Specific serum and CSF microRNA profiles distinguish sporadic behavioural variant of frontotemporal dementia compared with Alzheimer patients and cognitively healthy controls

Abstract

Information on circulating miRNAs in frontotemporal lobar degeneration is very limited and conflicting results have complicated an interpretation in Alzheimer's disease thus far. In the present study we I) collected samples from multiple clinical centers across Germany, II) defined 3 homogenous patient groups with high sample sizes (bvFTD n = 48, AD n = 48 and cognitively healthy controls n = 44), III) compared expression levels in both CSF and serum samples and IV) detected a limited set of miRNAs by using a MIQE compliant protocol based on SYBR-green miRCURY assays that have proven reliable to generate reproducible results. We included several quality controls that identified and reduced technical variation to increase the reliability of our data. We showed that the expression levels of circulating miRNAs measured in CSF did not correlate with levels in serum. Using cluster analysis we found expression pattern in serum that, in part, reflects the genomic organization and affiliation to a specific miRNA family and that were specifically altered in bvFTD, AD, and control groups. Applying factor analysis we identified a 3-factor model characterized by a miRNA signature that explained 80% of the variance classifying healthy controls with 97%, bvFTD with 77% and AD with 72% accuracy. MANOVA confirmed signals like miR-320a and miR-26b-5p at BH corrected significance that contributed most to discriminate bvFTD cases with 96% sensitivity and 90% specificity and AD cases with 89% sensitivity and specificity compared to healthy controls, respectively. Correlation analysis revealed that miRNAs from the 3-factor model also correlated with levels of protein biomarker amyloid-beta1-42 and phosphorylated neurofilament heavy chain, indicating their potential role in the monitoring of progressive neuronal degeneration. Our data show that miRNAs can be reproducibly measured in serum and CSF without pre-amplification and that serum includes higher expressed signals that demonstrate an overall better ability to classify bvFTD, AD and healthy controls compared to signals detected in CSF.

Fig 1. Unsupervised hierarchal clustering of miRNA levels measured in serum of healthy control, bvFTD and AD samples.

Using the DiffCorr package, the genes were grouped according to their expression patterns in each subtypes (cognitively healthy controls, bvFTD and AD) using the cluster.molecule function. We used (1 − correlation coefficient) as a distance measure (the cutoff value was a coefficient of 0.6) based on the cutree function. We then visualized the module network using the get.eigen.molecule and get.eigen.molecule.graph functions. MiRNAs that share similar seed sequences (miRNA families) are coloured. MiRNAs that are co-transcribed as a polycistronic unit (http://www.mirbase.org/, < 10 kb) are listed in S5 Table. AD = Alzheimer’s disease, bvFTD = behavioural variant frontotemporal dementia, M = Module.

Fig 2. Factor and discriminant analysis of miRNA signals detected in serum.

We have used factor and discriminant analysis to identify the variables that can best characterize and classify the bvFTD, AD and healthy control samples in our serum cohort. The following results are shown: (A) factor model of n = 29 serum miRNAs that load on the Factors 1–3 with factor loadings ≥ |0.5|, (B) observation plot showing coordinates of the observations resulting from the two-factor model (Factors 1–2), and result of the discriminant analysis using a re-factoring 2-factor model based on the n = 29 miRNAs from the original 3-factor model (Fig 2A) using (C) a training (n = 78) and validation (n = 53) set and (D) all cases (n = 131).

Fig 3. Differentially expressed miRNAs in bvFTD, AD and cognitively healthy control cases detected in CSF and serum.

Expression levels of n = 96 circulating miRNAs were measured in CSF (n = 140) and serum (n = 131) samples from bvFTD (n = 48/48) and AD patients (n = 48/47) as well as healthy controls (n = 44/38) and compared using MANOVA and ROC curves. Displayed are signals with significantly different expression levels after multiple comparisons detected in (A) CSF: group comparisons of AD, bvFTD and healthy controls with (B-C) corresponding ROC curves and (D) serum: group comparisons of AD, bvFTD and healthy controls (up- and downregulated miRNAs) with (E-F) corresponding ROC curves, (G) serum: group comparisons of AD, bvFTD and healthy controls (only upregulated miRNAs) with (H-I) corresponding ROC curves and (J) serum: group comparisons of AD, bvFTD and healthy controls (only downregulated miRNAs) with (K-L) corresponding ROC curves. Expression ratio: ddCt = mean dCt_{AD or bvFTD} − mean dCt_HC. Dotted lines indicate ddCt cut-off of |0.58|. Error bars indicate mean ± SEM. BH = Benjamini-Hochberg.

Fig 4. Correlations of miRNA expression levels in serum with CSF protein biomarkers.

Depicted are normalized expression levels dCt = Ct (Ct_{mean RefmiR} − Ct_miR) of (A) miRNAs from the original 3-factor model that correlated with Factor 1 vs CSF levels of amylod-beta_1-42 in the control group and (B) miRNAs from the original 3-factor model that correlated with Factor 2 vs CSF levels of pNfH in the bvFTD group. pNfH = phosphorylated neurofilament heavy chain.