DTI-RME: a robust and multi-kernel ensemble approach for drug-target interaction prediction.

Background: Drug-target interaction (DTI) refers to the specific mechanisms by which drug molecules interact with biological targets within a biological system. Computational methods are widely employed for DTI prediction, as they are time-efficient and resource-saving compared to experimental approaches. Although numerous DTI prediction methods have achieved promising results, accurately modeling DTIs remains challenging due to three key issues: noisy interaction labels, ineffective multi-view fusion, and incomplete structural modeling.

Results: We propose a novel method termed DTI-RME. The DTI-RME introduces an innovative $L_2-C$ loss function that combines the benefits of $L_2$ loss to reduce prediction errors and the robustness of C-loss in handling outliers. This method fuses multiple views through multi-kernel learning that assigns weights to different kernels. DTI-RME uses ensemble learning to assume and learn multiple structures, including the drug-target pair, drug, target, and low-rank structures.

Conclusions: We evaluated DTI-RME on five real-world DTI datasets and conducted experiments focusing on three key scenarios. In all experiments, DTI-RME demonstrated superior performance compared to existing methods. Furthermore, the case study confirmed DTI-RME's ability to identify novel drug-target interactions accurately, with 17 of the top 50 predicted interactions being validated.