Identification of distinct miRNA target regulation between breast cancer molecular subtypes using AGO2-PAR-CLIP and patient datasets
- Thalia A Farazi†1,
- Jelle J ten Hoeve†2,
- Miguel Brown1,
- Aleksandra Mihailovic1,
- Hugo M Horlings3,
- Marc J van de Vijver3,
- Thomas Tuschl1Email author and
- Lodewyk FA Wessels2, 4Email author
© Farazi et al.; licensee BioMed Central Ltd. 2014
Received: 24 July 2013
Accepted: 7 January 2014
Published: 7 January 2014
Various microRNAs (miRNAs) are up- or downregulated in tumors. However, the repression of cognate miRNA targets responsible for the phenotypic effects of this dysregulation in patients remains largely unexplored. To define miRNA targets and associated pathways, together with their relationship to outcome in breast cancer, we integrated patient-paired miRNA-mRNA expression data with a set of validated miRNA targets and pathway inference.
To generate a biochemically-validated set of miRNA-binding sites, we performed argonaute-2 photoactivatable-ribonucleoside-enhanced crosslinking and immunoprecipitation (AGO2-PAR-CLIP) in MCF7 cells. We then defined putative miRNA-target interactions using a computational model, which ranked and selected additional TargetScan-predicted interactions based on features of our AGO2-PAR-CLIP binding-site data. We subselected modeled interactions according to the abundance of their constituent miRNA and mRNA transcripts in tumors, and we took advantage of the variability of miRNA expression within molecular subtypes to detect miRNA repression. Interestingly, our data suggest that miRNA families control subtype-specific pathways; for example, miR-17, miR-19a, miR-25, and miR-200b show high miRNA regulatory activity in the triple-negative, basal-like subtype, whereas miR-22 and miR-24 do so in the HER2 subtype. An independent dataset validated our findings for miR-17 and miR-25, and showed a correlation between the expression levels of miR-182 targets and overall patient survival. Pathway analysis associated miR-17, miR-19a, and miR-200b with leukocyte transendothelial migration.
We combined PAR-CLIP data with patient expression data to predict regulatory miRNAs, revealing potential therapeutic targets and prognostic markers in breast cancer.
Breast cancer is a heterogeneous disease involving various tumorigenesis mechanisms manifesting at the DNA, RNA, and protein level. Patients are classified by estrogen receptor (ESR/ER), progesterone receptor (PGR/PR), and ERBB2/HER2 amplified oncogene expression based on immunohistochemistry, molecular subtypes based on mRNA expression signatures (luminal, basal-like, HER2, normal-like), or integrated clusters based on combination of mRNA expression and DNA copy number alteration . Prognostic mRNA expression signatures have been defined for specific sets of breast tumors [2, 3], but given the heterogeneity of patient outcomes within the same subtype, it is clear that pathways regulating tumor aggressiveness remain to be further elucidated. miRNAs have shown promise as therapeutic targets in cancer, suggested by the recent introduction of the first miRNA mimic in Phase I cancer clinical trials, and as diagnostic/prognostic markers, suggested by their cell-type specificity. Oncogenic and tumor suppressive miRNAs have been implicated in the regulation of critical cellular pathways, such as differentiation and apoptosis, across several tumor types [4–6], but identifying miRNA target regulation/repression in tumor samples remains challenging.
Multiple studies have examined the correlation between miRNA and mRNA expression in breast tumors as well as the role of miRNA expression in prognosis, using samples from variable molecular subtypes, but a clear conclusion has yet to be reached (Additional file 1: Table S1) [7–12]. The Cancer Genome Atlas (TCGA) published same-sample miRNA and mRNA expression profiles for a large patient collection (n = 797) determined by sequencing but has not commented on miRNA targeting activity and prognosis . Finally, a recent study including 1,302 breast tumors, utilizing miRNA and mRNA expression by microarrays, did not determine direct miRNA target repression . The variability of findings, some of which is due to technical limitations of quantification methods, highlights the need for further studies and detailed examination of approaches used for correlation analysis aimed at establishing regulatory relationships between miRNAs and their targets in patient samples.
Correlations between miRNA families and their targets depend on mRNA and miRNA abundance
We conducted correlation analysis of the same-sample miRNA-mRNA expression from 161 patient samples from our earlier study , and a selection of 444 samples from the TCGA study . Our samples included normal breast, ductal carcinoma in situ (DCIS), and invasive ductal carcinoma (IDC), comprising a variety of molecular subtypes. TCGA samples included invasive breast carcinomas also comprising a variety of molecular subtypes. In our dataset miRNA abundance was measured as relative read frequency (RRF) and mRNA abundance as the average fluorescence intensity from both channels of Operon arrays (A-value, see Materials and methods). In the TCGA dataset miRNA and mRNA expression levels were determined by sequencing; the miRNA abundance reported as RRF and mRNA abundance as reads per kilobase per million (RPKM). We confirmed that intronic miRNAs and their host protein-coding genes were positively correlated and established thresholds for miRNA abundance, selecting a threshold of 1e-4 RRF (see Materials and methods; Additional file 2: Figure S1 and S2).
To assess direct miRNA-target repression, we investigated whether correlations between expression of miRNAs with their computationally predicted-targets were more negative compared to all remaining miRNA-mRNA correlations, and explored whether mRNA abundance thresholds influenced the strength of the correlations. There are many miRNA target prediction algorithms, previously reviewed in depth [19–21]. TargetScan  and miRanda  demonstrated similar performance when evaluating the significance of enrichment of negative correlations between miRNAs and their targets in datasets from TCGA . In addition to canonical miRNA targets defined by both algorithms, miRanda also determines non-canonical miRNA targets, computing a miRSVR score as the weighted sum of a number of sequence and context features of the predicted miRNA-mRNA duplex . Our analysis showed that a larger set of conserved TargetScan-predicted targets performed similarly to a smaller set of stringent miRSVR scoring miRanda-predicted targets (Additional file 2: Figure S3) . Thus, we chose to conduct our analysis using conserved TargetScan-predicted targets focusing on miRNA seed families to group miRNAs with similar regulatory potential. When we refer to miRNA correlations with their respective targets we refer to miRNA seed families as defined by TargetScan (referenced by the miRNA member of the lowest number).
Correlation analysis within molecular subtypes reveals varying degrees of miRNA target repression
Molecular subtypes with variability in expression of their dominant miRNAs, but less variability in their mRNA expression, are more likely to display negative miRNA-TargetScan-target correlations. Therefore, we conducted miRNA-mRNA correlation analyses by molecular subtypes of breast cancer  using the miRNA/mRNA abundance thresholds defined above. Our dataset  included 78 basal-like, 23 HER2, 25 luminal A, six luminal B, and 21 normal-like samples (10 carcinomas and 11 normal breast); eight samples could not be assigned to a particular subtype . The 444 TCGA samples were subdivided to molecular subtypes using the PAM-50 classification scheme based on Agilent microarray data (84 basal, 52 HER2, 205 luminal A, 103 luminal B) .
AGO2-PAR-CLIP-defined biochemical miRNA targets in MCF7 breast cancer cell line
To identify which miRNA-target pairs are more likely to display regulation, we used AGO2-PAR-CLIP  to capture biochemical miRNA targets and define their specific location within 3′ UTRs and CDSs, in the MCF7 luminal subtype and ER-positive/HER2-negative breast cancer ductal cell line . Even though MCF7 cells display distinct mRNA profiles compared to cell lines belonging to the basal-like subtype (cell line subtypes defined in ), they share many abundant miRNAs with other breast cancer cell lines and tumors across all molecular subtypes . MCF7 cells exhibit a drastic upregulation of miR-21, similar to breast tumors when compared to normal breast tissue .
Top expressed miRNA TargetScan families in MCF7 cells
miRNA TargetScan family
All other miRNAs
Regression model predicts additional miRNA targets
TargetScan lists theoretically possible target sites within annotated 3′ UTRs, whereas PAR-CLIP provides evidence for expressed targets within MCF7 cells, and depending on sequencing depth may not have covered low-level-expressed miRNAs that may be more abundant in patient samples within different molecular subtypes. Using PAR-CLIP, we identified 3,597 canonical miRNA-target interactions (assuming seed sequence complementarity, including targets in the 3′ UTR and CDS), 2,584 of which were predicted by TargetScan (1,507 conserved and 1,077 non-conserved). To identify additional-subtype-specific miRNA targets from the large number of miRNA-TargetScan-target interactions (72,770 conserved and approximately 3.5 million non-conserved) and prioritize them, we followed a supervised machine learning approach (elastic net regression model; combination of LASSO and ridge regression). The goal of this approach was to build a model that can predict, based on characteristics of the miRNAs and their targets, whether a miRNA-target interaction is, in fact, a true interaction as determined by PAR-CLIP. As inputs to this model we used characteristics of the PAR-CLIP identified targets (number of 7-mer and 8-mer sites, conservation and context score derived from TargetScan) and their expression levels in patient subtypes (Additional file 5: Table S4 and Materials and methods for description). The training and test sets were constructed using all miRNA-TargetScan-target pairs that are: (1) expressed according to our miRNA and mRNA abundance thresholds in patients for each subtype; and (2) include an AGO2-crosslinked mRNA target (n = 10,200 for luminal A subtype). We used 5,106 for training the model and the remainder for testing model performance. As positive set we employed the crosslinked and PAR-CLIP-site seed-matched miRNA-TargetScan-target pairs (n = 561 for luminal A subtype). As negative set we employed crosslinked, but not PAR-CLIP-site seed-matched, miRNA-TargetScan-target pairs (n = 4,545) (Additional file 2: Figure S5). Our trained model allowed us to predict and rank miRNA-TargetScan-target pairs based on their likelihood of being ‘PAR-CLIP-like’ interactions (further details in Materials and methods).
For the luminal A subtype (which is the closest match to the MCF7 cell line in which the PAR-CLIP targets were determined), we obtained an area under the curve (AUC) of 0.73 for both training and test sets (Additional file 2: Figure S5). We chose a 0.5 threshold on the posterior probability, resulting in an FDR of approximately 0.5 (Figure 4C). We evaluated 12,925 conserved and 45,293 non-conserved miRNA-TargetScan-target interactions (meeting our miRNA and mRNA thresholds). We predicted 283 interactions from all TargetScan interactions, 41 of which were supported by PAR-CLIP, thus identifying 233 conserved and 9 non-conserved additional target interactions (additional 14%) . These interactions involved 23 miRNA families, mainly let-7 and miR-29a. Model-predicted targets not identified by PAR-CLIP exhibited a median RPKM expression of 5 in MCF7 cells, compared to 14 for targets supported by PAR-CLIP (expression from ). This suggested that the regression model adds not only targets for highly expressed miRNAs in patient tissues (38 interactions including miR-125, miR-142-3p, miR-145, miR-199a, miR-21 and miR-34a), but also miRNA targets abundant in patient tissues missed from PAR-CLIP due to their lower abundance in MCF7 cells.
We observed a greater difference of the medians of the distribution of correlations for miRNA families and their model-predicted targets compared to the distribution of correlations of remaining miRNA-mRNA pairs, as opposed to miRNA-Targetscan targets and PAR-CLIP targets, supporting our approach (Figure 4D). The TCGA dataset showed similar results (Additional file 2: Figure S6).
We defined miRNA targets by taking the union of the biochemical PAR-CLIP and regression model-predicted targets calculated within each molecular subtype to focus on experimentally tractable targets. Irrespective of their behavior in patient data (inherent with variability due to sample annotation and profiling method, as well as feedback regulation) PAR-CLIP targets are supported by crosslinking evidence in a breast cancer cell line at a binding site resolution, while model-predicted targets resemble PAR-CLIP targets and result in a greater difference of the medians of the two correlation distributions. We will refer to this set of miRNA-target pairs as the Model Predicted and PAR-CLIP (MP-PCLIP) pairs (n = 2,008 in the luminal A subtype: 1,766 from PAR-CLIP and an additional 242 from model prediction).
To understand the contribution of each individual input to predict PAR-CLIP targets we conducted univariate correlation analyses (Additional file 5: Table S4). TargetScan total context score, aggregate conservation score, and number of conserved 7-mer and 8-mer sites showed the highest correlation to PAR-CLIP status, hence providing the most predictive power in the model [18, 31, 34]. We also observed that miRNA abundance in patient samples correlated with PAR-CLIP status, supporting a threshold in miRNA abundance required for measureable regulation of mRNAs.
miRNA pathway associations across molecular subtypes
After selecting miRNA targets expressed in the different patient subtypes from the MP-PCLIP pairs, we used the Global Test (GT) to analyze miRNA-mRNA associations in the context of KEGG pathways . The GT can be used to determine whether the global expression pattern of a group of gene sets is significantly related to a variable, as supported by either negative or positive correlations. We assessed whether miRNA expression significantly associated with expression of genes belonging to KEGG pathways (obtaining a GT P value for the association; results for each individual subtype and dataset can be obtained at . The majority of miRNA-pathway associations that included MP-PCLIP targets, included a negative correlation between the miRNA and at least one of its respective targets. For the majority of miRNAs, miRNA-pathway associations that included an MP-PCLIP target showed lower P values compared to miRNA-pathway associations that did not (t-test P value <0.05), further validating our approach (Additional file 6: Table S5).
As expected, most pathways were targeted by more than one miRNA. There was a large number of significant pathway associations for the miR-17, miR-19a, and miR-25 families in the basal-like subtype, with very few significant associations in the HER2 subtype in our dataset. The most significant miRNA-pathway association in the basal-like subtype was the association of miR-17 family with leukocyte transendothelial migration (P value = 3.5e-8), including a negative correlation between miR-17 family and its PAR-CLIP identified target CXCL12  (Additional file 2: Figure S8). In the TCGA dataset, similarly to our dataset, miR-17 and miR-25 families showed many pathway associations within the basal-like subtype but not in the HER2 subtype.
Ranking miRNA regulatory activity and tumor phenotype association across molecular subtypes
To elucidate miRNA-mediated regulation in the context of tumorigenesis, we performed an overall ranking of miRNAs by combining a number of evidence sources . There are three components we considered in prioritizing miRNA regulatory activity: (1) association with its respective targets; (2) association with pathways - indicative of the miRNA’s ability to regulate its targets and in turn the pathways they regulate; and (3) association with cancer-related genes. A miRNA ranks high if achieving a high score (low P value) for each of the following statistical tests: (1) association of miRNA expression to expression of its respective targets based on the GT P value; (2) association of miRNA expression with expression of genes belonging to a KEGG pathway containing at least one MP-PCLIP target displaying either a negative or positive correlation with the miRNA (indicating functional relevance) (smallest GT P value out of all targeted pathways in KEGG); and (3) association of miRNA expression with expression of the gene set representing the Cancer Genome Census, modeling cancer relevance (GT P value) (see Materials and methods for further details). Each of the three tests is weighted equally in the ranking .
Top scoring miRNA TargetScan families in the Farazi and TCGA datasets
Expression mean (RRF)
Expression min (RRF)
Expression max (RRF)
Rank test P value
Luminal A (25)
Luminal B (6)
Luminal A (205)
Luminal B (103)
At the same time, to elucidate miRNA tumor phenotype association in each subtype, we performed a second overall ranking of miRNAs by combining a set of evidence sources associated with patient histopathological and clinical characteristics, using the rank test described above . These are GT P values assessing whether expression of miRNA families and their respective targets significantly related to development of distant metastasis and overall survival, number of positive lymph nodes, tumor size, lymphovascular invasion, and histological grade. The highest scoring miRNA family in our dataset was miR-130a in the basal-like subtype (Additional file 7: Table S6), regulating angiogenesis . In the NKI295 dataset, which was used for validation of these results, miR-130a family ranked third, but did not reach statistical significance (Additional file 7: Table S6). Expression of miR-130a targets was also associated with distant metastasis-free survival and relapse-free survival in the basal-like subtype in a large cohort of breast cancer samples (see analysis in following section). Expression of miR-203 targets (implicated in cancer stem cell characteristics ) significantly correlated with lymphovascular invasion in the basal-like subtype in our dataset, a finding also supported in the luminal A subtype in the NKI295 dataset. It is interesting to note that the top ranked miRNAs according to regulatory activity do not necessarily overlap with the top ranked miRNAs according to association with tumor phenotype, but may be more interesting candidates for targeted therapy as they have a detectable regulatory role.
Expression of miR-182 targets predicts metastasis
To determine whether expression levels of miRNAs and their respective targets predicted metastasis and overall survival, we used the GT with Cox-regression in our and the NKI295 study  (Additional file 7: Table S6). The NKI295 study includes mRNA microarray expression for 295 samples (55 luminal B, 123 luminal A, 29 normal-like, 53 basal-like, and 35 HER2). We selected 283 samples from patients with metastasis as first event to compare to our dataset. TCGA only reports overall survival with a short follow-up (average = 736 days), so we did not use it in this analysis. Expression of miR-182 targets, recently reported to be involved in breast cancer metastasis , was significantly associated with overall survival when considering all NKI295 patients. This prognostic signature comprised 12 genes with expression in the NKI295 series (XBP1, IGF1R, THBS1, PLAGL2, YWHAG, ZFP36, PSMC2, CCNG1, HSPA8, PFN1, ADCY6, NUP50). MiR-182 regulatory activity ranked fourth in the HER2 subtype in the TCGA dataset. None of the results within individual subtypes in our and the NKI295 dataset reached statistical significance after multiple testing correction and multivariate analysis accounting for histologic grade, tumor size, and lymph node status. However, we did notice weak concordance in the ranking of metastasis prognostic signatures between our and the NKI295 datasets in the basal-like and HER2 subtypes (correlation 0.35 and 0.43, P value <0.05). Finally, we further assessed the miRNA target prognostic signatures in two additional datasets (n = 623 (distant metastasis-free survival) and n = 1,616 (relapse-free survival)), using normalized mRNA expression from a large cohort of breast cancer samples [42, 43]. The clinical and histopathological characteristics were unavailable, so we could not conduct multivariate analysis for these datasets. miR-183, which is co-expressed with miR-182, was the top prognostic signature in these datasets, with miR-182 still maintaining significance, providing some support for our results (Additional file 7: Table S6).
Functional studies in breast cancer cell lines and mouse models have suggested multiple roles played by miRNAs in the development of breast carcinomas and their metastatic potential involving targets regulating many cellular pathways. However, which miRNA-target pair(s) is (are) important in human disease progression is not always predicted by cell culture or animal model studies alone. Here we examined the extent of correlation in mRNA and miRNA expression in large sample collections by prioritizing the effects of miRNAs on many targets.
High miRNA abundance is critical for experimental analysis of transcriptome-wide seed-dependent target mRNA repression [44–47]. In our study we showed the importance of miRNA and mRNA abundance thresholds to focus on more reliably quantified and molecularly validated miRNA targets to conduct computational analysis of miRNA-mRNA correlations in tumor samples. The recent study by Dvinge et al.  did not impose sequence-based derived thresholds for miRNA expression and did not document miRNA repression in breast cancer, as suggested by lack of enrichment of negative correlations for miRNA-target pairs. Our approach documented miRNA and mRNA expression changes consistent with miRNA target regulation and focused on miRNA-target pairs based on their crosslinking to AGO2 through PAR-CLIP. This limited the large number of possible miRNA-TargetScan-target pairs to experimentally tractable pairs.
Even though miR-21 is highly expressed both in MCF7 cells and patient breast tumor samples, we were only able to identify a small number of its targets crosslinked by AGO2-PAR-CLIP. A recent article sheds some light into the targeting behavior of miR-21 . They showed that miR-21 exhibited poor mRNA silencing activity in healthy mouse liver, despite being one of the top expressed miRNAs in this tissue, and suggested that reduced thermodynamic stability of seed pairing and target binding may contribute to this effect. At the same time, they were able to document target miR-21 regulation in HeLa cells, suggesting that the effect may be modulated by competition from AU-rich-RNA-binding proteins differentially expressed in distinct cell types.
We showed that conducting the analysis in each tumor subtype pointed to miRNAs and associated pathways that may represent therapeutic targets for specific groups of patients. Members of the mir-17~92 cluster had high miRNA regulatory activity (Table 2) in the basal-like subtype both in our and the TCGA dataset. MiR-17 and miR-19a families were associated with the leukocyte transendothelial migration pathway, with similarities to metastasis, and were negatively correlated with their PAR-CLIP target CXCL12. CXCL12, involved in metastasis , was also a PAR-CLIP target of other miRNA families (miR-7, miR-23a, miR-182, and miR-183) (Additional file 2: Figure S8).
Our prioritization of miRNA regulatory activity selects for miRNAs that show regulation through association with their respective targets and regulated pathways, as well as genes implicated in cancer, in distinct molecular subtypes. We consistently observed regulation by miRNAs in the basal-like subtype across two independent datasets. Detecting miRNA activity and cancer association does not necessarily predict whether inhibiting or over-expressing the miRNA will have therapeutic benefit - it simply points to the relevance of the prioritized miRNA as evidenced by repression of its targets in patient tissues. Two recent manuscripts also point to the importance of two of our top prioritized miRNA families: miR-200 and miR-22 [50, 51] (Table 2). Song et al. found that miR-22 regulated breast cancer stemness and metastasis via TET-family-dependent chromatin remodeling. In vitro and in vivo experiments showed that miR-22 promoted epithelial mesenchymal transition and tumor invasion and metastasis. Our results point to high miR-22 activity in the luminal B subtype in the TCGA dataset, as well as the HER2 subtype in both datasets (ranked second with P value >0.05 in TCGA and P value <0.05 in our dataset). Another study by Pecot et al. showed that the miR-200 family blocked cancer angiogenesis specifically in the basal-like subtype. Our results point to high miR-200b family activity in the basal-like subtype in our dataset.
Abundant miRNAs repress their respective targets in breast tumor-related processes, as documented by regulation of their targets in patient tissues. This regulation is subtle and may not be readily revealed in global analysis with a moderately large number of patient samples, but only by using approaches involving data curation and biochemical evidence, relying on miRNA sequencing-derived abundance. Moreover, this regulation may only be evident when conducting the analysis within individual molecular subtypes: for example, the extent of regulation as supported by pathway association in the HER2 subtype is less pronounced compared to the other subtypes.
We can only detect regulation for few highly abundant miRNAs, and can only validate three of these miRNAs across two independent datasets. Challenges and caveats to interpretation of our results include: (1) patient heterogeneity between the different patient datasets examined; (2) noise in the patient mRNA profiles due to the different platforms used for their detection (that is, sequencing vs. microarray); (3) assumptions made for the detection of miRNA targets, mainly focusing on targets that exhibit a negative correlation between their respective regulating miRNAs to derive thresholds for miRNA and mRNA abundance and negative or positive correlations for miRNA pathway association. Lack of detection of miRNA activity using our methodology does not necessarily rule out miRNA-mediated regulation; the analysis, instead, focuses on providing support from patient data for a few miRNAs that could be considered promising candidates for therapeutic manipulation. Finally, the challenges in validating prognostic signatures across datasets are not unique to our study, but represent frequent complexities arising from breast cancer heterogeneity and the different sets of genes detected by microarray and/or sequencing methodologies not allowing a direct comparison of gene expression signature performance.
In conclusion, we provide a list of miRNA targets, associated pathways, tumor phenotypes, and miRNA regulatory activity derived from patient samples as well as supported by biochemical evidence, to allow generation of clinically relevant hypotheses. Our analysis allows definition of a few specific miRNAs as potential therapeutic targets and prognostic markers in breast cancer and can be applied to other patient datasets.
Materials and Methods
Datasets and analysis
Our miRNA dataset was reported in . mRNA abundance values (A) correspond to the fluorescence intensity averaged from both dye swap NKI Operon array experiments: defined as log2(sqrt(R*G)), where R and G are the red and green fluorescent channels. mRNA expression was normalized to a set of 100 tumors (log2(fold-change)). Probes correlating >0.8 were condensed to genes by averaging probe log2(fold-change). The TCGA dataset is described in  and was downloaded from (; 2013-02). miRNA counts correspond to the most abundant isoform read measured for each miRNA within each sample and normalized to RRF. Detected miRNAs were defined as having more than 10 reads in at least 5% of the samples. Detected mRNAs were defined as having more than 20 reads in at least 5% of the samples. mRNA RPKM values of 0 were set to the lowest non-zero RPKM value measured in a given sample and subsequently log2-transformed. The NKI295 dataset is described in  and downloaded from , with an updated median follow-up of 12 years.
Intronic miRNAs were obtained from Table S2 in . We excluded miRNAs with multiple copies, as they cannot be assigned to a single host gene. We used TargetScan version 6.2  (context score and evolutionary conservation scores aggregated per gene and miRNA; Summary Counts file) and miRanda-miRSVR August 2010 release  (miRSVR scores aggregated per gene and miRNA). KEGG pathways were obtained from BioConductor , the CGC from  (Table_1_full_2012-01-18.xls). GT 5.12.0 and glmnet 1.9-3 packages were obtained from BioConductor version 2.11 (R version 2.15.3; 2013-03-01). Rank test for miRNA regulatory activity and phenotype association as described in . Figure 1 and Additional file 2: Figure S9 describes the analysis outline and provides examples of the tables generated.
miRNA and mRNA abundance thresholds for patient data
We evaluated thresholds for miRNA and mRNA expression to focus on higher confidence correlations. We established that overall expression of intronic miRNAs and their protein-coding host genes displayed a positive Pearson correlation, as described in [23, 59] (Additional file 2: Figure S1A-B; Additional file 8: Table S7). We next investigated whether miRNA abundance influenced the positive correlations observed between expression of intronic miRNAs and their host genes. In our dataset, the correlation results for poorly expressed intronic miRNAs near the detection limit were more variable as compared to higher expressed miRNAs, which displayed stronger positive correlations with their host genes (P = 0.001) (Additional file 2: Figure S1C). mRNA abundance did not influence the correlation between intronic miRNAs and host genes, likely due to the non-linear variation in our array-based measurements (Additional file 2: Figure S1D). Hybridization-based mRNA arrays do not display the same linear variations for detection of lower expressed mRNAs, and may also reach saturation during detection of highly expressed mRNAs. We therefore set the miRNA expression threshold to an RRF of 1e-4 (corresponding to an average correlation of 0.28). Given that TCGA was sequenced deeper than our dataset (750,000 compared to 5,000 minimum reads per sample), almost all correlations between expression of intronic miRNAs and their host genes were positive (Additional file 2: Figure S2A).
Applying more stringent TargetScan thresholds for aggregate conservation/PCT or total context score resulted in an even greater difference between the medians of the two correlation distributions at our selected miRNA and mRNA abundance thresholds (Additional file 2: Figure S10), further supporting the use of TargetScan.
We conducted the following GTs  for miRNA regulatory activity. First, we conducted a GT evaluating the association of miRNA expression with expression of its MP-PCLIP targets (miR ~ target1 + … + targetN). Second, we conducted a GT evaluating the association of miRNA expression with expression of gene sets corresponding to KEGG pathways (miR ~ kegg1.gene1 + … + kegg1.geneN,…, miR ~ keggK.gene1 + … + keggK.geneN) (examples can be found in Additional file 2: Figure S8). Third, we conducted a GT evaluating the association of miRNA expression with expression of the genes comprising the Cancer Gene Census (miR ~ cgc.gene1 + … + cgc.geneN). For tumor phenotype association, we conducted GTs evaluating the association of expression of a miRNA along with expression of its respective targets (miRNA target expression signature) to a particular tumor clinical or histopathological characteristic. We used logistic regression for association with lymph node status and lymphovascular invasion (yes or no), multinomial regression for tumor size (<2, 2-5, >5 cm) and histologic grade (good, moderate, poor), and Cox-regression for association with time to metastasis and overall survival (patient characteristics described in ). Multiple testing correction was conducted using the Benjamini-Hochberg method.
We used a combination of LASSO and ridge multivariate regression (glmnet package) to predict whether a given miRNA-TargetScan-target is a PAR-CLIP identified pair (true or false). As input to the model we employed the following variables: (A) TargetScan: aggregate conservation/PCT score, total context score, total number of conserved/non-conserved sites, total number of 7-mer m8, 7-mer 1A, and 8-mer conserved/non-conserved sites; (B) Patient data: miRNA/mRNA abundance/variance, miRNA-mRNA interaction terms (miRNA abundance multiplied by mRNA abundance considering sign of mRNA log2(fold-change), or irrespective of sign). We viewed the predictive model as hypothesis generating and not as a final set of high confidence pairs to have a larger set of miRNA-target pairs to include in further enrichment and association studies. Thus, we used a posterior probability prediction cutoff of 0.5 because it resulted in the best model performance, as judged by the positive predictive value (PPV) or FDR of 50%, yielding 283 miRNA-target pairs (Additional file 2: Figure S5). Increasing the posterior probability prediction cutoff to 0.7 for the TCGA dataset allowed us to reach an FDR of approximately 25%, but resulted in prediction of only 23 miRNA-target pairs (Additional file 2: Figure S6). Increasing the mRNA abundance threshold did not result in improvement in model performance (Additional file 2: Figure S11). Additional file 2: Figure S12 depicts the distribution of low- and high-expressed genes in the patient luminal A samples as a function of the MCF7 cell RPKM expression levels.
Biochemical identification of miRNA targets using AGO2-PAR-CLIP
MCF7 cells were obtained from ATCC and grown at 37ºC in an atmosphere containing 5% CO2 in Dulbecco’s modified Eagle’s medium (1X D-MEM/high-glucose/L-glutamine/sodium pyruvate) supplemented with 10% heat inactivated fetal bovine serum, 100 unit/mL penicillin, 100 mg/mL streptomycin (Invitrogen, Sigma, and Gibco). Cells were grown in the presence of 100 μM 4-thiouridine (4SU) for 24 h and AGO2 complexes were immunoprecipitated using a monoclonal antibody against AGO2 (Millipore clone 9E8.2; used in [28, 29]), according to . We used lysis buffer in lieu of high-salt wash buffer to not disrupt the monoclonal antibody-bead interaction. Crosslinked RNA of 20-40 nt in length was recovered from the 100 kDa AGO2 immunoprecipitated protein complexes separated on SDS gel, confirmed by Western blot probing with a polyclonal antibody recognizing AGO2 (Millipore 07-590). The isolated RNA was converted into cDNA libraries, and sequenced by Illumina at the Rockefeller University Genomics Center. We analyzed the data similarly to . The sequence reads were aligned to the human genome and transcript sequences from public databases, allowing for up to one mismatch. Overlapping reads >20 nt were clustered, and clusters containing <5 sequence reads or those with a content of <20% crosslinked sequences were not considered. A T-to-C conversation rate of 80% and 40% was noted for mRNA and miRNA reads, respectively. The lower T-to-C conversion rate for miRNAs was noted in our previous publication  and is likely due to the association of AGO2 with background abundant non-crosslinked miRNAs (such as, miR-21). miRNA targets were defined for the 69 top-expressed miRNAs in MCF7 cells (95% of miRNA sequence reads) by searching the sequences for complementary miRNA seed sequence sites (position 2-8, 1-7 perfect match, or allowing A at position 1), that are enriched within the isolated mRNAs compared to random sequences of the same di-nucleotide composition. The raw sequencing file is deposited with the Sequence Read Archive (SRX388831; ). Finally, we compared the number of conserved TargetScan and high miRSVR scoring Miranda miRNA-target interactions validated by PAR-CLIP. Accounting for expression of potential targets in MCF7 cells (RPKM >14), PAR-CLIP validated 8.3% of conserved TargetScan-predicted targets (3,104) and 9.9% of high miRSVR (<-1.2) scoring Miranda-predicted targets (1,970).
AGO2-Photoactivatable-ribonucleoside-enhanced crosslinking and immunoprecipitation
Area under the curve
Coding DNA sequence
Ductal carcinoma in situ
False discovery rate
Invasive ductal carcinoma
Positive predictive value
Reads per kilobase per million
Relative read frequency
The Cancer Genome Atlas
We thank Anders Jacobsen for assistance with the processed level 3 TCGA miRNA and mRNA dataset and discussions; Olivier Elemento, Pavel Morozov and Doron Betel for helpful discussions. We thank Alexander Kotlyar for technical assistance at the beginning of the project.
- Curtis C, Shah SP, Chin SF, Turashvili G, Rueda OM, Dunning MJ, Speed D, Lynch AG, Samarajiwa S, Yuan Y, Graf S, Ha G, Haffari G, Bashashati A, Russell R, McKinney S, Langerod A, Green A, Provenzano E, Wishart G, Pinder S, Watson P, Markowetz F, Murphy L, Ellis I, Purushotham A, Borresen-Dale AL, Brenton JD, Tavare S, Caldas C, et al: The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. Nature. 2012, 486: 346-352.Google Scholar
- Wang Y, Klijn JG, Zhang Y, Sieuwerts AM, Look MP, Yang F, Talantov D, Timmermans M, Meijer-van Gelder ME, Yu J, Jatkoe T, Berns EM, Atkins D, Foekens JA: Gene-expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer. Lancet. 2005, 365: 671-679.View ArticleGoogle Scholar
- van de Vijver MJ, He YD, van’t Veer LJ, Dai H, Hart AA, Voskuil DW, Schreiber GJ, Peterse JL, Roberts C, Marton MJ, Parrish M, Atsma D, Witteveen A, Glas A, Delahaye L, van der Velde T, Bartelink H, Rodenhuis S, Rutgers ET, Friend SH, Bernards R: A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med. 2002, 347: 1999-2009. 10.1056/NEJMoa021967.View ArticleGoogle Scholar
- Ventura A, Jacks T: MicroRNAs and cancer: short RNAs go a long way. Cell. 2009, 136: 586-591. 10.1016/j.cell.2009.02.005.View ArticleGoogle Scholar
- Farazi TA, Spitzer JI, Morozov P, Tuschl T: miRNAs in human cancer. J Pathol. 2011, 223: 102-115. 10.1002/path.2806.View ArticleGoogle Scholar
- Gurtan AM, Sharp PA: The role of miRNAs in regulating gene expression networks. J Mol Biol. 2013, 425: 3582-3600. 10.1016/j.jmb.2013.03.007.View ArticleGoogle Scholar
- Blenkiron C, Goldstein LD, Thorne NP, Spiteri I, Chin SF, Dunning MJ, Barbosa-Morais NL, Teschendorff AE, Green AR, Ellis IO, Tavare S, Caldas C, EA M: MicroRNA expression profiling of human breast cancer identifies new markers of tumor subtype. Genome Biol. 2007, 8: R214-10.1186/gb-2007-8-10-r214.View ArticleGoogle Scholar
- Smeets A, Daemen A, Vanden Bempt I, Gevaert O, Claes B, Wildiers H, Drijkoningen R, Van Hummelen P, Lambrechts D, De Moor B, Neven P, Sotiriou C, Vandorpe T, Paridaens R, Christiaens MR: Prediction of lymph node involvement in breast cancer from primary tumor tissue using gene expression profiling and miRNAs. Breast Cancer Res Treat. 2011, 129: 767-776. 10.1007/s10549-010-1265-5.View ArticleGoogle Scholar
- Enerly E, Steinfeld I, Kleivi K, Leivonen SK, Aure MR, Russnes HG, Ronneberg JA, Johnsen H, Navon R, Rodland E, Makela R, Naume B, Perala M, Kallioniemi O, Kristensen VN, Yakhini Z, Borresen-Dale AL: miRNA-mRNA integrated analysis reveals roles for miRNAs in primary breast tumors. PLoS One. 2011, 6: e16915-10.1371/journal.pone.0016915.View ArticleGoogle Scholar
- Van der Auwera I, Limame R, van Dam P, Vermeulen PB, Dirix LY, Van Laere SJ: Integrated miRNA and mRNA expression profiling of the inflammatory breast cancer subtype. Br J Cancer. 2010, 103: 532-541. 10.1038/sj.bjc.6605787.View ArticleGoogle Scholar
- Buffa FM, Camps C, Winchester L, Snell CE, Gee HE, Sheldon H, Taylor M, Harris AL, Ragoussis J: microRNA-associated progression pathways and potential therapeutic targets identified by integrated mRNA and microRNA expression profiling in breast cancer. Cancer Res. 2011, 71: 5635-5645. 10.1158/0008-5472.CAN-11-0489.View ArticleGoogle Scholar
- Cascione L, Gasparini P, Lovat F, Carasi S, Pulvirenti A, Ferro A, Alder H, He G, Vecchione A, Croce CM, Shapiro CL, Huebner K: Integrated MicroRNA and mRNA signatures associated with survival in triple negative breast cancer. PLoS One. 2013, 8: e55910-10.1371/journal.pone.0055910.View ArticleGoogle Scholar
- TCGA: Comprehensive molecular portraits of human breast tumours. Nature. 2012, 490: 61-70. 10.1038/nature11412.View ArticleGoogle Scholar
- Dvinge H, Git A, Graf S, Salmon-Divon M, Curtis C, Sottoriva A, Zhao Y, Hirst M, Armisen J, Miska EA, Chin SF, Provenzano E, Turashvili G, Green A, Ellis I, Aparicio S, Caldas C: The shaping and functional consequences of the microRNA landscape in breast cancer. Nature. 2013, 497: 378-382. 10.1038/nature12108.View ArticleGoogle Scholar
- Farazi TA, Horlings HM, Ten Hoeve JJ, Mihailovic A, Halfwerk H, Morozov P, Brown M, Hafner M, Reyal F, van Kouwenhove M, Kreike B, Sie D, Hovestadt V, Wessels LF, van de Vijver MJ, Tuschl T: MicroRNA sequence and expression analysis in breast tumors by deep sequencing. Cancer Res. 2011, 71: 4443-4453. 10.1158/0008-5472.CAN-11-0608.View ArticleGoogle Scholar
- Lewis BP, Burge CB, Bartel DP: Conserved seed pairing, often flanked by adenosines, indicates that thousands of human genes are microRNA targets. Cell. 2005, 120: 15-20. 10.1016/j.cell.2004.12.035.View ArticleGoogle Scholar
- Hafner M, Landthaler M, Burger L, Khorshid M, Hausser J, Berninger P, Rothballer A, Ascano M, Jungkamp AC, Munschauer M, Ulrich A, Wardle GS, Dewell S, Zavolan M, Tuschl T: Transcriptome-wide identification of RNA-binding protein and microRNA target sites by PAR-CLIP. Cell. 2010, 141: 129-141. 10.1016/j.cell.2010.03.009.View ArticleGoogle Scholar
- MicroRNA target explorer for breast cancer data. [http://mp-pclip.nki.nl]
- Bartel DP: MicroRNAs: target recognition and regulatory functions. Cell. 2009, 136: 215-233. 10.1016/j.cell.2009.01.002.View ArticleGoogle Scholar
- Rajewsky N: microRNA target predictions in animals. Nat Genet. 2006, Suppl: S8-S13.View ArticleGoogle Scholar
- Alexiou P, Maragkakis M, Papadopoulos GL, Reczko M, Hatzigeorgiou AG: Lost in translation: an assessment and perspective for computational microRNA target identification. Bioinformatics. 2009, 25: 3049-3055. 10.1093/bioinformatics/btp565.View ArticleGoogle Scholar
- Betel D, Wilson M, Gabow A, Marks DS, Sander C: The microRNA.org resource: targets and expression. Nucleic Acids Res. 2008, 36: D149-D153.View ArticleGoogle Scholar
- Creighton CJ, Hernandez-Herrera A, Jacobsen A, Levine DA, Mankoo P, Schultz N, Du Y, Zhang Y, Larsson E, Sheridan R, Xiao W, Spellman PT, Getz G, Wheeler DA, Perou CM, Gibbs RA, Sander C, Hayes DN, Gunaratne PH: Integrated analyses of microRNAs demonstrate their widespread influence on gene expression in high-grade serous ovarian carcinoma. PLoS One. 2012, 7: e34546-10.1371/journal.pone.0034546.View ArticleGoogle Scholar
- Perou CM, Sorlie T, Eisen MB, van de Rijn M, Jeffrey SS, Rees CA, Pollack JR, Ross DT, Johnsen H, Akslen LA, Fluge O, Pergamenschikov A, Williams C, Zhu SX, Lonning PE, Borresen-Dale AL, Brown PO, Botstein D: Molecular portraits of human breast tumours. Nature. 2000, 406: 747-752. 10.1038/35021093.View ArticleGoogle Scholar
- Hu Z, Fan C, Oh DS, Marron JS, He X, Qaqish BF, Livasy C, Carey LA, Reynolds E, Dressler L, Nobel A, Parker J, Ewend MG, Sawyer LR, Wu J, Liu Y, Nanda R, Tretiakova M, Ruiz Orrico A, Dreher D, Palazzo JP, Perreard L, Nelson E, Mone M, Hansen H, Mullins M, Quackenbush JF, Ellis MJ, Olopade OI, Bernard PS, et al: The molecular portraits of breast tumors are conserved across microarray platforms. BMC Genomics. 2006, 7: 96-10.1186/1471-2164-7-96.View ArticleGoogle Scholar
- Cerami E, Gao J, Dogrusoz U, Gross BE, Sumer SO, Aksoy BA, Jacobsen A, Byrne CJ, Heuer ML, Larsson E, Antipin Y, Reva B, Goldberg AP, Sander C, Schultz N: The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov. 2012, 2: 401-404. 10.1158/2159-8290.CD-12-0095.View ArticleGoogle Scholar
- Kao J, Salari K, Bocanegra M, Choi YL, Girard L, Gandhi J, Kwei KA, Hernandez-Boussard T, Wang P, Gazdar AF, Minna JD, Pollack JR: Molecular profiling of breast cancer cell lines defines relevant tumor models and provides a resource for cancer gene discovery. PLoS One. 2009, 4: e6146-10.1371/journal.pone.0006146.View ArticleGoogle Scholar
- Lipchina I, Elkabetz Y, Hafner M, Sheridan R, Mihailovic A, Tuschl T, Sander C, Studer L, Betel D: Genome-wide identification of microRNA targets in human ES cells reveals a role for miR-302 in modulating BMP response. Genes Dev. 2011, 25: 2173-2186. 10.1101/gad.17221311.View ArticleGoogle Scholar
- Skalsky RL, Corcoran DL, Gottwein E, Frank CL, Kang D, Hafner M, Nusbaum JD, Feederle R, Delecluse HJ, Luftig MA, Tuschl T, Ohler U, Cullen BR: The viral and cellular microRNA targetome in lymphoblastoid cell lines. PLoS Pathog. 2012, 8: e1002484-10.1371/journal.ppat.1002484.View ArticleGoogle Scholar
- Wang Y, Juranek S, Li H, Sheng G, Tuschl T, Patel DJ: Structure of an argonaute silencing complex with a seed-containing guide DNA and target RNA duplex. Nature. 2008, 456: 921-926. 10.1038/nature07666.View ArticleGoogle Scholar
- Friedman RC, Farh KK, Burge CB, Bartel DP: Most mammalian mRNAs are conserved targets of microRNAs. Genome Res. 2009, 19: 92-105.View ArticleGoogle Scholar
- Helwak A, Kudla G, Dudnakova T, Tollervey D: Mapping the human miRNA interactome by CLASH reveals frequent noncanonical binding. Cell. 2013, 153: 654-665. 10.1016/j.cell.2013.03.043.View ArticleGoogle Scholar
- Wang ET, Sandberg R, Luo S, Khrebtukova I, Zhang L, Mayr C, Kingsmore SF, Schroth GP, Burge CB: Alternative isoform regulation in human tissue transcriptomes. Nature. 2008, 456: 470-476. 10.1038/nature07509.View ArticleGoogle Scholar
- Grimson A, Farh KK, Johnston WK, Garrett-Engele P, Lim LP, Bartel DP: MicroRNA targeting specificity in mammals: determinants beyond seed pairing. Mol Cell. 2007, 27: 91-105. 10.1016/j.molcel.2007.06.017.View ArticleGoogle Scholar
- Goeman JJ, van de Geer SA, de Kort F, van Houwelingen HC: A global test for groups of genes: testing association with a clinical outcome. Bioinformatics. 2004, 20: 93-99. 10.1093/bioinformatics/btg382.View ArticleGoogle Scholar
- Aerts S, Lambrechts D, Maity S, Van Loo P, Coessens B, De Smet F, Tranchevent LC, De Moor B, Marynen P, Hassan B, Carmeliet P, Moreau Y: Gene prioritization through genomic data fusion. Nat Biotechnol. 2006, 24: 537-544. 10.1038/nbt1203.View ArticleGoogle Scholar
- Olive V, Jiang I, He L: mir-17-92, a cluster of miRNAs in the midst of the cancer network. Int J Biochem Cell Biol. 2010, 42: 1348-1354. 10.1016/j.biocel.2010.03.004.View ArticleGoogle Scholar
- Brabletz S, Brabletz T: The ZEB/miR-200 feedback loop–a motor of cellular plasticity in development and cancer?. EMBO Rep. 2010, 11: 670-677. 10.1038/embor.2010.117.View ArticleGoogle Scholar
- Chen Y, Gorski DH: Regulation of angiogenesis through a microRNA (miR-130a) that down-regulates antiangiogenic homeobox genes GAX and HOXA5. Blood. 2008, 111: 1217-1226.View ArticleGoogle Scholar
- DeCastro AJ, Dunphy KA, Hutchinson J, Balboni AL, Cherukuri P, Jerry DJ, DiRenzo J: MiR203 mediates subversion of stem cell properties during mammary epithelial differentiation via repression of DeltaNP63alpha and promotes mesenchymal-to-epithelial transition. Cell Death Dis. 2013, 4: e514-10.1038/cddis.2013.37.View ArticleGoogle Scholar
- Lei R, Tang J, Zhuang X, Deng R, Li G, Yu J, Liang Y, Xiao J, Wang HY, Yang Q, Hu G: Suppression of MIM by microRNA-182 activates RhoA and promotes breast cancer metastasis. Oncogene. 2013, doi:10.1038/onc.2013.65Google Scholar
- Gyorffy B, Schafer R: Meta-analysis of gene expression profiles related to relapse-free survival in 1,079 breast cancer patients. Breast Cancer Res Treat. 2009, 118: 433-441. 10.1007/s10549-008-0242-8.View ArticleGoogle Scholar
- Staiger C, Cadot S, Gyorffy B, Wessels LF, Klau W: Current composite-feature classification methods do not outperform simple single-genes classifiers in breast cancer prognosis. Frontiers in Genetics. 2013, 4: 289-View ArticleGoogle Scholar
- Krutzfeldt J, Rajewsky N, Braich R, Rajeev KG, Tuschl T, Manoharan M, Stoffel M: Silencing of microRNAs in vivo with 'antagomirs'. Nature. 2005, 438: 685-689. 10.1038/nature04303.View ArticleGoogle Scholar
- Baek D, Villen J, Shin C, Camargo FD, Gygi SP, Bartel DP: The impact of microRNAs on protein output. Nature. 2008, 455: 64-71. 10.1038/nature07242.View ArticleGoogle Scholar
- Selbach M, Schwanhausser B, Thierfelder N, Fang Z, Khanin R, Rajewsky N: Widespread changes in protein synthesis induced by microRNAs. Nature. 2008, 455: 58-63. 10.1038/nature07228.View ArticleGoogle Scholar
- Linsley PS, Schelter J, Burchard J, Kibukawa M, Martin MM, Bartz SR, Johnson JM, Cummins JM, Raymond CK, Dai H, Chau N, Cleary M, Jackson AL, Carleton M, Lim L: Transcripts targeted by the microRNA-16 family cooperatively regulate cell cycle progression. Mol Cell Biol. 2007, 27: 2240-2252. 10.1128/MCB.02005-06.View ArticleGoogle Scholar
- Androsavich JR, Chau BN, Bhat B, Linsley PS, Walter NG: Disease-linked microRNA-21 exhibits drastically reduced mRNA binding and silencing activity in healthy mouse liver. RNA. 2012, 18: 1510-1526. 10.1261/rna.033308.112.View ArticleGoogle Scholar
- Wendt MK, Cooper AN, Dwinell MB: Epigenetic silencing of CXCL12 increases the metastatic potential of mammary carcinoma cells. Oncogene. 2008, 27: 1461-1471. 10.1038/sj.onc.1210751.View ArticleGoogle Scholar
- Pecot CV, Rupaimoole R, Yang D, Akbani R, Ivan C, Lu C, Wu S, Han HD, Shah MY, Rodriguez-Aguayo C, Bottsford-Miller J, Liu Y, Kim SB, Unruh A, Gonzalez-Villasana V, Huang L, Zand B, Moreno-Smith M, Mangala LS, Taylor M, Dalton HJ, Sehgal V, Wen Y, Kang Y, Baggerly KA, Lee JS, Ram PT, Ravoori MK, Kundra V, Zhang X, et al: Tumour angiogenesis regulation by the miR-200 family. Nat Commun. 2013, 4: 2427-View ArticleGoogle Scholar
- Song SJ, Ito K, Ala U, Kats L, Webster K, Sun SM, Jongen-Lavrencic M, Manova-Todorova K, Teruya-Feldstein J, Avigan DE, Delwel R, Pandolfi PP: The oncogenic microRNA miR-22 targets the TET2 tumor suppressor to promote hematopoietic stem cell self-renewal and transformation. Cell Stem Cell. 2013, 13: 87-101. 10.1016/j.stem.2013.06.003.View ArticleGoogle Scholar
- BioPortal for Cancer Genomics. [http://www.cbioportal.org/public-portal/]
- Computational Cancer Biology: Division of Molecular Carcinogenesis, Netherlands Cancer Institute. [http://ccb.nki.nl/data.php]
- Radfar M, Wong W, Morris QD: Predicting the target genes of intronic microRNAs using large-scale gene expression data. Conf Proc IEEE Eng Med Biol Soc. 2010, 2010: 791-794.Google Scholar
- TargetScan Human: Prediction of microRNA targets. [http://www.targetscan.org]
- microRNA.org – Targets and Expession. [http://microRNA.org]
- COSMIC: Catalogue of somatic mutations in cancer. [http://www.sanger.ac.uk/genetics/CGP/Census/]
- Baskerville S, Bartel DP: Microarray profiling of microRNAs reveals frequent coexpression with neighboring miRNAs and host genes. RNA. 2005, 11: 241-247. 10.1261/rna.7240905.View ArticleGoogle Scholar
- The Sequence Read Archive (SRA). http://www.ncbi.nlm.nih.gov/sra/?term=SRX388831,
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.