Capsule Networks Showed Excellent Performance in the Classification of hERG Blockers/Nonblockers.

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Capsule Networks (CapsNets) are advanced deep learning models with demonstrated success in image and natural language processing.
• Their application in drug discovery, particularly for predicting cardiotoxicity risks associated with hERG channel blockade, remains unexplored.

Purpose of the Study:

• To investigate the efficacy of CapsNets for classifying hERG blockers and nonblockers, identifying potential cardiotoxic drug candidates.
• To develop and evaluate novel CapsNet architectures for drug safety prediction.

Main Methods:

• Two CapsNet architectures were developed: Convolution-CapsNet (Conv-CapsNet) and Restricted Boltzmann Machine-CapsNet (RBM-CapsNet), utilizing convolution and RBMs as feature extractors, respectively.
• These models were trained on Doddareddy's dataset (2,389 compounds) and validated on an independent test set (255 compounds).

Main Results:

• The Conv-CapsNet and RBM-CapsNet models achieved high prediction accuracies of 91.8% and 92.2%, respectively, on the independent test set.
• Performance was compared against various established machine learning methods (DBN, CNN, MLP, SVM, kNN, LR, LightGBM), demonstrating superior or competitive results.

Conclusions:

• Capsule Networks, specifically Conv-CapsNet and RBM-CapsNet, exhibit significant potential as powerful tools for drug discovery, particularly in predicting hERG blockade and assessing cardiotoxicity risks.
• The study highlights CapsNets as a promising avenue for developing robust drug safety prediction models.