In a recent study published in the Proceedings of the National Academy of Sciences, researchers from the United States of America developed and validated "TimeMachine," an algorithm that predicts the circadian phase of patients using gene expression in peripheral blood mononuclear cells (PBMCs) from a single blood sample. They found that the algorithm was flexible and accurate in its predictions and performed well on new data without retraining or renormalization.
Study: Platform-independent estimation of human physiological time from single blood samples. Image Credit: Nuva Frames / Shutterstock
Background
The circadian rhythm, a conserved and endogenous timekeeping system, influences various biological processes. Dysregulation of this rhythm is associated with health issues like obesity, diabetes, cancer, and cardiovascular disease. Aligning drug dosing with circadian cycles may enhance treatment efficacy and reduce side effects. Dim-light melatonin onset (DLMO), the current gold standard measure of the circadian phase, is time-consuming and costly, hindering broad implementation in research and clinical settings. Transcriptomic profiling and machine learning offer a promising alternative, using gene expression as a readout of circadian rhythm. Previous algorithms developed in this regard lacked generalizability and were limited by batch correction, retraining, or platform-specific challenges. To address this gap, the present study introduced the "TimeMachine" algorithm to predict the human circadian phase from a single blood draw while focusing on simplicity and generalizability.
About the study
The TimeMachine algorithm's framework involves three steps: feature selection, within-sample rescaling, and fitting the predictor. It was initially trained and tested on a human circadian gene expression dataset (TrTe). Further, it was applied without batch correction or retraining to sample data from three independent, published datasets (V1, V2, and V3) of the human whole-blood transcriptome. All the datasets used in the present study were obtained from the National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO) repository and involved the use of various platforms for microarray and ribonucleic acid sequencing (RNA-seq). A total of 7,615 genes common to all the datasets were used for analysis.
The algorithm begins by selecting genes containing phase information as potential circadian biomarkers. Applying existing methods (ZeitZeiger, PLSR-based, TimeSignature) to training data helped identify 135 candidate genes. From these, 37 genes exhibiting robust cycling patterns, determined by JTK_Cycle analysis, were utilized as inputs for the TimeMachine predictor.
The researchers proposed two normalization approaches: pairwise gene ratios and z-score transformation, leading to the development of two algorithm variants— ratio TimeMachine (rTM) and z-score TimeMachine (zTM), respectively. These variants aim to ensure consistent and comparable expression data across platforms by emphasizing relative gene expression rather than absolute magnitudes. Both variants use bivariate regression with elastic net regularization for predicting physiological time as a function of gene expression. The evaluation considered the median absolute error and the normalized area under the error cumulative distribution function curve (AUC). Further, TimeMachine was compared to the state-of-the-art method based on PLSR (short for partial least squares regression) across multiple datasets and platforms.
Results and discussion
For TrTe, rTM achieved a median absolute error of 1.39 h, with 55.7% of predictions within 2 h and 83.8% within 4 h. When applied to V1 and V3, rTM demonstrated accurate predictions despite differences in experimental conditions and profiling platforms, yielding a median absolute error of 2:41 h for V1 and 1:53 h for V3. The zTM variant displayed comparable performance, suggesting that both normalization methods are equivalent for inferring the circadian phase from blood transcriptomics across platforms.
In comparison studies, TimeMachine exhibited similar or superior performance to PLSR, with a median absolute error of 2:13 h to 2:55 h. Importantly, TimeMachine outperformed PLSR while requiring fewer predictor genes. Both ratio TimeMachine and z-score TimeMachine variants achieved comparable accuracy, outperforming PLSR in datasets TrTe and V3. Generalizability assessments across all four studies demonstrated TimeMachine's consistent performance in predicting local time, even across different platforms. Compared to two-timepoint methods, rTM and zTM showed a mean absolute error more significant by 20–40 min. Overall, TimeMachine's one-time point predictions of circadian phase and local time exhibited robust and competitive performance across diverse datasets and platforms.
Additionally, the study investigated the factors influencing TimeMachine's performance, focusing on the relationship between prediction accuracy and predicted amplitude. Samples with predicted amplitudes below 0.5 consistently exhibited significantly higher errors for rTM and zTM, providing insights into prediction confidence. Categorizing samples by phase intervals revealed an inverse relationship between predicted amplitude and error.
Conclusion
In conclusion, TimeMachine addresses the challenges of assaying circadian biomarkers, offering accurate predictions of the circadian phase using a single-timepoint gene expression profile of PBMCs. Its practicality, agreement with in-lab data, and generalizability for prospective and retrospective analyses make it a valuable tool for diverse applications in medical research, clinical settings, and the exploration of circadian rhythms' roles in various diseases, including cancer.
Groundbreaking study unveils sex-specific genetic influences on blood pressure