DeepMLP: A Proteomics-Driven Deep Learning Framework for Identifying Mis-Localized Proteins across Pan-Cancer

Accurate protein subcellular localization (PSL) is essential for proteins to perform their biological functions, whereas protein mis-localization can disrupt cellular processes and contribute to various diseases, particularly cancer. Although spatial proteomics technologies have enabled systematic studies of PSL, their high cost and technical complexity limit large-scale analysis of mis-localized proteins (MLPs) in cancer. To address this challenge, we propose DeepMLP, a proteomics-driven deep learning framework that identifies MLPs across cancers using large-scale mass spectrometry-based proteomics data. It constructs pathway-aware protein representations via a cross-attention mechanism and, by integrating these representations with dynamic protein-protein interaction (PPI) networks through graph attention networks, synergistically enhances the ability to uncover potential cancer MLPs. Benchmarking demonstrated that DeepMLP consistently outperformed state-of-the-art methods in both accuracy and stability for PSL prediction in normal and tumor conditions. Furthermore, we systematically identified potential MLPs across pan-cancer types and further uncovered a subset of potential mis-localized protein kinases. Functional enrichment analyses of these MLPs revealed their significant involvement in cancer-related metabolic and signaling pathways, highlighting their potential biological relevance in tumorigenesis.