MSA-free protein-complex structure prediction from one sequence per chain, fine-tuned from AtlasFold on a 3B-parameter protein language model.
No providers recorded yet. Browse all providers
A protein language model is pretrained on single sequences. Its attention maps carry strong signals about which residues touch inside one fold, but nothing in that training tells it how two separate chains meet. AtlasFold-Multimer (AtlasFold-M) keeps the language model to what it learned: before the pair representation enters the folding trunk, the language-model pair features are masked so that only intra-chain residue pairs survive. Every inter-chain pair is then formed and refined by the trunk's triangular update layers, which learn interfaces from solved complexes rather than from sequence statistics.
AtlasFold-M is the protein-complex member of the Atlas family, developed by Woo Youn Kim's group at KAIST with HITS Inc. and released in September 2026. It is a fine-tune of the monomer folder AtlasFold: the same frozen AtlasLM-3B masked language model (3.06 billion parameters, pretrained on roughly 1.56 billion UniRef, MGnify and MetaClust sequences) feeds a folding stack derived from AlphaFold 3, extended to about 220 million parameters. It sits in the single-sequence co-folding lineage alongside ESMFold2, and it takes on the role that AlphaFold-Multimer played for AlphaFold 2, without the alignment search.
atlasfold multimer command.Fine-tuning runs in three stages of 21,000, 2,000 and 2,000 updates at crop lengths of 384, 640 and 768 residues, with up to six chains per crop in the first two stages and eight in the last; the final stage freezes the trunk from its exponential-moving-average weights. Training data mix 177,363 experimental PDB complexes released by 2021-09-30, 29,431 AlphaFold-Multimer predictions of PDB sequences with disordered regions, and the MGnify AlphaFold 2 distillation monomers inherited from AtlasFold. Chains are clustered at 40% identity and interfaces are sampled at twice the weight of single chains. The loss follows AlphaFold 3's weighting of distogram, diffusion and confidence terms, except that the PAE term is active in all three stages. The authors tested adding AlphaFold-Multimer-predicted homo- and heterodimers from an expanded AlphaFold Database; it helped general protein–protein prediction but hurt antibody–antigen accuracy, so the released model omits it.
On FoldBench, evaluated without MSAs or templates over 172 antibody–antigen and 278 protein–protein targets, AtlasFold-M reaches confidence-ranked DockQ success rates (DockQ > 0.23) of 46.96% and 63.36%. The antibody–antigen figure sits between AlphaFold 3 (49.29%) and Protenix-v1 (44.95%), both run with MSAs, and above Boltz-1 (34.97%). On protein–protein targets it trails every co-folding comparator tested, including ESMFold2 at 69.22%. At 1,024 total residues it generates five samples in 33.24 seconds using 16.05 GiB, against 39.17 seconds and 45.24 GiB for ESMFold2.
The model suits workflows where many protein–protein pairings need a structural answer and building paired alignments for each one is the bottleneck: screening antibody or nanobody candidates against an antigen, triaging interaction hypotheses from proteomics, or folding complexes of metagenomic proteins with few homologs. Its ipTM and interface scores give a ranking signal for sampled poses. The framework handles proteins only; ligands, nucleic acids and modified residues fall outside it, and the same group directs general biomolecular complexes to K-Fold.
AtlasFold-M's main contribution is a demonstration of transfer cost: a monomer folder built on a pretrained language model became a competitive antibody–antigen predictor with a modest additional training budget, which lowers the barrier for groups that want to study co-folding without training a language model and trunk from scratch. Code, weights, the processed multimer training set with its template resources, the stage-handoff checkpoints and the FoldBench prediction archives are all published under the MIT License, though the project notes that the underlying RCSB PDB, MGnify and AlphaFold records remain subject to their original providers' terms, and the roughly 1.56 billion sequences AtlasLM-3B was pretrained on are described but not released. The candid limit is on general protein–protein interfaces, where MSA-based co-folding models and the larger-backbone ESMFold2 remain ahead, and the authors attribute the gap to backbone size and fine-tuning scale. The work is a preprint and has not been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.