Performance of Blood-Based Biomarkers for Human Circadian Pacemaker Phase: Training Sets Matter As Much As Feature-Selection Methods

Biomarkers are valuable tools in a wide range of human health areas including circadian medicine. Valid, low-burden, multivariate molecular approaches to assess circadian phase at scale in people living and working in the real world hold promise for translating basic circadian knowledge to practical applications. However, standards for the development and evaluation of these circadian biomarkers have not yet been established, even though several publications report such biomarkers and claim that the methods are universal. Here, we present a basic exploration of some of the determinants and confounds of blood-based biomarker development for suprachiasmatic nucleus (SCN) phase by reanalysing publicly available data sets. We compare performance of biomarkers based on three feature-selection methods: Partial Least Squares Regression, ZeitZeiger, and Elastic Net, as well as performance of a standard set of clock genes. We explore the effects of training sample size and the impact of the experimental protocols from which training samples are drawn and on which performance is tested. Approaches based on small sample sizes used for training are prone to poor performance due to overfitting. Performance to some extent depends on the feature-selection method, but at least as much on the experimental conditions from which the biomarker training samples were drawn. Performance of biomarkers developed under baseline conditions does not necessarily translate to protocols that mimic real-world scenarios such as shiftwork in which sleep may be restricted or desynchronized from the endogenous circadian SCN phase. The molecular features selected by the various approaches to develop biomarkers for the SCN phase show very little overlap although the processes associated with these features have common themes with response to steroid hormones, that is, cortisol being the most prominent. Overall, the findings indicate that establishment of circadian biomarkers should be guided by established biomarker-development concepts and foundational principles of human circadian biology.