We constructed a comprehensive dataset by integrating pharmacological data from authoritative databases (Yang et al., 2023; Chandak et al., 2023; Ochoa et al., 2023) and real-world pharmacovigilance reports (U.S. Food and Drug Administration, 2024). As summarized in Table 1, molecular associations (e.g., Compound-Target, Target-Target) were curated from structured biological repositories to form the backbone of our heterogeneous graph.

To capture post-marketing safety evidence, we mined the FDA Adverse Event Reporting System (FAERS) quarterly releases from Q1 2004 to Q1 2025. We parsed the FAERS quarterly DEMO, DRUG, and REAC tables. CMM exposure mentions were identified by matching reported product names in the DRUG table against our standardized CMM lexicon (harmonized to Latin scientific names as described below), and ADR terms were taken from the REAC table as MedDRA Preferred Terms. All analyses were conducted on a locked snapshot of FAERS (through Q1 2025), which ensures reproducibility of the reported results. Figure 4 summarizes the long-term reporting trend of CMM-related adverse event records. For comparability across calendar years, the trend is presented in an annualized form; data from 2025 include Q1 only and are therefore not shown in the year-level plot.

Because spontaneous reporting data are noisy and heterogeneous, we applied a structured curation workflow before constructing CMM–ADR signals. We harmonized CMM names using TCMBank (Lv et al., 2023) and the Chinese Pharmacopoeia Commission (2025), and used standardized Latin scientific names as unique identifiers to reduce ambiguity introduced by synonyms and homonyms. Following common pharmacovigilance practice, we handled duplicate case versions by retaining the most recent version when applicable, and treated the resulting CMM–ADR associations as signals for prioritization rather than causal confirmation.

Starting from 759,122 CMM-related FAERS report–reaction records, we excluded non-medical or administrative entries by removing records mapped to 378 non-ADR MedDRA Preferred Terms (e.g., “Product issues”), which eliminated 70,120 records. We then aggregated the remaining records at the CMM–ADR level to obtain 25,734 FAERS-derived CMM–ADR associations, and supplemented them with 1,328 expert-curated associations from the literature (Lin et al., 2013). The resulting union constituted 27,062 curated positive CMM–ADR associations used as ground truth for model development; consistent with pharmacovigilance practice, these associations should be interpreted as safety signals for prioritization rather than causal confirmation.

We modeled the pharmacological space as a heterogeneous graph G = ( V , E ) comprising four node types (CMM, Compound, Target, ADR) and six relation types. Node features were initialized using domain-specific pre-trained language models: textual nodes (CMM, Target, ADR) were encoded via BioBERT (Lee et al., 2020). For compound nodes, canonical SMILES strings were retrieved from PubChem. To ensure structural consistency, we performed chemical standardization using RDKit (version 2024.03). This process involved salt removal, charge neutralization, and tautomer canonicalization before encoding sequences via ChemBERTa (Chithrananda et al., 2020). All adverse reaction terms were systematically mapped to MedDRA (Brown et al., 1999), and target protein identifiers were unified to HGNC gene nomenclature (Tweedie et al., 2021).

To differentiate pharmacovigilance-derived signals from binary mechanistic relations, we augmented each CMM–ADR edge with a six-dimensional evidence vector encoding: (1) semantic similarity between CMM and ADR descriptions (BioBERT cosine similarity); (2) FAERS evidence as log-transformed case report counts for the CMM–ADR pair; (3) a source provenance indicator (FAERS-derived vs. literature-only); and (4–6) topology-derived metrics including the CMM degree, ADR degree, and a meta-path connectivity score capturing the density of mechanistic CMM → Compound → Target → ADR paths between the pair (log-transformed and normalized). For literature-only associations without FAERS support, the FAERS count feature was set to 0 and the provenance indicator was set accordingly, while the remaining features were computed in the same manner. This evidence vector is used to condition attention during message passing, allowing evidence strength and provenance to directly modulate information propagation rather than acting as an auxiliary covariate.

The core representation learning module utilizes a stack of three Multi-Scale Attention layers. Each layer is designed for evidence-rich heterogeneous graphs and combines edge-conditioned attention with an internal non-linear transformation that supports multi-scale signal propagation from molecular interactions to clinical phenotypes.

The final stage of MSAT is designed to bridge the translational gap between computational outputs and clinical practice.

We first compared MSAT with nine baseline methods using stratified 10-fold cross-validation on the CMM–ADR prediction task. As shown in Table 2, MSAT achieved the best performance on every metric against all baseline models. It obtained the highest AUC, AUPRC, and Matthews correlation coefficient, indicating strong discriminative ability in ranking CMM–ADR associations and a favorable balance between sensitivity and specificity. Compared with the strongest heterogeneous baseline (Simple-HGN) and the strongest homogeneous baseline (GCN), MSAT achieved higher recall and MCC, and the improvements in AUC were statistically significant across all baselines ( p < 1 0 − 5 ; Table 2).

In comparison to other models, MSAT showed different strengths. Against heterogeneous GNNs, MSAT achieved similar precision but substantially higher recall, indicating that standard heterogeneous models tend to under-predict true ADRs. Compared with homogeneous GNNs and traditional classifiers, MSAT had a clear performance margin, suggesting that casting CMM–ADR prediction as a heterogeneous graph problem can help leverage biomedical context.

Removal of any single component degraded performance relative to the full model (Table 3). Notably, removing HSP-Layer caused the largest drop, suggesting that hierarchical signal propagation is important for modeling multi-scale patterns from molecular binding to clinical phenotypes. Dropping ESA-Gate also led to a clear performance decline, supporting the importance of adaptive edge encoding for heterogeneous edge semantics. The model without HCI-Module saw a smaller yet consistent reduction (mostly in recall and MCC), reflecting HCI-Module’s role as a hub-calibrated inference mechanism.

The single-component variants further illustrate these roles. Variants with only ESA-Gate or only HSP-Layer performed better than only HCI-Module, but all three remained clearly inferior to the full MSAT model, and the differences in AUC relative to the full model were statistically significant ( p < 0.001 ) . Overall, the ablation results indicate that ESA-Gate, HSP-Layer and HCI-Module provide complementary benefits, and their combined use yields the strongest performance for CMM–ADR prediction in our evaluation.

Beyond aggregate performance metrics, we carried out a multi-dimensional evaluation to assess MSAT’s external validity, robustness to degree-related bias and computational efficiency, which are all relevant for potential clinical deployment.

Real-world pharmacovigilance data are highly imbalanced, with far fewer confirmed ADRs than non-events. To assess robustness under different imbalance levels, we varied the positive:negative sampling ratio from 1:1 to 1:10 and evaluated MSAT alongside three representative GNN baselines (HGT, Simple-HGN and GAT) on the CMM–ADR prediction task (Figure 6).

Across all ratios, MSAT maintained the highest AUC, AUPRC, F1-score and MCC among the compared models. As the proportion of negative samples increased, the performance of all methods declined, but MSAT showed a smaller loss than the baselines, particularly in AUPRC and MCC. This indicates that MSAT degrades more gracefully under increasing class imbalance and is better suited to the skewed CMM–ADR distributions encountered in practical pharmacovigilance settings.

While the above analysis evaluates robustness to test-set imbalance under balanced training, real-world pharmacovigilance must learn from inherently skewed data. We therefore performed an additional end-to-end imbalanced experiment where the training, validation, and test splits were all sampled at a 1:10 positive-to-negative ratio (Table 4). Under this setting, MSAT consistently ranked first and maintained a clear advantage over all baselines on AUC, AUPRC, and MCC (Table 4), indicating more reliable retrieval of rare adverse-reaction signals under end-to-end skew. Notably, R-GCN exhibited a high-precision but low-recall profile, suggesting an overly conservative strategy that would miss many true ADR signals in practice. In contrast, MSAT maintained a more balanced precision–recall trade-off, which is preferable in pharmacovigilance where both missed signals and false alarms carry tangible clinical costs. Traditional tabular classifiers degraded markedly under this skew, highlighting the advantage of graph-based models that can exploit relational structure when labels are sparse.

To assess the plausibility of MSAT’s ranked predicted results, we first reviewed the Top 15 highest-confidence predictions using the protocol in Section 3.5.6 (Table 5). Overall, 13 of 15 (86.7%) had support from either database verification or literature/mechanistic evidence. Because this assessment is defined relative to an incomplete labeled edge list (rather than a claim of real-world novelty), some clinically recognized effects can appear in the prediction list if they are unlabeled under FAERS-driven extraction and conservative term normalization/filtering. For example, Changshan (Dichroa febrifuga Lour.)–vomiting is supported by the literature (Sun et al., 2022), yet it was not captured as a labeled positive under our curated supervision; we therefore interpret this instance as evidence of label incompleteness rather than as a “new” finding. To mitigate the selection bias inherent to a Top- K -only analysis, we further evaluated two uniformly randomly sampled strata of additional predictions. In the mid-confidence stratum (ranks 21–100), 8 of 15 predictions (53.3%) were supported by at least one validation channel (Supplementary Table S5). In the lower-confidence stratum (ranks 101–200), 3 of 15 predictions (20.0%) were supported (Supplementary Tables S7). This decline across strata is consistent with score-based prioritization: higher-ranked predictions are enriched for externally supported associations, whereas lower-ranked predictions increasingly reflect uncertain hypotheses. Mechanistic context can further aid interpretation when direct CMM-level evidence is sparse. For example, prior evidence suggests that salicylate constituents related to Ciwujia may contribute to tinnitus (Sheppard et al., 2014), offering a plausible rationale for the predicted phenotype. Together, these analyses suggest that MSAT can help prioritize candidate CMM–ADR signals for downstream review, including associations not captured as labeled positives in our curated supervision.

A central motivation of this work was that accurate CMM–ADR prediction requires models that reflect both the semantic heterogeneity of biomedical associations and the multi-scale structure of pharmacological mechanisms. The ablation study supports this view. In particular, removing HSP-Layer led to the largest performance drop, consistent with the need to model hierarchical and non-linear transduction from molecular binding to clinical phenotypes. As a hierarchical signal propagation mechanism, HSP-Layer appears important for capturing the long-range dependencies linking molecular components to clinical outcomes.

ESA-Gate’s contribution suggests that evidence-bearing CMM–ADR edges should be treated differently from binary biological links. By injecting edge evidence as an attention bias, ESA-Gate helps preserve quantitative FAERS signals that might otherwise be diluted by relation-level projections. Removing ESA-Gate reduced discrimination and recall, indicating that uniform transformations across edge types can underutilize the rich edge attributes. In our graph, CMM–ADR edges carry multi-dimensional attributes derived from FAERS statistics, literature evidence, and structural context, which are fundamentally different from binary biological edges. ESA-Gate allows the model to adaptively balance non-linear transformation and direct information preservation, supporting detection of subtle but clinically meaningful risk signals.

HCI-Module provided a complementary benefit. Although removing HCI-Module had a smaller effect than removing HSP-Layer or ESA-Gate, the model’s performance still declined (particularly in recall and MCC). This outcome aligns with HCI-Module’s role as a hub-calibrated inference mechanism combining multiple scoring functions, capturing both symmetric similarity and directional (asymmetric) association patterns. The degree-aware weighting in HCI-Module further calibrates predictions for both high- and low-degree nodes. Overall, the ablation results suggest that MSAT’s performance advantage arises from the combined action of these three components rather than from any single module in isolation.

From a clinical perspective, an important contribution of this work is the mapping of predicted MedDRA-coded ADRs into sixteen TCM functional systems. Conventional pharmacovigilance models output signals in regulatory terminology that may not align with how TCM clinicians conceptualize organ systems and syndromes. By translating ADR predictions into categories based on Zang–Fu organs and key functional systems (such as Qi-Blood-Fluid and meridian systems), MSAT produces outputs that can be more directly integrated into TCM decision-making.

The case analyses illustrate this potential. For example, predicted vomiting events for specific CMMs were correctly assigned to the Stomach system, and tinnitus was associated with the Kidney system, consistent with the classical principle that the Kidney “opens into the ears.” Similarly, dizziness mapped to the Liver system, aligning with patterns such as liver-yang rising in TCM theory. These examples suggest that, when MSAT identifies a potential ADR, the corresponding TCM system mapping can help clinicians interpret the signal within their diagnostic framework, guiding syndrome differentiation, formula adjustments, and targeted monitoring of high-risk organ systems.

Pharmacologically, MSAT recovered both well-documented and literature-reported CMM–ADR associations and also produced high-confidence predicted signals for downstream review. This pattern suggests that heterogeneous pathways (e.g., compound–target–ADR chains) carry meaningful predictive information. The cold-start evaluation, in which most CMMs were unseen during training, is particularly relevant. In this setting, MSAT maintained high precision and MCC, suggesting that it can extrapolate risk patterns based on chemical and network context rather than relying solely on historical co-occurrence. This capability may be important for proactive safety assessment of emerging CMMs and formula components.

From a pharmacovigilance operations standpoint, MSAT is best positioned as a signal-triage and prioritization layer, rather than a replacement for standard disproportionality screening. In a periodic monitoring setting, MSAT can produce a ranked list of candidate CMM–ADR signals that are supported by spontaneous reporting evidence and coherent with mechanistic paths in the heterogeneous graph (e.g., compound–target–ADR). Such rankings can help safety teams prioritize manual case review, targeted literature assessment, and downstream pharmacoepidemiological or experimental follow-up, particularly for long-tail CMMs where historical evidence is sparse and rule-based screening is less informative. Compared with marginal-count disproportionality analyses, MSAT incorporates FAERS-derived reporting evidence (log-transformed case report counts and provenance) as edge features while jointly modeling mechanistic connectivity, enabling context-aware prioritization under confounding and sparsity.

Methodologically, this work contributes to a growing line of research applying graph representation learning to pharmacovigilance. Existing heterogeneous GNNs such as R-GCN (Schlichtkrull et al., 2018), HGT (Hu et al., 2020), and Simple-HGN (Lv et al., 2021) provide strong baselines, but they are primarily designed around relation-type–specific message passing on graph topology, with limited emphasis on high-dimensional edge attributes and degree-aware calibration. These results suggest that, in settings where edges carry semantically rich features and where degree distributions are highly skewed, explicit edge encoders and late fusion mechanisms can yield measurable gains.

Our multi-dimensional evaluation further highlights the importance of going beyond average test metrics. Degree-stratified analysis revealed a “head-hard” phenomenon in the CMM–ADR graph: high-degree CMMs were more difficult to predict, likely due to noisy and heterogeneous interaction patterns. MSAT improved performance in this head group without sacrificing accuracy on long-tail CMMs, suggesting a pathway to fairer and more robust graph models in other biomedical domains with similar degree distributions. Likewise, robustness across class-imbalance ratios suggests that MSAT is relatively stable under the skewed positive–negative ratios that characterize many real-world safety datasets.

Although this study focuses on CMM–ADR prediction, the underlying ideas of ESA-Gate, HSP-Layer and HCI-Module are more general. Similar architectures could be explored for tasks such as drug–drug interaction (DDI) prediction, target deconvolution, or adverse event forecasting in other complementary and integrative medicine systems, where heterogeneous evidence and theory-specific clinical frameworks coexist.

Several limitations should be acknowledged. First, the ground-truth CMM–ADR associations are derived primarily from FAERS and other spontaneous reporting sources, supplemented by expert-curated literature. Such data are prone to under-reporting, stimulated reporting, indication and channeling bias, co-medication confounding, and duplicate or incomplete reports, and they do not provide causal confirmation (Harpaz et al., 2012). Moreover, FAERS lacks exposure denominators, so the predicted scores should be interpreted as relative signal strength for prioritization rather than incidence or absolute risk. Co-medication and indication-related confounding are particularly salient in FAERS. We do not perform causal adjustment in this study. As a report-level sensitivity analysis, we restricted to monotherapy-dominant TCM-containing reports (defined as reports with exactly one unique drug in the exported table). Under this proxy definition, the strongest CMM–ADR report-count edges remained highly stable relative to the full set (Supplementary Figure S4), suggesting that the most prominent signals are not solely driven by polypharmacy. In addition, a single standardized CMM name can correspond to heterogeneous commercial preparations and formulations with variable constituent profiles in FAERS, which can dilute attribution and introduce label noise. Finally, curated resources used to instantiate mechanistic edges (e.g., Target–ADR) may be semantically correlated with FAERS-derived signals. We interpret this as real-world knowledge overlap and residual confounding rather than cross-validation leakage under our fold-wise, bidirectional test-edge removal protocol. Second, our supervised evaluation relies on type-constrained negative sampling, where negatives represent unobserved CMM–ADR pairs rather than confirmed non-associations; while standard in link prediction, this setting may not fully reflect real-world decision thresholds. For comparability, we retain uniform type-constrained negatives in the main protocol and additionally report a mechanism-aware hard-negative sensitivity analysis (Supplementary Table S3). More principled hard-negative construction that incorporates richer exposure or covariate information remains an important direction for future work. Third, because FAERS often lacks reliable dosage and formulation details for CMM products, MSAT currently focuses on hazard identification (signal detection) rather than dose-response or risk characterization. We encoded FAERS evidence primarily via report-count–based features; incorporating classical disproportionality metrics (e.g., ROR/PRR) will be explored in future work to further characterize signal strength under confounding. Future work should incorporate complementary data sources (e.g., electronic health records, prospective registries, or active surveillance systems) and temporal analyses to strengthen validation and enable more clinically actionable risk stratification.

The TCM functional system mapping is based on expert-defined rules and authoritative references, and was evaluated using a limited number of case studies. The current mapping operates at the level of organ and functional systems rather than full syndrome patterns. Constrained by the granularity of the source data, it does not account for inter-individual variability, formula composition, or dosing regimens. Extending the mapping to finer-grained TCM entities (e.g., syndrome patterns and meridians) and validating it in clinical cohorts will be important steps toward routine use in practice.

Regarding interpretability, although MSAT improves clinical alignment at the level of outputs, the internal reasoning process of the model remains only partially transparent. This work did not systematically apply post hoc explanation techniques for GNNs, and therefore cannot fully characterize which specific pathways (e.g., particular compounds, targets, or subgraphs) drive individual predictions. Future research could integrate pathway-level explanation methods or attention-based path-tracing techniques to highlight the most influential components in each CMM–ADR association and to generate mechanistic hypotheses.

Finally, the present study focuses on single CMMs. In clinical practice, CMMs are most often used as part of multi-CMM formulas following principles such as Jun–Chen–Zuo–Shi (sovereign, minister, assistant, and courier), and interactions among ingredients can alter both efficacy and safety. Extending the heterogeneous graph to include formula nodes and explicit formula–CMM relationships would enable MSAT-like architectures to predict ADRs at the formula level and to study how different combinations modify risk. Incorporating patient-level covariates and temporal information would further support personalized and dynamic safety assessment.

Overall, this work illustrates how clinically aligned machine learning can connect computational pharmacovigilance with traditional medicine frameworks. While further validation and refinement are needed, especially in prospective and formula-level settings, our results suggest a practical template: combining semantic edge encoding, multi-scale representation learning, and theory-grounded clinical mapping can move models beyond black-box prediction toward outputs that are more interpretable in clinical context. Such alignment between data-driven modeling and domain knowledge may help support translation of safety informatics into integrative medicine practice.