BalancedDiff: Balanced Diffusion Network for High-Quality Molecule Generation

Traditional drug discovery and development are time-consuming and expensive. Deep learning-based molecule generation techniques can reduce costs and improve efficiency, helping to generate high-quality molecules with desirable properties. However, existing deep learning-based methods focus on designing complex network structures to extract key features, which ignore the impact of sample bias and rarely take biochemical principles into account. To solve the above problems, a Balance Loss is proposed to balance sample bias. Second, we designed a KAN-based Balanced Feature Filtering (KBFF) module that balances molecular feature information with spatial location data, effectively filtering out unrelated groups. This approach ensures that the model considers both the chemical properties of functional groups and their spatial arrangements, minimizing noise while preserving critical biochemical relationships. By achieving this balance, the module improves the generated molecular quality. Besides, while diffusion models generate numerous molecules, their effectiveness and reliability remain uncertain, limiting their practical utility. To overcome this limitation, we introduce a QikProp module that predicts ADME properties, filtering out molecules with poor drug-like characteristics or potential safety risks, thereby enhancing the quality and applicability of generated molecules. Experiments on the CrossDocked2020 data set demonstrate the superiority of our method.