ToxACoL: an endpoint-aware and task-focused compound representation learning paradigm for acute toxicity assessment

Area of Science:

• Toxicology
• Computational Chemistry
• Machine Learning

Background:

• Multi-species acute toxicity assessment is crucial for chemical safety, classification, and risk management.
• Current deep learning models face challenges with diverse experimental conditions, imbalanced datasets, and limited target data, impacting prediction accuracy for data-scarce endpoints.
• Accurate prediction of chemical toxicity is essential for regulatory compliance and human health protection.

Purpose of the Study:

• To develop a novel machine learning paradigm, Adjoint Correlation Learning (ToxACoL), for robust multi-condition acute toxicity assessment.
• To address limitations of existing methods in handling diverse experimental conditions and scarce data.
• To improve the prediction of data-scarce toxicological endpoints and facilitate knowledge transfer across conditions.

Main Methods:

• Adjoint Correlation Learning (ToxACoL) models endpoint associations using graph topology and knowledge transfer via graph convolution.
• The adjoint correlation mechanism synchronously encodes compounds and endpoints, generating endpoint-aware and task-focused representations.
• ToxACoL was evaluated for its performance in multi-condition acute toxicity prediction, including its ability to handle data scarcity.

Main Results:

• ToxACoL demonstrated significant improvements, achieving 43%-87% enhancement for data-scarce human endpoints.
• The model achieved substantial reductions in training data requirements, needing 70% to 80% less data.
• Visualization of learned representations provided insights into structural alert mechanisms and potential extrapolation of animal to human toxicity.

Conclusions:

• ToxACoL offers a powerful and data-efficient solution for multi-condition acute toxicity assessment.
• The developed platform can aid in predicting chemical toxicities, potentially bridging the gap between animal and human data.
• This approach enhances the accuracy and reliability of chemical safety evaluations, particularly for endpoints with limited available data.