ADMET予測のための解釈可能な分子表現学習手法の研究 Interpretable Molecular Representation Learning for Accurate ADMET Property Prediction

本研究は、コンピュータ支援創薬において重要な ADMET（吸収・分布・代謝・排泄・毒性）特性の高精度予測を目的とし、分子表現学習手法の高度化に取り組むものである。従来の有向メッセージパッシングニューラルネットワーク（D-MPNN）が抱える、特徴量の重み付けの不均衡、分子構造に対する感度不足、ならびに重要な部分構造の識別能力の制約といった課題を体系的に解決することを目指す。

本研究では、注意機構を導入したメッセージ伝播手法、トポロジカルな位置情報を考慮した原子表現、さらに群知能最適化に基づく構造モチーフ探索を統合することで、分子構造情報の精緻な表現と機能的に重要な部分構造の可視化を実現する。これにより、特に毒性および代謝予測といった複雑な ADMET タスクにおいて予測性能とモデルの解釈性を大きく向上させることが期待される。

This research focuses on the accurate prediction of ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties, which is a critical challenge in computer-aided drug design. Aiming to improve both predictive performance and model interpretability, this study investigates advanced molecular representation learning methods based on directed message passing neural networks (D-MPNNs). Specifically, it addresses three fundamental limitations of conventional D-MPNNs: uneven feature weight allocation, insufficient sensitivity to molecular structural variations, and limited capability in identifying key functional substructures.

To overcome these challenges, the proposed framework integrates an attention-guided message passing mechanism, topology-aware atomic representations that encode structural location information, and structural motif mining based on swarm intelligence optimization. By jointly modeling fine-grained molecular structural features and identifying critical substructures relevant to ADMET properties, the framework significantly enhances prediction accuracy, particularly for complex tasks such as toxicity and metabolism prediction. This research not only advances interpretable ADMET prediction methodologies but also introduces a novel molecular representation learning paradigm that holds strong potential for accelerating drug discovery and reducing development costs.

Reference:

1. S. Lin, G. Cui, C. Tang, C. Zhang, Y. Nagata and H. Yang, "ATLAS-DMPNN: An Attention-Guided Topological Framework for Enhanced ADMET Property Prediction," 2025 International Conference on New Trends in Computational Intelligence (NTCI), Jinan, China, 2025, pp. 142-147, doi: 10.1109/NTCI67886.2025.11308181.