FDA’s rare disease pathway: where AI fits next

A one-patient genetic therapy used to mean years of paperwork, bespoke manufacturing, and an approval process that simply wasn’t built for “n=1.” In December 2025, senior FDA officials laid out a new pathway for individualized genetic technologies—an explicit admission that the standard model doesn’t move fast enough for rare genetic diseases.

That matters for anyone building in AI in pharmaceuticals and drug discovery. Not because regulation is “finally catching up,” but because the FDA is describing what evidence counts when you can’t run a traditional trial. And that evidence map lines up unusually well with what AI does best: integrating genomic data, learning from sparse clinical signals, and tightening the loop between hypothesis, design, and patient impact.

This post breaks down what the proposed pathway is really asking for, why it’s different, and where AI teams in pharma and biotech can create real advantage—without hand-waving.

What the FDA is proposing (and why it’s a big deal)

The key point: FDA leadership is signaling a practical route to clinic for individualized genetic medicines when traditional development timelines don’t match disease urgency.

The proposal was inspired by a highly publicized success: baby KJ, diagnosed with a rare disorder caused by a CPS1 gene mutation, treated within months using a base editor designed to correct the mutation. The FDA’s message isn’t “skip evidence.” It’s “bring the right evidence for the right context.”

Here’s the evidence pattern they highlight, which effectively becomes a checklist for future submissions:

1. Identify a specific molecular/cellular abnormality (e.g., a known pathogenic variant)
2. Target the biological alteration (e.g., a base editor addressing that variant)
3. Know the natural history (what happens without treatment, and how fast)
4. Prove the intended target is edited (target engagement / on-target activity)
5. Show evidence of clinical improvement (even if endpoints must be pragmatic)

In baby KJ’s case, target editing was demonstrated in mouse models because a liver biopsy was considered too risky. The child was reported healthy seven months later.

The stance I like here: the FDA is describing a pathway that treats mechanism and target engagement as first-class citizens when large randomized datasets don’t exist. That’s a meaningful shift for rare disease drug development.

The core problem: traditional trials don’t scale to rare genetic diseases

The direct answer: Rare genetic diseases break the assumptions behind conventional clinical development.

Most trial designs assume you can:

• recruit enough patients to power statistical endpoints,
• randomize ethically,
• wait long enough to observe outcomes,
• standardize a single investigational product.

For ultra-rare conditions, you often can’t do any of that. The issues show up operationally (patients dispersed globally), ethically (rapidly progressive disease), and scientifically (heterogeneous variants, heterogeneous baseline severity).

Why “n=1” isn’t a loophole—it's a different evidence problem

Individualized genetic medicines aren’t inherently less scientific. They’re less compatible with population-level evidence frameworks. So the FDA’s new pathway is effectively saying: if you can’t prove benefit through scale, you must prove benefit through causal coherence:

• the variant is real and pathogenic,
• the intervention hits the correct biological target,
• the disease trajectory is well-characterized,
• observed clinical changes line up with the mechanism.

That’s exactly where modern AI methods can help—especially when the available data is messy and small.

Where AI fits: turning FDA’s checklist into an execution plan

The direct answer: AI can compress the time between diagnosis, design, evidence generation, and regulatory-ready documentation—particularly when patient counts are tiny.

But only if teams focus on the FDA’s actual asks: abnormality → target → natural history → target engagement → clinical improvement.

1) Variant interpretation and “abnormality confirmation” at clinical speed

For individualized therapies, the clock starts at diagnosis. AI-assisted variant interpretation helps teams:

• prioritize candidate causal variants (especially in cases with multiple findings),
• link variants to known functional consequences,
• assess confidence levels when literature is sparse.

What works in practice is a hybrid approach: ML models to rank and summarize, plus domain experts to adjudicate. The deliverable isn’t a fancy score—it’s a regulator-readable rationale for why this molecular abnormality is the right target.

Practical tip: build a structured “variant dossier” template early (gene/variant, predicted impact, known cases, functional data, uncertainty). If you wait until after design begins, you’ll end up backfilling the most important part.

2) AI-guided design for base editors, guides, and payload constraints

Once you know the target, design speed matters. AI can assist with:

• guide design optimization (sequence constraints, predicted efficiency),
• off-target risk triage,
• manufacturability and formulation constraints,
• prioritizing among multiple feasible editing strategies.

For rare disease drug discovery, this is one of the few moments where hours and days matter, not just quarters.

A strong stance: if your AI team isn’t connected to CMC realities (vector capacity, editing window limits, tissue delivery constraints), you’re optimizing the wrong objective. “Most accurate model” is not a regulatory endpoint.

3) Natural history modeling when the dataset is thin

FDA explicitly calls out knowing the natural history. For ultra-rare diseases, that often means:

• few published cases,
• inconsistent phenotype documentation,
• different care standards across sites,
• missing timelines.

AI can help standardize and extract natural history signals from fragmented sources (registries, EHR narratives, published case reports, caregiver logs). The goal isn’t perfection—it’s a defensible baseline trajectory that helps interpret whether observed changes are meaningful.

Concrete ways teams are doing this well:

• NLP pipelines to structure symptom onset and progression from clinical text
• Bayesian models that handle uncertainty explicitly instead of hiding it
• Synthetic control arms built from matched historical patients (with clear caveats)

Natural history is also where you can create compounding advantage: every curated record becomes future evidence for the next patient.

4) Proving “the intended target is successfully edited” without risky biopsies

The FDA’s baby KJ example matters because it acknowledges reality: sometimes you can’t sample the most relevant tissue.

AI becomes useful in two ways:

• Translational bridging: learning relationships between accessible biomarkers (blood, imaging, metabolites) and target tissue editing based on preclinical and limited clinical data.
• Assay design prioritization: selecting the smallest set of assays that best supports a target-engagement claim under safety and feasibility constraints.

This is a place where teams should be careful. AI can strengthen inference, but it can’t replace measurement. Regulators will still ask, “What is measured, how reliable is it, and how does it connect to the mechanism?”

A good pattern is to pre-specify a target engagement evidence ladder:

1. direct evidence (tissue editing) when feasible,
2. validated surrogate evidence when direct isn’t feasible,
3. robust preclinical confirmation + converging clinical biomarkers.

5) Detecting clinical improvement in sparse, noisy real-world endpoints

For individualized therapies, clinical improvement may be captured through:

• metabolic markers,
• hospitalization frequency,
• neurodevelopmental milestones,
• caregiver-reported outcomes,
• wearable data.

AI can help denoise and contextualize these endpoints—especially when the patient is changing rapidly (as infants do) and care pathways shift.

The best applications I’ve seen use AI to:

• model expected developmental trajectories,
• normalize for care changes (new feeding support, new meds),
• identify early warning signals for safety monitoring.

This aligns directly with clinical trial optimization: you’re not optimizing recruitment; you’re optimizing signal detection and decision speed.

What pharma and biotech should do in 2026: a practical playbook

The direct answer: Treat the new pathway as a product development system, not a one-off regulatory event.

If you’re building or partnering in rare genetic disease programs, here are moves that pay off quickly.

Build an “FDA-ready evidence package” pipeline

Don’t wait for a patient. Pre-build templates and tooling for:

• variant dossier generation,
• natural history summary + uncertainty statement,
• target engagement plan (assays + feasibility),
• safety rationale and monitoring plan,
• clinical improvement measurement plan.

The competitive advantage is operational: faster assembly of coherent, auditable evidence.

Treat data standardization as a first-class deliverable

Most rare disease AI efforts fail because the underlying data is inconsistent. Invest early in:

• phenotype normalization (common vocabularies and timelines),
• lab unit harmonization,
• versioned data provenance (what changed, when, and why).

Regulatory conversations go better when your AI outputs are traceable back to source records.

Choose AI use cases that survive regulatory scrutiny

Some AI applications are exciting but fragile in front of a regulator. Prioritize use cases that map cleanly to the pathway:

• “This model helps confirm the causal variant”
• “This model reduces off-target risk and documents why”
• “This model supports natural history estimation with quantified uncertainty”

Avoid “black box efficacy prediction” as your core story. It’s rarely the strongest pillar when n is tiny.

Common questions teams ask about the new pathway

Does this mean the FDA will approve more one-patient therapies?

It points in that direction, but the bigger change is that FDA is describing a repeatable evidentiary logic for individualized genetic technologies. Approval still depends on the totality of evidence, safety, and follow-up.

Will AI replace clinical trials for ultra-rare diseases?

No. AI can reduce uncertainty and improve decision-making, but it doesn’t replace the need for measurement. What it can do is make the best use of limited patient data and help justify endpoints when traditional trials are impossible.

What’s the fastest way to show value with AI here?

Start with natural history + endpoint strategy. If you can’t define what “improvement” looks like and how fast deterioration occurs without treatment, everything downstream becomes speculative.

What this means for AI in drug discovery: regulation is becoming an enabler

The primary keyword here—FDA rare disease pathway—isn’t just a policy headline. It’s a signal that the agency is willing to evaluate individualized genetic medicines using a framework that fits their constraints.

For AI in pharmaceuticals and drug discovery, this is a chance to build systems that make “n=1” development more disciplined: faster variant-to-therapy cycles, better evidence coherence, and more credible measurement when biopsies and big trials aren’t options.

If you’re deciding where to place your next AI bet in 2026, pick the work that turns the FDA’s checklist into a repeatable engine: data standardization, natural history modeling, target engagement inference, and endpoint signal detection. That’s where teams will win.

One forward-looking question to sit with: When individualized genetic medicines become operationally routine, will your organization’s bottleneck be biology—or evidence assembly?