Spatial transcriptomics informed tumor purity estimation from histology slides for triple negative breast cancer

Introduction: Tumor purity, defined as the proportion of malignant cells within a tumor region, is a critical factor in cancer research and clinical practice. Accurate tumor purity estimates (TPEs) are crucial in triple-negative breast cancer (TNBC), where tumor heterogeneity complicates diagnosis, biomarker interpretation, and therapeutic decisions. Traditional pathological assessment of tumor purity is limited by observer variability and scalability. Spatial transcriptomics (ST) integrates whole-transcriptome data with spatial context, enabling high-resolution and scalable estimation of tumor purity directly from H&E-stained slides. This study develops and validates ST-supervised deep learning models for spatially resolved, reliable TPEs that support more precise clinical evaluation and treatment guidance in TNBC. Methods: Visium ST data and matched 40× H&E whole-slide images from 25 TNBC patients were collected from Dartmouth-Hitchcock Medical Center (DHMC), yielding 120,973 50‑µm Visium spots with co-registered 512×512-pixel H&E patches. Tumor regions were segmented using a validated DeepLabv3 model. Cell-type reference profiles were derived by integrating an external breast single-cell atlas with in-house single-cell RNA-seq data from seven TNBC patients using SCANVI. The integrated single-cell data were mapped onto in-house Visium sections with Cell2Location to estimate spot-level tumor cell-type proportions, serving as supervisory labels. A deep learning model (VIDCellTyper), built on the Virchow 2 model, was trained and cross-validated to predict spot-level tumor proportions and produce spatially resolved tumor purity maps. HoVerNet-derived cell counts from each patch were used for slide-level purity computation via cell-count-weighted averaging. The workflow was validated on an independent TNBC cohort (n=29; DHMC and Cedars-Sinai Medical Center), where slide-level purity was derived from aggregated patch-level predictions by the ST-informed model and HoVerNet. Results: Across tumor regions, the ST-supervised deep learning model achieved a spot-level purity correlation of 0.88 (p < 0.001) with spot-level Hovernet-derived TPE. When aggregating across the internal and external cohorts (n=29), slide-level ST-informed TPE showed a Spearman correlation of 0.83 ( p < 0.001) with Hovernet-derived TPE. Conclusion: This proof-of-concept study shows that ST can guide computational models to derive TPE directly from H&E slides, yielding accurate and spatially resolved results. The ST-guided model generalized across independent TNBC cohorts, demonstrating robustness across tissue workflows. Future work will refine and validate this approach in clinically relevant contexts, including therapy response, prognosis, and pre- versus post-chemotherapy evaluation, to advance precision in treatment assessment and biomarker development.