HiST: Histological Images Reconstruct Tumor Spatial Transcriptomics via MultiScale Fusion Deep Learning

Spatial transcriptomics (ST) provides valuable insights into the tumor microenvironment by integrating molecular features with spatial context; however, its clinical utility is limited by high costs. To address this, we develop a multi-scale convolutional deep learning framework, HiST, which utilizes ST to learn the relationship between spatially resolved gene expression profiles (GEPs) and histological morphology. HiST accurately predicts tumor regions across multiple cancer types (e.g., breast cancer; area under the curve: 0.96), demonstrating high concordance with pathologist annotations. Moreover, HiST reconstructs spatially resolved GEPs from histological images with an average Pearson correlation coefficient of 0.74 across five cancer types, outperforming existing models by about two-fold. These high-fidelity spatial GEPs enable tumor heterogeneity assessment from histological images, including identification of tumor subtypes with distinct DNA copy number variations. We demonstrate the clinical utility of the predicted GEPs, which robustly stratify patient prognosis across five cancer types from The Cancer Genome Atlas (e.g., breast cancer; concordance index: 0.78). The predicted profiles further facilitate immunotherapy response prediction and enrichment analyses of relevant biological pathways and markers. Collectively, HiST achieves state-of-the-art performance in spatial GEP reconstruction, providing a reliable molecular representation that enhances downstream tasks such as tumor profiling and clinical analyses.