DeepMind says AlphaGenome 2 identifies 30% of causal variants that alter gene expression in single-cell tests, compared with 16% for its predecessor fine-tuned for single cells, per the company.
That gap is the pitch.
AlphaGenome 2 drops the parallel output design of the first AlphaGenome, which predicted thousands of bulk tissue tracks at once from a 1-million-letter DNA sequence, according to DeepMind. The new model is conditional. It takes one tissue, cell type or state at a time and uses that condition through the architecture, so the same DNA can be read differently in gut macrophages versus other immune cells and missing readouts can be imputed from RNA alone.
Sarah Teichmann, co-founder of the Human Cell Atlas consortium and professor at Cambridge, watches single-cell mapping closely. She said AlphaGenome 2 adapts to new single-cell data without retraining, per DeepMind.
In practice that means a snapshot from a few dozen cells can condition predictions for an unseen state, a claim DeepMind illustrates with vaccine-activated T cells where a severe COVID-linked variant hits harder after activation, per the company.
On direction, AlphaGenome 2 correctly called whether expression goes up or down for 43% of effects compared with 23% before on the TenK10K map of 28 immune cell types, per DeepMind.
Training used more than 107 million cell profiles across 68 tissues in human and mouse plus bulk data, according to DeepMind. That training set supports precomputed scores across hundreds of cell conditions and extends earlier work including AlphaGenome Atlas, which covers genome-wide single-letter changes, per the company.
The numbers do not show false positive rates or performance outside blood and immune differentiation, where DeepMind centered validation including an ETS2 macrophage effect in inflammatory bowel disease. Replication in primary cells such as brain astrocytes would change the read, as would prospective testing of state-specific effects reported by DeepMind.
Access is non-commercial research via the AlphaGenome API today, with commercial use in private preview, per DeepMind.
Peter Donnelly, CEO of Genomics Limited and a genetics-to-drug-target specialist, said sequence models help move from association signals to drug programs and plans to use the model in the Mystra platform, per DeepMind. Will that 30% recall hold outside blood and immune cells?
