Abstract 6260: Characterization of the tumor microenvironment’s histologic landscape through histology-based deep learning spatial transcriptomic cell-type deconvolution of colon tumors

Abstract The tumor microenvironment (TME) is an essential component of cancer biology. TME-derived measures such as tumor-infiltrating lymphocytes proportions (TILs) and tumor-stroma ratio (TSR) are used to evaluate TME composition. We leveraged spatial transcriptomics (ST) data from colon tumors to deconvolve cell-type populations and developed a deep learning (DL) algorithm to infer local ST-derived TIL and TSR estimates reflecting TME heterogeneity from held-out H&E whole slide images (WSIs). A total of 44 colon tumors surgical sections were profiled using matched H&E WSI and Visium ST, with 10 patients profiled for single-cell RNASeq. A pathologist annotated tumor microarchitecture within each slide. A total of 284,770 512x512 pixel2 image patches were extracted and paired with corresponding Visium ST spots. Using SCANVI, cell type labels from the Human colon cancer atlas (c295) were lifted over to our in-house scRNASeq data. Cells were mapped spatially to estimate spot-level cell-type proportions using Cell2Location (C2L). Cell-type mapping was validated through human-in-the-loop (HITL) testing, where 512-pixel image crops around 400 randomly selected spots were annotated at each pixel. Spot-level TIL/TSR measures were estimated using cell-type proportions from the C2L-transformed ST data and proportion of pixels within each patch assigned to lymphocytes and tumor. Next, Virchow, a state-of-the-art deep learning pathology foundation model, was used to extract spot-level histological features- a Random Forest model was trained on spot-level features to infer cell-type proportions. Spot-level performance for cell-type mapping (evaluated through HITL) and cell-type proportion inference from histology was measured using the area under the curve (AUC) on a held-out test set of 9 patients. HITL tests demonstrated high agreement between H/ST-TIL (AUC=0.71; 95% CI: [0.60-0.82]) and H/ST-TSR (AUC=0.86; 95% CI: [0.81-0.92]). DL spot-level cell-type deconvolution within tumor-only slide regions achieved a macro-averaged AUC of 0.812 (Tumor: 0.85, Mesenchymal: 0.91, Lymphocytes: 0.79, Monocytes: 0.76, Mast: 0.79, Plasma: 0.78). Previous work inferred spatial expression patterns across broader regions, reporting cell-type AUCs of Tumor: 0.93, Mesenchymal: 0.82, Lymphocytes: 0.76, Monocytes: 0.74, and Mast: 0.91. Our results show that DL can infer ST-labeled cell types from H&E WSI, capturing TME features beyond tissue architecture. HITL testing demonstrated strong concordance between derived spot-level ST-TSR/TIL metrics with their histology-based counterparts (H-TSR/TIL), with utility for pathologist validation of spatially mapped cell-type data. This study highlights future potential of DL to capture spatial distribution of TME measures as prognostic biomarkers through ST inference from histology. Citation Format: Minh-Khang Le, Gokul Srinivasan, Fred Kolling, Keluo Yao, Alos Diallo, Zarif Azher, Louis J. Vaickus, Xiaoying Liu, Joshua Levy. Characterization of the tumor microenvironment’s histologic landscape through histology-based deep learning spatial transcriptomic cell-type deconvolution of colon tumors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6260.