How Does Biohub's Protein Model Design Antibodies?
Alex Rives says Biohub's new ESM fold model designs single-chain antibodies on a computer, and a lab test of 96 proteins found nanomolar binders, the level needed for therapeutic activity, even though the team did not build the model for antibodies.
About a week before the conversation, the team announced ESM fold as an open system for scientific discovery in protein biology. It is a world model with a language-model base, trained on billions of protein sequences, and it learns emergent representations of protein biology. It predicts atomic-resolution protein structure fast enough to illustrate a Pareto optimal frontier of speed and accuracy. The team folded over 1.1 billion proteins, predicted their structures, and identified features connecting them through mechanistic interpretability. ESM fold hits state of the art across structure prediction benchmarks, especially on protein-protein interactions and protein-antibody interactions, which matter for therapeutic design.
Searching the world model runs hundreds of thousands of trajectories digitally. The team then synthesizes 96 proteins on a 96-well plate and tests them in a short lab cycle. The tests found nanomolar binders. A general-purpose model of protein biology can do that design work. The same system maps vast stretches of the protein universe and moves from digital design to binders with therapeutic-level activity in one small experimental cycle.
From the transcript
Maybe I could say something about our most recent launch, because I think it also kind of illustrates some of this. Yeah, so, you know, because I guess it was just a week ago, about now, so we announced the new ESM fold. And so this is basically...
an open system for scientific discovery in protein biology. It's a world model of protein biology that's been trained. It's a language model base. So it's been trained on billions of protein sequences, kind of learns these emergent representations of protein biology. And then we can use it to make predictions of atomic resolution protein structure.
And we can use it to... And it's really fast. So it's blazing fast. So it's kind of...
you know, illustrating this Pareto optimal frontier of kind of speed and accuracy in structure prediction. And so this allows us to kind of characterize, you know, really vast kind of stretches of the protein universe. So we folded over 1.1 billion proteins and predicted their structures and and identified kind of features connecting all of them through mechanistic interpretability.