Graph Attention Network Model with Defined Applicability Domains for Screening PBT Chemicals

Area of Science:

• Computational chemistry
• Environmental toxicology
• Machine learning applications in chemical safety

Background:

• Screening for environmentally persistent, bioaccumulative, and toxic (PBT) substances is crucial for chemical management.
• Existing in silico models often lack applicability domain (AD) characterization or interpretability due to complex structure-activity landscapes (SALs).

Purpose of the Study:

• To develop novel, interpretable in silico models for screening PBT chemicals using graph attention networks (GATs).
• To introduce a new AD characterization (AD_FP-AC) to enhance model reliability.

Main Methods:

• Utilized graph attention networks (GATs), a deep learning architecture, for PBT chemical screening.
• Developed an AD characterization (AD_FP-AC) incorporating molecular fingerprint (FP) similarities and activity cliffs (ACs).
• Optimized attention weight parameters (P_AW) to interpret atom contributions to PBT attributes.

Main Results:

• The GAT model demonstrated superior performance compared to previous PBT screening models.
• The model provided interpretability by highlighting atom contributions to PBT properties.
• Identified eight previously unrecognized classes of PBT chemicals from the Inventory of Existing Chemical Substances in China.

Conclusions:

• The GAT model combined with AD_FP-AC offers an efficient approach for PBT chemical screening.
• This modeling methodology can be extended to assess other chemical parameters for risk assessment and management.