MTCPINet: A Unified Multi-Task Deep Learning Framework for Prediction of Compound-Protein Interaction in Drug Discovery

The accurate prediction of molecular interactions and binding affinities between compounds and their target proteins are a pivotal focus in the field of drug discovery. Traditional single-task models for predicting such compound-protein interaction (CPI) require chaining multiple specialized models to perform compound screening, which can lead to error accumulation and inefficiencies. In this study, we propose a novel multi-task deep learning framework MTCPINet, which integrates classification and regression tasks into a unified model, enabling simultaneous prediction of compound-protein interaction activity and the corresponding binding affinity. To address data heterogeneity between different tasks, this study employs a newly curated high-quality dataset of compound-protein pairs sourced from ChEMBL, paired with a “cold-start for compounds” data preparation strategy. In MTCPINet, the network architecture synergistically combines Graph Isomorphism Network Convolution (GINConv), Convolutional Neural Networks (CNN), a one-dimensional Convolutional Block Attention Module (1D CBAM), and residual modules. And a masking mechanism is incorporated during training to ensure that regression loss is applied exclusively to active compound-target pairs. This design streamlines the drug discovery process through the precise identification of high-quality effective target proteins. Leveraging robust data reutilization strategies, our MTCPINet not only demonstrates potential for drug repositioning, but also facilitates the prediction of target proteins for natural products from traditional Chinese medicine (TCM).