Characterizing internal circadian time using omics data and computational tools

Many computational tools have been developed to infer circadian phase from gene expression data (microarray/bulk RNA-seq). These methods provide circadian pattern metrics (i.e., period, amplitude, and phase) quantified using statistical approaches. Period refers to the time between peaks in expression (tau), amplitude is the difference between the highest and the lowest point over a cycle (sometimes divided by 2), and phase shift refers to a horizontal shift in the peaks of expression from a reference while maintaining the same period difference between peaks. Since many genes show rhythmic patterns (but with peaks at different times), rhythmic parameters can be inferred through their joint analysis. Generating predictors that can be applied to single timepoint data is an active area of research trained on data sets that have longitudinal sampling within individuals (not possible for many tissues in human), or on large cross-sectional data sets where tissue from different individuals is sampled across the 24-hour period. Data training methods are classified as supervised or unsupervised.

Supervised learning methods (e.g., molecular time-table (Ueda et al. Reference Ueda, Chen, Minami, Honma, Honma, Iino, Hashimoto and Waterman2004), ZeitZeiger (Hughey et al. Reference Hughey, Hastie and Butte2016), BIO_CLOCK (Agostinelli et al. Reference Agostinelli, Ceglia, Shahbaba, Sassone-Corsi and Baldi2016)) use ground-truth data sets with known time of sampling to generate prediction models tested in other independent ground-truth data sets and then applied in data sets where sampling time was unknown. These methods identify a core set of “time-indicating genes,” (e.g., 13 genes in ZeitZeiger) and train statistical models on the expression of these core genes. A limitation of supervised learning methods is the need for ground-truth samples since few data sets have samples documented with a collection “timestamp.” Unsupervised machine learning methods (e.g., CYCLOPS (Anafi et al. Reference Anafi, Francey, Hogenesch and Kim2017), CHIRAL (Talamanca et al. Reference Talamanca, Gobet and Naef2023)) use underlying modeling and joint analysis across multiple genes to infer circadian phase i.e., since many genes are rhythmic in their expression in a coordinated manner, the expression levels of many genes infer circadian phase more accurately than can be inferred from expression of a single gene. For example, CHIRAL uses a new mathematical method to infer circadian time applied to the Genotype-Tissue Expression (GTEx) data, which allows a comprehensive analysis of rhythmic patterns on the whole organism (since GTEx data comprise gene expression data from multiple organs from the same post-mortem donors). Briefly, the algorithm first assigns tissue internal phase (TIP) for each sample in each tissue type with a selected set of seed genes. Donor internal phase is estimated from TIPs, which assumed that each TIP is determined by the donor and the tissue. With this approach, differences across donors and tissues can be captured. Technological advances that allow generation of large data sets that can be interrogated for circadian phase, have been accompanied by an explosion in computational methods (Table 1). With adequate variation in sampling times across individuals, methods such as CYCLOPS and CHIRAL can achieve a high accuracy with a mean absolute error between 1∼2 h (Talamanca et al. Reference Talamanca, Gobet and Naef2023). When new methods are presented, comparisons with standard methods are provided (and show improvements in accuracy of assignments in ground-truth data sets or computational efficiency), but independent systematic comparisons of methods across multiple data sets are now needed. Hughes et al. (Reference Hughes, Abruzzi, Allada, Anafi, Arpat, Asher, Baldi, de Bekker, Bell-Pedersen, Blau, Brown, Ceriani, Chen, Chiu, Cox, Crowell, DeBruyne, Dijk, DiTacchio, Doyle, Duffield, Dunlap, Eckel-Mahan, Esser, FitzGerald, Forger, Francey, Fu, Gachon, Gatfield, de Goede, Golden, Green, Harer, Harmer, Haspel, Hastings, Herzel, Herzog, Hoffmann, Hong, Hughey, Hurley, de la Iglesia, Johnson, Kay, Koike, Kornacker, Kramer, Lamia, Leise, Lewis, Li, Li, Liu, Loros, Martino, Menet, Merrow, Millar, Mockler, Naef, Nagoshi, Nitabach, Olmedo, Nusinow, Ptacek, Rand, Reddy, Robles, Roenneberg, Rosbash, Ruben, Rund, Sancar, Sassone-Corsi, Sehgal, Sherrill-Mix, Skene, Storch, Takahashi, Ueda, Wang, Weitz, Westermark, Wijnen, Xu, Wu, Yoo, Young, Zhang, Zielinski and Hogenesch2017) provide “Guidelines for genome-scale analysis of biological rhythms” and a web-based application (CircaInSilico) to generate ground-truth synthetic genome biology data to facilitate benchmarking of methods.

Table 1. Computational methods to assign circadian rhythm parameters to samples based on omics data

^a Supervised (uses ground-truth data to train the model)/Unsupervised.

Both supervised and unsupervised methods need samples collected across the 24-hour period, with better prediction accuracy achieved if each time point has more samples. While such time series sample collections are achieved in sleep laboratory studies or in post-mortem (i.e., across samples the full 24-hour period is represented), samples collected in clinics or volunteer studies are likely to represent only a fraction of the 24-hour period. A new method, has been developed to predict circadian time from bulk RNA-seq without needing to collect samples in a complete time series.

Beyond using bulk tissue gene expression data to study circadian biology, chronobiology at the single-cell level will give much deeper insights into potential differences in circadian pattern across different cell types. Many computational tools have been developed to infer cell-cycle state, which is a periodic biological process (Leng et al. Reference Leng, Chu, Barry, Li, Choi, Li, Jiang, Stewart, Thomson and Kendziorski2015; Liu et al. Reference Liu, Lou, Xie, Wang, Chen, Aparicio, Zhang, Jiang and Chen2017; Riba et al. Reference Riba, Oravecz, Durik, Jiménez, Alunni, Cerciat, Jung, Keime, Keyes and Molina2022). However, circadian rhythms and the cell cycle are independent processes. Characterizing circadian rhythm using single-cell data has been rarely documented but is likely to be an area of active research as more and more single-cell gene expression data sets are generated. One published method has used scRNA-seq data applying an unsupervised Bayesian algorithm developed to estimate circadian phase with a prior knowledge of core clock genes to initialize the model (Auerbach et al. Reference Auerbach, FitzGerald and Li2022). The model uses the posterior distribution of cell circadian phase to identify new rhythmic genes. This offers an opportunity to identify cell-type specific circadian patterns and ultimately cell therapies for chronobiological disorders. Disruption of circadian clocks in tumor vs non-tumor tissue in adenocarcinomas has already been demonstrated (Ananthasubramaniam & Venkataramanan Reference Ananthasubramaniam and Venkataramanan2024).

Although gene expression data are currently the most abundant, computational methods have been developed or adapted for applications to proteomic (Weger et al. Reference Weger, Gobet, David, Atger, Martin, Phillips, Charpagne, Weger, Naef and Gachon2021; Specht et al. Reference Specht, Kolosov, Cederberg, Bueno, Arrona-Palacios, Pardilla-Delgado, Ruiz-Herrera, Zitting, Kramer, Zeitzer, Czeisler, Duffy and Mignot2023) and metabolomics (Minami et al. Reference Minami, Kasukawa, Kakazu, Iigo, Sugimoto, Ikeda, Yasui, van der Horst, Soga and Ueda2009) data. Understanding circadian rhythms in genes, proteins and metabolites targeted directly or indirectly by drugs underpins the emerging field of chronopharmacology (Li et al. Reference Li, Lu and Androulakis2024).

Potential for application of computational methods to infer sleep and circadian disruption in depression

Methods to infer circadian phase from a single routinely collected blood sample (the easiest tissue to access) measured for gene expression offer the possibility of cost-effective investigation of circadian parameters in people with severe mental illness compared to those without. These scalable technologies will allow generation of large data sets which are needed to draw robust conclusions. To take advantage of these new methods, key data sets need to be generated. First, post-mortem brain samples are needed to demonstrate circadian disruption in a cell type relevant to depression (e.g., excitatory neurons, inhibitory neurons, astrocytes). Second, matched post-mortem blood and brain samples are needed to demonstrate that circadian disruption is detectable in blood. If these two steps were to be established, then it would pave the way to identification of disruption of circadian functions in a subset of living patients. Investigations could be conducted longitudinally within a person comparing mental ill health episodes to periods of euthymia and good health. Since a meta-analysis of randomized clinical trials (Scott et al. Reference Scott, Webb, Martyn-St James, Rowse and Weich2021) has shown that improved sleep quality leads to better mental health, quantification of the changes in circadian rhythm before and after sleep improvement using omics data could provide a direct test of the causal role of circadian disruption separated from sleep disruption. Over the next 5-10 years we expect to see a rapid growth both in statistical and computational methods and of data sets. This emerging area of research could help provide the evidence base to address the question posed by Hickie et al. (Reference Hickie, McCarthy, Crouse and Carpenter2024).