AtlasFold

A protein language model is usually consulted once. The sequence goes in, an embedding comes out, and the folding network works from that fixed snapshot for every recycling iteration that follows. AtlasFold does the opposite: at each recycle it re-runs its language model on a freshly masked copy of the input, with 15% of residues hidden anew each time, and folds from the representations that come back. The structure module therefore sees a family of slightly different readings of the same sequence rather than one frozen encoding, and the pair representation it builds on is refreshed alongside it.

AtlasFold is the monomer-folding member of the Atlas family, developed by Woo Youn Kim's group at KAIST with HITS Inc. and released in September 2026. Its backbone is AtlasLM-3B, a 3.06-billion-parameter masked language model pretrained on roughly 1.56 billion sequences drawn from UniRef and the metagenomic collections MGnify and MetaClust. On top of those representations sits a 215-million-parameter folding trunk adapted from AlphaFold 3 and narrowed to proteins alone.

That narrowing is the second design decision worth understanding. AlphaFold 3 carries a general molecular tokenization capable of describing nucleic acids, ligands and modified residues; AtlasFold restricts the problem so that every token is exactly one amino acid. The resulting representation is homogeneous, which is what lets the model run four recycles and a short diffusion schedule where a general-purpose co-folding model needs ten recycles and 200 fixed steps. It places AtlasFold in the single-sequence folding lineage that runs from ESMFold through SimpleFold and ESMFold2, and it was built as part of the same KAIST effort that produced K-Fold.

#Key Features

• Stochastic language-model encoder: AtlasLM recomputes both the single and pair representations at every recycling iteration from a newly masked input, rather than embedding the sequence once and caching it.
• Protein-only folding trunk: a 48-block Pairformer feeding a 12-block diffusion head, with one token per residue, replaces the general biomolecular representation of AlphaFold 3.
• Length-adaptive diffusion: 20 sampling steps for chains up to 512 residues, 30 up to 1,024 and 100 beyond, instead of a fixed schedule applied to every target.
• Calibrated confidence: on a 10,000-structure holdout, mean pLDDT tracks measured lDDT-Cα at Pearson r = 0.829 and pTM tracks TM-score at r = 0.909, making the scores usable for selecting among sampled structures.
• Memory-lean, batchable inference: 39.81 GiB peak memory at 2,048 residues against 160.32 GiB for ESMFold2, and batching lifts throughput on 64-residue monomers from 0.775 to 9.663 sequences per second.

#Technical Details

AtlasLM-3B is evaluated without any structural supervision on long-range contact precision over partition 4 of the ESM structural split, where it exceeds ESM-2 at both 3B and 15B scale, though ESMC-6B remains higher. The folding stack is trained on experimental PDB structures with a 2020-05-01 release cutoff plus roughly 16 million AlphaFold2 predictions of MGnify sequences distributed through the OpenFold Portal, in four stages at crop sizes of 256, 384, 512 and 640 residues for 45,000, 45,000, 8,000 and 2,000 optimizer steps. On CAMEO22 (183 targets) it reaches a mean TM-score of 0.865 and lDDT of 0.838, the strongest of the PLM-based group and close to AlphaFold 2 at 0.879 TM-score. The gap widens on the harder CASP sets: 0.732 on CASP14 (70 targets) and 0.701 on CASP15 (56 targets), where AlphaFold 2 reaches 0.844 and 0.751. Speed is where the architecture pays off — five structures for a 1,024-residue chain in 11.57 seconds, against 34.38 seconds for a single AlphaFold2 model.

#Applications

The combination of no alignment search, modest memory and batched execution suits AtlasFold to work where the target count rather than the individual prediction is the constraint: annotating proteome-scale sequence sets, folding metagenomic proteins that have too few homologs for an alignment to help, and generating structural input for downstream screening pipelines. Because the training code, staged configurations and processed datasets ship with the model, it also functions as a base that groups can retrain or extend — a fine-tune of AtlasFold on protein complexes produces the AtlasFold-Multimer sibling.

#Impact

AtlasFold's contribution is as much a release practice as a result. The code, the model weights, the intermediate pretraining and staged folding checkpoints, and the processed LMDB training corpora are all published under the MIT License, which makes the full trajectory from a raw sequence corpus to a working folder reproducible rather than merely described — an unusual posture among structure predictors of this scale. The honest ceiling is that alignment-based prediction remains more accurate on difficult targets, so the case for AtlasFold rests on throughput, licensing and trainability rather than on topping a leaderboard. The work is a preprint and has not yet been peer reviewed.