NAM data management for IND submissions: what regulators need from organoid and computational evidence

The data-management problem NAMs create

NAMs are intrinsically heterogeneous. An organoid, a PBPK model, a QSAR classifier, and an organ-on-chip device may all qualify as NAMs, but they generate fundamentally different evidence types. Without structured metadata, a reviewer may see a result but not the scientific context needed to evaluate it.

A liver-on-chip toxicity readout is not interpretable unless the submission package captures: cell origin, maturity and donor characteristics, chip architecture, media composition, flow and shear conditions, exposure duration, nominal versus measured compound concentration, protein binding, endpoint definitions, positive and negative controls, replicate structure, assay acceptance criteria, and known limitations. For an ML or PBPK model the equivalent list covers model version, training data provenance, applicability domain, feature definitions, validation split, uncertainty quantification, parameter sources, sensitivity analysis, and model-lock status.

This is where ontology-backed metadata infrastructure — the kind NAMO is designed to provide — becomes directly useful. FAIRness is a necessary condition, but it is not sufficient. NAM data management should be designed around regulatory confidence, not just FAIRness.

What FDA's March 2026 draft guidance actually requires

FDA's March 2026 draft guidance on alternatives to animal testing frames NAM validation around four principles: context of use, human biological relevance, technical characterization, and fit-for-purpose use in regulatory decision-making. This is almost exactly a data-management problem translated into regulatory language. Each principle maps to a concrete data requirement:

• Context of use: Explicitly define the regulatory question — e.g., "support hepatotoxicity risk assessment before Phase 1." This must be stated upfront and must drive every metadata and validation decision downstream.
• Human biological relevance: Structured biological metadata — human cell and tissue type, disease context, organ system, endpoint pathway, ontology-mapped terms (Cell Ontology, UBERON, MONDO). The model does not need to be perfect; it needs to recapitulate the biology relevant to the risk claim.
• Technical characterization: Protocol versioning, QC rules, acceptance criteria, intra- and inter-lab reproducibility, reference compounds, controls, deviation records.
• Fit-for-purpose: Each NAM result should be linked to a specific risk claim, an uncertainty statement, and a weight-of-evidence argument. Not "this model is generally good" — "this model, at clinically relevant free concentrations, shows no hepatocellular injury signal."

FDA has also stated explicitly that NAM data is encouraged in IND applications — including AI toxicity models, cell lines, organoids, and organ-on-chip systems — as part of its plan to phase out mandatory animal testing requirements for drug development.

Five ways better data management facilitates IND submission

Good NAM data management helps in concrete, not abstract, ways. Here are the five that matter most for IND preparation:

1. Reviewability. The IND reviewer should not need to reconstruct experimental context from disconnected PDFs, spreadsheets, CRO reports, and code files. A well-managed NAM dataset links raw data, protocol, test article, biological system, assay endpoints, statistical analysis, model outputs, and conclusions — readable as a coherent package.
2. Traceability. Under 21 CFR Part 58 (GLP), nonclinical study reports must document methods, test systems, dose regimen, statistical methods, transformations, data-quality issues, and where raw data and reports are stored. Good data management encodes this from the beginning rather than trying to reconstruct it at submission time.
3. Cross-study comparability. If every organoid or chip experiment describes cell type, organ, endpoint, exposure, and controls differently, you cannot build cross-study evidence. Ontology-backed metadata — shared terms for cell types, tissues, compounds, endpoints — enables comparison across platforms, laboratories, and disease models.
4. Weight-of-evidence reasoning. NAMs rarely replace animal data outright; they contribute to a broader safety argument. A good data model connects each NAM result to a specific claim: risk domain, organ system, exposure level, uncertainty, and the gap it fills or supports in the overall nonclinical package.
5. eCTD and study-data readiness. FDA requires commercial INDs in eCTD format. Supported study-data standards include SEND for nonclinical data, SDTM for clinical data, ADaM for analysis data, and Define-XML for accompanying metadata. SEND was designed around traditional animal studies and does not accommodate many NAM output types. An upstream harmonisation layer — what NAMO could provide — is needed before crosswalking to SEND where applicable.

Where NAMO fits in this picture

The previous note in this series introduced NAMO as a unified ontology for NAM metadata. Its role here is more specific: NAMO is a pre-submission semantic and data-quality layer that helps transform heterogeneous NAM outputs into reviewer-ready evidence packages. It is not "the IND standard." It sits upstream of eCTD, SEND, study reports, and regulatory summaries.

Practically, a NAMO-to-IND implementation would need five artefacts:

• Context-of-Use card — a structured document per NAM defining the regulatory question, organ system, risk domain, and decision supported
• NAM study metadata schema — one record per experiment: platform, biological system, test article, exposure protocol, endpoints, controls, QC
• Validation evidence matrix — concordance across reference compounds, inter-lab reproducibility, known limitations, fit-for-purpose conclusion
• Weight-of-evidence claim graph — each NAM result linked to a risk claim and uncertainty statement
• eCTD Module 4 export map — crosswalk from NAMO metadata to IND Module 4 (nonclinical pharmacology/toxicology) section structure

A critical caution worth stating plainly

Better data management does not make a weak NAM acceptable. It makes the evidence assessable. A poorly validated organoid assay with perfect metadata is still weak evidence. A strong human-relevant model with poor provenance, missing controls, no reproducibility data, and undefined acceptance criteria may fail reviewer confidence even if the underlying biology is sound.

The regulatory value of NAM data infrastructure comes from the combination of scientific validity and data integrity. One without the other is insufficient.

NAMs will not enter regulatory decision-making simply because they are innovative or animal-free. They will enter when their context of use, biological relevance, technical reliability, limitations, and decision impact are captured in a structured, traceable, and reviewable form.