Google Research turns AI agents toward scientific discovery

Google Research’s science AI group, led by Lizzie Dorfman, is applying AI systems across genomics, neuroscience, epidemiology and climate science. Early work centered on DeepVariant, a variant-detection tool that won the Precision FDA competition for overall accuracy and has since helped process something like 2.5 million human exomes and genomes.

Gemini has shifted the team toward Empirical Research Assistants, or ERA, which use coding agents and tree-search methods to test many computational approaches. In one epidemiological forecasting example, the team generated 200,000 models to evaluate, allowing weak candidates to be discarded quickly and promising directions to surface faster than traditional serial experimentation.

The systems are also being used to reproduce scientific papers from short method summaries and will appear within Gemini for Science, with computational discovery as a product experiment. Human oversight remains central: a curved-solar-panel example produced a physically impossible design until researchers added validation checks.

Google Research is also pursuing long-term challenges such as brain mapping, where reducing computational cost by multiple orders of magnitude is a prerequisite. Breakthroughs were framed as cumulative progress on bottlenecks, citing DeepConsensus with PacBio and something like 45 papers in Nature and Science in the last four years.