Machine learning - Predicting Ames mutagenicity of small molecules.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Toxicology and risk assessment

Background:

• The Ames test, a traditional method for detecting compound mutagenicity, is expensive, time-consuming, and can yield inaccurate results.
• There is a critical need for robust in silico models to predict mutagenicity early in the drug discovery process, saving resources and improving accuracy.

Purpose of the Study:

• To develop and evaluate in silico models for predicting compound mutagenicity.
• To identify key molecular properties that contribute to mutagenicity prediction.
• To compare the performance of developed models against existing prediction tools.

Main Methods:

• Curated a compound mutagenicity library of over 5000 molecules.
• Calculated chemical fingerprints and molecular properties for each compound.
• Constructed 112 predictive models using 8 classification algorithms (e.g., SVM, RF, XGB).
• Assessed model performance using 10-fold cross-validation, a hold-out test set, and y-randomisation.

Main Results:

• Support Vector Machine (SVM) and Extreme Gradient Boosting (XGB) models demonstrated strong performance in cross-validation (AUROC >0.90) and on the test set (AUROC >0.65).
• Identified influential molecular properties that enhance prediction accuracy when combined with chemical fingerprints.
• Validated models showed superior performance in correctly predicting mutagens compared to StarDrop and TEST prediction tools.

Conclusions:

• Developed in silico models, particularly SVM and XGB, offer a reliable and efficient alternative to traditional mutagenicity testing.
• The identified molecular properties can guide the design of safer compounds in early drug discovery.
• These predictive models have the potential to significantly streamline the drug discovery pipeline by reducing the need for costly experimental testing.