Systematic profiling identifies PDLIM2 as a novel prognostic predictor for oesophageal squamous cell carcinoma (ESCC)

2.1 Data mining in GEO datasets

The normalized data of one previous Affymetrix HG-U133A 2.0 gene expression array that compared gene expression in 17 micro-dissected ESCC tumours and matched adjacent normal tissue pairs were downloaded from: (https://www.ncbi.nlm.nih.gov/sites/GDSbrowser?acc=GDS3838).12 The data were loaded into NetworkAnalyst for re-analysis, which is a visual analytics platform for comprehensive gene expression profiling.13 Only the genes had log2 fold change ≥2 and adjusted P < 0.05 were identified as the candidate DEGs.

2.2 Data mining in TCGA-ESCA

The level-3 data in TCGA-ESCA were acquired with the use of the UCSC Xena browser (https://xenabrowser.net/).14 Only the ESCC subset (TCGA-ESCC) that with RNA-seq data of gene expression was extracted and used in this study. None of the patients in this subset received neoadjuvant treatment. After screening with these criteria, 95 cases of primary ESCC were identified. The clinicopathological data, including age at initial diagnosis, gender, histological grade, the history of oesophageal cancer, pathological nodal (pN) status, pathological stages, reflux history, smoking history, post-operative drug therapy, radiation therapy, alcohol consumption per week (defined as frequency of alcohol consumption/week * amount of alcohol consumption/day) and OS survival data were downloaded for re-analysis.

The genomic data of the ESCC cases, including the RNA-seq data of total gene expression, the RNA-seq data of exon expression of a specific gene, the methylation status, DNA copy number alterations (CNAs) and somatic mutations of targeting genes were also extracted. Briefly, the methylation status of target genes was examined using Infinium HumanMethylation450 BeadChip, which covers over 450,000 methylation sites across the genome per sample. Beta values were calculated to determine the methylation level, with the formula: β = M/(U+M+100). In the equation, M and U refer to the fluorescence level of the methylation probe and unmethylated probe respectively. Genomic CNAs were calculated by an algorithm called gene-level thresholded Genomic Identification of Significant Targets in Cancer 2.0 (GISTIC2), which aims to assign discrete numbers to the gene fragments by thresholding the data.15 The alterations in a gene were generally classified into five groups, including copy neutral (0), low-level copy gain (+1), high-level amplification (+2), heterozygous loss (−1) or homozygous deletion (−2). Somatic mutations include both single-nucleotide polymorphisms and (SNPs) small insertions and deletions (INDELs).

2.3 In silico analysis of the alternative transcription of PDLIM2

The alternative transcripts of PDLIM2 in oesophageal cancer were analysed using ISOexpresso (http://wiki.tgilab.org/ISOexpresso/), which is a web-based platform for isoform-level expression analysis among samples in TCGA.16 The genes were matched with HGNC gene, Ensembl release 74 and UCSC refGene, while the isoform information of the genes is collected from UniProt ID mapping data and UCSC knownGene tables 16. The relative expression level of each transcript was quantified using transcripts per kilobase million.

2.4 Statistical analysis

Data integration and analysis were conducted using spss 25.0 software package (SPSS Inc, Chicago, IL), together with GraphPad Prism 7.04 (GraphPad Inc, La Jolla, CA). The difference in gene expression between groups was compared using unpair t test with Welch's correction. Receiver operating characteristic (ROC) curve was generated and area under the curve (AUC) was calculated to assess the predictive value of PDLIM2 and its exon expression for OS. Log-rank test was performed to examine the significance of the difference between the Kaplan-Meier OS curves. Stepwise regression was conducted to compare the predictive value of total PDLIM2 expression and its exon 7/8/9/10 expression in terms of OS. The prognostic value of PDLIM2 and its exon expression were assessed using the univariate and multivariate Cox regression models. Linear regression analysis was performed and the Pearson's correlation coefficient was calculated to assess the correlation between total PDLIM2 expression and the expression of its exons or the methylation status of the CpG sites in its DNA locus. P < 0.05 was considered statistically significant.

3.1 Screening of significantly dysregulated genes in ESCC

Using the microarray data from GDS3838, we screened the significant DEGs between the 17 cases of ESCC tumour tissues and matched normal tissues, with the following criteria: log2 fold change ≥2 and adjusted P < 0.05 (Figure 1A). Results showed that there were 59 up-regulated and 163 down-regulated genes in ESCC compared to adjacent normal tissues (Figure 1B,C). The details of the 222 DEGs were provided in Table S1.

3.2 Screening of OS-related DEGs in ESCC

In the ESCC subset of TCGA-ESCA, we found that 220 of the 222 DEGs had RNA-seq data of gene expression. Lymph nodal invasion is a critical hallmark of the progression of ESCC and direct increases the risk of unfavourable survival.17 Then, we examined the expression profiles of the 220 DEGs in ESCC patients grouped according to their nodal invasion status (Figure 2A) and OS status (Figure 2B). There were nine genes that were significantly dysregulated in nodal positive cases, including BBOX1, ACPP, ECT2, COBL, TRIP10, GPD1L, PDLIM2, KCNS3 and PLAC8 (Figure 2A), while the dysregulation of 18 genes (COL10A1, COL11A1, ABLIM1, THBS2, PDLIM2, VCAN, COL1A2, COL5A2, MMP2, POSTN, ATAD2, DUSP5, ZNF185, SLC39A14, MFAP2, MGLL, KIF14 and BLNK) were associated with OS status (Figure 2B). Interestingly, among these genes, only PDLIM2 was associated with nodal invasion and survival outcome at the same time. The nodal positive cases and dead cases had significantly down-regulated PDLIM2 expression compared to the respective control groups (P = 0.01 and P < 0.01 respectively) (Figure 2C,D).

We also generated ROC curves and calculated the AUCs of the 18 OS-related genes in Figure 2B, by setting living as the status variable. The AUCs of PDLIM2 (0.684) ranked third among the genes, while ATAD2 and COL10A1 showed higher AUCs (0.692 and 0.691 respectively) (Figure S1). However, we then checked their expression profile and found that these two genes were significantly up-regulated in ESCC compared with adjacent normal tissues (Table S1) and thus were abandoned for analysis.

The clinicopathological and survival data extracted from TCGA-ESCC were provided in Table S2. Ninety-five ESCC patients were divided into two groups according to median PDLIM2 expression. Their clinicopathological parameters were compared in Table 1. Results showed that these two groups had similar clinicopathological profiles, expect the ratio of survival. The high PDLIM2 expression had a significantly lower ratio of death (9/48 vs 23/47, P < 0.01) (Table 1).

By generating Kaplan-Meier survival curves according to the median expression of PDLIM2, we found that the high PDLIM2 expression group had significantly longer OS (P = 0.01) (Figure 3A). Univariate analysis showed that male patients, advanced pathological stages and decreased PDLIM2 expression were the risk factors of unfavourable survival (Table 2). The following multivariate analysis confirmed that total PDLIM2 expression was independently associated with better OS in ESCC patients (HR: 0.64, 95% CI: 0.43-0.95, P = 0.03) (Table 2), after the adjustment for gender and pathologic stages.

3.3 The expression of PDLIM2 exon 7/8/9/10 showed better prognostic value than total PDLIM2 expression

The accuracy of a prognostic marker can be generally divided into five categories on the basis of the AUC values 0.50-0.60 = fail，0.60-0.70 = poor, 0.70-0.80 = fair, 0.80-0.90 = good and 0.90-1 = excellent.18 Although we found that PDLIM2 expression might have potential prognostic value in terms of OS, ROC analysis showed that its expression only had an AUC value of 0.684 (Figure S1), which suggests a poor accuracy of a prognostic test. PDLIM2 gene contains 12 exons and spans about 20 kb in human genome.19 Besides, alternative splicing and subsequent transcript variants encoding multiple isoforms have been observed for this gene.19 Therefore, the prognostic value of total PDLIM2 expression might be impaired by the variation of its transcripts. By checking alternative splicing events in TCGA-ESCA, we found that among the four major transcripts of PDLIM2 (uc003xbx.2, uc003xby.4, uc003xca.3 and uc003xcc.2) (Figure 3B), the canonical uc003xby.4 isoform, which contains exon 1-10 is the dominant transcript in TCGA-ESCA (Figure 3C). Then, we checked the expression profile of PDLIM2 exons and analysed their correlation with total PDLIM2 expression (Figure 3D,E). Data in TCGA showed that among the 12 exons of PDLIM2, exon 7/8/9/10 and exon 11/12 tend to be transcribed together and their expression data were provided as a whole. Therefore, the expression of these two segments was treated as two expression units. Correlation analysis showed that except chr8:22452043-22455538:+, which represent the expression of exon 11/12, all other exons were very strongly correlated with total PDLIM2 expression (Pearson's r > 0.9, Figure 3D,E). The exon region, chr8:22446663-22451810:+, which represents the expression of exon 7/8/9/10 had the strongest correlation with total PDLIM2 expression (Pearson's r = 0.97; Figure 3E). ROC analysis showed that the expression of exon 7/8/9/10 had the highest AUC value (0.724) (Figure 3F, bold purple), which suggests that it might be a fair prognostic marker. Compared with the AUC value of total PDLIM2 expression (0.684), this region might be a better marker. To test this hypothesis, we further performed a stepwise regression analysis by setting OS as the dependent variable. Results showed that the expression of exon 7/8/9/10 is a better predictor in the model than total PDLIM2 expression in predicting the OS of ESCC patients (Table S3). Multivariate analysis further confirmed the independent prognostic value of exon 7/8/9/10 expression in terms of OS (HR: 0.43, 95% CI: 0.22-0.83, P = 0.01) (Table 2).

3.4 In silico analysis of the potential genetic and epigenetic alterations associated with PDLIM2 dysregulation

In TCGA-ESCC, all 95 ESCC cases included in this study had PDLIM2 CNA data available. Data showed that PDLIM2 CNAs were quite common. Among the 95 ESCC cases, 18 cases (18.9%) had PDLIM2 amplification (+1), 45 cases (47.4%) had heterozygous deletion (−1) and 4 cases (4.2%) had homozygous deletion (−2). Only 28 cases (29.5%) were copy neutral (Figure 4A). Then, we examined the correlation between PDLIM2 expression and its DNA CNAs. PDLIM2 amplification did not necessarily result in up-regulated transcription, compared to copy neutral cases (P = 0.82, Figure 4B). In comparison, its DNA deletion was associated with significantly lower gene expression, compared with the copy neutral cases (P < 0.01, Figure 4B). Kaplan-Meier survival curves showed that the group with PDLIM2 DNA deletion had significantly inferior OS, compared to the amplification/copy neutral group (P < 0.01, Figure 4C). We also checked the mutation status of PDLIM2 in 94 ESCC cases with mutation data available (Figure S2). Among 94 ESCC patients, two had missense mutations (Figure S2A). However, no survival difference was observed between the mutation and no mutation group (Figure S2B).

We then checked methylation 450 k data of 95 ESCC cases in TCGA. The methylation status of 19 CpG sites in PDLIM2 DNA locus was measured (Figure 5A). The correlation between the methylation level of PDLIM2 CpG sites and total PDLIM2 expression was calculated (Figure 5B). Results showed that the methylation level of two CpG site (cg23696886 and cg20449614) had a moderately negative correlation with PDLIM2 expression (Pearson's r = −0.47 and r = −0.46 respectively; Figure 5B). cg23696886 is located at the intron before exon 2, while cg20449614 is in exon 2 (red-dotted frame, Figure 5A), where the PDLIM2 promoter maps to Ref.19.

Then, we analysed the correlation between PDLIM2 CNAs and the methylation of the two CpG sites. Results showed that the copy loss group (−1/−2) had significantly higher methylation at these two CpG sites (P = 0.02 and P = < 0.01 respectively, Figure 5C,D). To reduce the potential mixing effect of CNAs and methylation on gene expression, we analysed the correlation between PDLIM2 expression and the methylation of the two CpG sites in copy neutral/amplification group and copy loss group respectively (Figure 5E,F). Results showed that, in copy neutral/amplification group, there were still moderately negative correlations (Figure 5E). However, the correlation coefficients were much smaller in the DNA −1/−2 groups (Figure 5F). These findings imply that DNA hypermethylation might be a potential mechanism leading to PDLIM2 down-regulation in copy neutral/amplification group. In comparison, PDLIM2 copy loss might be the dominant cause of reduced gene expression in the copy loss group.