The Cellular Energy Stack: what the human trials actually found

The Performance and Longevity Series, No.04 . Premium preview

Spermidine and urolithin A are the two most confidently marketed molecules in longevity supplementation. Between them they carry a mortality signal worth roughly 5.7 years, a mitophagy programme visible in human muscle biopsies, and four failed primary endpoints. This preview opens what the randomised trials measured, what they missed, and which compounds in the wider stack work against a training block. Built on evidence, not affirmations.

Why this matters

The tension nobody selling these molecules states plainly

Two compounds anchor the category. Spermidine is sold on autophagy. Urolithin A is sold on mitophagy. Both have genuine human biology behind them. Neither has a positive flagship efficacy trial.

The best-powered spermidine trial ever run, 12 months in 100 adults with subjective cognitive decline, missed its primary cognitive endpoint at P = .47, and one exploratory secondary moved the wrong way, with Trail Making Test B slower by 13.9 seconds, P = .03 (Schwarz et al., JAMA Netw Open 2022). Both flagship urolithin A trials also missed: cycling peak power was not significant in ATLAS (Singh 2022, Cell Rep Med), and both co-primaries, six-minute walk distance and maximal ATP production, were null in ENERGIZE (Liu 2022, JAMA Netw Open). That is not grounds for dismissing either molecule. It is grounds for reading the trials.

The numbers worth knowing

Five numbers that frame the category

Sources: Kiechl 2018; Senekowitsch 2023; Singh 2022, Eur J Clin Nutr; Gliemann 2013, J Physiol; Chilibeck 2017.

The grading

The four-tier verdict, condensed

Thirty-three compounds were graded on their best human functional evidence. Proven in humans requires replicated, adequately powered, primary-endpoint-positive evidence in a relevant population. Plausible but underpowered means positive signals that are small-n, secondary-endpoint, unreplicated or sponsor-dependent. Mechanism-only means a biomarker moved with no confirmed functional outcome. Marketing narrative means the claim is common in commerce and unsupported or contradicted by human data.

In the bottom two tiers the human evidence is often not weak but absent: no human interventional trial of isolated urolithin B or isourolithin A exists, no human trials of NRH were located, and none of apigenin as an NAD+ raising agent (Romo-Vaquero 2019; Rumpler 2026; Escande 2013).

Anchor molecule one

Spermidine: a superb epidemiological signal attached to a failing capsule

What is established

Dietary spermidine intake tracks with lower all-cause mortality across two independent cohorts, HR 0.74 (0.66 to 0.83), P < 0.001 in Bruneck and HR 0.71 (0.53 to 0.95), P = 0.019 in SAPHIR, equating to roughly 5.7 years (3.6 to 8.1) of mortality age between the top and bottom third of intake (Kiechl 2018). Fasting also raises endogenous spermidine in human volunteers, and blocking synthesis abolished the lifespan and cardioprotective effects of fasting in vivo, which is preclinical causal work (Hofer, Nat Cell Biol 2024). Both findings argue for wheat germ, soy ferments, mushrooms and fasting. Neither argues for capsules.

The failed primary endpoint

In SmartAge, the only adequately sized, registered, 12-month randomised trial, mnemonic discrimination between groups was -0.03 (-0.11 to 0.05), P = .47 in the intention-to-treat analysis and -0.04, P = .25 per protocol, despite compliance of 95.3%. The Trail Making Test B result of +13.9 s (1.5 to 26.2), P = .03, meaning slower executive function, came from the per-protocol-plus set of 74 participants, a different analysis population from the primary, in a trial with no multiplicity correction on secondaries; the one favourable secondary, sICAM-1 at -56.2 ng/mL, P = .03, sits in that same uncorrected set (Schwarz 2022).

Three independent pharmacokinetic failures

The premise that oral spermidine raises systemic spermidine has been tested three times, at three doses, by three groups, and failed each time.

The dose arithmetic closes the argument. Median habitual dietary spermidine in older adults is 11.5 mg/day (IQR 9.5 to 14.3), each doubling of intake raises plasma spermidine only 18%, p = 0.02, and skeletal muscle spermidine is not associated with intake or plasma, p > 0.05 (Thorup 2026). The EU-authorised supplement ceiling of 6 mg/day therefore sits below habitual dietary intake (FSA novel-130).

Sponsor concentration is heavy. The keystone mechanism paper's authors cofounded Samsara Therapeutics and hold equity in The Longevity Labs, SmartAge extract development was funded by TLL The Longevity Labs, and the most positive registered result names TLL and Oxford Healthspan, a spermidine retailer, in its conflict statement (Hofer 2024; Alsaleh 2026).

Anchor molecule two

Urolithin A: real target engagement, missing outcomes

What it genuinely does to human muscle

This is the strongest mechanistic package in the stack. Muscle mitophagy and autophagy transcripts including PARK2, GABARAPL1 and ULK1 rose at 28 days in healthy sedentary elderly adults (Andreux 2019, Nature Metabolism). At four months the Parkin-mediated ubiquitination pathway ranked first in muscle proteomics at 500 mg, with phospho-Parkin Ser65 increased and OXPHOS complexes I, II and III up dose-dependently (Singh 2022). Acylcarnitines, ceramides and CRP fall consistently, with zero serious adverse events across four months (Liu 2022).

The two flagship misses

In ATLAS the primary endpoint of cycling peak power was not significant, at +4.3% and +3.7% for 500 and 1000 mg (Singh 2022). In ENERGIZE both co-primaries were null: six-minute walk +60.8 m versus +42.5 m, and maximal ATP production 0.07 versus 0.06 mM/s, with the pre-specified low-baseline-ATP subgroup also non-significant at p = 0.36 (Liu 2022).

Most circulating positive numbers are within-group changes, not between-group contrasts. Peak VO2 rose +10.2% within-group at 1000 mg, p < 0.01, but p = 0.058 versus placebo, and six-minute walk +33.43 m within-group, p = 0.008, but p = 0.098 versus placebo (Singh 2022). The +5.4% VO2max figure used in marketing is a within-group change in a study where placebo also rose 3.6% during the same altitude camp, and in that trained-runner trial the 3000 m time trial did not improve, p = 0.116, with the VO2max time-by-treatment interaction not significant, p = 0.138 (Whitfield 2025, Sports Medicine).

Three structural problems follow. Hamstring average peak torque rose +12% at 500 mg, p = 0.027, while quadriceps were null (Singh 2022). The clearest muscle-endurance benefit, FDI +95.3 versus +11.6 contractions at two months, P < .01, had disappeared by four months, and at steady state parent urolithin A was 1038 pg/mL against UA-glucuronide at 1014 ng/mL, a thousand-fold conjugate excess (Liu 2022). A pharmacokinetic modelling analysis challenged "the potential for readily achieving beneficial effects by postbiotic supplementation" (Aichinger 2023).

The sponsor concentration, quantified

Amazentis SA, which sells the branded ingredient Mitopure, is lead sponsor of 13 of the 62 studies returned by a ClinicalTrials.gov query for urolithin, and as of 26 July 2026 not one of those 13 has posted registry results, so all sponsor outcome data reach readers through sponsor-authored journal papers (ClinicalTrials.gov). The only non-sponsor positive human data are two trials of n = 20, one single-blind and one carrying an internal statistical inconsistency in its CRP data (Zhao 2024; Monsalve Acevedo 2025).

The producer metabotype

Urolithin A is not present in food. It is a gut-microbial metabolite formed when dietary ellagitannins are hydrolysed to ellagic acid and converted by specific colonic bacteria. Only 12% of adults had detectable UA-glucuronide at baseline, and six hours after pomegranate juice only 4% exceeded the 100 ng/mL high-producer threshold, rising to roughly 40% at 24 hours with 33% still non-converters; no Asian or Singaporean estimate exists, recorded as n.a. (Singh 2022, Eur J Clin Nutr). Estimates disagree widely, with a Brazilian cohort splitting 32.2% UM-A, 52.5% UM-B, 15.3% UM-0 (Leite 2026). Metabotype is modifiable: a synbiotic plus 400 mg pomegranate extract took a cohort from 44.4% to 100% converters, p < 0.01 (Napier 2025).

The tier-one compounds

What actually has the evidence

Creatine monohydrate

Creatine is the only compound in the corpus with an authorised regulatory efficacy claim for adults over 55: a cause and effect relationship was established between 3 g/day or more plus resistance training three times weekly and improvement in muscle strength (EFSA NDA Panel). The effect is lean tissue mass +1.37 kg (0.97 to 1.76), p < 0.00001, with chest press SMD 0.35, p = 0.0002 (Chilibeck 2017). Memory carries a separate GRADE-moderate signal at SMD 0.31 (0.18 to 0.44) (Xu 2024), and side effects appeared in 13.2% of placebo versus 13.7% of creatine arms, p = 0.776, across 685 trials (Kreider 2025).

Two honest limits. Creatine is ineffective for endurance in trained populations, SMD -0.07, p = 0.47 (Fernandez-Landa 2023), and in U23 professional cyclists at 20 g/day there was no significant group by time interaction for any outcome (Barranco-Gil 2024). It does not, however, impair heat dissipation or fluid balance (Lopez 2009).

Omega-3 EPA and DHA

The best-evidenced recovery agent in the corpus. Across 41 RCTs and more than 1800 participants: IL-6 SMD -0.52 (-0.74, -0.29), p = 0.001; TNF-alpha -0.46; creatine kinase -0.57; DOMS -0.49; performance +0.21 (0.03, 0.41), p = 0.028, with GRADE moderate certainty for IL-6 and TNF-alpha. The dose-response breaks at 2 g/day or more of EPA plus DHA, the effect roughly doubles past six weeks, and EPA-dominant formulas outperform DHA-dominant ones at -0.58 versus -0.31 (Li and Zhang, FASEB J 2026).

The ceiling is a safety one. At 4 g/day there is a replicated atrial fibrillation signal, HR 1.69 (1.29 to 2.21) in STRENGTH and 3.1% versus 2.1%, P = 0.004, in REDUCE-IT (AHA STRENGTH summary; JAHA 2023), against no excess bleeding in 25,871 people over 5.3 years at 840 mg/day (VITAL, NEJM 2019).

Measure before buying, and eat the sprouts

Vitamin D3 is tier one only for correcting a measured deficiency, and 42% of healthy Singapore residents were deficient below 20 ng/mL despite 2,022.4 hours of annual sunshine (Bi, PLoS One 2016). Supplementing the replete does nothing: cardiorespiratory fitness SMD 0.03 (-0.12 to 0.17), P = .8959, I2 = 0% (Chen 2026).

One compound improves training rather than interfering with it. Sulforaphane from broccoli sprouts raised time to exhaustion from 380 to 426 s against placebo 401 to 406 s, interaction p = 0.02 (Bratt 2023), with NRF2 activation additive with exercise, 2.1-fold combined versus 1.5-fold each, p < 0.01 (Rodriguez 2025).

The most consequential evidence in the dossier

The blunting section: what not to take during a build

Much of this category is sold on antioxidant or AMPK-agonist mechanisms that intervene at exactly the nodes endurance adaptation depends on: transient reactive oxygen species, AMPK activation and inflammatory signalling converging on PGC-1alpha and mitochondrial biogenesis. This is the evidence that almost never appears in the marketing copy.

The honest synthesis. Blunting is real at the molecular level and consistently demonstrated in PGC-1alpha, COX4 and the mitochondrial proteome, yet in both direct human vitamin trials the performance outcome was unaffected over 3 to 11 weeks, and nobody has run the 6 to 12 month masters-athlete trial that would settle whether chronic blunting eventually costs performance (Paulsen 2014; Wyckelsma 2025). Resveratrol and metformin go further, showing whole-outcome harm relative to placebo in randomised humans. Timing does not rescue adaptation: the Paulsen protocol split doses before and after training and still blunted signalling.

The build-block rule

• Never during a build block, on randomised human evidence of harm: resveratrol, metformin, rapamycin, pterostilbene (Gliemann 2013; Konopka 2019; Walton 2019; Drummond 2009; Riche 2014).
• Never acutely before a key session or race: 80 mg MitoQ (Hughes 2023).
• Never run high-dose NMN, or any NAD+ precursor, through an adaptation-critical block (Yang 2026).
• Periodise the direct scavengers such as GlyNAC and high-dose omega-3 into recovery weeks and travel, not continuously through builds (Kumar 2022).
• Run the Nrf2 inducer and structural cofactors continuously, because sulforaphane improved rather than blunted training outcomes (Bratt 2023; Rodriguez 2025).
• Creatine is the exception that must be taken daily, including rest days, because the authorised effect failed when the same weekly dose was given only on training days (EFSA).
• Stop magnesium before performance testing if status is replete (Bomar 2025).
• For spermidine, urolithin A, GlyNAC, fisetin, dasatinib, pterostilbene and ALA, whether they blunt adaptation is n.a. That is an absence of evidence, not a clearance (Singh 2022; Isenmann 2020).

Evidence and price run in opposite directions. Creatine costs a fraction of urolithin A for the only authorised efficacy claim in this demographic, and what gets avoided during a build matters more than what gets added.

Colophon and method

Scope: this preview draws on a master dossier synthesised from six evidence files covering the polyamine, mitophagy and NAD+ axes, bioenergetic support, redox and substrate support, and senolytics. Thirty-three compounds were graded, with more than 200 individual human trial entries tabulated across the compound sections, alongside registry records, regulatory opinions and observational cohorts.

Evidence bar: human interventional data drive every verdict. Preclinical work appears only to explain mechanism and is labelled as such. Nulls carry equal weight with positives, and every compound carries both its strongest human study and its most damaging one. Within-group changes are separated from between-group contrasts throughout. Where a value could not be confirmed from a primary source it is written n.a. rather than estimated.

Independent fact-check: a pass dated 26 July 2026 reviewed the dossier against its sources. All 165 unique PubMed identifiers cited resolve at NCBI, with author, year and journal matching the anchor text in every case, and no dose figure anywhere in the document was found to be wrong. Eleven precision corrections were applied, and five items remain recorded as unverifiable, including retail prices behind HTTP 403 responses and registry status lines that drift.

Standing caution: this is evidence reporting, not medical advice. Dasatinib, sirolimus and metformin are prescription-only medicines requiring a physician, a diagnosis and monitoring, and doses exceeding HSA limits or tolerable upper intake levels, or interacting with prescription medication, require clinical supervision (HSA; MIMS Singapore).

Selected sources are linked inline throughout. The full reference carries the complete citation set across 165 resolved identifiers. Prepared July 2026. Built on evidence, not affirmations.