Expert-Guided Substructure Information Bottleneck for Molecular Property Prediction
- 23 July 2025
- journal article
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
Abstract
Molecular property prediction plays a crucial role in cheminformatics, yet existing methods are constrained by data scarcity and molecular structural heterogeneity. The Mixture of Experts (MoE) framework adopts a divide-and-conquer approach by partitioning the input space and employing expert models. However, current methods primarily rely on scaffold or atomic-level information, often neglecting fine-grained features such as functional groups. Moreover, existing MoE models lack effective mechanisms to filter redundant and noisy information, limiting prediction accuracy and generalization. To address these challenges, we propose a novel Expert-Guided Substructure Information Bottleneck (ESIB-Mol) framework that integrates MoE learning with the Information Bottleneck (IB) principle to optimize molecular representation learning. ESIB-Mol employs substructure-specific experts to focus on key molecular scaffolds and functional groups, which play a crucial role in determining molecular properties such as bioactivity and pharmacokinetics. Meanwhile, the IB principle is leveraged to filter out redundant and irrelevant information, thereby enhancing prediction accuracy and interpretability. Additionally, a dynamic gating mechanism adaptively assigns molecules to the most relevant expert, optimizing computational efficiency. Extensive experiments on benchmark data sets demonstrate the effectiveness of ESIB-Mol, highlighting its superior performance in molecular property prediction.Keywords
Funding Information
- Key Research and Development Program of Zhejiang Province (2022C01018, 2024C01025)
- National Natural Science Foundation of China (U21B2001)
- Natural Science Foundation of Zhejiang Province (LBMHZ25F020002, ZCLY24F0302)
- Key Technology Research and Development Program Project of Hangzhou (2024SZD1A23)
This publication has 36 references indexed in Scilit:
- MOL-MOE: Learning Drug Molecular Characterization Based on Mixture of Expert Mechanism2024
- Machine learning small molecule properties in drug discoveryArtificial Intelligence Chemistry, 2023
- HiGNN: A Hierarchical Informative Graph Neural Network for Molecular Property Prediction Equipped with Feature-Wise AttentionJournal of Chemical Information and Modeling, 2022
- Deep learning methods for molecular representation and property predictionDrug Discovery Today, 2022
- Machine learning directed drug formulation developmentAdvanced Drug Delivery Reviews, 2021
- A compact review of molecular property prediction with graph neural networksDrug Discovery Today: Technologies, 2020
- The Most Common Functional Groups in Bioactive Molecules and How Their Popularity Has Evolved over TimeJournal of Medicinal Chemistry, 2020
- Molecular property prediction: recent trends in the era of artificial intelligenceDrug Discovery Today: Technologies, 2019
- Ensemble Methods in Machine Learning2000
- The Generation of a Unique Machine Description for Chemical Structures-A Technique Developed at Chemical Abstracts Service.Journal of Chemical Documentation, 1965