Most AI biology announcements follow a familiar shape. A model gets released, a benchmark gets beaten, a press release calls it a leap forward, and nothing verifiable happens in a laboratory for another two years.
Biohub’s release on 27 May 2026 broke that pattern in one specific way, and it is the only part of this story that genuinely matters.
They designed proteins on a computer, took them into a lab, and a large fraction of them worked.
That sentence deserves scrutiny rather than applause, because “worked” is doing heavy lifting and the numbers behind it are more interesting than the headline. So let us go through what was actually released, what the results show, and where the term “world model” is earning its keep versus where it is borrowed marketing.
What Biohub released
Biohub is a nonprofit research institute co-founded by Priscilla Chan and Mark Zuckerberg. Its stated goal is to cure or prevent all disease, which is the kind of mission statement that invites eye-rolling until you look at what the organisation is actually shipping.
The release is not one model. It is three tools that work together, all freely available to researchers through the Biohub Platform.
ESMC is the foundation. It is a language model trained on roughly 2.8 billion protein sequences drawn from across the breadth of life, including organisms adapted to extreme environments, plus more than 20,000 types of protein found in the human body. It provides the base layer for modelling protein sequence, structure and function.
ESMFold2 is the prediction and design engine. It takes the evolutionary information encoded in ESMC and translates it into atomic-resolution structures and interactions.
ESM Atlas is the map. It makes ESMC’s representations navigable across 6.8 billion protein sequences and 1.1 billion predicted structures, which Biohub describes as the largest application of AI to protein biology to date.
Alex Rives, Biohub’s head of science and formerly chief scientist at EvolutionaryScale, presented the work at the Cold Spring Harbor Laboratory’s “AI in Biology” symposium.
The results that actually count
Here is where this release separates itself from the general noise in AI biology.
Biohub used ESMFold2 to design protein binders against five disease targets. Two were receptor tyrosine kinases implicated in tumour growth, EGFR and PDGFRβ. Two were immune checkpoints that cancer cells exploit to evade immune surveillance, PD-L1 and CTLA-4. The fifth, CD45, regulates immune cell signalling.
These are not toy targets. They sit at the centre of real oncology and immunology programmes.
The designs were then validated in the laboratory. Compact mini-binders achieved hit rates between 36 and 88 percent. Antibody-derived formats came in lower, between 15 and 29 percent. The successful designs showed nanomolar binding affinity, high specificity and stability profiles consistent with potential clinical utility.
To understand why those numbers are notable, you need to know what the baseline looks like. Computational protein design has historically been a discipline where you generate a large number of candidates, test them, and expect most to fail. Hit rates in the low single digits were long considered respectable.
A 36 to 88 percent hit rate for mini-binders is a different regime entirely.
The honest caveat is that the range is wide, and the low end and high end almost certainly correspond to targets of very different difficulty. Biohub has not, in the material released so far, broken down which target produced which rate. That matters, because a design method that hits 88 percent on an easy target and 36 percent on a hard one tells a more nuanced story than a headline average would.
The antibody-derived numbers are also worth sitting with. At 15 to 29 percent, they are meaningfully lower than the mini-binder results. Antibodies are the format most of the pharmaceutical industry actually builds drugs around, so the gap between those two figures is arguably the most commercially relevant number in the entire release.
Why they are calling it a world model
The phrase “world model” comes from AI research, not biology, and its use here is deliberate.
In AI, a world model is a system that has internalised the underlying rules governing an environment rather than memorising patterns of behaviour within it. The distinction is between a system that has seen many examples and a system that has learned why those examples look the way they do.
Biohub’s argument is that ESMC has learned the biological rules governing how proteins fold, function and interact, rather than simply cataloguing known structures. Rives framed it directly, saying the models have learned such a high-fidelity world model of biology that protein interfaces can be designed computationally, taken into the laboratory, and function as predicted.
Shana Kelley, the Northwestern University scientist who serves as Biohub’s president of bioengineering, put it in broader terms, describing a move from an era of reading the book of life to an era of writing it.
Is the terminology justified?
Partially. The strongest evidence for it is not any benchmark score but the lab validation itself. A system that had merely memorised known protein structures would struggle to design novel interfaces that behave as predicted against specific targets. Something closer to rule-learning is a reasonable inference from designs that function on first attempt at those rates.
The weaker part of the claim is scope. Proteins are one layer of biology. A world model of protein biology is not a world model of biology, and the distance between those two things is enormous. Cellular context, metabolic state, tissue environment and the immune system all shape whether a protein that binds beautifully in a dish does anything useful in a body.
Biohub has been reasonably careful about this distinction. Much of the coverage has not.
The architecture choice worth noting
One technical detail deserves attention because it addresses a genuine constraint in the field.
ESMFold2 uses a looped transformer architecture, which allows compute to scale at inference time. The practical benefit is that it avoids the overfitting that typically arises when training is limited by the relatively small number of experimentally determined protein structures available.
This is a real bottleneck. Experimentally solved protein structures are expensive and slow to produce, and the total set of them is tiny compared with the number of known protein sequences. Models trained directly on that structural data run into a ceiling.
Being able to push more compute at inference rather than depending on more training structures is a sensible way around a problem that is not going to resolve itself quickly.
What the ESM Atlas actually enables
The Atlas gets the least coverage of the three tools and is arguably the most useful for working scientists.
It organises proteins by relationships the model has learned rather than by the categories existing databases already encode. That surfaces connections nobody catalogued, including evolutionary links between gene-editing enzymes distributed across distant branches of the tree of life.
The significance is that a large amount of biology has never been annotated. Researchers working on diseases where the underlying biology is poorly characterised have historically been searching a map with most of the territory blank.
The Atlas makes uncharacterised biology searchable. For a researcher studying a rare disease with no established protein targets, that is a more immediately practical contribution than a state-of-the-art structure prediction score.
The open release is the strategic decision
All three tools are freely available to the global scientific community.
This is not incidental. It is the deliberate consequence of Biohub being a nonprofit rather than a company, and it stands in contrast to how comparable capability has been commercialised elsewhere in the field.
The ESM lineage itself illustrates the point. The work traces back through EvolutionaryScale, a company, before landing at a nonprofit institute that is giving it away. That institutional journey is unusual and worth watching, because open release of frontier biological design tools raises questions the field has not fully worked through.
Protein design capability is dual-use technology. Tools that make it easier to design a therapeutic binder also make aspects of biological design more accessible generally. The scientific case for open release is strong and well established, and it is also true that the governance conversation around it is less mature than the technology.
Biohub has not addressed this publicly in the release material.
What this does not mean
Some calibration, because the framing in circulation has drifted.
This is not a drug. Designing a binder with good affinity and specificity is an early step in a process that runs through pharmacokinetics, toxicity, manufacturing, and years of clinical trials, where the overwhelming majority of candidates fail for reasons that have nothing to do with binding.
This does not compress drug development from years to weeks, despite that claim appearing in coverage. It compresses one early stage of discovery. That stage is a genuine bottleneck and improving it has real value, but the timeline of a drug programme is dominated by the clinical phases, which this does not touch.
This is also not a solved problem. The antibody-format hit rates are substantially lower than the mini-binder rates, and antibodies are where most of the industry’s therapeutic value sits.
Why it still matters
Set the overclaiming aside and the underlying result holds up well.
A computational method that designs functional protein binders against clinically relevant targets, at hit rates high enough to change how early discovery is resourced, released openly to anyone who wants to use it, is a meaningful contribution.
The reason to take it seriously is not the “world model” framing or the mission statement about curing all disease. It is the narrow, checkable fact that designs made in silico behaved as predicted when someone took them to a bench.
That is the standard the rest of the field should be held to.
Next steps worth taking
If you work in computational biology: The tools are live on the Biohub Platform. The mini-binder versus antibody hit-rate gap is the most informative thing to probe if you have targets of your own.
If you follow AI research: The looped transformer approach to scaling inference compute in a data-constrained domain is transferable thinking, and worth reading beyond the biology context.
If you are tracking the field commercially: Watch whether any of the five targets progress toward preclinical development. That is the signal that separates a strong methods paper from a shift in how drugs get discovered.
If you write about this: Check whether coverage distinguishes protein biology from biology. Most of it currently does not, and the gap between those claims is where the credibility of the whole field gets spent.
Leave a Reply