The Food and Drug Administration opened a docket on Monday asking the public to nominate approved drugs for new uses. Most coverage read it as a clinical labeling story. The line that matters sits in the third bullet, and it points back at preclinical.

Repurposed: Why the FDA’s new docket is a preclinical evidence ask in a clinical labeling envelope.

On Monday, May 11, the FDA opened  and asked clinicians, researchers, patients, and other stakeholders to nominate approved drugs for new indications. Comments are due June 11. The press release led with the patient framing. “Too many patients lack effective treatment options, even when promising science exists,” . Most coverage echoed it.

The interesting line is buried.

The docket asks for three things. Candidates with sufficient evidence already in hand. Candidates with promising preliminary clinical data: case reports, case series, observational studies. And, in bullet three, “candidates with promising preliminary preclinical data, including findings from emerging tools such as artificial intelligence and machine learning.”

Read that bullet twice. Then read it again.

The FDA is asking the public to submit preclinical AI and machine-learning evidence as a basis for new clinical uses of approved drugs. Not validated NAMs. Not a qualified context of use. Not a docket on validation methodology. Just: send us the preclinical signals you think point somewhere. We will sort them.

This is a regulator inviting fit-for-purpose preclinical evidence into a labeling-decision pipeline before the validation discipline around that evidence has been fully built. It is unusual. It is welcome. And it is the application-layer problem the Digital Preclinical Society was founded to address.

Three things to notice.

First, the docket is technology-neutral by design, not by accident. The press release tags AI and machine learning as one example of “preliminary preclinical data,” not as the standard. It refuses to privilege a particular method. It puts the decision first and asks what evidence the field has that could improve it.

This is the right framing. It is also a stress test. Because the docket is open to “preliminary” preclinical evidence, the agency will receive submissions of wildly varying quality. Some will be in silico signals from AI models trained on TG-GATEs or DrugMatrix. Some will be observations from organoid screens, microphysiological systems, or other non-animal modalities. Some will be retrospective analyses of in vivo data. Some will be vendor pitches with a clinical anecdote stapled on. The FDA has not told the public what counts. That is the application gap.

Second, this docket lands in the same week another one closes. The comment period on , the agency’s general considerations for the use of New Approach Methodologies in drug development, closes Monday, May 18. The repurposing docket sits one floor up. It asks for the evidence the validation framework is supposed to govern. The two are not coordinated in the press releases. They should be. If the field treats the repurposing docket as a labeling exercise and the NAMs docket as a methods exercise, we end the year with two stacks of comments and no architecture connecting them.

Third, and this is the part that should keep preclinical leaders up at night, the docket explicitly invites AI-derived signals on equal footing with case reports and observational studies. The  put it plainly: the agency is asking for nominations based on “emerging AI and machine learning data.” That is a category that did not exist as a formal docket-submission class three years ago. It exists now because the field built it and because regulators are willing to evaluate it.

And yet.

We do not yet have a shared evidentiary grammar for what “promising preliminary preclinical data” from an AI model actually means. We do not have agreed test articles. We do not have agreed reference standards. We do not have a context-of-use template. We do not have a structured way for a sponsor or a clinician to say, “here is the signal, here is the model that produced it, here is what the model has and has not been validated against, here is the decision this evidence is being submitted to support.” We have not codified what fit-for-purpose looks like for an AI-derived preclinical signal entering a repurposing docket.

If we do not codify it before June 11, the agency will codify it for us, from the submissions that land. The first hundred docket entries will shape the next decade.

This is what DiPS exists to do. We define the decision. We name the modality. We connect the evidence to the decision context. We work with regulators while they are still drafting, not after the guidance is final.

For preclinical safety and translational science leaders at top-20 pharma, the question is not whether to nominate a repurposing candidate. The question is whether the AI and computational evidence supporting that nomination is structured enough to land in the docket as a credible submission, rather than as a footnote the reviewer skips.

For CROs and CDMOs, the question is whether your client-facing evidence packages include the application-layer logic the repurposing docket implicitly demands: decision context, evidence requirements, context of use, modality-symmetric framing. If not, you will lose the next round of repurposing-adjacent work to the providers who codify first.

For regulator-adjacent scientists at FDA, EMA, OECD, NIH ORIVA, ICCVAM, NICEATM, and EURL ECVAM, the question is whether the repurposing docket and the NAMs validation docket are read together inside your agency. They should be. The architecture only holds if they are.

The success of this transition will be measured not by how many drugs get new indications, but by whether the preclinical evidence that supports those indications survives the clinic. Not by the volume of the docket. By the predictivity of what lands inside it.

That is the standard.

If you run preclinical safety, toxicology, or translational research at a sponsor and you are deciding what to submit by June 11, I want to hear from you. If you are a regulator reading this docket the way I do, let DiPS be in the room. If you build the AI tools generating these signals, codify the context of use before you ship the submission.

The European framing on repurposed evidence: read-across is the lesson Europe has already learned.

Szczepan is right that the FDA’s repurposing docket is a preclinical evidence ask in a clinical labeling envelope. The question to add from the European side is whether the field has already done a version of this exercise, under a different name, and what we learned.

In European chemical safety assessment, the concept is called read-across. Read-across is the use of data from one substance to inform the assessment of another, structurally or mechanistically similar substance, for a specific regulatory decision. It is, in essence, evidence repurposing. The chemicals community has been refining the discipline for more than fifteen years, through ECHA, OECD, EURL ECVAM, and the Integrated Approaches to Testing and Assessment framework.

Three lessons from that experience apply directly to the FDA repurposing docket.

The first is that the evidence question is always nested inside a decision context. Read-across is not “are these substances similar.” It is “are these substances similar enough to support this specific regulatory decision.” The OECD has spent a decade codifying that distinction through . The same logic applies to repurposing. The question is not “does this drug have signal for this new condition.” It is “does this evidence support a regulatory decision to approve this drug for this new use in this specific patient population.” Context of use governs in both directions.

The second is that the data underneath the evidence claim must be interoperable, and interoperability is not the same as format compliance. SEND-compliant nonclinical data from two different facilities is structurally identical and scientifically not yet comparable, because the contextual metadata that determines comparability lives outside SEND. The European chemicals community learned this through the OECD QSAR Toolbox and the EURL ECVAM databases. Without minimal metadata standards at the point of data generation, repurposing evidence assembled across sources looks rigorous and is not. The  at the Pistoia Alliance, which I co-lead, was built to address this gap in the in vivo context. The same gap exists in AI-derived preclinical signals. The model output is only as transferable as the metadata describing what the model was trained on, and the system the model was applied to.

The third lesson is operational. The European agencies do not evaluate read-across submissions in isolation. They evaluate them inside a structured weight-of-evidence framework that integrates animal-based, non-animal, and computational evidence under a single decision logic. The EMA reflection paper on non-human primates, the EMA work plan for the Non-clinical domain 2026 to 2028, and the OECD IATA documents all use the same vocabulary. The FDA’s repurposing docket does not yet name a parallel framework.

This is not a criticism of the FDA. It is an invitation to convergence. The agencies are not in disagreement on the science. They are in different stages of codifying the application layer. The FDA opened a docket on what evidence to submit. The EMA, EURL ECVAM, OECD, and ICCVAM have already published most of the scaffolding for how to assemble it. The two stacks need to be read together.

DiPS exists, in part, to make that reading easier. We do not validate NAMs, AI models, or read-across methods. We connect. We codify the application layer that sits above all of them, where decision, evidence, modality, and context of use meet.

Read the repurposing docket and the  together this week, and the scaffolding of a transatlantic application layer for digital preclinical evidence becomes visible. That is what the next decade of preclinical decision-making looks like, if the field codifies it now.

Qualified: Why Virtual Control Groups Are a Decision Problem, Not a Statistics Problem.

VCG concordance with concurrent controls hits 51.5 percent in legacy rat tox reanalyses, and the gap is judgment, not data. By Szczepan Baran.

The Infrastructure Nobody Wants to Build.

SEND is the language regulators speak; metadata is the language data speaks; without both, VCG databases are not interoperable. By Stefano Gaburro.

AI and Computational Methods

Agentic AI compresses proteomic mechanism-of-action discovery from months to days. [cross-cutting]

A May 7 bioRxiv preprint from reports a closed-loop platform combining semi-automated mass spectrometry sample preparation with an AI-agent data analysis layer. In a screen of 172 compounds in HepG2 cells, the platform produced 1,232 proteomes covering more than 8,700 quantified proteins in approximately three weeks. The agentic AI layer reduced data analysis and interpretation time to under one day and produced structured, testable mechanism hypotheses, two of which the authors validated experimentally. The open application question is whether agentic AI mechanism summaries, generated this fast, are decision-grade for an IND-enabling program, or whether they remain a hypothesis-generation tool with no formal evidence weight in a regulatory submission.

Context of use: Mechanism-of-action characterization in early discovery and lead optimization; pre-IND hypothesis generation.

Source: Jiang Y, Movassaghi CS, et al., bioRxiv, May 2026. doi.org/10.64898/2026.05.04.722718

Federation Watch

AAPS National Biotechnology Conference convenes the NAMs and AI plenary this week in San Diego. [cross-cutting]

The American Association of Pharmaceutical Scientists , with Thomas Hartung delivering the opening plenary on the operational maturity of NAMs, AI, and human-relevant models in drug development. The closing plenary, by Lingle Wang of Schrödinger, frames physics-based computational modeling. The convening is one of the largest US gatherings this year where preclinical, regulatory, and AI methods communities meet under one roof. The open application question is whether the conversation moves from “NAMs are coming” toward agreed application-layer expectations: how the decision is named, what context of use is documented, and what evidence the field expects to see attached to a submission.

Context of use: Pre-competitive convening of preclinical, regulatory, and AI methods communities on application standards.

Source: AAPS NBC 2026 program, BioPharm International preview, May 13.

Performance of preclinical models in predicting drug-induced liver injury: systematic review evidence integration.

Dirven H, Vist GE, Bandhakavi S, et al., Scientific Reports, 2021; methodology echoed in 2026 NAMs validation literature.

A retrospective systematic review across nine large datasets that the FDA NAMs general considerations draft guidance lists in its bibliography as a worked example of evidence integration. The 2021 paper concluded that neither animal-based nor in vitro models in isolation produced a strong hazard signal for two reference DILI compounds, but a structured weight-of-evidence integration across three evidence streams did. Bridge to the Lead: this is what a repurposing-grade preclinical evidence package actually looks like, before the AI layer is even added. Read →

Context of Use, or CoU

A formal regulatory term-of-art that names exactly what decision a piece of evidence is being asked to support. EMA, FDA, ICH, and OECD all use it. The DiPS charter treats it as governing.

A context of use specifies, at minimum, four things: the question the evidence is being submitted to answer, the population to which the answer applies, the conditions under which the evidence was generated, and the consequences of getting the answer wrong. A context of use is not a description of a method. It is a description of a decision. A given method can have many qualified contexts of use, each requiring its own evidence. The same method can be appropriate for one context and inadequate for another.

In drug development, the term first entered FDA practice through the Drug Development Tool qualification program and is the spine of the . The EMA reflection paper on NAMs uses the same governing logic. ICH and OECD documents extend it across modalities and jurisdictions.

The term fails when it is used as a label rather than a discipline. Calling a study “context of use specific” without naming the decision, the population, the conditions, and the consequences is decorative. The FDA repurposing docket opened this week implicitly assumes a context of use for every nomination, and explicitly defines none. That is the gap the field will fill, or the agency will fill for it from whatever the docket receives. The shape of the gap is exactly the shape of the work DiPS was founded to do.

When the FDA’s repurposing docket closes on June 11, what one preclinical AI or machine-learning signal would you trust enough to nominate publicly as a basis for a new clinical use? And what evidence would you attach to it to make the nomination land? Join the conversation in the comments.

Was this forwarded to you? Subscribe to The Digital Record for free, every Saturday. Know someone in preclinical who should read this? Forward this issue. Ready to do more than read? This week’s on-ramp is Nominate a Founding Board Member. DiPS is recruiting a balanced founding board across animal-based, non-animal, and in silico expertise, and we want your shortlist.