Google Research is connecting AI agents to laboratory robots so molecular designs can be tested, measured and revised in a closed loop rather than judged only in software. The group says prototype systems discovered single-digit nanomolar protein binders for cancer targets, a measure indicating strong binding at very low concentrations.
The work tackles a practical weakness in scientific AI: models can propose experiments, but they do not reliably understand three-dimensional geometry, liquid volumes or the changing physical state of a wet lab. Google’s VERID framework places a verified control layer between an agent’s high-level instructions and specialized equipment. It tracks resources and measurements so experiments can be run more safely and efficiently.
Two additional projects use repeated physical feedback for aptamer and newly designed protein research. Aptamers are short molecules selected because they bind to a target. Instead of treating an initial model proposal as the answer, the systems test candidates and use the results to guide the next design round.
Google presents these projects as building blocks, not a fully autonomous laboratory. Human scientists remain responsible for goals and oversight, while slow and costly real-world experiments still limit how quickly an agent can learn. The immediate advance is a tighter, verifiable link between computational proposals and physical evidence.