hERG-toxicity prediction using traditional machine learning and advanced deep learning techniques.

Area of Science:

• Computational chemistry and toxicology
• Artificial intelligence in drug discovery

Background:

• Artificial intelligence (AI) is increasingly vital in pharmaceutical development.
• Predicting human ether-á-go-go related gene (hERG) channel blockage is crucial for drug safety, as hERG liability can lead to cardiotoxicity.

Purpose of the Study:

• To evaluate various machine learning and deep learning models for predicting hERG-derived toxicity.
• To identify the most effective AI approach for predicting hERG liability in new molecular structures.

Main Methods:

• Utilized traditional machine learning (Random Forest, SVM, XGBoost, DNN) and advanced deep learning (GRU-DNN, GNN) techniques.
• Trained and tested models on the largest hERG dataset to date, comprising 203,853 compounds for training and 87,366 for testing.
• Employed Graph Neural Network (GNN) with minimal feature engineering and human intervention.

Main Results:

• All evaluated models demonstrated strong performance, with AUC ROC scores ranging from 0.94 to 0.96.
• The Graph Neural Network (GNN) model achieved the highest predictive power and generalizability, with an AUC ROC score of 0.96.
• GNN models require minimal feature engineering, streamlining the prediction process.

Conclusions:

• Advanced AI techniques, particularly GNNs, are highly effective for predicting hERG toxicity.
• The GNN approach facilitates comprehensive automation in predictive toxicology, reducing the need for manual intervention.
• These AI models serve as valuable tools for academic institutions and pharmaceutical industries in assessing hERG liability early in drug development.