TraceRNA: A Web Application for ceRNAs Exploration

Background

Gene expression silencing at mRNA level by microRNAs (miRNA) is a well-established form of post-transcriptional regulation^{1, 2}. Such silencing is achieved through miRNA binding to miRNA response elements (MRE) residing mainly in the 3’ untranslated regions (UTRs) of the target mRNA. Over 1000 human miRNAs have been identified³, and the prevalence of miRNA regulation in a broad range of biological processes and disease often attributes to the fact that a single miRNA can repress hundreds of different mRNAs⁴. Interestingly, a single target mRNA often possesses MREs of distinct miRNAs in its 3’UTR⁵. Questions have been raised regarding the need for this redundancy in regulation and these multiple MREs were once thought to serve as regulatory buffers of different miRNAs. In a recent seminal work, a novel theory, termed the competing endogenous RNA (ceRNA), was proposed to provide a plausible explanation for this interesting phenomenon from a new perspective of gene regulation⁶. According to the ceRNA theory, MREs function as “letters” of this new regulatory system, and ceRNAs, or sets of RNAs including mRNA, pseudogenes, and long noncoding RNAs, can communicate, or regulate each other, through competition for common MREs. As such, ceRNA regulatory networks provide a unifying system for regulations among transcriptome-wide RNAs, greatly expanding the functions of RNAs⁶. Alteration of this competition between ceRNAs could modify normal state gene expression and in return alter the status of biological pathways to promote an oncogenic program for example. To that end, a PTEN ceRNA network was uncovered and shown to potentially regulate oncogenesis⁶.

The fact that this new level of RNA regulations could be prevalent in cells has prompted research to identify ceRNAs of genes related to disease. However, the complexity of ceRNA regulations and an incomplete knowledge of miRNA binding have hampered the prediction of ceRNAs, which often requires the use of computational tools and databases that are not readily available to the users. Thus far, two algorithms for human ceRNA predictions have been proposed. MuTaMe, proposed in⁶, aims to predict ceRNAs of a GOI. It starts by selecting a set of, ideally experimentally validated, miRNAs that target the given GOI in its 3’ UTR region. Predicted ceRNAs by sequence-pairing are the mRNAs that are also targeted by these miRNAs and the prediction is made based on scores generated from binding affinity statistics. While MuTaMe succeeded in predicting several ceRNAs of PTEN, it is not accessible for predicting other GOIs because experimentally validated miRNAs targeting a new GOI are mostly unavailable and binding affinity statistics used in MuTaMe are insufficient for accurate predictions. Furthermore, MuTaMe has not been implemented as a software tool yet and cannot be accessed by the general public. The second algorithm, Hermes, proposed in⁷, infers ceRNAs from expression profiles of genes and miRNAs by using conditional mutual information (CMI). While Hermes combines ceRNA/miRNA/target triplets via tissue specific gene expression, however, it does not provide an implementation that combines sequence-binding statistics with gene expression. There is a shortage of user-friendly tools that can be easily used for anyone interested in ceRNA research.

To address the need for user-friendly tools, we developed here TraceRNA, a web-based application for transcriptome-wide ceRNA discovery. TraceRNA is flexible, powerful, and user-friendly. It includes MiRTarBase⁸, a database of experimentally validated miRNA-target pairs, and miRNA binding scores and related data (sites position, length, etc.) from three prediction algorithms (SVMicrO⁹, BCMicrO¹⁰, and SiteTest) with different emphasis. TraceRNA provides the user with the flexibility to perform ceRNA predictions using one of three algorithms to meet different study objectives. Currently, TraceRNA maintains a database that includes genome-wide targets of >700 human miRNAs predicted by three algorithms. The user can compare among the prediction results from these different algorithms to either complement or reach a consensual prediction.

Two important observations have been integrated into the TraceRNA for context-specific ceRNA discovery. The first is that the miRNA expression is condition-specific. That is, if a miRNA is not expressed in a tissue environment or disease state, one can ignore its target-binding specificity. The second is that, GOI and its ceRNAs’ expressions are positively correlated because of the competition for miRNA binding. Therefore, an increased/decreased GOI expression will attract more/less miRNA binding away its ceRNAs, resulting in increased/decreased ceRNA expression level due to the decreased/increased repression effect of miRNAs. As another unique feature, TraceRNA can construct ceRNA interaction networks to help delineate complex interactions of ceRNAs and gain further insight into this novel ceRNA regulation-modulation mechanism. Finally, TraceRNA is developed to be user-friendly web application with an accessible interface. It generates predictions including statistics such as p-values and false discovery rate (FDR) both online and in spreadsheets available for download.

Methods

The goal of TraceRNA is to predict ceRNAs of a GOI, which are mRNAs that share MREs from a set of miRNAs that also target the GOI. In this paper, we named these miRNAs as GOI targeting miRNAs, or GTmiRs. ceRNAs’ competition for GTmiRs binding to GOI will alter the expression of GOI and its ceRNAs in a coordinated fashion and co-expression can be observed, where expressions of GOI and its ceRNAs are expected to be correlated. Predictions of ceRNAs can be done by examining miRNA:mRNA sequence pairing and/or GOI-ceRNAs co-expression. TraceRNA includes three main processing sections in its pipeline (Fig. 1): 1) sequence-based prediction of ceRNAs, 2) co-expression analysis of GOI and ceRNAs’ expression levels, and 3) generation of ceRNA regulatory network. Additionally, miRNA expression data is also included in TraceRNA for the user to select context-specific GTmiRs (Supplemental Fig. S1).

Figure 1.

TraceRNA Pipeline. The user initiates TraceRNA predictions with a gene of interest (GOI). Experimentally validated miRNAs and/or predicted miRNAs that target this GOI can be selected as GTmiRs. These GTmiRs are then fed to one of the three sequence-level target prediction algorithms (SVMicrO, SiteTest or BCmicrO) to generate a list of predicted ceRNAs by sequence-pairing together with the p-values and FDRs (Section 1). In addition, the user can select one of the provided expression sets (GMB and Breast Cancer datasets) to evaluate the expression correlation between the predicted ceRNAs by sequence-pairing and the GOI under the specific tissue/tumor condition and obtain predicted ceRNAs by co-expression test. Multiple prediction scores are consolidated with Borda method (Section 2). To generate a ceRNA network, top 20 ceRNAs will be selected as a set of the new GOIs, or cGOIs, each then subject to a round of new predictions to obtain their corresponding ceRNAs or cGOI-ceRNAs pairs. All resulting GOI-ceRNAs and cGOI-ceRNAs pairs with their scores are used to generate a ceRNA mediated regulatory network (Section 3).

Sequence-based prediction of ceRNAs

Selection of GOI targeting miRNAs (GTmiRs)

Given a GOI provided by the user, the first step in TraceRNA is to identify GTmiRs, or miRNAs that target GOI. TraceRNA provides two alternatives for GTmiRs identification (Fig. 1). First, TraceRNA maintains a local copy of experimentally validated miRNAs: targets pairs curated by miRTarBase Release 2.5 (downloaded on July 2012). Second, genome-wide SVMicrO⁹ predictions for >700 miRNAs were pre-calculated.. SVMicrO⁹ was developed previously to predict miRNA targets. It uses a support vector machine with sequence-based features including binding secondary structure, energy, binding conservation, number of predicted sites, and site densities. SVMicrO was tested to achieve improved performance compared to several popular algorithms including TargetScan, miRanda, Pictar, etc. The predicted miRNAs are displayed to the user in descending order of p-values (See Section 1.3). The user may select a subset or all of the miRNAs from these two sources as GTmiRs.

Prediction of ceRNAs

Once GTmiRs are selected, TraceRNA predicts ceRNAs as the mRNAs that are also targeted by these GTmiRs, by using one of three miRNA target prediction algorithms: 1) SVMicrO⁹, 2) BCMicrO¹⁰, or 3) SiteTest, depending on the user’s selection. SVMicrO⁹ and BCMicrO¹⁰ are two in-house developed algorithms which were published previously. As discussed above, SVMicrO makes predictions by utilizing a large number of miRNA binding features. BCMicrO uses a Bayesian approach that integrates prediction scores from 6 popular algorithms: TargetScan¹¹, miRanda¹², PicTar¹³, mirTarget2¹⁴, PITA¹⁵, and DIANA micro-T¹⁶. Both algorithms provide more accurate predictions than existing algorithms. The prediction scores of SVMicrO and BCMicrO were pre-calculated and stored in a MySQL database. In addition, a new algorithm, SiteTest, inspired by MuTaMe⁶, was also developed and its pseudo code is included in Supplemental Materials.

In order to show the scores calculation, let S_i be the score of GTmiR i targeting an mRNA by either algorithms and K be the total number of GTmiRs. Then, the score, S, for the mRNA to be a ceRNA predicted by sequence-pairing is calculated as

$$S=\frac{1}{K}{\sum }_{i=1}^K{S}_i.$$ (1)

We discuss next the calculation of the predictions significance.

Statistical Significance of Predicted ceRNAs

We first discuss the calculation of statistical significance for the SVMicrO scores. According to (1), S is calculated as the average of the sequence-pairing scores of each GTmiR and the mRNA. To calculate the p-value for S, the distribution of S under the null hypothesis, i.e., the mRNA is not predicted by sequence-pairing as a ceRNA, needs to be obtained. Because S is the average S_i, then the distribution of S_i under the null hypothesis needs to be evaluated first. Adopting the method developed in BCMicrO, the empirical distribution S_i under the null hypothesis was observed as a mixture of two distributions, one clustered around smaller scores and the other around larger scores (Fig. 2). Given that most genes are not miRNA targets and they should have smaller SVMicrO scores, we hypothesized that the distribution around smaller scores characterizes the scores derived from genes not targeted by any miRNA, which was further assumed to follow the i.i.d. Gamma distribution, or S_i~ Gamma(α,β) whose parameters α and β were obtained from fitting the empirical scores S_i (Supplemental Figure S1). Subsequently, due to (1), S is also Gamma-distributed under the null hypothesis as:

$$S~\mathit{\text{Gamma}}(K\alpha ,\beta ).$$ (2)

Therefore, the probability (p-value) of a sequence-pairing prediction score S can be evaluated analytically by (2). The same method was applied to BCMicrO and SiteTest by fitting the Gamma distributions directly to their scores. Once p-values of all predicted ceRNAs by sequence-pairing are calculated, the corresponding False Discovery Rates (FDRs) are computed using the Benjamini-Hochberg method¹⁷.

Figure 2.

Illustration of p-value calculation for ceRNA prediction scores. (a) Histogram of SVMicrO scores for genome-wide targets of 500 miRNAs. (b) Zoom-in view of the histogram in (a) and the fitted (shifted) Gamma distribution (solid line). The parameters of the fitted Gamma distribution are α=0.7234 and β=0.3594.

Co-expression based prediction of ceRNAs

Test for co-expression between GOI and predicted ceRNAs by sequence-pairing

TraceRNA can also integrate a tissue or disease specific expression dataset to predict tissue or disease specific ceRNAs of the GOI and potentially further improve the prediction specificity (Fig. 1). Currently, expression datasets of glioblastoma multiforme (GBM)¹⁸ and Breast Cancer¹⁹ from TCGA (http://cancergenome.nih.gov/) are included. The users may contact the web-master to upload their own expression datasets if needed. Because higher GOI expression competitively attracts more miRNA binding and thus reduces the possibility of the same miRNA binding to ceRNAs, leading to higher ceRNA expression, the co-expression analysis first computes the Pearson correlation coefficients between GOI expression levels and predicted ceRNAs by sequence-pairing and then removes the mRNAs with negative correlation coefficients. The p-values were calculated by Fisher transformation²⁰ and the resultant predictions have two scores: those by sequence-pairing and those by co-expression test. We discuss their consolidation in the next section.

Score consolidation

To fuse these two scores, we utilized the Borda counting method²¹, which essentially sums ranks of scores. The resultant ceRNAs list can be downloaded from the website as a common delimited text file that contains the gene symbols ordered based on the Borda scores from highest to lowest, their sequence-pairing scores, co-expression scores, and their rankings.

Generation of regulatory network based on a GOI

TraceRNA also aims to provide a tool that allows biologists to discover new regulatory networks that are potentially modulated by a set of GTmiRs and gain insight into this novel gene regulation-modulation mechanism. To generate a GOI-ceRNA regulatory network, the user can select top predicted ceRNAs by co-expression test for a given GOI and then treat each predicted ceRNA as a new GOI (or cGOI). TraceRNA performs new rounds of predictions for each cGOI iteratively using the same number of predicted miRNAs that target each cGOI as described before. The resulting list (containing GOI, ceRNAs, and scores for all cGOIs) is used to generate a regulation network using Cytoscape plug-in²² and it is saved as a file that can be downloaded for further analysis.

Biological Functional Enrichment

To examine the functional association of ceRNAs for a given GOI, we used DAVID²³ (http://david.abcc.ncifcrf.gov/), which uses a modified Fisher’s exact test to evaluate the functional enrichment of 40 annotation categories, including GO terms, protein-protein interactions, disease associations, pathways, homologies and other gene sets in a given gene list. In this paper, the enrichment results for p-value < 0.01 are reported.

Final remarks on methods

Discussion of TraceRNA implementation can be found in Supplemental Material. Table 1 summarizes the algorithms and databases used in TraceRNA. MiRTarBase, SVMicrO, and BCMicrO were implemented as databases queried by SQL commands. SiteTest accesses SVMicrO database and calculates binding scores for each ceRNA. All the computations including statistical significance and Borda fusion were implemented by R (http://www.r-project.org/).

Table 1.

Summary of TraceRNA Algorithms

Algorithm or Database	Output	Function of Algorithm
miRTarBase	Experimentally verified miRNA:target interactions	Identify validated miRNAs
SVMicrO	Scores of miRNA-gene binding prediction by using a large number of binding site and 3’UTR features	Predict miRNAs binding
BCMicrO	Scores of miRNAs-genes binding predictions by fusion of 6 algorithms
SiteTest	Scores of miRNAs-genes binding predictions using 3’UTR features
ceRNA-SVMicrO ceRNA-BCMicrO ceRNA-SiteTest	Scores for predicted ceRNAs by sequence-pairing	Predict ceRNAs by sequence-pairing
Expression Correlation	Pearson correlation of GOI and predicted ceRNAs by sequence-pairing	Calculate correlation between predicted ceRNAs by sequence-pairing
SeqExp Fusion (Borda merging method)	Scores from fusing sequence-level predictions and expression correlation	Fuse sequence-level predictions and co-expression
ceRNA-Net	List of gene pairs (nodes) and scores of their directed interactions	Discover regulatory networks

Results and Case Studies

TraceRNA integrates databases, sql queries, real-time predictions, and generation of ceRNA interaction network under a unified web interface, enabling ceRNA predictions and discovery of novel biological regulation. We illustrate its features and capabilities next.

TraceRNA web interface

The TraceRNA web interface (Fig. S1), starts with a query GOI by the user. Currently, TraceRNA only supports official gene symbols from the UCSC annotation. Given a GOI, a set of validated miRNAs that target the GOI derived from miRTarBase will be displayed under the checker box "Select validated miRNAs." Other miRNAs predicted by SVMicrO are listed under "Select Predicted miRNAs" in an increasing order of binding p-values of targeting GOI. The user can select from these two sources a set of miRNAs to form GTmiRs. A rule of thumb: one can require binding p-value<0.01, which is expected to produce less than 43 miRNAs for 50% of genes (Fig. S2), or miRNAs log2 expression level in GBM > 6 (Fig. S3).

After selection, the user can choose from SVMicrO, BCMicrO, or SiteTest and further integrate gene expression data. To evaluate ceRNA-mediated gene-gene interactions, the user can perform ceRNA prediction iteratively by treating top K (20 by default) ceRNAs as GOI, (cGOI). The resulting interactions will be displayed within the web-interface and can be also saved in a file to be imported into Cytoscape.

GTmiR determination

Determination of GTmiRs that target the GOI is an important step that can significantly affect the final prediction. GTmiRs are ideally determined by experiments, but the complete set of experimentally validated GTmiRs are rarely available. ceRNA predictions based on a subset of validated GTmiRs will have low specificity and predicted GTmiRs are needed to increase the specificity. However, high false positives associated with current miRNA target prediction algorithms could introduce false positives in ceRNA predictions, thus potentially harming rather than improving the ceRNA predictions specificity. Furthermore, a different number of candidate GTmiRs can also affect the prediction performance, where a lower number will likely produce lower prediction specificity, whereas a number too high can, on the contrary, harm the prediction sensitivity. This observation was captured in Supplemental Fig. S4, in which we varied the number of miRNAs from 2 to 70. It clearly demonstrated the low specificity with a small number of miRNAs and a loss of sensitivity when too many miRNAs were selected. Therefore, care needs to be taken in choosing GTmiRs. To this end, TraceRNA provides flexibility to choose between validated and predicted GTmiRs or a combination of both.

As an example, Table 2 includes the top 20 experimentally validated (from miRTarBase) and predicted GTmiRs for PTEN, ESR1, and BRCA1, respectively. In all three cases, the numbers of experimentally validated GTmiRs are less than 20, and there are only 4 for BRCA1. Apparently, using the validated GTmiRs alone will result in low specificity in ceRNA predictions.

Table 2.

Top 20 experimentally validated and predicted GTmiRs for PTEN, ESR1 and BRCA1.

PTEN		ESR1		BRCA1
20 predicted by SVMicrO	miRTarBase	20 predicted by SVMicrO	miRTarBase	20 predicted by SVMicrO	miRTarBase
miR-16	miR-106b	miR-518c	miR-18a	miR-1915	miR-15a
miR-512-3p	miR-141	miR-558	miR-18b	miR-588	miR-16
miR-30c-1*	miR-17	miR-1914	miR-193b	miR-324-3p	miR-212
miR-142-5p	miR-18a	miR-1305	miR-19a	miR-29b-2*	miR-24
miR-323-3p	miR-19a	miR-551b*	miR-19b	miR-548p
miR-548c-3p	miR-19b	miR-934	miR-206	miR-1304
miR-143	miR-20a	miR-509-5p	miR-20b	miR-1913
miR-580	miR-21	miR-142-3p	miR-22	miR-345
miR-26a	miR-214	miR-1233	miR-221	miR-545
miR-510	miR-216a	miR-1915	miR-222	miR-498
miR-212	miR-217	miR-582-3p	miR-29b	miR-146a
miR-26b	miR-221	miR-590-3p	miR-302c	miR-30e*
miR-194	miR-222	miR-302a*		let_7g*
miR-202*	miR-26a	miR-605		miR-1914
miR-494	miR-494	miR-647		miR-1910
miR-1253		miR-1913		miR-760
miR-340		miR-34a*		miR-615-5p
miR-410		miR-362-3p		miR-9
miR-335*		miR-629		miR-516a-3p
miR-513a-3p		miR-34b		miR-516b*

Significance of ceRNA Prediction Score

Fig. 2-(a) depicts the empirical distribution of the genome-wide SVMicrO scores for 772 human miRNAs and a mixture of two distributions can be clearly observed, one clustered around smaller scores with a much larger mass and the other around larger scores. As discussed in the Methods section, the peak around smaller scores was considered to represent the null distribution and was fitted with a Gamma distribution, whose parameters are α=0.7234 and β=0.3594 with 95% confidence interval (0.7229, 0.7239) and (0.3591, 0.3598) for α and β, respectively. Fig. 2-(b) shows the histogram and the fitted Gamma distribution (with a constant shift). Table S1 lists the fitted parameters of the Gamma distributions for SVMicrO, BCMicrO and SiteTest.

Case Study 1

We applied TraceRNA to predict the ceRNAs of PTEN. PTEN is a gene related to the development of many cancers, where it often functions as a key tumor suppressor, whose abundance determines the critical outcomes in tumorigenesis⁶. PTEN is also known to regulate cell cycle, particularly in preventing cells from growing and dividing too rapidly. PTEN ceRNAs have also been predicted and reported^{6, 7}.

To predict ceRNAs of PTEN with TraceRNA, we selected only the predicted miRNAs as GTmiRs (Table 2) and further chose SVMicrO to predict the ceRNAs by sequence-pairing. TraceRNA returned a total of 761 predicted ceRNAs by sequence-pairing for p-value < 0.05, and the 20 best predictions (in descending order of prediction score) together with p-values are provided in Table 3a. One important feature of TraceRNA is the possibility to increase the predictions specificity and predict context-specific ceRNAs by integrating an expression dataset of a disease condition. In this case, 400 GBM expression samples from TCGA project were included, based on which co-expression correlations against PTEN were evaluated and a total of 466 genes were obtained for Pearson correlation greater than zero as the GBM-specific ceRNAs of PTEN. Table 3b shows the results for the top 20 predicted ceRNAs based on Borda score. To examine their functions, pathway enrichment was performed using DAVID²³ on these 466 ceRNAs, and the 3 enriched pathways are shown in Table 4 (under TraceRNA+GBM). Well known cancer related pathways including MAPK signaling and WNT signaling were significantly enriched, indicating an important involvement of PTEN ceRNAs in cancer. Another enriched pathway, TGF- signaling, is known to utilize intracellular SMADs to mediate growth suppression and PTEN down-regulation simultaneously to induce growth proliferation. Here, the prediction result provided a third possible regulatory mechanism of TGF regulation by PTEN via its ceRNAs.

Table 3.

Predicted ceRNAs of PTEN

A: TraceRNA (sequence-based)			B:TraceRNA (w/ GBM Gene Expression)				C: ceRDB
Symbol	SVMicrO score	p-value	Symbol	SVMicrO score	p-value	Correlation coefficient	Symbol	Score
QKI	1.199	0.0031	GSPT1	0.9105	0.0065	0.5534	TNRC6B	26
CPEB2	1.099	0.0040	CD2AP	0.8590	0.0074	0.5632	TET3	25
OTUD4	1.079	0.0042	USP15	1.0550	0.0045	0.5167	NFIB	23
PPP6C	1.071	0.0043	PPP6C	1.0706	0.0043	0.4913	ATXN1	21
NOVA1	1.059	0.0044	RAD23B	0.9124	0.0065	0.5167	ANKRD52	20
USP15	1.055	0.0045	NIPBL	0.7567	0.0097	0.6163	tcag7.1228	20
LEMD3	1.031	0.0048	CNOT6	0.8730	0.0072	0.4888	NUFIP2	17
SUZ12	1.027	0.0048	ARF6	0.7706	0.0093	0.5561	CPEB4	16
NARG1	1.022	0.0049	MED14	0.8159	0.0083	0.5100	ZFHX4	16
ZFHX3	1.014	0.0050	KIAA0240	0.9639	0.0057	0.4572	BACH2	15
C18orf25	1.009	0.0050	SOCS5	0.7573	0.0096	0.5438	CLCN5	15
FNDC3B	1.007	0.0051	CSNK1G3	0.7361	0.0102	0.5675	CUGBP2	15
BCL11A	0.967	0.0056	NAB1	0.7867	0.0089	0.4986	ONECUT2	14
ST18	0.967	0.0056	CBFB	0.7516	0.0098	0.5273	AFF1	14
KIAA0240	0.964	0.0057	RPS6KB1	0.7437	0.0100	0.5387	BNC2	14
RSBN1	0.963	0.0057	RPS6KA3	0.7126	0.0108	0.5780	KLF12	14
TET2	0.961	0.0057	TBL1XR1	0.7553	0.0097	0.5019	NFAT5	14
ICK	0.952	0.0058	RBM12	0.8482	0.0076	0.4516	OTUD4	14
SCN2A	0.951	0.0058	ELL2	0.8582	0.0074	0.4472	ZBTB34	14
PAN3	0.925	0.0062	ZFX	0.7595	0.0096	0.4752	ZEB2	14

(a) Best 20 predicted ceRNAs by sequence-pairingof PTEN predicted by SVMicrO based on miRNAs in Table 2 (first column, top 20 predicted miRNAs).

(b) Top 20 predicted ceRNAs by co-expression testof PTEN under the context of GBM. Column “p-value” shows the prediction p-values obtained in (a). Column “Correlation coefficient” lists Pearson correlations between a ceRNA and PTEN calculated using GBM gene expression data (see Materials and Methods). The ranking is based on the Borda merging of the rankings in (a) and that based on the Pearson correlation. ceRNAs shared between Tables 3a and 3b are in b font.

(c)The ceRNAs obtained from ceRDB. ceRNAs that were predicted by both algorithms (within top 50) are in italic. Note: tcag7.1228 is a synonym of UBN2.

Table 4.

Enriched pathways of the predicted PTENceRNAs

KEGG Pathways	TraceRNA	TraceRNA + GBM	ceRDB
Long-term potentiation	0.00002	8.8×10−8	0.00844
Oocyte meiosis	0.00066	1.9×10−6
MAPK signaling pathway	0.00239	0.00096
Wnt signaling pathway	0.00496	0.00005
Neurotrophin signaling pathway		0.00019
Axon guidance		0.00460	0.00998
Insulin signaling pathway		0.00073
TGF-beta signaling pathway		0.00036	0.00543
ErbB signaling pathway			0.00110
Endocytosis			0.00210
Ubiquitin mediated proteolysis		0.00180
Focal adhesion		0.00074
Melanogenesis		0.00390

The table includes the top enriched (p-values <0.01) pathways for three ceRNA prediction results. Enrichment analysis was performed by DAVID with entire human genome as the background gene set.

It would be also interesting to examine if these GBM specific ceRNAs by co-expression test also have higher specificity than those by sequence-pairing alone (Table 3a). However, direct comparison was infeasible due to a lack of true PTEN ceRNAs. Alternatively, pathway enrichment of the predictions result was conducted to make an indirect comparison. Intuitively, true ceRNAs should be functionally more significant than false positive predictions, and therefore, the predictions with higher specificity should be accompanied by a larger number of more enriched pathways. Pathway enrichment of sequence-based predictions is shown in Table 4 (under TraceRNA) and it is apparent that the GBM-specific ceRNAs by co-expression test are of higher functional enrichment (9 enriched functions vs 4 weakly enriched functions in sequence-pairing predictions alone), thus a higher predictions specificity. Expression (GBM data) scatter-plots of PTEN vs. three predicted ceRNAs (QKI, NOVA1, and BCL11A) by sequence predictions (Table 3a) are shown in Fig. 3(a)–(c). The correlations are clearly very low, suggesting that they are not GBM-specific ceRNAs. As expected, they were not among the predicted GBM-specific ceRNAs (Table 3b). Expression scatter-plots of PTEN vs. the 3 predicted GBM-specific ceRNAs (GSPT1, PPP6C, and USP15; Table 3b) are shown in Fig. 3(d)–(f). Their correlations are much higher. Notice that USP15 was also ranked No. 6 in sequence-based predictions. As expected, its ranking improved after integrating gene expression data.

Figure 3.

Expression scatter plots of PTEN vs. top predicted ceRNAs in GBM. (a)–(c) The three genes were top ranked ceRNAs predicted by sequence-pairing listed in Table 3a. However, their gene expression correlations to GOI (PTEN) are low and insignificant. Consequently, they were not among the top ranked ceRNAs after considering co-expression with PTEN (see Table 3b). (d)–(e) Three top ranked GBM-specific ceRNAs predicted by sequence-pairing and co-expression test (Table 3b). Clearly, their higher correlation coefficients enhance their prediction p-values.

As a comparison, ceRNAs predicted by Competitive Endogenous mRNA DataBase (ceRDB)²⁴ were also retrieved and top 20 are listed in Table 3c. We observed only 1 overlaps between ceRDB and TraceRNA predictions in the top 20 predictions. To examine the functional significance of the ceRDB predictions, pathway enrichment was conducted (Table 4, column ceRDB). 9 pathways in TraceRNA+GBM are enriched compared to 5 pathways in ceRDB. For example, enrichment p-values of “Long-term Potentiation” that includes genes such as MAPK1, NRAS, RPS6KA3, KRAS, and CREBBP, are 8.8×10⁻⁸, 0.00002 and 0.0084 for TraceRNA+GBM, TraceRNA sequence-pairing alone, and ceRDB, respectively.

Case Study 2

Breast cancer (BC) is a common disease in women and its incidence is still increasing²⁵ despite great improvement in therapies and earlier screening²⁶. Hormone receptor (such as estrogen receptor, ER) ERpositive breast cancer and ERBB2-positive breast cancer (about 50% co-expressed with ER+ tumors) currently account for about 75% and 15% of all breast cancer cases, respectively. The remaining 10% are so-called triple-negative breast cancers (TNBC), as defined by absent expression of ER, progesterone receptor (PR) and ERBB2 proteins^{27, 28}. In this study, our objective is to identify genes mediated by the estrogen receptor alpha, ESR1, through ceRNA regulatory network in breast cancer. To this end, we selected ESR1 as the GOI, and then predicted miRNAs as GTmiRs (Table 2, column 2). A total of 730 predicted ceRNAs by sequence-pairing were obtained by SVMicrO at p-value<0.05. Top 20 predictions were provided in Table 5a. Predicted BC-specific ESR1 ceRNAs by co-expression test were subsequently obtained by including TCGA gene expression of 590 breast cancer tumor samples (described in Materials and Methods). A total of 378 BC-specific ceRNAs were obtained and top 20 are shown in Table 5b.

Table 5.

Predicted ceRNAs of ESR1

A. TraceRNA (sequence-based)			B. TraceRNA (w/ BRCA Gene Expression)
Symbol	SVMicrO Score	p-value	Symbol	SVMicrO score	p-value	Correlation
RUNX1	0.9158	0.0064	FOXP1	0.6874	0.0116	0.5082
MAP2	0.8791	0.0070	NRIP1	0.6654	0.0122	0.5419
MIER3	0.8748	0.0071	MIER3	0.8748	0.0071	0.3434
CTDSPL2	0.8649	0.0073	CPEB3	0.6562	0.0125	0.4838
LOC440354	0.7936	0.0088	ZBTB4	0.6587	0.0124	0.4226
LOC595101	0.7923	0.0088	ARL3	0.5801	0.0153	0.5627
DDX3Y	0.7771	0.0092	RUNX1	0.9158	0.0064	0.2674
GUCY1B2	0.7674	0.0094	FCHO2	0.5846	0.0151	0.3999
RNF44	0.7628	0.0095	NOVA1	0.5455	0.0167	0.4788
CTTNBP2NL	0.7586	0.0096	BSN	0.5917	0.0148	0.3294
C20orf112	0.7572	0.0097	OTUD7B	0.6230	0.0137	0.2962
WAPAL	0.7551	0.0097	MAX	0.5842	0.0151	0.3260
SKI	0.7352	0.0102	ALCAM	0.5379	0.0170	0.4151
E2F3	0.7311	0.0103	CTDSPL	0.5544	0.0163	0.3245
NIPBL	0.7307	0.0103	SP1	0.6709	0.0121	0.1925
UBE2Z	0.7168	0.0107	KCNMA1	0.5680	0.0157	0.2949
G3BP2	0.7061	0.0110	STRBP	0.5903	0.0149	0.2605
MYT1L	0.7050	0.0110	BTRC	0.4881	0.0194	0.5566
DDX3X	0.7029	0.0111	MYST3	0.6165	0.0139	0.1916
KIAA0240	0.6981	0.0112	POMGNT1	0.5768	0.0154	0.2208

(a) Best 20 predicted ceRNAs by sequence-pairing of ESR1 predicted by SVMicrO alone using top 20 miRNAs in Table 3b. ceRNAs are sorted by SVMscore.

(b) BRCA-specific ceRNAs obtained by integrating predictions in Table 5a) with TCGA breast cancer gene expression. Rows in italic are cross-listed in both (a) and (b).

To substantiate our finding, we examined the gene regulation networks modulated by ESR1 ceRNAs in different breast cancer subtypes. As classified by earlier studies¹⁹, 4 major subtypes, determined by molecular signatures, are luminal A (ER+, PR+, Her2−), luminal B (ER+, PR+, Her2+), basal-like (mostly TNBC), and Her2 (amplified or over-expressed ERBB2). To construct subtype-specific ESR1 mediated ceRNA networks, we prepared 4 TCGA expression datasets for the corresponding 4 subtypes, which included 93, 56, 228, and 123 samples for Basal-like, HER2, Luminal A and Luminal B, respectively. Considering that genes may express constantly within each subtype, we added all normal reference samples to each subtype to increase the dynamic range for correlation analysis. For each subtype, co-expression analysis was performed and integrated with sequence-paring predictions (Table 5a). To generate the interaction network, the process was repeated on top 10 predicted ceRNAs. Figs. 4A–4D illustrate the resulting subtype-specific ceRNA networks. Among these networks, only 2 first layer ceRNAs (NOVA1 and CPEB3) are shared. NOVA1, neuro-oncological ventral antigen 1, has been implicated in breast cancer²⁹ and shown correlated in gene expression with ESR1³⁰, and CPEB3, a regulator of EGFR³¹, has been shown to be important in breast cancer³². While these two genes’ expression levels are mediated by ESR1 via GTmiRs in all four subtypes, other unique ceRNAs are also very important to each subtype. For example, ceRNA MAX in Basal-like regulation (Fig. 4C) is an important partner of proto-oncogene Myc in driving cell proliferation in variety of tumors. In the case of Her2 (Fig. 4D), ESR1 interacts with ANK2 and then PAX2, another gene that plays critical role in breast cancer³³.

Figure 4.

ESR1 ceRNA interactions networks for four breast cancer subtypes. Each network consists of three layers. The top layer includes the GOI (the largest node), the second layer includes 10 ceRNAs (medium size nodes) predicted in the first iteration, and the third layer includes predictions (smallest nodes) from the second iterations.

Discussion

Here we presented TraceRNA, an easy to use web application for predictions of ceRNAs of a GOI and their interaction network. This web application is motivated by lack of ready-to-use tools for ceRNA predictions and, to the best of our knowledge, TraceRNA is the only web application other than ceRDB that is specialized for ceRNA predictions. Compared with ceRDB, TraceRNA has richer functionality designed to meet different research needs. Because of the high false positive rate and low sensitivity of existing miRNA target prediction algorithms, TraceRNA provides the users with 3 different algorithms so that they can compare and/or complement results as a remedy to the potentially poor predictions from a single algorithm. TraceRNA also includes the validated targets from mirTarBase, which can be selected to potentially improve the specificity of ceRNA predictions. TraceRNA also enables predictions of context-specific ceRNAs by integrating co-expression with sequence-level predictions. In the current version, two expression datasets from TCGA have been preloaded into the web database. All prediction results include p-values and FDR as statistical significance. What is also unique about TraceRNA is its ability to construct and plot ceRNA interaction networks. This network can reveal important interactions that might not be easily perceived with the list of predicted ceRNAs. The generated network plots can be downloaded and are ready to be used for scientific publication. In contrast, ceRDB includes only one algorithm for miRNA target predictions and is also devoid of the afore-mentioned functions in TraceRNA.

Two case studies, prediction of PTEN ceRNAs and that of ESR1 ceRNAs, were presented to demonstrate the effectiveness of TraceRNA in making biological meaningful predictions. Because both genes are important cancer associated genes, their ceRNAs in the context of cancer were also predicted. In the case of PTEN, the GBM specific ceRNAs were shown to be functionally more enriched than the sequence-level predictions alone and important signaling pathways known to be related with PTEN regulation were also predicted among the most enriched pathways. When compared with ceRDB, TraceRNA predictions were functional much more enriched, indicative of higher predictions specificity. For ESR1, unique ceRNA interaction networks for four breast cancer subtypes were constructed. While ceRNAs common to four networks were observed, considerable differences exist among these four networks in ceRNAs and their interactions. Examples were provided to show possible links between the unique ceRNA interactions and the subtypes, which suggests that these differences in ceRNA interactions may very well be used to explain the genomics mechanisms underlying the subtypes. If proven true, the ceRNA networks could provide an alternative to the genomics markers for disease treatment. Taken together, TraceRNA has been shown as an effective tool for context-specific ceRNA predictions and discovery of ceRNA interactions modulated by GTmiRs.

Context-specific ceRNAs are closely dependent on miRNA expressions. A predicted ceRNA by sequence-pairing could not compete with GOI for binding of weakly expressed miRNAs. As a result, only highly expressed miRNAs should be considered in ceRNA predictions. Current version of TraceRNA does not yet consider miRNA expression for ceRNA predictions but displays expression values to help the user to select GTmiRs. On the other hand, mRNA expression profiles are still much more accessible than miRNA expression data, ceRNA predictions based on mRNA expression will still be of higher interest in practice. However, as miRNA profiles become increasingly available, there will be more demand to include miRNA expression in ceRNA predictions to achieve more accurate context-specific predictions. Future work should allow us to incorporate this function in TraceRNA to enable predictions in a miRNA expression dependent fashion.