Abstract B065: Deep learning inference of the gene expression of specific cell types from histopathology of breast tumors
Andrew T Wang, Saugato Rahman Dhruba, Kun Wang, Eldad David Shulman, Eytan Ruppin · Cancer Immunology Research · 2025
Abstract Motivation: The immune-stromal microenvironment plays a crucial role in breast cancer precision oncology, offering insights that could guide targeted therapies. However, high-resolution methods like single-cell RNA sequencing remain prohibitively expensive and are rarely feasible in clinical settings. While recent deconvolution methods can estimate cell type-specific expression from bulk RNA sequencing (RNA-seq) data, even bulk assays are costly and have long turnaround times, limiting their widespread use. To address these challenges, we developed SLIDE (SLide-based Inference of Deconvolved gene Expression), a deep-learning model that accurately predicts cell type-specific expression profiles directly from standard histopathology slides of breast cancer tumors, making high-resolution immune profiling accessible for clinical use. Approach: To train SLIDE we first used our deconvolution tool, CODEFACS, to analyze bulk RNA-seq data from the TCGA breast cancer cohort to obtain deconvolved gene expression profiles across 8 immune-stromal cell types as well as cancer cells. SLIDE was then trained on these data to predict expression profiles from corresponding H&E images. Results: SLIDE demonstrates robust predictive performance in cross-validation on the 1,106 breast cancer patients from the TCGA dataset and in an independent validation cohort of 160 cases. Across the 9 cell types in both cohorts, it robustly infers the expression of >1,000 genes on average, achieving a Pearson correlation >0.4 between predicted and deconvolved expression. For the top 2 cell types (cancer epithelial cells & cancer-associated fibroblasts (CAFs)), SLIDE robustly infers >2,000 genes across both cohorts. To further evaluate SLIDE’s clinical utility, we applied DECODEM, a computational tool for predicting chemotherapy response from cell type-specific expression, to the SLIDE-inferred cell type-specific expression. When applied to an external validation cohort of 63 triple-negative breast cancer patients, five cell type-specific models, including 4 immune-stromal cell types (CAFs, endothelial cells, perivascular-like cells, and myeloid cells) outperformed both direct classification models from H&E images and models based on inferred bulk expression. Remarkably, despite using this cohort solely for external validation, the five cell types achieved performance comparable to a previously published H&E-based model that was both trained and cross-validated on the same dataset. Conclusions: In summary, SLIDE offers a cost-effective and accessible solution for inferring cell type-specific gene expression, with clinical applications including predicting chemotherapy treatment response with higher accuracy than is achievable from the inferred bulk expression. This framework enables cell type specific expression analysis at a scale and speed previously unattainable with traditional transcriptomic assays, paving the way for more personalized treatment strategies and enhancing immune-stromal profiling, potentially leading to improved and timely prediction of patient outcomes. Citation Format: Andrew T Wang, Saugato R Dhruba, Kun Wang, Eldad D Shulman, Eytan Ruppin. Deep learning inference of the gene expression of specific cell types from histopathology of breast tumors [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 B065.