The ethical debt of preclinical research: when poor data stewardship wastes animal lives

An implicit contract between every animal and science

We accept the use of live animals only under strict ethical conditions: the question must matter, the suffering must be minimized, the number of animals must be justified, and the knowledge generated must be robust enough to justify the cost. This is the moral architecture behind the 3Rs: Replacement, Reduction and Refinement [2,3].

But there is a fourth obligation that preclinical research has too often treated as optional: the obligation to care for the data. A mouse can be handled carefully, housed properly, monitored ethically, and included in a protocol approved by every committee — and still be ethically wasted if the data it generates are incomplete, poorly annotated, impossible to reuse, or lost after a single publication. This is what we should call the ethical debt of preclinical research.

It is not fraud, and not necessarily negligence. It is often the result of normal scientific habits: data saved on a hard drive, metadata scattered across notebooks, behavioural labels defined differently from one laboratory to another, negative results left unpublished, and protocols described just enough to pass peer review but not enough to enable reuse. The problem is that normal is no longer defensible.

The animal is not only in the cage — it is also in the dataset

The recent article “Data welfare is animal welfare: Building a WellFAIR research ecosystem” by Petit-Demoulière and Huzard makes the argument explicit: data stewardship is not merely a technical concern but part of the ethical responsibility attached to animal research [1]. The paper proposes the WellFAIR ecosystem: a model where animal-derived data are planned, captured, structured, federated and reused in ways that directly support the 3Rs [1].

A preclinical experiment does not end when the animal is euthanized, when the paper is submitted, or when the grant report is filed. It ends only when the information generated from that animal has been preserved with enough context to be understood, audited, reused and integrated into the broader scientific record. Without that, the animal’s contribution is artificially shortened.

Modern preclinical research produces more data than ever before — home-cage monitoring, video tracking, telemetry, high-resolution behavioural phenotyping, omics, imaging and automated physiological systems all generate vast volumes of data. But volume is not value. Data without metadata are not evidence; they are biological residue. A locomotor trace without strain, sex, age, housing condition, device configuration, light cycle, protocol version and analysis pipeline is weak evidence; a raw video without provenance is not an asset but an archaeological object. The ethical issue is simple: when data cannot be reused, animals may be used again.

The 3Rs are not only about the number of animals used today

The 3Rs were introduced by Russell and Burch in 1959 as the foundation of humane experimental technique [2]: Replacement avoids or substitutes animal use where possible, Reduction minimizes animal numbers while preserving validity, and Refinement reduces pain, distress and lasting harm. But Reduction is often interpreted too narrowly. It is not only about calculating the smallest sample size for one study; it is also about making sure that each animal contributes maximally to cumulative knowledge.

A study that uses the “correct” number of animals but produces unusable data has not truly respected Reduction. It may have reduced animals locally while increasing animal use globally, because future researchers must repeat what could have been reused. A missing metadata field may look trivial, but if it prevents a dataset from being included in a meta-analysis, harmonized across laboratories, used as historical control data, or integrated into a machine-learning model, the consequence is a lost opportunity to reduce future animal use.

When negative and neutral findings remain unpublished or inaccessible, other laboratories may unknowingly repeat the same experiment, expose new animals to the same procedures, and rediscover the same absence of effect [4]. Publication bias is therefore not only a statistical problem. In animal research, it is also a welfare problem.

Poor reproducibility is not only inefficient — it is ethically expensive

Freedman and colleagues estimated that irreproducible preclinical research costs the United States alone roughly US$28 billion per year [5]. Begley and Ellis famously argued that preclinical cancer research needed higher methodological standards after industry scientists failed to reproduce many landmark findings [6]. Those concerns are valid, but they are incomplete.

In animal research, irreproducibility also means that animal lives may have been spent without producing reliable knowledge. This does not mean that every failed study is unethical — science includes uncertainty, and negative findings and biological variability are part of honest research. The ethical problem is not failure; it is preventable non-informativeness. The moral question is not “Did the experiment work?” but “Did the experiment leave behind reliable, reusable evidence proportional to the animal cost?”

ARRIVE and PREPARE were necessary — they are no longer sufficient alone

The ARRIVE guidelines improved the field by defining what should be reported in animal research publications [7]. PREPARE moved the discussion upstream, emphasizing planning, dialogue between scientists and animal facilities, and quality control before the study begins [8]. But a publication checklist is not the same thing as a reusable data ecosystem: a paper can be ARRIVE-compliant and still leave behind a dataset that cannot be integrated into future work.

This is why the FAIR principles matter. FAIR means Findable, Accessible, Interoperable and Reusable [9]. In preclinical research, FAIR should not be treated as an administrative label added at the end of a project; it should be a condition of ethical data generation. FAIR is not simply “open data”: sensitive, proprietary and industrial datasets may require controlled access. FAIR means that data and metadata are structured so authorized humans and machines can find, interpret and reuse them under appropriate governance.

The key point is that FAIR must begin before the experiment. Retrofitting metadata after the fact is often fragile, incomplete and sometimes impossible: once personnel leave, devices are replaced, software versions change and informal laboratory knowledge disappears, the context of the experiment decays. The dataset may still exist, but its scientific meaning evaporates.

Metadata are the memory of the animal experiment

The 2024 Minimal Metadata Set (MNMS) paper made this problem operational, proposing a minimal metadata set to repurpose nonclinical in vivo data, aligned with ARRIVE 2.0 and designed to make animal-derived data FAIR-compliant [10]. This is a practical advance because it recognizes a central constraint: a useful metadata standard must capture enough to enable reuse without becoming so heavy that it collapses under its own administrative weight.

The right metadata are not bureaucratic decoration — they are the memory of the experiment. They tell future users what was done, to whom, under what conditions, with which device, according to which protocol, using which analysis, with which exclusions, and under which biological constraints. That detachment is especially dangerous in behavioural neuroscience and home-cage monitoring, where behaviour is exquisitely context-dependent: strain, sex, circadian phase, housing density, cage enrichment, prior handling, device layout, group composition and analysis definitions can all change interpretation [13]. Without controlled vocabularies, ontologies and shared metadata, a behavioural event called “social interaction” in one system may not map cleanly onto the same label in another, and apparent comparability may be an illusion. This is why WellFAIR is a call to make animal-derived data Born-FAIR: structured at the point of capture, not rescued after the project is over [1].

Shadow data: the invisible evidence base

One of the most important ideas in the WellFAIR framing is the value of “shadow data” [1]: pilot data, baseline measures, control-group observations, negative findings, welfare observations, device calibration outputs, failed assay attempts, environmental records and routine phenotyping. Much of this information is scientifically useful. Almost none of it is systematically preserved.

Shadow data can help detect laboratory-specific effects, improve experimental design, support historical controls, refine endpoints, train AI systems, and identify patterns invisible in single studies — in an era where AI and statistical modelling increasingly interrogate biological systems, it may become one of the most valuable assets in preclinical research. But only if it is curated. Unstructured shadow data are not a hidden treasure; they are a landfill. Structured shadow data, by contrast, can become an ethical multiplier: one animal’s data can inform many future analyses.

Virtual control groups make the ethical stakes concrete

Virtual control groups use historical control data, statistical modelling and sometimes AI-based approaches to reduce or replace concurrent control animals in future studies — now an active area in toxicology and regulatory science [11]. The VICT3R project, for example, aims to reduce animal use in toxicology by developing virtual control groups based on high-quality historical control data [11].

The promise is substantial: fewer animals used as controls, better reuse of existing data, and more efficient study designs. But the prerequisite is non-negotiable — virtual control groups require structured, high-quality, well-described datasets. No metadata, no virtual controls. This makes the ethical debt visible: every poorly annotated control animal from the past is a missed opportunity for Reduction in the future, and every historical dataset that cannot be reused because context was lost is an animal-welfare opportunity that expired.

The culture problem: data care is still not rewarded

The barriers are not only technical; they are cultural. Preclinical researchers are rewarded for papers, grants and novelty, and rarely for clean metadata, reusable datasets, controlled vocabularies, negative results, protocol versioning or long-term data governance. Data stewardship is often treated as extra work rather than core scientific labour. Institutions should therefore stop treating data care as a personal virtue and start treating it as infrastructure — a shared responsibility of animal facilities, research institutes, funders, journals, technology vendors and ethics committees.

Ethics committees should ask not only how many animals will be used, but how the resulting data will be preserved, described and made reusable. Funders should evaluate whether a data-management plan is realistic, resourced and connected to standards. Journals should require structured availability of protocols, metadata and analysis code. Technology providers should support interoperable formats, APIs, ontologies and provenance tracking rather than exporting black-box tables. And laboratories should capture metadata from the first day of the experiment.

Toward a new ethical standard: no animal-derived data without context

The field does not need another slogan; it needs a new default: no animal-derived data without context. Every dataset generated from live animals should carry enough metadata to support interpretation, auditability and possible reuse. This does not require that every dataset be fully open to everyone — it requires that data should not become useless because we failed to describe them.

This standard would transform ethical review. The harm-benefit balance of animal research depends on the expected knowledge gain — but if the data are not reusable, the knowledge gain is artificially limited, and poor data stewardship therefore weakens the ethical justification of the study after the fact. That is the ethical debt: the unpaid obligation created when animals contribute to research but the research system fails to preserve the full value of that contribution.

Paying the debt

The solution is not to shame individual scientists. Most researchers are working inside systems that were not designed for data reuse. The solution is to redesign the system. A practical WellFAIR approach would include:

• Define a minimal metadata set before the experiment begins.
• Capture metadata automatically where possible, directly from devices, ELNs and analysis pipelines.
• Use controlled vocabularies and ontologies for animals, procedures, devices, behaviours and outcomes.
• Preserve raw or minimally processed data with provenance.
• Publish or register negative and neutral findings where possible.
• Use federated models when data cannot be openly centralized.
• Assign data stewards as scientific collaborators, not administrative afterthoughts.
• Make data quality part of animal ethics, peer review and institutional evaluation.
• Build repositories that support reuse, meta-analysis, AI-readiness and virtual control groups.
• Treat every dataset as part of a cumulative evidence base, not as the disposable exhaust of a single paper.

The principle is not radical — it is the logical extension of the 3Rs into the data age. If we believe that animal research is justified only when it produces meaningful, reliable and necessary knowledge, then data stewardship is part of the ethical contract. The animal’s life cannot be reused. Its data can. And if we fail to make that possible, we are not only wasting information — we are wasting lives.