Google Research reports that AI agents improved the share of newly designed DNA strands that bound their intended proteins across repeated laboratory tests. In one study, the agents used failed results to change later designs; in a separate project, researchers built software to check liquid-handling robot commands against the equipment’s live physical state.
The October 16 overview brings together three separate projects: agent-guided aptamer design, VERID robot-state checks, and a human-and-agent protein-binder study.
Lab results changed the next DNA design round
Aptamers are short DNA strands that fold into shapes capable of binding a particular protein. In a Google Research study, agents designed aptamer candidates for two test proteins, myoglobin and alpha-thrombin. Researchers tested each batch in the lab, sequenced the selected DNA, and used those results to prepare the next batch. The reported design-test cycles took about a week each.
The study measured the share of new designs that bound the intended protein after correcting for sequences that stuck to bare beads or other proteins. For myoglobin, that hit rate fell from 10.2% in the first round to 3.7% in the second, then rose to 33.5% (184 of 549 designs) in the third. For alpha-thrombin, it increased from 0% to 26.0% and then 42.1% (131 of 311). Excluding candidates derived from known literature motifs, the alpha-thrombin rate was 37.1% (106 of 286).
The second myoglobin round exposed a misleading pattern. Many candidates used guanine-rich DNA folds called G-quadruplexes. They appeared to bind, but many also stuck to unrelated proteins and bare beads. The paper’s counter-screen showed on-target specificity falling from 75.0% to 19.1% as cross-reactivity rose. Specificity measures the share of statistically significant binders in that counter-screened set that reacted only with the intended protein; the hit rate instead measures new designs that passed the bead-corrected binding test. In the next round, the agents reduced those motifs and explored other DNA folds. Specificity recovered to 72.4%, and the new-design hit rate reached 33.5%.
A separate robot system checked the physical lab state
A second project, called VERID, gives an AI agent a persistent digital model of a liquid-handling workcell. It tracks the robot’s layout, reagent identities, well volumes, and available pipette tips. Before a command reaches the robot, VERID simulates it against that state and blocks actions that would collide with equipment or draw from an empty well. Synchronized video and execution logs help researchers trace what happened during a run.
The VERID report describes robot experiments with colored dyes and water, chosen to test liquid-handling mechanics without biological reagents. In one test, the agent adjusted its mixing protocol after measurements showed that dye had stratified near a well’s bottom. In another, the saved digital state let the system resume after a plate-reader fault without repeating completed pipetting. Those tests demonstrate a safety and record-keeping layer for lab automation; they do not establish that the aptamer experiments ran through VERID.
Four FAP binder clones showed lab binding; Google says combining VERID with protein design is a next step
A third study used a human-and-agent workflow to design protein binders for fibroblast activation protein alpha, or FAP, a target found in some solid tumors. After yeast-display screening, four selected clones showed apparent binding affinities between 3.0 and 9.1 nanomolar. The measurement indicates binding in a laboratory assay. The report contains no clinical or treatment result.
Google says combining VERID with closed-loop protein design is a next step. The overview also says researchers review designs and protocols and limits its demonstrations to controlled biological systems.
Sources
- Google Research overview, October 16, 2026
- Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery
- Semantic Representation for Agentic Self-Play in Automated Laboratories
- Agent-guided epitope excision enables economical de novo binder design
Representative image: NCATS liquid-handling robot photograph by the National Center for Advancing Translational Sciences, via Wikimedia Commons; public domain. The photograph is illustrative and does not show Google Research’s equipment.