This analysis, published in BMC Genomic Data, leverages a comprehensive dataset from the Molecular Transducers of Physical Activity Consortium (MoTrPAC) to evaluate the predictive power of transcriptomic data.
The investigation utilized a cohort of 50 inbred Fischer 344 rats, which were subjected to endurance exercise training over periods of one, two, four, and eight weeks. A control group of rats was maintained without exercise for the same eight-week duration. The MoTrPAC dataset encompasses hundreds of assays across 19 different tissues, providing a whole-organism molecular map of the temporal effects of endurance exercise. This extensive dataset allows for the evaluation of gene expression, isoform-level changes, and associations with other physiological parameters.
The primary objective of the study was to assess whether gene expression could serve as a predictor of the extent of exercise. Previous analyses by the MoTrPAC authors focused on differential expression of genes across tissues but did not evaluate the predictive power of these genes for the duration of training. The current study addresses this gap by employing modern computational tools to analyze the RNA-seq data, aiming to understand how molecular measurements can provide a roadmap for precision health and the optimization of exercise regimes.
To process the RNA-seq data, the researchers used kallisto-bustools, a suite of tools for bulk and single-cell RNA-seq analysis that allows for counting unique molecules via unique molecular identifiers (UMIs). A total of 21 samples were unavailable for reprocessing due to prior discards by the original study, leaving 879 different bulk RNA samples. The team discarded an additional 14 samples whose UMI counts deviated by more than three standard deviations from the mean for that tissue type, ensuring high-quality data for downstream analyses.
Quality control procedures included Leiden clustering of the samples, which identified four samples clustered in tissues different from their labels. Three of these had been previously discarded as outliers, while one was identified as a clerical error involving swapped labels between a hippocampus and a hypothalamus sample from the same rat. The researchers relabeled these samples, rescuing one for inclusion in the final analysis, which comprised 861 samples with at least 23 individuals per sex for all tissues except the Vena Cava.
The predictive model was developed using a generalized linear model regularized with ridge regression constraints. The model utilized the transcriptional profile of a randomly selected subset of 36 rats across 16 different tissues simultaneously, rather than analyzing each tissue individually as in previous studies. The input matrix was created by flattening the gene count matrix for each tissue into a vector for each individual rat, preserving tissue-specific expression profiles. The model achieved a concordance correlation coefficient of 0.81 on the testing set, with an average of 0.77 across 100 Monte Carlo cross-validation trials.
Analysis of the top 600 most significant predictors revealed that gene ontology terms associated with muscle contraction and filament sliding were positively correlated with exercise, while those involved in fatty acid metabolism were negatively correlated. Among the physiological measurements collected, only the change in percent body fat could be effectively predicted by gene expression. Myoglobin expression in skeletal muscle predicted low body fat and more weeks of exercise, whereas expression of the Scd gene in brown adipose tissue predicted high body fat and fewer weeks of exercise.
When comparing different omics data, proteomics emerged as the best predictor of exercise duration, followed by acetylomics. In contrast, Methyl-seq and ATAC-seq data showed the worst predictive outcomes, aligning with the original study’s findings that these modalities had the fewest differentially expressed features. The use of scVI, a variational autoencoder, to predict exercise duration resulted in a loss of predictive power, suggesting that while such tools are useful for data compression, they may remove biological signal relevant to prediction.
The study also investigated isoform-level responses to exercise using differential gene analysis with pyDeseq2. The model accounted for tissue, time, sex, and interaction terms. In a case study of gastrocnemius skeletal muscle, the researchers identified 18 genes that were significantly differential at the isoform level but not at the gene level. These findings would have been missed by a purely gene-level analysis, which identified 125 differentially expressed genes. An example of such a gene is SUMO, which showed differential expression at the isoform level despite no noticeable changes at the gene level.
Furthermore, the researchers applied a method for identifying RNA virus sequences in RNA-seq data via translated alignments to the PalmDB database. A positive correlation was found between total virus counts in a sample and the amount of time before sample collection. One virus-like sequence, u234187, was found to be differentially expressed between control and eight-week exercise rats. This sequence correlated with the expression of the Tigit gene, an immune expression gene previously reported as differentially expressed. The rat with the highest expression of this virus-like sequence also exhibited the largest negative change in VO2 max among the control rats, suggesting that viral infection may have confounded physiological measurements.
The findings indicate that gene expression can accurately predict the amount of exercise rats undergo, and this result is robust to the machine learning approach used. Linear regression proved effective for this task, whereas more sophisticated machine learning tools like variational autoencoders may not be ideal for prediction. The integration of data across the entire dataset uncovered gene expression programs that may be missed when each tissue is interrogated individually, highlighting the importance of multi-tissue analysis.
The study underscores the necessity of using tools that can reach the resolution of isoform-level changes to better explain expression changes and inspire further investigation. The discovery of a potential novel virus that may have negatively impacted physiological measurements in the infected rat highlights the importance of assessing animal health in studies linking genomics to physiological perturbations. Viruses are an underexplored aspect of RNA sequencing and may constitute a potential confounding variable in many studies.
Future experiments could be refined by more tightly controlling and tracking the amount of food consumed. This comprehensive analysis provides a blueprint for directing similar analyses with respect to physiological perturbations across organisms.
Beyond gene expression, the experimental design tracked multiple physiological metrics and molecular layers across the rat cohort. For the control, four-week, and eight-week groups, investigators monitored changes in body weight, percent lean mass, and maximal oxygen consumption alongside body fat. The broader corpus also captured acetylproteomics, global proteomics, phosphoproteomics, metabolomics, lipidomics, ubiquitylome profiles, Methyl-seq, bulk ATAC-seq, and multiplexed immunoassays. Each time point originally evaluated five male and five female rats, though identical animals were not sampled across every assay. Because male animals lacked omics data across several time points, the combined regression across all nine modalities was restricted to female rats and evaluated solely in heart tissue, the only organ sampled across the complete molecular panel.
Methodological choices in the regression pipeline directly influenced how cross-tissue dynamics emerged. When constructing transcriptomic input matrices, read counts were normalized for depth but left unlogged to ensure linear regression accurately reflected direct proportional relationships between expression levels and training time. When evaluating tissues individually, both types of skeletal muscle yielded the most impactful coefficients among top predictors. However, individual tissue predictive power did not correlate with tissue contributions to the full model, demonstrating that whole-organism modeling detects coordinated programs obscured during isolated single-tissue evaluations. In contrast, training a variational autoencoder across four hundred epochs to condense features into a ten-dimensional latent space eliminated predictive performance, with sample clustering remaining dominated by tissue type despite batch correction.
The analytical framework also provided qualifications regarding isoform-level discoveries. To identify transcript differences, researchers aggregated transcript p-values for each gene using the Lancaster method weighted by transcript count logarithms. While the model isolated eighteen genes with significant isoform alterations despite stable gene-level expression, the authors cautioned that rat genome annotations relied heavily on sequence similarity rather than direct empirical validation. Consequently, detected transcript shifts could reflect gene family expression dynamics, incomplete annotations, or ambiguous read assignments rather than definitive post-transcriptional regulation. Nevertheless, because predictive RNA signatures mirrored protein abundance but showed poor concordance with chromatin accessibility data, investigators hypothesized that fitness adaptations primarily reflect post-transcriptional activity, with transcript counts functioning largely as proxies for cellular protein products.
The viral detection workflow applied strict computational controls to isolate biological signals from technical artifacts. To prevent false matches, the viral reference index employed host masking before aligning reads against the PalmDB database, and sequences matching known laboratory reagent contaminants were systematically removed. Summed viral counts across tissues correlated with elapsed time before sacrifice, prompting a differential comparison between control and eight-week exercise animals to account for this temporal progression. When raw reads from sequence u234187 were queried against the BLAST database, alignments matched viral genomes including ORF and SARS.
Subject of Research: Transcriptomic prediction of endurance exercise training duration in rats
Article Title: Transcriptomic responses to endurance exercise training in rats
Article References: Oakes, C., & Pachter, L. (2026). Transcriptomic responses to endurance exercise training in rats. BMC Genomic Data, 27(1), Article 74. https://doi.org/10.1186/s12863-026-01440-8
Image Credits: AI Generated
DOI: 10.1186/s12863-026-01440-8
Keywords: transcriptomics, endurance exercise, gene expression, rats, RNA-seq, MoTrPAC, isoform analysis, viral infection, physiological prediction, molecular biology, Transcriptomic, responses
News Source: Juliet Wilcox. (November 10, 2026). Gene Expression Predicts Endurance Training Levels in Rats. Scienmag.
