Does Caffeine Mimic Exercise at the Cellular Level? A Comprehensive Transcriptomic Crossexamination of MoTrPAC Human and Rodent Gene Sets

Abstract

Background: Caffeine is the most widely consumed psychoactive substance worldwide. Growing evidence suggests that it may share molecular effects with physical exercise, including enhanced mitochondrial biogenesis, improved metabolic health, and modulation of inflammatory pathways. However, a systematic comparison of their transcriptomic signatures using comprehensive exercise data has not been performed at scale.

Methods: We queried the PerturbSeqR database for caffeine-induced gene expression signatures across multiple cell lines and perturbation platforms (37 gene sets, yielding 2,630 caffeine-upregulated and 2,602 caffeine-downregulated genes). These were compared against 176 MoTrPAC gene sets: 36 from primary human tissues (muscle, blood, adipose) and 140 from the full MoTrPAC 2023 library spanning 14 tissues including rodent data. Hypergeometric enrichment tests assessed concordant (same-direction) and antagonistic (opposite-direction) overlaps with Benjamini-Hochberg FDR correction. Combined Fisher's method assessed overall significance.

Results: Among 36 MoTrPAC Human comparisons, 16 showed significant same-direction overlap and 21 showed significant opposite-direction overlap (FDR < 0.05), yielding a combined Fisher's p = 1.46 × 10⁻¹⁴⁴. Among the full MoTrPAC 2023 library (140 gene sets), 19 comparisons survived FDR correction with predominantly same-direction enrichment (11 same vs. 8 opposite, combined p = 7.64 × 10⁻⁸). Muscle showed the strongest individual signal across both datasets (human best p = 3.24 × 10⁻¹³; rodent-inclusive best p = 9.6 × 10⁻¹⁰). Consistently overlapping genes included CST3, HMOX1, BST2, MYLK, CXCL9, and multiple mitochondrial genes.

Conclusions: Caffeine induces a transcriptional program that substantially overlaps with both human and rodent exercise responses. The human-specific data reveals a more balanced pattern of synergistic and antagonistic effects compared to the rodent-inclusive library, where same-direction predominance was stronger. This suggests that caffeine's molecular mimicry of exercise is partially conserved across species but attenuated in primary human tissues, particularly in skeletal muscle where mechanical loading drives unique adaptations.

Keywords: caffeine, exercise, transcriptomics, MoTrPAC, molecular mimicry, gene signature overlap, hypergeometric enrichment, systems biology

---

Introduction

Caffeine (1,3,7-trimethylxanthine) is the most widely consumed psychoactive substance globally, with over 80% of adults consuming it daily^1-3. Beyond central nervous system stimulation via adenosine receptor antagonism, caffeine influences diverse physiological processes: lipid metabolism, calcium signaling, mitochondrial function, and immune modulation⁴. Emerging evidence suggests caffeine supplementation can enhance exercise performance and may share molecular adaptations with physical training^5-7.

The Molecular Transducers of Physical Activity Consortium (MoTrPAC) systematically cataloged exercise-induced transcriptomic changes across human and rodent tissues⁸. The MoTrPAC 2023 library contains 140 gene sets spanning 14 tissues, while the MoTrPAC Human GMT contains 36 gene sets from primary human tissues (muscle, blood, adipose), providing complementary perspectives on exercise transcriptomics.

The hypothesis that caffeine may mimic exercise at the molecular level is attractive: if caffeine recapitulates key exercise-induced gene expression patterns, it might serve as a pharmacological exercise mimetic for populations unable to engage in physical activity. Exercise mimetics such as AICAR activate AMPK and promote mitochondrial biogenesis^9-11, but whether caffeine reproduces the exercise transcriptome in human and rodent tissues remains untested at scale.

Here we directly compare caffeine-induced gene signatures from PerturbSeqR against both the MoTrPAC Human gene sets (36 comparisons) and the full MoTrPAC 2023 library (140 gene sets), testing for both concordant and antagonistic gene-level effects across 176 total comparisons.

Methods

Caffeine Gene Signatures

We queried PerturbSeqR (https://perturbseqr.maayanlab.cloud) for all perturbation experiments involving caffeine. PerturbSeqR aggregates data from LINCS L1000¹², Ginkgo Bioworks, NIBR DRUG-seq, and RummaGEO Chem platforms. We retrieved 37 caffeine-related gene sets spanning 8 cell lines (A549, HCC515, U2OS, HEpg2, MCF10A.TP53.M, A375, MCF7, and mouse primary cells). For each experiment, up-regulated and down-regulated genes were separated, yielding aggregate gene sets of 2,630 caffeine-upregulated and 2,602 caffeine-downregulated genes.

MoTrPAC Gene Set Libraries

MoTrPAC Human. The MoTrPAC Human GMT file contains 36 gene sets derived from primary human tissues following exercise interventions, spanning three tissue types:

• Muscle: 12 gene sets from skeletal muscle biopsies (vastus lateralis, gastrocnemius) following acute and chronic exercise
• Blood: 14 gene sets from whole blood RNA following exercise
• Adipose: 10 gene sets from white and brown adipose tissue following exercise

MoTrPAC 2023 Library. The full MoTrPAC 2023 Enrichr library contains 140 gene sets spanning 14 tissues: skeletal muscle (vastus lateralis and gastrocnemius), heart, liver, blood RNA, lung, white adipose tissue, brown adipose tissue, hippocampus, cortex, hypothalamus, and others. This library includes both human and rodent data from acute (single session) and chronic (1, 2, or 8 weeks) exercise interventions^8,13.

Statistical Analysis

Hypergeometric enrichment testing. For each MoTrPAC gene set, we tested enrichment for caffeine-upregulated or caffeine-downregulated genes using the hypergeometric test with a background of 20,000 protein-coding human genes:

• Same-direction test: For exercise-up sets, test enrichment for caffeine-up genes; for exercise-down sets, test enrichment for caffeine-down genes.
• Opposite-direction test: For exercise-up sets, test enrichment for caffeine-down genes; for exercise-down sets, test enrichment for caffeine-up genes.

Multiple testing correction. Benjamini-Hochberg FDR procedure at q < 0.05.

Combined significance. Fisher's method for combining independent p-values across all categories, performed separately for the human and rodent-inclusive datasets.

Results

Human Data: Massive Combined Signal with Balanced Directionality

Among 36 MoTrPAC Human comparisons, 16 showed significant same-direction overlap and 21 showed significant opposite-direction overlap (FDR < 0.05). The combined Fisher's test yielded a p-value of 1.46 × 10⁻¹⁴⁴ (statistic = 906.43), far exceeding random expectation.

The near-equal balance of same-direction (16) versus opposite-direction (21) significant hits indicates that caffeine both mimics and opposes exercise effects in primary human tissues. Muscle showed the strongest individual signal (best p = 3.24 × 10⁻¹³) but was dominated by opposite-direction hits (7 of 12 significant opposite vs. 2 of 12 significant same). Blood showed a balanced pattern (8 same, 9 opposite, best p = 3.55 × 10⁻⁹), while adipose showed modest same-direction predominance (6 same, 5 opposite, best p = 1.98 × 10⁻⁶).

Table 1. MoTrPAC Human tissue-specific overlap statistics.

Tissue	N Tests	Sig Same	Sig Opp	Best P (Same)
Muscle	12	2	7	3.24 × 10⁻¹³
Blood	14	8	9	3.55 × 10⁻⁹
Adipose	10	6	5	1.98 × 10⁻⁶
Total	36	16	21	3.24 × 10⁻¹³

Figure 1. Same-direction enrichment between caffeine and MoTrPAC Human exercise gene sets. Horizontal bars show -log₁₀(p-value) for each gene set. Steel blue indicates FDR-significant overlaps (q < 0.05).

Figure 2. Opposite-direction enrichment between caffeine and MoTrPAC Human exercise gene sets. Horizontal bars show -log₁₀(p-value) for each gene set. Red indicates FDR-significant overlaps (q < 0.05).

Figure 3. MoTrPAC Human tissue-specific summary. Same-direction (steel blue) versus opposite-direction (red) significant comparisons by FDR < 0.05. Muscle shows the strongest individual signal, while blood and adipose show the most balanced same-direction versus opposite-direction patterns.

Rodent-Inclusive Data: Stronger Same-Direction Predominance

Among 156 MoTrPAC 2023 library comparisons (78 same-direction + 78 opposite-direction), 19 survived FDR correction with same-direction predominance (11 same vs. 8 opposite). The combined Fisher's test yielded p = 7.64 × 10⁻⁸.

The stronger same-direction bias in the rodent-inclusive library compared to the human-specific data is notable. The top signals came from adipose tissue (white adipose consensus: 335 overlapping genes, p = 4.4 × 10⁻¹⁴; brown adipose consensus: 316 genes, p = 1.1 × 10⁻¹³), heart (158 genes, p = 2.5 × 10⁻¹¹), and skeletal muscle (gastrocnemius consensus: 126 genes, p = 9.6 × 10⁻¹⁰).

Table 2. Top tissue-specific overlaps from MoTrPAC 2023 library (FDR-significant, same direction).

Rank	Tissue	Gene Set	Overlap	p-value	FDR
1	White Adipose	Consensus	335	4.4 × 10⁻¹⁴	6.9 × 10⁻¹²
2	Brown Adipose	Consensus	316	1.1 × 10⁻¹³	9.9 × 10⁻¹⁰
3	Heart	Consensus	158	2.5 × 10⁻¹¹	3.0 × 10⁻⁸
4	Gastrocnemius	Consensus	126	9.6 × 10⁻¹⁰	8.3 × 10⁻⁸
5	Lung	Consensus	189	4.3 × 10⁻⁹	1.8 × 10⁻⁷
6	Blood RNA	Consensus	283	9.2 × 10⁻⁹	2.5 × 10⁻⁴
7	Vastus Lateralis	Consensus	141	1.5 × 10⁻⁵	4.1 × 10⁻⁴
8	Liver	Consensus	61	2.6 × 10⁻⁵	2.2 × 10⁻³
9	Cortex	Consensus	28	1.6 × 10⁻⁴	2.2 × 10⁻³

Figure 4. Gene-level overlap between caffeine and MoTrPAC 2023 exercise signatures by direction. Horizontal bars represent overlapping genes in each category, color-coded by directionality (steel blue = same direction, red = opposite direction). P-values and odds ratios are annotated on each bar.

Figure 5. Tissue-specific enrichment of caffeine-exercise transcriptomic overlap from the MoTrPAC 2023 library. Heatmap shows the top 15 tissues ranked by -log₁₀(p-value) of same-direction overlap. White adipose tissue shows the strongest overlap, followed by brown adipose tissue and heart.

The difference between human and rodent-inclusive results has two possible explanations:

1. Species differences: Rodent exercise responses may more closely recapitulate cell-line caffeine responses than human primary tissue responses do, potentially reflecting evolutionary divergence in exercise adaptation pathways.

2. Data type differences: The MoTrPAC 2023 library includes consensus gene sets that aggregate across multiple experiments, potentially amplifying consistent signals. The MoTrPAC Human GMT includes individual gene sets that may be noisier but more specific.

Consistently Overlapping Genes

Several genes appeared in significant overlapping sets across multiple tissues and both datasets, forming a "core" set of consistently affected genes.

Table 3. Top 20 most consistently overlapping genes across MoTrPAC 2023 tissues.

Rank	Gene	Gene Name	Tissues with Overlaps	Significant	Primary Category
1	CST3	Cystatin C	14/14	11	Stress Response
2	HMOX1	Heme Oxygenase 1	14/14	7	Stress Response
3	BST2	Tetherin	12/12	11	Immune
4	MYLK	Myosin Light Chain Kinase	12/12	10	ECM/Remodeling
5	AQP1	Aquaporin 1	12/12	8	Transport
6	ALDH2	Aldehyde Dehydrogenase 2	10/10	8	Metabolic
7	FBN1	Fibrillin 1	10/10	8	ECM/Remodeling
8	MT-CO3	Cytochrome c Oxidase III	12/12	8	Mitochondrial
9	COL15A1	Collagen Type XV Alpha 1	10/10	9	ECM/Remodeling
10	S100A6	S100 Calcium Binding Protein A6	10/10	9	Stress Response

Figure 6. Most consistently overlapping genes across tissues. Horizontal bars show tissues in which each gene appeared in a significant overlap, color-coded by functional category (mitochondrial = steelblue, stress response = darkorange, ECM/remodeling = purple, immune = seagreen, other = slategray). CST3 and HMOX1 appeared in all 14 tissues tested.

Notably, mitochondrial genes (MT-CO3, MT-ATP6, MT-ND1, MT-ND3, MT-ND5, UQCRQ) were consistently represented in overlapping sets across 10 of 14 tissues in the MoTrPAC 2023 library, consistent with both caffeine and exercise being known modulators of mitochondrial function through AMPK and PGC-1α activation^9,11,14-16. Stress response genes (HMOX1, CST3, PRDX5, HSPD1) were prominent across all tissues, reflecting shared activation of cellular stress adaptation pathways including the integrated stress response and oxidative stress tolerance^17,18. Extracellular matrix/remodeling genes (COL15A1, FBN1, MYLK, Vimentin) suggest overlap in tissue remodeling processes, particularly in cardiovascular and muscular tissues where both interventions promote structural adaptation.

Discussion

Caffeine Partially Mimics Exercise Across Species

The combined analysis of human and rodent-inclusive data reveals a consistent pattern: caffeine shares a substantial transcriptional program with exercise, but the degree of overlap depends on the data source. The rodent-inclusive MoTrPAC 2023 library shows stronger same-direction predominance (11 vs. 8 significant hits), while the human-specific data reveals a more balanced pattern (16 same vs. 21 opposite).

This difference has important implications for the molecular mimicry hypothesis. The strong same-direction signal in the rodent-inclusive library supports the idea that caffeine can recapitulate exercise effects in certain contexts, particularly in metabolic tissues (adipose, liver) and cardiovascular tissues (heart, blood). However, the human-specific data reveals that the mimicry is less complete when restricted to primary human tissues, particularly in skeletal muscle where mechanical loading drives unique adaptations.

Tissue-Specificity of the Mimicry

The strength of the overlap varies substantially by tissue:

1. Adipose tissue (both white and brown) shows the largest raw overlaps (335 and 316 genes respectively) and the most significant p-values in the rodent-inclusive library. In the human data, adipose shows a modest same-direction bias (6 vs. 5 significant). Both agents promote lipolysis and thermogenesis in adipose tissue^19,20.

2. Heart and cardiovascular tissues show consistent overlap across both datasets. In the rodent-inclusive library, heart shows the 3rd strongest signal (158 genes, p = 2.5 × 10⁻¹¹). In the human data, blood shows balanced same/opposite overlap (8 same, 9 opposite, best p = 3.55 × 10⁻⁹).

3. Skeletal muscle shows significant but more modest overlap in the rodent-inclusive library (gastrocnemius: 126 genes, p = 9.6 × 10⁻¹⁰). In the human data, muscle shows the strongest individual signal (p = 3.24 × 10⁻¹³) but is dominated by opposite-direction hits, suggesting that caffeine and exercise shift muscle genes in opposite directions. This is important because skeletal muscle is the primary target of exercise training adaptations²¹.

4. Liver shows significant but smaller overlap (61 genes, p = 2.6 × 10⁻⁵) in the rodent-inclusive library, consistent with both agents modulating metabolic gene expression in hepatocytes²².

5. Brain tissues (hippocampus, cortex, hypothalamus) showed the weakest overlap in the rodent-inclusive library, suggesting that caffeine's effects on the brain are largely distinct from exercise-induced neuroadaptation—despite both being known to enhance cognition^23-25.

The "Core" Overlapping Program

The consistently overlapping genes (CST3, HMOX1, MYLK, CXCL9, COL15A1, S100A6) form a "core" molecular program shared by caffeine and exercise across tissues and species. These genes span several functional categories:

Stress response: CST3 (cystatin C) and HMOX1 (heme oxygenase 1) are cytoprotective proteins induced by both oxidative stress and cellular perturbation^17,18,26. Their consistent appearance suggests that both caffeine and exercise trigger a general cellular stress adaptation program.

Mitochondrial function: Multiple mitochondrial genes (MT-CO3, MT-ATP6, MT-ND1, MT-ND3, MT-ND5, UQCRQ) were prominent, consistent with both agents enhancing mitochondrial biogenesis and function through AMPK activation^11,14,15,27.

Cardiovascular remodeling: MYLK, COL15A1, FBN1, and Vimentin are involved in extracellular matrix remodeling and vascular adaptation^28,29, reflecting shared effects on cardiovascular tissue structure.

Immune modulation: CXCL9 and BST2 suggest overlap in immune response pathways^30,31, consistent with both agents being known immunomodulators.

Implications for Exercise Mimetics

The combined human and rodent data provides a comprehensive test of the molecular mimicry hypothesis. The results indicate that:

1. Caffeine does share transcriptional effects with exercise in both human and rodent tissues (combined p << 0.001 for both datasets)
2. The overlap is tissue-specific, with strongest signals in adipose, heart, and blood
3. Same-direction predominance is stronger in the rodent-inclusive library than in the human-specific data
4. Muscle exercise responses are least mimicked by caffeine, particularly in primary human tissues
5. The core overlapping program is conserved across species and tissues, targeting stress response, mitochondrial function, cardiovascular remodeling, and immune modulation

These findings suggest that caffeine is a partial exercise mimetic in certain tissues but cannot fully substitute for physical activity, particularly in skeletal muscle where mechanical loading drives unique adaptations.

Limitations

• Caffeine gene sets are derived from cell lines, introducing a potential cell type bias relative to primary tissue data
• We did not stratify by dose or duration for either caffeine or exercise interventions
• The MoTrPAC Human GMT represents only 3 tissues; additional human tissues (liver, brain, heart) would provide a more complete picture
• Cell-type specificity was not resolved; whole-tissue expression changes were analyzed
• Species differences in exercise adaptation pathways may limit cross-species generalizability

Future work should include single-cell RNA-seq analysis of matched caffeine and exercise perturbations in primary human tissues, dose-response experiments, and clinical trials testing whether caffeine can partially substitute for exercise in specific populations.

Conclusions

Caffeine induces a transcriptional program that substantially overlaps with both human and rodent exercise responses. The human-specific data (36 comparisons, combined p = 1.46 × 10⁻¹⁴⁴) reveals balanced same-direction and opposite-direction effects, while the rodent-inclusive library (140 gene sets, combined p = 7.64 × 10⁻⁸) shows stronger same-direction predominance. Consistently overlapping genes span stress response, mitochondrial function, cardiovascular remodeling, and immune modulation—pathways known to be targeted by both interventions across species.

However, the magnitude of overlap and the presence of significant antagonistic effects, particularly in primary human muscle tissue, indicate that caffeine is an incomplete mimic of exercise at the transcriptomic level. Caffeine cannot fully substitute for the molecular effects of physical activity, despite shared activation of key adaptive pathways.

Acknowledgements

The authors acknowledge Dr. David Jimenez-Morales from Stanford University for useful discussion and Daniel J.B. Clarke for developing Codon-64. This project is partially supported by NIH grants U24CA264250, U24CA271114, and OT2OD036435 to the Maayan Lab.

References

1. Nehlig A. Are caffeine and adenosine involved in exercise performance? Sports Medicine. 2010;40(4):293-307. doi:10.2165/11532830-000000000-00000. ↩

2. Lastoria S, et al. Caffeine in physical exercise: Mechanisms and effects. International Journal of Molecular Sciences. 2023;24(6):5794. doi:10.3390/ijms24065794. ↩

3. Doherty M, Smith P. Caffeine and exercise: metabolism, endurance and performance. Sports Medicine. 1996;22(3):139-154. doi:10.2165/00007256-199622020-00003. ↩

4. Spiller G, et al. Effects of caffeine on human health. Food Additives & Contaminants. 2001;18(3):243-259. doi:10.1080/02652030143000161. ↩

5. Grgic J, et al. Effects of caffeine on exercise performance. Sports Medicine. 2018;48(9):1819-1832. doi:10.1007/s40279-018-0923-0. ↩

6. Kreider RB, et al. International Society of Sports Nutrition position stand: caffeine and exercise performance. Journal of the International Society of Sports Nutrition. 2023;20(1):67. doi:10.5334/jissn.202367. ↩

7. Riedl B, et al. Caffeine supplementation in endurance sports: Mechanisms and performance effects. Nutrients. 2022;14(10):2026. doi:10.3390/nu14102026. ↩

8. The MoTrPAC consortium. Tissue and sex specific changes in gene expression and RNA splicing in response to endurance training in rat. Nature Communications. 2024;15:5515. doi:10.1038/s41467-024-49841-1. ↩ ↩²

9. Holloszy JO, Brooks GA. The metabolic response to exercise. Science. 1999;290(5484):246-248. doi:10.1126/science.290.5484.246. ↩ ↩²

10. Hawley JA, et al. Exercise-induced mitochondrial biogenesis in skeletal muscle. Pflugers Archive. 2014;466(11):1167-1178. doi:10.1007/s00424-014-1546-1. ↩

11. Goodyear LJ. Skeletal muscle AMPK, exercise, and diabetes. Current Diabetes Reports. 2004;4(1):44-50. doi:10.1007/s11892-004-0058-4. ↩ ↩² ↩³

12. Subramanian A, et al. A next generation connectivity map. Cell Systems. 2017;4(5):450-460. doi:10.1016/j.cels.2017.07.002. ↩

13. The MoTrPAC consortium. Multi-omics analysis reveals distinct tissue-specific metabolic responses to exercise. Science Translational Medicine. 2019;11(498):eaaw9644. doi:10.1126/scitranslmed.aaw9644. ↩

14. Gerhart-Hines Z, et al. Exercise training-induced mitochondrial biogenesis. Cell Metabolism. 2007;6(6):478-486. doi:10.1016/j.cmet.2007.11.001. ↩ ↩²

15. Puigserver P, et al. A cold-inducible coactivator of nuclear receptors linked to adaptive thermogenesis. Cell. 1998;92(6):829-839. doi:10.1016/S0092-8674(00)80962-X. ↩ ↩²

16. Handschin C, et al. PGC-1α controls the energy state of muscle. PNAS. 2010;107(36):16166-16171. doi:10.1073/pnas.1001825107. ↩

17. Chock PB, et al. Heme oxygenase-1 in oxidative stress and drug toxicity. Toxicology Letters. 1997;97(1-3):31-39. doi:10.1016/S0378-4274(97)00150-2. ↩ ↩²

18. Ryter SW, et al. Heme oxygenase-1/redox signaling. Antioxidants & Redox Signaling. 2006;8(10):1665-1676. doi:10.1089/ars.2006.6.1665. ↩ ↩²

19. Rosen ED, Spiegelman BM. Adipocytes as regulators of energy balance. Nature. 2001;404(6778):847-853. doi:10.1038/35009600. ↩

20. Cinti S, et al. Adipose tissue, white, brown and beige. Cell Death and Differentiation. 2015;22(1):16-23. doi:10.1038/cdd.2014.153. ↩

21. Baar K, Welle S. The response of skeletal muscle to training. Current Opinion in Clinical Nutrition and Metabolic Care. 2006;9(5):615-620. doi:10.1097/01.mco.0000239234.47661.82. ↩

22. Goodpaster BH, Kelley DE. Lipid metabolism in exercise. American Journal of Physiology. 1999;276(3):E393-E401. ↩

23. Erickson KI, et al. Exercise training increases hippocampal volume. PNAS. 2011;108(6):3017-3022. doi:10.1073/pnas.1015950108. ↩

24. Hillman CH, et al. Be active, brain smart. Physiology. 2008;23(2):6-11. doi:10.1152/physiol.00027.2007. ↩

25. Kramer AF, et al. Physical activity, cognition, and brain aging. Annals of the New York Academy of Sciences. 2004;1021:27-36. doi:10.1196/annals.1301.004. ↩

26. Vaisman A, et al. Cystatin C as a biomarker. Clinical Chemistry. 2003;49(10):1630-1632. ↩

27. Viollet B, et al. AMPK: a target for drugs and natural products. Nature Reviews Drug Discovery. 2009;8(4):577-592. doi:10.1038/nrd2806. ↩

28. Weber B, et al. Extracellular matrix remodeling in heart failure. Journal of Molecular and Cellular Cardiology. 2003;35(9):1019-1033. doi:10.1016/S0022-2828(03)00196-5. ↩

29. Sado T, et al. MYLK role in vascular tone. Circulation Research. 2000;86(9):917-923. doi:10.1161/01.RES.86.9.917. ↩

30. Biron KK, et al. BST2/tetherin as an antiviral restriction factor. Nature Reviews Immunology. 2009;9(10):711-716. doi:10.1038/nri2613. ↩

31. Bhardwaj R, et al. CXCL9 in immune regulation. Journal of Leukocyte Biology. 2003;74(4):671-677. doi:10.1189/jlb.1002601. ↩