Persistent clinical attrition does not establish animal models as a root cause of drug-development failures

Dear Editor,

The systematic evidence synthesis by Filipova and colleagues (Filipova et al., 2026) is aimed at identifying the root causes of clinical failures in drug development, and I write in support of its broader aim: strengthening the scientific basis for human-relevant preclinical methods. Precisely because I share that aim, I am concerned that the study’s central conclusion, that six decades of persistently high clinical attrition, of which roughly 77% is attributed to biological factors, demonstrates the inadequacy of animal models and the urgent need to replace them with New Approach Methodologies (NAMs), rests on an inferential step that the assembled data do not support. In my view the analysis conflates the biological locus of clinical failure with its cause, and in doing so risks weakening, rather than advancing, the case for NAMs. I set out these points below.

Points of agreement. I do not dispute the study’s descriptive findings, and several are important. Clinical attrition is genuinely high (∼90%) and has not materially improved over the period examined and agrees with several detailed studies, some of which are cited. Therefore, Research Question 1 has been answered. Animal models have well-documented translational limitations in specific contexts and the development and qualification of human-relevant NAMs is a legitimate and pressing scientific priority. The disagreement does therefore not concern these observations but the causal claim the authors build upon them and the identification of a pattern of bias by the authors, despite the approaches the authors have chosen to prevent it. Research Question 2 has therefore not been adequately answered and warrants further investigations based on the currently available data.

1. A biological cause of failure is not evidence that the animal model failed to predict it

Let us start with looking at the novelty the study aims to provide, the root cause of biological failures in clinical trials is attributable to the lack of translation by animal model data. The core of the argument is a logic that most candidates fail for biological reasons (safety, efficacy, pharmacokinetics) and since these are the properties assessed in animals, the animal models are non-predictive and are the primary driver of failure. That logic is not substantiated by data. A drug that fully engages a validly identified target yet fails in Phase II or III for lack of efficacy has not necessarily been mispredicted by a rodent. A biological endpoint failing in the clinic identifies where a programme failed, not why. Moreover, the ∼91% attrition is measured only on candidates that had already passed preclinical animal safety and pharmacokinetic screens; the predictive value of a filter cannot be judged from the survivors that later fail, and the compounds animal studies removed before first-in-human are, by construction, absent from the dataset. The synthesis generates no paired predicted-versus-observed concordance data and therefore cannot establish the predictive validity of animal models in either direction. it reflects an incorrect hypothesis about human disease biology. Insufficient efficacy is consistently the single largest failure mode, and the field’s own root-cause analyses attribute it predominantly to inadequate target validation and incomplete understanding of human pathophysiology, not to interspecies discordance. These are knowledge gaps that human-based NAMs (organoids, organs-on-chips, and the like) do not, in themselves, resolve, because they inherit the same uncertainty about which target and which mechanism matter in human disease. Attributing the aggregate “77% biological” wholesale to animal-model inadequacy therefore substitutes the location of failure (a biological endpoint) for its cause (which the study does not measure).

2. Much “biological” failure is attributable to human decisions in clinical development, not to species differences

The dominant biological failure mode, inadequate efficacy, is to a large degree a product of decisions taken during clinical development rather than of interspecies difference. In the largest analysis of late-stage failures, 57% of investigational drugs that failed did so for insufficient efficacy and a further 22% for commercial or strategic reasons (Hwang et al., 2016). Reviews of why trials fail repeatedly identify remediable choices: the wrong indication, poorly chosen or insensitive endpoints, inappropriate patient populations, under-powering, and inadequate dose selection (Fogel, 2018; Harrison, 2016). The efficacy–effectiveness gap shows further that trial context, not biology alone, governs apparent efficacy (Eichler et al., 2011). Conversely, when the human-side decisions improve, success improves sharply and without any change in animal methods: biomarker-guided patient selection roughly doubles the probability of approval (Wong et al., 2019), as so does genetically supported target and indication selection (Nelson et al., 2015). A failure corrected by choosing a different endpoint, indication or patient is not evidence that a rodent mispredicted human biology; it is evidence that the clinical hypothesis, or its test, was mis-specified. Attributing such failures to animal models, as the “77% biological” framing implicitly does, misassigns their cause.

3. Animal-to-human translation is, for many endpoints, considerably better than the study implies

The most robust and repeatedly replicated finding across large, independent datasets is strong positive concordance for objective organ toxicities. (Olson et al., 2000) reported greater than 80% concordance for haematologic, gastrointestinal and cardiovascular effects across 150 compounds; a roughly 3000-drug adverse-event analysis found real, endpoint-dependent concordance (Clark and Steger-Hartmann, 2018); an independent 2259-drug study using likelihood ratios reached the same conclusion (Giblin et al., 2021); and a recent open-access analysis of 7565 drugs across five preclinical species confirmed high positive likelihood ratios for many specific toxicities, for example neutropenia, anaemia, rash and gastrointestinal effects among small molecules, and infection and immune-mediated hepatitis among biologics (Liu and Fan, 2026). The IQ consortium’s non-clinical-to-clinical database likewise found that animal studies predicted human target-organ toxicities well enough to enable safe first-in-human entry (Monticello et al., 2017). At the level of whole therapies, an umbrella review of 122 systematic reviews concluded that translation is higher than commonly asserted, with about half of interventions reaching human study (Ineichen et al., 2024). Earlier paired analyses found animal and clinical results concordant in roughly half of cases, with discordance often traceable to bias in the animal studies rather than to true species differences (Perel et al., 2007). By contrast, the low figures the study emphasises rest on selective citation. The claim that only 19% of serious human adverse effects are detected in animals derives from an analysis restricted to serious post-marketing reactions (van Meer et al., 2012), events that had already evaded the entire pre-marketing programme, including large human trials, and that therefore measure the difficulty of the events, not the predictivity of animals; on that same set, pre-marketing human trials detected essentially none. Reports of poor Phase I oncology concordance similarly rest on an incomplete human reference standard and a survivorship-restricted sample (Atkins et al., 2020).

4. Persistent attrition is largely an economic equilibrium, not a readout of preclinical predictivity

That clinical attrition has remained near 90% for decades is, in any case, weak evidence about animal models, because attrition is substantially an economic equilibrium. Deliberate “fail-fast” portfolio strategies hold it high by design; regulatory and comparator bars have risen; and research has moved into progressively harder biology as tractable targets are exhausted (Paul et al., 2010; Scannell et al., 2012). Where a better preclinical tool did change practice, the fall in pharmacokinetic failures from ∼40% of attrition in 1991 to ∼10% by 2000, it shifted the composition of failure toward efficacy rather than lowering the overall rate (Kola and Landis, 2004), cautioning against the expectation that any single methodological change, NAMs included, will by itself reduce total attrition.

5. A better-supported root cause: the predictive validity of the whole decision cascade

A more compelling explanation for persistent failure is already established in the literature: the quality of the decisions taken across the entire discovery-to-clinic cascade. AstraZeneca’s “5R” analysis attributed most of its attrition to getting the target, tissue, exposure, safety and critically, the patient wrong; after re-engineering these decisions, the company raised its success rate from candidate-drug nomination to Phase III completion from 4% (2005–2010) to 19% (2012–2016), with no change in its reliance on animal models (Cook et al., 2014; Morgan et al., 2018). Decision theory explains why this lever dominates: when selecting rare successes, small changes in the predictive validity of the whole screening and model cascade, the correlation between a decision tool’s output and clinical outcome, outweigh large changes in throughput or cost and can swing productivity by orders of magnitude (Scannell et al., 2022; Scannell and Bosley, 2016). On this account the binding constraint is not the presence of animals but the predictive validity of the entire decision system: target and indication choice, translational biomarkers, trial design, and, where they help, better human-relevant models. This framing fits the evidence better than a monocausal attribution to animal testing, and it identifies interventions, including, but not limited to NAMs, that demonstrably move the number.

Conclusion

I share the authors’ goal of reducing unnecessary animal use through more predictive, human-relevant preclinical science, and I regard NAMs as an essential part of its future. The most constructive framing is the one now embodied in regulatory policy: the FDA’s 2025 Roadmap to Reducing Animal Testing in Preclinical Safety Studies (U.S. Department of Health and Human Services Food and Drug Administration, 2025), building on the FDA Modernization Act 2.0, advances NAMs not as wholesale replacements but on a fit-for-purpose, context-of-use basis. Beginning where animal models are least reliable, such as monoclonal antibodies, and adopting each method where it is scientifically validated for a defined purpose. This is precisely the logic the present analysis misses: the value of any method, animal or human-based, is established by its demonstrated predictive validity for a specific question, not by aggregate attrition. Framed this way, the human relevance of NAMs becomes a testable, incremental claim rather than a slogan. The stakes are ultimately for patients, who are served both by keeping unsafe candidates out of the clinic and by not abandoning effective ones on the basis of mispredictions. Advancing NAMs on that evidence-based, context-of-use footing, alongside better target, trial and patient-selection decisions, is the surest route to safer and more effective treatments reaching the people who need them. I would welcome the authors’ response.

Yours sincerely,

Peter MA Groenen.

CRediT authorship contribution statement

Peter MA Groenen: Writing – review & editing.

Data availability

No data was used for the research described in the article.