An integrative clinical and bioinformatic analysis identifies MicroRNAs as biomarkers of ischemic stroke severity

Abstract

Identifying reliable circulating biomarkers is crucial for improving the diagnosis and risk stratification of patients with ischemic stroke. In this study, we evaluated several whole-blood circulating miRNAs (miR-106b-5p, miR-16-5p, miR-15b-5p, let-7e-5p, and miR-125a-3p/-5p) to determine their diagnostic and disease severity in acute ischemic stroke (AIS). Sixty AIS patients and thirty age- and sex-matched controls were included. Whole-blood miRNAs were quantified at admission and on day 7. Statistical analyses included ROC curves, multivariate logistic regression, and SHAP-based machine learning. Bioinformatic analyses assessed predicted miRNA targets, pathway enrichment, and interaction networks. MiR-125a-3p was significantly reduced in AIS at both time points, while miR-125a-5p was elevated at admission and decreased by day 7. Both miRNAs showed moderate diagnostic value (AUC 0.675 and 0.712, respectively). Higher admission levels of miR-16-5p were strongly associated with greater neurological deficit (NIHSS) and unfavorable outcome (mRS ≥ 3). Multivariate analyses confirmed high miR-16-5p and elevated CRP as independent predictors of poor outcome. Bioinformatic analyses revealed that miR-16-5p targets were enriched in pathways relevant to ischemic injury, including hypoxia response, platelet activation, coagulation, TGF-β and BDNF signaling. A target-interaction network highlighted IL6, FN1, TGFB1, ICAM1, and TLR4 as central nodes linking miR-16-5p to ischemia-inflammatory mechanisms in AIS. Circulating miRNAs display distinct expression patterns in the acute phase of AIS. miR-16-5p emerges as a promising biomarker associated with stroke severity and unfavorable outcome, while miR-125a-3p and miR-125a-5p show potential diagnostic utility. These findings strengthen mechanistic links between platelet-derived miRNAs and ischemic stroke biology. Larger, longitudinal studies integrating functional validation are warranted to confirm their clinical value.

Introduction

Ischemic stroke accounting for 85% of all types of stroke is one of the most common cardiovascular diseases and one of the leading causes of death worldwide¹. Despite the availability of therapeutic methods such as thrombolysis or mechanical thrombectomy only about 13% of ischemic strokes patients may be treated. The main reason is the short time to recognize a stroke and start the treatment. For it to be effective, this time should not exceed 4.5 h¹. Over the years, advancements in imaging techniques, such as computed tomography (CT) and magnetic resonance imaging (MRI), have significantly improved the diagnosis of stroke and the prognosis of patients². However, not readily available in many hospitals, and transporting the patient to other facilities consumes valuable time^1,3. Therefore, new methods of quick and effective diagnostics are still being sought. The blood-based biomarkers appear promising due to their ease and rapid collection⁴.

Non-coding RNA (ncRNA), including microRNA (miRNA), have shown potential utility as diagnostic and disease severity biomarkers of ischemic stroke in previous in vitro, animal, and human studies⁵. MiRNAs are regulators of mRNA transcription that influence the translation process and protein synthesis by targeting the 3′-untranslated region^6–8. MiRNAs may affect gene expression in the cell of their origin, but they can also be released into the extracellular space and internalized by other cells^9,10. This mechanism is thought to play an important role in brain tissue regeneration following ischemic/reperfusion injury¹¹. Circulating miRNAs can be reliably detected in the serum/plasma/whole blood of patients with ischemic stroke and are stable enough for laboratory analysis^12,13. Several studies have shown that the expression levels of specific circulating miRNAs are significantly altered in stroke patients and often correlate with the severity of the condition, highlighting their potential as biomarkers. Moreover, both animal models and in vitro experiments support the involvement of miRNAs in stroke pathogenesis. Their ability to cross the blood–brain barrier and modulate gene expression in damaged brain tissue also suggests their promise as future therapeutic agents¹⁴.

We previously presented that miRNAs; such as miR-106b-5p, miR-16-5p, and miR-15b-5p are associated with cardiovascular and cerebrovascular diseases^15,16. Moreover, previous two important publications showed that let-7e and miR-125a family have high potential as blood biomarkers in ischemic stroke patients^17,18. Hence, in this study, we aimed to investigate the association between circulating miRNAs and ischemic stroke severity, assessing their potential as early biomarkers for predicting neurological outcomes. Furthermore, through integrative bioinformatics and machine learning, we elucidate the regulatory networks and molecular pathways of these miRNAs to uncover mechanisms underlying stroke pathophysiology.

Methodology

Patients

The study protocol, designed in compliance with the Declaration of Helsinki, was approved by the The Ethics Committee of the Institute of Psychiatry and Neurology in Warsaw, Poland approved our study protocol and consent forms. We obtained written informed consent from all 60 participants, who were selected from an existing dataset. These patients had a diagnosis of acute ischemic stroke, established through WHO clinical criteria and confirmed by brain imaging (CT or MRI).¹⁹. Stroke etiology was classified according to the TOAST (Trial of Org 10,172 in Acute Stroke Treatment) criteria based on clinical assessment, neuroimaging (CT/MRI), vascular imaging, cardiac evaluation, and laboratory findings. All major ischemic stroke subtypes were included in the study. The final distribution of TOAST categories in our cohort was as follows: large-artery atherosclerosis (37%), ischemic stroke of undetermined etiology with ipsilateral carotid stenosis ≥ 50% (43%), cardioembolic stroke (10%), small-vessel occlusion (5%), and other determined etiologies (5%). This prevented the exclusion of stroke cases where marked carotid atherosclerosis was present but not considered the definitive cause by TOAST standards. Duplex Doppler imaging of extracranial arteries and 24-h Holter ECG monitoring were performed routinely²⁰. Patients with a history of stroke, peripheral arterial occlusive disease or cancer were excluded from the study. For controls, we enrolled 30 age- and gender-matched patients without stroke but with a high cardiovascular risk profile, evidenced by either multiple cardiovascular risk factors or a diagnosis of coronary artery disease (CAD).

Sample collection and handling

Whole venous blood was collected as early as possible after hospital admission (for all patients, within 12 h). The follow-up blood sample was collected on day 7 of hospitalization after the acute ischemic stroke. Approximately 5 mL of venous blood was drawn into Tempus™ Blood RNA tubes (ThermoFisher) and stored at − 80 °C until analysis.

MiRNAs selection, RNA preparation, detection, and quantification of miRNAs by Quantitative PCR

MiR-106b-5p, miR-16-5p, miR-15b-5p miRNAs were selected based on bioinformatic analysis that was previously published by our team^15,16. Let-7e and miR-125a-3p and -5p, on the other hand, were selected based on previous important publications^17,18. Whole blood RNA was extracted using the Tempus™ Spin RNA Isolation Kit (Thermofisher). Total RNA was obtained as outlined above and diluted 1:10. Five microliters of diluted RNA were reverse transcribed using the TaqMan miRNA Reverse Transcription kit, following the manufacturer’s instructions. Subsequently, 3 μL of the product was used for detecting miRNA expression by quantitative polymerase chain reaction using TaqMan miRNA Assay kits for the corresponding miRNAs on a CFX384 Touch Real-Time PCR Detection System (BioRad Inc., Hercules, California, USA).

All samples were normalized to the synthetic spike-in cel-miR-39, which was added in equal amounts during RNA isolation. In addition, one randomly selected sample was used as an internal reference across all qPCR runs. Reactions were performed in triplicate, and the mean value was used for all analyses to minimize variability due to methodological factors. MiRNA expression levels were normalized to cel-miR-39 for an internal reference sample. All reactions were performed in triplicate, and mean values were used in statistical analysis as described before^21,22. MiRNA expressions were expressed as 2-ΔΔCT²³, and the results were log-10 transformed for statistical analysis.

Statistical analysis and results illustrations

All results for categorical variables were reported as numbers and percentages. Continuous variables were presented as either mean ± standard deviation (SD) or median and interquartile range (IQR), based on the normality of their distribution assessed using the Shapiro–Wilk test. The chi-square test was employed for dichotomous data. Depending on the normality of the distribution, the Student’s t-test or Mann–Whitney test was used for unpaired samples, and the Wilcoxon test was used for paired samples. For comparisons involving more than two groups, the Kruskal–Wallis, with Dunn’s post hoc and Bonferroni correction test, was applied. Receiver operating characteristic (ROC) analysis was performed to assess the diagnostic value of miR-125a-3p and -5p, and predictive potential of moderate-to-severe neurological deficit outcome (based on mRS < 3 NIHSS < 5 at day-7 and day-30 follow-up) of miR-16-5p. To determine independent variables affecting increased stroke severity, we applied multivariate logistic regression analysis. For this purpose, we included the following variables: high miR-16-5p (> 10.6 expression based on Log-transformed data), age, gender, CRP, antiplatelet therapy, platelet count, smoking status, and CAD. Model stability was confirmed by calculating Variance Inflation Factors (VIFs); all VIF values were below accepted thresholds, indicating no multicollinearity. A sensitivity analysis using a reduced model yielded comparable results, supporting the robustness of the associations. The Hosmer–Lemeshow test was applied to assess the goodness of fit and calibration of the logistic regression models. All tests were two-sided with the significance level of p < 0.05. Calculations were performed using SPSS software version 22.0 (IBM Corporation, Chicago, USA). Graphs were refined using Adobe Illustrator version 24.0.2, and the graphical abstract was created with BioRender (biorender.com). ChatGPT and Grammarly were used to assist in improving grammar and correcting typographical errors. Based on the observed difference in the proportion of patients with increased stroke severity; 65.6% (19/29; 95% CI: 47.3–80.1%) in the high miR-16-5p group and 28.6% (9/31; 95% CI: 16.1–46.6%) in the low miR-16-5p group, we estimated that, with a total sample size of 60 patients, the study had 84% power to detect this effect at a two-sided alpha level of 0.05.

Bioinformatic analysis

Selection of gene lists

Genes associated with coagulation, platelet, inflammation, hemopoiesis, hypoxia, and platelet activation receptor activity were obtained from Gene Ontology (GO) database through the biomaRt R package²⁴. Similarly, gene lists for specific platelet-related receptors. Platelet activation pathway was obtained from the KEGG database. Genes associated with stroke-related phenotypes were retrieved from Disgenet database²⁵. Genes related to BDNF and TGFB signalling pathways were retrieved after enrichment analysis from EnrichR database.

miRNAs target predictions

We conducted target-miRNA predictions using multimiR R package for all genes regulated by the selected miRNA hsa-miR-16-5p, 4633, hsa-miR-106b-5p, hsa-miR-15b-5p. We looked for the top 20% of predictions in 14 miRNA databases²⁶. MiRNA-gene interactions were aggregated using our wizbionet R package²⁷. In effect, we obtained information on how many genes from each list are regulated by a given miRNA and stacked the list of those genes. In the next step, we used our R function from wizbionet based on a machine learning algorithm OneR, which divided each numerical column with the number of targets into four clusters²⁸. In this context, OneR performs non-arbitrary numeric classification based on Jenks natural breaks optimization, which divides continuous numeric values (here, the number of regulated genes per miRNA) into four clusters that minimize within-cluster variance and maximize between-cluster separation. Further, we counted the occurrences of the top 1 clusters for each miRNA in each gene list and sorted them from highest to lowest. Top miRNAs were selected based on the number of targets from each gene list assigned to cluster 1.

Enrichment analysis

Custom gene lists related to platelet activity and ischemic stroke were analyzed using Fisher’s Exact test with FDR correction (p < 0.05). We performed pathway enrichment analysis via the EnrichR API, using the human genome as a reference and a hypergeometric test with Benjamini–Hochberg correction. Statistical significance for all tests was defined as an adjusted p-value ≤ 0.05. For data aggregation and ranking, we used our wizbionet R package²⁸. The visualization of ontological results was carried out in R using the ggplot2 and ggrepel libraries.

Construction of interaction networks

Target-target interactions were retrieved from the complete Human interactome using stringApp v2.1.1²⁹ through Cytoscape software v3.9³⁰. Onto the obtained networks, we mapped information of their participation in specific gene lists of interest. Further, we selected those associated with ischemic stroke or present on any of the lists of interest. Obtained genes were ordered by the number of connections with other nodes, then we selected the top 10 interactors, and all genes associated with ischemic stroke. Visualisation of the networks was done in Cytoscape software.

SHAP analysis

Our study used laboratory and demographic predictors extracted from a clinical dataset. Relevant variables were identified separately for the laboratory and demographic domains and coerced into numeric values to form model inputs. Missing values were handled with a custom preprocessing function (NoNA.df) from the wizbionet R package, which standardized and removed nonstandard NA encodings to yield a complete input matrix. The outcome (stroke severity) was binarized (0 = good, 1 = bad) based on mRS cut-off < 3. Models were trained with gradient-boosted trees (XGBoost, binary logistic) on a stratified 70/30 train-test split with early_stopping_rounds = 10, and selected by maximizing validation AUC with an overfitting guard (training AUC < 0.99). For the laboratory feature subset (45 predictors), the optimal configuration was η = 0.01, max_depth = 3, γ = 3, subsample = 0.8, and colsample_bytree = 0.8 (other parameters default). This model achieved a validation AUC = 0.8313 and training AUC = 0.9736, converging after 37 boosting iterations (best_ntreelimit = 37). The resulting model was subsequently applied to the held-out test set for SHAP value computation and feature attribution. The best configuration for the demographic predictor (97 predictors) set was η = 0.01, max_depth = 3, γ = 3, subsample = 0.8, colsample_bytree = 0.8, yielding a validation AUC = 0.994 and test accuracy = 0.944. The optimal probability threshold for binary classification (≈ 0.497) was determined using the Youden index from ROC analysis (pROC). These configurations were fixed for interpretation. SHAP values were computed on the independent test set using TreeSHAP (via SHAPforxgboost), which decomposes each prediction on the margin (log-odds) scale into a bias term plus per-feature contributions that sum to the model output, f(xᵢ) = φ₀ + ∑ⱼ φᵢⱼ. Positive SHAP values increase the log-odds of the “bad” class, whereas negative values decrease it. Global importance was summarized as the mean absolute SHAP value across test samples, and distributions/directionality were visualized with a SHAP summary (beeswarm) plot. The workflow was automated with custom R functions to support reproducibility. R packages which were used: xgboost (training/prediction), SHAPforxgboost (SHAP; with ppcor and WGCNA in sensitivity analyses), ggplot2 (graphics), dplyr and wizbionet (data handling), and additionally Hmisc, cluster, caret, and pROC.

Results

Participants

Patient characteristics are presented in Table 1. Cardiovascular disease includes a history of CAD, MI, and T2D. The ischemic stroke patients and healthy controls were well-matched in baseline demographics, with no significant differences in sex, age, or body mass index (BMI) (p = 0.176, p = 0.724, and p = 0.125, respectively). Of the 60 patients enrolled, six (10%) died during hospitalization; all were ICU patients with a median time to death of 3 days. Patients were divided into two separate groups according to stroke outcome, assessed using the mRS (mRS < 3) (Table 2).

Table 1.

Participants characteristics.

	Control (n = 30)	AIS (n = 60)	P value
Sex (male)	14 (47%)	37 (62%)	0.176
Age	70.90 ± 8.6	71.72 ± 11.0	0.724
BMI	28.7 ± 4.1	27.4 ± 4.7	0.125
Hypertension	19 (63.3%)	39 (65%)	0.876
History of HF	3 (10%)	7 (11.7%)	0.813
History of CAD	4 (13%)	12 (20%)	0.436
History of AF	–	3 (5%)	–
History of MI	3 (10%)	8 (13.3%)	0.649
History of T2D	11 (36%)	19 (32%)	0.635
Current smoking	4 (13%)	16 (30%)	0.093
Total cholesterol	154.5 ± 31.2	173.5 ± 40	0.156
Triglyceride	138.3 ± 57.1	124.7 ± 50.7	0.445
NIHSS at admission
(1–4) (5–25)	–	28 (47%) 32 (53%)	–
NIHSS at day 30
(1–4) (5–25)	–	36 (60%) 24 (40%)	–
modified Rankin Scale (mRS)
(1–2) (3–6)		31 (52%) 29 (48%)	–
TOAST subtype
– Large-artery atherosclerosis		22 (37%)
– Undetermined with ipsilateral carotid stenosis ≥ 50%		26 (43%)
– Cardioembolic		6 (10%)
– Small-vessel		3 (5%)
– Other determined		3 (5%)

Data are presented as number and percentage or mean and standard deviation. Student’s t test or Chi square test was used appropriately.

Table 2.

Demographics of stroke patients based on mRS outcome.

	mRS < 3 (n = 31, 52%)	mRS ≥ 3 (n = 29, 48%)	Completed data	p
Sex (male)	18 (58%)	19 (66%)	100%	0.553
Age (years)	69.19 ± 11.08	74.94 ± 8.79	100%	0.035
BMI	27.89 ± 4.96	27.56 ± 4.72	100%	0.755
Hypertension	15 (58%)	21 (72%)	100%	0.244
History of HF	3 (10%)	4 (14%)	100%	0.620
History of CAD	5 (16%)	7 (24%)	100%	0.438
History of AF	1 (3%)	2 (7%)	100%	0.514
History of MI	3 (10%)	5 (17%)	100%	0.389
History of T2D	8 (26%)	11 (38%)	100%	0.313
Current smoking	9 (30%)	7 (24%)	93%	0.808
Total cholesterol (mg/dL)	187.26 ± 42.23	165.29 ± 43.38	98%	0.154
Triglyceride (mg/dL)	109.90 ± 30.68	116.70 ± 42.43	98%	0.926
HDL-cholesterol (mg/dL)	55. 30 ± 20.27	46.53 ± 11.72	98%	0.377
LDL -cholesterol (mg/dL)	105.76 ± 41.10	95.45 ± 37.38	98%	0.302
HCT (%)	43.45 ± 2.77	42.24 ± 4.56	100%	0.050
WBC (103/µl)	8.01 (7.5–9.15)	8.81 (6.9–10.4)	100%	0.657
CRP (mg/l)	2.72 (1.1–3.8)	6.60 (1.3–14.5)	87%	0.013
PLT (103/µl)	241 (217–280)	209 (169–253)	100%	0.056
INR	1.01 (0.9–1.1)	1.03 (0.1–1.1)	87%	0.562
APTT (s)	29.95 (26.2–34.8)	31.10 (26.7–33.8)	86%	0.985
Antiplatelet treatment	9 (29%)	10 (35%)	100%	0.650
Statin	6 (19%)	11 (38%)	100%	0.195

P values marked with bold indicate statistical significance <0.05.

Alteration of circulating miRNAs expressions and diagnostic ability of miR-125a-3p and -5p

Figure 1 shows the comparison of circulating miRNAs relative expression between control group and patients with ischemic stroke at 2 different timepoints (at admission and 7-days after admission). The expression of miR-125a-3p was significantly lower in stroke patients both at admission and on day 7 compared with healthy controls (p = 0.013 and p = 0.003, respectively). MiR-125a-5p expression was higher at admission in stroke patients compared with controls; however, its levels decreased significantly by day 7 relative to admission (both p = 0.001). For let-7e and miR-106b-5p, no differences were observed between controls and stroke patients at admission, whereas both miRNAs showed a significant reduction on day 7 compared with day 1 (p < 0.001 for both).

Fig. 1.

Alterations of miRNAs expression among the groups (a) Let-7e; (b) miR-15b-5p; (c) miR-16-5p; (d) miR-106b-5p; (e) miR-125a-3p; (f) miR-125a-5p. Mann–Whitney U test for unpaired comparison, and Wilcoxon test were used between stroke day-1 and day-7. Kruskal–Wallis test followed by Dunn’s post hoc test with Bonferroni correction shows the difference among the three groups. Red p value represents significant difference < 0.05.

Diagnostic values of miR-125a-3p expression at admission (Fig. 2a) and miR-125a-5p (Fig. 2b) were evaluated with ROC analysis. The area under the ROC curve for miR-125a-3p was 0.675 (95% CI, 0.55–0.80) p = 0.013, whereas for miR-125a-5p, it was 0.712 (95% CI, 0.59–0.83) p = 0.001. ROC curve analysis did not show diagnostic potential for the other studied miRNAs (data not shown).

Fig. 2.

Diagnostic utility of miRNAs determined by ROC curve analysis. (a) miR-125a-3p at admission; (b) miR-125a-5p at admission.

Higher miR-16-5p expression at admission is associated with greater stroke severity in patients with ischemic stroke

Patients were divided into two separate groups according to stroke outcome and severity, assessed using the mRS (mRS ≥ 3), the NIHSS score at admission (NIHSS < 5), and the 30-day follow-up evaluation. Based on the NIHSS score at admission, 28 patients (47%) had minor stroke and 32 patients (53%) had moderate stroke. When classified using the mRS score (≥ 3), 31 patients (52%) were categorized as having favorable outcome, whereas 29 patients (48%) presented with unfavorable outcome. Admission levels of miRNAs were analyzed to determine whether these miRNAs could stratify the risk of stroke severity. Patients with moderate-to-severe stroke had significantly higher miR-16-5p levels both at admission (p < 0.001; Fig. 3c) and at the 30-day NIHSS evaluation (p = 0.022; Fig. 3e). Higher miR-16-5p levels were also associated with unfavorable outcomes based on mRS (≥ 3) (p = 0.026; Fig. 3a). Corresponding AUC values with confidence intervals are presented in Fig. 3b, d, and f. In contrast, miR-15b-5p levels were negatively correlated with disease severity. While no significant difference was observed with respect to NIHSS scores, lower miR-15b-5p expression was associated with unfavorable outcome on mRS (≥ 3) (p = 0.007; Fig. 3g). The ischemic stroke cohort was divided into two subgroups based on ROC curve analysis of miR-16-5p levels and mRS outcomes, resulting in a low- and high-expression group. A cut-off value of 10.5914 for miR-16-5p at admission provided 62% sensitivity and 61% specificity for identifying patients with an unfavorable outcome (mRS ≥ 3).

Fig. 3.

Day of admission MiR-16-5p expression association on stroke severity (NIHSS) and favorable outcome (mRS). (a) miR-16-5p box plots for mRS; (b) miR-16-5p ROC for mRS; (c) miR-16-5p box plots for NIHSS at admission; (d) miR-16-5p ROC for NIHSS at admission; (e) miR-16-5p box plots for NIHSS at day-30; (f) miR-16-5p ROC for NIHSS at day-30; (g) miR-15b-5p box plots for mRS; (h) miR-15b-5p ROC for mRS; (i) miR-15b-5p box plots for NIHSS at admission; (j) miR-15b-5p ROC for NIHSS at admission; (k) miR-15b-5p box plots for NIHSS at day-30; (l) miR-15b-5p ROC for NIHSS at day-30. The initial stroke severity was assessed using the NIHSS, with minor stroke defined as a score of 1–4 and moderate stroke as 5–15. A favorable outcome was defined as an mRS ≥ 3.

According to the multivariate logistic regression model, a high miR-16-5p expression at admission, together with CRP level at admission, were independent predictors of unfavorable stroke outcome based on mRS ≥ 3 (OR: 5.75; 95% CI, 1.12–29.54; p = 0.036 and OR: 1.47; 95% CI, 1.03–2.08; p = 0.032, respectively) (Table 3).

Table 3.

Multivariate logistic regression model for prediction of unfavorable stroke outcome based on mRS ≥ 3 by expression of high miR-16-5p at admission.

Variable	OR	95% CI		p-value
		Lower	Upper
High miR-16-5p vs low miR-16-5p (at admission)	6.174	1.132	33.663	0.035
Gender (female)	0.550	0.075	4.050	0.557
Age (years)	1.021	0.944	1.105	0.598
Smoking	0.4581	0.070	3.319	0.457
CAD	0.525	0.055	5.007	0.576
CRP	1.480	1.042	2.101	0.028
Platelet-lowering drugs	1.931	0.303	12.289	0.486
PLT	1.001	0.984	1.017	0.934

P values marked with bold indicate statistical significance < 0.05.

Results of the bioinformatic analysis

SHAP analysis

Efficient classification of clinical data, particularly in the context of disease severity, is essential for timely diagnosis and informed clinical decision-making. Delays in assessment can increase healthcare costs and patient risk. Machine learning (ML) offers enhanced predictive accuracy and efficiency, improving the detection of clinically relevant patterns. In this study, we utilized SHapley Additive exPlanations (SHAP) to enhance interpretability and identify the most influential features associated with drug response. SHAP assigns Shapley values, a game-theoretic measure that quantifies each feature’s contribution to model predictions, thereby improving transparency and understanding of complex ML outputs.

SHAP analysis was applied to both laboratory and demographic data, allowing us to identify key predictors influencing patient favorable outcomes. By calculating the contribution of each variable, SHAP reveals how deviations from values affect individual predictions, thus addressing the interpretability challenge common in ML models. This approach not only clarifies feature importance but also supports optimization of classification performance (Fig. 4).

Fig. 4.

Results of SHAP method applied by using XGBoost classifier to the panel; (a) laboratory; (b) Demographic part of clinical data. Results show the local contribution of each feature to the prediction for each sample and the most important features associated with the mRS outcome. The features with the highest importance for a given outcome are localized at the top of the graph.

The most influential laboratory features included high CRP, miR-15a-5p lower expression, high dysarthria, high miR-16–5, and low WBC linked to high severity mRS (Fig. 4a). Among demographic variables, the principal contributors were low Barthel index of activities on day 30, high stroke type, and high Scandinavian stroke scale on day 7 (Fig. 4b).

Target prediction and enrichment analysis

In order to characterize functions of analyzed miRNAs, we performed a series of bioinformatic analyses. Firstly, we identified targets of the analyzed miRNAs. We found 4366 targets for hsa-miR-15b-5p, 4633 for hsa-miR-106b-5p, 2185 for hsa-miR-125a-3p, 2611 for hsa-miR-125a-5p, 3882 for hsa-let-7e-5p and 7975 targets for hsa-miR-16-5p. Enrichment analysis of stroke-related phenotypes showed a significant association of targets only for hsa-miR-16-5p. Ischemic stroke was ranked in the first place, followed by reperfusion injury (Fig. 5a). For hsa-let-7e-5p, it was a reperfusion injury. For hsa-miR-125a-5p, it was Cerebral ischemia, while for hsa-miR-125a-3p, it was cardioembolic shock. For hsa-miR-106b-5p most significant was cerebral ischemia, while for hsa-miR-15b-5p, it was coronary arteriosclerosis.

Fig. 5.

Ranked enrichment analysis of targets regulated by selected stroke-related miRNAs. (a) Enrichment analysis of selected ontological terms; (b) Enrichment analysis of selected stroke-related phenotypes obtained from DisGeNET database. On the graph we highlighted top 5 terms for each miRNA. Analysis was performed using Fisher’s exact test with FDR correction. The adjusted p-values show categories that are more likely to have biological meanings. The color gradient corresponds to the adjusted p-values: blue indicates high p-values (low enrichment), while red indicates low p-values (high enrichment). The size of the dots represents the number of enriched genes.

Enrichment analysis of ontological terms revealed strongest association of hsa-miR-16-5p and hsa-miR-106b-5p with hypoxia. In the second place for hsa-miR-16-5p were combined platelet-related terms. Other terms for this miRNA were as follows hemopoiesis, coagulation, platelet activation signaling and aggregation (Fig. 5b). Enrichment analysis of pathways showed highest overrepresentation of terms like cell cycle, TGF-beta regulation of extracellular matrix and BDNF signaling for hsa-miR-16-5p (Fig. 5a). Interaction network analysis revealed that the top target of hsa-miR-16-5p was IL-6, a key molecule associated with platelet function and inflammatory pathways (Fig. 5b).

Interaction network analysis

In order to identify through which targets miR-16-5p influences processes related to ischemic stroke, we constructed a target-target interaction network. Further, we selected the top ten genes present on our gene lists of interest, with the highest degree of connections with other nodes, and all genes associated with ischemic stroke. Ordering targets based on decreasing degree of connections showed IL6 as the top interactor. It was followed by IL6, FN1, TGFB1, ICAM1 and TLR4 (Fig. 6.).

Fig. 6.

(a) Ranked pathway enrichment analysis of targets regulated by selected stroke-related miRNA’s terms are sorted by FDR value. On the graph, we highlighted the top 5 terms for each miRNA. Analysis was performed using Fisher’s exact test with FDR correction. The adjusted p-values show categories that are more likely to have biological meanings. The color gradient corresponds to the adjusted p-values: blue indicates high p-values (low enrichment), while red indicates low p-values (high enrichment). The size of the dots represents the number of enriched genes; (b) Interaction network of hsa-miR-16-5p top targets potentially playing a role in ischemic stroke. Targets were ordered by decreasing degree of connections.

Discussion

Circulating miRNAs have been identified as potential biomarkers for cardiovascular and cerebrovascular diseases. In our study, circulating miRNAs showed distinct expression patterns in acute ischemic stroke and were associated with both disease presence and early clinical severity. MiR-125a-3p was consistently lower, and miR-125a-5p was transiently elevated at admission in stroke patients compared with controls, with both exhibiting moderate diagnostic value. Among all examined miRNAs, miR-16-5p demonstrated the strongest relationship with stroke severity: higher admission levels were associated with moderate-to-severe neurological deficits based on NIHSS and unfavorable outcomes based on mRS scoring. In contrast, miR-15b-5p showed an inverse association, with lower levels linked to worse functional status.

MiR-16-5p has previously been assessed in ischemic stroke patients and shown potential diagnostic and prognostic value^31–34. The levels of circulating miR-16 on the first day after ischemic stroke diagnosis was increased more than eightfold compared to non-stroke controls. However, serum miR-16 expression was not an independent factor and showed strong negative correlation with HDL cholesterol levels and ApoA1, and a positive correlation with age. These risk factors are common not only in the stroke population but also in other cardiovascular patients, which may indicate the low specificity of miR-16 as an ischemic stroke biomarker. The value may be improved by using a panel of stroke-related miRNAs, including miR-16³¹. Moreover, Tian et al. revealed that circulatory miR-16 levels in the hyperacute phase depend on the etiology and are the most specific and sensitive for detecting large-artery atherosclerotic stroke. In the overall stroke population, higher levels of miR-16 in the hyperacute phase were associated with poor prognosis³². Another study showed that miR-16 may be useful for differentiating ischemic from hemorrhagic stroke, especially when combined with miR-124-3p. In ischemic stroke patients, the expression of miR-16 was increased, whereas miR-124-3p was decreased. The opposite situation occurred in the group of patients with hemorrhagic stroke³⁴. Thus, miR-16 may serve as a diagnostic and prognostic biomarker in the future, but as it was seen in previous studies, rather as part of a panel of miRNAs than individually.

The role of miR-16-5p in ischemic stroke pathophysiology remains unclear. It was revealed that miR-16-5p may bind adiponectin receptor 2 (AdipoR2), abundant in brain tissue and involved in the activation of peroxisome proliferator-activated receptor γ (PPARγ)³⁵. PPARγ modulates inflammatory processes, and its overexpression attenuates neuroinflammation in rats with cerebral ischemia³⁶. Another study showed that circulating RNA circ_0000831 downregulates miR-16-5p and reduces inflammation and apoptosis in oxygen-glucose-deprived (OGD) microglia. The circ_0000831/miR-16-5p/AdipoR2 axis was found to be an important signaling pathway in ischemic stroke pathophysiology³⁵. Moreover, miR-16-5p and miR-125-5p were contained in astrocyte-derived extracellular vesicles (EVs) shed in response to proinflammatory cytokines IL-1 and TNFα. It was shown that both miRNAs may target and downregulate genes involved in the neurotrophin-signaling pathway, namely NTKR3 and Bcl2. Inhibition of NTKR3/Bcl2 in neurons was associated with reduced dendritic growth and complexity and attenuated excitability³⁷. Furthermore, NTRK3 may activate the transcription factor CREB and thereby regulate the expression of genes promoting neuronal survival and differentiation³⁸. Thus, miR-16-5p may be secreted via EVs in response to brain tissue damage and may be involved in neuronal protection. Altered expression of miR-16 has also been noted in other diseases associated with neuronal damage, e.g., amyotrophic lateral sclerosis³⁹, Alzheimer’s disease⁴⁰, and Creutzfeldt-Jakob disease⁴¹. In line with our previous results miR-16-5p is one of the top miRNAs associated with platelet reactivity¹⁵. Since the activation of the coagulation system is one of the main mechanisms leading to brain vessel occlusion and ischemia the role of platelets should be carefully investigated, especially when activated, they can secrete molecules, e.g., miRNAs that are internalized and affect gene expression in nearby cells¹⁶. MiR-16-5p shows promise as a biomarker for ischemic stroke, but its specificity may be improved when used in conjunction with other miRNAs.

Although in the presented study no significant differences were found in miR-106b-5p levels between the stroke group and the control group, a decrease was observed during hospitalization. The significance of this observation remains unclear. A systematic review focused on miRNAs as biomarkers in the acute phase of stroke revealed that miR-106b-5p was differentially expressed, but the detailed results are contradictory⁴². Notwithstanding, miR-106b-5p silencing decreased infarct volume and reduced neurological deficits in rat models of ischemic stroke indicating its significant contribution to the pathophysiology of stroke⁴³.

MiR-125a-5p was previously studied as a promising blood-based biomarker of ischemic stroke. It was shown that circulating miR-125a-5p had a high specificity and sensitivity in discriminating between ischemic stroke patients and controls. When combined with two other stroke-related miRNAs, the sensitivity was greater than in multimodal CT scanning. However, the source of miR-125a-5p is not neural tissue but platelets; thus, the levels may be affected by changes in platelet count and antiplatelet therapy¹⁸. MiR-125a-5p may inhibit angiogenesis and play a role in the function of endothelium in the blood–brain barrier (BBB)^44,45. In vitro and animal studies revealed that miR-125-5p expression is increased in microglial cells after oxygen–glucose deprivation/reoxygenation, as well as in the rat ischemic stroke model. Moreover, inhibition of miR-125-5p resulted in attenuation of microglial apoptosis through targeting IGFBP3⁴⁶. The role of miR-125a-3p in the pathophysiology of ischemic stroke has not been described so far. However, in the animal model, miR-125a-3p expression in the acute phase of stroke has been shown to change after the administration of recombinant tissue plasminogen activator, which is the standard treatment of ischemic stroke⁴⁷. Importantly, we compared our data with the transcriptomic analysis in the study GSE291442⁴⁸ based on RNA-sequencing of circulating neutrophils collected from patients with transient ischemic attack (TIA), and identified several miRNAs with differential expression between control and ischemic stroke samples. Among them, hsa-miR-16-5p showed the strongest statistical significance (adj.P = 0.0116, P = 3.5 × 10⁻⁶) and a large negative logFC (− 3.41), indicating a marked increase in ischemic stroke relative to controls. Other miRNAs showing the same directional change (increased in stroke, though not passing FDR correction) include hsa-miR-106b-5p, hsa-miR-125a-5p, and hsa-miR-15b-5p, all with moderate negative logFC values (− 0.5 to − 0.6). In contrast, hsa-let-7e-5p showed a small negative logFC (− 0.25) with no statistical significance, and hsa-miR-125a-3p exhibited a slight positive logFC (+ 0.066), indicating no relevant differential expression.

Beyond the miRNAs analyzed in our study, several previous investigations have explored other stroke-related circulating miRNAs, with miR-124 being among the most frequently examined. In diabetic patients with acute ischemic stroke, plasma miR-124 levels were markedly downregulated, and this reduced expression showed a significant negative correlation with elevated lncRNA NEAT1 and inflammatory markers (CRP, TNF-α). These findings suggest that suppressed miR-124 expression may contribute to the heightened inflammatory response and more severe stroke phenotype observed in patients with diabetes⁴⁹. Similarly, another study reported that serum miR-124 was significantly downregulated in patients with acute cerebral infarction, demonstrating strong diagnostic performance (AUC 0.95). miR-124 levels were negatively correlated with IL-6, IL-8, and CRP, and reduced expression independently predicted poor clinical outcomes, supporting its potential as a dual biomarker for both diagnosis and prognosis⁵⁰. A recent meta-analysis (preprint) synthesizing data from 27 studies identified several high-confidence biomarkers across different time points after stroke onset. The largest effect sizes were observed for S100B, GFAP, and IL-6. Across 220 differentially expressed miRNAs (132 upregulated, 108 downregulated), a subset of robust markers (miR-16-5p, let-7e-5p, miR-107, miR-451a, and miR-126-3p) were validated in three or more independent cohorts. The analysis also reported that miR-124-3p was associated with dysregulated excitotoxicity and NF-κB pathway activation. Overall, the study proposed a temporally dynamic pattern of biomarker release following stroke: glial activation (GFAP, S100B), and inflammation (IL-6)⁵¹. In addition, findings from an emergency-department study conducted in a setting without on-site neurology support the potential utility of glial markers for rapid triage. GFAP levels were significantly higher in stroke patients compared with stroke mimics within the first 4.5 h of symptom onset. Notably, GFAP differentiated hemorrhagic from ischemic stroke with 94.2% accuracy, underscoring its promise as a point-of-care biomarker to guide early management decisions⁵².

Bioinformatic enrichment performed for this study showed enrichment for hsa-miR-16-5p of several signaling pathways, including cell cycle, TGF-beta regulation of extracellular matrix, and BDNF signalling. Stroke results from a temporary or permanent reduction in brain blood flow. The mechanisms leading to neuronal death after an ischemic event are complex and not fully understood. As a result, neurons in the affected brain region, or throughout the brain in the case of a cardiac arrest, suffer from a lack of oxygen and glucose. One factor that may influence this process is the abnormal activation of cell cycle regulators like cyclins, cyclin-dependent kinases (CDKs), and cyclin-dependent kinase inhibitors (CDKIs)⁵³. It has been proposed that neuronal death may be regulated by the improper activation of certain cell cycle elements⁵⁴. While these regulators promote cell proliferation in dividing cells, their inappropriate activation in mature neurons can lead to neuronal death. Evidence supports this “cell cycle/neuronal death hypothesis,” and understanding how these pathways contribute to ischemic neuronal death could lead to effective therapeutic strategies for stroke.

TGF-β1 may reduce the expression of matrix-degrading proteases, such as matrix metalloproteinases (MMPs), produced by hypoxic brain endothelial cells⁵⁵. TGF-β1 interacts with the extracellular matrix, leading to collagen accumulation in the vascular wall, reduced vascular elasticity, vascular remodeling, and increased blood pressure⁵⁶. It plays a role in the pathological processes of vascular development and remodeling in stroke patients. The Brain-Derived Neurotrophic Factor (BDNF) pathway plays a crucial role in ischemic stroke by promoting neuronal survival and plasticity. Studies have shown that BDNF is involved in neuroprotection and repair mechanisms following cerebral ischemic injury⁵⁷. Activation of the BDNF pathway can enhance synaptic plasticity, neurogenesis, and neuronal survival, contributing to functional recovery after stroke⁵⁷. The BDNF pathway induces Nrf2 nuclear translocation via ROS, RyR-mediated Ca2 + signals, ERK, and PI3K, potentially mitigating post-ischemic neuroinflammation and oxidative stress in the hippocampus⁵⁸. By modulating the expression of antioxidant and defense proteins, BDNF may help mitigate the pro-oxidative events that occur during the ischemic cascade, potentially offering neuroprotection in stroke settings⁵⁹. Platelets are a major peripheral reservoir of BDNF, storing it in α-granules and releasing it upon activation⁶⁰. Following ischemic stroke, circulating and platelet BDNF levels fluctuate, reflecting platelet activation and secondary neurovascular responses⁶¹. Released BDNF promotes angiogenesis, neurogenesis, and synaptic plasticity, supporting post-stroke recovery^61,62. Low platelet or plasma BDNF is associated with worse neurological outcomes and higher infarct volume in patients⁶³. Thus, platelet-derived BDNF serves as both a biomarker of platelet activation and a potential mediator of neurorepair in ischemic stroke. BDNF also reduces inflammation, which helps limit secondary damage⁶⁴. Its neuroprotective and restorative effects make BDNF a promising target for therapies aimed at improving stroke recovery.

Using bioinformatic tools, we identified the potential role of hsa-miR-16-5p in the regulation of crucial processes associated with ischemic stroke, like hypoxia, platelet activation, hematopoiesis, and coagulation. We also observed its regulation by targets associated with ischemic stroke and reperfusion injury phenotypes. Our bioinformatic analysis pointed out several gene targets regulated by hsa-miR-16-5p, such as IL6, FN1, TGFB1, ICAM1, and TLR4. IL-6 is a key cytokine in ischemic stroke, driving inflammation and neuronal injury⁶⁵. It is released by neurons, microglia, astrocytes, and endothelial cells, peaking 24 h after ischemia and remaining elevated for up to 7 days, particularly in the ischemic penumbra⁶⁶. Elevated IL-6 correlates with vascular dysfunction, higher body temperature, early neurological decline, larger infarcts, and poor outcomes⁶⁵, though it may also support neurogenesis and tissue repair⁶⁷. IL-6 enhances platelet responsiveness and amplifies thrombo-inflammatory cascades, promoting platelet activation and aggregation⁶⁸. Through platelet-leukocyte interactions and immune cross-talk, IL-6 exacerbates reperfusion injury and secondary infarct expansion^69,70. It is a valuable biomarker for stroke severity and recovery, with potential as a therapeutic target.

FN1 encodes fibronectin, a multifunctional glycoprotein essential for cell adhesion, migration, and tissue repair, with critical roles in ischemic stroke⁷¹. After stroke, fibronectin supports angiogenesis, extracellular matrix remodeling, and BBB stability while modulating inflammatory signaling^72,73. However, it can also promote cerebrovascular amyloid-β accumulation with aging or post-stroke⁷⁴. Platelets serve as a major reservoir of fibronectin (stored in α-granules and released upon activation), creating an autocrine source that enhances local clot formation and vessel repair⁷⁵. Both plasma and platelet-derived fibronectin, including EDA⁺ isoforms, bind to platelet receptors (α5β1, αIIbβ3, GPIb-V-IX, GPVI, and TLR4) to promote adhesion, fibrillogenesis, and thrombus stabilization^76,77. The EDA domain directly engages platelet TLR4, triggering prothrombotic signaling in ischemic settings⁷⁸. Furthermore, plasma fibronectin forms complexes with fibrin and fibrinogen, influencing clot architecture and mechanical stability^79,80. Through these platelet-dependent mechanisms, FN1 acts as both a mediator of thrombo-inflammation and a potential therapeutic target in ischemic stroke. Given its involvement in these processes, FN1 is considered a potential therapeutic target for improving stroke outcomes.

Transforming growth factor-β1 (TGF-β1) is a pleiotropic cytokine that modulates inflammation, supports tissue repair, and promotes angiogenesis after ischemic stroke⁸¹. Its levels rise in brain tissue and cerebrospinal fluid following ischemia, exerting neuroprotective and anti-inflammatory effects that limit neuronal death^82,83. Platelets are a major reservoir of latent TGF-β1, which is released and activated during platelet aggregation at vascular injury sites, contributing to local immune modulation and vessel remodeling^84,85. Platelet-derived extracellular vesicles (pEVs) deliver bioactive TGF-β1 to endothelial and neural cells, activating Smad2/3 signaling that enhances angiogenesis, neuroprotection, and blood–brain barrier repair^86–88. While TGF-β1-enriched pEVs improve recovery in experimental stroke models^87,89, excessive or prolonged TGF-β1 signaling may promote fibrosis, underscoring the need for controlled therapeutic targeting^89,90. Due to these roles, TGFB1 is considered a potential therapeutic target for enhancing recovery after ischemic stroke.

ICAM1 plays a central role in ischemic stroke by promoting leukocyte adhesion, endothelial activation, and BBB disruption, thereby amplifying inflammation and secondary injury. Serum ICAM-1 levels are elevated in patients with poor stroke outcomes, indicating prognostic value⁹¹. Endothelial ICAM-1 upregulation in the ischemic penumbra enhances platelet adhesion and microvascular obstruction, while its inhibition reduces thrombocyte recruitment and improves recovery⁹². ICAM-1 also mediates platelet-leukocyte interactions through β2-integrin binding, facilitating platelet–monocyte aggregate formation and reperfusion failure^93,94. Conversely, platelet-derived EVs carrying miR-223 can downregulate endothelial ICAM-1, forming a feedback loop that modulates vascular inflammation⁹⁵. Overall, ICAM-1 acts as both a mediator and amplifier of platelet-driven thrombo-inflammation in ischemic stroke.

TLR4 (Toll-like receptor 4) is a key mediator of innate immune and thrombo-inflammatory responses in ischemic stroke. Activated by damage-associated molecular patterns (DAMPs) released from injured neurons, TLR4 amplifies neuroinflammation and tissue injury^96,97. Platelet-expressed TLR4 enhances platelet activation, aggregation, and coagulopathy, while global or platelet-specific TLR4 deficiency reduces infarct size and inflammation^98,99. Platelet microvesicles carrying HMGB1 further activate neutrophil extracellular traps (NETs) and thrombosis, exacerbating ischemic injury¹⁰⁰. Conversely, therapeutic extracellular vesicles and miRNAs targeting the TLR4/NF-κB pathway mitigate inflammation, protect the BBB, and improve outcomes^101,102. Overall, TLR4 represents a crucial platelet-immune signaling hub and promising therapeutic target in ischemic stroke^103–105. Pharmacological modulation of TLR activation holds promise for developing therapies aimed at preventing and treating stroke. Due to its central role in amplifying inflammation and contributing to neuronal damage, TLR4 is a potential therapeutic target for reducing inflammation and improving outcomes after a stroke.

Limitations

Our study has several limitations. First, the sample size was relatively modest, which may limit the generalizability of our findings. Second, only a selected panel of platelet-related miRNAs was analyzed using RT-qPCR, without transcriptome-wide sequencing to capture additional regulatory non-coding RNAs. Third, while our bioinformatic predictions and enrichment analyses identified key pathways and target genes (IL6, FN1, TGFB1, ICAM1, TLR4), these targets were not experimentally validated in platelets or plasma extracellular vesicles. To mitigate this, we reanalyzed several GEO datasets (GSE22255¹⁰⁶, GSE255087¹⁰⁷, GSE241280⁴⁸, GSE202518¹⁰⁸) related to ischemic stroke, but didn’t find consistent results for analyzed proteins; they were neither significant nor present after filtering of signals. We explain this small sample size in this study. Future studies integrating larger cohorts, multi-omics profiling, and functional validation are warranted to confirm these findings.

Conclusions

In this study, we evaluated several circulating miRNAs and demonstrated that their expression levels differ in the acute phase of ischemic stroke. Among them, miR-16-5p, miR-125a-3p, and miR-125a-5p showed the most consistent associations, with higher admission levels of miR-16-5p correlating with greater stroke severity at presentation. On the other hand, our study emphasizes that miR-16-5p is unlikely to be stroke specific and is likely to be most useful as part of a biomarker panel rather than as a standalone marker as it is also known to be associated with age, inflammation, platelet activity and multiple cardiovascular conditions. Our bioinformatic analyses further supported the involvement of these miRNAs in biological pathways relevant to ischemic injury, including inflammation, hypoxia, and vascular dysfunction. While these findings highlight the potential of circulating miRNAs as biomarkers reflecting early pathophysiological processes in acute ischemic stroke, our data primarily describe associations with severity at admission. They do not assess subsequent clinical course or long-term outcomes; therefore, any prognostic interpretation should be made with caution. Overall, this work contributes to the understanding of circulating miRNAs in acute ischemic stroke and provides a foundation for future studies that will require larger, longitudinal cohorts to determine their diagnostic utility and potential value in disease severity. Further mechanistic and longitudinal studies are warranted to validate these biomarkers and elucidate their functional roles in stroke pathophysiology.

Author contributions

Writing—original draft preparation, CE, PC, ZW, MP; writing—review and editing, CE, PC, ZW, AS, DMG, AC, MP, ID, AWC, SA; bioinformatic analysis– ZW, visualization, CE, ZW; supervision, CE, DMG, MP; laboratory and computer analysis, CE, SA; Data preparation; CE, ID, AWC. All authors have read and agreed to the published version of the manuscript.

Funding

MP was supported financially as part of the research grant ‘OPUS’ from National Science Center, Poland (grant number 2018/31/B/NZ7/01137) and Medical University of Warsaw (grant number 1M9/2/M/MBM/N/21).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

Authors declare no conflict of interest related to this work.

Ethical approval

The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Ethics Committee of Medical University of Warsaw, Warsaw, Poland (protocol code KB/148/2017, date of approval: 4 July 2017).

Informed consent

Informed consent was obtained from all subjects involved in the study and written informed consent has been obtained from the patient(s) to publish this paper.