A preprint compared 10 algorithms for ordering samples within the daily cycle: two models based on clock genes ranked behind only CHIRAL

A preprint compared 10 algorithms for ordering samples within the daily cycle: two models based on clock genes ranked behind only CHIRAL

A preprint published on October 15 tested whether gene expression could reveal a sample’s position within the 24-hour cycle when its collection time was unknown. Across 25 datasets, CHIRAL had a mean error of 1,80 hours, while two models based on clock genes had mean errors of 2,15 and 2,20 hours.

Transcriptomics measures the expression of thousands of genes in a tissue sample. Public datasets often lack sample collection times, even though the expression of some genes varies over the course of the day. Circadian phase describes a sample’s position within this 24-hour cycle and can be used to place samples in relative temporal order.

In this preprint, the authors compared ten algorithms that infer this order from gene expression tables without timestamps against two simple models. One model used the expression of 12 clock genes, whose expression varies rhythmically throughout the day; the other used just two. This comparison tested whether a complex model improved accuracy beyond what could be achieved using the known circadian signal.

The main evaluation included 25 datasets: 16 real datasets from five studies and nine simulations. Each algorithm could choose its own starting point and direction around the cycle, so the authors aligned these scales before calculating the error. After alignment, the metric measured how accurately each algorithm recovered the relative order of samples and the intervals between their phases.

CHIRAL had the lowest mean error at 1,80 hours. The two simple models had mean errors of 2,15 and 2,20 hours, ranking behind only CHIRAL. A statistical test confirmed that CHIRAL outperformed COFE, PENN and PROTECT; the ranges of mean error estimates for CHIRAL, TEMPO, CYCLOPS and CYCLOPS 2.0 overlapped.

CHIRAL also uses 12 clock genes, but jointly fits the samples’ positions within the cycle and the oscillation curves of these genes. In the GTEx project, a large archive of samples from many human tissues, the method first estimated sample phases within each tissue, then used those estimates to infer an overall phase for each donor. The new benchmark compared this approach with competing algorithms across heterogeneous archival datasets.

Data preprocessing also affects the results. For ten datasets with documented differences between groups in age, disease or measurement batch, the authors applied ComBat-seq, a method for adjusting data to account for such differences. In these datasets, the mean error for CYCLOPS fell from 4,00 to 2,02 hours, remained almost unchanged for CHIRAL at 1,83 versus 1,84 hours, and rose from 2,68 to 2,96 hours for TEMPO. The authors recommend starting with CHIRAL when working with archival transcriptomic datasets and separately assessing how circadian signal strength and data processing affect the results.