Structure prediction · Biohub · 2026

ESMFold2

Predicts all-atom structures of proteins and their complexes from ESM C representations, with an optional alignment for difficult targets. It does not reuse the AlphaFold2-style structure module that ESMFold ended in.

In plain language

The current end of the language-model-to-structure line: a large encoder in front, an all-atom generative decoder behind, and the alignment demoted from requirement to option.

One way to picture it

The same pipeline as ESMFold, with a better reader at the front and a decoder that sketches rather than measures.

Commonly misread as

It is a preprint. The checkpoint you can download today is a small experimental variant whose own card tells you to use the main model for real work.

01 / Why it is here

Standing

The preprint reports complex prediction above established methods, including antibody and antigen interactions, and binder discovery with nanomolar affinities for miniproteins and single-chain antibodies.

02 / What sets it apart

Distinctions

The alignment becomes optional rather than absent or required.
The decoder is diffusion over atoms, which is the same shift AlphaFold3 made.
Inverting the model is used as a design procedure, not only a prediction one.

03 / Where it stops

Limits

A preprint, not yet peer reviewed. The published experimental checkpoint is a 0.2B parameter fast variant released for reproducibility, and its card recommends the main model for research use.

Weights and code

Checked against the registry, not from memory

Repository	Size	Licence	Note
biohub/ESMFold2-Experimental-Fast	0.2B	MIT	Experimental release; single-sequence or MSA-conditioned

Sources

Each number above comes from one of these

Candido et al. 2026, preprint. Language Modeling Materializes a World Model of Protein BiologyPreprint ESMFold2 experimental weightsModel card

Same task, other answers

Structure prediction

Structure prediction2021

AlphaFold 2

Evolution in, coordinates out.

Predicts the structure of a protein from its sequence, an alignment of its homologues and optional templates.

Explore the architecture Structure prediction2024

AlphaFold 3

One model for the whole complex.

Predicts the joint structure of complexes containing proteins, nucleic acids, small molecules, ions and modified residues.

Explore the architecture Structure prediction2023

ESMFold

Fold without searching for relatives.

Predicts a structure from a single sequence with no alignment search at query time.

Explore the architecture

Back to the catalogueCompare it with its neighbours