Researchers have trained a small neural module to extract molecular signals from standard tumor slides - without running a single sequencing assay.
The system sits atop two frozen foundation models, one trained on histopathology images and one on RNA sequencing data, and learns to bridge them using contrastive learning on 1,720 cancer cases. Once trained, it lets users query hematoxylin-and-eosin-stained whole-slide images using gene-set signatures - asking, in effect, whether a given biological pathway is active based on tissue appearance alone. The team reports a 25-fold improvement in retrieval accuracy over baseline methods. Crucially, the foundation models themselves are never retouched; only the small alignment layer trains.
Bulk RNA sequencing is expensive, slow, and largely absent from older tissue archives - meaning a vast store of clinical slides sits molecularly dark. A method that predicts pathway activity directly from H&E images could unlock that archive for research and, eventually, treatment planning. The validation on a real clinical trial cohort, POSEIDON, is the detail that lifts this above typical benchmark papers: predicted immune-signaling scores aligned with PD-L1 expression groups, which matters for immunotherapy decisions.
The honest limit is also spelled out: pathways that leave no visible mark on tissue morphology remain hard to predict, and that ceiling is unlikely to move regardless of how much data gets added. What the paper establishes is a credible floor - cell-cycle and immune programs hit R-squared above 0.5 - and a practical path to domain adaptation when a new hospital's slides look different from the training set. Whether clinical labs will trust an H&E proxy for a sequencing result is a regulatory and cultural question this paper cannot answer.