NZ537160A - Medical applications of adaptive learning systems - Google Patents

Abstract

A system for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more conditions comprises: a) an input capable of receiving one or more sets of gene expression data categorised into one or more predetermined conditions; b) a neural network module adapted to receive the gene expression data and one or more predetermined conditions; c) an extraction component for extracting rules from the rule base layer and d) an identifier for identifying over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more conditions. The neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions. The adaptive component is arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data. The rules represent the relationships between the gene expression data and the one or more conditions.

Description

53 7 1 6 0 NEW ZEALAND PATENTS ACT 1953 No: Divided out of No. 517817 Date: Dated 15 March 2002 COMPLETE SPECIFICATION MEDICAL APPLICATIONS OF ADAPTIVE LEARNING SYSTEM USING GENE EXPRESSION DATA We, PACIFIC EDGE BIOTECHNOLOGY LIMITED, of The Centre for Innovation, 87 St David Street, Dunedin, New Zealand, do hereby declare the invention for which we pray that a patent may be granted to us, and the method by which it is to be performed, to be particularly described in and by the following statement: la MEDICAL APPLICATIONS OF ADAPTIVE LEARNING SYSTEMS USING GENE EXPRESSION DATA FIELD OF INVENTION The invention relates to medical applications of adaptive learning systems using gene expression data, including but not limited to methodologies for disease profiling and gene expression profiling.

BACKGROUND TO INVENTION Many diseases and disorders, such as cancer, have very complex genetic and phenotypic abnormalities and an unpredictable biological behaviour. The cancer cell represents the end-point of successive generations of clonal cell evolution, multiple gene mutations, 15 genomic instability, and erroneous gene expression.

The biological behaviour of cancer is determined by multiple factors, most importantly the biological characteristics of the individual cancer, but also the biology of the patient such as age, sex, race, genetic constitution and the like, and the location of the cancer. This 20 biological and genetic complexity of cancer means that any individual cancer may follow an unpredictable clinical course, with an uncertain outcome for the patient Where multiple treatment options are available for a particular cancer, it is necessary to have an accurate prognosis for the patient, so that treatment can be tailored to the 25 individual disease of that patient The clinical and information tools currently available to clinicians for the classification . and prognostic evaluation of cancer have serious limitations, especially when applied to an individual patient. It would be desirable to integrate the clinical and biological 30 information for a given cancer in an individual patient's particular cancer.

Gene expression data is available using standard microarray data. Gene expression microarrays, such as available from Affymetrix™, provide a large volume of data and can be used to characterise a particular disease or condition in a patient by comparing diseased 2 PCT/N Z03/00045 or abnormal tissue with healthy normal tissue. However, the data obtained can be difficult to process to obtain meaningful infomiation about a particular condition or disease.

This problem is particularly acute in medical applications relating to patient treatment. In 5 order to influence patient management in a clinical environment, clinical decision support systems must have a high level of confidence. Shipp et al have elegantly demonstrated the potential of machine learning techniques for prognostic stratification of patients, however their approach misclassified 30% of the patients in terms of predicting the outcome of their treatment. They achieved 70% correct prognosis of cured cases of B-cell lymphoma 10 cancer, and wrongly predicted 12% of the cases as cured in contrast to the actual fatal outcome. This accuracy is not appropriate for a clinical application of the model.

Another difficulty with prior art approaches using machine learning is that they often do 15 not provide an easy means for updating the model should new data become available. Instead, complete retraining of the model is required. This is time-consuming and potentially expensive as it involves intensive computational resources. It is desirable that any system used be able to adapt to the addition of new data without complete retraining of the system.

It has been found by the inventors that in a general view, techniques utilising Evolving Connectionsist Systems (ECOS) techniques have the following advantages when compared with the traditional statistical and neural network techniques: (i) they have a flexible structure that reflects the complexity of the data used for their training; (ii) they 25 perform both clustering and classification/prediction; (iii) the models can be adapted on new data without the need to be retrained on old data; (iv) they can be used to extract rules (profiles) of different sub-classes of samples. The rules (profiles) are fuzzy with some statistical coefficients attached.

It is therefore an object of the present invention to provide a method for determining a relationship between gene expression data and one or more conditions or prognostic outcome, or at least to provide the public with a useful choice, intellectual property office of n.z. 1 3 JUN 2006 Sicsiygft 3 SUMMARY OF INVENTION In broad terms in one aspect of the invention comprises a neural network module comprising an input layer comprising one or more input nodes configured to receive gene 5 expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions.

In another aspect, the present invention provides a generic method for determining a relationship between gene expression data and one or more conditions including at least the steps of: a) providing sets of gene expression data categorised into one or more predetermined 15 classes of condition; b) training a neural network module on said gene expression data and said one or more predetermined classes of condition, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more classes of condition; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more classes of condition; and c) extracting rules from the rule base layer, said rules representing relationships 25 between the gene expression data and the one or more classes of condition. hi another aspect, the present invention provides a system for determining a relationship between gene expression data and one or more conditions comprising: a) an input capable of receiving sets of gene expression data categorised into one or 30 more predetermined classes of condition; b) a neural network module trainable on said gene expression data and said one or more predetermined classes of condition, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one WO 03/079286 PCT/NZ03/00045 4 or more output nodes configured to oulput one or more classes of condition; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more classes of condition; and c) means for extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more classes of condition.

In another aspect, the present invention provides a generic method for diagnosing a condition in a patient comprising determining whether gene expression data extracted 10 from a biological sample from said patient satisfies rules representing relationships between the gene expression data and one or more classes of condition, said rules determined by a method of the invention.

In another aspect, the present invention provides a generic system for diagnosing a 15 condition in a patient comprising means for determining whether gene expression data extracted from a biological sample from said patient satisfies rules representing relationships between the gene expression data and one or more classes of condition, said rules determined by a method of the invention.

In another aspect, the present invention provides a method for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more conditions comprising: a) providing one or more sets of gene expression data categorised into one or more predetermined conditions; b) training a neural network module on said gene expression data and one or more predetermined conditions, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more oulput nodes configured to output one or more conditions; and an adaptive component configured to extract one or 30 more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) permitting the adaptive component to aggregate selected two or more rule nodes in the rule-base layer; d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; e) identifying over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more conditions. hi another aspect, the present invention provides a system for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more conditions comprising: a) an input capable of receiving one or more sets of gene expression data categorised into one or more predetermined conditions; b) a neural network module adapted to receive said gene expression data and one or more predetermined conditions, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer 15 comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rales from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; 20 d) an extraction component for extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; e) an identifier for identifying over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-25 expressed genes linked to the one or more conditions.

Preferably, the rules extracted permit specific gene expressions to be linked to a particular condition.

In another aspect, the present invention provides a method for gene expression set reduction comprising: a) providing one or more sets of gene expression data categorised by one or more conditions; 6 PCT/N Z03/00045 b) training a neural network module on said gene expression data and the one or more classes of condition, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) permitting the adaptive component to aggregate selected two or more rule nodes in the rule-base layer; d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; e) identifying genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more conditions.

In another aspect, the present invention provides a system for gene expression set reduction comprising: a) input means for receiving one or more sets of gene expression data categorised by one or more conditions; b) a neural network module trainable on said gene expression data and the one or more classes of condition, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and die one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; d) rule extraction means adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; e) an identifier adapted to identify genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more conditions. 7 In another aspect, die present invention provides a neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive 5 component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the gene expression data.

In another aspect, the present invention provides rules representing relationships between gene expression data and one or more conditions when extracted from a neural network according to a method of the invention.

. In another aspect, the present invention provides a method of training a neural network to 15 diagnose a condition, the method including at least the steps of: a) providing gene expression data categorised by one or more conditions; b) training a neural network module on the gene expression data and the one or more conditions, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions; c) testing the data using a leave one out method; d) reducing input gene expression data to give best test accuracy; e) modifying the neural network module to accept the reduced gene expression data as its input layer; f) training the modified neural network module; g) extracting rules from the adaptive component; h) optionally repeating the method from the reduction step. hi another aspect, the present invention provides a system for training a neural network to diagnose a condition, the system comprising: 8 a) an input able to receive providing gene expression data categorised by one or more conditions; b) a neural network module trainable on the gene expression data and the one or more conditions, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions; d) an input gene expression data reducer operable to reduce gene expression data to give best test accuracy; e) a neural network modifier to modify the neural network module to accept the reduced gene expression data as its input layer; g) a rule extracting means for extracting rules from the adaptive component; Preferably, the above extracted rules are represented in human readable form. In a preferred embodiment, the neural network module further comprises a pruning algorithm arranged to prune nodes in the rule base layer not demonstrating a sufficient link to the one or more conditions.

In broad terms in another aspect of the invention comprises a neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an oulput layer comprising one or more output nodes configured to output one or more prognostic 25 outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes.

In another aspect, the present invention provides a generic method for determining a 30 relationship between gene expression data and prognostic outcome including at least the stqps of: a) providing sets of gene expression data classified by a predetermined prognostic outcome; 9 b) training a neural network module on said gene expression data and prognostic outcome, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an oulput layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules fiom the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes; c) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and one or more prognostic outcomes.

In another aspect, the present invention provides a generic system for condition prognosis in a patient comprising a correlating engine adapted to correlate gene expression data extracted from a biological sample from said patient with rules representing relationships between the gene expression data and prognostic outcomes, said rules determined by a 15 method of the invention.

In another aspect, the present invention provides a generic system for determining a relationship between gene expression data and prognostic outcome comprising: a) input for receiving sets of gene expression data classified by a predetermined 20 prognostic outcome; b) a neural network module trainable on said gene expression data and prognostic outcome, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes; c) a rule extractor adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and one or more prognostic outcomes.

In another aspect, the present invention provides a generic method for condition prognosis in a patient comprising correlating gene expression data extracted from a biological sample from said patient with rules representing relationships between the gene expression data and prognostic outcomes, said rules determined by a method of the invention.

In another aspect, the present invention provides a generic system for condition prognosis 5 in a patient comprising a correlating means adapted to correlate gene expression data extracted from a biological sample from said patient with rules representing relationships between the gene expression data and prognostic outcomes, said rules determined by a method of the invention.

In another aspect, the present invention provides a method for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more prognostic outcomes comprising: a) providing one or more sets of gene expression data categorised into one or more predetermined prognostic outcomes; b) training a neural network module on said gene expression data and one or more predetermined prognostic outcomes, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an oulput layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive 20 component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) permitting the adaptive component to aggregate selected two or more rule nodes in 25 the rule-base layer; d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes; e) identifying over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more prognostic outcomes.

In another aspect, the present invention provides a system for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more prognostic outcomes comprising: 11 a) input for receiving one or more sets of gene expression data categorised into one or more predetermined prognostic outcomes; b) a neural network module trainable on said gene expression data and one or more predetermined prognostic outcomes, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to oulput one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, 10 said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; d) a rule extractor adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes; e) an identifier able to identify over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more prognostic outcomes.

In another aspect, the present invention provides a njethod for gene expression set 20 reduction comprising: a) providing one or more sets of gene expression data categorised by one or more prognostic outcomes; b) training a neural network module on said gene expression data and the one or more classes of condition, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component 30 arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) permitting the adaptive component to aggregate selected two or more rule nodes in the rule-base layer; 12 d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes; e) identifying genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more prognostic outcomes.

In another aspect, the present invention provides a system for gene expression set reduction comprising: a) input to receive or more sets of gene expression data categorised by one or more prognostic outcomes; b) a neural network module trainable on said gene expression data and the one or more classes of condition, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to 15 extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; d) a rule extractor adapted to extract rules from the rule base layer, said rules 20 representing relationships between the gene expression data and the one or more prognostic outcomes; e) an identifying means adapted to identify genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more prognostic outcomes.

In another aspect, the present invention provides a neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more oulput nodes configured to output one or more prognostic outcomes; and an 30 adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the gene expression data. 13 Iii another aspect, the present invention provides rules representing relationships between gene expression data and one or more prognostic outcomes when extracted from a neural network according to a method of the invention.

In another aspect, the present invention provides a method of training a neural network to provide a prognostic outcome, the method including at least the steps of: a) providing gene expression data categorised by one or more prognostic outcomes; b) training a neural network module on the gene expression data and the one or more prognostic outcomes, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes; 15 c) testing the data using a leave one out method; d) reducing input gene expression data to give best test accuracy; e) modifying the neural network module to accept the reduced gene expression data as its input layer; f) training the modified neural network module; g) extracting rules from the adaptive component; h) optionally repeating the method from the reduction step.

In another aspect, the present invention provides a system of training a neural network to provide a prognostic outcome, the system comprising: a) input for receiving gene expression data categorised by one or more prognostic outcomes; b) a neural network module trainable on the gene expression data and the one or more prognostic outcomes, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule 30 base layer comprising one or more rule nodes; an oulput layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes; d) a reducer for reducing input gene expression data to give best test accuracy; 14 e) a modifier adapted to modify the neural network module to accept the reduced gene expression data as its input layer; g) a rule extractor for extracting rules from the adaptive component; Preferably, the above extracted rules are represented in human readable form.

In a preferred embodiment, the neural network module further comprises a pruning algorithm arranged to prune nodes in the rule base layer not demonstrating a sufficient link to the one or more prognostic outcomes.

In one embodiment, the rules extracted permit profiling of breast adenocarcinoma, prostate adenocarcinoma, lung adenocarcinoma, colorectal adenocarcinoma, lymphoma, bladder transitional cell carcinoma, melanoma, uterine adenocarcinoma, leukemia, renal cell carcinoma, pancreatic adenocarcinoma, ovarian adenocarcinoma, pleural mesothelioma, and central nervous system cancer.

In another preferred embodiment, the methods and systems of the invention are applied to profiling breast adenocarcinoma, prostate adenocarcinoma, lung adenocarcinoma, colorectal adenocarcinoma, lymphoma, bladder transitional cell carcinoma, melanoma, uterine adenocarcinoma, leukemia, renal cell carcinoma, pancreatic adenocarcinoma, 20 ovarian adenocarcinoma, pleural mesothelioma, and central nervous system cancer.

In another embodiment, the rules extracted permit profiling of DLBCL (Diffuse Large B-Cell Lymphoma).

In another preferred embodiment, the methods and systems of the invention are applied to profiling DLBCL (Diffuse Large B-Cell Lymphoma).

In another aspect, the present invention provides a gene profile for any one of 14 types of cancer: Breast-, Prostate-, Bladder -, Leukemia-, Lymphoma-, Central nervous system-, 30 Lung-, Colorectal- Melanoma, Uterine-, Renal cell-, Pancreatic-, Ovarian-, and Pleural Cancers as set forth in the Figures as read with the description herein. hi another aspect, the present invention provides a gene profile for any one of 14 types of cancer: Breast-, Prostate-, Bladder Leukemia-, Lymphoma-, Central nervous system-, intellectual property office of n.z. m'mwm REVIVED Lung-, Colorectal- Melanoma, Uterine-, Renal cell-, Pancreatic-, Ovarian-, and Pleural Cancers comprising at least one of the genes as set forth in the Figures as read with the description herein.

Preferably the profile comprises at least two genes in the profile, more preferably at least three genes in the profile, most preferably at least four genes in the profile.

This specification also describes a diagnostic kit for diagnosing a cancer selected from the group comprising Breast-, Prostate-, Bladder -, Leukemia-, Lymphoma-, Central nervous system-, Lung-, Colorectal- Melanoma, Uterine-, Renal cell-, Pancreatic-, Ovarian-, and Pleural Cancers comprising at least one nucleic acid sequence that selectively ligates to at least one of the gene expression products of the genes as set forth in the Figures for each of the respective cancers in question as read with the description herein.

Preferably, the kit also comprises an amount of the at least one nucleic acid sequence able to quantitatively determine the amount of the gene expression products present in a sample.

Also described is a diagnostic kit for diagnosing a cancer selected from the group comprising Breast-, Prostate-, Bladder Leukemia-, Lymphoma-, Central nervous system-, Lung-, Colorectal- Melanoma, Uterine-, Renal cell-, Pancreatic-, Ovarian-, and Pleural Cancers comprising a ligand capable of selectively binding to a peptide expressed from a nucleic acid sequence expressed from at least one of the genes as set forth in the Figures for each of the respective cancers in question as read with the description herein.

Preferably, the ligand is an antibody or an immunomolecule, such as a fab fragment, hi a preferred embodiment, the ligand is specifically capable of binding to the peptide in question. Preferably, the kit also comprises an amount of ligand able to quantitatively determine the amount of peptide present in a sample.

Also described is a method for diagnosing whether a patient is suffering from a cancer selected from the group comprising Breast-, Prostate-, 16 PCT/N Z03/00045 Bladder -, Leukemia-, Lymphoma-, Central nervous system-, Lung-, Colorectal-Melanoma, Uterine-, Renal cell-, Pancreatic-, Ovarian-, and Pleural Cancers comprising the steps of: a) isolating a sample from the patient; b) determining whether the sample contains expression levels consistent with a disease profile of at least one gene in the as set forth in the Figures for each of the respective cancers in question as read with the description herein.

Preferably the sample is a tissue, more preferably a tissue suspected of being cancerous.

The diagnostic kits and the method for diagnosing whether a patient is suffering from a cancer are described and claimed in NZ patent specification which is divided from this specification.

BRIEF DESCRIPTION OF THE FIGURES Preferred forms of the method and system for disease profiling will now be described with reference to the accompanying figures in which: Figure 1 is a schematic block diagram view of a neural network module useful in the invention; Figure 2 is a flow chart illustrating a preferred method of the invention; Figure 3 shows rules for each cancer class extracted from a trained EFuNN; Figure 4a shows gene expression level of genes for each of the genes identified as relevant by the EFuNN. Shaded areas indicate areas of underexpression; Figure 4b shows gene expression level of genes for each of the genes identified as relevant by the EFuNN. Shaded areas indicate areas of overexpression; Figure 5a shows class profiles for 14 types of cancer extracted through applying the methods of the invention highhghting genes that are underexpressed; 17 Figure 5b shows class profiles for the 14 types of cancer in Figure 5a highlighting genes that are overexpressed; Figure 6 shows a gene expression class profile of Breast adenocarcinoma genes.

Figure 7a shows 10 gene expression subgroup profiles for Breast adenocarcinoma highlighting underexpressed genes.

Figure 7b shows 10 gene expression subgroup profiles for Breast adenocarcinoma highlighting overexpressed genes.

Figure 8 shows a gene expression profile of prostate cancer genes.

Figure 9a shows 5 gene expression subgroup profiles for prostate cancer highlighting underexpressed genes.

Figure 9b shows 5 gene expression subgroup profiles for prostate cancer highlighting overexpressed genes.

Figure 10 shows the gene expression profile of Lung adenocarcinoma genes.

Figure 11a shows 8 gene expression subgroup profiles for Lung adenocarcinoma highlighting underexpressed genes.

Figure lib shows 8 gene expression subgroup profiles for Lung adenocarcinoma highlighting overexpressed genes.

Figure 12 shows the gene expression profile for Colorectal adenocarcinoma genes.

Figures 13a and 13b show the 5 gene expression subgroup profiles for Colorectal adenocarcinoma highlighting underexpression and overexpression respectively.

Figure 14 shows the gene expression profile for Lymphoma genes. 18 Figures 15a and 15b show the 6 gene expression subgroup profiles for Lymphoma highlighting underexpression and overexpression respectively.

Figure 16 shows the gene expression profile for Bladder transitional cell carcinoma genes.

Figures 17a and 17b show the 7 gene expression subgroup profiles for Bladder transitional cell carcinoma highlighting underexpression and overexpression respectively.

Figure 18 shows the gene expression profile fat Melanoma genes.

Figures 19a and 19b show the 5 gene expression subgroup profiles for Melanoma highlighting underexpression and overexpression respectively.

Figure 20 shows the gene expression profile for Uterine adinocarcinoma genes.

Figures 21a and 21b show the 4 gene expression subgroup profiles for Uterine adenocarcinoma highlighting underexpression and overexpression respectively.

Figure 22 shows the gene expression profile for Leukemia genes.

Figures 23a and 23b show the 5 gene expression subgroup profiles for Leukemia highlighting underexpression and overexpression respectively.

Figure 24 shows the gene expression profile for Renal cell carcinoma.

Figures 25a and 25b show the 6 gene expression subgroup profiles for Renal cell carcinoma highlighting underexpression and overexpression respectively.

Figure 26 shows the gene expression profile far Pancreatic adenocarcinoma genes. 19 Figures 27a and 27b show the 6 gene expression subgroup profiles for Pancreatic adenocarcinoma highlighting underexpression and overexpression respectively.

Figure 28 shows the gene expression profile for Ovarian adenocarcinoma genes.

Figures 29a and 29b show the 7 gene expression subgroup profiles for Ovarian adenocarcinoma highlighting underexpression and overexpression respectively.

Figure 30 shows the gene expression profile for Pleural mesothelioma genes.

Figures 31a and 31b show the 4 gene expression subgroup profiles for Pleural mesothelioma highlighting underexpression and overexpression respectively.

Figure 32 shows the gene expression profile for the Central nervous system genes.

Figures 33a and 33b show the 3 gene expression subgroup profiles for Central nervous system cancer highlighting underexpression and overexpression respectively.

Figure 34a and 34b show prognostic outcomes of various cancer types using methods of the invention highlighting fatal and cured conditions respectively.

DETAILED DESCRIPTION OF PREFERRED FORMS In the present specification, "biological sample" means a sample including at least a nucleic acid molecule or a polypeptide molecule. Such samples may conveniently be sourced from cells. The cells may, in turn, be extracted from tissues. Pre-processing of 30 samples may take place by, for example, extracting DNA, RNA and/or polypeptides from a raw sample using well-known techniques in the art. Other pre-processing may be conducted by, for example, pre-digesting the sample using DNAses, RNAses and proteinases.

A "condition" is a disease, disorder or trait selected from the group comprising cancer, a congenital disease and a hereditary disease. A "condition" also includes non-disease causing traits. A "predisposition to a condition" may also include a genetic predisposition 5 to a condition. A non-limiting example of a non-disease causing trait is alopecia.

Evolving connectionist systems ("ECOS") are multi-modular, connectionist architectures that facilitate modelling of evolving processes and knowledge discovery (PCT, WO 01/78003) An ECOS may consist of many evolving connectionist modules. An ECOS is a neural network system that operates continuously in time and adapts its structure and functionality through a continuous interaction with the environment and with other systems according to: (i) a set of parameters P that are subject to change during the system operation; (ii) an incoming continuous flow of information with unknown 15 distribution; (iii) a goal (rationale) criteria (also subject to modification) that is applied to optimise the performance of the system over time.

The evolving connectionist systems have the following specific characteristics: (1) They evolve in an open space, not necessarily of fixed dimensions. (2) They learn in on-line, pattern mode, incremental learning, fast learning - possibly by 20 one pass of data propagation. (3) They learn in a life-long learning mode. (4) They learn as both individual systems, and evolutionary population systems. (5) They have evolving structures and use constructive learning. (6) They learn locally and locally partition the problem space, thus allowing for a fast 25 adaptation and tracing the evolving processes over time. (7) They facilitate different kinds of knowledge, mostly combined memory based, statistical and symbolic knowledge.

There are two distinct phases of ECOS operation. During the first, the learning phase, data vectors are fed into the system one by one with their known oulput values. Li the 30 second phase (recall) a new vector is presented to the system and it calculates the output values for it 21 PCT/N Z03/00045 There are different models of ECOS protected by (PCT, WO 01/78003 Al). One of them, evolving fuzzy neural networks (EFuNN) is presented in fig.l and its algorithm given in fig.2. EFuNN can be used for both classification tasks and prediction tasks.

While '^neural network module" may refer to any neural network satisfying the requirements of the aspects of the invention above, the use of an ECOS neural network is preferred. Particularly preferred is the neural network exemplified in PCT publication WO 01/78003 (incorporated herein by reference). The algorithm describing the neural network from WO 01/78003 is set out below: Hie EFuNN learning algorithm (from PTO WO 01/78003 Al).

Set initial values for the system parameters: number of membership functions; initial sensitivity thresholds 15 (default Sj=0.9); error threshold E; aggregation parameter Nagg - number of consecutive examples after each aggregation is performed; pruning parameters OLD an Pr; a value for m (in m-of-n mode); maximum radius limit Rmax; thresholds Ti and T2 for rule extraction.

Set the first rule node ro to memorise 1he first example (x,y): { Evaluate the local normalised fuzzy distance D between Xf and the existing rule node connections W1 (formulae (1)) Calculate the activation Al of the rule node layer. Find fhe closest rule node rk (or the closest m rule Wl(r0)=xf, and W2(r0)=yg 20 Loop over presentations of new input-output pairs (x,y) (10) nodes in case of m-of-n mode) to the fuzzy input vector Xf for which Al(rt) >= St (sensitivity threshold for the node r^), //"there is no such a node, create a new rule node for (x&y^ else Find the activation of the fuzzy output layer A2=W2.Al(l-D(Wl,xf))j and the normalised output error Err= || y- y'|| / Nout l/Err >E create a new rule node to accommodate the current example (x&yt) else 40 Update W1 (r^) and W2(ri) according to (2) and (3) (in case of m-of-n system update all them rule nodes with the highest Al activation).

Apply aggregation procedure of rule nodes after each group of Nagg examples are presented Update the values for the rule node it parameters St, R& Agefrt), TA (ft).

Prune rule nodes if necessary, as defined by pruning parameters.

Extract rules from the rule nodes ( > 22 Another example of a particularly preferred neural network module for some aspects of the invention is an ECF.

Evolving classification function (''ECF"), can be used to classify data. The learning sequence of each iteration of ECF is described in the following steps: 1) if all vectors have been inputted, finish the current iteration; otherwise, input a vector from the data set and calculate the distances between the vector and all rule nodes already created; 2) if all distances are greater than a max-radius parameter, a new rule node is created, the position of the new rule node is the same as the current vector in the input data space and its radius is set to the min-radius parameter, and then go to step 1; otherwise: 3) if there is a rule node with a distance to the current input vector less then or equal to its radius and its class is the same as the class of the new vector, nothing will be changed and go to step 1; otherwise: 4) if there is a rule node with a distance to the input vector less then or equal to its radius and its class is different from those of the input vector, its influence field should be reduced, the radius of the new field is set to the larger value from the distance minus the min-radius, and the min-radius. 5) if there is a rule node with a distance to the input vector less then or equal to Ihe max-radius, and its class is the same to the vector's, enlarge the influence field by taking the distance as the new radius if only such enlarged field does not cover any other rule node which has the different class; otherwise, create a new rule node the same way as in step 2, and go to step 1.

The recall (classification phase of new input vectors) in ECF is performed in the following way: 1) if the new input vector lies within the field of one or more rule nodes associated with one class, the vector belongs to this class; 2) if the input vector lies within the fields of two or more rule nodes associated with different classes, the vector will belong to the class corresponding the closest rule node. 3) if Ihe input vector does not lie within any field, then there are two cases: (1) one-of-n mode: the vector will belong to the class corresponding the closest rule node; (2) m-of-n mode: take m highest activated by the. new vector rule nodes, and calculate the average 23 distances from the vector to the nodes with the same class; the vector will belong to the class corresponding the smallest average distance.

The above-described ECF for classification has several parameters that need to be 5 optimized according to the data set used. These are: 1) maximum radius 2) minimum radius 3) number of membership functions (mf) 4) m-of-n value 5) number of iterations for the data presentation during learning phase.

These parameters can be optimized with the use of evolutionary computation methods, or other statistical methods.

On one embodiment, the invention employs an adaptive method for gene reduction, model creation and profiling. The methodology employs ECOS. The methodology is comprised of the following main phases: (1) Train continuously an evolving connectionist system (ECOS) on incoming data thus creating a "mother" system that accommodates all available data; 20 (2) Extract features (genes) relevant to the output classes from the "mother" system; (3) Create a model based on the selected features and the output classes. (4) Extract profiles (rules) Feature selection in the adaptive component of an ECOS is preferably performed through 25 the extraction of rules from the ECOS created by supervised training on all available data. The ECOS training parameters are optimized so that the classification error is minimised and the ECOS models most closely the features present in the data. Each node in the hidden layer of the ECOS represents the center of a cluster of similar samples and can be expressed semantically as a rule. Each rule relates to the pattern of input feature levels for 30 one or more samples belonging to a particular class from the data set An example of what a rule might look like when extracted from the EFuNN is shown below: IF VAR1 is LOW (0.80) and VAR3 is HIGH (0.76) and 24 VAR12 is HIGH (0.91) and VAR25 is LOW (0.80) and VAR31 is LOW (0.87) and then CLASS_Z is VERY LIKELY (with a membership degree of 0.92), Accommodated Training Examples in this rule are 10 out of 50, Radius of the cluster for this rule is 0.15.

The rules are then analysed in order to identify a set of variables that are significant in distinguishing between classes. This is achieved by ranking each variable according 10 to its importance in the rules for each class using a standard signal-to-noise ration method as used in (Ramaswami et al; Dudoit et al)).

Using this method each input variable is assigned a value between -1 and 0 for each class. For each class, variables are then selected if their rank value is above a set threshold value. 15 This value is altered in order to select an optimal set of input features. Any variables that are selected through more than one class are only included once in the feature (gene) set thus representing a reduced gene set.

Once the feature selection (gene selection) phase is complete, the original data set is minimised by removing any features not present in the feature set. This new data is then 20 used to train a new ECOS. With the reduced feature space the time for training will be significantly reduced. The performance of the ECOS should be evaluated and training parameters modified so that the classification error is minimized and the generalization ability of the model is maximized.

The invention provides an information system which can integrate and interpret complex gene expression data and which can be adapted to the diagnostic evaluation, prognostic assessment and clinical management of patients. The method is based on knowledge-based neuro-computing which uses the learning ability of a supervised learning neural . network to learn patterns from input-output data pairs and then extract rules from the 30 structure.

Once a profile for a particular disease is extracted, the genes used in the profile may be used as input variables in a new ECOS model trained and tested in a leave-one-out method (see Dudoit et all, Shipp et al) on all available data to evaluate accuracy.

We have found that ECOS described in WO 01/78003 are particularly suited for complex disease profiling based on a variety of information sources, including gene expression data.

In one form, the invention can be applied to develop new methods for disease classification, outcome of treatment prognosis and drug response prognosis. Disease profiles can also be extracted through the method described above. Profiles extracted from the network may define tumor subgroups which can be interpreted by determination of a 10 function of the genes that constitute the rules. These genes will correspond to over-expressed and under-expressed genes that contribute to tumorigenesis, or are a consequence of tumorigenesis. Different rules would be extracted to include different genes contributing to the same functional unit, for example signaling pathway or cell adhesion complex, or different genes from redundant functional units, for example parallel 15 signaling pathways that lead to the same phenotypic outcome.

These general functional units could include the evasion of apoptosis, growth signaling, angiogenesis, tissue invasion/metastasis and replicative potential.

Analysis of the identity and function of genes in these rules would therefore be predicted to identify the functional units that are critically perturbed in the tumors and the consequences of these perturbed functions.

The genes which form the rules most correspond to transcripts that are expressed in the 25 tumor tissue-specific patterns. These may be tissue-specific genes that are over-expressed in the tumor tissue but do not contribute to the disease progression, for example insulin in an insulinoma or certain differentiation markers, or over-expressed and under-expressed genes that play a role in tumourigenesis but in a tumor specific pattern these genes would be a subset of those identified above, namely these genes will exclude genes which 30 contribute to functional units common to multiple cancers.

In one form of the invention the input data comprises raw gene expression data. The input data is first filtered as described above using a suitable filtering, normalisation or log transformation process. A standard statistical test, for example a T test or other correlation 26 could be applied to this pre-processed data and the genes obtained from this data that best distinguish the output classes. Ni genes could be selected from an initial set of N where Ni is less than N that are above a chosen threshold. Each gene could also be ranked.

The input data could comprise the Ni gene inputs as the individual expression values of the N selected genes, such as those mentioned above. The outputs could be, for example, 14 different types of cancer. The model or models would be trained on this input and output data as described above with reference to Figure 2. The model is then tested on a reduced input data set of features Ni.

After training each feature, for example a gene belonging to a membership function defining a high value Gh is ranked for each class CI based on calculated values Gh, I as follows: Gh,I = (Gh,mi-Gh,ma)(Gh,mr-Gh,ma) Gh,mi is the mean value for gene G is high for all rule nodes from the trained network that support class that CI, Gh, ma is the mean value for gene G is high for all rule nodes from the trained network across output classes, and Gh, mr is the mean value for gene G is high 20 for all rule nodes from the trained network that support other classes than class CL A threshold Gthr is applied and a new set of features, in this case genes, is selected for each class includes all genes that have at least one membership degree above the threshold. The features of all classes are combined together to form a new feature set of N2 genes.

Rules are then extracted from the model, each rule representing a profile of a group or cluster from the N-dimensional gene expression space.

In some cases there will be new input data available and the structure of ECOS provides 30 means for the addition of new data, therefore updating the model with new input data with the possibility of gaining improved accuracy.

Early diagnostic applications 27 Identification of specific antigens in body fluids including, for example, blood, urine, peritoneal washes and stool extracts can provide a valuable approach for the early diagnosis of disease and other conditions, leading to early treatment and improved prognosis. Specific antigens also can provide a means for monitoring disease or condition 5 progression, enabling the efficacy of surgical, radiotherapeutic and chemotherapeutic treatments to be tracked. However, for a number of diseases and conditions, the available markers suffer from insufficient sensitivity and specificity.

In the case of a number of diseases and conditions, proteins can be present in body fluids 10 at elevated levels compared to individuals without malignant disease, and can be sufficiently stable to enable immunodetection. Body fluids include blood, urine, sputum, semen, gastric fluids and stool. Where such body fluids are not useful, biopsies of suspect tissues may be used. Overexpression or underexpression can also be detected by either nucleic acid detection or protein detection techniques in fluids if they contain cells, or cell 15 lysates that are released from suspect tissues.

Proteins of interest can also be detected in body fluids. For example, immuno-detection techniques using monoclonal or polyclonal antibodies raised against either whole proteins 20 peptides of interest Peptides of interest can be either synthetic or expressed in in vitro or in vivo systems. Immunodetection techniques can include, but not be limited to, ELISA/EIA, radioimmunoassay, nephelometry, immunoturbidometric assays, chemiluminescence, immunofluorescence (by microscopy or flow cytometry), immunohistochemistry and Western blotting. It can be readily appreciated that other 25 methods for detecting proteins can be used.

Kits Based on the discoveries of this invention, several types of test kits can be produced. 30 First, kits can be made that have a detection device pre-loaded with a detection molecule (or "capture reagent"). In embodiments for detection of mRNA, such devices can comprise a substrate (e.g., glass, silicon, quartz, metal, etc) on which oligonucleotides as capture reagents that hybridize with the mRNA to be detected. In some embodiments, direct detection of mRNA can be accomplished by hybridizing mRNA (labeled with cy3, 28 cy5, radiolabel or other label) to the oligonucleotides on the substrate. In other embodiments, detection of mRNA can be accomplished by first making complementary DNA (cDNA) to the desired mRNA. Then, labeled cDNA can be hybridized to the oligonucleotides on the substrate and detected.

Regardless of the detection method employed, comparison of test expression with a standard measure of expression is desirable. For example, RNA expression can be standardized to total cellular DNA, to expression of constitutively expressed RNAs (for example, ribosomal RNA) or to other relatively constant markers.

Antibodies can also be used in kits as capture reagents. In some embodiments, a substrate (e.g., a multiwell plate) can have a specific capture reagent attached thereto. In some embodiments, a kit can have a blocking reagent included. Blocking reagents can be used to reduce non-specific binding. For example, non-specific oligonucleotide binding can 15 reduced using excess DNA from any convenient source that does not contain oligonucleotides for detection, such as salmon sperm DNA. Non-specific antibody binding can be reduced using an excess of a blocking protein such as serum albumin. It can be appreciated that numerous methods for detecting oligonucleotides and proteins are known in the art, and any strategy that can specifically detect associated molecules can be 20 used and be considered within the scope of this invention.

In embodiments relying upon antibody detection, proteins or peptides of interest can be expressed on a per cell basis, or on the basis of total cellular, tissue, or fluid protein, fluid volume, tissue mass (weight). Additionally, XXX in serum can be expressed on the basis of a relatively high-abundance serum protein such as albumin.

In addition to a substrate, a test kit can comprise capture reagents (such as probes), washing solutions (e.g., SSC, other salts, buffers, detergents and the like), as well as detection moieties (e.g., cy3, cy5, radiolabels, and the like). Kits can also include 30 instructions for use and a package.

Methods for Making Antibodies to proteins in profile To make antibodies against the proteins, any method known in the art can be used. For examole. oolvclonal antisera can be made by injecting isolated proteins, peptides or 29 mixtures of proteins and peptides into a suitable animal, such as a rabbit In some embodiments, an adjuvant can be used to augment the immune response. After suitable booster injections, serum can be collected and either used as serum, or, in alternative embodiments, IgG fractions can be produced from the serum.

In alternative embodiments, monoclonal antibodies can be made against the proteins or peptides using standard methods known in the art. Briefly, an isolated protein or peptide preparation is injected in to a suitable animal (e.g., a mouse), and an immune response is . elicited. The spleen of the animal is removed, and splenocytes can be fused with myeloma cells to produce hybridomas. Hybridomas producing antibodies directed towards the target 10 proteins or peptides can be selected and cell cultures expanded to produce desired amounts of monoclonal antibodies. Antibodies can be further selected that have desired affinity and cell lines can be selected that have desirable growth characteristics and antibody production.

Peptides can be made from intact protein isolated from a tissue of interest (e.g., gastric 15 tumor tissue) or can be chemically synthesized. In other embodiments, protein can be made using recombinant methods from cDNA specific for full-length target proteins or oligonucleotides encoding (in frame) a portion of the target proteins.

The antibody kit would consist of antibodies raised against the minimum number of 20 proteins that characterize the profile. It could also include all the reagents for carrying out the assay.

RNA based Idts Kits can be made that have a detection device pre-loaded with a detection molecule (or 25 "capture reagent"). In embodiments for detection of mRNA from genes within a genetic profile, such devices can comprise a substrate (e.g., glass, silicon, quartz, metal, etc) on which oligonucleotides as capture reagents that hybridize with the mRNAs to be detected. In some embodiments, direct quantitative detection of mRNA can be accomplished by hybridizing mRNAs (labeled with cy3, cy5, radiolabel or other label) to the 30 oligonucleotides on the substrate. In other embodiments, detection of mRNAs can be accomplished by first making complementary DNA (cDNA) to the desired mRNAs. Then, labeled cDNAs can be hybridized to the oligonucleotides on the substrate and detected.

PCT/N Z03/00045 cDNAs can also be detected using the Applied Biosystems Taqman™ procedure in which specific fluorescently-labelled oligonucleotide probes are hybridized to the target cDNA as it is amplified in a real-time PCR assay. Other quantitative detection methods including, but not limited to, molecular beacons and SyBr green labeling of PCR product, can also be 5 used.

In one technique useful herein, relevant RNA can be detected by RT-PCR using oligonucleotide primers specific for conserved sequences within the gene or genes identified by methods of the present invention. Total RNA can be extracted from body 10 fluid by standard techniques, converted to cDNA using reverse transcriptase and then amplified using PCR. Desirable PCR primers flank intronic DNA to prevent the PCR amplification of genomic DNA. Results would be detected using any conventional method, for example, either by gel electrophoresis or by quantitative real-time PCR techniques. Real-time PCR can be carried out using either fluorescently-labelled 15 Taqman™ probes or direct binding of fluorescent dyes such as Sybr Green to the PCR product However, it can be appreciated that other methods of detecting nucleic acids of interest can be used.

In addition to a substrate, a test kit can comprise capture reagents (such as probes), 20 washing solutions (e.g., SSC, other salts, buffers, detergents and the like), as well as detection moieties (e.g., cy3, cy5, radiolabels, and the like). Kits can also include instructions for use and a package.

Drug target validation Once a gene set is identified by a method of the invention, the individual genes can be developed as drug targets following validation as described below. The drug target is typically a protein that is over-expressed within a diseased or malfunctioning cell. Drug candidates would typically then be manufactured against the drug targets. Monoclonal antibodies produced against the protein and synthetic chemicals that bind the protein. The 30 invention therefore extends to methods of identifying drug targets and to drug candidates. 31 One method of determining the presence of a drug target is to stably transfect a gene identified as implicated in a condition or disease by a method of the invention into cancer lines with an inducible promoter. The protein expressed by the gene is then expressed and the effect of over expression on the cell line's viability is determined. The expression of 5 the gene in cell lines is knocked out rising either small interfering RNAs (siRNAs) or antibodies raised against the protein. The effect on cell viability is determined. If the gene in question has a desired effect on the cell either by killing the cell or by causing the cell to revert to a normal form then it is a drug target.

Disease management markers The methods of the present invention enable the creation of disease management markers. Such markers permit a practitioner to determine whether a disease is present, and if it is how, it is responding to treatment. Profiles identified by a method of the invention that correspond to a specific disease characteristic would be developed as disease management 15 markers as follows: Detection kits for Disease management markers would typically consist of a panel of antibodies raised against the gene profile identified by EfuNN or a kit capable of testing expression at the RNA level. The present invention extends to such kits.

An antibody kit can be developed using the genes identified in the methods of the present invention that are over- or under-expressed in a patient with a specific disease or condition. Antibodies or other immunomolecules against the proteins encoded by the identified genes are manufactured by standard techniques in the art Biopsy samples of 25 diseased or abnormal tissue would be taken from a patient and then analysed using the antibodies or immunomolecules by immunodetection techniques including but not be limited to, ELISA/EIA, radioimmunoassay, nephelometry, immunoturbidometric assays, chemiluminescence, immunofluorescence (by microscopy or flow cytometry), immunohistochemistry and Western blotting for levels of the proteins that are over-or 30 under-expressed in the diseased tissue. The presence of abnormal levels of the proteins in question are indicative of the presence of the specific disease or condition.

WO 03/079286 PCT/NZ03/00045 32 Using similar techniques, a patient with the specific disease or condition can be monitored to determine whether treatment is having the desired effect.

In an alternative embodiment, a RNA based assay may be employed. Total RNA or 5 mRNA would be extracted from the biopsy sample using standard techniques, then applied in a standard nucleic acid kit by ligating the RNA to one or more antisense markers indicative of the presence of a gene identified by a method of the present invention as over- or under-expressed in a specific disease or condition and determining the level of bound RNA. The level of bound RNA in a sample would be indicative of the absence or 10 presence of the specific disease or condition. Using this technique, a patient with the specific disease or condition can be monitored to determine whether treatment is having the desired effect. * Methods for Making Antibodies to proteins in profile To make antibodies against the proteins, any method known in the art can be used. For example, polyclonal antisera can be made by injecting isolated proteins, peptides or mixtures of proteins and peptides into a suitable animal, such as a rabbit. In some embodiments, an adjuvant can be used to augment the immune response. After suitable booster injections, serum can be collected and either used as serum, or, in alternative 20 embodiments, IgG fractions can be produced from the serum.

In alternative embodiments, monoclonal antibodies can be made against the proteins or peptides using standard methods known in the art. Briefly, an isolated protein or peptide preparation is injected in to a suitable animal (e.g., a mouse), and an immune response is elicited. The spleen of the animal is removed, and splenocytes can be fused with myeloma 25 cells to produce hybridomas. Hybridomas producing antibodies directed towards the target proteins or peptides can be selected and cell cultures expanded to produce desired amounts of monoclonal antibodies. Antibodies can be further selected that have desired affinity and cell lines can be selected that have desirable growth characteristics and antibody production.

Peptides can be made from intact protein isolated from a tissue of interest (e.g., gastric tumor tissue) or can be chemically synthesized. In other embodiments, protein can be 33 made using recombinant methods from cDNA specific for full-length target proteins or oligonucleotides encoding (in frame) a portion of the target proteins.

The antibody kit would consist of antibodies raised against the minimum number of 5 proteins that characterize the profile. It could also include all the reagents for carrying out the assay.

RNA based kits Kits can be made that have a detection device pre-loaded with a detection molecule (or 10 "capture reagent"). In embodiments for detection of mRNA from genes within a genetic profile, such devices can comprise a substrate (e.g., glass, silicon, quartz, metal, etc) on which oligonucleotides as capture reagents that hybridize with the mRNAs to be detected, hi some embodiments, direct quantitative detection of mRNA can be accomplished by hybridizing mRNAs (labeled with cy3, cy5, radiolabel or other label) to the 15 oligonucleotides on the substrate. In other embodiments, detection of mRNAs can be accomplished by first making complementary DNA (cDNA) to the desired mRNAs. Then, labeled cDNAs can be hybridized to the oligonucleotides on the substrate and detected. cDNAs can also be detected using the Applied Biosystems Taqman™ procedure in which 20 specific fluorescently-labelled oligonucleotide probes are hybridized to the target cDNA as it is amplified in a real-lime FOR assay. Other quantitative detection methods including, but not limited to, molecular beacons and SyBr green labeling of PCR product, can also be used.

In addition to a substrate, a test kit can comprise capture reagents (such as probes), washing solutions (e.g., SSC, other salts, buffers, detergents and the like), as well as detection moieties (e.g., cy3, cy5, radiolabels, and the like). Kits can also include instructions for use and a package.

The invention will be described below with reference to non-limiting examples: 34 EXAMPLE 1 Figure 1 shows a neural network module 22. The preferred structure is a fuzzy neural network which is a connectionist structure which implements fuzzy rules. The neural 5 network module 22 includes input layer 40 having one or more input nodes 42 arranged to receive input data. This input data will depend on the particular application to which the neural network module or modules are directed.

The neural network module also includes output layer 56 having one or more output nodes 10 58. The oulput nodes represent the real values of the output variables.

The input nodes and output nodes are configured depending on the type of information to be retrieved from the system. In one application involving gene expression profiling and classification of disease, the input data is a set of relevant variables, such as gene 15 expression data, and the output variables are categories of diseases or prognostic outcomes over a time scale.

In another configuration for disease prognostic profile development and disease prognosis over a time scale, the input data comprises individual expression values of selected genes 20 and Ihe output data includes different types of disease, such as different types of cancer.

In a further configuration in relation to drug response prognosis and profile development, the inputs are configured as gene expression data for a particular patient group given a particular treatment regime and the output data is prognosis.

The neural network module 22 may further comprise fuzzy input layer 44 having one or more fuzzy input nodes 46. The fuzzy input nodes 46 transform data from the input nodes 42 for the further use of the system. Each of the fuzzy input nodes 46 could have a different membership function attached to it, for example a triangular membership 30 function, Gaussion function or any other known function suitable for the purpose. The main purpose of the fuzzy input nodes 46 is to transform the input values from the input nodes 42 into membership degrees to which the values belong to the membership function.

WO 03/079286 PCT/NZ03/00045 The neural network module 22 may further comprise a fuzzy output layer 42 having one or more fuzzy output nodes 54. Each fuzzy node 54 represents a fuzzy quantisation of the output variables, similar to the fuzzy input nodes 46 of the fuzzy input layer 54. Preferably, a weighted sum input function and saturated linear activation function are used 5 for the nodes to calculate the membership degrees to which the output vector associated with the presented input vector belongs to each of the output membership functions.

The neural network module 22 may also include a short term memory layer 60 having one or more memory nodes 62. The purpose of the short term memory layer 60 is to memorise 10 structurally temporal relationships of the input data. The short term memory layer is preferably arranged to receive information from and send information to the rule base layer 48.

As more particularly described in WO 01/78003, each rule node 50 represents an 15 association between a hyper sphere from the fuzzy input space and a hyper sphere from the fuzzy output space. Each rule node rj has a sensitivity threshold parameter Sj which defines the minimum activation threshold of the rule node rj to a new input vector x from a new example or input (x, y) in order for the example to be considered for association with this rule node. A new input vector x activates a rule node if x satisfies the minimum 20 activation threshold and is subsequently considered for association with the1 rule node. The radius of the input hypersphere is defined as Rj = 1-Sjj Sj being the sensitivity threshold parameter.

Fuzzy logic rules and other types of knowledge are able to be extracted from the trained 25 neural network or combination of neural networks in an easily accessible form. One example of a fuzzy rule extracted from a train network is: R1: IF [gene 1 is High to a degree of 0.9] and [gene 3 is High to a degree of 0.9] and [gene 8 is Low to a degree of 0.8] {radius of the receptive field = 0.109} 30 THEN Disease A {accommodated training examples = 15 out of 24} These types of rules represent relationships between the input data and the output data. They provide a profile from which knowledge can be gained about the classification or the prognostic process of a disease. The rules point to profiles of genes and clinical 36 information which is strongly associated with a specific disease, for example different types of cancer, and can be used for the development of new tests and treatments.

EXAMPLE 2 Figure 2 illustrates one method of disease profiling.

The input data optionally has preprocessing 100 performed on it. Such preprocessing includes filtering, correlation evaluation, normalisation, log transformation and/or noise 10 reduction.

Input data and desired output data is used to train 110 the model. The training is performed by a supervised learning algorithm, for example one of the supervised learning algorithms described in WO 01/78003.

Hie model is tested 120 on reduced input data. In this example, the input data comprises 198 tumor samples and a separate model could be trained for each tumor sample, resulting in 198 models. Each model is trained on a reduced input data set, for example 197 of the 198 samples and then tested on the sample not included in the reduced input data set. The error could then be evaluated and an average classification error calculated.

Table 1 sets out sample results from testing on reduced input data. It shows two different neural network models having different parameter values, trained on 139 selected genes out of 16,036 micro array gene expression data with Hie use of correlation statistic analysis to evaluate the relation between genes and classes for 14 classes of cancer data. The results show that the neural network of the invention is effective in reducing the genes linked to a particular condition. The results also show that a high degree of accuracy may be obtained in predicting cancer outcome and cancer type.

Referring back to Figure 2, following testing of the model or models on reduced input data, the reduced input data set that gives the best accuracy is chosen for a final classification system development and for extracting the profiles.

Tablet 37 PCT/N Z03/00045 Number of Errthr- error Aggregation Activation Average Leave-one-out membership threshold after every function: 1- Number of method junctions MF number of linear;2-RBF Rules accuracy examples 2 0.9 - 1 37.9 79.3 1 0.9 200 2 48.8 81.9 The model is then trained 130 based on the selected reduced input data as will be more particularly described in different applications below. Various rules may then be extracted 140 from the model. If there is new input data available 150, this new input data can then 5 be used to further train the model or models.

Figure 3 illustrates a set of 38 rules over 60 genes. Each group of samples representing a class of cancer is represented by at least one rule and in many cases by several rules, each rule representing samples of a distinct cluster. A rule represents a profile of the input features, for example genes, that is characteristic for a cluster.

To find a common pattern between all clusters of the same class, an aggregation procedure is applied that results in a single profile for each class, depending on a threshold Qhr used to select genes that have a higher value than this threshold value for any of their 15 membership degrees. These higher values could include a high expression value or a low expression value.

A profile of class CI will include gene G is high, Gh with a value of Gh, I, if Gh appears in all rule nodes j that support class CI above the threshold Qhr and the value Gh, I is 20 calculated as the mean of all Gh values across these rule nodes.

Described below are techniques and a methodology based on adaptive, learning, evolving connectionist systems (ECOS) for modelling, prognosis and rule extraction on the same gene expression data as in Shipp et al but with the addition of clinical information that is 25 available for the patients — the length of survival and the IPI (International Prognostic Index) number. Two experiments are presented.

In the first one, the same set of 11 genes as in Shipp et al is used to evolve a prognostic model. Using cross-validation (leave-one-out method) on the whole set of 58 samples (32 cured, and 26 fatal), 90% prediction rate (93% versus 14%, respectively) was obtained. The experiment employs evolving fuzzy neural networks (EFuNN) having the following 38 parameter values: 3 membership functions for the input variables and no fuzzy representation for the output variables; max radius=l; 1 iteration of training.

After the model was evolved from the data, each cluster of data that form a sub-class or a 5 prototype is subsequently mapped with clinical information - the length of survival represented in months trough a fuzzy representation - short (1, less than 12 months), medium (2, between 20 and 40 months), long (between 40 and 60 months). Rules (profiles) of gene expression that characterise each of these prototypes are extracted. The profiles identify different patterns of gene expression for each outcome class and each 10 length of survival category.

The results can be visualised as in Table 2.

Table 2 Gene 1 Gene 2 Gene 3 Gene 4 Gene5 Gene6 Name Dystrophi n related protein Protein kinase C gamma MINO R/NOR 1 PDE4B Protein kinase C beta-1 Zink-finger protein C2H2-150 Class Cured 0 58 58 58 58 43 Class Fatal 58 4 54 23 1 57 Figure 34a shows in greyscale clusters of the fetal class, while Figure 34b shows in greyscale clusters of the cured class. It can be seen that fatal cases of category 2 and 3 are closer in terms of gene expression profile to some of the cured. There is a significant difference between some of the fatal cases of category 1 and the fetal cases of category 2 20 and 3, which can be detected through gene expression and can be correctly predicted by the model. In addition to predicting a new case, the model can give as a result the highly matched prototype which will indicate a prognosis about the survival time for the new case.

As an example, one of the extracted rules from the trained EFuNN that defines a profile of the class fetal, category 1, is given below: 39 Rule 4: IF G3 is(Low 0.83 ) and G4 is(High 0.83 ) and G8 is( Medium 0.91 ) and G9 is (Low 0.83 ) and G11 is(Low 0.83 ) THEN Class is [fatal], where Gl, etc are the genes listed in the same order as in Shipp et al (incorporated hereing 5 in full by reference).

In a second experiment, the same raw Affymetrix gene expression data of 6,817 genes for 58 patients as in Shipp et al is used, but separate models for each IPI class are created with the use of the same EFuNN technique with different parameter values. For each 10 model, 5 genes are selected out of the initial 6,817 using the same signal to noise ratio analysis as in Shipp et al. The accuracy of the prognosis (cured vs fatal )four each IPI class model is as follows: for IPI==1 (low), 91.9±3.9; for IPI=2 (low intermediate), 96.4 ± 5.7; for IPI=3 (high intermediate), 90.4 ±6.1; for 1PI=4 the prognosis is 100% as all classes belong to the class 'fatal'.

The samples were stratified according to their IPI: IPI = low: 26 samples, of which 19 belong to the cured class and 7 to the fatal class.

IPI = low intermediate: 11 samples, of which 7 belong to the cured class and 4 to the fatal class.

IPI = high intermediate: 17 samples, of which 4 belong to the cured class and 13 to the fatal class.

IPI = high : 2 samples , of which 2 belong to the fatal class. No further classification by microarray data necessary.

DPI = unknown: 2 samples. These samples were not considered here.

For each of the 3 IPI categories, gene selection and prognosis was performed on a N-cross validation basis. First, 5 genes were selected for each IPI in each cross validation run. Classification was performed by EFuNN with the following parameter values: membership functions MF=2, error threshold Errth= 0.1, no aggregation applied; 30 Euclidean distance for measuring the similarity between input vectors; radial basis activation function for the rule nodes in the EFuNN model; 3 iterations over the training data set in each cross validation model. Running 30 N-cross validation procedures for each IPI group, the classification performance was assessed as shown in Table 3. 40 Table 3 .Test accuracy in the N-cross validation experiments for the three IPI models IPI Low Low intermediate High intermediate Original group 91.9 ±3.9 96.4 ±5.7 90.4 ±6.1 As a baseline, groups with number of samples and the same class percentage with random IPI selection were formed. In this case the accuracy is lower and variation is higher, e.g. 5 the variation for each of the three models is respectively: ± 6.4, ± 9.9, ±6.5.

After the cross-validation of the EFuNN models, all examples were used to build the final prognostic models with the use of reduced gene set of 5 genes as defined in the previous 10 step (table 4). Rules were extracted from the trained EFiiNNs, each rule representing a profile of a group (cluster) of samples from the 5 dimensional space. Gene expression profiles can be used not only to predict the outcome of DLBCL patient treatment in one of the 3 IPI categories in a clinical environment, but to direct research of pharmaceutical companies towards new drug targets.

Table 4. Selected genes in the cross validation EFuNN modelling for each of the IPI categories Category A set of selected genes IPI=1 (low) 'KIAA0278 gene, partial cds' 26 'Gene encoding prepro form of corticotropin releasing factor' 25 T>BH Dopamine beta-hydroxylase (dopamine beta-monooxygenase)'25 'SP4 Sp4 transcription factor1 14 'CCAAT BOX-BINDING TRANSCRIPTION FACTOR 1' 14 IPI=2 (low intermediate) * LEM domain protein CLP-36 mRNA' 10 'CES2 Carboxylestease 2 (liver)1 9 "KIAA0197 gene, partial cds' 7 *LYZ Lysozyme' 3 'ApM2 mRNA for GS2374 (unknown product specific to 41 adipose tissue)' 2 IPI=3 ( high intermediate) 'HPrpl8mRNA' 15 'KIAA0036 gene' 15 'D-aspartate oxidase1 15 'Partial mRNA for pyrophosphatase' 12 'JNK activating kinase (JNKK1) mRNA' 5 EXAMPLE 3 The gene expression database published by (Ramaswamy et al.) and publicly available 5 from (http://www-genome.wi.mit.edu/MPR/GCMJitml) is used in this example. This database is incorporated herein by reference. This database contains gene expression values for 90 normal tissue samples and 218 tumor samples from 14 common tumor types. Each sample has the expression level of 16,063 genes and expressed sequence tags (ESTs). 20 of the tumor samples were shown by Ramaswamy et al. to be poorly 10 differentiated resulting in unsatisfactory classification. As we intend to identify patterns in gene expression values that differentiate between tumor types, we used the tumor samples from the database excluding the poorly differentiated samples.

A system to identify patterns (profiles) of gene expression values in each of the 14 tumor 15 classes considered here was developed by applying the feature selection method above and minimised the feature space of the case study data from (Ramaswami et at) from 16,063 to 399 genes. This data set was then used to train an EFuNN to model patterns of gene expression in each of the 14 tumor classes. Rules were extracted from the EFuNN that describe the expression levels of genes for a particular class. A class can be represented by 20 more than one rule, but each rule describes only one class.

A pattern for each tumor class was then generated by combining all the rules relating to that class. A gene is included in the pattern if the membership degree of its expression level (HIGH/LOW) is above a given threshold in all rules relating to that particular class. 25 The threshold gives an indication of the characteristics of the features in the pattern. In this instance the threshold indicates the importance of the genes in the pattern. For example if gene g is present in the profile of Breast Adenocarcinoma when the threshold is 0.9, it is likely that this is an important gene for this class. That is, the membership degree of this 42 gene in the rules extracted from the EFuNN is above 0.9 in all rules related to this class, hi some cases the threshold must be set relatively low. This does not mean that the feature (gene) is not important, just that it is not a strong feature in this pattern even though it is common to all rules related to that class.

The graphical representation is given in figures 5a and 5b. Class profiles of different types of cancer 300 are extracted. The profiles of each class can be modified through a threshold Cthr defining a membership degree above which a gene should be either over-expressed or under-expressed in order to appear in the class profile.

A graduated scale 320 indicates levels of underexpression of genes shown in greyscale in Ihe left top large pane 340, whereas the other graduated scale 330 indicates levels of overexpression of genes in the left top large pane 350. The profiles of each class can be modified through a threshold CW tuned for each individual class that defines the 15 membership degree above which a gene should be either over-expressed or under-expressed in all rules of this class in order for this gene to appear in the profile. Each class of cancer is represented by only one pattern that represents the common genes in all tissue samples of this class either under-expressed, or over-expressed. In addition to this combined class patterns, for each class several subgroups (clusters) can be identified by 20 the method and the characteristic pattern of each sub-group of the same class can be extracted. In the tables below, first the class profiles are extracted, and second for each class - the sub-group profiles are extracted too for a defined threshold that can be modified. Through lowering the threshold, more detailed profiles are extracted.

A pattern is shown on the graph for each of the 14 tumor classes. A line represents in pane 350 a high expression level for a gene and a line in pane 340 represents a low expression level for a gene. The relative strength of these colours represents the average membership| degree for a gene, indicating the strength of the gene's participation in the rule. The graphical interface also allows the user to view the individual rules, as shown in figures (b) 30 afterwards. This allows a visual interpretation of the rules that make up a particular pattern. This may be useful, especially when a pattern appears not to show any common genes. When viewing the individual rules it may be possible to identify genes of interest that do not necessarily appear in all the rules. 43 Accuracy was evaluated in the case of Lymphoma, when the used threshold is 0.6, there are 13 genes selected. The accuracy of the model is 97%. In order to select a minimum number of genes, different models are created out of the 13 genes, starting from 1 gene and their accuracy is evaluated, so that the ratio (Number of Genes)/Accuracy is 5 optimized. For example, using gene D64142 gives and accuracy of 64%, but adding two other genes brings the accuracy (confidence) to 95%.

It is not the intention to limit the scope of the invention to the abovementioned examples only. As would be appreciated by a skilled person in the art, many variations are possible 10 without departing from the scope of the invention (as set out in the accompanying claims).

Figure 6 shows a gene expression class profile of Breast adenocarcinoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number, which is a unique number corresponding 15 to the gene name. Genes numbered on the list with the following order numbers are over-expressed: 2,3. The rest of the genes are under-expressed. Minimum level of fuzzy membership degree of over-, or under expression is 0.55. Figure 7a shows the 10 gene expression subgroup profiles for Breast adenocarcinoma highlighting underexpressed genes. Figure 7b shows the same data but highlighting overexpressed genes.

Figure 8 shows a gene expression profile of prostate cancer genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are under-expressed: 1. The rest of the genes are over-expressed. The minimum 25 level of fuzzy membership degree of over-, or under expression is 0.62. Figures 9a and 9b show the 5 gene expression subgroup profiles for prostate cancer highlighting underexpressed and overexpressed genes respectively.

Figure 10 shows the gene expression profile of Lung adenocarcinoma genes as opposed to 30 the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are under-expressed: 4,8,9,15. The rest of the genes are over-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.6. Figures 44 11a and lib show 8 gene expression subgroup profiles for Lung adenocarcinoma highlighting underexpressed and overexpressed genes respectively.

Figure 12 shows the gene expression profile for Colorectal adenocarcinoma genes as 5 opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 1,2,3,6,8,9,10,17,22. The rest of the genes are under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.65. Figures 13a and 13b show the 5 gene expression subgroup profiles for 10 Colorectal adenocarcinoma highlighting underexpressed and overexpressed genes respectively.

Figure 14 shows the gene expression profile for Lymphoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by 15 their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 3,6,8,9,10. The rest of the genes are under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.6. Figures 15a and 15b show the 6 gene expression subgroup profiles for Lymphoma highlighting underexpressed and overexpressed genes respectively.

Figure 16. shows the gene expression profile for Bladder transitional cell carcinoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 1,2,3,5. Hie rest of the genes are 25 under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.6. Figures 17a and 17b show the 7 gene expression subgroup profiles for Bladder transitional cell carcinoma highlighting underexpressed and overexpressed genes respectively.

Figure 18 shows the gene expression profile for Melanoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 1,3,4,5,6,7. Gene 2 is under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.65. Figures 19a and 19b 45 show the 5 gene expression subgroup profiles for Melanoma highlighting underexpressed and overexpressed genes respectively.

Figure 20 shows the gene expression profile for Uterine adinocarcinoma genes as opposed 5 to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 1,2,3,4,5,6,8,9,10. The rest of the genes on the list below are under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.65. Figures 21a and 21b show the 4 gene expression subgroup profiles for 10 Uterine adenocarcinoma highlighting underexpressed and overexpressed genes respectively. The minimum level of fuzzy membership degree of over-, or under expression is 0.7.

Figure 22 shows the gene expression profile for Leukemia genes as opposed to the profiles 15 of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 1,2,3,4,5,6,7,9. The rest of the genes on the list below are under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.65. Figures 23a and 23b show the 5 gene expression subgroup profiles for Leukemia 20 highlighting underexpressed and overexpressed genes respectively.

Figure 24 shows the gene expression profile for Renal cell carcinoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order 25 numbers are oVer-expressed: 6,7. The rest of the genes on the list below are under-expressed. Minimum level of fuzzy membership degree of over-, or under expression is 0.63. Figures 25a and 25b show the 6 gene expression subgroup profiles for Renal cell carcinoma highlighting underexpressed and overexpressed genes respectively.

Figure 26 shows the gene expression profile for Pancreatic adenocarcinoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 1,2,3,4,5,6,8. The rest of the genes on the list below are under-expressed. Minimum level of fuzzy membership degree of over-, or under 46 expression is 0.6. Figures 27a and 27b show the 6 gene expression subgroup profiles for Pancreatic adenocarcinoma highlighting underexpressed and overexpressed genes respectively.

Figure 28 shows the gene expression profile for Ovarian adenocarcinoma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 9. The rest of the genes on the list below are under-expressed. Minimum level of fuzzy membership degree of over-, or under 10 expression is 0.6. Figures 29a and 29b show die 7 gene expression subgroup profiles for Ovarian adenocarcinoma highlighting underexpressed and overexpressed genes respectively. The minimum level of fuzzy membership degree of over-, or under expression is 0.65.

Figure 30 shows the gene expression profile for Pleural mesothelioma genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 2,13,14,16,17,18,19,20,21. The rest of the genes on the list 20 below are under-expressed. The minimum level of fuzzy membership degree of over-, or under expression is 0.7. Figures 31a and 31b show the 4 gene expression subgroup profiles for Pleural mesothelioma highlighting underexpressed and overexpressed genes respectively.

Figure 32 shows the gene expression profile for the Central nervous system genes as opposed to the profiles of the other 13 types of cancer considered here. The genes are represented by their GenBank accession number. Genes numbered on the list with the following order numbers are over-expressed: 6,7,8,10,11,12,13,14,15. The rest of the genes on the list below are under-expressed. The minimum level of fuzzy membership 30 degree of over-, or under expression is 0.8. Figures 33a and 33b show the 3 gene expression subgroup profiles for Central nervous system cancer highlighting underexpressed and overexpressed genes respectively. The minimum level of fuzzy membership degree of over-, or tinder expression is 0.7. 47 INDUSTRIAL APPLICABILITY The neural networks and methods of this invention are useful in diagnosis, management of disease, evaluating efficacy of therapy, producing reagents and test kits suitable for diagnosing diseases or conditions, evaluating drug targets and the development of drugs 5 for the treatment of a variety of diseases and conditions.

THE FOLLOWING PUBLICATIONS ARE INCORPORATED HEREIN BE REFERENCE: S. Dudoit, J. Fridlyand, and T. P. Speed, Comparison of Discrimination Methods for the Classification of Tumors Using Gene Expression Data, Journal of the American Statistical Association, vol.97, no.457, March, pp.77-87,2002.

T.R. Golub, D. K. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J. P. Mesirov, H. Coller, 15 M. L. Loh, J. R. Downing, M. A. Caligiuri, C. D. Bloomfield, and E. S. Lander, Molecular classification of cancer: class discovery and class prediction by gene expression monitoring, Science, vol.286,15 October, pp.531-537,1999.

J. Khan, J.S.Wei, M. Ringner, L. H. Saal, M. Ladanyi, F. Westermann, F. Berthold, M. 20 Schwab, C. R. Antonescu, C. Peterson, and P. S. Meltzer, Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks, Nature Medicine, Nature Publishing Group, vol.7, no.6, pp.673-679,2001.

N. Kasabov, Evolving Connectionist Systems: Methods and Applications in 25 Bioinformatics, Brain Study and Intelligent Machines, Springer Veriag, 2002.

N. Kasabov, Evolving Connectionist Systems for Adaptive On-line Knowledge-Based Learning, TREE Transactions of Systems, Man and Cybernetics—part B: Cybernetics, vol.31, no.6, pp.902-918,2001.

. N.Kasabov, Adaptive method and system, PCT, WO 01/78003 Al.

S. Ramaswamy, P. Tamayo, R. Rifkin, S. Mukheq'ee, C-H. Yeang, M Angelo, C. Ladd, M. Reich, E. Latulippe, J. P. Mesirov, T. Poggio, W. Gerald, M. Loda, E. S. Lander, and 48 T. R. Golub, Muliclass cancer diagnosis using tumor gene expression signatures, Proceedings of the National Academy of Sciences, vol.98, no.26, pp.15149-15154,2001.

F. P. Roth, Bringing out the best features of expression data, Genome Research, Cold 5 Spring Harbor Laboratory Press, vol.11, no.ll, pp.1878-1887,2001.

M. A. Shipp, K. N. Ross, P. Tamayo, A. P.Weng, J. L. Kutok, R. C. T. Aguiar, M. Gaasenbeek, M. Angelo, M. Reich, G. S. Pinkus, T. S. Ray, M. A. Koval, BC W. Last, A. Norton, T. A. Lister, J. Mesirov, D. S. Neuberg, E. S. Lander, J. C. Aster, and T. R. Golub 10 Diffuse large B-cell lymphoma outcome prediction by gene expression profiling and supervised machine learning, Nature Medicine, vol.8, no. 1, pp.68-74,2002.

T. D. Wu, Analysing gene expression data from DNA microarrays to identify candidate genes, Journal of Pathology, JohnWiley and Sons, Ltd., vol.195, pp.53- 65,2001.

. C-H. Yeang, S. Ramaswamy, P. Tamayo, S. Mukhegee, R. M. Rifkin, M. Angelo, M. Reich, E. Lander, J. Mesirov, and T. Golub, Molecular classification of multiple tumor types, Bioinfortnatics, Oxford University Press, vol.17, Suppl.l, pp.S316-S322,2001. 49

Claims (32)

1. CLAIMS oj:,. 'I'c A neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more oulput nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions.
2. A generic method for determining a relationship between gene expression data and one or more conditions including at least the steps of: a) providing sets of gene expression data categorised into one or more predetermined classes of condition; b) training a neural network module on said gene expression data and said one or more predetermined classes of condition, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more classes of condition; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more classes of condition; and c) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more classes of condition.
3. A system for determining a relationship between gene expression data and one or more conditions comprising: a) an input capable of receiving sets of gene expression data categorised into one or more predetermined classes of condition; b) a neural network module trainable on said gene expression data and said one or more predetermined classes of condition, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more classes of condition; and an adaptive component configured to WO 03/079286 50 PCT/N Z03/00045 extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more classes of condition; and c) means for extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more classes of condition.
4. 4. A generic method for diagnosing a condition in a patient comprising: determining one or more rules by the method of claim 2; determining whether gene expression data extracted from a biological sample from the product satisfies one or more of the determined rules; and diagnosing the condition based at least partly on whether the one or more determined rules are satisfied.
5. 5. A generic system, for diagnosing a condition in a patient comprising means for determining whether gene expression data extracted from a biological sample from said patient satisfies rules representing relationships between the gene expression data and one or more classes of condition, said rules determined by a method of claim 2.
6. 6. A method for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more conditions comprising: a) providing one or more sets of gene expression data categorised into one or more predetermined conditions; b) training a neural network module on said gene expression data and one or more predetermined conditions, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an oulput layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or 51 PCT/NZ03/00045 which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more conditions.
7. A system for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more conditions comprising: a) an input capable of receiving one or more sets of gene expression data categorised into one or more predetermined conditions; b) a neural network module adapted to receive said gene expression data and one or more predetermined conditions, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) an extraction component for extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; and d) an identifier for identifying over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more conditions.
8. A system for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more conditions as claimed in claim 7, wherein the rules extracted permit specific gene expressions to be linked to a particular condition.
9. A method for gene expression set reduction comprising: a) providing one or more sets of gene expression data categorised by one or more conditions; b) training a neural network module on said gene expression data and the one or more classes of condition, the neural netwoik comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base 52 PCT/NZ03/00045 layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) permitting the adaptive component to aggregate selected two or more rule nodes in the rule-base layer; d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; and e) identifying genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more conditions.
10. A system for gene expression set reduction comprising: a) input means for receiving one or more sets of gene expression data categorised by one or more conditions; b) a neural network module trainable on said gene expression data and the one or more classes of condition, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; d) rule extraction means adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more conditions; and e) an identifier adapted to identify genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more conditions.
11. A neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rale base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions, WO 03/079286 53 PCT/NZ03/00045 said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the gene expression data.
12. 12.
13. 13.
14. 14. A method of training a neural network to diagnose a condition, the method including at least the steps of : a) providing gene expression data categorised by one or more conditions; b) training a neural network module on the gene expression data and the one or more conditions, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions; c) testing the data using a leave one out method; d) reducing input gene expression data to give best test accuracy; e) modifying the neural network module to accept the reduced gene expression data as its input layer; f) training the modified neural network module; and g) extracting rules from the adaptive component.
15. A method of training a neural network to diagnose a condition of claim 12 further comprising the step of repeating the method from the reduction step.
16. A system for training a neural network to diagnose a condition, the system comprising: a) an input able to receive providing gene expression data categorised by one or more conditions; b) a neural network module trainable on the gene expression data and the one or more conditions, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising WO 03/079286 54 PCT/N Z03/00045 15. 16. 17. 18. one or more output nodes configured to output one or more conditions; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more conditions; c) an input gene expression data reducer operable to reduce gene expression data to give best test accuracy; d) a neural network modifier to modify the neural network module to accept the reduced gene expression data as its input layer; and e) a rule extracting means for extracting rules from the adaptive component.
17. A system for training a neural network to diagnose a condition as claimed in claim 14, wherein the extracted rules are represented in human readable form.
18. A system for training a neural network to diagnose a condition as claimed in claim 14 or claim 15, wherein the neural network module further comprises a pruning algorithm arranged to prune nodes in the rule base layer not demonstrating a sufficient link to the one or more conditions.
19. A neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more oulput nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes.
20. A generic method for determining a relationship between gene expression data and prognostic outcome including at least the steps of: a) providing sets of gene expression data classified by a predetermined prognostic outcome; b) training a neural network module on said gene expression data and prognostic outcome, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an oulput layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule WO 03/079286 PCT/NZ03/00045 55 19. 20. 21. base layer representing relationships between the gene expression data and the one or more prognostic outcomes; c) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and one or more prognostic outcomes.
21. A generic system for condition prognosis in a patient comprising a correlating engine adapted to correlate gene expression data extracted from a biological sample from said patient with rules representing relationships between the gene expression data and prognostic outcomes, said rules determined by a method of claim 18.
22. A generic system for determining a relationship between gene expression data and prognostic outcome comprising: a) input for receiving sets of gene expression data classified by a predetermined prognostic outcome; b) a neural network module trainable on said gene expression data and prognostic outcome, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes; c) a rule extractor adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and one or more prognostic outcomes.
23. A generic method for condition prognosis in a patient comprising: determining one or more rules by the method of claim 18; determining whether gene expression data extracted from a biological sample from the patient satisfies one or more of the determined rules; and determining condition prognosis based at least partly on whether the one or more determined rules are satisfied.
24. A generic system for condition, prognosis in a patient comprising a coirelatiiig means adapted to correlate gene expression data extracted from a biological sample from: WO 03/079286 56 PCT/NZ03/00045 li. 24. expression data and prognostic outcomes, said rules determined by a method of claim 18.
25. A method for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more prognostic outcomes comprising: a) providing one or more sets of gene expression data categorised into one or more predetermined prognostic outcomes; b) training a neural network module on said gene expression data and one or more predetermined prognostic outcomes, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data; c) permitting the adaptive component to aggregate selected two or more rule nodes in the rule-base layer; d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes; and e) identifying over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more prognostic outcomes.
26. A system for selecting a set of distinguishing over-expressed or under-expressed genes linked to one or more prognostic outcomes comprising: a) input for receiving one or more sets of gene expression data categorised into one or more predetermined prognostic outcomes; b) a neural network module trainable on said gene expression data and one or more predetermined prognostic outcomes, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more oulput nodes configured to oulput one or more prognostic outcomes; and an adaptive component configured to extract one or WO 03/079286 57 PCT/NZ03/00045 25. \e> % & *;more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data;;c) a rule extractor adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes; and d) an identifier able to identify over-expressed or under-expressed genes from the extracted rules, which genes represent a set of distinguishing over-expressed or under-expressed genes linked to the one or more prognostic outcomes.;A method for gene expression set reduction comprising:;a) providing one or more sets of gene expression data categorised by one or more prognostic outcomes;;b) training a neural network module on said gene expression data and the one or more classes of condition, the neural network comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to oulput one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data;;c) permitting the adaptive component to aggregate selected two or more rule nodes in the rule-base layer;;d) extracting rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes;;e) identifying genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more prognostic outcomes.;A system for gene expression set reduction comprising:;a) input to receive or more sets of gene expression data categorised by one or more prognostic outcomes;;b) a neural network module trainable on said gene expression data and the one or more classes of condition, the neural network comprising an input layer;V;WO 03/079286;58;PCT/NZ03/00045;comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the input data;;d) a rule extractor adapted to extract rules from the rule base layer, said rules representing relationships between the gene expression data and the one or more prognostic outcomes; and e) an identifying means adapted to identify genes from the extracted rules, which genes represent a reduced gene expression set linked to the one or more prognostic outcomes.;
27. 27. A neural network module comprising an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes, said adaptive component arranged to aggregate selected two or more rule nodes in the rule base layer based on the gene expression data.;
28. 28. A method of training a neural network to provide a prognostic outcome, the method including at least the steps of:;a) providing gene expression data categorised by one or more prognostic outcomes;;b) training a neural network module on the gene expression data and the one or more prognostic outcomes, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an output layer comprising one or more output nodes configured to output one or more nsral p^operty office OF NX;I 3 J UN 2806;RECEIVED;WO 03/079286;59;PCT/NZ03/00045;prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes;;c) testing the data using a leave one out method;;d) reducing input gene expression data to give best test accuracy;;e) modifying the neural network module to accept the reduced gene expression data as its input layer;;f) training the modified neural network module; and g) extracting rules from the adaptive component;
29. 29. A method of training a neural network to provide a prognostic outcome as claimed in claim 28 further comprising the step of repeating the method from the reducing step.;
30. 30. A system of training a neural network to provide a prognostic outcome, the system comprising:;a) input for receiving gene expression data categorised by one or more prognostic outcomes;;b) a neural network module trainable on the gene expression data and the one or more prognostic outcomes, wherein the neural network module comprises an input layer comprising one or more input nodes configured to receive gene expression data; a rule base layer comprising one or more rule nodes; an oulput layer comprising one or more oulput nodes configured to output one or more prognostic outcomes; and an adaptive component configured to extract one or more rules from the rule base layer representing relationships between the gene expression data and the one or more prognostic outcomes;;c) a reducer for reducing input gene expression data to give best test accuracy;;d) a modifier adapted to modify the neural network module to accept the reduced gene expression data as its input layer; and e) a rule extractor for extracting rules from the adaptive component.;
31. 31. A system of training a neural network to provide a prognostic outcome as claimed in claim 30, wherein the extracted rules are represented in human readable fbmi.;'3 JUN 2006;I!!* IVED , MAR-29-04 05:16PM FROM- T PCTNZ03/00045 Received 29 March 2004 60
32. 32. a system of training a neural network to provide a prognostic outcome as claimed in claim 30 or claim 31, wherein the neural network module further comprises a pruning algorithm arranged to prune nodes in the rule base layer not demonstrating a sufficient link to the one or more prognostic outcomes. A system of training a neural network to provide a prognostic outcome as claimed in any one of claims 30 to 32 wherein the disease profiled is selected from the group of cancers comprising breast adenocarcinoma, prostate adenocarcinoma, lung adenocarcinoma, colorectal adenocarcinoma, lymphoma, bladder transitional cell carcinoma, melanoma, uterine adenocarcinoma, leukemia, renal cell carcinoma, pancreatic adenocarcinoma, ovarian adenocarcinoma, pleural mesothelioma, and central nervous system cancer. PACIFIC EDGE BIOTECHNOLOGY LIMITED By the authorised agent 33.