Metadata as a NAM: from the library on the floor to virtual control groups

1. The library on the floor

Imagine a library where every book has been pulled off the shelves and left in piles on the floor. The knowledge is all there — every volume, every page. But the catalogue is gone. The shelves have no labels. The books have no spines facing outward. You can stand in that library and know the answer exists somewhere in front of you, while being completely unable to retrieve it in any useful timeframe.

This is the situation most preclinical research data is in today. Experiments are conducted. Results are recorded. Animals are used. Files accumulate on shared drives and local servers. But the context that would make those files into reusable evidence — who ran the assay, which strain, under which housing conditions, using which version of the protocol, at what time of year — is often missing, inconsistent, or locked in someone's memory rather than in a machine-readable record.

The analogy is not rhetorical. A library becomes a knowledge resource through exactly the same mechanism as a preclinical dataset: structured descriptors attached to each item, organized under a shared schema, with persistent identifiers that allow any authorised user to retrieve the right item given the right query. Metadata is the cataloguing system. Without it, even the largest collection of data is not a knowledge resource — it is a pile.

2. What metadata actually is — and what it is not

Metadata is not the same as data. This distinction matters practically, not just philosophically. Data is the measurement: the latency value, the RFID pulse count, the blood corticosterone concentration, the video frame. Metadata is the frame that defines what that measurement means and under what conditions it was obtained: the species, the strain, the sex, the age at the time of measurement, the housing density, the cage position in the rack, the assay version, the operator identity, the date and time of day, the device model and firmware version.

A behavioral result recorded as "latency to find platform: 42 seconds, group difference p = 0.03, n = 10 per group" is data. The same result recorded alongside the strain (C57BL/6J), sex (male), age (12 weeks at treatment), protocol version (Morris water maze SOP v3.2, 25°C water temperature, 60-second ceiling), experimenter (blinded; Zeitgeber time ZT7–10), and deposited with a DOI and a CC BY 4.0 license is evidence. The difference is not in the number. The difference is in whether the number can be verified, replicated, or compared against results from another study.

The five fields that most consistently determine whether a preclinical dataset is reusable are: the animal (species, strain, sex, age, genotype), the cage (type, housing density, rack position, bedding), the assay (protocol name, SOP version, apparatus model, parameters), the operator (experimenter identity, blinding status, training record), and the date (procedure date, time of day, light-phase, season) [4][5]. These are not comprehensive — they are minimal. Missing any one of them is sufficient to block cross-study comparison in most scenarios.

3. FAIR: the standard for data that outlasts the experiment

In 2016, Wilkinson and colleagues published a four-word standard for making scientific data assets useful beyond the experiment that generated them: Findable, Accessible, Interoperable, Reusable [1]. The FAIR principles are now central to funding body requirements, journal data policies, and reproducibility initiatives worldwide. They are not a checklist; they are an integrated system.

Findable means the data has a globally unique persistent identifier and that its metadata is machine-readable and deposited in a searchable repository. Accessible means the data and its metadata can be retrieved by anyone using a standard protocol, under defined access conditions where required. Interoperable means the data uses shared vocabularies and ontologies that allow it to be integrated with other datasets without bespoke translation work. Reusable means the data has clear provenance, a license, and sufficient documentation that a second researcher can understand and act on it without contacting the original team [1].

Most preclinical datasets currently satisfy none of these properties fully. They may satisfy reporting requirements — a paper describing the study is published, the methods section describes the strain and protocol — but the data itself is not deposited, not machine-readable, not findable by a system rather than a person who already knows it exists. The gap between reporting and FAIR is not small. Reporting is the floor. FAIR is the ceiling that actually enables reuse.

The operational implication is that FAIR must be designed into the protocol at the same time as the hypothesis. Post-hoc curation — attempting to reconstruct metadata after the study is complete — cannot recover what was never recorded. Cage position in the rack is not in the data files. Firmware version of the recording device is not in the paper. The experimenter who ran the cohort during a week they were covering for a colleague is not in the methods section. These are the fields that determine VCG eligibility, and they cannot be fabricated retroactively [4].

4. Unusable data wastes animal lives — the WellFAIR argument

The ethical dimension of data quality is underacknowledged. If a preclinical dataset cannot be reused — because its metadata is incomplete, its files are in a proprietary format, its provenance is unverifiable — then the animals in that experiment contributed to a dead end. The experiment happened. The animals were used. But the evidence cannot be built on, cannot be included in a meta-analysis, cannot serve as a historical control for a future study. The scientific contribution is lost.

Petit-Demoulière and Huzard (2026) articulate this directly in a paper framing the WellFAIR concept: data welfare is animal welfare [2]. The argument is not metaphorical. It follows from the 3Rs framework — Replacement, Reduction, and Refinement — that has governed animal research ethics since the 1960s. If FAIR metadata enables historical control data to be reused in a subsequent study, fewer control animals are needed. If FAIR metadata enables a dataset to be cross-compared with studies from other sites, a replication experiment may be unnecessary. If FAIR metadata enables confounds to be detected and corrected, fewer refinement-motivated follow-up experiments are required [2].

The structural overlap between FAIR and the 3Rs is not coincidental. Both frameworks are asking the same question: how do we extract the maximum scientific value from each animal used? Reporting requirements answer part of this question — the study is described well enough to read. FAIR answers the rest — the data is structured well enough to reuse. The WellFAIR ecosystem proposed by Petit-Demoulière and Huzard frames this as a four-stage workflow: planning the metadata schema before the study begins; acquiring context at the source during the study; federating across studies using shared vocabularies; and producing outcomes — reusable, citable, VCG-eligible evidence — rather than isolated results [2].

5. Virtual control groups: what makes historical data a valid comparator

A virtual control group is not the same as historical data. This is the most important practical distinction in the NAM framing of metadata. Historical data is what you have from past experiments. A virtual control group is historical data whose comparability has been verified — through metadata completeness, contextual matching, and statistical correction — for a specific experimental comparison [3].

The European Medicines Agency's 2023 public consultation on virtual control groups explicitly positioned VCGs as innovative New Approach Methodologies supporting the 3Rs [3]. The EMA's framing is precise: historical control data that is selected, contextually matched, statistically corrected, and prospectively validated for a specific use case can reduce the number of concurrent control animals needed in drug development studies. But validity requires metadata. The matching process cannot proceed without knowing, for each historical animal, the strain, age, sex, room, welfare status, assay version, diet, cage type, housing density, time window (season, batch), and operator identity [3][5].

Each of these ten eligibility dimensions is a filter. A historical dataset that is missing even two or three of them cannot be confidently included in a VCG pool without introducing uncontrolled confounds. A rack position missing from the metadata means an uncontrolled ventilation gradient. An operator identity missing means an uncontrolled handling-stress effect. A firmware version missing means an uncontrolled sensor calibration confound. The metadata gaps documented as most common in Home Cage Monitoring studies — cage position, housing density, experimenter identity, device firmware, light cycle timing — are exactly the fields that VCG eligibility algorithms require [4].

The Pistoia Alliance's minimal metadata set for repurposing nonclinical in vivo data (MNMS) makes the same operational point from a different direction [5]. Their recommendation is that the metadata schema should be minimal and mandatory rather than comprehensive and optional. A small, enforced checklist that every study actually completes is worth more than a thorough framework that most studies ignore. The MNMS is aligned with ARRIVE 2.0 and FAIR, and its design reflects the practical reality that metadata compliance scales only when compliance is cheap and unavoidable [4][5].

6. AI does not fix missing context — it exposes it

A recurring assumption in conversations about AI for preclinical research is that language models and machine learning pipelines will eventually be able to infer missing metadata from the data itself. This assumption is wrong in a specific and important way. AI amplifies available structure; it cannot fabricate experimental conditions that were never recorded.

A well-structured dataset with complete, schema-consistent metadata becomes dramatically more powerful when an LLM can query it, extract candidate eligibility criteria, match across studies, and propose VCG pools for human review. An underspecified dataset with missing fields, free-text protocol descriptions, and no persistent identifiers remains inert regardless of the model's capabilities. The model cannot tell you which rack row the cages were on if that information was never captured. It cannot tell you whether the experimenter that week was the principal investigator or a rotating technician if operator identity was not logged. Garbage in, garbage out remains the most durable principle in computational science.

The practical consequence is that AI tools for preclinical evidence — whether for VCG matching, meta-analysis, or regulatory evidence synthesis — should be treated as infrastructure for well-structured data, not as a rescue mechanism for poorly structured data. The value they add scales directly with metadata completeness. This is why the data infrastructure argument is upstream of the AI argument, not downstream from it.

7. From sensors to evidence: the DVC implementation path

One practical implementation path for automated metadata capture in Home Cage Monitoring is the Digital Ventilated Cage system by Tecniplast. The DVC® integrates RFID tracking, load cells, and sensor arrays to capture locomotion, feeding, and individual presence data automatically across standard IVC racks at facility scale [6]. Neuronautix is a listed scientific partner of Tecniplast.

What the DVC® illustrates is a structural point about metadata and data. The raw signal from the DVC® — load cell voltages, RFID pulse timestamps, beam-break counts — is data. It becomes evidence only when it is framed by the metadata that defines the conditions under which it was collected: the animal registry (strain, sex, DOB, cohort ID), the study design (protocol version, treatment dates, group assignments), the housing record (cage type, rack position, housing density, bedding, cage change schedule), and the device record (firmware version, calibration date, rack ID). The sensor handles the data layer automatically. The metadata layer requires deliberate schema design and capture infrastructure — and it must be in place before data collection begins.

Metadatapp provides the interface layer for this kind of schema-driven, controlled-vocabulary metadata capture. The principle is the same regardless of which HCM platform or ELN is in use: structured capture at the source, from protocol design to file deposit, using a schema that maps onto the fields required for cross-study comparison and VCG eligibility.

8. Multi-site federation: what shared metadata makes possible

The payoff of the metadata investment becomes clearest at scale. Consider three datasets from three different sites — a CRO, an academic laboratory, and a pharma facility — each using a different HCM system (DVC®, LMT, DOME), collected in different years, by different operators. These datasets are heterogeneous in instrumentation, site, and timing. They cannot be meaningfully combined without shared metadata labels.

If all three datasets document the same strain (C57BL/6J), sex (male), age at treatment (10 weeks), SOP version (v2.1), and time window (aligned by batch), a VCG pool can be constructed from their control arms that is more statistically powerful than any single-site control group, more representative of biological variance across facilities, and defensible in regulatory submissions [3]. The instrumentation heterogeneity becomes a manageable covariate rather than an insurmountable obstacle. The metadata is what makes the federation possible.

This is the argument for metadata as the infrastructure for NAMs. Virtual control groups are the NAM — EMA positions them explicitly in that regulatory framing [3]. Structured, FAIR metadata is what makes VCGs scientifically credible, auditable, and useful. The metadata does not replace the animals already used; it makes their contribution compound across studies, institutions, and years, rather than dissipating after a single publication.

What this means in practice

The practical steps that follow from this argument are not complex, but they require decision-making at the start of a study rather than its end. Define the metadata schema — which fields, in which format, using which controlled vocabularies — at the same time as the experimental hypothesis. Identify which fields need to be captured automatically (device and rack metadata), which need to be captured at the protocol level (SOP version, housing conditions), and which need to be captured at the individual animal level (DOB, cohort ID, group assignment). Build or adopt a capture interface — structured forms, ELN templates, or a system like Metadatapp — that enforces the schema rather than suggesting it.

The return on this investment is not only regulatory. Reusable data reduces the time and cost of future meta-analyses. Findable data attracts collaboration. Interoperable data makes it possible to answer a question across five studies in a day rather than in a grant cycle. Reusable data makes each animal's contribution permanent rather than provisional. That is what it means to treat data welfare as animal welfare — and it is available now, with current technology, in any laboratory that decides the metadata question deserves as much attention as the experimental design question.

A companion presentation covering the full 33-slide visual argument — from the library analogy through FAIR, WellFAIR, virtual control groups, and the DVC implementation example — is available at neuronautix.com/presentations/2026-05-metadata-nam-slideshow/.