Proteins are the molecular machinery of the human body. Every function, from fighting infection to regulating blood sugar, depends on them. Nearly every drug in existence works by targeting a specific protein, blocking it, activating it, or mimicking it. The problem is that figuring out the shape of a protein and designing something that binds to it reliably has always been brutally slow and expensive work.
Biohub, the nonprofit funded by Mark Zuckerberg and Priscilla Chan, just released a toolkit called ESMFold2 that takes direct aim at that bottleneck. The system was trained on approximately 2.8 billion protein sequences drawn from organisms across all of life, including bacteria from extreme environments and the full range of proteins found in the human body. From that training, it learned the underlying rules of how proteins fold and interact, without being explicitly taught those rules.
The practical output is three tools working together. ESMC is the language model that understands proteins the way a text AI understands sentences. ESMFold2 translates that understanding into precise three-dimensional models of proteins and how they connect with other molecules. ESM Atlas is a map of 6.8 billion protein sequences with over one billion predicted structures, the largest open catalogue of its kind ever built.
In early lab tests, ESMFold2 designed protein binders against five specific targets tied to cancer and immune disease, including PD-L1 and CTLA-4, which are two of the most important targets in modern cancer immunotherapy. Success rates in those tests ran from 36% to 88%, and crucially, the designs were validated in actual laboratory experiments, not just computer simulations. The computational search for those binders was completed in days, a process that traditionally takes months or years.
On accuracy, the claims are significant. ESMFold2 outperforms AlphaFold 3, Google DeepMind's flagship protein model, on key benchmarks including predicting how antibodies bind to their targets. AlphaFold itself was considered a landmark when it arrived and won its creators a Nobel Prize in chemistry in 2024. ESMFold2 entering this space with competitive or superior results, and doing so as a free open tool, is a meaningful shift in who can access this kind of capability.
For anyone in healthcare, pharmaceuticals, insurance, or medical devices, the strategic read is this: the early, expensive phase of drug discovery is getting dramatically cheaper and faster. Traditional drug development costs over $2.6 billion per approved drug and takes an average of 10 to 15 years from start to finish. AI platforms have already shown they can cut preclinical timelines by up to 60% and reduce early-stage costs by 30% to 70%. The first fully AI-designed drug approvals are projected to arrive within the next two years.
The open access angle matters beyond research. When powerful tools like this become free, the organisations that win are not necessarily the ones with the biggest R&D budgets. They are the ones that move fastest once the tools exist. That includes smaller biotech firms, academic hospitals, and research teams in countries that previously could not afford to compete in protein science.
Biohub is also putting $500 million behind a broader Virtual Biology Initiative, which aims to build open datasets and AI simulations of the entire human cell. ESMFold2 is the first major public release from that programme. The competition is real: Isomorphic Labs, Google DeepMind's drug discovery spinout, has raised over $600 million and is running its own platform. Recursion Pharmaceuticals and others are advancing fully AI-designed drugs through clinical trials right now.
The clinical trial hurdle has not gone away. As one industry observer put it, identifying a promising protein binder is one piece of the puzzle, not the whole answer. Safety testing, human trials, and regulatory approval remain as demanding as ever. But the front end of that pipeline, where most time and money historically disappeared, is changing faster than most industries have noticed.