Abstract B105: Predicting spatial transcriptomics from histopathology with <i>Path2Space:</i> AI charts breast cancer patients tumor immune microenvironment on an unprecedented scale

Abstract Background: Spatial transcriptomics (ST) is revolutionizing our understanding of tumor heterogeneity by providing high-resolution, location-specific gene expression mapping across the tumor microenvironment (TME). Despite its transformative potential, ST’s high cost limits its widespread clinical application, particularly for large-scale biomarker discovery. Methods: To address these challenges, we developed Path2Space, a deep learning model trained on ST data, to predict spatial gene expression directly from histopathology slides, offering an accessible alternative to experimental ST assays. First, Path2Space is validated using three independent spatial transcriptomics (ST) breast cancer cohorts, demonstrating robust and consistent gene expression prediction across these datasets. Second, the model’s predictions allow for detailed TME characterization, including the accurate inference of cell-type abundances and immune cell distributions within spatial domains. Thirdly and most importantly, we applied Path2Space to the TCGA breast cancer cohort to assess the prognostic potential of immune and TME spatial patterns. Results: Path2Space achieves unprecedented accuracy and coverage in robustly predicting the spatial expression of 4,500 genes across three independent spatial transcriptomics (ST) cohorts, markedly outperforming current methods by a factor of 10. Applied to the TCGA breast cancer cohort, we identify three patient subsets with distinct immune tumor microenvironments (TME) having distinct survival outcomes. The immune TME of the poor-prognosis subset has dense cancer cell populations with low immune infiltration and marked inactivation of specific key immune interactions. In two trastuzumab-treated breast cancer cohorts, Path2Space identifies spatial domains strongly associated with treatment response, with high tumor-infiltrating lymphocytes (TIL) presence, activation of immune pathways, and robust immune-driven cell-cell interactions, enabling the prediction of patient response more accurately than that based on the measured gene expression. Conclusions: Path2Space represents a groundbreaking approach in spatially mapping the tumor microenvironment directly from histopathology slides, paving the way for the study of the immune TME of different cancer types on an unprecedented scale. By enabling the cost-effective inference of the TME, Path2Space holds potential to transform cancer care across diverse healthcare settings, facilitating the development of targeted treatments and improving patient survival globally. Citation Format: Roshan Lodha, Eldad D Shulman, Emma M Campagnolo, Thomas Cantore, Tom Hu, MacLean Nasrallah, Danh-Tai Hoang, Kenneth Aldape, Eytan Ruppin. Predicting spatial transcriptomics from histopathology with Path2Space: AI charts breast cancer patients tumor immune microenvironment on an unprecedented scale [abstract]. In: Proceedings of the AACR IO Conference: Discovery and Innovation in Cancer Immunology: Revolutionizing Treatment through Immunotherapy; 2025 Feb 23-26; Los Angeles, CA. Philadelphia (PA): AACR; Cancer Immunol Res 2025;13(2 Suppl):Abstract nr B105.