Applicability Domains Based on Molecular Graph Contrastive Learning Enable Graph Attention Network Models to...

Area of Science:

• Computational chemistry and toxicology
• Environmental science and risk assessment
• Machine learning applications in chemical safety

Background:

• Accurate prediction of physicochemical properties and environmental fate is crucial for chemical management.
• Existing in silico models face limitations in performance and applicability domain definition.
• Need for robust computational tools to screen large chemical inventories for environmental hazards.

Purpose of the Study:

• To develop and validate advanced graph attention network (GAT) models for predicting 15 chemical endpoints.
• To introduce and integrate a novel structure-activity landscape (SAL)-based applicability domain (AD_SAL) for enhanced model reliability.
• To assess the performance of GAT models coupled with AD_SAL for screening large chemical datasets.

Main Methods:

• Construction of GAT models utilizing molecular graph representations for 15 environmental endpoints.
• Calculation of molecular similarity density (ρs) and activity inconsistency (IA) for applicability domain characterization.
• Integration of molecular graph contrastive learning to define AD_SAL{ρs, IA} for model refinement.

Main Results:

• GAT models demonstrated superior performance compared to state-of-the-art methods across all endpoints.
• The novel AD_SAL{ρs, IA} significantly improved the prediction coefficient of determination (R²) by an average of 14.4%, achieving R² > 0.9 for all endpoints.
• Models successfully screened approximately 10^6 chemicals for persistence, mobility, and bioaccumulation potential.

Conclusions:

• The developed GAT models coupled with AD_SAL represent a significant advancement in predictive toxicology and environmental risk assessment.
• The enhanced models provide reliable predictions and robust applicability domains, setting a new benchmark for future in silico modeling.
• This approach facilitates efficient and accurate screening of large chemical inventories, supporting sound chemical management strategies.