DeepMind opens AlphaGenome 2 API with 9B-variant atlas

Google DeepMind has shipped the AlphaGenome Python client on GitHub, opening programmatic access to AlphaGenome 2 and a companion dataset the lab calls the AlphaGenome Atlas. The release pairs a multimodal DNA sequence model with a whole-genome, pre-computed variant atlas.

The README describes AlphaGenome 2 as "our latest multimodal DNA sequence model, with improved cell-type specificity, generalization to your own human scRNA-seq data (expression conditioning), and imputation of unobserved modalities across 1 Mb sequences." The Atlas, per the same page, is "genome-wide in silico mutagenesis across 9 billion human variants, providing pre-computed variant effect scores and the AlphaGenome Variant Impact (AVI) score on a zero-code website and through the API." The original AlphaGenome paper, Avsec et al. (2026), appeared in Nature 649(8099), pp. 1206–1218.

Access terms are tight. The API is "offered free of charge for non-commercial use," with commercial access routed through Google Cloud, and DeepMind warns that "predictions are for theoretical modelling and research purposes only" and must not be used for clinical decision-making. Outputs cannot be used to train other machine learning models. Two researchers in our Who's Who tracker shared the repo link on the day it surfaced.