Use ESMFold2 Online
Commercially Available ESMFold2 No-Code Web Server
ESMFold2: Language-Model-Based All-Atom Structure Prediction for Complexes, Antibodies and Ligands
ESMFold2 is a protein structure prediction model that builds on ESM Cambrian (ESMC) language model representations instead of relying primarily on multiple sequence alignments (MSAs). It predicts all-atom structures of single chains and complexes, including protein-protein, antibody-antigen and protein-ligand systems. It is described in "Language Modeling Materializes a World Model of Protein Biology" by Salvatore Candido, Thomas Hayes, Alexander Rives and colleagues, with work performed at Biohub or EvolutionaryScale. The preprint text we reviewed does not state code, weights or license terms.
High-accuracy structure predictors for complexes have generally leaned on MSAs, which are slow to build and often unavailable for antibody-antigen pairs and designed proteins. The authors argue that language modeling trained on very large sequence collections, including metagenomic data, learns general representations of protein biology that can replace much of that dependence.
A faster variant, ESMFold2-Fast, halves the number of folding layers for speed-sensitive use.
Item | Detail from the paper |
|---|---|
Language model backbone | ESMC (300M, 600M and 6B parameters); 6B representations are frozen |
Folding trunk | 4-layer pair encoder plus 48 recurrent folding layers (24 in ESMFold2-Fast) |
Structure output | Diffusion transformer with sliding-window atom attention |
Input | Protein sequence by default, with optional MSA |
Output | Atomic coordinates with confidence scores such as ipTM |
Speed | 1,024 residues in 15.8 s, about 1.3x faster than AlphaFold3 |
How ESMFold2 Predicts Structures from Language Model Representations
Language model features: activations from all ESMC layers are combined and projected into a two-dimensional pair representation.
Recurrent folding: the trunk uses triangle multiplication and feedforward layers without attention, and more loops at inference improve accuracy.
Diffusion decoding: two final folding layers condition a diffusion transformer that denoises to atomic coordinates.
Sampling and ranking: multiple diffusion samples and seeds per target are ranked by confidence scores such as ipTM.
ESMFold2 Accuracy on Antibody-Antigen, Protein-Protein and Protein-Ligand Benchmarks
Benchmark | Reported result |
|---|---|
Antibody-antigen (FoldBench, DockQ pass rate) | 50% single sequence versus 47% for AlphaFold3 with MSA; 53% with MSA |
Protein-protein | 70% single sequence; 76% with MSA versus 73% for AlphaFold3 |
Protein-ligand (Runs N' Poses) | 66% single sequence |
Inference scaling, antibody-antigen | 49% with one seed rising to 65% with 1,000 samples |
ESMFold2-Fast | 50% antibody-antigen, 68% protein-protein, 63% protein-ligand |
Binder Design Results
Using a simple search procedure with ESMFold2, the authors tested minibinders and scFvs against PDGFRb, EGFR, PD-L1, CD45 and CTLA-4. Higher-compute hit rates were 36 to 88% for minibinders and 15 to 29% for scFvs, with measured affinities from 68 pM to 70 nM. A cryo-EM structure of an EGFR minibinder complex matched the prediction with 1.204 A RMSD.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with ESMFold2 on Tamarind Bio
Antibody-antigen modeling: predict complexes of antibodies and targets without building MSAs.
Binder design filtering: rank minibinder and scFv candidates by predicted confidence, as the authors did.
Protein-ligand complexes: model ligand-bound structures from sequence.
Large-scale structure annotation: the authors applied the model to build an atlas of 1.1 billion predicted structures.
How to Use ESMFold2 on Tamarind Bio
Open the tool: log in to tamarind.bio and open ESMFold2.
Enter sequences: provide one protein sequence or several chains for a complex, plus ligands if exposed.
Choose the variant: select the full model or ESMFold2-Fast, if exposed.
Set sampling: choose the number of seeds, diffusion samples and folding loops, if exposed. More loops and samples improved accuracy in the paper.
Add an MSA: optionally supply one, if exposed.
Run the job: submit and wait for processing.
Download and interpret: retrieve the structures and rank them by confidence such as ipTM.
Things to Keep in Mind
Single-sequence antibody-antigen success is about 50 to 55%, so roughly half of targets still fail.
scFv hit rates are much lower than minibinder hit rates.
Functional assays were reported for PD-L1 only, and epitope competition for a subset of binders.
Some comparisons use subsets because other models did not produce predictions for every target.
Atlas structures are computational, and many clusters remain unannotated.
Source: Candido S, Hayes T, Derry A, Rao R, Lin Z, Verkuil R, et al. Language Modeling Materializes a World Model of Protein Biology. bioRxiv 2026. DOI: 10.64898/2026.06.03.729735. No code repository is listed in the text reviewed.