Today marks the announcement of AlphaGenome 2, a next-generation sequence model designed to predict how DNA sequences and single-letter variants affect specific human cell types and states. The model aims to help researchers identify the cellular contexts in which genetic variants drive traits and disease.
Why this matters
Every cell in the body contains the same DNA, yet cell types and states activate different genes to perform specialized functions. Large efforts like the Human Cell Atlas map gene expression across millions of cells, but predicting how subtle DNA differences change that expression in particular cells remains an open challenge — and key to linking variants to disease mechanisms.
How AlphaGenome 2 builds on earlier work
The original AlphaGenome predicted molecular effects of variants across broad tissues, and AlphaGenome Atlas provided precomputed predictions for all possible single-letter changes in the human genome. AlphaGenome 2 moves from tissue-average predictions to cell-type and cell-state resolution by redesigning the model architecture.
Technical highlights
- Conditional modeling: rather than predicting many experiments in parallel, AlphaGenome 2 conditions on a single "condition" at a time (a tissue, cell type, or state). The target condition is fed into the model along with the million-base DNA sequence at every stage, allowing the model to learn how the same DNA is interpreted differently across cells.
- Cross-modality imputation: by separating a cell’s biological state from measured outputs, the model can impute missing experimental modalities. For example, given single-cell gene activity the model can predict unmeasured properties such as chromatin accessibility, helping to prioritize follow-up experiments.
- Training data: AlphaGenome 2 was trained on bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles spanning 68 tissues.
- Predicted readouts: it predicts the same core DNA function readouts as AlphaGenome (gene expression, histone modifications, chromatin accessibility) and additionally DNA methylation.
Improved cell-type predictions and variant recall
AlphaGenome 2 delivers marked improvements in predicting gene regulation in specific cell types:
- Cell-type specificity: on genomic regions outside its training set, the model explains substantially more of the gene expression variance across cell types within a tissue and can track continuous developmental trajectories (for instance, blood stem cell maturation into erythroid and immune lineages).
- Variant effect prediction: on the TenK10K dataset — a large single-cell map covering 28 primary immune cell types — AlphaGenome 2 identified nearly twice as many causal expression-changing variants (30% vs. 16%) and more directional gene effects (43% vs. 23%) compared with the original AlphaGenome fine-tuned on single-cell data and other specialized tools.
- Pinpointing disease-relevant variants: AlphaGenome 2 reproduced a published finding that a distal regulatory variant drives inflammatory bowel disease (IBD) by overactivating the ETS2 gene in gut macrophages. Computationally, the model singled out the causal variant and showed its selective effect on ETS2 in gut macrophages while leaving other cell types and genes largely unaffected.
Generalizing to new cell types and states without retraining
The model can generalize to entirely new single-cell experiments without retraining: researchers can upload a small single-cell RNA‑seq snapshot (a few dozen cells); the model compresses gene activity into a cell-state representation and conditions on that representation. AlphaGenome 2 can match or outperform models trained or fine-tuned directly on those datasets.
This capability enables predictions for disease- or perturbation-induced cell states. For example, using unseen single-cell data from an mRNA COVID‑19 vaccine study, AlphaGenome 2 distinguished vaccine-activated T cells from resting ones and indicated that a variant linked to severe COVID‑19 disrupts gene activity more strongly in the activated state — an effect that would be missed in resting cells.
Linking human genetics and single-cell biology
To illuminate disease mechanisms, it is crucial to connect GWAS-identified variants to single-cell maps. AlphaGenome 2’s precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions address two major needs:
- Mapping enhancer–gene links in primary human cell types: many disease variants lie in distant enhancers whose target genes are difficult to trace, particularly in primary cells not well represented in CRISPR screens. AlphaGenome 2 extends enhancer-to-gene mapping to primary cell types — for example, brain astrocytes — enabling researchers to link regulatory variants to target genes in the actual cells involved in disease.
- Pinpointing causal GWAS variants and their cellular contexts: GWAS loci are broad and can obscure the causal variant, gene, and cell type. AlphaGenome 2’s precomputed scores help narrow likely causal variants and the cell types where they act; when confident, the model also predicts whether a variant increases or decreases a gene’s activity.
Availability, community engagement, and next steps
AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API and is in private preview for commercial research. The team plans to release AlphaGenome 2 Skills in Google Antigravity, integrating the model with agentic development tools to accelerate scientific workflows.
The developers acknowledge a broad community of testers and dataset contributors and invite researchers to submit interesting AlphaGenome 2 use cases by 30 November; selected projects may receive additional quota and credits. Community feedback will help shape future development priorities.
Research-only notice
AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.
Acknowledgements
The core research credits include Adam Gayoso, Anne Mottram, Alexander Karollus, Guido Novati, Pavol Drotár, John Reid, Laetitia Meng-Papaxanthos, Adam Kosiorek, Eirini Arvaniti, Vincent Dutordoir, Soham De, Tom Ward, Lai Hong Wong, Matteo Perino, Kyle Taylor, Jun Cheng, Natasha Latysheva, Joshua Pan, Dhavanthi Hariharan, Clare Bycroft, Lauren Nicolaisen, Demis Hassabis, Pushmeet Kohli and Žiga Avsec. Additional contributors and communicators are also acknowledged in the project announcement.
