Integrated model for drug-target interaction prediction leveraging multiple protein language embeddings

The prediction of drug-target interaction (DTI) is crucial for drug design and development, reducing costs and boosting discovery efficiency. However, a core bottleneck of existing methods is their inability to fully leverage the rich and complementary semantic information from diverse protein language models (PLMs), which restricts the accuracy and generalization capability of DTI prediction. To address this, an Integrated Model for Drug-Target Interaction (IM-DTI) has been developed. It enhances DTI prediction accuracy through multi-module architecture and multi-source information integration. Initially, the model uses a topology-based feature extraction approach combined with a Parallel Feature Encoder to encode multi-level representations of protein sequences. Subsequently, the Dual-modal Embedding Module deconstructs the deep features of drugs and targets, generating three prediction outcomes based on multiple PLMs. These predictions are integrated using a Bayesian Optimization-enhanced Random Forest (BO_RF) algorithm, further improving accuracy and generalization. Furthermore, residual structures in the model accelerate convergence and reduce overfitting. Experiments show that IM-DTI outperforms seven benchmark methods with performance gains of 4%-10%. These findings confirm the effectiveness of the IM-DTI model for DTI prediction, offering new insights and foundations for future drug discovery.