Google Research details VERID digital-twin framework and closed-loop agentic aptamer design experiments for autonomous wet labs

Google Research published a blog post covering three pieces of work on autonomous wet labs. First, VERID, a digital twin built on the open-source PyLabRobot, gives AI agents live 3D state tracking, fast virtual dry runs, hardware safety checks and synchronised video and action logs for liquid-handling robots. Second, agents ran three closed design-test-iterate rounds on aptamers (short molecules that bind a target), reaching confirmed on-target hit rates of 33.5% for myoglobin and 42.1% for α-thrombin, with 98.5% specificity on thrombin. Third, a paper on de novo protein binder design reportedly produced single-digit nanomolar binders for cancer targets, but the text we have is truncated, so we cannot verify the detail. In a dye-mixture self-play test, agent accuracy rose more than fivefold.

The main audience is labs and research teams building automated or AI-driven experiments. VERID's dry runs and safety checks address a real obstacle: letting software agents drive physical robots without wasting reagents or damaging hardware. The aptamer results suggest agents can improve designs from real experimental readouts, but this is a research prototype with no sign of outside adoption. It does not yet change purchasing or hiring decisions, and any drug discovery benefit is a long way off.

All the numbers are Google's own and have not been independently replicated. The hit rates come from two well-studied targets, and there is no stated baseline, such as human-designed or conventional screening, so we cannot tell how impressive 33.5% and 42.1% are. The fivefold gain comes from a simple dye-mixing task, which says little about real biology. The framing of agents that "learn how to learn" and the drug discovery implications are aspiration, not demonstrated results.

1. 01Independent replication or third-party use of VERID, for example PyLabRobot community adoption, GitHub activity and labs reporting results on their own hardware.
2. 02Full paper details: comparison with human-designed or standard baselines for the aptamer hit rates, number of experiments per round, and the full methods and data for the nanomolar protein binders.
3. 03Whether the approach transfers to harder targets or to wet-lab validation of therapeutic candidates, such as follow-up papers, partnerships with pharma or biotech, or preclinical data in the next 6 to 12 months.