On the Accuracy of Internal Circadian Time Prediction Methods from a Single Sample

Abstract Biological processes ranging from gene expression to sleep-wake cycles display oscillations with an approximately 24-hour period, or circadian rhythms. A challenge in analyzing circadian rhythms is that these rhythms vary across individuals and are based on an individual’s internal circadian time (ICT), which is uniquely offset relative to the 24-hour day-night cycle time (zeitgeber time, or ZT). Many model-based methods have been proposed to predict ICT given an individual’s biomarker measurements. However, the prediction accuracy of these methods is rarely validated using known ICT. In this article, we evaluate this accuracy for three state-of-the-art model-based methods: COFE, partial least squares regression, and TimeSignature. We find that if a single sample is obtained from each individual and a model is fit using only biomarker measurements as predictors, then ZT predicts ICT more accurately than any of the model-based ICT predictions. However, we also find that TimeSignature can outperform ZT when the model incorporates sine and cosine transforms of sample collection ZT as two additional predictors. These findings are based on analysis of three circadian transcriptome datasets as well as simulation studies, and highlight the importance of accounting for individual-level differences in biomarker oscillations to improve ICT prediction. Competing Interest Statement The authors have declared no competing interest.