Hypothesis
Cancer-trained predictors and machine-learning tools with comprehensive feature spaces should separate cancer-driving missense mutations from passengers better than simple conservation-based methods.
What failed
Tested with high-frequency polymorphisms as neutrals and recurrent COSMIC and newly reported somatic mutations as positives, no method or combination of methods exceeded 81% accuracy; machine-learning predictors swung with the choice of positive set, cancer-trained methods did not improve on general predictors, and metapredictors did not beat MutationAssessor used alone.
Broken assumption
Training a predictor on cancer mutations makes it better at identifying drivers.
Why
The authors conclude that more sophisticated feature integration is still needed, and report performance disparities including differences in coverage and tool interdependence.
From the published paper: Gnad et al. BMC Genomics, 2013. ↗CC BY