Google says AI agents improved DNA binders by learning from lab setbacks

Google Research says AI agents improved protein-binding DNA designs after physical experiments exposed flaws in their choices. Its October 16 research disclosure pairs those results with robot safety software and protein-design experiments. The work demonstrates human-supervised steps toward autonomous laboratories, with agents revising designs based on what actually happened at the bench.

When stronger binding was the wrong signal

The clearest example involves aptamers: DNA sequences designed to bind proteins. Agents completed three design-and-test rounds against two targets, myoglobin and α-thrombin. Each cycle takes seven to 10 days. Candidate DNA sequences are synthesized, screened against proteins attached to magnetic beads, then sequenced to supply data for the next round.

For myoglobin, the second round went backward. Agents favored structures called G-quadruplexes, which made up 63% of their designs. Raw binding increased, but many candidates stuck to unrelated proteins and the magnetic beads used in screening. Specificity—the measure of how selectively they bound the intended target—fell from 75.0% to 19.1%. The bead-corrected hit rate dropped from 10.2% to 3.7%.

Using that experimental feedback, the agents changed their design rules. They limited the troublesome structures, accounted for bead binding and shifted toward other shapes. Google says the third round recovered specificity to 72.4% and raised the on-target hit rate to 33.5%. Optimizing raw binding alone had rewarded the wrong candidates.

The α-thrombin experiments exposed a different failure. The first round produced no binders; fixed DNA segments used in the workflow disrupted folding in 48% of sequences. Agents added stabilizing flanking structures and later diversified their designs. Hit rates rose to 26.0% in round two and 42.1% in round three, with final specificity of 98.5%.

Checking the robot’s world before it moves

The infrastructure project, VERID, addresses another gap: valid code can still describe an impossible laboratory action. Built on the open-source PyLabRobot framework, it maintains a digital twin—a live software representation of the robot’s workspace. It tracks robot geometry, liquid volumes, reagent identities and available pipette tips. Simulated dry runs and hardware checks block invalid commands before execution.

In one case, its simulator rejected an eight-channel pipette operation because the tool would not fit inside a narrow trough. Synchronized video and action records also exposed uneven dye mixing caused by dispensing too close to a well’s floor. In that campaign, persistent state tracking let the agent recover from a plate-reader failure without repeating the pipetting.

Faster protein design, with humans setting the strategy

Protein discovery adds biological expression and folding steps that aptamer synthesis does not require. Google’s protein-binder experiments targeted fibroblast activation protein alpha, or FAP, a receptor overexpressed in solid tumors. The top experimental designs achieved single-digit nanomolar affinity, a measure of tight binding, after a few experimental runs.

The computational shortcut was to model only the local binding pocket during the expensive structure-prediction step, rather than the full target. Google says this accelerated that step more than eightfold without losing structural accuracy. At equal computing budgets, a scientist-agent team nearly doubled the predicted hit rate of a scientist working alone. The scientist set strategy; the agent handled tactical execution.

These remain building blocks, not an integrated autonomous laboratory. Google identifies combining VERID with closed-loop protein discovery as its next step. The researchers say every designed sequence undergoes biosecurity screening. Humans authorize and review sequences, verify protocols and supervise robot runs; protocols must pass simulation and digital-twin validation before execution.