MMRCL: An interpretable multi-modal deep learning framework for predicting hERG blockers.

Area of Science:

• Computational chemistry
• Cardiovascular pharmacology
• Drug discovery

Background:

• The human ether-a-go-go-related gene (hERG) encodes a potassium channel critical for cardiac repolarization.
• Inhibition of hERG channels by drugs can lead to QT interval prolongation, torsade de pointes, and fatal arrhythmias.
• Early identification of hERG blockers is vital in pharmaceutical development to prevent cardiotoxicity, reduce drug attrition, and minimize economic losses.

Purpose of the Study:

• To develop an interpretable multi-modal molecular representation cross-learning framework (MMRCL) for accurate prediction of hERG channel blockers.
• To integrate diverse molecular features, including fingerprints and graphs, for enhanced predictive power.
• To provide actionable insights for medicinal chemists through model interpretability.

Main Methods:

• Developed MMRCL, integrating multi-dimensional molecular fingerprints and molecular graphs.
• Employed a dual-channel message passing neural network (MPNN) for atom- and bond-level features and a multi-layer perceptron for fingerprint semantics.
• Utilized a multi-head cross-attention mechanism for adaptive feature fusion and a fully connected neural network for classification.

Main Results:

• MMRCL demonstrated superior performance over seven state-of-the-art models on internal and external datasets.
• Achieved high performance metrics: AUC of 0.8895, PRC of 0.9073, and MCC of 0.6146 on the internal dataset.
• Interpretability analysis identified key toxic substructures associated with hERG-blocking activity.

Conclusions:

• MMRCL offers superior prediction accuracy and generalization for identifying hERG blockers.
• The framework enhances model interpretability, aiding structure-activity relationship studies.
• MMRCL provides valuable insights for medicinal chemists to mitigate cardiotoxicity risks in drug discovery.