A novel Bayesian hierarchical model for detecting differential circadian pattern in transcriptomic applications.

Circadian rhythm plays a critical role in regulating various physiological processes, and disruptions in these rhythms have been linked to a wide range of diseases. Identifying molecular biomarkers showing differential circadian (DC) patterns between biological conditions or disease status is important for disease prevention, diagnosis, and treatment. However, circadian pattern is characterized by three key components: amplitude, phase, and MESOR, which poses a great challenge for DC analysis. Existing statistical methods focus on detecting differential shape (amplitude and phase) but often overlook MESOR difference. Additionally, these methods lack flexibility to incorporate external knowledge such as differential circadian information from similar clinical and biological context to improve the current DC analysis. To address these limitation, we introduce a novel Bayesian hierarchical model, BayesDCirc, designed for detecting differential circadian patterns in a two-group experimental design, which offer the advantage of testing MESOR difference and incorporating external knowledge. Benefiting from explicitly testing MESOR within the Bayesian modeling framework, BayesDCirc demonstrates superior FDR control over existing methods, with further performance improvement by leveraging external knowledge of DC genes. Applied to two real datasets, BayesDCirc successfully identify key circadian genes, particularly with external knowledge incorporated. The R package "BayesDCirc" for the method is publicly available on GitHub at https://github.com/lichen-lab/BayesDCirc.