Introduction

MicroRNAs (miRNAs) are endogenous RNA transcripts that regulate diverse patterns of gene expression [1]. Most human miRNAs are transcribed as long precursors known as pri-miRNAs. Starting in the nucleus, pri-miRNAs undergo a series of processing events that ultimately result in the cytoplasmic release of mature transcripts ∼22 nucleotides in length. Mature miRNAs catalyze translational inhibition by directly binding to messenger RNAs (mRNAs) and promoting their degradation [2]. Recent data indicate that miRNAs can inhibit translation independent of their ability to induce mRNA degradation.

Patterns of miRNA expression have been extensively profiled in human tissues. It is now clear that dysregulated miRNA expression is a feature of many different cancers, including carcinomas of the breast, ovary and lung [3]–[5]. However, determining the mechanisms by which individual miRNAs contribute to cancer outcomes remains a key challenge for biologists hoping to exploit their power. Recently, the Cancer Genome Atlas Consortium (TCGA) reported that ovarian cancers cluster into distinct molecular subtypes based on their patterns of gene and microRNA expression [6]. However, we have discovered an alarming lack of consistency between the microRNA (miRNA) expression profiles initially used by the TCGA and a subsequent profile of miRNA expression generated by this group for the same ovarian cancer specimens using miRNA-Seq. As these observations challenge the validity of the underlying data, they also suggest that scientific discoveries based solely on this data should be interpreted with caution.

Results

To delineate miRNAs associated with ovarian cancer patient survival, we performed a univariate Cox regression analysis using Level 3 TCGA miRNA data for 469 ovarian cancers profiled using Agilent microarray technology. Initial regression analysis was further refined by use of the Benjamini–Hochberg (BH) procedure to adjust for multiple hypothesis testing [7]. We found that 16 mature miRNAs are significantly associated with ovarian cancer survival (FDR<0.01) (Figure 1A). Of these, miR-505, miR-652 and miR-551b* demonstrate the most robust associations. Hazard ratios (HR) calculated for these miRNAs were −1.73, −1.8, and 9.3, respectively. This result indicates that each of these miRNAs potentially plays an important role in determining ovarian cancer survival.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. MicroRNAs associated with ovarian cancer survival.

P-value plots of univariate Cox regression for microRNAs associated with ovarian cancer survival identified by microarray (A) or miRNA-Seq (B) data. P-value<0.01 (Solid line). False discovery rate (FDR)<0.1 (Dotted line). In both A&B, blue dots indicate miRNAs associated with survival by miRNA array, while red dots indicate miRNAs associated with survival by miR-Seq. Green stars are miRNAs associated with survival in both datasets. (C) percentage of overlapping miRNAs between the array and NGS seq platform at different cut-off threshold for Cox p-values, BH adjusted FDR, and Storey q-values.

https://doi.org/10.1371/journal.pone.0087782.g001

To validate these observations, we next interrogated a second dataset of miRNA expression generated for the same ovarian cancer specimens using Next Generation Sequencing (miRNA-Seq). The TCGA ovarian cancer project is unique in that miRNA expression has been profiled using both miRNA array and miRNA-Seq. These technically distinct platforms create a unique opportunity to validate discoveries made using one dataset against the other. Ideally, the results obtained should correlate well. Using Cox Proportional Hazards analysis, we found that 4 miRNA transcripts are associated with survival when miRNAs were profiled in ovarian cancers using miRNA-Seq at an identical FDR level (Figure 1B). There is no overlap between the results obtained from these two platforms, despite the fact that both datasets were generated from the same samples.

To determine whether the microarray and Next Gen platforms will give more consistent results when analyzed using a relaxed threshold, we reduced the p-value threshold used for our analyses to 0.01. This resulted in more miRNAs significantly associated with patient survival in both datasets. For example, we identified 61 miRNAs from data generated using the array platform. However, the hazard ratios estimated for the 12 miRNAs identified from miRNA-Seq data are all very close to 1.0. Only miR-652 is associated with survival in both the miRNA-Seq and microarray datasets. To correct for multiple hypothesis testing, we adjusted our Cox model p-values using Benjamini–Hochberg procedure [7]. After completing these analyses, no miRNAs are correlated with survival in both datasets when the false discovery rate was set at 10%.

To determine whether choice of a multiple hypothesis adjustment procedure contributes to these results, we re-analyzed the TCGA data using an alternative q-value estimation procedure [8]. In addition, we computed the percentage of overlapping miRNAs at different FDR or p-value cut-off. Our results indicate that the limited number of overlapping miRNAs between the two platforms is independent of the choice of multiple hypothesis adjustment procedure or cut-off thresholds (Figure 1C).

To elucidate potential causes for this unexpected discrepancy, we examined the reproducibility of miRNA expression between the two TCGA files that describe this data. Pearson correlation coefficients (r) were calculated for each of the 359 mature human miRNAs for which Level 3 expression data was available in both the miRNA-Seq and microarray databases. We found that correlation coefficients for levels of individual miRNAs reported by each technique varied widely. For example, miR-505 is the miRNA most robustly associated with patient outcome in our analyses of the miRNA array data (HR = −1.7, p<9e-5). However, when assessed using sequencing data, the hazard ratio for mir-505 was 0.998 (p = 0.03). Levels of miR-505 measured by miRNA-array and miRNA-Seq data correlated only modestly (r = 0.59) (Figure 2B). Discrepancies were also observed in a number of other miRNAs that have been previously implicated in ovarian cancer, such as miR-143 [9]. The correlation coefficient for miR-143 in our analyses was 0.39 (Figure 2C). Another miRNA well-studied in ovarian cancer is miR-141, which has been previously reported to target p38α and modulate the oxidative stress response [10], [11]. However, the correlation between levels of miR-141 in TCGA microarray and miRNA-Seq expression data is only 0.32 (Figure 2D). Overall, we found that correlation coefficients for ∼72% of miRNAs profiled in both datasets were ≤0.5 (Figure 3A, 3C), indicating poor reproducibility. Only 22% of the mRNAs measured by Agilent microarray and Illumina HiSeq using the same ovarian cancer specimens correlate poorly (r≤0.5; Figure 3B, 3C). Thus, the discrepancy we report here appears to be limited to the TCGA miRNA dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Scatter-plots of microRNA expression measured by microarray and miRNA-Seq.

(A) miR-98, (B) miR-505 (C) miR-143 and (D) miR-141.

https://doi.org/10.1371/journal.pone.0087782.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Distribution of correlations between microarray and sequencing profiles for miRNA and gene expression.

(A) Histogram of correlation coefficients for individual miRNAs measured by miRNA-Seq and miRNA array. (B) Histogram of correlation coefficients for mRNAs profiled by Illumina HiSeq and mRNA array. (C) The empirical cumulative distribution function (ECDF) of the correlation between array and sequencing for miRNA (black), filtered miRNA (color) and mRNA (gray) measurements. Nearly, 72% of miRNAs demonstrate a correlation coefficient ≤0.5 whereas 22% of RNAs have a correlation coefficient ≤0.5. When filtered based on expression level, the percentage of miRNAs with correlation ≤0.5 saturated to 56%.

https://doi.org/10.1371/journal.pone.0087782.g003

One potential cause for poor reproducibility may be the signal detection algorithm used to report levels of miRNA expression. Level 3 TCGA miRNA data are reported in two formats. The first, labeled as a “Quantification Data,” reports levels for individual human miRNAs. However, one of the advantages of miRNA-Seq is that transcripts retrieved by this technique can be precisely mapped. A second file, labeled as “Isoform Data," has also been released by the TCGA. This file reports read counts for transcripts according to their genomic location. As part of this file, transcripts are identified as either mature miRNA, miRNA* (3p arms of human miRNAs), stem-loop transcript or precursor. While working through this data, we learned that miRNA levels reported in the TCGA quantification file include read counts for miRNA precursors as well as mature miRNAs. Because miRNA precursors are currently thought to lack biologic activity, inclusion of precursors with counts for mature miRNAs could confound survival analyses. To address this issue, we retrieved read counts for mature miRNAs only from the isoform data file and repeated our analyses. However, the proportion of miRNA correlation coefficients ≤0.5 remained as high as 71% despite the use of this more precisely defined data.

A second possible explanation for the observed discrepancy might be that correlations between measures of miRNA expression depend on the frequency with which individual miRNA transcripts are expressed. If so, infrequently expressed miRNAs might be reported by one or both of the platforms used to profile miRNA expression randomly or inaccurately. To explore this hypothesis, we re-calculated correlation coefficients for each miRNA identified by both platforms after excluding any transcript in the miRNA-Seq dataset with a read count less than 5. This reduced the number of distinct miRNAs available for analysis in the miRNA-Seq data file from 705 to 380. However, the proportion of miRNAs with correlation coefficients ≤0.5 also decreased from 72% to 56%. Similarly removing poorly expressed transcripts from the pool of mRNAs profiled by Illumina HiSeq reduces the proportion of mRNAs whose correlation coefficients ≤0.5 from 22% to 20%. These observations indicate that problems detecting infrequently expressed miRNA may impact the ability or one or both platforms to reliably report miRNA expression. However, the fact that more than half of miRNA transcripts still had correlation coefficients ≤0.5 even after correcting for this issue indicates that poorly expressed transcripts are not solely responsible for the discordant patterns of miRNA expression reported by the two platforms.

To explore this issue more in depth, we calculated the range of log2 transformed expression levels for all microRNAs in the two datasets. We also developed an algorithm that allowed us to vary the threshold of expression acceptable for inclusion for analysis from a minimum value (0) to the mean log2 transformed expression level of all transcripts. For each threshold, we only considered microRNAs expressed above the threshold and recomputed the correlation between the two platforms. This analysis reveals that the exclusion of miRNA transcripts expressed less frequently than the mean only slightly improves the overall correlation between the two platforms used to profile miRNA expression (Figure 3C). As shown graphically, we found that 71% of the miRNA demonstrate correlation less than 0.5 without the use of any filtering. By utilizing an expression level filter as described, we found that the proportion of transcripts with correlation coefficients across the two platforms saturated at 56%. This is still far higher than the 22% observed with mRNA expression profiling systems.

Discussion

Much to our surprise, our analyses indicate that the microRNAs associated with survival in ovarian cancer depend highly on whether specimens were profiled by the TCGA using microarray or miRNA-Seq. Our analyses indicate that this discrepancy exists because miRNA-Seq and microarray have generated very different profiles of miRNA expression, even though the data is based on the same ovarian cancer specimens. We do not currently have a clear explanation for why miRNA expression profiles reported by the TCGA are discordant. However, understanding this discrepancy will ultimately be important for identifying which miRNAs if any are important for determining ovarian cancer outcomes.

A variety of DNA microarray technologies have been previously validated by investigators examining within platform and cross-platform reproducibility [12]–[14]. Spearman correlation coefficients reported in these studies range from 0.59 to 0.94 with a mean of 0.82. These results are similar to what we have observed for correlations between patterns of gene expression profiled using microarray and Illumina HiSeq platforms by the TCGA. Both miRNA-Seq and microarray technologies are associated with multiple technical limitations that might account for the differences we have observed. For example, cross-hybridization is a well-recognized issue that can reduce signal specificity when profiling RNA transcripts by microarray [15]. However, it seems unlikely that cross-hybridization is a primary cause of the discrepancy we observed, as the number of transcripts correlated with survival by array is greater than the number associated with survival by miRNA-Seq. One alternate explanation might be that the signal extraction algorithm used to analyze miRNA-Seq data does not accurately report miRNA levels. In general, miRNA-Seq allows for precise transcript mapping with much greater confidence. The signal extraction algorithm currently used by the TCGA to report miRNA levels includes read counts for both a mature miRNA and its corresponding precursor. Our analyses indicate that precursors account for fewer than 1% of the total miRNA counts in the TCGA isoform file. This likely reflects the use of size-fractionated RNA to prepare libraries for miRNA-Seq [5]. Thus, their inclusion or exclusion in analyses of the TCGA dataset likely has little bearing on which miRNAs are associated with ovarian cancer survival.

Collectively, these observations underscore the urgent need for well-defined algorithms for processing signals generated by miRNA-Seq and transcriptional profiling platforms. Our understanding is that the same analyses have been performed by TCGA for other cancers, including colon, breast and lung [16]–[18]. Because miRNA expression in these other cancers has not been profiled by microarray, it is not possible to repeat our analyses to determine whether the discrepancy we report is observed in other cancers. Ultimately, consistent and reliable genomic data is critical for constructing testable hypotheses and achieving the full potential of the TCGA. Our observations identify an important hazard of which investigators should be aware as they utilize the TCGA miRNA data to study ovarian cancer. For the short term, knowledge of this hazard underscores the need to validate observations made with one or both of TCGA miRNA datasets. However, for the long term, resolution of this discrepancy will be important for determining the most effective platform and signal extraction algorithms for profiling miRNA expression as part of large scale genomic profiling efforts.

Materials and Methods

Acknowledgments

The authors gratefully acknowledge communication from David Wheeler, Rehan Akban, Gordon Robertson and Andy Chu regarding TCGA miRNA data analysis algorithms.