The gene and miRNA expression profiles were downloaded from The Cancer Genome Atlas (TCGA¹) database release 10.0, which provided miRNASeq and HTSeq data. The lncRNA and mRNA annotation were downloaded from GENCODE (V22, GRCh38). Only tumor types with sufficient adjacent normal samples were considered (N > 30), including breast invasive carcinoma (BRCA), head and neck squamous cell carcinoma (HNSC), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), liver hepatocellular carcinoma (LIHC), lung squamous cell carcinoma (LUSC), prostate adenocarcinoma (PRAD), and thyroid carcinoma (THCA) (Supplementary Table 1). The expressed genes (FPKM ≥ 1 in >70% samples) were selected for subsequent analyses. All of the expression profiles were log2 transformed.

We also downloaded the HM450 DNA methylation profile of eight cancer types from TCGA (Supplementary Table 1). The probes with missing values in more than 30% of samples were removed, and other missing values were replaced by the mean value of the corresponding probe across samples.

The independent datasets were downloaded from the GEO database², including 15 datasets acquired by the Affymetrix Human Genome U133 plus 2.0 array and Illumina HumanMethylation450 BeadChip across eight cancer types (Supplementary Table 2).

The experimental human miRNA–mRNA/lncRNA interactions were collected from four datasets, including miRTarBase 7.0 (Chou et al., 2018), miRecords 2013 (Xiao et al., 2009), starBase 2.0 (Li J. H et al., 2014), and lncRNASNP2 (Miao et al., 2018). Through redundancy analysis and standardization, 729,240 miRNA–mRNA pairs and 7092 miRNA–lncRNA pairs were obtained.

The cancer hallmark-associated GO terms were derived from a previous study (Plaisier et al., 2012). Cancer-related lncRNAs/mRNAs were collected from several databases, including COSMIC v89 (Forbes et al., 2015), NCG 6.0 (Repana et al., 2019), LncRNADisease (Chen et al., 2013), and lnc2Cancer v2.0 (Gao et al., 2019). Besides, we searched miRCancer (Xie et al., 2013) and used the eight caner types as keywords to filter cancer-related miRNA. In total, we obtained cancer-related mRNAs, miRNAs, and lncRNAs for 2362, 461, and 756 separately.

To identify dysregulated ceRNA interactions, the miRNA-target regulations as well as expression associations among miRNA, lncRNA, and mRNA were considered. Then, we constructed the dysregulated ceRNA network according to the following three qualification rules (Figure 1).

Epigenetic regulation is one of the important mechanisms utilized to control lncRNA expression. To explore the association between lncRNA expression and methylation in the dysregulated ceRNA network, we identified ER lncRNA according to the method of Wang Z. et al. (2018). We first searched the probes in the promoter region of each lncRNA to acquire lncRNA–probe pairs. Then, Spearman correlation coefficient between the methylation and expression levels for each lncRNA–probe pair was calculated in each cancer type. The discrete categories included strongly negatively correlated (SNC, correlation coefficient: [−1, −0.5]), weakly negatively correlated (WNC, correlation coefficient: [−0.5, −0.25]), and no negative correlation (NNC, correlation coefficient: [−0.25, 1]), which were assigned based on the correlation coefficient. The probe with the highest coefficient was selected for the lncRNA if multiple probes were annotated to the same lncRNA promoter. Next, we defined lncRNA status according to the observed 30th and 70th beta values across tumor (T) and normal (N) samples. Then, we scored each lncRNA gene per cancer type according to the following rules:

Finally, we assigned a “call” for each of the possible combinations [3 (SNC, WNC, NNC) × 4 (CUN, CMN, VMN, IMN) × 4 (CUT, CMT, VMT, IMT)] per platform. In this way, the global trend of each lncRNA in one cancer type was acquired. Through combining the obtained pattern with the methylation level of this lncRNA in each cancer sample, we further determined the role of this lncRNA in a single sample. The epigenetic regulation types comprised EA and ES, and the other cases were not considered (Supplementary Table 3). According to the manually defined classifier, if the combination for the lncRNA and methylation probe was SNC × CMN × CUT and the beta value of the cancer sample at this probe was less than 0.25, we called the cancer sample EA at this probe or lncRNA. Next, epigenetic statuses of lncRNAs were characterized based on the percentage of regulated patients. If the number of EA samples was more than twice as the ES samples in a single cancer type, we determined it as EA lncRNA in this cancer type. Similarly, if the number of ES samples was more than twice as the EA samples in a single cancer type, we determined it as ES lncRNA in this cancer type. The others will be defined as multi-ER lncRNAs. The detailed information of these ER lncRNAs that occurred in a single cancer type is shown in Supplementary Table 4.

The characterizing process of ER lncRNAs in pan-cancer was similar as mentioned above. If the number of EA samples was more than twice as the ES samples in at least 75% of cancer types, we determined it as pan-cancer EA lncRNA. For example, the lncRNA was ER in four cancer types and the number of EA samples was twice greater than the ES sample in three cancer types; the percentage was 75%, and we considered this lncRNA as an EA lncRNA in pan-cancer. Next, if the number of ES samples was more than twice as the EA samples in at least 75% of cancer types, we determined it as an ES lncRNA in pan-cancer. The others will be defined as multi-ER lncRNAs, and these lncRNAs played different roles of epigenetic activation and epigenetic silencing in different cancer types. The detailed information of these ER lncRNAs that occurred in pan-cancer is shown in Supplementary Table 5.

To evaluate the potential diagnosis capacity of ER lncRNAs, the scoring classifier was constructed using two-dimensional data (expression and methylation, Figure 2). We first build four classifiers that separated tumor and normal samples using the mean value of expression and methylation of EA and ES lncRNA, respectively. Next, the four cutoffs acquired from the abovementioned receiver operating characteristic (ROC) curve were collected, and an ER lncRNA-based score was constructed for each sample. We will assign a point when the sample meets either of the following situations: (1) the sample showed higher expression than the EA expression cutoff; (2) the sample showed lower methylation than the EA methylation cutoff; (3) the sample showed lower expression than the ES expression cutoff; (4) the sample showed higher methylation than the ES methylation cutoff. Then, we summarized these points for each sample of the ER lncRNA-based score (range from 0 to 4). A higher score denoted that the sample was subjected to epigenetic regulation in cancer. Finally, the scores of all tumor and normal samples in each cancer type were collected, and ROC curve analyses were conducted to investigate the diagnosis performance of the classifier.

The Cox regression was performed to evaluate the prognosis of each ER lncRNA based on its expression or methylation level. Then, the ER lncRNAs with potential prognosis (cox-P < 0.3) were combined to obtain the survival score for each cancer sample according to the following formula:

where X_i is the expression or methylation level of lncRNA i in sample S, HR_i is the overall hazard ratio of lncRNA i, se(HR_i) is the standard estimates of HR_i, and n is the number of ER lncRNAs.

The median score was used as the cutoff point to divide the patients into low-risk and high-risk groups. The overall survival (OS) of these groups was compared using log-rank test.

lncRNA has been found to act as ceRNA that indirectly regulates mRNA via shared miRNAs, and the dysregulation of the crosstalk between ceRNAs could promote the development of cancers (Zhang S. et al., 2018; Li P. et al., 2020). To assess dysregulated lncRNA-associated ceRNA patterns in cancer process, we identified the gain and loss ceRNA interactions and further constructed dysregulated ceRNA networks for eight cancer types (Figure 1). In total, we identified 6381 mRNAs and 154 lncRNAs participating in 47,714 dysregulated ceRNA interactions. In DysCeNets, there were 2807–4589 mRNA and 51–102 lncRNA involved in ceRNA dysregulation (Table 1 and Figure 3A). We then explored the distribution of gain and loss dysregulated patterns of ceRNA interaction across networks. As a result, there were 3360–7885 gain interactions and 1921–6846 loss interactions across eight cancer types (Figure 3B). These dysregulated ceRNA pairs were either specifically competing with miRNAs in cancer/normal context or significant differences in the intensity of competitive capacity between both statuses. These results suggest that ceRNA dysregulation was common in the cancer process.

The global patterns of lncRNA-associated competing triplets and the characteristics of ceRNAs in the network across different cancer types have been revealed (Wang et al., 2015). However, few studies have focused on the dysregulated ceRNA interactions in pan-cancer. Through topological feature analysis, properties of DysCeNets were revealed (Figure 3C and Supplementary Figure 1). Firstly, the node degree distribution of the networks was investigated. We found that these DysCeNets revealed power-law distribution with R² ranging from 0.57 to 0.62, suggesting that the networks displayed scale-free characteristics typical of biological networks. In each DysCeNet, most ceRNAs had few interacting dysregulated ceRNA partners, while a small subset of ceRNAs had a relatively large number of interacting dysregulated ceRNAs. In general, the characteristic path length, average node, and edge betweenness were significantly increased when compared with random networks (P-value < 0.001), implying that the DysCeNet had reduced global efficiency. In addition, we found that node degree and betweenness of lncRNAs were significantly higher than mRNAs (Wilcoxon test, P-value < 0.05), suggesting the leading roles of lncRNAs in the dysregulated networks.

Herein, we compared the attributes (including mRNAs, lncRNAs, and edges) of the DysCeNets and found that any two DysCeNet shared a large proportion of mRNA and lncRNA, implying that the lncRNA-associated ceRNA dysregulation was widespread in the cancer environment (Figure 3D). Besides, we found that the lncRNAs in the dysregulated ceRNA pairs were more consistent than mRNAs, suggesting that lncRNAs may play more crucial roles in ceRNA dysregulation. Moreover, DysCeNets obtained from similar original tissues tended to share more lncRNAs, which were consistent with previous studies (Xu et al., 2015; Zhang Y. et al., 2016). For instance, KIRC and KIRP are two types of kidney carcinomas, and approximately 88% of lncRNAs in KIRC also worked in KIRP. This analysis revealed that the molecular characterization of cancers with similar tissue of origin was more relevant than the others. Although there was a considerable mean of 1314 common pairs between any two cancer types, their Jaccard indexes ranged from 0.057 to 0.103, which were far less than the nodes’ indexes.

To explore the lncRNA properties as miRNA sponge in DysCeNets, the related characteristics including transcript length, number of exons, and expression level were compared to those lncRNAs that were not involved in DysCeNets. As a result, the lncRNAs in DysCeNets were found to have longer transcript length, own more exons, and express higher than other lncRNAs in eight cancer types (Wilcoxon test, P-value < 0.05, Figure 3E), suggesting that they are more adaptable to act as miRNA sponges. The detailed comparative information between lncRNA in and out networks was provided in Supplementary Table 6. Together, these results validated that lncRNAs are key components involved in ceRNA dysregulation.

A common core of ceRNA regulatory interactions was defined as a component whose ceRNA triplets occurred in multiple cancer types. The core component could maintain the architecture of ceRNA networks across cancers and those ceRNAs in the component were found to be highly enriched in basic cellular processes to cancer (Xu et al., 2015). To determine the core component that exists in dysregulated ceRNA networks, we focused on the dysregulated ceRNA interactions that occurred in at least four cancer types. In total, 1713 edges were extracted to construct the core component, involving 1291 mRNAs and 43 lncRNAs (Figure 4A). We further defined the edges with consistent dysregulated type in more than 75% of cancers as gain or loss interactions and the others as multi-interactions in the pan-cancer dysregulated core component. We found that 57.62% of the edges in the core component showed the same dysregulated direction across multiple cancers. There were 590 gain interactions (59.78%) and 397 loss interactions (40.22%), indicating that a stable portion of RNA molecules tended to gain competitive relationships during cancer process. In addition, 42.38% of the edges showed different status among cancers, suggesting the complexity of ceRNA dysregulation.

Next, the proportion of different dysregulated interactions that lncRNAs linked in the core component is explored in Figure 4B. We found that lncRNAs with coincident property in the core component may associate with multiple cancer processes. For example, the lncRNA MALAT1, which owned the largest subnetwork and largest loss proportion, was found to regulate cancer glucose metabolism by enhancing mTOR-mediated translation of TCF7L2 in hepatocellular carcinoma (Malakar et al., 2019). Moreover, the lncRNA MALAT1 could mediate cisplatin resistance via the miR-101-3p/VEGF-C pathway in bladder cancer (Liu et al., 2019) and promote cell proliferation and inhibit apoptosis by sponging miR-101 in colorectal cancer (Si et al., 2019). Another case is the lncRNA HCG18, possessing the most gain interactions, which could cooperate with NOTCH1 to regulate the proliferation and migration of bladder cancer cells (Xu et al., 2019). In addition, HCG18 was identified as an immune-related signature and showed prognostic efficacy for anaplastic gliomas (Wang W. et al., 2018). The NEAT1–MET axis was identified as gain interaction in our pan-cancer core component, suggesting that the competitive relation between NEAT1 and MET happened in cancer environment. Several studies have proved that NEAT1 can regulate c-met via ceRNA mechanism in different cancer types. For instance, NEAT1 suppressed sorafenib sensitivity of hepatocellular carcinoma cells via regulating miR-335/c-Met (Chen and Xia, 2019). NEAT1 was also found to regulate the growth, invasion, and migration of pancreatic cancer cells through microRNA-335-5p/c-met (Cao et al., 2016). These results further verified the validity of our study, and the utilization of the method could provide potential cancer biomarkers.

Using a cohort of publicly available datasets including COSMIC, NCG, LncRNADisease, and lnc2Cancer, cancer-related mRNAs and lncRNAs were collected. We found that each DysCeNet owned 17.24–18.47% cancer-related mRNAs and 28.43–42.03% cancer-related lncRNAs. Through comparing the core component with single DysCeNet, we found that the percent of cancer-related ceRNAs (especially lncRNAs) in the core component was higher than that in single dysregulated network (Figure 4C). The large proportion of cancer-related lncRNAs in dysregulated networks and core component further confirmed the crucial position of lncRNAs, which was consistent with previous results. A previous study has demonstrated the relationship between the number of common miRNAs and the intensity of ceRNA competitive capacity; co-expression of ceRNAs in the network could increase with the number of common miRNAs (Xu et al., 2015). In our study, we found that the numbers of common and cancer-related miRNAs that ceRNAs compete with in the core component were significantly increased than those in a single dysregulated network (Wilcoxon test, all P-value < 0.05, Figure 4D and Supplementary Figure 2). This result revealed that the stable ceRNA pairs were inclined to dysregulate in the pan-cancer level. Due to the limitation of annotated information for lncRNAs, lncRNA functions were frequently presumed based on known functions of related mRNAs (Wang et al., 2015, 2019; Song et al., 2017). Thus, the biological process and KEGG pathway enrichments were tested using mRNAs that occurred in the component. Processes for cell proliferation (such as cell death and cell cycle) and cancer-related pathways (such as TGF-beta signaling pathway, MAPK signaling pathway, and pathways in cancer) were highly enriched (Figures 4E,F). Overall, these observations suggest that the core dysregulated ceRNA component was strongly related to cancer processes and further proved the importance of lncRNAs.

Growing evidences suggest that DNA methylation, a fundamental feature of epigenomes, can affect lncRNA expression, and there are intricate regulatory relationships between DNA methylation and lncRNA (Hadji et al., 2016; Yang et al., 2017; Zhi et al., 2018). Among these studies, Wang et al. characterized the epigenetic landscape of lncRNAs and identified recurrent ER lncRNAs in 20 cancer types (Wang Z. et al., 2018). However, the function of ER lncRNAs and the effect of lncRNA alternations on relevant mRNAs have not yet been explored. Here, we combined expression and methylation data to identify ER lncRNA involved in a single-ceRNA dysregulated network. All ER lncRNAs showed a negative correlation between their expression and promoter DNA methylation status. For EA lncRNAs, they showed hypermethylation and low expression in normal samples, while their methylation level decreased and expression was upregulated in tumor samples. For ES lncRNAs, they showed hypomethylation and high expression in normal samples, while their methylation level increased and expression was downregulated in tumor samples. Based on these principles, we totally identified 49 ER lncRNAs with a rate of 31.82% (154 lncRNAs in total) involved in DysCeNets. Through analyzing the epigenetic status of each ER lncRNA in the different samples of single cancer type, it was found that the vast majority of ER lncRNAs in single cancer were either inclined to EA (EA/ES > 2) or ES (ES/EA > 2) (Figure 5A and Supplementary Table 4). Only SVIL-AS1 in HNSC and RP11-218M22.1 in PRAD were subjected to complex regulation. These results revealed that the epigenetic regulation of lncRNAs showed a tendency in a single cancer type. Next, we explored the role of each ER lncRNA across cancer types and found that there were 22 ER lncRNAs that occurred in unique cancer type (10 EA and 12 ES lncRNAs), while 27 ER lncRNAs were regulated by DNA methylation in multiple carcinomas (Figure 5A). We further investigated the regulatory tendency of ER lncRNAs in pan-cancer based on the status of ER lncRNAs in each cancer type. For ER lncRNAs that occurred in pan-cancer, we identified consistently EA and ES lncRNA in pan-cancer (at least 75% of cancer types), including eight EA lncRNAs and six ES lncRNAs. In addition, the function of some ER lncRNAs in different cancer types was rewired. We recorded these lncRNAs as multi-ER lncRNAs based on their epigenetic regulation across cancer types. In summary, there were 18 EA lncRNAs, 18 ES lncRNAs, and 13 multi-ER lncRNAs involved in DysCeNets. From the landscape of lncRNAs, we could clearly know their epigenetic status in different cancer types. For instance, lncRNA PVT1 was EA in BRCA, HNSC, KIRC, LUSC, and PRAD cancer types, while lncRNA HCG18 was ES in only KIRC carcinoma.

To explore the effects of ER lncRNAs on ceRNA dysregulation, we next characterized the proportion of edges linked to ER lncRNAs. As shown in Figure 5B, a large scale of EA and ES lncRNAs were inclined to possess gain interaction while a small part of lncRNAs linked with loss interaction in each DysCeNet. We further compared the number of gain and loss interactions in which EA or ES lncRNAs regulated. The result showed that EA lncRNAs in KIRC, LIHC, and THCA DysCeNets owned more gain interactions than loss interactions, and ES lncRNAs in HNSC DysCeNet had the same phenomenon (Figure 5C and Supplementary Figure 3). Together, these results indicated that EA and ES lncRNAs were inclined to possess gain interaction in most cancer types (other cancers showed similar tendency but their P-values were not significant, which may be due to the limited number of ER lncRNAs in comparative groups). In addition, 17 ER lncRNAs and 945 related mRNAs were highly enriched in the core dysregulated ceRNA component above (Hypergeometric test, Figure 5D), which implied the vital function of ER lncRNAs in common cancer processes. Notably, the ER lncRNAs were also highly enriched in the top 100 nodes with the largest degree in each DysCeNet (Fisher test, Supplementary Figure 4). For example, 12 of 15 ER lncRNAs in BRCA DysCeNet were included in the top 100 of 3671 nodes (P-value = 2.87e−12, 12 of 15 vs. 100 of 3671, Fisher test, Figure 5E). All these observations suggest that the ER lncRNA might influence the stability of ceRNA interactions and further affect the cancer process.

Epigenetic alterations have been established as one of the hallmarks of tumorigenesis, and the ER lncRNAs may provide new insight into the cancer diagnosis. We first filtrated ER lncRNAs with continuous status in multiple cancers (EA samples/ES samples > 2 or <0.5 in at least three cancer types) and obtained six EA lncRNAs (PVT1, PSMD5-AS1, FAM83H-AS1, MIR4458HG, HCP5, and GAS5) and three ES lncRNAs (CTD-2201E18.3, HCG11, and AC016747.3) (Figure 6A and Table 2). The association between expression and DNA methylation of these lncRNAs has been revealed in several studies. For instance, the EA lncRNA PVT1 expression was strongly and negatively correlated with its methylation status in uveal melanoma (Xu et al., 2017). Hypomethylation within another EA lncRNA HCP5 involves a CpG site that contains a single-nucleotide polymorphism in linkage disequilibrium with HLA-B^∗27 and that controls DNA methylation at this locus in an allele-specific manner in ankylosing spondylitis (Coit et al., 2019). Overall, these pan-cancer ER lncRNAs were associated with multiple complex diseases. Next, we developed a frame to understand the relation between ER lncRNAs and cancers by connecting ER lncRNAs with cancer hallmark associated GO terms derived from a previous study (Plaisier et al., 2012). In the hierarchical model, the ER lncRNAs were firstly linked to mRNAs through ceRNA dysregulation. Then, ER lncRNA-related mRNAs were associated with biological processes and finally connected to cancer hallmarks. Through mapping hallmark genes to the ER lncRNAs, four cancer hallmarks including sustained angiogenesis, self-sufficiency in growth signals, insensitivity to antigrowth signals, and tissue invasion and metastasis were found to be associated with six ER lncRNAs in eight cancer types (Figure 6B). These results can help us comprehend how pan-cancer ER lncRNAs regulate mRNAs through ceRNA dysregulation and further influence cancer biological processes.

Cancer-related genes usually showed significant differences in cancer and normal tissues and thus could distinguish carcinoma and normal samples as biomarkers (Smolle et al., 2019; Yoon et al., 2020). We have previously discovered the important roles of ER lncRNAs in DysCeNet and their guidance on the dysregulation of ceRNA interactions. Next, we expected to predict the status of cancer through these important nodes involved in DysCeNets. To determine whether these ER lncRNAs have the diagnostic capacity in multiple cancer types, we developed a classifier based on these ER lncRNAs as described in section “Materials and Methods.” This method systematically integrated the methylation and expression levels of ER lncRNAs. If the EA lncRNAs showed hypomethylation and high expression and the ES lncRNAs showed hypermethylation and low expression in one sample (based on four cutoffs obtained by ROC curves), we would assign the maximal score to that sample (scores of samples ranged from 0 to 4). Based on the scoring principle, the cancer samples with ER status would get higher scores, while the normal samples would hold lower scores. To verify the predictive validity of the score, we calculated the areas under the curve (AUCs) of our method and found that the AUCs distributed in 0.7412–0.9917 across eight cancer types (Figure 6C). We then compared the AUCs of our method in eight cancer types with the method that simply considered methylation or expression level of the same ER lncRNAs. As shown in Figure 6D and Supplementary Table 7, the classifier performed better when two-dimensional data instead of single-platform data were considered, and the single methylation dimension-based classifier performed apparently better than the single expression dimension-based classifier. Furthermore, we randomly selected nine lncRNAs from DysCeNets and expression/methylation profiles and conducted two classifiers using our method based on these random lncRNAs for 1000 times. We found that the performance of nine ER signatures across eight cancer types was significantly higher than those nine lncRNAs randomly selected (Figure 6E). These results proved the validity of our classifier based on two-dimension data in distinguishing cancer and normal samples. To further test the predictive effect of these nine ER lncRNAs on cancer, we downloaded 15 sets of independent data from the GEO database. Due to the lack of a complete omics study like TCGA, we separately obtained eight sets of expression profiles and seven sets of methylation profiles (lacking the KIRP methylation dataset). Using these 15 external GEO datasets, we developed two classifiers that simply considered methylation or expression level of our pan-cancer ER lncRNAs as before. The AUCs of the classifiers based on the external methylation data of these ER lncRNAs ranged from 0.7124 to 1 (Figure 6F). Similar to the result of TCGA datasets, the lncRNA methylation status separated the tumor and normal sample better than expression data (Figures 6G,H). In conclusion, these nine ER lncRNAs could serve as predictive biomarkers for multiple cancers. Moreover, the prediction effect of ER lncRNAs would reach best when combining expression and methylation data. For cases of lacking paired omics data, it is better to utilize methylation level data than expression level data to construct the classifier.

The ability of ER lncRNAs to cancer diagnosis has been examined herein before. We then evaluated the effect of these ER lncRNAs on cancer progression, that is, to determine whether patients with different OS could be distinguished based on expression or methylation data of these ER lncRNAs. As described in section “Materials and Methods,” the TCGA paired samples were regarded as a train set, and other TCGA samples were regarded as the test set. We first estimated the survival difference between high and low score groups in the train set and then validated the effectiveness of these ER lncRNAs in the test set using the parameters [including median score, HR, and se(HR)] of the train set. The low survival score based on ER lncRNAs predicted poor prognosis in most cancer train sets, and there were three test sets that performed well, including KIRC expression dataset, KIRP, and LIHC methylation datasets. The genome information and related CpG probes of the ER lncRNAs that occurred in three datasets are shown in Figure 7A. There are six ER lncRNAs whose methylation levels in the promoter region changed, thus affecting lncRNA expression in these three datasets, including four EA lncRNAs (FAM83H-AS1, MIR4458HG, HCP5, and GAS5) and two ES lncRNAs (HCG11 and AC016747.3). In particular, FAM83H-AS1, HCP5, GAS5, and HCG11 were identified in the pan-cancer core component mentioned above. As shown in the left panel of Figures 7B–D, the survival score based on ER lncRNAs could predict the prognosis of cancer patients. Next, the multivariate Cox regression model was used to verify the prognostic efficacy of multiple clinical factors. The survival score based on ER lncRNAs was found to be positive as a protective factor, and the stage conversely showed prognostic efficacy as a risk factor (all P-value < 0.05, Figures 7B–D, middle panel). Then, patients with stage I and II were merged into the low-stage group, and patients with stage III and IV were merged into the high-stage group. We integrated the score based on ER lncRNA and stage information and estimated the survival curves by the similar method above. It was found that the prognosis capacity was stronger than the method using molecular-level data alone. In particular, in the case of the high-stage group, the OS of patients with high scores showed significantly better than those with low scores (Figures 7B–D, right panel and Supplementary Figure 5). These results implied that the combination of molecular and clinical data could better predict the survival of these cancer patients. Collectively, the data suggest that ER lncRNAs involved in ceRNA dysregulated network could not only act as cancer diagnostic markers but also influence cancer progression and outcome in some cancer types.