Optimizing machine-learning models for mutagenicity prediction through better feature selection.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning in drug discovery

Background:

• Traditional mutagenicity tests are costly and time-consuming.
• In silico methods offer a more efficient alternative for predicting compound mutagenicity.
• Machine learning (ML) models show promise for improving mutagenicity prediction accuracy.

Purpose of the Study:

• To investigate the impact of feature selection on ML model accuracy for mutagenicity prediction.
• To identify novel structural and molecular features that maximize predictive performance.
• To evaluate the effectiveness of combining diverse feature sets for enhanced accuracy.

Main Methods:

• Systematic evaluation of novel structural and molecular features.
• Training and rigorous assessment of multiple classification models, including deep neural networks.
• Performance evaluation using 5- and 10-fold cross-validation on the Hansen et al. benchmark dataset.

Main Results:

• The proposed approach, utilizing molecule structure, molecular properties, and structural alerts as features, achieved an area under the receiver operating characteristic curve of 0.93.
• This performance surpasses state-of-the-art methods for mutagenicity prediction on the benchmark dataset.
• Combining different feature sets demonstrably benefits model accuracy improvements.

Conclusions:

• Optimized feature selection is crucial for enhancing the accuracy of machine learning models in mutagenicity prediction.
• The developed framework provides a robust method for improving in silico mutagenicity assessments.
• This approach has significant implications for accelerating drug discovery and development by enabling more reliable early-stage safety evaluations.