Raw human microarray expression profiles were retrieved and downloaded from Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih. gov/geo/). First of all, the key words 'high-grade serous ovarian cancer' and 'chemotherapeutic resistance' were used to search for samples. Secondly, as more lncRNAs could be detected using the Affymetrix Human Genome U133 Plus2.0 micro-array (Affynetrix; Thermo Fisher Scientific, Inc., Waltham, MA, USA), the samples not from that platform were filtered. Finally, the samples with data associated with chemotherapeutic response were selected. In total, 328 microarray profiles with corresponding clinical information [i.e. age, tumor grade, International Federation of Gynecology and Obstetrics (FIGO) stage (17), residual tumor size and platinum sensitivity] were downloaded. The accession numbers of GEO for the datasets were as follows: GSE63885 (18) (n=70); GSE51373 (19) (n=28) and GSE15372 (20) (n=10). Another dataset GSE9891 (21) (n=220) was used as a GEO validation dataset. The expression levels were normalized using the robust multi-array average (RMA) algorithm (22). Batch effects were removed using the Combat algorithm (23).

The expression profiles and clinical information of 233 patients with HGS-OvCa were downloaded from The Atlas of Noncoding RNAs in Cancer (24) and cBioportal (25) databases, respectively, as The Cancer Genome Atlas (TCGA) validation dataset. The expression levels were processed using the same methods as aforementioned. All datasets were analyzed to determine the differences in expression (Table I).

LNCipeida (26) is a publicly available database for annotated human lncRNA transcript sequences and structures. The public repository produced by Du et al (27) provides a resource of clinically relevant lncRNAs for the development of lncRNA biomarkers and the identification of lncRNA therapeutic targets. The present study integrated the records of lncRNAs from GENCODE (28), LNCipedia, NONCODE (29), TCGA and the public repository by Du et al (27). In total, 24,372 lncRNAs were included in the present study. To avoid inaccurate annotations, the probes were cross-referenced by gene name and Ensembl transcript ID (release 89). For multiple probes of a single lncRNA, the median expression of the unique transcript ID was preserved. Finally, the expression profiles of 6,955 lncRNAs were obtained.

The platinum-based sensitivity of patients with HGS-OvCa varies greatly, and the DFS ranged from 180 days to several thousand days. In this context, unsupervised hierarchical clustering based on Pearson distance for platinum-sensitive samples were performed using the ConcensusClusterPlus package in R (30).

The differential expression analysis was performed using R/Limma package (31). The common differentially expressed lncRNAs between each sensitive subgroup and resistant group were screened [differential expressed (DE) set A]. The differentially expressed lncRNAs between the complete sensitive group and the resistant group were obtained (DE set B); DE sets A and B were then merged.

To investigate the platinum-based chemoresistance-associated lncRNAs and the clinical characteristics of patients with HGS-OvCa, an independent univariate logistic regression was performed on 6,955 lncRNAs and clinical characteristics. The lncRNAs and clinical characteristics with P<0.05 were considered statistically significant and associated with chemoresistance. The survival package in R (32) was used to investigate DFS-associated lncRNAs. The Kaplan-Meier survival curves of highly expressed and weakly expressed lncRNAs were constructed. The median expression level was used as a cut-off for determining high and low expression.

The lncRNAs and clinical characteristics that were statistically significant in the univariate logistic regression were used as candidate parameters for constructing the multivariate logistic regression model. A best multivariate logistic regression model was generated after fitting the model thousands of times. The expression levels of each lncRNA and its coefficient in the best multivariate logistic model were used to calculate the platinum-based chemoresistance risk score of patients with HGSC-OvCa. The risk score was computed by the following formula: where n is the number of prognostic lncRNAs in the model, 'expr_i' is the expression level of lncRNA_i and coef_i is the estimated regression coefficient of lncRNA_i in the multivariate logistic regression model. Patients who have higher risk scores are expected to have higher probability of platinum-based chemoresistance. The median risk score was used as a cut-off to divide the patients in the GEO validation dataset into low-risk and high-risk groups. Furthermore, the multivariate logistic regression analysis was performed to test whether the risk score was independent of clinical covariates.

The 100-gene set containing four gene signatures corresponding to four HGS-OvCa subtypes (differentiated, immunoreactive, proliferative and mesenchymal) (33) were used to train the support vector machine for subtype classification. The univariate and multivariate logistic regression were performed to investigate whether the HGS-OvCa subtypes were associated with platinum-based chemoresistance, and whether these subtypes associated with our signature, respectively.

To assess the sensitivity and specificity of the risk scores in prediction of chemoresistance in patients with HGS-OvCa, ROC (34) and ROCR package (35) in R software were used to construct the receiver operating characteristic curve and perform the area under curve (AUC) analysis, respectively. Furthermore, the predictive performance of the risk score and three signatures previously developed by Zhou et al (36), Liu et al (37) and TCGA mRNA signatures (38) was compared.

Microarrays and RNA-seq data were used to validate the expression of lncRNAs. The RNA-seq data of 136 patients with HGS-OvCa and platinum-based chemosensitivity was derived from the Gaucher Disease Advisory Committee (39). The reads per kilobase per million mapped reads were obtained by chromosome location (GRCH37, hg19) of lncRNAs and normalized using Z-scores.

First, the gene expression in the Cancer Cell Line Encyclopedia (CCLE) (40) database had been pre-processed using the RMA algorithm. The chemotherapeutic status of 52 ovarian cancer cell lines was investigated. The data regarding the cell lines of clear cell, granulosa cell, undifferentiated cell, mixed cell, endometrioid cell, Brenner cell, mucinous cell and cross-contaminated cell were filtered. The cell lines that were treated with cisplatin or carboplatin under-went >3 cycles of treatment. The cell lines exhibited resistance to platinum and the degree of resistance varied. Secondly, gene expression levels in 46 ovarian cancer cell lines with or without cisplatin treatment were obtained from GSE47856 (41). To investigate the changes in the expression of lncRNAs that was induced by platinum, cell lines from the CCLE were analyzed using the Integrative Genomics Viewer (IGV) software (version 2.3) (42). In addition, cell lines from GSE47856 with ≥3 replicates were analyzed. For multi-probe lncRNA, the median expression was calculated.

To investigate the biological processes and cellular components implicated in the drug resistance mechanisms, the enriched Gene Ontology (GO) (43) terms of 7 lncRNAs were identified using the NONCODE data-base and were visualized using Cytoscape (version 3.2.1) (44). The potential functions of the lncRNAs were further analyzed using non-coding RNA in the Drug Resistance (ncDR) data-base (45); P<0.05 was used to identify resistant compounds of lncRNAs in cancer.

The empirical Bayes method and a two-tailed Student's t-test were used in differential expression analysis. The log-rank test was used for survival analysis, with Kaplan-Meier curves. Clinical characteristics and molecular subtypes of three datasets were tested using the χ² tests or two-way analysis of variance, where appropriate. In multivariate logistic regression, the χ² test was used to add or filter features and avoid overdispersion. The features were also filtered using the methods of Watt et al (46). P<0.05 was considered to indicate a statistically significant difference. All statistical analyses were conducted using the R software (version 3.2.1).

The results of unsupervised hierarchical clustering were obtained when clustering center k was set from 2 to 5 (Fig. 1). The specimens were perfectly divided into two clusters (Fig. 1A); a third cluster was observed when k=3 (Fig. 1B). However, fourth stable clusters were not found when the number of cluster centers k continued to increase (Fig. 1C and D). These results indicate that there may be three stable subgroups with different expression patterns in chemosensitive patients.

In the present study, 386 differential expressed lncRNAs were obtained subsequent to merging 262 lncRNAs in DE set A and 365 lncRNAs in DE set B (Fig. 2A). In univariate regression, 337 lncRNAs (288 plus 43 plus 7 minus 1) were associated with platinum-based chemoresistance after merging the lncRNA transcripts corresponding to the same lncRNA gene, 50 lncRNAs (43 plus 7) (Fig. 2B) were associated with DFS (log-rank P<0.05). Next, a 7-lncRNA signature was identified (Table II and Fig. 2C), which was associated with platinum-based chemoresistance in the univariate and multivariate logistic regression model. The risk scores based on the 7-lncRNA signature for the prediction of platinum-based chemoresistance in patients with HGS-OvCa were also calculated. According to the risk scores, patients in the GEO training dataset were divided into low-score and high-score groups using the median risk score as a cut-off in the multivariate logistic regression model (Fig. 2D). The high-score group exhibited a higher probability of drug resistance [univariate model: Odds ratio (OR), 2.675; 95% confidence interval (CI), 1.841-3.897; P=6.36×10⁻⁷; multivariate model: OR, 2.859; 95% CI, 1.801-5.257; P=9.36×10⁻⁵). The results of Kaplan-Meier curves also indicated that the high risk score was significantly associated with the poor survival of patients with HGS-OvCa and drug resistance (n=47; log-rank, P<0.05; Fig. 2E).

In the multivariate logistic regression model of the training datasets, platinum-based chemosensitivity was a dependent variable. Platinum-based chemosensitivity has multiple covariates, including tumor stage, grade, tumor size, molecular subtypes and risk score. However, the results indi-cated that the risk score was an independent variable when adjusting using the covariates as aforementioned (OR, 2.859; 95% CI, 1.801-5.257; P=9.36×10⁻⁵; Table III).

The results of the univariate and multivariate logistic regression analysis for the GEO validation set (n=220) and the TCGA validation set (n=233) confirmed that the risk score of the training set was associated with platinum-based chemore-sistance of patients with HGS-OvCa (univariate model of the GEO dataset: OR, 4.190; 95% CI, 2.974-4.685; P<4.68×10⁻¹⁴; multivariate model of the GEO dataset: OR, 4.210; 95% CI, 2.967-6.366; P<1.15×10⁻¹³; univariate model of TCGA dataset: OR, 2.315; 95% CI, 1.852-2.956; P=2.33×10⁻⁵; and multivariate model of TCGA dataset: OR, 2.514; 95% CI, 1.817-3.354; P=2.65×10⁻⁵; Table III). Furthermore, clinical characteristics and HGS-OvCa subtypes (differentiated, immunoreactive, proliferative and mesenchymal) were not associated with platinum-based resistance in two validation datasets. Collectively, these results suggest that risk score has the potential to be an independent predictor for chemoresistance of HGS-OvCa cancer.

The risk score of the 7-lncRNA signature indicated a high sensitivity and specificity in predicting platinum-based chemoresistance in the GEO training (AUC=0.848), GEO validation (AUC=0.901) and TCGA validation (AUC=0.818) datasets (Fig. 3A). Compared with the lncRNA signature of the study by Liu et al (37), the lncRNA signature developed in the study by Zhou et al (36) and the TCGA mRNA signature, the 7-lncRNA signature exhibited an improved predictive performance (Fig. 3B–D).

A total of 7 lncRNAs were observed in the GEO validation dataset, while F11-AS1 and RP11-348N5.7 were not observed in the TCGA validation dataset and RNA-seq dataset, respectively. However, all of the 7 lncRNAs were covered by validation datasets. Notably, the expression levels of 7 lncRNAs were concordant in at least two validation datasets (P<0.05; Fig. 3E).

With several cycles of cisplatin or carboplatin treatment, ZBED1-AS1 was upregulated, while F11-AS1 was downregulated by platinum in OV56 and OVKATE cell lines (P<0.05; Fig. 3F). Growth arrest specific 5 (GAS5) was downregulated by cisplatin in A2008, DOV13B and HeyA8 cell lines (P<0.05; Fig. 3G). This indicates that these lncRNAs may be involved in drug resistance mechanisms. Therefore, the functional prediction analysis of 7 lncRNAs based on NONCODE database was further performed. The results showed that a total of 42 GO terms were involved in the biological functions of 7 lncRNAs, which were distributed in 'biological processes' (91.3%) and 'cellular components' (8.7%). Furthermore, 16 of the 42 GO terms was associated with >1 lncRNA (Fig. 4). Based on ncDR, the aberrant regulation of lncRNAs was predicted to be resistant to dozens of compounds across various cancer types. For example, the upregulation of ZBED3-AS1 is associated with resistance to cisplatin, dasatinib and bicalutamide in low-grade glioma. The downregulation of RNF144A-AS1 is associated with resistance to dacinostat in sarcoma, and the downregulation of GAS5 is associated with resistance to cisplatin and cytarabine in pancreatic adenocarcinoma.