Binding Mode Matters: Residue-Guided Drug Discovery via Explorative Preferences

Dingshuo Chen, Hao Yang, Kuangqi Zhou, Qiang Liu, xiangxiang Zeng, Shu Wu, Yaosen Min, Liang Wang

Keywords: fragment-based drug discovery, reinforcement learning

Abstract:

The discovery of novel hit or lead molecules requires navigating a vast chemical space to identify compounds with optimal binding modes, which are typically unknown beforehand. Despite various generative approaches, they have predominantly relied on optimizing a monolithic scalar docking score to guide generation, masking the distinct contributions of key binding determinants.
In this work, we introduce a paradigm shift by formulating target-based drug design as a multi-objective exploration task, where each objective explicitly corresponds to enhancing interactions with a specific key residue. To this end, we introduce **BindMol**, a novel generative framework that integrates a fragment-based generator with a customized multi-objective reinforcement learning algorithm. By incorporating explorative preferences during training, our approach efficiently uncovers molecules with distinct and desirable binding profiles.
Empirical evaluations demonstrate that **BindMol** facilitates the discovery of structurally novel, high-affinity compounds across five protein targets and establishes new state-of-the-art records on the multi-property optimization tasks in GuacaMol benchmarks, thereby providing a versatile paradigm for goal-directed drug discovery.

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

Submission Number: 3056