Google deploys AI agents to automate wet-lab discovery

Google Research has demonstrated new systems that bridge computational AI design with real-world wet-lab robotics through its VERified Infrastructure for Discovery Laboratories (VERID) framework. Operating on top of the open-source PyLabRobot tool, VERID provides a live 3D digital twin to track fluid volumes, simulate protocols in seconds, and execute hardware safety checks before running automated pipetting tools. In operational tests, an AI agent running self-play trials autonomously fixed liquid mixing depth errors, recovered from plate-reader outages without re-pipetting, and improved execution accuracy by over fivefold when evaluating overhead camera data.

To test continuous physical feedback, the team established a closed-loop testing environment using DNA aptamers across 7–10 day iteration cycles targeting myoglobin and α-thrombin. When initial α-thrombin runs yielded 0% success because primer overhangs disrupted folding in 48% of designs, the agent adapted by adding clamped stems, raising the hit rate to 42.1% with 98.5% specificity. On myoglobin, an over-enrichment of G-quadruplex motifs at 63% dropped target specificity from 75.0% to 19.1%. Recognizing the off-target binding artifact from lab data, the agent altered its design constraints to recover specificity to 72.4% and achieve a ninefold surge in on-target hit rate to 33.5%.

For complex protein engineering, Google targeted fibroblast activation protein alpha (FAP), a major solid tumor marker. By introducing surgical epitope excision during backbone folding predictions, the team accelerated compute speeds by 8.6x. Combined with yeast surface display and fluorescence-activated cell sorting (FACS), the agent-guided workflow nearly doubled the in-silico hit rate relative to a human researcher working alone, successfully yielding FAP binders with single-digit nanomolar binding affinity.

These advancements demonstrate how grounding software agents in real-world lab feedback transforms drug discovery. By combining simulated dry runs with real-time experimental readouts, practitioners can efficiently search vast sequence landscapes—such as the 20^100 possible variations of a 100-amino-acid protein—while bypassing typical physical bottlenecks and compute limitations.