AI agents gain physical lab grounding in autonomous wet-lab research
Google Research reports VERID, a digital twin for wet labs, plus closed-loop aptamer and protein binder design where agents run and refine experiments.
The story in focus
Google Research described three projects linking AI agents to physical wet-lab work: VERID, a 3D digital twin of a lab workcell that simulates and validates robot protocols; a closed-loop aptamer design campaign over three cycles; and an agent-guided pipeline yielding single-digit nanomolar protein binders for a cancer target.
- VERID is a semantic digital twin built on PyLabRobot that tracks volumes, tips and robotic geometry, and dry-runs protocols before execution.
- In a three-round aptamer campaign, agents recognized G-quadruplex sticky binders from readouts and revised design rules, raising on-target hit rate to 33.5% for myoglobin.
- For FAP, excising the local binding pocket during folding accelerated cycles more than eightfold; top designs reached single-digit nanomolar affinity.
Grounding agents in the bench
Google Research outlined work aimed at letting AI agents run experiments rather than only propose them. Its VERID framework builds a live, stateful 3D model of a laboratory workcell, tracking robotic arm geometry, well volumes, reagent identities and pipette tip inventory. Agents check state before acting, and protocols are simulated in seconds to catch collisions or depleted wells before reagents are touched.
The post describes case studies: synchronized video and execution logs revealed uneven dye stratification from dispensing too close to the well floor; the digital twin let an agent recover from a plate-reader failure without re-pipetting; and a simulator rejected an 8-channel aspiration that would not fit a narrow trough.
Aptamers and protein binders
Using DNA aptamers as a fast testbed, agents completed three design-test-iteration cycles on two protein targets. Confirmed on-target hit rates reached 33.5% for myoglobin and 42.1% for α-thrombin. On myoglobin, agents enriched G-quadruplex motifs after prioritizing raw affinity, then saw specificity fall and revised rules for the next round.
For fibroblast activation protein alpha, a receptor overexpressed in solid tumors, the team reports top de novo designs with single-digit nanomolar affinity. Local binding-pocket excision accelerated each folding cycle more than eightfold. Human–agent collaboration nearly doubled the in-silico hit rate of a scientist working alone at equal compute, the post states. The stated next step is combining VERID with closed-loop protein discovery.
Oversight
The post says agents never send unverified code to liquid-handling robots, sequences face biosecurity screening, and humans authorize designs and supervise runs. It lists open problems including verifying multi-step protocols and ensuring agents optimize for true target binding.
How much did the aptamer campaigns raise on-target hit rates?
Across three rounds on two targets, confirmed on-target hit rates reached 33.5% for myoglobin and 42.1% for α-thrombin, according to the post, with each target requiring different wet-lab fixes.
Sources & further reading
Help keep the record accurate. Send a correction.
