Chemical Information Modeling for Component-Level Review Prioritization of Adverse-Reaction Signals in...

Area of Science:

• Pharmacology and Toxicology
• Computational Chemistry
• Data Science in Medicine

Background:

• Postmarketing safety signals for multicomponent medicines are challenging to attribute to specific components.
• Existing methods often assign signals to the entire product, hindering targeted safety reviews.

Purpose of the Study:

• To develop a chemical information modeling approach for prioritizing component-level safety reviews in multicomponent medicines.
• To create a system that ranks the review priority of individual crude-drug components within Kampo formulas based on safety signals.

Main Methods:

• Utilized a heterogeneous safety graph connecting Kampo formulas, components, targets, and adverse reactions.
• Developed HerbPairIAM, a graph neural network model employing target-centered profiles and attention mechanisms.
• Curated high-confidence safety signals from JADER and FAERS databases using a disproportionality rule.

Main Results:

• HerbPairIAM achieved an AUROC of 0.823 and AUPRC of 0.517 in a benchmark dataset, outperforming baseline models.
• Ablation studies confirmed the importance of unordered component pairs and graph-derived profiles for performance.
• Model demonstrated transferability to unseen formula compositions but showed weaker generalization for novel adverse reactions.

Conclusions:

• HerbPairIAM effectively prioritizes component-level safety reviews for multicomponent medicines, aiding in safety signal triage.
• The model's reliability is higher within known adverse reaction spaces, suggesting limitations for novel adverse reaction phenotypes.
• Findings support targeted safety review and mechanism-focused follow-up rather than direct causal attribution of toxicity.