In-silico predictive mutagenicity model generation using supervised learning approaches.

Area of Science:

• Computational chemistry and cheminformatics
• Toxicology and drug safety assessment
• Machine learning applications in bioinformatics

Background:

• Experimental screening for chemical biological activity is costly and time-consuming.
• In silico predictive models offer rapid virtual screening to prioritize compounds.
• Previous mutagenicity prediction focused on toxicophores, not whole-molecule models.

Purpose of the Study:

• To evaluate various in silico classification models for predicting compound mutagenicity.
• To apply machine learning methods like J48 and SMO in cheminformatics for mutagenicity prediction.
• To develop and validate accurate predictive models for mutagenic properties.

Main Methods:

• Utilized Bursi (Set 1) and Benchmark (Set 2) mutagenicity datasets, combined into Set 3.
• Employed classification algorithms: Naïve Bayes, Random Forest, J48, and SMO with 10-fold cross-validation.
• Validated models on external datasets (AID1189, AID1194) and applied to DrugBank and Zinc databases.

Main Results:

• The combined dataset (Set 3) yielded better model performance than Set 2 alone.
• Random Forest achieved the highest accuracy (89.27%), precision (89%), and ROC (95.3%) on Set 3.
• External validation showed robust performance, with one test achieving 91% accuracy, 91% precision, and 96.3% ROC.

Conclusions:

• A novel mutagenicity benchmark dataset of approximately 8,000 compounds was created.
• Highly accurate machine learning models for mutagenicity prediction were developed, complementing toxicophore-based methods.
• Random Forest models consistently outperformed other machine learning methods for this predictive task.