Microsoft Research unveils Quine: a world model of biology that already picks cancer drugs

Microsoft Research has unveiled Quine, its most ambitious attempt yet to give artificial intelligence a working model of biology itself. Announced on September 29 in a Microsoft Research blog post by Nicolo Fusi, VP and distinguished scientist, and Jonathan M. Carlson, vice president, Quine pairs a multimodal world model of biology with an interactive harness that connects the model to scientific tools, the research literature, the wet lab, and the scientists using it. And it is already doing real science: with researchers at the Broad Institute of MIT and Harvard, the system ranked compounds predicted to force therapeutic shifts in pancreatic tumor cells — with several of its top picks validated across multiple wet-lab assays.

A world model, built across biological scales#

Microsoft defines the world model it is chasing as one that represents the state of a biological system, forecasts how that state would change under an intervention, and reasons through the downstream consequences several steps ahead. The point is not to replace experimentation, the company says — it is to let researchers weigh and rank candidate directions computationally before committing limited lab resources.

Quine's core model learns representations shared across biological scales and data types: sequence, structure, function, cellular state, and imaging. Its training spans genomics, proteins, chemistry, RNA and cell state, and bioimaging. The claim that matters most here is about joint training: Microsoft reports that training across these modalities together strengthens rather than dilutes performance — evidence from one modality can inform predictions in another, capturing relationships that would stay siloed or be lost entirely if separate single-domain specialist models were simply orchestrated by an agent.

The system is designed to run as a loop that starts and ends with the scientist: a question yields proposals, proposals become experiments worth running, and the resulting measurements refine both the next question and the model itself. As Microsoft puts it, the model need not be perfect — it only needs to guide experimental design usefully.

Pancreatic cancer, in the wet lab#

The announcement's proof case is pancreatic ductal adenocarcinoma (PDAC) — the most common form of pancreatic cancer and among the hardest to treat. Inside the joint Microsoft–Broad Institute effort known as Project Ex Vivo (with support from the Dana-Farber Cancer Institute), researchers have spent years building patient-derived ex vivo models to test the hypothesis that how a tumor behaves and responds to drugs depends on its transcriptional cell state as well as its genetics.

Microsoft used Quine to score and rank thousands of compounds for their potential to push tumor cells between states relevant to therapy. In wet-lab studies centered on the shift from classical to basal cell states, the compounds Quine ranked highest moved cells furthest in the intended direction across the assays — and several of the strongest effects involved compounds with unexpected mechanisms of action, which Microsoft frames as an early hint that AI can surface opportunities for drug repurposing and discovery.

The striking detail is the pace: narrowing the compound search space to a short list of candidates for lab validation took a single weekend. Microsoft describes that as potentially saving months of experimental work and considerable research cost. There is also an honest failure mode in the report: the reverse shift, from basal back to classical, proved harder — matching Quine's own prediction that available compounds would produce weaker effects in that direction. The experiments additionally confirmed a separate prediction that several compounds would repeatedly push cells toward a distinct third phenotype, suggesting the pancreatic cancer cell-state landscape is more complex than a simple classical–basal axis.

Fellows, phased access, and the road ahead#

With the announcement, Microsoft opened applications for the first Quine Fellows cohort. The 16-week fellowship is hosted by Microsoft Research in Cambridge, Massachusetts, and is open to PhD candidates, postdocs, research scientists, and academic or independent researchers. Areas of interest include protein design and engineering, enzyme design and optimization, genetic and chemical perturbation of cell state, and early-stage therapeutic research in under-resourced disease areas. Applications opened September 29 and run through November 2, 2026; the fellowship itself runs June 7 through September 24, 2027.

The caveats are explicit and worth quoting in spirit: Quine is experimental research technology, intended for research only and not for clinical or medical use. Its outputs can be incomplete or inaccurate and require review by qualified researchers. Access is deliberately phased — the Fellows program and select research collaborations first, with wider availability expected later through products such as Microsoft Discovery as the technology matures.

The bigger picture: the world-model race is no longer only about robots and self-driving cars. From protein folding to cell-state engineering, the labs that can simulate the physical substrate — and close the loop with real experiments — are positioning themselves at the center of AI-for-science. Quine is Microsoft's bid to own that loop in biology. The wet-lab results, if they replicate beyond pancreatic cancer, would be the first evidence that a generalist biological world model can actually shorten the path from hypothesis to candidate drug.