Microsoft's Quine searches biology with a world model

Microsoft Research introduced Quine, an experimental AI research system built for biology. It has two parts: a multimodal world model of biology trained jointly across genomics, proteins, chemistry, RNA and cell state, and bioimaging, plus an interactive harness that connects orchestration and reasoning models to scientific tools, literature, wet labs and researchers.

The joint training is the design argument worth noting. Microsoft’s claim is that evidence at one scale informs predictions at another, and that training across modalities together strengthens rather than dilutes performance, rather than blurring them.

What it actually did

Working with the Broad Institute of MIT and Harvard on pancreatic ductal adenocarcinoma, Quine predicted and ranked thousands of compounds for pushing tumor cells between therapeutically relevant states. The top ranked compounds produced the largest classical to basal transcriptional shifts in wet lab assays. The full cycle of narrowing the search space, prioritizing candidates and validating them in the lab took one weekend, against an estimate of months of experimental work.

The system also predicted that the reverse basal to classical transition would be weaker, and it was. More interestingly, the experiments surfaced a distinct third phenotype the team had not anticipated, which is evidence that the cell state landscape is richer than a single axis.

Access is gated to the Quine Fellows program and select research collaborations, with wider availability planned through Microsoft Discovery.

Microsoft’s own framing is careful and worth repeating: Quine is “experimental research technology intended only for research, not clinical or medical use,” its outputs may be incomplete or inaccurate and require review by qualified researchers, and the world model “will never perfectly model biology.” It also notes that some of the strongest effects came from mechanisms it did not anticipate, which cuts both ways on how much of the reasoning is understood.

(Original source: Microsoft Research)