Google DeepMind has released AlphaGenome 2, a genomics AI model designed to predict how genetic variants affect particular cell types rather than produce predictions across many tissues at once. The system is available through an API for noncommercial research, with commercial access offered through a closed preview.
The model shifts the focus toward the cellular context in which a DNA change operates. Most cells in a person share the same genome, but they use different sets of genes. A variant that alters gene activity in an activated immune cell may have little detectable effect in a resting cell, making those distinctions important for understanding disease.
AlphaGenome 2 takes both a DNA sequence and a specified biological state as inputs, then generates predictions for that context. Its predecessor processed sequences of roughly one million DNA letters and predicted properties across thousands of experimental settings. DeepMind says that broad, parallel approach limited the earlier model’s ability to distinguish closely related cell types.
The new system was trained on more than 107 million individual cell profiles spanning 68 tissues in humans and mice. Its outputs cover gene expression, the accessibility of DNA within the nucleus and DNA methylation, a chemical modification involved in regulating gene activity. Together, these measurements describe several layers of the machinery that determines how cells use their DNA.
The model can also use one kind of measurement to predict others. Given gene activity data, for example, it can estimate which genomic regions are accessible. Such predictions could help researchers prioritize follow-up experiments, although they remain computational estimates rather than substitutes for laboratory measurements.
On the TenK10K benchmark, which includes 28 primary immune cell types, AlphaGenome 2 identified about 30 percent of causal variants, compared with 16 percent for a fine-tuned version of its predecessor. It correctly predicted whether a variant increased or decreased gene activity in 43 percent of cases, versus 23 percent for the earlier system. Those results describe performance on that benchmark, rather than a general accuracy rate for all genetic variants.
In one demonstration, AlphaGenome 2 computationally reproduced a previously studied mechanism associated with inflammatory bowel disease. It identified a variant in a regulatory region distant from ETS2 that selectively increases the gene’s activity in gut macrophages, while leaving other tested cell types and genes largely unaffected.
Another application used data from a COVID-19 mRNA vaccine study. The model distinguished activated T cells from resting ones and predicted that a variant associated with severe COVID-19 primarily disrupts activity in the activated state. The example illustrates why analyzing an average across cell populations can obscure a context-dependent effect.
DeepMind says the architecture can adapt to previously unseen cell states without retraining, using gene activity measurements from a few dozen cells. The company has also prepared predictions for hundreds of cell and tissue conditions to support research, including interpretation of genome-wide association studies.
Those studies identify statistical links between genomic regions and traits or diseases, but often leave uncertainty about which variant is responsible and where it acts. AlphaGenome 2 is intended to help narrow those possibilities by prioritizing variants, identifying relevant cell types and, when sufficiently confident, predicting the direction of a change in gene activity. The system is not approved for clinical diagnostics.
Sources: The Decoder