Prediction of P-tau/Aβ42 in the cerebrospinal fluid with blood microRNAs in Alzheimer’s disease

Four datasets were acquired in this study (Fig. 1, Tables 1, 2, 3, and 4). The data for the pilot study (Dataset 1) were obtained from a Beijing center (n = 44; controls, 21; AD, 23); those for the development of our model (Dataset 2) were collected from centers in the provinces of Shandong, Henan, and Guangxi (n = 406, controls, 216; AD, 190); those for the validation of the model (Dataset 3) were acquired from centers in Guizhou, Hebei, Jilin, and the Inner Mongolia Autonomous Region (n = 304; controls, 153; AD, 151). Those for the application of the model (Dataset 4) were acquired from Beijing center (n = 503; control, 139; AD, 155; amnestic mild cognitive impairment [aMCI], 55; VaD, 51; PDD, 53; bvFTD, 53; DLB, 52). Diagnoses of AD were based on the criteria published by the National Institute on Aging and Alzheimer’s Association (NIA-AA) [1]. Informed by previously published data [10, 11], a cutoff value for the P-tau/Aβ42 ratio of 0.14 was used to differentiate patients with AD from normal controls (Fig. 1a–d). In addition, based on the ATN framework, low CSF Aβ42 is the key “Alzheimer’s pathological change” [17]. We, therefore, used a reported CSF Aβ42 cutoff of 500 pg/ml (Fig. 1c) [18] as another inclusion criterion. Diagnoses of aMCI [19], VaD [20], PDD [21], bvFTD [22], and DLB [23] were based on previously published criteria. In addition, other neurodegenerative diseases may share overlapping clinical manifestations and pathology with AD, which often result in difficulties in clinical diagnosis. To avoid the mixture of other dementias and AD, patients with VaD, PDD, bvFTD, and DLB who had AD-like cutoff values of P-tau/Aβ42 and Aβ42 were excluded (Fig. 1d, Additional file 1: Fig. S1). A previously published paper has showed the very different levels of CSF P-tau and Aβ42 in VaD, PDD, FTD, and DLB [24]. It would reasonable to use P-tau/Aβ42 and Aβ42 as biomarkers for these diseases. The details of diagnostic criteria were included in Additional file 2: materials and methods. Written informed consent was obtained from all participants or their legal guardians. This study was approved by the Institutional Ethics Board of Xuanwu Hospital, Capital Medical University (LYS[2017]004).

Blood samples were collected in the morning after a 12-h fast. Twenty milliliters of whole blood were drawn from each subject and stored in a polypropylene tube containing EDTA. The whole-blood samples were immediately processed at the Beijing center (Xuanwu Hospital). At the other centers, the collected samples were immediately centrifuged at 4200×g for 10 min at room temperature to obtain the plasma, which was then kept at 4 °C and shipped in dry ice to the Beijing central laboratory within 12 h. Total RNA was isolated using the miRNeasy Serum Kit (Qiagen, USA) following the manufacturer’s instructions. For the preparation of the sequencing library, we used 1 μg of total RNA (quantified with Nano Drop 8000 [Thermo Fisher Scientific, USA] and Agilent 2100 bioanalyzer [Agilent, USA]). Total RNA was purified by electrophoretic separation with 15% urea denaturing polyacrylamide gel electrophoresis (PAGE). Small RNA regions corresponding to the 18–30 nt bands in the marker lane (14–30 ssRNA Ladder Marker, TAKARA, Japan) were excised and recovered. The small 18–30 nt RNAs were ligated to adenylated 3′ adapters annealed to unique molecular identifiers (UMI), followed by the ligation of 5′ adapters. The adapter-ligated small RNAs were subsequently transcribed into cDNA by SuperScript II Reverse Transcriptase (Invitrogen, USA). Several rounds of PCR amplification with PCR Primer Cocktail and PCR Mix were then performed to enrich the cDNA fragments. The PCR products were subsequently purified with PAGE gel, and the recycled products were dissolved in EB solution. The double-stranded PCR products were validated on an Agilent Technologies 2100 bioanalyzer. They were then heat-denatured and circularized with the splint oligo sequence. The single-strand circle DNA (ssCir DNA) was formatted as the final library. The library was amplified with phi29 to generate DNA nanoball (DNB), which has more than 300 copies of one molecule. The DNBs were loaded onto the patterned nanoarray, and single-end 50-base reads were generated with combinatorial Probe-Anchor Synthesis (cPAS). The final ligation PCR products were sequenced using the BGISEQ-500 platform (BGI-Shenzhen, China). Samples for quantitative real-time PCR analyses were added with synthetic Caenorhabditis elegans miR cel-miR-39 (Qiagen) after homogenization by QIAzol Lysis reagent, as an external calibration for RNA extraction, reverse transcription, and miRNA amplification [25].

The raw tags—i.e., the raw sequencing data—were processed according to the following procedure: removal of low-quality tags, removal of tags with 5′-primer contaminants, removal of tags without three primers, removal of tags without insertion, removal of tags with poly A, and removal of tags whose lengths were shorter than 18 nt. After filtration, the remaining tags were mapped to the reference genome NCBI GRCh38 and other databases, including miRbase with Bowtie2 [26]. Cmsearch [27] was performed for Rfam mapping. miRDeep2 [28] was used to predict novel miRNAs by exploring secondary structures. The level of miRNA expression was calculated by counting the absolute numbers of molecules using unique molecular identifiers [29]. Differential expression analysis was performed using the DEseq2 [30]; Q value ≤ 0.001 and the absolute value of Log2Ratio ≥ 1 were used as the default thresholds to judge the significance of the differences in expression.

Quantitative real-time PCR (qPCR) analyses were performed to confirm the altered miRNAs in the validation study. The miRNA levels were quantified using the NCode™ VILO™ miRNA qRT-PCR kit (Invitrogen, USA) and normalized to synthetic Caenorhabditis elegans miR-39-3p. All reactions were triplicated in three independent experiments. The 2^-ΔΔCt method was used to calculate miRNA expression [31]. All miRNA primers are listed in Additional file 2: Table S1. To improve experimental precision, triplicates for qPCR were performed. Coefficient of variation (CV) was calculated using standard deviation divided by mean value of a group of replicates. All CV% in the study were lower than 5%, indicating that the data were of high quality (Additional file 2: Table S2).

CSF was collected immediately after the collection of blood samples according to international guidelines [32]. Specifically, 15 mL of CSF were collected from each subject using lumbar puncture while they were positioned in a left lateral position. The participants were monitored for any signs of discomfort for at least 12 h following the lumbar puncture. The CSF samples were centrifuged at 2000×g for 10 min at room temperature and stored in a polypropylene tube at − 80 °C. The levels of Aβ42, total-tau (T-tau), and P-tau (tau phosphorylated at Thr 181) in the CSF were then measured using an enzyme-linked immunosorbent assay (ELISA) kits, all of which are listed in Additional file 2: Table S2.

Statistical analyses were performed using SPSS v.22 and Stata 13.0. Datasets 1, 2, and 3 were analyzed independently. Group differences in categorical data, such as sex, clinical subgroups, and p ε4 allele (ApoE ε4) carrier distributions, were analyzed with the χ² test. Group differences in continuous data, such as the concentrations of biomarkers, were analyzed with Welch’s t test or analyses of variance (ANOVAs). In Dataset 1, the false discovery rate (FDR) correction was performed to select the differential miRNAs. Q values were used to show the analysis results. In Datasets 2 and 3, the correlative analysis of miRNAs and P-tau/Aβ values was performed using a linear regression model. The tolerance, variance inflation factor (VIF), eigenvalue, and condition index were calculated to assess multicollinearity [33]. After the generation of the linear regression model, the predicted values of P-tau/Aβ42 were calculated with miRNA levels. Receiver operating characteristic (ROC) curves were established using the predicted P-tau/Aβ42 ratio. All tests were two-tailed, and the level of significance was set at P < 0.05.

Four datasets were included (Fig. 1). Tables 1, 2, 3, and 4 list the characteristics of the participants. There were no differences in the ages, or ratio of males/females between the AD and control groups in Datasets 1, 2, 3, and 4. The education years, percentages of ApoE ε4, and findings of the Mini-Mental State Examination (MMSE) differed significantly (P < 0.05) between AD patients and controls in Datasets 1, 2, and 3. MMSE scores were also reduced in VaD, PDD, bvFTD, and DLB compared to controls in Dataset 4 (all P < 0.05).

A pilot study was performed in a relatively small sample (Dataset 1). The RNA-sequencing results revealed 860 miRNAs in the blood of patients with AD and the controls. The miRNAs whose read counts were lower than 100 were excluded from subsequent analyses. Differences in fold changes of ≥ 1.2 or ≤ 0.80 between patients with AD and controls were selected as significant miRNAs; we thus identified 29 upregulated and 31 downregulated miRNAs in the AD group (all Q < 0.05 with FDR correction, Fig. 2).

The extended samples (Dataset 2) were included to investigate the miRNAs that were differentially expressed between the patients with AD and controls. All 29 upregulated and 31 downregulated miRNAs in Dataset 1 were confirmed in Dataset 2, supporting the significance of the sequencing data obtained in the pilot study. We then analyzed the potential association between miRNA expression and the ratio of P-tau/Aβ42. The 29 upregulated and 31 downregulated miRNAs were included as independent variables in the linear regression model, and P-tau/Aβ42 as the dependent variable. After adjusting for age, sex, and ApoE ε4, the P-tau/Aβ42 ratio was found to be associated with a panel of seven miRNAs: miR-139-3p, miR-143-3p, miR-146a-5p, miR-485-5p, miR-10a-5p, miR-26b-5p, and miR-451a-5p (Fig. 3). Among these, miR-139-3p, miR-143-3p, miR-146a-5p, and miR-485-5p were decreased in patients with AD, while miR-10a-5p, miR-26b-5p, and miR-451a-5p were increased (P < 0.001; Fig. 3a–g). The linear regression analysis yielded P values of > 0.05 for the variables of age, sex, and ApoE ε4, indicating that the linear regression model was independent of these variables. Performing a linear regression analysis of only the panel of seven miRNAs, we found that the seven miRNAs were significantly associated with the values of the P-tau/Aβ42 ratio (adjusted R² = 0.64, P < 0.001, Fig. 4a). The predictive equation of P-tau/Aβ42 was established and showed in Additional file 2: materials and methods. The equation was applied to Datasets 3 and 4 to predict P-tau/Aβ42 in the following analyses. We performed analyses to estimate the multicollinearity between the seven miRNAs in AD and controls. All tolerances were > 0.1, VIFs were < 10, eigenvalues were > 0, and condition index < 30, indicating that there is no significant multicollinearity between each miRNA. By applying the linear regression model, the predicted values of P-tau/Aβ42 ratio in patients with AD and the controls were calculated (Fig. 4c). Further ROC analyses showed that the P-tau/Aβ42 ratio predicted from the panel of seven miRNAs had a significantly high area under the curve (AUC; 0.90, P < 0.001; Fig. 4e) that far exceeded random chance (AUC of 50%).

An independent dataset (Dataset 3) was used to confirm the above findings. We found that miR-139-3p, miR-143-3p, miR-146a-5p, and miR-485-5p were decreased in patients with AD relative to the controls, while miR-10a-5p, miR-26b-5p, and miR-451a-5p were increased in patients with AD (P < 0.001; Fig. 5a–g). Multicollinearity analyses showed that there is no significant multicollinearity between each miRNA in AD and controls (all tolerances > 0.1, VIFs < 10, eigenvalues > 0, and condition index < 30). The predictive equation generated from Dataset 2 was used to predict P-tau/Aβ42. The predicted P-tau/Aβ42 was highly associated with the actual P-tau/Aβ42 ratio in CSF (adjusted R² = 0.67, P < 0.001, Fig. 4b). By using the predicted P-tau/Aβ42 ratio to distinguish the controls and AD (Fig. 4d), ROC analyses revealed a very high AUC (0.90, P < 0.001; Fig. 4f), which was the same as the AUC calculated from Dataset 2. Taken together, our model generated by the panel of 7 miRNAs in the blood may help to predict the P-tau/Aβ42 ratio in CSF, and diagnose of AD.

To assess the diagnostic capacity of the model when applying to subjects in clinical practice that may include controls, AD, and other neurodegenerative diseases, such as aMCI, VaD, PDD, bvFTD, and DLB, a fourth dataset was used. We obtained similar results to Datasets 1, 2 and 3; levels of miR-139-3p, miR-143-3p, miR-146a-5p, and miR-485-5p were decreased, while miR-10a-5p, miR-26b-5p, and miR-451a-5p were increased in patients with AD (P < 0.001; Fig. 6a–g). For aMCI, although all the seven miRNAs were altered, the alteration of miRNAs was slight in aMCI than in AD or in controls. This is reasonable, since aMCI is considered to be the early stage of AD. Thus, miRNAs may start changing at this stage, with changes in miRNAs in the AD stage becoming more significant as the disease progresses. All the seven miRNAs were not altered in patients diagnosed with VaD, PDD, bvFTD, and DLB (all P > 0.05), suggesting that these miRNAs are AD-specific. The predictive equation generated from Dataset 2 was used to predict P-tau/Aβ42. The predicted P-tau/Aβ42 was highly associated with the actual P-tau/Aβ42 ratio in CSF (adjusted R²= 0.62, P < 0.001, Fig. 7a). The predicted P-tau/Aβ42 ratio in AD patients was robustly increased compared to non-AD (combination of controls, VaD, PDD, bvFTD, and DLB) (P < 0.001, Fig. 7b). Further ROC analysis showed a very high AUC (0.90, P < 0.001, Fig. 7c), indicating that the panel of seven miRNAs is highly effective to identify AD from healthy controls and other neurodegenerative diseases. In addition, the predicted P-tau/Aβ42 ratio was also compared between aMCI and Non-AD (0.76, P < 0.001, Fig. 7d), and AD and aMCI (0.72, P < 0.001, Fig. 7e).