IBEX: Information-Bottleneck-EXplored Coarse-to-Fine Molecular Generation under Limited Data

Dong Xu, Zhangfan Yang, Shuangbao Song, Zexuan Zhu, Jianqiang Li, Junkai Ji

Keywords: Molecular Generation, Diffusion Model, Information Bottleneck

TL;DR: The superiority of scaffold-hopping–focused training is validated both theoretically and experimentally.

Abstract: The potential of three-dimensional molecular generation for structure-based drug discovery is hampered by the scarcity of public protein-ligand complexes, which causes models to overfit and fail to learn generalizable geometric priors. To address this challenge, we employ the PAC-Bayes information bottleneck framework to systematically quantify the information density of three generation paradigms: Scaffold Hopping (SH), Side-Chain Decoration (SC), and De Novo Design (DN). Our analysis reveals that SH possesses the highest information density, which tightens the model's generalization bound and enhances its transferability compared to conventional de novo generation. Motivated by this finding, we propose IBEX, a novel decoupled generation framework. IBEX is trained exclusively on the information-rich SH task to structure its latent representation of chemical space, which is then directly applied to de novo generation in a zero-shot transfer setting. Subsequently, a rapid physical refinement module utilizes the L-BFGS algorithm to optimize each conformer's geometry and binding compatibility by adjusting five short-range interaction terms and six degrees of freedom. Evaluated in a rigorous zero-shot setting on the CBGBench CrossDocked2020-based dataset, IBEX demonstrates substantial improvements over the TargetDiff baseline. It increases the docking success rate from 53% to 64% and improves the average Dock score from -7.41 to -8.09 kcal/mol. Notably, IBEX achieves a superior median Vina energy in 57 out of 100 binding pockets. Furthermore, IBEX enhances drug-likeness by approximately 25% while maintaining state-of-the-art validity and diversity, all corresponding to a demonstrably reduced generalization error. Our results validate that this decoupled approach, which synergizes information-dense pre-training with physical refinement, enables robust zero-shot structure generation and cross-pocket generalization in data-limited regimes.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 6127