Fasting and Human Muscle: The Age-Related Gene Signal We Found

What We Measured in Human Muscle

An Age Reference Built Separately from Fasting

To ask whether fasting resembles an age-associated pattern, we needed to define that pattern without using the fasting responses. One study compared muscle from 12 younger and 12 older men. The other examined age across 53 people, taking sex into account. Both compared people of different ages; neither followed the same people as they grew older. [1,2]

The two analyses shared 10,603 uniquely mapped genes. From these, we retained 1,074 whose expression pointed in the same nonzero direction in both studies and had sufficient statistical support in at least one, not necessarily both. The reference contained 504 genes expressed more highly with age and 570 expressed less highly. The exact selection rule appears in the Method summary.

We then kept the gene list and its directions fixed. This is what “frozen reference” means throughout the article. A gene could not enter the reference because it happened to make the fasting response look more age-aligned. Table 1 shows how the other datasets were used to examine that independently defined pattern.

Table 1. The nine datasets and their roles. People and samples are shown separately. Only the first two datasets constructed the age reference; the others did not select its genes or directions. These studies are not nine replications of fasting. [1–9]

Scroll horizontally to read all four columns.

Dataset	Question or role	People and samples	Comparison and key constraint
GSE175495	Build the age reference	24 men; 24 samples	12 older versus 12 younger people; cross-sectional.
GSE164471	Build the age reference	53 people; 53 samples; corrected metadata: 33 men and 20 women	Continuous age per decade, adjusted for sex; cross-sectional.
GSE55924	Fasting discovery	12 men; 48 samples	Nominal 4/10/24 h versus 1.5 h; the paper reports the first biopsy at 105 min.
GSE28016	Fasting/refeeding validation	Seven people; 14 paired samples	40 h FASTED minus six h REFED; a different comparator from discovery.
GSE252357	Fed specificity control	Five people; 24 samples	Repeated biopsies versus PRE; nominal 8 h sample missing for one person; external control.
GSE108539	Independent rhythmicity sensitivity	Ten donors; 57 samples	Six scheduled clock times; three samples absent in donor 7; not a fasting replication.
GSE25941	Test age-reference transport	36 people; 36 baseline samples	21 older versus 15 younger people, adjusted for sex; no fasting contrast.
GSE8479	Age and training challenge	51 people; 65 arrays; 14 training pairs	Baseline age model; training POST minus PRE fully confounded with chip batch.
GSE129843	Eating-window challenge	11 men; 125 samples; nine eligible for joint A test	TRF/URF crossover; seven biopsies absent; period/order unavailable; not acute fasting validation.

Reading the Alignment Score

The score, called A, asks a straightforward question: did expression, on balance, move toward the directions associated with age in the independent reference?

For a gene expressed more highly with age, an increase points toward the reference. For a gene expressed less highly with age, a decrease points toward it. Changes in the reverse directions contribute the other way. A summarizes this balance across the measured genes, using changes within the same participant.

That makes A an expression-alignment score, not a measurement of biological age. The equation below uses log₂, the expression scale on which the changes were analyzed.

A positive score does not require every gene to follow the age-associated direction. Different combinations of changes can produce the same sign. Conversely, a score near zero can conceal substantial changes that balance one another. This is why we keep both the participants’ individual patterns and the broader biological programs in view.

The samples also represent muscle tissue as a whole, often called bulk muscle. They contain a mixture of muscle fibers and other cells, rather than isolated fibers. An expression change can arise within cells, from differences in the abundance of cell types, or from both. These data do not identify the cellular source of every change.

Nine Datasets with Different Jobs

The studies let us ask several questions, not repeat one question nine times. Two built the age reference. One supplied the fasting discovery and another the separate fasting/refeeding comparison. The remaining datasets asked whether the sign also appeared when people were fed, whether removing rhythmic genes changed the result, whether the age reference worked elsewhere, and what happened in other experimental settings.

The technologies used in these studies did not measure exactly the same genes. We calculated each score from the genes that study actually measured and also checked the 711 genes measurable across discovery, validation and the fed control. Missing genes were not filled in.

That preserves a way to compare directions while recognizing the differences between studies. Their score sizes are not interchangeable: the measurements, processing and comparison states differ. A larger number in one study is not evidence that its participants’ muscles aged more.

A Clear Discovery Signal and an Imprecise Validation

Following Twelve Participants

The discovery study followed 12 men through four muscle biopsies, yielding 48 expression profiles. Each man’s later samples were compared with his own early sample. The deposited post-meal labels are 1.5, 4, 10 and 24 hours. We retain those labels, although the publication places the first biopsy at 105 minutes rather than exactly 1.5 hours. [3]

At the final point, all 12 participants had moved in the age-reference direction. Figure 1 shows that agreement without hiding the differences in the size of their responses. Mean A was 0.04030, with a 95% confidence interval of 0.02458 to 0.05601. The entire interval is positive: allowing for variation among these men, the estimated average still points toward the reference.

The statistical analysis strongly supports a positive average shift within this protocol under its stated assumptions. Its evidence comes from the participants, not from treating hundreds of genes as hundreds of independent people.

The earlier samples show that the participants did not all follow the same path to that final direction. Alignment was already detectable at four hours. At ten hours, the average was similar to the four-hour value, but responses were more scattered and the estimate was less precise. The final agreement therefore sits alongside variation over time, not a uniform rise in every person.

There is no evidence here for a switch that turns on at precisely 24 hours. No threshold was estimated, and even the increase from ten to 24 hours did not meet its statistical criterion after accounting for the temporal comparisons. The study does not establish a smooth relationship in which more hours without food produce progressively more alignment.

The resemblance also has a defined scope. A focuses on genes selected for consistent age associations. When we compared expression across the broader set of measurable genes in the age reference, the resemblance was much weaker. A coherent overlap in a targeted gene set is not the same as the whole expression profile becoming age-like.

A Different Fasting and Refeeding Comparison

Would the same direction appear in another fasting setting? The validation study provided paired measurements from seven people, comparing muscle after 40 hours of fasting with muscle six hours after refeeding. The analyzed difference was FASTED minus REFED. This was a related question, but not a second group following the discovery schedule. [4]

Here the individuals did not agree. Four had positive A and three had negative A. The average, 0.01713, pointed in the discovery direction, but its 95% interval ran from −0.01028 to 0.04455, spanning both negative and positive values. The positive average is therefore not a precise independent confirmation.

That is also different from showing that the discovery signal never appears in other fasting settings. This small cohort changes several things at once: fasting duration, the comparison against refeeding, and the measurement technology. We cannot assign the difference between studies to any one of them. Combining the two groups into a single 19-person average would hide those distinctions rather than resolve them.

The discovery remains consistent within its own protocol, with uncertain extension to this second comparison. The next question is not how large the score is, but whether its positive direction identifies fasting specifically.

The Fed Control Changes the Interpretation

A Positive Shift without the Fasting Protocol

A positive score also appeared in muscle from people who were fed. These five participants belonged to the biopsy-only control group of a separate study; its exercise group was excluded from our analysis. Their series contained 24 samples, with one biopsy missing at the nominal eight-hour point. [5]

At the last point, all five scores were positive. Mean A was 0.06027. The sample label is nominally 24 hours, while the documented elapsed time is 25 hours after the first biopsy, labeled PRE. Figure 2 shows the individual paths and the broad uncertainty around their average.

The key observation is already clear: the age-reference direction appeared without the fasting protocol. A positive A, by itself, cannot identify a fasting-specific response.

Five people provide little precision about the size of an average effect. The shared direction still matters. The formal test has particularly coarse resolution at this sample size, so its failure to cross the usual significance threshold is not a reason to discard what all five participants show. Equally, the positive confidence interval does not make this a precisely measured effect. The technical note explains why the interval and test can give apparently different impressions.

What This Control Can Resolve

The fed group changes what can be attributed to fasting, not the discovery values. It shows that the same positive direction can appear in another physiological and experimental setting. It does not establish that the two settings produce equal effects.

This was an external study, not a group randomly assigned to be fed alongside the discovery participants. The people, timing, measurement scale and measured genes differ. The fed group’s numerically larger average therefore cannot be read as a greater aging effect, nor subtracted from the discovery average to isolate a “corrected” fasting effect.

Time of day, meals, repeated sampling and changes in the tissue being sampled could each contribute to the fed pattern. The study does not separate them. In particular, the biopsy-only design does not demonstrate that biopsies caused the alignment.

What it does make necessary is a closer biological comparison. If the scores point the same way in fasting and fed muscle, do the expression programs underneath them also look alike?

Shared Directions Do Not Mean a Shared Biological State

The Oxidative Phosphorylation Overlap

One of the clearest overlaps involved genes linked to oxidative energy metabolism. The group called oxidative phosphorylation, or OXPHOS, concerns mitochondrial machinery associated with oxidative energy production. Its expression pattern pointed downward in both age-reference studies and in discovery at 24 hours.

To examine that overlap alongside the rest of the biology, we used all 50 gene sets in the frozen MSigDB Hallmark collection. Each set groups genes associated with a process or response, such as oxidative phosphorylation or interferon signaling. These are collections of genes, not the conceptual “hallmarks of aging.” [10]

For each group, we asked whether its genes tended toward higher or lower expression relative to the other genes measured in that comparison. The statistical method used for this question is CAMERA. [11] Figure 3 brings those results together across the age references, fasting studies and fed control.

The shared OXPHOS direction gives the age-alignment finding a concrete metabolic context. The discovery response partly resembles age-associated expression in genes linked to oxidative metabolism. It is not simply a score carrying an age-related label.

But the measured overlap is in RNA expression. It is not a measurement of respiration, ATP production or mitochondrial damage, and it does not require every OXPHOS gene to decrease uniformly. The same expression direction does not establish the same mitochondrial function.

The other cohorts also show where this pattern stops being consistent. OXPHOS pointed upward in the fasting/refeeding comparison and in the late fed control, but neither upward result had sufficient statistical support after correction. Those observations do not establish an opposite functional response. They show that the strongly supported downward pattern in discovery was not established across all three settings.

Genes involved in fatty acid metabolism provide a more partial overlap. Their expression pattern pointed downward in discovery at 24 hours and in both age references, but the statistical support was sufficient in only one age reference. This concerns genes associated with handling fatty acids, not a direct measurement of fat oxidation. It cannot tell us that muscle burned less fat or failed to change fuel use during fasting.

Interferon Responses Separate the Contexts

The immune-related programs tell a different part of the story. At 24 hours in discovery, genes associated with the interferon alpha and gamma responses pointed upward. Neither response had a statistically supported match in either age reference. In the late fed control, both pointed downward, with statistical support.

Discovery and the fed control both had positive A, yet their interferon expression patterns went in different directions. Their shared score sign did not describe one uniform state of muscle. Looking at the gene programs reveals a distinction that the score alone leaves out.

The age references sharpen that point further. The interferon changes seen in discovery were not a statistically supported feature shared with those references. The overlap with age was selective: the oxidative-metabolism expression pattern aligned more clearly than these immune-related programs did.

A group of genes associated with inflammatory signaling, TNF signaling through NF-kappa B, also pointed upward in discovery. It had sufficient statistical support in only one age reference. That is evidence about an inflammatory expression program, not a diagnosis of pathological inflammation or the age-related inflammatory process called inflammaging in the fasting participants.

These comparisons describe patterns from separate studies. We did not directly test differences between the studies’ effects, or calculate how much OXPHOS, interferon or any other pathway contributed to A. The pathway results help interpret the score; they are not a breakdown of what caused it.

What Enrichment Can and Cannot Measure

The term “enrichment” refers here to evidence that a group of related genes tends toward higher or lower expression compared with the other measured genes. It helps identify organized patterns that a single score would obscure. It does not turn an expression pattern into a measurement of the biological process itself.

The figure’s colors follow that distinction. Red and blue show direction. Darker color reflects stronger statistical support, not a larger expression change or a more active pathway. Beyond the display’s cap, still stronger statistical results receive the same saturated color.

A pathway name can otherwise invite a conclusion before its evidence has been read. The mTORC1 signaling set pointed upward at 24 hours, but without convincing statistical support. It established neither activation nor inhibition of mTORC1, and autophagic flux was not measured. The analysis therefore cannot support a chain from fasting through mTOR inhibition and autophagy to rejuvenation.

The biological finding is the combination of overlap and divergence: a shared oxidative-metabolism expression direction, a less consistent fatty-acid-metabolism overlap, and immune-related programs that distinguish the settings. All 50 groups remain visible in the figure, so the examples can be judged against the complete pattern. The next step is to ask whether the discovery alignment survives changes to how it is measured.

What the Sensitivity Analyses Leave Standing

Removing Predefined Rhythmic Genes

Could genes that change across the day be creating the signal? In the discovery protocol, time without further food, clock time and the history of biopsies all advance together. To examine one part of that problem, we removed genes identified as rhythmic in an independent muscle time series and checked whether the alignment remained.

That series contained 57 samples from ten donors, with three scheduled samples missing from one donor. We used three separately defined lists of rhythmic genes: one estimated independently from the series and two based on distinct classifications supplied by the original authors. None was selected according to how genes behaved during fasting. [6]

The final discovery alignment remained positive after all three exclusions and retained statistical support under each definition’s correction. Mean A ranged from 0.03057 to 0.03534. Figure 4 shows the result alongside the different gene counts and earlier-time patterns, rather than presenting the exclusions as identical tests.

These checks make a specific explanation less persuasive: the discovery result did not disappear when those predefined rhythmic genes were removed. They do not separate the effects of clock time, time without food and repeated biopsies. Which genes qualify as rhythmic also depends on the source data and model. Removing a gene list is not the same as experimentally controlling when the samples were taken.

Changing the Reference Estimate

Could the result depend on one particular construction of the age reference? We repeatedly rebuilt that reference to ask whether a different estimate would change the answer, while keeping the observed fasting data fixed.

Figure S1 shows the first 1,000 accepted reconstructions. Discovery alignment at 24 hours stayed positive in 997 of them. The fasting/refeeding average stayed positive in all 1,000. The direction therefore generally survived this way of rebuilding the reference.

The distinction between rebuilding a reference and measuring more people is essential here. The measurements from the seven validation participants stayed fixed. A consistently positive average under changes to the reference does not resolve the uncertainty associated with variation among those people or turn their imprecise result into a precise confirmation. The distributions in Figure S1 describe uncertainty about the reference, not the participant confidence intervals in Figure 1.

Other checks asked whether the result depended on the choice of genes or how their changes were summarized. The final discovery alignment remained positive after the specified removal of the largest contributors, after matching genes on baseline expression, and when using the genes common to the core datasets. These checks challenge particular analytical explanations; none adds an independent fasting participant.

The Limits That Remain

Some additional demographic sensitivity analyses failed their statistical checks, so they cannot be used to claim that the result is robust across demographic adjustments. These analyses did not define the primary reference, and their technical limitations are documented in the Methods.

The discovery signal has survived several defined challenges, not every possible one. With that distinction in place, the next question concerns the reference itself: does it recognize age-associated expression in people outside the studies that built it?

What the Other Cohorts Add

An Independent Age Comparison

A reference built from age differences should also be tested in another age comparison. GSE25941 supplied baseline muscle samples from 36 adults: 15 younger and 21 older. Although its source publication also studied exercise, the comparison used here concerns age before that intervention. [7]

The reference distinguished the older from the younger group in the expected direction, taking sex into account. Across the 1,028 measured signature genes, approximately 91.9% pointed in the direction defined by the original reference. Figure 5 shows both the overall age-group difference and the individual gene effects.

That agreement concerns genes, not extra participants. The study still contains 36 people. Its uncertainty estimate comes from resampling those people while keeping the reference fixed, unlike the earlier exercise of rebuilding the reference while keeping the fasting observations fixed.

The result supports an age-associated distinction beyond the two datasets that defined it, and on another measurement platform. It strengthens the meaning of the reference used to interpret discovery. It does not validate fasting or supply a conversion from A into years of age.

Why the Training Result Stays Qualified

A different question is what happens to A in a muscle-remodeling setting. GSE8479 contains an age comparison and paired measurements before and after training: 65 expression profiles from 51 people, including 14 older participants with both measurements. [8]

In those pairs, 13 of 14 scores were negative, indicating an expression shift opposite to the age-reference direction, but every post-training sample was measured on chip batches absent from baseline. Figure S2 shows the negative average, −0.03272, together with that constraint.

The biological intervention and the laboratory measurement batch changed together. There is no way in this comparison to separate their contributions to the before-and-after difference. Following the same people does not solve that problem, and a strong statistical result cannot solve it either.

The separate age comparison was positive after taking sex and chip into account. That adjustment belongs to a different comparison; it does not rescue attribution of the paired shift to training. The negative score remains an observation, but not evidence that exercise specifically reversed the signature or rejuvenated muscle.

A Different Test of Eating Time

Returning to food timing, does changing the daily eating window produce a clear difference in A? Eleven men followed both schedules in a crossover study: five days with an eight-hour eating window and five days with a fifteen-hour window. The shorter schedule is labeled TRF, for time-restricted feeding. We retain URF for the extended schedule, as in the deposited data; the paper calls it EXF. This is a different experiment from the acute 24- or 40-hour fasting comparisons. [9]

The upper panel of Figure 6 follows each schedule from its own 07:00 sample. Later expression is compared with the same man’s expression at 07:00 in that condition. The lower panel then compares those changes between schedules within the same person. Starting at the same clock time does not mean the men had fasted for the same duration in both conditions.

The comparison did not establish a precise difference between schedules in this score. The overall test used nine eligible participants, and none of the individual later-time differences remained statistically supported after the planned correction. That is not a finding that the schedules had identical effects: equivalence was not tested.

The missing samples help explain why the figure requires careful reading. There were 125 observed samples and seven missing biopsies, including two condition-specific 07:00 starting samples. Without a starting sample, that man’s corresponding within-condition change cannot be calculated. Comparing schedules requires both starting samples and both later samples; the eligible group therefore changes across times. Missing values were not filled in.

For the same reason, subtracting the two upper averages will not necessarily reproduce the lower panel. The upper averages and paired differences can contain different participants. Individual records needed to adjust for the order of the schedules and the study period were also unavailable.

The result for A does not mean gene expression was unchanged. Figure S3 preserves the complete external pathway comparison, with 50 programs across 24 contrasts. Expression directions and statistical support vary across the age comparisons, training and eating-time settings. Changes in particular genes, coordinated gene programs and the overall age-alignment score answer different biological questions.

What This Can Tell Us about Fasting and Muscle Aging

From Expression Patterns to Mechanisms

The investigation leaves a coherent discovery result, with a clearer account of what it resembles. Every participant’s final score pointed toward an age reference that had been built without the fasting responses. The shared downward OXPHOS expression pattern places part of that resemblance in oxidative metabolism. The tested changes to gene selection and reference construction did not simply make it vanish.

The other observations narrow the explanation. A positive average in the separate fasting/refeeding study remained imprecise. Positive scores in the fed group showed that the direction was not unique to fasting. And the interferon results showed that a common score direction could accompany different expression programs.

Taken together, this is more specific than a global “aging response.” Muscle in an acute experimental setting can share some age-associated expression features without reproducing the whole age-associated pattern. The metabolic overlap and immune-related divergence both belong in that description. Neither is an inconvenience to be explained away.

What produces the overlap remains unresolved. Nutritional state, clock time, repeated sampling and the mixture of cells present in the biopsy are possible contributors. These explanations could operate together; the data do not identify which dominates. Recognizing the expression pattern and identifying its cause are different achievements. This analysis reaches the former without claiming the latter.

Prior work already connects muscle gene expression with age, nutrient sensing and daily rhythms. The original fasting and circadian studies provide part of that context. [3,6] Malhan and colleagues also examined muscle expression in relation to aging and circadian regulation, but their fasting component used mouse data, not a human replication of this comparison. [19] The contribution here is the combination of an independently built human age reference, within-person fasting comparisons and tests that constrain specificity, not a claim to be the first to connect these subjects.

What Would Distinguish the Explanations

The remaining questions point to particular missing evidence. A fed group studied alongside the fasting participants, with balanced clock times and sampling schedules, would help separate food exposure from the shared progress of the protocol. A more precise independent fasting comparison would clarify how consistently the alignment appears elsewhere. These are requirements suggested by the current findings, not additional analyses performed here.

Measurements that distinguish cell types would help locate the source of the expression changes. Do they occur within particular muscle fibers, or reflect contributions from other cells in the biopsy? Measurements of muscle function would address a different question: whether those expression changes accompany a change in what the tissue does. A downward oxidative-metabolism expression pattern cannot stand in for measured respiration, and an immune-related pattern cannot stand in for demonstrated tissue injury.

The course of the response also matters. An acute observation does not reveal how long a change lasts. Describing it as a harmless transient adaptation would require evidence of both its course and its consequences. The fasting/refeeding study did not follow recovery of the same 12 discovery participants, so it cannot establish that their signal reversed after eating.

The age reference has a parallel limitation. Differences between younger and older people are not a record of how one person’s muscle changed over a lifetime. Recognizing those differences in a third age cohort strengthens the reference but leaves that longitudinal question open. This is why a biological-age score and a demonstrated change in aging require different evidence.

The Practical Boundary

Nothing in these results selects a fasting duration or provides a reason to start, prolong or abandon a fast. The times describe sampling comparisons, not durations shown to be safe or beneficial. A was not evaluated as a measure of individual benefit, risk or treatment guidance.

The same standard applies to reassuring interpretations. Muscle adaptation is useful context for changing expression, but benefits in strength, muscle mass, metabolic outcomes or long-term function require those outcomes to be measured. This analysis did not measure them, and it does not negate evidence from studies that did.

For a reader assessing a molecular aging claim, the useful questions are concrete: what tissue was measured, what was it compared with, did the people respond similarly, and do the biological programs and functional measurements support the proposed meaning? Here, those questions preserve the discovery while keeping its interpretation tied to the muscle data.

What the Analysis Establishes

At the final point of the discovery fasting protocol, every participant showed muscle gene-expression changes that aligned, on balance, with the independently built age reference. The resemblance included genes linked to oxidative energy metabolism and remained visible through the reported checks of gene selection and reference construction.

The other studies made its meaning more specific. The separate fasting/refeeding result was positive on average but imprecise. Fed participants also showed positive alignment. And similar score directions accompanied different underlying expression programs, rather than one common biological state.

The data therefore support a partial and context-dependent resemblance in muscle gene expression, not a demonstrated change in biological age or a clinical benefit or harm. The finding survives as a defined expression overlap. What becomes narrower, and more informative, is the explanation: the pattern cannot be understood apart from the setting in which it appears.

Method summary

We performed an original cross-dataset analysis using nine public human skeletal-muscle datasets from GEO, with the roles listed in Table 1. Two studies built the age reference; the others supplied fasting discovery, fasting/refeeding validation, a fed serial-biopsy control, independent rhythmicity information, a third age comparison, an age/training challenge and an eating-window crossover. We kept their different questions separate rather than pooling them into one effect. The original studies are credited in the references. [1–9]

The reference was built without using the fasting responses. We identified genes whose age-associated expression pointed in the same direction in both reference studies, with statistical support in at least one. We then fixed that list and its directions. For each paired comparison, A summarized whether expression moved toward or away from those directions within the same person. Only genes actually measured in that dataset contributed to its score.

We also examined groups of genes associated with biological programs, asking whether they tended toward higher or lower expression relative to the other measured genes. The sensitivity analyses asked whether the discovery result depended on rhythmic genes, particular contributors, different representations of the measurements, gene coverage, baseline expression or the estimated age reference. These checks addressed specified choices in the analysis, not new fasting experiments.

Technical implementation. The age analyses used expression filtering, TMM normalization and robust edgeR quasi-likelihood models. [12,13] The shared universe contained 10,603 uniquely mapped genes. Selection required concordant nonzero age-effect directions and Benjamini-Hochberg false discovery rate (BH FDR) < 0.05 in at least one reference, yielding 1,074 genes: 504 age-up and 570 age-down. No other cohort selected the reference genes. Discovery measured 887 of them, validation 832 and the fed control 1,037; their common-feature sensitivity used 711.

Data preparation followed each platform and deposited scale: log₂ transformation followed by quantile normalization for discovery arrays; joint raw-CEL RMA for the two Affymetrix cohorts; and the specified log-scale treatment for already normalized matrices. Participant-aware limma models represented repeated observations and paired comparisons. [14–16] Missing biopsies and unavailable crossover period/order information were not imputed.

A was half the difference between the mean log₂ changes in measured age-up and age-down genes, with separate measured denominators. Participants were the units of inference for paired scores. Pointwise 95% t intervals described uncertainty around participant means; exhaustive two-sided sign-flip tests evaluated the absolute group mean under reversals of participant score signs. The predefined comparison groups used distinct BH correction families, not one global correction over the analysis. [17,18] The joint crossover A test kept one sign per participant across available later times and was separate from the gene-level condition-by-time interaction. Its figure note retains the specific upper-tail test definition.

All 50 frozen MSigDB Hallmark sets were examined with competitive CAMERA tests and correction within each contrast. [10,11] Gene exclusions, alternative feature representations, common coverage, contributor trimming and baseline-expression matching tested the specified analytical choices. Their distributions do not replace uncertainty among participants.

The two bootstrap procedures varied different inputs. For the third age cohort, participants were resampled while the reference stayed fixed. For age-reference reconstruction, the two reference studies were rebuilt while the observed fasting cohorts stayed fixed, retaining the first 1,000 accepted attempts. That second procedure describes uncertainty in the reference estimate, not a joint interval incorporating fasting-participant sampling. Neither procedure is interchangeable with the participant t intervals or the gene-selection sensitivities.

Two sensitivities adjusting for the recorded race variable produced invalid quasi-likelihood statistics. Restricting one downstream calculation to 1,070 genes with finite statistics did not validate the complete adjusted analysis. These variants did not define the primary reference. Complete per-fit bootstrap model diagnostics were not saved; accepted and rejected attempts remain documented, but those missing diagnostics limit retrospective verification of each fit.

The accompanying Full Methods describe the data processing, statistical models, sensitivity analyses and resampling procedures used in this analysis. Analysis scripts and associated scientific materials are provided in the Scientific Supplement, version 2.0.

References

1. Trim et al. (2022 (online 2021)). Divergent immunometabolic changes in adipose tissue and skeletal muscle with ageing in healthy humans. doi:10.1113/JP280977.
2. Tumasian et al. (2021). Skeletal muscle transcriptome in healthy aging. doi:10.1038/s41467-021-22168-2.
3. Wijngaarden, Bakker et al. (2014). Regulation of skeletal muscle energy/nutrient-sensing pathways during metabolic adaptation to fasting in healthy humans. doi:10.1152/ajpendo.00215.2014.
4. Kunkel et al. (2011). mRNA expression signatures of human skeletal muscle atrophy identify a natural compound that increases muscle mass. doi:10.1016/j.cmet.2011.03.020.
5. Edman et al. (2024). The 24-hour molecular landscape after exercise in humans reveals MYC is sufficient for muscle growth. doi:10.1038/s44319-024-00299-z.
6. Perrin et al. (2018). Transcriptomic analyses reveal rhythmic and CLOCK-driven pathways in human skeletal muscle. doi:10.7554/eLife.34114.
7. Raue et al. (2012). Transcriptome signature of resistance exercise adaptations: mixed muscle and fiber type specific profiles in young and old adults. doi:10.1152/japplphysiol.00435.2011.
8. Melov, Tarnopolsky, Beckman, Felkey and Hubbard (2007). Resistance Exercise Reverses Aging in Human Skeletal Muscle. doi:10.1371/journal.pone.0000465.
9. Lundell, Parr, Devlin et al. (2020). Time-restricted feeding alters lipid and amino acid metabolite rhythmicity without perturbing clock gene expression. doi:10.1038/s41467-020-18412-w.
10. Liberzon et al. (2015). The Molecular Signatures Database (MSigDB) hallmark gene set collection. doi:10.1016/j.cels.2015.12.004.
11. Wu and Smyth (2012). Camera: a competitive gene set test accounting for inter-gene correlation. doi:10.1093/nar/gks461.
12. Chen, Chen, Lun, Baldoni and Smyth (2025). edgeR v4: powerful differential analysis of sequencing data with expanded functionality and improved support for small counts and larger datasets. doi:10.1093/nar/gkaf018.
13. Robinson and Oshlack (2010). A scaling normalization method for differential expression analysis of RNA-seq data. doi:10.1186/gb-2010-11-3-r25.
14. Ritchie et al. (2015). limma powers differential expression analyses for RNA-sequencing and microarray studies. doi:10.1093/nar/gkv007.
15. Irizarry et al. (2003). Exploration, normalization, and summaries of high density oligonucleotide array probe level data. doi:10.1093/biostatistics/4.2.249.
16. Bolstad et al. (2003). A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. doi:10.1093/bioinformatics/19.2.185.
17. Benjamini and Hochberg (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. doi:10.1111/j.2517-6161.1995.tb02031.x.
18. Hemerik and Goeman (2018 (online 2017)). Exact testing with random permutations. doi:10.1007/s11749-017-0571-1.
19. Malhan et al. (2023). Skeletal muscle gene expression dysregulation in long-term spaceflights and aging is clock-dependent. doi:10.1038/s41526-023-00273-4.

Data and code availability

Analysis code, full methods, source-file information, frozen derived identifier mappings, supplementary results and figure source tables are provided in the accompanying Scientific Supplement, version 2.0. The Full Methods are also available as a separate PDF.

The underlying expression datasets are publicly available from GEO under the accession numbers listed in Table 1. The source-file catalog identifies the external data and resources used in the analysis.

Supplementary figures

Frequently asked questions