Imagine a future where AI can design DNA switches that work inside mammals, and AI agents can take on computational jobs modeled on entire biology studies.

Actually, you know what? Don’t imagine it. Because those are two examples present in this very issue.

Week after week, AI × bio keeps producing things that would have sounded fanciful not very long ago. It’s hard not to be a little awestruck by the ingenuity and sheer pace of it.

But being awestruck doesn’t cure diseases. Capabilities need to be measured, and computational designs eventually have to survive contact with real biology. One of today’s stories tries to put a number on that progress: Edison Scientific’s BixBench3 scores AI agents against the concrete results produced in published computational biology research.

Wonder is great. Evidence is better. This newsletter has both.

Let’s dive in.

A heart cell and a brain cell contain essentially the same DNA. One reason they behave so differently is that different genetic switches help determine which genes get turned on.

Researchers at the Research Institute of Molecular Pathology (IMP) in Vienna and UC Irvine have now used AI to write entirely new versions of those switches - and shown that they work inside developing mice. The findings are published in Nature Genetics. The longer-term ambition is to make gene expression more programmable: specify the tissue where you want a gene active, then design a DNA switch for that job.

“For almost twenty years, our vision has been to understand gene regulation well enough that we could eventually write it,” Alexander Stark, one of the leading researchers in gene regulation, said in an IMP article accompanying his lab’s new paper. “This study shows that this is possible in mammals. I think it marks the beginning of a new era in which we can design genetic switches for virtually any tissue and cell type, including those of humans.”

The switches are called enhancers: stretches of regulatory DNA that help control when and where genes are active. Proteins called transcription factors recognize patterns in enhancer DNA, and different tissues have different combinations of those proteins at work.

Stark and his team first trained relatively small deep-learning models on maps of chromatin accessibility in heart, limb and nervous-system tissue from mouse embryos. These maps show which stretches of DNA are accessible to regulatory proteins - an important clue to where enhancers may be active.

The researchers then fine-tuned those models on experimentally validated enhancers from mice and humans. There were only a few hundred known examples for each tissue, but that was enough for the models to learn patterns associated with enhancer activity in the heart, limbs and central nervous system.

Starting from random DNA, a model-guided design system generated completely synthetic sequences predicted to work as enhancers in each tissue. The researchers selected five designs for the heart, five for the limbs and five for the central nervous system, attached each to a reporter gene that makes enhancer activity visible, and inserted them into mouse embryos.

All fifteen switched the reporter gene on in the tissue they had been designed for. And these weren’t close copies of known switches: none of the 15 significantly matched the mouse or human genome or the known enhancers used in the study.

“We were surprised that all fifteen enhancers worked,” said Vincent Loubiere, a former postdoc in Stark’s lab, to IMP. “That tells us that the regulatory language encoded in mammalian DNA is much more systematic and learnable than many people had expected.”

The results were strongest for the heart and nervous system. Most of those stayed within their intended tissue. Four of the five heart enhancers switched on only in the heart, and four of the five nervous-system enhancers only in the nervous system. The limb enhancers, however, also activated related connective tissues, and attempts to target specific brain regions were less successful.

In addition to that, the researchers wanted to find out whether the approach could work for tissues where few or no experimentally validated enhancers are available. Instead of relying on known enhancers, they used genome-wide measurements associated with enhancer activity - especially regions where the DNA was accessible in the target tissue but not in others.

In tests against known enhancers, models trained this way still performed well, suggesting the method could eventually be extended to many more tissues and cell types. That part, though, has not yet been validated in animals.

Studying organ development is one obvious direction, as is using synthetic enhancers to understand how disrupted gene regulation contributes to disease. Another is gene therapy.

Getting a therapeutic gene into the right organ is only part of the challenge. Once the DNA arrives, you also want the gene to switch on in the right cells - and remain off elsewhere.

An engineered enhancer could provide that extra layer of control. A heart-directed gene therapy, for example, might also reach the liver. Pairing the therapeutic gene with a heart-specific enhancer could help keep it active in heart cells while largely silent in liver cells. For now that remains a future application.

This is an important step toward more programmable control of gene expression. While designing regulatory DNA is not new, doing it predictively inside a mammal is a first.

Adaptyv Bio has raised a $40 million Series A to scale its automated protein-testing operation, doubling its team in Lausanne and opening a new lab in London later this year.

Just last week, BAIO covered Anthropic using Adaptyv’s lab to test 1,320 proteins designed by Claude. “We believe this is where biology is going: AI agents designing experiments, running them in automated wet labs, learning from the results and iterating,” Adaptyv says in a press release.

The new round is separate from that collaboration, but it shows investors are willing to fund the physical infrastructure behind this kind of AI-driven experimentation.

The Lausanne company lets researchers and AI agents submit protein sequences through an API. Its lab turns them into physical proteins, runs binding assays and returns experimental data.

The round, led by Highland Europe, comes as Adaptyv says throughput has grown more than fivefold in the past year and that it now has over 100 customers, including Chai Discovery, Boltz, Roche and Novo Nordisk.

The lab-in-the-loop race is heating up. As AI agents take on more of the design work, companies are competing to build the automated laboratories that can turn those designs into experiments - and feed the results back into the next round.

Edison Scientific says AI agents are “approaching the ability to complete computational biology work at the scale of entire research studies.” Its new BixBench3 benchmark is an attempt to measure that.

The benchmark takes 20 published biology studies and gives AI agents the research objective, a broad analysis plan and the underlying data. From there, the agents have to do the computational work themselves: write and run code, process the data, manage long pipelines and reproduce the outputs of the original study.

Thirteen frontier models took the test. GPT-5.6 Sol came first, reproducing about half of the key computational results from the original studies on average. Runs lasted almost seven hours on average and sometimes reached the full 24-hour limit.

AI agents are moving from solving scientific tasks to carrying substantial chunks of a research workflow. They are not yet reproducing whole studies reliably, but “give the AI several hours and a paper-sized computational job” has become something we can actually benchmark.

Insilico Medicine has launched O3DC, a free index for finding and comparing AI drug-discovery benchmarks. It covers hundreds of benchmarks across everything from docking and binding affinity to molecule generation, clinical prediction and AI agents. Alongside what each benchmark measures, who maintains it and where the code lives, O3DC collects published flaws and biases that can make impressive scores misleading.

☑️ Noetik hits first milestone after deploying OCTO-VC models at GSK.

☑️ Biotech needs “AI translators” between bench scientists and software engineers.

☑️ Single-cell perturbation models can look right while missing the drug-response mechanism.

☑️ BenchSci brings AlphaFold 3 and AlphaGenome into its agentic R&D environment.

☑️ Atlas starts preclinical testing of an AI-identified drug-repurposing candidate.

☑️ AI-designed proteins move into commercial research assays.