Abstract C060: Spatial cell-identity and gene expression inference directly from histopathology slides in colorectal cancer with Path2SpaceHD

Background: High-resolution spatial profiling of the tumor microenvironment (TME) offers a promising approach for advancing immunotherapy in solid tumors. However, spatial transcriptomics remains expensive, low-throughput, and difficult to integrate into routine pathology workflows. In contrast, hematoxylin and eosin (H&E) slides are universally available and routinely generated during diagnostic evaluation. We developed Path2SpaceHD, a deep learning framework that predicts both spatial gene expression and cell-type composition directly from standard H&E images. Methods: Path2SpaceHD leverages Virchow2, a ViT-huge transformer pre-trained on histopathology, to extract fine-grained features from 60-pixel patches approximating single-cell resolution. We trained two separate models: one predicting spatial gene expression (∼6,000 genes, regression) and one predicting cell types (9-class classification), using lightweight classifier heads. Training and validation were performed on publicly available VisiumHD data from Oliveira et al., Nature Genetics, 2025. The training set included three VisiumHD slides from three patients (two colorectal carcinoma [CRC], one normal adjacent tissue [NAT]). The validation set consisted of two slides from a fourth patient (one CRC, one NAT). Cell-type ground truths were derived using the VisiumHD annotation protocol described in the same study. Each model was trained via leave-one-slide-out cross-validation (LOOCV), and ensemble predictions were generated by mean pooling across the three resulting models. To benchmark performance, we compared Path2SpaceHD against CellViT, a state-of-the-art vision transformer for morphology-based cell-type inference. Results: Path2SpaceHD robustly predicted the expression of 546 genes in external validation (Pearson correlation > 0.4 between measured and predicted values across spots), including very high prediction accuracy for clinically relevant biomarkers such as CEACAM5, CEACAM6, and EPCAM (correlations > 0.8). For cell-type classification, Path2SpaceHD achieved 77% overall accuracy and 89% top-2 accuracy across nine biologically relevant cell types, including T cells, B cells, and myeloid subsets. Notably, the model resolved morphologically similar immune subtypes such as B cells vs. T cells at near single-cell resolution (Figure 1). A simplified 4-class version of the model, trained for head-to-head comparison with CellViT, demonstrated consistent performance advantages across epithelial, neoplastic, inflammatory, and connective tissue classes. Conclusions: Path2SpaceHD delivers robust, high-resolution spatial insights directly from H&E, enabling immune profiling and biomarker inference without transcriptomic assays. It can be adapted for other cancer types and immune contexts, offering a powerful tool for digital pathology-driven immuno-oncology at single-cell resolution. Citation Format: Roshan Lodha, Amos Stemmer, Victoria Rogness, Alan Shen, Eldad Shulman, Emma Campagnolo, Danh-Tai Hoang, Eytan Ruppin. Spatial cell-identity and gene expression inference directly from histopathology slides in colorectal cancer with Path2SpaceHD [abstract]. In: Proceedings of the AACR Immuno-Oncology Conference (AACR IO): Discovery and Innovation in Cancer Immunology: Revolutionizing Treatment through Immunotherapy; 2026 Feb 18-21; Los Angeles, CA. Philadelphia (PA): AACR; Cancer Immunol Res 2026;14(2 Suppl):Abstract nr C060.