AlphaGenome 2 Found 30% of Causal Variants in a Single-Cell Benchmark
DeepMind's AlphaGenome 2 predicts how DNA variants affect specific cell types and states. It recovered 30% of causal variants in a single-cell benchmark versus 16% for its comparison model, but remains a research-only system that needs independent and experimental validation.
Google DeepMind has released AlphaGenome 2, a DNA sequence model designed to predict how genetic variants alter gene regulation in a specific tissue, cell type or cell state. The company says the model improves on its 2025 predecessor by adding single-cell resolution and by adapting to newly supplied cell profiles without additional training.
The 30-second summary
- What happened? AlphaGenome 2 analyzes up to one million DNA letters together with one biological condition, then predicts gene expression and several regulatory signals for that condition.
- Why does it matter? In DeepMind's TenK10K benchmark, it identified 30% of causal variants, compared with 16% for an adapted version of the original AlphaGenome.
- What is the catch? The result comes from a preprint and company-led benchmarks. The tool generates research hypotheses, not diagnoses, and it still missed most causal variants in that test.
One DNA sequence, one cellular context
Most cells contain nearly the same genome, yet they read it differently. A variant that changes gene activity in an immune cell may do little in a liver cell, and the same cell can behave differently when activated or diseased. That context is central to linking noncoding DNA variants with biological effects.
AlphaGenome 2 accepts a DNA sequence as long as one million base pairs and a selected condition such as a tissue, cell type or cell state. It predicts gene expression, chromatin accessibility, histone modifications and DNA methylation across the sequence. Unlike the first AlphaGenome, the new model is conditioned on one cellular context at a time.
DeepMind says it trained the system with bulk and single-cell datasets representing more than 107 million individual cell profiles across 68 human and mouse tissues. Researchers can also supply single-cell RNA sequencing measurements from a cell state that was not present during training. The model uses those measurements to infer additional regulatory signals without being retrained.
Key number: On the TenK10K single-cell benchmark, AlphaGenome 2 recovered 30% of tested causal variants and predicted the correct direction of their gene effects for 43%, versus 16% and 23% for the comparison model.
A benchmark gain, not a solved problem
The TenK10K dataset links naturally occurring genetic differences to gene-expression changes measured in single cells. DeepMind compared AlphaGenome 2 with the original AlphaGenome adapted for single-cell prediction and with specialized tools. The reported gains suggest that explicitly modeling cellular context can narrow the search for variants worth testing in the laboratory.
The team also revisited an inflammatory bowel disease finding involving the ETS2 gene. Earlier experiments had implicated a regulatory region active in macrophages. DeepMind reports that AlphaGenome 2 reproduced the cell-specific pattern computationally, while Nature quoted one of the original researchers saying the analysis took about 30 minutes instead of the roughly 18 months required for the experimental work.
That comparison illustrates speed, not equivalence. The earlier experiments established biological evidence; the model reconstructed a known result after learning from large public datasets. Its more useful future test is whether it can prioritize previously unknown variants that experiments later confirm.
Where researchers can use it
DeepMind has opened an application programming interface for noncommercial research and offers a private preview for commercial research. The service is intended for thousands rather than millions of live queries, so genome-wide industrial pipelines may still require other versions or local methods.
The release sits between two areas NewTqnia has covered recently. A single-cell atlas of colorectal tumors showed how separating cell populations can uncover signals hidden in bulk tissue. Separately, claims that AI produced major scientific results highlighted why independent verification must follow impressive model outputs. AlphaGenome 2 combines those themes: finer cellular resolution and a need for careful experimental testing.
Before treating predictions as biological answers
The paper is a preprint and the central performance results were produced by the model's developer. Thirty percent recall is a substantial improvement over the reported baseline, but it also means seven in ten causal variants in that benchmark were not recovered. Predictions can reflect gaps and biases in training data, and supplying a new cell profile does not guarantee accurate behavior outside familiar biology. An independent 2026 study of the first AlphaGenome also found that local promoter changes did not always propagate coherently across transcripts. That limitation does not directly test AlphaGenome 2, but it reinforces the need for independent, generation-specific benchmarks.
What comes next
The strongest evidence will come from researchers applying AlphaGenome 2 prospectively, before laboratory results are known. Useful evaluations should test diverse ancestries, rare cell states and disease settings, then measure whether the model improves experimental hit rates over existing tools.
DeepMind explicitly says the system is not validated or approved for clinical or diagnostic use. For now, AlphaGenome 2 is best understood as a faster way to form cell-specific hypotheses about DNA, with the laboratory still responsible for deciding which hypotheses become knowledge.
Verified topics and entities
Sources and citations4 sources
External references used to support the reporting in this article.
Published by
NewTqnia Artificial Intelligence Desk
An institutional editorial team within NewTqnia
