Modelling In vitro Mutagenicity Using Multi-Task Deep Learning and REACH Data.

Area of Science:

• Computational toxicology
• Chemical safety assessment
• In vitro toxicology

Background:

• Mutagenicity assessment under REACH regulation relies on a tiered approach of in vitro and in vivo testing.
• Existing in vitro assays for mutagenicity assessment include bacterial gene mutation tests and mammalian cell assays.
• Exploring correlations between in vitro assays can potentially improve predictive model performance.

Purpose of the Study:

• To investigate the use of multi-task deep learning models for predicting chemical mutagenicity based on in vitro assay data.
• To compare the performance of multi-task deep learning models against single-task models and classical machine learning methods.
• To assess the generalizability of developed models using extensive external test sets.

Main Methods:

• Compiled a large genotoxicity dataset (>12,000 substances) from REACH, ToxValDB, and literature.
• Developed and evaluated various single-task and multi-task deep learning models, including graph neural networks.
• Utilized classical machine learning techniques and chemical fingerprints for comparison.
• Constructed external test sets for rigorous model validation.

Main Results:

• Deep learning single-task models achieved 73-84% balanced accuracy in cross-validation for in vitro assays, outperforming classical methods by 2-8%.
• Specific bacterial gene mutation and metabolic activation models showed 82-85% balanced accuracy, with 7-12% improvement.
• Multi-task models demonstrated an average 8% higher cross-validation accuracy than single-task models for specific assays.
• External validation showed 72-78% balanced accuracy for best models with sufficient data.
• Graph neural network embeddings identified structural alerts and correlated structural moieties with genotoxicity outcomes.

Conclusions:

• Multi-task deep learning models show promise for improving the accuracy and efficiency of in vitro mutagenicity assessments.
• The developed models can predict genotoxicity and identify structure-activity relationships, aiding in chemical safety evaluations.
• These computational approaches offer a valuable complement to traditional testing strategies under regulatory frameworks like REACH.