New AI Reads Ordinary Biopsy Slides to Predict Gene Activity

Spatial transcriptomics is a lab technique that measures which genes are switched on in a piece of tissue while keeping track of exactly where in the tissue those genes are active. For cancer researchers, that's gold: it shows which parts of a tumor are aggressive, which are dormant, and how surrounding cells influence it. The problem is cost and complexity. The test is expensive, needs special equipment, and isn't part of routine care — so the vast majority of biopsies never get one.

Meanwhile, nearly every biopsy already produces an H&E slide: tissue stained pink and purple and examined under a microscope. That's cheap, fast, and universal. SpaFactor asks a simple question — can AI squeeze gene-level information out of those ordinary images? Its trick is two-fold. First, it looks not just at a single spot on the slide but also at the surrounding neighborhood, so it picks up on the tissue's local environment. Second, instead of predicting thousands of genes one by one, it groups them into "gene programs" — sets of genes that naturally work together — which reduces noise and makes the model lighter and faster than the heavy alternatives.

In tests across five public datasets, SpaFactor delivered the best overall accuracy, with the clearest gains on spatially variable genes (the ones whose activity changes depending on location, and which matter most biologically). It also reproduced more realistic spatial patterns, suggesting it's capturing real biology rather than statistical noise — and it does so without the bulky computing power that competing methods require.

The catch: this is a research paper, not a product. It's validated on existing datasets, not in real hospitals, and a prediction is not the same as a measurement. Think of it as a promising screening hint that could one day help decide who needs the expensive test — not a replacement for the test itself.