EACR25-3030

Path2Space: An AI Approach for Cancer Biomarker Discovery Via Histopathology Inferred Spatial Transcriptomics

E. Campagnolo1, E. Shulman1, R. Lodha1, A. Stemmer1, P. Jiang1, C. Caldas2, S. Knott3, D. Hoang1, K. Aldape4, E. Ruppin1
1National Cancer Institute (NIH), Cancer Data Science Laboratory, Center for Cancer Research, Bethesda, United States
2University of Cambridge, Department of Clinical Biochemistry and Institute of Metabolic Science, Cambridge, United Kingdom
3Cedars-Sinai Medical Center, Department of Biomedical Sciences, Los Angeles, United States
4National Cancer Institute (NIH), Laboratory of Pathology, Center for Cancer Research, Bethesda, United States
Introduction:

Spatial transcriptomics (ST) is transforming our understanding of tumor heterogeneity by enabling high-resolution, location-specific mapping of gene expression across tumors and their microenvironment. However, the high cost of ST currently limits its translational potential, hindering the discovery of spatial biomarkers for treatment response. To address this, we developed Path2Space, an AI approach that predicts spatial gene expression directly from histopathology (H&E) slides.

Material and method:

Path2Space is a deep learning framework trained on paired H&E and spatial gene expression data from breast cancer tumors. The model was trained and validated using cross-validation on a large ST breast cancer dataset and externally evaluated on two independent datasets. Predicted gene expression was then used to infer cell-type abundances and identify spatial domains in over 1,000 TCGA breast tumors. Finally, the inferred spatial expression patterns were used to predict treatment response in three independent trastuzumab and chemotherapy cohorts.

Result and discussion:

Path2Space robustly predicts the spatial expression of over 4,300 genes in external validation cohorts, significantly outperforming existing ST predictors. It accurately infers cell-type composition in the tumor microenvironment (TME), with performance comparable to that of models using measured ST data. Applied to over 1,000 breast cancer slides from TCGA, Path2Space reveals three spatially distinct tumor subgroups which significantly stratify patient survival. Furthermore, in three independent clinical cohorts, TME features inferred by Path2Space from H&E slides outperform existing models based on bulk multi-omics data in predicting response to chemotherapy and trastuzumab.

Conclusion:

Path2Space is a fast, cost-effective, and generalizable framework for predicting spatial gene expression and TME composition directly from routine pathology slides. By enabling high-throughput spatial analysis at a large scale, it facilitates the discovery of spatially grounded biomarkers and advances precision oncology for breast cancer and beyond.