Mechanistic Task Groupings Enhance Multitask Deep Learning of Strain-Specific Ames Mutagenicity

Area of Science:

• Computational toxicology
• Quantitative structure-activity relationship (QSAR) modeling
• Deep learning in cheminformatics

Background:

• The Ames test is a standard assay for detecting chemical mutagens.
• Multitask deep learning can enhance QSAR model performance by jointly training related tasks.
• Integrating domain knowledge into multitask learning can optimize predictive accuracy.

Purpose of the Study:

• To investigate the impact of toxicology-informed task groupings on multitask deep learning for Ames mutagenicity prediction.
• To compare the performance of grouped versus ungrouped multitask models against single-task controls.

Main Methods:

• Utilized 16 Salmonella typhimurium strain tasks from the Ames test.
• Developed multitask neural networks, both with and without task groupings based on mechanistic correlations.
• Employed correlation data analysis to inform task grouping strategies.

Main Results:

• Both grouped and ungrouped multitask models outperformed single-task controls in predicting Ames mutagenicity.
• Grouped multitask models consistently showed incremental performance gains over ungrouped models.
• Mechanistic task groupings enhanced synergistic training signals.

Conclusions:

• Multitask learning provides a significant performance boost for Ames mutagenicity prediction.
• Toxicology domain knowledge, used for task grouping, further refines multitask QSAR models.
• This approach leads to more transparent and accurate mutagenicity predictions.