Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Substructure-Aware Expert Model for Molecular Property Prediction

Tianyi Jiang, Zeyu Wang, Shanqing Yu, Qi Xuan

TL;DR

ASE-Mol presents a graph neural network framework that uses BRICS-based substructure decomposition and a substructure-aware Mixture-of-Experts to distinguish positive and negative motifs for molecular property prediction. By routing motif embeddings through specialized experts via two routers, the method mitigates negative motifs and enhances adaptability, achieving state-of-the-art results across eight MoleculeNet datasets. The approach also provides interpretability through substructure attribution and router visualizations, linking key BRICS fragments to predicted properties. This work advances practical molecular property prediction with improved accuracy and clearer chemical insight, though it currently targets classification tasks and could be extended to regression in future work.

Abstract

Molecular property prediction is essential for applications such as drug discovery and toxicity assessment. While Graph Neural Networks (GNNs) have shown promising results by modeling molecules as molecular graphs, their reliance on data-driven learning limits their ability to generalize, particularly in the presence of data imbalance and diverse molecular substructures. Existing methods often overlook the varying contributions of different substructures to molecular properties, treating them uniformly. To address these challenges, we propose ASE-Mol, a novel GNN-based framework that leverages a Mixture-of-Experts (MoE) approach for molecular property prediction. ASE-Mol incorporates BRICS decomposition and significant substructure awareness to dynamically identify positive and negative substructures. By integrating a MoE architecture, it reduces the adverse impact of negative motifs while improving adaptability to positive motifs. Experimental results on eight benchmark datasets demonstrate that ASE-Mol achieves state-of-the-art performance, with significant improvements in both accuracy and interpretability.

Adaptive Substructure-Aware Expert Model for Molecular Property Prediction

TL;DR

ASE-Mol presents a graph neural network framework that uses BRICS-based substructure decomposition and a substructure-aware Mixture-of-Experts to distinguish positive and negative motifs for molecular property prediction. By routing motif embeddings through specialized experts via two routers, the method mitigates negative motifs and enhances adaptability, achieving state-of-the-art results across eight MoleculeNet datasets. The approach also provides interpretability through substructure attribution and router visualizations, linking key BRICS fragments to predicted properties. This work advances practical molecular property prediction with improved accuracy and clearer chemical insight, though it currently targets classification tasks and could be extended to regression in future work.

Abstract

Molecular property prediction is essential for applications such as drug discovery and toxicity assessment. While Graph Neural Networks (GNNs) have shown promising results by modeling molecules as molecular graphs, their reliance on data-driven learning limits their ability to generalize, particularly in the presence of data imbalance and diverse molecular substructures. Existing methods often overlook the varying contributions of different substructures to molecular properties, treating them uniformly. To address these challenges, we propose ASE-Mol, a novel GNN-based framework that leverages a Mixture-of-Experts (MoE) approach for molecular property prediction. ASE-Mol incorporates BRICS decomposition and significant substructure awareness to dynamically identify positive and negative substructures. By integrating a MoE architecture, it reduces the adverse impact of negative motifs while improving adaptability to positive motifs. Experimental results on eight benchmark datasets demonstrate that ASE-Mol achieves state-of-the-art performance, with significant improvements in both accuracy and interpretability.

Paper Structure

This paper contains 29 sections, 17 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Molecular Property Prediction
  4. Expert Models
  5. Preliminaries
  6. Problem Definition
  7. Graph Neural Networks
  8. Mixture of Experts
  9. Methods
  10. Molecular Substructure Recognition and Learning
  11. BRICS Recognition
  12. Learning Optimization
  13. Substructure-based MoE Module
  14. Multiple Experts Representation
  15. Substructure-based Expert Routing
  16. ...and 14 more sections

Figures (5)

  • Figure 1: The T-SNE visualization on different substructures.
  • Figure 2: Overview of the proposed ASE-Mol.
  • Figure 3: The substructure attribution visualization on the BBBP dataset.
  • Figure 4: T-SNE visualization with the router assignment on the BACE dataset.
  • Figure 5: The hyper-parameter sensitivity analysis on the ClinTox and BBBP dataset.