Google tests AI agents in closed-loop wet-lab discovery

Connecting reasoning with physical laboratory state

Google Research has outlined three projects intended to connect AI-agent planning with real wet-lab experiments. The broader goal is not an unsupervised laboratory, but faster human-directed design-test-learn cycles in which an agent proposes experiments, receives physical measurements and revises its next step.

The first building block is VERID, a semantic digital twin of a laboratory workcell built on the open-source PyLabRobot framework. It keeps a live three-dimensional record of deck geometry, well volumes, reagent identities and pipette-tip inventory. Before code reaches a liquid-handling robot, VERID can simulate the protocol, reject collisions or impossible aspirations and preserve synchronized video and action logs.

Google reports several case studies: the system detected uneven dye mixing caused by dispensing too close to a well floor, recovered from a plate-reader failure without repeating prior pipetting, and rejected an eight-channel command that could not physically fit into a narrow trough. In a blinded dye-reconstruction task, iterative experiments improved accuracy by more than fivefold. These are controlled research demonstrations rather than evidence of a general autonomous laboratory.

Learning from failed aptamer rounds

A second project used DNA aptamers—short DNA sequences selected to bind a target—as a relatively fast closed-loop testbed. Candidate sequences were designed, chemically synthesized in pooled libraries, screened against target proteins and sequenced so that the results could guide the next round. One full cycle took seven to ten days.

Across three rounds, Google reports confirmed on-target hit rates of 33.5 percent for myoglobin and 42.1 percent for alpha-thrombin. The route was not linear. Fixed primer overhangs disrupted many alpha-thrombin candidates in the first round; modified flanking stems helped later rounds reach 42.1 percent at 98.5 percent specificity.

For myoglobin, optimizing raw affinity enriched G-quadruplex motifs that often stuck to off-target proteins and magnetic beads. Specificity fell from 75.0 to 19.1 percent and the bead-corrected hit rate dropped from 10.2 to 3.7 percent. After the agents changed their design rules, specificity recovered to 72.4 percent and the on-target hit rate rose to 33.5 percent.

Protein design and the remaining safeguards

The third project targeted fibroblast activation protein alpha, a receptor overexpressed in solid tumors. Google says a human-agent workflow used local epitope excision to make structure prediction 8.6 times faster without losing structural accuracy. At equal compute, the collaboration nearly doubled the in-silico hit rate versus a scientist working alone, and laboratory screening produced binders with single-digit nanomolar affinity.

The team stresses that these workflows remain supervised. Researchers selected non-infectious test systems without gain-of-function risk, screened every designed sequence, simulated robot actions before execution and required humans to authorize sequences, verify protocols and supervise runs. Open problems include preventing agents from exploiting experimental artifacts, validating multi-step procedures and continuously auditing both reasoning and execution. The work therefore demonstrates components for autonomous discovery, not a finished autonomous wet lab.