AtlasFold: Protein structure prediction with metagenomic-scale language models

Abstract

Protein language models (PLMs) trained on evolutionary sequences learn representations that encode protein structure, enabling direct structure prediction without multiple-sequence alignments (MSAs). Here we present the Atlas model family, an open and trainable system spanning protein language modeling, monomer folding, and protein-complex prediction. AtlasLM-3B is a 3B-scale language model trained with masked language modeling on approximately 1.56 billion sequences, including metagenomic data, and outperforms the similarly sized ESM2-3B in unsupervised contact prediction. Building on these representations, AtlasFold predicts all-atom protein structures and achieves state-of-the-art accuracy among PLM-based folding models. Fine-tuning AtlasFold for protein-complex prediction produces AtlasFold-Multimer (AtlasFold-M), whose antibody–antigen prediction performance is comparable to that of AlphaFold3 and ESMFold2. This protein-specific folding architecture enables fast, memory-efficient inference with AtlasFold and AtlasFold-M. By releasing the training code and data, stage checkpoints, and model weights under the MIT License, we provide a foundation for advancing PLM-based protein structure prediction.1