Grounding AI agents in biological reality: Building blocks for autonomous wet labs

Google Research published a research overview describing three projects that connect AI agents to physical laboratory equipment, with the stated goal of running closed-loop molecular design cycles. The authors are Pooja Rao, Petar Sirkovic and Annalisa Pawlosky of Google. The work spans a robotics infrastructure framework called VERID, a closed-loop aptamer discovery prototype, and a de novo protein binder design workflow targeting fibroblast activation protein alpha (FAP).
The motivation the authors give is the scale of biological search spaces: a single 100-amino-acid protein spans at least 20^100 possible sequences, which they say calls for dynamic multi-candidate exploration. Their long-term vision is that agents will not only propose and analyze designs but also run experiments, learn from readouts and improve, all under human scientific direction. They identify two blockers: frontier models still do not reason reliably about 3D geometry and physical state, which complicates automating liquid-handling robots that move reagents and track volumes; and real wet-lab experimentation has not kept pace with in-silico design, making experimental readouts the bottleneck.
VERID, short for VERified Infrastructure for Discovery Laboratories, is described as a semantic digital twin of a laboratory workcell. It is a live, stateful 3D volumetric entity tree tracking robotic arm geometry, individual well volumes, reagent identities and pipette tip inventory. The authors argue that the safety of a command depends on live physical context as well as code, since an agent can write syntactically valid code that is physically impossible, such as lowering an 8-channel pipette into a trough that is too short, logistically invalid, such as aspirating from a depleted well, or scientifically compromised by unmodeled fluidic effects.
Built on the open-source PyLabRobot framework, VERID offers four capabilities: live 3D state tracking so agents can check deck layout, well volumes and available tips before acting; fast virtual dry runs that simulate protocols in seconds to catch collisions or empty wells; hardware safety checks that block invalid commands before they reach the robot; and synchronized video plus action logs for reproducibility and debugging.
The report includes case studies. Synchronized video and execution traces caught uneven dye stratification caused by dispensing too close to the well floor, and the agent refined its mixing strategy. In the same campaign, the persistent digital twin state let the agent recover mid-run from a plate-reader failure without re-pipetting. When the simulator rejected an 8-channel aspiration that would not fit in a narrow trough, the agent replanned and designed a factorial diagnostic isolating a bias from residual tip-wall carryover. In a closed-loop self-play campaign, an agent iteratively designed and ran robot experiments to reconstruct blinded dye mixtures from overhead camera images, identifying its own optical and fluidic artifacts and improving accuracy more than fivefold.
The aptamer work appears in a paper titled Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery. Protein-binding DNA aptamers are used to prototype the lab in a closed loop over three design-test-iteration cycles. A complete design-test-analyze cycle takes 7 to 10 days, because DNA can be chemically synthesized in pooled libraries and read out directly by sequencing. Each round, agents design candidate sequences that are screened against target proteins on magnetic beads to isolate functional binders, then sequenced to return quantitative data for the next cycle. Across three rounds on two targets of differing difficulty, myoglobin and alpha-thrombin, AI-driven rational and de novo design raised confirmed on-target hit rates to 33.5% and 42.1% respectively, though each target required overcoming a different wet-lab failure mode.
On alpha-thrombin, round 1 produced 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 reached 26.0% and then 42.1% at 98.5% specificity. On myoglobin, prioritizing raw binding affinity in round 2 enriched G-quadruplex motifs to 63% of designs. Raw pull-down rose, but the authors report that most of these were non-specific sticky binders adhering to off-target proteins and bare magnetic beads, dropping target specificity from 75.0% to 19.1% and the bead-corrected hit rate from 10.2% to 3.7%. The agents recognized the artifact from the round 2 readout and revised design rules for round 3, capping promiscuous motifs, subtracting bead binders and pivoting to non-G4 scaffolds, recovering specificity to 72.4% and lifting on-target hit rate ninefold to 33.5%. The authors also note that agents sometimes reason like scientists and sometimes explore sequence spaces structural biologists might not.
Protein binders close the loop differently, since proteins must be expressed biologically rather than chemically synthesized, adding selection, expression and folding steps to each cycle, alongside a larger sequence space. In Agent-guided epitope excision enables economical de novo binder design, the team worked on FAP, described as a large, challenging receptor overexpressed in solid tumors. In a few experimental runs, top designs yielded binders with single-digit nanomolar affinity, which the authors characterize as tight, drug-like binding typically reached only after extensive optimization. Computationally, de novo binder design involves generating candidate backbones, designing sequences to fit them, and re-folding the resulting complexes to predict binding. That folding step consumes most of the GPU compute because it scales quadratically with target length. Excising just the local binding pocket during folding accelerated each cycle more than eightfold without losing structural accuracy, the authors report, turning decisions about where to bind and how much surrounding structure to include into tactical levers for LLM agents. At equal compute, a human-agent setup, with a scientist setting strategy and an agent executing tactics, nearly doubled the in-silico hit rate of a scientist working alone while finding hits across a broader range of target sites. For wet-lab screening, the team operationalized yeast surface display with fluorescence-activated cell sorting (FACS) to screen large libraries of protein variants in parallel by fluorescence.
The team describes safeguards aligned with Google's Frontier Safety Framework: deliberately selecting well-controlled, non-infectious biological systems without gain-of-function risk; screening every designed sequence for biosecurity; requiring all protocols to be dry-run in simulation and validated against the digital twin before contacting reagents; and keeping humans authorizing sequences, verifying protocols and supervising robot runs. They list open problems as agents take on more of the cycle: ensuring agents optimize for true target binding rather than experimental shortcuts, verifying multi-step wet-lab protocols, and continuously monitoring agent reasoning and execution traces.
Why it matters: this is a concrete signal that agentic AI in the life sciences is moving from in-silico design toward instrumented physical execution, which matters to teams building lab automation, molecular design pipelines or agent infrastructure. The most transferable piece for developers is probably VERID's pattern of a live state model plus simulation gating before hardware commands, since the same problem of validating agent actions against changing physical state recurs in other robotics settings. The biological results are early and target-specific; the reported hit-rate swings, especially the G-quadruplex specificity collapse on myoglobin, suggest agent loops still need careful readout design and artifact correction to avoid optimizing the wrong signal, which is an inference rather than a claim by the authors.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
General Science