Motif-Level Graph Learning Enables Interpretable Prediction of Drug-Induced QT Prolongation via Cooperative...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Pharmacology

Background:

• Drug-induced QT interval prolongation is a major safety concern in pharmaceutical development.
• Existing structure-based prediction methods lack sufficient substructural resolution and mechanistic interpretability.

Purpose of the Study:

• To develop a motif-level graph learning framework for accurate and interpretable prediction of QT interval prolongation risk.
• To provide a more chemically specific interpretation of QT liability beyond traditional structural alerts.

Main Methods:

• Decomposition of molecules into chemically meaningful motifs.
• Encoding motif features using a pre-trained chemical language model.
• Employing attention-based graph learning for inter-motif relationship modeling with cross-scale integration.

Main Results:

• Achieved strong and consistent predictive performance on regulatory (DIQTA) and pharmacovigilance (FAERS) datasets.
• Demonstrated robust generalization across different data sources.
• Identified specific motif organizations (cationic centers, heteroatom-rich scaffolds) associated with QT liability.

Conclusions:

• The motif-level graph learning framework offers a generalizable and interpretable approach for QT risk prediction.
• This strategy effectively models adverse drug reactions based on chemical structure at the motif level.
• Findings provide insights into hERG channel blockade determinants beyond isolated functional groups.