This study enrolled 141 participants. All procedures were conducted in accordance with the Declaration of Helsinki and were approved by the Clinical Ethics Committee of The First Affiliated Hospital of Shantou University Medical College (2020-115-XZ2, B-2022-232). Written informed consent was obtained from all participants before inclusion. AD was diagnosed using a combination of clinical assessments, neuropsychological evaluations, and neuroimaging examinations, following the revised NINCDS-ADRDA criteria endorsed by the International Working Group in 2007. Participants were excluded if they had experienced a clinical stroke within three months before enrollment; presented with MRI-confirmed regional cerebral infarction; had autoimmune or systemic inflammatory diseases; had a history of malignancy; or had received treatment with non-steroidal anti-inflammatory drugs or corticosteroids, as such medications may influence circulating cytokine concentrations.

Cognitive function was assessed in all participants using locally adapted versions of the Mini-Mental State Examination (MMSE, 30-point scale) and the Montreal Cognitive Assessment (MoCA). The MMSE comprised assessments of orientation, registration, recall, attention and calculation, and language. The MoCA consisted of seven domains: visuospatial/executive function, naming, memory and delayed recall, attention, language, abstraction, and orientation. All participants were allowed to use personal visual aids (e.g., corrective glasses) as needed and completed the test in their preferred language to ensure optimal performance and communication. The MMSE was administered first, followed by the MoCA to minimize potential learning or fatigue effects.

MRI examinations were performed on a 3.0-T scanner. Comprehensive imaging sequences were used, including T1 fluid-attenuated inversion recovery (FLAIR), T2 FLAIR, T2-weighted imaging, sagittal T2 fast spin-echo, and diffusion-weighted imaging, to exclude unrelated intracranial abnormalities. MTA was independently assessed by a neurologist and a radiologist. Any discrepancies in scoring were resolved by consensus. Abnormal MTA was defined according to established age-adjusted criteria: MTA score ≥ 1 for individuals < 65 years, ≥ 2 for those 65–74 years, and ≥ 3 for individuals ≥ 75 years.

Peripheral blood samples were collected from all participants and processed immediately. Samples were centrifuged at 300 × g for 5 minutes at 4 °C. A 1-mL aliquot of whole blood was reserved for APOE genotyping, while the remaining 3 mL was used for plasma and peripheral blood mononuclear cell isolation. The supernatant obtained after centrifugation, corresponding to the plasma fraction, was carefully harvested and stored for subsequent quantification of inflammatory proteins.

Plasma concentrations of inflammatory biomarkers were quantified using Luminex liquid bead–based multiplex technology according to the manufacturer’s protocols. The panel included TNF-α, IFN-γ, IL-1β, IL-5, IL-6, IL-7, IL-8, IL-10, IL-13, IL-17, IL-18, IL-33, MCP-1, soluble TREM-1, TSLP, and CCL11.

APOE genotyping was performed using the SNaPshot Multiplex System, following procedures described previously. The assay targeted two single-nucleotide polymorphisms, rs429358 and rs7412, located in exon 4 of the APOE gene. APOE genotype was determined based on the combination of ϵ2, ϵ3, and ϵ4 alleles. Individuals carrying at least one ϵ4 allele (ϵ4/ϵ4, ϵ3/ϵ4, or ϵ2/ϵ4) were classified as APOE ϵ4 carriers, whereas all other genotypes were categorized as non-APOE ϵ4 carriers.

Classification models were constructed using least absolute shrinkage and selection operator (LASSO) regression within a nested cross-validation (CV) framework, which inherently reduces the risk of multicollinearity through coefficient shrinkage and variable selection. The inner 5-fold CV was used to determine the optimal regularization parameter (λ), and the outer 5-fold CV was applied for independent model evaluation. The dataset was randomly divided into five approximately equal folds; in each iteration, one fold was used as the test set and the remaining four folds as the training set. Predictors with missing values were excluded. Categorical variables were one-hot encoded, and all predictors were centered and scaled prior to model fitting. Model performance was assessed using receiver operating characteristic (ROC) analysis. Final models were refitted using the optimized λ identified by cv.glmnet on the full dataset.

Single-nucleus RNA-sequencing (snRNA-seq) data from fresh-frozen frontal cortex samples of individuals with AD carrying the APOE4/4 genotype, individuals with AD carrying the APOE3/3 genotype, and age- and sex-matched cognitively normal APOE3/3 controls were retrieved from the GSE254205 dataset. Gene expression matrices were log-normalized and scaled prior to downstream analyses. Highly variable features were identified using Seurat’s variance-stabilizing transformation. Principal component analysis (PCA) was conducted on the top 2,000 highly variable genes using the RunPCA function. The top 50 principal components were then used to construct a shared nearest-neighbor graph with FindNeighbors. Dimensionality reduction was further visualized in two dimensions using the Uniform Manifold Approximation and Projection algorithm (RunUMAP). Major cell-type annotations were obtained from the original study (20).

Subsequent analyses focused specifically on microglia. To minimize inter-sample variability, Harmony was applied to correct batch effects within the microglial subset. Microglial subclusters were identified using a clustering resolution of 0.4 and were annotated according to canonical marker genes and cluster-specific transcriptional signatures. To quantify pathway activity, inflammatory response and IFN-γ response scores were computed for each major cell type and for each microglial subcluster using gene sets from the HALLMARK collection (h.all.v7.4.symbol, https://www.gsea-msigdb.org/gsea/msigdb/, Supplementary Table 1) and the AUCell algorithm.

HMC3 were purchased from the Procell Life Science & Technology Co. (Procell, Wuhan, China). Cells were authenticated by STR profiling and tested negative for mycoplasma contamination. HMC3 were cultured in MEM supplemented with non-essential amino acids, 10% fetal bovine serum (Sigma-Aldrich, cat#F0193, USA), and 1% penicillin/streptomycin (100 U/mL, GIBCO, cat#15140122, USA). For IFN-γ-related experiments, 3 x 10⁵ cells were seeded onto 6-well plates (Guangzhou Jet Bio-Filtration Co., Ltd., cat#TCP010006) and treated with 100 ng/mL IFN-γ (MedChemExpress, cat#HY-P7025, USA) for 24 h before RNA and protein extraction.

Lentiviral expression plasmids pLVX-Ctrl-Flag-ZsGreen-Puro, pLVX- APOE2-Flag-ZsGreen-Puro, pLVX-hAPOE3-Flag-ZsGreen-Puro, and pLVX-hAPOE4-Flag-ZsGreen-Puro were constructed by Dongze Co. Each plasmid was transfected into HEK293T packaging cells to generate high-titer lentiviral particles carrying the corresponding transgenes. Viral supernatants were collected after 48 h, filtered, and concentrated. HMC3 were then transduced with the respective lentiviruses in the presence of 8 μg/mL polybrene, followed by 2 μg/mL puromycin selection. Successful transduction was confirmed by ZsGreen fluorescence, real-time qPCR, and immunoblotting of Flag-tagged proteins.

Total RNA was extracted by Total RNA Extraction Kit (Goonie, cat# 400-100, Guangzhou, China), and cDNA was synthesized with the HiScript III All-in-one RT SuperMix (Vazyme, cat# R333-01, Nanjing, China) using 1 μg RNA as input. Real-time qPCR was performed with the Fast Taq SYBR Green qPCR Mix (Goonie, cat# 500-102). GAPDH was used as an internal control. Primer sequences used for real-time qPCR were listed in Supplementary Table 2.

Cells were washed with ice-cold PBS and lysed in RIPA buffer supplemented with protease and phosphatase inhibitors. Equal amounts of protein were separated by 7.5% SDS-PAGE and transferred onto polyvinylidene difluoride membranes (0.22 μm pore size; Merck Millipore, cat# ISEQ00010, Germany). Membranes were blocked with 5% skim milk for 1 h at room temperature, incubated with primary antibodies overnight at 4 °C, washed six times with TBST, and then incubated with horseradish peroxidase–conjugated secondary antibodies (1:8000) for 1 h at room temperature. Protein signals were visualized using enhanced chemiluminescence reagents on an automated imaging system (Bio-Rad ChemiDoc XRS). Primary antibodies applied in this study include antibodies against APOE4 (1:1000, CST, cat# 8941), ACSL1 (1:1000, CST, cat# 9189), Flag (1:20000, Proteintech, cat# 20543-1-AP), and GAPDH (1:10000, Proteintech, cat# 60004-1-Ig).

All statistical analyses and data visualizations were performed using GraphPad Prism (version 7.0) and R (version 4.4.3). Receiver operating characteristic (ROC) curve analysis and calculation of the area under the curve (AUC) for binary classification were conducted using the “pROC” package in R. Spearman’s rank correlation was applied to assess associations between continuous variables. For comparisons of continuous variables between 2 groups, either Student’s t-test or the Wilcoxon rank-sum test was used. All statistical tests were two-sided, and P < 0.05 was considered statistically significant.

A total of 143 participants were included for analyses, with 71 patients diagnosed with AD, 44 patients with mild cognitive impairment (MCI), and 28 cognitively healthy controls (HC) in the First Affiliated Hospital of Shantou University Medical College. Participant characteristics are summarized in Table 1. As expected, both MMSE and MoCA scores, which reflect global cognitive performance, were markedly reduced in individuals with AD. Consistent with previous reports, patients with AD were generally older, had fewer years of formal education, and exhibited a higher proportion of APOE ϵ4 genotype. Sex distribution did not differ significantly between the AD and non-AD groups. Altogether, these demographic patterns closely mirror those described in other cohorts, supporting the representativeness and reliability of our study population.

Next, we measured plasma levels of 16 inflammatory biomarkers (Figure 1A). The results revealed significant increases in IFN-γ, IL-33, and IL-18, accompanied by significant reductions in IL-7, IL-6, and CCL11 in patients with AD (Table 1; Figure 1B). We then further interrogated their associations with cognitive levels (Figure 2A) and observed remarkable negative correlations between IFN-γ and MMSE (R = -0.47, P < 0.001) or MOCA (R = -0.46, P < 0.001) scores (Figure 2B). In addition, both IL-33 and IL-18 were negatively correlated with MMSE (IL33: R = -0.41, P < 0.001; IL-18: R = -0.25, P < 0.01) or MOCA (IL33: R = -0.47, P < 0.001; IL-18: R = -0.29, P < 0.001). In contrast, higher IL-7, IL-8, and TSLP were significantly associated with higher MMSE or MOCA scores (Figure 2A). Together, these alterations indicate a profound dysregulation of systemic inflammatory mediators in AD, potentially reflecting an underlying shift in neuroinflammatory activity.

We then applied the LASSO algorithm within a 5 × 5 nested CV framework to optimize λ and objectively compare model performance. As shown in Figure 3A, the optimized model incorporating clinical variables and APOE genotype distinguished AD from HC with a mean out-of-fold AUC of 0.863. In comparison, the model based solely on inflammatory biomarkers achieved a higher mean AUC of 0.890, and the integrated model combining clinical variables, APOE genotype, and plasma protein measurement yielded the highest mean AUC of 0.953, indicating that inflammatory biomarker alterations provide additional discriminative information beyond conventional clinical variables. Subsequently, final models were established using all included participants and demonstrated a consistent pattern of predictive performance (Figure 3B; Supplementary Figure S1).

Next, the inferred LASSO coefficients were used to quantify the relative contribution of each feature to the finalized joint model. As shown in Figure 3C, a set of protective factors, such as longer duration of education, IL-7, and CCL11, was associated with a reduced probability of AD. In contrast, several variables exerted strong risk-enhancing effects, including IFN-γ, older age, APOEϵ4 carrier status, IL-13, IL-17, and IL-18, all of which shifted model predictions toward AD when increased or present. Notably, IFN-γ exhibited the largest coefficient magnitude, indicating its dominant role in driving model predictions. In line with this, ROC analysis identified IFN-γ as the most predictive marker for AD among 16 proteins (AUC = 0.913, Figures 3D, E), and further demonstrated its ability to distinguish AD from MCI individuals (AUC = 0.789, Supplementary Figure S2). Collectively, these findings support a potential role for IFN-γ in mediating inflammatory processes in AD pathogenesis.

Recent studies have highlighted a strong link between neuroinflammation and APOE4 genotypes in AD. We observed significantly higher plasma IFN-γ levels in APOE ϵ4 carriers compared with noncarriers (Figure 4A). Subgroup analyses revealed that this elevation was specific to individuals with AD, as no significant differences were observed in HC or MCI subjects (Figure 4B). These findings led us to hypothesize that IFN-γ signaling may be selectively upregulated in specific brain cell types of APOE4/4 AD patients.

To investigate this, we analyzed published snRNA-seq data generated from 26 human cortex samples (Figures 5A-C), including AD patients with APOE4/4 (n = 10) or APOE3/3 (n = 8) genotypes, and cognitively normal APOE3/3 controls (n = 8) (20). Major brain cell types were annotated according to the original study and further verified using canonical marker genes (Figure 5D), which included astrocytes (AQP4, GFAP), endothelia (CLDN5, VWF), neurons (RBFOX3, SYT1), microglia (P2RY12, CD74), and oligodendrocytes (MBP, PLP1). Notably, microglia were the primary cell type (6.63% of all annotated cells) exhibiting significant enrichment in the inflammatory response (Figure 5E). Moreover, microglia from APOE4/4 AD patients displayed markedly higher activity in the inflammatory and IFN-γ pathways compared with those from APOE3/3 AD patients or HC (Figures 5F, G), suggesting a genotype-specific amplification of IFN-γ-driven inflammatory signaling in APOE4-associated AD.

Further subclustering of microglia identified three transcriptionally distinct subtypes: (i) disease-associated microglia (DAM), characterized by high expression of SPP1, CD63, TREM2, and APOE; (ii) lipid droplet-accumulating microglia (LDAM), marked by elevated ACSL1, NAMPT, DPYD, and CD163; and (iii) homeostatic microglia, defined by P2RY12, P2RY13, and CX3CR1 (Figures 5H-J). In APOE4/4 AD, LDAM were markedly expanded, whereas homeostatic microglia were reduced (Figure 5K). LDAM has been validated to promote AD-related pathology (20, 23). Our result showed that LDAM exhibited significantly increased IFN-γ pathway activity (Figures 5L, M), suggesting that IFN-γ may contribute to APOE4/4-driven AD pathology by promoting the expansion or maintenance of the LDAM population.

To validate our findings, we first established HMC3 microglial cell lines stably overexpressing empty vector, ApoE2, ApoE3, or ApoE4 (Figure 6A). LDAMs are primarily defined by the expression of the lipid droplet-associated enzyme ACSL1, and overexpression of ACSL1 is sufficient to induce lipid droplet formation in multiple cell types, including microglia (20, 24). We confirmed that ApoE4 overexpression markedly increased both ACSL1 mRNA and protein levels in HMC3 cells (Figures 6B-D). We next evaluated whether IFN-γ modulates the LDAM phenotype by treating the cells with exogenous IFN-γ. As anticipated, IFN-γ stimulation robustly induced canonical IFN-γ-responsive genes, including CXCL10 and IRF1 (Figures 6E, F), indicating effective pathway activation in HMC3. Importantly, we found that IFN-γ further augmented ACSL1 transcription and protein expression specifically in ApoE4-overexpressing cells (Figures 6G-I). Together, these data demonstrate that IFN-γ amplified ACSL1 expression and promoted an ACSL1-driven LDAM phenotype in the context of ApoE4.