Google Research's VERID blocks impossible robot commands as its lab agents reach 33.5% and 42.1% aptamer hit rates

On Oct 16 Google Research described three projects that put AI agents in the wet lab, including a checker that blocks impossible robot commands and an aptamer campaign whose confirmed hit rates reached 33.5% and 42.1% after three rounds. Annalisa Pawlosky leads the research.

The gains came from wet-lab feedback that a model alone could not supply. On myoglobin, the agents' round-2 designs looked better on raw binding, but specificity fell from 75.0% to 19.1% and the true hit rate from 10.2% to 3.7%; only after reading that readout did they revise their rules and recover to 72.4% specificity and a 33.5% hit rate. VERID covers the robotics side: it checks every agent-written protocol against a live digital twin, so physically impossible or logistically invalid commands never reach the liquid-handling robot.

An agent can write valid code the robot cannot run

Google says an agent can write syntactically valid code that is physically impossible, such as lowering an 8-channel pipette into a narrow trough, or logistically invalid, such as aspirating from a depleted well. VERID, VERified Infrastructure for Discovery Laboratories, is a semantic digital twin of the lab workcell built on the open-source PyLabRobot framework. It tracks robotic arm geometry, well volumes, reagent identities and tip inventory as a live 3D entity tree, and its dry runs simulate protocols in seconds to catch collisions or empty wells before reagents are touched. In one case the simulator rejected an 8-channel aspiration that would not fit a narrow trough, and the agent replanned into a factorial diagnostic that isolated a bias from residual tip-wall carryover.

The round-2 drop taught the agents their rules

Each target hit a different wet-lab failure. Each design-test-analyze cycle took 7 to 10 days, so the three rounds added up to weeks of wet-lab time. Round 1 on alpha-thrombin produced 0% binders because fixed primer overhangs disrupted folding in 48% of sequences; clamped flanking stems in round 2 and diversification in round 3 lifted hit rates to 26.0% and then 42.1%, at 98.5% specificity. Myoglobin failed the other way. Prioritizing raw binding and enriching G-quadruplex motifs to 63% of designs raised pull-down, but most of those motifs were non-specific sticky binders, and the round-2 drop described above followed. The agents capped promiscuous motifs, subtracted bead binders and pivoted to non-G4 scaffolds, reaching 72.4% specificity and a ninefold rise in on-target hit rate, to 33.5%.

On fibroblast activation protein alpha, a receptor overexpressed in solid tumors, Google reports single-digit nanomolar binders "in just a few experimental runs", a level it says usually takes extensive lab optimization. Excising the rest of the target and keeping only the local binding pocket during structure prediction sped that step 8.6x without losing structural accuracy, and at equal compute a scientist-agent setup nearly doubled the in-silico hit rate of a scientist working alone.

Humans still authorize designed sequences, verify protocols and supervise robot runs, and Google lists agents optimizing for unintended experimental shortcuts as an open problem. The next step it names is to combine VERID's traceable automation with closed-loop learning for protein discovery.