The Model that Taught Itself Antibodies

In this brief episode segment, a speaker describes the announcement of “the new ESM fold”, framing it as an open system intended to support scientific discovery in protein biology. The speaker claims the system has “folded over 1.1 billion proteins”, predicting their structures and identifying features that connect them.

They characterise the model as a general-purpose system for understanding proteins — effectively a “world model” for protein biology — which they say can be used to search the space of possible proteins and design new ones. The speaker argues that protein design can arise as an “emergent property” of a model trained to broadly understand proteins, rather than being built for a single purpose.

As an example, the speaker says the model was not designed specifically for antibodies or for binding a particular target, but can nonetheless be used to generate antibodies computationally. They contrast this with conventional antibody discovery workflows, which they describe as requiring intensive laboratory screening of hundreds of thousands to millions of candidates via high-throughput experiments. The speaker suggests the approach could shift some of that work towards compute by “spin[ning] up an instance” to generate candidates.

No underlying data, benchmarking, or external validation is provided in the excerpt.

• A speaker says a new version of ESM Fold has been announced.
• They describe it as an open system for scientific discovery in protein biology.
• They claim the system folded over 1.1 billion proteins and predicted their structures.
• They argue the model is a general protein ‘world model’ usable for protein design.
• They say antibody generation can emerge without training a model specifically for antibodies.
• They contrast this with large-scale wet-lab antibody screening workflows and suggest compute could replace some of that work.