AlphaGenome 2 release notes: New conditional DNA sequence model for cell-type specific variant prediction

Overview of AlphaGenome 2

AlphaGenome 2 is a next-generation DNA sequence model designed to predict the impact of genetic variants on specific human cell types and states. By transitioning from the bulk-tissue predictions of its predecessor to a conditional architecture, the model enables researchers to identify how subtle DNA variations influence gene expression within specialized cells, providing a critical tool for understanding human traits and disease mechanisms.

Conditional Architecture and Cell-Type Specificity

Unlike the original AlphaGenome, which predicted properties across multiple tissues in parallel, AlphaGenome 2 utilizes a conditional architecture. The model takes a million-letter DNA sequence as input alongside a specific biological condition—such as a cell type or state—at every stage of its architecture. This design allows the model to learn how the same stretch of DNA is interpreted differently across various cell types.

Key technical capabilities include:

  • Cell Encoder: Scientists can input single-cell RNA-seq data to generate a cell state representation, allowing the model to make predictions for unseen cell states without requiring retraining.
  • Cross-modality Imputation: The model can predict missing experimental data, such as chromatin accessibility, based on a single measurement like gene activity.
  • Expanded Readouts: In addition to gene expression, histone modifications, and chromatin accessibility, AlphaGenome 2 now predicts DNA methylation.

Performance and Variant Recall

AlphaGenome 2 demonstrates significant improvements in precision and recall compared to the original model. Trained on over 107 million individual cell profiles across 68 tissues, the model shows enhanced performance in several key areas:

  • Variant Identification: On the TenK10K dataset, AlphaGenome 2 identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) compared to the original model.
  • Developmental Tracking: The model tracks continuous changes in gene expression, such as the maturation of blood stem cells into immune lineages, with higher fidelity.
  • Disease-Relevant Insights: The model successfully reproduced findings from landmark studies, such as the identification of a variant in a distant regulatory switch that drives Inflammatory Bowel Disease (IBD) by overactivating the ETS2 gene specifically in gut macrophages.

Bridging Genetics and Single-Cell Biology

AlphaGenome 2 helps bridge the gap between Genome-Wide Association Studies (GWAS) and functional cell biology. By precomputing predictions for trait-associated variants, the model assists researchers in two primary ways:

  1. Enhancer-Gene Mapping: It extends the ability to map regulatory switches (enhancers) to target genes in primary human cell types, such as brain astrocytes, where experimental data is often limited.
  2. Causal Variant Pinpointing: It helps identify the specific causal variants and the cell types in which they act for complex traits, ranging from height to type 2 diabetes, while predicting whether a variant increases or decreases gene activity.

Availability and Future Directions

AlphaGenome 2 is currently available for non-commercial research via the AlphaGenome API, with a private preview available for commercial research. Google DeepMind also announced plans to integrate AlphaGenome 2 Skills into their agentic development platform, Google Antigravity, alongside Gemini 4.

Note: AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.

Sources