AlphaGenome: DNA sequence to regulatory function, in one model

DeepMind’s Nature paper on a unifying DNA model: 1 Mb context, base-resolution multimodal tracks, and strong regulatory variant-effect scores, the positive-science sequel to AlphaFold, with clear limits.

Primary source: Nature (DOI 10.1038/s41586-025-10014-0): https://www.nature.com/articles/s41586-025-10014-0

What’s new: AlphaGenome (Nature, DOI 10.1038/s41586-025-10014-0; Avsec et al.) is Google DeepMind’s unifying DNA sequence model for gene regulation. It takes up to 1 Mb of DNA and predicts thousands of functional genomic tracks at up to single-base-pair resolution across many cell types. Reported modalities include gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage, and splice junction coordinates and strength. The model is trained on human and mouse genomes (public consortia such as ENCODE, GTEx, 4D Nucleome, and FANTOM5 in DeepMind’s accompanying description).

On the published benchmarks, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 variant-effect prediction evaluations, and the paper also reports state-of-the-art track prediction on 22 of 24 genome-track tasks. Scoring a variant means contrasting predictions for mutated vs unmutated sequence across modalities at once. A worked example near the TAL1 oncogene shows the model recapitulating mechanisms of clinically discussed non-coding variants. Tools (and a non-commercial API preview) support genome-track and variant-effect queries from sequence.

Why it matters: AlphaFold changed how biology uses AI by mapping sequence to protein structure. AlphaGenome aims at a different, equally concrete problem: most disease-linked variants sit in the non-coding majority of the genome, where “what does this letter change do to regulation?” is hard. Prior DNA models traded long context against resolution, so they could not jointly cover expression, chromatin, contacts, and splicing well. One model that takes a megabase, emits multimodal tracks, and scores variants across those tracks is scientific AI that does something beyond chat: it turns sequence into testable regulatory hypotheses.

It is complementary to AlphaMissense (coding-region missense effects). Coding is roughly 2% of the genome; regulatory sequence is where AlphaGenome’s bet sits. For Bitware’s beat, that is the positive applied-AI story: structure was Act I; regulatory genomics is a serious Act II.

Caveats: Research tool, not a clinical diagnostic. DeepMind’s own write-up flags limits that matter: influence from very distant elements (on the order of >100 kb) remains hard; cell- and tissue-specific patterns still need work; the system was not designed or validated for personal-genome prediction; molecular track and variant scores are not a full account of complex traits or disease (development and environment sit outside the model). Benchmark wins are against published external models on defined tasks, not a blank check for every locus. Use AlphaGenome to prioritize and explain regulatory hypotheses, then test them in the wet lab or clinic with ordinary scientific caution.