Peptide basics

A paper is not a trial: why animal work is not human evidence

This page is about how to read evidence, not about whether any compound works. It names compounds only as examples of citation patterns.

What a citation hides

A reference is a pointer. It records that something was published, in a journal, by authors, in a year. It does not record whether the study was done in people, in rodents, in cultured cells, or in a computational model, and it does not record whether it was original research or a review of somebody else's.⁸

That flattening is why citation counts are such poor guides in this field. Twenty references beneath a marketing claim look like twenty pieces of evidence. If nineteen are rodent studies and the twentieth is a review of the nineteen, the claim rests on one line of work whose relevance to humans is an open question rather than a settled one.

How far animal results carry

This has been measured directly. A BMJ systematic review selected six interventions where clinical trials had produced unambiguous evidence of benefit or harm, then went back to the animal experiments on the same interventions to see whether the two agreed.¹

The answers were mixed in a specific and instructive way. Corticosteroids for head injury showed benefit in animal models and none in clinical trials. Tirilazad reduced infarct volume by 29 per cent and improved neurobehavioural scores by 48 per cent in animal models of ischaemic stroke, and was associated with a worse outcome in patients. Thrombolysis and bisphosphonates agreed across both. The review's own conclusion is that discordance may be due to bias or to the failure of animal models to mimic clinical disease adequately.¹

The animal literature is skewed before you read it

A 2010 analysis in PLoS Biology examined 16 systematic reviews of interventions tested in animal models of acute ischaemic stroke, covering 525 unique publications. Only ten of those publications — 2 per cent — reported no significant effect on infarct volume, and only six, 1.2 per cent, failed to report at least one significant finding anywhere.²

A literature in which 98 per cent of papers report a positive result is not describing a world in which almost everything works. It is describing a filter. The authors' trim-and-fill analysis suggested publication bias might account for around a third of the efficacy reported in those systematic reviews, with pooled reported efficacy falling from 31.3 per cent to 23.8 per cent after adjustment.²

A companion PLoS Medicine review sets out the internal-validity problems that compound this: selection bias where allocation is not randomised, performance and detection bias where investigators are not blinded, attrition bias where deviations from protocol are handled unequally, and sample sizes too small to support the estimates drawn from them.³

Whether the finding is even stable

The Reproducibility Project: Cancer Biology set out to repeat 193 experiments from 53 high-impact preclinical papers, with protocols peer reviewed and published before any experiment ran. It managed 50 experiments from 23 papers. Of the replications that could be scored, the combined success rate was 46 per cent, and for positive effects the median replication effect size was 85 per cent smaller than the original, with 92 per cent of replication effect sizes smaller than their originals.⁴

The reasons it managed only a quarter of what it planned are their own finding. The data needed to compute effect sizes and run power analyses were publicly accessible for just 4 of the 193 experiments, and none of the 193 was described in enough detail in the original paper to design a repeat without contacting the authors.⁹

A theoretical account published sixteen years earlier had predicted where this would bite hardest: a research finding is less likely to be true when studies in a field are smaller, when effect sizes are smaller, when there is greater flexibility in designs, definitions, outcomes and analytical modes, and when more teams are chasing significance in the same area. Peptide preclinical work has most of those characteristics.⁷

And then the attrition

Even compounds that survive preclinical work and reach human testing mostly do not arrive. An analysis of more than 400,000 development-phase records estimated the overall probability of a compound entering phase 1 reaching approval in the low double digits, with the largest absolute losses at phase 3 — the last and most expensive stage, where the question finally being asked is whether the thing helps people.⁵

Stack the filters. A preclinical result may not replicate. If it replicates, the literature it sits in is skewed toward positives. If the effect survives that, it may reverse in humans. If it reaches a human trial, the trial will probably not end in an approved product. A page of animal citations is separated from a human conclusion by every one of those steps, none of which the citations have taken.¹²⁴⁵

A worked example that counted itself

In 2025 a systematic review in HSS Journal examined the BPC-157 literature from an orthopaedic sports medicine perspective. It searched PubMed, Cochrane and Embase from database inception to June 2024, screened 544 articles, and included 36 studies. Thirty-five were preclinical. One was clinical: a retrospective report of 12 patients treated for chronic knee pain. The review states that no clinical safety data were found.⁶

That ratio was produced by the review's own authors under a stated protocol, which is what makes it usable here. It is not our reading of the literature; it is a count of the literature by people who read all of it. And it describes a compound with several hundred PubMed records to its name.⁶⁸

How to tell them apart in ninety seconds

1. Open the abstract, not the citation. Study type is almost always in the first two sentences of the methods.
2. Look for the subjects. Rats, mice, dogs, rabbits, cells, or a named cell line mean the study is not about people.
3. Look for a design word. Randomised, double-blind, placebo-controlled and crossover are human-trial vocabulary; in vitro, in vivo in a named species, and knockout are not.
4. Check the publication type field on PubMed. Review, systematic review, case reports and comment are all labelled there, and a review is not new evidence.
5. If it is a human study, look for a registration number. An NCT identifier in the abstract means the plan was filed before the result existed.⁸