Beyond AlphaFold: ESMFold2 Launches with a 1‑Billion‑Protein Open Atlas

Model overview

ESMFold2 is an open‑source protein‑structure predictor announced in Nature on 27 May 2026. It is trained on the 6.8 billion‑sequence ESM Atlas and can generate predictions for roughly 1 billion proteins. The model replaces AlphaFold’s dependence on multiple‑sequence alignments with a large protein language model that learns statistical relationships between amino‑acid sequences and three‑dimensional folds. A recurrent architecture feeds later‑stage representations back to earlier stages, and a simplified pairwise layer retains only triangular multiplication and feed‑forward transitions. Custom CUDA kernels accelerate each inference loop.

Benchmark performance

On standard structure‑prediction benchmarks ESMFold2 attains state‑of‑the‑art accuracy while being markedly faster, especially for protein‑protein interaction tasks such as antibody‑antigen complexes. Direct comparisons show ESMFold2 surpassing AlphaFold 3 on several key antibody‑design metrics.

Antibody and functional‑protein design workflow

The design pipeline consists of two stages:

Candidate generation in the model’s representation space, producing tens of thousands of designs over roughly two days.

Confidence‑scored ranking of binding affinity and structural stability, completed in under one day. Both stages are trivially parallelizable.

When the inference‑compute budget was increased and the top 84 designs per target were selected, the average success rate of mini‑binders rose from 54 % to 70 % and that of single‑chain antibodies from 12 % to 21 %.

ESM Atlas

The ESM Atlas is a searchable map linking 68 billion protein sequences to 1 billion predicted structures. Using the Atlas, researchers identified a CRISPR‑like protein in a soil fungus and other eukaryotes, suggesting previously unknown gene‑editing capabilities.

Limitations and community perspective

MIT computational biologist Sergey Ovchinnikov cautions that the ESM Atlas should be viewed as a strategic complement rather than a replacement for AlphaFold, which still excels at high‑resolution predictions of small‑molecule ligands and ion binding. While ESMFold2 greatly improves macro‑genomic generalization and high‑throughput screening, its performance on highly atypical protein folds remains modest and requires careful community evaluation.

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
全开源无商用限制的结构库重塑抗体设计与生物学发现。