AI for Biopharma, Part I

For some time, “healthcare” has been the reflexive answer of AI optimists pressed on the promise of AI. This series tests that instinct, from the perspective of a fellow optimist.

Today’s post is about AI in drug development: what is happening, what has been produced, and what it might mean. A subsequent piece takes up why progress might look slower than expected, and what changes are (or should be) on the horizon.

What is happening in AI for biopharma?

In 2026, AI is standard tooling in drug R&D.

One sign of how routine this has become: AlphaFold’s open structure database has been used by more than three million researchers across 190 countries - a tool that barely existed five years ago.

Drugmakers spend roughly $300 billion a year on R&D, and every one of the world’s top-20 pharma companies now uses AI somewhere in the pipeline (Zhang et al., Nature Medicine, 2025).

The regulators see it, too. FDA drug and biologic submissions with an AI or machine-learning component jumped from a handful in the late 2010s to 132 in 2021 - roughly a tenfold leap in a single year - and passed 500 cumulatively by 2025.

Capital has followed. Private investment in AI-focused drug-discovery firms has run into the tens of billions over the past decade, and the number of AI-originated molecules in trials has gone from near-zero in 2020 to well over 150.

The activity is not spread evenly: it clusters upstream, in discovery, and the successes there have been notable.

AlphaFold predicts a protein’s 3D structure from its amino-acid sequence at near-experimental accuracy. AlphaGenome (2025) predicts how single-letter DNA variants change gene regulation across the non-coding genome. RFdiffusion generates novel protein backbones from scratch with a diffusion model, while Chai and AlphaFold3 predict how proteins bind small molecules, DNA, and one another. Generative-chemistry models propose new drug-like scaffolds atom by atom.

What lies downstream - trials, approval, surveillance - is far quieter, though that is starting to change.

This pattern is not accidental: discovery, relative to wet-lab experimentation, is a fairly tractable problem for machine learning. Structure prediction has a clean, gradable objective (molecular accuracy against known structures) as well as a large labeled dataset in the Protein Data Bank. Molecular design offers fast in-silico feedback. As Demis Hassabis has put it, he has “always thought of biology as an information processing system at a fundamental level.”

It also fits how the industry innovates: much early experimentation happens in small, nimble biotechs that adopt new methods fast, then license or sell to the majors. What follows is a brief survey of how these capabilities have been built and deployed - not exhaustive, but enough to make the point.

Figure 1 — AI capabilities across development stages (2018-2026). Interactive; hover any dot, or open the full chart ↗.

What has it yielded to date?

Plenty of activity in the pipeline, and very little on pharmacy shelves - largely what you would expect at this point.

Clinical trials consume the majority of drug-development capital and time: about 69% of R&D cost, and on average close to eight years of testing versus under three for preclinical work.

“AI drugs” have not reached the market - nor, given those timelines, could they reasonably be expected to have yet.

The table below lists the drugs discovered or designed with AI that have reached human testing at all.

Figure 2 — AI-discovered and AI-designed drugs in the clinic. Open the full table ↗.

A few observations worth stating plainly.

• “Zero approvals” is not a verdict. The first AI-designed molecules entered trials around 2020, and first-in-human to market runs about eight years on average. The leaders - Insilico’s rentosertib (Phase III for pulmonary fibrosis) and Relay’s RLY-4008 (an FDA decision expected around September 2026) - are close, not done.
• The misses are informative, and therapeutically diverse. Exscientia’s DSP-1181 (OCD), BenevolentAI’s BEN-2293 (eczema), and the ulotaront schizophrenia program all washed out - different diseases, one lesson: AI-found molecules fail where the biology does not cooperate. There is no identifiable class of methodological failure or obvious blind spot.
• It is not even clear what an “AI drug” is. Definitions span “AI helped screen,” “AI designed the molecule against a known target,” and “AI found the target and designed the molecule” - rentosertib’s claim.

So: real molecules, real momentum, no approvals yet.

What does this all mean for public health?

A harder question - and I’ll afford myself some authorial license to oversimplify. Let’s say for a moment that the social value biopharma delivers is a function of three things:

• Diversity & efficacy - distinct diseases for which an effective therapy exists.
• Access - affected populations’ ability to get and afford the treatment.
• Clinical deployment - how effectively the treatments we have are actually used.

AI’s clearest effect is on the first: more targets, more hits, more leads, faster.

But that is also the slowest lever to reach a patient. Clinical development and review still run close to a decade between a designed molecule and a market authorization.

Access is, for now, largely untouched. Affordability turns on patent timelines and manufacturing cost.

Regulatory change and lab automation could move both - a faster FDA, cheaper production - though the net direction is not obvious.

Clinical deployment may be the most underrated of the three.

We have more than 20,000 approved prescription drug products in the US today. The vast majority are prescribed using crude, population-average guidelines. The largest near-term gains in public health may come not from a new molecule, but instead from using existing ones better - the multi-omics / precision-medicine ambition of matching drug to patient based on a rich view of their personal biology. A topic for a later post.

So, will AI open the floodgates?

Yes, eventually. Leading labs are quite explicit about the ambition: Isomorphic Labs was founded to “solve disease” with AI, and Hassabis has predicted the technology could compress drug discovery from years to months. Initial results have been impressive, as attested above. The pipeline has gone from near-zero to 150-plus programs in five years, and every major biopharma is now a buyer.

But “eventually” is doing some work in the above. Part II is about the path to drug abundance: stay tuned.