Differential distribution improves gene selection stability and has competitive classification performance for patient survival.

A consistent difference in average expression level, often referred to as differential expression (DE), has long been used to identify genes useful for classification. However, recent cancer studies have shown that when transcription factors or epigenetic signals become deregulated, a change in expression variability (DV) of target genes is frequently observed. This suggests that assessing the importance of genes by either differential expression or variability alone potentially misses sets of important biomarkers that could lead to improved predictions and treatments. Here, we describe a new approach for assessing the importance of genes based on differential distribution (DD), which combines information from differential expression and differential variability into a unified metric. We show that feature ranking and selection stability based on DD can perform two to three times better than DE or DV alone, and that DD yields equivalent error rates to DE and DV. Finally, assessing genes via differential distribution produces a complementary set of selected genes to DE and DV, potentially opening up new categories of biomarkers.