Google Lets AI Agents Run Their Own Wet-Lab Experiments: Rehearsing in a Digital Twin of the Bench, Aptamer Hit Rates Rose From 0 to 42.1% in Three Rounds

The bottleneck is the pipette, not the idea

On October 16, Google Research scientists Pooja Rao, Petar Sirkovic and Annalisa Pawlosky described three projects that put AI agents into a design–experiment–revise loop in the wet lab, each with a paper. The post starts with the hard part: frontier models still do not reason reliably about 3D geometry and physical state. An agent can write syntactically valid code that tries to lower an eight-channel pipette into a trough that is too short, or aspirate from a well that is already empty.

Google's answer is VERID (VERified Infrastructure for Discovery Laboratories). Built on the open-source PyLabRobot, it keeps a live, stateful 3D digital twin of the workcell that records arm geometry, each well's volume, reagent identities and tip inventory. Agents' plans get a virtual dry run in seconds, invalid commands never reach the robot, and every run has synchronized multi-camera video and action logs. In the paper's case studies, the video caught uneven dye layering caused by dispensing too close to the well floor, and the agent changed how it mixed; when a plate reader failed mid-run, the agent carried on from the twin's state without re-pipetting.

A failed round along the way

The second project has agents design protein-binding DNA aptamers, with each design, synthesis, selection and sequencing cycle taking about 7 to 10 days. On α-thrombin, Round 1 produced no binders because fixed primer overhangs disrupted folding in 48% of sequences; the next two rounds reached 26.0% and 42.1%, at 98.5% specificity. On myoglobin, the agents leaned heavily on G-quadruplexes in Round 2 (63% of designs), and most turned out to be non-specific binders that even stuck to bare beads: specificity fell from 75.0% to 19.1%, and the bead-corrected hit rate dropped from 10.2% to 3.7%. In Round 3 the agents rewrote their own rules to cap those structures, recovering specificity to 72.4% and lifting the on-target hit rate to 33.5%, about nine times the previous round.

The third project targets FAP, a receptor overexpressed in solid tumors. Folding only the region around the binding pocket made each computational cycle 8.6 times faster. At equal compute, a scientist setting strategy with an agent executing nearly doubled the in-silico hit rate of the scientist working alone, and the best designs reached single-digit nanomolar affinity in the lab.

Google says it chose non-infectious test systems, screens every designed sequence for biosecurity, and requires human authorization for sequences, protocols and robot runs.