SpaPheno: Linking Spatial Transcriptomics to Clinical Phenotypes with Interpretable Machine Learning

Linking spatial transcriptomic data to clinically relevant phenotypes is essential for advancing spatially informed precision oncology. Here, we present SpaPheno, an interpretable machine learning framework that integrates spatial transcriptomics with clinically annotated bulk RNA-seq to identify spatially resolved biomarkers predictive of patient outcomes, including survival, tumor stage, and immunotherapy response. SpaPheno provides multi-scale interpretability from tissue regions to cell types and individual spatial spots, enabling clear biological insights from complex spatial data. We validate SpaPheno through extensive simulations and applications to multiple cancer cohorts--primary liver cancer, clear cell renal cell carcinoma, breast cancer, and melanoma--demonstrating robust predictive performance alongside biologically meaningful spatial patterns. SpaPheno offers a generalizable strategy to translate spatial omics data into clinically actionable knowledge, facilitating precision oncology informed by tumor spatial architecture.