lung cancer is highly heterogenous in classifications and outcomes -- how to separate those patients that would and would not benefit from chemotherapy after surgery - Andrew Su

previous studies based on gene expression analysis have not replicated. why? weak statistical methods - Andrew Su

method: "modified steepest descent" - Andrew Su

first assess genes individually, then take best gene and evaluate best pair, take top pair and screen for top triple, etc. - Andrew Su

identified 6 -gene marker set -- stx1a, hif1a, cct3, mafk, rnf5, hla_dpb1 - Andrew Su

classier in training data, survival difference 1E-5 (not surprising) - Andrew Su

8 separate studies: p = 0.02 (409 patients - Andrew Su

evaulation: take 10 million random six-gene sets, better than 99.9999% of backgorund, drops to ~90% on other training data, merged 99.98% - Andrew Su

but, what are the ~450,000 random sets that are better? Could this explain non-overlap of previous microarray studies? - Andrew Su

evaulate gene based on marker pluraily -- how many good predictors does a gene occur in? - Andrew Su

e.g., calca -- not a good predictor on its own, but a part of many prognostic gene sets - Andrew Su

top 10: calca, ccr7, stx1a, cct3, sprr1b, selp, pafa... cpe, ... - Andrew Su

pathway analysis: reinforces previous network by Jurisica et al. plurality genes in particular help to reinforce links - Andrew Su