MGCL-CAP: Masked Graph Contrastive Learning with Gated Cross-Attention for Chemical Allergenicity Prediction.

Area of Science:

• Computational toxicology
• cheminformatics
• machine learning

Background:

• Chemical allergens in consumer and industrial products pose public health risks.
• Traditional allergen screening methods are slow and resource-intensive.
• Current computational methods lack accuracy due to limitations in capturing molecular complexity.

Purpose of the Study:

• To develop an advanced deep learning framework, MGCL-CAP, for accurate chemical allergenicity prediction.
• To improve upon existing computational approaches by integrating graph-based learning and cross-modal attention.
• To provide a reliable tool for efficient allergen candidate prioritization and safer chemical formulation.

Main Methods:

• Utilized masked graph contrastive learning within a graph isomorphism network encoder.
• Employed random subgraph masking to learn structure-invariant graph embeddings.
• Integrated molecular fingerprints with graph embeddings using multihead gated cross-attention for modality fusion.

Main Results:

• MGCL-CAP demonstrated superior performance compared to state-of-the-art allergenicity predictors.
• The model exhibited stability across various hyperparameter settings.
• Interpretability analysis identified substructures linked to sensitization mechanisms.

Conclusions:

• MGCL-CAP provides a robust and accurate computational tool for chemical allergenicity assessment.
• The framework enhances efficiency in prioritizing chemical candidates and designing safer products.
• Mechanistic insights from interpretability analysis support future chemical safety evaluations.