Abstract
The study of ubiquitous circadian rhythms in human physiology requires regular measurements across time. Repeated sampling of the different internal tissues that house circadian clocks is both practically and ethically infeasible. Here, we present a novel unsupervised machine learning approach (COFE) that can use single high-throughput omics samples (without time labels) from individuals to reconstruct circadian rhythms across cohorts. COFE can simultaneously assign time labels to samples and identify rhythmic data features used for temporal reconstruction, while also detecting invalid orderings. With COFE, we discovered widespread de novo circadian gene expression rhythms in 11 different human adenocarcinomas using data from The Cancer Genome Atlas (TCGA) database. The arrangement of peak times of core clock gene expression was conserved across cancers and resembled a healthy functional clock except for the mistiming of a few key genes. Moreover, rhythms in the transcriptome were strongly associated with the cancer-relevant proteome. The rhythmic genes and proteins common to all cancers were involved in metabolism and the cell cycle. Although these rhythms were synchronized with the cell cycle in many cancers, they were uncoupled with clocks in healthy matched tissue. The targets of most of FDA-approved and potential anti-cancer drugs were rhythmic in tumor tissue with different amplitudes and peak times. These findings emphasize the utility of considering "time" in cancer therapy, and suggest a focus on clocks in healthy tissue rather than free-running clocks in cancer tissue. Our approach thus creates new opportunities to repurpose data without time labels to study circadian rhythms.