Large-Scale Modeling of Multispecies Acute Toxicity End Points Using Consensus of Multitask Deep Learning Methods.

Area of Science:

• Computational toxicology and cheminformatics.
• Drug discovery and development.
• Environmental chemical safety assessment.

Background:

• Computational methods accelerate safety and toxicology predictions, reducing drug development time and costs.
• There is a significant need for reliable computational toxicity prediction models.
• Deep neural networks and other machine learning approaches are being explored for toxicity prediction.

Purpose of the Study:

• To curate and integrate public data for acute systemic toxicity.
• To develop and evaluate various computational models for toxicity prediction.
• To create the largest publicly available dataset for acute systemic toxicity.

Main Methods:

• Data collection and curation from public sources for the ChemIDplus database.
• Development of single- and multitask models using random forest, deep neural networks, and graph convolutional neural networks.
• Creation and evaluation of consensus models based on multitask learning approaches.

Main Results:

• Generated the largest public dataset (>80,000 compounds) for 59 acute systemic toxicity endpoints.
• Developed novel prediction models for 36 previously unpublished endpoints.
• Achieved superior performance with a consensus model, particularly for smaller tasks (<300 compounds).

Conclusions:

• The developed computational models and curated dataset significantly advance toxicity prediction capabilities.
• Publicly available resources support regulatory and research applications in chemical safety.
• Multitask learning and consensus modeling show promise for improving toxicity predictions.