Compound Monograph  ·  No. 35  ·  Research Use Only

Humanin The message from the ruins — a peptide found in the one part of an Alzheimer’s brain that had not died

Almost all Alzheimer’s research looks at the damage. In 1998 a laboratory in Tokyo looked at what was left. It took the occipital lobe of an autopsied patient — the region the disease characteristically spares — built a library of every gene switched on there, and asked a blunt question: is anything in this tissue actively keeping cells alive? Four rounds of selection later it had one answer, and the answer was strange. The rescuing sequence was twenty-four amino acids long, and it was written inside a gene that is not supposed to encode a protein at all — the mitochondrion’s gene for a ribosomal RNA. The peptide was named humanin, after humanity. In the twenty-five years since, it has been shown to protect neurons, heart muscle, kidney, retina, lung, testis, ovary, bone and blood vessels; it founded the idea that mitochondria send signals as well as receive them; and it has never once been given to a human being in a registered clinical trial.

Compiled by South Beach Longevity · 3 August 2026
Copyright 2026
Corpus 494 scientific full texts · ~8,109 printed-page equivalents
Metadata layer 525 PubMed records screened from 823
Source project 05 · Therapeutic Peptide Research Library
Constraint No human use, dose, route or schedule is recommended anywhere in this document

01The part of the brain that did not die

Alzheimer’s disease does not consume a brain evenly. It begins in the entorhinal cortex and the hippocampus, spreads through the temporal and parietal association areas, and characteristically leaves the primary sensory regions comparatively intact. The occipital lobe — the visual cortex at the back of the head — is among the last places to go, and in many patients it never really does. A pathologist examining such a brain sees devastation at the front and something close to normality at the back.

That asymmetry is usually treated as a fact about where the disease goes. In the late 1990s, a group at Keio University School of Medicine in Tokyo treated it instead as a question about what the surviving tissue might be doing.

The laboratory belonged to Ikuo Nishimoto, and it had a specific reason to be interested in how neurons die. Six years earlier Nishimoto had published a paper in Nature arguing that the amyloid precursor protein — the molecule at the centre of Alzheimer’s genetics — behaves like a receptor, coupling through the G protein G_o (Nishimoto et al., 1993). If a familial Alzheimer’s mutation turned that receptor into a death signal, then the interesting question was not only what pulls the trigger but what, in some cells, blocks it.

The method they used was called a death trap, and its logic is attractively brutal. Take a library of complementary DNA — a snapshot of every gene being transcribed in a tissue — and put it into cells that are about to be killed. Kill them. Whatever survives is likely to have received a cDNA that protects. Recover the plasmids from the survivors and do it again.

The killing was done by a familial Alzheimer’s mutation. The cells were F11 hybrids of rat primary neurons and a mouse neuroblastoma line, engineered to switch on the V642I “London” mutant of the amyloid precursor protein when given the insect hormone ecdysone. Seventy-two hours of that is lethal. The library came from the occipital cortex of a patient who had died with Alzheimer’s disease.

Four rounds of selection reduced the library to 250 clones. Those clones fell into thirty-six groups that cross-hybridised with one another; the largest group held twenty-eight of them, all carrying versions of a single complementary DNA 1,567 bases long. Inside it was an open reading frame of seventy-five bases — enough for twenty-four amino acids and a stop (Hashimoto et al., 2001).

Synthesised and added back to dying cells, that twenty-four-residue peptide abolished the death. Not reduced it — abolished it, across a wide range of familial Alzheimer’s mutations in the amyloid precursor protein and in presenilin 1, and against amyloid-β itself. It did not work against glutamate excitotoxicity, and it did not work against a prion fragment, which mattered: the effect had a shape, and was not simply a peptide making cells generically hardier.

They called it humanin. The etymology given in the literature is “humanity” — the hope that a molecule which could keep neurons alive might give back what dementia takes. Reviews most often attribute the name to Nishimoto, the senior author; at least one credits Hashimoto, the first author. The 2001 paper itself offers no explanation at all.

02One gene, two readings

Here the story acquires the feature that has defined it ever since, and it is worth being precise about the order in which things were learned, because the literature routinely gets it wrong.

The 2001 paper did not report a mitochondrial peptide. It reported that the rescuing cDNA was 99 per cent identical — 1,545 bases out of 1,553 — to a human sequence already deposited in GenBank as a nuclear complementary DNA. Nothing in it says the word mitochondrion. The mitochondrial identity was published the following year, when the same group raised an antibody and noted that their 1,567-base cDNA matched the mitochondrial 16S ribosomal RNA (Tajima et al., 2002).

The gene is MT-RNR2. It occupies positions 1,671 to 3,229 of the human mitochondrial genome and it encodes the large subunit ribosomal RNA — a structural RNA, part of the machine that builds proteins. Both of the databases that curate human genes classify it as non-coding. It is not supposed to be a gene for a protein. The humanin reading frame sits inside it, beginning at position 2,633.

Read that frame out and something genuinely elegant happens, and it explains a discrepancy that runs through the whole literature. The same seventy-five bases give two different peptides depending on which machine reads them.

A ribosome in the cytoplasm uses the standard genetic code and produces twenty-four residues, MAPRGFSCLLLLTSEIDLPVKRRA, terminating at a TAA. A ribosome inside the mitochondrion uses the vertebrate mitochondrial code, in which the codons AGA and AGG are stop signals rather than arginine. The twenty-second codon of the humanin frame is AGG. An organelle translating its own gene therefore stops three residues early and produces twenty-one.

Both forms are biologically active, which is why the field has never been able to settle the question by function alone, and why the peptide is reported in the literature as both twenty-four and twenty-one residues long without either being an error.

The consequence is that humanin is a genuinely awkward object. It is a sequence with strong evidence of biological activity, nested inside a gene annotated as non-coding, in an organelle whose own translation machinery reads it differently from the rest of the cell. Every subsequent controversy about the molecule — where it is made, whether the thing measured in blood is really it, whether the nuclear copies matter — descends from this one structural fact.

03Found three times

A molecule discovered once by a functional screen is a candidate. A molecule that three laboratories arrive at independently, chasing three unrelated questions, is harder to dismiss. Humanin is usually described as the second kind, and that description needs one correction.

In May 2003 John Reed’s laboratory at the Burnham Institute in La Jolla published in Nature that humanin binds Bax (Guo et al., 2003). They had not been looking for a neuroprotective peptide. They had been looking for proteins that interact with Bax, the pro-apoptotic switch that punches holes in mitochondrial membranes, and humanin came out of a yeast two-hybrid screen with Bax as the bait. No author on that paper had any connection to Tokyo. This is a genuinely independent discovery.

The Nature paper also contains a detail that has shaped the field’s confusions ever since. Guo and colleagues cloned humanin from a nuclear library and treated the nuclear copy as the real one, noting only in passing that “the mitochondrial genome contains an identical open reading frame” and that the mitochondrial version also works. Their proposal was that humanin arose in mitochondria and transferred to the nucleus. Twenty-two years later, which copy the body actually uses is still not settled.

Five months later a second group published in the Proceedings of the National Academy of Sciences that humanin binds insulin-like growth factor binding protein 3 (Ikonen et al., 2003). Pinchas Cohen’s laboratory at UCLA had run a yeast two-hybrid screen with IGFBP-3 as bait against a HeLa library, and humanin fell out of it unprompted. The screen was genuinely unbiased and genuinely unrelated to Alzheimer’s disease.

The paper, however, was not independent of the original group. Hashimoto, Niikura and Nishimoto are all co-authors. Several widely read reviews state that two laboratories found humanin “independently of each other and of Nishimoto’s group”, or that there were “three independent laboratories”. That is true of Reed’s paper and false of Cohen’s, whose author list includes Nishimoto himself. What was independent was the route — an unrelated bait, an unrelated question — not the personnel. The distinction matters because convergent discovery is being used as an argument that the molecule is real, and an argument is only as strong as its weakest instance.

A fourth recovery is worth naming because it came from a different species and a different continent. A group in Rome, screening for neuroprotective factors, found a rat gene encoding a humanin-like peptide they called rattin (Caricasole et al., 2002). It is thirty-eight residues rather than twenty-four, shares 73 per cent identity across the conserved region, and — unlike humanin — protects against NMDA excitotoxicity.

Nishimoto did not see most of what followed. He died on 17 October 2003, aged forty-seven, five months after the Nature paper and eleven days before the PNAS paper appeared. His last review of the molecule he had named was published posthumously in Trends in Molecular Medicine in March 2004, running from page 102 to page 105 of that issue. His obituary, written by his Keio colleague Masaaki Matsuoka, is on page 105.

04Is it actually made?

This is the section most reviews of humanin skip, and it is the one a careful reader most needs, because everything downstream depends on it. A peptide that rescues cells when you add it to them is a pharmacological fact. Whether the body makes it, and from which genome, is a separate question with a genuinely unsettled answer.

The difficulty is specific and it is not going away. When mitochondrial DNA fragments insert themselves into the nuclear genome — a common event, and the resulting sequences are called NUMTs — they carry their reading frames with them. The human nuclear genome contains thirteen such humanin-like loci, MTRNR2L1 through MTRNR2L13, scattered across eleven chromosomes. At least ten of them are transcribed in human tissue (Bodzioch et al., 2009). Both of the authorities that curate human gene names classify all thirteen as pseudogenes.

And two of them — MTRNR2L8 on chromosome 11 and MTRNR2L12 on chromosome 3 — encode peptides that are identical to each other and differ from mitochondrial humanin by a single amino acid.

Against that background, the evidence that humanin is genuinely a mitochondrial product comes from two directions, and both are worth stating precisely.

The clean experiment was done in a rat. A group at UCLA reasoned that the rat humanin sequence is predicted to lack the confounding nuclear insertions, and used that to ask the question directly (Paharkova et al., 2015). Production of the rat peptide fell in cells depleted of mitochondrial DNA, which places its origin in the mitochondrial genome; and it was unaffected by cycloheximide, which blocks cytoplasmic ribosomes but not mitochondrial ones, which places its translation inside the organelle. The same paper found that adding a humanin analogue to isolated mitochondria from a rat β-cell line reduced hydrogen peroxide production by 55 per cent. That paper’s opening sentence concedes what it was written to fix: evidence for the mitochondrial origin of humanin “has been lacking”. The equivalent experiment has never been done in human cells, where the thirteen nuclear copies exist and ten are transcribed.

The evolutionary argument is the strongest one available, and it is clever. Because the humanin frame is embedded in a gene for a structural RNA, any selection acting on the RNA would constrain the DNA sequence — but it would not care which amino acids the frame happens to spell. Synonymous codon bias — a preference among codons that all specify the same residue — can therefore only arise if something is selecting on the peptide. Analysing vertebrate genomes, Gruschus and colleagues found exactly that for humanin, with robust evidence of purifying selection (Gruschus et al., 2023).

That result cuts two ways, and the monograph reports both edges. The same analysis found no significant bias for SHLP1, SHLP2, SHLP3 or SHLP5 — four of the peptides that the field describes as humanin’s family. Whatever the test establishes about humanin, it declines to establish it about most of its relatives. And humanin’s own conservation is patchier than the headline suggests: the initiating methionine is conserved in only 79 per cent of vertebrates, ranging from 96 per cent in birds to 46 per cent in reptiles, and the stop codon in only 58 per cent.

What is missing is the measurement that would settle it. Despite twenty-five years of work and several hundred papers reporting humanin concentrations in human blood, there is no published mass-spectrometric identification of endogenous human humanin — no measurement that reads the actual peptide sequence out of a real sample and distinguishes the mitochondrial product from its nuclear near-twins. Every mass-spectrometry paper in this corpus characterises synthetic peptide or maps an interaction using synthetic peptide. The nearest thing to a worked example concerns a sibling peptide, where a detected sequence turned out to match equally well a micropeptide translatable from a NUMT on chromosome 17 — an interval that falls inside MTRNR2L1.

One further complication, which anyone reading the animal literature should hold on to: in mice, the mitochondrial humanin gene is itself a pseudogene, and the humanin-like peptide is encoded by a nuclear gene. The most heavily used experimental species for this compound is not a model of the arrangement being studied.

None of this means humanin is not made. The codon-bias result is a real argument that something has been conserving this peptide sequence across vertebrate evolution, and the rat experiment is a real demonstration of mitochondrial translation in at least one mammal. It means that the sentence “humanin is a mitochondrial-derived peptide”, which appears in the opening line of most papers on the subject, is a well-supported inference rather than a settled measurement — and that the field has known this and said so in print for over a decade.

05Two receptors, and then four

If humanin is secreted and acts on other cells, it must bind something. The question of what has been open since 2004, and the honest answer in 2026 is that the field carries several incompatible accounts side by side and has never tested them against each other.

The first answer came from the National Cancer Institute. Humanin, they reported, draws mononuclear phagocytes across a membrane through FPRL1 — formyl peptide receptor-like 1 — and its murine counterpart FPR2 (Ying et al., 2004). The finding had an attractive corollary: amyloid-β₄₂ uses the same receptor, so humanin might work by competing with it for access.

A companion paper the same year supplied the only hard receptor potency numbers the humanin literature has ever produced. In cells transfected with human FPRL1, humanin suppressed cyclic AMP with a half-maximal concentration of 3.5 nM; N-formylated humanin did the same at 0.012 nM, a roughly 290-fold gain from a single chemical modification (Harada et al., 2004).

One year later the Keio group tested that receptor in the system where humanin had been discovered, and it failed. Knocking down FPR2 did not attenuate humanin’s rescue of neurons from Alzheimer-related insults; the authors concluded in print that “a receptor other than FPR2 exists” (Hashimoto et al., 2005). That negative result is the reason the second account exists at all.

The second account is a three-part receptor: ciliary neurotrophic factor receptor α, WSX-1, and gp130. Overexpressing the first two increased humanin binding to neurons, knocking them down reduced it, and a reconstituted assay showed humanin driving the three subunits into a complex (Hashimoto et al., 2009). Downstream it runs through JAK2 and STAT3.

Since then the picture has become more crowded rather than less. In the testis, humanin’s protection of germ cells is blocked by antibodies against gp130, IL-27R and EBI-3, and is absent in mice lacking either of the latter two — but it is not blocked by antibody against CNTFR (Jia et al., 2021). The testicular receptor is not the neuronal one. In glioblastoma, humanin acts through gp130 and the formyl peptide receptors are undetectable. And in 2024 a fourth mechanism arrived: humanin as a ligand for integrin αV, signalling through the TGF-β receptor to drive migration and invasion (Ha et al., 2024).

The field’s standard reconciliation is that different cell types use different receptors, phagocytes carrying the formyl peptide receptors and neurons not. That is a reasonable hypothesis and it has never been an experimental result. Two decades of reviews restate the conflict without adjudicating it; a 2026 review still writes the architecture as “humanin–FPR2/gp130”, two receptors joined by a slash.

The sharpest datum is easy to miss. The 2024 glioblastoma paper looked for expression of IGFBP3, BCL2, WSX-1, gp130, CNTFR and FPRL1 across cancer cell types that did and did not respond to humanin, and found no distinguishable pattern. None of the canonical receptors predicted which cells the peptide would act on.

06The other humanin, inside the cell

Alongside the receptor literature runs a second, largely separate account in which humanin never leaves the cell that made it.

Reed’s 2003 Nature paper established the first version: humanin binds Bax and prevents it moving from the cytosol to the mitochondrial membrane, where it would otherwise trigger the release of cytochrome c and commit the cell to apoptosis. Two years later the same group extended it to Bid and its truncated form, and showed that humanin retains protective activity on mitochondria from bax-null cells — so the effect is not solely Bax-dependent (Zhai et al., 2005). A third partner, BimEL, is bound with unusual specificity: humanin binds the extra-long splice form and not the two shorter ones, and protects against death induced by the one it binds (Luciano et al., 2005).

Then there is IGFBP-3, the binding protein that carries most of the circulating insulin-like growth factor. Humanin binds its heparin-binding domain, residues 215 to 232, and blocks the protein’s import into the nucleus. The F6A mutant, which does not bind IGFBP-3, is the tool that separates this arm from the others.

Since 2019 this account has been substantially reinterpreted, and the reinterpretation is uncomfortable. Studies of humanin with Bax and with Bid found that the two do not simply bind — they form fibres together, with Bax converting entirely to β-sheet as it is incorporated (Morris et al., 2019, 2020). The proposed mechanism is sequestration by co-fibrillation rather than classical inhibition. The internal control is reassuring: the anti-apoptotic protein BCL-x_L, which humanin has no reason to neutralise, does not form fibres with it.

In 2025 the same group reported that humanin itself forms amyloid-like β-sheet fibrils, and that the mutants which fail to fibrillate are the same ones previously catalogued as secretion-deficient (Morris et al., 2025). They also questioned the founding structural model directly, suggesting that results the original laboratory attributed to humanin dimerisation may reflect aggregation instead.

07The potency problem

Three numbers, taken together, describe a molecule whose pharmacology is stranger than its reputation.

The first: wild-type humanin protects cells in the founding assay at 1 to 10 micromolar. The second: substituting a single serine at position 14 with glycine — giving S14G-humanin, the analogue that carries most of the animal literature — drops that to 10 nanomolar or below, a thousandfold gain. The third: fusing a nine-residue fragment of another neurotrophic peptide to a shortened humanin derivative gives colivelin, which suppresses the same death at 100 femtomolar (Chiba et al., 2005).

That is nine orders of magnitude across four closely related molecules, and it is the strongest possible argument for the rule this monograph applies throughout: their evidence must not be pooled. A result obtained with colivelin says nothing about humanin.

The fourth number is the one the field does not put in its abstracts. In 2013 the laboratory that had identified and named the three-part humanin receptor published a paper stating that “the protective effect of Humanin via the htHNR is weak (EC50 = 1–10 µM)” (Hashimoto et al., 2013). The paper exists because that weakness sent them looking for a better agonist of their own receptor, and they found one — calmodulin-like skin protein, active at 10 to 100 picomolar and circulating in normal human blood at around 5 nanomolar. The discovering laboratory, in other words, concluded that humanin is a micromolar agonist at the receptor it had named for humanin, and that something else is probably the physiological ligand.

The fifth observation is an absence. There is no measured binding constant for humanin at any of its proposed cell-surface receptors. No dissociation constant, no Scatchard analysis, no surface plasmon resonance, no calorimetry — not for the three-part receptor, not for the formyl peptide receptors. The 3.5 nM figure above is a functional half-maximal concentration in a transfected cell line, not an affinity. The only rigorously measured affinity in the entire humanin literature is for a partner nobody predicted: apolipoprotein E4, at 12.7 nM for wild-type humanin (Miller et al., 2024). The only measured affinity for an intracellular partner is 5.05 micromolar, for IGFBP-3.

The structural work is correspondingly thin on the native molecule. There are four experimental structures. Three are solution NMR of synthetic humanin or its analogues in 30 per cent trifluoroethanol, a solvent used to coax helix out of peptides that do not have much of their own — in water humanin is largely unstructured, with transient turns. The fourth is a 2022 cryo-electron microscopy structure of FPR2 with humanin bound, and it is a genuinely important result. It also used N-formylated humanin, not the native peptide, and the physiological abundance of the formylated form has not been established.

08The brain

The first demonstration that a humanin analogue does anything in a whole animal came from a laboratory unconnected to the discoverers. At Meijo University in Nagoya, [Gly¹⁴]-humanin injected into the cerebral ventricles of mice reversed the memory impairment caused by scopolamine, and did so without changing how much the animals moved — a control that matters, because a drug that makes a mouse more active will improve a maze score without improving memory (Mamiya & Ukai, 2001).

What followed over two decades is one of the broader preclinical neurology literatures attached to any peptide in this series. Humanin analogues have been given into the ventricles, into the hippocampus, into the peritoneum and up the nose, in mice and rats, against amyloid-β infusion, against anticholinergic amnesia, and in the standard transgenic models of Alzheimer’s disease. Reported effects include restored performance in water mazes and object recognition, preserved cholinergic neurons, reduced amyloid burden, reduced tau phosphorylation, restored dendritic spine density, and — repeatedly and across several groups — rescue of long-term potentiation measured in the living animal, which is a more physiological endpoint than a behavioural score.

Two internal controls give this literature more weight than a simple list of positive findings would. In the amyloid-β_25-35 model, S14G-humanin prevented the deficit at 50 picomoles, while S14A-humanin — a neuroprotection-defective mutant — did nothing at 5 nanomoles, a hundredfold higher dose (Tajima et al., 2005). And in an alanine scan of the analogue series, truncated derivatives and most tert-leucine substitutions were simply inactive in vivo. The effect is sequence-specific, not an artefact of injecting a hydrophobic peptide into a brain.

Two findings on the other side belong here rather than in a late section.

The first is a clean failure. In the intracerebroventricular streptozotocin model — a rat model of sporadic rather than familial Alzheimer’s disease, built around disrupted brain insulin signalling — S14G-humanin was tested at 0.01, 0.05, 0.1 and 1 nanomole daily for eleven days and failed at every dose. It did not restore spatial memory and it did not restore hippocampal Akt phosphorylation. The paper is titled, without euphemism, “Humanin Does Not Protect Against STZ-Induced Spatial Memory Impairment” (Negintaji et al., 2015). Since sporadic disease is what almost all patients have, this is not a peripheral negative.

The second is a dissociation. In triple-transgenic mice given intranasal S14G-humanin for three months, brain amyloid fell markedly in both sexes — but cognition improved only in males (Niikura et al., 2011). A treatment that lowers the pathology without changing the behaviour in half the animals is a result worth holding on to when reading any summary that reports only the amyloid figure.

Colivelin deserves its own note, because it is the most potent molecule in the family and because its record is not uniformly favourable. It crosses the blood–brain barrier after intraperitoneal injection, reaches the central nervous system when given intranasally by way of the olfactory bulb, and in transgenic mice carrying the G93A superoxide dismutase mutation it dose-dependently improved motor performance and prolonged survival — where the neurotrophic fragment alone improved motor performance but did not extend life. But colivelin is also sold as an off-the-shelf STAT3 activator, and in that role it has produced the opposite result: in rats with chronic cerebral hypoperfusion, activating the pathway with colivelin exacerbated neuronal damage and cognitive decline, while inhibiting the same pathway helped. The contradiction is direct, it is within the same species, and it is not resolved.

09The heart, and the pig that did not follow

The cardiac work produced the cleanest dose–response curve in the entire humanin corpus, and then the clearest illustration of why a clean rodent curve is not a clinical result.

In mice subjected to forty-five minutes of coronary occlusion and twenty-four hours of reperfusion, S14G-humanin reduced infarct size as a fraction of the tissue at risk from 58 ± 4 per cent in vehicle animals to 44 ± 4 at 0.2 mg/kg, 39 ± 2 at 1 mg/kg and 28 ± 4 at 2 mg/kg. Four milligrams per kilogram gave no further benefit — the curve plateaus, which is what a real receptor-mediated effect looks like. At one week, ejection fraction had fallen from 78 to 25 per cent in vehicle animals and from 76 to 56 per cent in treated ones (Muzumdar et al., 2010).

Ten years later the same intervention was taken into a large animal, which is the step at which most cardioprotective agents fail. Female Yucatan minipigs underwent balloon occlusion of the left anterior descending coronary artery followed by forty-eight hours of reperfusion, with S14G-humanin at 2 mg/kg given intravenously ten minutes before reperfusion — a protocol designed to mirror what could actually be done to a patient arriving for angioplasty (Sharp et al., 2020).

With sixty minutes of ischaemia, infarct size fell from 56.8 to 33.7 per cent of the area at risk, a 41 per cent relative reduction (p = 0.017). This is a genuine large-animal result and it should be reported as one.

With seventy-five minutes of ischaemia, at the same dose, there was no significant reduction at all (p = 0.541).

And in both arms, the circulating markers of myocardial injury — troponin I and heart-type fatty acid binding protein — showed no significant reduction. The authors state plainly that by these clinically relevant measures the peptide provided no benefit.

Read together, those three results say something more useful than any of them alone. The effect is real, it is measurable by tissue staining in a species with a human-like heart, and it is fragile with respect to the one variable that is least controllable in a patient — how long the artery has been shut. Human door-to-balloon times are rarely sixty minutes.

The timing sensitivity recurs in the rodent work in a different form. In rats, 84 µg/kg was effective given before ischaemia and failed when given after ischaemia had begun; only 252 µg/kg — three times the dose — worked in the delayed protocol. And when the same group looked at the brain injury that follows cardiac ischaemia, treatment during ischaemia or at reperfusion gave no neuroprotection at all, and even at the tripled dose never reduced blood–brain-barrier breakdown or brain oxidative stress.

The vascular work is more straightforwardly positive. In apolipoprotein E-deficient mice on a high-cholesterol diet for sixteen weeks, the analogue HNGF6A reduced proximal aortic plaque area from 0.09 to 0.01 mm² — six of seven untreated animals had plaques against two of six treated — restored endothelium-dependent relaxation, and cut nitrotyrosine staining from 12.3 to 3.2 per cent of plaque area, without lowering cholesterol and without any direct vasoactive effect (Oh et al., 2011).

One human tissue finding sits awkwardly beside that and is reported here rather than buried. In thirty-four carotid endarterectomy specimens, humanin was higher in the unstable, symptomatic plaques than in the stable ones — 29.4 against 14.1 per cent of plaque area. That is consistent with humanin being induced by stress rather than preventing it, and it is a pattern that will recur throughout the human data.

10An effect on the liver that is not in the liver

The metabolic work contains the single most surprising mechanistic result in the humanin literature, and it is one that ought to change how the compound is thought about.

Under hyperinsulinaemic–euglycaemic clamp in rats, humanin infused into the cerebral ventricles increased the glucose infusion rate required to hold blood sugar steady — the standard readout of improved insulin sensitivity — principally by suppressing the liver’s own glucose output, which fell from 12.1 ± 0.9 to 4.7 ± 0.4 (Muzumdar et al., 2009).

Infused intravenously instead, the analogues reproduce the effect. The obvious inference is that circulating humanin acts on the liver.

It does not. Blocking STAT-3 in the hypothalamus — in the brain, not the liver — completely abolished the hepatic effect of intravenously administered humanin analogue. And humanin has no direct effect on isolated primary hepatocytes. The peptide reaches the brain, the brain instructs the liver, and a related study showed the instruction travelling by way of the vagus nerve: vagotomy blocked the effect of both intravenous and intracerebroventricular humanin on hepatic triglyceride secretion.

In Zucker diabetic fatty rats, a single intravenous bolus of 100 µg of the analogue HNGF6A lowered blood glucose by nearly half over four hours in a group of five animals. A related experiment showed that the analogue acts directly on the β-cell as well, increasing glucose-stimulated insulin secretion from islets of both normal and diabetic mice in a manner independent of the ATP-sensitive potassium channel. And in non-obese diabetic mice — an autoimmune model — humanin normalised glucose tolerance after six weeks, reduced the severity of islet infiltration, and over twenty weeks delayed or prevented the onset of diabetes.

Two negatives complete the picture. The analogue does not require IGFBP-3 binding to improve insulin sensitivity, since the non-binding analogue works — which detaches this arm from one of the founding mechanisms. And in the one detailed diabetic-nephropathy study to report a full cytokine panel, humanin significantly reversed two of seven measured cytokines and left the rest, including every anti-inflammatory one, unchanged.

11Everywhere else

Beyond brain, heart and metabolism, humanin and its analogues have been reported protective in reproductive tissue, kidney, lung, retina, cochlea, bone, skin, intestine and liver. Read as a set this is impressive; read carefully it is a statement about the breadth of preclinical interest, not about the breadth of demonstrated benefit, and this monograph reports it as the former.

The reproductive work is the most mechanistically informative, because it is where the receptor question was tested most rigorously and gave a different answer from the brain — gp130 and IL-27R and EBI-3, but not CNTFR. It also produced a rare functional endpoint: in mice with chemotherapy-induced premature ovarian insufficiency, S14G-humanin restored litter size from 4.6 ± 1.1 to 6.2 ± 0.8 against a control value of 8.0 (p = 0.029). Offspring counted is a harder endpoint than a stain.

The same literature contains a useful demonstration of specificity. Humanin alone had no effect on baseline germ-cell apoptosis — it acts only where death has been provoked. Its protection is stage-selective within the seminiferous epithelium. And the structure–function work in this tissue found the membrane-receptor route dominant and the Bax-binding route minor, with one analogue — HN-L12A — acting as an outright antagonist that blocked wild-type humanin’s protection.

The retinal work illustrates why the organ atlas needs its evidence tier attached. Humanin analogues protect retinal pigment epithelial cells in culture across many studies. When taken into a live animal — the Royal College of Surgeons rat, a genuine model of retinal degeneration — twice-weekly injection for four weeks produced no difference from sham on electroretinography at either dose tested. Optokinetic visual acuity improved at the high dose only. That is a partially negative in-vivo result behind a largely positive in-vitro literature, and both belong in the record.

Elsewhere the pattern repeats with variations: benefit in acute lung injury and in sepsis models; benefit in septic and diabetic kidney injury, including a recent mechanism in which humanin blocks mitochondrial Z-DNA leakage and the necroptosis it triggers; protection of cochlear hair cells against aminoglycoside damage in explant; prevention of glucocorticoid-induced bone growth impairment; prolonged anagen in hair follicles. In a haemorrhagic shock model the effect partitioned: humanin-G improved lung architecture in every group but reduced neutrophil infiltration only in animals with an intact AMPKα1, which is a clean demonstration that the peptide works through more than one pathway and that they can be separated.

12The tumour problem

A peptide whose principal action is to stop cells dying raises an obvious question, and the humanin literature contains two answers that point in opposite directions.

The reassuring literature is real and it is specific. In immunocompetent mice bearing melanoma, S14G-humanin rescued germ cells and white cells from cyclophosphamide while lung metastases were further suppressed by the combination than by chemotherapy alone. In mice bearing human medulloblastoma xenografts, the analogue did not blunt temozolomide’s effect on the tumour while protecting the animals from its toxicity. In a neuroblastoma model, bortezomib slowed bone growth from 0.19 to 0.09 mm/day and S14G-humanin restored it to 0.15 mm/day without interfering with the anticancer effect. Similar results exist for glucocorticoid-induced growth impairment and for doxorubicin cardiotoxicity.

The other literature is equally real and comes from different tumours and different groups.

In experimental triple-negative breast cancer, doxorubicin up-regulated humanin in the tumour cells. Exogenous humanin impaired the drug’s antitumour and anti-metastatic effect and accelerated tumour growth — and humanin given on its own, without any chemotherapy, significantly increased spontaneous lung metastases. Silencing humanin did the reverse, reducing viability and enhancing chemosensitivity. The authors’ own conclusion is quoted here because it is the field’s, not this document’s: “caution should be taken when using exogenous HN to treat degenerative diseases” (Moreno Ayala et al., 2020).

In glioblastoma the case is stronger still, and recent. Humanin is abundantly expressed in human glioblastoma, largely by the tumour-associated myeloid cells; nanomolar concentrations drive temozolomide resistance through gp130 and accelerate formation of the blood–tumour barrier that keeps chemotherapy out; immunodepleting humanin from the medium abolishes the pro-tumour effect; and blocking gp130 restores the drug’s efficacy (Cheng et al., 2024). A separate group found that chemotherapy up-regulates both humanin and FPR2 in glioblastoma cells, that the analogue HNGF6A boosts chemoresistance, and that silencing humanin reverses it. A third reported that humanin drives invasion through integrin αV, and that mice given humanin in an orthotopic intracranial model survived less long than controls (Ha et al., 2024). Their conclusion is that in glioblastoma humanin is a therapeutic target, not a therapeutic agent.

These two bodies of work are not, strictly, in contradiction: they used different tumours. The protective studies come predominantly from melanoma, medulloblastoma and neuroblastoma models, several from collaborating groups; the adverse studies come from breast, glioblastoma, gastric, pituitary and bladder systems, from groups with no overlap. In human gastric cancer, humanin isoforms were the most highly overexpressed genes found in a screen against normal gastric tissue. Knocking down the rat homologue in a pituitary tumour delayed its growth and improved survival.

Two mechanistic observations recur across the adverse literature and are what make it more than a disagreement about models. The first is that chemotherapy itself induces humanin in tumour cells — the tumour reaches for the peptide when attacked. The second is that silencing humanin makes tumours more chemosensitive, which is the reciprocal experiment and gives the same answer.

Two further adverse signals belong beside it. Humanin and MOTS-c are elevated in senescent human fibroblasts, and administering them increased components of the senescence-associated secretory phenotype — the inflammatory programme that senescent cells impose on their neighbours. A commentary on that work is titled, flatly, “Mitochondrial-Derived Peptides Exacerbate Senescence”. And across the whole preclinical corpus there is no dedicated toxicology, safety-pharmacology or repeat-dose study for humanin or any analogue. Every statement that the peptide caused no harm is a secondary observation inside a study designed to measure something else.

13Measuring humanin, and why the numbers cannot be compared

This section comes before the human findings rather than after them, because every one of those findings has to be read through it.

Humanin in human blood has been measured in something over thirty studies. All of them used an antibody. None used mass spectrometry. There is no published measurement anywhere that reads the peptide’s actual sequence out of a human sample.

That matters more here than it would for most peptides, for the reason set out in section 04: the human nuclear genome carries thirteen humanin-like sequences, and the products of two of them differ from humanin by a single amino acid. The most explicit statement of the problem in the primary literature is worth giving at length, because it is the field’s own assessment and not this document’s:

“There are many nuclear-encoded MTRNR2-like genes that produce humanin-like peptides, which have only one, two, or four different amino acids compared to mitochondrial-encoded humanin. Because cross-reactivity involving humanin and all humanin-like peptides with commercial humanin antibodies has not been completely confirmed to date, we still cannot definitively distinguish the sources — nuclear or mitochondrial — of peptides that are recognized as humanins.” (Kim et al., 2021)

The same paper closes the obvious escape route. Measuring the transcript instead does not help, because the reading frames sit inside ribosomal RNA genes and the standard reverse-transcription primers cannot tell a messenger RNA from the ribosomal RNA it is embedded in.

There is also a direct demonstration that the antibodies are promiscuous. In the course of characterising a different protein, the Keio group found that human calmodulin-like skin protein — a 146-residue molecule with no sequence relationship to humanin — “was cross reactively recognized by the Humanin antibody” in immunoprecipitation.

The practical consequence is visible in the numbers themselves. Reported concentrations in healthy or control adults, all purporting to measure the same twenty-four-residue peptide, span more than thirtyfold.

That spread is wider than any disease effect reported anywhere in the human humanin literature. A difference between patients and controls of thirty per cent, measured on one platform, cannot be interpreted against a between-platform disagreement of three thousand per cent.

Two further details complete the picture, and both come from the assay developers rather than from critics. The most-used in-house assay was raised against S14G-humanin, the synthetic analogue, not the native peptide. And its own methods sections concede that while plate-to-plate variability is under ten per cent at any one time, “the drift over time is greater” — which invalidates comparison across the years over which most of these cohorts were collected.

No study in this corpus compares two humanin immunoassays on the same samples, cross-validates any of them against mass spectrometry, or publishes recovery and dilution-linearity data for endogenous humanin in plasma. That absence is the reason the literature does not converge, and it is the cheapest thing the field could fix.

14Age, centenarians and growth hormone

Three claims about humanin circulate far beyond the specialist literature: that it declines with age, that it is high in the children of centenarians, and that it is high in people whose growth hormone receptors do not work. All three appear in review abstracts without qualification. All three deserve their evidence stated.

The age claim is contradicted by the largest dataset in the field.

The source of “humanin declines with age in humans” is a single sentence in a 2009 paper reporting that levels decreased with age in mice and in humans. The paper gives no human sample size, no age range and no concentration values; the only human datum is an unlabelled scatter in one figure panel. Subsequent papers from the same laboratory restate it rather than re-measuring it.

Against that, an Italian group measured humanin in 693 people aged 21 to 113 and found it increased in old age, with the highest levels in centenarians — and, notably, inversely correlated with survival among the oldest subjects (Conte et al., 2019). A second cohort of 569 from the same group found humanin positively correlated with age. A third study found resting humanin positively correlated with age. A fourth found no significant change either way.

The two literatures use different assay platforms, which is very likely part of the explanation, and neither has been reconciled with the other. The honest statement is that the direction of humanin’s relationship with human age is unresolved, and that the widely repeated version rests on the weaker evidence.

The centenarian claim rests on eighteen people and did not replicate. In the Einstein longevity cohort, circulating humanin was significantly higher in eighteen offspring of centenarians than in nineteen age-matched controls. The abstract describes the difference as “much greater”; no effect size is reported, only a graph. The paper is explicit that centenarians themselves were not compared, for want of a control group. When an Italian group ran the same comparison in 102 centenarian offspring against 92 controls, the peptide that came out highest in the offspring was FGF21 — humanin did not differ.

The growth hormone finding is the most mechanistically interesting and the smallest. In children being evaluated for short stature, plasma humanin was uncorrelated with peak stimulated growth hormone but strongly negatively correlated with IGF-I (r = −0.69). Treating them with recombinant growth hormone lowered humanin by about a fifth. And in an Ecuadorian cohort with growth hormone receptor deficiency — Laron syndrome, a condition associated with striking protection from cancer and diabetes — humanin was 80 per cent higher than in matched relatives (Lee et al., 2014). Mouse data in the same paper align: long-lived growth-hormone-deficient dwarfs have more humanin, short-lived growth-hormone-transgenics less.

The Laron result is the one most often quoted, and it is six people against six, on an assay whose own authors flag long-term drift, in a paper whose senior authors declare stockholding in a company founded on this peptide class. It is a genuinely interesting finding and it is not a robust one.

The cross-species work is more persuasive than any of the human cohorts. Humanin overexpression extends lifespan in C. elegans, dependent on the FOXO transcription factor daf-16; transgenic mice show overlapping phenotypes; middle-aged mice treated twice weekly with the analogue showed improved metabolic measures and lower inflammatory markers; and humanin levels are stable across the lifespan of the naked mole-rat, a species with negligible senescence (Yen et al., 2020). Whether any of that transfers to people is exactly what the human data above fail to establish.

15The biomarker record

Read with the assay caveat attached, the human observational literature describes a molecule that moves with illness in almost every system examined — usually downward, sometimes upward, occasionally both.

Cardiovascular. In 327 patients, humanin fell stepwise from controls through unstable angina to myocardial infarction. In 40 patients undergoing coronary function testing, those with endothelial dysfunction had lower humanin than those without. Against that, plaque tissue shows the opposite direction, with more humanin in the unstable lesions.

Diabetes and pregnancy. In 225 subjects, serum humanin was lower in type 2 diabetes and correlated negatively with glycated haemoglobin, glucose and triglycerides. It is lower in gestational diabetes. In polycystic ovary syndrome the result is unusually informative: serum humanin was significantly lower in 40 patients than in 40 matched controls, while humanin measured in their vastus lateralis muscle showed no difference at all — a dissociation the authors read as evidence that the circulating fall is not of muscle origin.

Kidney. In haemodialysis patients followed for two years, the relationship with cardiovascular mortality is U-shaped: both the lowest and the highest humanin concentrations carried elevated risk. Adding humanin to five established risk models improved discrimination.

Lung. In stable chronic obstructive pulmonary disease, humanin was higher than in smokers without disease, and high humanin independently predicted malnutrition, worse walking distance, and future exacerbations — severe ones with a hazard ratio of 3.4. This is a well-conducted prospective study in which more of the protective peptide predicted worse outcomes.

Mitochondrial disease is where a widely repeated statement inverts. The claim that humanin is “decreased in MELAS” traces to a paper whose MELAS data are cybrid cell lines and mitochondrial-DNA-depleted cells, not patients. In actual patient tissue the finding is the reverse: humanin is increased in MELAS skeletal muscle, staining strongly in every ragged-red fibre, and is expressed in the muscle of patients with chronic progressive external ophthalmoplegia. A peptide induced by mitochondrial crisis is what those biopsies show.

Cancer is contradictory. Two studies report humanin elevated in breast cancer; a third reports it decreased. In prostate cancer, higher humanin was associated with lower risk in European-American men and the association was null in African-American men, with a significant interaction.

Alzheimer’s disease deserves its own paragraph, because the contradiction there is usually presented as brain-versus-blood biology and is actually a disparity in evidence quality. The claim that humanin is elevated in the Alzheimer’s brain rests on one brain, one immunoblot, no quantification, in 2002. The claim that it is reduced in cerebrospinal fluid rests on seven subjects, in a paper whose text and figure legend disagree about how those seven were split. The only claim with a real cohort behind it is that plasma humanin is lower in Alzheimer’s disease, in 120 patients — and a 2025 study designed to test it found that transcript levels differed while plasma protein concentrations did not discriminate Alzheimer’s from mild cognitive impairment. Circulating humanin is not an established Alzheimer’s biomarker.

16Two letters in the mitochondrial genome

The strongest human evidence about humanin is not pharmacological. It is genetic, and it is strong precisely because nobody administered anything.

rs2854128 sits at the very end of the humanin reading frame. It does not change any amino acid; it changes the stop codon from TAA to TAG. Carriers have about 14 per cent less circulating humanin. The variant is present in roughly 42 per cent of Caucasian-Americans and is much rarer in other groups. Tested against cognitive scores in 10,158 participants of a national ageing study, it showed a significant association, with a race interaction: among African-American carriers the effect corresponded to about two years of additional cognitive age, against roughly two months in Caucasian-American carriers (Yen et al., 2018).

P3S is a change of one base at position 2,639 that turns the third residue from proline to serine. It is present in about 12 per cent of a cohort of 146 female centenarians, principally Ashkenazi, against under 0.2 per cent in other populations. Within those centenarians it was strongly enriched among the ones carrying APOE4, the major genetic risk factor for Alzheimer’s disease — six of twenty APOE4 carriers had it, against about seven per cent of the rest. Humanin P3S binds APOE4 far more tightly than wild-type humanin does: dissociation constants of 0.69 nM against 12.7 nM by surface plasmon resonance, which is the single most rigorous binding measurement in the humanin literature. In mice carrying both APP/PS1 and human APOE4, the P3S peptide reduced brain amyloid more than wild-type humanin did (Miller et al., 2024).

The published commentary on that paper supplies the counterweight, and this monograph gives it equal space. P3S is a haplogroup marker: it defines mitochondrial haplogroup N1b, and by the commentator’s arithmetic roughly 9.5 million non-Ashkenazi people carry it — in whom no benefit has been demonstrated. The same commentary points out that S14G-humanin, the most potent synthetic analogue in the whole literature, occurs naturally in haplogroup U6a7a1a, and that nobody has shown carrying it is advantageous either (Logan, 2024).

Set against these two positives is a clean, well-powered negative that is rarely cited. When all thirteen nuclear humanin-like gene regions were looked up in a large coronary artery disease meta-analysis, no association survived correction for multiple testing, and no significant expression quantitative-trait loci were found.

17Nobody has taken it

The following is reported as a registry finding and nothing more. Every search below was run directly against the registry on 3 August 2026, not taken from any paper’s summary.

ClinicalTrials.gov returns seven records for humanin. Every one of them measures humanin as a biomarker or an outcome; none administers it. Three are observational studies of acute kidney injury or cardiac surgery. Two are interventional, but their interventions are anaesthetic agents and an exercise programme respectively, with humanin as the endpoint. One is an observational study in polycystic ovary syndrome and one a long-running study of cerebral palsy muscle biopsies.

Searching for colivelin returns zero records. Searching for S14G-humanin returns zero records. The European Union Clinical Trials Register returns zero for humanin, for colivelin and for MOTS-c. The single relevant record on the ISRCTN registry is observational, with humanin as an assayed outcome.

No human being has been given humanin, S14G-humanin or colivelin in a registered clinical trial anywhere, at any time. Twenty-five years after the discovery, with a preclinical literature spanning more than a dozen organ systems, the compound has never entered a first-in-human study.

Two further facts complete the regulatory picture. Humanin has no marketing authorisation anywhere, for any indication — searches of the United States drug label and approval databases return no match. And one completed study that did measure it, a 106-patient investigation of humanin in atrial tissue and plasma after cardiac surgery, closed in 2019 and has never posted results; no corresponding publication could be located.

Finally, the residual gaps in this search are stated rather than glossed. The World Health Organization’s international registry portal could not be queried directly because its search interface has been retired; its three largest constituent registries were queried individually instead. And the United States orphan-drug designation database did not return an interpretable result for the query submitted, so orphan status is recorded here as unverified, not as absent.

18Sold, unstudied, and forming amyloid

The facts in this section are commercial and regulatory. None of them is evidence of efficacy or of safety, and none is presented as such.

Humanin and S14G-humanin are available for purchase. Two distinct channels exist and they should not be confused. The first is the ordinary laboratory-reagent trade: humanin analogues appear in the catalogues of established chemical suppliers, sold to research institutions as reagents, which is unremarkable and is how most of the studies in this document were done. The second is the research-chemical market, where the same peptides are offered as lyophilised powder in five- and ten-milligram vials, labelled “research use only” and “not for human consumption”, alongside websites that index them together with dosing content.

Set that against what has been established. There is no approved indication in any jurisdiction. There is no completed trial of any phase in which the compound was given to a person. There is no published human pharmacokinetic data — no measurement of what happens to injected humanin in a human body over time, at all. There is no published human safety or tolerability data. The only pharmacokinetic work in the corpus is in rodents.

And there is no formal toxicology. Across the whole preclinical literature — hundreds of animal studies over twenty-five years — this corpus contains no dedicated safety-pharmacology study and no repeat-dose toxicology study for humanin or any analogue. The reassuring observations that exist are real but they are incidental: that the analogue did not worsen dystrophic muscle in one study, that it did not prolong bleeding time in another, that it did not blunt an anti-inflammatory drug’s effect in a third. Each is a secondary endpoint inside an experiment designed to measure something else.

To this must be added the structural finding from 2025 described in section 06: humanin forms amyloid-like β-sheet fibrils, and the residues required for fibrillation are the residues required for secretion. What that implies for a peptide injected repeatedly into a living body has not been investigated by anyone.

The gap between the commercial availability of this molecule and the evidence base underneath it is, on the numbers assembled here, among the widest in the peptide market.

19What would have to be true

It is easy to end a document like this with a list of promising directions. More useful is to state what would actually have to be established, and in what order, for the claims made about humanin to become claims about people.

First, the measurement. The single cheapest and most consequential experiment the field has not done is to quantify endogenous circulating humanin by mass spectrometry, in a way that distinguishes the mitochondrial twenty-four-residue product from the twenty-one-residue form and from the one-residue-different peptides encoded by the nuclear paralogues. Until that exists, every human observational finding in section 15 rests on an uncalibrated proxy, and the thirtyfold disagreement between platforms cannot be adjudicated. A head-to-head comparison of the commercial kits on shared samples would cost very little and would resolve a great deal.

Second, the origin. The clean genetic demonstration that humanin is translated inside mitochondria was done in the rat because the rat lacks the confounding nuclear copies. The equivalent experiment in human cells — mitochondrial-DNA depletion, cytoplasmic translation blockade, and a readout that can tell the mitochondrial product from the nuclear ones — has never been performed. Everything the compound is named for depends on its answer.

Third, the receptor. Four accounts coexist, partitioned by experimental system rather than by evidence. What would settle it is not another system: it is one system in which the candidate receptors are all present and are knocked out in turn, with a measured binding constant attached. The field has no dissociation constant for humanin at any surface receptor, which is an unusual position for a molecule twenty-five years old.

Fourth, the tumour question. The reassuring chemoprotection data and the adverse tumour-promotion data used different malignancies, and no study has tested them against each other. The two facts that recur in the adverse literature — that chemotherapy induces humanin in tumour cells, and that silencing it restores chemosensitivity — are not incidental findings but a coherent mechanism, and they were reported by groups with no stake in the peptide as a drug. Any programme contemplating systemic administration to humans has to address this first, and the field’s own reviewers have said so in print.

Fifth, which animal findings are most likely to travel. On the evidence assembled here, the cardiac work is the strongest candidate: it has a plateaued dose–response, a mechanism, and a large-animal replication. It also carries the clearest warning, because that replication failed at seventy-five minutes of ischaemia and produced no change in circulating injury markers in either arm. The metabolic work is mechanistically the most interesting — a peripherally injected peptide acting on the liver through the hypothalamus and the vagus — and is the least likely to be reproduced by simply giving the peptide to a person, precisely because the route matters. The neurological work is the largest and the least secure, resting heavily on models of familial disease and containing a total failure in the one model of sporadic disease that was tried.

And what would refute the framing. If mass spectrometry fails to detect endogenous humanin in human plasma at the concentrations the immunoassays report, the entire human biomarker literature becomes a literature about something else. If the nuclear paralogues turn out to be the translated source, humanin remains a real and interesting peptide but stops being a mitochondrial-derived one, and the conceptual claim it founded — that mitochondria signal outward — would have to be re-established on MOTS-c and the others, half of which fail the same evolutionary test humanin passes.

20References

Generated from verified NCBI records rather than from recall. Author lists, journal names, volumes, pages and identifiers are taken from the PubMed record for each identifier, not written out by hand. The build refuses to run if any cited identifier fails to resolve.

Two retracted papers were excluded by a gate. This compound's harvest contains two primary papers that were retracted in 2016 and are still indexed as live records, both concerning S14G-humanin — the analogue that carries most of the animal literature. Nothing in the index links a retracted paper to its notice automatically. Their identifiers are held in a barred list that stage 05 checks before it will emit a reference list at all.

Sources without a PubMed record

Sequence databases, structural databases, trial registries and the mitochondrial reference genome have no PubMed record and are listed separately, so that the generated list above remains wholly machine-verified.

21How this document was assembled

The corpus was built against project 05, the Therapeutic Peptide Research Library. For this compound the interesting part of the arithmetic is how much had to be discarded, and why — because the name humanin looks like the safest retrieval key in this series and is not.

Four ways to retrieve the wrong document. The first is an overloaded abbreviation. HNG is the standard short form of S14G-humanin, so admitting it looks obviously correct; measured against the 823-record indexed surface, 113 records — 14.5 per cent — were admitted on that string alone and not one was about this peptide. They were papers on hybrid nanogenerators, human nasal gland cells, human neutrophil gelatinase, helium-rich natural gas, home nasogastric tubes, handgrip exercise and hydrogel poultry nutrition. Matching without regard to case made it worse, admitting HnG from a 1975 paper on mortality in Leghorn chickens. Removing the abbreviation cost nothing at all: every genuine S14G-humanin paper in the harvest names humanin in full somewhere in its title or abstract. The external check is the one that matters — humanin was described in 2001, and after the correction the earliest surviving record in the corpus is the 2001 paper itself.

The second is the reverse case: the right molecule in the wrong role. Colivelin is sold as an off-the-shelf STAT3 activator, and across pharmacology it is used as a positive-control reagent rather than studied. Of 176 records naming it, 161 never name humanin at all — work on evodiamine, berberine, coptisine, asiaticoside, tofacitinib and several cancers, in which the peptide appears once, in a methods sentence, to switch a pathway back on. Those 161 are genuinely about the molecule in the sense that they use it; they are not documents about it. The remedy is corroboration that is specific to the humanin family, because the obvious corroborators — STAT3, JAK2, ERK — are exactly the pathway the reagent is being used to probe, and would have admitted every one of them.

The third is a property of the corpus rather than of the molecule. Biomedical prose is saturated with the bigram human in — human in vitro, human in vivo, in human in-vivo studies — and wherever text extraction, de-hyphenation or column reflow drops that space, the token humanin appears with clean word boundaries on both sides. No lookbehind can help, because the left-hand boundary is real; the discriminator is the word that follows.

The fourth is an embedded superstring, and here it denotes different molecules entirely. Humanin-like peptide, small humanin-like peptide and the thirteen nuclear paralogues MTRNR2L1 to MTRNR2L13 all contain this compound's name in full. One of them, MTRNR2L12, has acquired an independent literature in which humanin is never the subject.

The numbers. Every file with a document extension in the project's stores was opened — 47,292 of them — and its extracted text searched with the same matcher used everywhere else in the pipeline, so that the sweep and the harvest cannot drift apart. 402 files were admitted. Of those, only 141 were peer-reviewed scientific full texts. 157 were vendor and commercial pages, 56 were this project's own earlier monographs and internal write-ups, 39 were reference sources, and 9 were bibliographic search caches. Reporting the larger number as a corpus would be true and useless.

That last category is worth one sentence of its own, because it was mis-classified at first. A source is normally classified by the store it came from; two files in a peer-reviewed store turned out to be saved database query results holding 500 and 261 bibliographic records, and between them they contributed 1,150 phantom printed pages — seventeen per cent of the apparent reading corpus. One of them was a cache for MOTS-c, the sibling compound, which names humanin throughout and therefore passed the identity gate legitimately. A container can hold more than one kind of thing, so the classifier now tests the shape of the file as well as its location.

The external harvest. A scoped PubMed query returned 823 records, of which 525 survived the identity screen — a rejection rate of thirty-six per cent on a compound whose name has no dictionary homograph. Because PubMed indexes only titles, abstracts and subject headings, a second route searched PubMed Central's full text and returned 1663 matches, of which 1352 were invisible to the first route. Stage 03 fetched the union: 1712 documents.

The far-side screen. Of those 1712 fetched documents, 537 never named the compound in their retrieved body at all — they had been returned because they cite a paper about it. 7 were about a sibling mitochondrial-derived peptide and 2 named something else by the same string. 690 mentioned it below the substantive-use threshold and were counted but not read, and 19 yielded no body text at all. That last group is not a retrieval failure: a substantial number of the deposited records in this field carry the abstract only, and several of the papers most central to the mechanism are among them.

The reading corpus is therefore 494 unique scientific full texts, approximately 8,109 printed-page equivalents. It is a union keyed on identifier rather than a sum: 104 documents were present in both the local library and the external harvest and are counted once.

What was derived rather than cited. The humanin reading frame, its position in the mitochondrial genome, the twenty-four-residue and twenty-one-residue translations, the molecular formulae and masses, and the codon arithmetic for the S14G and P3S variants were computed for this document directly from the revised Cambridge Reference Sequence, and checked against the corresponding sequence-database record. They are stated here as derivations, not attributed to any paper.

What was checked against the instrument rather than a summary. Every trial-registry and regulatory statement in section 17 was obtained by querying the registry directly on the compilation date. Where a registry could not be queried — the World Health Organization's international portal has retired its search interface — the fact is stated rather than converted into a negative finding, and the three largest constituent registries were queried individually instead.

22Evidence handling

Findings are labelled by the kind of study that produced them, in the sentence that reports them, and the label is taken from the source record rather than inferred. Six tiers are used: human interventional, human observational, animal in vivo, animal or human in vitro, in silico, and review or commentary. Registry and market facts are their own category and are never converted into outcome claims.

For this compound one of those tiers is empty. There is no human interventional evidence of any kind. That is not a gap in this document's retrieval; it is a fact about the compound, established in section 17 by direct registry search, and it is the reason no human dose appears anywhere in these pages.

Six molecules are kept apart throughout. Humanin, S14G-humanin, HNGF6A, AGA-(C8R)HNG17, colivelin and rattin differ in potency across roughly nine orders of magnitude, and a result obtained with one is never reported as a result about another. The same separation is enforced against MOTS-c and the small humanin-like peptides, which share this compound's reviews, its laboratories and much of its conceptual frame, and which are the subject of their own work in this series.

Conflicts are presented as conflicts. Where the literature disagrees — on whether circulating humanin rises or falls with age, on which receptor mediates its effects, on whether it is elevated or reduced in breast cancer, on whether it protects tumours or spares only normal tissue — both sides are given with their sample sizes and their methods, and where one side is substantially better evidenced that is said plainly rather than left for the reader to infer. Where a widely repeated claim rests on weaker evidence than its circulation suggests, the evidence is stated: an unlabelled figure panel, a single immunoblot, seven subjects, six subjects per group.

Negative and failed studies appear beside the positive ones, in the same section, not quarantined into a late paragraph. The failure of the analogue in the streptozotocin model sits in the neurological section; the collapse of the porcine infarct result at seventy-five minutes sits beside the result it qualifies; the absence of any change in circulating troponin sits with both.

Values that could not be verified are named as unverified rather than quietly omitted. Where a number appears in secondary sources but could not be traced to a primary record — the cause of the discoverer's death, the demographics of the patient whose tissue produced the founding library, a dissociation constant that is described as high affinity but never measured — the document says so.