Google Research Says AI Agents Ran Closed-Loop Wet Labs and Found Cancer-Target Binders
Google Research posted building blocks for autonomous wet labs: VERID, a live digital twin that lets agents plan and safely run liquid-handling robots, plus closed-loop aptamer and de novo protein campaigns that Google says produced single-digit nanomolar binders for a cancer-related target.
By 1EyeAI Desk ·
In an Oct. 16 Google Research post, the team describes three building blocks for autonomous wet labs. The first is VERID (VERified Infrastructure for Discovery Laboratories), a semantic digital twin of a wet-lab workcell built on open-source PyLabRobot. Google says it tracks robot geometry, well volumes, reagents and tips; dry-runs protocols in seconds; blocks physically impossible commands before they hit hardware; and logs synchronized video with actions.
The second is a DNA aptamer sandbox run over three design–test–analyze cycles on myoglobin and α-thrombin. Google reports that on α-thrombin, Round 1 produced 0% binders because fixed primer overhangs disrupted folding in 48% of sequences; after the agents introduced clamped flanking stems and diversified designs, confirmed hit rates rose to 26.0% then 42.1% at 98.5% specificity. On myoglobin, Round 2 enrichment of sticky G-quadruplex motifs crashed specificity; Round 3 revisions recovered specificity to 72.4% and an on-target hit rate of 33.5%. Those figures are Google’s reported results.
The third is agent-guided epitope excision for de novo binders to fibroblast activation protein alpha (FAP), a receptor overexpressed in many solid tumors. Google says the excision step accelerated structure prediction 8.6×, that a human–agent hybrid nearly doubled the in-silico hit rate versus a scientist alone at equal compute, and that top wet-lab designs reached single-digit nanomolar affinity—a company claim, not an independent assay report.
Google stresses limits and safety: agents do not send unverified code to robots; every designed sequence gets biosecurity screening; testbeds were chosen as non-infectious with no gain-of-function risk; and humans authorize sequences, verify protocols and supervise robot runs. The post frames these as building blocks for human-directed autonomous discovery, not a finished autonomous lab.
What this means for you: If you work in drug discovery or lab automation, read the Google Research post and watch for the underlying papers. Everyday users: nothing to try in a consumer product today—this is research infrastructure and early binder chemistry, not a new chatbot feature.