GraphDeep-hERG: Graph Neural Network PharmacoAnalytics for Assessing hERG-Related Cardiotoxicity.

Area of Science:

• Computational chemistry and pharmacology
• Artificial intelligence in drug discovery
• Cardiovascular safety pharmacology

Background:

• The human Ether-a-go-go Related-Gene (hERG) channel is crucial for cardiac repolarization.
• hERG channel blockade by drugs can cause lethal arrhythmias like long QT syndrome.
• Current drug screening methods for hERG inhibition are inefficient and time-consuming.

Purpose of the Study:

• To develop an automated method for learning molecular representations to improve in silico hERG screening.
• To overcome limitations of traditional models relying on manually defined atomic features.
• To accelerate the identification of potential hERG inhibitors for drug safety.

Main Methods:

• Developed a deep neural network (DNN) model for automated atom embedding using 118,312 compounds from ZINC.
• Trained a Graph Neural Network (GNN) model using 7,909 ChEMBL compounds for classification.
• Integrated the atom embedding and GNN models into a classifier to distinguish hERG inhibitors from non-inhibitors.

Main Results:

• The automated atom embedding model achieved 0.93 accuracy in structural representation.
• The best performing GNN model reached 0.84 accuracy in predicting hERG inhibition.
• The GNN model outperformed traditional machine learning and existing AI-driven models in external validation.

Conclusions:

• The automated atom embedding model provides a robust standard for molecular representations.
• Integrating this model with GNNs significantly aids in screening hERG inhibitors.
• This approach accelerates drug discovery and repurposing by enhancing computational safety assessments.