Google DeepMind has announced AlphaGenome 2, a model that predicts how DNA sequences and genetic variants affect specific human cell types. It is available now for non-commercial research through the AlphaGenome API, while commercial research access is in private preview.
The earlier AlphaGenome model predicted DNA activity across broad tissues. AlphaGenome 2 takes a DNA sequence and a chosen tissue, cell type or cell state as inputs. According to Google DeepMind, that lets researchers study effects that may be missed when predictions are averaged across a tissue.
The model predicts effects in individual cell types
Google DeepMind says it trained the model on more than 107 million individual cell profiles, alongside other human and mouse data. It predicts measures including gene expression, chromatin accessibility and DNA methylation. Researchers can also give it gene activity data from a few dozen cells to make predictions for a cell state the model has not seen before, without retraining it.
In a test using the TenK10K dataset, which maps 28 primary immune cell types, Google DeepMind says AlphaGenome 2 identified 30% of causal variants, compared with 16% for the earlier model fine-tuned on single-cell data. It predicted the direction of gene effects in 43% of cases, compared with 23%. The company also says the model reproduced a published finding about a variant that increases activity of the ETS2 gene in gut immune cells.
Google DeepMind has precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions. The company says these predictions can help researchers investigate which variants, genes and cell types may be involved in traits and disease. AlphaGenome 2 is intended for research only and has not been validated or approved for clinical or diagnostic use.
Google DeepMind says interactive Colab tutorials are available now alongside AlphaGenome 2 through the API. It plans to add AlphaGenome 2 Skills to its Antigravity development platform soon.
