Improved Statistical Methods Enable Greater Sensitivity in Rhythm Detection for Genome-Wide Data

Robust methods for identifying patterns of expression in genome-wide data are important for generating hypotheses regarding gene function. To this end, several analytic methods have been developed for detecting periodic patterns. We improve one such method, JTK_CYCLE, by explicitly calculating the null distribution such that it accounts for multiple hypothesis testing and by including non-sinusoidal reference waveforms. We term this method empirical JTK_CYCLE with asymmetry search, and we compare its performance to JTK_CYCLE with Bonferroni and Benjamini-Hochberg multiple hypothesis testing correction, as well as to five other methods: cyclohedron test, address reduction, stable persistence, ANOVA, and F24. We find that ANOVA, F24, and JTK_CYCLE consistently outperform the other three methods when data are limited and noisy; empirical JTK_CYCLE with asymmetry search gives the greatest sensitivity while controlling for the false discovery rate. Our analysis also provides insight into experimental design and we find that, for a fixed number of samples, better sensitivity and specificity are achieved with higher numbers of replicates than with higher sampling density. Application of the methods to detecting circadian rhythms in a metadataset of microarrays that quantify time-dependent gene expression in whole heads of Drosophila melanogaster reveals annotations that are enriched among genes with highly asymmetric waveforms. These include a wide range of oxidation reduction and metabolic genes, as well as genes with transcripts that have multiple splice forms.Author Summary: Much biomedical research focuses on how the expression of genes changes over time. Many genes’ activities vary periodically. For example, circadian rhythms repeat daily with the light-dark cycle. Understanding how such rhythms couple to biological processes requires statistical methods that can identify cycling time series in typical genome-wide data. In this paper, we improve on a method used to identify cycling time series by better estimating the statistical significance of periodic patterns and, in turn, by searching for a wider range of patterns than traditionally investigated. We apply these methods to a compilation of data on gene expression in fruit flies, an important model organism. We find that our method allows us to discover rhythmic biological activities that the other methods tested are unable to reveal.