The source and its scope
Gangwal and Lavecchia (2025) is an open-access review in Drug Discovery Today titled "Artificial intelligence in preclinical research: enhancing digital twins and organ-on-chip to reduce animal testing". It argues that machine learning, deep learning, AI-powered digital twins and AI-enhanced organ-on-a-chip platforms can simulate biological systems well enough to support early safety and efficacy decisions, in line with the 3Rs of replacement, reduction and refinement.
It is a narrative review. It brings together technologies, tools and companies across a fast-moving field, which makes it a useful map for a newcomer. It does not report a systematic search or a quantitative synthesis, so it cannot tell the reader how strong the evidence for each tool is. That is the main task of this critique.
What the review covers
The review opens with the limitations of animal studies, including species differences and false positive and false negative safety findings, and notes that the FDA Modernization Act 2.0 allows sponsors to submit non-animal data. It then surveys existing alternatives: in vitro and three-dimensional cultures, organoids, organ chips, human tissue, ex vivo slices, microdosing, molecular modelling and digital twins.
A long section describes AI tools offered as alternatives to animal studies. These include AnimalGAN and Tox-GAN, generative models from FDA scientists that produce synthetic clinical pathology and toxicogenomic data (Chen et al., 2023; Chen et al., 2022); deep learning models for drug-induced liver injury and carcinogenicity; and several commercial platforms. Separate sections cover AI-supported digital twins and organ chips, again mixing academic and company examples.
The review closes with challenges: biological complexity that models cannot fully capture; data bias, including datasets drawn mainly from young, healthy male subjects; scarce data for rare diseases; and the black-box nature of deep learning, for which it discusses explanation tools such as SHAP and LIME. Its conclusion is measured: AI is a complementary tool rather than a full replacement, and complete phasing out of animal testing is unlikely in the near future (Gangwal & Lavecchia, 2025).
Appraising the evidence behind the claims
Different kinds of evidence are presented side by side
The review's examples rest on very different foundations. Some are peer-reviewed studies with stated methods and metrics. Others are company descriptions of what a platform can do. Presenting them together, without grading, gives vendor statements the same weight as published validation. A reader should sort every example into one of the levels in the table below before citing it.
| Level of evidence | Typical form | What it can support |
|---|---|---|
| Vendor or press statement | Company website, funding announcement | That a product exists; not that it works |
| Retrospective benchmark | Model tested on historical compounds | Discrimination on known drugs (Chen et al., 2023) |
| Prospective or blinded test | Predictions made before outcomes known | Performance on genuinely new compounds |
| Independent replication | Another group, other compounds | Transferability beyond the developer |
| Regulatory acceptance | Qualified for a stated context of use | Use in decisions for that context (U.S. Food and Drug Administration, 2025a) |
Strong examples, read carefully
The best examples are informative precisely because their limits are visible. Chen et al. (2023) trained AnimalGAN on legacy rat studies and reported that its synthetic clinical pathology data gave hepatotoxicity assessments comparable to the real data and better than conventional quantitative structure-activity relationship models. The model learns from animal data, so its contribution is to reuse existing studies and reduce new ones, not to make animal data unnecessary. Its predictions also concern rat outcomes, so their relevance to people depends on how well rat findings translate in the first place.
Organ chips have a similar profile. Ingber (2022) reviews their development for disease modelling and drug testing. In one of the largest performance studies, Ewart et al. (2022) tested 870 human Liver-Chips against 27 drugs with known hepatotoxicity and reported 87 per cent sensitivity and 100 per cent specificity. This is strong evidence for one organ, one endpoint and one platform; it does not show that a multi-organ chip, or an AI layer on top of it, performs as well.
The review itself acknowledges the practical limits of organ chips: current devices approximate particular tissue structures, physiological relevance requires stable function over long periods, and standardisation and reproducibility between laboratories remain obstacles (Gangwal & Lavecchia, 2025). These are measurable problems, and they are where much of the useful doctoral work in this area lies.
A contested statistic
The review supports the case against animal studies with a figure for the proportion of animal-tested drugs that fail to reach the market. Figures of this kind are widely repeated, but they refer to candidates entering clinical trials and do not by themselves show that animal models caused the failures. Sun et al. (2022) attribute clinical failures mainly to lack of efficacy (40% to 50%), unmanageable toxicity (about 30%), poor drug-like properties (10% to 15%) and commercial or strategic reasons (about 10%). Better preclinical models would help with some of these causes, not all.
What counts as a digital twin
The term "digital twin" is used loosely in the review and across the field. The US National Academies define a digital twin as computational models coupled to a physical counterpart and updated dynamically through two-way data flows, and stress verification, validation and uncertainty quantification as foundations (National Academies of Sciences, Engineering, and Medicine, 2024). Many preclinical "digital twins" are better described as mechanistic or statistical simulation models: they are built from data but are not continuously updated from a specific organism or patient.
A scoping review of digital twins in health found the field active but still at an early stage of adoption (Katsoulakis et al., 2024). For a doctoral researcher, the practical point is to define the term in your own work and to say which elements of a twin your model has: a physical counterpart, data flowing in, predictions flowing back, and quantified uncertainty.
Regulation in 2026: from permission to credibility
The regulatory context has moved since the review was written. In the United States, the FDA Modernization Act 2.0 of December 2022 replaced references to animal tests with "nonclinical tests", which include cell-based assays, organ chips, computer models and animal studies (Consolidated Appropriations Act, 2023). The FDA's April 2025 roadmap aims to make animal studies the exception for preclinical safety testing, beginning with monoclonal antibodies (U.S. Food and Drug Administration, 2025b), and a direct final rule in September 2026 aligned the wording of FDA drug and biologics regulations (U.S. Food and Drug Administration, 2026). In the European Union, animal procedures remain governed by Directive 2010/63/EU, which requires non-animal methods to be used wherever they are scientifically satisfactory (European Parliament & Council of the European Union, 2010).
Permission to use non-animal evidence is not the same as acceptance of a particular model. The FDA's January 2025 draft guidance on AI in drug development proposes a risk-based credibility assessment, judged against the model's context of use, defined as the specific role and scope of the AI model in addressing a question of interest (U.S. Food and Drug Administration, 2025a). It excludes AI used purely in drug discovery, but covers models whose outputs bear on the reliability of non-clinical or clinical study results. The EMA's reflection paper, adopted in September 2024, covers AI across the medicinal product lifecycle from discovery to post-authorisation (European Medicines Agency, 2024). Both expect evidence proportionate to the decision the model informs, which is exactly what a narrative review cannot supply.
Overall appraisal
Strengths
- Broad, readable coverage that links AI methods, organ chips and digital twins in one place.
- A balanced conclusion that does not overstate the prospect of full replacement.
- Explicit discussion of data bias, data scarcity and interpretability as barriers.
- Useful pointers to peer-reviewed tools from regulatory science, such as AnimalGAN and Tox-GAN (Chen et al., 2023; Chen et al., 2022).
Weaknesses
- No reported search strategy or inclusion criteria, so coverage and selection cannot be judged.
- Vendor claims sit beside peer-reviewed results without any grading of evidence.
- Performance figures are reported without the information needed to interpret them, such as the test set, whether it was independent, and the prevalence of the outcome.
- The term "digital twin" is not defined strictly enough to distinguish twins from simulation models (National Academies of Sciences, Engineering, and Medicine, 2024).
- Validation and regulatory acceptance are mentioned but not analysed against a framework such as context of use (U.S. Food and Drug Administration, 2025a).
Worked example: appraising a claim
This example is hypothetical. Suppose a company states that its AI-enhanced liver chip "predicts human liver toxicity better than animal studies". Work through the following questions before citing it.
- Source. Is the claim in a peer-reviewed paper with methods, or only in marketing material?
- Context of use. Which decision is the model meant to inform: ranking candidates, setting a first-in-human dose, or replacing a specific study (U.S. Food and Drug Administration, 2025a)?
- Test set. How many drugs, with what reference classification, and were they chosen before testing?
- Comparator. Was the animal model tested on the same drugs, or is the comparison drawn from other studies?
- Metrics. Are sensitivity and specificity both reported, with uncertainty, as in Ewart et al. (2022)?
- Independence. Has anyone other than the developer reproduced the result?
- AI contribution. Does the AI layer add predictive value over the chip read-outs alone?
If most answers are missing, cite the claim as an example of a proposed approach, not as evidence of performance.
Research gaps and common mistakes
Gaps a doctoral project could address
- Prospective, blinded evaluation of AI-enhanced organ chips on compounds not used in development.
- Direct comparison of non-animal and animal predictions on the same compound set against human outcomes.
- Methods for quantifying uncertainty in organ-chip and in silico predictions (National Academies of Sciences, Engineering, and Medicine, 2024).
- Bias audits of training data for sex, age and ancestry, which the review identifies as a concern (Gangwal & Lavecchia, 2025).
- Credibility assessments of a specific model against a stated context of use (U.S. Food and Drug Administration, 2025a; European Medicines Agency, 2024).
Common mistakes
- Citing a narrative review as proof that a technology works.
- Repeating company performance figures as if they were peer-reviewed results.
- Calling a simulation model a digital twin without the features that define one.
- Assuming that regulatory permission to submit non-animal data means a given model will be accepted.
Key takeaways
- Gangwal and Lavecchia (2025) is a useful map of AI-enhanced non-animal methods, and its own conclusion is that AI complements rather than replaces animal studies.
- Grade every example by level of evidence; vendor statements and peer-reviewed benchmarks are not equivalent.
- The strongest organ-chip and generative-model evidence is specific to one organ, endpoint or species (Ewart et al., 2022; Chen et al., 2023).
- A digital twin requires a physical counterpart and two-way data flow, not only a simulation (National Academies of Sciences, Engineering, and Medicine, 2024).
- Regulators now ask how credible a model is for a stated context of use (U.S. Food and Drug Administration, 2025a; European Medicines Agency, 2024).
