SepPCNET: Deeping Learning on a 3D Surface Electrostatic Potential Point Cloud for Enhanced Toxicity Classification...

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Deep learning (DL) in quantitative structure-activity relationship (QSAR) studies is revolutionizing toxicity prediction.
• Current DL-QSAR models primarily use 2D structural representations, limiting their predictive power.
• There is a need for advanced methods to capture complex 3D chemical structures for improved toxicity assessment.

Purpose of the Study:

• To develop a novel 3D molecular representation for toxicity prediction.
• To introduce a deep learning architecture capable of processing this 3D representation.
• To evaluate the model's performance against existing methods and its ability to elucidate toxicity mechanisms.

Main Methods:

• A novel 3D molecular surface point cloud (SepPC) was created, incorporating 3D coordinates and molecular electrostatic potential for each surface point.
• A new deep learning architecture, SepPCNET, was designed to directly process unordered SepPC data for toxicity classification.
• The SepPCNET model was trained and validated on 1317 chemicals from the ToxCast program, focusing on estrogen receptor-related assays.

Main Results:

• The SepPCNET model achieved high accuracies: 82.8% for active chemicals, 88.9% for inactive chemicals, 88.3% on the internal test set, and 92.5% on the external test set.
• The model demonstrated superior performance compared to other state-of-the-art machine learning models.
• SepPCNET successfully differentiated the activity of isomers and provided insights into toxicity mechanisms through visualization and feature extraction.

Conclusions:

• The proposed 3D SepPC representation and SepPCNET model significantly advance toxicity prediction capabilities.
• This novel approach offers a powerful tool for understanding chemical toxicity and mechanisms of action.
• The method holds promise for future drug discovery and chemical safety assessments.