Machine learning–guided Huanglian Jiedu decoction targets STING in periodontitis-induced Alzheimer’s Disease

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder increasingly associated with peripheral inflammatory conditions such as chronic periodontitis (CP); however, the molecular mechanisms linking these conditions remain poorly understood. Here, we investigated the therapeutic effects of Huanglian Jieddu Decoction (HLJDD) on CP-induced AD using an integrative machine learning-guided multi-omics approach. Analysis of public single-cell RNA-sequencing data revealed pronounced inflammatory activation in microglia from AD samples. We further established a CP-induced AD rat model and performed hippocampal transcriptomic profiling. Multiple complementary machine learning strategies, including Random Forest-based feature selection, support vector machine-based refinement, network modeling, and interpretable model analysis, were applied to prioritize disease-relevant pathways from high-dimensional transcriptomic data. Across models, components of the cGAS–STING signaling pathway consistently exhibited strong and directional contributions to CP–AD pathology, indicating a central inflammatory axis linking peripheral infection to neurodegeneration. Guided by these data-driven insights, in vivo and in vitro experiments demonstrated that HLJDD suppressed cGAS–STING activation, attenuated neuroinflammation, and improved cognitive function in CP-induced AD models. Collectively, this study highlights the value of machine learning-assisted transcriptomic interpretation for mechanistic prioritization and identifies HLJDD as a multitarget therapeutic strategy for CP-induced AD.

Introduction

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by memory impairment, cognitive dysfunction, and neuroinflammation, with its incidence rapidly increasing with the global population's ages, placing a substantial burden on patients and society^1,2. Despite extensive research, the underlying mechanisms of AD remain incompletely unclear, and current treatments primarily provide symptomatic relief^3,4. In recent years, accumulating evidence has suggested a link between chronic periodontitis (CP) and AD⁵. CP, a common oral inflammatory disease, not only causes local tissue damage but also induces systemic inflammation, which can affect distant organs, including the brain^6–8. The periodontal pathogen Porphyromonas gingivalis (P. gingivalis) has been detected in the brains of patients with AD, where it triggers neuroinflammation and exacerbates amyloid-β (Aβ) accumulation and tau phosphorylation^9,10. Gingipains, including lysine-gingipain (Kgp) and arginine-gingipains (RgpA/B), are major virulence factors that enable P. gingivalis to breach the blood-brain barrier (BBB)¹¹ and translocate into the central nervous system (CNS)¹², Once in the CNS, these proteases induce chronic neuroinflammation, directly cleave tau, promote Aβ accumulation, and activate microglia, thereby further aggravating neuronal damage directly cleave tau, promote Aβ accumulation, and activate microglia, further aggravating neuronal damage^9,13–16. Systemic P. gingivalis infection also increases the inflammatory burden, activates neuroinflammation-related pathways, impairs immune regulation, and accelerates AD-like cognitive decline^17,18. These findings highlight the multifactorial role of P. gingivalis in linking CP to AD pathogenesis.

The cyclic GMP‒AMP synthase-stimulator of interferon genes (cGAS‒STING) pathway is an innate immune signaling pathway responsible for recognizing microbial or self-derived cytosolic DNA and inducing inflammatory responses^19,20. Accumulating evidence highlights the pivotal role of cGAS–STING signaling in neuroinflammation and neurodegenerative diseases^21,22. Activation of this pathway leads to the phosphorylation of downstream effectors, such as TANK-binding kinase 1 (TBK1) and interferon regulatory factor 3 (IRF3), resulting in type I interferons (IFN-I) production²³. Recent studies have demonstrated that chronic activation of cGAS–STING exacerbates microglial reactivity, accelerates AD-like pathology in multiple animal models, and may contribute to pathological features such as Aβ accumulation and abnormal tau phosphorylation^24–26. Moreover, P. gingivalis infection has been shown to activate STING signaling in peripheral tissues and in experimental periodontitis, whereas pharmacological blockade of STING mitigates local bone loss and inflammation^27,28. P. gingivalis may activate the cGAS–STING pathway directly via bacterial DNA or virulence factors entering the brain or indirectly through infection‑induced release of host mitochondrial or nuclear DNA, ultimately inducing AD-like neuroinflammation, which subsequently leads to cognitive impairment. However, whether the cGAS–STING pathway is the pathological link between CP and AD remains to be elucidated.

Huanglian Jiedu Decoction (HLJDD), a classic traditional Chinese medicine formula, is known for its anti-inflammatory and antimicrobial properties. It has been widely used for the treatment of cerebrovascular diseases, ischemic stroke, and AD^29–31. HLJDD has been shown to alleviate gut microbiota dysbiosis, reduce Aβ accumulation, improve sphingolipid and carbonyl compound metabolism, and decrease neuroinflammation and pyroptosis, ultimately reversing cognitive impairment^31–33. Additionally, HLJDD effectively alleviates bacterial infections and associated inflammatory complications, including oral ulcers, gingivitis, and periodontitis. Pharmacological studies indicate that HLJDD protects against gingival damage in CP by reducing inflammation and enhancing antioxidant capacity³⁴. Its active components, such as berberine and palmatine, inhibit P. gingivalis growth, suppress gingipain activity, and mitigate immune and tissue damage³⁵. Baicalin disrupts P. gingivalis structure and reduces the activation of inflammatory pathways, including Toll-like receptor signaling, p38 MAPK, and NF-κB, alleviating inflammation and alveolar bone loss³⁶. It is reasonable to hypothesize that HLJDD may mitigate CP-related, AD-like neuroinflammation by restricting P. gingivalis burden and inhibiting the cGAS–STING axis.

With the advancement of multi-omics technologies and machine learning, there has been a significant shift in the exploration of the mechanisms underlying traditional Chinese medicine (TCM) prescriptions. Traditional computational methods, such as molecular docking and virtual screening, have limitations in addressing the complex interactions between TCM formulations and their targets^37–39. However, machine learning-based approaches, including deep learning (DL) and hypergraph convolutional networks, have shown strong potential in predicting herb–disease associations and identifying key active compounds^40,41. These methods integrate large-scale biological data, providing deeper insights into the pharmacological effects of TCM.

In this study, we utilized public single-cell databases, transcriptomic sequencing, and machine learning techniques to explore the therapeutic targets of HLJDD in CP-induced AD. By integrating multi-omics data with machine learning approaches, we systematically identified key molecular mechanisms and therapeutic pathways of HLJDD (Fig. 1). We then validated these findings in a rat model of CP-induced AD-like pathology, assessing the impact of HLJDD on bacterial colonization, cognitive performance, neuroinflammation, and the cGAS–STING axis. These results highlight the role of machine learning and multi-omics technologies in advancing the mechanistic understanding of traditional Chinese medicine TCM and support HLJDD as a multitarget therapeutic candidate for AD.

Fig. 1. Study workflow.

This flow diagram illustrates the study on HLJDD in treating CP-induced AD, emphasizing a machine learning–based integration of multi-omics data. The process includes: (1) scRNA-seq analysis (using public databases) to examine inflammation and cGAS–STING pathway activation in AD and normal control microglia; (2) creation of a rat model for periodontitis-induced AD to validate experimental data; (3) bulk transcriptomics analysis to identify key genes and pathways; (4) application of machine learning models (Random Forest, Boruta, SVM, SHAP, BN and ANN) to prioritize disease-driving pathways; (5) experimental validation in vitro and in vivo to assess HLJDD’s effects on inflammation, cGAS–STING pathway suppression, microbial infection, and cognitive function; and (6) final results showing HLJDD’s potential to mitigate neuroinflammation and AD pathology.

Results

Activation of the cGAS–STING pathway in microglia of AD

To characterize disease-associated transcriptional changes in AD, we first analyzed publicly available scRNA-seq datasets from AD and normal control (NC) brain samples. After stringent quality control (Figs. S1 and S2), unsupervised clustering and UMAP-based dimensionality reduction identified multiple transcriptionally distinct cell populations (Fig. 2A). While AD and NC cells largely overlapped across major cell types (Fig. 2B), cell type annotation based on canonical marker genes clearly delineated astrocytes, endothelial cells, excitatory and inhibitory neurons, oligodendrocytes, microglia, and excitatory neuron (Fig. 2C and D).

Fig. 2. Single-cell RNA-seq profiling and cell-type annotation of AD and control brain samples.

A UMAP plot showing the clustering of cells based on high-resolution Louvain community detection, revealing distinct cell populations. B UMAP visualization of AD and NC samples, demonstrating integration without batch effects. C Cell-type annotation of major populations, including astrocytes, endothelial cells, excitatory neurons, interneurons, microglia, oligodendrocytes, and excitatory neurons. D Heatmap of key marker genes used for cell-type identification, with Z-score-scaled expression across cell types. E Differential gene expression heatmap between AD and NC samples, highlighting disease-associated transcriptional changes. F Violin plots showing the expression levels of microglial markers (C1QA, C1QB, C1QC, CX3CR1, P2RY12, TMEM119) in AD and NC groups.

Comparative analysis revealed that microglia exhibited pronounced transcriptional alterations in AD relative to NC. Heatmap visualization of differentially expressed genes highlighted disease-associated changes that were most prominent within the microglial compartment (Fig. 2E). Consistently, key microglial marker genes, including C1QA, C1QB, C1QC, CX3CR1, P2RY12, and TMEM119, displayed distinct expression patterns between AD and NC samples (Fig. 2F), indicating substantial microglial reprogramming in the AD brain.

Given the established role of innate immune signaling in neuroinflammation, we next investigated whether the cGAS–STING pathway is activated in AD-associated microglia. Differential expression analysis revealed upregulation of inflammation-related genes in AD microglia compared to NC (Fig. 3A). Gene set enrichment analysis further demonstrated significant enrichment of the cGAS–STING signaling pathway in AD microglia (adjusted p < 0.05), supporting pathway-level activation in the disease context (Fig. 3B). Feature plots revealed increased expression of core pathway components, including cGAS, TMEM173 (STING), TBK1, and IRF3, in AD microglia relative to NC, with spatial localization largely restricted to the microglial clusters (Fig. 3C). Dot plot analysis confirmed that activation of cGAS–STING-related genes was preferentially observed in microglia rather than other neural or non-neural cell types (Fig. 3D). Consistent with these findings, cell type-resolved heatmap analysis showed coordinated upregulation of cGAS–STING pathway genes in microglia from AD samples (Fig. 3E).

Fig. 3. Single-cell RNA-seq analysis of inflammation and cGAS–STING pathway activation in AD versus NC samples.

A Heatmap displaying the expression of inflammation-related genes across different cell types, highlighting differences in the inflammatory response between AD and NC samples. B Gene set enrichment analysis (GSEA) plot of the cGAS–STING pathway, showing significant enrichment in AD samples (p-value = 0.002). C UMAP plots showing the expression of selected cGAS–STING pathway markers (TMEM173, cGAS, TBK1, and IRF3) in AD and NC samples, illustrating distinct activation patterns in different cell populations. D Dot plot showing the differential expression of cGAS–STING pathway genes across major cell types, with significant expression differences (FDR < 0.05) highlighted. E Heatmap of cGAS–STING pathway gene expression across AD and NC samples, with gene expression levels normalized by Z-score, annotated with cell types and functional pathways.

Transcriptomic co-morbidity analysis of CP and AD using public databases

To further investigate the molecular overlap between CP and AD, we performed a transcriptomic comorbidity analysis using publicly available datasets. Given the observed activation of the cGAS–STING pathway in AD microglia, we examined whether genes involved in this pathway were also altered in periodontitis. Differential gene expression analysis between periodontitis and NC samples revealed extensive transcriptional alterations, with numerous upregulated and downregulated genes (Fig. 4A). Heatmap analysis of the top differentially expressed genes (DEGs) demonstrated distinct expression patterns between periodontitis and NC samples, with periodontitis exhibiting a unique molecular profile (Fig. 4B). Principal component analysis (PCA) further revealed clear separation between the periodontitis and NC groups, indicating marked differences in their transcriptomic landscapes (Fig. 4C).

Fig. 4. Transcriptomic co-morbidity analysis of periodontitis and AD using public databases.

A Volcano plot showing DEGs between periodontitis and NC samples. Upregulated genes are colored in red, downregulated genes in yellow, and non-significant genes in gray. B Heatmap of the top DEGs between periodontitis and NC, highlighting distinct gene expression profiles between the two groups. C PCA of periodontitis and NC samples, demonstrating clear separation between the two groups based on their transcriptomic signatures. D Venn diagram showing the overlap of DEGs between periodontitis and AD, with 4.5% of genes shared between the two diseases. E ROC curves for key pathway-related genes (cGAS, TMEM173, IRF3, TBK1) in distinguishing periodontitis from AD, demonstrating their diagnostic potential with AUC values of 0.873, 0.735, 0.7, and 0.729, respectively.

We next examined the overlap of DEGs between periodontitis and AD and found that 4.5% of genes were shared between the two conditions (Fig. 4D). This shared gene set suggests the presence of convergent molecular mechanisms, potentially involving inflammatory and immune-related processes. We then assessed the diagnostic potential of key cGAS–STING pathway genes, including cGAS, TMEM173 (STING), IRF3, and TBK1. Receiver operating characteristic (ROC) analysis showed that these genes exhibited robust discriminative performance, with AUC values of 0.873, 0.736, 0.7, and 0.729, respectively (Fig. 4E).

Together, these results support a shared involvement of the cGAS–STING pathway in periodontitis and AD, highlighting its potential relevance as a common therapeutic target.

HLJDD alleviates CP in rats

To evaluate the effects of HLJDD on CP, 8-week-old Sprague‒Dawley rats were subjected to ligation combined with P. gingivalis application to establish a CP model (Fig. 5A). Micro-computed tomography (micro-CT) reconstruction revealed pronounced alveolar bone loss and root furcation exposure in the CP group. HLJDD treatment markedly attenuated alveolar bone resorption in a dose-dependent manner, while the broad-spectrum antibiotic doxycycline also significantly reduced bone loss (Fig. 5B, C). Furthermore, HE staining showed that periodontal ligation resulted in significant resorption of both the vertical height and transverse width of the alveolar bone, accompanied by active osteoclastic resorption lacunae, epithelial downgrowth toward the root surface, and inflammatory cell infiltration (Fig. 5D). These pathological changes were substantially alleviated by HLJDD in a dose-dependent manner, indicating that HLJDD effectively reduces periodontal inflammation and alveolar bone destruction in experimental periodontitis.

Fig. 5. Effect of HLJDD on CP due to ligation combined with the application of P. gingivalis in rats.

A Timeline of the experiment. B Mesial‒distal micro-CT slices of maxillary molars and 3D reconstruction images. C The distance from the mesial enamel bone boundary of the second molar to the crest of the alveolar bone (CEJ-ABC) was used to quantify alveolar bone resorption (n = 3 per group; one-way ANOVA, **p < 0.01 and ***p < 0.001 compared with the CP group; ^#p < 0.05 and ^##p < 0^.01 compared with the control group). The graphs show the means with SEM error bars. D HE staining of alveolar bone (n = 3 per group).

HLJDD treatment mitigates AD-like pathology induced by CP in rats

To assess the therapeutic potential of HLJDD in alleviating AD-like pathology induced by CP, we performed a series of behavioral, histological, and molecular analyses. HLJDD treatment significantly improved spatial memory, as demonstrated by the Morris water maze (MWM) test, with HLJDD-treated rats performing better than the CP group (Fig. 6A). Escape latency during the MWM training phase was also significantly reduced in HLJDD-treated rats (Fig. 6B), indicating improved cognitive performance.

Fig. 6. Therapeutic effects of HLJDD on AD-like pathology induced by CP in rats.

A Representative swimming trajectories from the MWM spatial probe test. B Escape latency across the 4-day training phase. C HE staining was used to observe pathological changes (n = 3 per group) and Nissl staining was performed to assess neuronal damage in the cortex and hippocampal regions CA1. D Quantitative analysis of the number of surviving neurons in the CA1 (n = 6 per group). E Representative immunoblot bands of Aβ, tau and p-tau by western blotting. F Relative expression of Aβ (n = 6 per group) in the hippocampus of the rats. G Ratio of p-tau to total tau expression (n = 6 per group). H The expression of Aβ in the CSF of the rats was measured via ELISA (n = 6 per group). (I-M) The levels of the inflammatory cytokines TNF-α, IL-1β, IL-6, IL-4, and IL-10 in the serum and CSF were measured via ELISA (n = 8 per group). N and O Quantification of P. gingivalis DNA copy numbers in the hippocampus and CSF via qPCR. (P-R) Relative mRNA expression of the P. gingivalis virulence genes Kgp, RgpA, and RgpB in the hippocampus. The data are expressed as the means ± SEMs (one-way ANOVA, *p < 0.05, **p < 0.01 and ***p < 0.001 compared with the CP group; ^#p < 0.05 and ^##p < 0.01 compared with the control group).

Histological analysis using HE staining revealed that P. gingivalis-induced periodontitis caused pronounced neuronal damage in the hippocampal CA1 region, which was substantially alleviated by HLJDD treatment. The high-dose HLJDD group exhibited the most marked histological improvement, and Nissl staining further confirmed a dose-dependent rescue of neuronal survival (Fig. 6C and D). Western blot analysis of hippocampal tissue (Fig. 6E and F) and ELISA measurement of Aβ1–42 levels in cerebrospinal fluid (Fig. 6H) consistently supported these observations. HLJDD treatment reduced Aβ expression in a dose-dependent manner, whereas doxycycline exerted only limited effects. In addition, HLJDD treatment significantly decreased the p-tau/tau ratio, a key pathological hallmark of AD, as determined by Western blot analysis (Fig. 6E and G).

We further quantified inflammatory cytokine levels, including TNF-α, IL-1β, IL-6, IL-4, and IL-10, in both serum and CSF. HLJDD treatment markedly reduced the levels of these cytokines, indicating effective modulation of systemic and central inflammatory responses associated with CP-induced AD-like pathology (Fig. 6I–M). Moreover, HLJDD treatment significantly decreased P. gingivalis DNA copy numbers in the hippocampus and CSF, suggesting a robust antimicrobial effect (Fig. 6N and O). In addition, qPCR analysis showed that HLJDD treatment significantly reduced the hippocampal mRNA expression of key P. gingivalis virulence genes, including Kgp, RgpA, and RgpB (Fig. 6P–R), further supporting its role in mitigating AD-related pathological processes.

Machine learning-assisted transcriptomic modeling identifies key pathways modulated by HLJDD in CP-induced AD

To quantitatively characterize transcriptomic alterations associated with CP-induced AD and the therapeutic effects of HLJDD, we applied machine learning-assisted multivariate modeling to hippocampal RNA-seq data. Differential expression analysis revealed widespread transcriptional changes between CP and control groups, as visualized by volcano plots (Fig. 7A). These CP-induced alterations were markedly attenuated by HLJDD treatment. PCA demonstrated clear separation between control and CP samples, with HLJDD-treated samples shifting toward the control cluster, indicating partial restoration of the transcriptomic state (Fig. 7B). Consistently, heatmap analysis showed that CP-induced dysregulation of inflammation-related genes was partially reversed following HLJDD treatment (Fig. 7C and D).

Fig. 7. Transcriptomic analysis of HLJDD treatment in rats with CP.

A Volcano plot showing DEGs between Control vs CP and CP vs HLJDD-H groups. Upregulated genes are shown in red, downregulated genes in yellow, and non-significant genes in gray. B PCA of transcriptomic data from Control, CP, and HLJDD-H groups, showing distinct clustering of gene expression profiles in each group. C Heatmap of DEGs between Control and CP groups, with gene expression levels color-coded by Z-score. D Heatmap of DEGs between CP and HLJDD-H groups, showing changes in gene expression following HLJDD treatment in the CP model.

To further integrate high-dimensional gene expression patterns and maximize discrimination among experimental groups, partial least squares (PLS)-based modeling was employed. PLS score plots revealed distinct separation among control, CP, and HLJDD-treated groups, indicating that HLJDD induced a systematic and coordinated shift in hippocampal transcriptomic profiles compared with CP alone (Fig. 8A–E). Model diagnostics and performance metrics supported the robustness and predictive capacity of the PLS-based approach (Fig. 8B–D). Importantly, overlap analysis identified both shared and condition-specific gene sets (Fig. 8F, G). Functional enrichment of the overlapping genes revealed predominant involvement in immune and inflammatory processes (Fig. 8H, I), suggesting that HLJDD exerts its therapeutic effects through network-level modulation of inflammatory signaling rather than isolated molecular targets. Subsequent pathway enrichment analysis of PLS-prioritized genes revealed significant overrepresentation of innate immune and neurodegeneration-associated pathways, among which the cGAS–STING signaling pathway consistently emerged as a highly ranked axis (Fig. 8J, K).

Fig. 8. Transcriptomic analysis of HLJDD treatment in CP rats.

A Bar plot showing the R2Y and Q2 values for partial least squares regression (PLS-R) analysis, indicating the fit and predictive ability of the model. B Scatter plot of PLS-R analysis, showing the similarity between treatments and controls based on gene expression profiles. C Observation diagnostics plot from PLS-DA, showing the distance of individual observations from the model, indicating the model’s accuracy. D PLS-DA score plot displaying the separation between treatment and control groups, highlighting distinct gene expression patterns for each group. E PLS-DA scores plot showing the clustering of Control, CP, and HLJDD-H groups, with clear separation indicating distinct treatment effects. F Venn diagram showing the overlap of DEGs between CP and HLJDD-H, with gene numbers represented in each set. G Venn diagram showing the three-way intersection of DEGs between different treatment groups, highlighting the common and unique genes. H Bar plot of the number of DEGs identified in different gene sets, showing a higher number of DEGs in HLJDD-H-treated rats compared to CP alone. I GO term enrichment analysis, showing the top biological processes, cellular components, and molecular functions associated with DEGs in HLJDD-H treated rats. J KEGG pathway analysis of DEGs, showing the top enriched pathways, including the cGAS-STING pathway, neurodegenerative diseases, and immune-related pathways. K Gene-pathway chord diagram illustrating the relationship between key genes and their associated pathways, highlighting the impact of HLJDD treatment on inflammation and neuroinflammation-related pathways.

Together, these results demonstrate that multivariate machine learning modeling effectively distills complex transcriptomic data into a biologically coherent and mechanistically informative set of pathways responsive to HLJDD intervention.

Machine learning prioritizes the cGAS–STING axis as the dominant inflammatory driver in CP-induced AD

To further determine which pathway plays a dominant role in CP-induced AD pathology, we applied multiple complementary machine learning approaches to hippocampal transcriptomic data. Rather than relying on a single method, we integrated ensemble tree-based feature selection, margin-based classification, and network-based inference to robustly prioritize regulatory axes underlying disease progression and therapeutic intervention.

Feature selection using Random Forest and Boruta

Random Forest modeling was first applied to rank genes according to their contribution to disease classification, enabling unbiased identification of high-impact transcriptomic features. Boruta analysis further refined this ranking by retaining only features whose importance consistently exceeded that of randomized shadow attributes. Feature importance analysis identified a subset of genes with strong discriminatory power between CP and HLJDD-treated groups (Fig. 9A). The Random Forest out-of-bag (OOB) error rate plot demonstrated stable model performance across increasing numbers of trees, indicating robust classification accuracy for both experimental groups (Fig. 9B). Boruta feature selection further confirmed the relevance of these genes, as shown by the box plot displaying feature importance scores (Fig. 9C). In addition, UMAP visualization based on Boruta-selected genes revealed clear separation between CP and HLJDD-treated samples, supporting the effectiveness of Boruta-based feature selection (Fig. 9D).

Fig. 9. Machine learning models for feature selection and classification in CP and HLJDD treatment.

A Feature importance plot showing the top genes selected based on importance scores from the random forest model. B Random Forest error rate plot, showing the out-of-bag (OOB) error for the overall model (black line), Class 1 (blue dashed line), and Class 2 (red dashed line) across different numbers of trees. C Boruta feature selection results, showing the importance of each feature with a box plot, indicating which genes were selected as important by the Boruta algorithm. D UMAP plot based on Boruta-selected genes, showing distinct separation between the two groups (Group 0 and Group 1), indicating effective feature selection for classification. E Support SVM generalization error vs number of features plot, showing how the generalization error changes as more features are included in the model. F SVM cross-validation accuracy vs. number of features plot, showing the accuracy of the SVM model as the number of features increases, with the optimal number of features achieving the best classification performance.

Support vector machine-based refinement of prioritized gene features

To independently validate feature robustness, support vector machine-based recursive feature elimination (SVM-RFE) was employed to refine the prioritized gene set and identify the minimal feature combination required for accurate classification. The generalization error versus number of features plot showed progressive error reduction as informative features were incorporated, allowing identification of an optimal feature subset (Fig. 9E). SVM cross-validation accuracy increased accordingly, with the optimal feature set achieving the highest classification performance (Fig. 9F).

Gene set intersection analysis

Following feature selection, we performed gene set intersection analysis to assess overlap among genes identified by Boruta, SVM-RFE, and top-ranked Random Forest features (Top_RF). The bar plot illustrates both method-specific and shared gene sets, with Boruta identifying the largest number of unique genes (15 genes) (Fig. 10A). Notably, a core set of genes was consistently identified across all three methods, indicating high feature stability. Consistently, Venn diagram analysis revealed a substantial intersection of genes across Boruta, SVM-RFE, and Top_RF (Fig. 10B).

Fig. 10. Consensus gene selection and SHAP-based feature contribution analysis.

A Intersection analysis of genes identified by Random Forest, Boruta, and SVM-based recursive feature elimination (SVM-RFE). B Venn diagram showing the overlap among Boruta-selected genes, SVM-RFE genes, and top-ranked Random Forest genes. C SHAP summary plot of cGAS–STING pathway genes (CGAS, TMEM173 [STING], TBK1, and IRF3). Points represent individual samples and are colored by gene expression level (purple, high; yellow, low). Numbers next to gene names indicate the mean absolute SHAP value.

SHAP-based interpretation of core feature contributions

Across the prioritized feature sets, genes associated with the cGAS–STING signaling axis were consistently ranked among the top contributors by multiple machine learning approaches. To further interpret the contribution of these core genes to disease classification, SHapley Additive exPlanations (SHAP) analysis was applied to the optimized Random Forest model. The SHAP summary plot quantified the direction and magnitude of each gene’s contribution to CP–AD pathology classification, highlighting CGAS, STING (TMEM173), TBK1, and IRF3 as key contributors (Fig. 10C).

Network-based machine learning reveals regulatory centrality of the cGAS–STING pathway

To move beyond feature ranking and explore functional relationships, we applied artificial neural network (ANN) and Bayesian network (BN) inference to model interactions among key genes within the cGAS–STING pathway and their associations with behavioral and pathological phenotypes. ANN-based analysis revealed interaction patterns among core pathway components, including cGAS, STING, TBK1, and IRF3 (Fig. 11A), illustrating the connectivity structure of this signaling axis in the context of AD- and CP-associated pathology.

Fig. 11. Network analysis and gene expression in the cGAS–STING pathway.

A ANN analysis of the cGAS–STING pathway, depicting interactions between key genes (cGAS, STING, TBK1, and IRF3). B BN analysis illustrating the relationships between key genes (cGAS, STING, SARM1, NFkB1) and their influence on behavioral metrics (e.g., cognition, inflammation, neural damage) and pathology scores. C BN analysis of the expression distribution for genes (TBK1, SARM1, NFKB1, PTPN11, RELA, STING, IRF3, and cGAS).

In parallel, BN analysis was used to infer probabilistic relationships among cGAS, STING, SARM1, and NFKB1, as well as their associations with behavioral and pathological measures, including cognition, inflammation, and neural damage (Fig. 11B). In addition, BN modeling of gene expression distributions for TBK1, SARM1, NFKB1, RELA, STING, IRF3, and cGAS provided further insight into the contribution of these genes to disease-associated pathological alterations (Fig. 11C). Notably, BN inference linked cGAS–STING pathway components directly to cognitive impairment and neuroinflammatory markers, supporting their central regulatory role rather than a passive association within the CP-induced AD network.

HLJDD alleviates CP-induced AD-like neuroinflammation by suppressing the cGAS–STING pathway in vivo and in vitro

Building on our single-cell transcriptomic and bulk RNA-seq analyses, as well as machine learning–based modeling, which consistently identified the cGAS–STING axis as a central inflammatory driver linking CP to hippocampal dysfunction, we next performed targeted mechanistic validation experiments to determine whether HLJDD directly modulates this pathway.

Comprehensive characterization of HLJDD using UPLC–Q Exactive Orbitrap HRMS identified 34 major bioactive compounds (Fig. 12A, Table S1) that were enriched in innate immune and neuroinflammatory pathways, including cytosolic DNA-sensing pathways upstream of cGAS–STING and downstream NF-κB signaling pathway (Fig. 12B). To validate this mechanistic implication in vivo, CP markedly increased STING expression in hippocampal microglia, whereas medium- and high-dose HLJDD significantly attenuated STING activation, as demonstrated by immunofluorescence and phosphorylation analyses (Fig. 12C–I).

Fig. 12. HLJDD suppresses cGAS–STING pathway activation in the hippocampus of CP rats.

A Positive (red) and negative (green) ion mode chromatogram. B The significantly enriched KEGG pathways. C Immunofluorescence staining of STING (green) and Iba1 (red) in the hippocampus; DAPI counterstaining (blue). D Western blot of p-TBK1, TBK1, cGAS, p-IRF3, IRF3, and STING. E Quantification of STING fluorescence intensity. F–I Densitometric analysis of cGAS, STING, p-IRF3/IRF3, and p-TBK1/TBK1. The data are presented as the means ± SEMs (n = 6, *p < 0.05, **p < 0.01, ***p < 0.001 vs. the CP group; #p < 0.05, ##p < 0.01, ###p < 0.001 vs. the sham group).

In vitro, gingipain stimulation of BV2 microglia recapitulated CP-like activation of the STING pathway (Fig. 13A). HLJDD did not induce cytotoxicity (Fig. 13B) and dose-dependently reduced nitric oxide (NO) production (Fig. 13C) and iNOS expression (Fig. 13I). Moreover, HLJDD significantly suppressed gingipain-induced phosphorylation of cGAS, STING, IRF3, and TBK1 (Fig. 13D–H), with efficacy comparable to that of the STING inhibitor H-151. Consistently, HLJDD reduced transcription of downstream inflammatory mediators, including IFN-β, TNF-α, IL-1β, IL-6, and CXCL10 (Fig. 13J–N). To further confirm pathway specificity, co-treatment with H-151 enhanced HLJDD-mediated suppression of STING signaling, whereas the STING agonist DMXAA partially reversed these inhibitory effects (Fig. 13O–W).

Fig. 13. HLJDD inhibits gingipain-induced activation of the cGAS–STING–TBK1–IRF3 axis and downstream inflammatory mediators in BV2 cells.

A Experimental timeline of HLJDD and gingipain (10 μg/ml Kgp + 10 μg/ml RgpA) treatment. B CCK-8 assay of BV2 cell viability. C Griess assay for NO production. D Western blot analysis of iNOS and cGAS–STING pathway proteins. E–I Densitometric quantification of iNOS, cGAS, p-STING/STING, p-IRF3/IRF3, and p-TBK1/TBK1. J–N qPCR analysis of the mRNA expression levels of IFN-β, TNF-α, IL-1β, IL-6, and iNOS. O Western blot analysis of cGAS–STING pathway proteins following HLJDD and gingipain treatment, including p-TBK1, TBK1, p-IRF3, IRF3, p-STING, and STING. P–R Densitometric quantification of the ratios p-STING/STING, p-TBK1/TBK1, and p-IRF3/IRF3. (S–W) qPCR analysis of the mRNA expression levels of IL-1β, IL-6, TNF-α, IFN-β, and CCL2. The data are presented as the means ± SEMs (n = 3; *p < 0.05, **p < 0.01, ***p < 0.001 vs. Gingipain; ###p < 0.001 vs. Ctrl).

Taken together, both in vivo and in vitro findings consistently show that HLJDD exerts anti-inflammatory and neuroprotective effects by targeting the cGAS–STING–TBK1–IRF3 signaling axis, thereby attenuating neuroinflammatory activation and downstream proinflammatory mediator production, in agreement with multi-omics and machine learning predictions.

Discussion

In this study, we employed a machine-learning-guided analytical strategy to elucidate the mechanistic link between CP and AD and to identify the key inflammatory pathway modulated by HLJDD. By integrating public single-cell datasets with in vivo hippocampal transcriptomics, multiple machine-learning models consistently prioritized the cGAS–STING axis as the dominant driver of CP-exacerbated neuroinflammation. Guided by this data-driven prioritization, experimental validation confirmed that HLJDD mitigates AD-like pathology primarily through suppression of cGAS–STING signaling, thereby highlighting the value of machine learning in revealing core mechanistic targets of traditional Chinese medicine formulas.

Accumulating epidemiological evidence indicates that CP is a significant risk factor for AD^42,43. Consistent with this association, our analysis of public transcriptomic and single-cell datasets revealed shared inflammatory and immune-related alterations between CP and AD, strengthening the evidence for a mechanistic link between peripheral oral inflammation and neurodegeneration⁴⁴. Guided by this data-driven comorbidity signal, we subsequently established a ligature-induced CP rat model with P. gingivalis challenge to validate disease-relevant phenotypes in vivo. In this model, CP exacerbated AD-like pathological and cognitive impairments, accompanied by hippocampal bacterial colonization, enhanced neuroinflammation, and activation of AD susceptibility gene networks^14,17,18,45. Beyond previously emphasized roles of bacterial virulence factors, such as gingipains and lipopolysaccharide in promoting neuroinflammation and BBB^9,11,15, our findings suggest that chronic periodontal inflammation transcriptionally amplifies intrinsic neurodegenerative vulnerability⁴⁶. Together, these observations support a gene–environment interaction framework in which sustained peripheral inflammatory stimuli prime central immune responses and increase susceptibility to AD-related pathology.

The cGAS–STING pathway has been increasingly implicated in innate immune-driven neuroinflammation and neurodegeneration^19,47. In this study, its central role was not presupposed but emerged consistently from a multi-layered, data-driven analytical framework. Analysis of public single-cell RNA sequencing datasets revealed pronounced activation of cGAS–STING-related signaling in disease-associated microglial populations, indicating cell-type-specific engagement of this pathway in AD-relevant neuroinflammatory states. In parallel, pathway scoring in public comorbidity datasets demonstrated that cGAS–STING signaling was among the most consistently enriched inflammatory pathways shared between CP and AD, strengthening its relevance as a cross-disease mechanistic axis. Building on these population and cell-level observations, machine learning-guided prioritization of disease-relevant hippocampal transcriptomic data further identified cGAS–STING components as high-importance features across multiple complementary models, occupying central positions within inferred regulatory networks. This convergence across single-cell resolution, comorbidity-level pathway scoring, and in vivo transcriptomic modeling supports the designation of cGAS–STING as a dominant regulatory hub rather than a secondary inflammatory pathway^24,48. Biologically, this prioritization is supported by the capacity of P. gingivalis-associated signals to activate cGAS–STING through bacterial- and host-derived DNA, thereby amplifying neuroinflammatory responses and accelerating Aβ deposition and tau pathology. Consistent with this framework, CP-induced inflammation was accompanied by coordinated upregulation of cGAS, STING, TBK1, and IRF3, enhanced IFN-I signaling, and sustained microglial activation. These findings align with recent reports demonstrating that genetic or functional disruption of cGAS–STING signaling mitigates microglial dysfunction and amyloid-associated neurotoxicity in AD models^49–52. Together, these convergent lines of evidence position the cGAS–STING pathway as a machine-learning-prioritized, cell-resolved, and cross-disease inflammatory axis linking periodontal inflammation to central neuroimmune dysregulation in AD.

HLJDD is a classical multi-component formula traditionally used for inflammatory disorders and has been increasingly investigated in the context of neurodegenerative diseases^53–56. In the present study, its therapeutic relevance was interpreted through a mechanism-oriented, data-driven framework rather than a purely symptom- or pathogen-centered perspective. Notably, the cGAS–STING pathway, which was independently prioritized by comorbidity analysis, single-cell profiling, and machine learning-based transcriptomic modeling, emerged as a central regulatory axis targeted by HLJDD. Consistent with this prioritization, HLJDD markedly attenuated CP-induced neuroinflammatory activation at both tissue and cellular levels, including suppression of cGAS–STING–TBK1–IRF3 signaling, reduction of IFN-I responses, and normalization of microglial inflammatory states. The magnitude and directionality of these effects were comparable to those observed with pharmacological STING inhibition, supporting a specific modulatory action of HLJDD on this pathway rather than a nonspecific anti-inflammatory effect⁵⁷. Importantly, downstream amelioration of Aβ accumulation, tau pathology, and cognitive impairment is therefore best interpreted as a consequence of restoring immune network homeostasis rather than isolated antibacterial or cytokine-suppressive actions^58,59. Collectively, these findings reposition HLJDD as a systems-level immunomodulator that converges on a machine-learning-prioritized inflammatory hub linking CP to AD. This work illustrates how integrative computational strategies can refine the mechanistic interpretation of TCM formulas and facilitate their translation into evidence-based, network-informed therapeutic paradigms.

Beyond its biological findings, this study highlights the methodological value of integrating machine learning with multi-omics analyses to advance mechanism-oriented research on traditional Chinese medicine^38,60. By combining public comorbidity datasets, single-cell transcriptomics, and in vivo hippocampal transcriptomic profiling, we adopted a data-driven framework to prioritize disease-relevant regulatory pathways rather than prespecifying targets based on prior assumptions. Similar machine-learning-assisted strategies have been increasingly applied to uncover core mechanisms of complex diseases and to decode the multi-target actions of herbal formulas^61,62, demonstrating their utility in improving robustness and interpretability in TCM research^37,60,63. Conceptually, our work reflects a shift from descriptive or experience-based interpretations of TCM efficacy toward network-informed and mechanism-prioritized modeling^38,64. The identification of cGAS–STING signaling as a dominant inflammatory hub shared between CP and AD exemplifies how machine learning can guide experimental validation and refine the mechanistic attribution of multi-component formulas. This paradigm is consistent with emerging views that artificial intelligence provides a critical bridge between holistic TCM theory and modern systems biology, enabling scalable and reproducible discovery of key therapeutic axes.

Several limitations should be noted. Although convergent computational and experimental evidence support the involvement of cGAS–STING signaling, causal validation using genetic or pharmacological perturbation remains necessary. In addition, the interpretability of complex machine learning models and cross-species translational relevance warrant further investigation. Future studies incorporating explainable artificial intelligence, causal inference strategies, and larger human datasets may further enhance the precision and clinical relevance of machine learning-guided TCM research.

In summary, this study demonstrates that the multi-component formula HLJDD exerts therapeutic effects on CP-induced AD through coordinated regulation of a machine learning-prioritized immune network. This integrative framework highlights the advantage of traditional Chinese medicine formulas and provides a novel paradigm for mechanism-oriented TCM research.

Methods

Single-cell RNA-seq data processing and analysis

Single-cell RNA sequencing data from AD and NC samples were obtained from the Gene Expression Omnibus (GEO; accession GSE157827). Raw count matrices were processed in R using Seurat (v5.0.3). Cells expressing fewer than 300 genes or genes detected in fewer than five cells were excluded. Additional quality control filters were applied to remove cells with mitochondrial gene content >25%, ribosomal gene content >50%, or erythrocyte-associated gene expression >1%. Doublets were identified using DoubletFinder (v0.2.3), with the expected doublet rate set to 7%, and predicted doublets were excluded from downstream analyses. Data were normalized using SCTransform, and the top 2000 highly variable genes were selected for PCA. Batch effects associated with individual patients were corrected using Harmony (v0.1.0), enabling integrated analysis across samples. Cells were clustered using the Louvain algorithm (resolution = 0.8) on a k-nearest neighbor graph (K = 20) and visualized using Uniform Manifold Approximation and Projection (UMAP). Cell types were annotated based on established canonical marker genes, and neuronal clusters with overlapping transcriptional profiles were consolidated into major neuronal categories for downstream analyses. Differential gene expression between AD and NC groups was assessed using the Wilcoxon rank-sum test with Benjamini–Hochberg correction, and genes with an adjusted p-value < 0.05 were considered significant. Expression patterns were visualized using UMAP feature plots, dot plots, and heatmaps, with a focus on inflammatory mediators and key components of the cGAS–STING pathway (cGAS, TMEM173, TBK1, IRF3). The effectiveness of batch correction was evaluated by visual inspection of sample mixing in low-dimensional embeddings (Fig. S3).

RNA-seq processing and transcriptome analysis

Raw count data were imported into R and processed using standard pipelines. Duplicate gene identifiers were removed, and genes with zero variance across all samples were excluded. Expression values were normalized using a variance-stabilizing transformation. Samples were assigned to Control, Model, and Treatment groups (n = 6 per group), with group labels extracted from sample names and manually verified. PCA was performed for dimensionality reduction. PLS-DA was subsequently applied to identify transcriptional features that maximally separated predefined groups, with model performance evaluated using R²Y and Q² metrics. DEGs were defined using a threshold of |log2 fold change| > 0.585 and p < 0.05, and intersections of DEGs across experimental conditions were visualized using UpSet plots to identify core genes with consistent expression changes. Functional annotation of core DEGs was performed using Gene Ontology (GO) and KEGG pathway enrichment analyses with the clusterProfiler package. Enrichment results were visualized using bar plots, bubble plots, and network representations. All analyses were conducted in R (v4.1.0). Data preprocessing, visualization, and enrichment analyses were performed using ggplot2 (v3.4.0), clusterProfiler (v4.6.0), ropls (v1.30.0), and UpSetR (v1.4.0).

Machine learning analysis

Machine learning analyses were performed in R (v4.1.0). Gene expression matrices were preprocessed by removing duplicated gene identifiers and genes with zero variance. Samples were grouped according to experimental conditions, and stratified sampling was applied to preserve class balance. All models were evaluated using five-fold stratified cross-validation to minimize overfitting.

Random Forest classifiers were constructed to rank disease-relevant features, with model hyperparameters optimized by grid search combined with cross-validation. Feature importance was assessed using the mean decrease Gini index, and Boruta analysis was applied to identify robust and non-redundant gene features. Model performance was further evaluated using out-of-bag error estimates.

To further refine the feature set, support vector machine-recursive feature elimination (SVM-RFE) with a linear kernel was applied, and the optimal feature subset was selected based on cross-validation error minimization.

Artificial neural network (ANN) models were constructed using a single hidden layer to balance model complexity and generalization. Model performance was evaluated using classification accuracy and confusion matrices.

Model performance was assessed using accuracy, F1 score, and area under the receiver operating characteristic curve (AUC-ROC). To enhance model interpretability, SHapley Additive exPlanations (SHAP) analysis based on an XGBoost classifier was applied to quantify the direction and magnitude of individual gene contributions to disease classification.

Model performance evaluation and robustness analyses are provided in Supplementary Figs. S4–S6.

Preparation of HLJDD

All the herbal materials used were purchased from Xinhehua (Chengdu) Co., Ltd. A mixture of Huanglian (Coptis chinensis, Batch 2306085, 225 g), Huangqin (Scutellaria baicalensis, Batch 2305034, 225 g), Huangbai (Phellodendron amurense, Batch 2308321, 150 g), and Zhizi (Gardenia jasminoides, Batch 2310067, 225 g) at a ratio of 3:2:2:3 was soaked in 10 volumes of purified water for 1 h. The mixture was then boiled at high heat and simmered for three cycles, each lasting 40 min. After each cycle, the filtrate was filtered through filter paper. The combined filtrates were concentrated via a rotary evaporator and then freeze-dried into powder. The yield of the lyophilized powder was ~35%. Detailed information on the herbal ingredients and medicinal parts of HLJDD is provided in Table S1.

Qualitative analysis of HLJDD

The experiments were conducted on a Thermo Scientific UPLC system (Waltham, MA, USA) combined with high-resolution MS (HRMS, Q Exactive Orbitrap). Chromatographic separation was performed on an Ultimate UHPLC XB-C18 (2.1 mm × 100 mm, 1.8 µm). The mobile phase consisted of water, 0.1% formic acid (eluent A), and acetonitrile (eluent B) at a flow velocity of 0.3 mL min⁻¹, and the column temperature was 30 °C. The sample injection volume was 5 μL. The gradient program was as follows: 0–4 min, 95–75% A; 4–15 min, 75–20% A. MS was operated through an electrospray ionization source in positive and negative ion modes. The scanning mode was full-scan mode (Full-MS) with a scanning range of m/z 100–1500 Da (resolution: 35,000) and a data-dependent ms2 scan (dd-MS², resolution: 17,500). The data acquisition and manipulation were carried out with Xcalibur software (version 3.1). Metabolites were identified by matching with an online database (mzCloud) and a local traditional Chinese medicine ingredient database (OTCML). The components with a matching degree above 80 were defined and compared with the MS/MS information of the related literature and online databases, including PubChem, ChemSPEID, and SciFinder.

Animals

Healthy male Sprague–Dawley rats (200–230 g), obtained from Dashuo (Chengdu) Laboratory Animal Co., Ltd., SPF grade, were housed in the TCM Regulating Metabolic Diseases Key Laboratory of Sichuan Province. The rats were maintained at 20–23 °C with free access to food and water, allowed to acclimate for one week, and tested at a body weight of 250 ± 20 g (Approval No. 2024DL-014).

CP model and drug administration

The CP model was established using the ligature method. All surgical procedures were performed under general anesthesia to minimize animal suffering. Rats were anesthetized by intraperitoneal injection of 2% sodium pentobarbital (50 mg/kg; Merck, cat. no. 230816, NJ, USA) and were confirmed to be fully unconscious prior to ligation by the absence of pedal withdrawal reflex. Under sterile conditions, 4–0 silk sutures (Jinhuan Medical Products Co., Shanghai, China) were placed around the left and right second maxillary molars and maintained throughout the experimental period to facilitate bacterial accumulation. To enhance periodontal inflammation, 100 μL of Porphyromonas gingivalis suspension (1 × 10⁸ CFU) was applied locally to the ligation sites every other day. In the sham-operated group, sutures were placed but not ligated.

At the experimental endpoint, rats were deeply anesthetized with an overdose of sodium pentobarbital and subsequently euthanized for tissue collection, in accordance with institutional guidelines and approved animal care protocols. The ligated rats were randomly divided into the following groups: periodontitis (CP), periodontitis + doxycycline (CP+Doxy, 10 mg/kg, 412S031, Solarbio, Beijing, China), periodontitis + HLJDD-L (CP + HLJDD-L, 0.75 g/kg), periodontitis + HLJDD-M (CP + HLJDD-M, 1.5 g/kg), and periodontitis + HLJDD-H (CP + HLJDD-H, 3 g/kg), with 8 rats per group. Starting 4 weeks post-ligation, HLJDD or doxycycline was administered orally every day for 6 weeks. The sham-operated group received an equivalent volume of normal saline.

All animal experiments were conducted in accordance with the guidelines for the care and use of laboratory animals and complied with relevant national and institutional regulations. The experimental protocols were reviewed and approved by the Ethics Committee of the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine (Approval No. 2024DL-014).

HE staining

Rat brains were fixed in 4% paraformaldehyde (P1110; Solarbio, Beijing, China) for 24 h. Rat maxillary bone samples were decalcified by immersion in EDTA decalcification solution (E1171, Solarbio, Beijing, China) at 4 °C for 4 weeks. The samples subsequently underwent routine histological processing, including dehydration, trimming, embedding, sectioning, staining, and sealing. Histopathological changes in the rat hippocampus were observed after whole-slide scanning.

Nissl staining

Paraffin-embedded rat brain sections (5 μm) were deparaffinized, rehydrated, and stained with 0.1% toluidine blue (G3668, Solarbio, Beijing, China) solution at room temperature for 10 min. After being rinsed in distilled water, the sections were dehydrated through a graded ethanol series, cleared in xylene, and mounted with neutral resin. Neuronal morphology and Nissl substance distribution were observed via a digital pathology scanning system (SQS-12P, Shengqiang, China).

Micro-CT

The fixed rat maxillary specimens were scanned via micro-CT (QUANTUM GX2). All scans were redirected to the same location at 90 kV and 80 µA for bone loss assessment. The images were reoriented via anatomical landmarks to measure the distance between the odontoblastic-odontoblastic junction and the alveolar bone crest.

MWM test

The learning and memory functions of the rats were assessed via the WMT-100S system (TaiMeng, Sichuan, China). The maze consisted of a circular pool with a diameter of 120 cm, which was divided into four quadrants. The pool was filled with opaque water containing black, nontoxic dye and maintained at ~22 °C. In the place navigation test, a platform (12 cm in diameter) was submerged 2 cm below the surface in the center of the fourth quadrant. The rats were allowed to explore freely for 60 s, and the escape latency was recorded over 4 consecutive days via a video tracking system. On the 5th day, the platform was removed for the spatial probe test, and the number of platform crossings, number of entries into the target zone, and percentage of time spent in the target zone within 60 s were recorded.

Western blot

Proteins from rat hippocampal tissues and BV2 cells were extracted with RIPA buffer (R0010, Solarbio, Beijing, China) and centrifuged, and the resulting supernatants were quantified via a BCA (YK378668, Thermo, USA) assay. After denaturation at 100 °C for 10 min, the samples were subjected to SDS‒PAGE (PG112, Epizyme, Shanghai, China), transferred onto PVDF membranes (Millipore Corp., Bedford, MA, USA), blocked with 5% skim milk (EZ66E5BBA2, Biofroxx, Germany) for 1 h, and incubated overnight at 4 °C with primary antibodies (β-actin 1:10000 #AF7018; Tau 1:2000 #AF6141; p-Tau(Ser396) 1:1000 #AF3148; TBK1 1:2000 #DF7026; p-TBK1 1:2000 #AF8190; IRF3 1:2000 #DF6895; p-IRF3(Ser396) 1:1000 #AF2436; p-STING(Ser366) 1:1000 #AF7416; iNOS 1:2000 #AF0199 from Affinity, Jiangsu, China). Aβ 1:2000 Cat No.25524-1-AP; STING 1:10000, Cat No. 19851-1-AP; cGAS 1:2000, Cat No. 84045-1-RR from Proteintech, Wuhan, China). The blots were then incubated with secondary antibody (anti-rabbit IgG, 1:10000 #S0001 from Affinity, Jiangsu, China) for 1 h and analyzed via a gel imaging analysis system (Tanon-46008F).

Immunofluorescence

The brain tissue paraffin sections (6 µm) were processed for immunofluorescence. The sections were blocked with 5% BSA for 1 h, followed by an overnight incubation with primary antibodies (Iba1, 1:500; STING, 1:500; Proteintech, Wuhan, China) at 4 °C. After being washed with PBS, the sections were incubated with fluorophore-conjugated secondary antibodies (Cy3-conjugated goat anti-rabbit IgG (H + L) [red], FITC-conjugated goat anti-rabbit IgG (H + L) [green], 1:500; Proteintech, Wuhan, China) for 1 h at room temperature. The sections were then counterstained with DAPI and mounted. Images were captured via a laser scanning confocal microscope (TCS SP8, Leica) and analyzed via ImageJ software.

ELISA

The serum and cerebrospinal fluid of the rats were collected, and the levels of TNF-α (JM-01587R1), IL-1β (JM-10482R1), IL-6 (JM-01597R1), IL-4 (JM-01598R1), IL-10 (JM-01602R1), and Aβ_1-42 (JM-01630R1) were measured via commercial ELISA kits (JINGMEI, Jiangsu, China) according to the manufacturer’s instructions. The absorbance was recorded at 450 nm via a microplate reader, and the concentrations of the target proteins were calculated accordingly.

Real-time quantitative PCR

Total RNA from tissues and cells was extracted via the Bacterial and Tissue/Cell Total RNA Extraction Kit (B511361, Sangon Biotech, Shanghai, China) and reverse-transcribed into cDNA. qPCR was performed via the use of SYBR Green Premix (175037; Qiagen, Hilden, Germany). The cycling parameters included an initial step at 95 °C for 10 min, followed by 42 cycles of 95 °C for 15 s and 60 °C for 1 min. Relative gene expression was calculated via the 2^−ΔΔCt method, with normalization to the internal controls 16S (bacteria) and β-actin (tissues and cells). All the assays were performed in triplicate. The primers used are listed in Table S2.

Cell culture and treatment

Mouse microglia (BV2, CTCC-003-0003, Meisen, Zhejiang, China) were cultured in DMEM (Gibco, California, USA) supplemented with 10% fetal bovine serum (Biochannel, Nanjing, China) and 1% penicillin‒streptomycin (Cytiva, USA). The cells were incubated at 37 °C in a humidified atmosphere containing 5% CO₂. After 24 h of culture in dishes, the cells were pretreated for 1 h with the STING pathway inhibitor H-151 (2 µM, AbMole, USA) or with HLJDD at concentrations of 100 µg/mL, 200 µg/mL, and 400 µg/mL. The cells were subsequently stimulated with gingipains (RgpA 10 µg/mL and Kgp 10 µg/mL; Abcam, USA) for 16 h. The culture supernatants or cell lysates were then collected for further tests.

Cell viability

BV2 cells were seeded into 96-well plates and treated with 0, 50, 100, 200, 400, 800, or 1000 μg/mL HLJDD for 16 h. Subsequently, CCK-8 solution (AbMole, USA) was added to each well. After incubation for 1 h, the absorbance was measured at 450 nm via a multifunctional microplate reader.

NO assay

The concentration of nitric oxide (NO) in the cell culture supernatant was measured via a commercial NO detection kit (Beyotime, Shanghai, China) according to the manufacturer’s instructions. After the reaction, the absorbance was measured at 540 nm via a microplate reader, and NO levels were calculated on the basis of the standard curve.

Statistical analysis

The statistical analysis in this study was performed using GraphPad Prism 8 software. Data are presented as means ± SEM. For the MWM data, two-way analysis of variance (ANOVA) was used, while other data were analyzed with one-way ANOVA followed by Tukey’s post hoc test. A p-value < 0.05 was considered statistically significant.

For bioinformatics analysis, all enrichment analyses used the Benjamini–Hochberg method to control the false discovery rate, with a threshold of q < 0.05. Multiple comparisons were corrected using the p.adjust function to reduce false positives. In single-cell RNA sequencing, gene expression differences between the AD and NC groups for each cell type were compared using the Wilcoxon rank-sum test, with a significance threshold of p < 0.05. Gene expression levels were represented as Z-scores of normalized counts. For comparisons of continuous variables, the Mann–Whitney U test was used, and multiple test correction was performed using the Benjamini–Hochberg method. All analyses were conducted in the R (v4.1.0) environment, primarily using the Seurat (v4.3.0) and ggplot2 (v3.4.0) packages.

Author contributions

All the authors read and approved the manuscript. Wenbin Wu, Guohua Zhao, Yuzhen Xu, and Jie Li designed and supervised the studies. Jie Li and Mingqi Chen performed the experiments with the help of Pan Ren, Furong Zhong, Guangming Sun, Yue Zhu, Yiran Fan, Jinxin Chen, Manru Xu, Mengyuan Qiao, and Ganggang Li. Jie Li and Guangming Sun wrote the manuscript.

Data availability

The single-cell RNA sequencing data analyzed in this study were retrieved from the GEO database under accession number GSE157827. Bulk RNA sequencing data for periodontitis were obtained from GEO under accession number GSE23586. The rat hippocampal RNA sequencing data generated in this study have been deposited in the NCBI Sequence Read Archive (SRA) under accession number PRJNA1394501. All other data supporting the findings of this study are available from the corresponding authors upon reasonable request.

Code availability

All custom code used for data preprocessing, statistical analysis, and machine learning modeling in this study was written in R (v4.1.0). Analyses were performed using established R packages including Seurat (v4.3.0), ggplot2 (v3.4.0), clusterProfiler, randomForest, Boruta, e1071, caret, neuralnet, bnlearn, ropls, and UpSetR. The code is available from the corresponding authors upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-026-02468-x.