Finding associations in a heterogeneous setting: Statistical test for aberration enrichment

Abstract

Most two-group statistical tests are implicitly looking for a broad pattern such as an overall shift in mean, median or variance between the two groups. Therefore, they operate best in settings where the effect of interest is uniformly affecting everyone in one group versus the other. In real-world applications, there are many scenarios where the effect of interest is heterogeneous. For example, a drug that works very well on only a proportion of patients and is equivalent to a placebo on the remaining patients, or a disease associated gene expression dysregulation that only occurs in a proportion of cases whereas the remaining cases have expression levels indistinguishable from the controls for the considered gene. In these examples with heterogeneous effect, we believe that using classical two-group statistical tests may not be the most powerful way to detect the signal. In this paper, we developed a statistical test targeting heterogeneous effects and demonstrated its power in a controlled simulation setting compared to existing methods. We focused on the problem of finding meaningful associations in complex genetic diseases using omics data such as gene expression, miRNA expression, and DNA methylation. In simulated and real data, we showed that our test is complementary to the traditionally used statistical tests and is able to detect disease-relevant genes with heterogeneous effects which would not be detectable with previous approaches.