Clinicopathological value of long non-coding RNA profiles in gastrointestinal stromal tumor

GIST gene expression datasets preparation

Microarray data of GSE20710_RAW (Astolfi et al., 2010), GSE56670_RAW (Killian et al., 2014), GSE8167_RAW (Yamaguchi et al., 2008) and GSE17743_RAW (Ostrowski et al., 2009) were directly downloaded from Gene Expression Omnibus databases. These datasets corresponded to all available public datasets fulfilling the following criteria: available gene expression data obtained using the same chip platform (U133 Plus 2.0 chips, Affymetrix, Santa Clara, CA, USA) with raw data CEL files, and some associated clinicopathological data were available. SDH-deficient gastrointestinal stromal tumors (Killian et al., 2014) were excluded, because of the small percentage of GIST and totally different gene-expression style (Janeway et al., 2011; Miettinen et al., 2011). Gene expression’s background correction, normalization and summarizing were performed using Robust Multichip Average (RMA) (Irizarry et al., 2003b) by TCA software (Affymetrix Co., Santa Clara, CA, USA) (Irizarry et al., 2003a).

Microarray data processing and lncRNA profile mining

The approach of lncRNA profile mining mainly referred to Hu et al. (2014), Zhu et al. (2016) and Zhang et al. (2012). Briefly, the Affymetrix HG-U133 Plus 2.0 probe set IDs were mapped to the NetAffx Annotation Files and gene biotypes were further validated by BioMart databases (biomaRt R package). Based on the Refseq transcript ID and/or Ensembl gene ID, we only retained non-coding protein genes and further filtered them by eliminating pseudogenes including microRNAs, rRNAs and other short RNAs such as snoRNAs, snRNAs and tRNAs. Finally, 2,448 annotated lncRNA transcripts with corresponding Affymetrix probe IDs were generated. Some lncRNA signatures were well confirmed by previous study (Hu et al., 2014; Zhu et al., 2016).

LncRNA markers selection and expression style between subtypes

The subtypes of GIST were identified in row median centered microarray datasets using the non-negative matrix factorization (NMF) algorithm (Brunet et al., 2004). We used the Comparative Marker Selection algorithm (Gould et al., 2006) by GenePattern (https://genepattern.broadinstitute.org/gp). One-vs.-all comparisons were performed to identify differentially expressed genes that can discriminate between distinct lncRNA clustering subtypes. Strict criteria were used to select the lncRNA markers, including fold change of gene expression > 2, FDR (BH) < 0.05, Feature-Specific P-value < 0.05.

Identification of biological pathways distinguishing each subtype

We performed GSEA through the JAVA program (http://www.broadinstitute.org/gsea) by using MSigDB C2 CP: Canonical pathways gene set collection (1,320 gene sets available), and visualized the GSEA outputs in Cytoscape (version 3.4). Then we used the Enrichment Map software to identify the biological processes distinguishing one subtype from others. In order to simplify the network map, we chose a strict threshold of gene-set permutations with an FDR cutoff of 5% and P-value cutoff of 1% in Enrichment Map software.

SEER database exploration

SEER data was used to identify patients that were diagnosed as GISTs from 1973 to 2014 (Cronin, Ries & Edwards, 2014). The exclusion criterions included: age < 18, no evaluation of histological type, not the first diagnosed malignancy, an unknown cause of death or survival month. The primary outcomes of interest were overall survival (OS), which was calculated from the date of diagnosis to the date of attributed death.

To better explore gender influence on the survival rate, the propensity-score-matched analyses were performed with equal factors which may influence the prognosis. Pairing criteria included tumor sizes, grade, age, treatment. Survival function estimation was performed with the Kaplan–Meier method and the resulting curves compared with the log-rank test. All statistical analyses were computed using SPSS version 24 (IBM Corporation, Armonk, NY, USA).

lncRNA-mRNA co-expression network construction and gene annotation

R package LINC provided methods to compute co-expression networks of lincRNAs and protein-coding genes. Biological terms associated with the sets of protein-coding genes predicted the biological contexts of lincRNAs according to the ‘Guilty by Association’ approach (Gillis & Pavlidis, 2012). This R package LINC applied this idea on arbitrary expression matrices (Yu et al., 2015) and could reveal which functions, pathways or compartments were associated with the lncRNA co-expressed genes.

Somatic copy number alterations (CNA) frequency profiles

We downloaded CNA array data of patients via the NCBI Gene Expression Omnibus (accession number GSE20709), and analyzed the raw copy number data for each GIST sample provided by array comparative genomic hybridization (CGH) by using the Copy Number Inference Pipeline for SNP6 CEL files and GISTIC2 algorithm (Mermel et al., 2011) via web server GenePattern.

Classification of GISTs into four subtypes based on extracted lncRNA profiles

Due to the limited chipset data available for GIST, only two datasets GSE8167_RAW (Yamaguchi et al., 2008) and GSE17743_RAW (Ostrowski et al., 2009) were adapted as main data source in this study. The clinicopathological, immunohistochemical, and genetic characteristics of the 61 cases of GIST used in the chipset analysis are summarized in Table S1.

To explore the overall gene expression pattern, we first performed unsupervised analysis of all 54,613 probe sets and with 20,647 collapsed genes. Consensus NMF clustering (Brunet et al., 2004) separated the 61 GISTs into three principal classes (max coef. = 0.9929, when k = 3). To eliminate probes that had little or no variation across samples, we then keep the 50% most varying probes to run the non-negative matrix factorization (NMF) algorithm again. Next, we repeated the same unsupervised analysis with extracted lncRNA probe sets. According to the cophenetic coefficient scores from the consensus NMF clustering (k = 2 to k = 6), we chose to cluster 61 GISTs into four subtypes (max coef. = 0.9932, when k = 4) (Figs. 1A–1F). The four clustering subtypes were designated as G1, G2, G3, and G4. Metagenes analysis and linkage tree of extracted lncRNAs ordered by k = 4 clustering subtypes were showed in Figs. 1G–1I. Thus, as previously reported (Brunet et al., 2004), NMF with consensus clustering and hierarchical dendrogram gave strong evidence for four classes split of the lncRNAs with a correspondingly high cophenetic coefficient.

LncRNAs clustering GIST subtypes were associated with tumor size and gender

Subtypes clustered by global gene expression (max coef. = 0.9929, k = 3) did not show any meaning to the clinicopathological value, as well as extracted most varying probes. However, clustering by extracted lncRNA profile showed distinct clinical relevance. Average tumor diameter in G1 or G2 subtype were small size (10.45 ± 7.90 mm) and were referred as benign or early stage the GIST. G3 or G4 subtype showed larger size (52.17 ± 30.89 mm), and were referred as malignant GIST tumor. Fisher’s exact test showed that there were significant differences between class G1 and class G2 in the frequency of gender, as well as between subtypes G1/G2 and subtypes G3/G4 in the frequency of tumor size. LncRNAs’ expression profile did not correlate with c-kit and PDGFRA mutation status, as well as age and tumor location (Table 1).

Previous research showed that methylation pattern of GISTs arising from the stomach segregated apart from GISTs from the intestine (Huang et al., 2016). Unlike methylation pattern indicating that GIST tissue of origin has a strong imprint on the epigenetic profile of this disease (Huang et al., 2016), lncRNA expression profile does not have correlation to stomach or intestine origin. Instead, the gender has significant influence on clustering by lncRNA expression profile, especially in small size GIST. This may be due to that the same cajai cells originate from alimentary track no matter where the GISTs locate. The high expression abundance of gender-related lncRNA, such as XIST, FTX and TTY among the limited lncRNA-probes may also give much weight coefficient for this gender-oriented clustering.

Female GIST patients showed favorable survival time by SEER database

Female and male patient showed distinct lncRNA expression profile especially in early stage of GIST and rapidly lost the gender-oriented clustering during the tumor progress, which suggest gender may be prognosis relevant as well as tumor size. So, we further explored the SEER database to analyze the survival rate differences between male and female GIST patients.

A total of 7,983 eligible GIST patients were enrolled during the 41-year study period. The baseline characteristics of GIST patients stratified by gender were summarized in Table 2. The median survival time was 100 months (95% CI [94.16–105.84]) in male and 130 months (95% CI [121.38–138.62]) in female GIST patients, female patients had a better overall survival (OS) than male patients. (HR = 1.277, P < 0.0001, Fig. 2A).

Furthermore, 2,567 female and 2,567 male patients were perfect matched pairs for better analyses, baseline clinicopathological characteristics after perfect propensity-score-matched (PSM) analyses were also showed in Table 2, the differences between two groups were eliminated with regard to primary tumor site, age, therapy, marital status, race, grade, tumor size (all factors P = 1.0). After PSM, the median survival time of male patients was 106 months (95% CI [98.90–113.10]), while matched female patients were 162 months (95% CI [143.26–180.74]). Consistently, female patients had a significantly longer overall survival time than male patients in GIST (HR = 1.470, P < 0.0001, Fig. 2B).

Identification of biological processes and signaling pathways distinguishing certain subtype by gene set enrichment analysis (GSEA)

Significant gene sets were visualized as interaction networks with Cytoscape and Enrichment Map (Merico et al., 2010). For analysis of male and female expression differences, we chose 18 female vs. 18 male matched samples with equal tumor size for Cytoscape and Gene Set Enrichment Analysis. Compared to female patients, interferon signaling, cytokine signaling in immune system, immune checkpoint (PD1, TCR, CTLA4) and circadian expression were significantly up-regulated in subtype of male patients (Figs. 3A–3D). However, pathways of oxidative phosphorylation, TCA cycle and respiratory electron transport were up-regulated in subtype of female patients compared to male patients (Figs. 3A, 3E–3G).

Subtypes G1 and G2, G3 and G4 were combined for analysis, and referred as G1/2 (small size) and G3/4 (large size) respectively. Compared to subtype G1/2, most signaling pathways showed up-regulated in subtype G3/4, including DNA repair, carbohydrates metabolism, protein metabolism and lysosome pathway which might be involved in carcinogenesis or development of GIST (Figs. 3H, 3L–3N). A few signaling pathways such as olfactory signaling, olfactory transduction signaling and hedgehog pathway were up-regulated in subtype G1/2 compared to G3/4 (Figs. 3I–3K).

These analysis results suggested that different gender and lncRNAs subtypes of GISTs might arise from distinct abnormalities of biological signaling pathways.

lncRNA Markers selection and associated expression pattern between four subtypes

A total of 32 featured lncRNAs were identified to be significantly different expression between subtype G1/2 and subtype G3/4 by the Comparative Marker Selection algorithm, and their expression style were plotted in Fig. 4A. As lncRNAs’ functions are closely associated with their transcript abundance (Du et al., 2013), six lncRNAs LRRC75A-AS1, HYMAI, MALAT1, NEAT1, XIST and FTX with high expression value were recruited for further study.

The expression of most featured lncRNAs were significantly downregulated in GIST progression subtype G3/G4 compared with subtype G1/G2, whereas LRRC75A-AS1 showed the opposite upregulation (Fig. 4A). High expression of LRRC75A-AS1 was positively associated with tumor diameters and seemed to be oncogene in GIST. XIST was specifically expressed in female GIST patients and was dramatically down-regulated in G3/G4 subtypes, which suggested XIST was significantly inhibited during tumor progression and it seemed to have a protective effect in female GIST patients. The expression of HYMAI, NEAT1, XIST and FTX were negatively correlated with tumor diameter and significantly down-regulated during GIST progression (Table 3). Overall, XIST down-regulation may suggest the malignant transformation for female GIST patients, while down-regulation of HYMAI, NEAT1 and FTX, or up-regulation of LRRC75A-AS1 may suggest the malignant transformation for both female and male patient, which may be the promising biomarkers irrespective of small intestinal or gastric origin.

Protein-coding genes co-expressed with LRRC75A-AS1 and functional prediction

To further investigate the potential pathways associated with LRRC75A-AS1, the co-expressed protein-coding genes were computed using Pearson correlation coefficients (PCC) by R package LINC. There are 500 genes correlating with LRRC75A-AS1 (coefficient > 0.5). Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis showed that LRRC75A-AS1 co-expressed protein coding genes were mainly enriched in cadherin binding, unfolded protein and cell adhesion molecule binding, FK506 binding, isomerase and NADH activity (Fig. 4B). Taken together, these findings indicated that LRRC75A-AS1 may contribute to tumor malignancy by interacting with protein coding genes that involved in the above biological pathways.

lncRNA expression and CNA aberration

LRRC75A-AS1 as a novel candidate oncogene in chromosome region 17p11.2–p12 had been reported in human osteosarcoma to have a significant association between copy number and expression level (Both et al., 2012; Both et al., 2016). Although it also acted as an oncogene in GIST, LRRC75A-AS1 expression level showed no significant association with copy number alteration in our data analysis. Chromosome region 17p11.2–p12 showed a hotspot site in GIST, where LRRC75A-AS1 was significantly upregulated during tumor progression. Meanwhile chromosome region 17p11.2–p12 also contained protein-coding genes, RAI1, MSI2 and SMCR8, which were in co-expressed network of lncRNA HYMAI, NEAT1 and FTX (Fig. 5). Thus, chromosome region 17p11.2–p12 may contribute to the oncogenic procession in malignant GIST.