DeepHIT

DeepHIT predicts hERG-induced cardiotoxicity of small compounds by using deep learning to identify compounds that block the human ether-à-go-go-related gene (hERG) channel, a key determinant of QT interval prolongation.

Key Features:

• Gold-standard dataset: Trained on a large dataset comprising 6,632 hERG blockers and 7,808 hERG non-blockers.
• Model architecture: Employs three deep learning models that collectively prioritize sensitivity and negative predictive value (NPV) to reduce false-negative predictions.
• Performance on external test set: Reported overall accuracy 0.773, Matthews correlation coefficient (MCC) 0.476, sensitivity 0.833, and NPV 0.643.
• In silico chemical transformation module: Generates virtual compounds from a seed compound by applying known chemical transformation patterns.
• Proof-of-concept application: Identified novel urotensin II receptor (UT) antagonists derived from KR-36676 that do not inherit its hERG-blocking activity.

Scientific Applications:

• Early-stage cardiotoxicity screening: Prioritizes small molecules with reduced hERG-blocking risk during lead selection.
• Lead optimization via virtual chemistry: Generates chemically transformed virtual compounds to explore analogs with lower cardiotoxic potential.
• Risk reduction in drug development: Screens compounds to reduce the likelihood of late-stage failures due to hERG-mediated QT prolongation.
• Discovery of safer receptor antagonists: Supports identification of UT antagonists that avoid hERG blockade.

Methodology:

Models were trained on a dataset of 6,632 hERG blockers and 7,808 non-blockers using three deep learning models optimized for sensitivity and NPV, evaluated on an external test dataset (accuracy 0.773, MCC 0.476, sensitivity 0.833, NPV 0.643), and paired with an in silico chemical transformation module that generates virtual compounds from seed molecules using known transformation patterns.