Urinary Cytokines as Potential Biomarkers of Mild Cognitive Impairment and Alzheimer’s Disease: A Pilot Study

Study design and sample preparation

We measured the concentration of 37 inflammatory cytokines in urine samples obtained from AD patients (n = 25), MCI sufferers (n = 25), and age and gender-matched cognitively normal controls (n = 26). The participants were recruited from Beaumont Health Outpatients. The diagnosis and evaluation of patients with AD and MCI sufferers was carried out by a fellowship-trained geriatrician according to the criteria of the National Institute of Neurological and Communicative Disorders and the Stroke (NINCDS) and Alzheimer’s Disease and Related Disorders Association (ADRDA) [15]. Participants underwent a focused history and physical examination (assessing motor strength and tone, the existence of a tremor, sensation, balance (Romberg), and gait) to include an exhaustive cognitive testing battery routinely utilized in the Geriatric Clinic at Beaumont health to include: Mini– Mental State Examination (MMSE), Saint Louis University Mental Status (SLUMS), clock drawing tasks (CLOX-I, and CLOX-II), trial making tests (Trails making A, and Trails making B), and Geriatric Depression Scale. The study was approved by the Beaumont Health Institutional Review Board (IRB# 2014-038) and all collection/testing parameters were carried out in accordance with the approved guidelines. The collection was completed in the early morning and all subjects were asked to refrain from eating, drinking, or smoking for at least 1 h before urine collection. Urine was collected in a polypropylene container, then aliquoted into an Eppendorf tube and stored at –80°C for further analysis.

Cytokine assay

Urine samples were defrosted and centrifuged at 1000× g for 5 min at 4°C and 50μl of the supernatant was analyzed according to the manufacturer’s protocol. In brief, using the Bio-Plex Pro™ Human Inflammation Panel 1, 37-Plex #171AL001M (BIO-RAD-USA), we measured 37 inflammatory cytokines from the TNF superfamily of proteins, the IFN family of proteins, Treg cytokines, and Matrix metalloproteinases (MMPs). While the methodology for the cytokine assay was done robustly, it should be noted that the limitation of the study is that only one technical replicate was acquired. This may have impacted the reliability and reproducibility of the results, and further studies with multiple technical replicates may be necessary to confirm our findings. The assay plates were read using a Luminex 200^® analyzer system, and the raw concentration values for each inflammatory biomarker were calculated using the manufacturer’s software.

Univariate analysis

Before statistical analysis, any variable with >40% missing values were excluded from our analysis. Subsequently, the concentration for each cytokine was normalized using serially diluted internal calibrants, and standard curves were generated. The normality of the data was assessed by the Kolmogorov-Smirnov test of normality. Subsequently, using MetaboAnalyst (v 5.0) [16], an independent t-test and Mann Whitney U test were performed for all pair-wise comparisons for both parametric and non-parametric distributions, respectively. All demographic information was analyzed using the SPSS Statistics toolbox (ver 24.0).

Machine learning models

A total of 11 machine learning algorithms were evaluated in this study, including logistic regression, linear discriminant analysis, linear support vector machine (SVM), random forest, decision tree, xgboost, K-nearest Neighbor (knn), gaussian naïve bayes, and kernel SVM. Before examining the diagnostic performance of the models, each urinary cytokine was generalized log-transformed (glog) [17] and auto-scaled. Before performing pattern recognition, data from each group were analyzed using principal component analysis (PCA) to identify any potential outliers or systematic variation (p < 0.05). Subsequently, data were divided into training (60% of data) and testing (40% of data) sets. A Recursive Feature Elimination (RFE) method with Logistic Regression as the classifier was utilized to select the best predictor variables and eliminate redundancy in the variable space. Once the best set of markers was identified, model hyper-parameters [18] were optimized using 10-fold cross-validation. The trained models were assessed using the test set and a 10-fold CV method. Model performance was evaluated using the classification accuracy rate, the area under a receiver operating characteristic (AUROC) curve, and the identification of true discriminating features.

Pathway analysis

The differentially expressed inflammatory cytokines between NC and AD were categorized in FunRich software (https://www.funrich.org), an open access, functional enrichment, and network analysis tool [19]. The biological function was divided into four components: biological processes, transcription factors, cellular components, and molecular functions. The biological pathways and target miRNA for these cytokines were performed using the Enrichr software [20–22]. Differentially expressed cytokines (p < 0.05) between AD and NC were used for the pathway analysis. The miRtargetBase 2017 and targetScanmicroRNA2017 databases were employed to identify miRNAs linked to these cytokines. The LINCRNA database was used to identify corresponding Long intergenic non-coding RNA (LINCRNAs).

Biomarker panels and machine learning models

‘IFN-α2’, ‘IL-19’, ‘IL-12(p70)’, ‘IL-29/FN-λ1’, and ‘APRIL/TNFSF1’ were identified by RFE to be the most discriminative features when NC and AD cases were compared. In particular, K nearest neighbors algorithm (k = 5) with 200 oversampled points provided the best diagnostic outcome for all machine learning (ML) based models. The performance of 11 ML models used in this study was assessed by calculating a classification accuracy rate, AUROC, sensitivity, and specificity values. Based on these performance metrics, the best three classifiers diagnosing AD cases when compared to NC cases were found to be SVM kernel, logistic regression, and ridge clf (Fig. 2A) providing AUC values = 79.4, 69.3, and 69.1, respectively (Fig. 2B). When we compare MCI and NC individuals, ‘IFN-γ’, ‘IL-12(p40)’, ‘IFN-α2’, ‘IL-34’, and ‘MMP-2’ were found to be the most discriminative markers. Of the 11 different ML models we assessed for distinguishing MCI sufferers from NCs, the top three outperforming classifiers were found to be xgboost, knn, and gbm (Supplementary Figure 1A) providing AUC values = 78.6, 68.3, and 67.7, respectively (Supplementary Figure 1B). Finally, for the AD versus MCI comparison, of the 11 different ML models assessed and using RFE selected variables (IL-19, IFN-γ, IL-34, gp130/sIL-6Rβ, and sCD30/TNFRSF8) decision tree, svm kernel and logistic regression were found to be the most discriminative models (Supplementary Figure 2A) with AUC values of = 74.9, 65.4, and 65.3, respectively (Supplementary Figure 2B).

OPEN IN VIEWER

Fig. 2

Performance evaluation metrics for each ML-based model to include sensitivity, specificity, accuracy, and AUC using the RFE variable selection algorithm for distinguishing NC from AD and corresponding AUC curves top five models.

Pathway analysis

We report several biochemical pathways which may have been affected by the abnormal expression of the measured cytokines, and they include molecular functions, biological processes, biological pathways, and transcription factors when comparing AD cases with NC. Within the molecular function aspect, cytokine activity (p < 0.001) was identified as the most significantly different pathway when comparing AD and NC (Supplementary Figure 3A). In biological processes, the immune response pathway (p < 0.001) was the most significantly perturbed biochemical pathway (Supplementary Figure 3B) and in the biological pathway, validated transcriptional targets of the AP1 family (p = 0.017) and syndecan-2 mediated signaling events (p = 0.043) were identified as being significantly perturbed (Supplementary Figure 3C). Nuclear factor I C (NFIC) was the most significant transcription factor (TF) identified (p = 0.005) (Supplementary Figure 3D). We also identified several LINCRNAs whose expression could be modified by these cytokines. LINC01937 (p = 0.00001269) and LINC02376 (p = 0.00001269) were the top LINCRNAs we identified as being perturbed (Supplementary Table 3A). Subsequent Clustergram highlights MMP-2 as having significant co-expression with LINC01937 (p = 0.00001269), LINC01920 (p = 0.0004032), and LINC00687 (p = 0.0004032) (Supplementary Figure 4A).

Using an in-silico approach, we identified several miRNAs that may regulate these cytokines. FunRich provided gene hits on mmu-miRNAs; however, to maintain uniformity, we have converted all mmu-miRNAs to hsa-miRNAs based on the homolog search using Ingenuity Pathways Analysis [23]. Among the top 10 significant miRNAs (Supplementary Figure 4B), the human homolog of hsa-miR-503-3p (p = 0.00238) and hsa-miR-27a-3p (p = 0.0018) were the most significant. Of the top 10 miRNAs, hsa-miR-7152-3p (p = 0.015), hsa-miR-142a-3p (p = 0.020), hsa-miR-130a-3p (p = 0.025), and hsa-miR-4661-5p (p = 0.02) are regulated by IL6ST (Supplementary Table 3B). We also identified several miRNAs targeted by these cytokines. hsa-miR-4278 (p = 0.027) was the most significant target miRNA of IL6ST and MMP-2 (Supplementary Table 3C, Supplementary Figure 4C). Cytokines are widely known to modulate TF and affect molecular pathways. The in-silico analysis revealed an increase in the expression of several transcription factors, suggesting a possible regulatory role in our study (Supplementary Table 3D, (Supplementary Figure 4D). The pathway analysis further identified proteins MMP-2 and IL6ST in targeting the TF, FLI1.

In silico analysis

miRNA and LINCRNA play an important role in neurite outgrowth, neuronal differentiation, synaptic plasticity, dendritic spine morphology [51, 52], and various cellular functions by interacting at the mRNA level resulting in the inhibition of protein synthesis [53]. In the immune system, miRNA regulates immune cell function in response to several stimuli [54], contributing to neural dysfunction in AD brain [51]. We identified several miRNAs and LINCRNAs that these inflammatory biomarkers could target. The miRNAs fall into two groups: a) miRNA that could be causal regulatory factors for cytokine expression, and b) Differentially expressed cytokines that target miRNAs. However, both groups are likely to play a critical role in AD pathogenesis [51]. Among regulatory miRNAs, hsa-miR-7152-3p, hsa-miR-142a-3p, hsa-miR-130a-3p, and hsa-miR-4661-5p are regulated by cytokine IL6ST. Hence, these miRNAs could prove to be promising therapeutic candidates for AD pathogenesis. Similarly, hsa-miR-4278 targets proinflammatory cytokine IL6ST and MMP-2. These miRNAs have not been previously reported for their role in AD pathogenesis; however, this novel finding requires further investigation.

Moreover, we identified LINC01937, LINC01920, and LINC00687 target MMP-2. As previously, these could prove to be promising therapeutic targets. For instance, the development of anti-oligonucleotide strategies to regulate these in those individuals suffering from AD.

Conclusion

This cross-sectional study measured cytokine levels in urine samples from AD, MCI, and cognitively normal subjects. Our study demonstrates the potential of inflammatory biomarkers, particularly cytokines in urine, as noninvasive diagnostic biomarkers of AD. Further, using an in-silico approach, we link the changes as observed in the dementia spectrum to the potential onset and pathogenesis of AD. There are several limitations to this study. First, the sample size used herein is modest, but we believe with a much larger sample set we would have the power to demonstrate their potential clinical utility. Second, systemic inflammatory diseases and comorbidities might affect the bladder and related urine conditions which have the potential to confound the results. Thus, additional studies are warranted to determine the importance of these changes, how inflammation is involved in AD development, and whether these targets can slow or stop the progression of the disease.