Drug Combination Recommendation Based on Multi-View Drug Feature Learning

With the booming development of artificial intelligence technology and the medical industry, deep learning-based drug recommendation models have been playing an increasingly crucial role in the healthcare community. However, the existing approaches suffer from inadequate drug representations and the lack of multi-view drug information. To address these challenges, this paper proposes a novel model named Graph-based Multi-View Learning for Drug Recommendation (MVRM). Specifically, the semantic information of drugs are learned from electronic health records (EHRs) and drug molecular structures are extracted. Moreover, functionally differentiated drug molecular representations are leveraged to distinguish consistent 2D structures from distinct 3D geometric conformations. By incorporating both semantic information and structural relationships, comprehensive drug representations are generated and adverse drug-drug interactions are avoided. Furthermore, a cross-view contrastive learning mechanism is deployed to fully explore the EHR data from multiple views, thereby achieving safe and effective drug recommendations. Finally, the experimental results on EHR datasets demonstrate the outstanding performance of MVRM, compared with the state-of-the-art baseline models in several metrics.