Google’s Wet-Lab Agents Raise Aptamer Hit Rates to 42.1%

Google Research says AI agents working in a closed loop with wet-lab experiments raised confirmed on-target hit rates for designed DNA aptamers to 42.1% on one protein target and 33.5% on another over three design-test rounds, in a blog post published on 16 October 2026. In a separate project the agents helped find protein binders with single-digit nanomolar affinity for fibroblast activation protein alpha (FAP), a receptor that Google describes as overexpressed in solid tumors.

The post, by Google researchers Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, describes three projects: a system that lets agents run lab robots safely, the aptamer loop and the protein binder work. It says humans authorize the designed sequences, verify protocols and supervise robot runs, and that combining the robot system with closed-loop learning is its next step.

Business Pill · PHYSICAL AI

A Business Engineer explainer on AI systems that act in the physical world. It explains the concept in general and says nothing about Google Research.

What Google Research Built

The first project is VERID, the VERified Infrastructure for Discovery Laboratories. Google describes it as “a semantic digital twin of the laboratory workcell” that tracks robotic arm geometry, well volumes, reagent identities and pipette tip inventory. It is built on the open-source PyLabRobot framework.

Google lists four capabilities: live 3D state tracking, fast virtual dry runs that simulate protocols in seconds, hardware safety checks that block invalid commands before they reach the robot, and synchronized video and action logs of every run.

In one case study, the post says, the simulator rejected an 8-channel aspiration that would not physically fit in a narrow trough, and the agent replanned. In a closed-loop self-play campaign, an agent ran robot experiments to reconstruct blinded dye mixtures from overhead camera images and improved its accuracy “more than fivefold”, according to Google.

What Was Posted on X

Google Research announced the work on X on 16 October, describing agents grounded in physical lab robotics that run closed-loop molecular design cycles and found single-digit nanomolar protein binders for cancer targets. The piece is built from the blog post this post links.

Three Rounds of Aptamer Design

The second project uses protein-binding DNA aptamers as a sandbox. Google says a complete aptamer design-test-analyze cycle takes 7–10 days, because DNA can be synthesized in pooled libraries and read out by sequencing. In each round, agents designed candidate sequences, which were screened against target proteins on magnetic beads and sequenced for the next cycle.

On α-thrombin, Google says Round 1 yielded 0% binders because fixed primer overhangs disrupted folding in 48% of sequences. After the agents introduced clamped flanking stems in Round 2 and diversified in Round 3, hit rates rose to 26.0% and then 42.1%, at 98.5% specificity.

On myoglobin, the post says that in Round 2 the agents prioritized raw binding affinity and enriched G-quadruplex motifs in 63% of designs. The wet-lab data showed most were non-specific binders: target specificity fell from 75.0% to 19.1% and the bead-corrected hit rate from 10.2% to 3.7%. In Round 3 the agents revised their design rules, and Google reports specificity recovered to 72.4% and the on-target hit rate rose ninefold, to 33.5%.

Binders for a Tumor Target

The third project applies the workflow to de novo protein binders for FAP. Google says that “In just a few experimental runs, the top designs yielded binders with single-digit nanomolar affinity, a level of tight, drug-like binding typically achieved only after extensive rounds of laboratory optimization.”

The post says excising just the local binding pocket during folding made each computational cycle more than eightfold faster, and a figure caption puts the speed-up in structure prediction at 8.6x. At equal compute, Google says a scientist setting the strategy with an agent driving execution nearly doubled the in-silico hit rate of a scientist working alone.

We have reported on Anthropic opening a wet lab, and on Claude orchestrating protein-design tools to produce validated binders.

Safety Measures Google Lists

Google says it chose well-controlled, non-infectious biological systems without gain-of-function risk as testbeds, and that every designed sequence goes through biosecurity screening. Through VERID, it says, agents never send unverified code directly to the liquid-handling robots, and every protocol is dry-run in simulation first.

The post names open problems, including “ensuring agents optimize for true target binding rather than unintended experimental shortcuts”, verifying multi-step wet-lab protocols and monitoring agent reasoning and execution traces.

The Structural Read

As we read it, the post treats experiments as the check on the agents’ designs: Google says computational models “cannot replace empirical validation”, and the myoglobin round shows a design rule that scored well on raw binding and failed on specificity.

As we read it, the robot system and the design loops are still separate pieces: Google says its next step is to combine VERID’s automation with closed-loop learning for protein discovery.

As we read it, humans keep the decisions that carry risk in this setup: Google says humans authorize designed sequences, verify protocols and supervise robot runs, and that every sequence goes through biosecurity screening.

Three Implications

DRUG-DISCOVERY TEAMS Google says the top FAP designs reached single-digit nanomolar affinity in a few experimental runs, a level it says is typically achieved only after extensive rounds of laboratory optimization.

LAB AUTOMATION BUILDERS Google says VERID is built on the open-source PyLabRobot framework and checks every protocol against a digital twin of the workcell before it reaches the liquid-handling robot.

RESEARCH LEADERS Google says a scientist setting the strategy with an agent driving execution nearly doubled the in-silico hit rate of a scientist working alone, at equal compute.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework argues that what changes between paradigms is “where intelligence accumulates and who can capture it”, and describes “The Verifier Moat — built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”

As we read it, a wet lab that returns confirmed hit rates to agents each round works as a domain-specific signal of this kind. The framework was published in July 2026 and does not discuss Google Research’s wet-lab work.

What Is Not Established

The blog post does not give the exact binding constants for the FAP binders beyond “single-digit nanomolar”, the cost of the experiments, or a timeline for therapeutic use. It describes the work as building blocks toward human-directed autonomous discovery.

The figures come from Google’s post, which links three papers; we read the post, not the papers. We did not contact Google.

The Bottom Line

Google Research says agents learning from wet-lab readouts raised confirmed aptamer hit rates to 42.1% on α-thrombin and 33.5% on myoglobin over three rounds, and helped find single-digit nanomolar binders for the tumor-associated receptor FAP in a few experimental runs.

It says humans authorize designs, verify protocols and supervise the robots, and that every designed sequence is screened for biosecurity.