Multiple Instance Learning Improves Ames Mutagenicity Prediction for Problematic Molecular Species.

Area of Science:

• Toxicology
• Computational Chemistry
• Drug Discovery

Background:

• Ames mutagenicity prediction is crucial for regulatory and pharmacological toxicology.
• Traditional QSAR models struggle with compounds requiring metabolic activation.
• Metabolites formed during activation lack individual mutagenicity labels, hindering traditional modeling.

Purpose of the Study:

• To explore the utility of Multiple Instance Learning (MIL) for predicting Ames mutagenicity.
• To develop and evaluate MIL models capable of handling metabolically activated compounds.
• To assess MIL's performance on challenging chemical datasets.

Main Methods:

• Trained MIL models on Ames mutagenicity data, starting with aromatic amines (n=457) and expanding to a larger dataset (n=6505).
• Utilized MIL to group parent molecules and their potential metabolites under a single mutagenicity label.
• Compared MIL model performance against established prediction models.

Main Results:

• MIL models achieved a balanced accuracy of 0.778, comparable to existing methods.
• MIL demonstrated strong predictive performance on previously identified hard-to-predict chemical groups.
• The improved accuracy is attributed to MIL's consideration of metabolic activation pathways.

Conclusions:

• MIL offers a viable approach for Ames mutagenicity prediction, particularly for metabolically activated compounds.
• MIL can supplement existing models, enhancing predictions where implicit metabolite modeling is insufficient.
• This study highlights MIL's potential for improving toxicological risk assessment of xenobiotics.