Metformin: Do Metabolic Changes Improve Function in Older Adults Without Diabetes?

Metformin is not established as a general functional intervention for older adults without diabetes, but neither has every possible functional benefit been excluded. Population, training context, outcome and analysis denominator matter. Our original transcriptomic analysis contained pairing and multiple-testing errors, so we withdraw the attribution of the published gene counts to “likely pseudoreplication.” Corrected analyses still do not fully reproduce the published lists; the cause remains unresolved.

Correction, 20 September 2026: We corrected numeric subject coding, BH adjustment order and sample linkage, withdrawing the 17,648 adipose differential-gene count and the associated attribution to the original authors. We obtained the MILES supplement, actual MASTERS corrigendum and MET-PREVENT main report, and added omitted positive studies. We corrected gait-interval interpretation, randomized versus completer counts, the adherence estimand and TAME source identity. Figures and the reproduction package have been replaced.

01Define the population and outcome first

These are different trials. Connecting one trial’s molecular change to another trial’s walking outcome does not establish a tested mechanistic chain.s1s3s5s7s10s12

02MILES: correct our analysis before interpreting disagreement

GEO provides 60,554 genes across 100 sequencing-sample columns. Exact subject, replicate, period, treatment and tissue keys uniquely link all 100 columns to GSM records. Aggregating technical replicates gives 55 subject-period-tissue samples. There are 13 complete muscle pairs and 14 adipose pairs; subject 8 lacks the muscle placebo sample. One GSM characteristics field actually lacks crossover period, recoverable consistently from the title and count-column name; the original article incorrectly described a tissue-field problem.s2

Our purported paired DESeq2 model treated numeric subject ID as a continuous covariate, producing only intercept, subject number and drug columns. It did not create person-specific pairing effects. The alternative model retaining technical replicates instead converted subject ID to a category. Comparing them changed both pairing and replicate handling, so the difference cannot be attributed entirely to pseudoreplication.

The paired-t analysis contained a separate BH error: the cumulative minimum must operate on sorted p-values before restoring gene order. Our code applied it in the original gene order. Correcting the same stored p-values changes the adipose count from 17,648 to 5; muscle remains 0. The extreme count was a programming error, not evidence of inherent biological instability.

The newly obtained supplement explicitly specifies edgeR, negative-binomial GLMs controlling subject and period, likelihood-ratio tests and BH correction; GEO specifies TMM normalization. We therefore withdraw the claim that detailed methods were unavailable. The source supplement does contain 647 muscle and 146 adipose genes. Full scripts, technical aggregation and some filtering wording remain incompletely specified.s10

Our corrected analysis uses categorical subject, period and treatment terms, sums technical replicates, and retains only complete pairs. Fractional RSEM expected counts are preserved rather than silently truncated. With current edgeR, different filters give 0 muscle and 12–18 adipose genes at FDR<0.05. Applying the source’s literal wording—exclude genes below 1 CPM in at least two libraries—gives 0 and 14; the common alternative of retaining genes reaching 1 CPM in at least two libraries gives 0 and 18. These are different filtering rules.

This is not an exact replication: the authors used edgeR 3.18.1 and we used 4.6.3, without their complete scripts or fully specified technical-replicate handling. The remaining discrepancy requires further tracing. It neither establishes author error nor means that failure to pass a particular FDR threshold proves absence of biological change.

03Pathway scores: people are the replication unit

The original analysis tested correlated genes as if they were independent replicates, overstating evidence. We retained its 11 hand-selected small lists for exploratory review, used each participant’s mean list-expression change as the unit, adjusted for period, reported intervals, and applied BH correction across 11 comparisons within each preprocessing approach. These are not complete canonical pathways or an independently preregistered analysis.

The adipose SREBP/lipogenesis list decreases under both approaches, with q approximately 0.0032–0.0033. Under TMM processing, the treatment coefficient is approximately−0.292 mean log2-expression units, 95% interval−0.419 to−0.165. Muscle oxidative-phosphorylation direction is downward, but its participant-level interval includes zero. Adipose collagen/ECM has unadjusted p around 0.04 and does not pass 0.05 after correction; it cannot be presented as equally established.

The CPM and TMM approaches differ in normalization and pseudocount conventions, so differences cannot all be assigned to one setting. More fundamentally, tissue-average RNA can reflect cell composition, and pathway labels are not measurements of metabolic flux, DNA repair capacity or kinase activity. These data do not directly identify activation of an anti-ageing mechanism.

04Exercise trials: reduced hypertrophy is not universal functional harm

MASTERS randomized 109 participants and analyzed 94 completers. Both groups undertook 14 weeks of resistance training. Lean mass increased 0.41% with metformin and 1.95% with placebo: a−1.54-percentage-point difference. Approximate Welch intervals recomputed from published completer means, SDs and counts are−2.55 to−0.53 points; thigh muscle-mass difference is−3.45 points, interval−5.42 to−1.48. These support smaller muscle-mass gains in that setting.s3s11

They are not effects in the complete randomized population or evidence that metformin causes muscle loss without training. We correct 109 versus 94 and retain selection concerns from 15 withdrawals. Strength, mass and tissue signaling also differ. The actual 2020 correction gives relative-strength changes of 14.8% versus 18.9%, between-group p=0.30. The earlier archived correction URL returned a challenge page, not the correction; the correct original is now obtained. Some archived supplementary entries retain old values, so explicitly corrected values take precedence.s4

Konopka’s aerobic-training trial supports attenuation of insulin-sensitivity and some mitochondrial-respiration adaptations. However, the approximately 50% smaller aerobic-fitness gain had between-group p=0.08 and should not be described with equal certainty. Significance in one arm but not the other does not replace a between-arm test. Positive and negative changes over one interval do not establish stable individual responder types.s7

A 2025 four-arm MRI study in 42 adults with prediabetes detected no muscle-imaging differences between combined treatment and exercise alone, adding a boundary to extrapolation. Mean age was approximately 48; the study was open-label, small and involved multiple outcomes and post hoc comparisons. Nonsignificance is not formal equivalence, and a within-arm change significant only in one group does not prove combination superiority. Its population and measurements differ from older-adult training studies.s16

05MET-PREVENT: how much gait-speed benefit was excluded?

MET-PREVENT randomized 36 participants per arm; 70 completed the primary assessment and 71 received at least one dose and entered safety analyses. Main-report Table 2 gives an adjusted four-metre gait-speed difference of 0.001 m/s, 95% interval−0.06 to 0.06,p=0.96. The trial targeted a 0.10 m/s improvement, which the interval excludes. Failure to achieve the planned larger improvement is supported.s5

The interval still includes+0.05 m/s. Research in other older populations has treated approximately 0.05 m/s as a small meaningful individual change. This is not a universally validated between-group threshold for MET-PREVENT, but makes “all meaningful gains excluded” untenable. We retain Table 2/abstract values; the source’s overall Figure 2 row instead shows−0.050 to 0.058. The discrepancy remains unresolved and we do not construct a new combined interval.s15

The 2026 secondary estimate of 0.072 m/s (−0.292 to 0.445) asks what would happen if everyone reached medication possession of at least 80%, not 100% perfect adherence. Its parametric g-formula requires additional adherence, post-treatment confounding, measurement and modeling assumptions. The wide interval does not reveal a hidden beneficial “true effect, ” and its difference from the assignment estimate is not an identified mechanistic share.s6

Tolerability matters. The abstract reports participants admitted to hospital as 12/35 versus 3/36, yielding an approximate RR 4.11 (1.27–13.34) and a difference of 26 percentage points (Newcombe interval approximately 7–43). Table 4 labels the row as admission counts, so our proportion calculations explicitly follow the abstract’s participant interpretation. Events are few, intervals wide and comparisons not multiplicity-adjusted; not every admission can be declared drug-caused. Frequent adverse events, greater discontinuation and the authors’ poor-tolerability finding nevertheless matter in this population.

06The omitted positive results belong in the assessment

Laksmi 2017 completers walked approximately 1.18 versus 1.14 m/s at baseline, substantially faster than MET-PREVENT participants. Of 120 randomized, 43 metformin and 48 placebo participants completed 16 weeks. The outcome was fifteen-foot walking time: adjusted means 3.72 versus 4.23 seconds, between-group p=0.024, a positive signal worth retaining. The frequently cited+0.13 m/s is within-metformin change, not an adjusted treatment contrast. Inverting group mean times cannot reconstruct mean individual speed effects.s12

Grip strength, myostatin and overall EQ-5 D did not show corresponding adjusted between-group benefits. Approximately 24% attrition and completer-only functional analyses leave selection and precision concerns. The study should neither be omitted nor treated as general proof of functional benefit.

Two reports from Qaisar’s group in older men also contain positive findings: one reports grip and some sarcopenia-related quality-of-life improvements without gait improvement; another reports improved grip, gait and SPPB. We obtained formal abstracts only and have not resolved full analyses, diabetes exclusions, dosage wording or cohort overlap. We therefore do not count them as two verified independent replications or treat marker correlations as proof of intestinal-barrier or neuromuscular-junction mediation.s13s14

Baseline function, metabolic state, training, exposure and analysis choices could contribute to differences. These remain hypotheses: ordering healthier positive trials beside frailer negative trials does not identify an optimal treatment population.

07From molecules to function and healthspan

Metabolic improvement, tissue synthesis and training adaptation may involve context-dependent tradeoffs. Transcriptomes, phosphorylation and respiration provide mechanistic clues, but do not identify which pathway determines net functional effects across populations or establish quantifiable human lifespan extension.

The original article mistakenly cited MILES registration NCT02432287 as evidence for TAME registration. The AFAR page checked on 20 September 2026 still describes fundraising for launch and future work; we obtained no completed TAME outcome results. A planned trial cannot substitute for effect evidence.s8s9

Some molecular changes and specific positive functional signals exist, alongside negative or null findings in exercise adaptation and very old frail adults. Current evidence does not establish metformin as a general way to improve function or extend life in older adults without diabetes. This does not negate established disease-treatment uses or establish zero benefit in every setting.

We did not redo sequencing alignment or obtain complete original transcriptomic code or clinical participant data. Software versions, filtering/technical-replicate details, unavailable small-trial full texts and source-table inconsistencies constrain interpretation. Statistical reproduction and source audit are not clinical professional review. Codex performed revision, self-review and bilingual checking; no personal treatment advice is given.

The reproduction package provides pinned-input download checks, corrected sample linkage, categorical paired/period models, participant-level gene-list analyses, safety calculations and figures. Superseded erroneous calculations are retained privately as correction history, not current findings.