Comparison of Descriptor- and Fingerprint Sets in Machine Learning Models for ADME-Tox Targets.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Toxicology and pharmacology

Background:

• Accurate prediction of ADME-Tox properties is crucial for efficient drug design.
• Quantitative structure-property relationship (QSPR) models offer a faster alternative to experimental screening.
• Model performance is significantly influenced by the choice of molecular descriptors.

Purpose of the Study:

• To compare the effectiveness of various molecular descriptor groups for predicting six key ADME-Tox targets.
• To evaluate the performance of XGBoost and RPropMLP algorithms using different molecular representations.
• To identify the optimal descriptor set for QSPR modeling in ADME-Tox prediction.

Main Methods:

• Utilized literature-based datasets for Ames mutagenicity, P-glycoprotein inhibition, hERG inhibition, hepatotoxicity, blood-brain-barrier permeability, and CYP2C9 inhibition.
• Employed XGBoost and RPropMLP algorithms for binary classification model building.
• Compared Morgan, Atompairs, MACCS fingerprints, and 1D, 2D, 3D molecular descriptors, including their combinations.

Main Results:

• All QSPR models achieved performance comparable to typical benchmarks for ADME-Tox targets.
• Traditional 1D, 2D, and 3D molecular descriptors demonstrated superior performance with the XGBoost algorithm.
• 2D descriptors alone provided better predictive models for most datasets than combined descriptor sets.

Conclusions:

• Traditional molecular descriptors, particularly 2D, are highly effective for QSPR modeling of ADME-Tox properties.
• The XGBoost algorithm benefits significantly from traditional descriptor sets.
• Simpler descriptor sets, like 2D descriptors, can yield superior results compared to complex combinations.