Combined statistical modeling enables accurate mining of circadian transcription

Abstract

Circadian-regulated genes are essential for tissue homeostasis and organismal function, and are therefore common targets of scrutiny. Detection of rhythmic genes using current analytical tools requires exhaustive sampling, a demand that is costly and raises ethical concerns, making it unfeasible in certain mammalian systems. Several non-parametric methods have been commonly used to analyze short-term (24 h) circadian data, such as JTK_cycle and MetaCycle. However, algorithm performance varies greatly depending on various biological and technical factors. Here, we present CircaN, an <i>ad-hoc</i> implementation of a non-linear mixed model for the identification of circadian genes in all types of omics data. Based on the variable but complementary results obtained through several biological and <i>in silico</i> datasets, we propose a combined approach of CircaN and non-parametric models to dramatically improve the number of circadian genes detected, without affecting accuracy. We also introduce an R package to make this approach available to the community.