Google Research links AI agents to lab robots and finds nanomolar cancer binders

Google Research has described how it is putting AI agents into closed loops with real laboratory experiments. In its announcement, the team reports single-digit nanomolar protein binders for a cancer-linked target, found in a few experimental runs.

What Google announced

The post, written by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, presents three research projects. Together they aim to close the gap between AI exploration on a computer and experiments in a wet lab. Google calls them building blocks for human-directed autonomous discovery in the life sciences.

The first is VERID, short for VERified Infrastructure for Discovery Laboratories. It is a semantic digital twin of a lab workcell: a live 3D model that tracks robotic arm geometry, individual well volumes, reagent identities and pipette tip inventory. It is built on the open-source PyLabRobot framework. According to Google, it lets agents check the deck layout and available tips before acting, run virtual dry runs in seconds, and have invalid commands blocked before they reach the robot. Synchronized video and action logs record every run.

Google says the motivation is that frontier models still do not reason reliably about 3D geometry and physical state. An agent can write code that is syntactically valid but physically impossible, such as lowering an 8-channel pipette into a trough that is too short. In one case study, the simulator rejected exactly that aspiration, and the agent replanned and designed a factorial diagnostic. The diagnostic isolated a bias caused by residual tip-wall carryover. In a self-play campaign, an agent reconstructed blinded dye mixtures from overhead camera images and improved its accuracy more than fivefold.

Aptamers as a testing ground

The second project uses protein-binding DNA aptamers as a sandbox. A full design-test-analyze cycle takes 7 to 10 days, because DNA can be synthesized in pooled libraries and read out by sequencing. Agents ran three rounds on two targets, myoglobin and α-thrombin.

According to Google, confirmed on-target hit rates rose to 33.5% for myoglobin and 42.1% for α-thrombin. Each target exposed a different failure mode. On α-thrombin, round 1 produced no binders because fixed primer overhangs disrupted folding in 48% of sequences. Once the agents added clamped flanking stems, the share of confirmed binders rose to 26.0% in round 2 and 42.1% in round 3, where specificity reached 98.5%.

On myoglobin, agents that prioritized raw affinity in round 2 enriched G-quadruplex motifs, which made up 63% of designs. Wet-lab data showed many of these were non-specific binders that stuck to off-target proteins and bare beads. Specificity fell from 75.0% to 19.1%, and the bead-corrected hit rate fell from 10.2% to 3.7%. The agents then revised their design rules on their own for round 3. They capped promiscuous motifs, subtracted bead binders and moved to other scaffolds. Specificity recovered to 72.4% and the hit rate rose to 33.5%.

Protein binders for FAP

The third project targets fibroblast activation protein alpha (FAP), a large receptor overexpressed in solid tumors. Proteins must be expressed biologically, which adds steps to every loop. Google also says the final re-folding step in computational binder design consumes most of the GPU compute, because it scales quadratically with target length.

The team’s fix is to excise only the local binding pocket during folding. Google says this sped up each cycle 8.6 times without losing structural accuracy. At equal compute, a scientist setting strategy while an agent handled tactics nearly doubled the in-silico hit rate of a scientist working alone, and found hits across a wider range of target sites. In the wet lab, the team used yeast surface display with fluorescence-activated cell sorting. The top designs reached single-digit nanomolar affinity, which Google says usually takes extensive laboratory optimization.

Why it matters

The post continues a line of Google work that has moved from generating ideas toward touching equipment. The Co-Scientist system, introduced in February 2025, started as a hypothesis generator. According to The Decoder’s report from August 2026, a newer version now writes machine-readable lab protocols and has been paired with a semi-automated furnace. In that work, three semiconductor thin films were synthesized on the first try, but humans still loaded samples and precursors, and the atomic structure of a separate 2D material was still unconfirmed. The VERID digital twin, which blocks invalid commands before they reach a robot, addresses the manual steps and the physical-state errors that earlier work left open.

The same report is a caution on how far to trust agent output. It describes leftover errors, including one case where the system wrote methods that, in lead author Samuel Schmidgall’s words, did not match its actual code. That makes the myoglobin episode here, where agents caught a misleading metric against wet-lab data, a useful counterexample, though a single episode on non-infectious test targets.

Google is also no longer alone in building this. BioPharmaTrend reported that Lila Sciences, which builds autonomous labs, raised $235 million at a valuation above $1 billion. Confirmation would come from VERID being joined to the protein pipeline, as the post says is planned, and from results outside Google’s own papers.

Safety measures

Google says it follows its Frontier Safety Framework and bioresilience approach. It chose non-infectious systems with no gain-of-function risk, and every designed sequence goes through biosecurity screening. Protocols are dry-run in simulation and validated against the digital twin before touching reagents, and humans authorize designed sequences, verify protocols and supervise robot runs. Open problems include making sure agents optimize for true binding rather than experimental shortcuts, verifying multi-step protocols and monitoring agent reasoning traces.

The AI-agent push is not limited to labs. We recently covered an Anthropic report on how much longer AI agents now work than its staff. Google’s post shows the same shift toward longer, more autonomous work, this time with physical equipment attached.