Semantic knowledge improves molecular machine learning for chemical toxicity prediction

Area of Science:

• Computational toxicology
• Cheminformatics
• Bioinformatics

Background:

• Quantitative structure-activity relationship (QSAR) models predict chemical activity using molecular properties.
• Computational toxicology can benefit from integrating semantic knowledge, often structured as knowledge graphs, beyond molecular structure.
• Graph neural networks (GNNs) are increasingly used for learning molecular representations.

Purpose of the Study:

• To assess the impact of incorporating semantic knowledge into GNN-based QSAR models for improved toxicity prediction.
• To evaluate the performance of semantic-enhanced GNN models in binary classification of compound activity (active/inactive).
• To determine if semantic context enhances the interpretability and accuracy of toxicity predictions.

Main Methods:

• Utilized GNNs to learn molecular representations from structural and physico-chemical properties.
• Augmented GNNs with semantic knowledge from knowledge graphs, linking biomedical entities.
• Applied models to publicly available in silico toxicity datasets, including Tox21 assays and external endpoints, for binary classification tasks.

Main Results:

• Semantic-enhanced GNN models demonstrated improved discrimination in toxicity prediction compared to structure-only QSAR models.
• The enhanced models maintained calibration comparable to traditional structure-only QSAR models.
• Model explanations indicated that semantic knowledge guided the GNN encoder to focus on substructures with known toxicity effects.

Conclusions:

• Integrating mechanistic semantic context with molecular structure in GNN-based QSAR models leads to more accurate and interpretable toxicity predictions.
• This approach supports more robust chemical safety assessments by leveraging both structural and semantic information.
• Semantic knowledge integration offers a promising avenue for advancing computational toxicology.