← Back to feed

Biohub released ESMFold2 on May 27, 2026, as part of a three-artifact open-science stack it describes as a protein world model. The release includes ESMC, a protein language model; ESMFold2, a structure prediction and design engine; and ESM Atlas, a database of 1.1 billion predicted protein structures drawn from 6.8 billion sequences. All three are available under MIT license.

The central performance claim: ESMFold2 surpasses AlphaFold and other models on DockQ pass-rate for protein-protein and antibody-antigen complexes on FoldBench — and it does this in single-sequence mode, without multiple sequence alignments. That distinction matters for antibodies, which mutate rapidly. AlphaFold’s performance degrades on antibodies precisely because it depends on evolutionary conservation signals that antibody diversity disrupts. ESMFold2 does not have that dependency.

What Was Built and Validated

The core scientific claim behind ESMFold2 is that a vanilla transformer trained on enough diversity outperforms architecturally specialised models. ESMC was trained on approximately 2.8 billion protein sequences drawn from across the tree of life, without the hand-engineered features that specialised structural models rely on. The hypothesis, described by Biohub lead researcher Alex Rives as applying “the bitter lesson to proteins,” is that scale and diversity of training data carries more signal than domain-specific inductive biases.

ESMFold2 consumes ESMC’s sequence representations and converts them into atomic-resolution 3D structures of biomolecular complexes — including proteins together with DNA, RNA, small molecules, and modified residues. Its architecture uses a looped transformer that reuses the same blocks repeatedly, so compute can scale at inference time rather than being fixed by network depth. Biohub reports that inference time scaling works in practice: spending more compute at inference yields measurably better predictions, a property tested across five targets.

Those five targets are in cancer and immunology. Biohub ran inversion of ESMFold2 to generate novel protein binder designs, then tested them in wet-lab experiments. The designed binders worked. Biohub describes the search-to-validation time as days rather than the months or years typical of conventional protein design campaigns.

ESM Atlas

ESM Atlas gives researchers access to 6.8 billion protein sequences and 1.1 billion predicted structures, covering the protein universe across species and functional families. The scale is intended to make protein space searchable computationally in the way genome databases made DNA searchable — researchers rank hypotheses before committing resources to synthesis and testing.

Institutional Context

Biohub was formed by combining CZI Science, CZ Biohub, and the acquired EvolutionaryScale team. EvolutionaryScale was the company Alex Rives and colleagues founded to continue the ESM line of research after Meta AI, where the original ESMFold was developed. Chan Zuckerberg Biohub is the non-profit research organisation funded by Mark Zuckerberg and Priscilla Chan.

The $500M Virtual Biology Initiative announced by Biohub in April 2026 targets the data generation problem: building the open, multi-modal single-cell measurements needed to train predictive cellular models. ESMFold2 addresses a narrower and more tractable problem: protein structure and binder design. Both programmes share the premise that the limiting constraint in AI-driven biology is the quality and scale of training data, not model architecture.

What This Does Not Change

ESMFold2’s benchmark lead on antibodies is specific to protein-protein interaction prediction, not general protein folding. For many structural biology use cases, AlphaFold and its derivatives remain well-supported community infrastructure with broader ecosystem tooling. Biohub’s claim is that ESMFold2 occupies a distinct and clinically important niche: antibody therapeutics, where AlphaFold’s MSA dependency is a practical liability.

The MIT license and the scale of the ESM Atlas release position this as infrastructure rather than a proprietary capability. Whether that translates to therapeutic development outcomes depends on what drug discovery teams build with it over the next several years.