Google DeepMind released AlphaGenome 2 on November 11, 2026, marking a significant shift in genomic prediction tools. Unlike its predecessor, which averaged effects across broad tissues, this new model predicts how DNA variants affect gene regulation in specific cell types and states. It leverages over 100 million single-cell profiles to achieve a 30% recall of putatively causal expression-linked variants at 90% precision, nearly double the 16% recall of AlphaGenome 1.

Moving Beyond Tissue Averages

The core engineering breakthrough here is conditioning the model on cell identity. Traditional models suffer from signal dilution when a variant is active in only one specialized cell type, such as gut monocytes for inflammatory bowel disease or activated T cells for vaccine responses. AlphaGenome 2 allows researchers to input specific tissue, cell type, or state conditions, including those derived from their own single-cell experiments. This approach ensures that the model evaluates one condition at a time, providing readouts for gene expression, RNA splicing, and chromatin accessibility with added CpG DNA methylation data.

Practical Integration for Researchers

For developers and researchers, the most compelling feature is the ability to supply new single-cell RNA-seq data without retraining the model. A cell encoder processes this data into a condition, enabling predictions for unseen datasets and cell states. The model can also predict unmeasured properties, such as chromatin accessibility, from gene-expression data alone, helping labs prioritize experimental targets. However, this flexibility comes with a computational trade-off: each condition requires a separate prediction run, making large-scale comparisons across hundreds of cell types resource-intensive compared to fixed-track multitask models.

Benchmark Results and Limitations

On the TenK10K single-cell immune dataset, DeepMind reports AlphaGenome 2 recovered 43% of variant effect directions at 90% accuracy, up from 23% for the previous version. It also explained a median 34% of within-tissue, cell-type-specific gene-expression variance in excluded genome regions, compared to 23% for the predecessor. Despite these gains, the model still misses 70% of causal variants, and its predictions weaken in hard-to-access chromatin regions or for transient disease states. Crucially, all benchmark numbers are from DeepMind's own paper, and independent replication is still pending.

Availability and Use Cases

AlphaGenome 2 is currently available via the AlphaGenome API for non-commercial research, with commercial access in private preview. DeepMind has precomputed resources mapping enhancer-gene links across 640 conditions and scoring over 6.8 million variants across 510 human conditions. While Genomics PLC CEO Peter Donnelly expressed interest in incorporating the tool into its Mystra platform, the model remains strictly a research instrument. It has not been validated for clinical or diagnostic use, and cell state representations are based on observational correlations rather than causal mechanisms.

Key Takeaways

  • AlphaGenome 2 doubles variant recall (30% vs 16%) by predicting effects in specific cell states rather than tissue averages.
  • Researchers can use their own single-cell RNA-seq data for zero-shot condition prediction without model retraining.
  • The model is computationally intensive, requiring separate runs for each cell condition.
  • It remains a research tool only, with no clinical validation or approval for diagnostic use.

The Bottom Line

This is a sharp tool for narrowing the search space in drug discovery, but don't mistake better benchmarks for medical breakthroughs. If you're working with single-cell data, fire up the free API and test your hypotheses today; if you're waiting for clinical applications, keep waiting for independent validation.