Discovering non-linear dynamics of miRNAs in Alzheimer’s disease-related cognitive impairment: a cross-species approach with explainable machine learning

Abstract

Background

MicroRNA (miRNA) biomarker studies in Alzheimer’s disease (AD) typically assume monotonic relationships between expression levels and disease status, overlooking the non-linear, context-dependent nature of miRNA regulatory networks. This simplification limits mechanistic insight and clinical translation. We aimed to characterise non-linear contribution patterns of miRNAs across the AD continuum using explainable machine learning and to define stage-specific “operating windows” where individual miRNAs drive classification.

Methods

Candidate miRNAs were prioritised from APPtg/TAUtg mouse hippocampus (accession number GSE110743) using minimum-redundancy-maximum-relevance selection. A three-miRNA panel (miR-155-5p, miR-339-5p, and miR-455-5p) was validated in human serum (GSE120584; AD, MCI, and healthy controls). Linear and non-linear classifiers were compared, and SHAP dependence analysis was used to quantify sample-level contributions across expression ranges.

Results

Non-linear models (SVM-RBF and k-NN) consistently outperformed linear classifiers, with discrimination strongest for MCI vs. healthy controls (AUC: 0.844). SHAP analysis revealed that miR-155-5p functions as a stable primary driver across disease stages, whereas miR-339-5p and miR-455-5p act as context-dependent modulators contributing only within restricted expression ranges. Each miRNA exhibited distinct, stage-specific non-linear operating windows with threshold effects and inflection points rather than uniform dose-response patterns.

Conclusions

This study reframes circulating miRNAs as dynamic, interaction-governed signals rather than static biomarkers. The operating window framework provides interpretable, threshold-aware guidance for clinical decision-making and supports stage-sensitive early screening strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13062-026-00819-y.

Background

Alzheimer’s disease (AD), the most common neurodegenerative disorder worldwide, is characterised by progressive cognitive decline that severely impairs patients’ autonomy and overall quality of life [1–4]. Despite decades of research, no effective pharmacotherapy or curative strategy exists; thus, early diagnosis followed by pharmacological intervention remains the primary clinical approach to slow disease progression [4, 5]. A variety of biological indicators have been proposed to support early diagnosis and elucidate AD’s molecular mechanisms [6]. Among these, microRNAs (miRNAs) have attracted attention because they are stably detectable in biofluids such as blood and cerebrospinal fluid and can simultaneously regulate multiple genes and signalling pathways [7–10].

miRNAs are short non-coding RNAs that form complex, non-linear many-to-many regulatory networks with numerous mRNAs and other miRNAs, playing key roles in AD pathophysiology [11]. Numerous studies have reported aberrant expression of specific miRNAs across multiple tissues and biofluids and proposed diagnostic panels based on expression profiles [12–14]. However, linear analyses focused on individual expression levels struggle to capture non-linear, many-to-many interactions, limiting network-level characterisation and clinical translation [15–18].

Recent studies have applied machine learning and explainable artificial intelligence (XAI) to model miRNA networks in AD and identify hub miRNAs and genes [19, 20]. Yet many studies have not directly validated the cross-species diagnostic utility of the same miRNAs within a single study, and linear statistics remain limited in capturing multivariate non-linear interactions [21, 22]. To address these gaps, we prioritised candidates related to ageing-associated cognitive impairment in the hippocampi of APPtg and TAUtg mice, validated the same candidates in human serum, and applied XAI to quantify interactions and predictive contributions.

In this work, we aimed to construct and cross-species validate a concise circulating miRNA panel that marks ageing-related cognitive decline within the AD continuum and to interpret interactions and predictive contributions using Shapley additive explanations (SHAP). By employing both linear and non-linear models, we characterise the non-linearity of miRNA signals and delineate the intersection between ageing-related cognitive decline and AD pathophysiology. This approach links mechanistic coherence with clinical practicality, presenting a robust, low-dimensional, blood-based biomarker set suitable for early-stage screening and providing a scientific foundation for future precision-medicine applications.

Methods

Mouse samples

Mouse models expressing APPtg (APPswe/PS1 L166P) and TAUtg (THY-Tau22) transgenes were used to study amyloid-β and tau pathologies. All mice were backcrossed to a C57BL/6J background (for > 9 generations) to ensure genetic uniformity. Male wild-type and transgenic mice (n = 12/group) were sacrificed at 4 and 10 months, creating eight experimental groups. Hippocampal tissues were collected, total RNA was isolated, and miRNA expression was analysed using high-throughput sequencing. Data pre-processing included quality verification, trimming, filtering, alignment, and log transformation. After pre-processing, no missing values were detected across the samples or miRNAs; therefore, no imputation was performed. The miRNA dataset is available at the Gene Expression Omnibus (GEO) (GSE110743, http://www.ncbi.nlm.nih.gov/projects/geo/) [23].

Two APPtg mice (10 M) were excluded due to significantly lower transgene expression (APPswe: 46% lower; Psen1L166P: 26% lower). The remaining mice were grouped as follows: (1) Age-control groups: comparison between 4-month-old (APPwt-4 M, TAUwt-4 M, APPtg-4 M, and TAUtg-4 M) and 10-month-old (APPwt-10 M, TAUwt-10 M, APPtg-10 M, and TAUtg-10 M) mice, regardless of genotype; (2) Genetic control groups: comparison between transgenic (APPtg-4 M, APPtg-10 M, TAUtg-4 M, and TAUtg-10 M) and wild-type (APPwt-4 M, APPwt-10 M, TAUwt-4 M, and TAUwt-10 M) mice, across both age groups; and (3) Cognitive control groups: comparison between CI mice (APPtg-10 M, TAUtg-10 M) and PS mice (APPtg-4 M, TAUtg-4 M) to validate biomarkers of cognitive decline.

Clinical samples

Serum samples and clinical data from 1,341 participants, along with their miRNA expression profiles, were obtained from the National Center for Geriatrics and Gerontology Biobank. A total of 1,021 patients with AD and 32 individuals with MCI were diagnosed with probable or possible AD based on the criteria established by the National Institute on Aging–Alzheimer’s Association [24, 25]. A total of 288 HC participants with normal cognitive function had mini-mental state examination scores of ≥ 23 points. All diagnoses were made by specialists, such as neurologists, psychiatrists, geriatricians, or neurosurgeons, who were well versed in the diagnostic criteria for dementia. Diagnoses were based on medical history, physical and diagnostic examinations, neurological and neuropsychological assessments, and brain imaging using MRI or CT. All participants were aged ≥ 60 years. The miRNA expression data are publicly available from the GEO database (GSE120584, http://www.ncbi.nlm.nih.gov/projects/geo/) [26].

In the source cohort, serum miRNA expression was measured using the 3D-Gene Human miRNA Oligo Chip (Toray Industries, Inc.; miRBase release 21, 2,562 miRNAs), and signal processing and normalisation were performed using negative controls and internal control miRNAs (miR-149-3p, miR-2861, and miR-4463).

To ensure that all miRNA expression values were positive, the dataset was shifted by adding the absolute minimum value plus a small constant (ε = 0.001), followed by log₂ transformation to improve distributional normality [27, 28]. Subsequent z-score standardisation for model development was performed within each cross-validation iteration using the training fold only, and the resulting parameters were then applied to the corresponding validation or test fold. The data were standardised using z-score normalisation (mean = 0, standard deviation = 1) on a per-sample basis to enhance comparability across samples.

Feature selection

The minimum-redundancy-maximum-relevance (mRMR) method was used to select features exhibiting high relevance to the target class while minimising redundancy among selected features. Maximum relevance was achieved by prioritising features based on mutual information with the target class, whereas minimum redundancy ensured that the selected features provided distinct information [29–31]. In this study, the mRMR algorithm was independently applied to the age and genetic control groups using the mRMRe package in R version 4.4.2 (Vienna, Austria).

Machine-learning algorithms

To comprehensively assess the diagnostic potential of the selected miRNAs, we compared the machine-learning models based on their underlying decision boundaries. Specifically, three linear models, including LR, LDA, and SVM-linear, were used to evaluate whether the miRNA features exhibited linear separability. Additionally, two non-linear models, including SVM-RBF and K-NN, were applied to capture potential non-linear relationships among miRNAs. This approach enabled us to determine whether the identified miRNA biomarkers reflected simple linear associations or more complex non-linear patterns in the cognitive impairment classification. The LR, SVM, LDA, and K-NN, provided by the Scikit-learn library (version 1.4.2), were implemented in Python (version 3.12.3).

Model training and optimisation

To address class imbalance across all pairwise human classifications (AD vs. HC, MCI vs. HC, and AD vs. MCI), we applied the Synthetic Minority Over-Sampling Technique to the training split within each cross-validation fold only, thereby preventing information leakage and enabling learning from under-represented classes. Hyperparameter optimisation was performed using nested stratified cross-validation with an outer five-fold and an inner five-fold loop and evaluated using 500,000 bootstrap resamples in an out-of-fold (OOF) manner. Specifically, the optimal hyperparameter set identified in each fold was reapplied across all five folds, and the set producing the highest mean OOF performance across the five folds was selected as the final configuration. For the final performance reporting, we generated 10,000 bootstrap resamples.

Explainable artificial intelligence

To interpret the machine-learning models, we applied SHAP analysis. Stratified five-fold cross-validation for the mouse dataset and 10-fold cross-validation for the human dataset were applied. The SHAP values were calculated for the entire standardised dataset based on the trained models using background data selected from the training set via k-means clustering (k = 10). In addition, SHAP-dependence plots were generated by averaging SHAP values within 30 binned intervals of scaled expression levels, resulting in merged line plots that highlighted the non-linear contribution patterns of the selected miRNAs. To statistically assess these patterns, simple linear and segmented regression models were compared for each SHAP dependence profile using permutation testing, and 95% bootstrap uncertainty bands were added to the fitted curves.

Model evaluation

We evaluated five machine-learning models: SVM with two different kernels (SVM-linear and SVM-RBF), LR, LDA, and K-NN. For the mouse data, we assessed the generalisation performance of the models using 5-fold cross-validation. Subsequently, model performance was evaluated by calculating the average values across five folds for accuracy, sensitivity, specificity, positive predictive value, negative predictive value. All values are reported as means ± standard deviations.

For the human data, a 10-fold cross-validation was used to evaluate the generalisation performance of the models. Model performance was summarised using the mean test AUC across folds, and the corresponding 95% confidence intervals were reported together with these cross-validation summary results. ROC-AUC analyses were used to compare machine-learning performance across the three classification tasks, and DeLong’s test was applied to pooled out-of-fold predictions for statistical comparison of ROC-AUC values between models.

Statistical analysis

Data are presented as means ± standard deviations. Statistical analyses included the calculation of Pearson’s correlation coefficients among the identified miRNAs and Welch’s t-test to assess the statistical significance of the mouse data, and DeLong’s test for statistical comparison of ROC-AUC values in the human data. All the analyses were conducted using Python (v3.12.3). Welch’s t-tests were performed using SciPy (v1.11.4.) Stats package, and Pearson’s correlation coefficients were calculated using the Pandas package (v2.2.2).

Cross-species miRNA mapping and homology analysis

To evaluate the functional relevance of the miRNAs, the miRNAs identified from the mouse data were first mapped to their human homologs using miRBase [32]. Next, for each miRNA, experimentally validated target genes were programmatically retrieved with the R package multiMiR (v1.28.0). The main enrichment analysis was performed using the validated-target-only set. In the validated set, we retained only human entries and kept miRNA–gene pairs that were reported by at least two curated validation databases or experimentally demonstrated. We then standardised miRNA IDs, harmonised gene symbols to the HGNC nomenclature, and removed duplicate or empty records. Over-representation analysis (ORA) was performed after mapping gene symbols to Entrez IDs using Reactome pathways and Gene Ontology Biological Process (GO: BP). Multiple testing was controlled by the Benjamini–Hochberg procedure with FDR < 0.05, and gene-set size filters of minGS = 10 and maxGS = 500.

The detailed workflow of this study is shown in Fig. 1. Initially, to identify the miRNA biomarkers associated with AD-induced CI, we divided the mouse dataset into two groups based on age and genetic background. Using the mRMR method, the 100 miRNAs with the highest mutual information were selected from each group, resulting in 15 common candidate miRNAs. Subsequently, SHAP analysis was conducted on these 15 miRNAs using five different machine-learning models, and key biomarkers were identified using a filter-based approach. The selected key biomarkers were further validated by comparing their expression levels between the cognitively impaired and cognitively normal mouse groups to confirm their relevance to CI.

Fig. 1.

Workflow of the analysis pipeline. The diagram illustrates the overall analytical process, including miRNA expression data pre-processing, feature selection, classification using machine learning models, and subsequent interpretation using SHAP analysis. miRNA, micro RNA; AD, Alzheimer’s disease; MCI, mild cognitive impairment; HC, healthy control; ROC-AUC, area under the receiver operating characteristic curve; SHAP, Shapley additive explanations

Additionally, to determine the pathological stage at which these biomarkers were most significantly involved in AD progression, we analysed human serum miRNA expression data across three distinct groups: AD vs. MCI, MCI vs. HC, and AD vs. HC. Using the three biomarkers previously identified in the mouse model as features, we applied five machine-learning models to generate ROC-AUC curves, thereby confirming the translational validity of the mouse-identified biomarkers in humans. Finally, by employing SHAP dependence plot analysis, we elucidated the non-linear relationships of these miRNAs with AD, rather than merely describing changes in miRNA expression.

Results

Using the mRMR algorithms, the top 100 miRNAs with the highest mutual information scores were independently selected from the age and genetic control groups. Among these, 15 miRNAs were identified in both groups: mmu-miR-592-5p, mmu-miR-181d-5p, mmu-miR-146a-5p, mmu-miR-339-5p, mmu-miR-203-3p, mmu-miR-339-3p, mmu-miR-653-5p, mmu-miR-455-5p, mmu-miR-155-5p, mmu-miR-3084-3p, mmu-miR-212-3p, mmu-miR-212-5p, mmu-miR-17-3p, mmu-miR-140-3p, and mmu-miR-132-5p. These miRNAs were found to be associated with both age-related and genetically driven characteristics and were subsequently utilised as input features for downstream classification and interaction analysis.

Based on these 15 miRNAs, we trained several machine-learning models to classify group-specific conditions into the genetic and age-control groups (Tables 1 and 2). All machine-learning models achieved an average accuracy exceeding 0.85, suggesting that the miRNAs selected using the mRMR algorithm effectively captured the genetic and ageing-related characteristics associated with AD. In the genetic control group, the support vector machine (SVM)-radial basis function (RBF) model demonstrated the highest performance, with an average accuracy of 0.905 (Table 1), and it achieved the best performance in the age-control group, with an average accuracy of 0.894 (Table 2). Sensitivity, specificity, positive predictive value, and negative predictive value were well-balanced across the models.

Table 1.

Performance comparison of classification models in the genetic control group

Model	ACC	SEN	SPE	PPV	NPV
LR	0.873 ± 0.062	0.827 ± 0.150	0.916 ± 0.076	0.918 ± 0.076	0.862 ± 0.098
LDA	0.852 ± 0.084	0.804 ± 0.147	0.896 ± 0.063	0.880 ± 0.074	0.840 ± 0.105
SVM-Linear	0.883 ± 0.077	0.916 ± 0.076	0.856 ± 0.152	0.877 ± 0.125	0.917 ± 0.074
SVM-RBF	0.905 ± 0.061	0.891 ± 0.070	0.918 ± 0.117	0.925 ± 0.106	0.903 ± 0.055
K-NN	0.884 ± 0.051	0.827 ± 0.087	0.938 ± 0.081	0.938 ± 0.081	0.853 ± 0.057

Abbreviations: ACC, accuracy; SEN, sensitivity; SPE, specificity; PPV, positive predictive value; NPV, negative predictive value; LR, LASSO regression; LDA, linear discriminant analysis; SVM, support vector machine; SVM-RBF, support vector machine-radial basis function; K-NN, K-nearest neighbours

Table 2.

Performance comparison of classification models in the age-control group

Model	ACC	SEN	SPE	PPV	NPV
LR	0.883 ± 0.051	0.891 ± 0.004	0.876 ± 0.097	0.881 ± 0.087	0.892 ± 0.012
LDA	0.862 ± 0.041	0.847 ± 0.056	0.876 ± 0.097	0.878 ± 0.087	0.861 ± 0.029
SVM-Linear	0.883 ± 0.084	0.822 ± 0.089	0.940 ± 0.120	0.940 ± 0.120	0.853 ± 0.076
SVM-RBF	0.894 ± 0.047	0.891 ± 0.070	0.898 ± 0.063	0.893 ± 0.064	0.899 ± 0.058
K-NN	0.861 ± 0.088	0.847 ± 0.114	0.876 ± 0.080	0.863 ± 0.090	0.861 ± 0.097

Abbreviations: ACC, accuracy; SEN, sensitivity; SPE, specificity; PPV, positive predictive value; NPV, negative predictive value; LR, LASSO regression; LDA, linear discriminant analysis; SVM, support vector machine; SVM-RBF, support vector machine-radial basis function; K-NN, K-nearest neighbours

Mean absolute SHAP values were calculated to evaluate the contribution of each feature and identify potential biomarkers (Fig. 2a, b). The results showed that mmu-miR-592-5p and mmu-miR-17-3p had the highest average SHAP values in the genetic and age-control groups, respectively, indicating their dominant contributions in each context. Conversely, both miRNAs exhibited low SHAP values in the opposite group, suggesting that their relevance was limited to genetic or age factors rather than AD-related ageing.

Fig. 2.

Mean absolute SHAP values of miRNA features in the two comparison groups. (a) genetic-control group and (b) age-control group. The bar plots display the 15 miRNAs with the highest mean absolute SHAP values in each group, reflecting their contributions to the machine-learning model’s predictions. (c) Venn diagram illustrating the overlap between the top 10 SHAP-ranked miRNAs from the age and genetic control groups and the full set of annotated Homo sapiens miRNAs

Building on this, we selected the top 10 miRNAs with the highest mean absolute SHAP values from each control group and analysed their overlap (Fig. 2c). This comparison revealed seven miRNAs that were shared across both groups, implying their potential role in ageing-related AD pathophysiology in the mouse model. Among these, four miRNAs—mmu-miR-155-5p, mmu-miR-455-5p, mmu-miR-339-5p, and mmu-miR-181d-5p—were found to be conserved in humans through miRBase mapping and were selected as final candidate biomarkers.

We conducted a correlation analysis of the four selected miRNAs using Pearson’s correlation coefficients (Supplementary Data 1). The heatmap showed that mmu-miR-155-5p, mmu-miR-455-5p, and mmu-miR-339-5p exhibited strong positive correlations with one another (Pearson’s r > 0.66), with the highest correlation observed between mmu-miR-155-5p and mmu-miR-339-5p (r = 0.73).

To evaluate whether the expression levels of the miRNAs—mmu-miR-155-5p, mmu-miR-339-5p, mmu-miR-455-5p, and mmu-miR-181d-5p—differed significantly between the cognitively impaired (CI) and pre-symptomatic (PS) groups, Welch’s t-test with Bonferroni correction was performed (Fig. 3). The results showed that mmu-miR-155-5p (adjusted p = 3.85e-09), mmu-miR-339-5p (adjusted p = 2.53e-10), and mmu-miR-455-5p (adjusted p = 1.32e-06) were significantly upregulated in the CI group, suggesting that these miRNAs may reflect the pathophysiological changes associated with CI. In contrast, mmu-miR-181d-5p showed no significant difference in mean expression between the CI and PS groups (adjusted p = 1.00e + 00) but exhibited relatively high inter-sample variability, which led to its exclusion from subsequent comparative analyses. Taken together, mmu-miR-155-5p, mmu-miR-339-5p, and mmu-miR-455-5p consistently showed increased expression in the CI group and were statistically significant discriminative markers.

Fig. 3.

Boxplot of differentially expressed miRNAs associated with cognitive impairment and pre-symptomatic stages. The plots display the medians and interquartile ranges and highlight group-wise differences in miRNA expression patterns. Statistical comparisons were conducted using Welch’s t-test, followed by Bonferroni correction for multiple testing. Asterisks indicate statistical significance (p < 0.05, p < 0.01, p < 0.001)

We applied five machine-learning models using the three selected miRNA biomarkers to classify the CI and PS groups and evaluated their performance based on various metrics (Table 3). All models achieved an average accuracy of over 0.85 and consistently demonstrated high sensitivity and specificity. Notably, the SVM with an RBF kernel and the K-nearest neighbours (K-NN) classifier outperformed the others, achieving an average accuracy of 0.936, an average sensitivity of 0.960, and a specificity of 0.920.

Table 3.

Performance comparison of classification models in the cognitive impairment group

Model	ACC	SEN	SPE	PPV	NPV
LR	0.891 ± 0.004	0.860 ± 0.116	0.920 ± 0.098	0.927 ± 0.090	0.893 ± 0.088
LDA	0.871 ± 0.036	0.820 ± 0.093	0.920 ± 0.098	0.920 ± 0.098	0.853 ± 0.075
SVM-Linear	0.913 ± 0.044	0.910 ± 0.111	0.920 ± 0.098	0.927 ± 0.090	0.927 ± 0.090
SVM-RBF	0.936 ± 0.053	0.960 ± 0.080	0.920 ± 0.098	0.927 ± 0.090	0.960 ± 0.080
K-NN	0.936 ± 0.053	0.960 ± 0.080	0.920 ± 0.098	0.927 ± 0.090	0.960 ± 0.080

Abbreviations: ACC, accuracy; SEN, sensitivity; SPE, specificity; PPV, positive predictive value; NPV, negative predictive value; LR, LASSO regression; LDA, linear discriminant analysis; SVM, support vector machine; SVM-RBF, support vector machine-radial basis function; K-NN, K-nearest neighbours

Using SHAP analysis, we interpreted the contributions of each feature in the SVM model trained with mmu-miR-155-5p, mmu-miR-339-5p, and mmu-miR-455-5p as input variables. Feature importance was assessed and visualised using SHAP summary plots, bar graphs of the mean absolute SHAP values, force plots of representative samples, and a decision plot depicting the cumulative contributions across samples. According to the SHAP summary plot (Fig. 4a), mmu-miR-155-5p exhibited the highest absolute SHAP value among the three miRNAs. In addition, mmu-miR-455-5p and mmu-miR-339-5p showed clearly distinguishable SHAP distributions between the two groups, but their overall contributions were relatively small.

Fig. 4.

SHAP-based interpretation of miRNA contributions to classification performance across samples. (a) Summary plot showing overall feature importance and directionality. (b) Decision plot showing cumulative contribution patterns across samples ordered by similarity. (c–f) Force plots illustrating the contribution of individual miRNAs to representative sample predictions

In most samples, the three miRNAs acted in a similar manner. This trend was evident in the decision plot (Fig. 4b), where the cumulative SHAP values were arranged by sample similarity. For instance, mmu-miR-155-5p consistently increased the prediction probability for CI, whereas mmu-miR-455-5p and mmu-miR-339-5p showed more variable contributions depending on the expression levels in each sample. The SHAP force plots for individual samples (Fig. 4c–f) visually illustrate how each miRNA contributed to the predictions at the sample level. However, some samples exhibited atypical patterns that deviated from the overall trend (Fig. 4c, e). As shown in Figs. 3b and 4d, the primary miRNA contributing to the prediction shifted from mmu-miR-155-5p to either mmu-miR-339-5p or mmu-miR-455-5p.

ROC-AUC analyses with 95% confidence intervals were performed for the three classification tasks (Fig. 5). In AD vs. MCI, SVM-RBF yielded the highest AUC [0.651 (95% CI, 0.534–0.769)], followed by k-NN [0.617 (0.506–0.728)], SVM-Linear [0.555 (0.450–0.660)], LDA [0.471 (0.383–0.559)], and L1-logistic regression [0.429 (0.361–0.497)]. In MCI vs. HC, overall discrimination was the greatest; SVM-RBF again performed the best [0.844 (0.757–0.932)], followed by k-NN [0.800 (0.695–0.904)], LDA [0.500 (0.349–0.650)], SVM-Linear [0.474 (0.338–0.610)], and L1-logistic regression [0.466 (0.344–0.587)]. In AD vs. HC, SVM-RBF also showed the highest AUC [0.721 (0.677–0.766)], followed by k-NN [0.702 (0.665–0.739)], SVM-Linear [0.595 (0.541–0.649)], L1-logistic regression [0.594 (0.544–0.645)], and LDA [0.594 (0.544–0.644)]. Consistent with this pattern, OOF-based DeLong tests demonstrated statistically significant differences between SVM-RBF and the linear models in MCI vs. HC and AD vs. HC (all p < 0.001), whereas in AD vs. MCI, SVM-RBF significantly outperformed LDA (p = 0.0259) and L1-logistic regression (p = 0.0287), but not SVM-Linear (p = 0.0659) (Supplementary data 2). Additional calibration analyses showed that calibration patterns varied across tasks and models. SVM-RBF and K-NN showed the most favourable Brier-score profiles in MCI vs. HC and AD vs. HC, whereas this pattern was less clear in AD vs. MCI (Supplementary data 3).

Fig. 5.

ROC curves with AUC for the classification of AD, MCI, and HC based on serum microRNA expression. Receiver operating characteristic curves were generated for each pairwise comparison using three selected miRNAs: (a) AD vs. MCI, (b) MCI vs. HC, and (c) AD vs. HC. The area under the ROC curve was used to assess diagnostic performance. AD, Alzheimer’s disease; MCI, mild cognitive impairment; HC, healthy control; ROC, receiver operating characteristic; AUC, area under the ROC curve; miRNA, microRNA; LDA, linear discriminant analysis; SVM, support vector machine; RBF, radial basis function; k-NN, k-nearest neighbours

SHAP dependence plots were created to visualise the non-linear predictive contributions of the three miRNAs (Fig. 6). We report results for the SVM with an RBF kernel, which showed the best performance (Fig. 5). Each panel represents one miRNA within one pairwise task, and the displayed fitted curves include 95% bootstrap uncertainty bands. The left three columns correspond to miR-339-5p, miR-155-5p, and miR-455-5p, respectively; each row denotes one pairwise task in order: AD vs. mild cognitive impairment (MCI), MCI vs. healthy control (HC), and AD vs. HC. The rightmost column overlays smoothed curves for all three miRNAs on a common axis to facilitate direct comparison of operating windows. Overall, the direction and magnitude of contribution varied with expression level and classification context, yielding distinct non-linear profiles for each miRNA. These non-linear patterns were consistently better captured by segmented regression than by simple linear regression.

Fig. 6.

SHAP dependence plots of three miRNAs based on SVM with RBF kernel. Scatter plots display the non-linear contribution patterns of hsa-miR-155-5p, hsa-miR-339-5p, and hsa-miR-455-5p to the model’s output across different pairwise classifications: AD vs. MCI, MCI vs. HC, and AD vs. HC. The merged line plots on the right summarise the SHAP value trends by averaging across binned expression levels, highlighting miRNA-specific behaviours under each classification scenario. SHAP, Shapley additive explanations; miRNAs, microRNA; AD, Alzheimer’s disease; MCI, mild cognitive impairment; HC, healthy control

In AD vs. HC, miR-455-5p showed a steep rise in SHAP values in the high-expression range (z-score > 1), indicating a strong positive contribution to AD prediction. By contrast, miR-339-5p and miR-155-5p showed piecewise non-linear patterns with clear changes in slope, indicating that their predictive contributions depended on specific expression windows rather than on a simple monotonic increase or decrease. Thus, the most AD-predictive combination consisted of elevated miR-455-5p with decreased miR-339-5p and miR-155-5p within their respective non-linear operating ranges.

In MCI vs. HC, overall discriminability was the highest (Fig. 5). miR-455-5p again showed a pronounced non-linear relationship, with a sharp transition near the lower-to-intermediate expression range followed by a flatter slope at higher values. miR-155-5p also showed a strong non-linear pattern with a marked change in SHAP contribution, whereas miR-339-5p exhibited a weaker but still statistically supported segmented profile. These findings indicate that the strong separability of MCI vs. HC is driven not by simple linear shifts, but by distinct non-linear operating windows across the three miRNAs.

In AD vs. MCI, all three miRNAs showed non-monotonic curves with one or more inflection points. Notably, miR-455-5p showed a significant segmented relationship with a transition in the low-expression range, whereas miR-339-5p and miR-155-5p also displayed clear slope changes and non-monotonic contributions across expression levels. These patterns suggest that progression from MCI to AD is characterised not by a single linear shift but by combinatorial threshold effects across distinct pathological axes. Thus, the transition from MCI to AD is better represented by statistically supported non-linear inflection structures than by linear trends alone.

Taken together, the results from the SVM-RBF–based SHAP analysis indicate that all three miRNAs contribute in a context-dependent and non-linear manner across the three classification tasks. Consistent with the quantitative analyses, segmented regression provided a better fit than simple linear regression for all nine pair-feature combinations, with empirical permutation p-values of 0.0005 and ΔR² values ranging from 0.135 to 0.579. Together with the 95% uncertainty bands shown in Fig. 6, these findings indicate that the observed SHAP dependence patterns are not merely visual artefacts but statistically support non-linear relationships. The corresponding model-comparison statistics are provided in the Supplementary Data 4.

Additional exploration univariate cut-off analyses were performed for each individual miRNA based on ROC analysis and the Youden index (Supplementary Data 5). Although cut-offs could be formally defined, both the cut-off values and their discriminatory performance varied across pairwise comparisons, and the AUCs of individual miRNAs remained limited. These findings suggest that the observed group separation is not adequately captured by a single monotonic threshold for any one miRNA. Instead, together with the SHAP dependence analyses and model-comparison results, they further support the interpretation that the three-miRNA panel operates through context-dependent, non-linear expression windows.

Next, to identify genes regulated by the three miRNAs, we retrieved experimentally validated target genes using multiMiR. Using this validated-target set, we performed Reactome ORA and GO: BP enrichment analysis. The Reactome ORA of the validated target genes for the three miRNAs highlighted RHO GTPase-related signalling, growth factor/TGF-beta signalling, and transcriptional, apoptotic, and cell-cycle signalling pathways (Fig. 7). The top 15 pathways (FDR-adjusted) converged on the RHO GTPase cycle, diseases of signal transduction by growth factor receptors and second messengers, transcriptional regulation by TP53, signalling by TGF-beta family members and the TGF-beta receptor complex, transcriptional activity of the SMAD2/SMAD3:SMAD4 heterotrimer, programmed cell death/apoptosis, and mitotic G1/S and G2/M phase regulation, collectively indicating dysregulated signal transduction together with altered cell-fate and cell-cycle control. Top terms showed high GeneRatio with significant FDR (< 0.05). Overall, these findings suggest that validated targets of miR-155-5p, miR-339-5p, and miR-455-5p converge on RHO GTPase-related signalling, growth factor/TGF-beta signalling, and transcriptional regulation, apoptosis, and mitotic cell-cycle pathways relevant to AD pathophysiology. Permutation-based null analysis using experimentally validated targets only showed that the observed three-miRNA panel yielded 577 significant Reactome pathways, exceeding the null expectation from random miRNA assignment (empirical p = 0.04595) (Supplementary Data 5). The maximum enrichment signal was also greater than expected under permutation (empirical p = 0.04096), indicating that the validated-target Reactome enrichment reflected a non-random biological signal.

Fig. 7.

Reactome over-representation analysis dot plot for the validated target genes of the three miRNAs (hsa-miR-155-5p, hsa-miR-339-5p, and hsa-miR-455-5p). Gene symbols were mapped to Entrez IDs, and Reactome ORA was performed after Benjamini–Hochberg multiple-testing correction (FDR < 0.05) with gene-set size filters of minGS = 10 and maxGS = 500. The x-axis indicates the GeneRatio, the dot size represents gene count, and the dot colour denotes the adjusted p-value

Furthermore, to further contextualise the performance of our model, we compared the proposed panel with previously reported miRNA-based biomarker panels (Table 4). Although direct one-to-one comparison is inherently limited by differences in specimen type, biomarker composition, and study design, a contextual comparison in the clinically relevant MCI-versus-control setting remains informative. In this setting, Wen et al. reported a plasma single-miRNA model based on miR-145-5p with an AUC of 0.72, whereas Sandau et al. (2020) reported a 5-miRNA CSF panel with an AUC of 0.705. By comparison, our serum-based 3-miRNA panel achieved an AUC of 0.844, suggesting competitive performance, despite its relatively simple and parsimonious biomarker composition.

Table 4.

Performance comparison between the proposed and previous methods

Study	Specimen	Biomarker/panel	comparison	AUC
Wen et al., [33]	Plasma	miR-145-5p	MCI vs. HC	0.720
Sandau et al., [34]	CSF	5-miRNA panel	MCI vs. HC	0.705
proposed	Serum	3-miRNA	MCI vs. HC	0.844

Abbreviations: MCI, mild cognitive impairment; HC, healthy control

Discussion

This study presents a low-dimensional, interpretable three-miRNA panel (hsa-miR-155-5p, hsa-miR-339-5p, hsa-miR-455-5p) for capturing AD-related cognitive impairment using explainable machine learning and examines its translational applicability between mouse hippocampus and human serum. In the human cohort, discriminability was the highest for MCI vs. HC [AUC: 0.844 (95% CI, 0.757–0.932)], intermediate for AD vs. HC [AUC: 0.721 (95% CI, 0.677–0.766)], and the lowest for AD vs. MCI [AUC: 0.651 (95% CI, 0.534–0.769)]. Across five classifiers, the non-linear SVM-RBF consistently outperformed the linear models, indicating superior capacity to model the multivariate, non-linear structure of miRNA signals. DeLong tests further supported this pattern, showing statistically significant differences between SVM-RBF and the linear models in MCI vs. HC and AD vs. HC (all p < 0.001), whereas in AD vs. MCI, SVM-RBF significantly outperformed LDA and L1-logistic regression, but not SVM-Linear. Together, these findings strengthen the conclusion that non-linear modelling provides a more appropriate framework for this three-miRNA panel in clinically defined AD-related cognitive impairment.

SHAP dependence curves clarified that each miRNA exhibits distinct, stage-contextual non-linear operating windows. In AD vs. HC, miR-455-5p showed a sharp positive contribution at high expression (z ≈ 1 or higher), whereas miR-339-5p and miR-155-5p contributed negatively in low-to-mid ranges. In MCI vs. HC, miR-455-5p contributed positively only within a mid-to-high expression window; miR-339-5p retained an inverse trend with reduced magnitude, whereas miR-155-5p traced a shallow curve. In AD vs. MCI, all three markers displayed one or more inflection points, with miR-455-5p showing a threshold-like transition around z ≈ 1.2. These patterns suggest that progression is shaped not by a single linear shift but by cumulative threshold interactions among miRNAs representing distinct pathological axes.

The observed patterns align with prior work showing that reduced miR-339-5p increases AD prediction probability, that elevated miR-155-5p associates with neuroinflammation and cognitive decline, and that increased miR-455-5p accompanies metabolic/synaptic stress [35, 36]. Beyond this, sample-level SHAP curves quantified operating windows and curve crossings/divergences, revealing context-dependent modulation whereby miR-339-5p and miR-455-5p amplify or delimit the effect of miR-155-5p within restricted expression bands. Unlike earlier studies assuming monotonic or linear effects, our results indicate that marker influence is not uniform across the expression range—for example, a modest contribution of miR-155-5p in MCI vs. HC and a surge of miR-455-5p only beyond high-expression thresholds (z ≈ 0.8–1.2), with minimal contribution elsewhere. Collectively, these findings motivate a new interpretive frame that assigns roles to a driver (miR-155-5p) and context-dependent modulators (miR-339-5p/miR-455-5p), supports stage- and threshold-aware clinical cut-off design, and positions a hippocampus-anchored, low-dimensional serum panel for early screening.

Biological interpretation was supported by pathway enrichment of validated targets, indicating convergence on RHO GTPase-related signalling, growth factor and TGF-beta signalling, and transcriptional regulation, programmed cell death, and mitotic cell-cycle processes. Reactome highlighted the RHO GTPase cycle, diseases of signal transduction by growth factor receptors and second messengers, transcriptional regulation by TP53, signalling by TGF-beta family members, programmed cell death, apoptosis, and mitotic phase transitions, supporting a framework in which the three miRNAs may influence AD-related cognitive impairment through coordinated effects on cytoskeletal dynamics, stress-responsive transcriptional programmes, cell-fate regulation, and signalling pathways linked to neuronal vulnerability and tissue remodelling. These results are consistent with prior evidence linking miR-155-5p to neuroinflammation and cognitive decline, and implicating miR-339-5p and miR-455-5p in synaptic, metabolic, and cytoskeletal pathways [37–39], supporting a stage- and dose-sensitive regulatory model.

Methodologically, combining non-linear classification with sample-level explainability preserved interpretability for a compact panel. SHAP-defined operating windows translate directly to clinical decision-making, including threshold setting, cohort enrichment, and pharmacodynamic monitoring, while extensive bootstrap resampling quantified uncertainty and between-model differences to enhance reproducibility.

This study had some limitations. Because the human cohort was clinically defined rather than biomarker-confirmed, some etiological heterogeneity may remain within the AD and MCI groups, warranting caution in interpreting the present findings. Sample size, particularly in mice, was limited; external independent-cohort validation and experimental mechanism testing were not performed; and SHAP-inferred interactions are associative rather than causal. Serum-based performance may be dampened by dilution of CNS-specific signals, pre-analytic/batch heterogeneity, and systemic comorbidities or medications. Although the cross-species design enabled unbiased miRNA candidate discovery in mice and validation in human clinical data, direct comparison of downstream target and pathway enrichment across species was limited by insufficient data. Therefore, this framework should be interpreted as a miRNA discovery–validation pipeline rather than proof of conserved molecular pathways across species. Accordingly, the results should be regarded as hypothesis-generating rather than definitive.

Future work should pursue: (1) pre-analytic standardisation and multi-centre external validation, (2) clinical-utility assessments including calibration and decision-curve analysis, (3) multimodal expansion by integrating systemic blood markers (e.g. metabolic/vascular) to complement CNS-leaning miRNA signals, and (4) targeted in vitro/in vivo tests of inferred thresholds, co-operating operating windows, and key pathway nodes. If validated, this three-miRNA panel offers a concise, interpretable, blood-based candidate to support early screening and the implementation of precision-medicine strategies.

Conclusions

This study established a low-dimensional, interpretable three-miRNA panel (miR-155-5p, miR-339-5p, miR-455-5p) using explainable machine learning and demonstrated its translational applicability from mouse hippocampus to human serum. In human data, the panel showed the highest discriminability for MCI versus HC (AUC 0.844), indicating a stage-sensitive signal. The non-linear SVM-RBF consistently outperformed linear baselines, and SHAP dependence curves identified miR-155-5p as a stable driver, while miR-339-5p and miR-455-5p acted as context-dependent modulators that sharpen decision boundaries within specific expression windows. Pathway enrichment of validated targets indicated convergence along RHO GTPase-related signalling, growth factor and TGF-beta signalling, transcriptional regulation, programmed cell death, and mitotic cell-cycle processes, providing biological plausibility for the XAI-inferred thresholds and co-acting operating windows.

Clinically, this panel requires only a small blood volume and three features, yet provides per-feature directions of effect and thresholdable operating windows to inform cut-off setting, early-trial cohort enrichment, and pharmacodynamic monitoring. Methodologically, combining non-linear classification with sample-level explainability repositions miRNAs as dynamic, interaction-governed signals rather than static markers. Limitations include modest sample size (particularly in mice), absence of external cohort validation, and the associative (non-causal) nature of SHAP-inferred interactions. Future work should prioritise pre-analytic standardisation and multi-centre external validation, clinical-utility assessments (including calibration and decision-curve analyses), multimodal integration with systemic metabolic/vascular markers, and targeted in vitro/in vivo testing of inferred thresholds, co-acting windows, and key pathway nodes. If validated, this three-miRNA panel represents a concise, interpretable, blood-based candidate for early-stage screening and future precision-medicine-oriented stratification.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

AD

Alzheimer’s Disease

CI

Cognitive impairment

GEO

Gene expression omnibus

HC

Healthy control

K-NN

K-nearest neighbours

LDA

Linear discriminant analysis

LR

Lasso regression

MCI

Mild cognitive impairment

miRNA

microRNA

MMSE

Mini-mental state examination

mRMR

Minimum-redundancy maximum-relevance

NCGG

National Centre for Geriatrics and Gerontology

RBF

Radial basis function

ROC-AUC

Area under the receiver operating characteristic curve

SHAP

Shapley Additive explanations

SMOTE

Synthetic Minority Over-sampling Technique

SVM

Support vector machine

XAI

Explainable artificial intelligence

GO

BP: Gene ontology biological process

ORA

Over-representation analysis

RTK

Receptor tyrosine kinase

Author contributions

S. H. L.: Conceptualisation; Methodology; Software; Formal analysis; Investigation; Data curation; Visualisation; Writing – original draft. S. B. L.: Conceptualisation; Methodology; Supervision; Project administration; Writing – review & editing; Funding acquisition. S. K.: Resources; Supervision; Writing – review & editing; Funding acquisition. All authors have read and approved the final manuscript.

Funding

This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Ministry of Science and ICT (RS-2024-00439078).

Data availability

The datasets used in the preparation of this article are publicly available from the Gene Expression Omnibus (GEO) database under accession numbers GSE110743 and GSE120584 (http://www.ncbi.nlm.nih.gov/projects/geo/) following the respective database guidelines and access policies.

Declarations

Ethics approval and consent to participate

This study analysed publicly available mouse and human datasets. The original animal experimental procedures were reviewed and approved by the relevant animal ethics committee in the source study, and the original human study was approved by the corresponding institutional ethics committee with written informed consent obtained from all participants.

Consent to participate

Informed consent was obtained from all individual participants included in the study.

Consent for publication

The authors are responsible for the correctness of the statements provided in the manuscript.

Competing interests

The authors declare no competing interests.