Predicting hERG channel blockers with directed message passing neural networks

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacology and toxicology
• Machine learning in drug discovery

Background:

• Compounds inhibiting the human ether-à-go-go related gene (hERG) channel can lead to severe cardiotoxicity.
• Early identification of hERG liability is essential for safe drug development.
• In silico prediction models are actively researched to assess hERG blockade potential.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting hERG channel blockers.
• To identify optimal molecular descriptors and model architectures for hERG liability assessment.
• To compare the performance of the developed model against existing methods.

Main Methods:

• Application of the directed message passing neural network (D-MPNN) algorithm.
• Utilizing diverse datasets for training and validation of classification models.
• Testing various molecular descriptors and fingerprints, including MOE-generated moe206 descriptors.
• Performance evaluation using Area Under the Receiver Operating Characteristic Curve (AUC-ROC) with random and scaffold splits.

Main Results:

• The D-MPNN model combined with moe206 descriptors (D-MPNN + moe206) demonstrated superior performance.
• Achieved high AUC-ROC values of 0.956 ± 0.005 (random split) and 0.922 ± 0.015 (scaffold split) on Cai's hERG dataset.
• Comparative analysis showed the D-MPNN + moe206 model as one of the best-performing classification models for hERG blockers.

Conclusions:

• The D-MPNN + moe206 model is a highly effective tool for predicting hERG channel blockers.
• This model shows significant potential for application in early-stage drug discovery to mitigate cardiotoxicity risks.
• The study underscores the power of advanced machine learning techniques in identifying potential drug safety liabilities.