Google's AI agents run real wet-lab experiments, lifting DNA binder success from 0% to 42.1%

Dr_Singularity · x · 2026-10-17

Google Research details how it plugged AI agents into real wet-lab robotics to build closed-loop molecular design cycles.

Key points:

Agents drive liquid-handling robots, run experiments autonomously under human oversight, learn from failures, and refine molecular designs.
In one test, successful DNA binder designs jumped from 0% to 42.1% over three iterations.
A separate human-AI workflow discovered proteins binding a cancer-associated target with single-digit nanomolar affinity after just a few experimental runs.
Google flags two challenges: frontier models still reason unreliably about 3D geometry and physical state, and wet-lab throughput lags agentic design speed, making experimental readouts the bottleneck.
The work spans three papers: VERID, Agentic aptamer sandbox, and Epitope excision.

More from AGI Musings