Is caffeine an exercise mimicker at the cellular level? A transcriptomics deep dive

Abstract

Background: Caffeine can improve exercise performance, but performance enhancement does not establish that it reproduces exercise-induced cellular programs. We compared caffeine perturbation signatures with tissue-resolved exercise signatures from the Molecular Transducers of Physical Activity Consortium (MoTrPAC).

Methods: Two supplied gene-matrix-transposed files contained 36 directional human signatures and 225 rat signatures, including 206 directional and 19 unsigned consensus sets. Exact-name caffeine searches of Perturb-Seqr yielded 23 human-cell directional sets representing 13 exposure conditions across three libraries. Rat symbols were mapped to human orthologs using reciprocal unambiguous HCOP mappings supported by at least two resources. All exercise–caffeine comparisons were tested by one-sided Fisher/hypergeometric enrichment, with Benjamini–Hochberg correction across 5,566 directional tests. Enrichr analysis of Reactome and MSigDB Hallmark pathways characterized selected same-direction overlaps. Background sensitivity analyses restricted comparisons to shared orthologs and conditioned pathway tests on the input signatures.

Results: Human signatures produced 29 significant same-direction comparisons, versus 10 rat comparisons. The strongest match involved upregulated human skeletal-muscle genes 15–45 minutes after resistance exercise and an A375 caffeine signature: 30 shared genes, odds ratio 5.97, adjusted P = 1.34 × 10⁻⁹. Rat mimicry was concentrated in brown adipose, especially males after eight weeks. Same-direction findings favored upregulated genes: 28 of 29 human and seven of 10 rat comparisons. However, 25 human and 18 rat comparisons showed significant opposite-direction overlap. TNF/NF-κB-associated transcription was prominent in human muscle and blood; immune-system and interferon-response annotations characterized rat brown-adipose overlaps. Restriction to shared orthologs retained 17 human and three rat same-direction comparisons.

Conclusions: Caffeine signatures selectively resemble components of exercise-responsive transcription, most convincingly early human skeletal-muscle responses and a subset of rat brown-adipose responses. The evidence supports context-dependent transcriptional similarity rather than a universal exercise-mimicking effect. Differences in acute versus chronic exposure, gene-set sizes, tissue coverage, and cellular models prevent attributing the apparent human preference to species alone.

Keywords: caffeine; exercise mimetics; MoTrPAC; Perturb-Seqr; transcriptomics; skeletal muscle; brown adipose; inflammation; connectivity mapping.

Introduction

Exercise initiates molecular responses across organs, connecting local muscle activity to systemic adaptation. MoTrPAC was established to map these responses using coordinated molecular profiling¹. Its rat endurance-training study revealed extensive tissue- and sex-dependent remodeling². More recently, the initial human cohort characterized acute endurance and resistance exercise responses across skeletal muscle, adipose tissue, and blood³. Together, these resources make it possible to ask whether a chemical perturbation resembles particular components of exercise-induced transcription rather than treating exercise as one homogeneous signature.

Caffeine is a plausible candidate because it influences both performance and cellular metabolism. A randomized-trial meta-analysis found a small improvement in endurance running time-trial performance⁴. At customary exposure levels, adenosine-receptor antagonism is central to its pharmacology; higher concentrations can engage additional mechanisms, making dose an essential interpretive variable⁵. In C2C12 myotubes, caffeine enhanced oxidative metabolism and induced regulators of mitochondrial biogenesis, with effects partly dependent on PPARβ/δ⁶. These observations motivate transcriptomic comparison, but do not establish that caffeine duplicates the coordinated effects of exercise across tissues.

Connectivity mapping connects perturbations through shared gene-expression signatures^7,8. Perturb-Seqr extends this approach by integrating multiple chemical and genetic perturbation resources⁹. Its name should not be taken to imply that every included experiment uses single-cell Perturb-seq: the caffeine signatures analyzed here came from LINCS L1000, Ginkgo Bioworks, and NIBR DRUG-seq.

We therefore performed an exploratory, direction-aware comparison of the supplied MoTrPAC gene sets with caffeine signatures retrieved from Perturb-Seqr. The analysis addressed six related questions: which tissues show the greatest resemblance; whether human or rat exercise signatures match more strongly; whether induced or suppressed genes dominate; which pathways characterize the strongest overlaps; whether immune and inflammatory responses recur; and whether specific tissues favor concordance or reversal. Throughout, “mimicry” denotes statistically enriched same-direction gene-set overlap, not demonstrated equivalence in cellular function or organismal benefit.

Methods

Exercise signatures and study context

The human input, MoTrPAC_Human.gmt, contained 36 sets across blood, skeletal muscle, and adipose tissue: 18 upregulated and 18 downregulated sets. Labels encoded endurance or resistance exercise, post-exercise windows, and a delta_delta contrast against a non-exercise control. We interpreted these labels as control-adjusted changes rather than independently reconstructing the original differential-expression models. Available windows included 10 minutes, 15–45 minutes, 3.5–4 hours, and 24 hours, with tissue-dependent coverage.

The rat input, MoTrPAC_Endurance_Trained_Rats_2023.gmt, contained 105 upregulated, 101 downregulated, and 19 consensus sets across 19 tissues. Directional labels identified sex and one, two, four, or eight weeks of endurance training. The corresponding published MoTrPAC study used progressive treadmill training in adult Fischer 344 rats and collected tissues 48 hours after the last bout². The filename's “2023” label was retained as input provenance; the associated Nature article appeared in 2024. Consensus-set construction was not defined in the GMT, so these sets were treated as unsigned and excluded from directional conclusions.

Both GMT files supplied membership lists rather than expression matrices, effect sizes, replicate-level measurements, or complete tested-gene backgrounds. Differential-expression thresholds and the precise export provenance could not be recovered from those files. Analyses therefore used the supplied membership definitions without inventing additional fold-change or significance thresholds. Empty trailing GMT fields were removed, and each set was deduplicated.

Caffeine signature retrieval

Perturb-Seqr's GraphQL endpoint was queried on October 9, 2026 using geneSetTermSearch(terms:["caffeine"], first:1000). All 37 returned records were retrieved with gene memberships and library metadata; the returned count was checked against the reported total. We retained only records whose first ::-delimited field equaled “caffeine,” case-insensitively. Eleven records for 8-(3-chlorostyryl)caffeine were excluded because this derivative is a different perturbation. Three mouse records associated with GSE183323 were excluded from the primary analysis to keep the chemical-signature species consistent.

The resulting 23 sets comprised 14 LINCS L1000 CP, four Ginkgo Bioworks, and five NIBR DRUG-seq records. They represented 13 exposure conditions: 10 with both up and down sets and three with only an up set. Seven distinct cell-model labels were represented: A375, A549, HCC515, HEPG2, MCF7, and MCF10A.TP53.M in LINCS, A549 in Ginkgo, and U2OS in NIBR. Recorded exposures ranged from 0.01 to 100 μM and lasted 24 or 48 hours. Dose and model labels were accepted as database metadata, without independently auditing the underlying raw experiments. No caffeine signature was a matched exercise-tissue experiment.

Gene harmonization and analysis backgrounds

Human identifiers were resolved using Perturb-Seqr's geneMap2 function. Rat symbols were matched case-insensitively to the HCOP human–rat orthology table¹⁰. We retained relationships supported by at least two listed resources and unambiguous in both directions within that filtered table. This produced 15,839 symbol mappings, of which 15,597 human symbols were represented in Perturb-Seqr's background. The downloaded HCOP file carried a September 10, 2023 timestamp; its mappings therefore reflect that archived snapshot.

Unmapped or ambiguous rat identifiers were excluded rather than assumed to be orthologous on the basis of capitalization. The primary universe was the 22,285 human symbols in Perturb-Seqr's current background. All input sets were intersected with it. A sensitivity analysis repeated comparisons within the 15,597-symbol shared-ortholog universe. This assesses background and orthology sensitivity, but cannot substitute for unavailable tissue- and platform-specific assayed-gene universes.

Across directional set memberships, human mapping retained 12,881 of 13,598 entries (94.7%) and rat mapping retained 4,718 of 6,945 (67.9%). These counts sum memberships across signatures and do not represent unique genes or independent observations. Median mapped directional sizes were 539.5 genes for human up sets and 207 for human down sets, versus seven genes for each rat direction.

Directional overlap statistics

For every exercise set E and caffeine set C, we computed the shared count k, Jaccard similarity, the fraction of exercise genes shared, and the expected count |E||C|/N. One-sided Fisher enrichment was evaluated using the equivalent hypergeometric upper tail. The contingency table was:

	In caffeine set	Outside caffeine set
In exercise set	k	|E| − k
Outside exercise set	|C| − k	N − |E| − |C| + k

An up–up or down–down comparison was classified as same-direction overlap; up–down or down–up overlap was classified as reversal. Benjamini–Hochberg correction¹¹ was applied jointly across all 242 × 23 = 5,566 directional comparisons, including both species and both relations. Adjusted P values, denoted q, below 0.05 were considered significant under this background model. The 437 rat-consensus comparisons formed a separate unsigned testing family.

Where both directions existed in both datasets, we also tabulated the four overlap channels and calculated descriptive concordant and discordant totals. Missing directional sets were not interpreted as measured zero responses. These paired totals are descriptive and have no separate inferential P value. Comparisons across conditions, tissues, and species share genes and are not independent biological replicates; their significant-test proportions are descriptive, not species-level hypothesis tests.

Pathway enrichment

Pathway enrichment was performed through the Enrichr and Speedrichr APIs on October 9, 2026^12,13, using Reactome_2022 and MSigDB_Hallmark_2020^14,15. We selected the three lowest-q same-direction comparisons within each species and, separately, unions of genes from significant same-direction comparisons grouped by species, tissue, and exercise direction. This generated 15 overlap sets, with dependence and repeated biological content between individual and pooled sets. Submitted gene lists were checked against Enrichr's returned list memberships.

Each overlap list was analyzed against both libraries. Primary enrichment used Speedrichr's custom-background workflow with the 22,285-symbol Perturb-Seqr universe. Conditional analyses used the corresponding exercise or caffeine signature as background; pooled backgrounds were the union of all exercise sets for that tissue and direction or all retained caffeine sets in that direction. Standard Enrichr enrichment using its default background was also retained as a supplementary comparison. All enrichment P values, overlap genes, odds ratios, combined scores, and native adjusted P values came from the service responses.

For comparability across selected lists, we additionally applied Benjamini–Hochberg correction to the API-returned P values across all 15 × (1,818 Reactome + 50 Hallmark) = 28,020 terms, separately for each background mode. Terms omitted from a response were assigned P = 1. These global q values are reported in the text and figure; Enrichr's native adjusted P values, which apply within each query and library, are preserved separately in the supplementary results. These exploratory analyses characterize selected overlap content and do not remove all effects of selection or gene dependence.

Results

Early human skeletal muscle showed the strongest statistical resemblance

Of 414 human same-direction comparisons, 29 were significant, involving 13 of 36 exercise sets. The lowest q value occurred for the skeletal-muscle up set 15–45 minutes after resistance exercise and the A375 caffeine up set recorded at 0.01 μM for 24 hours. Their 30 shared genes gave an odds ratio of 5.97 and q = 1.34 × 10⁻⁹. The corresponding endurance-exercise muscle set shared 26 genes with the same caffeine set (odds ratio 5.04; q = 6.73 × 10⁻⁷).

These overlaps included immediate-response regulators and signaling-associated genes such as EGR3, KLF4, KLF6, NR4A1, SIK1, and ZC3H12A, alongside CX3CL1, CXCL2, TAP1, and other immune-associated genes. The resistance overlap also included VEGFA, PDK4, and RCAN1. Thus, the strongest correspondence extended across regulatory, immune-associated, vascular, and metabolic annotations rather than consisting solely of mitochondrial genes.

Blood also showed substantial matching. The endurance up set at 10 minutes shared 17 genes with the A549 LINCS caffeine up set at 10 μM for 24 hours (odds ratio 6.83; q = 3.48 × 10⁻⁶), including GNLY, GZMA, GZMB, NKG7, PRF1, and several chaperones. Human adipose had fewer significant comparisons, but large odds ratios in small early up sets: seven genes in the endurance overlap with A375 (odds ratio 19.89; q = 1.15 × 10⁻⁴), and five in the resistance overlap with A549 (odds ratio 25.38; q = 8.95 × 10⁻⁴). Consequently, muscle ranked first by statistical evidence and number of significant matching comparisons; adipose ranked highly by relative enrichment in a small subset of genes. These are different measures of “most.”

Table 1. Tissue distribution of significant directional comparisons. Counts are exercise-set × caffeine-set comparisons, not independent experiments. q values were adjusted across all directional tests. The unsigned consensus sets are excluded.

Species and tissue	Directional exercise sets	Same-direction tests	Significant same direction	Significant opposite direction	Best same-direction q
Human skeletal muscle	12	138	13	6	1.34 × 10⁻⁹
Human blood	14	161	12	19	3.48 × 10⁻⁶
Human adipose	10	115	4	0	1.15 × 10⁻⁴
Rat brown adipose	16	184	6	2	4.89 × 10⁻⁴
Rat white adipose	14	161	1	2	1.96 × 10⁻²
Rat lung	12	138	3	2	2.71 × 10⁻²
Other 16 rat tissues combined	164	1,892	0	12	—

Rat matching was concentrated in brown adipose and depended on sex and time

Ten of 2,375 rat same-direction comparisons were significant, spanning six of 206 directional exercise sets. Brown adipose accounted for six matches. Male brown adipose after eight weeks produced four significant up–up comparisons with caffeine sets from A375, A549, HEPG2, and MCF7. Its strongest comparison shared 15 genes with the A375 up set (odds ratio 4.79; q = 4.89 × 10⁻⁴), including C3, CD74, CSF1R, GPNMB, RAC2, TAP1, and TAPBP.

Female brown adipose showed a different pattern. Downregulated genes after one or two weeks matched the A549 caffeine down set recorded at 1.11 μM for 24 hours. The one-week overlap contained CCL20, CXCL9, PSMB9, and TAP1 (q = 0.0449); the two-week overlap contained CCL20, CCL5, CD44, KCNJ15, PSMB9, and TAP1 (q = 0.0196). Male late induction and female early suppression therefore both yielded same-direction caffeine overlaps, but in different exercise conditions.

Rat lung showed one significant down–down match in females after one week, involving ABHD5, BCL2A1, CD44, and SELE (q = 0.0271). Male lung after four weeks had two up–up matches, each involving only two genes, with q = 0.0499. Female white adipose after four weeks showed a three-gene up–up match—AQP3, CA2, and CRABP2—with A375 (q = 0.0196). These small overlaps warrant less biological weight than the larger muscle and brown-adipose results.

No directional rat mimicry passed q < 0.05 in adrenal, blood, colon, cortex, gastrocnemius, heart, hippocampus, hypothalamus, kidney, liver, ovaries, small intestine, spleen, testes, vastus lateralis, or vena cava. Sparse mapped sets restrict power, so these findings identify weaker evidence in this dataset rather than tissues proven unresponsive to caffeine.

Human signatures matched more frequently, but this was not an isolated species effect

The significant same-direction proportions were 29/414 (7.00%) in humans and 10/2,375 (0.42%) in rats. At the exercise-set level, 13/36 human sets (36.1%) and 6/206 rat sets (2.9%) had at least one significant match. Human muscle and blood drove much of this difference.

The direction of this contrast persisted in the shared-ortholog sensitivity analysis: 17 human and three rat same-direction comparisons remained significant. All three retained rat findings involved male brown-adipose upregulation after eight weeks. Human muscle remained the strongest match; its resistance overlap with A375 retained 23 genes and q = 2.71 × 10⁻⁴. The human adipose matches also persisted, whereas all down–down matches lost significance in this smaller universe.

Nevertheless, the human and rat datasets differed markedly in signature size, mapping retention, sampled tissues, and exercise timescale. Acute human exercise cannot be directly equated with chronic rat training. The supported conclusion is that the available human signatures matched the available human-cell caffeine signatures more strongly and more often; an intrinsic biological preference for human exercise has not been demonstrated.

Same-direction overlap was strongly biased toward induced genes

Twenty-eight of 29 significant human matches were up–up. Their test rate was 28/234 (12.0%), compared with 1/180 (0.56%) for down–down comparisons. The sole human down–down match involved muscle 3.5–4 hours after resistance exercise and Ginkgo A549 caffeine at 0.1 μM (21 genes; q = 0.0196).

Seven of 10 rat matches were up–up: 7/1,365 (0.51%) versus 3/1,010 (0.30%) for down–down comparisons. This bias was less pronounced than in humans, and the rat suppression matches were concentrated in early female brown adipose and lung. Unequal gene-set sizes and the 13-up/10-down composition of the caffeine collection affect these proportions, so they should not be interpreted as a universal property of caffeine pharmacology.

Figure 1. Percentage of comparisons significant under the primary background model, stratified by tissue and overlap relation (A), and by species and exercise direction for same-direction comparisons (B). Rat tissues with zero same-direction findings remain visible. Counts and percentages are descriptive because signatures are correlated.

Reversal accompanied mimicry, especially in human blood

There were 25 significant human and 18 significant rat opposite-direction comparisons. In human blood, reversals outnumbered matches, 19 to 12. One strong reversal connected blood upregulation 15–45 minutes after resistance exercise with A375 caffeine downregulation: 39 shared genes, q ≈ 1.18 × 10⁻⁶. Human muscle also showed six reversals, including later endurance responses.

Rat reversals occurred in testes, adrenal, blood, brown and white adipose, gastrocnemius, heart, kidney, and lung. Testes accounted for three; its strongest result involved the male two-week down set and MCF7 caffeine upregulation, with seven shared genes and q ≈ 1.91 × 10⁻⁵. Thus, some tissues had reversal evidence without significant directional mimicry.

Even a strong one-channel match did not imply complete directional reproduction. For early human resistance muscle versus A375, the paired totals were 30 concordant and 13 discordant memberships. For male eight-week brown adipose versus A375, they were 15 concordant and 12 discordant. The latter illustrates why a significant up–up overlap should be described as component-level similarity rather than whole-signature mimicry.

Immune and inflammatory programs characterized striking overlaps

The strongest individual human muscle overlap was enriched for the Hallmark TNF-alpha Signaling via NF-kB set: 16 of its 30 genes belonged to that annotation (global pathway q = 1.75 × 10⁻²¹). This finding survived both conditional backgrounds—q = 2.95 × 10⁻⁴ against the exercise set and q = 2.17 × 10⁻⁶ against the caffeine set—supporting preferential concentration of this program in the overlap rather than merely its presence in either input.

Pooled significant muscle-up overlaps contained 155 unique genes, including 37 TNF/NF-κB Hallmark members (q = 4.79 × 10⁻³⁸). Conditional q values remained 1.62 × 10⁻¹⁸ and 1.03 × 10⁻¹⁶. Pooled blood-up overlaps contained 150 genes, with 25 TNF/NF-κB members (q = 1.00 × 10⁻²⁰); this also survived both conditional backgrounds. Apoptosis-associated sets characterized both pooled muscle and blood overlaps, although annotation does not establish that apoptosis occurred. Muscle overlaps additionally implicated signal transduction and growth/remodeling-associated annotations.

Pooled rat brown-adipose-up overlaps contained 43 genes, of which 21 belonged to Reactome Immune System (q = 2.25 × 10⁻⁸). Interferon Gamma Response included eight genes (q = 5.98 × 10⁻⁶). These terms remained significant against the pooled caffeine-up background but not against the exercise background. The immune content is therefore a shared feature, while evidence that caffeine selectively captures it beyond the already immune-rich exercise signature is weaker. Complement and collagen-related annotations also appeared against the database universe, but failed the conditional-background checks.

The pooled early female brown-adipose-down overlaps contained eight genes. Four belonged to the TNF/NF-κB Hallmark set and four to Interferon Gamma Response; each had primary q = 1.98 × 10⁻⁴. Their exercise-conditioned q values were 0.00260 and 0.0360, respectively, while their caffeine-conditioned q values exceeded 0.05. Immune-related overlap can therefore occur through shared suppression as well as shared induction.

Figure 2. Selected pathway enrichments in pooled concordant overlap genes using Enrichr/Speedrichr with the Perturb-Seqr universe and conditional exercise or caffeine backgrounds. The dashed line marks q = 0.05. Conditional tests assess different null hypotheses and should not be treated as interchangeable estimates. Strong human TNF/NF-κB findings persist across backgrounds; broad rat immune annotations depend more strongly on background choice.

Discussion

This analysis supports selective transcriptional resemblance between caffeine perturbation signatures and exercise responses. The strongest evidence concerned early human skeletal muscle, with additional matching in blood and adipose. Rat findings were more concentrated, with late male brown-adipose induction providing the most reproducible directional signal. Upregulated genes dominated the human results, whereas rat results included both late induction and early female suppression.

The prominent TNF/NF-κB and immune annotations change the interpretation of caffeine as an exercise mimicker. They suggest convergence on regulatory and immune-associated responses alongside metabolic genes, rather than simple duplication of a mitochondrial adaptation program. A TNF/NF-κB signature contains response genes and feedback regulators; its enrichment does not demonstrate increased TNF protein, uniform inflammatory activation, or a beneficial or harmful inflammatory state. Similar caution applies to interferon and apoptosis annotations.

Bulk exercise-tissue profiles also combine changes within cells with changes in cell composition. The human blood overlap containing GNLY, granzymes, NKG7, and PRF1 could reflect redistribution of cytotoxic immune cells as well as altered transcription within them. The human MoTrPAC blood study independently reported exercise-mode differences in immune-cell responses¹⁶. In brown adipose, shared genes such as CD74, CSF1R, TAP1, and TAPBP are compatible with immune-related remodeling, but the gene sets alone cannot distinguish adipocyte-intrinsic effects from stromal or immune contributions. Matching cultured-cell caffeine signatures to these tissue profiles establishes shared genes, not a shared cell of origin.

The conditional pathway analysis reinforces this distinction. TNF/NF-κB concentration in the strongest human resistance-muscle overlap survived comparison with each parent signature. Broad immune-system enrichment in rat brown adipose did not survive conditioning on the exercise genes, suggesting that it partly inherited the immune-rich composition of that exercise set. This remains biologically informative, but provides weaker evidence for selective capture of immune remodeling by caffeine.

Cross-species interpretation is constrained by experimental design. Rat training signatures describe adaptation across weeks, whereas human signatures describe minutes to hours after a single bout. Moreover, rat mapping discarded many identifiers, and median directional rat sets contained only seven mapped genes. Human signatures were much larger. Statistical enrichment adjusts for list size under its null model, but detection power, selection thresholds, gene dependence, and mapping losses still affect comparisons. The apparent human preference should therefore be treated as a dataset-level result awaiting experiments that match species, tissue, dose, assay, and timescale.

The caffeine collection introduces further limitations. Twenty-seven of the 29 human matches and all 10 rat matches came from LINCS; two human matches came from Ginkgo, and none came from NIBR. Multiple matches to different LINCS cell models are not equivalent to replication across independent assay platforms. The leading signature came from A375 at a recorded concentration of 0.01 μM, much lower than several other included exposures. Its prominence does not establish dose responsiveness or justify extrapolation to a dietary dose. The exposure labels and underlying signature quality require examination in the source data before experimental prioritization.

The database-wide background is another limitation. Tissue- and platform-specific genes have unequal opportunities to enter a signature, particularly in a reduced-representation assay such as L1000⁸. Neither a database-wide universe nor its shared-ortholog subset fully represents those assay opportunities. Accordingly, the reported q values quantify overlap under an explicit exploratory model; they should be recalibrated using source-specific measured-gene universes when available. The GMT format also precludes rank-based connectivity, effect-size correlation, and adjustment for donor-level covariates. Closely related signatures and pooled pathway sets are dependent, and pathway selection followed observed overlap results.

Finally, cellular similarity must be separated from physiological substitution. Myotube experiments provide mechanistic support for caffeine-related metabolic changes⁶, but recent cell and mouse work also found attenuation of selected protein-synthesis and training-adaptation outcomes under the studied exposures¹⁷. The coexistence of matching and reversing transcriptional components here is consistent with a complex response. It does not establish that caffeine reproduces improvements in fitness, mitochondrial capacity, tissue repair, or long-term health.

A decisive follow-up would expose primary human myotubes, adipocytes, and relevant immune populations to a physiologically grounded caffeine dose series, alongside exercise-related stimuli and appropriate controls. Parallel rat models and matched early and late sampling would help disentangle species from exposure duration. Full differential-expression statistics, measured-gene backgrounds, cellular composition, respiration, cytokine outputs, and protein-synthesis endpoints would allow the proposed shared programs to be tested functionally.

Conclusions

Caffeine is a partial, context-dependent transcriptional exercise mimicker in the analyzed signatures. Early human skeletal muscle provides the strongest statistical overlap, while rat brown adipose—especially male tissue after eight weeks—provides the clearest rat signal. Human adipose shows high relative enrichment in small early induced sets, and human blood shows substantial matching alongside even more significant reversing comparisons.

The available human signatures match more strongly than the rat signatures, but this cannot be assigned to species alone. Concordance is biased toward upregulated genes, especially in humans. Immune and inflammatory programs recur through both induction and suppression, with robust TNF/NF-κB-associated overlap in human muscle and blood. These findings define testable cellular hypotheses; they do not establish caffeine as a replacement for exercise.

Acknowledgements

We thank Dr. David Jimenez-Morales of Stanford University for useful discussion and Daniel Clarke for developing Codon-64. We acknowledge the National Institutes of Health for funding this project through the CFDE Workbench grant OT2OD036435.

Data and code availability

The accompanying analysis archive contains the supplied exercise GMT files, retrieved caffeine memberships, Enrichr/Speedrichr requests and cached API responses, the HCOP snapshot, pathway libraries, mapping tables, analysis scripts, complete overlap results, and pathway results. The analysis can be rerun from these frozen inputs without requerying live services. Retrieval dates, input checksums, and software versions are included in the archive. No new expression measurements were generated.

References

1. Sanford JA, Nogiec CD, Lindholm ME, et al. Molecular Transducers of Physical Activity Consortium (MoTrPAC): Mapping the Dynamic Responses to Exercise. Cell. 2020;181(7):1464–1474. doi:10.1016/j.cell.2020.06.004. ↩

2. MoTrPAC Study Group; Lead Analysts; MoTrPAC Study Group. Temporal dynamics of the multi-omic response to endurance exercise training. Nature. 2024;629:174–183. doi:10.1038/s41586-023-06877-w. ↩ ↩²

3. MoTrPAC Study Group, Katz DH, Jin CA, et al. Multi-Omic, Multi-Tissue Responses to Acute Exercise in Sedentary Adults: Findings from the Molecular Transducers of Physical Activity Consortium. bioRxiv. 2026. Preprint. doi:10.64898/2026.02.27.702183. ↩

4. Wang Z, Qiu B, Gao J, Del Coso J. Effects of Caffeine Intake on Endurance Running Performance and Time to Exhaustion: A Systematic Review and Meta-Analysis. Nutrients. 2023;15(1):148. doi:10.3390/nu15010148. ↩

5. Fredholm BB, Yang J, Wang Y. Low, but not high, dose caffeine is a readily available probe for adenosine actions. Molecular Aspects of Medicine. 2017;55:20–25. PubMed PMID:27915051. ↩

6. Schnuck JK, Gould LM, Parry HA, et al. Metabolic effects of physiological levels of caffeine in myotubes. Journal of Physiology and Biochemistry. 2018;74(1):35–45. doi:10.1007/s13105-017-0601-1. ↩ ↩²

7. Lamb J, Crawford ED, Peck D, et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. Science. 2006;313(5795):1929–1935. doi:10.1126/science.1132939. ↩

8. Subramanian A, Narayan R, Corsello SM, et al. A Next Generation Connectivity Map: L1000 Platform and the First 1,000,000 Profiles. Cell. 2017;171(6):1437–1452.e17. doi:10.1016/j.cell.2017.10.049. ↩ ↩²

9. Gardner JK, Taub LD, Clarke DJB, Diamant I, Ma'ayan A. Perturb-Seqr: Comprehensive Signature Search Engine Integrating Connectivity Maps from Multiple Sources. Web resource. Perturb-Seqr. Accessed October 9, 2026. ↩

10. Yates B, Gray KA, Jones TEM, Bruford EA. Updates to HCOP: the HGNC comparison of orthology predictions tool. Briefings in Bioinformatics. 2021;22(6):bbab155. doi:10.1093/bib/bbab155. ↩

11. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B. 1995;57(1):289–300. Author-hosted article. ↩

12. Kuleshov MV, Jones MR, Rouillard AD, et al. Enrichr: a comprehensive gene set enrichment analysis web server 2016 update. Nucleic Acids Research. 2016;44(W1):W90–W97. doi:10.1093/nar/gkw377. ↩

13. Xie Z, Bailey A, Kuleshov MV, et al. Gene Set Knowledge Discovery with Enrichr. Current Protocols. 2021;1(3):e90. doi:10.1002/cpz1.90. ↩

14. Gillespie M, Jassal B, Stephan R, et al. The Reactome pathway knowledgebase 2022. Nucleic Acids Research. 2022;50(D1):D687–D692. doi:10.1093/nar/gkab1028. ↩

15. Liberzon A, Birger C, Thorvaldsdóttir H, Ghandi M, Mesirov JP, Tamayo P. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Systems. 2015;1(6):417–425. doi:10.1016/j.cels.2015.12.004. ↩

16. MoTrPAC Study Group, et al. Blood Biochemical Responses to Acute Exercise: Findings from the Molecular Transducers of Physical Activity Consortium (MoTrPAC). bioRxiv. 2026. Preprint. Full text and publication metadata. ↩

17. Steffen D, et al. Caffeine decreases muscle and tendon protein synthesis and engineered ligament strength in vitro and attenuates adaptation to exercise in mice. Journal of Applied Physiology. 2025;139(6):1569–1580. doi:10.1152/japplphysiol.00512.2025. ↩