Breast cancer transcriptomes reveal extensive co-expression involving lncRNAs

We analyzed the transcriptomes of 1,217 primary tumor samples from breast cancer patients (source: TCGA), with the goal of identifying regulatory relationships between lncRNAs and mRNAs. TCGA HTSeq-counts data was preprocessed to remove lowly expressed genes, normalize expression values and control for common confounding variables, such as age, sex, ethnicity and race. (See Methods for details of data processing and normalization.) The processed data included expression values for 13,963 mRNAs and 1,079 lncRNAs (Fig 1A). In a preliminary analysis, we calculated Pearson correlation coefficients (PCC) for each lncRNA-mRNA pair, and found extensive co-expression (Fig 1B), with 30% of all pairs at a PCC of absolute value greater than 0.17 (Bonferroni corrected p-value of 0.05). An example of a pair with strong correlation is the lncRNA MALAT1, a key regulator in breast cancer [22], and the mRNA MATR3, an RNA-binding protein with tumor suppressive function in breast cancer [36] (Fig 1C). The extensive co-expression was also noted among pairs of lncRNAs (Fig 1D), with ~36.5% of all pairs at a PCC of absolute value greater than 0.15 (Bonferroni corrected p-value of 0.05), with the majority of these co-expressed pairs being positively correlated. The extensive co-expression observed among lncRNAs and mRNAs is consistent with previous analyses of cancer data sets [8,9,11], including the TCGA breast cancer cohort [12,37], and provides a natural starting point for the rest of this work.

Download:

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

Fig 1. Co-expression among mRNAs and lncRNAs.

(A) Graphical Abstract of Workflow. RNA-Seq data for 1,217 primary tumor samples from 1,092 breast cancer patients were obtained from TCGA-BRCA cohort. The processed transcriptomic data (see Methods) include expression profiles for 1,079 lncRNAs and 13,963 mRNAs. Using ElasticNet regression, significant lncRNA, TF, and RBP regulators were identified. These significant lncRNAs with their supporting mechanisms or involved in interaction terms were then used in an Ordinary Least Squares (OLS) regression model to construct the final lncRNA-regulatory network. (B) Histogram of Pearson correlation coefficients of lncRNA-mRNA pairs. Among ~15M pairs analyzed, ~4.5M pairs had absolute value of Pearson correlation above 0.17, which corresponds to Bonferroni corrected p-value of 0.05. (C) Density scatter plot of expression of lncRNA MALAT1 and mRNA MATR3. MALAT1 and MATR3 are highly correlated with Pearson correlation of 0.58 (Bonferroni corrected p-value of 1.45e-109). (D) Histogram of pairwise Pearson correlations between lncRNA-lncRNA pairs. There are ~170K pairs positively correlated with correlation greater than 0.15 and 42K negatively correlated pairs with correlation lower than -0.15 (Bonferroni corrected p-value of 0.05).

https://doi.org/10.1371/journal.pcbi.1012103.g001

lncRNAs can explain a significant portion of gene expression variation

Our next goal was to estimate how much of the variation in gene expression can be explained using lncRNAs. This is an important step towards quantifying the potential regulatory influence of lncRNAs as a class. To this end, we first trained a linear model to predict each gene’s expression as a weighted sum of multiple lncRNAs’ expression levels. With over 1000 lncRNAs being candidate predictors and ~1,200 samples available for modeling, it was imperative that we constrain the model’s complexity and avoid “over-fitting” the data. We therefore used the Elastic Net algorithm to select ~10 informative lncRNAs for each target gene (mRNA) and used only the selected short-list of lncRNAs in modeling that gene (see Methods). The median R² (percentage of variance explained) was noted at 0.52 (Fig A in S1 Text), meaning that for most genes over half of the variance in expression can be explained using ~10 selected lncRNAs. To ensure that this observation is not the result of over-fitting, we performed a 5-fold cross-validation of the model. The test R² values (Fig 2A) had a median of 0.47, which is close to the 0.52 value noted above. Fig 2B shows that the test R² value of a gene is generally close to the overall R² value observed when training and evaluating the linear model on all samples, indicating that the results are likely not the result of over-fitting. The high R² values noted here are also mirrored in high values of Pearson Correlation Coefficient (PCC) between predicted and observed expression levels (Figs B and C in S1 Text), with a median PCC of 0.66.

Download:

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

Fig 2. LncRNA-mRNA co-expression network.

(A) Histogram of test R² when mRNA expression values were modeled using lncRNA. For each mRNA, Elastic Net algorithm was used to select ~10 most significant lncRNAs as predictors, which were then used to predict the mRNA’s expression using ordinary least squares (OLS) regression. Average test R² in a 5-fold cross-validation was calculated for each mRNA. (Each “fold” involved selecting predictor lncRNAs and training the model on 80% of samples and testing on the remaining 20%.) (B) Density scatter plot of test R² and overall (“training”) R² for the models. Overall R² is calculated by training the model on all samples and calculating its R² on all samples. Average test R² values were generally similar to the corresponding overall R² values of the same gene, indicating that overfitting is not a major concern. (C) Histogram of pairwise Pearson correlations for TF-mRNA pairs. About 6.6M of the pairs (37.57% of all pairs) have a PCC with absolute value greater than 0.17 (Bonferroni corrected p-value of 0.05). (D) Histogram of test R² when mRNA expression values were modeled using TFs selected using Elastic Net, in a five-fold cross validation. (E) Scatter plot of test R² values for models utilizing TFs or lncRNAs (x and y axes respectively) as predictors. Each point represents an mRNA whose expression is predicted by the models. PCC between two axes is 0.82 (p-value < 2.2e-16). (F) Histogram of test R² values (average over five folds of cross-validation) when lncRNAs were used to model mRNA expression values from which TF-based predictions have been subtracted away.

https://doi.org/10.1371/journal.pcbi.1012103.g002

The observation that a substantial portion of the variance of mRNA expression can be explained using small numbers of lncRNAs suggests extensive regulation by lncRNAs. But such lncRNA-mRNA covariation also reflects the effects of shared regulators such as transcription factors (TFs) that influence the expression of both the mRNA and the lncRNA. To arrive at a more conservative estimate of the extent of influence and to enrich the statistically identified lncRNA-mRNA pairs for those with direct regulatory relationships (lncRNA regulates mRNA), we considered two options: (a) impose mechanistic prior knowledge, as is done by demanding the presence of TF binding sites in gene promoters when reconstructing TF-gene regulatory networks [38], and/or (b) remove the effect of other regulatory mechanisms, e.g., TF-driven regulation of a lncRNA-mRNA pair, by treating them as statistical confounders.

We first chose to adopt the latter approach of controlling for shared regulators (TFs). A correlation analysis (Fig 2C) revealed that ~6.6M TF-mRNA pairs (37.6% of all pairs) have a PCC with absolute value greater than 0.17 (Bonferroni corrected p-value of 0.05), highlighting their extensive co-expression as expected. We then trained linear models to predict mRNA expression using ~10 selected TFs (from all 1,255 TFs included in the data set) and evaluated these using 5-fold cross validation. Test R² values, shown for all target genes in Fig 2D, have a median of 0.52, which is somewhat higher than the median R² value of 0.47 in comparable evaluations of lncRNA-based predictions. Direct comparison of R² values from TF-based and lncRNA-based models (Fig 2E) suggests that genes tend to be predictable to similar degrees using either class of predictors (TFs or lncRNAs). We also observed genes that are far more predictable using lncRNAs only or using TFs only (off-diagonal points in Fig 2E). Focusing on the 109 mRNAs that are more predictable using lncRNAs (y–x > 0.2 and y > 0.5 in Fig 2E) revealed that the most common significant lncRNA predictors of these target genes are AC108449.2 and AC073896.4 (S1 and S2 Tables), increasing our confidence in their functional regulatory roles. For example, AC073896.4 is sense overlapping to gene SMARCC2, which is one of the core subunits of human SWI/SNF complex, which is known to be associated with breast cancer [39]. Examining the most significant (p-value < 1E-10) lncRNA-mRNA relationships involving the above set of 109 mRNAs, we observed that 23.7% (65/274) were overlapping pairs and 39.8% (109/274) were pairs with genes within 1 Mbp of each other. These observations are consistent with reported mechanisms of lncRNA action that involve gene overlap [40,41], including in breast cancer context [42], as well as proximal chromatin localization [43]. We visit these observations again in a later section.

Finally, for each mRNA we computed the residuals from its TF-based model, i.e., subtracted from each original expression value the prediction from the TF-based model, and trained lncRNA-based models on these residuals. In other words, we now sought to use lncRNAs to explain the expression variance remaining after the putative effects of TFs had been subtracted away. Fig 2F shows the distribution of test R² values for all genes. (Also see Fig D in S1 Text for overall R² values.) Clearly, these estimates of percentage variance explained are substantially lower than those from Fig 2A (before TF effects were subtracted), with a median R² of 0.12 compared to 0.47, which is consistent with our expectation that much of the lncRNA-mRNA covariation can be explained as the result of TF-driven regulation. At the same time, we noted that the R² value was 0.2 or more for 1,869 mRNAs (16.2% of all); i.e., for a substantial number of mRNAs, at least 20% of the variance may be the result of lncRNA-driven regulation. While this threshold of 20% is somewhat arbitrary, it does represent a highly significant level of statistical association: a corresponding null distribution estimated empirically by permuting the expression levels of target genes across samples revealed that none of the 11,531 target genes can be modeled at a level better than this threshold (Fig E in S1 Text).

The high R² values observed above for a subset of target genes might still reflect mostly their co-expression with lncRNAs as opposed to regulatory influence of those lncRNAs, due to unknown shared regulators whose effects could not be subtracted away. We thus performed a “control” exercise where we used a random set of 1,079 enzyme-encoding mRNAs in place of the 1,079 lncRNAs as predictors and obtained R² values for all genes in an analogous manner. Since enzyme-encoding mRNAs do not represent a regulatory class, prediction of a target gene’s expression using these mRNAs likely reflects co-expression rather than regulation. Repeating this control exercise 100 times, each time with a different random choice of 1,079 enzyme-encoding mRNAs as predictors, we obtained a null distribution of R² values for each target gene; this allowed us to assign empirical p-values to the R² value achieved for each target gene when using lncRNAs as predictors. We observed 699 genes to have empirical p-value < = 0.01 (578 genes have p-value < 0.01, S3 Table), i.e., where at most one of the 100 control exercises using enzyme-encoding mRNAs achieved an R² value better than that seen with lncRNA predictors. These 699 genes comprise 6.1% of the 11,531 target genes analyzed, significantly higher than the chance expectation of 1% (Binomial test p-value <1.11e-302). This suggests that the high R² values observed using lncRNA predictors do not merely represent co-expression and are likely to reflect lncRNA-related regulation.

In summary, our analyses provide a rough estimate of the part of transcriptomic variation that may be under lncRNA regulation, and points to the existence of a large number of genes under such regulation, even after accounting for the confounding effects of common regulators.

Categorization of putative lncRNA regulators based on location relative to target gene

We next sought to characterize putative lncRNA regulators identified by our models above in terms of potential mechanisms of action. There have been several reports of cis-regulatory effects of lncRNAs (Fig 3A), e.g., antisense lncRNAs that regulate neighboring genes pre-transcriptionally, transcriptionally or post-transcriptionally [30]. Indeed, global maps of lncRNA-DNA interaction show that nucleus-retained lncRNAs tend to localize on proximal chromatin regions [43]. On the other hand, an often-reported mechanism of action is the so-called competing endogenous RNA (ceRNA) [44], which manifests in trans. It refers to the situation where a microRNA regulates a target mRNA post-transcriptionally and a lncRNA capable of binding the microRNA competes for the latter, thereby indirectly regulating the mRNA (Fig 3A). To assess the relative frequencies of cis and trans regulatory mechanisms, we considered, for each target gene, the following categories of associated lncRNAs: those (A) sense or antisense overlapping with target gene (“overlapping”), (B) located within the same topologically association domain (TAD) detected from Hi-C data from MCF10A breast cancer cell lines [45]; [46] as the target gene (“same-TAD”), (C) located within 1 Mbp of the target gene (“proximal”), (D) located on the same chromosome as the target gene (“same-chromosome”) and (E) competing endogenous RNAs (“ceRNAs”) predicted to share a microRNA regulator with the target gene, based on microRNA-mRNA and microRNA-lncRNA interactions [47]. In our context, competing endogenous RNA refers to a lncRNA-mRNA pair that are targets of the same miRNA in the same cellular contexts (see Methods). We focused on the target mRNAs whose expression could be well explained using lncRNA regulators (R² > = 0.2 and empirical p-value < 0.01, S3 Table), and examined their strongest putative lncRNA regulators (p-value < = 1E-10 in OLS regression, and pair-wise correlation greater than 0.1 in absolute value) in terms of the above categories. This yielded 643 lncRNA-mRNA associations (S4 Table), of which 500 (77.8%) were positive and 143 (22.2%) were negative relationships. Of these, 171 pairs have correlation lower than 0.2 and thus might not be detected by traditional co-expression based methods.

Download:

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

Fig 3. Contribution of lncRNAs from different “mechanistic” categories.

(A) Schematic of lncRNA regulation mechanisms. The figure shows the four classes of lncRNAs acting in cis with the target gene and one type of trans-acting lncRNAs (ceRNA). (B) Overlapping lncRNA-mRNA pairs (blue) are more correlated than overlapping pairs of mRNAs (red). X-axis shows the Pearson Correlation Coefficient (PCC) and the Y-axis shows for any value of PCC what fraction of lncRNA-gene pairs have correlation above that value. (CDF refers to cumulative distribution function.) (C) Number of genes whose expression was modeled using lncRNAs from each category. For instance, only 588 genes had an overlapping lncRNA to use as predictor, thus only these genes could be modeled. (D,E) Number of genes (D) and fraction of genes (E) that were modeled well (test R² > 0.2, on average across five folds of cross-validation) using lncRNAs from different categories. The fraction (E) is calculated as the ratio of respective numbers from C and D. (F) Normalized expression of gene LRP1 in four groups of samples where lncRNA TMPO-AS1 and the TF EBF1 are expressed at above (high) or below (low) their respective median level. In each category label, the first ‘H/L’ indicates high/low expression of TF and the second ‘H/L’ corresponds to the lncRNA. LRP1 expression is significantly lower in samples where both the TF and the lncRNA are highly expressed (HH) compared with the other three groups where either the TF or the lncRNA or both are lowly expressed (t-test p-values are shown for each between group comparison). The interaction term representing EBF1 and TMPO-AS1 has a p-value of 3.23E-10 in modeling gene LRP1. (G) Gene MYL12A expression is significantly higher in samples with high levels of RBP FXR1 and lncRNA OIP5-AS1 (HH) compared to the three other groups (HL, LH or LL) where either regulator has low expression. The interaction term representing FXR1 and OIP5-AS1 has a p-value of 7.8E-17 in modeling gene MYL12A.

https://doi.org/10.1371/journal.pcbi.1012103.g003

We noted that 177 (27.5%) of these pairs were overlapping lncRNA-mRNA pairs, all but one of which were positively correlated. Similarly, 228 (35.5%) of the 643 pairs were within the same TAD as each other; 54.0% and 62.7% were proximal and on the same chromosome (respectively) as each other, almost always positively correlated (Table 1). In other words, the discovered associations had large fractions of pairs located in “cis” and such pairs were overwhelmingly found to be positively related pairs. We also confirmed that these large fractions are not expected merely by chance (S2 Text). We noted a more balanced ratio of positively correlated pairs (51.5%) and negatively correlated pairs (48.5%) for the 293 trans lncRNA-mRNA pairs.

Download:

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

Table 1. lncRNA-mRNA pairs categorized by relative genomic locations.

For mRNAs that could be modeled using lncRNAs with R² > = 0.2 and empirical p-value < 0.01, we considered putative lncRNA regulators with OLS p-value < = 1E-10 and pairwise correlation coefficient > = 0.1 in absolute value. Shown are the number of pairs in each category and the number (of those) that are positively correlated. “Non-proximal” refers to pairs on different chromosomes or more than 1 Mbp apart on the same chromosome.

https://doi.org/10.1371/journal.pcbi.1012103.t001

A previous study by Chen et al. [48] has reported greater expression correlation between overlapping protein-coding genes than between non-overlapping pairs of genes. We therefore asked if the prevalence of overlapping pairs (and by generalization, the cis-located pairs) noted above may simply reflect co-expression rather than regulator-target relationships. To answer this, we considered 3,715 pairs of overlapping protein-coding genes (PCGs) and computed pairwise correlations to obtain a null distribution, which we then contrasted with a corresponding distribution of 615 overlapping lncRNA-mRNA pairs selected in an unbiased manner, without regard to their expression profiles. As shown in Fig 3B, the two distributions are significantly different (Kolmogorov-Smirnoff test p-value 2.1e-14). For instance, 250 (41%) of the 615 overlapping lncRNA-mRNA pairs have a correlation coefficient > = 0.3, a ~1.6 fold enrichment over the corresponding fraction of 24% seen among overlapping PCG pairs (Binomial test p-value of 1.8e-19). In other words, a random overlapping lncRNA-mRNA pair has a significantly higher expression correlation than a random overlapping PCG pair, making such overlaps a marker of potential regulatory relationship. This also suggests that the 177 overlapping lncRNA-mRNA pairs identified by us (Tables 1 and S4) are enriched for regulatory pairs.

Taken together, our statistical analyses suggest various modes of cis-regulatory action as a prominent theme in lncRNA-driven regulation. We also noted 288 lncRNA-mRNA pairs (of the 643 in S4 Table) whose genes were located more than 1 Mbp away on the same chromosome or on different chromosomes. These were evenly split between positive and negative correlations (146 and 142 respectively), in sharp contrast to cis-located pairs, which were predominantly positively correlated. (See S3 Text for selected examples of such pairs that have literature support for their regulatory relationship.) Overall, of the 288 significant lncRNA-mRNA pairs that were located distally or in trans, only 19 were designated as ceRNA in our analysis, leaving the vast majority of such pairs mechanistically unresolved.

cis-located lncRNAs explain a substantial part of gene expression variation

Above, we observed a prevalence of cis location in a lncRNA regulatory network that was obtained without regard to lncRNA location or mechanism. We next sought to quantify the percentage of mRNA expression variation that can be explained by lncRNA regulators limited to each of the above-defined categories of cis action. To this end, we repeated the above modeling exercise (Fig 2F) while limiting the candidate set of lncRNAs for a target gene to each of the five categories–overlapping, same-TAD, proximal, same-chromosome or ceRNA. The newly imposed restrictions on candidate lncRNAs used for predicting a given gene’s expression led to reductions in the number of genes analyzed (Fig 3C). We next examined the number (Fig 3C) and fraction (Fig 3D) of target genes that could be modeled “well” (average test R² > = 0.2 in 5-fold cross validation) using each of the five categories of candidate lncRNAs. The “overlapping” category yielded the highest fraction, with ~16.7% of the 588 genes modeled using this class of lncRNAs reaching the R² > = 0.2 threshold. This fraction is similar to the fraction of genes modeled well (16.2%) when the unrestricted class of all 1079 lncRNAs is used as candidates. Among the 588 genes that do have at least one overlapping lncRNA, 303 genes can be modeled well using all lncRNAs as candidates and 32% of these can also be modeled well using just the overlapping lncRNA(s) (Fig F in S1 Text). In other words, if an mRNA’s expression can be predicted using lncRNAs and the mRNA has an overlapping lncRNA, then that overlapping lncRNA is likely to be a significant predictor. (Also see Fig G in S1 Text) Note that it is possible for such pairs to be co-regulated by a shared regulator, rather than the lncRNA regulating the mRNA, but our strategy of subtracting the predicted effects of TFs prior to modeling makes this less likely. These results reinforce the observations made above and in the literature [49] that lncRNA-gene overlaps are a significant marker of and potentially tied to the underlying mechanisms of lncRNA regulatory function.

When modeling a gene using candidate lncRNAs in the “same-TAD” (typically, 1.8 candidates per gene) or “proximal” (typically, 2.0 candidates per gene) categories, the fraction of genes with R² > = 0.2 fell to 2–4% (Fig 3E), much smaller than the 16.7% observed with “overlapping” lncRNAs. This is not surprising: a TAD often has multiple genes and it is possible that a lncRNA within the same TAD regulates only one of those genes, driving the above-mentioned fraction down. Indeed, we found that ~20% of lncRNA-harboring TADs include at least one gene that can be modeled well using lncRNAs in the same TAD. In contrast, among TADs harboring enzyme-encoding genes, only 8.8% include at least one gene that can be modeled well using enzymes in the same TAD.

Enlarging the candidate sets even further, we noted that ~8.0% of analyzed genes can be predicted with R² > = 0.2 using lncRNAs from the same chromosome (typically, using ~11.5 regulators per gene). One way to interpret this observation is that while ~1,900 (~16%, Fig 2F) of analyzed genes could be modeled well using cis- as well as trans-acting lncRNAs, about 50% of these are amenable to expression prediction using only (putatively) cis-acting lncRNAs from the same chromosome.

Finally, gene expression prediction using lncRNAs in the ceRNA category led to only ~2% (246 out of 10556) of analyzed genes being modeled well, even though an average of 11 lncRNAs were considered as candidate regulators for each gene (S6 Table). We found 45 genes that have R² equal to 0.2 or greater and empirical p-value < 0.01 (S4 Text) in such modeling. Notable among these, lncRNA GAS5 was selected as one of the significant predictors for 18 out of the 45 mRNAs (Table 2), including 12 genes that encode ribosomal proteins, including PRS18, RPL31, RPL30, etc.. A previous study found that the overexpression of GAS5 facilitates cell apoptosis in breast cancer and increases the sensitivity to treatments [50], and it has been shown to act via the ceRNA mechanism [51]. It belongs to the 5’-TOP gene family that includes all ribosomal proteins [50], and five of its twelve putative ribosomal protein targets were also found to be among the top 20 genes related with GAS5 in kidney renal clear cell carcinoma [52].

Download:

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

Table 2. Predicted GAS5 targets with ceRNA mechanism of action.

Shown are target genes of GAS5 whose expression could be modeled using putative ceRNA regulators yielding R² of 0.2 or greater, with an empirical p-value < 0.01, i.e., where none of the 100 modeling attempts using equally many random lncRNAs yielded a better R². (* GAS5 target genes which are also included in the top 20 genes related to GAS5 in [27]).

https://doi.org/10.1371/journal.pcbi.1012103.t002

Statistical interaction analysis reveals potential protein mediators of lncRNA action

Beyond cis-regulatory and ceRNA mechanisms, lncRNAs are also known to function by acting as scaffolds for RNA- and/or DNA-binding proteins [34]. We hypothesized that such mechanisms may be detectable in the current data if we simultaneously examine expression variations of a target gene (mRNA), candidate lncRNA regulators and (mRNA of) potential protein mediators. Here, we report our findings from such an analysis where we detected significant statistical interactions between lncRNAs and either TFs or RNA-binding proteins (RBPs) in modeling mRNA expression.

Interactions with TFs: Above, we used a set of ~10 automatically selected lncRNAs and, separately, a set of ~10 selected TFs to model each target mRNA (Fig 2F and Fig 2D respectively). We now included both sets as covariates and re-trained a linear model to predict the target mRNA’s expression across all 1,217 samples as a function of these predictors. Starting with this combined “lncRNA + TF predictors” model as a baseline, we tested each lncRNA-TF pair for evidence of interaction by including their product as an additional covariate, re-training the model and extracting the statistical significance of the interaction term. Such interaction terms were added to the model iteratively, using the Akaike Information Criterion (AIC) as a measure of improved fit (Step-AIC method [53]). Inclusion of interaction terms improved the R² values achieved by the linear model, as expected, but the AIC criterion assured us that improvements were not merely due to additional free parameters. The median R² value achieved over all target mRNAs was now at 0.69, compared to 0.67 (t-test p-value = 2.68e-27) from the baseline model without interactions. Summarizing all significant interactions (S7–S10 Tables, limited to target genes whose R² values were above median), we noted that the TF SPI1 is involved in the most interactions (88, compared to a median of 5, see S11 Table). This is consistent with literature reports of this TF being associated with several lncRNAs [54–56]. We further focused on significant lncRNA-TF interaction terms where the target gene is differentially expressed (t-test p-value < = 0.05) between samples with high (above median) lncRNA and TF levels compared to samples that have low (below median) levels of either putative regulator. These 1,058 short-listed terms (S12 Table) present putative cases where the simultaneous presence of lncRNA and TF has an effect on target gene expression. In Fig 3F, we present the case of an interaction term involving the lncRNA TMPO-AS1 and the TF EBF1 (Bonferroni corrected p-value 3.23E-10), used in modeling the expression of gene LRP1. TMPO-AS1 has been implicated in several cancers, is activated by EBF1 and also indirectly activates EBF1 by sponging the microRNA miR-98-5p in bladder cancer [57]. Moreover, the predicted target, LRP1, is a repressor of the Wnt signaling pathway [58] and TMPO-AS1 has been shown to activate the Wnt pathway [59]. These observations together suggest that a repressive effect of TMPO-AS1 on LRP1 in the presence of TF EBF1 underlies the lncRNA’s activating influence on the Wnt pathway.

Interactions with RBPs: We performed a similar analysis to systematically detect lncRNA-RBP statistical interactions in prediction of mRNA expression (S13 Table). The inclusion of RBP interaction terms significantly enhanced the performance of the model, increasing the median R-squared value from 0.29 to 0.33 (t-test p-value: 1.51e-199). Summarizing all significant interactions (S14–S16 Tables, limited to target genes whose R² values were above median), we observed the RBPs ESRP1 and ESRP2 to be among the top four RBPs by frequency of detected interactions. ESRP1 has been reported as controlling ER+ breast cancer [60], and ESRP1/ESRP2 were noted as being the most strongly down-regulated RBPs in epithelial-to-mesenchymal transition (EMT) in a cell culture model of breast cancer [61]. The most frequent RBP (by frequency) per our analysis–PCBP1 –has also been shown to regulate EMT in breast cancer [62]. Next, as for TFs, we next short-listed 2,425 interaction terms with the above-mentioned pattern of differential target gene expression between categories of samples defined by high and low expression of lncRNA and RBP (S17 Table). One such term involves the lncRNA OIP5-AS1 and the RBP FXR1, with a p-value of 7.8E-17 when used for modeling the target gene MYL12A. As shown in Fig 3G, samples with high levels of the lncRNA and the RBP have significantly higher expression of the target gene MYL12A than those with low levels of either regulator, especially the lncRNA. Indeed, OIP5-AS1 has been shown to bind to FXR1 in order to regulate other genes, activate the Wnt signaling pathway and contribute to cell proliferation [63,64]. Another example is the interaction term for lncRNA SNHG11 and RBP SNRPB2, found significant when modeling the target gene RPRD1B (Bonferroni corrected p = 5.5E-17). We noted that the lncRNA (SNHG11) is located in cis to (< 400 Kbp separated from) the target gene (RPRD1B), adding support to the regulatory relationship. We also examined the interaction terms associated with the lncRNA MALAT-1, which is among the best studied lncRNAs in breast cancer context [22,65]. Our analysis (S17 Table) finds 12 such RBPs, including SRRM1, SRRM2, and SRSF1, which have been shown to interact with MALAT-1 [46,66].

To summarize, extending the linear model for gene expression to include interaction terms allowed us to generate hypotheses regarding potential mechanisms by which a lncRNA may exert its regulatory influence, mediated by or while competing with RNA binding proteins or transcription factors.

A multi-evidence lncRNA regulatory network

In the analyses above, we relied on expression modeling to identify potential lncRNA regulators of a gene, and then annotated them with features such as cis-location, ceRNA, or statistical interactions with proteins, that provide supporting evidence or plausible mechanisms for a subset of these regulators. Our final goal was to integrate these two complementary sources of information: to model each target gene using only the small set of candidate lncRNA regulators for which there exists supporting evidence found in the various annotation steps above. Specifically, we limited the candidate regulators for each target gene to those that were statistically significant in above-trained models for that gene (p-value of OLS coefficient < = 0.01) and had been annotated with any of the location categories of “overlapping”, “same-TAD”, “proximal” (S4 Table) or the mechanistic category of “ceRNA” (empirical p-value = 0, S6 Table) or had been found in a statistical interaction term involving a TF (S7 Table) or an RBP (S13 Table). As a result, for each target gene there were about two candidate lncRNAs, which were used in an ordinary least square regression to fit the gene’s expression profile. The R² values in this exercise were comparable to those obtained when utilizing all lncRNAs as candidates (Fig 4A), even though the number of candidates was two orders of magnitude smaller, with 271 (5.0%) of the 5,364 genes modeled meeting the R² > = 0.2 threshold. We focused on these target genes (mRNAs), constructing a regulatory network comprising their significant lncRNA predictors (p-value of coefficient < = 0.01). This network (S18 Table) includes 508 lncRNAs regulating 232 mRNAs, a total of 1157 lncRNA-mRNA regulatory relationships, with an average of 2.3 targets per lncRNA and 5.0 regulators per mRNA. We calculated the correlation between the lncRNA regulator and the mRNA target for the edges in the predicted network and, separately, for 1157 randomly sampled pairs, using RNA-seq data for 54 tissues from GTEx (S22 Table). The edges have significantly higher correlation compared to the random pairs in breast mammary tissue with t-test p-value of 2.1e-10, suggesting significantly higher co-expression for predicted lncRNA-mRNA regulatory pairs. Only 199 edges would be detected by the co-expression based regulatory network from the same transcriptomic dataset where 1157 most correlated lncRNA-mRNA pairs among the same 508 lncRNAs and 232 mRNAs are defined as lncRNA-mRNA associations. In the co-expression network, each lncRNA has 4.2 targets on average, which is significantly higher (p-value: 1.04e-7) than the average number of targets per lncRNA in the regulatory network, suggesting that co-expression-based approaches might be in favor of common regulators whose expression are highly correlated with many targets and fail to find regulatory lncRNAs with more specific function.

Download:

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

Fig 4. Expression modeling and regulatory network using selected lncRNAs with location- or mechanism-based supporting evidence.

(A) R² values using selected lncRNAs (on average ~2 lncRNAs per target gene) were comparable to those obtained when utilizing all lncRNAs as candidates. (B) Network edges (lncRNA-gene pairs) in the “overlapping” category are strongly enriched among the edges with the strongest p-values from expression modeling. X-axis represents the network edges sorted by p-value from expression modeling, and Y-axis shows the number of “overlapping” lncRNA-gene pairs among the top x network edges, for each value of x. (C) Breast cancer related lncRNAs are enriched among lncRNAs with high degrees in the regulation network. Shown is the standard graphical output from the GSEA tool. X-axis represents lncRNAs sorted by the number of putative gene targets (edges in the network). Breast cancer-related lncRNAs from the literature are marked by vertical bars in the middle panel, and the top panel (green curve0 shows the “enrichment score” statistic calculated by GSEA as a function of the rank order. This analysis shows that breast cancer-related lncRNAs tend to have significantly more targets compared to other lncRNAs, in our network. (D) Venn diagram of predicted GAS5 targets that are also differentially expressed genes (DEGs) between M1 and M4 cell lines. (E) Target genes of four high-degree lncRNAs (more than 25 targets each) were enriched in REACTOME pathways related to translation processes.

https://doi.org/10.1371/journal.pcbi.1012103.g004

We further validated a subset of the putative lncRNA-mRNA pairs using experimentally measured chromatin interaction from GRID-seq assays on the breast cancer epithelium cell line MDA231 [67,68]. These data provided RNA-DNA interaction information for 23 of the lncRNAs present in the multi-evidence network. The network includes 111 edges involving these lncRNAs, 29 (26.1%, hypergeometric test p-value: 5.7e-8, see Methods) of which are supported by the GRID-seq data as interactions between lncRNA and proximal region (1 Mbp) of the target gene (S18 Table Column “GRID-seq”). Notably, all of the nine predicted edges involving MALAT1 and six of the seven predicted targets of NEAT1 are supported by GRID-seq data.

The network included 549 ceRNAs and 514 cis-acting (overlapping, same-TAD or proximal) lncRNAs. We noted a clear enrichment of the “overlapping” category among the edges with the strongest p-values from expression modeling (Fig 4B). Each edge is supported by some form of evidence, e.g., location near or in cis with target gene, ceRNA evidence, statistical interaction with a TF or RBP, in addition to its expression being predictive of the target’s expression, making this a higher confidence, multi-evidence lncRNA regulatory network. Among the 388 lncRNAs with cis-acting edges in the network, 271 overlap the regulatory genetic regions of human mammary epithelial cells obtained from GenenHancer [69]. Given that 745 of the 1074 lncRNAs analyzed in this work overlap the regulatory regions, the lncRNAs reported with cis-edges in the network showing no enrichment for enhancer marks (Odds ratio: (271/388)/(745/1079) = 1.01, p-value:0.64), suggesting cis-acting lncRNAs reported in the network don’t merely mark active regulatory regions.

We observed 53 of the 508 lncRNAs in this network to have 25 or more targets (> 10% of the 232 target mRNAs in the network, see S19 Table) and significantly higher expression compared to other lncRNAs (p-value: 1.92e-6), suggesting a possible prioritization of lncRNAs with the most regulatory role. Indeed, these “hubs” included several lncRNAs well known for their role in breast cancer, e.g., MALAT1 [22,65] at rank 17, SNHG29 [70] at rank 2, SNHG16 [71] at rank 19, GAS5 [50,72] at rank 1, XIST [73,74] at rank 11, NEAT1 [75,76] at rank 23, NORAD [17,77] and others. (It is noteworthy that the majority of connections associated with these lncRNAs are predominantly trans-regulatory interactions, see S21 Table and Fig I in S1 Text) Pursuing the above observation systematically, we obtained 30 breast cancer-related lncRNAs from the LncRNADisease database [72], of which 10 were included in the high confidence regulatory network. Four of these 10 were among the 53 high-degree regulators noted above, and a Gene Set Enrichment Analysis (GSEA) revealed a significant association (p-value < 0.027) between the breast cancer-related lncRNAs and the degree of a lncRNA in the network, as shown in Fig 4C.

Next, we functionally characterized the predicted targets of each high-degree regulator (lncRNA) through gene set analysis. We first conducted enrichment tests against a previously published set of genes that are differentially expressed (DE) between benign (“M1”) and highly tumorigenic (“M4”) states of a widely used breast cancer progression cell line series [18] derived from MCF10A mammary epithelial cells. We found two of the lncRNAs—GAS5 and SNHG29—to have targets significantly enriched in the M1-M4 DE genes (FDR < 0.05), with Hypergeometric test p-value 5E-6 (Fig 4D) and 6E-4 respectively. Expanding the enrichment tests to iterate over an entire compendium of biological pathways from the REACTOME database [78], we observed a set of four lncRNAs (OIP5-AS1, GAS5, ZFAS1, SNHG29) to have their targets enriched for several pathways related to translation processes (Fig 4E, S20 Table, and Fig J in S1 Text). For instance, the lncRNA GAS5, a tumor suppressor that is down-regulated in mammary carcinoma [79], has 20 targets in the lncRNA-mRNA network and 12 of these putative targets belong to the pathway “peptide chain elongation” (Bonferroni corrected p-value 3.23e-15). Indeed, this lncRNA is believed to regulate translation of specific proteins such as c-Myc by interacting with the translation initiation complex protein eIF4F [80], which is one of its predicted targets in our network.

The regulatory network comprises edges with directionality, i.e., not only is each lncRNA-mRNA relationship assigned a significant p-value in expression modeling, the sign of the coefficient in the linear model tells us if the relationship is activating or inhibitory. We noted 782 activating and 355 inhibitory edges overall in the network and identified four lncRNAs with a significant departure from the global ratio in one direction or the other (Binomial test p-value < 0.01, see Figure H in S1 Text). These included the lncRNA SNHG29, that has 19 targets in the network, of which 18 (94.7%) represent activating relationships.

While the majority of lncRNAs were initially thought to lack protein-coding potential, emerging evidence suggests that certain lncRNAs can indeed produce functional peptides [1]. These peptides, often referred to as "micropeptides," are short sequences of amino acids encoded by specific regions within lncRNA transcripts. They can be translated from non-canonical open reading frames (ORFs) and have been shown to have regulatory roles in various biological processes. Micropeptides derived from lncRNAs can associate with ribosomes and be involved in translation regulation or other ribosome-associated functions.

Overall, while the majority of lncRNAs are non-coding, there is growing evidence supporting a connection between certain lncRNAs and the production of ribosome-associated peptides, expanding the functional repertoire of these RNA molecules beyond their non-coding roles.

In summary, the regulatory network constructed here is based on both expression modeling and prior evidence of plausible mechanisms (cis-action, ceRNA or protein-mediated action), and we found several additional lines of evidence supporting its compendium of regulatory relationships: enrichment of the highest degree regulators for reported involvement in breast cancer, association of target gene sets with DE genes from a breast cancer progression cell line series and with translation-related as well as other relevant biological processes, and statistically significant biases towards either activating or inhibitory relationships.

Expression data collection and pre-processing

Transcriptome profiling data from 1092 patients with breast invasive carcinoma were obtained from The Cancer Genome Atlas (TCGA) breast cancer cohort (TCGA-BRCA project). The dataset consisted of RNA-Seq expression measurements and clinical data for 1217 primary tumor samples. The HTSeq-counts data encompassed measurements for 60488 RNAs on the gene level. RNA-Seq data processing steps are documented at the GDC portal (URL https://docs.gdc.cancer.gov/Data/Bioinformatics_Pipelines/Expression_mRNA_Pipeline/). To ensure reliable analysis, RNAs with counts per million (CPM) greater than 1 in fewer than 500 samples were removed due to low expression, resulting in a filtered expression dataset containing 15766 RNAs with adequate expression levels. RNA annotations were retrieved from the Ensembl database (Ensembl 104: May 2021) [87]. The filtered matrix comprised expression data for 13963 protein-coding genes and 1079 long non-coding RNAs (lncRNAs). Note that since TCGA RNA-seq data collection involved a polyA enrichment step, lncRNAs that are not polyadenylated will not be suitable represented in the analyzed data. The raw HTSeq-counts were normalized using TMM normalization in the R package “EdgeR”. The TMM-normalized expression values were then log-transformed and standardized to have zero mean and variance equals to 1. Confounder variables including age, sex, ethnicity and race for patients were obtained from the clinic data for the same cohort. The transformed normalized expression values of all RNAs were modeled using the confounder variables via simple linear regression models and residuals from these confounder-based models were used for further modeling.

Regression models of expression data

We trained linear models to predict each gene’s expression using different sets of predictors, which included lncRNAs, TFs or confounder variables. ElasticNet algorithm was used to select ~10 predictors (features) when there were more than 10 candidates. An ordinary linear squares (OLS) regression model was then trained using the selected features. For genes with fewer than 10 predictors, an OLS model was used directly to predict the gene’s expression using all the predictors, without any feature selection. For most analyses here, 5-fold cross validation was used and test R² averaged across the five folds was used to measure the performance of the model. After cross validation, the same process (ElasticNet followed by OLS model) was repeated on the entire data set and the resulting R² was reported as the training or “overall” R².

Transcription factors as confounders

We obtained identities of 1665 human transcription factors (TFs) from HumanTFDB [88]. Expression profiles for 1255 of these TFs were available in the expression data set. We first used Elastic Net algorithm to select ~10 TF regulators to predict the expression of each gene. If the gene being modeled is a TF itself, it was removed from the predictor set when training the Elastic Net model. The gene’s expression was then modeled using the selected TFs. Residuals from the TF-based model were then subjected to modeling using lncRNA predictors as described above.

Statistical lncRNA-RBP interaction analysis

We first modeled mRNA residuals from the TF-based models using ~10 RNA-binding protein (RBP) genes selected by the Elastic Net algorithm out of a candidate set of 351 RBPs obtained from RBPDB V1.3.1 [89]. We then used the candidate set of ~10 lncRNAs and ~10 RBPs to model each mRNA’s residual from the TF-based model. The model with the lncRNA and RBP predictors forms the baseline model for each mRNA. We then identified significant lncRNA-RBP interaction terms iteratively using the Step-AIC method [53]. We reported the R² values from the lncRNA-based models, RBP-based models, lncRNA+RBP (baseline) models and lncRNA-RBP interaction models.

Location and mechanism categorization of lncRNAs

We categorized putative lncRNAs based on the following relationships to target genes: 1) Overlapping: LncRNA overlaps the gene on either DNA strand. Locations of protein coding genes and lncRNAs were obtained from Ensembl BioMart. 2) Same TAD: TAD boundaries from Hi-C data for MCF-10A mammary epithelial cell line were used to define TADs [45]. LncRNAs in the same TAD with the gene were defined as “same TAD” lncRNAs. 3) Same Chromosome: LncRNAs in the same chromosome as the target gene were included in the “same chromosome” category. 4) ceRNAs: Experimentally supported miRNA-gene interactions were downloaded from TarBase v.8 [47]. Relationships between miRNAs and lncRNAs were obtained from lncBase v.2 [90]. LncRNAs that are related to the same miRNAs as the target gene in the same cell lines were considered to be the competing endogenous RNAs (ceRNAs).

GRID-Seq enrichment test

RNA-chromatin interactions in the breast cancer epithelium cell line MDA231 were obtained from GRID-seq profile GSE82312. Twenty three lncRNAs are measured in both GRID-seq and TCGA RNA-seq data. Hypergeometric test was used to measure the overall enrichment of lncRNA interacted genes in predicted edges:

1. Universe: size 318849 = (23 lncRNA included in both our study and GRID-seq) X (13963 genes included with TCGA RNA-seq measurements)
2. Set A: size 27757 = lncRNA-mRNA interaction pairs identified by GRID-seq, through binding evidence of lncRNA on 1mb region around the gene.
3. Set B: size 111 = predicted edges for 23 lncRNAs included in both studies
4. Intersection between Set A and Set B: size 29.
5. P-value: 5.738724e-08

Pathway enrichment test

For lncRNAs with more than 25 putative target genes, associations between the target gene set and pathway gene sets from Reactome database [91] were detected via hypergeometric test of overlap between the two gene sets and FDR-corrected p-values were reported.