Chemical Space Networks Enhance Toxicity Recognition via Graph Embedding.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Chemical Space Networks (CSNs) offer a novel approach to uncovering latent chemical patterns.
• CSNs can enhance the assessment of potential adverse health effects of chemicals.
• Existing methods may have limitations in comprehensively characterizing chemical toxicity.

Purpose of the Study:

• To embed Chemical Space Network structures into a metric space using graph neural networks.
• To improve the discrimination between toxic and non-toxic chemicals for various human health endpoints.
• To provide interpretable results for toxicity prediction using an explainable AI framework.

Main Methods:

• Utilizing molecular descriptors and fingerprints to construct CSNs.
• Applying graph neural networks to embed CSN structures into a metric space.
• Employing an eXplainable Artificial Intelligence (XAI) framework for result interpretation.

Main Results:

• Improved classification performance for eight different toxicological human health endpoints.
• An average increase of +12% in the area under the ROC curve (AUC) for predictive performance.
• Identification of putative structural alerts associated with specific toxicities.

Conclusions:

• The proposed method enhances the prediction of chemical toxicity by leveraging CSN embeddings.
• This approach represents a significant advancement in alternative methods for chemical safety assessment.
• The findings could drive innovation in the design of safer chemicals and pharmaceuticals.