Google DeepMind has introduced AlphaGenome 2, a model designed to predict how differences in DNA affect particular human cell types, helping researchers investigate the cellular origins of disease. The company says it identifies nearly double the number of variants that alter gene expression compared with its previous model. Researchers can access it today for non-commercial work through the AlphaGenome API, while commercial research access is in private preview.
The model was trained using profiles of more than 100 million individual cells drawn from 68 tissues. Its predictions cover hundreds of human cell types, moving beyond the broad tissue-level activity that the original AlphaGenome was built to assess. Google DeepMind says the finer detail can reveal effects in cells relevant to disease that models working with tissue averages would miss.
Google DeepMind says the redesigned architecture also supports predictions for cellular conditions and disease states absent from the model’s training experience. The company positions that capability as a way for scientists to investigate how small genetic differences influence human traits and disease.
Cells share a DNA blueprint, but different types and states activate or silence different genes to perform their functions. A heart muscle cell and an immune cell responding to infection are examples of that distinction. Efforts including the Human Cell Atlas are charting gene expression across millions of cells, while predicting how subtle DNA changes affect that expression remains an unresolved challenge, according to Google DeepMind.
The earlier AlphaGenome Atlas supplied precomputed predictions for all 9 billion possible changes to a single DNA letter in the human genome, with a dataset spanning 1 petabyte. Its website required no coding, and its AlphaGenome Variant Impact (AVI) score combined predictions concerning regions that encode proteins with those that do not.
Google DeepMind launched the original AlphaGenome last year to assess how genetic variants influence molecular activity at the tissue level. The company says that model and AlphaGenome Atlas help researchers select variants for investigation and interpret their effects, supporting work on rare diseases, complex traits and basic biology.
