What changed
Google DeepMind released AlphaGenome 2, a DNA sequence model trained on more than 100 million cell profiles across 68 tissues. It predicts how genetic variants affect specific cell types and states and is accessible through the AlphaGenome API for non-commercial research.
Why it matters
Moving from tissue averages to cell-specific predictions can help researchers prioritize variants that broad models miss. The release is a research instrument, not clinical evidence or a diagnostic system.
What to do
Researchers can compare API predictions with a held-out assay in one cell type, document preregistered metrics and avoid clinical interpretation until prospective independent validation exists.
Caveat
DeepMind says it identified nearly twice as many expression-changing variants as its prior model on a named comparison. That is a first-party result; the model is not validated for diagnosis and commercial use remains a private preview.