What happened
Google Research published a blog post detailing three research projects aimed at enabling AI agents to operate in physical wet labs. The central component is VERID, a semantic digital twin that tracks the 3D state of laboratory equipment to prevent physical errors. The post also reports on closed-loop experiments where AI agents designed DNA aptamers and protein binders, achieving specific hit rates and single-digit nanomolar affinities by learning from real-world experimental feedback.
Google Research has outlined a framework for grounding AI agents in biological reality, focusing on the automation of wet labs. The primary technical contribution is VERID (VERified Infrastructure for Discovery Laboratories), a semantic digital twin built on the open-source PyLabRobot framework. VERID tracks the live, stateful 3D geometry of robotic arms, reagent volumes, and pipette tips to ensure that AI-generated commands are physically possible and safe. The source notes that without such a twin, agents might generate syntactically valid code that is physically impossible, such as lowering a pipette into a trough that is too short.
The research demonstrates the utility of this framework through case studies where agents recovered from equipment failures and identified subtle experimental biases, such as residual tip-wall carryover, by analyzing synchronized video and execution traces. In one closed-loop campaign, an agent improved its accuracy in reconstructing dye mixtures by more than fivefold by identifying its own optical and fluidic artifacts.
Beyond infrastructure, the post details two biological discovery workflows. First, using DNA aptamers as a sandbox, AI agents completed three design-test-iteration cycles for myoglobin and α-thrombin targets. The agents learned to overcome wet-lab failure modes; for example, on myoglobin, they initially enriched G-quadruplex motifs which caused non-specific binding, but they autonomously revised their design rules in subsequent rounds to recover specificity and increase the on-target hit rate to 33.5%.
Second, the team applied a similar agent-guided workflow to de novo protein binder design for fibroblast activation protein alpha (FAP). By surgically excising local binding pockets during the computational folding step, the accelerated cycles by more than eightfold. In a head-to-head comparison, human-agent collaboration nearly doubled the in-silico hit rate compared to a scientist working alone, yielding binders with single-digit nanomolar affinity in just a few experimental runs.
Source details: research.google ↗
Why it matters
This work addresses a critical bottleneck in AI-driven science: the gap between in-silico design and physical validation. By providing a framework for safe, traceable robotic execution and demonstrating that agents can correct for wet-lab artifacts (such as non-specific binding), Google Research moves AI from a purely computational tool to an active participant in the experimental cycle. This could significantly accelerate drug discovery and biological research by automating the iterative design-test-analyze loop, though the system is currently a research prototype rather than a commercial product.
The significance of this work lies in its potential to close the loop between AI hypothesis generation and empirical validation. While large language models are increasingly capable of in-silico design, they lack reliable reasoning about 3D geometry and physical state, which is essential for operating laboratory robotics. VERID addresses this by providing a safety layer that allows agents to interact with physical hardware without causing damage or invalid experiments.
The demonstrated ability of agents to learn from wet-lab feedback is a major step toward autonomous scientific discovery. The aptamer and protein design experiments show that AI can identify and correct for experimental artifacts that are not captured by computational models alone. This suggests that AI agents can act as collaborative scientists, not just code generators, by iteratively refining their strategies based on real-world data.
This research aligns with Google’s Frontier Safety Framework and emphasizes the importance of human agency and oversight in autonomous systems. By building safeguards into both the computational and wet-lab sides of the workflow, the team aims to ensure that agents optimize for true biological signals rather than unintended experimental shortcuts. This approach could make AI-driven drug discovery more reliable and efficient, potentially reducing the time and cost of bringing new therapeutics to market.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
Watch for the open-sourcing of the VERID framework or its integration into broader laboratory automation platforms. Additionally, monitor whether these closed-loop methods are applied to more complex biological targets or if they remain limited to specific protein families like FAP and myoglobin.
The next step for this research is to integrate VERID’s traceable automation with closed-loop learning for protein discovery. This would enable agents to directly navigate complex biological search spaces in the wet lab, potentially leading to fully autonomous discovery pipelines.
Researchers and industry partners should watch for the release of the VERID framework or related tools. If made available, this could lower the barrier to entry for labs looking to implement AI-driven automation, though the current implementation is specific to Google’s infrastructure.
The application of these methods to other biological targets will be a key indicator of their generalizability. The success with FAP, a challenging receptor overexpressed in solid tumors, is promising, but further validation on diverse targets will be necessary to establish the of the approach.