It's difficult, but important, to make negative predictions.

Area of Science:

• Computational toxicology
• Predictive modeling
• Chemical safety assessment

Background:

• Negative predictions in silico are vital for preventing population exposure to harmful substances.
• Ensuring confidence and robustness in these predictions is a key challenge in regulatory toxicology.

Purpose of the Study:

• To evaluate two novel computational approaches for making reliable negative in silico predictions of mutagenicity.
• To assess the performance of these methods across diverse chemical datasets.

Main Methods:

• Analysis of 12 datasets comprising over 13,000 compounds.
• Evaluation of computational models for predicting mutagenicity (Ames test).
• Identification of features impacting prediction accuracy and certainty.

Main Results:

• The best computational approach demonstrated high negative predictivity (approximately 90%).
• Features reducing prediction accuracy or certainty were identified as misclassified or unclassified compounds.
• Negative predictivity remained high even with these features, suggesting they do not indicate mutagenicity.

Conclusions:

• Novel in silico methods provide confident and robust negative predictions for mutagenicity.
• These approaches support regulatory decision-making and chemical safety evaluations.
• Identified features do not serve as reliable flags for mutagenicity, reinforcing the models' negative predictive power.