FDA dropped a draft guidance in January on what AI-discovered drugs actually need to show during IND review. Most of the coverage so far has framed it as a permission slip for the agentic-AI crowd. It isn't.
The draft is a transparency tax. FDA isn't lowering the bar for safety or efficacy. They're adding a new one. If a model picked your lead compound, you now owe FDA an explanation of the model. Not in the academic sense. In the "show me your training data, your validation set, your version control, and your decision logs" sense.
I read the whole thing. Here's what changes for a Series A/B biotech that didn't think this guidance applied to them — and why it probably does.
What FDA actually wrote
Three things in the draft worth your attention.
1. A pilot, not a rule. Ten companies are currently in the streamlined-IND pilot. They get to use computational evidence as part of their case for a Phase 1 start, but only if their AI discovery process meets FDA's transparency criteria. The pilot is closed for now. The criteria are not.
2. Reproducibility, not provenance. FDA isn't asking who built the model. They're asking whether somebody else, given the same inputs, could get the same outputs. That's a different question than what most discovery teams document today.
3. Documentation of the AI step, not the science. Your tox studies still need to follow ICH M3(R2). Your safety pharmacology still needs ICH S7A. What's new is a parallel package documenting how the AI picked the molecule: model architecture, training data lineage, validation cohorts, what was held out, what was prospectively tested.
That last one is the part nobody's ready for.
Why FDA wrote this now
Two numbers explain it.
First: 41% of R&D leaders surveyed in early 2026 said they're planning to hand entire discovery workflows to agentic systems — not just one task, the whole pipeline. Hypothesis to candidate, autonomously. That's a lot of moves FDA doesn't get to see if all they're shown is the final IND package.
Second: Nature Medicine published Phase IIa results for rentosertib (ISM001-055), the first fully AI-discovered drug to read out in a Phase 2 trial. Insilico Medicine. Idiopathic pulmonary fibrosis. The compound worked.
A working AI-discovered drug means more AI-discovered INDs are coming. FDA reviewers can either get visibility into the discovery process while it's still a pilot, or they can get a hundred filings next year that they have no framework for. They picked the first option.
What the pilot actually requires
The draft is light on procedure, heavy on principles. From what's been said publicly about the ten participants, FDA wants:
- Model card describing architecture, training corpus, and known failure modes
- Validation evidence, including prospective validation, not just retrospective
- Held-out test cohorts that didn't influence the model
- Decision logs showing which candidates the model proposed and which were pursued
- Reproducibility package: versioned weights or sufficient detail that a third party could re-derive the prediction
None of that lives in your standard IND. None of it is in your Module 2.4 nonclinical overview or Module 4 nonclinical study reports. Right now it lives in your computational team's Slack and a few Jupyter notebooks that haven't been opened since the lead was nominated.
That's the gap.
Who this affects
Three audiences should be paying attention.
AI-native discovery shops. If your pitch deck says "AI-discovered" in 24-point font, you're going to get asked. The streamlined-IND path is a real option, but only if you can produce the documentation. Start building the model card now, not at the pre-IND meeting.
Traditional biotechs using AI as a tool. You used a generative chemistry platform to narrow a library. Does this guidance apply? Probably yes if the AI made the call on the lead. Probably no if it was used for property prediction and a chemist picked the molecule. The draft doesn't define the line clearly. That's a comment opportunity.
CROs and tooling vendors. Your customers are going to start asking for "FDA-ready" outputs. The first vendor who ships a model card template that FDA actually accepts is going to win a lot of contracts.
What to do this quarter
Three things, in priority order.
-
Inventory your AI decisions. Walk back from your IND-track candidate to discovery. Was a model involved? At which step? Who has the notebooks? If the answer is "I'd have to ask," that's the project.
-
Lock provenance for active programs. Version your training data, validation cohorts, and model weights now, not at submission time. Reconstructing a 2024 model from memory is the kind of work that becomes a six-month delay.
-
File comments on the draft. FDA is taking comments through summer 2026. If the line between "AI as tool" and "AI as decision-maker" matters to your business, this is your one chance to push back before it sets.
The honest read
This guidance doesn't open the floodgates. It puts a meter on them.
FDA's signal is clear. AI-discovered drugs are welcome at IND. The agency isn't going to ask you to re-run discovery in a wet lab. But they will ask you to show your work, and the standard for "showing your work" is going to be higher than what discovery teams have been keeping.
The teams that win the next two years won't have the best models. They'll have the best documentation of their models.
We built RegFo's rules engine on the assumption that regulatory expectations get more specific every year, not less. This draft confirms it. If you want a sanity check on whether your current IND-track package would survive an FDA reviewer asking the new questions, paste it into our protocol checker. The rules engine cites the specific guideline behind every finding, so you can see exactly where the new transparency bar sits relative to where you are.
Source: this piece builds on Axis Intelligence's analysis of the FDA Jan 2026 AI drug development draft guidance, with additional context from the rentosertib Nature Medicine readout and our own work on regulatory documentation gaps.