When AlphaFold3 launched, it looked like specialized architecture had won the protein prediction race. DeepMind's system used domain-specific tricks - geometric constraints, evolutionary couplings, structural priors - to predict how proteins fold and bind.
BioHub's Alex Rives just released ESMFold2, and it beats AlphaFold3 on antibody design and binding interactions. Not through better tricks. Through scale.
ESMFold2 is a BERT-like transformer trained on 2.8 billion protein sequences. No geometric constraints. No evolutionary templates. Just a language model that learned patterns from raw data, then got scaled up until it started making predictions the specialized systems couldn't match.
The Bitter Lesson, Applied to Biology
Rich Sutton's "bitter lesson" from AI research: methods that scale with compute beat methods that encode human knowledge. It's been true in chess, in Go, in language, in vision. Researchers keep building systems with clever domain-specific features, and then someone scales up a general method and wins anyway.
Protein folding looked like the exception. AlphaFold2 worked because it combined neural networks with decades of biological knowledge - multiple sequence alignments, structure databases, physics-informed constraints. When it worked, it felt like validation that domain expertise still mattered.
ESMFold2 suggests otherwise. Train a transformer on enough sequences, and it learns the physics. The model sees how amino acid patterns co-occur across billions of examples, and from those patterns it infers the rules of protein structure. No need to hand-code the rules if the data is large enough to surface them.
What ESMFold2 Actually Does
The model takes a protein sequence as input and predicts its 3D structure. It also predicts binding interactions - how one protein will interact with another. On antibody design specifically, it outperforms AlphaFold3's metrics for accuracy and confidence.
Antibodies are hard because they're dynamic. They don't just fold into one shape - they flex and adapt when binding to targets. Predicting that requires understanding not just static structure, but how structure changes under interaction. ESMFold2 seems to have learned that from the data.
The system is open-source, which means every lab working on antibody design, enzyme engineering, or drug discovery can now use it. That's not just an academic detail - AlphaFold3 is closed. DeepMind licenses it through Isomorphic Labs. ESMFold2 is a public good.
Why This Changes the Builder Landscape
Protein design was becoming a walled garden. AlphaFold3's capabilities were impressive, but access meant negotiating with a commercial entity. Academic labs and small biotech companies were priced out or stuck waiting for partnerships.
ESMFold2 breaks that open. Any developer with a GPU cluster can run it. Any researcher can fork it, modify it, and publish improvements. The barrier to entry for computational protein work just dropped dramatically.
For builders, this means you can prototype antibody designs, test enzyme variants, or explore binding interactions without needing a DeepMind partnership. The models are there. The code is public. The only constraint is compute - and that's a solvable problem.
What It Means for the Bitter Lesson
Every few years, a domain seems immune to the bitter lesson. Protein folding was supposed to be one of them. Biology is complex, the reasoning went. You need domain knowledge to navigate it. Pure scale won't work here.
ESMFold2 is evidence that biology isn't special. Train a big enough model on enough data, and it learns the domain. Not through hard-coded rules, but through patterns in the data itself. The knowledge is there - you just need enough compute to extract it.
That doesn't mean domain expertise is worthless. The people who know biology are the ones who will ask the right questions, design the right experiments, and turn model predictions into real-world applications. But the predictive power? That's coming from scale now, not from encoding what we think we know.
The Open Question
AlphaFold3 is still better at some tasks. Full-complex prediction, large multi-protein assemblies, interactions involving small molecules - DeepMind's system still leads there. ESMFold2's advantage is narrow: antibodies and binding interactions.
But the trajectory is clear. Scale is closing the gap. And unlike AlphaFold3, ESMFold2 will keep improving as researchers fork it, train bigger versions, and publish their work openly. Open development compounds faster than closed development, especially in research tools.
The bitter lesson isn't that domain knowledge doesn't matter. It's that when you're choosing between encoding knowledge and learning it from data, bet on data. Biology just learned that the hard way.