In Vivo

Observations on AI in healthcare

AI for Biopharma, Part II

Bottlenecks

Part I argued that AI has become standard tooling in drug discovery, has produced a wave of candidate molecules with more likely on the way, and will probably deliver.

This next post is about what stands in the way: the bottlenecks between a model and a patient, and how the regulatory landscape might change.

What are the bottlenecks?

If discovery is where AI shines, why hasn’t it produced more? Some hypotheses, roughly in order of how binding they are.

The road beyond discovery

Discovery is a minority of drug-development cost (roughly 30% of pre-Phase-2 spending), while clinical trials account for about 69% of total R&D.

A recent law-review analysis found roughly 76% of AI use cases sit in early discovery. Part I also summarizes this - have a look.

This is the shape of Eroom’s Law (the decidedly less cheery counterpart to Moore’s Law). Since 1950, the number of new drugs approved per billion dollars of R&D has halved roughly every nine years, even as nearly every tool got cheaper (Scannell et al., 2012).

Cheaper and faster helps. But it’s not where cost or risk concentrate.

And depending on our choice of metrics, there is a scenario where AI makes pharma companies look “worse”: more cheap candidates entering trials, only to fail in Phase II.

Augmentation, or abdication?

AI tools can hand a researcher a candidate molecule without an explanation for why it should work. This is a practical problem: a candidate you cannot reason about is harder to optimize, harder to learn from when it fails, and harder to defend to a regulator.

The optimists frame this as augmentation. EDA software, for example, let a generation of engineers design far more complex chips than they could by hand, without tracking every transistor - the tool raised the level of abstraction.

One might call it, as one blogger does, an “abdication of reasoning”: if a model designs candidates that are poorly understood, the scientific feedback loop weakens (Pharma Fox).

The open question is which is happening with AI drug-design tools like Chai and Isomorphic - raising the level of abstraction, or removing the understanding.

Data

The more practical problem is not what models can do in expectation, but what data they have to work with. Even a model that made biology tractable would still need data to make it useful - and biology is both fiendishly complex and data-poor (at least for now).

Start with complexity. The core problem in drug development is identifying and exploiting biological cause. The genome is not a linear script: only ~2% of it codes for proteins (the “exome”). The rest is part of a vast system that regulates the activity of those protein-coding parts. Gene-to-protein expression is conditioned by many factors beyond my comprehension - chromatin folding, cell type, higher levels of the so-called multi-omics cascade - that models never see (Ball, Quanta, 2026). The magnitude of impact from those intervening layers, between genome and phenome, is massive.

Now scarcity. Biology does not have a massive corpus of freely generated data at its disposal. Large language models worked because the internet already existed; biology has no equivalent. AlphaFold is the exception that proves the rule - it was possible only because the Protein Data Bank had spent fifty years accumulating experimentally-solved structures. Most of biology has no Protein Data Bank.

So data has to be physically generated on an experiment-by-experiment basis. The rate-limiter is atoms, not GPUs - and atoms do not follow Moore’s Law.

There are actors trying to solve this: the “self-driving labs.” Insitro, for example, generates its own high-throughput functional-genomic data to train its models. Lila Sciences and others are building autonomous labs that run experiments end to end. Still others, like Emerald Cloud Lab and Benchling, are digitizing lab operations and the data they produce.

Note where they sit, though: upstream, in discovery and preclinical data generation. Industrializing data does not escape the pattern; it deepens it.

Even the labels are not infallible! On RECIST, the standard tumor-response yardstick, expert readers disagreed on whether a patient had progressed in about 30% of more than 13,000 paired assessments (cited in Khozin, 2026).

This is also why benchmarks do not tell the full story - a problem familiar to anyone following AI in other domains, where leaderboards are routinely saturated, contaminated, or gamed. Strong performance on CASP - the biennial Critical Assessment of Structure Prediction, where AlphaFold made its name - is not the same as making a drug.

These hypotheses share a shape: the binding constraints are downstream, biological, or infrastructural.

“AI will touch every drug” and “AI will transform human health” are different claims.

How might the regulatory landscape evolve?

Two forces will shape this, pulling in opposite directions.

Fragmentation

In September 2024 the FDA and EMA jointly set out ten common principles for AI across the medicines lifecycle.

In practice, approaches are diverging. A Journal of Law and the Biosciences analysis describes the FDA taking a flexible, case-by-case, risk-based line while the EMA builds structured, documented requirements by development phase (Lenarczyk, Minssen, Price & Rai, 2025). The EMA, for instance, leans on the EU AI Act’s high-risk classifications and expects model documentation up front, where the FDA judges each model by its “context of use” and the risk it poses to a specific decision.

Three potential futures follow:

  • Firms default to the stricter EMA standard, making it the de facto global rule.
  • They run parallel tracks per jurisdiction.
  • Friction rises, and falls hardest on small innovators who cannot afford two regulatory strategies.

The above factors may very well chill innovation to some degree. No one is quite sure how a regulator will treat a model inside a pivotal trial (frozen for the study, or allowed to keep learning?), for example, or whether “digital twin” control arms are admissible.

Regulators adopting AI themselves

In January 2025 the FDA issued its first guidance on AI in regulatory submissions. By mid-2025 it had deployed an internal generative-AI assistant, “Elsa,” to help its reviewers.

The symmetry is worth noting: AI is arriving on both sides of the desk - sponsors using it to generate applications and draft labeling, the agency using it to review them.

The wildcard: cheaper trials, and who owns the drug

Elliot Hershberg makes the argument that regulatory innovation could matter more than any innovation in model architecture:

If trials became dramatically cheaper and faster, biotechs could finance their own late-stage development instead of selling to pharma.

(It helps to be precise about who does what today. Large pharma are, in effect, program managers of extraordinary complexity. CROs run the trials. CDMOs manufacture. Small biotechs are where much of the methodological experimentation happens.)

Cheaper trials could rewire the division of labor in the biopharmaceutical value chain.

The EDA comparison cuts a second way, too: those firms stayed tools and never became chip companies.

AI-bio could very well reshape the value chain - sell the software (Schrödinger, Chai, NVIDIA), or integrate forward and run your own drug programs (Insilico, Recursion, Isomorphic, Xaira).

Reading list

A lot of excellent work has been written on this topic. Here are the five pieces I found most impactful.

← All writing