Transferring Computer Science Research to Mining DNA chip Protein Expression

Introduction

• DNA chips estimate concentration of individual gene products. A chip can measure 1000s of gene expression levels simultaneously

• Tissue samples (patients) are in short supply and chips are expensive. => Few training records with many attributes. => Over fitting.

• Genetic programming used for feature selection

• 3 Stage GP. 10,000 genes -> 100s -> 10s

• Last stage GP used to produce predictive model

GSK 578 examples 4548 -> 361 ->18 (=>6) genes

• 100 GP runs with genes
  • data split 50:50 train:verification
  • fitness = 1/2 true positive rate + 1/2 true neg (ROC)
  • binary arithmetic operators (+ - * /), 4548 constants
  • five trees per individual
  • pop 500, 50 generations
  • 50% size fair crossover, 50% 1 of 4 mutations

• Extract all genes used in best of run model

• 361 appeared more than once (1730 used)

18=>6 gene model

• 100 GP runs with 18 genes

• Over fitting.
  • Select model from generation 10
  • Highest fitness and smallest model, with least genes
  • run 40, 6 genes, size=23

• IF((gene1 + gene2 + gene3 -13.8 -gene4 (0.00918 * gene5 * gene1 -2230)/gene6) >= 0) THEN predict positive
  • 71% of positive examples correct and 81% of the negatives correct

Training Set vs. Test set fitness

ALL Leukemia 119 Patients 12625 ->2 139 ->17 (=>9) genes

• 100 GP runs with 12625 genes
  • no data split
  • pop 10,000 (all genes in gen 0), 5 generations

• Extract all genes used in best of run model

• 139 appeared more than once (1737 used)

Leukemia 2nd stage 139->17 genes

Leukemia GP linear 9 genes

• Leave one out 10 GP runs each with 17 genes 1180 runs.

• linear model with 9 (of 12,625) inputs

• IF((40371_at/31+ 34985_at + 1828_s_at + 451_at + 41419_at -34052_at -39535_at - 36730_at -32483_at) >= 0) THEN Predict positive

• Leave one out estimate
  • 76% of positive examples correct and 71% of the negatives correct

60 Childhood Cancers 7129 -> 404 => 2 genes

• 600 GP runs with 7129 genes
  • ten leave one out (10*60=600)
  • pop 500, 50 generations

• Extract all genes used in best of run model

• 404 appeared ten or more times (6970 used)

Cancer 404-> 2 genes

Cancer final 2 gene model

Summary

• Many variables few examples (578, 119, 60). Fear over fitting.

• Simple models, few (6, 9, 2) genes.

• Leave one out cross validation or holdout set

• Genetic programming used to select genes

• Fitness area under ROC. Cf class imbalance

• Population sized to get all variables n(log n+.6)

• Few generations to reduce over fitting

• Size fair crossover and mutation eliminate bloat

• ftp://cs.ucl.ac.uk/genetic/gp-code/GProc-1.7.tar.gz