Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration

Recent advances in Structure-based Drug Design (SBDD) have leveraged generative models for 3D molecular generation, predominantly evaluating model performance by binding affinity to target proteins. However, practical drug discovery necessitates high binding affinity along with synthetic feasibility and selectivity, critical properties that were largely neglected in previous evaluations. To address this gap, we identify fundamental limitations of conventional diffusion-based generative models in effectively guiding molecule generation toward these diverse pharmacological properties. We propose CBYG, a novel framework extending Bayesian Flow Network into a gradient-based conditional generative model that robustly integrates property-specific guidance. Additionally, we introduce a comprehensive evaluation scheme incorporating practical benchmarks for binding affinity, synthetic feasibility, and selectivity, overcoming the limitations of conventional evaluation methods. Extensive experiments demonstrate that our proposed CBYG framework significantly outperforms baseline models across multiple essential evaluation criteria, highlighting its effectiveness and practicality for real-world drug discovery applications.

Another key consideration is Synthetic Accessibility (SA) scores [14], combining structural complexity and fragment contributions into a single numerical value between 0 and 1, have frequently been employed to assess synthetic feasibility. However, high SA scores often do not guarantee practical synthetic routes, highlighting a critical gap in accurately evaluating real-world synthesizability. Moreover, selectivity (ensuring that molecules bind specifically to target proteins without off-target interactions) is equally important as affinity [47,26]. Poor selectivity can cause unwanted side effects, and while recent studies have proposed diffusion-based guidance methods [17,6], these rely heavily on pretrained classifiers predicting true binding molecules. The widely used CrossDocked2020 [15] dataset, initially intended for general docking research, is unsuitable for selectivity evaluations without substantial additional docking computations. Furthermore, unclear criteria for identifying true binders and significant data imbalance hinder obtaining reliable guidance signals. Consequently, establishing biologically relevant benchmark datasets and developing effective controllable generation strategies for selectivity is urgently required. A detailed discussion on this subject is provided in the Appendix J. The Contributions of this research addressing the aforementioned points are as follows:

• We introduce a novel approach that integrates BFN with diffusion models from a guidance perspective to circumvent the limitations of diffusion guidance in SBDD. This approach explicitly formulates a gradient-based guidance mechanism within a Bayesian update process and rigorously establishes its theoretical foundations.

• We provide a comprehensive analysis of how injecting guidance into generative models affects posterior sampling-based predictions and uncertainty in SBDD. This study offers new insights into the interplay between guided generation and uncertainty quantification.

• We revisit the limitations of existing evaluation metrics in SBDD and propose a new set of essential metrics for evaluating practical molecular generative models, including binding affinity, Synthetic Accessibility (SA), and selectivity.

• Through extensive and comprehensive experiments, we demonstrate that the proposed framework CBYG outperforms existing baseline models by a substantial margin. Notably, it achieves state-ofthe-art performance under both conventional and newly introduced evaluation criteria.