Time trajectories in the transcriptomic response to exercise – a meta-analysis

The study investigates how human tissues transcriptionally respond over time to an acute bout of exercise versus long-term training, and how moderators such as time post-exercise, training modality, age, and sex shape these responses. Although exercise confers broad health benefits and prevents many chronic diseases, the underlying molecular mechanisms remain incompletely defined. Prior transcriptome studies in skeletal muscle and blood identified differentially expressed genes but were limited by small sample sizes, heterogeneous designs, and incomplete coverage of moderators (sex, age, time). The authors aim to integrate and harmonize available datasets, explicitly model key moderators, and delineate time trajectories and regulatory networks underlying exercise adaptation in blood and skeletal muscle.

Previous transcriptomic studies of exercise in skeletal muscle and blood using microarrays and RNA-seq identified many responsive genes but suffered from small cohorts and heterogeneous sampling of moderators and time points. Meta-analysis has been successfully applied in genomics but naive approaches can inflate false positives, especially with high heterogeneity and small numbers of studies. Pillon et al. (2020) performed a random-effects meta-analysis of skeletal muscle exercise datasets and highlighted NR4A3 but considered only a single moderator and did not systematically model time dynamics, potentially limiting power and interpretability. The present work addresses these gaps by applying mixed-effects meta-regression with moderator selection and stringent filters to improve replicability and to capture temporal patterns across studies.

Data collection and curation: The authors searched GEO (2/22/2019) for human exercise/training transcriptomic datasets (arrays or RNA-seq) in skeletal muscle or blood with pre- and post-exercise samples and at least two time points per subject. Inclusion criteria: published data, ≥3 subjects, whole blood or PBMCs for blood, non-disease cohorts. Adipose tissue and disease-associated datasets were excluded. From 670 datasets screened, 43 datasets met criteria, yielding 59 cohorts (13 blood, 46 skeletal muscle) and 1724 samples from 739 individuals. Some studies contributed both acute and long-term cohorts. Sex was imputed for 116 subjects using an SVM trained on Y-chromosome gene expression (AUC 0.99), resulting in 310 females, 409 males; 18 missing; 2 discordant.

Cohort definition and moderators: Datasets were partitioned into homogeneous cohorts by tissue, study arm, and training modality (endurance, resistance, or untrained controls). Four meta-analyses were defined: acute muscle (15 cohorts), long-term muscle (26), acute blood (10; 8 trained, 2 untrained), and long-term blood (13; 11 trained, 2 untrained). Moderators retained where coverage allowed: for muscle, acute analysis included time, training modality, and age; long-term included time (binned <150 vs >150 days), training modality, age, and sex (proportion male). Blood datasets had insufficient coverage for moderator modeling; base models only were used.

Preprocessing and summary statistics: For each gene and cohort time point, the mean log2 fold-change (y) relative to baseline, its variance (v), and paired t-test p-value (p) were computed. Genes missing ≥25% values were excluded, retaining >18,000 genes per analysis (acute muscle 18,374; acute blood 18,621; long-term muscle 18,685; long-term blood 19,683). Genes proceeding to model selection had ≥2 p<0.05 across cohorts (acute muscle 13,016; long-term muscle 14,309; acute blood 8,650; long-term blood 508).

Modeling: Using the R package metafor, random-effects (RE) meta-analysis and mixed-effects meta-regression were performed per gene. The base RE model estimated overall effect (µ) with between-study heterogeneity (τ²) and I². A nested random-effects structure (cohort|study) was also tested. Meta-regression modeled y = β0 + Σβj xij + ui + εi with moderators x (time, modality, age, sex) where available.

Model selection and filters: For each gene, all combinations of eligible moderators and two random-effects structures were evaluated and ranked by corrected AIC (AICc). A model was selected if it improved over the base model by ΔAICc>5, had overall model p<0.001 and absolute effect size |β|>0.1. If no improved model met criteria, the base model could be selected if I²<50%, p<0.001, and |µ|>0.1. These dual thresholds (significance and effect size) were used to enhance reproducibility and control false positives.

Pathway and network analyses: Naive RE meta-analysis fold-changes were ranked and subjected to GSEA (Reactome pathways; fGSEA with 50,000 permutations; 10% BY-FDR). Selected genes were clustered by k-means on t-statistics within model-defined groups (cluster number via elbow method). For acute muscle time-only genes, temporal clustering and TF inference were performed using DREM with ENCODE TF-target networks and low node penalty (12). Network integration and visualization used GeneMANIA and ReactomePA within Cytoscape.

Validation: An independent human acute endurance cohort (n=16; 8 males, 8 females) performed 60 min cycling at 70% peak VO2 with vastus lateralis biopsies pre-, 2 h, and 6 h post-exercise. qRT-PCR measured SMAD3, NR4A1, HES1, ID1, SCN2B, SLC25A25, MTMR3, normalized to GAPDH and RPS18. Mixed-effects modeling with Dunnett’s test evaluated time effects.

Untrained controls: Separate RE meta-analyses on untrained cohorts (acute muscle 3 cohorts; acute blood 2; long-term muscle 2) served to test exercise specificity by comparing effect sizes of selected gene sets between exercise and untrained data via paired Wilcoxon tests.

• Naive pathway-level signals: Across four analyses (acute vs long-term; muscle vs blood), multiple Reactome pathways were enriched at 10% BY-FDR. Respiratory electron transport pathways were significant in all four. Example of ignoring moderators: mitochondrial translation showed negative NES in acute muscle (NES −2.17), but gene-level effects (e.g., MRPL34) depended on late time points (>20 h), illustrating time-dependent heterogeneity (I²≈65%).
• Model-selected differential genes: After moderator-aware model selection and stringent filters, detected genes were: • Acute exercise, muscle: 537 genes. • Long-term training, muscle: 441 genes. • Acute exercise, blood: 37 genes. • Long-term training, blood: 48 genes. Lower counts in blood reflected fewer cohorts, lower coverage, and smaller effect sizes.
• Temporal patterns in acute muscle: 159 genes were selected with time-only models and clustered into four temporal trajectories (early-only, early–mid, late up, late down), mapping to functions such as FOXO-mediated transcription, blood vessel development and stress response (early), myeloid activation/neutrophil regulation (late up), and fatty-acid metabolism/mitochondrial protein import (late down). Network integration revealed a large connected component with SMAD3 as a central early hub; additional notable genes included PPARGC1A (PGC-1α; up at 2–5 h, mild down >20 h), HES1 (early induced), NR4A1, PDGFB, VEGFA (angiogenesis), SPP1 (late up), MLYCD and CPT1B (late down).
• Upstream regulators (DREM): Early (0–1 h) putative regulators included BHLHE40 and HDAC2 (top), and NR3C2 and DBP (NELFE/negative elongation factor E), linking early temporal branches to TF activity.
• Long-term training in muscle: 104 genes upregulated and 10 downregulated with moderator-independent patterns; downregulated included MSTN. Upregulated genes were enriched for extracellular matrix (ECM) reorganization and laminin interactions. Strong induction observed for collagen genes (COL4A1, COL4A2, COL1A1, COL5A2). Regulatory hubs included ITGA1 and KDR (VEGFR2), underscoring angiogenesis and ECM remodeling.
• Overlap between acute and long-term muscle responses was limited (13 genes shared), consistent with distinct transcriptional programs.
• Moderator-specific effects: • Age (acute muscle): 76 genes associated with age; included NR4A2 and NR4A3. • Age (long-term muscle): 73 genes associated with age; older individuals showed greater induction of interferon/inflammatory genes (e.g., HLA-DRA, HLA-F, CD44, IFI44, IFI44L). • Sex (long-term muscle): 247 genes showed sex-associated regulation, enriched for chromatin organization; notable regulators included HIF1A and HDAC3. MTMR3 decreased more in male-predominant cohorts; not previously linked to exercise.
• Blood responses: Acute blood upregulated genes enriched for neutrophil degranulation; long-term blood showed downregulation of glycerophospholipid biosynthesis and upregulation of peptide chain elongation.
• Exercise specificity: Effect sizes of selected genes were significantly different in exercise cohorts versus untrained controls (paired Wilcoxon): acute muscle 537 genes, p=1.5×10^-88; long-term muscle 441 genes, p=3.1×10^-68; acute blood 37 genes, p=6.2×10^-08 (all p<1×10^-7).
• Experimental validation: In an independent cohort (n=16), SMAD3 and NR4A1 were significantly up at 2 h; HES1 and ID1 showed significant downregulation at 2 h and 6 h (consistent with quadratic early induction followed by downregulation); SCN2B trended up (p=0.09); SLC25A25 increased at 2 h and decreased at 6 h; MTMR3 showed no significant sex interaction in acute validation, consistent with a subtle training-associated effect identified by meta-analysis.
• Resource: All curated data and results are available at www.extrameta.org and GitHub (AshleyLab/motrpac_public_data_analysis).

By explicitly modeling moderators—especially time post-exercise—this study clarifies that acute and long-term adaptations are transcriptionally distinct, and it resolves contradictory signals arising from naive meta-analyses that ignore heterogeneity. The temporal clustering and network integration identify SMAD3 as a central early regulator, implying that TGF-β/SMAD signaling helps orchestrate downstream transcriptional programs following acute exercise. Long-term training adaptations emphasize ECM remodeling and angiogenesis, reflected by upregulated collagens, integrins, and KDR. Moderator-aware analyses reveal biologically meaningful heterogeneity: older individuals exhibit a more pronounced training-induced inflammatory signature in muscle, and numerous sex-associated gene regulation differences emerge, including chromatin-related genes and MTMR3. Blood transcriptomic responses show acute immune activation and modest long-term translational changes but are limited by sample size and coverage. The validation cohort corroborated several key acute regulators and temporal patterns, supporting the robustness of the meta-regression and model selection framework. Overall, the findings map time trajectories and regulatory networks underlying exercise adaptation and demonstrate that accounting for moderators improves inference and reproducibility.

This work delivers a comprehensive, moderator-aware meta-analysis of human exercise transcriptomes across skeletal muscle and blood, distinguishing acute and long-term responses and uncovering time trajectories, regulatory hubs (notably SMAD3), and moderator-specific effects of age and sex. The study validates key genes in an independent cohort and provides an open resource (ExTraMeta: www.extrameta.org) with curated datasets and results. Methodologically, it introduces a model selection pipeline using AICc, effect size thresholds, and heterogeneity control to improve replicability in heterogeneous genomic meta-analyses. Future research should expand underrepresented modalities (e.g., acute resistance in blood), include richer phenotypic moderators (BMI, training status, VO2max, heart rate), achieve more balanced sampling across moderators, and integrate individual physiological responses to further refine mechanistic insights and personalized exercise prescriptions.

• Limited moderator coverage across studies prevented inclusion of potentially important covariates (e.g., height, weight, BMI, training status, baseline VO2max, baseline heart rate, biopsy site). Blood datasets lacked sufficient coverage to perform moderator-specific analyses; no public acute resistance blood datasets were available.
• Heterogeneity across small-to-moderate numbers of studies can bias p-values, I² estimates, and weighting in random-effects meta-analysis; small studies may receive disproportionate weight. Unbalanced sampling across time and moderators can reduce power and inflate false positives if unaddressed.
• The approach cannot account for individual-level physiological variability in response to exercise, known to be substantial, because only study-level summaries and moderators were modeled.
• Blood analyses had fewer cohorts, fewer genes meeting inclusion criteria, and generally smaller effect sizes, limiting robustness compared to skeletal muscle.
• Meta-regression is exploratory and depends on included moderators; unmeasured confounders may remain.