Microsoft Research unveils Quine, an AI for biology labs

Put simply, it's a model that tries to guess what happens to cells when you poke them, and what happens after that.

Quine learns shared representations across different kinds of biological data and at different scales, including sequence, structure, function, cellular state and imaging. It trains on genomics, proteins, chemistry, RNA and cell state, and bioimaging. Because it learns all of those together, instead of stitching separate specialist models side by side, Microsoft says evidence at one level can inform predictions at another. The team says learning across connected representations made the model stronger, not weaker.

The authors are clear about what Quine isn't. It doesn't replace experiments. It's there to help scientists explore, rank and prioritise options before they spend scarce lab time. "Critically, the world model doesn't need to be perfect," they wrote. They went further, adding "indeed, it will never perfectly model biology." It "simply needs to usefully inform experimental design."

The other half of the system is the harness. Microsoft's project page describes a tool layer that helps researchers break questions into steps, use models and scientific tools, compare evidence and revise a plan. Results from the lab then feed back in, so each round of experiments can sharpen the next round of predictions.

The pancreatic cancer test

Pancreatic ductal adenocarcinoma, or PDAC, is the most common form of pancreatic cancer and one of the hardest to treat. Microsoft and the Broad have spent years testing an idea. It's that a tumour's behaviour and how it responds to drugs depend not just on its genetics, but also on the transcriptional state of its cells, which is roughly which genes the cells have switched on. In PDAC, tumour cells can sit in different states, and those states are linked to how they respond to treatment.

With Quine, the team predicted and ranked thousands of compounds on how likely they were to push tumour cells from one treatment-relevant state to another. In wet-lab studies of the shift from the "classical" state to the "basal" state, Quine's highest-ranked compounds produced the largest intended shifts across the assays. Microsoft says the whole process, from narrowing the search to picking a handful of candidates for lab checks, took one weekend, "potentially saving months of experimental work and significant research costs."

Some of the strongest effects came from compounds that work in unexpected ways. The team calls that early evidence that AI can find new openings for repurposing existing drugs and discovering new ones.

Going the other way, from basal back to classical, was harder. Quine had predicted that available compounds would only have a weaker effect there. Then came a result the team says it didn't fully expect. Quine predicted that several compounds would reliably push cells towards a distinct third state, and the lab backed that up. Microsoft says that suggests the pancreatic cancer cell-state picture is richer than a simple line from classical to basal. "The experiments not only tested the model's hypotheses but also generated new ones," the authors wrote.

Here's the important part. These are lab results in cell lines. They aren't treatments, and nobody's been treated with anything. Microsoft says Quine is "intended only for research, not clinical or medical use," and that its outputs may be incomplete or inaccurate and need to be reviewed and tested in experiments by qualified researchers.

Opening the doors to outside scientists

Microsoft isn't keeping Quine to itself. It's opened the Quine Fellows program, a 16-week paid fellowship hosted by Microsoft Research in Cambridge, Massachusetts. Applications are open from 29 September to 2 November 2026, and the fellowship runs from 7 June to 24 September 2027.

PhD candidates, postdocs, research scientists, and academic or independent researchers can apply. Fellows will work with Microsoft researchers and get access to Quine and computing power, plus experimental support where it fits their projects. Areas of interest include protein design and engineering, enzyme design and discovery, changing cell state with genetic and chemical tweaks, and early-stage therapy research in disease areas that don't get much funding.

Access is being opened slowly on purpose. Microsoft says it's starting with the Fellows program and select research collaborations, with internal review and built-in safeguards, and that it expects to widen access later through products like Microsoft Discovery. "Building responsibly will always come first," the authors wrote.

Part of a busy stretch for AI in the lab

Quine builds on years of Microsoft work across proteins, regulatory genomics, tumour cell state, tissue pathology, microscopy and biological data, according to its project page. Microsoft says the system was already central to how its own teams approach discovery in cancer biology, protein engineering, genomics and bioimaging before it opened Quine up.

It's not the only push to bring AI closer to the lab bench. In August, Anthropic opened a research preview of a shared standard so AI agents can run lab and factory instruments. On the same day Quine launched, a team led by Baylor College of Medicine published work in Nature Communications that used AI-based protein structure modelling to speed up the discovery of molecular glues aimed at blood cancers and autoimmune diseases, Phys.org reported.

Microsoft's own yardstick is a practical one. The authors say benchmarks and leaderboards matter, but the real test is whether Quine helps when the evidence is patchy and the questions are new. "The protagonists are not the model or the platform," they wrote. "They are the scientists, the experiments, and the discoveries that follow."

Microsoft Research has posted an official introduction to Quine on its YouTube channel.