AlphaGenome 2 Finds Nearly Twice as Many Causal DNA Variants in Cells — via API

All cells in the body share the same genome, yet they work differently: a heart muscle cell and an immune cell fighting an infection switch on different genes. A DNA mutation can break the function of just one cell type and show no effect in the others. Older models that predicted the effects of mutations from averaged tissues often missed such effects. DeepMind has released AlphaGenome 2, which aims to close this gap.

What changed in the architecture

The first AlphaGenome took a DNA sequence a million letters long and immediately produced thousands of parallel predictions for different tissues and cell lines. That is convenient, but the model's capacity was spread thin, and it distinguished closely related cell types poorly.

The second version takes a different approach. The model is now conditional: its input is DNA plus a description of a biological condition — a tissue, a cell type, or a cell state. This condition is fed in at every stage of the architecture, so the model learns how the same stretch of DNA is read differently in different cells. The condition can be chosen from a catalog or built from your own single-cell experiment using a dedicated cell encoder.

The model was trained on human and mouse data, including more than 107 million single-cell profiles from 68 tissues. In addition to the earlier predictions (gene expression, histone modifications, chromatin accessibility), it now covers DNA methylation — the most frequently requested addition.

What the numbers and examples show

On the TenK10K dataset, a single-cell map of 28 immune cell types, AlphaGenome 2 correctly identifies 30% of causal variants versus 16% for the first model fine-tuned on single-cell data. It guesses the direction of the effect (whether a gene is switched on more strongly or more weakly) in 43% of cases versus 23%.

There are also real-world examples. A recent study showed that a variant in a distant regulatory region causes inflammatory bowel disease by rewiring the ETS2 gene in macrophages. The model reproduced this computationally: it singled out the causal variant and showed that it boosts ETS2 specifically in gut macrophages while barely affecting other cells.

Another capability is working with unknown states without retraining. A snapshot of gene activity (scRNA-seq) from just a few dozen cells is enough. On data from a COVID-19 vaccine study, the model distinguished activated T cells from resting ones. It also showed that a variant linked to severe COVID-19 disrupts gene function more strongly specifically in the activated state, while in resting cells this effect would go unnoticed.

Separately, DeepMind precomputed predictions for variants found in GWAS (the search for associations between DNA regions and traits such as height or type 2 diabetes) across hundreds of conditions. This helps researchers identify the causal variant, gene, and cell without running heavy computations themselves.

Limitations and what to do in practice

There are important caveats. The model is intended for research only and has not been validated for clinical or diagnostic tasks. The metrics are reported by DeepMind itself, and there is no independent verification yet. And even 30% of causal variants found means that most still remain unseen: this is a tool for prioritizing hypotheses, not a replacement for experiments.

For those working in and around biotech, the takeaways are as follows:

• If you are building a platform for target discovery or GWAS analysis, it is worth testing the API: it is open for non-commercial research, and a closed preview is available for commercial use. The first integrations are already being announced, for example by Genomics Limited with its Mystra platform.
• If you have your own single-cell data, pay attention to conditional learning without retraining: even a small snapshot of cells can set the context for predictions.
• The model can fill in missing measurements, for example chromatin accessibility from expression data — a way to decide which experiments to run next.

One more detail for those following agentic tools: AlphaGenome 2 Skills is planned for release in Google Antigravity, where the specialized model will work in tandem with Gemini. So narrow scientific models are gradually becoming tools that an agent calls.