Biological Molecules As Language

2026-06-26 · Source: The Batch | DeepLearning.AI | AI News & Insights - www.deeplearning.ai · Field: Science & Research — Life Sciences & Biology, Artificial Intelligence & Machine Learning, Health & Medical Research · Depth: Advanced, short

Summary

ESMFold2, a 6.2 billion parameter model developed by Biohub and EvolutionaryScale, infers the shapes of biologically active molecules like proteins, DNA, RNA, and their binders. It refines Google's AlphaFold 3 architecture by treating molecular components as natural language, uniquely using a separate transformer (ESMC) to embed individual molecules directly, unlike AlphaFold 3 which primarily relies on multiple sequence alignments (MSAs). ESMFold2 outperforms AlphaFold 3 and other models when MSAs are not provided, achieving 0.85 lDDT for proteins compared to Chai-1's 0.81 lDDT, and an 80 percent DockQ pass rate for protein-DNA complexes versus Chai-1's 71 percent. When MSAs are available, its performance is comparable, matching AlphaFold 3's 0.89 lDDT for proteins. The model and its weights are freely available via Biohub's website and HuggingFace.

Key takeaway

For research scientists or drug developers working with novel or synthetic biological molecules, ESMFold2 offers a significant advantage by predicting molecular shapes without requiring multiple sequence alignments. This capability streamlines your workflow, especially when related molecular data is scarce. You should consider integrating ESMFold2's open-source model and API into your computational biology pipelines to accelerate discovery and reduce data preparation overhead.

Key insights

ESMFold2 predicts molecular shapes using language model-like embeddings, reducing reliance on multiple sequence alignments.

Principles

• Transformer embeddings can replace MSAs for molecular structure prediction.
• Iterative refinement of embeddings improves inference performance.

Method

ESMFold2 embeds input sequences, atoms, and optional MSAs, then iteratively refines a distance embedding, and finally uses a diffusion model to deduce atom coordinates and estimate errors.

In practice

• Use ESMC for embedding individual protein or nucleic acid sequences.
• Apply ESMFold2 for novel or synthetic molecules lacking MSA data.

Topics

• Molecular Structure Prediction
• Protein Folding
• Large Language Models
• AlphaFold 3
• ESMFold2
• Biohub

Code references

• chaidiscovery/chai-lab

Best for: AI Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Move over, AlphaFold: open-source model predicts shape of 1 billion proteins — ESMFold2, an open-source AI, predicts over a billion protein structures, outperforming AlphaFold3 and enabling novel protein design.
• [R] Genomic Large Language Models — Genomic LLMs can identify functional similarities in DNA beyond sequence alignment, indicating learned regulatory patterns.
• Viral Proteins Reveal Geometry of Protein Language Models — Protein language models organize sequence embeddings by a "nativeness axis" while preserving specific biological group signals.
• Augmenting Molecular Language Models with Local $n$-gram Memory — Integrating local $n$-gram memory via MolGram efficiently resolves the locality gap in molecular language models, improving performance with fewer parameters.
• Bridging three-dimensional molecular structures and artificial intelligence with a conformation description language — ConfSeq enables large language models to process 3D molecular structures by converting them into discrete token sequences.
• Data-driven classification of Escherichia coli using protein language model ascertains O-serotype determining genes — Protein Language Models can accurately classify *E. coli* O serotypes, improving upon traditional database-dependent methods.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Batch | DeepLearning.AI | AI News & Insights - www.deeplearning.ai.