Data Mining in Genomics: the dawn of personalized medicine Gregory Piatetsky-Shapiro KDnuggets www.KDnuggets.com/gps.html Connecticut College, October 15, 2003
Overview Data Mining and Knowledge Discovery Genomics and Microarrays Microarray Data Mining
Trends leading to Data Flood More data is generated: Bank, telecom, other business transactions ... Scientific Data: astronomy, biology, etc Web, text, and e-commerce More data is captured: Storage technology faster and cheaper DBMS capable of handling bigger DB
Knowledge Discovery Process Integration Interpretation & Evaluation Knowledge Data Mining Patterns and Rules Knowledge RawData __ ____ Transformation Selection & Cleaning Understanding Transformed Data DATA Ware house Target Data
Major Data Mining Tasks Classification: predicting an item class Clustering: finding clusters in data Associations: e.g. A & B & C occur frequently Visualization: to facilitate human discovery Summarization: describing a group Estimation: predicting a continuous value Deviation Detection: finding changes Link Analysis: finding relationships
Major Application Areas for Data Mining Solutions Advertising Bioinformatics Customer Relationship Management (CRM) Database Marketing Fraud Detection eCommerce Health Care Investment/Securities Manufacturing, Process Control Sports and Entertainment Telecommunications Web
Genome, DNA & Gene Expression An organism’s genome is the “program” for making the organism, encoded in DNA Human DNA has about 30-35,000 genes A gene is a segment of DNA that specifies how to make a protein Cells are different because of differential gene expression About 40% of human genes are expressed at one time Microarray devices measure gene expression
Molecular Biology Overview Nucleus Cell Chromosome Gene expression Gene (DNA) Protein Gene (mRNA), single strand Graphics courtesy of the National Human Genome Research Institute
Affymetrix Microarrays 1.28cm 50um ~107 oligonucleotides, half Perfectly Match mRNA (PM), half have one Mismatch (MM) Gene expression computed from PM and MM
Affymetrix Microarray Raw Image Gene Value D26528_at 193 D26561_cds1_at -70 D26561_cds2_at 144 D26561_cds3_at 33 D26579_at 318 D26598_at 1764 D26599_at 1537 D26600_at 1204 D28114_at 707 Scanner raw data enlarged section of raw image
Microarray Potential Applications New and better molecular diagnostics New molecular targets for therapy few new drugs, large pipeline, … Outcome depends on genetic signature best treatment? Fundamental Biological Discovery finding and refining biological pathways Personalized medicine ?!
Microarray Data Mining Challenges Avoiding false positives, due to too few records (samples), usually < 100 too many columns (genes), usually > 1,000 Model needs to be robust in presence of noise For reliability need large gene sets; for diagnostics or drug targets, need small gene sets Estimate class probability Model needs to be explainable to biologists
False Positives in Astronomy cartoon used with permission
CATs: Clementine Application Templates CATs - examples of complete data mining processes Microarray CAT Preparation Multi- Class Clustering 2-Class
Key Ideas Capture the complete process X-validation loop w. feature selection inside Randomization to select significant genes Internal iterative feature selection loop For each class, separate selection of optimal gene sets Neural nets – robust in presence of noise Bagging of neural nets
Microarray Classification Train data Feature and Parameter Selection Data Model Building Evaluation Test data
Classification: External X-val Gene Data Train data Feature and Parameter Selection T r a i n Data Model Building Evaluation Test data FinalTest Final Model Final Results
Measuring false positives with randomization Class Gene Class 178 105 4174 7133 1 2 2 1 Randomize 500 times Gene Class Bottom 1% T-value = -2.08 Select potentially interesting genes at 1% 178 105 4174 7133 2 1
Gene Reduction improves Classification most learning algorithms look for non-linear combinations of features -- can easily find many spurious combinations given small # of records and large # of genes Classification accuracy improves if we first reduce # of genes by a linear method, e.g. T-values of mean difference Heuristic: select equal # genes from each class Then apply a favorite machine learning algorithm
Iterative Wrapper approach to selecting the best gene set Test models using 1,2,3, …, 10, 20, 30, 40, ..., 100 top genes with x-validation. Heuristic 1: evaluate errors from each class; select # number of genes from each class that minimizes error for that class For randomized algorithms, average 10+ Cross-validation runs! Select gene set with lowest average error
Clementine stream for subset selection by x-validation
Microarrays: ALL/AML Example Leukemia: Acute Lymphoblastic (ALL) vs Acute Myeloid (AML), Golub et al, Science, v.286, 1999 72 examples (38 train, 34 test), about 7,000 genes well-studied (CAMDA-2000), good test example ALL AML Visually similar, but genetically very different
Gene subset selection: one X-validation Single Cross-Validation run
Gene subset selection: multiple cross-validation runs For ALL/AML data, 10 genes per class had the lowest error: (<1%) Point in the center is the average error from 10 cross-validation runs Bars indicate 1 st. dev above and below
ALL/AML: Results on the test data Genes selected and model trained on Train set ONLY! Best Net with 10 top genes per class (20 overall) was applied to the test data (34 samples): 33 correct predictions (97% accuracy), 1 error on sample 66 Actual Class AML, Net prediction: ALL other methods consistently misclassify sample 66 -- misclassified by a pathologist?
Pediatric Brain Tumour Data 92 samples, 5 classes (MED, EPD, JPA, EPD, MGL, RHB) from U. of Chicago Children’s Hospital Outer cross-validation with gene selection inside the loop Ranking by absolute T-test value (selects top positive and negative genes) Select best genes by adjusted error for each class Bagging of 100 neural nets
Selecting Best Gene Set Minimizing Combined Error for all classes is not optimal Average, high and low error rate for all classes
Error rates for each class Genes per Class
Evaluating One Network Averaged over 100 Networks: Class Error rate MED 2.1% MGL 17% RHB 24% EPD 9% JPA 19% *ALL* 8.3%
Bagging 100 Networks Class Individual Error Rate Bag Error rate Bag Avg Conf MED 2.1% 2% (0)* 98% MGL 17% 10% 83% RHB 24% 11% 76% EPD 9% 91% JPA 19% 81% *ALL* 8.3% 3% (2)* 92% Note: suspected error on one sample (labeled as MED but consistently classified as RHB)
AF1q: New Marker for Medulloblastoma? AF1Q ALL1-fused gene from chromosome 1q transmembrane protein Related to leukemia (3 PUBMED entries) but not to Medulloblastoma
Future directions for Microarray Analysis Algorithms optimized for small samples Integration with other data biological networks medical text protein data Cost-sensitive classification algorithms error cost depends on outcome (don’t want to miss treatable cancer), treatment side effects, etc.
Acknowledgements Eric Bremer, Children’s Hospital (Chicago) & Northwestern U. Greg Cooper, U. Pittsburgh Tom Khabaza, SPSS Sridhar Ramaswamy, MIT/Whitehead Institute Pablo Tamayo, MIT/Whitehead Institute
Thank you Further resources on Data Mining: www.KDnuggets.com Microarrays: www.KDnuggets.com/websites/microarray.html Contact: Gregory Piatetsky-Shapiro: www.kdnuggets.com/gps.html