DeepMind's AlphaGenome 2 doubles detection of gene-altering DNA variants

Summary

• Google DeepMind released AlphaGenome 2, predicting DNA variant effects in hundreds of individual human cell types.
• The model was trained on over 107 million single-cell profiles spanning 68 tissues.
• It is free for non-commercial research via API; commercial access stays in private preview.

The latest

Google DeepMind on November 11, 2026 released AlphaGenome 2, a DNA sequence model that predicts how genetic variants behave inside specific cell types rather than broad tissue averages. On one immune-cell benchmark it flagged nearly twice as many causal variants as its predecessor. DeepMind said the model turns static cell atlases into a predictive engine for human genetics.

Details

• The benchmark: On TenK10K, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 reliably identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools, according to DeepMind. Directional gene effects rose to 43% from 23%.
• Training scale: DeepMind said the model was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. The company describes the shift from tissue averages to individual cell types and states as a step-change in variant predictive performance.
• Architecture change: Rather than predicting in parallel, AlphaGenome 2 is a conditional model handling one condition at a time — a tissue, cell type or state. Alongside a million-letter DNA sequence, it takes the target biological condition as an input at every stage of its architecture.
• Filling data gaps: The conditional design enables cross-modality imputation, DeepMind said, letting the model predict unmeasured properties from a single input. A scientist supplying only single-cell gene activity can obtain predictions for readouts such as chromatin accessibility without running the experiment.
• New readout: The model predicts the same core readouts of DNA function as AlphaGenome — gene expression, histone modifications and chromatin accessibility — and adds DNA methylation, which DeepMind described as a widely requested capability among researchers.
• Unseen cell states: DeepMind said the model generalizes to entirely new single-cell experiments without retraining. Researchers can upload an RNA-seq snapshot of gene activity from a few dozen cells, compressed into a cell state representation used to condition the model on states it has not encountered.
• Precomputed results: To spare researchers large-scale computation, DeepMind precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions. The earlier AlphaGenome Atlas carried predictions for all 9 billion possible single-letter changes in the human genome.
• Industry reaction: Peter Donnelly, CEO of Genomics Limited, called the model "an impressive advance in the field" and said his company expects to incorporate it into its Mystra platform. He said genome-wide association studies rarely translate directly into what a drug program needs.
• Access terms: AlphaGenome 2 and interactive Colab tutorials for expression conditioning are available for non-commercial research through the AlphaGenome API, with commercial research limited to private preview. DeepMind said AlphaGenome 2 Skills will be released in Google Antigravity, its agentic development platform, without giving a date.
• Clinical limits: DeepMind stated that both AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.

Background

DeepMind launched the first AlphaGenome last year to predict how DNA variants affect molecular activity across broad tissues, later adding AlphaGenome Atlas. The company said those tools helped researchers prioritize and interpret variant effects in rare diseases, complex traits and fundamental biology.

Between the lines

The commercial boundary is the signal here. Free API access for academics alongside a gated private preview for industry points to where DeepMind expects value to accrue — in drug discovery, the use case Donnelly described. The model's ability to predict unmeasured readouts and unseen cell states also lowers the experimental cost of genetics work, the lever pharmaceutical buyers would pay for.

What's next

Watch for the release date of AlphaGenome 2 Skills in Google Antigravity, the terms on which commercial access leaves private preview, and the first independent benchmarks against datasets other than TenK10K.

Source: Google DeepMind