Performance of Validated MicroRNA Biomarkers for Alzheimer’s Disease in Mild Cognitive Impairment

Participants

All procedures were approved by the Institutional Review Board of the University of California San Diego (UCSD IRB 80012). All participants provided written informed consent and underwent detailed evaluations consisting of medical history, physical and neurological examinations, laboratory tests, and neuropsychological assessments. Cognitively unimpaired NC participants were volunteers in good health with no symptoms of cognitive impairment or neurological disease, and normal performance on a detailed battery of neuropsychological tests. The age distribution and sex ratio of NC samples to that of the MCI and AD diagnostic groups were matched to the best of our ability. Participant samples were banked at the UCSD Shiley-Marcos Alzheimer’s Disease Research Center (UCSD ADRC).

CSF collection

The USCD ADRC uses a standardized CSF collection protocol corresponding to that of other AD centers engaged in biomarker research [14]. Lumbar punctures were done in the morning under fasting conditions, in the lateral decubitus position with a 24-gauge Sprotte spinal needle. The first 2 mL of CSF collected was used for clinical tests, then 10– 20 mL of CSF was collected, gently mixed, and transferred to polypropylene tubes in 500μL aliquots. The tubes included a subject number but no other identifying information. The CSF aliquots were flash frozen on dry ice and stored at – 80°C.

APOE ɛ4 genotyping

APOE ɛ4 genotypes were determined by polymerase chain reaction (PCR) amplification and restricted fragment length polymorphism [15], as previously described [16]. Venous blood was drawn from participants and genomic DNA was extracted using the QIAamp DNA Blood Mini Kit (Qiagen, Valencia, CA, USA) followed by PCR amplification. The APOE ɛ4 gene sequences were amplified using forward (5^′-ACGCGGGCACGGCTGTCCAAGGA-3^′) and reverse (5^′-GCGGGCCCCGGCCTGGTACAC-3^′) primers. The amplification products were digested with HhaI (restriction enzyme site GCG∧C) and subjected to electrophoresis on polyacrylamide gels. After electrophoresis, the gels were stained with ethidium bromide and the digested fragments were visualized by ultraviolet illumination. A unique combination of HhaI fragment sizes enabled unambiguous typing of all homozygotic and heterozygotic combinations: HhaI cleaves at GCGC encoding 112arg (E4) and 158arg (E3, E4), but does not cut at GTGC encoding 112cys (E2, E3) and 158cys (E2).

CSF Aβ₄₂ and T-Tau measures

Measurements of CSF Aβ₄₂ and T-Tau levels were performed at the UCSD ADRC using enzyme-linked immunosorbent assays (ELISAs) as previously reported [17]. CSF Aβ₄₂ was measured using the Euroimmun ELISA kit EQ 6521-9601-L (ADx Neurosciences, Ghent, Belgium). CSF T-Tau was measured using the ELISA kit EQ 6531-9601-L (ADx Neurosciences).

RNA isolation and qPCR

We instituted safeguards to improve quality control based on findings and outcomes from our initial AD miRNA discovery studies. We included multiple positive controls (CSF miRNAs verified in our previous studies [12, 13] that were unchanged between controls and AD) and negative controls (miRNAs not detected in CSF in our previous studies). We also imposed strict uniformity over manufacturing variables: all kit and reagents were matched by lot, with no exceptions, to minimize variation from sources unrelated to the differentiation of diagnostic groups.

We generated Custom TaqMan® Array MicroRNA Cards which included probes for 17 AD miRNA biomarkers, as well as 4 non-changing CSF miRNAs (positive controls), 2 miRNAs not detected in CSF (negative controls), and U6 snRNA (Supplementary Table 1). Of the 17 biomarker miRNAs in the arrays, 14 appeared as strong biomarker candidates in our validation study [13], and an additional 3 (miR-1291, miR-142-3p, and miR-519b-3p) were strong biomarker candidates in our initial discovery study [12], but did not perform well in the validation study. Each array card was designed to assay 4 individual CSF samples, with 4 technical replicates for all RNA probes on the card. We analyzed miRNA expression in 102 new CSF samples obtained from the UCSD ADRC, and 63 repeated CSF samples from our previous validation studies [13] for reliability testing. As a further reliability check, we obtained duplicate aliquots for 5 CSF samples (4 NC, 1 MCI) and ran each of the duplicates on a different array card. The run order of the samples was arranged so that every single array card contained at least one NC sample and, whenever possible, also included at least one MCI sample, AD sample, or both. This was done to avoid confounding technical card batch differences with sample diagnosis differences, and only 10% of cards failed to meet this arrangement (due to the larger number of NC samples relative to the other two diagnoses). Within diagnosis type, the plating order of samples was fully randomized (under the constraint that we would only accept a randomization where duplicates fell on separate cards). On every card we also included a sample from a homogeneous reservoir of a pooled CSF reference standard obtained from NCs at the Oregon Health & Science University Layton Aging and Alzheimer’s Disease Center in order to monitor the stability of the process over time.

Total RNA was extracted from 500μL of CSF using the mirVana™ PARIS™ Kit (AM1556, Thermo Fisher Scientific (TFS), Waltham, MA) then concentrated using the RNA Clean and Concentrator-5 Kit (R1013, Zymo Research, Irvine CA), as previously described [12, 13]. We previously reported that 500μL of CSF yields a very low amount of RNA (∼1 ng) such that it precludes measuring the concentration [12]. Thus, we instituted a protocol to isolate total RNA from an absolute (500μL) amount of CSF, then concentrate the RNA to the smallest amount needed for the arrays in order to maximize the amount of RNA going into the PCR. We then controlled for variance in the normalization step. Thus, we used 4.1μL of total RNA from each sample for reverse transcription (RT) using a custom primer pool generated by TFS to specifically reverse transcribe 23 miRNAs (17 AD miRNA biomarkers, 4 non-changing positive control CSF miRNAs, and 2 negative control miRNAs not detected in CSF), and U6. Based on the recommended protocol by TFS for Custom TaqMan® Array MicroRNA Cards (publication part #4478705), to include a pre-amplification step for RNA inputs <350 ng of total RNA, we then used 7.5μL of the resulting cDNA for pre-amplification using a custom Pre-Amp primer pool generated by TFS to specifically pre-amplify the 23 miRNAs and U6. The pre-amplification products were diluted 1 : 4 in RNase/DNase-free water, then 9μL of the diluted samples were mixed with TaqMan^® master mix (TFS) and RNase/DNase-free water to a 225μL final volume, then loaded onto the Custom TaqMan® Array MicroRNA Cards (TFS) containing probes for the 23 miRNAs and U6. We also included parallel reactions that contained only water (no RNA)±RT enzyme, or RNA with no RT enzyme, as controls for spurious amplifications in these experiments (Supplementary Table 1). The quantitative PCR (qPCR) amplifications and acquisition of the resulting data were carried out on a QuantStudio™ 12K Flex Real-Time PCR System (TFS). All NC, MCI, and AD qPCR data files were imported into ExpressionSuite Software v.1.1 (TFS) and processed together to generate automatic baseline and threshold values used to calculate quantification cycle (Cq) values. The Cq value for each well was reported along with the amplification quality score (AmpScore) and Cq confidence (CqConf) score, which serve as metrics for the quality of each amplification.

Preprocessing of Cq values

We applied Cq validity checks to each well individually, and implemented acceptability criteria across the 4 replicate wells per probe per sample that had to be met in order to consider the miRNA measurement valid. We required AmpScore ≥1.0 and CqConf ≥0.8, and based on the rapid decline (with increasing Cq value) that we observed in the fraction of wells meeting these criteria for Cq >34, we set the Cq censoring threshold at 34. If a well failed to meet either the AmpScore or CqConf criteria, we set the Cq value to missing if the assigned Cq was <34; if Cq was ≥34 the well was considered censored instead. After this step, if at least 2 of the 4 replicates for a probe for a sample had a non-missing Cq value, then we considered the measurement valid; otherwise the entire measurement was removed (just 0.16% of measurements failed in this way). The rationale for this approach was that if at least half of the replicates for a sample for a probe indicated some measurable expression, it would be treated as having been measured at a value rather than censored at a lower limit. However, usually either all four replicate wells had valid Cq values, or none of them did. Final Cq values were taken as the median of the miRNA values in the quadruplicate wells; miRNAs with median value ≥34 were labeled censored for the sample, as they were considered to have expression too low for reliable detection in the assay. Censored values were included in the final statistical analysis.

Preprocessing of miRNAs

Our assay included probes for negative controls, positive controls, U6 snRNA, and the AD miRNA biomarkers. We required the negative control miRNAs to show low expression profiles (i.e., uniformly AmpScore <1.0, CqConf <0.8, and Cq ≥34 or missing for every well). Due to their importance for normalization, we applied greater stringency to the positive control miRNAs. Thus, for the positive controls and U6 snRNA we required robust expression in at least 90% of samples, with AmpScore ≥1.0 and CqConf ≥0.8, but with a more stringent average Cq <32 across the cohort than for the biomarker miRNAs. As U6 was used for batch correction of cards and positive controls were used for normalization (see Batch correction and normalization below), we required that every sample have a valid U6 measurement and valid measurements for at least 2 of the 4 non-changing positive control miRNAs. We also verified that the maximum of pairwise mean Cq contrasts between diagnostic groups was no larger than 1 Cq in magnitude for any positive control miRNA (in nearly all cases the maximum contrast magnitude was 0.5 Cq or less). By these rules, all positive controls were retained and no experimental samples were excluded. For biomarker miRNAs we required robust expression in at least 20% of samples with average Cq <34, with AmpScore ≥1.0 and CqConf ≥0.8 for all expressed wells.

Array card batch correction, normalization, and transformation of expression scale

U6 snRNA has shown weaknesses as a normalizer in previous investigations by us and others [18]. However, U6 snRNA showed excellent stability of expression across cards when averaged over all samples on the card and for that reason was suitable as a monitor for variation at the card level. Thus, median Cq values were corrected for array card batch variation by aligning mean values of U6 snRNA across cards using a 3-level random-effects mixed model [19]: sample nested in card nested in RNA isolation batch. We generated best linear unbiased predictions for the batch and card effects from the fitted mixed model and subtracted these from all Cq values in the corresponding batches and cards. Following batch correction, Cq values were normalized relative to a linear combination of our verified “non-changing” positive controls (miR-30e-3p, miR-574-3p, miR-638, and miR-92a) using a crossed random-effects mixed model (miRNAs crossed with samples) for the Cq values of all positive control miRNAs. (Supplementary Table 2 includes the Cq distributional characteristics of the positive control miRNAs, and evidence that the per-card mean of U6, and the distributions of each of the other positive controls, did not differ by diagnosis group.) For normalization we calculated a best linear unbiased prediction for each sample’s offset from the overall mean (across all samples) of positive controls using fitted values from the mixed model, then subtracted this offset from the measured Cq values for the sample.

As in our discovery and validation studies [12, 13], we transformed the normalized Cq values onto an “expression” (i.e., positively oriented) scale so that higher values indicated relatively greater quantities of miRNA expression. (Expression = Cqnorm _max – Cqnorm, where Cqnorm _max was the largest non-censored normalized Cq value in the data, rounded up to the nearest integer.) Censored Cq values were assigned a value of zero on this expression scale and included in the final analysis. All qPCR data and donor-specific metadata is being prepared for submission into the NCBI Gene Expression Omnibus (GEO) repository.

Exclusion of aberrant samples

We initially examined CSF miRNA expression in 165 samples (49 AD, 38 MCI, and 78 NC). However, after normalization, 32 CSF samples showed normalization offsets that were extremely large in magnitude (∼7 Cq), with normalizer values from 3- to 10-fold more extreme than the standard deviation of typical samples. These 32 samples had unusually extreme Cq values across all miRNAs in the array, a finding that was replicated when the assays were rerun using stored RT samples. Investigating further, we identified 14 of the 32 CSF samples as technical failures in the RT that resulted in extremely low miRNA expression. However, no explanation could be found for the remaining 18 CSF samples, most of which showed extremely high miRNA expression. There was also no apparent relationship between an aberrant expression profile and any patient characteristic or design factor, including collection site, RNA isolation batch or date, array card lot or run date, participant diagnosis, sex, age, or any AD-related characteristic such as Aβ₄₂ or T-Tau levels in the CSF. (The minimum p-value across all factors tested for association with the expression profile was >0.2, using a variety of parametric and nonparametric statistical tests.) Thus, we excluded the 14 technical failures and the 18 aberrant samples from further analysis to avoid the risk of bias in the results. Due to the lack of association with any patient factors, these exclusions had no meaningful effect on the distributions of those factors, and had only negligible effect on the Cq values for the remaining 133 samples when these were recalculated in ExpressionSuite; the concordance between calculated Cq values before and after the exclusion of the 32 samples was 0.98 (Lin’s concordance coefficient), showing nearly perfect linear correlation and zero bias. Thus, after exclusion of these samples, 133 unique samples (37 AD, 31 MCI, 65 NC) and 3 duplicates (1 MCI, 2 NC) remained in the analytic cohort.

Statistical analysis

To determine whether miRNA information correlated with diagnostic progression and group ordering (NC to MCI to AD), we first confirmed that the diagnostic groups were different according to established measures of disease severity by visually comparing group distributions of Mini-Mental State Examination (MMSE) scores and Aβ₄₂:T-Tau levels, and tested for natural ordering using ordered logistic regression. Next, we used several visual and analytic methods to assess whether the miRNA expression profiles were naturally ordered across diagnosis groups. We initially plotted the Cq distributions of the miRNAs for each group side by side and looked for trending of the median Cq values. Outcomes were either a linear trend (NC to MCI to AD), a nonlinear trend (2 of 3 groups different from the remaining group), or no trend (an ambiguous pattern). We then calculated approximate Bayesian posterior probabilities (based on approximate Bayes factors from harmonic-mean combining of p-values) that the AD and MCI were both significantly different from the NC, and that the 3 groups were strictly ordered.

We then used receiver operating characteristic (ROC) curve analysis [20] to assess the predictive power of the linear trending miRNAs as a group for discriminating AD and MCI from NC. We report the area under the ROC curve (AUC), a measure of classification performance, for the miRNA classifiers with and without APOE ɛ4 status, with and without Aβ₄₂:T-Tau levels, and with and without the combination of Aβ₄₂:T-Tau levels and APOE ɛ4 status to differentiate performance under alternative sets of information. In calculating the ROC curves, we used two different methods to build classifiers: logistic discriminant analysis (similar to logistic regression, but with more than one group) and a smooth k-nearest-neighbor classifier (k = 9) based on Euclidean distance. We then averaged the predicted probabilities using the harmonic mean (the appropriate average for values between zero and one). The ROC curves were smoothed for easier visualization of the differences. (Average spacing between curves and AUC were both well preserved under the smoothing.) We assessed the strength of the canonical (i.e., best linear) correlations between the multivariate set of trending miRNA expression values and diagnosis rank, and with Aβ₄₂:T-Tau levels, to further validate the panel of trending miRNAs, not just for distinguishing AD and MCI from NC, but also for correctly ordering the groups with respect to disease severity.

MiRNA target prediction

We used TargetScan 7.2 [21] and miRDB [22, 23] to predict target mRNAs of the top five trending miRNAs (miR-142-3p, miR-146a-5p, miR-146b-5p, miR-365a-3p, miR-193a-5p). TargetScan and miRDB were chosen as they are both widely used and frequently updated. As subsequent pathway analysis is most effective with a limited gene set, predicted targets were excluded if they had a Cumulative Weighted Context Score above – 0.3 in TargetScan or a target score of below 60 in miRDB. The use of a union between the two target prediction algorithm outputs, as opposed to an intersection where only targets predicted by both algorithms are included in the analysis, may increase the sensitivity of the predicted targets and the likelihood of predicting novel targets [24, 25]. Thus to determine whether a union, an intersection, or an individual target list should be used in subsequent pathway analysis for the trending miRNAs, we calculated sensitivity, specificity, and precision values using validated miRNA-target pairs from miRTarBase, an experimentally validated miRNA-target interactions database [26]. The miRTarBase values ranged from 0– 1, with high-quality results indicated as closer to 1 [24, 25]. The union of TargetScan and miRDB showed the highest values of sensitivity (0.35), specificity (0.833), and precision (0.98). Thus, pathway analysis was performed on the union set of the unique targets using QIAGEN Ingenuity Pathway Analysis (IPA). For this pathway analysis we excluded cancer related tissue and cell lines to avoid the knowledge bias towards cancer in IPA.

Participant characteristics

Participant characteristics for the 133 total participants in this study (37 AD, 31 MCI, and 65 NC) are shown in Table 1. In the final set of 133 participants the sex ratio was close to a 1 : 1 male to female ratio overall, but varied somewhat from group to group (NC ∼2 : 3 but AD and MCI both ∼3 : 2). Participants in all 3 groups were well matched for age (mean 73.2±7.1, range 53– 89), but the AD group had more variance than the other groups and contained some sex bias (with females tending to be younger than males in that group). NC participants were in good health with a mean MMSE [27] score of 29.1±1.2, Clinical Dementia Rating (CDR) scores of 0, no evidence or personal or family history of cognitive or functional decline, and also performed normally on a neuropsychological test battery. AD patients were diagnosed with probable AD according to ADRDA-NINDS criteria [6, 28], with a mean MMSE score of 21.6±4.5 and CDR scores of 1– 2. MCI patients were in between, with mean MMSE score of 26.9±1.5 and CDR scores of 0– 0.5. CSF Aβ₄₂ and T-Tau measurements were as expected, with NC having high Aβ₄₂ and low T-Tau, AD having low Aβ₄₂ and high T-Tau, and MCI falling somewhere in the middle (medium values of both Aβ₄₂ and T-Tau). APOE ɛ4 genotyping was available for 131 of the 133 participants in the study. Within the NC group, 66.2% of participants had no APOE ɛ4 allele, 27.7% had 1 allele, and 6.2% had 2 alleles. In the AD group 50.0% had no APOE ɛ4 alleles, 33.3% had 1 allele, and 16.7% had 2 alleles. Thus as expected, APOE ɛ4 genotype was over-represented in AD [29]. Unsurprisingly, the MCI group was more similar to AD in APOE ɛ4 allele frequencies: 40.0% with no alleles, 46.7% with 1, and 13.3% with 2. Figure 1 shows the distributions of MMSE scores (A) and Aβ₄₂:T-Tau measures (B) for each of the 3 diagnostic groups.

OPEN IN VIEWER

Fig. 1

Distributions of MMSE and Aβ₄₂:T-Tau ratio with AD progression. The diagnosis groups trend as expected for measures of disease severity: average MMSE (A) was decreased modestly in MCI and dramatically in AD, and average Aβ₄₂:T-Tau ratio (B) was much lower for MCI and AD than for NC. For both, the MCI profile was intermediate to the NC and AD profiles. Individual participants denoted by points and dotted lines represent group medians. The trend of medians was very apparent for both measures.

Verification of biomarker relevance in the cohort

As a global check on the relevance of the panel of miRNAs for the cohort in the present study (which included both new NC, MCI, and AD samples, as well as repeated NC and AD samples), we verified that the miRNA profiles (taken across the entire panel) for MCI and AD samples consistently differed from those of typical NC samples. Thus, the miRNA profile as a whole is a marker of AD-like disease in our cohort. Details of this analysis can be found in the Supplementary Results.

Quantitative and statistical measures for the individual miRNAs

Next, we report the quantitative and statistical measures of 15 of the 17 miRNAs examined, along with their trend classifications, in Table 2. The 15 miRNAs included all 14 of the validated biomarkers from the previous study, plus 1 of the additional miRNAs (miR-142-3p). The 2 excluded miRNAs were miR-1291, which failed to show expression in most samples (appearing robustly in just 4% of the cohort), and miR-519b-3p, which exhibited universal failure in all samples. Thirteen of the miRNAs were well represented across the samples at close to 80% attestation or better in all 3 diagnostic groups. However, miR-15b-5p and miR-331-3p were poorly attested at <50% each in every group, with the lowest attestation in AD. Inspection of the trend of median fold change across groups showed that, overall, many of the miRNAs had a pattern of decreased expression with increasing disease.

Expression trends of the individual miRNAs across diagnoses

We show the trend plots for the 15 viable miRNAs in Fig. 2. Five of the AD markers (miR-142-3p, miR-146b-5p, miR-146a-5p, miR-365a-3p, and miR193a-5p) showed a linear trend of median expression across the ordered diagnoses, where miRNA expression decreased as disease severity increased. Four of the remaining markers showed nonlinear trends: miR-19b-3p had higher expression in NC that decreased to similar levels in MCI and AD, whereas miR-484 had lower expression in NC that increased to similar levels in MCI and AD. Both miR-140-5p and miR-331-3p showed similar levels in NC and MCI that decreased in AD. The other 6 markers were classified as having no trend, as their expression in MCI tended to be either higher or lower than in both AD and NC. For each of the trending miRNAs we calculated a Bayesian posterior probability of the hypothesis that both MCI and AD differ from NC and are naturally ordered with respect to average levels of the miRNA. The probabilities were generally close to 2/3 (i.e., approximately 2 : 1 odds in favor).

OPEN IN VIEWER

Fig. 2

Trend plots of 15 validated miRNAs. All normalized Cq values for samples and median profile across groups are shown for each of the 15 miRNAs that remained viable for this study. The connected diamonds represent medians for 4 kinds of apparent profiles: linear (“TREND”), two kinds of nonlinear (“AD + MCI versus NC” and “AD versus MCI + NC”), and an ambiguous profile showing a lack of natural diagnosis ordering (“NO TREND”). Multivariate correlation of the trending miRNAs with diagnosis order was – 0.36 (p = 0.003).

Classification performance of trending miRNA biomarkers versus APOE ɛ4 and Aβ₄₂:T-Tau

For detailed evaluation of classification performance (AUC) we used only the 5 trending miRNAs. The nonlinear patterns were interesting, but not directly relevant to our primary goal, which was the identification of CSF miRNAs as markers of disease progression from MCI to AD. We evaluated classification models for MCI versus NC and for AD versus NC. Predictors comprised the 5 trending miRNAs alone and in combination with: 1) APOE ɛ4, 2) Aβ₄₂:T-Tau, and 3) both Aβ₄₂:T-Tau and APOE ɛ4. (For comparison, we also assessed AUC for APOE ɛ4 alone, Aβ₄₂:T-Tau alone, and Aβ₄₂:T-Tau + APOE ɛ4.) The trending miRNAs achieved an AUC of 0.705 for predicting MCI and 0.770 for predicting AD, compared to NC (Fig. 3 and Table 3). In MCI, adding the trending miRNAs to APOE ɛ4 showed a large AUC increase (+0.074) over using APOE ɛ4 alone (Fig. 3A). In AD there was an even more dramatic improvement in AUC (+0.152) from adding the trending miRNAs to APOE ɛ4 (Fig. 3B). Thus, the miRNA information was not redundant with the APOE ɛ4 genotype, but provided substantial independent power to differentiate MCI and AD from NC, presumably because the miRNAs reflect the dynamics of the disease state and not mere propensity for disease. Next, adding the trending miRNAs to Aβ₄₂:T-Tau over using Aβ₄₂:T-Tau alone showed substantial AUC increases in both MCI (Fig. 3C) and AD (Fig. 3D). Notably, the boost in classification performance from the miRNAs was larger for MCI (+0.055) than for AD (+0.036) because Aβ₄₂:T-Tau alone was a weaker predictor of MCI (AUC 0.813) than of AD (AUC 0.903). If clinical MCI indeed reflects latent neuropathology, it appears that the information from the CSF protein ratio showed more overlap with the miRNA information when disease was already present, and less overlap when disease was still nascent. Finally, adding the miRNAs to the combination Aβ₄₂:T-Tau + APOE ɛ4 robustly improved classification performance for MCI (+0.063, Fig. 3E) and modestly for AD (+0.017, Fig. 3F), relative to Aβ₄₂:T-Tau + APOE ɛ4 alone. However, note that the combination of all 3 markers (trending miRNAs + Aβ₄₂:T-Tau + APOE ɛ4) provided no meaningful improvement over the combination of trending miRNAs + Aβ₄₂:T-Tau (without APOE ɛ4): 0.819 versus 0.813 respectively for MCI, and 0.898 versus 0.903 respectively for AD. Our previous discovery and validation studies also showed that adding APOE ɛ4 or Aβ₄₂:T-Tau to the miRNAs improved classification performance for AD [12, 13]. Together these data suggest that the miRNAs combined with other biomarkers offer a larger gain for predicting MCI, suggesting less overlap and hence greater novelty of relevant information provided by the miRNAs at that stage. Thus, these miRNAs may carry substantial additional information as biomarkers for MCI.

OPEN IN VIEWER

Fig. 3

Classification performance of AD biomarkers and trending miRNAs. The effects of combining trending miRNAs with APOE ɛ4 or Aβ₄₂:T-Tau on classification performance (ROCs) are shown for MCI (A, C, E) and AD (B, D, F). AUC increased from 0.639 for predicting MCI using APOE ɛ4 alone to 0.713 with the trending miRNAs (A), and from 0.620 for predicting AD using APOE ɛ4 alone to 0.772 with the trending miRNAs (B). AUC increased from 0.758 using Aβ₄₂:T-Tau alone to predict MCI to 0.813 with trending miRNAs (C), and from 0.867 using Aβ₄₂:T-Tau alone to predict AD to 0.903 with trending miRNAs (D). Combining all 3 markers (Aβ₄₂:T-Tau + APOE ɛ4 + trending miRNAs) showed no meaningful improvement in classification performance over Aβ₄₂:T-Tau + trending miRNAs: AUC increased from 0.813 to 0.819 for predicting MCI, but decreased from 0.903 to 0.898 for predicting AD (F), in both cases only negligible differences. miRNAs, expression levels of the 5 trending miRNAs; APOE, APOE ɛ4 status; Aβ:Tau, Aβ₄₂:T-Tau ratio.

Robustness of findings to exclusion of repeated samples

As a sensitivity analysis, we temporarily excluded the repeated 21 NC and 23 AD samples (retaining 44 NC, 31 MCI, and 14 AD), and on the reduced cohort performed the same trend and classification analyses as described above. We found that essentially the same trend profiles emerged for all of the miRNAs, with some attenuation or distortion in a few cases due to the drastically reduced NC and especially AD sample sizes. We further found that the classification performance showed the same rank ordering of models as in the full cohort, as well as strong agreement in AUC values over the scenarios considered. Therefore, we believe these findings are robust whether or not repeated samples are included. Further details of this sensitivity analysis can be found in Supplementary Results.

Weak correlation between Aβ₄₂:T-Tau and trending miRNAs

We employed a multivariate correlation of the trending markers with Aβ₄₂:T-Tau to examine whether there was a correlation between the trending miRNAs and Aβ₄₂:T-Tau. The scatterplot (Fig. 4) showed a modest multivariate correlation of 0.38 (p = 0.004) that exists primarily due to a modest trend in control samples (gray). However, in the “diseased” range of Aβ₄₂:T-Tau levels (between 0 and 2, where most of the AD and MCI samples lie), there was no appreciable correlation between the trending miRNAs and Aβ₄₂:T-Tau levels (0.17, p = 0.922). Empirical trend lines in the plot indicated that a positive relationship only existed for NC samples. This suggests that information provided by the trending miRNAs as markers for MCI was mostly independent of that provided by Aβ₄₂:T-Tau levels, so that the miRNAs carry novel information relevant for prediction of AD-like pathology.

OPEN IN VIEWER

Fig. 4

Weak correlation between Aβ₄₂:T-Tau and five trending miRNAs for MCI. The canonical linear combination of miRNAs (i.e., the linear projection of miRNA expression values that best predicts Aβ₄₂:T-Tau) is shown against Aβ₄₂:T-Tau levels. What correlation exists (0.38) was primarily due to a modest trend in control samples (gray). In the “diseased” range of Aβ₄₂:T-Tau between 0 and 2, where most of the AD (orange) and MCI (blue) samples lie, there was no appreciable correlation (0.17) between the trending miRNAs and Aβ₄₂:T-Tau levels. Lines show the empirical trends: for NC there was a weak, but consistent positive relationship, and for AD and MCI the lines show random sampling fluctuations, but no apparent correlation.

Predicted targets of the trending miRNAs

We queried the predicted targets for the top 5 trending miRNAs (miR-142-3p, miR-146a-5p, miR-146b-5p, miR-365a-3p, miR-193a-5p) using TargetScan 7.2 [21] and miRDB [22, 23], as depicted in Fig. 5. Target prediction returned a total of 1,214 predicted mRNA: 367 predicted by TargetScan, 1,096 predicted by miRDB. The two target lists from each algorithm overlapped by 249 targets. We then used IPA to identify pathways that include the 1,214 total predicted targets. IPA reported that 353 of the targets were found in 214 significant pathways. Notably, Ras-related C3 botulinum toxin substrate 1 (RAC1) was predicted to be targeted by 4 of the 5 top miRNAs (miR-142-3p, miR-146a-5p, miR-146b-5p, miR-365a-3p). The significant neural, glial, and immune pathways identified by IPA that include RAC1 are listed in Table 4.

OPEN IN VIEWER

Fig. 5

Target prediction pipeline for 5 trending miRNAs. Workflow of the target prediction analysis. The 5 trending miRNAs were queried using TargetScan 7.2 with a CWCS of < – 0.3 and miRDB with a target score >60. The Venn diagram shows the overlap in miRNA targets predicted by TargetScan 7.2 and miRDB. The pathways of the 1214 predicted mRNA targets were identified using IPA: 353 of the targets were found in 214 significant pathways.