Data Mining of Gene Expression Profiles for the Diagnosis and Understanding of Diseases Limsoon Wong Institute for Infocomm Research Copyright 2003 limsoon wong Plan • Some accomplishments and challenges in knowledge discovery from biological and clinical data • Data mining in microarray analysis – diagnosis of disease state and subtype – derivation of treatment plan – understanding of gene interaction network Copyright 2003 limsoon wong Knowledge Discovery from Biological and Clinical Data: MOTIVATION Copyright 2003 limsoon wong Driving Forces: Genes, Proteins, Interactions, Diagnosis, & Cures • Complete genomes are now available • Proteins, not genes, • Proteins function by are responsible for interacting with other proteins and • Knowing the genes is many cellular activities biomolecules not enough to understand how biology functions INTERACTOME GENOME PROTEOME Copyright 2003 limsoon wong If we figure out how these work, we get these Benefits To the patient: Better drug, better treatment To the pharma: Save time, save cost, make more $ To the scientist: Better science Copyright 2003 limsoon wong To figure these out, we bet on... “solution” = Data Mgmt + Knowledge Discovery Data Mgmt = Integration + Transformation + Cleansing Knowledge Discovery = Statistics + Algorithms + Databases Copyright 2003 limsoon wong Knowledge Discovery from Biological and Clinical Data: ACCOMPLISHMENT Copyright 2003 limsoon wong 8 years of bioinformatics R&D in Singapore Integration Technology (Kleisli) MHC-Peptide Protein Interactions Extraction (PIES) Binding (PREDICT) Gene Expression Molecular Cleansing & Connections & Medical Record Warehousing Datamining (PCL) (FIMM) Gene Feature Recognition (Dragon) Venom Informatics GeneticXchange 1994 ISS 1996 1998 KRDL 2000 Biobase 2002 LIT/I2R Copyright 2003 limsoon wong Predict Epitopes, Find Vaccine Targets • Vaccines are often the only solution for viral diseases • Finding & developing effective vaccine targets is slow and expensive process • Develop systems to recognize protein peptides that bind MHC molecules • Develop systems to recognize hot spots in viral antigens Copyright 2003 limsoon wong Recognize Functional Sites, Help Scientists • Effective recognition of initiation, control, and termination of biological processes is crucial to speeding up and focusing scientific experiments • Data mining of bio seqs to find rules for recognizing & understanding functional sites Dragon’s 10x reduction of TSS recognition false positives Copyright 2003 limsoon wong Diagnose Leukaemia, Benefit Children • Childhood leukaemia is a heterogeneous disease • Treatment is based on subtype • 3 different tests and 4 different experts are needed for accurate diagnosis Curable in USA, fatal in Indonesia • A single platform diagnosis based on gene expression • Data mining to discover rules that are easy for doctors to understand Copyright 2003 limsoon wong Understand Proteins, Fight Diseases • Understanding function and role of protein needs organised info on interaction pathways • Such info are often reported in scientific paper but are seldom found in structured databases • Knowledge extraction system to process free text • extract protein names • extract interactions Copyright 2003 limsoon wong Data Mining in Microarray Analysis: MICROARRAY BACKGROUND Copyright 2003 limsoon wong What’s a Microarray? • Contain large number of DNA molecules spotted on glass slides, nylon membranes, or silicon wafers • Measure expression of thousands of genes simultaneously Copyright 2003 limsoon wong Affymetrix GeneChip Array Copyright 2003 limsoon wong Making Affymetrix GeneChip quartz is washed to ensure uniform hydroxylation across its surface and to attach linker molecules exposed linkers become deprotected and are available for nucleotide coupling Copyright 2003 limsoon wong Gene Expression Measurement by GeneChip Copyright 2003 limsoon wong A Sample Affymetrix GeneChip File (U95A) Copyright 2003 limsoon wong Data Mining in Microarray Analysis: DISEASE SUBSTYPE DIAGNOSIS Copyright 2003 limsoon wong Pediatric Acute Lymphoblastic Leukemia • A heterogeneous disease with more than 12 subtypes, e.g., T-ALL, E2A-PBX1, TELAML1, BCR-ABL, MLL, and Hyperdip>50. • Treatment response is subtype dependent • 80% continuous remission if subtype is correctly diagnosed and the corresponding treatment plan is applied Copyright 2003 limsoon wong Subtype Diagnosis • Require different tests: – immunophenotyping – cytogenetics – molecular diagnostics • Require different experts: – hematologist – oncologist – pathologist – cytogeneticist Copyright 2003 limsoon wong Difficulties and Implications • The different tests and experts are not commonly available within a single hospital, especially in less advanced countries An 80%-curable disease in USA can be a fatal disease in Indonesia! Is there a single diagnostic platform that does not need multiple human specialists? Copyright 2003 limsoon wong A Potential Solution by Microarrays Yeoh et al., Cancer Cell 1:133--143, 2002 BCR-ABL T-ALL Hyperdiploid >50 MLL Novel TEL-AML1 E2A-PBX1 Genes for class distinction (n=271) Diagnostic ALL BM samples (n=327) E2APBX1 MLL 1 0 -3 -2 -1 = std deviation from mean T-ALL 2 3 Hyperdiploid >50 BCRABL Novel TEL-AML1 Copyright 2003 limsoon wong Some Caveats • Study was performed on Americans • May not be applicable to Singaporeans, Malaysians, Indonesians, etc. • Large-scale study on local populations currently in the works Copyright 2003 limsoon wong Typical Procedure in Analysing Gene Expression for Diagnosis • • • • Gene expression data collection Gene selection Classifier training Classifier tuning (optional for some machine learning methods) • Apply classifier for diagnosis of future cases Copyright 2003 limsoon wong Feature Selection Methods A refresher of feature selection methods Copyright 2003 limsoon wong Signal Selection (Basic Idea) • Choose a signal w/ low intra-class distance • Choose a signal w/ high inter-class distance Copyright 2003 limsoon wong Signal Selection (eg., t-statistics) Copyright 2003 limsoon wong Signal Selection (eg., 2) Copyright 2003 limsoon wong Signal Selection (eg., CFS) • Instead of scoring individual signals, how about scoring a group of signals as a whole? • CFS – Correlation-based Feature Selection – A good group contains signals that are highly correlated with the class, and yet uncorrelated with each other Copyright 2003 limsoon wong Gene Expression Profile Classification An introduction to gene expression profile classification by the example on ALL subtype diagnosis Copyright 2003 limsoon wong Subtype Classification of ALL A tree-structured diagnostic workflow was recommended by the doctors, as per Yeoh et al., Cancer Cell 1:133--143, 2002 Copyright 2003 limsoon wong Training and Testing Sets Copyright 2003 limsoon wong Our procedure for ALL subtype diagnosis • • • • Gene expression data collection Gene selection by entropy Classifier training by emerging pattern Classifier tuning (optional for some machine learning methods) • Apply classifier for diagnosis of future cases by PCL Copyright 2003 limsoon wong Signal Selection (eg., entropy) Copyright 2003 limsoon wong Emerging Patterns (EPs) • An EP is a set of conditions – usually involving several features – that most members of a class satisfy – but none or few of the other class satisfy • A jumping EP is an EP that – some members of a class satisfy – but no members of the other class satisfy • We use only most general jumping EPs Copyright 2003 limsoon wong PCL: Prediction by Collective Likelihood Copyright 2003 limsoon wong Accuracy (using 20 genes of lowest entropy) PCL 0:1 0:2 1:1 0:0 4 0:1 1:1 1:1 1:6 14 5 0:1 2:2 1:1 7 0:2 5 Copyright 2003 limsoon wong Comprehensibility Copyright 2003 limsoon wong Gene Expression Profile Classification How about other feature selection and classification methods? Copyright 2003 limsoon wong Some gene selection heuristics • • • • • • all-CFS: all features from CFS top20-2: 20 features w/ highest 2 stats top20-t: 20 features w/ highest t-stats top20-mit: 20 features w/ highest MIT stats entropy: 20 features w/ lowest entropy all-2: all features meeting 5% significance level of 2 stats Copyright 2003 limsoon wong Some other classification methods • k-NN (k=1) – majority votes of the k nearest neighbours determined by Euclidean distance • C4.5 – widely used decision tree method. • Naïve Bayes (NB) – probabilistic prediction using Bayes’ rule • SVM – (linear) discriminant function that maximizes separation of boundary samples Copyright 2003 limsoon wong Accuracy • Feature selection improves performance • Entropy+PCL has consistent high performance Copyright 2003 limsoon wong When 20 genes are selected randomly Average over 100 experiments Cf. 7-15 mistakes total with good feature selection Copyright 2003 limsoon wong Data Mining in Microarray Analysis: TREATMENT PLAN DERIVATION A pure speculation! Copyright 2003 limsoon wong Can we do more with EPs? • Detect gene groups that are significantly related to a disease • Derive coordinated gene expression patterns from these groups • Derive “treatment plan” based on these patterns Copyright 2003 limsoon wong Colon Tumour Dataset Alon et al., PNAS 96:6745--6750, 1999 • We use the colon tumour dataset above to illustrate our ideas – 22 normal samples – 40 colon tumour samples Copyright 2003 limsoon wong Detect Gene Groups • Feature Selection – Use entropy method – 35 genes have cut points • Generate EPs – 19501 EPs in normals – 2165 EPs in tumours • EPs with largest support are gene groups significantly co-related to disease Copyright 2003 limsoon wong Top 20 EPs Copyright 2003 limsoon wong Observation 1 • Some EPs contain large number of genes and still have high freq • E.g., {2, 3, 6, 7, 13, 17, 33} has freq 90.91% in normal and 0% in cancer samples Nearly all normal sample’s gene expr. values satisfy all conds. implied by these 7 items Copyright 2003 limsoon wong Observation 2 • Freq of singleton EP is not necessarily larger than EP having multiple genes • E.g., {5} is EP in cancer samples and has freq 32.5% • E.g., {16, 58, 62} is EP in cancer samples and has freq 75.5% Groups of genes and their correlation's could be more impt than single genes Copyright 2003 limsoon wong Observation 3 • M33680 has lowest entropy of the 35 genes if cutpoint is set at 352 • 18/40 of cancer samples shift expr level of M33680 from its normal range to its abnormal range Copyright 2003 limsoon wong Treatment Plan Idea • Increase/decrease expression level of particular genes in a cancer cell so that – it has the common EPs of normal cells – it has no common EPs of cancer cells Copyright 2003 limsoon wong Treatment Plan Example • From the EP {2,3,6,7,13,17,33} – 91% of normal cells express the 7 genes (T51560, T49941, M62994, R34701, L02426, U20428, R10707) in the corr. Intervals – a cancer cell never express all 7 genes in the same way – if expression level of improperly expressed genes can be adjusted, the cancer cell can have one common EP of normal cells – a cancer cell can then be iteratively converted into a normal one Copyright 2003 limsoon wong Choosing Genes to Adjust Copyright 2003 limsoon wong Doing more adjustments... • Down regulating T49941 leads to 2 more top 10 EPs of normal cells to show up in the adjusted T1 • Down regulating X62153 to below 396 and T72403 to below 296 leads to T1 having 9 top 10 EPs of normal cells • Ave. no. of EPs in normal cells is 9 • So the adjusted T1 now has impt features of normal cells Copyright 2003 limsoon wong Next, eliminate common EPs of cancer cells in T1 • 6 more genes (K03001, T49732, U29171, R76254, D31767, L40992) are adjusted • All top 10 EPs of cancer cells now disappear from T1 • Ave. no. of top 10 EPs contained in cancer cells is 6 • The adjusted T1 now holds enough common features of normal cells and no features of cancer cells T1 is converted to normal cellsCopyright 2003 limsoon wong “Treatment Plan” Validation • “Adjustments” were made to the 40 colon tumour samples based on EPs as described • Classifiers trained on original samples were applied to the adjusted samples It works! Copyright 2003 limsoon wong A Big But... • Effective means for identifying mechanisms and pathways through which to modulate gene expression of selected genes need to be developed Copyright 2003 limsoon wong Data Mining in Microarray Analysis: GENE INTERACTION PREDICTION Copyright 2003 limsoon wong Beyond Classification of Gene Expression Profiles • After identifying the candidate genes by feature selection, do we know which ones are causal genes and which ones are surrogates? Genes for class distinction (n=271) Diagnostic ALL BM samples (n=327) E2APBX1 MLL 1 0 -3 -2 -1 = std deviation from mean T-ALL 2 3 Hyperdiploid >50 BCRABL Novel TEL-AML1 Copyright 2003 limsoon wong Gene Regulatory Circuits • Genes are “connected” in “circuit” or network • Expression of a gene in a network depends on expression of some other genes in the network • Can we reconstruct the gene network from gene expression data? Copyright 2003 limsoon wong Key Questions For each gene in the network: • which genes affect it? • How they affect it? – Positively? – Negatively? – More complicated ways? Copyright 2003 limsoon wong Some Techniques • Bayesian Networks – Friedman et al., JCB 7:601--620, 2000 • Boolean Networks – Akutsu et al., PSB 2000, pages 293--304 • Differential equations – Chen et al., PSB 1999, pages 29--40 • Classification-based method – Soinov et al., “Towards reconstruction of gene network from expression data by supervised learning”, Genome Biology 4:R6.1--9, 2003 Copyright 2003 limsoon wong A Classification-based Technique Soinov et al., Genome Biology 4:R6.1-9, Jan 2003 • Given a gene expression matrix X – each row is a gene – each column is a sample – each element xij is expression of gene i in sample j • Find the average value ai of each gene i • Denote sij as state of gene i in sample j, – sij = up if xij > ai – sij = down if xij ai Copyright 2003 limsoon wong A Classification-based Technique Soinov et al., Genome Biology 4:R6.1-9, Jan 2003 • To see whether the state of gene g is determined by the state of other genes – we see whether sij | i g can predict sgj – if can predict with high accuracy, then “yes” – Any classifier can be used, such as C4.5, PCL, SVM, etc. • To see how the state of gene g is determined by the state of other genes – apply C4.5 (or PCL or other “rule-based” classifiers) to predict sgj from sij | i g – and extract the decision tree or rules used Copyright 2003 limsoon wong Advantages of this method • Can identify genes affecting a target gene • Don’t need discretization thresholds • Each data sample is treated as an example • Explicit rules can be extracted from the classifier (assuming C4.5 or PCL) • Generalizable to time series Copyright 2003 limsoon wong Acknowledgements Vladimir Brusic See-Kiong Ng Jinyan Li Vladimir Bajic Huiqing Liu Copyright 2003 limsoon wong Data Mining in Microarray Analysis: NOTES Copyright 2003 limsoon wong References • J.Li, L. Wong, “Geography of differences between two classes of data”, Proc. 6th European Conf. on Principles of Data Mining and Knowledge Discovery, pp. 325--337, 2002 • J.Li, L. Wong, “Identifying good diagnostic genes or gene groups from gene expression data by using the concept of emerging patterns”, Bioinformatics, 18:725--734, 2002 • J.Li et al., “A comparative study on feature selection and classification methods using a large set of gene expression profiles”, GIW, 13:51--60, 2002 Copyright 2003 limsoon wong References • E.-J. Yeoh et al., “Classification, subtype discovery, and prediction of outcome in pediatric acute lymphoblastic leukemia by gene expression profiling”, Cancer Cell, 1:133--143, 2002 • U.Alon et al., “Broad patterns of gene expression revealed by clustering analysis of tumor colon tissues probed by oligonucleotide arrays”, PNAS 96:6745--6750, 1999 • L.A.Soinov et al., “Towards reconstruction of gene networks from expression data by supervised learning”, Genome Biology 4:R6.1--9, 2003. Copyright 2003 limsoon wong > > > > > > > > > > > Data Mining of Gene Expression Profiles for the Diagnosis and Understanding of Diseases This talk is divided into two parts. In Part I, I will provide a brief overview of some accomplishments and challenges in Bioinformatics. In Part II, I will discuss the data mining in the analysis of microarray gene expression profiles for (a) diagnosis of disease state or subtype, (b) derivation of disease treatment plan, and (c) understanding of gene interaction networks. Copyright 2003 limsoon wong