From cells to patients

Architecture diagram: patient samples of cells go through a cell embedder, cell aggregator, and patient embedder to produce a single patient-level embedding and disease severity score

Bag of cells in, patient-level embedding out: disease severity and biology follow directly.

Foundation models like SCimilarity work at the level of the single cell, millions of them. But most of the decisions in drug development happen at the level of the patient, and we have a very different scale of data there: tens of thousands of tissue samples, not millions of cells. This project, which started as a single summer intern's project, was our first real attempt to bridge that gap: take a tissue sample, a "bag" of thousands of individual cells with no inherent order, and compress it into one patient-level embedding, a list of numbers that describes that patient's disease biology as a whole, the same way SCimilarity compresses a single cell. From that one representation, the model can also read off things like disease severity directly, instead of needing a separate model built for each new question.

Interpretability diagrams showing which cell types and which genes most influenced a patient-level prediction

The aggregation step doubles as interpretability: for any patient prediction, the model can point back to which cell types, and which genes within them, moved the needle most.

That matters in practice: a prediction is far more useful to a researcher when it comes with a reason attached, not just a number. I think this direction, patient-level modeling, is one of the places where AI in biology can have transformational impact on a near-term timeline, in contrast to new-target discovery, which has to survive a decade-long development pipeline before it means anything. Getting the representation right at the patient level could change how quickly we understand two open questions: why patients with the same diagnosis can have very different disease courses, and why the same drug helps one patient but not another. Both matter well before any of that translates into a new molecule.

Liu, T., et al. Learning multi-cellular representations of single-cell transcriptomics data enables characterization of patient-level disease states. Cell Systems, 2026.