Inferring Tumor Spatial Gene Expression from Routine H&E Images: Pan-Cancer Validation and Spatial Profiling of Primary Breast Tumors and Lymph Node Metastases

Abstract

Spatial transcriptomics (ST) enables gene expression profiling while preserving tissue architecture, but its high cost and limited throughput constrain large-scale applications. This study evaluated the biological utility of VirtualST, a conditional diffusion model for inferring spatial gene expression from H&E images, in colon adenocarcinoma (COAD), primary invasive ductal carcinoma of the breast (IDC), and breast cancer lymph node metastasis (LYMPH_IDC). Twelve samples with paired H&E images and ground-truth ST data were evaluated using leave-one-sample-out cross-validation. Inference started from pure noise, with ground-truth expression used only for post hoc evaluation, and the final prediction for each sample was obtained by averaging ten sampling runs. The mean gene-wise Pearson correlation coefficients (PCCs) were 0.467, 0.439, and 0.496 for COAD, IDC, and LYMPH_IDC, respectively, with better reconstruction of tumor epithelial, secretory, and proliferative programs. Using ground-truth ST epithelial scores from the held-out samples as an independent molecular reference, the predicted and ground-truth scores achieved within-sample Pearson correlations of 0.79 and 0.66 in IDC and LYMPH_IDC, respectively, with corresponding hotspot Dice coefficients of 0.64 and 0.44. These results indicate that VirtualST can recover within-sample spatial epithelial signals but is not suitable for absolute quantification across tissues or direct pathological region detection.