Abstract:The rise of spatial transcriptomics (ST) is transforming our understanding of tumor heterogeneity by enabling high-resolution, location-specific mapping of gene expression across the tumor microenvironment. However, the translational potential of spatial transcriptomics is still limited by its high cost, hindering the assembly of large patient cohorts needed for robust biomarker discovery. Here we present Path2Space, a deep learning model trained on spatial transcriptomics data to predict spatial gene expression directly from histopathology slides. Studying breast cancer, the model was first trained on one ST breast cancer patients cohort and further independently tested and validated on two other ST cohorts. Path2Space identifies 4,500 genes whose spatial expression is robustly predicted in all three cohorts, markedly more accurate than existing ST predictors. Second, we show that based on the inferred ST, Path2Space accurately infers cell-type abundances in the tumor microenvironment (TME). Thirdly, Applying Path2Space to the TCGA breast cancer cohort we identify three TME spatially grounded subsets of patients that have markedly different survival rates. Fourth, analyzing two large breast cancer patients cohorts treated with trastuzumab, we show that the heterogeneity of the inferred HER2 levels across the tumor microenvironment is predictive of response, concordant with the bystander hypothesis. Finally, we developed a spatially interpretable model for predicting trastuzumab response directly from the H&E-slides. This model's performance surpasses that of existing response prediction models that use measured bulk multi-omics data. Path2Space is the first approach capable of robustly delineating the tumor microenvironment on an unprecedented scale directly from the tumor slides. It heralds the upcoming development of robust spatially grounded biomarkers of patients survival and treatment that promises to transform precision oncology in both the developed and developing worlds.

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