Augmenting Expert Knowledge-Based Toxicity Alerts by Statistically Mined Molecular Fragments.

Area of Science:

• Computational toxicology and cheminformatics.
• Development of predictive models for chemical safety assessment.

Background:

• Expert knowledge-based structural alerts in in silico toxicology often lack predictivity and coverage.
• Existing methods struggle with specificity and satisfactory generalization for diverse toxic effects.

Purpose of the Study:

• To develop and evaluate hybrid quantitative structure-activity relationship (QSAR) models.
• To assess if combining expert alerts with statistically mined fragments improves predictive performance.
• To identify features that activate or mitigate toxicity and discover novel alerts.

Main Methods:

• Hybrid QSAR models were constructed by integrating expert knowledge-based alerts and statistically mined molecular fragments.
• Lasso regularization was employed for variable selection, allowing elimination only from molecular fragments.
• The approach was validated on three distinct toxicity endpoints: skin sensitization, acute Daphnia toxicity, and Ames mutagenicity.

Main Results:

• Hybrid models significantly outperformed models based solely on expert alerts or statistically mined fragments.
• The method successfully identified features influencing toxicity activation and mitigation.
• New structural alerts were discovered, enhancing model accuracy and reducing false outcomes.

Conclusions:

• Combining expert knowledge with data-driven fragments creates more predictive and robust QSAR models.
• This hybrid approach offers a promising strategy to enhance the reliability of in silico toxicology assessments.
• The methodology aids in refining existing alerts and discovering novel ones, improving prediction accuracy.