Multimodal molecular representation learning via text retrieval augmentation

Summary Molecular representation learning plays a critical role in drug design and personalized therapy. However, most existing methods overlook the abundant information present in vast chemical text corpora. To address this limitation, this study proposes a text retrieval‐augmented, multimodal molecular representation learning framework (MolReTA), aiming to utilize retrieved chemical text knowledge to enhance molecular representation. First, MolReTA retrieves multiple structurally similar molecule–text pairs from databases based on fingerprint similarity and encodes these textual descriptions via a specialized chemical language model. Subsequently, a hierarchical contrastive module incorporating instance‐ and global‐level perspectives is designed, integrating abundant complementary semantic information across different granularities while effectively aligning the feature spaces of text and molecular graph structure. Ultimately, a cross‐attention fusion module is constructed to achieve deep integration of textual and structural features. Experimental results show that MolReTA achieves competitive performance on the MoleculeNet benchmark and effectively regularizes molecular distributions in latent space.