[R] Genomic Large Language Models

2026-03-17 · Source: Machine Learning · Field: Science & Research — Life Sciences & Biology, Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Intermediate, quick

Summary

Arc Institute's genomic foundation model, Evo2, trained on 9.3 trillion nucleotides, was explored to assess its ability to capture biological relationships beyond simple sequence similarity. Researchers extracted embeddings from Evo2's intermediate layers for 512bp windows across 25 human genes and compared these against BLAST, a standard sequence alignment tool. While many strong matches were initially driven by common repeat elements like Alu, stricter filtering revealed a significant pairing: a section of the VIM gene (chr10) and a section of the DES gene (chr2) showed high embedding similarity (cosine = 0.948) despite lacking detectable sequence alignment. Both regions are active promoters in muscle and connective tissue, share regulatory proteins, and originate from related, co-expressed genes, suggesting Evo2 can recognize gene regulation patterns.

Key takeaway

For AI Scientists exploring genomic data, Evo2 demonstrates a promising capability to identify functional relationships in DNA that traditional sequence alignment tools miss. While the signal currently requires heavy filtering to distinguish from noise, your team should consider integrating genomic LLM embeddings into early-stage research to uncover non-obvious regulatory connections between genes, potentially accelerating discovery in gene function and disease mechanisms.

Key insights

Genomic LLMs can identify functional similarities in DNA beyond sequence alignment, indicating learned regulatory patterns.

Principles

• Embeddings can capture functional similarity.
• Sequence alignment detects raw DNA similarity.

Method

Extract embeddings from a genomic LLM for DNA windows, then compare these embeddings to identify similarities not detectable by traditional sequence alignment tools like BLAST.

In practice

• Use Evo2 to find regulatory patterns.
• Filter for meaningful biological signals.

Topics

• Genomic Large Language Models
• Evo2 Model
• Gene Regulation
• Biological Embeddings
• Sequence Alignment

Best for: AI Scientist, AI Researcher, Research Scientist, Data Scientist

Related on AIssential

• Explicit dynamic cross-strand interactions for DNA sequence language modelling — Explicitly modeling DNA's double-strand interactions improves genomic sequence language model performance and efficiency.
• AI can write genomes — how long until it creates synthetic life? — The Evo2 genomic language model, an AI designed to read, interpret, and generate DNA, RNA, and protein sequences, marks a significant advancement in synthetic biology by designing whole genome sequenc…
• Viral Proteins Reveal Geometry of Protein Language Models — Protein language models organize sequence embeddings by a "nativeness axis" while preserving specific biological group signals.
• GenoMAS: A Multi-Agent Framework for Scientific Discovery via Code-Driven Gene Expression Analysis — GenoMAS automates gene expression analysis using a multi-agent LLM framework that balances structured workflows with adaptive, code-driven problem-solving.
• Biological Reasoning-Informed Regression for Interpretable Regulatory DNA Activity Prediction — Integrating structured biological knowledge with LLMs via two-stage training improves interpretable DNA regulatory activity prediction.
• Single-round evolution of RNA aptamers with GRAPE-LM — GRAPE-LM enables single-round RNA aptamer evolution using AI, outperforming multi-round human methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.