Google works towards AI agents that design and run laboratory experiments

Google Research presented three projects on 16 October towards AI agents that design and run laboratory experiments, always under human supervision, the company says. In one, agents designed DNA aptamers over three rounds with real tests: according to the paper, hit rates rose from 0% to 42.1% with α-thrombin. Another, VERID, is a digital twin of the robotic lab that simulates every protocol and blocks invalid commands before they run.

The text admits a failure: with myoglobin, the agents first favoured "sticky" molecules and specificity fell from 75% to 19%. Google says it uses non-infectious systems and screens every sequence for biosecurity. All the figures are the company's own.