FGTN: Fragment-based graph transformer network for predicting reproductive toxicity

Area of Science:

• Computational toxicology
• Chemical safety assessment
• Quantitative structure-activity relationship (QSAR) modeling

Background:

• Reproductive toxicity poses significant challenges in chemical safety.
• Traditional toxicity testing is expensive, time-consuming, and ethically problematic.
• Existing in silico models often fail to fully utilize compound topological information.

Purpose of the Study:

• To develop a novel fragment-based graph transformer network (FGTN) for predicting human reproductive toxicity.
• To incorporate internal topological structure information into QSAR modeling.
• To improve upon existing atom-based graph neural network methods.

Main Methods:

• Representing compounds as graphs with fragments as nodes and bonds as edges.
• Introducing a super molecule-level node to capture global molecular features.
• Utilizing fragment embeddings and graph transformer architecture for prediction.

Main Results:

• Achieved high performance with an accuracy (ACC) of 0.861 and an AUC of 0.914 on blind tests.
• Outperformed traditional fingerprint-based models and atom-based graph convolutional network (GCN) models.
• Demonstrated the ability to attribute toxic predictions to specific fragments, identifying structural alerts.

Conclusions:

• The FGTN model provides a reliable and effective tool for predicting human reproductive toxicity.
• FGTN enhances chemical safety assessment by offering a more efficient and informative in silico approach.
• The model's ability to identify toxic fragments and potentially distinguish isomers advances computational toxicology.