The translation gap: why most animal results never help people

From animals to approval. An umbrella review from Zurich and Edinburgh collected 122 systematic reviews on 367 treatments. Of the 281 treatments whose first animal study was more than 10 years old, 50% had entered any study in people, 40% a randomised trial, and 5% had been approved by the US Food and Drug Administration; for those that got there, approval came a median of 10 years after the first animal study.⁴ The same review cuts the other way too: among 62 treatments with at least five animal studies, 79% of animal studies were positive, against 62% of clinical studies and 50% of randomised trials. The authors call this 86% concordance, argue that translation may be better than often claimed, and link the low approval rate to weaknesses in both animal studies and early trials.⁴

Highly cited animal studies. Of 76 highly cited animal studies from leading journals, only just over a third were confirmed by randomised trials in people, and about one in ten treatments was approved.^5,6,8 These had a median of 889 citations; less visible animal research probably translates less often.⁸

Once trials in people start. Of 41,040 industry programmes from 2000 to 2015, 66.4% moved from phase 1 to phase 2, 58.3% from phase 2 to phase 3 and 59.0% from phase 3 to approval: 13.8% in all, and 3.4% for cancer.⁷

None of this makes animal research useless: it is how promising leads are found, and how many dangers are caught before anyone is exposed. The gap is the reason human trials exist.

Receptors in different places. GLP-1 medicines raised calcitonin and caused C-cell overgrowth in rodents, whose thyroid C-cells carry GLP-1 receptors; human and monkey C-cells had low receptor levels, and 20 months of liraglutide caused no C-cell overgrowth in monkeys Animals(Evidence level: Shown in animal studies).⁹ The semaglutide page explains what this means for people, which is not fully settled.

Immune systems. Mouse and human immune systems differ in the balance of white-cell types, in receptors that detect germs, in signalling molecules (cytokines) and in the signals that switch immune cells on.¹¹ In one comparison of gene responses to trauma, burns and bacterial toxin, the mouse models matched the human conditions little better than chance.¹² A later re-analysis of the same data, looking only at genes that changed in both species, found much closer agreement, so how well these mouse models mirror human inflammation is still argued over.¹⁶

Drug handling. How long a drug stays in the blood depends on protein binding, breakdown and transport, and all three differ between species. Drugs that bind strongly to blood proteins, are broken down extensively or carried actively, or whose targets differ between species, are unlikely to suit simple scaling by size.¹³ Some peptide medicines are built to bind blood proteins on purpose: semaglutide is more than 99% bound to albumin.¹⁰

A short life. Laboratory rats live about 2 to 3.5 years; in adulthood, one rat month corresponds to roughly three human years.¹⁴ So time does not carry over simply: how long a rat study lasted says little about what years of use would do in a person.

A model is built for a clear result: the same injury, in animals of the same age, strain and sex, treated at the same moment. That clarity makes a small study possible, and makes the result hard to carry over. The stroke review lists the causes of poor external validity: disease induced in young, otherwise healthy animals; a uniform group of animals instead of varied patients; timing and dosing impossible in a clinic; and outcome measures unlike those that matter in trials.⁸ The Zurich umbrella review names the same problem.⁴

A tendon cut in a young rat and treated within half an hour is a long way from a sports injury that has troubled someone for a year.

• Randomisation and blinding. In 290 animal studies, those without randomisation were more likely to report a positive outcome (odds ratio 3.4), and so were those without blinding (3.2).¹⁹ In most systematic reviews of animal studies, fewer than half of the studies reported random allocation, and 0 to 3% a sample size worked out in advance.⁸ In 2,671 animal studies of drugs collected for systematic reviews, 25% reported randomisation and under 1% a sample size calculation.²¹
• What gets published. Of 525 publications of animal stroke experiments, only 2% reported no significant effect on the size of the brain damage. Correcting for the studies that seem to be missing lowered the average benefit by about a third.²⁰
• What holds up. Scientists at one company could confirm only 6 of 53 landmark preclinical cancer studies.²² When 50 experiments from high-profile cancer papers were repeated, the typical (median) positive effect was 85% smaller than first reported.²³

For BPC-157, most of the animal literature comes from a single group; see the Croatian lab.

The US guidance describes how to choose the first dose for healthy volunteers “regardless of the projected clinical use”. It starts from the highest dose without observed harm in animal toxicity studies, not from a dose that worked, and it “does not address dose escalation”. Its safety margin exists because of known gaps between species: people may be more sensitive; some toxic effects, “e.g., headache, myalgias, mental disturbances”, are hard to detect in animals; receptor numbers and binding strength can differ; toxicity can be unexpected; and absorption, breakdown and excretion differ. It warns that “human bioavailability and metabolism may differ significantly from that of animals”.²⁴

The European guideline states that the expected exposure in people “is considered more relevant than the relative dose levels between animals and humans”, and asks for estimates from human cells where possible, because sensitivity can differ. For healthy volunteers, it asks that the first dose give an exposure below the level expected to have a drug effect (the pharmacologically active dose), unless there is a robust reason to go higher.²⁵ Differences in transport, breakdown and dose-response “can override the effect of size alone”.¹³

So a scaled animal dose answers one narrow question: roughly how much of this pure substance a closely watched volunteer can probably tolerate once. It does not say how much works, how much can be taken week after week without harm, or what is actually in a vial from outside the legal supply chain; tests of such products have found amounts that differ from the label (see product risk).²⁶

Six healthy young men each received a single dose into a vein of TGN1412, a monoclonal antibody designed as a superagonist of CD28 (a switch on T cells). Within 90 minutes all six had a body-wide inflammatory reaction; within 12 to 16 hours they were critically ill, with lung injury, kidney failure and widespread clotting problems Humans, limited(Evidence level: Seen in humans in case reports or small uncontrolled studies). Two needed intensive organ support for 8 and 16 days. All six survived.²⁷

The cytokine storm was “not predicted by pre-clinical safety testing”. In 2010, scientists at the UK National Institute for Biological Standards and Control reported the likely reason: in human cells TGN1412 switched on CD4+ effector memory T cells, and in the species used for safety testing these cells do not carry CD28, so the antibody could not act on them Cells(Evidence level: Shown in cells or tissue in the laboratory).¹⁵ The animals lacked the part of the human system the drug acted on.

The European guideline on first-in-human trials, first issued in 2007 and revised in 2017, asks sponsors to show that the species tested is relevant to people, to take “a cautious approach” with medicines that are highly specific to humans, to choose the starting point for the first dose (the lowest expected effect, the active dose or the highest dose without harm in animals) according to how much is uncertain, and to dose a first person and watch before dosing the rest of a group.²⁵

Some peptides sold outside pharmacies have never had a published first trial in people. For KPV, the FDA found no study in people by any route.³ For BPC-157, the developer’s early human studies survive only as conference abstracts, and the FDA found the human safety information too thin to describe its safety profile.¹

When a page or a post says a peptide “heals” or “repairs”, ask:

• Which species, and which model? A fresh injury in a young rat, treated within minutes, is the easiest case there is.⁸
• Which target, and is it the same in people? If no receptor has been identified, nobody can check.
• Randomised and blinded? Without these, positive results are more likely whether or not the effect is real.¹⁹
• Repeated by an independent laboratory, and tested in people? If not, the honest answer to “does it work?” is “unknown” (see evidence grading).

Promising animal results deserve proper trials in people. Where that is left undone for years, users carry the uncertainty; see the case for reform.