What Hassabis means by "six AlphaFold-level models"

When Demis Hassabis told Two Minute Papers that Isomorphic Labs is building "six to a dozen AlphaFold-level models" across the drug discovery pipeline, he was making a specific technical claim that is easy to miss in the headline. He was not saying DeepMind is building bigger AlphaFolds. He was saying that drug discovery is not one problem, it is roughly six to eight problems in a trench coat, and Isomorphic is trying to solve each of them at the same level of precision that AlphaFold 2 brought to protein folding.

That distinction is the whole story. I want to walk through why.

What AlphaFold actually did, and why it was one thing. AlphaFold solved a single sub-problem: given the amino acid sequence of a protein (the linear string of building blocks), predict the 3D shape it will fold into. Shape determines function in biology, so this matters enormously. The problem had been open for fifty years. AlphaFold 2 hit roughly 90% of experimental accuracy on the CASP14 benchmark in 2020 — close enough to crystallography that biologists started using it as a substitute for the wet-lab method.¹

But predicting protein shape is one stop on a much longer road. To turn a folded protein into a drug, you need to do something like the following: identify which protein is actually causing the disease (target identification), find or design a small molecule that binds it in the right place (hit discovery and docking), tune that molecule so it binds tightly and selectively (lead optimisation), predict how it moves through the body without poisoning the liver (ADMET — absorption, distribution, metabolism, excretion, toxicity), and figure out how to actually make it at scale (synthesis planning).

Each of these is its own scientific problem with its own data, its own benchmarks, and its own physics. AlphaFold touches one of them.

The platform claim, restated plainly. Hassabis is saying Isomorphic is building a separate foundation model for each of these stages, at AlphaFold-class precision. Not one giant model. A coordinated suite. The press headline writes itself as "DeepMind building six AlphaFolds"; the more accurate version is "DeepMind building six different kinds of AlphaFold, one per stage of the pipeline."

Why this is a much harder claim than it sounds. AlphaFold worked because the Protein Data Bank exists: roughly 200,000 experimentally determined protein structures, curated over fifty years, sitting in the public domain as training data. That dataset is the reason the problem could be solved by a deep learning model at all. No equivalent public corpus exists for ADMET. No equivalent public corpus exists for synthesis planning. The data that does exist is largely locked inside pharmaceutical companies' internal assay databases, generated at enormous expense and treated as core IP.

This is why the Eli Lilly and Novartis partnerships matter more than the dollar figures suggest. The 2024 deals were reported at up to $1.7 billion and $1.2 billion respectively in milestone payments, but the upfront cash was $45 million and $37.5 million — modest by pharma standards.² What those deals almost certainly include, and what Isomorphic actually needs, is access to proprietary experimental data. The moat for the next AlphaFold is not compute. It is labelled biology at scale, and only pharma has it.

Where the metaphor breaks. Calling each model "AlphaFold-level" is a useful shorthand, but it papers over a real technical tension. AlphaFold predicts a static structure — a single answer to a single question. ADMET prediction is fundamentally a prediction about dynamics: how does a molecule behave over time, in a body, across tissues. Synthesis planning is a search problem over discrete chemical reactions. These are not the same shape of problem as folding, and there is no guarantee that the architecture that worked for one will work for the others.

The honest version of the field's current state: physics-based simulation tools, particularly from Schrödinger and D.E. Shaw Research, remain competitive with and in some cases superior to deep learning models on lead optimisation tasks, especially for chemical scaffolds outside the training distribution.³ AlphaFold 3, which extended to small molecules and nucleic acids, has been criticised by structural biologists for accuracy issues on novel ligands relative to established docking tools. "AlphaFold-level" for these other stages is an aspiration, not an achieved benchmark.

The falsification timeline. This is the part I find most useful for cutting through the hype. Hassabis said pre-clinical testing is already underway. Drug development has a well-defined cadence: pre-clinical (animal studies) typically runs two to four years before a company can file an Investigational New Drug application with the FDA. Phase I human trials then take another year or two before any readout. If Isomorphic's platform started generating real candidates in 2023 or 2024, the earliest plausible Phase I readout for an AI-designed-end-to-end drug is somewhere in 2027 or 2028.

That is the public falsification test. Hassabis is making a strong claim about a coordinated platform. If no novel candidate from that platform clears Phase I by 2028, the "AlphaFold-level across the pipeline" framing weakens considerably. If one does, the claim holds up. Either way, we will know.

What to watch. Three concrete things, in roughly increasing order of how much they would change the picture.

First, watch for Isomorphic to publish benchmarks. AlphaFold 2 was credible because it was evaluated on CASP14, a blind community benchmark. If Isomorphic's ADMET or synthesis models are real, they will eventually need to post numbers against accepted benchmarks like Therapeutics Data Commons or USPTO reaction datasets. Silence on benchmarks is informative.

Second, watch the publication patterns. AlphaFold 2 came with an open paper and open weights. AlphaFold 3 came with restrictions on commercial use and limited reproducibility, which the structural biology community noticed. Whether the next models are published, gated, or kept entirely internal will tell you what Isomorphic thinks they are worth.

Third, watch the partner pipelines. If Lilly or Novartis nominates an Isomorphic-designed compound as a development candidate, that is a real signal — pharma companies do not put their names on AI-designed molecules lightly. The 2027–2028 window is when this becomes observable.

The interesting thing about Hassabis's framing is that he gave a number. "Half a dozen to a dozen" is specific enough to be checked. That is a different mode from "we're at the foothills of the singularity," which he said in a separate interview the same week.⁴ One claim can be falsified by a clinical readout. The other cannot be falsified by anything. Track the first.

Footnotes and links

Further reading

• Demis Hassabis interview, Two Minute Papers, 25 May 2026: https://www.youtube.com/watch?v=huAwz_BR8WM
• AlphaFold 3 paper, Nature, May 2024: https://www.nature.com/articles/s41586-024-07487-w
• Therapeutics Data Commons (open benchmarks for drug discovery ML): https://tdcommons.ai

¶