SPARK: deciphering tumor-specific signaling networks through an integrative predictive model

While kinase-substrate associations (KSAs) are fundamental to cancer signaling, their rewiring patterns and functional roles across different cancers during cancer progression remain generally poorly characterized. Current predictive tools often fail to integrate three-dimensional structural constraints and cancer-specific dynamic information, limiting their biological relevance and translational potential. To address this, the present study constructs SPARK, a machine learning framework that integrates multi-modal features within cancer-specific contexts. These include data-driven features (such as phosphoproteomic co-expression, mutual information, protein-level abundance, and phosphosite-level abundance), knowledge-driven features (e.g., STRING protein-protein interactions), and structural features (ESM-2 embeddings compressed via an autoencoder). The training of the model was done with the adoption of the XGBoost classifier on data from eight CPTAC cancer cohorts. SPARK demonstrates strong predictive performance across different cancer types, with AUROC values ranging from 0.929 to 0.957 (mean: 0.944) in normal tissues and 0.941 to 0.958 (mean: 0.948) in tumor tissues. The accuracy of these predictions was further validated against experimentally determined kinase-substrate specificities. Using SPARK’s predictions, tissue-specific phosphorylation signaling networks are systematically reconstructed. These SPARK-derived networks reveal distinct phosphorylation patterns and kinase activation landscapes across cancers while identifying several highly tissue-specific kinases as promising therapeutic targets.