The paper proposes a new deep learning framework, DeepPhosPPI, for predicting functional phosphorylation sites and regulatory effects on protein-protein interactions (PPIs). The method aims to accurately construct a comprehensive framework for the two-stage prediction problem: one for functional site identification (DeepPhosPPI-1) and another for its regulatory effect on PPIs (DeepPhosPPI-2). DeepPhosPPI-1 identifies functional phosphorylation sites using a CNN-based attention model, while DeepPhosPPI-2 classifies their regulatory effects on PPIs using an ensemble of CNNs, Transformers, and soft voting. Model performance was compared with competing methods based on phosphorylation site identification, regulatory effect classification on PPIs, and features such as global dependencies, local features, and voting mechanisms. Additionally, the BAD protein (Uniprot ID: Q92934) was used in a case study to explore DeepPhosPPI predictions. BAD regulates cell survival and apoptosis through partner protein interactions, with its abnormal phosphorylation being linked to cancer and neurodegenerative disorders. Experimental results indicate that DeepPhosPPI surpasses the ability of existing and accepted models in identifying phosphorylation sites and classifying regulatory effects. The model was also effective in identifying potential phosphorylation sites in the BAD protein and visualising a key PPI involved in suppressing apoptosis. This study presents DeepPhosPPI as a promising framework for advancing disease understanding and therapeutic development.