Stacked Generalization with Applicability Domain Outperforms Simple QSAR on in Vitro Toxicological Data.

Area of Science:

• Computational toxicology
• cheminformatics
• Machine learning in drug discovery

Background:

• In silico tools are crucial for early toxicity assessment of chemical substances.
• The ToxCast program provides extensive in vitro bioactivity data for computational model development.
• Predicting bioactivity and toxicity requires robust machine learning approaches.

Purpose of the Study:

• To evaluate the machine learning capabilities of ToxCast data.
• To build predictive models for in vitro bioactivities.
• To explore advanced ensemble methods for improved model performance.

Main Methods:

• Large-scale analysis of ToxCast in vitro bioactivity data.
• Application of classical quantitative structure-activity relationship (QSAR) algorithms (ANN, SVM, LDA, random forest, Bayesian).
• Implementation and evaluation of the Stacked generalization ensemble method.
• Integration of an applicability domain filter for prediction reliability assessment.

Main Results:

• Classical QSAR algorithms showed limitations with datasets containing many inactive compounds.
• Stacked generalization significantly improved model performance for 61% of 483 models tested.
• Combining Stacked generalization with an applicability domain filter enhanced prediction reliability, improving ROC scores for 50% of models.

Conclusions:

• Ensemble methods, particularly Stacked generalization, offer superior performance for predicting bioactivity from ToxCast data, especially with inactive compounds.
• The integration of an applicability domain filter enhances the reliability of in silico predictions, aiding in compound prioritization for toxicity assessment.
• These findings support the development of more accurate and reliable in silico tools for chemical safety evaluation.