Google Research builds AI agents for autonomous wet-lab experiments

In one study, AI agents used VERID to run closed-loop experiments, recovering from hardware failures and refining mixing strategies based on video feedback. In another, agents designed DNA aptamers in a 7–10 day cycle, raising on-target hit rates to 33.5% for myoglobin and 42.1% for α-thrombin after learning to avoid non-specific binding artifacts. The agents autonomously revised their design rules when initial strategies failed, demonstrating a capacity for scientific reasoning and self-correction.

For protein design, the team applied this workflow to fibroblast activation protein alpha (FAP), a cancer target. By using a technique called epitope excision to reduce computational load, they accelerated the design cycle more than eightfold. Human-agent collaboration nearly doubled the in-silico hit rate compared to human-only work, yielding binders with single-digit nanomolar affinity. These efforts form a foundation for human-directed autonomous discovery in the life sciences.