0
0The research article introduces STimage, a specialized deep learning framework designed to predict gene expression and cell types directly from standard H&E histological images. By linking tissue morphology with spatial transcriptomics, the platform enables researchers to visualize molecular data without the high costs of traditional sequencing. A key innovation of the system is its focus on robustness and interpretability, utilizing an ensemble approach and foundation models to quantify both data-driven and model-based uncertainty. The authors validated the tool across diverse datasets, including breast, skin, and kidney cancers, demonstrating its ability to accurately identify cancer markers even in out-of-distribution clinical samples. Ultimately, STimage provides a powerful tool for digital pathology by allowing for single-cell resolution predictions that can help stratify patient survival and inform drug response.
References:
Tan X, Mulay O, Xie J, et al. Robust and interpretable prediction of gene markers and cell types from spatial transcriptomics data[J]. Nature Communications, 2026, 17(1): 1781.
