Toward Metabolic Similarity in Read-Across: A Case Study Using Graph Convolutional Networks to Predict Genotoxicity...

Area of Science:

• Computational toxicology
• cheminformatics
• predictive modeling

Background:

• Metabolic similarity is crucial for read-across (RAx) in chemical safety assessment.
• Current methods for characterizing metabolism in RAx are still developing.
• Metabolic similarity involves metabolic trees, simulated metabolites, and transformation pathways.

Purpose of the Study:

• To compare metabolic graph representations with structural similarities for predicting genotoxicity.
• To evaluate the performance of graph convolutional networks (GCNs) in encoding metabolic information for RAx.
• To identify the most effective approach for predicting genotoxicity using metabolic pathway data.

Main Methods:

• Predicted xenobiotic metabolism using TIssue MEtabolism Simulator (TIMES) and BioTransformer.
• Converted metabolic pathways into graphs and trained GCNs to generate chemical embeddings.
• Compared classification performance of GenRA, RF, LR, and MLP using GCN embeddings versus chemical fingerprints (Morgan, MACCS).

Main Results:

• GCN embeddings with logistic regression (LR), using TIMES metabolism predictions and MACCS fingerprints, achieved the highest AUC (0.807).
• This GCN-based approach outperformed GenRA and LR with MACCS fingerprints by 14.47% and 5.49%, respectively.
• GCN embeddings of predicted metabolism pathways demonstrated superior performance over parent chemical structural features.

Conclusions:

• GCN embeddings of predicted metabolic pathways are highly effective for genotoxicity prediction.
• This method offers a systematic way to encode metabolic information for improved analogue identification in RAx.
• The findings support the use of GCNs for enhancing predictive toxicology and chemical safety assessments.