And just like that AI agents perform brain surgery in mice

In partnership with

An AI agent helping perform brain surgery on a mouse sure sounds like a milestone. An AI struggling to plan a routine biology experiment sounds like a disappointment. Both appear in today’s issue.

The frontier of AI in biology is jagged, too: impressive capabilities in one task don’t necessarily translate into dependable performance in another.

Let’s dive in.

LESS IS MORE

A few examples guide Evo 2’s DNA designs

Arc Institute has shown that just a few dozen examples of DNA with known functions can guide an AI to design new sequences that control gene activity.

The method is called in-context learning: an already-trained AI learns what you want from examples in its prompt, without further training. Familiar from language models since GPT-3, it has also been explored in biology.

Arc, the Palo Alto research institute, tested this with its well-known Evo 2 model. Trained on vast amounts of DNA, Evo 2 can already generate new sequences. But training it to design DNA for a particular job can require thousands of sequences whose functions have been measured experimentally.

And that data can be scarce. Take, for example, enhancers: stretches of DNA that help turn genes on in particular cells. Researchers may know only a small number that work in the cell type they want to target.

In a new preprint, the Arc team reports giving Evo 2 prompts containing tens of these enhancer sequences. It generated new designs without retraining, and tests in cultured human cells confirmed that they worked.

The team also tested designs in living mouse brains. For this, they used additional training, with other AI models scoring and helping improve Evo 2's designs. Seven of nine designs activated a fluorescent marker in mouse brains. Two mainly labeled the intended neuron type.

If this generalizes, biological AI could become useful across many more specialized tasks, including ones with too little experimental data to train a dedicated model. A small collection of well-understood sequences could become enough to start designing.

OF MICE AND AGENTS

AI agents perform brain injections in living mice

“Last night an AI agent performed brain surgery on a mouse.”

That was the announcement from Gabriel Ocana-Santero, founder of San Francisco biotech startup Exin Therapeutics. His team connected Claude to a surgical rig. GPT and Grok have now done it too.

The team gave the agents 19 candidate gene therapies and a goal: reduce seizures. Each chose a treatment and a brain region, then operated the injection needle while watching through cameras.

A veterinary surgeon still made the incision, opened the skull and stitched the animal up. The vet also approved irreversible steps. But the agents were making scientific choices and physically carrying them out in living animals - what Ocana-Santero calls “in vivo lab in the loop.”

Getting the bio-averse Claude to participate was not straightforward. Opus 5.5 refused outright; Opus 5 completed a dry run but declined the live procedure. The team used Sonnet 5 instead.

The Claude, GPT and Grok agents chose different treatments and brain targets. There was only one mouse per model, and whether the therapies reduce seizures remains to be seen.

“We’ve realized that we need to connect these agents to a lab to get the maximum value from them,” Ocana-Santero told WIRED. His argument is that agents need experience that papers alone cannot provide: encountering failed ideas, unexpected results and equipment problems, then working out how to respond.

A NOT-SO-SIMPLE CLONE

AI struggles to plan DNA-building experiments

Molecular cloning - assembling and copying DNA for experiments - is routine work in biology labs. Planning it involves choosing materials, designing sequences and arranging reactions that must work together. But even experienced scientists make mistakes that can take days or weeks to surface.

Reliable AI assistance could save time and prevent wasted experiments. It would offer “enormous scientific and commercial benefit,” researchers at biology software companies Benchling and DNAForge write in a new preprint.

Their BenchBench-Cloning benchmark tests that ability through 65 tasks. Eight AI systems had to design DNA or plan how to build it. Their answers were assessed using simulation and automated grading, not wet-lab experiments.

The best model, GPT-6 Astra, scored 30.6%. Models often handled individual steps correctly but failed to produce a complete, workable plan.

The “AI scientist” pitch is evolving and BenchBench-Cloning reflects a more demanding expectation: earning that title requires competence in everyday scientific work. And that competence is still lacking.

☑️ RNA Foundry offers personalized cancer-vaccine manufacturing and AI-assisted design.

☑️ TBC releases a video model it says draws on discoveries from living neurons.

☑️ OpenAI offers GPT-Rosalind-6 to vetted life-science researchers.

☑️ Parse and Nomic begin building a 100-million-cell immune-response dataset.

☑️ Opentrons raises $40 million for AI-driven laboratory automation.

☑️ BioGeometry brings cumulative Series A funding above $100 million.

☑️ Bruce Booth argues that AI drug abundance could worsen R&D productivity in the near term.

☑️ ReproBAIT tests whether AI agents can recreate biological AI tools from published papers.