# MSAT: a FAERS-informed heterogeneous graph neural network for pharmacovigilance prediction of Chinese materia medica–associated adverse drug reactions

## TL;DR

MSAT is a new AI framework that improves safety monitoring of Chinese herbal medicines by combining real-world safety reports with biological data to predict adverse drug reactions.

## Contribution

MSAT introduces a novel heterogeneous graph neural network that integrates pharmacovigilance data with multi-scale biomedical mechanisms for ADR prediction.

## Key findings

- MSAT achieved high performance (AUC = 0.9792, AUPRC = 0.9766) in predicting CMM-associated ADRs.
- MSAT remained robust under severe class imbalance and showed good cold-start generalization.
- 13 out of 15 high-confidence predictions were supported by external evidence, including a liver injury risk for Artemisia argyi.

## Abstract

Post-marketing safety surveillance of Chinese Materia Medica (CMM) is challenged by multi-component chemical heterogeneity and the limited mechanistic interpretability of signals derived solely from spontaneous reports. The FDA Adverse Event Reporting System (FAERS) provides large-scale pharmacovigilance evidence, yet it is noisy, susceptible to reporting bias, and weakly linked to underlying biological mechanisms. We aimed to develop an FAERS-informed, clinically oriented framework to predict CMM-associated adverse drug reactions (ADRs).

We constructed an evidence-rich heterogeneous graph integrating CMMs, compounds, protein targets, and ADRs. To differentiate pharmacovigilance-derived statistical associations from binary molecular interactions, we augmented each CMM–ADR edge with a six-dimensional evidence feature vector (including semantic similarity, FAERS evidence as log-transformed report counts, source provenance, and topology-derived structural metrics) and used it to condition attention during message passing. We propose MSAT, a multi-scale heterogeneous graph neural network comprising: (i) an Evidence-Semantic Adaptive Gate to inject evidence-conditioned attention bias, (ii) a Hierarchical Signal Propagation layer to model cross-scale transduction from molecular mechanisms to clinical phenotypes, and (iii) a Hub-Calibrated Inference module to mitigate hub-driven bias. We evaluated MSAT using stratified 10-fold cross-validation, stress-tested robustness under increasing class imbalance up to a 1:10 positive:negative ratio, and assessed cold-start generalization. High-confidence predicted results were further examined via external database concordance and literature support.

In stratified 10-fold cross-validation on 27,062 curated CMM–ADR associations, MSAT achieved strong performance (AUC = 0.9792, AUPRC = 0.9766) and outperformed representative heterogeneous GNN baselines. MSAT remained robust under severe class imbalance (up to 1:10) and demonstrated favorable generalization in cold-start settings. Among the top 15 high-confidence predicted results absent from the labeled positives, 13/15 (86.7%) were supported by independent database or literature evidence. For example, MSAT prioritized a potential liver-injury signal for Aiye (Artemisia argyi) (predicted ADR: drug-induced liver injury, DILI), consistent with external evidence.

By unifying FAERS pharmacovigilance evidence with multi-scale biomedical mechanisms in a heterogeneous graph learning framework, MSAT enables robust prediction and prioritization of CMM-associated ADR risks. This framework can support hypothesis generation and risk triage for post-marketing safety surveillance of complex Chinese Materia Medica products.