DeepHIT: a deep learning framework for prediction of hERG-induced cardiotoxicity

Area of Science:

• Computational chemistry
• Cardiovascular pharmacology
• Drug discovery and development

Background:

• Blockade of the human ether-à-go-go-related gene (hERG) channel by small compounds can cause severe cardiotoxicity, leading to drug development failures.
• Accurate evaluation of hERG-blocking activity is crucial for safe and successful drug development.
• Existing computational prediction tools require improved sensitivity and negative predictive value (NPV) to minimize false negatives.

Purpose of the Study:

• To develop a computational framework, DeepHIT, for predicting hERG blockers and non-blockers.
• To enhance the prediction accuracy, sensitivity, and NPV of hERG channel activity assessment.
• To reduce false negative predictions in the early stages of drug discovery.

Main Methods:

• Development of DeepHIT, a computational framework utilizing three deep learning models.
• Generation of a large-scale gold-standard dataset comprising 6632 hERG blockers and 7808 hERG non-blockers.
• Implementation of an in silico chemical transformation module to generate virtual compounds.

Main Results:

• DeepHIT demonstrated superior performance on an external test dataset with an accuracy of 0.773, MCC of 0.476, sensitivity of 0.833, and NPV of 0.643.
• The framework successfully identified novel urotensin II receptor antagonists without hERG-blocking activity.
• DeepHIT outperforms existing computational tools in key predictive metrics.

Conclusions:

• DeepHIT serves as a valuable tool for predicting hERG-induced cardiotoxicity in small compounds during early drug discovery.
• The framework's improved predictive performance can aid in mitigating drug development failures due to cardiotoxicity.
• DeepHIT facilitates the identification of safer drug candidates by assessing hERG channel interactions.