Chiral Molecular Graph Encoder for Medication Recommendation

Drug recommendation as an auxiliary medical tool has garnered significant attention from researchers in recent years, primarily due to its potential in identifying effective combination therapy drugs for patients. However, a major challenge arises with naturally occurring chiral drugs, which possess the property that molecules with identical structures can exhibit entirely opposite pharmacological effects. Existing methods often relying on graph structure learning techniques such as graph neural networks, fail to address this issue, leading to safety concerns when recommending chiral drugs. To address the limitations of graph neural network methods in distinguishing chiral drugs with identical molecular structures, we propose a novel chiral molecular graph encoder named ChiMedRec. This encoder employs a fine-grained molecular representation method, incorporating both the original and chiral graph perspectives, thereby effectively differentiating the pharmacological properties of chiral drugs. Experimental results on two public datasets demonstrate that the proposed method significantly improves the performance of drug combination recommendations by considering the molecular chirality of drugs.