Abstract
Drug discovery relies on effective representation learning to capture the complex structural and semantic characteristics of molecules and biological targets. However, existing deep learning methods often focus on single-view representations and fail to fully exploit complementary information from heterogeneous molecular descriptors and structural views. In this paper, we propose a multi-view deep representation learning framework for computational drug discovery that integrates multiple molecular and target representations into a unified embedding space. The proposed approach jointly learns view-specific encoders and a shared latent representation through cross-view consistency regularization, enabling the model to capture both local structural patterns and global biochemical semantics. Extensive experiments on benchmark drug discovery datasets demonstrate that the learned representations exhibit strong discriminative capability and generalization performance in downstream screening and similarity retrieval tasks. The proposed framework provides an effective and scalable representation learning paradigm for large-scale drug discovery and molecular analysis.