Google DeepMind has unveiled AlphaGenome 2, a next-generation AI model designed to predict how genetic variants affect specific human cell types and biological states. Building on the original AlphaGenome released last year, the new model was trained on more than 100 million individual cell profiles across 68 tissues to address the challenge of identical DNA expressing differently across varied cell types.

According to Google DeepMind, AlphaGenome 2 identifies nearly twice as many gene-expression-altering variants as its predecessor. The model demonstrated the ability to make predictions about unseen cell types and biological states without requiring retraining, effectively converting static cell atlases into predictive genetics engines. DeepMind researchers emphasized that this cell-specific granularity is crucial for identifying viable drug targets and understanding disease mechanisms at the cellular level.

Looking ahead, DeepMind plans to expand its genomics capabilities by generating datasets for underrepresented cell states and exploring multi-species training data. The ultimate long-term objective is to contribute to digital simulations of biological systems, often referred to as virtual cells, to streamline laboratory research.

Why it matters

  • Biotech and pharma founders can leverage granular cellular predictions to prioritize drug targets and optimize expensive wet-lab experimentation.

  • Zero-shot generalization to un-trained cell types indicates underlying biological representation capabilities in multi-modal life sciences models.

  • Investors should note DeepMind’s aggressive expansion into computational biology platforms that transform static genomic data into predictive software.

Source: theverge.com