Circulating brain-enriched microRNAs as novel biomarkers for detection and differentiation of neurodegenerative diseases

Thirty-seven miRNAs were analyzed in two 0.5-ml aliquots (1 ml total) of K₂-ethylenediaminetetraacetic acid-treated frozen plasma samples from 250 research participants (n = 50 for each ND) recruited and clinically evaluated at the University of Pennsylvania Health System, including the Alzheimer’s Disease Center (in 2003–2014), the Parkinson’s Disease and Movement Disorders Center (in 2010–2015), the Frontotemporal Degeneration Center (in 2010–2013), and the Amyotrophic Lateral Sclerosis Clinic (in 2010–2014). Fifty samples from age- and sex-matched control subjects were collected in 2009–2015. The ND conditions were clinically diagnosed according to published procedures [50‐58]; demographic characteristics of the study groups, including control subjects, are summarized in Table 1 and Additional file 1. The study participants in the AD group were selected on the basis of AD phenotype and tau and amyloid biomarkers as determined by CSF analysis. The biofluid samples were collected and processed on the day they were obtained following the same standard operating procedures in all specialty clinical practices at the Center for Neurodegenerative Disease Research, University of Pennsylvania [59]. Blood samples were centrifuged at 3000 × g for 15 minutes at 4 °C and aliquoted into 2.0-ml polypropylene cryovials (Corning Life Sciences, Corning, NY, USA). The aliquots were stored at −80 °C. De-identified samples were sent to the Asuragen laboratory (Austin, TX, USA) for RNA isolation and qRT-PCR (see below). Clinical information, including demographics, diagnosis and cognitive scores, and identifiers to match the samples, were sent to the DiamiR principal investigator, who kept the information blinded from the laboratory personnel.

All groups were randomly divided into two sets of equal size (25 samples each, balanced by age, sex, and comorbidities) and used in training and confirmation experiments. The PD and ALS groups had fewer female participants, and the ALS group was younger than other ND and control groups. Age may affect miRNA expression, and thus differences in the age composition of different groups may contribute to the differentiation of the groups with specific miRNAs. In the present study, however, the average ages were not significantly different between various groups (see Additional file 1); we found that the exclusion of several outliers did not change the results. Correlations between miRNA ratios and age in all pathology cohorts are low (r < 0.2) and have low statistical significance (p > 0.04). To determine the source of miRNA ratio differences in disease groups and the control group, we performed linear regression analysis of miRNA ratios using disease group identity and age as covariates. p Values for the significance of the disease group identity for all miRNA pairs selected were < 0.0001, indicating that the differentiation of two groups depended on diagnosis. Only in the ALS group did some miRNA pairs correlate with age with p < 0.05. Age adjustment for these pairs slightly increased the accuracy (1–2%); age adjustment for all other pairs either did not change or slightly improved the accuracy (by < 2%). Similarly, to determine the effect of sex on differentiation of PD and ALS (the two NDs with smaller numbers of female participants; see Table 1 and Additional file 1) from control subjects, we performed a linear regression analysis of miRNA ratios using sex as a covariate. For most pairs, the effect of sex adjustment was insignificant. Additional stratified analysis of all NDs (described in the “miRNA biomarker sex-dependent effects” subsection of the Results section below) revealed miRNA pairs differentiating male and female participants separately better than all (male and female) participants.

RNA was extracted from 1 ml of plasma using a TRIzol treatment (Life Technologies, Carlsbad, CA, USA) and silica (Ambion Glass Fiber Microcolumn; Fisher Scientific, Pittsburgh, PA, USA) binding protocol (http://​asuragen.​com/​wp-content/​uploads/​2016/​05/​biomarkers.​pdf). Single-target qRT-PCR was performed using the TaqMan® Reverse Transcription Kit and miRNA-specific stem-loop primers (Applied Biosystems, Foster City, CA, USA). The RT step for miRNA measurements was performed in triplicate and each RT reaction was then used separately for PCR. 2µl plasma equivalents were present in the final PCR. (Additional file 2 presents the qRT-PCR variability for each tested miRNA.) Placental RNA was used as a positive control and no template as a negative control. Calibration curves for each miRNA were generated to facilitate comparing and combining of training and confirmation datasets. Quality control of miRNA preparations was performed by testing two ubiquitous miRNAs, miR-16 and miR-27a, in each plasma preparation; all samples with values within 2 SD of the average value were qualified as acceptable for analysis. The average SDs of miR-16 and miR-27a were 1.40 and 1.01 in the training and confirmation sets, respectively. miRNAs with cycle threshold (C _t) > 37 were excluded from the analysis of a respective sample. Although small RNA yield was not directly measured after RNA extraction, the control data reported in Additional file 2 provide a good indicator of miRNA yield.

In our approach, an effective biomarker is a ratio (pair) of two brain-enriched miRNAs. Software developed at DiamiR [48, 49] was used for all calculations, including determination of miRNA pairs capable of differentiating NDs from controls and from each other on the basis of their concentration ratios (∆C _t) in plasma. The application was designed using .NET technology with a set of .NET statistical packages. The selection of miRNA pairs as well as their combinations was based on the AUC and p values as follows. For every pair, ROC curves were constructed to calculate the area under the ROC curve (AUC). In the training set analysis, pairs with AUC ≥ 0.75 for AD, PD, and ALS, as well as AUC ≥ 0.70 for FTD, were initially accepted. Training and confirmation sets were of the same size. A lower cutoff for FTD was used because of high heterogeneity of the group (presence of several FTD syndromes). In the combined set, there were 23 behavioral variant frontotemporal dementia (bvFTD), 17 primary progressive aphasia (PPA), and 10 progressive supranuclear palsy (PSP) cases. The acceptance p value threshold was set on the basis of significance of differences between any two experimental groups, calculated by the Mann-Whitney U test with Bonferroni correction. Validation of miRNA pair selection and testing for overfitting were performed using random forest classification with bootstrapping. For 37 miRNAs used, the total number of pairs was 666 (36 × 37/2). The ratios A/B and B/A are equivalent as potential biomarkers because they have the same variance, difference of means, correlation coefficients, p values, and ROC parameters. In cases where disease was compared with control, the ratio whose numerical value was larger in disease samples was selected. According to the Bonferroni correction, statistically significant results should then have a p value of 0.05/666 = 7.5 × 10⁻⁵. Because we were using highly correlated miRNAs for creating pairs (C _t Spearman’s rank correlation coefficient, ≥ 0.8), the number of pairs analyzed was reduced to about 160, and thus the threshold for statistical significance rose to ~ 3.0 × 10⁻⁴. Pair sensitivity and specificity are reported for the cutoff points on the ROC curves that provided the best overall accuracy. For selecting effective pair combinations, miRNA classifiers, a stepwise algorithm for logistic regression using a linear model with no interaction [60] was applied. The miRNA classifiers having AUC ≥ 0.9 for AD, PD, and ALS, as well as AUC ≥ 0.8 for FTD, in the combined (training + confirmation) set were accepted.

In this work, we performed targeted preselection of miRNAs present in synapses, enriched in different brain regions, and also detectable in plasma. The levels of all preselected miRNAs in plasma samples were determined by qRT-PCR. The miRNAs analyzed in the study are listed in Additional file 3 and included (1) miRNAs present in synapses [33, 35, 36, 61, 62] and enriched in different brain regions affected by the target pathologies [26‐30, 63‐67]; (2) miRNAs associated with inflammatory processes [68‐71]; (3) miR-206 highly enriched in muscle tissue and in cerebellum [68, 72, 73]; (4) ubiquitous apoptosis-associated miR-16 [74]; and (5) miR-451, which is more effectively excreted from pathologic than normal cells [39]. As shown in Additional file 3, certain brain-enriched miRNAs are expressed in several brain regions; typically, the levels of these miRNAs in specific brain regions (shown in bold in Additional file 3) are significantly higher than in others. Potential roles of many of the miRNAs listed in Additional file 3 in NDs are being investigated by several groups [62, 75‐88].

As described above, each group (AD, FTD, PD, ALS, and age- and sex-matched control group) was divided in two equal sets and analyzed independently in the training and confirmation studies. miRNA pairs and their combinations (miRNA classifiers) capable of differentiating each ND from controls with the highest accuracy were assessed in the training set and verified in the confirmation set analysis, followed by the analysis of the combined dataset (Figs. 1 and 2 and Additional file 4). Combining two or more qRT-PCR experiments is often challenging because of possible technical variations in RNA extraction, RT and/or PCR kits, and instrumental fluctuations. Furthermore, samples for the confirmation study were kept frozen for an additional 6 months. Nonetheless, certain miRNA pairs and classifiers effectively differentiated AD, FTD, PD, and ALS from controls in both training and confirmation studies (Additional file 4). The accuracy of ND detection in the combined dataset by miRNA classifiers for AD, PD, FTD, and ALS was 0.89, 0.90, 0.88, and 0.83, respectively. The comparison of the differentiation obtained with individual miRNAs and with miRNA pairs is reported in Additional file 5. Only muscle-enriched miR-206, whose average concentration is about eight times higher in patients with ALS than in control subjects, differentiates these two groups with accuracy similar to that of miRNA pairs. In all other cases, the use of miRNA pairs produces much better differentiation between NDs and controls than individual miRNAs.

The same approach was used for establishing miRNA biomarkers capable of differentiating the four NDs (AD, PD, FTD, and ALS) from each other. Figure 3 and Additional file 6 present miRNA classifiers effectively distinguishing two NDs in the training and confirmation studies as well as in the combined dataset. The overall accuracy in the range of 0.75–0.93 was obtained. The lowest accuracy (0.77 with most effective miRNA biomarkers) was observed for differentiation of AD and FTD, for which partial anatomic overlap between the affected brain regions could be larger than between other NDs. The miRNA pairs most effectively differentiating disease groups from each other and from control are listed in Additional file 7.

In our other studies focused on distinguishing very mild/mild cognitive impairment from control, we observed sex-specific differences in plasma concentrations of some brain-enriched miRNAs in patients and controls (unpublished data). Thus, in this study, we performed the additional stratified analysis for the four pathologies (AD, FTD, PD, and ALS) and control in the combined dataset as follows:

Table 2, section A, shows the performance of effective miRNA classifiers for differentiating each ND from control in all (male and female), male, and female groups. There are no significant differences in AUC and accuracy between these groups, which indicates that these miRNAs behave similarly in controls and male and female patients. However, other tested miRNAs demonstrated sex-associated effects, and miRNA pairs that included these miRNAs more effectively differentiated male or female participants from sex-matched controls. Table 2, section B, presents the most effective miRNA classifiers with additional miRNA biomarker candidates for differentiating AD, FTD, PD, and ALS vs. control separately in males and females. In this case, both AUC and accuracy were significantly higher than effective miRNA classifiers for all (male plus female) participants. In addition, we compared male and female subgroups in each disease group. The data presented in Additional file 9 demonstrate that certain miRNA pairs are capable of differentiating sex-specific subgroups from each other in every disease group with 66–75% accuracy. These data demonstrate sex-specific patterns in levels of circulating brain-enriched miRNAs in both patient and control samples and suggest that sex-specific assays can potentially yield higher accuracy in diagnosing NDs.

A limitation of this study is that all participants were recruited at a single clinical site. Larger, multicenter studies are needed to further evaluate the utility of the approach described herein. A further limitation of the study is that the blood samples were collected from symptomatic patients only. Longitudinal studies are needed to assess the prognostic value of the biomarker candidates in presymptomatic patients.