AlphaGenome 2 adds molecular readouts and variant resources

Google DeepMind says AlphaGenome 2 predicts how DNA sequences and genetic variants affect specific human cell types, with the aim of helping scientists identify cellular drivers of disease.[1] Unite.AI reports that its predicted molecular properties include gene expression, RNA splicing, chromatin accessibility and CpG DNA methylation.[4] It also reports that the model can use gene-expression data to predict properties that were not measured, including chromatin accessibility.[4] These capabilities broaden the kinds of molecular questions researchers can explore with the model; they do not, on their own, establish a disease mechanism.

Google DeepMind says the model zooms in from broad tissue averages to individual cell types and states, where disease-relevant variant effects may otherwise be missed.[1] The company says it can make predictions for cell and disease states it has not previously encountered.[1] Researchers can upload single-cell RNA-seq data to adapt the model to unseen biological states, according to Google DeepMind.[2] Nature also reported that researchers can provide new single-cell gene-activity data and generate gene-variant predictions.[5] This offers a way to test predictions in a particular cellular context, rather than relying only on tissue-level averages.

Google DeepMind says AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 tissues and identifies nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model.[1] Separate reporting found that it recovered 30% of putatively causal expression-linked variants at 90% precision, compared with 16% for AlphaGenome 1 fine-tuned on single-cell data.[4] The Verge reported that the model’s more detailed predictions are more computationally expensive than its predecessor’s.[3] These reported results indicate a performance-cost trade-off, not a guarantee that every prediction will be useful in a given study.

Google DeepMind says the AlphaGenome Atlas contains predictions for all 9 billion possible single-letter changes in the human genome.[1] Nature reported that the earlier Atlas predictions did not specify which cell type a variant was likely to affect.[5] Other resources reported by Unite.AI map enhancer-gene links across 640 primary single-cell reference conditions and 39,323 genes, and score more than 6.8 million single-letter variants across 510 human conditions.[4] Google DeepMind makes AlphaGenome 2 available for non-commercial research through its API and offers a private preview for commercial research; Unite.AI reports non-commercial use is free, with query rates varying according to demand.[1][4]