AMCF-RDP: a self-attention-based multi-source and cascade framework for the identification of drug-protein relationships.

The identification of relationships between drugs and proteins not only helps in the study of pathological mechanisms but also in drug repositioning studies. However, conventional wet-lab methods are often plagued by issues such as being time-consuming, labour-intensive, and characterized by low accuracy. Therefore, the development of a theoretical computational method is imperative for the expeditious and precise identification of drug-protein relationships. In this study, a self-attention-based multi-source and cascade framework (AMCF-RDP) is developed to identify the drug-protein relationships. Embedded features and network topology features derived from the knowledge graph and complex network were employed to characterize the drug-protein relationships. A two-layer model was constructed using attention mechanism and fully connected layers and was used to predict whether a drug interacts with a protein and what type of interaction it is. The efficacy of the proposed method was evaluated and confirmed based on the non-redundant datasets, ablation experiments, and comparisons with machine learning algorithms and other state-of-the-art methods. Results from fivefold cross-validation demonstrate that the developed method can quickly and accurately recognize drug-protein interactions with an accuracy of 90.21%, a sensitivity of 90.35%, and a Matthews correlation coefficient of 0.8043. Furthermore, it can also distinguish the types of drug-protein interaction, achieving a macro-recall of 93.43% and a macro-F1 score of 0.9381. Compared to the methods described in the literature, the proposed method achieved an area under the receiver operating characteristic curve of 0.9176, representing an improvement of 0.4746. A total of 100,000 drug-protein associations were identified, some of which were confirmed through molecular docking, KEGG, and gene ontology analyses. The AMCF-RDP has been demonstrated to significantly improve the identification of drug-protein relationships. It is anticipated that this will serve as a valuable tool in the domains of drug development and the investigation of mechanisms of action.