normal tissues using a parametric clustering method. Three groups of genes were discovered. Some of the genes were shown to not only correlate with the differences between tumor and normal tissues but also the presence and absence of distant metastasis. U.S. Patent Application Publication No. 2005/0048542A1, published on March 3, 2005, describes a noninvasive, quantitative test for prognosis determination in cancer patients. The test makes use of measurements of the tumor levels of certain messenger RNAs (mRNAs). These mRNA levels are inserted into a polynomial formula (algorithm) that yields a numerical recurrence score, which indicates recurrence risk.

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

3. SUMMARY OF THE INVENTION The invention provides a method for determining a prognosis of colorectal cancer in a colorectal cancer patient, comprising classifying said patient as having a good prognosis or a poor prognosis using measurements of a plurality of gene products in a cell sample taken from said patient, said gene products being respectively products of at least 5 of the genes listed in Table 1, or in any of Tables 2-5, 7, and 8 or any subset of these tables, or respective functional equivalents thereof, wherein said good prognosis predicts survival of a patient within a predetermined time period from obtaining a tumor sample from said patient by surgery or from diagnosis of colorectal cancer, and said poor prognosis predicts non-survival of a patient within said time period. In a specific embodiment, the predetermined time period is not 3 years. For example, in one embodiment, the predetermined time period is longer than 3 years. In other embodiments, the time period is 4 or 5 years or is between 3 and 5 years. In another particular embodiment, measurements of gene products of all or in the range of 8 to 19, in the range of 21 to 25, in the range of 27 to 42, in the range of 44 to 52, or in the range of 54 to 130 of the genes listed in Table 1 are used. In yet another embodiment, both the predetermined time period and the number of genes is as described in these embodiments.

The invention also provides a method for evaluating whether a colorectal cancer patient should be treated with chemotherapy, comprising (a) classifying said patient as having a good prognosis or a poor prognosis using any one of the prognosis methods of the invention; and (b) determining that said patient's predicted survival time favors treatment of the patient with chemotherapy if said patient is classified as having a poor prognosis.

The invention also provides a method for enrolling colorectal cancer patients for a clinical trial of a chemotherapeutic agent for colorectal cancer, comprising (a) classifying each patient as having a good prognosis or a poor prognosis using any one of the prognosis methods of the invention; and (b) assigning each patient having a good prognosis to one patient group and each patient having a poor prognosis to another patient group, at least one of said patient group being enrolled in said clinical trial.

The invention also provides a method for identifying a set of genes for prognosis of colorectal cancer, comprising: (a) determining for each of a plurality of genes a metric of correlation between abundance level of a gene product of said gene and survival outcome in a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples; and (b) selecting one or more genes based on said metric of correlation.

In another embodiment, the invention provides a method for identifying a set of genes for prognosis of colorectal cancer, comprising: (a) generating a subset of patients by leaving out one or more patients in a plurality of patients having known outcomes at a predetermined time after obtaining tumor samples; (b) determining for each of a plurality of genes a metric of correlation between abundance level of said gene and survival outcome in said subset of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples; (c) selecting one or more genes based on said metric of correlation; (d) repeating steps (a)-(c) for a plurality of iterations, each with a different subset of patients by leaving out one or more patients in said plurality, wherein said one or more patients are different from any previous iteration; and (e) selecting one or more genes that are selected in at least a predetermined percentage of all iterations.

Also provided is a method of identifying genes that discriminate between colorectal cancer patients that have a poor prognosis and colorectal patients that have a good prognosis comprising analyzing survival data and RNA levels of colorectal patients using SAM, clustering analysis, or a neural network to select genes whose RNA levels correlate with a selected survival time.

The invention further provides a method for constructing a prognosis predictor for prognosis of colorectal cancer, comprising: (a) generating a subset of patients by leaving out one or more patients in a plurality of patients having known outcomes at a predetermined time after obtaining tumor samples; (b) determining for each of a plurality of genes a metric of correlation between expression level of said gene and survival outcome in a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples from a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples; (c) selecting one or more genes based on said metric of correlation; (d) training a prognosis predictor, wherein said prognosis predictor receives an input comprising a marker profile comprising expression levels of said one or more genes selected in step (c) and provides an output comprising data indicating a good prognosis or a poor prognosis, with training data from said subset of patients, wherein said training data comprise for each of said subset of patients a marker profile comprising measurements of said one or more genes in a tumor cell sample taken from said patient and prognosis information; (e) determining a prognosis for at least one of said one or more patients who are left out in step (a); (f) repeating steps (a)-(e) for a plurality of iterations, each with a different subset of patients by leaving out one or more patients in said plurality, wherein said one or more patients are different from any previous iteration; (g) selecting one or more genes that are selected hi at least a predetermined percentage of all iterations; and (h) training a prognosis predictor, wherein said prognosis predictor receives an input comprising a marker profile comprising expression levels of said one or more genes selected in step (g) and provides an output comprising data indicating a good prognosis or a poor prognosis, with training data from said subset of patients, wherein said training data comprise for each of said plurality of patients a marker profile comprising measurements of said one or more genes in a tumor cell sample taken from said patient and prognosis information.

The invention also provides a computer system comprising a processor, and a memory coupled to said processor and encoding one or more programs, wherein said one or more programs cause the processor to carry out any of the methods of the invention.

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

In another embodiment, the invention provides a microarray comprising for each of a plurality of genes, said genes being at least 5 of the genes listed in Table 1, one or more polynucleotide probes complementary and hybridizable to a sequence in said gene, wherein polynucleotide probes complementary and hybridizable to said genes constitute at least 50% of the probes on said microarray. In one embodiment, the invention provides a kit comprising the microarray of the invention in a sealed container.

4. BRIEF DESCRIPTION OF THE DRAWINGS

Figure IA shows cluster analysis of the 53 SAM selected genes. Darker levels represent over-expressed genes relative to (lighter levels) under-expressed genes. The data suggest that genes can be identified that discriminate good from poor prognosis. Figure IB shows that cluster analysis of SAM selected genes, grouped by Dukes' stage B and C, does not demonstrate a discriminating pattern.

Figure 2 A shows Kaplan-Meier survival curves corresponding to gene clusters. Figure 2B shows Kaplan-Meier survival curves for Dukes' B and C patients.

Figure 3A shows survival curves for molecular classifiers using 78 samples tested by Leave-One-Out-Cross- Validation (LOOCV). Figure 3B illustrates the survival curve for the U133A-limited cDNA classifier (LOOCV).

Figures 4A-4C show an independent Test Set Evaluation (Denmark Test Set) using the U133A data set. A) Survival curves generated using probe sets corresponding to 26 of the Molecular Classifier genes. Using these translated probe sets, 95 tumors were clustered and censored survivorship was evaluated (P < 0.001). B) Survival curves using Dukes' staging criteria show no significant difference in outcome. C) Survival curves grouped by both Dukes' stage and molecular signature show that both Dukes' B and C cases can be further subdivided into good and poor prognosis groups.

Fig. 5 details an exemplary computer system that supports the functionality described herein.

5. DETAILED DESCRIPTION OF THE INVENTION

The invention provides markers, i.e., genes, the expression levels of which discriminate between a good prognosis and a poor prognosis for colorectal cancer. The identities of these markers and the measurements of their respective gene products, e.g., measurements of levels (abundances) of their encoded mRNAs or proteins, can be used by application of a pattern recognition algorithm to develop a prognosis predictor that discriminates between a good and poor prognosis in colorectal cancer using measurements of such gene products in a sample from a patient. Colorectal cancer includes colon cancer and rectal cancer. Such molecular markers, the expression levels of which can be used for prognosis of colorectal cancer in a colorectal cancer patient, are listed in Table 1, infra. Measurements of gene products of these molecular markers, as well as of their functional equivalents, can be used for prognosis of colorectal cancer. A functional equivalent with respect to a gene, designated as gene A, refers to a gene that encodes a protein or rnRNA that at least partially overlaps in physiological function in the cell to that of the protein or rnRNA encoded by gene A. In particular, prognosis of colorectal cancer in a colorectal cancer patient is carried out by a method comprising classifying the patient as having a good or poor prognosis based on a profile of measurements (e.g., of the levels) of gene products of (i.e., encoded by) at least 5, 10, 20, 30, 40, 50, 60, 70, 80, or 100 of the genes in Table 1 or in any of Tables 2-5, 7, and 8 or any subset of Tables 2-5, 7, and 8, or functional equivalents of such genes; or of at least 2%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, or 90% of the genes in Table 1 or in any of Tables 2-5, 7, and 8 or any subset of Tables 2-5, 7, and 8, or functional equivalents of such genes; or all or in the range of 8 to 19, in the range of 21 to 25, in the range of 27 to 42, in the range of 44 to 52, or in the range of 54 to 130 of the genes listed in Table 1 or in any of Tables 2-5, 7, and 8 or any subset of Tables 2-5, 7, and 8, or functional equivalents of such genes, in an appropriate cell sample from the patient, e.g., a tumor cell sample obtained from biopsy or after surgical resection. Preferably, the tumor sample is contaminated with less than 50%, 40%, 30%, 20%, or 10% of normal cells. Such a profile of measurements is also referred to herein as an "expression profile." In some embodiments, "at least some of the genes listed" in a table refers to at least 5, 10, 20, 40, 50, 70 or 100 of the genes listed in the table. In other embodiments, 26, 43, or 53 genes from Table 1 are used. In still another embodiment, all or in the range of 8 to 19, in the range of 21 to 26, in the range of 27 to 42, in the range of 44 to 52, or in the range of 54 to 130 of the genes listed in Table 1 are used. Different subcombination of genes from Table 1 may be used as the marker set to carry out the prognosis methods of the invention. For example, in various embodiments, the markers that are the genes listed in Table 2, 3, 4, 5, 7 or 8 are used.

In a preferred embodiment, the plurality of gene products for which measurements are used according to the invention comprises gene products of osteopontin and neuregulin 2 isoform 4, respectively.

In a specific embodiment, the classifying of the patient as having good or poor prognosis is carried out using measurements of gene products of less than 30, 40, 50, 70, 100, 200, 300, 400, or 500 total genes, in which all or at least 20%, 30%, 40%, 50%, 60%, 70%, 80%, or 90% of the genes are from Table 1 or any of Tables 2-5, 7, and 8 their functional equivalents, or at least 5, 10, 20, 30, 40, 50, 60, 70, 80, or 100 of the genes are from Table 1 or any of the Tables 2-5, 7, and 8 or their functional equivalents, or in the range of 8 to 19, or in the range of 21 to 25, in the range of 27 to 42, in the range of 44 to 52, or in the range of 54 to 130 of the genes listed in Table 1 or any of the Tables 2-5, 7, and 8 or their functional equivalents.

The measurements in the profiles of the gene products that are used can be any suitable measured values representative of the expression levels of the respective genes. The measurement of the expression level of a gene can be direct or indirect, e.g., directly of abundance levels of RNAs or proteins or indirectly, by measuring abundance levels of cDNAs, amplified RNAs or DNAs, proteins, or activity levels of RNAs or proteins, or other molecules (e.g., a metabolite) that are indicative of the foregoing. In one embodiment, the profile comprisies measurements of abundances of the transcripts of the marker genes. The measurement of abundance can be a measurement of the absolute abundance of a gene product. The measurement of abundance can also be a value representative of the absolute abundance, e.g., a normalized abundance value (e.g., an abundance normalized against the abundance of a reference gene product) or an averaged abundance value (e.g., average of abundances obtained at different time points or from different tumor cell samples from the patients, or average of abundances obtained using different probes, etc.), or a combination of both. As an example, the measurement of abundance of a gene transcript can be a value obtained using an Affymetrix^® GeneChip^® to measure hybridization to the transcript.

In another embodiment, the expression profile is a differential expression profile comprising differential measurements of a plurality of transcripts in a sample derived from the patient versus measurements of the plurality of transcripts in a reference sample, e.g., a cell sample of normal cells. Each differential measurement in the profile can be but is not limited to an arithmetic difference, a ratio, or a log(ratio). As an example, the measurement of abundance of a gene transcript can be a value for the transcript obtained using a cDNA array in a two-color measurement.

The invention also provides methods and systems for predicting prognosis of colorectal cancer in a colorectal cancer patient based on a measured marker profile comprising measurements of the markers of the present invention, e.g., an expression profile comprising measurements of transcripts of at least some of the genes listed in Table 1, e.g., at least 5, 10, 20, 30, 40, 50, 60, 70, 80, or 100, or in the range of 8 to 19, in the range of 21 to 25, in the range of 27 to 42, in the range of 44 to 52, or in the range of 54 to 130 of the genes in Table 1 or funtional equivalents of such genes. The methods and systems of the invention use a prognosis predictor (also termed herein a "classifier") for predicting prognosis. The prognosis predictor can be based on any appropriate pattern recognition method (such as those described in Section 5.4) that receives an input comprising a marker profile and provides an output comprising data indicating a good prognosis or a poor prognosis. The prognosis predictor is trained with training data from a plurality of colorectal cancer patients for whom marker profiles and prognosis outcomes are known. The plurality of patients used for training the prognosis predictor is also referred to herein as the training population. The training data comprise for each patient in the training population (a) a marker profile comprising measurements of gene products of a plurality of genes, respectively, in an appropriate cell sample, e.g., a tumor cell sample, taken from the patient; and (b) prognosis outcome information (Le., information regarding whether or not survival occurred over a predetermined time period, for example, from diagnosis or from surgical resection of the cancer). Various prognosis predictors that can be used in conjunction with the present invention are described in Section 5.3., infra. In preferred embodiments, an artificial neural network or a support vector machine is used as the prognosis predictor. In some embodiments, additional patients having known marker profiles and prognosis outcomes can be used to test the accuracy of the prognosis predictor obtained using the training population. Such additional patients are also called "the testing population."

The markers in the marker sets are selected based on their ability to discriminate prognosis of colorectal cancer in a plurality of colorectal cancer patients for whom the prognosis outcomes are known. Various methods can be used to evaluate the correlation between marker levels and cancer prognosis. For example, genes whose expression levels are significantly different in tumor samples from patients who exhibit good prognosis and in tumor samples from patients who exhibit poor prognosis can be identified using an appropriate statistical method, e.g., t-test or significance analysis of microarray (SAM).

As used herein, a good prognosis predicts survival of a patient within a predetermined time period from surgical removal of tumor or from diagnosis of colorectal cancer, and a poor prognosis predicts non-survival of a patient within the time period. The predetermined time period is preferably 2, 3, 4, or 5 years. In a specific embodiment, the predetermined time period is not 3 years. For example, in one embodiment, the predetermined time period is longer than 3 years. In other embodiments, the time period is 4 or 5 years or is between 3 and 5 years.

5.1. DIAGNOSTIC AND PROGNOSTIC MARKER SETS

The invention provides molecular marker sets (of genes) that can be used for prognosis of colorectal cancer in a colorectal cancer patient based on a profile of the markers in the marker set (containing measurements of marker gene products). Table 1 lists markers that can be used to discriminate between good and poor prognosis of colorectal cancer according to the method of the invention.

Table 1 Marker genes that can be used for prognosis of colorectal cancer

Genes that are not listed in Table 1 or any of Tables 2-5, 7, and 8 (see Section 6, infra) but which are functional equivalents of any gene listed in Table 1 or any one of Tables 2-5, 7, and 8 can also be used with or in place of the gene listed in the table. A functional equivalent of a gene A refers to a gene that encodes a protein or mRNA that at least partially overlaps in physiological function in the cell to that of the protein or mRNA of gene A.

In various specific embodiments, different numbers and subcombinations of the genes listed in Table 1 are selected as the marker set, whose profile is used in the prognostic methods of the invention, as described in Section 5, supra. In various embodiments, such subcombinations include but are not limited to those genes listed in Table 2, 3, 4, 5, 7, or 8 infra in Section 6, or at least 5, 10, 15, 20, 25, 30, 40, 50, 60 or 70, or in the range of 8 to 19, in the range of 21 to 25, in the range of 27 to 42, in the range of 44 to 52, or in the range of 54 to 130 of the genes listed in Table 2, 3, 4, 5, 7, or 8, as applicable, or their respective functional equivalents.

In one embodiment, one or more genes that cluster together with one or more genes listed in a table can be selected to represent the cluster such that the marker set contains genes representing a plurality of different clusters. For example, among the 53 SAM-identified genes listed in Table 2, one gene can be selected from each cluster (see Figure IA) to constitute a marker set.

In a specific embodiment, measurements of gene products of the genes, respectively, shown in Table 2 (which is a subset of the genes listed in Table 1), or the + marked subset thereof, or their respective functional equivalents, are used for prognosis according to the invention, hi a particular embodiment, measurements of gene products of all or at least 5, 10, 15, 20, 25, 30, 40, or 50, or in the range of 8 to 52 of the genes listed in Table 2, or the + marked subset thereof, are used.

In another specific embodiment, measurments of gene products of the genes shown in Table 3 (which is a subset of the genes listed in Table 1) or their respective functional equivalents are used for prognosis according to the invention. In a particular embodiment, measurement of gene products of all or at least 5 of the genes listed in Table 3, or their respective functional equivalents, are used. In another embodiment, measurements of the gene products of the genes, respectively, shown in Table 4 (which is a subset of the genes listed in Table 1) or their respective functional equivalents are used for prognosis according to the invention. In a particular embodiment, measurements of gene products of all or at least 5, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90, or 100 of the genes listed in Table 4, or the asterisk (*) marked subset thereof, or the genes marked by M, are used. In another particular embodiment, measurements of gene products of all or in the range of 8 to 42 or in the range of 44 to 130 of the genes listed in Table 4 are used. In another particular embodiment, genes selected in at least a given number of iterations, e.g., at least 20, 30, 40, 50, 60, or 70 iterations as provided in the table, or selected in at least a given percentage of iterations, e.g., 20%, 40%, 50%, 75% or 90% of iterations, are used. Genes appearing in both the cDNA classifier and U133A-limited cDNA classifier are marked by * in Table 4. In another particular embodiment, measurements of the gene products of the genes, respectively, shown in Table 4 or their respective functional equivalents are measured using a cDNA microarray and used for prognosis according to the invention.

In another specific embodiment, measurements of gene products of the set of 43 genes shown in Table 5 (which is a subset of the genes listed in Table 1) or their respective functional equivalents are used for prognosis according to the invention. In another specific embodiment, measurements of the gene products of the 26 genes (identified by an asterisk in Table 5), which are mapped to the U133A Affymetrix^® GeneChip^® oligonucleotide-array based platform, are used. In a particular embodiment, measurements of gene products of all or at least 5, 10, 15, 20, 25, 30, or 40 of the genes listed in Table 5, or the asterisk marked subset thereof, are used. In another particular embodiment, measurements of gene products of all or in the range of 5 to 25 or in the range of 27 to 42 of the genes listed in Table 5 are used.

In another specific embodiment, measurements of gene products of the genes shown in Table 7 (which is a subset of the genes listed in Table 1) or their respective functional equivalents are used for prognosis according to the invention. In a particular embodiment, measurements of gene products of all or at least 5, 10, or 15 of the genes listed in Table 7, or their respective functional equivalents, are used.

In still another embodiment, measurements of the gene products of the genes, respectively, shown in Table 8 (which is a subset of the genes listed in Table 1) or their respective functional equivalents are used for prognosis according to the invention. In a particular embodiment, measurements of gene products of all or at least 5, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90, or 100 of the genes listed in Table 8, or the asterisk marked subset thereof, or the genes marked by M, are used. In another particular embodiment, measurements of gene products of all or in the range of 5 to 25 or in the range of 27 to 130 of the genes listed in Table 8 are used. In another particular embodiment, genes selected in at least a given number of iterations, e.g., at least 20, 30, 40, 50, 60, or 70 iterations as provided in the table, or selected in at least a given percentage of iterations, e.g., 20%, 40%, 50%, 75% or 90% of iterations, are used. Genes appearing in both the cDNA classifier and U133A- limited cDNA classifier are marked by * in Table 8. In another particular embodiment, measurements of the gene products of the genes, respectively, shown in Table 8 or their respective functional equivalents are measured using an Affymetrix^® GeneChip^® oligonucleotide-array and used for prognosis according to the invention.

In a specific embodiment, one or more of the genes listed in Table 3 can be used to subdivide a patient population into subgroups according to the expression levels of such genes, with the prognostic methods of the invention then applied to such a patient subgroup. For example, in a specific embodiment, the prognostic methods of the invention are applied to patients that have an osteopontin level higher than a predetermined threshold, e.g., the average level in subjects not having colorectal cancer or the average level in colorectal cancer patients.

In another specific embodiment, the prognostic methods of the invention are applied to patients that have a neuregulin 2 isoform 4 level higher than a predetermined threshold, e.g., the average level in subjects not having colorectal cancer or the average level in colorectal cancer patients.

In yet another embodiment, the pregnostic methods are applied to patients that have both osteopontin and neuregulin 2 isoform 4 levels higher than predetermined thresholds.

In one embodiment, a leave-one-out cross-validation method (LOOCV) (see Section 5.3., infra) is used to obtain a marker set using cDNA data of a training population of patients at 36 months of follow-up. By way of example, Table 4 lists genes selected by the LOOCV approach via t-test as discussed in Section 5.3. and Section 6. In another specific embodiment, measurements of products of a set of genes that are selected in about 75% of the training population in a leave-one-out cross-validation (LOOCV) (see Section 5.3., infra) or their respective functional equivalents are used for prognosis according to the invention.

In a specific embodiment, cross-platform mapping of marker genes can also be carried out. For example, translation of cDNA gene signature into available Affymetrix^® probe sets is carried out using the Resourcerer program (WWW.TIGR.org).

In another embodiment, SAM is used to identify a set of genes most correlated with censored survival time.

5.2. METHODS OF PREDICTING CANCER OUTCOME

The invention provides methods for predicting prognosis of colorectal cancer in a colorectal cancer patient using a measured marker profile comprising measurements of the gene products of genes, e.g., the sets of genes described in Section 5.1., supra. The prognosis indicates the patient's predicted survival at a predetermined time after surgery, e.g., at 2, 3, 4 or 5 years.

In preferred embodiments, the methods of the invention use a prognosis predictor, also called a classifier, for predicting prognosis. The prognosis predictor can be based on any appropriate pattern recognition method that receives an input comprising a marker profile and provides an output comprising data indicating a good prognosis or a poor prognosis. The prognosis predictor is trained with training data from a training population of colorectal cancer patients. Typically, the training data comprise for each of the colorectal cancer patients in the training population a marker profile comprising measurements of respective gene products of a plurality of genes in a tumor cell sample taken from the patient and prognosis outcome information. In a preferred embodiment, the training population comprises patients from each of the different stages of colorectal cancer, e.g., from adenomas (precancerous polyps), and Dukes stages A, B, C, and D. In another preferred embodiment, the training population comprises patients from each of the different TNM stages of colorectal cancer.

hi a preferred embodiment, the prognosis predictor is an artificial neural network (ANN). An ANN can be trained with the training population using any suitable method known in the art, e.g., a method described in Section 5.4.1., infra. In a specific embodiment, the ANN is a feed-forward back-propagation neural network with a single hidden layer of 10 units, a learning rate of 0.05, and a momentum of 0.2.

In another embodiment, the prognosis predictor is a support vector machine (SVM). In a specific embodiment, the SVM is a linear SVM having a dot product kernel. In still another specific embodiment, the SVM is a nonlinear SVM having a nonlinear kernel, e.g., a d-degree dot product kernel or a Gaussian kernel. An SVM can be trained with the training population using any suitable method known in the art, e.g., a method described in Section 5.4.2., infra. Kernels that can be used in conjunction with the present invention are also described in Section 5.4.2., infra.

In still other embodiments, the prognosis predictor can also be based on other classification (pattern recognition) methods, e.g., logic regression (Section 5.4.3., infra), linear or quadratic discriminant analysis (Section 5.4.4., infra), decision trees (Section 5.4.5., infra), clustering (Section 5.4.6., infra), principal component analysis (Section 5.4.7., infra), nearest neighbor classifer analysis (Section 5.4.8., infra). Such prognosis predictors can be trained with the training population using methods described in the relevant sections, infra.

The marker profile can be obtained by measuring the plurality of gene products in a tumor cell sample from the patient using a method known in the art, e.g., a method described in Section 5.5., infra.

In a specific embodiment, the prognosis method of the invention can be used for evaluating whether a colorectal cancer patient may benefit from chemotherapy. The benefit of adjuvant chemotherapy for colorectal cancer appears limited to patients with Dukes stage C disease where the cancer has metastasized to lymph nodes at the time of diagnosis. For this reason, the clinicopathological Dukes' staging system is critical for determining how adjuvant therapy is administered. Unfortunately, as noted above, Dukes' staging is not very accurate in predicting overall survival and thus its application likely results in the treatment of a large number of patients to benefit an unknown few. Alternatively, there are a number of patients who would benefit from therapy that do not receive it based on the Dukes' staging system. Accordingly, an important use of the prognosis/survival classifier of the present invention is the ability to identify those Dukes' stage B and C cases for which chemotherapy may be beneficial. Thus, in one embodiment, the invention provides a method for evaluating whether a colorectal cancer patient should be treated with chemotherapy, comprising (a) classifying said patient as having a good prognosis or a poor prognosis using a method described above; and (b) determining that said patient's predicted survival time favors treatment of the patient with chemotherapy if said patient is classified as having a poor prognosis. In one embodiment, the patient is further staged using Dukes staging.

The prognosis method of the invention can also be used in selecting patients for enrollment for a clinical trial of a chemotherapeutic agent for colorectal cancer. In one embodiment, this can be achieved using a method comprising (a) classifying each patient as having a good prognosis or a poor prognosis using a method described above; and (b) selecting patients having a poor prognosis for the clinical trial. By only enrolling patients having a poor prognosis, the efficacy of the chemotherapeutic agent can be more reliably evaluated. In a specific embodiment, the invention provides a method for enrolling colorectal cancer patients for a clinical trial of a chemotherapeutic agent for colorectal cancer, comprising (a) classifying each patient as having a good prognosis or a poor prognosis using a method described above; and (b) assigning each patient having a good prognosis to one patient group and each patient having a poor prognosis to another patient group, at least one of said patient group being enrolled in said clinical trial.

The patient is preferably a mammal, e.g., a primate or a human, and can be a non- human animal such as a dog, cat, horse, cow, mouse, rat, rabbit, etc.

5.3. METHODS FOR IDENTIFYING MARKER SETS

The invention provides methods for identifying a set of genes for prognosis of colorectal cancer. The methods make use of measured expression profiles of a plurality of genes (e.g., measurements of abundance levels of the corresponding gene products) in tumor samples from a plurality of patients whose prognosis outcomes are known. As used herein, a patient is animal inflicted with colorectal cancer. The patient can be but is not limited to a human, or, in a veterinary context, from non-human animals such as ruminants, horses, swine or sheep, or from domestic companion animals such as felines and canines. The prognosis outcomes are preferably those at a predetermined time after surgery. In one embodiment, for each of the plurality of genes a metric of correlation between expression level of the gene and survival outcome in the plurality of colorectal cancer patients is determined. One or more genes are then selected based on their metrics of correlation. The predetermined time can be any convenient time period, e.g., 2, 3, 4, or 5 years.

Prognosis markers can be obtained by identifying genes whose expression levels in good prognosis patients group are significantly different from those in poor prognosis patients. In preferred embodiments, the tumor samples from the plurality of patients are separated into a good prognosis group and a poor prognosis group for the predetermined time period. Genes whose expression levels exhibit differences between the good and poor prognosis groups to at least a predetermined level are selected as the genes whose expression levels correlate with patient survival, i.e., patient prognosis. In one embodiment, the metric of correlation of a gene with survival is an absolute t- value of a t-test. The absolute t- value can be calculated using expression levels of the gene in tumor samples from patients in the good prognosis group and expression levels of the gene in tumor samples from patients in the poor prognosis group. Genes whose t-value is higher than a certain threshold value can be selected as markers. In one embodiment, a gene is selected if the p- value of the gene corresponds to a predetermined significance level, e.g., a p-value less than 0.05.

In another embodiment, the metric of correlation of a gene is a relative difference in expression levels between the good and poor prognosis patients obtained using significance analysis of microarray (SAM). The relative difference can be calculated using expression levels of the gene in tumor samples from patients in the good prognosis group and expression levels of the gene in tumor samples from patients in the poor prognosis group according to Tusher (Tusher et al., 2001, Proc. Natl. Acad. Sci. 98:5116-5121, which is incorporated by reference herein in its entirety). In SAM, the relative difference d(i) for gene i is defined based on the ratio of change in gene expression and standard deviation in the data for that gene. The relative difference d(i) in gene expression is a t-value between expression levels in the two sample groups. Relative differences 4X0' s from permutations of the hybridizations for the samples in one prognosis group and samples in the other prognosis group can be calculated to generate a large number of controls. To find significant changes in gene expression, genes are ranked by magnitude of their d(i) values so that d(l) is the largest relative difference, d(2) is the second largest relative difference, and d(ϊ) is the ith largest relative difference. For each of the permutations, the genes are again ranked such that d_p(i) is the /th largest relative difference for permutation/?. The expected relative difference, d^i), is defined as the average over the permutations. To identify potentially significant changes in expression, a scatter plot of the observed relative difference d{ϊ) vs. the expected relative difference ddS) ^can be used. For the vast majority of genes, d(ϊ) > d_E(i), but some genes are represented by points displaced from the d(ϊ) = J^(O line by a distance greater than a threshold Δ. Genes that are within a threshold Δ are ' 'called significant.' ' To determine the number of falsely significant genes generated by SAM, horizontal cutoffs are defined as the smallest d(i) among the genes called significantly induced and the least negative d(i) among the genes called significantly repressed. The number of falsely significant genes corresponding to each permutation is computed by counting the number of genes that exceeded the horizontal cutoffs for induced and repressed genes. The estimated number of falsely significant genes is the average of the number of genes called significant from all permutations. In one embodiment, a gene is selected such that a median false detection rate is less than 40%, 28%, or 15%.

In still another embodiment, difference in expression levels of a gene in good prognosis patients versus poor prognosis patients is determined using ANOVA. A gene is selected if the difference in expression levels of the gene between the good and poor prognosis groups corresponds to a predetermined significance level.

A set of genes for prognosis of colorectal cancer can also be identified using an iterative approach. In one embodiment, a subset of patients is created by leaving out one or more patients in a plurality of patients having known outcomes after obtaining tumor samples. A metric of correlation between expression level of each gene and survival outcome in the subset of colorectal cancer patients can then be determined using a method described above. One or more genes are then selected based on the metric of correlation. The process is repeated for a plurality of iterations, each with a different subset of patients by leaving out one or more patients who are different from the one or more patients left out in any other iteration. Genes that are selected in at least a predetermined percentage of all iterations are included for the marker set.

In one embodiment, for each iteration, the plurality of genes is ranked according to the correlation metric. A given number of genes that are ranked with the highest correlation are selected. In preferred embodiments, a set of 20, 50, 70 or 100 genes are selected from the rank list. The sets of genes obtained in different iterations are then compared to identify one or more genes that are selected in at least a predetermined percentage of all iterations. In one embodiment, the predetermined percentage is 50%, 75% or 90%. In a preferred embodiment, the subset of patients is created by leaving out one of the plurality of patients. The embodiment is also termed a "leave-one-out" method.

hi one embodiment, a total of T tumor samples from patients having known prognosis (the training population of patients) is used to identify marker genes that can be used for prognosis of colorectal cancer. Preferably, T is at least 50, 100, 200 or 500. In one embodiment, the samples are frozen colorectal cancer samples selected from based on evidence for good (survival > a predetermined number of months) or poor (survival < the predetermined number of months) prognosis. Among the T samples of the training population, M samples are poor prognosis and N samples are good prognosis cases. In one embodiment, the training population comprises samples in all different Dukes' stages, i.e., A, B, C, and D cases. Dukes' stage A cases are very rare. Thus, in one embodiment, P samples of adenomas are included in the T samples to represent very good prognosis cases, hi other embodiment, the training population consists of samples from patients belonging to one or more specific Dukes stages, e.g., Dukes stage B, Dukes stage C, or Dukes stages B and C. Such training population can be used to identify genes that can be used for prognosis of colorectal cancer in patients belonging to the respective one or more Dukes stages. Survival is measured as last contact minus collection date for living patients, or date of death minus collection date for patients who had died. Samples are microdissected ( >80% tumor cells) by frozen section guidance and RNA was extracted using Trizol followed by secondary purification on RNAEasy columns.

In one embodiment, identification of marker genes can be combined with construction and validation of a prognosis predictor. In one embodiment, after selecting a set of genes in each iteration, a prognosis predictor that receives an input comprising a marker profile comprising expression levels of the one or more selected genes and provides an output comprising data indicating a good prognosis or a poor prognosis is trained with training data from the subset of patients. The training data comprise for each patient a marker profile comprising measurements of the one or more selected genes in a tumor cell sample taken from the patient and prognosis information. The prognosis predictor is then used to determine a prognosis for at least one of the one or more patients who are left out. Thus, the accuracy of the prognosis predictor can be determined based on rate of true or false predictions from the plurality of iterations, hi one embodiment, after the ^iterations have been completed, another prognosis predictor is constructed using the one or more genes that are selected in at least a predetermined percentage of all iterations. This prognosis predictor can be trained with training data from the subset of patients or all patients and used as the prognosis predictor for predicting prognosis for new patients.

In a preferred embodiment, the prognosis predictor is an artificial neural network (ANN). In a specific embodiment, the ANN is a feed-forward back-propagation neural network with a single hidden layer of 10 units, a learning rate of 0.05, and a momentum of 0.2.

In another preferred embodiment, the prognosis predictor is a support vector machine (SVM). In another specific embodiment, the SVM is a linear SVM having a dot product kernel, hi still another specific embodiment, the SVM is a nonlinear SVM having a nonlinear kernel, e.g., a J-degree dot product kernel or a Gaussian kernel. Exemplary kernels are described in Section 5.3.2.

hi a specific embodiment, a leave-one-out cross-validation (LOOCV) technique is used for constructing and validating a neural network-based classifier. The test samples are classified as having "good" or "poor" prognosis based on survival for more or less than 36 months, respectively. Using the leave-one-out cross-validation approach also provides the ability to rank the selected genes. The number of times a particular gene is chosen can be an indicator of the usefulness of that gene for general classification and may imply biological significance. Therefore, genes that are consistently selected by the t-test are selected for the marker set. m one embodiment, a set of 43 core genes is identified in about 75% of the LOOCV iterations. The set of 43 genes is listed in Table 5, infra.

hi another specific embodiment, a leave-one-out cross-validation (LOOCV) technique is used for constructing and validating a SVM-based classifier. The test samples are classified as having "good" or "poor" prognosis based on survival for more or less than 24 months, respectively. The SVM classifier is a linear SVM classifier. A set of 20 genes is identified in about 14% of the LOOCV iterations. The set of 20 genes is listed in Table 7, infra.

In one embodiment, the molecular classifier is obtained using an iterative approach using iterations of two distinct steps: gene selection using an appropriate statistical method, e.g., a t-test, and classification using an appropriate prognosis predictor, e.g., a neural network. Both steps are taken after the one or more test samples are left out, e.g., from the leave-one-out cross-validation, to avoid bias from the gene selection step. A predetermined number of the top genes as ranked by absolute value of the t statistic using a t test are selected for each cross-validation step. In a specific embodiment, a feed-forward back-propagation neural network with a single hidden layer of 10 units, learning rate of 0.05, and momentum of 0.2 is constructed. Training occurred for a maximum of 500 epochs or until a zero misclassification error is achieved on the training set.

The obtained prognosis predictor can be evaluated for its accuracy. In one embodiment, differences between Kaplan-Meier curves are evaluated using the log-rank test, which is well known to a skilled person in the art. This can be performed both for the initial survival analysis and for the classifier results. In accordance with the present invention, the classifier can split the samples into two groups: those predicted as having good or poor prognosis. Classifier accuracy can then be reported to the user both as overall accuracy and as specificity/sensitivity. In one embodiment, a McNemar's Chi-Square test is used to compare the molecular classifier with the use of a Dukes' staging classifier. In a related embodiment, several permutations of the dataset (i.e., 1,000 permutations) are used to measure the significance of the classifier results as compared to chance.

Thus, the invention provides a method for identifying a set of genes for prognosis of colorectal cancer, comprising: (a) generating a subset of patients by leaving out one or more patients in a plurality of patients having known outcomes at a predetermined time after obtaining tumor samples; (b) determining for each of a plurality of genes a metric of correlation between abundance level of said gene and survival outcome in said subset of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples; (c) selecting one or more genes based on said metric of correlation; (d) repeating steps (a)-(c) for a plurality of iterations, each with a different subset of patients by leaving out one or more patients in said plurality, wherein said one or more patients are different from any previous iteration; and (e) selecting one or more genes that are selected in at least a predetermined percentage of all iterations.

In one embodiment, said step (c) is carried out by a method comprising (cl) ranking said plurality of genes according to said metric; and (c2) selecting a given number of genes that ranked with the highest correlation. In a specific embodiment, said given number is 50 and said predetermined percentage is about 75%. In one embodiment, said step (b) is carried out by a method comprising (bl) dividing said plurality of colorectal cancer patients into a first group consisting of one or more patients who are living at said predetermined time and a second group consisting of one or more patients who are not living at said predetermined time; and (b2) determining a difference in expression levels between said first group and said second group, wherein said difference represents said metric of correlation, and wherein said difference for each of said plurality of genes is a t- value of a t-test of expression levels of said gene between said first group and said second group. In a specific embodiment, said selecting is carried out by a method comprising selecting a given number of top-ranked genes according to the absolute t- value of the genes. In another specific embodiment, said selecting is carried out by a method comprising selecting a gene if the p-value of said gene corresponds to a predetermined significance level, e.g., a p-value less than 0.05. In one embodiment, said predetermined time is 3 years.

The invention also provides a method for constructing prognosis predictor for prognosis of colorectal cancer, comprising: (a) generating a subset of patients by leaving out one or more patients in a plurality of patients having known outcomes at a predetermined time after obtaining tumor samples; (b) determining for each of a plurality of genes a metric of correlation between expression level of said gene and survival outcome in a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples from a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples; (c) selecting one or more genes based on said metric of correlation; (d) training a prognosis predictor, wherein said prognosis predictor receives an input comprising a marker profile comprising expression levels of said one or more genes selected in step (c) and provides an output comprising data indicating a good prognosis or a poor prognosis, with training data from said subset of patients, wherein said training data comprise for each of said subset of patients a marker profile comprising measurements of said one or more genes in a tumor cell sample taken from said patient and prognosis information; (e) determining a prognosis for at least one of said one or more patients who are left out in step (a); (f) repeating steps (a)-(e) for a plurality of iteractions, each with a different subset of patients by leaving out one or more patients in said plurality, wherein said one or more patients are different from any previous iteration; (g) selecting one or more genes that are selected in at least a predetermined percentage of all iterations; and (h) training a prognosis predictor, wherein said prognosis predictor receives an input comprising a marker profile comprising expression levels of said one or more genes selected in step (g) and provides an output comprising data indicating a good prognosis or a poor prognosis, with training data from said subset of patients, wherein said training data comprise for each of said plurality of patients a marker profile comprising measurements of said one or more genes in a tumor cell sample taken from said patient and prognosis information.

In one embodiment, said step (c) is carried out by a method comprising (cl) ranking ' said plurality of genes according to said metric; and (c2) selecting a given number of genes that ranked with the highest correlation.

In a preferred embodiment, said prognosis predictor is an artificial neural network (ANN).

In one embodiment, said step (b) is carried out by a method comprising (bl) dividing said plurality of colorectal cancer patients into a first group consisting of one or more patients who are living at said predetermined time and a second group consisting of one or more patients who are not living at said predetermined time; and (b2) determining a difference in expression levels between said first group and said second group, wherein said difference represents said metric of correlation. In a specific embodiment, said selecting is carried out by a method comprising selecting a given number of top-ranked genes according to the absolute t- value of the genes. In another specific embodiment, said selecting is carried out by a method comprising selecting a gene if the p- value of said gene corresponds to a predetermined significance level, e.g., a p-value less than 0.05. In one embodiment, said predetermined time is 3 years.

The invention also provides a computer system comprising a processor, and a memory coupled to said processor and encoding one or more programs, wherein said one or more programs cause the processor to carry out a method described above.

The invention also provides a computer program product for use in conjunction with a computer having a processor and a memory connected to the processor, said computer program product comprising a computer readable storage medium having a computer program mechanism encoded thereon, wherein said computer program mechanism may be loaded into the memory of said computer and cause said computer to carry out a method described above. 5.4. STATISTICAL METHODS

Various known statistical pattern recognition methods can be used in conjunction with the present invention. A prognosis predictor based on any of such methods can be constructed using the marker profiles and prognosis data of training patients. Such a prognosis predictor can then be used to predict prognosis of a colorectal patient based on the patient's marker profile. The methods can also be used to identify markers that discriminate between a good and poor prognosis using a marker profile and prognosis data of training patients.

5.4.1. ARTIFICIAL NEURAL NETWORK

In some embodiments, a neural network is used. A neural network can be constructed for a selected set of molecular markers of the invention. A neural network is a two-stage regression or classification model. A neural network has a layered structure that includes a layer of input units (and the bias) connected by a layer of weights to a layer of output units. For regression, the layer of output units typically includes just one output unit. However, neural networks can handle multiple quantitative responses in a seamless fashion.

In multilayer neural networks, there are input units (input layer), hidden units (hidden layer), and output units (output layer). There is, furthermore, a single bias unit that is connected to each unit other than the input units. Neural networks are described in Duda et al., 2001, Pattern Classification, Second Edition, John Wiley & Sons, Inc., New York; and Hastie et at, 2001, The Elements of Statistical Learning, Springer- Verlag, New York.

The basic approach to the use of neural networks is to start with an untrained network, present a training pattern, e.g., marker profiles from training patients, to the input layer, and to pass signals through the net and determine the output, e.g., the prognosis of the training patients, at the output layer. These outputs are then compared to the target values; any difference corresponds to an error. This error or criterion function is some scalar function of the weights and is minimized when the network outputs match the desired outputs. Thus, the weights are adjusted to reduce this measure of error. For regression, this error can be sum-of- squared errors. For classification, this error can be either squared error or cross-entropy (deviation). See, e.g., Hastie et al., 2001, The Elements of Statistical Learning, Springer- Verlag, New York. Three commonly used training protocols are stochastic, batch, and on-line. In stochastic training, patterns are chosen randomly from the training set and the network weights are updated for each pattern presentation. Multilayer nonlinear networks trained by gradient descent methods such as stochastic back-propagation perform a maximum- likelihood estimation of the weight values in the model defined by the network topology. In batch training, all patterns are presented to the network before learning takes place. Typically, in batch training, several passes are made through the training data. In online training, each pattern is presented once and only once to the net.

In some embodiments, consideration is given to starting values for weights. If the weights are near zero, then the operative part of the sigmoid commonly used in the hidden layer of a neural network (see, e.g., Hastie et al, 2001, The Elements of Statistical Learning, Springer- Verlag, New York) is roughly linear, and hence the neural network collapses into an approximately linear model. In some embodiments, starting values for weights are chosen to be random values near zero. Hence the model starts out nearly linear, and becomes nonlinear as the weights increase. Individual units localize to directions and introduce nonlinearities where needed. Use of exact zero weights leads to zero derivatives and perfect symmetry, and the algorithm never moves. Alternatively, starting with large weights often leads to poor solutions.

Since the scaling of inputs determines the effective scaling of weights in the bottom layer, it can have a large effect on the quality of the final solution. Thus, in some embodiments, at the outset all expression values are standardized to have mean zero and a standard deviation of one. This ensures all inputs are treated equally in the regularization process, and allows one to choose a meaningful range for the random starting weights. With standardization inputs, it is typical to take random uniform weights over the range [-0.7, +0.7].

A recurrent problem in the use of networks with one hidden layer is the optimal number of hidden units to use in the network. The number of inputs and outputs of the network are determined by the problem to be solved. In the present invention, the number of inputs for a given neural network can be the number of molecular markers in the selected set of molecular markers of the invention. The number of output for the neural network will typically be just one. However, in some embodiment more than one output is used so that more than just two states can be defined by the network. If too many hidden units are used in a neural network, the network will have too many degrees of freedom and is trained too long, there is a danger that the network will overfit the data. If there are too few hidden units, the training set cannot be learned. Generally speaking, however, it is better to have too many hidden units than too few. With too few hidden units, the model might not have enough flexibility to capture the nonlinearities in the data; with too many hidden units, the extra weight can be shrunk towards zero if appropriate regularization or pruning, as described below, is used. In typical embodiments, the number of hidden units is somewhere in the range of 5 to 100, with the number increasing with the number of inputs and number of training cases.

One general approach to determining the number of hidden units to use is to apply a regularization approach. In the regularization approach, a new criterion function is constructed that depends not only on the classical training error, but also on classifier complexity. Specifically, the new criterion function penalizes highly complex models; searching for the minimum in this criterion is to balance error on the training set with error on the training set plus a regularization term, which expresses constraints or desirable properties of solutions:

J ⁼ 'pat "^" "J reg-

The parameter λ is adjusted to impose the regularization more or less strongly. In other words, larger values for λ will tend to shrink weights towards zero: typically cross-validation with a validation set is used to estimate λ. This validation set can be obtained by setting aside a random subset of the training population. Other forms of penalty can also be used, for example the weight elimination penalty (see, e.g., Hastie et ah, 2001, The Elements of Statistical Learning, Springer- Verlag, New York).

Another approach to determine the number of hidden units to use is to eliminate - prune - weights that are least needed. In one approach, the weights with the smallest magnitude are eliminated (set to zero). Such magnitude-based pruning can work, but is nonoptimal; sometimes weights with small magnitudes are important for learning and training data. In some embodiments, rather than using a magnitude-based pruning approach, WaId statistics are computed. The fundamental idea in WaId Statistics is that they can be used to estimate the importance of a hidden unit (weight) in a model. Then, hidden units having the least importance are eliminated (by setting their input and output weights to zero). Two algorithms in this regard are the Optimal Brain Damage (OBD) and the Optimal Brain Surgeon (OBS) algorithms that use second-order approximation to predict how the training error depends upon a weight, and eliminate the weight that leads to the smallest increase in training error.

Optimal Brain Damage and Optimal Brain Surgeon share the same basic approach of training a network to local minimum error at weight w, and then pruning a weight that leads to the smallest increase in the training error. The predicted functional increase in the error for a change in full weight vector δw is:

B²J where — r- is the Hessian matrix. The first term vanishes because we are at a local dw minimum in error; third and higher order terms are ignored. The general solution for minimizing this function given the constraint of deleting one weight is:

Here, u_q is the unit vector along the qth direction in weight space and L_q is approximation to the saliency of the weight q - the increase in training error if weight q is pruned and the other weights updated δw. These equations require the inverse of H. One method to calculate this inverse matrix is to start with a small value, H^¹ = Cf¹I, where α is a small parameter - effectively a weight constant. Next the matrix is updated with each pattern according to

TT^-1Y Y^r TT"¹

Η-l _ ττ-1 ^■"^■»! ^Am-H^A»H-lⁿm ^lm ,„.+.1l — ■£•. ,„

Λ- \T^T TT-¹Y ^l~ ^Λm+l^X1m ^Λm+l a

where the subscripts correspond to the pattern being presented and a_m decreases with m. After the full training set has been presented, the inverse Hessian matrix is given by H^{"1 ;} H^"1. In algorithmic form, the Optimal Brain Surgeon method is:

begin initialize n_#, w, θ train a reasonably large network to minimum error

do compute H^"1 by Eqn. 1

q* <^■- arg min w² /(2[H^-1J Λqq ) * ( vsaliency L_q) q 9 ^x L —

(saliency L_q)

until J(W) > θ

return w

end

The Optimal Brain Damage method is computationally simpler because the calculation of the inverse Hessian matrix in line 3 is particularly simple for a diagonal matrix. The above algorithm terminates when the error is greater than a criterion initialized to be θ. Another approach is to change line 6 to terminate when the change in J(w) due to elimination of a weight is greater than some criterion value.

In some embodiments, a back-propagation neural network (see, for example Abdi, 1994, "A neural network primer", J. Biol System. 2, 247-283) containing a single hidden layer of ten neurons (ten hidden units) found in EasyNN-Plus version 4.Og software package (Neural Planner Software Inc.) is used. In a specific example, parameter values within the EasyNN-Plus program are set as follows: a learning rate of 0.05, and a momentum of 0.2. In some embodiments in which the EasyNN-Plus version 4.Og software package is used, "outlier" samples are identified by performing twenty independently-seeded trials involving 20,000 learning cycles each.

5.4.2. SUPPORT VECTOR MACHINE

In some embodiments of the present invention, support vector machines (SVMs) are used to classify subjects using expression profiles of marker genes described in the present invention. General description of SVM can be found in, for example, Cristianini and Shawe- Taylor, 2000, An Introduction to Support Vector Machines, Cambridge University Press, Cambridge, Boser et al, 1992, "A training algorithm for optimal margin classifiers, in Proceedings of the 5^l Annual ACM Workshop on Computational Learning Theory, ACM Press, Pittsburgh, PA, pp. 142-152; Vapnik, 1998, Statistical Learning Theory, Wiley, New York; Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc.; Hastie, 2001, The Elements of Statistical Learning, Springer, New York; and Furey et at, 2000, Bioinformatics 16, 906-914. Applications of SVM in biological applications are described in Jaakkola et al., Proceedings of the 7* International Conference on Intelligent Systems for Molecular Biology, AAAI Press, Menlo Park, CA (1999); Brown et al, Proc. Natl. Acad. ScL 97(l):262-67 (2000); Zien et al, Bioinformatics, 16(9):799-807 (2000); Furey et al, Bioinformatics, 16(10):906-914 (2000)

In one approach, when a SVM is used, the gene expression data is standardized to have mean zero and unit variance and the members of a training population are randomly divided into a training set and a test set. For example, in one embodiment, two thirds of the members of the training population are placed in the training set and one third of the members of the training population are placed in the test set. The expression values for a selected set of genes of the present invention is used to train the SVM. Then the ability for the trained SVM to correctly classify members in the test set is determined. In some embodiments, this computation is performed several times for a given selected set of molecular markers. In each iteration of the computation, the members of the training population are randomly assigned to the training set and the test set. Then, the quality of the combination of molecular markers is taken as the average of each such iteration of the SVM computation.

Support vector machines map a given set of binary labeled training data to a high- dimensional feature space and separate the two classes of data with a maximum margin hyperplane. In general, this hyperplane corresponds to a nonlinear decision boundary in the input space. Let I e i^ c 3i" be the input vectors, ye {-1,+1} be the labels, and φ : R₀ → F be the mapping from input space to feature space. Then the SVM learning algorithm finds a hyperplane (w,b) such that the quantity

r= mmy,{(w,^(X,.))-b} i

is maximized, where the vector w has the same dimensionality as F, b is a real number, and γ is called the margin. The corresponding decision function is then f(X) = sign({w,φ(X))-b)

This minimum occurs when

w = ∑a_iy_iφ(X_i) i

where { cζ} are positive real numbers that maximize

subject to

∑ CXJ₁ = 0,01_; > 0 i

The decision function can equivalently be expressed as

From this equation it can be seen that the (%. associated with the training point Xi expresses the strength with which that point is embedded in the final decision function. A remarkable property of this alternative representation is that only a subset of the points will be associated with a non-zero Oi. These points are called support vectors and are the points that lie closest to the separating hyperplane. The sparseness of the a vector has several computational and learning theoretic consequences. It is important to note that neither the learning algorithm nor the decision function needs to represent explicitly the image of points in the feature space, ^(X₁) , since both use only the dot products between such images,

\ώ{X _t),φ(X _jlj . Hence, if one were given a function K(X, Y) = (^(X), φ(X)) , one could learn and use the maximum margin hyperplane in the feature space without ever explicitly performing the mapping. For each continuous positive definite function AT(X, Y) there exists a mapping φ such that K(X,Y) = {φ(X), φ(X)) for all X₅Ye R₀ (Mercer's Theorem). The function UT(X, Y) is called the kernel function. The use of a kernel function allows the support vector machine to operate efficiently in a nonlinear high-dimensional feature spaces without being adversely affected by the dimensionality of that space. Indeed, it is possible to work with feature spaces of infinite dimension. Moreover, Mercer's theorem makes it possible to learn in the feature space without even knowing φ and F. The matrix

K_y = (^(X_i),^(X_y)\ is called the kernel matrix. Finally, note that the learning algorithm is a quadratic optimization problem that has only a global optimum. The absence of local minima is a significant difference from standard pattern recognition techniques such as neural networks. For moderate sample sizes, the optimization problem can be solved with simple gradient descent techniques. In the presence of noise, the standard maximum margin algorithm described above can be subject to overfitting, and more sophisticated techniques should be used. This problem arises because the maximum margin algorithm always finds a perfectly consistent hypothesis and does not tolerate training error. Sometimes, however, it is necessary to trade some training accuracy for better predictive power. The need for tolerating training error has led to the development the soft-margin and the margin-distribution classifiers. One of these techniques replaces the kernel matrix in the training phase as follows:

K ±- K + λl

while still using the standard kernel function in the decision phase. By tuning λ, one can control the training error, and it is possible to prove that the risk of misclassifying unseen points can be decreased with a suitable choice of λ.

If instead of controlling the overall training error one wants to control the trade-off between false positives and false negatives, it is possible to modify K as follows:

K <r- K + AD

where D is a diagonal matrix whose entries are either J⁺ or d^~ , in locations corresponding to positive and negative examples. It is possible to prove that this technique is equivalent to controlling the size of the «;^• in a way that depends on the size of the class, introducing a bias for larger Cζ- in the class with smaller d. This in turn corresponds to an asymmetric margin; i.e., the class with smaller d will be kept further away from the decision boundary. In some cases, the extreme imbalance of the two classes, along with the presence of noise, creates a situation in which points from the minority class can be easily mistaken for mislabelled points. Enforcing a strong bias against training errors in the minority class provides protection agaist such errors and forces the SVM to make the positive examples support vectors. Thus, choosing J⁺ =— and d^~ =—_z provides a heuristic way to automatically n n adjust the relative importance of the two classes, based on their respective cardinalities. This technique effectively controls the trade-off between sensitivity and specificity.

In the present invention, a linear kernel can be used. The similarity between two marker profiles X and Y can be the dot product X-Y. In one embodiment, the kernel is

K(X, Y) = X-Y + 1

In another embodiment, a kernel of degree d is used

K(X, Y) = (X-Y + if , where d can be either 2, 3, ...

In still another embodiment, a Gaussian kernel is used

where σ is the width of the Gaussian.

5.4.3. LOGISTIC REGRESSION

In some embodiments, the prognosis predictor is based on a regression model, preferably a logistic regression model. Such a regression model includes a coefficient for each of the molecular markers in a selected set of molecular markers of the invention. In such embodiments, the coefficients for the regression model are computed using, for example, a maximum likelihood approach, hi particular embodiments, molecular marker data from the two prognosis groups is used and the dependent variable is the prognosis of the patient for which molecular marker characteristic data are from.

Some embodiments of the present invention provide generalizations of the logistic regression model that handle multicategory (polychotomous) responses. Such embodiments can be used to discriminate an organism into one or three or more prognosis groups. Such regression models use multicategory logit models that simultaneously refer to all pairs of categories, and describe the odds of response in one category instead of another. Once the model specifies logits for a certain (J-I) pairs of categories, the rest are redundant. See, for example, Agresti, An Introduction to Categorical Data Analysis, John Wiley & Sons, Inc., 1996, New York, Chapter 8, which is hereby incorporated by reference.

5.4.4. DISCRIMINANT ANALYSIS

Linear discriminant analysis (LDA) attempts to classify a subject into one of two categories based on certain object properties. In other words, LDA tests whether object attributes measured in an experiment predict categorization of the objects. LDA typically requires continuous independent variables and a dichotomous categorical dependent variable. In the present invention, the expression values for the selected set of molecular markers of the invention across a subset of the training population serve as the requisite continuous independent variables. The prognosis group classification of each of the members of the training population serves as the dichotomous categorical dependent variable.

LDA seeks the linear combination of variables that maximizes the ratio of between- group variance and within-group variance by using the grouping information. Implicitly, the linear weights used by LDA depend on how the expression of a molecular marker across the training set separates in the two groups (e.g. , a group that has osteoarthritis and a group that does not have osteoarthritis) and how this gene expression correlates with the expression of other genes. In some embodiments, LDA is applied to the data matrix of the N members in the training sample by K genes in a combination of genes described in the present invention. Then, the linear discriminant of each member of the training population is plotted. Ideally, those members of the training population representing a first subgroup (e.g. those subjects that do not have osteoarthritis) will cluster into one range of linear discriminant values (e.g., negative) and those member of the training population representing a second subgroup (e.g. those subjects that have osteoarthritis) will cluster into a second range of linear discriminant values (e.g., positive). The LDA is considered more successful when the separation between the clusters of discriminant values is larger. For more information on linear discriminant analysis, see Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc; and Hastie, 2001, The Elements of Statistical Learning, Springer, New York; Venables & Ripley, 1997, Modern Applied Statistics with s-plus, Springer, New York.

Quadratic discriminant analysis (QDA) takes the same input parameters and returns the same results as LDA. QDA uses quadratic equations, rather than linear equations, to produce results. LDA and QDA are interchangeable, and which to use is a matter of preference and/or availability of software to support the analysis. Logistic regression takes the same input parameters and returns the same results as LDA and QDA.

5.4.5. DECISION TREES

In some embodiments of the present invention, decision trees are used to classify patients using expression data for a selected set of molecular markers of the invention.

Decision tree algorithms belong to the class of supervised learning algorithms. The aim of a decision tree is to induce a classifier (a tree) from real-world example data. This tree can be used to classify unseen examples which have not been used to derive the decision tree.

A decision tree is derived from training data. An example contains values for the different attributes and what class the example belongs. In one embodiment, the training data is expression data for a combination of genes described in the present invention across the training population.

The following algorithm describes a decision tree derivation:

Tree(Examples ,Class Attributes)

Create a root node

If all Examples have the same Class value, give the root this label

Else if Attributes is empty label the root according to the most common value

Else begin

Calculate the information gain for each attribute

Select the attribute A with highest information gain and make this the root attribute

For each possible value, v, of this attribute

Add a new branch below the root, corresponding to A = v

Let Examples(v) be those examples with A = v If Examples(v) is empty, make the new branch a leaf node labeled with the most common value among Examples

Else let the new branch be the tree created by

Tree(Examples(v),Class, Attributes - {A})

end

A more detailed description of the calculation of information gain is shown in the following. If the possible classes vj of the examples have probabilities P(vO then the information content I of the actual answer is given by:

1(P(V₁)_^P(V_n)) = J] - P(V₁)IOg₁ P(v₍) *^=I

The I- value shows how much information we need in order to be able to describe the outcome of a classification for the specific dataset used. Supposing that the dataset contains p positive (e.g. has osteoarthritis) and n negative (e.g. healthy) examples (e.g. individuals), the information contained in a correct answer is:

I{-^—_t ) = --^- iog₂ -^- IGg₂ p + fi p + n p + n p+n p+n p + n

where log₂ is the logarithm using base two. By testing single attributes the amount of information needed to make a correct classification can be reduced. The remainder for a specific attribute A (e.g. a gene) shows how much the information that is needed can be reduced.

Ttemaindm^>(Λ)

"v" is the number of unique attribute values for attribute A in a certain dataset, "i" is a certain attribute value, "pi" is the number of examples for attribute A where the classification is positive (e.g. cancer), "ni" is the number of examples for attribute A where the classification is negative (e.g. healthy). The information gain of a specific attribute A is calculated as the difference between the information content for the classes and the remainder of attribute A:

GaIn(A) = J(— — ,— ^- ) -Re mamd&iΛ) p+n p÷ n

The information gain is used to evaluate how important the different attributes are for the classification (how well they split up the examples), and the attribute with the highest information.

In general there are a number of different decision tree algorithms, many of which are described in Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc. Decision tree algorithms often require consideration of feature processing, impurity measure, stopping criterion, and pruning. Specific decision tree algorithms include, cut are not limited to classification and regression trees (CART), multivariate decision trees, ID3, and C4.5.

In one approach, when an exemplary embodiment of a decision tree is used, the gene expression data for a selected set of molecular markers of the invention across a training population is standardized to have mean zero and unit variance. The members of the training population are randomly divided into a training set and a test set. For example, in one embodiment, two thirds of the members of the training population are placed in the training set and one third of the members of the training population are placed in the test set. The expression values for a select combination of genes described in the present invention is used to construct the decision tree. Then, the ability for the decision tree to correctly classify members in the test set is determined. In some embodiments, this computation is performed several times for a given combination of molecular markers. In each iteration of the computation, the members of the training population are randomly assigned to the training set and the test set. Then, the quality of the combination of molecular markers is taken as the average of each such iteration of the decision tree computation.

5.4.6. CLUSTERING

In some embodiments, the expression values for a selected set of molecular markers of the invention are used to cluster a training set. For example, consider the case in which ten genes described in the present invention are used. Each member m of the training population will have expression values for each of the ten genes. Such values from a member m in the training population define the vector:

where Xj_m is the expression level of the 1^th gene in organism m. If there are m organisms in the training set, selection of i genes will define m vectors. Note that the methods of the present invention do not require that each the expression value of every single gene used in the vectors be represented in every single vector m. hi other words, data from a subject in which one of the i^ώ genes is not found can still be used for clustering. In such instances, the missing expression value is assigned either a "zero" or some other normalized value. In some embodiments, prior to clustering, the gene expression values are normalized to have a mean value of zero and unit variance.

Those members of the training population that exhibit similar expression patterns across the training group will tend to cluster together. A particular combination of genes of the present invention is considered to be a good classifier in this aspect of the invention when the vectors cluster into the trait groups found in the training population. For instance, if the training population includes patients with good or poor prognosis, a clustering classifier will cluster the population into two groups, with each group uniquely representing either good or poor prognosis.

Clustering is described on pages 211-256 of Duda and Hart, Pattern Classification and Scene Analysis, 1973, John Wiley & Sons, Inc., New York. As described in Section 6.7 of Duda, the clustering problem is described as one of finding natural groupings in a dataset. To identify natural groupings, two issues are addressed. First, a way to measure similarity (or dissimilarity) between two samples is determined. This metric (similarity measure) is used to ensure that the samples in one cluster are more like one another than they are to samples in other clusters. Second, a mechanism for partitioning the data into clusters using the similarity measure is determined. Similarity measures are discussed in Section 6.7 of Duda, where it is stated that one way to begin a clustering investigation is to define a distance function and to compute the matrix of distances between all pairs of samples in a dataset. If distance is a good measure of similarity, then the distance between samples in the same cluster will be significantly less than the distance between samples in different clusters. However, as stated on page 215 of Duda, clustering does not require the use of a distance metric. For example, a nonmetric similarity function s(x, x') can be used to compare two vectors x and x'. Conventionally, s(x, x') is a symmetric function whose value is large when x and x' are somehow "similar". An example of a nonmetric similarity function s(x, x') is provided on page 216 of Duda.

Once a method for measuring "similarity" or "dissimilarity" between points in a dataset has been selected, clustering requires a criterion function that measures the clustering quality of any partition of the data. Partitions of the data set that extremize the criterion function are used to cluster the data. See page 217 of Duda. Criterion functions are discussed in Section 6.8 of Duda.

More recently, Duda et al., Pattern Classification, 2^nd edition, John Wiley & Sons,

Inc. New York, has been published. Pages 537-563 describe clustering in detail. More information on clustering techniques can be found in Kaufman and Rousseeuw, 1990, Finding Groups in Data: An Introduction to Cluster Analysis, Wiley, New York, NY; Everitt, 1993, Cluster analysis (3d ed.), Wiley, New York, NY; and Backer, 1995, Computer-Assisted Reasoning in Cluster Analysis, Prentice Hall, Upper Saddle River, New Jersey. Particular exemplary clustering techniques that can be used in the present invention include, but are not limited to, hierarchical clustering (agglomerative clustering using nearest-neighbor algorithm, farthest-neighbor algorithm, the average linkage algorithm, the centroid algorithm, or the sum-of-squares algorithm), k-means clustering, fuzzy k-means clustering algorithm, and Jarvis-Patrick clustering.

5.4.7. PRINCIPAL COMPONENT ANALYSIS

Principal component analysis (PCA) has been proposed to analyze gene expression data. Principal component analysis is a classical technique to reduce the dimensionality of a data set by transforming the data to a new set of variable (principal components) that summarize the features of the data. See, for example, Jolliffe, 1986, Principal Component Analysis, Springer, New York. Principal components (PCs) are uncorrelate and are ordered such that the &^ώ PC has the Mh largest variance among PCs. The k⁰¹ PC can be interpreted as the direction that maximizes the variation of the projections of the data points such that it is orthogonal to the first k - 1 PCs. The first few PCs capture most of the variation in the data set. In contrast, the last few PCs are often assumed to capture only the residual 'noise' in the data.

PCA can also be used to create a prognosis predictor in accordance with the present invention. In such an approach, vectors for a selected set of molecular markers of the invention can be constructed in the same manner described for clustering above. In fact, the set of vectors, where each vector represents the expression values for the select genes from a particular member of the training population, can be considered a matrix. In some embodiments, this matrix is represented in a Free- Wilson method of qualitative binary description of monomers (Kubinyi, 1990, 3D QSAR in drug design theory methods and applications, Pergamon Press, Oxford, pp 589-638), and distributed in a maximally compressed space using PCA so that the first principal component (PC) captures the largest amount of variance information possible, the second principal component (PC) captures the second largest amount of all variance information, and so forth until all variance information in the matrix has been accounted for.

Then, each of the vectors (where each vector represents a member of the training population) is plotted. Many different types of plots are possible. In some embodiments, a one-dimensional plot is made. In this one-dimensional plot, the value for the first principal component from each of the members of the training population is plotted. In this form of plot, the expectation is that members of a first group (e.g. good prognosis patients) will cluster in one range of first principal component values and members of a second group (e.g., poor prognosis patients) will cluster in a second range of first principal component values.

In one example, the training population comprises two groups: good prognosis patients and poor prognosis patients. The first principal component is computed using the molecular marker expression values for the select genes of the present invention across the entire training population data set. Then, each member of the training set is plotted as a function of the value for the first principal component. In this example, those members of the training population in which the first principal component is positive are the good prognosis patients and those members of the training population in which the first principal component is negative are poor prognosis patients. In some embodiments, the members of the training population are plotted against more than one principal component. For example, in some embodiments, the members of the training population are plotted on a two-dimensional plot in which the first dimension is the first principal component and the second dimension is the second principal component. In such a two-dimensional plot, the expectation is that members of each subgroup represented in the training population will cluster into discrete groups. For example, a first cluster of members in the two-dimensional plot will represent subjects with mild osteoarthritis, a second cluster of members in the two-dimensional plot will represent subjects with moderate osteoarthritis, and so forth.

In some embodiments, the members of the training population are plotted against more than two principal components and a determination is made as to whether the members of the training population are clustering into groups that each uniquely represents a subgroup found in the training population. In some embodiments, principal component analysis is performed by using the R mva package (Anderson, 1973, Cluster Analysis for applications, Academic Press, New York 1973; Gordon, Classification, Second Edition, Chapman and Hall, CRC, 1999.). Principal component analysis is further described in Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc.

5.4.8. NEAREST NEIGHBOR CLASSIFIER ANALYSIS

Nearest neighbor classifiers are memory-based and require no model to be fit. Given a query point XQ, the k training points X(_r), r, ..., k closest in distance to xo are identified and then the point XQ is classified using the k nearest neighbors. Ties can be broken at random. In some embodiments, Euclidean distance in feature space is used to determine distance as:

Typically, when the nearest neighbor algorithm is used, the expression data used to compute the linear discriminant is standardized to have mean zero and variance 1. In the present invention, the members of the training population are randomly divided into a training set and a test set. For example, in one embodiment, two thirds of the members of the training population are placed in the training set and one third of the members of the training population are placed in the test set. Profiles of a selected set of molecular markers of the invention represents the feature space into which members of the test set are plotted. Next, the ability of the training set to correctly characterize the members of the test set is computed. In some embodiments, nearest neighbor computation is performed several times for a given combination of genes of the present invention. In each iteration of the computation, the members of the training population are randomly assigned to the training set and the test set. Then, the quality of the combination of genes is taken as the average of each such iteration of the nearest neighbor computation.

The nearest neighbor rule can be refined to deal with issues of unequal class priors, differential misclassification costs, and feature selection. Many of these refinements involve some form of weighted voting for the neighbors. For more information on nearest neighbor analysis, see Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc; and Hastie, 2001, The Elements of Statistical Learning, Springer, New York.

5.4.9. EVOLUTIONARY METHODS

Inspired by the process of biological evolution, evolutionary methods of classifier design employ a stochastic search for an optimal classifier. In broad overview, such methods create several classifiers - a population - from measurements of gene products of the present invention. Each classifier varies somewhat from the other. Next, the classifiers are scored on expression data across the training population. In keeping with the analogy with biological evolution, the resulting (scalar) score is sometimes called the fitness. The classifiers are ranked according to their score and the best classifiers are retained (some portion of the total population of classifiers). Again, in keeping with biological terminology, this is called survival of the fittest. The classifiers are stochastically altered in the next generation - the children or offspring. Some offspring classifiers will have higher scores than their parent in the previous generation, some will have lower scores. The overall process is then repeated for the subsequent generation: The classifiers are scored and the best ones are retained, randomly altered to give yet another generation, and so on. In part, because of the ranking, each generation has, on average, a slightly higher score than the previous one. The process is halted when the single best classifier in a generation has a score that exceeds a desired criterion value. More information on evolutionary methods is found in, for example, Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc.

5.4.10. BAGGING. BOOSTING AND THE RANDOM SUBSPACE METHOD Bagging, boosting and the random subspace method are combining techniques that can be used to improve weak classifiers. These techniques are designed for, and usually applied to, decision trees. In addition, Skurichina and Duin provide evidence to suggest that such techniques can also be useful in linear discriminant analysis.

In bagging, one samples the training set, generating random independent bootstrap replicates, constructs the classifier on each of these, and aggregates them by a simple majority vote in the final decision rule. See, for example, Breiman, 1996, Machine Learning 24, 123-140; and Efron & Tibshirani, An Introduction to Bootstrap, Chapman & Hall, New York, 1993.

In boosting, classifiers are constructed on weighted versions of the training set, which are dependent on previous classification results. Initially, all objects have equal weights, and the first classifier is constructed on this data set. Then, weights are changed according to the performance of the classifier. Erroneously classified objects (molecular markers in the data set) get larger weights, and the next classifier is boosted on the reweighted training set. In this way, a sequence of training sets and classifiers is obtained, which is then combined by simple majority voting or by weighted majority voting in the final decision. See, for example, Freund & Schapire, "Experiments with a new boosting algorithm," Proceedings 13^th International Conference on Machine Learning, 1996, 148-156.

To illustrate boosting, consider the case where there are two phenotypic groups exhibited by the population under study, phenotype 1 (e.g., poor prognosis patients), and phenotype 2 (e.g. , good prognosis patients). Given a vector of molecular markers X, a classifier G(X) produces a prediction taking one of the type values in the two value set: { phenotype 1, phenotype 2}. The error rate on the training sample is

where N is the number of subjects in the training set (the sum total of the subjects that have either phenotype 1 or phenotype 2). For example, if there are 30 good prognosis patients and 48 good prognosis patients, N is 78.

A weak classifier is one whose error rate is only slightly better than random guessing. In the boosting algorithm, the weak classification algorithm is repeatedly applied to modified versions of the data, thereby producing a sequence of weak classifiers G_m(x), m, = 1, 2, ..., M. The predictions from all of the classifiers in this sequence are then combined through a weighted majority vote to produce the final prediction:

M

G(x) = sign \ ∑ Σa_mG_m(x)

HJ=I

Here Ci₁, α_2> ..._, a^ are computed by the boosting algorithm and their purpose is to weigh the contribution of each respective G_m(x). Their effect is to give higher influence to the more accurate classifiers in the sequence.

The data modifications at each boosting step consist of applying weights wi, w₂, ..., W_n to each of the training observations fa, yi), i = 1, 2, ..., N. Initially all the weights are set to w,- = 1/N, so that the first step simply trains the classifier on the data in the usual manner. For each successive iteration m = 2, 3, ..., M the observation weights are individually modified and the classification algorithm is reapplied to the weighted observations. At stem m, those observations that were misclassified by the classifier G_m.i(x) induced at the previous step have their weights increased, whereas the weights are decreased for those that were classified correctly. Thus as iterations proceed, observations that are difficult to correctly classify receive ever-increasing influence. Each successive classifier is thereby forced to concentrate on those training observations that are missed by previous ones in the sequence.

The exemplary boosting algorithm is summarized as follows:

1. Initialize the observation weights Wi = 1/N, i = 1, 2, ..., N.

2. For m = 1 to M:

(a) Fit a classifier G_m(x) to the training set using weights Wu

(b) Compute

(c) Compute α_m=log((l-err_m)/err_m).

(d) Setw_t < W₁. -exp[α_m -I(y_t ≠ G_m(x,))],i = 1,2,...,N.

3 . Output G(X) = sign |∑^α_mG_M(x)]

In the algorithm, the current classifier G_m(x) is induced on the weighted observations at line 2a. The resulting weighted error rate is computed at line 2b. Line 2c calculates the weight a_m given to G_m(x) in producing the final classifier G(x) (line 3). The individual weights of each of the observations are updated for the next iteration at line 2d. Observations misclassified by G_m(x) have their weights scaled by a factor exp(α_m), increasing their relative influence for inducing the next classifier G_m+;(x) in the sequence. In some embodiments, modifications of the Freund and Schapire, 1997, Journal of Computer and System Sciences 55, pp. 119-139, boosting method are used. See, for example, Hasti et al., The Elements of Statistical Learning, 2001, Springer, New York, Chapter 10. In some embodiments, boosting or adaptive boosting methods are used.

In some embodiments, modifications of Freund and Schapire, 1997, Journal of Computer and System Sciences 55, pp. 119-139, are used. For example, in some embodiments, feature preselection is performed using a technique such as the nonparametric scoring methods of Park et al., 2002, Pac. Symp. Biocomput. 6, 52-63. Feature preselection is a form of dimensionality reduction in which the genes that discriminate between classifications the best are selected for use in the classifier. Then, the LogitBoost procedure introduced by Friedman et al., 2000, Ann Stat 28, 337-407 is used rather than the boosting procedure of Freund and Schapire. In some embodiments, the boosting and other classification methods of Ben-Dor et al, 2000, Journal of Computational Biology 7, 559-583 are used in the present invention. In some embodiments, the boosting and other classification methods of Freund and Schapire, 1997, Journal of Computer and System Sciences 55, 119- 139, are used.

In the random subspace method, classifiers are constructed in random subspaces of the data feature space. These classifiers are usually combined by simple majority voting in the final decision rule. See, for example, Ho, "The Random subspace method for constructing decision forests," IEEE Trans Pattern Analysis and Machine Intelligence, 1998; 20(8): 832-844.

5.4.11. OTHER ALGORITHMS

The pattern classification and statistical techniques described above are merely examples of the types of models that can be used to construct a model for classification.

Moreover, combinations of the techniques described above can be used. Some combinations, such as the use of the combination of decision trees and boosting, have been described. However, many other combinations are possible. In addition, in other techniques in the art such as Projection Pursuit and Weighted Voting can be used to construct a prognosis predictor.

5.5. SAMPLE COLLECTION

In the present invention, gene products, such as target polynucleotide molecules or proteins, are extracted from a sample taken from an individual afflicted with colorectal cancer. The sample may be collected in any clinically acceptable manner, but must be collected such that marker-derived polynucleotides (Le., RNA) are preserved (if gene expression is to be measured) or proteins are preserved (if encoded proteins are to be measured), hi one embodiment, samples can be microdissected (>80% tumor cells) by frozen section guidance and RNA extraction performed using Trizol followed by secondary purification on RNAEasy columns. In another embodiment, samples can be paraffin- embedded tissue sections (see, e.g., U.S. Patent Application Publication No.

2005/0048542A1, which is incorporated by reference herein in its entirety). The mRNA profiles of paraffin-embedded tissue samples are preferably obtained using quantitative reverse transcriptase polymerase chain reaction qRT-PCR (see Section 5.6.7., infra).

In a specific embodiment, mRNA or nucleic acids derived therefrom (Le., cDNA or amplified RNA or amplied DNA) are preferably labeled distinguishably from polynucleotide molecules of a reference sample, and both are simultaneously or independently hybridized to a microarray comprising some or all of the markers or marker sets or subsets described above. Alternatively, mRNA or nucleic acids derived therefrom may be labeled with the same label as the reference polynucleotide molecules, wherein the intensity of hybridization of each at a particular probe is compared.

A sample may comprise any clinically relevant tissue sample, such as a tumor biopsy or fine needle aspirate, or a sample of body fluid, such as blood, plasma, serum, lymph, ascitic fluid, cystic fluid, or urine. The sample may be taken from a human, or, in a veterinary context, from non-human animals such as ruminants, horses, swine or sheep, or from domestic companion animals such as felines and canines.

Methods for preparing total and poly(A)+ RNA are well known and are described generally in Sambrook et al, MOLECULAR CLONING - A LABORATORY MANUAL (2ND ED.), VoIs. 1-3, Cold Spring Harbor Laboratory, Cold Spring Harbor, New York (1989)) and Ausubel et ah, CURRENT PROTOCOLS IN MOLECULAR BIOLOGY, vol. 2, Current Protocols Publishing, New York (1994)). Preferably, total RNA, or total mRNA (ρoly(A)+ RNA) is meausured in the methods of the invention directly or indirectly (e.g., via measuring cDNA or cRNA).

RNA may be isolated from eukaryotic cells by procedures that involve lysis of the cells and denaturation of the proteins contained therein. Cells of interest include wild-type cells (Le., non-cancerous), drug-exposed wild-type cells, tumor- or tumor-derived cells, modified cells, normal or tumor cell line cells, and drug-exposed modified cells. Preferably, the cells are breast cancer tumor cells.

Additional steps may be employed to remove DNA. Cell lysis may be accomplished with a nonionic detergent, followed by microcentrifugation to remove the nuclei and hence the bulk of the cellular DNA. In one embodiment, RNA is extracted from cells of the various types of interest using guanidinium thiocyanate lysis followed by CsCl centrifugation to separate the RNA from DNA (Chirgwin et ah, Biochemistry 18:5294-5299 (1979)). PoIy(A)+ RNA is selected by selection with oligo-dT cellulose (see Sambrook et al, MOLECULAR CLONING - A LABORATORY MANUAL (2ND ED.), VOIS. 1-3, Cold Spring Harbor Laboratory, Cold Spring Harbor, New York (1989). Alternatively, separation of RNA from DNA can be accomplished by organic extraction, for example, with hot phenol or phenol/chloroform/isoamyl alcohol. If desired, RNase inhibitors may be added to the lysis buffer. Likewise, for certain cell types, it may be desirable to add a protein denaturation/digestion step to the protocol.

For many applications, it is desirable to preferentially enrich mRNA with respect to other cellular RNAs, such as transfer RNA (tRNA) and ribosomal RNA (rRNA). Most mRNAs contain a poly(A) tail at their 3' end. This allows them to be enriched by affinity chromatography, for example, using oligo(dT) or poly(U) coupled to a solid support, such as cellulose or Sephadex™ (see Ausubel et al., CURRENT PROTOCOLS IN MOLECULAR BIOLOGY, vol. 2, Current Protocols Publishing, New York (1994). Once bound, poly(A)+ mRNA is eluted from the affinity column using 2 mM EDTA/0.1% SDS.

In a specific embodiment, total RNA or total mRNA from cells is used in the methods of the invention. The source of the RNA can be cells of an animal, e.g., human, mammal, primate, non-human animal, dog, cat, mouse, rat, bird, etc. In specific embodiments, the method of the invention is used with a sample containing total mRNA or total RNA from 1 x 10⁶ cells or less. In another embodiment, proteins can be isolated from the foregoing sources, by methods known in the art, for use in expression analysis at the protein level.

Probes to the homologs of the marker sequences disclosed herein can be employed preferably when non-human nucleic acid is being assayed.

5.6. DETERMINATION OF ABUNDANCE LEVELS OF GENE PRODUCTS

The abundance levels of the gene products of the genes in a sample may be determined by any means known in the art. The levels may be determined by isolating and determining the level (i.e., amount) of nucleic acid transcribed from each marker gene. Alternatively, or additionally, the level of specific proteins encoded by a marker gene may be determined.

The levels of transcripts of specific marker genes can be accomplished by determining the amount of mRNA, or polynucleotides derived therefrom, present in a sample. Any method for determining RNA levels can be used. For example, RNA is isolated from a sample and separated on an agarose gel. The separated RNA is then transferred to a solid support, such as a filter. Nucleic acid probes representing one or more markers are then hybridized to the filter by northern hybridization, and the amount of marker-derived RNA is determined. Such determination can be visual, or machine-aided, for example, by use of a densitometer. Another method of determining RNA levels is by use of a dot-blot or a slot- blot, hi this method, RNA, or nucleic acid derived therefrom, from a sample is labeled. The RNA or nucleic acid derived therefrom is then hybridized to a filter containing oligonucleotides derived from one or more marker genes, wherein the oligonucleotides are placed upon the filter at discrete, easily-identifiable locations. Hybridization, or lack thereof, of the labeled RNA to the filter-bound oligonucleotides is determined visually or by densitometer. Polynucleotides can be labeled using a radiolabel or a fluorescent (Le., visible) label.

These examples are not intended to be limiting; other methods of determining RNA abundance are known in the art.

The levels of transcripts of particular marker genes may also be assessed by determining the level of the specific protein expressed from the marker genes. This can be accomplished, for example, by separation of proteins from a sample on a polyacrylamide gel, followed by identification of specific marker-derived proteins using antibodies in a western blot. Alternatively, proteins can be separated by two-dimensional gel electrophoresis systems. Two-dimensional gel electrophoresis is well-known in the art and typically involves isoelectric focusing along a first dimension followed by SDS-PAGE electrophoresis along a second dimension. See, e.g., Hames et al, 1990, GEL ELECTROPHORESIS OF PROTEINS: A PRACTICAL APPROACH, IRL Press, New York; Shevchenko et al., Proc. Nat'l Acad. Sci. USA 93:1440-1445 (1996); Sagliocco et al, Yeast 12:1519-1533 (1996); Lander, Science 274:536-539 (1996). The resulting electropherograms can be analyzed by numerous techniques, including mass spectrometric techniques, western blotting and immunoblot analysis using polyclonal and monoclonal antibodies.

Alternatively, marker-derived protein levels can be determined by constructing an antibody microarray in which binding sites comprise immobilized, preferably monoclonal, antibodies specific to a plurality of protein species encoded by the cell genome. Preferably, antibodies are present for a substantial fraction of the marker-derived proteins of interest. Methods for making monoclonal antibodies are well known (see, e.g., Harlow and Lane, 1988, ANTIBODIES: A LABORATORY MANUAL, Cold Spring Harbor, New York, which is incorporated in its entirety for all purposes). In one embodiment, monoclonal antibodies are raised against synthetic peptide fragments designed based on genomic sequence of the cell. With such an antibody array, proteins from the cell are contacted to the array, and their binding is assayed with assays known in the art. Generally, the expression, and the level of expression, of proteins of diagnostic or prognostic interest can be detected through immunohistochemical staining of tissue slices or sections.

Finally, levels of transcripts of marker genes in a number of tissue specimens may be characterized using a "tissue array" (Kononen et al, Nat. Med 4(7):844-7 (1998)). In a tissue array, multiple tissue samples are assessed on the same microarray. The arrays allow in situ detection of RNA and protein levels; consecutive sections allow the analysis of multiple samples simultaneously.

5.6.1. MICROARRAYS

In preferred embodiments, polynucleotide microarrays are used to measure expression so that the expression status of each of the markers above is assessed simultaneously. Generally, microarrays according to the invention comprise a plurality of markers informative for prognosis, or outcome determination, for a particular disease or condition, and, in particular, for individuals having specific combinations of genotypic or phenotypic characteristics of the disease or condition {i.e., that are prognosis-informative for a particular patient subset).

The invention also provides a microarray comprising for each of a plurality of genes, said genes being all or at least 5, 10, 20, 30, 40, 50 or 70 of the genes listed in Table 1 or any of Tables 2-5, 7 and 8, one or more polynucleotide probes complementary and hybridizable to a sequence in said gene, wherein polynucleotide probes complementary and hybridizable to said genes constitute at least 50%, 60%, 70%, 80%, 90%, 95%, or 98% of the probes on said microarray. In a particular embodiment, the invention provides such a microarray wherein the plurality of genes comprises the 43 genes listed in Table 5 or the 53 genes listed in Table 2 or the 7 genes listed in Table 3. The microarray can be in a sealed container.

The microarrays of the invention preferably comprise at least 2, 3, 4, 5, 7, 10, 15, 20,

25, 30, 35, 40, 45, 50, 75, 100, 150, 200 or more of markers, or all of the markers, or any combination of markers, identified as prognosis-informative within a patient subset, e.g., within Table 1 or any of Tables 2-5, 7 and 8. The actual number of informative markers the microarray comprises will vary depending upon the particular condition of interest. In specific embodiments, the invention provides polynucleotide arrays in which the prognosis markers identified for a particular patient subset comprise at least 50%, 60%, 70%, 80%, 85%, 90%, 95% or 98% of the probes on the array. In another specific embodiment, the microarray comprises a plurality of probes, wherein said plurality of probes comprise probes complementary and hybridizable to at least 75% of the prognosis-informative markers identified for a particular patient subset. Microarrays of the invention, of course, may comprise probes complementary and hybridizable to prognosis-informative markers for a plurality of the patient subsets, or for each patient subset, identified for a particular condition. In another embodiment, therefore, the microarray of the invention comprises a plurality of probes complementary and hybridizable to at least 75% of the prognosis-informative markers identified for each patient subset identified for the condition of interest, and wherein the probes, in total, are at least 50% of the probes on said microarray.

In yet another specific embodiment, the microarray is a commercially-available cDNA microarray that comprises probes to at least five markers identified by the methods described herein. Preferably, a commercially-available cDNA microarray comprises probes to all of the markers identified by the methods described herein as being informative for a patient subset for a particular condition. However, such a microarray may comprise at least 5, 10, 15 or 25 of such markers, up to the maximum number of markers identified.

In one embodiment, the invention provides oligonucleotide or cDNA arrays comprising probes hybridizable to the genes corresponding to each of the marker sets described above, e.g., as shown in Table 1 or any one of Tables 2-5, 7 and 8. In another embodiment, a whole-genome cDNA array can be used. Any of the microarrays described herein may be provided in a sealed container in a kit.

The invention provides microarrays containing probes useful for the prognosis of colon cancer patients. In particular, the invention provides polynucleotide arrays comprising probes to a subset or subsets of at least 5, 10, 15, 20, 25 or more of the genetic markers, or up to the full set of markers, in Table 1, which distinguish between patients with good and poor prognosis. In certain embodiments, therefore, the invention provides microarrays comprising probes for a plurality of the genes for which markers are listed in Table 1. In a specific embodiment, the microarray of the invention comprises 1, 2, 3, 4, 5 or 10 of the markers in Table 1. In other embodiments, the microarray of the invention contains each of the markers in Table 1. In another embodiment, the microarray contains all of the markers shown in Table 1.

In specific embodiments, the invention provides polynucleotide arrays in which the colon cancer prognosis markers described herein in Table 1 comprise at least 50%, 60%, 70%, 80%, 85%, 90%, 95% or 98% of the probes on said array. In another specific embodiment, the microarray comprises a plurality of probes, wherein said plurality of probes comprise probes complementary and hybridizable to transcripts of at least 75% of the genes for which markers are listed in Table 1.

In yet another specific embodiment, the microarray is a commercially-available cDNA microarray that comprises probes to at least five of the markers listed in Table 1.

Preferably, a commercially-available cDNA microarray comprises all of the markers listed in Table 1. However, such a microarray may comprise probes to at least 5, 10, 15 or 25 of the markers in Table 1, up to the maximum number of markers in Table 1, and may comprise probes to all of the markers in Table 1. hi a specific embodiment of the microarrays used in the methods disclosed herein comprise probes to the markers that are all or a portion of Table 1 make up at least 50%, 60%, 70%, 80%, 90%, 95% or 98% of the probes on the microarray.

General methods pertaining to the construction of microarrays comprising the marker sets and/or subsets above are described in the following sections.

In a specific embodiment, the Affymetrix^® Human Genome U133 (HG-U133) Set, consisting of two GeneChip® arrays, is used in accordance with known methods. The

Human Genome U133 (HG-U133) Set contains almost 45,000 probe sets representing more than 39,000 transcripts derived from approximately 33,000 well-substantiated human genes. This set design uses sequences selected from GenBank®, dbEST, and RefSeq. The sequence clusters were created from the UniGene database (Build 133, April 20, 2001). They were then refined by analysis and comparison with a number of other publicly available databases including the Washington University EST trace repository and the University of California, Santa Cruz Golden Path human genome database (April 2001 release).

hi another embodiment, the HG-Ul 33 A array is used in accordance with the methods of the invention. The HG-U133 A array includes representation of the RefSeq database sequences and probe sets related to sequences previously represented on the Human Genome U95Av2 array. The HG-U133B array contains primarily probe sets representing EST clusters. In another embodiment, the U133 Plus 2.0 GeneChip® is used in the invention. The U133 Plus 2.0 GeneChip® represents over 47,000 transcripts.

In another embodiment, a cDNA based microarray is used. In one embodiment, TIGR' s 32,488-element spotted cDNA arrays is used. The TIGR cDNA array contains 31,872 human cDNAs representing 30,849 distinct transcripts: 23,936 unique TIGR TCs and 6,913 ESTs, 10 exogenous controls printed 36 times, and 4 negative controls printed 36-72 times.

5.6.2. CONSTRUCTION OFMICROARRAYS

Microarrays are prepared by selecting probes which comprise a polynucleotide sequence, and then immobilizing such probes to a solid support or surface. For example, the probes may comprise DNA sequences, RNA sequences, or copolymer sequences of DNA and RNA. The polynucleotide sequences of the probes may also comprise DNA and/or RNA analogues, or combinations thereof. For example, the polynucleotide sequences of the probes may be full or partial fragments of genomic DNA. The polynucleotide sequences of the probes may also be synthesized nucleotide sequences, such as synthetic oligonucleotide sequences. The probe sequences can be synthesized either enzymatically in vivo, enzymatically in vitro (e.g., by PCR), or non-enzymatically in vitro.

The probe or probes used in the methods of the invention are preferably immobilized to a solid support which may be either porous or non-porous. For example, the probes of the invention may be polynucleotide sequences which are attached to a nitrocellulose or nylon membrane or filter covalently at either the 3' or the 5' end of the polynucleotide. Such hybridization probes are well known in the art (see, e.g., Sambrook et al., MOLECULAR CLONING - A LABORATORY MANUAL (2ND ED.), VOIS. 1-3, Cold Spring Harbor Laboratory, Cold Spring Harbor, New York (1989). Alternatively, the solid support or surface may be a glass or plastic surface, hi a particularly preferred embodiment, hybridization levels are measured to microarrays of probes consisting of a solid phase on the surface of which are immobilized a population of polynucleotides, such as a population of DNA or DNA mimics, or, alternatively, a population of RNA or RNA mimics. The solid phase may be a nonporous or, optionally, a porous material such as a gel.

In preferred embodiments, a microarray comprises a support or surface with an ordered array of binding (e.g., hybridization) sites or "probes" each representing one of the markers described herein. Preferably the microarrays are addressable arrays, and more preferably positionally addressable arrays. More specifically, each probe of the array is preferably located at a known, predetermined position on the solid support such that the identity (Le., the sequence) of each probe can be determined from its position in the array (i.e., on the support or 'surface). In preferred embodiments, each probe is covalently attached to the solid support at a single site.

Microarrays can be made in a number of ways, of which several are described below. However produced, microarrays share certain characteristics. The arrays are reproducible, allowing multiple copies of a given array to be produced and easily compared with each other. Preferably, microarrays are made from materials that are stable under binding (e.g., nucleic acid hybridization) conditions. The microarrays are preferably small, e.g., between 1 cm² and 25 cm², between 12 cm² and 13 cm², or 3 cm². However, larger arrays are also contemplated and may be preferable, e.g., for use in screening arrays. Preferably, a given binding site or unique set of binding sites in the microarray will specifically bind (e.g., hybridize) to the product of a single gene in a cell (e.g., to a specific mRNA, or to a specific cDNA derived therefrom). However, in general, other related or similar sequences will cross hybridize to a given binding site.

The microarrays of the present invention include one or more test probes, each of which has a polynucleotide sequence that is complementary to a subsequence of RNA or DNA to be detected. Preferably, the position of each probe on the solid surface is known. Indeed, the microarrays are preferably positionally addressable arrays. Specifically, each probe of the array is preferably located at a known, predetermined position on the solid support such that the identity (i.e., the sequence) of each probe can be determined from its position on the array (i.e., on the support or surface).

According to the invention, the microarray is an array (i.e., a matrix) in which each position represents one of the markers described herein. For example, each position can contain a DNA or DNA analogue based on genomic DNA to which a particular RNA or cDNA transcribed from that genetic marker can specifically hybridize. The DNA or DNA analogue can be, e.g., a synthetic oligomer or a gene fragment. In one embodiment, probes representing each of the markers is present on the array. In a preferred embodiment, the array comprises probes for each of the markers listed in Table 1 or any one of Tables 2-5, 7 and 8. 5.6.3. PREPARINGPROBES FORMICROARRAYS

As noted above, the "probe" to which a particular polynucleotide molecule specifically hybridizes according to the invention contains a complementary genomic polynucleotide sequence. The probes of the microarray preferably consist of nucleotide sequences of no more than 1,000 nucleotides. In some embodiments, the probes of the array consist of nucleotide sequences of 10 to 1,000 nucleotides. In a preferred embodiment, the nucleotide sequences of the probes are in the range of 10-200 nucleotides in length and are genomic sequences of a species of organism, such that a plurality of different probes is present, with sequences complementary and thus capable of hybridizing to the genome of such a species of organism, sequentially tiled across all or a portion of such genome. In other specific embodiments, the probes are in the range of 10-30 nucleotides in length, in the range of 10-40 nucleotides in length, in the range of 20-50 nucleotides in length, in the range of 40- 80 nucleotides in length, in the range of 50-150 nucleotides in length, in the range of 80-120 nucleotides in length, and most preferably are 60 nucleotides in length.

The probes may comprise DNA or DNA "mimics" (e.g., derivatives and analogues) corresponding to a portion of an organism's genome, hi another embodiment, the probes of the microarray are complementary RNA or RNA mimics. DNA mimics are polymers composed of subunits capable of specific, Watson-Crick-like hybridization with DNA, or of specific hybridization with RNA. The nucleic acids can be modified at the base moiety, at the sugar moiety, or at the phosphate backbone. Exemplary DNA mimics include, e.g., phosphorothioates .

DNA can be obtained, e.g., by polymerase chain reaction (PCR) amplification of genomic DNA or cloned sequences. PCR primers are preferably chosen based on a known sequence of the genome that will result in amplification of specific fragments of genomic DNA. Computer programs that are well known in the art are useful in the design of primers with the required specificity and optimal amplification properties, such as Oligo version 5.0 (National Biosciences). Typically each probe on the microarray will be between 10 bases and 50,000 bases, usually between 300 bases and 1,000 bases in length. PCR methods are well known in the art, and are described, for example, in Innis et al. , eds., PCR PROTOCOLS: A GUIDE TO METHODS AND APPLICATIONS, Academic Press Inc., San Diego, CA (1990). It will be apparent to one skilled in the art that controlled robotic systems are useful for isolating and amplifying nucleic acids. An alternative, preferred means for generating the polynucleotide probes of the microarray is by synthesis of synthetic polynucleotides or oligonucleotides, e.g., using N- phosphonate or phosphoramidite chemistries (Froehler et al , Nucleic Acid Res. 14:5399-5407 (1986); McBride et al, Tetrahedron Lett. 24:246-248 (1983)). Synthetic sequences are typically between about 10 and about 500 bases in length, more typically between about 20 and about 100 bases, and most preferably between about 40 and about 70 bases in length. In some embodiments, synthetic nucleic acids include non-natural bases, such as, but by no means limited to, inosine. As noted above, nucleic acid analogues may be used as binding sites for hybridization. An example of a suitable nucleic acid analogue is peptide nucleic acid (see, e.g., Egholm et al, Nature 363:566-568 (1993); U.S. Patent No. 5,539,083).

Probes are preferably selected using an algorithm that takes into account binding energies, base composition, sequence complexity, cross-hybridization binding energies, and secondary structure. See Friend et al., International Patent Publication WO 01/05935, published January 25, 2001; Hughes et al, Nat. Biotech. 19:342-7 (2001).

A skilled artisan will also appreciate that positive control probes, e.g., probes known to be complementary and hybridizable to sequences in the target polynucleotide molecules, and negative control probes, e.g., probes known to not be complementary and hybridizable to sequences in the target polynucleotide molecules, should be included on the array. In one embodiment, positive controls are synthesized along the perimeter of the array. In another embodiment, positive controls are synthesized in diagonal stripes across the array. In still another embodiment, the reverse complement for each probe is synthesized next to the position of the probe to serve as a negative control. In yet another embodiment, sequences from other species of organism are used as negative controls or as "spike-in" controls.

5.6.4. ATTACHINGPROBES TOTHESOLID SURFACE The probes are attached to a solid support or surface, which may be made, e.g., from glass, plastic (e.g., polypropylene, nylon), polyacrylamide, nitrocellulose, gel, or other porous or nonporous material. A preferred method for attaching the nucleic acids to a surface is by printing on glass plates, as is described generally by Schena et al, Science 270:467-470 (1995). This method is especially useful for preparing microarrays of cDNA (See also, DeRisi et al, Nature Genetics 14:457-460 (1996); Shalon et al , Genome Res. 6 :639-645 (1996); and Schena et al, Proc. Natl. Acad. ScL U.S.A. 93:10539-11286 (1995)). A second preferred method for making microarrays is by making high-density oligonucleotide arrays. Techniques are known for producing arrays containing thousands of oligonucleotides complementary to defined sequences, at defined locations on a surface using photolithographic techniques for synthesis in situ (see, Fodor et al, 1991, Science 251:767- 773; Pease et al, 1994, Proc. Natl. Acad. Sci. U.S.A. 91:5022-5026; Lockhart et al, 1996, Nature Biotechnology 14:1675; U.S. Patent Nos. 5,578,832; 5,556,752; and 5,510,270) or other methods for rapid synthesis and deposition of defined oligonucleotides (Blanchard et al, Biosensors & Bioelectronics 11:687-690). When these methods are used, oligonucleotides {e.g., 60-mers) of known sequence are synthesized directly on a surface such as a derivatized glass slide. Usually, the array produced is redundant, with several oligonucleotide molecules per RNA.

Other methods for making microarrays, e.g., by masking (Maskos and Southern, 1992, Nuc. Acids. Res. 20:1679-1684), may also be used. In principle, and as noted supra, any type of array, for example, dot blots on a nylon hybridization membrane (see Sambrook et al. , MOLECULAR CLONING - A LABORATORY MANUAL (2ND ED.), VOIS. 1-3, Cold Spring Harbor Laboratory, Cold Spring Harbor, New York (1989)) could be used. However, as will be recognized by those skilled in the art, very small arrays will frequently be preferred because hybridization volumes will be smaller.

In one embodiment, the arrays of the present invention are prepared by synthesizing polynucleotide probes on a support. In such an embodiment, polynucleotide probes are attached to the support covalently at either the 3' or the 5' end of the polynucleotide.

In a particularly preferred embodiment, microarrays of the invention are manufactured by means of an ink jet printing device for oligonucleotide synthesis, e.g., using the methods and systems described by Blanchard in U.S. Pat. No. 6,028,189; Blanchard et al., 1996, Biosensors and Bioelectronics 11:687-690; Blanchard, 1998, in Synthetic DNA Arrays in Genetic Engineering, Vol. 20, J.K. Setlow, Ed., Plenum Press, New York at pages 111-123. Specifically, the oligonucleotide probes in such microarrays are preferably synthesized in arrays, e.g., on a glass slide, by serially depositing individual nucleotide bases in "microdroplets" of a high surface tension solvent such as propylene carbonate. The microdroplets have small volumes (e.g., 100 pL or less, more preferably 50 pL or less) and are separated from each other on the microarray (e.g., by hydrophobic domains) to form circular surface tension wells which define the locations of the array elements (i.e., the different probes). Microarrays manufactured by this ink-jet method are typically of high density, preferably having a density of at least about 2,500 different probes per 1 cm . The polynucleotide probes are attached to the support covalently at either the 3 ' or the 5' end of the polynucleotide.

5.6.5. TARGET LABELING AND HYBRIDIZATION TQ MICROARRAYS

The polynucleotide molecules which may be analyzed by the present invention (the "target polynucleotide molecules") may be from any clinically relevant source, but are expressed RNA or a nucleic acid derived therefrom (e.g., cDNA or amplified RNA derived from cDNA that incorporates an RNA polymerase promoter), including naturally occurring nucleic acid molecules, as well as synthetic nucleic acid molecules. In one embodiment, the target polynucleotide molecules comprise RNA, including, but by no means limited to, total cellular RNA, poly(A)⁺ messenger RNA (mRNA) or fraction thereof, cytoplasmic mRNA, or RNA transcribed from cDNA (i.e., cRNA; see, e.g., Linsley & Schelter, U.S. Patent Application No. 09/411,074, filed October 4, 1999, or U.S. Patent Nos. 5,545,522, 5,891,636, or 5,716,785). Methods for preparing total and poly(A)⁺ RNA are well known in the art, and are described generally, e.g., in Sambrook et al., MOLECULAR CLONING - A LABORATORY MANUAL (2ND ED.), VOIS. 1-3, Cold Spring Harbor Laboratory, Cold Spring Harbor, New York (1989). In one embodiment, RNA is extracted from cells of the various types of interest in this invention using guanidinium thiocyanate lysis followed by CsCl centrifugation (Chirgwin et al, 1979, Biochemistry 18:5294-5299). In another embodiment, total RNA is extracted using a silica gel-based column, commercially available examples of which include RNeasy (Qiagen, Valencia, California) and StrataPrep (Stratagene, La Jolla, California). In an alternative embodiment, which is preferred for S. cerevisiae, RNA is extracted from cells using phenol and chloroform, as described in Ausubel et al, eds., 1989, CURRENT PROTOCOLS IN MOLECULAR BIOLOGY, Vol. Ill, Green Publishing Associates, Inc., John Wiley

& Sons, Inc., New York, at pp. 13.12.1-13.12.5). PoIy(A)⁺ RNA can be selected, e.g., by selection with oligo-dT cellulose or, alternatively, by oligo-dT primed reverse transcription of total cellular RNA. In one embodiment, RNA can be fragmented by methods known in the art, e.g. , by incubation with ZnCl₂, to generate fragments of RNA. In another embodiment, the polynucleotide molecules analyzed by the invention comprise cDNA, or PCR products of amplified RNA or cDNA. In one embodiment, total RNA, mRNA, or nucleic acids derived therefrom, is isolated from a sample taken from a colorectal cancer patient. Target polynucleotide molecules that are poorly expressed in particular cells may be enriched using normalization techniques (Bonaldo et al, 1996, Genome Res. 6:791-806).

As described above, the target polynucleotides are detectably labeled at one or more nucleotides. Any method known in the art may be used to detectably label the target polynucleotides. Preferably, this labeling incorporates the label uniformly along the length of the RNA, and more preferably, the labeling is carried out at a high degree of efficiency. One embodiment for this labeling uses oligo-dT primed reverse transcription to incorporate the label; however, conventional methods of this method are biased toward generating 3' end fragments. Thus, in a preferred embodiment, random primers {e.g., 9-mers) are used in reverse transcription to uniformly incorporate labeled nucleotides over the full length of the target polynucleotides. Alternatively, random primers may be used in conjunction with PCR methods or T7 promoter-based in vitro transcription methods in order to amplify the target polynucleotides.

In a preferred embodiment, the detectable label is a luminescent label. For example, fluorescent labels, bioluminescent labels, chemiluminescent labels, and colorimetric labels may be used in the present invention. In a highly preferred embodiment, the label is a fluorescent label, such as a fluorescein, a phosphor, a rhodamine, or a polymethine dye derivative. Examples of commercially available fluorescent labels include, for example, fluorescent phosphoramidites such as FluorePrime (Amersham Pharmacia, Piscataway, NJ.), Fluoredite (Millipore, Bedford, Mass.), FAM (ABI, Foster City, Calif.), and Cy3 or Cy5 (Amersham Pharmacia, Piscataway, N.J.). In another embodiment, the detectable label is a radiolabeled nucleotide.

In a further preferred embodiment, target polynucleotide molecules from a patient sample are labeled differentially from target polynucleotide molecules of a reference sample. The reference can comprise target polynucleotide molecules from normal tissue samples {i.e., tissues from those not afflicted with colorectal cancer).

Nucleic acid hybridization and wash conditions are chosen so that the target polynucleotide molecules specifically bind or specifically hybridize to the complementary polynucleotide sequences of the array, preferably to a specific array site, wherein its complementary DNA is located.

Arrays containing double-stranded probe DNA situated thereon are preferably subjected to denaturing conditions to render the DNA single-stranded prior to contacting with the target polynucleotide molecules. Arrays containing single-stranded probe DNA (e.g., synthetic oligodeoxyribonucleic acids) may need to be denatured prior to contacting with the target polynucleotide molecules, e.g., to remove hairpins or dimers which form due to self complementary sequences.

Optimal hybridization conditions will depend on the length (e.g., oligomer versus polynucleotide greater than 200 bases) and type (e.g. , RNA, or DNA) of probe and target nucleic acids. One of skill in the art will appreciate that as the oligonucleotides become shorter, it may become necessary to adjust their length to achieve a relatively uniform melting temperature for satisfactory hybridization results. General parameters for specific (i.e., stringent) hybridization conditions for nucleic acids are described in Sambrook et al., MOLECULAR CLONING - A LABORATORY MANUAL (2ND ED.), VOIS. 1-3, Cold Spring Harbor Laboratory, Cold Spring Harbor, New York (1989), and in Ausubel et al, CURRENT PROTOCOLS IN MOLECULAR BIOLOGY, vol. 2, Current Protocols Publishing, New York (1994). Typical hybridization conditions for the cDNA microarrays of Schena et al. are hybridization in 5 X SSC plus 0.2% SDS at 65⁰C for four hours, followed by washes at 25°C in low stringency wash buffer (1 X SSC plus 0.2% SDS), followed by 10 minutes at 25°C in higher stringency wash buffer (0.1 X SSC plus 0.2% SDS) (Schena et al, Proc. Natl. Acad. Sci. U.S.A. 93:10614 (1993)). Useful hybridization conditions are also provided in, e.g., Tijessen, 1993, HYBRIDIZATION WITH NUCLEIC ACID PROBES, Elsevier Science Publishers B.V.; and Kricka, 1992, NONISOTOPIC DNA PROBE TECHNIQUES , Academic Press, San Diego, CA.

Particularly preferred hybridization conditions include hybridization at a temperature at or near the mean melting temperature of the probes (e.g., within 51°C, more preferably within 21⁰C) in 1 M NaCl, 50 mM MES buffer (pH 6.5), 0.5% sodium sarcosine and 30% formamide.

5.6.6. SIGNAL DETECTION AND DATA ANALYSIS When fluorescently labeled gene products are used, the fluorescence emissions at each site of a microarray may be, preferably, detected by scanning confocal laser microscopy. In one embodiment, a separate scan, using the appropriate excitation line, is carried out for each of the two fluorophores used. Alternatively, a laser may be used that allows

5 simultaneous specimen illumination at wavelengths specific to the two fluorophores and emissions from the two fluorophores can be analyzed simultaneously (see Shalon et al, 1996, "A DNA microarray system for analyzing complex DNA samples using two-color fluorescent probe hybridization," Genome Research 6:639-645, which is incorporated by reference hi its entirety for all purposes). In a preferred embodiment, the arrays are scanned

0 with a laser fluorescent scanner with a computer controlled X-Y stage and a microscope objective. Sequential excitation of the two fluorophores is achieved with a multi-line, mixed gas laser and the emitted light is split by wavelength and detected with two photomultiplier tubes. Fluorescence laser scanning devices are described in Schena et al, Genome Res. 6:639-645 (1996), and in other references cited herein. Alternatively, the fiber-optic bundle

5 described by Ferguson et al , Nature Biotech. 14: 1681-1684 (1996), may be used to monitor mRNA abundance levels at a large number of sites simultaneously.

5.6.7. OTHER ASSAYS FOR DETECTING AND QUANTIFYING RNA hi addition to microarrays such as those described above any technique known to one of skill for detecting and measuring RNA can be used in accordance with the methods of the i0 invention. Non-limiting examples of techniques include Northern blotting, nuclease protection assays, RNA fingerprinting, polymerase chain reaction, ligase chain reaction, Qbeta replicase, isothermal amplification method, strand displacement amplification, transcription based amplification systems, nuclease protection (SI nuclease or RNAse protection assays), SAGE as well as methods disclosed in International Publication Nos. WO

15 88/10315 and WO 89/06700, and International Applications Nos. PCT/US87/00880 and PCT/US89/01025.

A standard Northern blot assay can be used to ascertain an RNA transcript size, identify alternatively spliced RNA transcripts, and the relative amounts of mRNA in a sample, in accordance with conventional Northern hybridization techniques known to those 0 persons of ordinary skill in the art. hi Northern blots, RNA samples are first separated by size via electrophoresis in an agarose gel under denaturing conditions. The RNA is then transferred to a membrane, crosslinked and hybridized with a labeled probe. Nonisotopic or high specific activity radiolabeled probes can be used including random-primed, nick- translated, or PCR-generated DNA probes, in vitro transcribed RNA probes, and oligonucleotides. Additionally, sequences with only partial homology {e.g., cDNA from a different species or genomic DNA fragments that might contain an exon) may be used as

5 probes. The labeled probe, e.g., a radiolabeled cDNA, either containing the full-length, single stranded DNA or a fragment of that DNA sequence may be at least 20, at least 30, at least 50, or at least 100 consecutive nucleotides in length. The probe can be labeled by any of the many different methods known to those skilled in this art. The labels most commonly employed for these studies are radioactive elements, enzymes, chemicals that fluoresce when

.0 exposed to ultraviolet light, and others. A number of fluorescent materials are known and can be utilized as labels. These include, but are not limited to, fluorescein, rhodamine, auramine, Texas Red, AMCA blue and Lucifer Yellow. A particular detecting material is anti-rabbit antibody prepared in goats and conjugated with fluorescein through an isothiocyanate. Proteins can also be labeled with a radioactive element or with an enzyme.

L 5 The radioactive label can be detected by any of the currently available counting procedures. Non-limiting examples of isotopes include ³H, ¹⁴C, ³²P, ³⁵S, ³⁶Cl, ⁵¹Cr, ⁵⁷Co, ⁵⁸Co, ⁵⁹Fe, ⁹⁰Y, ¹²⁵1, ¹³¹I, and ¹⁸⁶Re. Enzyme labels are likewise useful, and can be detected by any of the presently utilized colorimetric, spectrophotometric, fluorospectrophotometric, amperometric or gasometric techniques. The enzyme is conjugated to the selected particle by reaction with

20 bridging molecules such as carbodiimides, diisocyanates, glutaraldehyde and the like. Any enzymes known to one of skill in the art can be utilized. Examples of such enzymes include, but are not limited to, peroxidase, beta-D-galactosidase, urease, glucose oxidase plus peroxidase and alkaline phosphatase. U.S. Patent Nos. 3,654,090, 3,850,752, and 4,016,043 are referred to by way of example for their disclosure of alternate labeling material and 5 methods.

Nuclease protection assays (including both ribonuclease protection assays and Sl nuclease assays) can be used to detect and quantitate specific mRNAs. In nuclease protection assays, an antisense probe (labeled with, e.g., radiolabeled or nonisotopic) hybridizes in solution to an RNA sample. Following hybridization, single-stranded, unhybridized probe 0 and RNA are degraded by nucleases. An acrylamide gel is used to separate the remaining protected fragments. Typically, solution hybridization is more efficient than membrane-based hybridization, and it can accommodate up to 100 μg of sample RNA, compared with the 20- 30 μg maximum of blot hybridizations. The ribonuclease protection assay, which is the most common type of nuclease protection assay, requires the use of RNA probes. Oligonucleotides and other single-stranded DNA probes can only be used in assays containing Sl nuclease. The single- stranded, antisense probe must typically be completely homologous to target RNA to prevent cleavage 5 of the probe:target hybrid by nuclease.

Serial Analysis Gene Expression (SAGE), which is described in e.g., Velculescu et al., 1995, Science 270:484-7; Carulli, et al., 1998, Journal of Cellular Biochemistry Supplements 30/31:286-96, can also be used to determine RNA abundances in a cell sample.

Quantitative reverse transcriptase PCR (qRT-PCR) can also be used to determine the .0 expression profiles of marker genes (see, e.g., U.S. Patent Application Publication No. 2005/0048542A1). The first step in gene expression profiling by RT-PCR is the reverse transcription of the RNA template into cDNA, followed by its exponential amplification in a PCR reaction. The two most commonly used reverse transcriptases are avilo myeloblastosis virus reverse transcriptase (AMV-RT) and Moloney murine leukemia virus reverse L 5 transcriptase (MLV-RT). The reverse transcription step is typically primed using specific primers, random hexamers, or oligo-dT primers, depending on the circumstances and the goal of expression profiling. For example, extracted RNA can be reverse-transcribed using a GeneAmp RNA PCR kit (Perkin Elmer, Calif., USA), following the manufacturer's instructions. The derived cDNA can then be used as a template in the subsequent PCR ZO reaction.

Although the PCR step 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. Thus, TaqMan^® PCR typically utilizes the 5 '-nuclease activity of Taq or Tth polymerase to hydrolyze a hybridization probe 5 bound to its target amplicon, but any enzyme with equivalent 5' nuclease activity can be used. Two oligonucleotide primers are used to generate an amplicon typical of a PCR reaction. A third oligonucleotide, or probe, is designed to detect nucleotide sequence located between the two PCR primers. The probe is non-extendible by Taq DNA polymerase enzyme, and is labeled with a reporter fluorescent dye and a quencher fluorescent dye. Any laser-induced 0 emission from the reporter dye is quenched by the quenching dye when the two dyes are located close together as they are on the probe. During the amplification reaction, the Taq DNA polymerase enzyme cleaves the probe in a template-dependent manner. The resultant probe fragments disassociate in solution, and signal from the released reporter dye is free from the quenching effect of the second fluorophore. One molecule of reporter dye is liberated for each new molecule synthesized, and detection of the unquenched reporter dye provides the basis for quantitative interpretation of the data.

TaqMan RT-PCR can be performed using commercially available equipment, such as, for example, ABI PRISM 7700™. Sequence Detection System™ (Perkin-Elmer- Applied Biosystems, Foster City, Calif., USA), or Lightcycler (Roche Molecular Biochemicals, Mannheim, Germany). In a preferred embodiment, the 5' nuclease procedure is run on a realtime quantitative PCR device such as the ABI PRISM 7700™ Sequence Detection System™. The system consists of a thermocycler, laser, charge-coupled device (CCD), camera and computer. The system includes software for running the instrument and for analyzing the data.

5'-Nuclease assay data are initially expressed as Ct, or the threshold cycle. Fluorescence values are recorded during every cycle and represent the amount of product amplified to that point in the amplification reaction. The point when the fluorescent signal is first recorded as statistically significant is the threshold cycle (Ct).

To minimize errors and the effect of sample-to-sample variation, RT-PCR is usually performed using an internal standard. The ideal internal standard is expressed at a constant level among different tissues, and is unaffected by the experimental treatment. RNAs most frequently used to normalize patterns of gene expression are mRNAs for the housekeeping genes glyceraldehyde-3-phosphate-dehydrogenase (GAPDH) and β-actin.

A more recent variation of the RT-PCR technique is the real time quantitative PCR, which measures PCR product accumulation through a dual-labeled fluorigenic probe (i.e., TaqMan^® probe). Real time PCR is compatible both with quantitative competitive PCR, where internal competitor for each target sequence is used for normalization, and with quantitative comparative PCR using a normalization gene contained within the sample, or a housekeeping gene for RT-PCR. For further details see, e.g. Held et al., Genome Research 6:986-994 (1996).

5.6.8. DETECTION AND QUANTIFICATION OF PROTEIN Measurement of the translational state may be performed according to several methods. For example, whole genome monitoring of protein (e.g., the "proteome,") can be carried out by constructing a microarray in which binding sites comprise immobilized, preferably monoclonal, antibodies specific to a plurality of protein species encoded by the cell genome. Preferably, antibodies are present for a substantial fraction of the encoded proteins, or at least for those proteins relevant to the action of a drug of interest. Methods for making monoclonal antibodies are well known (see, e.g., Harlow and Lane, 1988, Antibodies: A Laboratory Manual, Cold Spring Harbor, New York, which is incorporated in its entirety for all purposes). In one embodiment, monoclonal antibodies are raised against synthetic peptide fragments designed based on genomic sequence of the cell. With such an antibody array, proteins from the cell are contacted to the array and their binding is assayed with assays known in the art.

Immunoassays known to one of skill in the art can be used to detect and quantify protein levels. For example, ELISAs can be used to detect and quantify protein levels. ELISAs comprise preparing antigen, coating the well of a 96 well microtiter plate with the antigen, adding the antibody of interest conjugated to a detectable compound such as an enzymatic substrate (e.g., horseradish peroxidase or alkaline phosphatase) to the well and incubating for a period of time, and detecting the presence of the antigen. In ELISAs the antibody of interest does not have to be conjugated to a detectable compound; instead, a second antibody (which recognizes the antibody of interest) conjugated to a detectable compound may be added to the well. Further, instead of coating the well with the antigen, the antibody may be coated to the well. In this case, a second antibody conjugated to a detectable compound may be added following the addition of the antigen of interest to the coated well. One of skill in the art would be knowledgeable as to the parameters that can be modified to increase the signal detected as well as other variations of ELISAs known in the art. In a preferred embodiment, an ELISA may be performed by coating a high binding 96- well microtiter plate (Costar) with 2μg/ml of rhu-IL-9 in PBS overnight. Following three washes with PBS, the plate is incubated with three-fold serial dilutions of Fab at 25°C for 1 hour. Following another three washes of PBS, lμg/ml anti-human kappa-alkaline phosphatase-conjugate is added and the plate is incubated for 1 hour at 25 C. Following three washes with PBST, the alkaline phosphatase activity is determined in 50μl/AMP/PPMP substrate. The reactions are stopped and the absorbance at 560 nm is determined with a VMAX microplate reader. For further discussion regarding ELISAs see, e.g., Ausubel et al, eds, 1994, Current Protocols in Molecular Biology, Vol. 1, John Wiley & Sons, Inc., New York at 11.2.1.

Protein levels may be determined by Western blot analysis. Further, protein levels as well as the phosphorylation of proteins can be determined by immunoprecitation followed by Western blot analysis. Immunoprecipitation protocols generally comprise lysing a population of cells in a lysis buffer such as RIPA buffer (1% NP-40 or Triton X-100, 1% sodium deoxycholate, 0.1% SDS, 0.15 M NaCl, 0.01 M sodium phosphate at pH 7.2, 1% Trasylol) supplemented with protein phosphatase and/or protease inhibitors {e.g., EDTA, PMSF, aprotinin, sodium vanadate), adding the antibody of interest to the cell lysate, incubating for a period of time {e.g., 1 to 4 hours) at 40° C, adding protein A and/or protein G sepharose beads to the cell lysate, incubating for about an hour or more at 40° C, washing the beads in lysis buffer and resuspending the beads in SDS/sample buffer. The ability of the antibody of interest to immunoprecipitate a particular antigen can be assessed by, e.g., western blot analysis. One of skill in the art would be knowledgeable as to the parameters that can be modified to increase the binding of the antibody to an antigen and decrease the background {e.g., pre-clearing the cell lysate with sepharose beads). For further discussion regarding immunoprecipitation protocols see, e.g., Ausubel et al, eds, 1994, Current Protocols in Molecular Biology, Vol. 1, John Wiley & Sons, Inc., New York at 10.16.1.

Western blot analysis generally comprises preparing protein samples, electrophoresis of the protein samples in a polyacrylamide gel {e.g. , 8%- 20% SDS-PAGE depending on the molecular weight of the antigen), transferring the protein sample from the polyacrylamide gel to a membrane such as nitrocellulose, PVDF or nylon, incubating the membrane in blocking solution {e.g., PBS with 3% BSA or non-fat milk), washing the membrane in washing buffer {e.g., PBS-T ween 20), incubating the membrane with primary antibody (the antibody of interest) diluted in blocking buffer, washing the membrane in washing buffer, incubating the membrane with a secondary antibody (which recognizes the primary antibody, e.g., an anti- human antibody) conjugated to an enzymatic substrate {e.g., horseradish peroxidase or alkaline phosphatase) or radioactive molecule {e.g., ^^P or 125i) diluted in blocking buffer, washing the membrane in wash buffer, and detecting the presence of the antigen. One of skill in the art would be knowledgeable as to the parameters that can be modified to increase the signal detected and to reduce the background noise. For further discussion regarding western blot protocols see, e.g., Ausubel et al, eds, 1994, Current Protocols in Molecular Biology, Vol. 1, John Wiley & Sons, Inc., New York at 10.8.1.

Protein expression levels can also be separated by two-dimensional gel electrophoresis systems. Two-dimensional gel electrophoresis is well-known in the art and typically involves isoelectric focusing along a first dimension followed by SDS-PAGE electrophoresis along a second dimension. See, e.g., Hames et al., 1990, GeZ Electrophoresis of Proteins: A Practical Approach, IRL Press, New York; Shevchenko et al., 1996, Proc. Natl. Acad. ScL USA 93:1440-1445; Sagliocco et al., 1996, Yeast 12:1519-1533; Lander, 1996, Science 274:536-539. The resulting electropherograms can be analyzed by numerous techniques, including mass spectrometric techniques, Western blotting and immunoblot analysis using polyclonal and monoclonal antibodies, and internal and N- terminal micro- sequencing.

5.7. DETERMINING THERAPEUTIC REGIMENS FOR PATIENTS

The benefit of adjuvant chemotherapy for colorectal cancer appears limited to patients with Dukes stage C disease where the cancer has metastasized to lymph nodes at the time of diagnosis. For this reason, the clinicopathological Dukes' staging system is critical for determining how adjuvant therapy is administered. Unfortunately, as noted above, Dukes' staging is not very accurate in predicting overall survival and thus its application likely results in the treatment of a large number of patients to benefit an unknown few. Alternatively, there are a number of patients who would benefit from therapy that do not receive it based on the Dukes' staging system.

Thus, the methods of the prognosis prediction can be used for determining whether a colorectal cancer patient may benefit from chemotherapy. In one embodiment, the invention provides a method for determining whether a colorectal cancer patient should be treated with chemotherapy, comprising (a) classifying the patient as having a good prognosis or a poor prognosis using a method as described in Section 5.2.; and (b) determining that said patient's predicted survival time favors treatment of the patient with chemotherapy if said patient is classified as having a poor prognosis. In another embodiment, the methods are used in conjuction with Dukes staging. For example, the prognosis methods of the invention can be used to identify those Dukes' stage B and C cases for which chemotherapy may be beneficial. If a patient is determined to be one likely to benefit from chemotherapy, a suitable chemotherapy may be prescribed for the patient. Chemotherapy can be performed using any one or a combination of the anti-cancer drugs known in the art, including but not limited to any topoisomerase inhibitor, DNA binding agent, anti-metabolite, ionizing radiation, or a combination of two or more of such known DNA damaging agents.

A topoisomerase inhibitor that can be used in conjunction with the invention can be, for example, a topoisomerase I (Topo I) inhibitor, a topoisomerase II (Topo II) inhibitor, or a dual topoisomerase I and II inhibitor. A topo I inhibitor can be from any of the following classes of compounds: camptothecin analogue (e.g., karenitecin, aminocamptothecin, lurtotecan, topotecan, irinotecan, BAY 56-3722, rubitecan, GI14721, exatecan mesylate), rebeccamycin analogue, PNU 166148, rebeccamycin, TAS-103, camptothecin (e.g., camptothecin polyglutamate, camptothecin sodium), intoplicine, ecteinascidin 743, J- 107088, pibenzimol. Examples of preferred topo I inhibitors include but are not limited to camptothecin, topotecan (hycaptamine), irinotecan (irinotecan hydrochloride), belotecan, or an analogue or derivative thereof.

A topo II inhibitor that can be used in conjunction with the invention can be, for example, from any of the following classes of compounds: anthracycline antibiotics (e.g., carubicin, pirarubicin, daunorubicin citrate liposomal, daunomycin, 4-iodo-4- doxydoxorubicin, doxorubicin, n,n-dibenzyl daunomycin, morpholinodoxorubicin, aclacinomycin antibiotics, duborimycin, menogaril, nogalamycin, zorubicin, epirubicin, marcellomycin, detorubicin, annamycin, 7-cyanoquinocarcinol, deoxydoxorubicin, idarubicin, GPX-100, MEN-10755, valrubicin, KRN5500), epipodophyllotoxin compound (e.g., podophyllin, teniposide, etoposide, GL331, 2-ethylhydrazide), anthraquinone compound (e.g., ametantrone, bisantrene, mitoxantrone, anthraquinone), ciprofloxacin, acridine carboxamide, amonafide, anthrapyrazole antibiotics (e.g., teloxantrone, sedoxantrone trihydrochloride, piroxantrone, anthrapyrazole, losoxantrone), TAS-103, fostriecin, razoxane, XK469R, XK469, chloroquinoxaline sulfonamide, merbarone, intoplicine, elsamitrucin, CI- 921, pyrazoloacridine, elliptinium, amsacrine. Examples of preferred topo II inhibitors include but are not limited to doxorubicin (Adriamycin), etoposide phosphate (etopofos), teniposide, sobuzoxane, or an analogue or derivative thereof.

DNA binding agents that can be used in conjunction with the invention include but are not limited to DNA groove binding agent, e.g., DNA minor groove binding agent; DNA crosslinking agent; intercalating agent; and DNA adduct forming agent. A DNA minor groove binding agent can be an anthracycline antibiotic, mitomycin antibiotic (e.g., porfiromycin, KW-2149, mitomycin B, mitomycin A, mitomycin C), chromomycin A3, carzelesin, actinomycin antibiotic (e.g., cactinomycin, dactinomycin, actinomycin Fl), 5 brostallicin, echinomycin, bizelesin, duocarmycin antibiotic (e.g., KW 2189), adozelesin, olivomycin antibiotic, plicamycin, zinostatin, distamycin, MS-247, ecteinascidin 743, amsacrine, anthramycin, and pibenzimol, or an analogue or derivative thereof.

DNA crosslinking agents include but are not limited to antineoplastic alkylating agent, methoxsalen, mitomycin antibiotic, psoralen. An antineoplastic alkylating agent can

0 be a nitrosourea compound (e.g., cystemustine, tauromustine, semustine, PCNU, streptozocin, SarCNU, CGP-6809, carmustine, fotemustine, methylnitrosourea, nimustine, ranimustine, ethylnitrosourea, lomustine, chlorozotocin), mustard agent (e.g., nitrogen mustard compound, such as spiromustine, trofosfamide, chlorambucil, estramustine, 2,2,2- trichlorotriethylamine, prednimustine, novembichin, phenamet, glufosfamide, peptichemio,

.5 if osf amide, defosf amide, nitrogen mustard, phenesterin, mannomustine, cyclophosphamide, melphalan, perfosfamide, mechlorethamine oxide hydrochloride, uracil mustard, bestrabucil, DHEA mustard, tallimustine, mafosfamide, aniline mustard, chlornaphazine; sulfur mustard compound, such as bischloroethylsulfide; mustard prodrug, such as TLK286 and ZD2767), ethylenimine compound (e.g., mitomycin antibiotic, ethylenimine, uredepa, thiotepa,

£0 diaziquone, hexamethylene bisacetamide, pentamethylmelamine, altretamine, carzinophilin, triaziquone, meturedepa, benzodepa, carboquone), alkylsulfonate compound (e.g., dimethylbusulfan, Yoshi-864, improsulfan, piposulfan, treosulfan, busulfan, hepsulfam), epoxide compound (e.g., anaxirone, mitolactol, dianhydrogalactitol, teroxirone), miscellaneous alkylating agent (e.g., ipomeanol, carzelesin, methylene dimethane sulfonate,

25 mitobronitol, bizelesin, adozelesin, piperazinedione, VNP40101M, asaley, 6- hydroxymethylacylfulvene, EO9, etoglucid, ecteinascidin 743, pipobroman), platinum compound (e.g., ZD0473, liposomal-cisplatin analogue, satraplatin, BBR 3464, spiroplatin, ormaplatin, cisplatin, oxaliplatin, carboplatin, lobaplatin, zeniplatin, iproplatin), triazene compound (e.g., imidazole mustard, CB 10-277, mitozolomide, temozolomide, procarbazine, 0 dacarbazine), picoline compound (e.g., penclomedine), or an analogue or derivative thereof. Examples of preferred alkylating agents include but are not limited to cisplatin, dibromodulcitol, fotemustine, ifosfamide (ifosfamid), ranimustine (ranomustine), nedaplatin (latoplatin), bendamustine (bendamustine hydrochloride), eptaplatin, temozolomide (methazolastone), carboplatin, altretamine (hexamethylmelamine), prednimustine, oxaliplatin (oxalaplatinum), carmustine, thiotepa, leusulfon (busulfan), lobaplatin, cyclophosphamide, bisulfan, melphalan, and chlorambucil, or analogues or derivatives thereof.

Intercalating agents can be an anthraquinone compound, bleomycin antibiotic, rebeccamycin analogue, acridine, acridine carboxamide, amonafide, rebeccamycin, anthrapyrazole antibiotic, echinomycin, psoralen, LU 79553, BW A773U, crisnatol mesylate, benzo(a)pyrene-7,8-diol-9,10-epoxide, acodazole, elliptinium, pixantrone, or an analogue or derivative thereof, etc.

DNA adduct forming agents include but are not limited to enediyne antitumor antibiotic (e.g., dynemicin A, esperamicin Al, zinostatin, dynemicin, calicheamicin gamma II), platinum compound, carmustine, tamoxifen (e.g., 4-hydroxy-tamoxifen), psoralen, pyrazine diazohydroxide, benzo(a)pyrene-7,8-diol-9,10-epoxide, or an analogue or derivative thereof.

Anti-metabolites include but are not limited to cytosine, arabinoside, floxuridine, fluorouracil, mercaptopurine, Gemcitabine, and methotrexate (MTX).

In addition to identifying those patients for whom therapy is most beneficial, the classifier of the subject invention can identify those genes that are most biologically significant based on their frequency of appearance in the classification set. In one embodiment, those genes that are most biologically significant to colorectal cancer were identified using the classifier provided in the Example (Section 6). Specifically, osteopontin and neuregulin reported biological significance in the context of colorectal cancer.

5.8. KITS

The invention provides kits that are useful in predicting prognosis of colorectal cancer in a colorectal caner patient. The kits of the present invention comprise one or more probes and/or primers for each of at least 5, 10, 20, 30, 40, 50, 60, 70, 80, or 100 gene products that are encoded by the respectively marker genes listed in Table 1 or funtional equivalents of such genes. The probes of marker genes may be part of an array, or the biomarker(s) may be packaged separately and/or individually. In one embodiment, the invention provides kits comprising probes that are immobilized at an addressable position on a substrate, e.g., in a microarray. In a particular embodiment, the invention provides such a microarray.

The kits of the present invention may also contain probes that can be used to detect protein products of the marker genes of the invention. In a specific embodiment, the invention provides a kit comprises a plurality of antibodies that specifically bind a plurality of at least 5, 10, 20, 30, 40, 50, 60, 70, 80, or 100 protens that are encoded by the respectively marker genes listed in Table 1 or funtional equivalents of such genes. In accordance with this embodiment, the kit may comprise a set of antibodies or functional fragments or derivatives thereof (e.g., Fab, F(ab')₂, Fv, or scFv fragments). In accordance with this embodiment, the kit may include antibodies, fragments or derivatives thereof (e.g., Fab, F(ab')₂, Fv, or scFv fragments) that are specific for these proteins. In one embodiment, the antibodies may be detectably labeled.

The kits of the present invention may also include reagents such as buffers, or other reagents that can be used in obtaining the marker profile. Prevention of the action of microorganisms can be ensured by the inclusion of various antibacterial and antifungal agents, for example, paraben, chlorobutanol, phenol sorbic acid, and the like. It may also be desirable to include isotonic agents such as sugars, sodium chloride, and the like.

In some embodiments of the invention, the kits of the present invention comprise a microarray. The microarray can be any of the microarrays described above, e.g., in Section 5.6.1, optionally in a sealed container. In one embodiment this microarray comprises a plurality of probe spots, wherein at least 20%, 40%, 60%, 80%, or 90% of the probe spots in the plurality of probe spots correspond to marker genes listed in Table 1.

In still other embodiments, the kits of the invention may further comprise a computer program product for use in conjunction with a computer system, wherein the computer program product comprises a computer readable storage medium and a computer program mechanism embedded therein. In such kits, the computer program mechanism comprises instructions for prediction of prognosis using a marker profile obtained with the reagents of the kits.

In still other embodiments, the kits of the present invention comprise a computer having a central processing unit and a memory coupled to the central processing unit. The memory stores instructions for for prediction of prognosis using a marker profile obtained with the reagents of the kits.

5.9. COMPUTER-FACILITATED ANALYSIS

The analytic methods described in the previous sections can be implemented by use of 5 the following computer systems and according to the following programs and methods. A computer system comprises internal components linked to external components. The internal components of a typical computer system include a processor element interconnected with a main memory. For example, the computer system can be based on an Intel 8086-, 80386-, 80486-, Pentium™, or Pentium IV™-based processor with preferably 512 MB or more of LO main memory. The computer system may also be a Macintosh or a Macintosh-based system, but may also be a minicomputer or mainframe.

The external components preferably include mass storage. This mass storage can be one or more hard disks (which are typically packaged together with the processor and memory). Such hard disks are preferably of 10 GB or greater storage capacity. Other L 5 external components include a user interface device, which can be a monitor, together with an inputting device, which can be a "mouse", or other graphic input devices, and/or a keyboard. A printing device can also be attached to the computer.

Typically, a computer system is also linked to network link, which can be part of an Ethernet link to other local computer systems, remote computer systems, or wide area 20 communication networks, such as the Internet. This network link allows the computer system to share data and processing tasks with other computer systems.

Loaded into memory during operation of this system are several software components, which are both standard in the art and special to the instant invention. These software components collectively cause the computer system to function according to the 5 methods of this invention. These software components are typically stored on the mass storage device. A software component comprises the operating system, which is responsible for managing computer system and its network interconnections. This operating system can be, for example, of the Microsoft Windows^® family, such as Windows 3.1, Windows 95, Windows 98, Windows 2000, Windows NT, or Windows XP, or may be of the Macintosh OS 0 family, or may be UNIX, a UNIX derivative such as LINUX, or an operating system specific to a minicomputer or mainframe. The software component represents common languages and functions conveniently present on this system to assist programs implementing the methods specific to this invention. Many high or low level computer languages can be used to program the analytic methods of this invention. Instructions can be interpreted during runtime or compiled. Preferred languages include C/C++, FORTRAN and JAVA. Most preferably, the methods of this invention are programmed in mathematical software packages that allow symbolic entry of equations and high-level specification of processing, including some or all of the algorithms to be used, thereby freeing a user of the need to procedurally program individual equations or algorithms. Such packages include Mathlab from Mathworks (Natick, MA), Mathematica^® from Wolfram Research (Champaign, IL), or S- Plus^® from Math Soft (Cambridge, MA). Specifically, the software component includes the analytic methods of the invention as programmed in a procedural language or symbolic package.

The software to be included with the kit comprises the data analysis methods of the invention as disclosed herein. In particular, the software may include mathematical routines for marker discovery, including the calculation of similarity values between clinical categories (e.g., prognosis) and marker expression. The software may also include mathematical routines for calculating the similarity between sample marker expression and control marker expression, using array-generated fluorescence data, to determine the clinical classification of a sample.

Additionally, the software may also include mathematical routines for determining the prognostic outcome, and recommended therapeutic regimen, for an individual with a cancer. The mathematical routines determine the prognostic outcome and recommended therapeutic regimen for an individual. Such software can include instructions for the computer system's processor to receive data structures that include the levels of expression of at least 5, 10, 20, 30, 40, 50, 60, 70, 80, or 100 of the genes (as appropriate) listed in Table 1, or in any of

Tables 2-5, 7, and 8 or any subset of these tables, or respective functional equivalents thereof, in a colorectal cancer tumor sample obtained from the patient. The software may additionally include mathematical routines for assigning the patient a prognosis using one or more classifiers of the invention.

The software preferably would include decisional routines that integrate the patient's clinical and marker gene expression data, and recommend a course of therapy. In one embodiment, for example, the software causes the processor unit to receive expression data for prognosis-related genes in the patient's tumor sample, assign the patient a prognosis, and, on the basis of the prognosis, assign a recommended therapeutic regimen.

Correlating genes to clinical outcomes in accordance with the subject invention can be performed using software on a computing means. The computing means can also be 5 responsible for maintenance of acquired data as well as the maintenance of the classifier system itself. The computing means can also detect and act upon user input via user interface means known to the skilled artisan (i.e., keyboard, interactive graphical monitors) for entering data to the computing system.

In one embodiment, the computing means further comprises means for storing and .0 means for outputting processed data. The computing means includes any digital instrumentation capable of processing data input from the user. Such digital instrumentation, as understood by the skilled artisan, can process communicated data by applying algorithm and filter operations of the subject invention. Preferably, the digital instrumentation is a microprocessor, a personal desktop computer, a laptop, and/or a portable digital device. The L 5 computing means can be general purpose or application specific.

The subject invention can be practiced in a variety of situations. The computing means can directly or remotely connect to a central office or health care center. In one embodiment, the subject invention is practiced directly in an office or hospital. In another embodiment, the subject invention is practiced in a remote setting, for example, personal ZO residences, mobile clinics, vessels at sea, rural villages and towns without direct access to healthcare, and ambulances, wherein the patient is located some distance from the physician.

In a related embodiment, the computing means is a custom, portable design and can be carried or attached to the health care provider in a manner similar to other portable electronic devices such as a portable radio or computer.

5 The computing means used in accordance with the subject invention can contain at least one user-interface device including, but not limited to, a keyboard, stylus, microphone, mouse, speaker, monitor, and printer. Additional user-interface devices contemplated herein include touch screens, strip recorders, joysticks, and rollerballs.

Preferably, the computing means comprises a central processing unit (CPU) having 0 sufficient processing power to perform algorithm operations in accordance with the subject invention. The algorithm operations, including the microarray analysis operations (such as SAM or binary classification), can be embodied in the form of computer processor usable media, such as floppy diskettes, CD-ROMS, zip drives, non-volatile memory, or any other computer-readable storage medium, wherein the computer program code is loaded into and executed by the computing means. Optionally, the operational algorithms of the subject invention can be programmed directly onto the CPU using any appropriate programming language, preferably using the C programming language.

In certain embodiments, the computing means comprises a memory capacity sufficiently large to perform algorithm operations in accordance with the subject invention. The memory capacity of the invention can support loading a computer program code via a computer-readable storage media, wherein the program contains the source code to perform the operational algorithms of the subject invention. Optionally, the memory capacity can support directly programming the CPU to perform the operational algorithms of the subject invention. A standard bus configuration can transmit data between the CPU, memory, ports and any communication devices.

In addition, as understood by the skilled artisan, the memory capacity of the computing means can be expanded with additional hardware and with saving data directly onto external mediums including, for example, without limitation, floppy diskettes, zip drives, non-volatile memory and CD-ROMs.

Further, the computing means can also include the necessary software and hardware to receive, route and transfer data to a remote location.

In one embodiment, the patient is hospitalized, and clinical data generated by a computing means is transmitted to a central location, for example, a monitoring station or to a specialized physician located in a different locale.

In another embodiment, the patient is in remote communication with the health care provider. For example, patients can be located at personal residences, mobile clinics, vessels at sea, rural villages and towns without direct access to healthcare, and ambulances. An expression profile of the patient can be measured on-site, and communicated to the health care provider. By using the classifier of the invention, the health care provider can carry out prognosis remotely. Advantageously, mobile stations, such as ambulances, and mobile clinics, can monitor patient health by using a portable computing means of the subject invention when transporting and/or treating a patient.

To ensure patient privacy, security measures, such as encryption software and firewalls, can be employed. Optionally, clinical data can be transmitted as unprocessed or "raw" signal(s) and/or as processed signal(s). Advantageously, transmitting raw signals allows any software upgrades to occur at the remote location where a computing means is located. In addition, both historical clinical data and real-time clinical data can be transmitted.

Communication devices such as wireless interfaces, cable modems, satellite links, microwave relays, and traditional telephonic modems can transfer clinical data from a computing means to a healthcare provider via a network. Networks available for transmission of clinical data include, but are not limited to, local area networks, intranets and the open internet. A browser interface, for example, NETSCAPE NAVIGATOR or INTERNET EXPLORER, can be incorporated into communications software to view the transmitted data.

Advantageously, a browser or network interface is incorporated into the processing device to allow the user to view the processed data in a graphical user interface device, for example, a monitor. The results of algorithm operations of the subject invention can be displayed in the form of interactive graphics.

An exemplary computer system is shown in Fig. 5. The exemplary system is a computer system 10 having: a central processing unit 22; a main non-volatile storage unit 14, for example, a hard disk drive, for storing software and data, the storage unit 14 controlled by storage controller 12; a system memory 36, preferably high speed random-access memory (RAM), for storing system control programs, data, and application programs, comprising programs and data loaded from non-volatile storage unit 14; system memory 36 may also include read-only memory (ROM); a user interface 32, comprising one or more input devices (e.g., keyboard 28) and a display 26 or other output device; a network interface card 20 for connecting to any wired or wireless communication network 34 (e.g., a wide area network such as the Internet); an internal bus 30 for interconnecting the aforementioned elements of the system; and a power source 24 to power the aforementioned elements. Operation of computer 10 is controlled primarily by operating system 40, which is executed by central processing unit 22. Operating system 40 can be stored in system memory 36. In addition to operating system 40, in a typical implementation system memory 36 includes: file system 42 for controlling access to the various files and data structures used by the present invention.

6. EXAMPLE The following example is presented by way of illustration of the present invention, and is not intended to limit the present invention in any way.

6.1. TUMOR SAMPLES AND MEASUREMENT OF EXPRESSION PROFILES

A set of 78 tumor samples, including 3 adenomas and 75 cancers were used for the development of the cDNA classifiers. These were informative frozen colorectal cancer samples selected from the Moffitt Cancer Center Tumor Bank (Tampa, Florida) based on evidence for good (survival > 36 mo) or poor prognosis (survival < 36 mo) from the Tumor Registry (the "Moffitt set"). Dukes' stages can include B, C, and D. Survival was measured as last contact minus collection date for living patients, or date of death minus collection date for patients who have died.

The 75 cancer samples in the Moffitt set include 23 Dukes stage B samples, 22 Dukes stage C samples, and 30 Dukes stage D samples. Just as adenomas can be included to help train the classifier to recognize good prognosis patients, Dukes D patients with synchronous metastatic disease can be used to train the classifier to recognize poor prognosis patients.

All samples in the Moffitt set were selected to have at least 36 months of follow-up. The follow-up results in this embodiment showed that thirty-two of the patients survived more than 36 months, while 46 patients died within 36 months. The median follow-up time for all 78 patients was 27.9 months. The median follow-up for the poor prognosis cases (<36 months survival) was 11.7 months, and for the good prognosis cases (>36 months survival) was 64.2 months.

Since the NIH consensus conference in 1990, chemotherapeutic application in the

United States has been relatively homogeneous, with nearly all Dukes stage B avoiding chemotherapy, and nearly all Dukes stage C receiving 6 months of adjuvant 5-fluorouracil (5- FU) and leucovorin.

Samples in the Moffitt set were microdissected (>80% tumor cells) by frozen section guidance and RNA extraction performed using Trizol followed by secondary purification on RNAEasy columns. The samples were profiled on cDNA arrays (i.e., TIGR's 32,488- element spotted cDNA arrays, containing 31,872 human cDNAs representing 30,849 distinct transcripts - 23,936 unique TIGR TCs and 6,913 ESTs, 10 exogenous controls printed 36 times, and 4 negative controls printed 36-72 times).

5 Tumor samples were co-hybridized with a common reference pool in the Cy5 channel for normalization purposes. The reference pool contained a mixture of RNAs derived from multiple cell lines. cDNA synthesis, aminoallyl labeling and hybridizations were performed according to previously published protocols (see Hegde, P. et al., "A concise guide to cDNA microarray analysis," Biotechniques; 29:552-562 (2000) and Yang, LV, et al., "Within the

.0 fold: assessing differential expression measures and reproducibility in microarray assays," Genome Biol; 3:research0062 (2002)). For example, labeled first-strand cDNA was prepared, and co-hybridized with labeled samples are prepared, from a universal reference RNA consisting of equimolar quantities of total RNA derived from three cell lines, CaCO2 (colon), KM12L4A (colon), and U118MG (brain). Detailed protocols and description of the array are

L5 available at <http://cancer.tigr.org>. Array probes were identified and local background were subtracted in Spotfinder (Saeed, A.I. et ah, "TM4: a free, open-source system for microarray data management and analysis," Biotechniques; 34:374-8 (2003)). Individual arrays can be normalized in MIDAS (see Saeed, A.I. ibid.) using LOWESS (an algorithm known to the skilled artisan for use in normalizing data) with smoothing parameter set to 0.33.

20 In addition to the 78-sample Moffitt training set, a set of eighty-eight patients with

Dukes' stage B and C colorectal cancer and a minimum follow-up time of 60 months were used for the development of the U133 classifiers. This set of samples is also referred to as the "Denmark set" of samples.

In the Denmark set, there were 28 patients with stage B and 60 patients with stage C 5 colorectal cancers. All Dukes' stage B patients were treated by surgical resection alone whereas all C patients received 5-FU/leucovorin adjuvant chemotherapy in addition to surgery. Colorectal tumor samples were obtained fresh from surgery and were immediately snap-frozen in fluid nitrogen but were not microdissected, with the potential for inclusion of samples with < 80% purity. Total RNA was isolated from 50-150 mg tumor sample using 0 RNAzol (WAK-Chemie Medical) or using spin column technology (Sigma) according to the manufacturer's instructions. Results were noted (i.e., fifty-seven of the patients survived more than 36 months, while 31 died within 36 months). Ten micrograms of total RNA were used as starting material for the cDNA preparation and hybridized to Affymetrix^® U133A GeneChips^® (Santa Clara, CA) by standard protocols supplied by the manufacturer. The U133A gene chip is disclosed in U.S. Patent Nos. 5,445,934; 5,700,637; 5,744,305; 5,945,334; 6,054,270; 6,140,044; 6,261,776; 6,291,183; 6,346,413; 6,399,365; 6,420,169; 6,551,817; 6,610,482; and 6,733,977; and in European Patent Nos. 619,321 and 373,203, all of which are hereby incorporated in their entirety. The first and second strand cDNA synthesis was performed using the Superscript II System (Invitrogen) according to the manufacturer's instructions except using an oligodT primer containing a T7 RNA polymerase promoter site. Labeled cRNA was prepared using the BioArray High Yield RNA Transcript Labeling Kit (Enzo). Biotin labeled CTP and UTP (Enzo) were used in the reaction together with unlabeled NTP 's. Following the IVT reaction, the unincorporated nucleotides were removed using RNeasy columns (Qiagen). Fifteen micrograms of cRNA were fragmented at 94°C for 35 min in a fragmentation buffer containing 40 mM Tris-acetate pH 8.1, 100 mM KOAc, 30 mM MgOAc. Prior to hybridization, the fragmented cRNA in a 6xSSPE-T hybridization buffer (1 M NaCl, 10 mM Tris pH 7.6, 0.005% Triton) was heated to 95°C for 5 min and subsequently to 45°C for 5 min before loading onto the Affymetrix^® HG_U133A probe array cartridge. The probe array was then incubated for 16 h at 45°C at constant rotation (60 rpm). The washing and staining procedure were performed in an Affymetrix^® Fluidics Station.

The probe array was exposed to several washes (Le., 10 washes in 6xSSPE-T at 25°C followed by 4 washes in 0.5xSSPE-T at 50°C). The biotinylated cRNA was stained with a streptavidinphycoerythrin conjugate, final concentration 2 mg/ml (Molecular Probes, Eugene, OR) in 6xSSPE-T for 30 min at 25°C followed by 10 washes in 6xSSPE-T at 25°C. An antibody amplification step followed, using normal goat IgG as blocking reagent, final concentration 0.1 mg/ml (Sigma) and biotinylated anti-streptavidin antibody (goat), final concentration 3 mg/ml (Vector Laboratories). This was followed by a staining step with a streptavidin-phycoerythrin conjugate, final concentration 2 mg/ml (Molecular Probes, Eugene, OR) in 6xSSPE-T for 30 min at 25°C and 10 washes in 6xSSPE-T at 25⁰C. The probe arrays were scanned (Le., at 560 nm using a confocal laser-scanning microscope (Hewlett Packard GeneArray Scanner G2500A)). The readings from the quantitative scanning were analyzed by the Affymetrix^® Gene Expression Analysis Software (MAS 5.0) and normalized to a common mean expression value of 150. 6.2. IDENTIFICATION OF PROGNOSIS-RELATED GENES USING SAM The first analysis of the colon cancer survival data was performed using censored survival time (in months) and 500 permutations. Significance analysis of microarrays (SAM) was used to select genes most closely correlated to survival. The subset of genes that 5 correspond to an empirically derived, estimated false discovery rate (FDR) was then chosen. This subset of genes was used in subsequent cluster analyses. Cluster 3.0 and Java Tree View 1.03 were used to cluster and visualize the SAM-selected genes.

SAM survival analysis was used to identify a set of genes most correlated with censored survival time using the training set tumor samples. A set of 53 genes was found,

0 corresponding to a median expected false discovery rate (FDR) of 28%. These genes are listed in the following Table 2, wherein genes denoted with (+) indicate a positive correlation to survival time and genes without the (+) notation indicate a negative correlation in survival time (over expression in poor prognosis cases). Included in this list of genes in Table 2 are several genes believed to be biologically significant, such as osteopontin and neuregulin.

L5 Table 3 illustrates seven genes selected by SAM survival analysis with a FDR of 13.5%, where osteopontin and neuregulin are noted to be present and in common with the gene lists for all classifiers. In Table 3, genes denoted with (+) indicate a positive correlation to survival time and genes without the (+) notation indicate a negative correlation in survival time (over expression in poor prognosis cases).

ZO Table 2 — Censored survival analysis using SAM, resultant 53 genes selected with median 28% FDR

Table 3 Censored survival analysis using SAM; seven genes selected with median estimated FDR of 13.5%

A hierarchical clustering algorithm was used to analyze the expression profiles of the 53 genes, with complete linkage and the correlation coefficient (Le., Pearson correlation coefficient) as the similarity metric. Dukes' staging clusters were manually created in the appropriate format. Clustering software produced heatmap (see Figures IA and IB) and dendrograms. The highest level partition of the SAM-selected genes was chosen as a survival grouping. Given two clusters of survival times, Kaplan-Meier curves were plotted (see Figures 2A and 2B).

Figure IA presents a graphical representation of the 53 SAM-selected genes (as described above) as a clustered heat map. Darker areas represent over-expressed genes whereas lighter areas represent under-expressed genes. Figure IA shows only the Dukes' stage B and C cases, whose outcome Dukes' staging predicts poorly. Since only genes correlated with survival were used in clustering, the distinctly illustrated clusters in the heatmap correspond to very different prognosis groups.

The 53 SAM-selected genes were also arranged by annotated Dukes' stage in Figure IB. Unlike Figure IA, where two gene groups were apparent, there was no discernible gene expression grouping when arranged by Dukes' stage. Figure 2A shows the Kaplan-Meier plot for two dominant clusters of genes correlated with stage B and C test set tumor samples. These genes separated the cases into two distinct clusters of patients with good prognosis (cluster 2) and poor prognosis (cluster 1) (P<0.001 using a log rank test). Figure 2B presents a Kaplan-Meier plot of the survival times of Dukes' stage B and C tumors grouped by stage, showing no statistically significant difference.

As noted above, Dukes' staging provides only a probability of survival for each member of a population of patients, based on historical statistics. Accordingly, the prognosis of an individual patient can be predicted based on historical outcome probabilities of the associated Dukes' stage. For example, if a Dukes' C survival rate was 55% at 36 months of follow up, any individual Dukes' C patient would be classified as having a good prognosis since more than 50% of patients would be predicted to be alive.

As illustrated in figures IA, IB, 2A, and 2B, gene expression profiles separate good and poor prognosis cases better than Dukes' staging. This suggests that a gene-expression based classifier described in this example may be used for more accurate prediction of patient prognosis than the traditional Dukes' staging.

The identified genes may also have biological significance. For example, osteopontin, a secreted glycoprotein and ligand for CD44 and αvβ3, appears to have a number of biological functions associated with cellular adhesion, invasion, angiogenesis and apoptosis (see Fedarko NS et ah, "Elevated serum bone sialoprotein and osteopontin in colon, breast, prostate, and lung cancer," Clin Cancer Res, 7:4060-6 (2001); Yeatman TJ and Chambers AF, "Osteopontin and colon cancer progression," Clin Exp Metastasis, 20:85-90 (2003)). Using an oligonucleotide microarray platform, osteopontin was identified as a gene whose expression was strongly associated with colorectal cancer stage progression (Agrawal D et al. , "Osteopontin identified as lead marker of colon cancer progression, using pooled sample expression profiling," J Natl Cancer Inst, 94:513-21 (2002)). INSIG-2, one of the 43 core classifier genes provided in Example 1, was recently identified as an osteopontin signature gene, suggesting that an osteopontin pathway may be prominent in regulating colon cancer survival.

Similarly, neuregulin appeared to have biological significance in the context of colorectal cancer based on frequency of appearance in the classification set of the present invention. Neuregulin, a ligand for tyrosine kinase receptors (ERBB receptors), may have biological significance in the context of colorectal cancer where current data suggest a strong relationship between colon cancer growth and the ERBB family of receptors (Carraway KL, 3rd, et ah, "Neuregulin-2, a new ligand of ErbB3/ErbB4-receptor tyrosine kinases," Nature, 387:512-6 (1997)). Neuregulin was recently identified as a prognostic gene whose expression correlated with bladder cancer recurrence (Dyrskjot L, et al., "Identifying distinct classes of bladder carcinoma using microarrays," Nat Genet, 33:90-6 (2003)).

Accordingly, the identification of such genes may be significant in terms of gene therapy. For example, a therapeutic gene may be identified, which when reintroduced into tumor cells, may arrest or even prevent growth in cancer cells. Additionally, using the classifier of the present invention, a therapeutic gene may be identified that enables increased responsiveness to interventions such as radiation or chemotherapy.

6.3. CONSTRUCTION QF COLORECTAL CANCER CLASSIFIERS USING CDNA

ARRAY DATA A leave-one-out cross-validation (LOOCV) technique was used for constructing and validating a neural network-based classifier using the Moffitt set of 78 tumor samples (see Section 6.1). The samples were classified as having "good" or "poor" prognosis based on survival for more or less than 36 months, respectively. Using the leave-one-out cross- validation approach also provided the ability to rank the selected genes. The number of times a particular gene was chosen was used as an indicator of the usefulness of that gene for general classification and may imply biological significance. Table 4 lists genes used in a cDNA classifier (selected by t-test) and ranked by selection frequency using LOOCV. A set of 43 core genes were identified in 75% of the LOOCV iterations. Table 5 lists the set of 43 genes.

Table 4 Genes used in a cDNA classifier (selected by t-test) and ranked by selection frequency using LOOCV

M denotes genes that were used to classify 75% of all tumors, and genes appearing in both the cDNA classifier and the U133A-limited cDNA classifier are marked by *

Table 5 The 43 enes

The molecular classifier was obtained using a procedure of two distinct steps: gene selection using a t-test and classification using a neural network. Both steps were taken after the test sample was left out (from the leave-one-out cross-validation) to avoid bias from the gene selection step. The top 50 genes as ranked by absolute value of the t statistic using a t test were retained for each cross-validation step. A feed-forward back-propagation neural network with a single hidden layer of 10 units, learning rate of 0.05, and momentum of 0.2 was constructed. Training occurred for a maximum of 500 epochs or until a zero misclassification error was achieved on the training set. It was found that neural networks were extremely robust to both the number of genes selected and the level of noise in these genes.

Using LOOCV, the classifier was evaluated in predicting prognosis for each patient at 36 months follow-up as compared to Dukes' staging predictions. The results of LOOCV demonstrated that the cDNA classifier was 90% accurate (93% sensitivity/84% specificity) in predicting the correct prognosis for each patient at 36 month of follow-up. A log-rank test of the two predicted groups (good and poor prognosis) was significant (P<0.001), demonstrating the ability of the cDNA classifier to distinguish the two outcomes (Figure 2A). Permutation analysis demonstrates the result is better than possible by chance (P<0.001 - 1000 permutations).

This result was also significantly higher than that observed using Dukes' staging as a classifier (77%) for the same group of patients (P=0.03878). The results for both Dukes' staging and molecular staging are summarized in Tables 6A-6C below. Table 6A shows the relative accuracies of Dukes' staging and the cDNA classifier (molecular staging) for all tumors and then a comparison by Dukes' stage. Table 6B shows that Dukes' staging was particularly inaccurate at predicting outcome for patients with poor prognosis (70% and 55% for all stages and B and C, respectively). In contrast, molecular staging, as provided by the present invention, identified the good prognosis cases (the "default" classification using Dukes' staging), but also identified poor prognosis cases with a high degree of accuracy. Table 6C shows the detailed confusion matrix for all samples in the dataset, showing the equivalent misclassification rate of both good and poor prognosis groups by the cDNA classifier.

Table 6A — LOOCV Accuracy of Dukes' vs. Molecular Staging for all tumors.

* Dukes' staging vs. cDNA Classifier, P=0.03878, one-sided McNemar's test.

Leave-one-out cross-validation technique was also utilized to construct a SVM classifier using a set of 72 samples. The classifier was constructed in two steps: first a gene selection procedure was performed with SAM and then a support vector machine was constructed on each fold using all but one sample that was left out.

The gene selection approach used was a univariate selection. SAM (significance analysis of microarrays) was the method chosen for selecting genes. Since gene selected was to be based on two classes (good vs. poor prognosis), the two-class SAM method was used for selecting genes with the best d values. SAM calculates false discovery rates empirically through the use of permutation analysis. SAM provides an estimate of the false discovery rate (FDR) along with a list of genes considered significant relative to censored survival. This feature of SAM was used with this particular embodiment to select the number of genes that resulted in the smallest FDR possible, e.g., zero FDR.

Once the genes were selected using the SAM method, a linear support vector machine (SVM) was constructed. The software used for this approach was implemented in a weka machine learning toolkit. A linear SVM was then chosen to reduce the potential for overfitting the data, given the small sample sizes and large dimensionality. One further advantage of this approach is the transparency of the constructed model, which is of particular interest when comparing the classifier of the subject invention on two different platforms (see below).

A prognostic classifier was also constructed based on an analysis of 20 genes capable of determining survival that is significantly more accurate than Dukes staging. Methods: Seventy-two patients were selected from the Moffitt Cancer Center Tumor Bank and Registry who had been followed for greater than 24 months. Gene expression profiles were created using a 32,000 cDNA microarray and then subjected to Significance Analysis of Microarray (SAM) analysis to identify large sets of discriminating genes. Construction of a Support Vector Machine was then undertaken to develop a classifier capable of predicting outcome. The accuracy of this classifier was assessed by leave one out cross validation and compared to that of standard Dukes staging. Results: Hierarchical clustering and principal component analysis identified two populations of genes distinguishing the majority of patients that had survived greater than 24 months versus those that did not. These observations strongly suggested the potential for underlying gene expression information to drive a prognostic classifier. SVM analysis identified a set of approximately 20 genes whose expression discriminated good from bad survival with better accuracy than Dukes staging at a minimum of 24 months of follow-up (82% vs 77%). Interestingly, one of the 20 genes identified by the SVM as important in survival prediction was osteopontin, a gene we previously reported to be the gene most strongly correlated with advancing tumor stage in colorectal cancer. Conclusion: a molecular classifier was constructed which is capable of predicting outcome for colon cancer that exceeds the accuracy of Dukes staging, particularly in stages B and C where discrimination is critical. This classifier is based on a 20 gene set of which osteopontin, a known gene associated with colon cancer progression, plays a prominent role. The list of 20 genes is provided in the following Table 7. Table 7 The 20 genes used in the SVM classifier

6.4. U133 COLORECTAL CANCER CLASSIFIER

A colorectal cancer survival classifier was obtained using U133A-limited genes selected by LOOCV via statistical analytic tools (Le., t-test). The list of U133A-limited genes selected using LOOCV via t-test is provided in Table 8. The named genes common to both the original classifier (a set of 43 genes) and the U133A-limited classifier are marked with an asterisk in Table 5.

Table 8 Genes used in U133A-limited cDNA classifier (selected by t-test) and ranked by selection fre uenc using LOOCV

M denotes genes used to classify 75% of all tumors, and genes appearing in both the cDNA classifier and U133A-liraited cDNA classifier are marked by *.

The cDNA classifier was tested by applying a classifier to an immediately available, well-annotated, independent test set of colon cancer tumor samples (Denmark set, as described in Section 6.1.) run on the Affymetrix^® platform. Using database software such as the Resourcer software from TIGR (see also Tsai J et al., "RESOURCER: a database for annotating and linking microarray resources within and across species," Genome Biol, 2: software0002.1-0002.4 (2001)), genes were mapped out from the cDNA chip to a corresponding gene on the Affymetrix^® platform. The linkage was done by common Unigene IDs.

12,951 genes (out of 32,000) were mapped to an Affymetrix^® U133A GeneChip. In certain instances, probes on the cDNA chip were unknown expressed sequence tag markers (ESTs) which can reduce the number of usable genes identified. A U133 A-limited cDNA classifier was constructed in accordance with the subject invention by using the identical approach on this reduced set of overlapping genes.

With the U133A-limited cDNA classifier, only those cDNA probes were chosen that (according to Resourcerer) mapped to an Affymetrix^® probe set. This approach enabled cross-platform comparison. For example, the training set samples were used together with the test set tumor samples in a flip-dye design. The end expression value from a cDNA probe was then the Iog2 of the training set to test set sample ratio. This same reference RNA was used on two U133A Affymetrix^® chips. Once the U133A-limited cDNA classifier was constructed, a linear scaling factor based on the expression of a common training set (H. Lee Moffitt Cancer Center & Research Institute, Tampa, Florida) sample applied to both the cDNA microarrays and the Ul 33 A GeneChips, was applied equally to all Affymetrix^® samples (the Moffitt training set as well as the Denmark test set). Using this assumption, the U133A chip value corresponding to a cDNA probe was the ratio of training set to test set sample (on U133A chips). Each of the. Affymetrix^® U133A arrays (both the test set and the reference samples) was scaled to a constant average intensity (150) prior to taking the ratio and the test sample chip values were averaged.

The results of a full LOOCV for the U133A-limited classifier on the test set sample (Moffitt Cancer Center cDNA microarray data set; original 78 samples) are shown in Tables 9A-9C. The accuracy of the U133A-limited classifier was 72% (80% sensitivity/59% specificity), which contrasted from the original cDNA classifier results (90%, P=0.001154). Many ESTs were selected both in the SAM survival analysis and in the original cDNA-based classifier, indicating unknown genes (ESTs) may be very important to colorectal cancer outcome. The U133A-limited classifier was not significantly different, however, than the Dukes' staging (77%), P=0.4862 using a two-sided McNemar's test, and still significantly discriminated the two groups, as can be seen in Figure 3B (P<0.001).

Figure 3A illustrates the survival curve for a cDNA classifier of the subject invention on the 78 training set samples (LOOCV). Figure 3B illustrates the survival curve for the U133A-limited cDNA classifier (LOOCV). Figures 4A-C illustrate the survival curves for an independent test set classification (Denmark test set sample). A large difference in sensitivity can be seen between the Dukes' method and the classifier (Tables 9A-9C). The confusion matrix and accuracy rates by Dukes' stage are also presented in Tables 9A-9C.

With respect to comparing the predictive power of a classifier of the subject invention to Dukes' staging, the U133A-limited classifier was tested on the test set of colorectal cancer samples from Denmark that were profiled on the Affymetrix^® U133A platform. The normalized and scaled test-set data were evaluated with the U133A-limited cDNA classifier. Because the Denmark cases included only Dukes' stages B and C, classification of outcome by Dukes' staging would predict all samples to be of good prognosis. The accuracy of the cDNA classifier was reduced from 72% in LOOCV of the training set (Tables 9A-9C) to 68% in the Denmark cross-platform test set (Tables 10A- 10C). A reduction in accuracy (4%) was expected due to the limitations imposed by cross-platform analyses, however this reduction was very small compared to that caused by limiting the classifier gene set to U133A content. This result is not significantly different from that achieved by classification using Dukes' staging (64%, P=0.7194 using a two sided McNemar's test) and is better than other reported results (47%) (see Sorlie T et al., "Repeated observation of breast tumor subtypes in independent gene expression data sets," Proc Natl Acad Sci USA, 100:8418-23 (2003)) for cross-platform analyses where scaling was required. Moreover, the classifier of the subject invention was able to predict the outcome for poor prognosis patients (sensitivity) with an accuracy of 55% whereas 0% would be predicted correctly by Dukes' staging.

Figures 4A-4C show results from the independent Test Set Evaluation (Denmark Test Set) using the U133A data set. A) Survival curves generated using probe sets corresponding to 26 of the Molecular Classifier genes. Using these translated probe sets, 95 tumors were clustered and censored survivorship was evaluated (P < 0.001). B) Survival curves using Dukes' staging criteria show no significant difference in outcome. C) Survival curves grouped by both Dukes' stage and molecular signature shows that both Dukes' B and C cases can be further subdivided into good and poor prognosis groups.

7. REFERENCES CITED

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

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

Claims

WHAT IS CLAIMED IS:

1. A method for determining a prognosis of colorectal cancer in a colorectal cancer patient, comprising classifying said patient as having a good prognosis or a poor prognosis using measurements of a plurality of gene products in a cell sample taken from said patient, said gene products being respectively products of at least 5 of the genes listed in Table 1 or respective functional equivalents thereof, wherein said good prognosis predicts survival of a patient within a predetermined time period from obtaining a tumor sample from said patient by surgery or from diagnosis of colorectal cancer, and said poor prognosis predicts non- survival of a patient within said time period.

2. The method of claim 1, wherein said plurality of gene products are of at least 5 of the genes listed in Table 1.

3. The method of claim 1, further comprising obtaining said marker profile by a method comprising measuring said plurality of gene products in said tumor cell sample.

4. The method of claim 1, wherein said time period is 3 years.

5. The method of claim 1, wherein said time period is 4 years.

6. The method of claim 1, wherein said time period is 5 years.

7. The method of claim 1, wherein said classifying is carried out by a method comprising using a prognosis predictor, wherein said prognosis predictor receives an input comprising said measurements and provides an output comprising data indicating a good prognosis or a poor prognosis.

8. The method of claim 4, wherein said prognosis predictor is trained with training data from a plurality of colorectal cancer patients, wherein said training data comprise for each of said plurality of colorectal cancer patients (a) measurements of said plurality of gene products in a cell sample taken from said patient and (b) information with respect to whether survival for said time period occurred or not.

9. The method of claim 8, wherein said prognosis predictor is an artificial neural network (ANN).

10. The method of claim 9, wherein said ANN is a feed-forward back-propagation neural network with a single hidden layer of 10 units, a learning rate of 0.05, and a momentum of 0.2.

11. The method of claim 1, wherein said prognosis predictor is a support vector machine (SVM).

12. The method of claim 11, wherein said SVM is a SVM having a dot product kernel.

13. The method of claim 11, wherein said SVM is a SVM having a kernel selected from the group consisting of a ^-degree dot product kernel and a Gaussian kernel.

14. The method of claim 1, wherein said gene products are products of at least 10 of the genes, respectively, listed in Table 1.

15. The method of claim 14, wherein said gene products are products of at least 20 of the genes, respectively, listed in Table 1.

16. The method of claim 14, wherein said gene products are products of at least 40 of the genes, respectively, corresponding to the markers listed in Table 1.

17. The method of claim 14, wherein said gene products are products of at least 70 of the genes, respectively, corresponding to the markers listed in Table 1.

18. The method of claim 1, wherein said gene products are products of the 53 genes, respectively, listed in Table 2.

19. The method of claim 1, wherein said gene products are products of the 7 genes, respectively, listed in Table 3.

20. The method of claim 1, wherein said gene products are products of the 43 genes, respectively, listed in Table 5.

21. The method of claim 1, wherein said gene products are products of the genes, respectively, identified by a * in Table 5.

22. The method of claim 18, 19 or 20, wherein said time period is 3 years, and wherein each of said gene products is a gene transcript.

23. The method of clam 22, wherein a support vector machine or neural network is used in said classifying step."

24. The method of claim 1, wherein each of said gene products is a gene transcript.

25. The method of claim 24, wherein measurement of each said gene transcript is obtained by a method comprising contacting a positionally-addressable microarray with nucleic acids from said cell sample or nucleic acids derived therefrom under hybridization conditions, and detecting the amount of hybridization that occurs, said microarray comprising one or more polynucleotide probes complementary to a hybridizable sequence of each said gene transcript.

26. The method of claim 25, wherein said microarray is selected from the group consisting of cDNA microarray, ink-jet synthesized microarray, and oligonucleotide microarray.

27. The method of claim 1, wherein each of said plurality of gene products is a protein.

28. The method of claim 1, wherein said patient has an increased level of a gene product of osteopontin or neuregulin 2 isoform 4 relative to the average level of said gene product in patients without colorectal cancer.

29. A method for evaluating whether a colorectal cancer patient should be treated with chemotherapy, comprising (a) classifying said patient as having a good prognosis or a poor prognosis using a method as claimed in any one of claims 1, 2, 3, 7, 15 and 21; and (b) determining that said patient's predicted survival time favors treatment of the patient with chemotherapy if said patient is classified as having a poor prognosis.

30. The method of claim 29, further comprising additionally staging said patient using Duke staging.

31. A method for enrolling colorectal cancer patients for a clinical trial of a chemotherapeutic agent for colorectal cancer, comprising (a) classifying each patient as having a good prognosis or a poor prognosis using a method as claimed in any one of claims 1, 2, 3, 7, 15 and 21; and (b) assigning each patient having a good prognosis to one patient group and each patient having a poor prognosis to another patient group, at least one of said patient group being enrolled in said clinical trial.

32. A method for identifying a set of genes for prognosis of colorectal cancer, comprising:

(a) determining for each of a plurality of genes a metric of correlation between abundance level of a gene product of said gene and survival outcome in a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples;

(b) selecting one or more genes based on said metric of correlation.

33. The method of claim 32, wherein said step (a) is carried out by a method comprising

(al) dividing said plurality of colorectal cancer patients into a first group consisting of one or more patients who are living at said predetermined time and a second group consisting of one or more patients who are not living at said predetermined time; and

(a2) determining a difference in abundance levels between said first group and said second group, wherein said difference represents said metric of correlation.

34. The method of claim 33, wherein said difference for each of said plurality of genes is a relative difference of a significance analysis of microarray (SAM) of expression levels of said gene between said first group and said second group.

35. The method of claim 32, wherein said predetermined time is 3 years.

36. A method for identifying a set of genes for prognosis of colorectal cancer, comprising:

(a) generating a subset of patients by leaving out one or more patients in a plurality of patients having known outcomes at a predetermined time after obtaining tumor samples;

(b) determining for each of a plurality of genes a metric of correlation between abundance level of said gene and survival outcome in said subset of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples; (c) selecting one or more genes based on said metric of correlation;

(d) repeating steps (a)-(c) for a plurality of iterations, each with a different subset of patients by leaving out one or more patients in said plurality, wherein said one or more patients are different from any previous iteration; and

(e) selecting one or more genes that are selected in at least a predetermined percentage of all iterations.

37. The method of claim 36, wherein said step (c) is carried out by a method comprising

(cl) ranking said plurality of genes according to said metric; and

(c2) selecting a given number of genes that ranked with the highest correlation.

38. The method of claim 37, wherein said given number is 50 and said predetermined percentage is about 75%.

39. The method of claim 38, wherein said difference for each of said plurality of genes is an absolute t-value of a t-test of expression levels of said gene between said first group and said second group.

40. The method of claim 39, wherein said selecting is carried out by a method comprising selecting a gene if the p- value of said gene corresponds to a predetermined significance level of p-value less than 0.05.

41. The method of claim 40, wherein said predetermined time is 3 years.

42. A method of identifying genes that discriminate between colorectal cancer patients that have a poor prognosis and colorectal patients that have a good prognosis comprising analyzing survival data and RNA levels of colorectal cancer patients using SAM, clustering analysis, or a neural network to select genes whose RNA levels correlate with a selected survival time.

43. A method for constructing a prognosis predictor for prognosis of colorectal cancer, comprising: (a) generating a subset of patients by leaving out one or more patients in a plurality of patients having known outcomes at a predetermined time after obtaining tumor samples;

(b) determining for each of a plurality of genes a metric of correlation between expression level of said gene and survival outcome in a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples from a plurality of colorectal cancer patients having known outcomes at a predetermined time after obtaining tumor samples;

(c) selecting one or more genes based on said metric of correlation;

(d) training a prognosis predictor, wherein said prognosis predictor receives an input comprising a marker profile comprising expression levels of said one or more genes selected in step (c) and provides an output comprising data indicating a good prognosis or a poor prognosis, with training data from said subset of patients, wherein said training data comprise for each of said subset of patients a marker profile comprising measurements of said one or more genes in a tumor cell sample taken from said patient and prognosis information;

(e) determining a prognosis for at least one of said one or more patients who are left out in step (a);

(f) repeating steps (a)-(e) for a plurality of iterations, each with a different subset of patients by leaving out one or more patients in said plurality, wherein said one or more patients are different from any previous iteration;

(g) selecting one or more genes that are selected in at least a predetermined percentage of all iterations; and

(h) training a prognosis predictor, wherein said prognosis predictor receives an input comprising a marker profile comprising expression levels of said one or more genes selected in step (g) and provides an output comprising data indicating a good prognosis or a poor prognosis, with training data from said subset of patients, wherein said training data comprise for each of said plurality of patients a marker profile comprising measurements of said one or more genes in a tumor cell sample taken from said patient and prognosis information.

44. The method of claim 43, wherein said step (c) is carried out by a method comprising (cl) ranking said plurality of genes according to said metric; and

45. The method of claim 44, wherein said prognosis predictor is an artificial neural network (ANN).

46. The method of claim 43, wherein said step (b) is carried out by a method comprising

(bl) dividing said plurality of colorectal cancer patients into a first group consisting of one or more patients who are living at said predetermined time and a second group consisting of one or more patients who are not living at said predetermined time; and

(b2) determining a difference in expression levels between said first group and said second group, wherein said difference represents said metric of correlation.

47. A computer system comprising a processor, and a memory coupled to said processor and encoding one or more programs, wherein said one or more programs cause the processor to carry out the method of any one of claims 1, 32, 35, and 43.

48. A computer program product for use in conjunction with a computer having a processor and a memory connected to the processor, said computer program product comprising a computer readable storage medium having a computer program mechanism encoded thereon, wherein said computer program mechanism may be loaded into the memory of said computer and cause said computer to carry out the method of any one of claims 1, 32, 35, and 43.

49. A microarray comprising for each of a plurality of genes, said genes being at least 5 of the genes listed in Table 1, one or more polynucleotide probes complementary and hybridizable to a sequence in said gene, wherein polynucleotide probes complementary and hybridizable to said genes constitute at least 50% of the probes on said microarray.

50. The microarray of claim 49, wherein said plurality of genes is at least 10 of the genes listed in Table 1.

51. The microarray of claim 49, wherein said plurality of genes is at least 20 of the genes listed in Table 1.

52. The microarray of claim 49, wherein said plurality of genes comprises the 43 genes listed in Table 5.

53. A kit comprising the microarray of claim 49 in a sealed container.

54. The method of claim 1, wherein said plurality of gene products comprises gene products of osteopontin and neuregulin 2 isoform 4, respectively.