AI models are very good at predicting how molecules should behave. Biology is famously bad at taking direction.
Google Research is trying to close that gap. In a blog post dated October 16, 2026, the team laid out building blocks for autonomous wet labs, where AI agents design experiments, run them on physical robotics, read the results, and try again.
The goal is AI that learns from what actually happens in a test tube, not just from what a simulation says should happen.
Meet VERID, the lab’s digital stunt double
The centerpiece is a framework called VERID, short for VERified Infrastructure for Discovery Laboratories. Google describes it as a semantic digital twin of the lab.
According to Google, VERID supports three core functions. Agents can query the system to inspect its state, run time-compressed simulations, and execute protocols safely on physical hardware.
With those guardrails in place, the agents can run closed-loop cycles. They design a batch of candidates, test them, analyze the outcomes, and refine the next batch.
From zero hits to 42.1%
The headline result involves aptamers, short strands of DNA that fold into shapes capable of gripping a target molecule. Google’s agents were tasked with finding aptamers that bind α-thrombin, a protein involved in blood clotting.
The first round did not go well. Out of 167 candidates tested, none hit, for a success rate of 0%.
By the third iteration, the picture looked very different. The agents landed 131 hits out of 311 tested, a hit rate of 42.1%.
Each round involved approximately 1,000 candidates, at a cost of $1 to $3 per candidate.
The aptamer work is deliberately cheap. Google frames DNA aptamers as a low-cost, high-throughput sandbox, a place where agents can fail quickly and inexpensively while the methods mature.
The team then moved to harder targets. Using de-novo protein design, meaning proteins built from scratch rather than adapted from nature, the agents produced binders for cancer-related proteins with single-digit nanomolar binding affinities.
For non-chemists: binding affinity measures how tightly a molecule holds onto its target, and lower numbers mean a stronger grip. Single-digit nanomolar puts these designs in the tight-binding range that drug developers tend to care about.
Why simulations alone keep falling short
The project targets a stubborn problem in computational biology. Predictions made in silico, on a computer, often fail to survive contact with the wet lab.
Real experiments come with measurement errors, contamination, temperature drift, and variables nobody thought to model.
The work also extends Google’s broader AI-for-science push. It builds on Co-Scientist, the company’s multi-agent system designed to help researchers generate and evaluate hypotheses.
The money is moving in the same direction
The research landed shortly after a much larger funding announcement. On October 7, 2026, a coalition including Google DeepMind, Meta, Isomorphic Labs, and US government partners committed nearly $1.8 billion to the Virtual Biology Initiative.
That effort focuses on expanding biological datasets and building virtual cells, computational models meant to simulate how living cells behave.
What this means for biotech and AI
If agents can run tight design-test-refine loops at $1 to $3 per candidate and improve hit rates this quickly, the cost of early-stage exploration could fall.
There are real caveats. The aptamer results come from a sandbox chosen precisely because it is cheap and fast, and the jump from DNA aptamers to complex protein therapeutics is not trivial.
Strong binding affinity is also only one early hurdle. A molecule that grips a cancer protein tightly still has to prove it is safe, stable, and effective in living systems.
Google’s emphasis on safe protocol execution acknowledges that letting AI agents control physical lab equipment raises new questions about oversight and error handling.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.