Google Advances Its Autonomous Wet Lab, Achieving Closed-Loop Integration of Agents and Robots

Google Is Also Advancing an Autonomous Wet Lab, Achieving a Closed Loop Between Agents and Robots

On October 16, Google published a technical blog detailing progress on building an autonomous wet lab. The team aims to have AI agents design molecules, schedule experiments, analyze results, and then adjust the next round of plans based on feedback. In one study, they have already identified protein molecules that tightly bind to solid tumor–related targets.

The work is led by scientist Annelise Polosky and includes three research areas: robotic operation, molecular screening, and protein design, which have not yet been fully integrated.

Compared with predictive models like AlphaFold, this time Google has focused more on experimentation: A molecule that looks promising in a computer may not be so on the bench; figuring out how to quickly identify problems and then decide the next steps is becoming a missing piece that AI research needs to fill.

Google’s goal goes beyond having agents perform experiments. According to the team’s vision, the model proposes candidates from the vast molecular space, the agent carries out tests using experimental equipment, the experimental readouts flow back to the model, forming a closed-loop “self‑improving search,” where the agent gradually “learns how to learn.”

Let experimental results feed back to improve the model itself, Google bringsthe recursive self‑improvement (RSI) loop into a wet lab, using real biological feedback to drive it.

(Source: Google)

Build a “Digital Twin” for the Lab

A wet lab is where reagents and biological samples are handled and biochemical experiments are conducted. To have an intelligent agent work here, it must first understand what’s happening on the workbench. How much reagent remains in the well prepared for aspiration, whether the plate has a lid—these all affect the next steps. Even if the code has no syntax errors, it may still encounter problems during actual execution.

Google built VERID for this purpose. It is based on the open‑source robot control framework PyLabRobot and maintains a continuously updated “digital twin” of the workstation, recording the status of equipment, reagents, and samples. The agent first checks the protocol in a simulator, then lets the robot execute it, while video and action logs preserve the execution process for researchers to troubleshoot anomalies.

Laboratories already have automation equipment, but it usually requires people to write and debug programs in advance. After the agent gets involved, the team hopes to hand over the work of adjusting protocols based on results to AI as well. This requires the system, when changing experimental steps, to still verify the status of equipment and samples and to check whether the new protocol can be executed.

Experimental results can also overturn the agent’s own judgment. In a myoglobin screening, the AI initially thought it had found better molecules, but after adding controls it turned out that many molecules were just sticking everywhere without accurately identifying the target. It then revised its design rules, and the next round showed improvement. The control data helped it uncover the issues in its original judgment.

Currently, wet experiments in this type of molecular research are still performed by researchers, while the agent handles design and analysis. Validation of robot automation mainly uses dyes and water. Another protein design study has identified molecules that tightly bind tumor‑related targets, but the full workflow has not yet been integrated into VERID.

What Google showcased this time are several capabilities needed for an independent lab; to deploy them in the real world, they still need to be integrated.

One direction, multiple pathways

Before Google published its blog, DeepMind had already announced two related studies in October, focusing on new molecule design and the search for potential drug leads. On October 5, DeepMind collaborated with the research teams of Frances Arnold and Peng Liu to release the enzyme design pipeline AlphaProtein Novo.

Enzymes can catalyze chemical reactions, and the team aims to design new enzymes from scratch for tasks such as drug synthesis. This pipeline was first validated in DeepMind’s own wet lab, then collaborated with researchers to design and experimentally optimize for specific reactions, and finally used the experimentally obtained clues to improve subsequent designs.

At the same time, the nonprofit research institute Biohub announcedexpanding the “Virtual Biology Program”, with the U.S. Department of Energy and the National Institutes of Health joining. DeepMind, Alphabet’s drug‑discovery company Isomorphic Labs, and Meta together pledged $300 million for biological data and measurement technologies. This investment focuses on foundational data: to enable models to predict cellular and organismal behavior, we first need to measure how they perform under various conditions.

In mid‑October, DeepMind once again showcased the workflow for discovering antimicrobial peptides on the Antigravity platform. The agent starts from protein sequences in the database with unknown functions, combines functional prediction, structural analysis, and literature mining to select candidates, and then generates experimental protocols. The main candidates are currently undergoing wet‑lab validation.

Each line of work has its own focus: enzyme design tries to create new molecules, antimicrobial peptide searches look for clues in existing sequences, and data programs supplement models with experimental data. Yet the underlying logic is very similar: after computers generate an increasing number of possibilities, experiments must keep pace to determine which ones are worth further investment.For Google, predictive models, research tools, experimental equipment, and data accumulation are beginning to be organized around the same research workflow.

(Source: Google)

 

Why now?

Google noted in a blog post that the speed of hypothesis generation and molecular design by agents is accelerating rapidly, while experiments are lagging behind. In the past, linking these steps relied mainly on researchers. They searched literature, selected tools, wrote control programs, processed readouts, and scheduled the next round of experiments.

Nowadays, AI can take on part of this work, and open-source frameworks have also lowered the difficulty of connecting devices. As a result, Google has begun to gradually try to let agents continuously enter the entire process, rather than staying only at prediction or analysis.

Other tech giants are also betting on this field. In April this year, OpenAI launched GPT‑Rosalind, covering tasks such as experimental design and data analysis. In September, Anthropic announced early results from its in‑house biology lab, where Claude searches for novel enzyme systems, which scientists then validate experimentally.

Both aim to bring AI onto the lab bench, but Google and Anthropic have different focuses.

Regarding how to control instruments, Google “from reality to virtual” focuses more on specific workstations, and VERID is responsible for fully modeling the states of devices, reagents, and samples, with every instruction from the agent first undergoing simulation and verification.Anthropic chooses “virtual-real integration,” participating in setting model hardware standards (MHS) and attempting to use a unified interface to connect microscopes, liquid handling devices, robotic arms, and other instruments to AI, achieving cross‑device universal connectivity.

In terms of the depth of AI involvement in experiments, Anthropic still relies on scientists to conduct experiments, with Claude responsible for finding clues in massive datasets and orchestrating external specialized models to complete designs, thereby accelerating validation and accumulating first‑hand research experience. Google goes a step further, hoping that AI can conduct experiments, feed the readings directly back into the model, and allow the agent to adjust its strategy within the loop.

As for commercial translation, Anthropic has explicitly stated that it will not conduct commercial clinical trials itself, to avoid direct competition with pharmaceutical clients; within the Google ecosystem, drug development is handled by the spin‑off Isomorphic Labs, while the headquarters focuses more on method development.

AI vendors are moving rapidly in the life sciences field, and they have almost simultaneously reached the same conclusion: the value of experimental data is rising quickly.

When a molecule shows no effect, is it because the design was poor, the synthesis failed, or the assay was flawed? With only the final result, attribution is difficult. However, if the operational and measurement processes are retained, researchers can more easily determine which failures merit retesting and which pathways should be abandoned.

This information is hard to fill in using only publicly available papers: papers describe important methods and results, but rarely present every detour taken during a study in full.By putting models and experiments into the same workflow, the team has the opportunity to continuously accumulate these records. When a similar problem arises later, they can look back instead of starting from scratch.

Overall, the significance of wet labs for AI companies lies in allowing them to test models more directly: the gaps between predictions and actual measurements, and which designs repeatedly fail, can be exposed earlier.

Will pharmaceutical R&D change because of this?

When AI agents enter the lab, the earliest beneficiaries are likely the discovery phase: designing and optimizing candidate molecules requires repeated trial‑and‑error, and after each round of experiments researchers must analyze data, interpret anomalies, and plan the next round.If an agent can take over part of this work, it can reduce waiting times and free researchers to explore more options. Moreover, it could help teams abandon unproductive paths earlier.

But there is still a long way from identifying a target to developing a drug: whether the molecule can produce the desired effect, whether it is stable in the body, whether it can reach the target site, and whether it is safe—all still need to be validated… Google’s current results are not enough to demonstrate that the entire R&D chain will save a substantial amount of time, nor have they proven that these early molecules already have therapeutic efficacy.

Even with higher automation, researchers still need to discern which results are trustworthy. A robot can follow a protocol step‑by‑step, but that doesn’t guarantee the assay is sufficiently thorough; an AI can interpret a data set, but that doesn’t mean the interpretation is complete. Experiments are faster, so controls, verification, and anomaly handling must keep pace, otherwise a single misjudgment could be carried into subsequent designs.

Currently, Google still retains human intervention: researchers review sequences, verify protocols, and supervise the robots. Decisions about how to set parameters, which results require alternative validation methods, and whether the next round is worth further investment all still require the researchers’ own judgment.

A more realistic division of labor at this stage is to let the intelligent agents take on more design, operation, and analysis, while researchers spend their time on judgment. By bringing AI into the lab, Google can compress the often time‑consuming workflow between prediction and validation. Whether it can accelerate drug discovery will have to be answered by subsequent experiments.

Reference:

https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/

https://www.gstatic.com/science-ai/verid_semantic_lab.pdf

https://www.gstatic.com/science-ai/agentic_aptamer_sandbox.pdf

https://www.anthropic.com/news/claude-discovers-novel-enzyme-system

https://openai.com/index/introducing-gpt-rosalind/ 

Note: Cover/featured image generated with AI assistance