Sparse Regression-Based Spatial Deconvolution Recovers Cellular Tumor Ecology and Subtypes in Breast Cancer: Methodological Study

Objective: This study aimed to analyze the weightinduced sparse regression (WISpR) predicted cellular landscapes of spatial breast cancer data, and to determine whether WISpRderived cell-type compositions preserve subtype-informative structure as effectively as full spatial gene-expression profiles. Material and Methods: 4 triple-negative and 2 estrogenreceptor-positive breast cancer spatial transcriptomics (ST) data were analyzed. WISpR was applied to infer cell-type proportions. Spatial organization was quantified using Moran's I and Geary's C. Cross-sample heterogeneity was evaluated using the nonparametric statistical tests. To assess the discriminative value of WISpR predictions, 5 classifiers as Random Forest, ExtraTrees, Gradient Boosting, Multinomial Logistic Regression, and LinearSVC were trained on ST gene expression and WISpR cell-type compositions under identical spatial cross-validation conditions. Results: WISpR revealed consistent spatial cellular patterns in cancer tissue. Strong spatial autocorrelations in both cancer and stromal populations were confirmed. Nonparametric tests identified significant cross-sample variation, particularly in CAF/myCAF, basal SC, luminal B, and immune subsets. Treebased classifiers achieved near-perfect accuracy on subtype prediction on both ST and WISpR inputs, while linear models performed poorly yet consistently better on WISpR than on ST. Conclusion: WISpR preserves both spatial cellular organization and subtype-informative structure, showing that proportion-based deconvolution can serve as a reliable, interpretable alternative to full transcriptomic profiles.