Developing novel in silico prediction models for assessing chemical reproductive toxicity using the naïve Bayes...

Area of Science:

• Computational toxicology
• Drug discovery and development
• cheminformatics

Background:

• Assessing reproductive toxicity is crucial for drug safety.
• Existing methods can be time-consuming and resource-intensive.
• Predictive models can accelerate safety evaluations.

Purpose of the Study:

• To develop a robust computational model for predicting chemical reproductive toxicity.
• To identify key molecular features associated with reproductive toxicity.
• To facilitate early-stage screening of drug candidates.

Main Methods:

• Applied the Naïve Bayes (NB) classifier to build binary classification models.
• Utilized a genetic algorithm to select six significant molecular descriptors.
• Developed 110 models using various fingerprints and maximum diameters, focusing on the LCFC_20 fingerprint.
• Validated models on training, external Test Set I, and a rat multi-generation reproductive toxicity dataset (Test Set II).

Main Results:

• The NB-1 model, using six descriptors and LCFC_20 fingerprints, achieved the best performance.
• Achieved a 0.884 ROC score and 91.8% accuracy on the Training Set.
• Obtained 0.888 ROC and 83.0% accuracy on Test Set I.
• Showed 0.806 ROC and 85.1% concordance on the rat multi-generation dataset (Test Set II).

Conclusions:

• The NB-1 model provides reliable predictions for chemical reproductive toxicity.
• The model can effectively filter early-stage molecules for potential reproductive adverse effects.
• Identified key molecular descriptors and structural alerts to guide drug optimization for safety.