source: Google Research: Grounding AI agents in biological reality: Building blocks for autonomous wet labs

level: research

google research shared progress on ai agents that run wet lab experiments. they built a framework called verid that acts as a digital twin of the lab workcell. it tracks robot arm positions, well volumes, and pipette tips in real time. agents can simulate protocols before execution, catching collisions or empty wells. this grounding lets agents refine their actions based on physical measurements, not just code.

in one case, verid caught uneven dye mixing from dispensing too close to the well floor. the agent then adjusted its strategy. in another, the simulator rejected an 8-channel aspiration that would not fit a narrow trough. the agent replanned on the fly. across three closed loop rounds on two protein targets, ai driven design raised confirmed hit rates to 33.5% and 42.1%. on one target, round 1 had 0% binders due to primer overhang issues.

the team also applied this to protein binder design for fap, a cancer target. by excising just the local binding pocket during folding, they sped up structure prediction 8.6 times. human agent collaboration nearly doubled in silico hit rate versus a scientist alone. top designs reached single digit nanomolar affinity. safety measures include biosecurity screening, simulation before execution, and human oversight of all robot runs.

why it matters: this work shows how ai agents can learn from real lab feedback, potentially speeding up drug discovery and reducing experimental bottlenecks.


source: Google Research: Grounding AI agents in biological reality: Building blocks for autonomous wet labs