SRSF1 as a promising biomarker and key player in the pathogenesis of Alzheimer's disease

Introduction

Alzheimer's disease (AD), a gradual neurodegenerative condition, is the leading cause of dementia and places a notable burden on health worldwide (1). Initially presenting as mild memory deficits, AD advances to cause marked cognitive impairment, personality changes and language dysfunction, severely compromising the quality of life and independence of patients (2).

Advancements in biomarker research have markedly enhanced AD diagnosis. Positron emission tomography (PET) scans, along with plasma measurements of amyloid β (Aβ) and p-tau proteins, have emerged as notable diagnostic tools (3). The Aβ42/Aβ40 ratio, along with p-tau181 and p-tau217, have emerged as pivotal plasma biomarkers for AD diagnosis, prognosis and pathological understanding (4). Additionally, neurofilament light chain (NfL) and glial fibrillary acidic protein (GFAP) have emerged as pivotal plasma biomarkers for AD (5). These biomarkers act not just as key resources for identifying the condition early on, but also enable tracking of disease progression, whilst yielding important insights into the pathological mechanisms underlying AD. In doing so, they aid in developing therapeutic methods that are more precisely targeted and more effective (6).

Beyond the well-characterised amyloid and tau pathologies, AD pathophysiology involves a complex interplay of multiple biological systems (7). Microglial activation and astrocyte reactivity trigger neuroinflammatory cascades, whereas vascular dysfunction and blood-brain barrier disruption exacerbate neuronal damage (8). The peripheral immune system also plays a notable role, with T lymphocytes and B cells infiltrating the central nervous system, thereby promoting neuroinflammation and synaptic loss (9). Additionally, emerging evidence suggests that the glymphatic system, responsible for the clearance of ‘brain waste’, and the gut microbiome may contribute to AD development and progression (10). Substantial findings underscore the vital function of immune dysregulation in AD (11). Specifically elevated T cell infiltration into the brain promotes crosstalk with microglia, intensifying neuroinflammation (12). Peripheral B lymphocytes can cross the blood-brain barrier and interact with resident microglia and other brain immune cells, further activating the immune response (13). Natural killer cells, dendritic cells and mast cells have likewise been linked to a higher risk of AD, which underscores how crucial both innate and adaptive immune reactions are in the development of the disease (14).

Current AD treatments focus on managing symptoms and slowing disease progression, including cognitive enhancement therapies, treatment of neuropsychiatric symptoms and disease-modifying therapies (15). Disease-modifying treatments for AD have emerged as a marked advancement, demonstrating remarkable potential to alter the disease trajectory. Among these, lecanemab and donanemab, both cutting-edge monoclonal antibodies, have garnered notable attention. These agents specifically target and bind to β-amyloid proteins, initiating a cascade of events that facilitate the clearance of these pathogenic aggregates (16). However, despite ongoing research efforts, no cure exists, and several experimental drugs have shown limited efficacy (17). This highlights the pressing requirement for innovative approaches to diagnosis and treatment.

Despite these advancements, the lack of curative treatments and limited early diagnostic tools underscores the need for robust biomarkers capable of detecting AD at pre-symptomatic or prodromal stages. In the present study, Weighted Gene Co-expression Network Analysis (WGCNA) was used in combination with machine learning methods to identify new candidate hub genes associated with AD progression. A comprehensive flowchart illustrating the study's procedure is presented in Fig. 1. The aim of the present study was to identify novel targets for clinical diagnosis and treatment, potentially paving the way for more effective interventions to combat AD.

Figure 1. Workflow of the present study. LASSO, Least Absolute Shrinkage and Selection Operator; SVM-RFE, Support Vector Machine-Recursive Feature Elimination; XGboost, Extreme Gradient Boosting; ROC, receiver operating characteristic; GSVA, gene set variation analysis; AD, Alzheimer's disease; ssGSEA, single-sample gene set enrichment analysis; PBMCs, peripheral blood mononuclear cells.

Materials and methods

Data acquisition and WGCNA

To verify the diagnostic value of the candidate biomarkers, the GSE95587 dataset (18) was selected from Gene Expression Omnibus. This dataset consists of 117 fusiform gyrus tissue samples: 33 from neurologically normal, age-matched controls and 84 from autopsy-confirmed patients with AD.

To elucidate the potential association between clinical information and key genes, disease-related modules and key driver genes, from the perspective of gene co-expression networks, were analysed. First, to ensure analytical reliability, rigorous quality control was performed on the raw expression matrix using the ‘goodSamplesGenes’ function from the WGCNA package (19). This step filtered out samples and genes with missing values, zero variance or identified as outliers (specifically, sample GSM2516858 was detected and removed as an outlier, while all genes passed the aforementioned criteria), thereby ensuring the robustness of subsequent network construction. Subsequently, a scale-free gene co-expression network was constructed using the WGCNA method. This approach effectively captured coordinated expression patterns among genes, clustering functionally related genes into modules, to serve as a powerful computational biology tool for identifying core gene sets in disease mechanisms.

During WGCNA, the scale-free topology fit index was evaluated under different soft-thresholding powers. The optimal soft-thresholding power was determined to be β=7, which achieved a scale-free topology fit index (R2)>0.9, satisfying the prerequisite for constructing a scale-free network. To capture network topology beyond direct connections, the adjacency matrix was converted into a Topological Overlap Matrix. Gene modules were then preliminarily identified through hierarchical clustering combined with a dynamic tree-cutting algorithm. To streamline the network and enhance biological interpretability, the module eigengene dissimilarity was set to a cutoff of 0.25 (corresponding to a correlation of >0.75), and markedly similar modules were merged. Finally, Pearson's correlation coefficients were calculated between each module eigengene and the AD status (case vs. control). Modules demonstrating a significant association with the disease (|R|>0.25 and P<0.01) were selected. The genes within these significantly correlated modules were designated as candidate key genes for subsequent in-depth study. Furthermore, single-sample gene set enrichment analysis was conducted to quantify the immune-related pathway enrichment levels across individual samples, utilizing the GSVA R package. This analysis aimed to explore the potential correlation between AD-associated co-expression modules and the local immune microenvironment, thereby providing a biological rationale for subsequent immune infiltration validation. Genes within the significantly correlated modules were subsequently used for protein-protein interaction (PPI) network construction and hub gene identification. The STRING database (http://string-db.org) was used to establish the PPI network structure. The topological properties of the PPI network were identified using the cytoHubba plugin for Cytoscape (20). A medium confidence threshold (>0.9) was applied to guarantee the reliability of the results. The top 200 genes were selected to intersect through four algorithms (Degree, Maximum Neighbourhood Component, Maximal Clique Centrality and Stress Centrality), and 108 key genes were obtained, which were used as the focus for subsequent research.

Screening of candidate hub genes via machine learning

In the present study, four machine learning algorithms were employed to identify key hub genes associated with AD: Least Absolute Shrinkage and Selection Operator (LASSO), Support Vector Machine-Recursive Feature Elimination (SVM-RFE), Random Forest (RF) and Extreme Gradient Boosting (XGBoost).

LASSO regression refines the least squares method, excelling at selecting minimal features from available data, thus reducing overfitting (21). SVM stands as a flexible method applicable to both regression and classification tasks. Leveraging the kernel trick, it maps input vectors to a higher-dimensional space, constructing an optimal hyperplane for data separation (22). XGBoost, a leading boosting algorithm, minimises the loss function by iteratively adjusting sample and feature weights. This enables it to efficiently process large datasets with numerous features (23). RF, a type of ensemble method, constructs numerous decision trees using different subsets of the data. The final prediction is derived from majority voting, and it can quantify feature importance while being robust to feature scaling and easy to tune (24). The intersecting genes across the four machine learning models were identified as AD-related hub genes.

Validation of candidate genes and development of diagnostic models

The pROC package (25) was used to build diagnostic models for single genes and gene combinations. The discriminatory power of hub genes as assessed by receiver operating characteristic (ROC) curves. By calculating the area under the curve (AUC), the diagnostic value was assessed; AUCs >0.7 were deemed indicative of a superior diagnostic marker.

Immune infiltration assessment

CIBERSORT is a computational approach that enables accurate quantification of cellular composition using extensive tissue gene expression data generated by RNA sequencing. In the present study, the CIBERSORT tool (https://cibersort.stanford.edu/) was used to estimate the relative abundance of 22 distinct immune cell types in the dataset. The findings were visualised as bar charts to illustrate the proportion of each immune cell type across different samples. Furthermore, a t-test was used to compare how immune cell distribution differed between patients with AD and healthy control subjects. P<0.05 was considered to indicate a statistically significant difference.

Gene set variation analysis (GSVA)

GSVA is a nonparametric, unsupervised method that evaluates shifts in gene set activity across diverse samples. In the present study, the enrichment levels of these pathways in normal vs. AD samples were examined. To quantify pathway scores within the samples, enrichment analyses were conducted using the R package GSVA (26). Subsequently, a two-tailed Student's t-test was used to assess differences in GSVA scores between the two experimental groups. Additionally, Pearson's correlation coefficient was calculated to evaluate the linear association between gene expression levels and pathway scores.

Animal experiments

The present study complied with the National Institutes of Health guidelines for the ethical use of animals in research. All animal studies were implemented in accordance with protocols approved by the Animal Care and Use Committee of Harbin Medical University (approval no. 2024047), and all experiments adhered to the ARRIVE guidelines for in vivo research. Experiments were performed using 90-day-old male C57BL/6J mice (25–30 g), purchased from Liaoning Changsheng Biotechnology Co., Ltd., and housed in the Animal Facility of The First Affiliated Hospital of Harbin Medical University (Harbin, China). Consistent with previous studies (27), male mice were used in the present study to minimize variability associated with the estrous cycle and to reduce confounding factors during behavioural testing. The animals were maintained under a 12-h light/dark cycle, with unrestricted access to food and water; all procedures were performed during the light period. The total experimental duration was ~3 to 4 weeks. A total of 24 male mice were randomly assigned to two groups (n=12 per group): Sham and model. Throughout the study, mice were monitored daily for signs of distress or abnormal behaviour. Humane endpoints were defined as marked weight loss (>20%), persistent lethargy, inability to ambulate or feed or other signs of severe morbidity, at which point animals would be euthanized; however, no animals died or were euthanized prematurely during the study. All mice were euthanized at the experimental endpoint for tissue collection. For western blot analysis, mice were deeply anesthetized with isoflurane and euthanized by cervical dislocation. For histological and immunofluorescence analyses, mice were deeply anesthetized with isoflurane and euthanized by transcardial perfusion with saline followed by 4% paraformaldehyde (PFA). Death was confirmed by cessation of heartbeat and systemic rigidity (following perfusion) or by absence of respiration and pupillary dilation (following cervical dislocation), in accordance with standard protocols. Every effort was made to minimise animal distress and limit the number of animals used.

Animal grouping and establishment of the AD model

Mice were randomly assigned to one of two groups: The sham group received an intracerebroventricular (i.c.v.) injection of PBS, and the model group received an i.c.v. injection of Aβ1-42 peptide solution.

Aβ1-42 (MedChemExpress) was solubilised in sterile 0.1 M PBS (pH 7.4) at a concentration of 1 mg/ml, aliquoted, and stored at −20°C. Prior to use, the peptide was aggregated by incubation at 37°C for 4 days, as previously described (28). Aggregated Aβ1-42 (400 pmol per mouse in 3 µl PBS) or an equal volume of PBS for sham controls was delivered via i.c.v. injection. For the surgery, mice were anesthetised with 60 mg/kg intraperitoneal sodium pentobarbital and positioned in a stereotaxic frame. A unilateral injection into the right lateral ventricle was performed using a 28-gauge Hamilton microsyringe (needle length: 3.0 mm) (29). Injection coordinates relative to bregma were anterior-posterior=0.1 mm, medial-lateral=1 mm and dorsal-ventral=3 mm; the 3 µl volume was delivered at 1 µl/min (30). The needle was left in place for an additional 5 min before slowly withdrawing to prevent backflow. Sham-operated mice received an i.c.v. injection of an equal volume of PBS.

Morris water maze (MWM) test

MWM test, consisting of a 5-day spatial acquisition phase and a 1-day probe trial, was performed to evaluate Aβ-induced spatial learning and memory deficits in mice (n=6 per group) (31). During spatial acquisition, mice were trained 3 times daily for 5 days (20 min intervals between trials). In each trial, they were placed in a 180-cm-diameter circular tank and tasked with finding a 10-cm-diameter platform submerged 2 cm below the water surface in the target quadrant. Once found, mice were allowed to stay on the platform for 10 sec; each trial was limited to 2 min. The escape latency (sec) and path length (cm) were recorded. During the probe trial, the platform was removed. Mice were positioned on the side opposite to the target quadrant and allowed to explore for 1 min. The time spent (%) and number of crossovers in the target quadrant were measured with the same tracking system.

Haematoxylin and eosin (H&E) and Nissl staining

Following the behavioural tests, mice were deeply anaesthetized with isoflurane (5% for induction, maintained at 1.5% via a nose cone) and transcardially perfused with ~20 ml ice-cold 0.9% saline, followed by 50 ml 4% PFA in 0.1 M phosphate buffer (pH 7.4). Brains were post-fixed in 4% PFA at 4°C for 24 h, cryoprotected in 30% sucrose at 4°C until they sank and then sectioned coronally at 5 µm using a freezing microtome.

For H&E staining, paraffin-embedded tissue sections were deparaffinised in xylene and rehydrated through a graded ethanol series (100, 95 and 70%) at room temperature. Sections were stained with Harris haematoxylin for 5 min at room temperature, rinsed in tap water, differentiated in 0.5% hydrochloric acid in 70% ethanol for 10 sec, briefly washed in tap water and counterstained with eosin Y for 2 min at room temperature. Sections were then dehydrated through graded ethanol (70, 95 and 100%) at room temperature, cleared in xylene for 5 min, and mounted with coverslips using a resinous mounting medium. For Nissl staining, sections were stained with 0.1% cresyl violet acetate solution for 10 min at room temperature, rinsed briefly in distilled water, differentiated in 95% ethanol containing acetic acid until background staining was clear, dehydrated through 95% and absolute ethanol at room temperature, cleared in xylene for 5 min and mounted on coverslips. All stained sections were examined under a light microscope to assess general tissue morphology and neuronal cytoarchitecture.

Western blot analysis

Mice from each group (n=3 per group) were deeply anaesthetized with isoflurane and euthanized by cervical dislocation. The brains were rapidly removed, and the prefrontal cortex and hippocampus were dissected on a chilled plate. Tissues were immediately snap-frozen in liquid nitrogen and stored at −80°C until analysis. Tissues were homogenized in ice-cold RIPA lysis buffer (Beyotime Biotechnology) supplemented with protease and phosphatase inhibitors. After centrifugation at 12,000 × g for 15 min at 4°C, the supernatants were collected as protein extracts. Protein concentration was determined using a BCA protein assay kit (Beyotime Biotechnology) according to the manufacturer's instructions. Equal amounts of protein (30 µg per lane) were fractionated by 10% SDS-PAGE and transferred to PVDF membranes. After blocking with 5% non-fat dry milk (Biosharp Life Sciences) in Tris-buffered saline with 0.1% (v/v) Tween-20 buffer for 1 h at room temperature, membranes were incubated overnight at 4°C with primary antibodies against ASF/SF2 (1:1,000; cat. no. 12929-2-AP; ProteinTech Group, Inc.) and β-actin (1:10,000; cat. no. 66009-1-Ig; ProteinTech Group, Inc.). Following incubation with HRP-conjugated secondary antibodies, (goat anti-rabbit; 1:5,000; cat. no. 31460; Invitrogen; Thermo Fisher Scientific, Inc.) for 1 h at room temperature, protein bands were visualised using an enhanced chemiluminescence reagent kit (Suzhou Xinsaimei Biotechnology Co., Ltd.). Band intensities were quantified using ImageJ (version 1.53t; National Institutes of Health), with ASF/SF2 expression normalised to β-actin.

Immunofluorescence analysis

For immunofluorescence analysis, tissue collection was performed via transcardial perfusion as previously described, using mice from each group (n=3 per group). Mice brains were dehydrated in 15 and 30% sucrose solutions for 72 h, embedded in optimal cutting temperature compound and stored frozen at −80°C. The brains were sectioned into 5 µm slices for immunofluorescence staining; these sections were incubated overnight at 4°C with the anti-ASF/SF2 antibody (1:300; cat. no. 12929-2-AP; ProteinTech Group, Inc.). Following mounting with Antifade Mounting Medium containing DAPI (cat. no. P0131; Beyotime Biotechnology), the sections were treated with a FITC-conjugated AffiniPure goat anti-rabbit IgG antibody (H+L; 1:100; cat. no. 33107ES60; Shanghai Yeasen Biotechnology Co., Ltd.) at room temperature for 60 min. Images were acquired using a fluorescence microscope (DMi8; Leica Microsystems GmbH).

Patients

Between May 2024 and March 2025, 73 participants were recruited from The First Affiliated Hospital of Harbin Medical University: 36 healthy controls and 37 patients with AD. Inclusion criteria for patients with AD included: i) met the 2011 NIA-AA core diagnostic criteria (32) for probable or possible AD; ii) aged ≥50 years; iii) Mini-Mental State Examination (MMSE) score ≤26; iv) Clinical Dementia Rating (CDR) score ≥0.5; and v) ability to cooperate with neuropsychological assessments. Exclusion criteria for all participants were as follows: i) History of traumatic brain injury, epilepsy, stroke or other neurological disorders; ii) severe psychiatric disorders; iii) inability to complete cognitive scale assessments; iv) brain tumours or other structural brain lesions; v) abnormal thyroid function or other severe systemic diseases that may affect cognition; and vi) use of medications that may affect cognitive function within the past 3 months.

All protocols conformed to the ethical norms established by institutional and national research boards, in line with the 1964 Declaration of Helsinki. The present study was approved by the Ethics Committee of the First Affiliated Hospital of Harbin Medical University (approval no. 2024374). Written informed consent was acquired from the patients or their legal representatives, and all participants underwent clinical evaluations, neuropsychological evaluations and blood sampling.

Collection of general information and evaluation of cognitive scales

Data on age, sex, educational background, body mass index (BMI), hypertension, diabetes and apolipoprotein E ε4 carrier status were extracted from medical records. The age of patients with AD ranged from 51 to 87 years, with a median age of 68 years. The control group ranged from 49 to 79 years, with a median age of 66 years. All participants were evaluated using the MMSE, Montreal Cognitive Assessment (MOCA) and CDR. The CDR assessments were conducted by professional raters at The First Affiliated Hospital of Harbin Medical University.

Blood sampling and peripheral blood mononuclear cells (PBMCs)

Blood samples were collected in 3-ml K2-EDTA tubes (Becton, Dickinson and Company). PBMCs were separated using a Ficoll-Hypaque gradient (cat. no. LTS1077; Tianjin Haoyang Biological Products Technology Co., Ltd.), and total RNA was extracted using a commercial kit (cat. no. M5106; Suzhou Xinsaimei Biotechnology Co., Ltd.). RNA concentrations were measured using a NanoDrop ND-2000, and 1 µg RNA was reverse transcribed into cDNA using a reverse transcriptase kit (Tiangen Biotech Co., Ltd.) following the manufacturer's instructions: incubation at 42°C for 15 min, followed by termination at 85°C for 5 sec.

Quantitative (q)PCR

qPCR was performed using a Gene 9600 system (Applied Biosystems; Thermo Fisher Scientific, Inc.) with 2X SYBR Green qPCR Master Mix (cat. no. FP205; Tiangen Biotech Co., Ltd.). The thermocycling conditions were as follows: initial denaturation at 95°C for 10 min, followed by 40 cycles of denaturation at 95°C for 15 sec and annealing/extension at 60°C for 1 min. For each sample, Serine/Arginine Rich Splicing Factor 1 (SRSF1) mRNA expression was measured; GAPDH served as the internal control, and relative expression levels were determined using the 2−ΔΔCq method (33). The primer sequences used were: SRSF1 forward, 5′-AGGGAACAACGATTGCCGCATCTAC-3′ and reverse, 5′-ATGTCGCGGATAGCGCCGTATTTGT'; and GAPDH forward, 5′-CTGGGCTACACTGAGCACC-3′ and reverse, 5′-AAGTGGTCGTTGAGGGCAATG-3′.

Plasma collection and biomarker analysis

Plasma p-tau181, p-tau217, GFAP, NfL, Aβ42 and Aβ40 were measured using a chemiluminescent immunoassay on a Shine i2910 (Shenzhen Yingkai Biotechnology Co., Ltd.) with commercially available kits (Nanjing Novozymes Biotechnology Co., Ltd.): p-tau181 (cat. no. 251201A), p-tau217 (cat. no. 260203A), GFAP (cat. no. 251223A), NfL (cat. no. 260128A), Aβ42 (cat. no. 251223A) and Aβ40 (cat. no. 251217A). Blood samples were collected as previously described. Within 30 min of blood sample collection, samples were centrifuged at 1,500 × g for 10 min at room temperature. After centrifugation, 1 ml plasma was transferred into 1.5 ml Protein Lobind tubes (Eppendorf SE), which are designed to minimise protein binding. The samples were then stored at −80°C until use. All samples were thawed, processed and subjected to standardised processing followed by analysis using a single batch to avoid inter-batch variability.

Statistical analysis

Each experiment was repeated at least three times. For normally distributed data, data are presented as the mean ± standard deviation. For non-normally distributed data, are presented as the median and interquartile range and frequencies are presented as n (%). Group comparisons for frequencies were compared using a χ2 test. The expression of SRSF1 across different Braak stage classifications was analysed using a rank-sum test. Spearman's correlation analysis was used to assess the relationship between SRSF1 mRNA expression levels in PBMCs and scores from MMSE, MoCA and CDR. Additionally, to further validate the association between SRSF1 and CDR, Kendall's tau rank correlation analysis was performed to assess the monotonic relationship. Statistical analyses were performed using GraphPad Prism version 8.0 (Dotmatics). Differences between two groups were compared using a Student's t-test for normally distributed data, or a Mann-Whitney U test for non-normally distributed data. P<0.05 was considered to indicate a statistically significant difference.

Results

Identification of AD-associated gene modules

To elucidate the associations between clinical information and key genes, WGCNA was used to analyse the GSE95587 dataset. After rigorous quality control, a scale-free co-expression network was successfully constructed using an optimal soft-thresholding power of β=7, achieving a scale-free topology fit index R2>0.9. Hierarchical clustering and dynamic tree cutting identified 19 initial gene modules, which were subsequently merged using a module eigengene dissimilarity cut-height of 0.25 (corresponding to a correlation of >0.75) to streamline the network.

Following the construction and merging of similar modules, eight distinct gene co-expression modules were identified and assigned the following colours: Dark turquoise, light yellow, red, dark grey, salmon, blue, turquoise and cyan. Among these, the dark grey module exhibited the strongest positive correlation with the normal control group (r=0.42; P=2×10−6), while the turquoise module showed the strongest negative correlations with AD (r=−0.41; P=4×10−6) (Fig. 2). These findings highlight the potential central roles of these modules in AD pathogenesis.

Figure 2. Weighted Gene Co-expression Network Analysis was used to identify AD-associated core modules. (A) Soft-thresholding analysis for scale-free network topology. (B) Module-trait correlation heatmap in AD. (C) Gene clustering and module assignment. (D) Visualization of gene expression and network metrics. (E) Venn diagram of overlapping genes across network-based metrics. AD, Alzheimer's disease; MNC, Maximal Clique Centrality; MCC, Maximal Clique Centrality.

Subsequently, to identify key driver genes within the AD-associated modules, key genes were defined using stringent criteria: Absolute gene significance (GS)>0.3 and absolute module membership (MM)>0.8. Through this process, 1,714 genes showed significant associations with AD in terms of both GS and MM. These genes were then subjected to PPI network analysis and key gene identification as described in the methods (Fig. S1). Intersection of the top 200 genes from each of the four cytoHubba algorithms resulted in 108 common key genes for further investigation.

Identification of AD-associated candidate genes using multiple machine learning algorithms

All models were implemented using the dataset after quality control, consisting of 116 samples (83 AD and 33 controls) and an initial feature set of 108 candidate genes derived from prior WGCNA and PPI network analyses. LASSO, SVM-RFE, RF and XGBoost algorithms were applied to the refined dataset to pinpoint key AD-associated hub genes.

The LASSO regression model used 10-fold cross-validation to determine the optimal penalty coefficient λ, and the final gene subset was selected using the ‘1 standard error’ criterion (lambda.1se) to promote a more parsimonious model and prevent overfitting. For the SVM-RFE algorithm, feature selection was conducted within a 5-fold cross-validation framework. Features were iteratively eliminated based on their weights, and the optimal subset, consisting of 104 genes, was identified as the one yielding the lowest classification error. The RF model was constructed with 1,000 decision trees, and feature importance was evaluated using the ‘IncNodePurity’ (mean decrease in node impurity) metric. A threshold of 0.4 was applied to select key genes, a cut-off value determined from the distribution of feature importance scores, and aligned with common practices in the field (34). Finally, an XGBoost model was trained and optimised via 10-fold cross-validation. Features were ranked according to their ‘Gain’ value, and those with a gain >0.03 were deemed significant, following established methodologies for threshold setting (23).

The LASSO regression analysis identified seven potential hub genes: ACTG1, NFKB1, SRSF1, PPP2R1B, KIF5B, NDUFB5 and IKBKB (Fig. 3A). The SVM-RFE algorithm determined that a subset of 104 genes yielded the minimum root mean square error, representing an optimal feature set for classification (Fig. 3B). The RF algorithm, based on importance scores, provided 10 top-ranked candidate genes implicated in AD-related biological processes: PRKN, SMARCC1, KIF5B, NDUFB5, PXN, HDAC1, PRPF38A, SRSF1 and NFKB1 (Fig. 3C). The XGBoost algorithm highlighted seven additional potential genes based on feature gain: PRKN, NDUFB5, SRSF1, SMARCC1, CDC42, TAF4 and U2AF2 (Fig. 3D). To enhance the robustness of the findings, the gene lists produced by all four algorithms were intersected. This approach yielded two high-confidence candidate hub genes: SRSF1 and NADH:Ubiquinone Oxidoreductase Subunit B5 (NDUFB5) (Fig. 3E). These genes were considered to demonstrate marked potential as biomarkers or therapeutic targets for AD.

Figure 3. Screening for diagnostic biomarkers in AD through machine learning algorithms. (A) LASSO regression identified seven core feature genes associated with AD via 10-fold cross-validation. (B) SVM-RFE model for feature selection in AD. The x-axis represents the number of selected features, and the y-axis shows prediction accuracy and error. (C) The XGBoost algorithm determined seven feature genes and ranked them by importance. (D) RF model for AD genes. RF analysis ranked overlapping candidate genes by relative importance, highlighting the top 10 significant genes. (E) A Venn diagram showing intersections and unique genes across different machine-learning models. AD, Alzheimer's disease. LASSO, Least Absolute Shrinkage and Selection Operator; SVM-RFE, Support Vector Machine-Recursive Feature Elimination; RF, Random Forest; XGBoost, Extreme Gradient Boosting.

Candidate biomarkers for AD identified via ROC and differential expression analyses

Univariate evaluation was performed on the two identified hub genes, SRSF1 and NDUFB5, within the screening dataset using ROC and differential expression analyses. The results revealed that SRSF1 was significantly upregulated in AD samples (Fig. 4A), whereas NDUFB5 was notably downregulated (Fig. 4C). These genes demonstrated remarkable predictive capabilities in the screening dataset: SRSF1 achieved an AUC value of 0.746, while NDUFB5 attained an AUC of 0.777 (Fig. 4B and D).

Figure 4. Expression patterns of SRSF1 and NDUFB5 and their predictive performance for AD. Violin plots showing (A) SRSF1 and (B) NDUFB5 expression levels in control and AD groups from GSE95587. Violin width indicates data density. ROC curves evaluating the diagnostic performance of (C) SRSF1 and (D) NDUFB5 for AD. x-axis, specificity; y-axis, sensitivity. ***P<0.001. SRSF1, Serine/Arginine Rich Splicing Factor; NDUFB5, NADH: ubiquinone oxidoreductase subunit B5; AD, Alzheimer's disease; ROC, receiver operating characteristic.

Immune cell dysregulation and its association with characteristic genes in AD

To comprehensively analyse immune cell dysregulation, a hallmark of AD, the CIBERSORT algorithm was used. Based on linear support vector regression, CIBERSORT is a well-established computational method for deconvoluting bulk tissue gene expression data and estimating the relative proportions of 22 human immune cell subtypes. As shown in Fig. 5A, a marked divergence in immune profiles was observed between AD and normal samples. Notably, 20 distinct immune cell types were detected in both groups, establishing a basis for comparative analyses. Upon in-depth analysis, it was observed that the abundance of several immune cell types was significantly higher in AD samples compared with normal controls. This list included activated B cells, activated CD4+ T cells, activated CD8+ T cells, CD56 bright natural killer cells, central memory CD4+ T cells, central memory CD8+ T cells, effector memory CD8+ T cells, eosinophils, immature B cells, mast cells, myeloid-derived suppressor cells, memory B cells, natural killer cells, natural killer T cells, neutrophils, plasmacytoid dendritic cells, T-helper type 1 cells, and T-helper type 17 cells (all P<0.01). The findings revealed notable variations in the levels of immune cell infiltration between patients with AD and healthy individuals. A detailed heat map analysis was used to explore the interrelationships among immune cell types. Notably, in AD samples, natural killer cells and effector memory CD8+ T cells exhibited the strongest positive correlation (r=0.89), suggesting a potential cooperative role in AD immune dysregulation (Fig. 5B). Additionally, correlation analysis (Fig. 5C) revealed a close association between the two signature genes, SRSF1 and NDUFB5, and immune cell abundances. Specifically, SRSF1 showed a positive correlation with CD56 bright natural killer cells, mast cells, memory B cells, and natural killer T cells, suggesting a potential regulatory role in AD-related immune responses.

Figure 5. Immune characteristics between the control and AD groups. (A) Histogram of immune cell infiltration levels in control and AD groups based on ssGSEA results. Blue represents control samples and red represents AD samples. (B) Heatmap of immune infiltration patterns in AD. Colores (light to dark red) indicate immune cell infiltration levels or related molecular markers across samples/regions. (C) Correlations between SRSF1, NDUFB5 and differentially infiltrated immune cells. Solid lines, positive correlations; dashed lines, negative correlations. Colours indicate the strength of correlations between cells. Left label, P-value; R, correlation coefficient. (D) Violin plots of SRSF1 and NDUFB5 expression across Braak stages (III–VI) in AD. The shape of the violin plot represents the distribution of gene expression, with the central box indicating the interquartile range and the median. (E) gene set variation analysis results exploring the relationship between SRSF1 and two signalling pathways. *P<0.05, **P<0.01, ***P<0.001. ns, not significant; AD, Alzheimer's disease; ssGSEA, single-sample gene set enrichment analysis; GSVA, gene set variation analysis.

Altered SRSF1 and NDUFB5 expression across Braak stages and their association with AD-related signalling pathways

As shown in Fig. 5D, notable differences in the expression levels of SRSF1 and NDUFB5 were observed across various Braak stages. Compared with Braak Stage III, SRSF1 expression was significantly upregulated in Braak Stages IV, V and VI, suggesting a potential association with AD pathological progression. Conversely, NDUFB5 showed significant downregulation at Braak Stage VI compared with Braak Stages III, IV and V, highlighting its distinct expression pattern during later disease stages. Moreover, GSVA analysis demonstrated a marked positive correlation between increased SRSF1 expression and the activation of the p38MAPK signalling pathway (r=0.59; P<0.001). Conversely, SRSF1 overexpression was associated with inhibition of the AKT/mTOR signalling pathway (r=−0.63; P<0.001; Fig. 5E).

Preliminary validation of bioinformatics findings via an Aβ1-42-induced AD mouse model

To validate the bioinformatics analysis findings, an Aβ1-42-induced AD mouse model was successfully established via i.c.v. injection. A total of 2 weeks after brain stereotactic injection, MWM tests were conducted to evaluate mouse cognitive function. Notably, Aβ1-42-injected mice exhibited significantly longer swimming distance and escape latency compared with the control group, which is a key indicator of spatial memory impairment. Consistently, the number of platform crossings and time spent in the target quadrant were also significantly reduced (P<0.01; Fig. 6A-D), further confirming cognitive decline in the AD model group. Following the identification of behavioural abnormalities, morphological examinations were performed using the mouse hippocampus. H&E and Nissl staining revealed marked neuronal loss in the hippocampal and cortical regions of AD model mice (Fig. 6E-G). These histological changes directly associated with the memory impairment observed in MWM tests, indicating successful establishment of the Aβ1-42-induced AD mouse model.

Figure 6. Establishment of Aβ1-42-induced AD mice model. Aβ1-42 induced spatial memory impairment in the MWM task 2 weeks after intracerebroventricular injection. (A) Plots showing the escape latencies of mice in two groups when finding the hidden platform over 5 consecutive training days. (B) The number of platform location crossings during the probe trial, and (C) the percentage of time spent in the target quadrant of MWM. (D) Representative swimming path of mice on the test day. Results of (E) H&E, (F) Nissl staining and (G) quantitative analysis of Nissl staining of mice in each group. Data are presented as the mean ± SEM. **P<0.001 vs. sham-treated control group. n=6. AD, Alzheimer's disease; MWM, Morris water maze; H&E, haematoxylin and eosin; DG, dentate gyrus; CA1, Cornu Ammonis 1.

Moreover, western blot and immunofluorescence assays were used to measure the expression levels of the target gene SRSF1. Western blot analyses revealed a significant upregulation of SRSF1 expression in the hippocampus and cortex of AD model mice when compared with control mice (P<0.01; Fig. 7A-C). Immunofluorescence analysis additionally revealed significant SRSF1 upregulation in the mouse cerebral cortex, suggesting a potential role for SRSF1 in AD pathogenesis (P<0.01; Fig. 7D and E).

Figure 7. Expression of SRSF1 in the AD mouse model and the sham group. Mice in each group received PBS (sham) or Aβ1-42 (model) for 14 days. Following the MWM test, mice were euthanised, and hippocampal and cortical tissues were collected. (A) Western blot bands and quantitative analysis of SRSF1 and β-actin expression in (B) the hippocampus and (C) the cortex; n=3/group. (D) Representative immunofluorescence images and (E) quantitative analysis of SRSF1. Scale bar, 50 µm. n=3/group; **P<0.01. AD, Alzheimer's disease; SRSF1, Serine/Arginine Rich Splicing Factor; MWM, Morris water maze.

Baseline characteristics and biomarker analysis of AD and control cohorts

Table I presents a comprehensive summary and comparative analysis of the baseline characteristics of participants in the AD and control groups. Demographic and clinical parameters, including age, sex distribution, educational attainment, BMI and the prevalence of diabetes and hypertension, were statistically comparable between the two groups (P>0.05), indicating a well-matched cohort for subsequent analyses. Notably, cognitive assessment using the MMSE and MOCA scores revealed significant disparities. Patients with AD exhibited substantially lower MMSE and MOCA scores compared with healthy controls (P<0.05), underscoring impaired cognitive function. Moreover, molecular analyses showed that SRSF1 mRNA expression was notably increased in the AD group compared with the control cohort (P<0.05). This upregulation of SRSF1 may contribute to the pathophysiological mechanisms underlying AD, warranting further investigation into its role in disease progression.

Table I. Demographic and clinical parameters of patients with Alzheimer's disease and healthy controls.

Table I.

Demographic and clinical parameters of patients with Alzheimer's disease and healthy controls.

Characteristic	Alzheimer's disease, n=37	Controls, n=36	P-value
Age, n (%)			0.281
<65 years	10 (27.03)	14 (38.89)
≥65 years	27 (72.97)	22 (61.11)
Sex, n (%)			0.395
Male	9 (24.33)	12 (33.33)
Female	28 (75.67)	24 (66.67)
Education, n (%)			0.871
<9 years	13 (35.14)	12 (33.33)
≥9 years	24 (64.86)	24 (66.67)
Body mass index, kg/m2b	20.12±2.52	21.57±3.51	0.812
Diabetes, n (%)	6 (16.22)	6 (16.67)	0.959
Hypertension, n (%)	9 (24.32)	11 (30.55)	0.551
Apolipoprotein E4 allele carriers, n (%)	4 (10.81)	-	-
Mini-mental state examinationc	20 (15.5, 23.5)	27.5 (26,28.75)	<0.001a
Montreal cognitive assessmentc	15 (8.5, 19)	25 (23.25, 26)	<0.001a
Clinical dementia rating stage, n (%)			-
0.5	6 (16.21)
1	15 (40.54)
2	14 (37.84)
3	2 (5.41)
SRSF1 mRNA levels in the peripheral blood mononuclear cellsc	2.59 (1.76, 3.72)	1.07 (0.833, 1.37)	<0.001a

a P<0.001;

b Mean ± SD,

c Median (P25, P75). SRSF1, Serine/Arginine Rich Splicing Factor 1.

Correlation between SRSF1 and cognitive scores and plasma biomarkers

Compared with healthy controls, SRSF1 expression in PBMCs was significantly higher in patients with AD (P<0.01; Fig. 8A), highlighting the potential role of SRSF1 in AD pathogenesis. Notably, in the PBMCs of patients with AD, SRSF1 mRNA expression was strongly negatively correlated with cognitive function as measured by MMSE (r=−0.677; P<0.001) and MOCA scores (r=−0.678; P<0.001) (Fig. 8B and C). Moreover, SRSF1 mRNA expression showed a positive correlation with CDR score (r=0.739; P<0.001; Fig. 8D). This positive association was further validated by Kendall's rank correlation analysis, which revealed a significant positive correlation (τ=0.580; P<0.01). These findings suggest that increased SRSF1 expression may be associated with cognitive decline. Furthermore, as shown in Fig. 8E-I, SRSF1 mRNA levels in the PBMCs in patients with AD showed significant positive correlations with plasma disease-related biomarkers, including p-tau217 (r=0.64; P<0.001), p-tau181 (r=0.51; P=0.001) and GFAP (r=0.344; P=0.04). By contrast, no significant correlations were observed between SRSF1 expression levels and plasma Aβ42/Aβ40 ratios or NfL levels. These findings further elucidate the intricate relationship between SRSF1 expression and AD-associated molecular and clinical phenotypes.

Figure 8. SRSF1 mRNA expression in PBMCs of control and AD groups and correlations with cognitive scores and plasma biomarkers. (A) SRSF1 expression was upregulated in PBMCs from patients with AD compared with controls; **P<0.01. Cognitive state parameters in AD, such as (B) MMSE and (C) MOCA scores, showed inverse correlations with SRSF1 expression, whereas (D) CDR scores were positively correlated with it. No significant correlations were found between SRSF1 expression and (E) plasma Aβ42/Aβ40 ratios or (F) NfL levels. SRSF1 mRNA levels positively correlated with plasma (G) GFAP levels, (H) p-tau181 and (I) p-tau217. AD, Alzheimer's disease; SRSF1, Serine/Arginine Rich Splicing Factor; PBMC, peripheral blood mononuclear cells; MMSE, mini-mental state examination; MOCA, Montreal cognitive assessment; NfL, neurofilament light chain; GFAP, glial fibrillary acidic protein; CDR, clinical dementia rating.

Discussion

AD, which is characterised by an insidious, progressive and ultimately fatal course, poses a notable challenge to global public health systems. The intricate pathophysiology of AD remains incompletely understood, with several leading hypotheses having been proposed to explain its pathogenesis (35). Prominent among these are the immune-inflammatory response, the accumulation of Aβ peptides, and the aberrant hyperphosphorylation of Tau proteins (36,37). Given the limited efficacy of current therapies and the growing global prevalence of AD, further elucidation of its molecular mechanisms is urgently needed. These insights are critical for discovering and validating novel biomarkers that can enable early disease detection, guide personalised treatment strategies, and accurately predict patient outcomes (38).

In the present study, using WGCNA, 19 AD-associated modules were identified. Based on the criteria of GS>0.3 and MM>0.8, using four topological ranking algorithms, 108 hub genes were identified. Next, by applying four distinct models, LASSO, SVM-RFE, RF and XGBoost, two candidate genes that demonstrated strong correlations with AD pathological biomarkers in the dataset were identified. The fact that SRSF1 and NDUFB5 were consistently selected by all four methods markedly strengthens the evidence for their association with AD, suggesting their potential to predict disease progression, although this requires validation in larger independent cohorts. In diagnostic test evaluation, an AUC of 0.7–0.8 is generally considered acceptable for diagnostic accuracy (39). The AUC values of SRSF1 (0.746) and NDUFB5 (0.777) in the present study indicated acceptable discriminatory power between AD and control samples. Overall, these results highlighted the potential of SRSF1 and NDUFB5 as promising biomarkers for AD. The observed discriminatory power suggested that a diagnostic model incorporating both genes may have acceptable utility for aiding AD diagnosis, with potential value for early detection and clinical management, and warrants further validation in larger cohorts. It is hypothesized that immune system dysfunction underlies the inflammatory responses linked to AD, although the underlying mechanisms remain incompletely elucidated (40,41). In the present study, bioinformatics analysis revealed that SRSF1 expression was significantly upregulated compared with healthy controls and implicated in p38MAPK and AKT1/mTOR signalling pathways. In vivo experiments confirmed elevated SRSF1 expression in the hippocampus and cerebral cortex of Aβ1-42-induced AD model mice. Given that the immune-inflammatory response is a core pathological hypothesis of AD, and that SRSF1 has been shown to regulate immune-related genes, the association between SRSF1 and immune cell infiltration in patients with AD was further explored. These observations lay the foundation for further investigating the role of SRSF1 in immune cell infiltration, a core pathological feature of AD.

Consistent with prior studies (42), the present study found that the AD group exhibited significant alterations in immune cell infiltration ratios compared with the control group. These alterations primarily included activated B cells, activated CD4+ and CD8+ T cells, T-helper type 1 cells, T-helper type 17 cells and other cell types (43,44). A previous study explored the roles of peripheral immune dysfunctions and peripheral-central immune crosstalk in the pathogenesis and progression of AD (45). Peripheral B lymphocytes, a key element of the adaptive immune system, can cross the blood-brain barrier in patients with AD and enhance immune activation via interactions with resident brain cells (46). Type 1 and 17 T cells, two subsets of CD4+ T cells, have been shown to accelerate AD progression by inducing glial pro-inflammatory responses (47). Indeed, under different physiological and pathological conditions, distinct patterns of altered gene expression can either slow or accelerate disease progression by influencing immune cells or activating multiple synergistic regulatory mechanisms simultaneously (48). In the present study, SRSF1 was positively associated with immune cell abundances, specifically correlating with CD56bright natural killer cells, memory B cells, mast cells and natural killer T cells. Thus, it was hypothesized that SRSF1 expression may be associated with peripheral immune cell infiltration, although the underlying regulatory mechanisms require further functional validation. To verify this hypothesis, the findings revealed not only elevated SRSF1 expression in the PBMCs of patients with AD but also a significant correlation with specific immune cell infiltration. This implicated a previously underappreciated connection between a central splicing regulator and peripheral immune dysregulation in AD, highlighting a systemic aspect of the disease.

SRSF1 is a notable RNA splicing factor that regulates pre-mRNA splicing (49,50). The Microtubule-Associated Protein Tau (MAPT) gene encodes the tau protein, and abnormal alternative splicing of this gene is a hallmark of AD pathogenesis, leading to the accumulation of pathological tau isoforms in the brain. It has been demonstrated that SRSF1 acts as a marked splicing factor that binds to the exonic splicing enhancer (ESE) sequences within exon 10 of MAPT pre-mRNA. Mechanistically, SRSF1 promotes the inclusion of exon 10 in the mature MAPT transcript, which generates the 4-repeat (4R) tau isoform. Notably, the imbalance between 3-repeat and 4R tau isoforms is closely associated with tau hyperphosphorylation and neurofibrillary tangle formation in the brains of patients with AD (51). The CD33 gene is a well-established risk factor for late-onset AD, and its alternative splicing generates two major isoforms: The full-length isoform (CD33-M1) and the truncated isoform (CD33-M2), which lacks the immunoreceptor tyrosine-based inhibitory motif (ITIM) domain. SRSF1 has been identified as a key regulator of CD33 splicing by recognizing the ESE elements in the region flanking exon 2 (52). Specifically, SRSF1 enhances exon 2 inclusion, leading to increased CD33-M1 expression in microglia. The CD33-M1 isoform inhibits microglial phagocytosis of Aβ plaques by activating the ITIM-dependent signalling pathway, thereby promoting Aβ accumulation and neuroinflammation in AD (53). By influencing CD33 splicing, SRSF1 and PTBP1 may alter microglial functional properties and affect Aβ clearance, neuroinflammation, and synaptic pruning (52). Additionally, bioinformatics analyses demonstrated that SRSF1 orchestrated distinct gene regulatory programs in regulatory T cells and effector T cells in vivo. Genes regulated by SRSF1 include Interleukin-17A, Interleukin-17F, colony-stimulating factor and C-X-C motif chemokine ligand 10 (54). These genes are markedly enriched in pathways associated with inflammatory responses and cytokine-cytokine receptor interactions (55). The p38MAPK and AKT1/mTOR signalling pathways regulate immune-inflammatory responses and have been validated as promising therapeutic targets for AD (56). Identified as key players in SRSF1-related bioinformatics analysis, they play notable roles in regulating immune cell function and metabolism (57). The p38MAPK pathway modulates the activation, differentiation, and functional regulation of multiple immune cell types, whereas the AKT1/mTOR pathway drives metabolic reprogramming to meet the energy demands of immune cell activation (58). These two pathways are closely interconnected rather than functioning independently, and their crosstalk fine-tunes immune cell functions and maintains immune microenvironment balance, further linking SRSF1 to AD pathogenesis through immune regulation (59). The present findings offer novel perspectives on the molecular mechanisms through which SRSF1 may be involved in AD pathogenesis, suggesting a potential association with the regulation of multiple signalling pathways that warrants further functional validation. Collectively, the present results suggest that SRSF1 expression is associated with the inflammatory microenvironment and the dynamics of central and peripheral immune cells in AD, possibly through links to the splicing of these key genes, although causal relationships remain unconfirmed.

In patients with AD, neuronal degeneration and death occur as the disease progresses (60). Activated astrocytes attempt to repair damaged neural tissue, leading to increased GFAP expression (61). Plasma GFAP concentrations show a notable correlation with cognitive deterioration, making it a useful biomarker for tracking AD progression (62). Phosphorylated tau proteins, such as p-tau217 and p-tau181, are effective at identifying AD pathology and distinguishing AD from other neurodegenerative disorders (63). It has been demonstrated that Lumipulse G plasma p-tau217 and the p-tau217/Aβ42 ratio precisely recognised abnormal Aβ and tau PET statuses in both clinical and community groups (64). The diagnostic capability of plasma p-tau217 and the p-tau217/Aβ42 ratio was on par with cerebrospinal fluid (CSF) assessments (64). Given the core value of these biomarkers in AD diagnosis and progression monitoring, the correlation between SRSF1 mRNA levels in PBMCs and these indicators was further analysed to verify their clinical relevance. In the present study, a novel clinical correlation was identified: SRSF1 mRNA levels in PBMCs showed a significant positive association with key AD plasma biomarkers, including p-tau217, p-tau181 and GFAP. Notably, SRSF1 expression also showed a negative correlation with parameters reflecting cognitive status in patients with AD, such as MMSE and MOCA scores, while positively correlating with CDR scores. This dual correlation links SRSF1 not only to core AD pathological biomarkers but also to the severity of clinical cognitive decline, reinforcing its potential as a comprehensive biomarker reflecting both pathological changes and cognitive impairment. SRSF1 expression was significantly upregulated at Braak Stage IV relative to Stage III, which is commonly associated with mild cognitive impairment, a recognized prodromal stage of AD. Braak Stage IV corresponds to mild dementia, indicating that SRSF1 may undergo dynamic changes during this early dementia phase, warranting further exploration of its role in disease mechanisms. This pattern suggested that SRSF1 may be involved in the early pathological progression of AD within the studied clinical spectrum; however, further validation in cohorts including pre-symptomatic or prodromal individuals is required to confirm its potential as a biomarker for earlier disease stages. Mechanistically, this suggests that SRSF1-mediated splicing defects may be functionally tied to the misprocessing of proteins involved in tau phosphorylation and glial activation (65). Notably, neurons and glial cells maintain intimate crosstalk in the brain (66), whereby changes in neuronal SRSF1 levels can be sensed by glial cells to trigger downstream glial responses (55,67), ultimately leading to GFAP upregulation. Concurrently, SRSF1 dysregulation in neurons may drive abnormal tau phosphorylation, contributing to elevated p-tau217 and p-tau181 levels. Beyond this, SRSF1-mediated regulation of inflammation-related genes (68) can activate pro-inflammatory pathways, which in turn exacerbate tau hyperphosphorylation and glial activation, forming a pathological feedback loop.

The results of the present study support the potential of SRSF1 as a diagnostic biomarker for AD. Its elevated expression is easily assessable in PBMCs, and its good diagnostic accuracy and strong correlation with cognitive function scores suggest that SRSF1 could be used to develop a relatively non-invasive blood-based test. Such a test could serve as a practical complement or alternative to more expensive and invasive methods, such as PET scans and CSF analysis.

It is worth noting that the present study has certain limitations. The current analysis is cross-sectional, focusing on associations between SRSF1 and baseline AD phenotypes. However, longitudinal data are lacking to verify whether SRSF1 levels can predict disease progression or response to interventions, and long-term follow-up studies are required to evaluate its utility as a prognostic biomarker. The value of SRSF1 as a diagnostic biomarker has been preliminarily validated in a single dataset and a small clinical cohort. The generalizability of these findings is limited by the single-centre design and the cohort's lack of diversity in ethnicity, age and comorbidities. Accordingly, future work will include conducting a formal power analysis to determine the necessary sample size and pursuing multicentre collaborations to independently validate these results. Additionally, the generalizability of the present preclinical findings may be constrained by the exclusive use of male mice in the animal experiments. Numerous studies have shown that female mice often exhibit more severe cognitive impairments and greater amyloid plaque deposition in AD models compared with males, which is consistent with the higher prevalence of AD in women among human populations (20). In the present study, male mice were used; however, it is acknowledged that sex is a notable biological variable in AD research, and the exclusive use of male mice may introduce a potential bias, particularly in terms of the generalizability of the present findings to female subjects. Female mice may exhibit differences in disease susceptibility, progression rates or treatment responses due to hormonal influences, especially oestrogen-related neuroprotection. In future studies, it is planned to include both male and female mice to more comprehensively evaluate sex-specific differences in disease mechanisms. Although it is hypothesized that SRSF1 contributes to AD by modulating immune cell infiltration and the p38MAPK and AKT1/mTOR signalling pathways, a causal relationship between SRSF1 and AD pathogenesis remains unverified by functional experiments. The present study identified a significant association between SRSF1 expression and key AD biomarkers, including p-tau (p-tau217, p-tau181) and GFAP; however, the functional relationship and potential regulatory role of SRSF1 in tau hyperphosphorylation and astrocyte activation have yet to be experimentally confirmed. In particular, whether SRSF1 directly influences AD pathological features and the direction and molecular mechanisms underlying its involvement in disease progression require further investigation. To address these gaps, future work should employ gene editing or knockdown strategies in AD models to clarify the direct impact of SRSF1 on AD pathology and elucidate the underlying molecular mechanisms.

Collectively, the results of the present study suggest that SRSF1 is associated with AD pathogenesis and shows potential as a candidate diagnostic biomarker. Its suitability as a therapeutic target warrants further investigation through functional studies. Its associations with immune dysregulation, signalling pathway modulation and clinical disease progression underscore its multifaceted role in AD, providing a foundation for future mechanistic and translational research to advance AD diagnosis and treatment.

Supplementary Material

Supporting Data

Acknowledgements

Not applicable.

Funding

The present study was funded by The Fundamental Research Funds for Provincial Universities in Heilongjiang Province (grant no. 2024-KYYWF-0198) and The Fund of Scientific Research Innovation of The First Affiliated Hospital of Harbin Medical University (grant no. 2025M04).

Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

Authors' contributions

LJC designed the study. YCL wrote the manuscript. YCL, YZ and YTY performed the experiments. DL, YLS, WSX and HLY analysed the data. All authors revised the manuscript. All authors have read and approved the final version of the manuscript. YCL and DL confirm the authenticity of all the raw data.

Ethics approval and consent to participate

The animal study was reviewed and approved by the Animal Care and Use Committee of Harbin Medical University and Authority (approval no. 2024047), and all experiments adhered to the ARRIVE guidelines for in vivo research. The studies involving human participants were approved by the Ethics Committee of The First Affiliated Hospital of Harbin Medical University (approval no. 2024374). All protocols conformed to the ethical norms established by institutional and national research boards, in line with the 1964 Declaration of Helsinki. Written informed consent was obtained from all patients or their legal guardians for the collection of samples and participation in the study.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

References