ESMFold2 moves protein folding closer to writing biology

Latent Space has published “ESMFold2: The Bitter Lesson is Coming for Proteins”, a conversation with BioHub’s Alex Rives, and the title lands on the real pressure point. This is not just another installment in the race to draw protein structures faster. The deeper issue is whether biology, like language and vision in modern AI, will lean more heavily on large datasets and models that learn regularities for themselves, instead of systems where humans pre-wire what should matter.

In that sense, ESMFold2 sits at the same kind of transition point large language models occupied before it became obvious that scale was not a detail, but a method. Protein folding has long been a domain where structural biology, evolutionary signals and physical assumptions mattered enormously. Systems such as the AlphaFold Protein Structure Database showed how far that approach can go when models, data and biological knowledge are joined carefully. Rives’ framing points to the next question: what happens when protein models start learning broader biological regularities from the data itself?

That is where the “bitter lesson” enters. The old AI lesson is that methods which exploit computation and learning from data often outlast elegant hand-built shortcuts. In proteins, the stakes are sharper than in text, because an error is not merely a bad chatbot answer. A protein is a physical object, with shape, interactions and consequences inside a cell. That is why the debate over inductive bias matters: less hand design can make models more flexible, but only if datasets, evaluation and biological validation carry enough weight.

The ESM line is already known for treating protein sequences as a kind of evolutionary language, and the open trail of that work is visible through facebookresearch/esm. In this context, ESMFold2 reads less like a successor to a single tool and more like part of a broader ambition: to build models that do not merely predict shape, but capture patterns useful for design, function and biological manipulation. That is why the same conversation naturally reaches for terms such as world models and programmable biology.

A biological “world model” does not mean a magic box that understands life. It means a model that can infer consequences from enough biological data: what changes when a sequence is edited, when a domain is swapped, when the target is not just a stable structure but a function. That is a much stricter task than producing an attractive protein rendering. If such models become reliable enough, they shift the boundary between reading biology and writing biology.

That makes ESMFold2 worth watching beyond the narrow structural-biology circle. The point is not that one model “wins” over another. The point is the direction of research and industry: from isolated predictors toward foundation models for molecular design. Comparison with AlphaFold remains useful, but the new front is different. The question is no longer only how well a model sees a protein. It is how well it can predict what can be done with one.