Grounding AI Agents in Biological Reality: Building Blocks for Autonomous Wet Labs

Grounding AI Agents in Biological Reality: Building Blocks for Autonomous Wet Labs

The intersection of artificial intelligence and biological research is evolving from theoretical ideation to physical execution. Google Research is advancing this frontier by grounding AI agents in biological reality, aiming to navigate complex biological landscapes through physical lab robotics. The long-term vision posits that AI agents will not only ideate and analyze but also run lab experiments to test hypotheses, learn from results, and self-improve to solve complex problems, all driven by human scientific agency.

Bridging the Gap with the VERID Framework

A primary challenge in realizing autonomous wet labs is that frontier models often do not reason reliably about three-dimensional geometry and physical state. This limitation poses significant problems for automating specialized liquid-handling robots that must move reagents, track volumes, and execute molecular biology protocols. To address this, Google Research has developed the VERified Infrastructure for Discovery Laboratories (VERID) framework.

VERID acts as a semantic digital twin of the laboratory workcell. It is a live, stateful 3D volumetric entity tree that tracks everything from robotic arm geometry to individual well volumes, reagent identities, and pipette tip inventory. According to Google Research, the safety of a command depends on this live physical context as well as the code itself. An agent might generate syntactically valid code that is physically impossible, logistically invalid, or scientifically compromised by unmodeled fluidic effects.

Built on top of the open-source PyLabRobot framework, VERID provides four core capabilities to bridge the gap between high-level reasoning and physical reality:

• Live 3D state tracking: Allows agents to check deck layout, well volumes, and available tips before acting.
• Fast virtual dry runs: Simulates protocols in seconds to catch collisions or empty wells before touching reagents.
• Hardware safety checks: Blocks invalid commands before they reach the robot.
• Synchronized video and action logs: Captures a complete record of every run for research rigor, reproducibility, and debugging.

Google Research demonstrates these capabilities through specific case studies. In one instance, synchronized video and execution traces identified uneven dye stratification caused by dispensing too close to the well floor, allowing the agent to refine its mixing strategy. Furthermore, the persistent digital twin state enabled an agent to recover mid-run from a plate-reader failure without re-pipetting. In another scenario, when a simulator rejected an 8-channel aspiration that physically would not fit in a narrow trough, the agent replanned on the fly. This led to a targeted factorial diagnostic that isolated a subtle bias caused by residual tip-wall carryover, a failure mode that replication alone would have missed.

In a closed-loop self-play campaign, an agent iteratively designed and ran robot experiments to reconstruct blinded dye mixtures from overhead camera images. By identifying its own optical and fluidic artifacts across runs, the agent improved accuracy more than fivefold. Google Research notes that reproducible execution and full trace logs—capturing procedural variables like dwell times, evaporation, and plate layout—lay the groundwork for agents to eventually separate biological signals from experimental noise.

Aptamers as a Sandbox for Lab-in-the-Loop Discovery

While VERID handles the physical execution, agents require real experimental readouts to refine their hypotheses, as computational models cannot capture the full picture of wet-lab reactions. To prototype this closed-loop system, Google Research utilized protein-binding DNA aptamers.

In the project titled "Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery," agents iteratively design, synthesize, select, and sequence aptamers over three design-test-iteration cycles. A complete aptamer design-test-analyze cycle takes just 7–10 days, as DNA can be chemically synthesized in pooled libraries and read out directly by sequencing.

Google Research reports significant improvements across three rounds on two protein targets with varying difficulty levels: myoglobin and α-thrombin.

• α-Thrombin: In Round 1, the yield was 0% binders because fixed primer overhangs disrupted folding in 48% of sequences. Once agents introduced clamped flanking stems in Round 2 and diversified in Round 3, hit rates climbed to 26.0% and then 42.1% at 98.5% specificity.
• Myoglobin: In Round 2, agents prioritized raw binding affinity, heavily enriching G-quadruplex motifs (63% of designs). While raw pull-down increased, wet-lab data revealed that most of these were non-specific "sticky" binders adhering to off-target proteins and bare magnetic beads. This caused target specificity to drop from 75.0% to 19.1% and the true, bead-corrected hit rate to plunge from 10.2% to 3.7%. Recognizing this artifact, agents autonomously revised their design rules for Round 3, capping promiscuous motifs, subtracting bead binders, and pivoting to non-G4 structural scaffolds. This recovery resulted in 72.4% specificity and a ninefold surge in on-target hit rate to 33.5%.

Google Research highlights that across three rounds, agents learn from wet-lab readouts to recover target specificity and improve confirmed de novo hit rates. The study also examines how agents reason like scientists, analyzing data and suggesting tweaks to wet-lab protocols, while sometimes exploring sequence spaces that structural biologists might not.

Accelerating De Novo Protein Binder Design

Closing the loop for proteins is more complex than for aptamers. Unlike chemically synthesized aptamers, proteins must be expressed biologically, adding cellular or cell-free selection, expression, and folding steps. This creates bottlenecks on both computational and wet-lab sides.

Google Research applied its workflow to fibroblast activation protein alpha (FAP), a large, challenging receptor overexpressed in solid tumors. In just a few experimental runs, the top designs yielded binders with single-digit nanomolar affinity. Google Research states that this level of tight, drug-like binding is typically achieved only after extensive rounds of laboratory optimization.

The computational pipeline accelerates structure prediction 8.6x by surgically excising just the local binding pocket during folding. This approach enables a human–agent hybrid pipeline to nearly double the in-silico hit rate compared to a scientist working alone. To scale wet-lab screening, the team operationalized yeast surface display with fluorescence-activated cell sorting (FACS), allowing them to screen large libraries of engineered protein variants in parallel by fluorescence.

Safety and Future Directions

In line with Google's Frontier Safety Framework and their approach to bioresilience, safeguards are built across both computational and wet-lab workflows. Google Research deliberately selected well-controlled, non-infectious biological systems without gain-of-function risk as testbeds to evaluate agent behavior and catch drift early. Every designed sequence goes through biosecurity screening.

The immediate next step involves weaving these threads together, combining VERID's fully traceable automation with closed-loop learning for protein discovery. This integration aims to enable agents to directly navigate complex, rugged biological search spaces in the wet lab. Through simulation before execution, agents never send unverified code directly to liquid handling robots; every protocol is dry-run in simulation and validated against the digital twin before physical action.

By grounding AI agents in biological reality, Google Research is establishing the building blocks for autonomous wet labs, moving toward a future where AI can intelligently navigate vast combinatorial search spaces and accelerate drug discovery.

Source:

• Grounding AI agents in biological reality: Building blocks for autonomous wet labs — Google Research — https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs