Multi-modal contrastive learning based on molecular and textual data for drug response prediction

Accurate prediction of drug response is crucial for advancing precision medicine and optimizing therapeutic strategies. However, current deep learning-based drug response prediction (DRP) methods still suffer from significant limitations. They predominantly rely on molecular features while overlooking biomedical knowledge embedded in textual corpora that can provide complementary biological insights. Moreover, these methods inadequately integrate attribute-based representations (focusing on intrinsic properties) with structure-based representations (capturing heterogeneous relations) for modeling drug-cell interactions. To address these challenges, we propose a multi-modal contrastive learning method that integrates biological knowledge with dual-view molecular features. First, we incorporate textual descriptions of biological pathways and molecular mechanisms from authoritative databases as a complementary modality. Next, attribute-based representations are extracted from multi-omics profiles and drug molecular graphs, while structure-based representations are captured from drug-cell line interactions using a heterogeneous graph neural network. Finally, a contrastive learning module is designed to align these multi-modal representations through cross-modal learning (textual information vs. attribute view and intra-modal learning (attribute view vs. structure view), thereby establishing a unified semantic space for enhanced feature integration. Experiments conducted on the GDSC and CCLE datasets demonstrate that our proposed model outperforms previous state-of-the-art methods in DRP. Source code and datasets are available at https://github.com/myduan18/CLMT.