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

Google DeepMind's AlphaGenome 2 reads a one-megabase stretch of human or mouse DNA together with a target cell state and predicts how a genetic variant changes activity in that specific cell type. Its predecessor averaged across bulk tissue. The new model can also adapt to unseen single-cell datasets without fine-tuning: researchers pass raw single-cell RNA counts and get a condition to run against.

For anyone trying to connect a genetic variant to a disease, the cell type is often the whole question. A variant can matter in one immune cell and not at all in its neighbor.

This is the kind of AI we are glad to see labs spend on. By DeepMind's own evaluation, reported by MarkTechPost, the gains are substantial. On single-cell eQTLs across 28 immune cell types, recall of causal variants at 90% precision rose to 30%, from 16% for a fine-tuned AlphaGenome 1. Recall for direction of effect rose from 23% to 43%. On an unseen Parkinson's cohort it reached 41% recall, versus 17% for a cohort-tuned rival model. If those numbers hold in independent hands, that is a meaningful step toward narrowing which variants deserve expensive lab follow-up.

We would read the results with the usual caveat: these are DeepMind's own benchmarks. Independent replication is what turns a strong paper into a trusted tool. The weak spots are also disclosed. Outside accessible chromatin, sign prediction reaches an AUROC of only 0.63, so large stretches of the genome remain hard.

Our bigger concern is access. AlphaGenome 2 is API-only, free for non-commercial use, with commercial research in private preview and weights promised only on final publication. Open models like Borzoi ship under permissive licenses. Science benefits from being able to inspect, run and audit a model locally, especially when its predictions may shape clinical research. A hosted model also means a single company sees what the research community is querying.

Still, the released atlases, covering roughly 7 million common and GWAS variants across 510 conditions, are a real public good. Used well, this shortens the path from a genetic association to a biological explanation. Used carelessly, model predictions risk being treated as findings. They are hypotheses, and should be labeled that way.

AlphaGenome 2 conditions a U-Net style model through FiLM layers, and adds CpG methylation to outputs including RNA coverage, splicing, chromatin accessibility and contact maps. Its single-cell encoder was trained on more than 100 million human profiles, largely from CELLxGENE, and training used 2,048 TPU v4 chips. A Python package and Colab tutorials are available.