MUMultiple speciesBmoderateT1hypothesis only
Only about 5 per cent of therapies tested in animals ever obtain approval. What animal studies say about humans and horses, and what they do not.
Executive summary
Animal studies generate hypotheses, not clinical evidence. An overview of 122 systematic reviews found that roughly half of the therapies tested in animals ever reach a human study, and that only about 5 per cent obtain approval. A finding transfers only when mechanism, exposure, outcome measure and population match up, and when the study was randomised, blinded and reported in full.
Key points
Transferability is not a feeling. It is a chain of four tests. Internal validity asks whether the measured result can be traced to the treatment at all within the study; randomisation, allocation concealment and blinded assessment are what secure it. External validity asks whether the same result holds outside that one experimental set-up, so at a different age, under different husbandry, in a different sex, in a different season. Construct validity asks whether the model reproduces the disease one actually wants to treat. Predictive validity asks, last of all, whether a positive animal result predicts a positive result in the target species. If any one of the four tests fails, the chain breaks, and nothing in the result itself will show it.
Because a mechanism establishes only that an effect is possible, not its size and not its direction in a living organism. Between receptor and recovery lie distribution, breakdown, counter-regulation, concurrent disease and time. A shared mechanism is a hypothesis; clinical benefit is a finding. The two must never be fused into a single sentence.
The most robust estimate comes from a 2024 review that pooled 122 systematic reviews covering 54 diseases and 367 therapeutic approaches. Roughly half of the approaches tested in animals reached a human study at all, around 40 per cent a randomised clinical trial and about 5 per cent regulatory approval. The road is long: a median of five years to the first human study, seven to the randomised trial, ten to approval. At the same time, the same analysis found concordance of positive results between animal and human of around 86 per cent.
A broad systematic scoping review of 121 analyses, each comparing at least two species including humans, arrived at a very different picture: the transfer success rates reported there ranged from 0 to 100 per cent. The authors judged the risk of bias in the included work to be high and much of the material to be out of date, and concluded that the success of a transfer can barely be predicted in advance.
A systematic review compared six interventions whose effect in humans had been settled beyond doubt with the corresponding animal data. For the free radical scavenger tirilazad, the animal models reduced infarct volume by around 29 per cent and improved neurological scores by around 48 per cent; in patients with ischaemic stroke, the same substance produced a worse course. Corticosteroids in traumatic brain injury looked favourable in the animal model and showed no benefit in clinical testing. Conversely, thrombolysis and bisphosphonates pointed the same way in both worlds; for antenatal corticosteroids the direction held for respiratory distress in the newborn, while the effect on mortality remained open in the animal model. The direction is therefore often right, but not reliably so.
A success shrinks along the way as well. A systematic review of the oncolytic virus talimogene laherparepvec, approved in 2015, traced five preclinical studies in 150 animals and seven clinical studies in 589 patients. Preclinically, complete regression was reported in up to 100 per cent of injected and 80 per cent of non-injected tumours; the highest complete response rate achieved clinically was 24 per cent, and in one study 0 per cent. Every preclinical study carried a high or unclear risk of bias.
Before reaching for biology, it is worth looking at method. An overview of 31 systematic reviews of animal experiments spanning stroke, spinal cord disease, tumours, multiple sclerosis and emergency medicine found this: only around 29 per cent of the studies reported randomisation, around 15 per cent allocation concealment and around 35 per cent a blinded assessment. In the main analysis, missing randomisation went together with markedly larger effects; in the secondary analyses, randomisation, allocation concealment and blinding lowered effect sizes above all where the outcome was a judgement rather than a measurement.
An analysis of 549 studies in 9,535 animals on acute spinal cord injury puts a precise figure on the price. Across all interventions, an improvement in neurological recovery of around 26 per cent was reported. More than half of the studies, 51.2 per cent, described no blinded assessment, and the effect thus inflated came out 7.2 percentage points higher. The publication bias analysis suggested that between 2 and 41 per cent of the experiments were never published; counting them in lowers the estimated effect by 0.9 to 14.3 percentage points.
| Methodological feature | Share of studies reporting it | Observed consequence |
|---|---|---|
| Randomised allocation | around 29 per cent | Missing randomisation went together with markedly larger effects |
| Allocation concealment | around 15 per cent | Where present, smaller effects, above all with subjective outcomes |
| Blinded assessment | around 35 per cent | Without blinding, the effect in the spinal cord analysis came out 7.2 percentage points higher |
| Full publication | an estimated 2 to 41 per cent of experiments remain unpublished | Counting them in lowers the effect by 0.9 to 14.3 percentage points |
The second block is biology, and it concerns exposure above all. A review of dose extrapolation between species records that simple scaling by body size fails wherever a substance binds strongly to proteins, is extensively metabolised or actively transported, is excreted in the bile, or wherever target structures differ between species in quantity, binding strength and distribution. The same amount per kilogram then does not mean the same concentration at the site of action, and the same concentration does not mean the same response. That is why no dose is ever converted from one species to another on this platform.
The third block is paradoxical: too much order. A simulation study based on 440 preclinical studies of 13 interventions in models of stroke, myocardial infarction and breast cancer compared single-laboratory designs with multi-laboratory designs. Single-laboratory studies mostly failed to estimate the true effect size accurately, and larger animal numbers made the estimate less accurate still, because they measure a narrow special case ever more precisely. Just two to four participating laboratories raised accuracy by up to 42 percentage points, without needing more animals. The bottleneck is not the species; it is the narrowness of the sample.
Since 2010 the ARRIVE guideline has set out the minimum information a publication reporting an animal experiment must contain. The revised version of 2020 arranges 21 items and singles out ten of them as an indispensable core, among them the species and origin of the animals, group size, randomisation, blinding and the full results. Numerous journals and funding bodies endorse it; the authors of the 2020 version note themselves, however, that the hoped-for improvement in reporting quality has failed to materialise since 2010.
Whether that improves the reports has actually been tested in randomised fashion. In a trial covering 1,689 manuscripts submitted to a large journal, half of the author teams were asked to complete an ARRIVE checklist. Not one manuscript in either group met the requirements in full; the only thing that improved was the description of housing conditions, from 52.1 to 74.1 per cent. A comparison of journals with and without official endorsement of the guideline found around 60 per cent of items fully reported in both groups in 2015, and no difference between the groups; it was precisely the bias-relevant information that was poorly reported.
The situation is not hopeless. A cross-sectional analysis of 943 animal experiments taken from the journals that originally published the guideline shows a clear trend across three time windows: the share of poorly reported studies fell from 46.05 to 26.06 and finally to 9.55 per cent. Even so, not a single study was reported in full. Not one of the 943 covered all 38 sub-items, and only in the most recent window did as much as 0.25 per cent reach the grade ‘excellent’, with none at all in the two earlier ones. Reports came out better where the journal made compliance mandatory.
A completed checklist is not a seal of quality. Check in the full text whether randomisation, blinding and sample size planning really are described. If they are missing, the result is not wrong, but it carries less weight than the abstract suggests. The phrase ‘reported in accordance with ARRIVE’ is no substitute for that check.
The horse occupies a rare dual position. For articular cartilage it counts, among the models in common use, as the animal whose cartilage and subchondral bone thickness comes closest to that of the human being; defects can be created and assessed arthroscopically, and loading after treatment can be steered under control. In the tendons the parallel is closer still: the superficial digital flexor tendon of the horse becomes diseased spontaneously and is regarded as the counterpart of human Achilles tendinopathy, tears of the deep digital flexor tendon as the counterpart of rotator cuff tears within the tendon sheath. These diseases arise of their own accord; they are not induced.
Being a good model does not, however, release the horse from the burden of proof. In a randomised, blinded study in 17 horses with an induced lesion of the superficial digital flexor tendon, a locally injected microRNA lowered collagen 3 activity in the tissue after two weeks; the cross-sectional area of the lesion came out smaller and the histological score better than under sham treatment. In a controlled trial in six horses, by contrast, mesenchymal stromal cells produced a temporary increase in inflammation and lesion size between weeks 3 and 6, and after 24 weeks no clear advantage over serum treatment alone could be found.
The same weaknesses apply in equine medicine as in preclinical research. A systematic review with meta-analysis of platelet-rich plasma for joint disease in the horse found 21 publications, of which only five could be pooled; the pooled estimate came out clearly in favour of treatment, yet most studies were rated as highly prone to bias, and the authors call for randomised, blinded comparative studies. A systematic review of the treatment of equine sarcoids included ten studies: the reported regression rates ranged from 28 to 100 per cent per lesion and from 9 to 100 per cent per horse, with a risk of bias running from ‘some concerns’ to ‘critical’. No recommendation for any particular treatment could be derived from it.
The following eight questions can be answered of almost any publication, and they place a result faster than any debate about whether animal experiments make sense.
Anyone wanting to proceed more systematically will find, in the risk of bias instrument for animal studies available since 2014, ten ordered items adapted from the established tool for clinical trials. It covers selection, performance, detection, attrition and reporting bias, and makes the assessment traceable rather than intuitive.
The sober balance is this: animal research is no proof for the human being, and human research is no proof for the horse. Both produce knowledge that travels in the other direction only as a hypothesis. Keeping that separation clean costs little and gains precision.
Original analysis
| Claim type | What animal data can do | What animal data never do | Minimum needed in the target species | Most common fallacy |
|---|---|---|---|---|
| Mechanism | Make the signalling pathway, cell type and time course visible | Establish the relevance of the pathway to the course of the disease | Evidence that the same structure is present and active in the target species | The mechanism is passed off as an effect |
| Safety | Deliver coarse warning signals at organ level | Rule out rare and late adverse effects | Systematic recording of adverse events in the target species | ‘Unremarkable in animals’ is read as ‘harmless’ |
| Dose | Narrow down a range for the first trial | Set an effective amount for another species | Measured concentration and effect in the target species | Milligrams per kilogram are simply converted across |
| Clinical benefit | Establish a testable hypothesis | Show the benefit in another species | Randomised, blinded study in the target species | The effect size from the model is quoted as though it held clinically |
| Prevention at herd level | Isolate individual risk factors | Reproduce the interplay of husbandry, feeding and use | Observational data from the real population | A laboratory finding becomes a husbandry recommendation |
No, that is precisely what the data do not show. The same review that counts only around 5 per cent approvals finds concordance of positive results between animal and human of around 86 per cent, and several treatments taken for granted today, such as thrombolysis in stroke or antenatal corticosteroids, have been confirmed in both worlds. The criticism is not aimed at the animal model as such, but at two things: at studies run without randomisation and blinding, and at the habit of presenting a result from a model as a clinical finding.
Because they count different things. The 5 per cent refers to all the modes of action tested and asks how many of them obtain approval. The range of 0 to 100 per cent comes from a scoping review of 121 analyses, each of which defined very differently what ‘successful transfer’ even means: in one case prediction within a factor of two, in another the share of correctly predicted outcomes, in another the overlap of statistical ranges. Without a common definition, comparing the percentages is pointless.
No, and this site does not do it on principle. Conversion by body weight or body surface area fails as a rule for substances that bind strongly to proteins, are extensively metabolised or actively transported, or are excreted in the bile. On top of that come differences between species in the quantity, binding strength and distribution of the target structures. The same amount per kilogram therefore does not produce the same concentration at the site of action. Questions about use belong in the hands of the attending veterinary surgeon or the attending doctor.
It improves construct validity, that is the question of whether the model reproduces the disease at all. A spontaneously arising tendon disease in the superficial digital flexor tendon of the horse develops over years under real loading, with age, prior strain and concurrent disease, whereas an artificially created defect leaves all of that out. About the transferability of a treatment effect it says nothing, however: with spontaneous disease too, randomised, blinded studies in the target species remain the basis of any statement about efficacy.
By three signals. First: the species is not named, or only deep in the text. Second: the words randomised and blinded are missing even though a treatment is being compared. Third: the outcome is a judgement, a rating scale for instance, with no indication of who recorded it. This information is missing more often than one would expect: in an analysis of 943 animal experiments not a single study covered all the items, and even in the most recent window only 0.25 per cent reached the grade ‘excellent’. Missing information does not refute a result, but it lowers its weight.
Yes, in both directions and under the same rules. A treatment effect established in humans is, in the horse, a hypothesis to begin with, because digestion, metabolism, loading profile, lifespan and husbandry all differ. The experience of human medicine in designing clinical trials, by contrast, does transfer perfectly well, and that is exactly what specialists from both fields recommend for tendon research. The method, then, transfers more reliably than the result.
Evidence Lab
Evidence levels grade confidence, not truth: what randomised, observational and animal studies support, how GRADE rates them and where the pyramid bre
MUMultiple speciesBmoderateT0not transferable Open access
Comparative Health
Radiocarbon data suggest the Achilles tendon core forms in growth and is barely renewed thereafter: what that means for healing, scar and the equine t
MUMultiple speciesBmoderateT2limited target data Open access
Comparative Health
Serum amyloid A rose about 227-fold in an equine experiment, haptoglobin only 1.1-fold. How inflammation starts, how it ends, what CRP and SAA measure
MUMultiple speciesBmoderateT2limited target data Open access
Once a month: new syntheses, checked studies, and what has changed in the state of the evidence.