Methods and matrices: approaches to identifying miRNAs for Nasopharyngeal carcinoma

The GWU IRB determined that the study samples used in this study did not meet the definition of human subjects research; i.e., a living individual about whom an investigator conducting research obtains: a) data through intervention or interaction with the individual or b) private identifiable information. This determination was made since the samples were limited to preexisting, de-identified specimen analysis labeled with a random code (see attached PDF).

Adapter sequences were clipped from deep sequencing reads using FastqMcf (http://​code.​google.​com/​p/​ea-utils/​wiki/​FastqMcf) and initial quality assessment performed using FastQC (http://​www.​bioinformatics.​babraham.​ac.​uk/​projects/​fastqc/​). To analyze miRNA expression profiles both miRDeep 2.0.0.5[42] and miRExpress 2.0[43] were used. Briefly, short reads were mapped to the human (UCSC hg19) and the Human herpes virus 4 (Epstein-Barr virus) genome (NCBI NC_007605.1) allowing a minimum read length of 18, zero mismatches in the seed region and a maximum of five genomic loci. Known human and EBV miRNAs were identified and quantified based on miRBase Release 19[44] entries. Using miRExpress known human and EBV miRNAs were identified from miRBase Release 19 with an alignment identity of 1% a tolerance range of four and a similarity threshold of 0.8 in the analysis. Differential expression analysis was performed separately for miRDeep and miRExpress using a negative binomial distribution in EdgeR[45]. Only miRNAs with at least one count per million in at least two samples were used in expression analysis and counts were normalized using the trimmed mean of M-values normalization method[46]. The analysis was performed using moderated tagwise dispersions. Differentially expressed miRNAs were defined as having a Benjamini and Hochberg corrected p value of ≤ 0.05.

cDNA was generated from 32–125 ng RNA using the miScript RT II kit (Qiagen) and the qPCR was performed using the miScript SYBR Green PCR Kit (Qiagen) on custom printed 96 well miScript miRNA arrays (SABiosciences). Selected miRNAs and normalization controls printed on the plate are shown in Additional file2. The qPCRs were performed using a BioRad iCycler iQ5 with an initial activation step of 95°C for 15 minutes followed by 40 cycles of 3-step cycling (denaturation, 15 sec, 94°C; annealing, 30 sec, 55°C; and extension, 30 sec, 70°C) followed by a melting curve analysis for 81 cycles at 55°C and 20 sec dwell time. C_t values were exported and analyzed using SABiosciences tool (http://​pcrdataanalysis.​sabiosciences.​com/​mirna) and relative quantitation was performed using the ΔΔC_t method[47]. SNORD and RT controls were utilized for normalization of samples.

RNA sequence data have been submitted to the Sequence Read Archive (SRA, National Center for Biotechnology Information, U.S. National Library of Medicine, Bethesda, MD) under accession number SRP029599. Microarray data were prepared according to MIAME standards and deposited in the GEO (Gene Expression Omnibus Database, National Center for Biotechnology Information, U.S. National Library of Medicine, Bethesda, MD) under accession number GSE46172.

Using the Qiagen miRNeasy FFPE kit, starting material of 2 × 10 μm sections provided RNA yields of ~100 ng/μm. The purified RNA exhibited 260/280 and 260/230 ratios of ~2.0 and ~1.9, respectively, which is considered an acceptable level of purity for the downstream applications in our program, including RNA-Seq. Both electrophoresis, using TBE-urea gels, and analysis with the Agilent 2100 BioAnalyzer (not shown) were used to monitor RNA profiles. Electropherograms of RNA isolated from FFPE showed broad peaks at < 100 nt, which indicated that the sample included small RNA species (not shown). The integrity (or RIN score) of the samples ranged between 2–3. When taken with the absence of 28S and 18S ribosomal RNA peaks, this suggested the degradation of larger RNA species. However, given the robustness of miRNAs in FFPE tissue[48] and reports from other groups[49] that RIN values have negligible effect on miRNA results, the purified RNA was considered suitable for further analysis.

Microarray and RNA-Seq exhibited similar miRNAexpression profiles in FFPE tissue.

High-throughput analysis of miRNA expression profiles typically utilizes small RNA microarrays (i.e. targeted approach) or RNA-Seq (i.e., untargeted approach)[20]. To compare the utility of the two techniques for biomarker discovery, both approaches were used to profile miRNA expression in NPC FFPE tissue compared to non-neoplastic nasorespiratory FFPE control tissue. For microarray analysis, 250 ng purified RNA from eight FFPE samples (four NPC FFPE and four non-neoplastic nasorespiratory tissue FFPE) were analyzed using the Agilent human miRNA microarray (miRBase Release 16.0), which includes 1,205 human and 144 viral miRNA targets. After hierarchical clustering and statistical analysis of differential expression, 31 miRNAs (13 down-regulated and 18 up-regulated) exhibited a fold change (FC) greater than two (p < 0.05) in tumor tissue icompared to non-neoplastic nasorespiratory control tissue (Figure 2). Four EBV miRNAs, (Bart4*, Bart5, Bart6-3p and Bart6-5p) were significantly up-regulated in the four FFPE from the histologically confirmed NPC cases versus the non-neoplastic nasorespiratory tissue controls. Absolute fold changes along with p-values for all dysregulated miRNAs obtained via microarray are shown in Additional file3.

For RNA-Seq, one μg RNA from five FFPE tissue samples (3 NPC FFPE and 2 non-neoplastic nasorespiratory tissue FFPE) was sequenced using the Illumina platform. (Two non-neoplastic nasorespiratory tissue FFPE (Control 2 and Control 4) and one NPC FFPE (Tumor) were not sequenced due to insufficient material0. Approximately 36 million reads were obtained across all FFPE samples and, after quality filtering and short read removal, >32 million reads were retained. The majority of reads (63%) mapped to the human or EBV genomes, with miRNAs constituting the predominant species of small RNAs identified (Figure 3). Analyses with miRDeep and miRExpress provided expression data for 984 and 847 known human and EBV miRNAs, respectively, each with greater than one count per million in at least two of the samples. Using EdgeR, 99 dysregulated miRNAs were identified in NPC tumor tissue versus non-neoplastic nasorespiratory control samples (Table 3). Approximately one-third (37) of these were of viral origin, all of which were up-regulated in the NPC tumor samples (Table 3).

MicroRNA expression levels obtained in microarray and RNA-Seq experiments were similar, with FC values obtained from the two methods showing a positive association (Pearson correlation 0.43; Figure 4A). Eighty-seven (87) human miRNA appeared to be dysregulated using either discovery method technology, with only 11 found in both analyses. There were 32 and 44 miRNAs identified independently by microarray and RNA-Seq, respectively (Figure 4A). We observed a strong positive correlation (Pearson correlation 0.85) between RNA-Seq and microarray FC values for miRNAs found to be significantly dysregulated in microarray analysis, and a weaker but still strong correlation for those identified as significantly dysregulated in the RNA-Seq analysis (Pearson correlation 0.60) (Figure 4A). All but three miRNAs identified as significantly dysregulated in the microarray analysis exhibited stronger up-regulation than in RNA-Seq, suggesting that cross-hybridization with closely related members of miRNA families could have inflated their intensities (Figure 4A). A similar effect, i.e. substantial differences in called differentially expressed miRNAs despite overall similarity in FC values, was recently reported when comparing the two platforms[56].

We evaluated three different RNA purification kits using excess sera from healthy controls: the Qiagen miRNeasy, Qiagen miRNeasy serum/plasma, and Qiagen QIAamp Circulating Nucleic Acid kits. Using the manufacturers’ protocols, the Qiagen QIAamp kit was the most effective of these kits with yields between 10–27 ng/μl serum compared to 0.5-1.0 ng RNA/μl from the other kits tested. This provided at least 1 μg of total RNA from the volumes available, typically ~0.25 ml, which is an amount adequate for downstream applications such as RNA-Seq (≥1 μg total RNA usually required). The purified RNA exhibited 260/280 and 260/230 ratios of 1.6-1.9 and ~1.0, respectively. Analysis of recovered RNA using the Agilent BioAnalyzer 2100 showed 28S and 18S ribosomal RNA bands. However, the enrichment of small RNA species lower than 100 nt in size was not observed (data not shown). As was the case for FFPE samples, the RIN score (~3.0) suggested degradation of larger RNA species.

To verify miRNA expression levels in tumor tissue and provide a tool for the measurement of miRNA levels in sera, custom quantitative PCR (qPCR) plates were printed with primers for 40 selected miRNAs found to be significantly dysregulated in the FFPE tissue by microarray and RNA-Seq (Additional file2). Given that only small volumes of sera were available, a preliminary examination of the effects of RNA concentration on qPCR was performed. Quantitative PCR experiments using both a small quantity of total RNA (30 ng), and an amount recommended by the manufacturers (250 ng) were conducted on RNA from both FFPE tissue and sera. A comparison of the expression values obtained using both concentrations revealed that, with one exception, no miRNAs displayed a FC difference greater than two when using these starting concentrations (Additional file4). Given these results, 30 ng total RNA was used for subsequent experiments and a threshold of greater than two-fold dysregulation, in addition to significance value of p ≤ 0.05 was imposed when determining significantly dysregulated miRNA expression levels via qPCR.

Quantitative PCR was performed on RNA from tumor and control tissue and expression levels showed good agreement with results from the microarray and RNA-Seq analyses. Of the 40 miRNAs included on the qPCR plate on the basis of their dysregulation in tissue, all were identified using qPCR. Using linear regression, ratios determined using RNA-Seq provided better correlation with qPCR results (R² 0.65), while a number of outliers reduced the overall correlation of the microarray analysis with qPCR (R² 0.43) (Figure 4B). Also evident were respective shifts of +0.70 and +0.13 in the y-intercept of the linear regression line for microarray and RNA-Seq analysis (Figure 4B). Shifts such as these have been observed when comparing ratios from qPCR to microarray or RNA-Seq ratios and they have been attributed to the use of external references in the normalization of qPCR results[56]. The use of external references for normalization makes qPCR sensitive to miRNA abundance when it varies between samples in relation to the external reference. For example, when miRNA abundance, as a proportion of total RNA, varies between samples and ribosomal RNA (which makes up the majority of total RNA) is used as the external reference. Microarray and RNA-Seq use normalization methods internal to the miRNA population and are not susceptible to this effect. In this analysis, SNORD was used to normalize qPCR results and thus positive shifts in the y-intercept could reflect the presence of less miRNA in tumor compared with control tissue.

To quantify circulating miRNAs in NPC, 12 NPC positive sera were compared to sera from four healthy controls from a Malaysian sample set using qPCR. Test samples were from age (+/− 10 years), and sex (male) matched to the NPC patients, and each sample possessed associated serological analyses providing anti-viral capsid antigen (VCA) IgG titers[34]. When these samples were compared to healthy controls, only three significantly dysregulated miRNAs were identified; miR-486-5p and miR-451 were up-regulated, and miR-100 was down-regulated (fold change ≥ 2, p ≤ 0.05) (not shown). All these miRNAs had been significantly dysregulated in tumor tissue, using either RNA-Seq and/or microarray (Figure 2, Table 3). However, ratios for miR-451 and miR-486-5p showed that expression levels were inverted, indicating significant up-regulation in sera, but down-regulation in tumor tissue. Based on serology, the NPC cases were also analyzed as three individual groups: low (VCA IgG 40–160), medium (VCA IgG 320–640), and high (VCA IgG >640) antibody titers. In addition to miR-451, -486-5p, and −100, two additional miRNAs were identified as significantly up-regulated in the low titer group (miR-25 and let-7b) both of which were also identified as significantly dysregulated in tumor tissue, although let-7b had been identified as down-regulated (Figure 5A). Finally, additional sera from the U.S (five NPC sera and three matched healthy control sera) and Germany (10 NPC sera and six healthy control sera) were analyzed using the qPCR plates and the combined the results with those from the Malaysian sera. When combined, only miR-486-5p was identified as significantly up-regulated (Figure 5A).

Despite the qPCR plates being designed using miRNAs found to be dysregulated in NPC tumor tissue, few of these miRNAs were subsequently found in the sera of NPC cases. Therefore, RNA-Seq was used as an untargeted approach to profile the miRNAs in the Malaysian sera discussed above. Sixteen (16) Malaysian serum samples from individuals with histologically confirmed NPC were pooled into four groups corresponding to low, medium and high levels of EBV VCA titers as well as a control group with no detectable VCA (above). These four groups were then analyzed using the Illumina platform. Approximately 18 million reads were obtained from the four groups, with 33% of these reads (~5.7 million) mapped to miRNAs in miRBase (Figure 3). In total, 463 miRNAs were identified in these sera. Of these miRNAs, 416 were also identified in the RNA-Seq analysis of the FFPE tissue. Differential analysis of serum miRNA counts using EdgeR resulted in the identification of 20 miRNAs that were significantly dysregulated in serum from histologically confirmed NPC cases; 7 up-regulated and 13 down-regulated (Figure 5B, Table 4). A comparison of the relative expression levels of miRNAs in sera and FFPE showed little correlation (Pearson correlation 0.20; Figure 5C), indicating that the relative expression of miRNAs in NPC FFPE tissue was not reflected in sera from NPC cases. Six miRNAs that had been previously identified as dysregulated in sera using qPCR (above) were also identified using RNA-Seq and three of these, let-7b, miR-451a and miR-486, were shown to be significantly up-regulated in sera despite their significant down-regulation in tumor tissue (Figure 5C). The other 14 significantly dysregulated sera miRNAs had not been identified as dysregulated in tumors (Tables 3 and4). Moreover, numerous miRNAs strongly dysregulated in tumor samples were absent from the sera (for example miR-205, miR-199a/b, and miR-139) (Tables 3 and4). FC values of the miRNAs that were identified as significantly dysregulated in sera by RNA-Seq and were also on the custom qPCR plate used in analysis of sera showed similar values (Figure 5D). Finally, despite the significant dysregulation of 37 EBV miRNAs in tumor, no EBV miRNAs were found to be significantly dysregulated in sera.

Both microarray and RNA-Seq analysis identified significant up-regulation of EBV miRNAs in tumor tissue. Using RNA-Seq, 37 EBV miRNAs were identified as significantly up-regulated and microarray analysis of the same samples identified four up-regulated miRNAs (ebv-miR-BART4*, ebv-miR-BART6-5p, ebv-miR-BART6-3p, ebv-miR-BART5). Though close association between EBV and NPC suggests that EBV miRNAs could serve as NPC biomarkers, no EBV miRNAs were found to be significantly dysregulated in serum by RNA-Seq. While the average depth of RNA sequencing achieved on FFPE samples was approximately 3.5 million reads, the average depth for the serum samples was significantly lower at ~1.5 million. This difference may reflect the lower amount of miRNA contained in sera and, therefore, low abundance reads may not have been detected. To validate this finding, qPCR was also used to assay EBV miRNAs in sera. Nine (9) EBV specific primers were used to screen 40 total sera; 13 sera from healthy controls and 27 sera from NPC cases from three geographic locations (U.S., Germany, and Malaysia). While EBV miRNAs were detected in all sera, no miRNAs were significantly dysregulated when case sera were compared to control sera by qPCR. This is likely the result of the extensive variation observed in sera for these miRNAs (Figure 6). When the Malaysian sera were analyzed by VCA IgG titers strata (as described above), ebv-BART-15 was found to be significantly up-regulated in the mid-VCA (VCA IgG 320–640 titre) (FC = 3.51, p = 0.04) and Ebv-BART-7* in the high VCA sample (VCA IgG >640 titre) (FC = 1, p = <0.001) (Figure 7). Moreover, there appeared to be an inverse correlation between serology (VCA IgG) and EBV miRNA levels (Figure 7), i.e. the lowest NPC VCA titer group (IgG 40–160) displayed the highest positive FC in EBV miRNAs (FC 8–491) and the highest VCA titer group (>1280) showed the lowest FC (0.39-1.02) (values < 1.0 considered a negative FC). Overall, EBV levels showed great variability, even within sera collected from a single sample population (e.g., Malaysian sera) and even when these populations were stratified based on the strength of the VCA titer. More specifically, the higher the Ct for a particular miRNA, the lower the observed VCA titer (Figure 7). In this regard, we found that lowest NPC VCA titer group (IgG 40–160) displayed the highest positive FC in EBV miRNAs (FC 8–491) and the highest VCA titer group (>1280) showed the lowest FC (0.39-1.02) (values < 1.0 considered a negative FC) (not shown).