Validation of microRNAs in Cerebrospinal Fluid as Biomarkers for Different Forms of Dementia in a Multicenter Study

Samples

CSF samples (Table 1) of 60 AD, 39 MCI-AD, 37 FTD, 37 DLB cases, and 40 controls were obtained from three different European centers: Radboud University Medical Center, Nijmegen, The Netherlands (AD = 20, DLB = 12, controls = 19); Clinica Neurologica, Perugia, Italy (AD = 20, DLB = 5, FTD = 12, MCI-AD = 39, controls = 21); BIODEM, University of Antwerp, Belgium (AD = 20, DLB = 20, FTD = 25). Patients were assessed by thorough clinical and neuropsychological examination, including the Mini-Mental State Exam (MMSE). Additional information was obtained by routine laboratory investigation of CSF Aβ₄₂, total tau, and phosphorylated tau levels. The diagnosis of AD patients had been established using NINCDS-ADRDA criteria before 2011 [24] and NIA-AA criteria after 2011 [25]. Published consensus criteria were used for the diagnosis of DLB [26] and FTD [27] cases. The MCI-AD group was diagnosed according to Albert criteria [28] and MCI patients had been followed for a maximum of four years to monitor the progression from MCI to AD [29]. The control group consisted of patients that had been subjected to lumbar puncture for various diagnostic reasons such as headache, seizures, or visual impairment. They were diagnosed with either not having a disease (healthy control), with a systemic disease without neurological manifestations, or with a neurological disease (for instance metabolic encephalopathy or transient global amnesia), but did not have any cognitive decline or any neurodegenerative disease. All patients were informed that their data could be used for further scientific purposes and were given the option to object against this use, in which case their data was not included. CSF had been collected by lumbar puncture in sterile polypropylene tubes. We aimed to include only CSF that contained no or only low numbers of cells (< 200 erythrocytes/μl; < 4 leukocytes/μl). These cell number cut points are based on our previous observations [15]. Some samples contained a slightly higher number of leukocytes (maximally 4 leukocytes; Table 1), but since the number of erythrocytes in these samples was very low these samples were included into analysis. One CSF sample (AD) with a very high number of erythrocytes (19,200 erythrocytes/μl) was excluded from analysis. Sample groups were matched for the number of leukocytes (Table 1). All samples from Nijmegen and Perugia were centrifuged for 10 min at 3,000×g, after which the supernatant was transferred to clean polypropylene tubes. According to a formerly used pre-analytical protocol, samples from Antwerp were either centrifuged (AD = 10, DLB = 2, FTD = 1) for 10 min at 4,000×g followed by transfer of supernatant to clean polypropylene tubes, or were immediately frozen and stored without centrifugation in case of non-blood contaminated samples (AD = 10, DLB = 18, FTD = 25) [30]. All samples were stored at –80°C until analysis. For transport, samples were put on dry ice and reached their destination in frozen condition where they were immediately stored again at –80°C until analysis.

RNA isolation and qPCR

RNA isolation, reverse transcription (RT), pre-amplification, and qPCR were performed as previously described by us [15]. For RNA isolation, we randomly divided CSF samples from each patient category over groups of 20 samples that were used for RNA isolation on several days. RT and pre-amplification were each performed in three separate runs on the same day. For qPCR, the samples from each patient category were evenly distributed over five runs per miRNA target. Three additional AD CSF samples that had been analyzed previously were included in each run to correct for inter-plate differences. Since the coefficient of variation (% CV) for the five qPCR plates was very low (% CV < 0.5 for the target miRNAs; % CV < 2.1 for the reference miRNAs) all data were pooled. The following miRNAs and snRNA were analyzed: hsa-miR-27a-3p, hsa-miR-29a, hsa-miR-125b-5p, hsa-miR-146a-5p, hsa-miR-16-5p, hsa-miR-24-3p, and U6 snRNA. Primer sequences of these RNAs can be found at http://appliedbiosystems.com.

Data analysis

For normalization, snRNA U6 and miRNAs miR-16 and miR-24 had been selected based on our previous study [16], where we had observed equal expression levels of these reference miRNAs in control and AD CSF samples. However, in this study differences due to sample origin and centrifugation status were observed, which were especially evident in U6 snRNA levels (see result section). Therefore, miRNA levels were normalized by the geometric mean (GM) of two reference RNAs, miR-16 and miR-24, and the relative expression levels (RELs) were calculated using following formulas: REL = 2^–ΔCt, ΔCt = Ct_miRNA–GM, and GM=Ct mi R - 16 · Ct miR - 24 2. If analyses included samples from only one center, data were normalized by the GM of U6, miR-16 and miR-24: GM=Ct U 6 · Ct miR - 16 · Ct miR - 24 3. Data were analyzed using GraphPad PRISM, version 5 (San Diego, CA, USA) and SPSS version 20 (Chicago, NY, USA). RELs of miRNAs were compared using Mann Whitney U test and Kruskal-Wallis test with Dunn’s post-hoc test for multiple comparison as non-parametric tests or Student’s t test and ANOVA with Bonferroni’s post hoc test as parametric tests. To perform a statistical test with covariates, most data was log-transformed and ANCOVA with Bonferroni’s post-hoc test for multiple comparison was applied. Correlations were determined using Pearson r for normally distributed data or the Spearman test when data were not normally distributed. p-values less than 0.05 were considered significant.

Selecting reference RNAs

SnRNA U6 was detected in all samples, but miR-16 was not detected in two samples and miR-24 was not detected in four samples (AD: n = 2, MCI-AD: n = 2, DLB: n = 2). These samples were therefore excluded from further analysis. Differences due to sample origin and centrifugation status were observed in the reference RNAs. Notably, when comparing only AD samples, Ct-values of miR-16 and miR-24 were increased in the Perugia samples compared to Nijmegen and/or Antwerp, while Ct-values of U6 were similar in the AD samples from all centers (Fig. 1A). The Ct-values of all three reference RNAs were decreased (i.e., higher expression) in non-centrifuged samples compared to centrifuged samples (Fig. 1B). We therefore tested the comparability of the reference RNAs between the groups in the three centers separately, including in the Antwerp samples only non-centrifuged samples. MiRNAs miR-16 and miR-24 were present at similar levels in all groups within one center, except for a small increase of miR-24 Ct-values in DLB patients compared to AD patients in the Nijmegen center (p = 0.026; Supplementary Figure 1). Ct-values of U6 snRNA were similar between groups in Perugia and Antwerp, but were also increased in the Nijmegen DLB group compared to AD (p = 0.009; Supplementary Figure 1). We then tested the correlation of the reference RNAs with each other. Mean Ct-values of miR-16 and miR-24 correlated strongly in all samples (r = 0.76; p < 0.0001) whereas a weaker correlation was observed for miR-24 and U6 (r = 0.43; p < 0.0001) or miR-16 and U6 (r = 0.28; p < 0.0001; Fig. 1C). Because U6 clearly showed the weakest correlation with the other reference RNAs, and showed a stronger difference in DLB samples compared to AD samples in one center, compared to the other reference RNAs, we decided to exclude U6 as a reference RNA and normalized miRNA levels to the geometric mean of miR-16 and miR-24.

Confounding factors

Differences between the groups were observed in gender, age, and sample storage (Table 1, Supplementary Figure 2A-C). The target miRNAs miR-29a and miR-125b were detected in all samples, whereas miR-146a was not detected in one sample of the AD group and miR-27a was not detected in two samples, one of the AD and one of the MCI-AD group. As for the reference RNAs, miR-16 and miR-24, profound differences were detected in Ct-values of the targets caused by sample origin and centrifugation status. All four target miRNAs had lower levels (increasedCt-values) in samples from the Perugia center, similar as the reference RNAs miR-16 and miR-24. This was most evident when AD samples from the different centers were compared (Fig. 2A). After normalization, levels of miR-27a and miR-29a were similar in the AD patients of all centers, whereas differences between AD groups in levels of miR-125b (p = 0.02) and miR-146a (p < 0.0001) remained. Furthermore, levels of miR-29a and miR-146a were higher (decreased Ct-values) in non-centrifuged samples compared to centrifuged samples. This was most evident in comparing centrifuged and non-centrifuged samples from AD patients (Fig. 2B). After normalization, these differences due to the centrifugation protocol were only resolved for miR-146a, but differences between centrifuged and non-centrifuged samples were still evident for miR-29a (p = 0.003) and now also for miR-27a (p = 0.004) and miR-125b (p = 0.003). Of note, storage time correlated significantly with Ct-values of all four miRNA targets and the reference RNAs, but after normalization, only miR-27a (r = 0.16; p = 0.02) and miR-125b (r = 0.25; p = 0.0003) weakly correlated with storage time.

For these reasons, gender, age and sample storage, center of origin and centrifugation status of the samples were included as confounding factors in the ANCOVA to compare miRNA expression between the groups.

miRNAs in early and late stage AD

To confirm earlier findings we first compared AD patients (n = 20) to controls (n = 19) of only the Nijmegen center. In this independent cohort, we could confirm the decrease in expression levels of miR-146a in AD (p = 0.02; Mann-Whitney test), and the similar levels of miR-27a. However, in contrast to our earlier findings where we detected increased levels of miR-29a and miR-125b in AD, we did not find changes in the current cohort.

Next, expression levels of miRNAs were compared between dementia-free controls and AD patients of all three centers. Levels of all four miRNAs were similar in the AD and dementia-free control group (Fig. 3). Levels of miR-125b displayed a trend toward a decreased expression in AD patients (p = 0.06; Mann-Whitney test), but when results were controlled for confounding factors this trend disappeared (p = 0.1; ANCOVA).

To identify potential early AD biomarkers, we further investigated the four target miRNAs for differences in MCI-AD versus controls. Expression levels of miR-27a, miR-29a, and miR-125b were similar (Fig. 3). The levels of miR-146a were increased in MCI-AD patients compared to dementia-free controls (p = 0.03; Mann-Whitney test; Fig. 3), but when confounding factors were included in the analysis no differences were observed (p = 0.3; ANCOVA). To exclude origin and centrifugation protocol as confounding factors, we also compared MCI-AD patients (n = 37) to controls (n = 21) of only the Perugia center. No differences in the four target miRNAs were identified. Finally, we compared MCI-AD patients to AD patients of all centers to identify differences between early and advanced stages of AD. Levels of miR-27a, miR-125b, and miR-146a (p = 0.047; p = 0.001 and p = 0.01, respectively; Mann Whitney test) were increased in MCI-AD patients. However, after correction for the confounding factors, these differences were lost (p = 0.18, p = 0.3, p = 0.29, respectively, ANCOVA).

miRNAs in AD and non-AD dementia

We investigated the four miRNAs for differential expression between late stage AD, FTD, and DLB (Fig. 4). CSF miRNA levels of miR-27a, miR-29a and miR-125b were comparable in these groups. Levels of miR-146a were similar in AD and DLB groups, but were increased in FTD patients compared to AD patients (p = 0.04; Kruskal-Wallis test). However, this difference disappeared when the covariates were included in the group comparisons (p = 0.9; ANCOVA). Furthermore, when specifically looking at the FTD and DLB groups, no significant differences were identified for FTD patients compared to dementia-free controls (data not shown). When comparing DLB patients to dementia-free controls, we found decreased miR-125b levels in the DLB group (p = 0.03; Mann-Whitney test), which changed into a trend toward decreased levels after inclusion of the covariates (p = 0.056; ANCOVA; data not shown). Levels of miR-27a, miR-29a and miR-146a were not different between the control and DLB group.

Also all groups (controls, MCI-AD, AD, FTD, and DLB) were compared to each other. Differences were detected in levels of miR-125b (p = 0.003; Kruskal-Wallis test). However, again after correction for confounding factors, no differences were detected between the groups (p = 0.14; ANCOVA).

Expression of miRNAs in AD versus controls

In this study, we were able to confirm our previous results of miR-27a and miR-146a in an independent cohort of the Nijmegen center. However, when comparing AD patients and controls from three different centers, we could only confirm findings for miR-27a which was again comparable between AD patients and controls. In contrast, the levels of miR-146a, miR-29a, and miR-125b were not different between AD patients and controls. The levels of miR-125b even showed a tendency toward a decrease in AD, which, although not in line with our own previous findings, corresponds to the results of another study in which miR-125b was shown to be decreased in the CSF of AD patients compared to controls [17]. However, this trend was only evident when the covariates had not been included in the analysis. Others have reported increased levels of miR-29a and miR-146a, and decreased levels of miR-27a in the CSF of AD patients versus controls [14 , 18], which is also partly different from our current and previousresults.

Expression of miRNAs in MCI-AD versus controls

We did not identify differences in levels of miR-27a, miR-29a miR-125b, and miR-146a between MCI-AD patients and controls. Also, when comparing MCI-AD patients to the controls in only one center, thus excluding origin and centrifugation protocol as confounding factors, no differences were detected. Expression of miRNAs has been investigated in MCI patients only in a small number of studies [31 –33], and even in fewer studies that make use of CSF [34, 35]. MiR-125b has been investigated before in amnestic MCI patients compared to AD patients and controls, but this study was performed in plasma [32]. Notably, in this study several promising miRNAs were identified and the ratios of miRNA-pairs (in particular the miR-132 and miR-134 families paired with miR-491-5p and miR-370, respectively) demonstrated high sensitivities (79–89%) and specificities (83–100%) in differentiating MCI from age-matched controls. Similar percentages were obtained when these ratios were tested in a larger follow-up study with MCI and control subjects [31]. It might be interesting to also study these miRNA-pairs longitudinally in MCI patients, to study their possible predictive value for developing AD.

Expression of miRNAs in AD versus non-AD dementia

As with MCI, the literature on miRNAs in FTD and DLB is scarce, and the present study is to our knowledge the only study in which miRNA expression in FTD and DLB has been investigated in CSF. Using this particular panel of miRNAs, we did not identify significant differences among the groups. In other studies, miRNAs that regulate levels of proteins that are involved in FTD pathology have been identified. For instance, miR-29b regulates progranulin which is a glycoprotein encoded by the GRN gene and mutations of this gene are present in a subset of familial FTD cases [36]. In an FTD mouse model, a decrease of miR-124 levels was associated with deregulation of AMPA receptor expression, which possibly is involved in the onset of behavioral deficits in FTD [37]. Furthermore, in a recent pilot study, deep sequencing of small RNAs in the cerebral neocortex of AD, DLB, FTD, hippocampal sclerosis, and controls identified miR-132-3p and miR-100 as differentially expressed miRNAs [21]. It would be interesting to test these miRNAs in CSF as possible biomarkers for FTD or DLB.

Variations across studies and possible confounding factors

Our results in AD patients and controls show once more how challenging it is to reproduce the findings of previous miRNA studies, especially when studying miRNA levels in body fluids such as CSF. Also in cancer studies, conflicting results have been reported about the levels of circulating miRNAs relating to the same type of tumor (reviewed in [1]). Such variations between studies might lie in technical differences such as different methods for RNA isolation or quantification, but even when the same techniques are used within the same laboratory, inter-study differences cannot be ruled out, as we showed. Previously, we also showed that blood contamination of CSF samples has a major effect on levels of miR-146a, miR-27a, and miR-29a [15, 16]. Based on these previous observations, one of the major criteria in selecting CSF samples for this study had been the number of blood cells in CSF, which had to be kept as low as possible to exclude any influence (here maximal number was 50 erythrocytes/μl and 4 leukocytes/μl). Many other CSF miRNA studies, however, do not provide information on blood cell count and whether groups are matched according to the number of cells. Results of miRNA levels might therefore be biased by samples that were possibly contaminated with blood. Other pre-analytical confounders such as differences in RNA quality of the sample and CSF processing may play important roles. The quality of RNA is usually determined by quantifying ribosomal RNA (rRNA) integrity using, for instance, a bioanalyzer [38]. However, in our experience RNA concentration in CSF is too low to quantify rRNA. Furthermore, it would be questionable whether rRNA integrity can be interchangeably used for miRNA integrity, since miRNA stability can be different from rRNA stability as has been observed in tissue samples [39]. We therefore used CSF volume as input measure, as has also been done previously in other miRNA studies using CSF [9 , 40]. CSF pre-analytical protocols such as time of processing and temperature, centrifugation protocol and storage might also influence the outcome of miRNA studies. As we observed in this multicenter study, centrifugation of CSF and subsequent removal of the spin-down-product can lead to differences in miRNA levels when compared with non-centrifuged samples. The increased miRNA levels detected in non-centrifuged versus centrifuged samples might result from unremoved cells that release their miRNA content when subjected to RNA extraction procedures. Furthermore, by only comparing AD samples, we identified differences between the three centers, which might be ascribed to differences in storage time or other unidentified pre-analytical factors (Supplementary Table 1).

Most of these factors might be controlled for by normalization using endogenous reference RNAs. However, an important concern with miRNA studiesremains the lack of an established endogenous miRNA control. Also commonly used endogenous miRNA controls such as snRNAs may be differentially expressed in various conditions [15]. Recently, it has been suggested to use more than one endogenous RNA as reference and calculate the mean or geometric mean for normalization purposes [41], a strategy which we also applied in our current and previous study [16]. In our previous study we chose three endogenous controls (U6, miR-16, and miR-24)that had similar levels in the investigated groups (AD and dementia-free controls). In the current study, miR-16 and miR-24 were comparable in almost all groups when analyzed per center, and also U6 was nearly similar. However, miR-16 and miR-24 correlated better with each other than with U6, and further showed the same patterns in differences due to origin as was observed for the target miRNAs. We therefore chose to exclude U6 as a reference, and calculated the geometric mean of miR-16 and miR-24 for normalization. However, even after normalization, miRNA levels were influenced by confounding factors such as storage time or centrifugation protocol. Therefore, all identified confounders were included in the statistical analyses.

All these factors might explain why miRNA studies are often not in agreement with each other, and might also be the reason why we could not repeat previous results. However, when only comparing AD patients and controls of Nijmegen, which excludes most of the pre-analytical factors, we could only repeat previous findings of differential expression of miR-146a, but not of miR-29a and miR-125b. Possibly, differences observed in miR-29a and miR-125b levels were too small to be reliably identified, particularly when testing small study populations.

Nonetheless, we did observe a downregulation of miR-146a in AD patients compared to controls in two independent study cohorts, measured under comparable conditions. This miRNA therefore deserves further attention.

Conclusion and future directions

In this study, we have measured several miRNAs with biomarker potential in a large cohort of patients, enrolled in three European centers. The use of large cohorts is a desirable feature to be considered for future miRNA studies in CSF samples. The use of samples from different centers to enlarge groups and diversity within groups introduced several confounding factors and possibly such a great bias that previously obtained results could not be replicated. Considering all the confounding factors found in this study, there are several recommendations for the quantification of miRNAs in CSF to be applied in future studies (Box 1). We suggest that samples should be well-controlled for the number of blood cells which would be preferably as low as possible, for instance < 50 erythrocytes/μl and < 4 leukocytes/μl, to exclude any effects of blood contamination. Furthermore, pre-analytical protocols should be standardized, as is also suggested when measuring other CSF biomarkers [30, 42]. In particular, samples should be controlled for storage time and centrifugation status. For normalization purposes, we suggest to use the geometric mean of several miRNAs or other small RNAs that are present at similar levels between groups. This may exclude any bias which could arise by including only one RNA as reference. All these issues may also account for differences in miRNA measurements in other biofluids than CSF, such as blood, urine, or saliva or in studies using postmortem tissue. For the latter, it has also been shown that a short postmortem interval of 1 hour maximally may be required for robust quantification of specific brain-enriched miRNAs [43].

We identified once again decreased levels of miR-146a in AD samples versus controls in the Nijmegen center. This strengthens the possible biomarker value for miR-146a. In the future, this miRNA should be investigated in larger studies under well-controlled pre-analytical conditions. However, miRNAs regulate many processes and a single miRNA usually has several protein targets, and might be differentially regulated in several diseases. An individual miRNA might therefore not be specific enough as biomarker for a disease, and also the size of effect of one miRNA to differentiate two groups might be too small. Therefore, we suggest focusing future research on a panel of several differentially expressed miRNAs, as other studies for different cancers already did, to increase specificity and power [9, 44]. Particularly, in differentiating different stages of AD versus controls, or AD from non-AD dementia types, a panel of miRNAs, possibly in combination with protein biomarkers, might be the most suitable approach. Finally, many circulating miRNAs are enclosed into exosomes, which might be specifically secreted by neuronal cells and be present in CSF, but it is not certain if these are lost during precipitation or centrifugation steps of the RNA extraction. Methods of extracting and measuring such exosomal miRNAs in CSF are just beginning to emerge [45], and this might be a new approach to test miRNAs as biomarkers.