Google Research Grounds AI Agents in Wet Labs for Autonomous Molecular Discovery

Bridging the Gap Between AI and Wet Labs

The core challenge lies in reconciling the abstract reasoning of advanced AI models with the tangible, often error-prone, realities of a physical laboratory. Frontier AI systems, while capable of navigating vast biological search spaces, still struggle with precise 3D geometry and physical state reasoning, which is crucial for operating specialized liquid-handling robots. Simultaneously, the pace of real-world wet-lab experimentation has lagged behind rapid advancements in AI-driven hypothesis generation and in-silico design, creating a bottleneck for validating AI proposals and achieving self-improvement.

The VERID Framework for Physical Grounding

To address these challenges, Google Research has developed the VERIFIED Infrastructure for Discovery Laboratories (VERID) framework. VERID acts as a semantic digital twin of the laboratory workcell, providing a live, stateful 3D volumetric representation that tracks all critical elements, from robotic arm geometry to reagent volumes and pipette tip inventory. This framework ensures the safety and feasibility of AI-generated commands by incorporating live physical context. VERID offers four key capabilities: live 3D state tracking for agents to assess the lab environment, fast virtual dry runs to simulate protocols and catch errors, hardware safety checks to block invalid commands, and synchronized video and action logs for research rigor and debugging. Case studies demonstrate VERID's effectiveness in refining mixing strategies, recovering from mid-run failures, and enabling agents to autonomously replan experiments to isolate subtle experimental biases.

Learning to Learn with Aptamer Discovery

Real-world experimental feedback is essential for AI agents to refine their hypotheses, as computational models cannot fully capture the nuances of wet-lab reactions. In a prototype closed-loop system using protein-binding DNA aptamers, AI agents went through three design-test-iteration cycles. Aptamer discovery, a common technique for clinical biomarker research, allows for rapid synthesis and readout, making it an ideal sandbox. Across these cycles, AI-driven design improved on-target hit rates significantly for targets like myoglobin and alpha-thrombin. Notably, agents learned to overcome wet-lab failure modes, such as identifying and mitigating non-specific binding artifacts caused by G-quadruplex motifs, demonstrating an ability to autonomously revise design rules based on experimental outcomes.

Accelerating De Novo Protein Binder Design

Closing the loop for de novo protein binder design is more complex due to the biological expression and folding steps required. The research highlights a workflow applied to fibroblast activation protein alpha (FAP), a challenging cancer target. By employing a technique called surgical epitope excision, the computational pipeline accelerated structure prediction by over eightfold, significantly speeding up the design-test cycle. This human-agent collaboration nearly doubled the in-silico hit rate compared to a scientist working alone, leading to the discovery of binders with single-digit nanomolar affinity in just a few experimental runs. The team also operationalized yeast surface display with fluorescence-activated cell sorting (FACS) to scale wet-lab screening.

Safety and Future Directions

Safety is paramount, with safeguards built into both computational and wet-lab aspects of the workflows. This includes using well-controlled biological systems, rigorous simulation before execution via VERID, and continuous human oversight. Future work will focus on weaving VERID's traceable automation with closed-loop learning for protein discovery, enabling AI agents to navigate complex biological search spaces autonomously. These efforts represent crucial building blocks for human-directed autonomous discovery in the life sciences.