Prediction of key genes and pathways involved in trastuzumab-resistant gastric cancer

The gene expression profile, GSE77346, was retrieved from the Gene Expression Omnibus (GEO) database (http://​www.​ncbi.​nlm.​nih.​gov/​geo/​) [18]. The profile was generated by GPL10558, Illumina Human 48 K gene chips (Illumina HumanHT-12 V4.0 Expression BeadChip). The GSE77346 dataset consisted of one trastuzumab-sensitive NCI-N87 cell line and four trastuzumab-resistant cell lines (N87-TR1, N87-TR2, N87-TR3, N87-TR4). Briefly, all the cell lines were maintained in Roswell Park Memorial Institute (RPMI) 1640 medium with 10% heat-inactivated FBS. The green fluorescent protein (GFP) +/luciferase+ NCI-N87 cell lines were harvested and injected into the gastric walls of a nude mice. The tumor-bearing mice were received 20 mg/kg trastuzumab i.p. twice per week when the resulting tumors were detectable (Living Image Software program, Xenogen). The trastuzumab treatments were stopped when the tumors were relapsed. By repeated GFP flow cytometric sorting (FACSAria II sorter, Becton Dickinson), four trastuzumab-resistant cell lines were established [17]. Next, total RNA was retrieved by TRIzol reagent (Ambion, Warrington, UK). The synthesis of biotinylated cRNA (Illumina TotalPrep RNA Amplification Kit, Ambion) and the hybridization (Human HT-12 V4 BeadChip) were performed according to the manufacturer protocols. Probe intensity was obtained and normalized by the Illumina GenomeStudio software (Genome Studio V2011.1) [17]. The gene expression profiles GSE13861, including 84 samples (65 tumors and 19 normal tissues), were used for investigation of mRNAs expression of the hub genes between tumor and normal tissues (Illumina Human V3) [19]. For external validation on gene expression profiles with other target drugs, we further included GSE19043 and GSE95414. GSE19043 contained 21 samples from DiFi and GTL-16 cell lines, of which biological triplicates of DiFi cells with gefitinib (EGFR inhibition) and DMSO (control) were used in this study for validation. The platform was GPL5104, Sentrix HumanRef-8 v2 Expression BeadChip [20]. GSE95414 contained one parental NCI-N87 cell line and one trastuzumab-DM1 (T-DM1, trastuzumab emtansine)-resistant cell line. T-DM1 is designed to achieve a combinational therapy of trastuzumab and DM1 (a potent microtubule-disrupting drug, a maytansine derivative) [21]. The RNA was processed by Human Transcriptome Array 2.0 arrays (Affymetrix, GPL17586). Given the absence of biological replicates, the fold change between the T-DM1-resistant cell line and parental cell line was used for investigation (original study of GSE95414 is not yet published).

The differentially expressed genes (DEGs) between the trastuzumab-resistant cell lines and sensitive control were identified by the GEO2R analytical tool [22]. Benjamini and Hochberg method was used for false discovery rate (FDR). The cut-off values of DEGs were defined as adj.p value < 0.05 and log2 fold change (log FC) > 2 or < − 2. The DEG expression data were processed for a bidirectional hierarchical clustering plot (FunRich, http://​www.​funrich.​org) [23].

The Database for Annotation, Visualization, and Integrated Discovery (DAVID, http://​david.​abcc.​ncifcrf.​gov/​) was employed for the gene ontology (GO) consortium reference, including biological processes (BP), cellular components (CC), and molecular functions (MF) [24, 25]. In addition, DAVID was also employed for pathway enrichment annotations with the data resources from Kyoto Encyclopedia of Genes and Genomes (KEGG, http://​www.​genome.​jp/​kegg/​) pathway enrichment analysis [24, 26].

The interaction networks of the DEG-coded proteins were determined by the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING, http://​www.​string-db.​org/​) [27]. Node degree ≥ 5 was defined as the cut-off values for further PPI networks visualization by Cytoscape software (version 3.6.0; http://​www.​cytoscape.​org/​) [28]. The Molecular Complex Detection (MCODE) program embedded in Cytoscape was used to subcluster the PPI networks with predefined cutoff criterions (max. depth = 100, node score = 0.2 and k-score = 2) [29]. Hub genes were defined by the degree value (paired connections between each node). In addition, the betweenness centrality (defining the fraction of shortest paths involved in a given node) of the hub genes were also added.

Kaplan-Meier (KM) plotter enables comprehensive analysis of the prognostic values among lists of genes in various cancers based on multiple genomic profiles, including GSE14210, GSE15459, GSE22377, GSE29272, GSE51105, and GSE62254 [30]. The prognostic values of overall survivals (OS) for hub genes were displayed with the hazard ratios (HR) and log-rank p values.

The gene expression profiling interactive analysis (GEPIA, http://​gepia.​cancer-pku.​cn) was established for customized genomic analysis based on The Cancer Genome Atlas (TCGA) database [31]. The top 20 hub genes were extracted for interactive networks based on paired gene correlations of the stomach adenocarcinoma (STAD) cohort in TCGA (Pearson correlation coefficients). In addition, the mRNA expressions of the hub genes were also investigated between tumor and normal tissues.

Moreover, the stage-specific expression of each hub gene was also generated by GEPIA. The mRNA expressions of the hub genes of TCGA (STAD) were also retrieved from the Xena system, University of California, Santa Cruz (UCSC) for prognostic analysis [32].

A total of 849 DEGs were identified to be associated with trastuzumab resistance, with 374 genes upregulated and 475 downregulated (Fig. 1). A bidirectional hierarchical clustering heat map of the DEGs was illustrated (Fig. 2).

The GO enrichment analysis was conducted by the DAVID tool. A total of 193 BP terms significantly enriched, including epithelium development/cell surface receptor signaling pathway/locomotion (Table 1). A total of 23 CC terms were significantly enriched, including membrane-bounded vesicle/extracellular region part/extracellular vesicle (Table 1). A total of nine MF terms were significantly enriched, including top-ranked cell adhesion molecular binding/glycoprotein binding/growth factor binding (Table 1). Specifically, in each term, top ranked 10 most significantly enriched gene-ontologies of upregulated and downregulated DEGs were compared (Fig. 3). In BP term, nervous system development and response to type I interferon were significantly enriched in up/downregulated DEGs, respectively (Fig. 3a). In CC term, proteinaceous extracellular matrix and extracellular region part were significantly enriched in up/down regulated DEGs, respectively (Fig. 3b). In MF term, protein dimerization activity and cell adhesion molecule binding were significantly enriched in up/down regulated DEGs, respectively (Fig. 3c).

Noteworthy, only two significant signaling pathways were identified in KEGG pathway analysis with cut-off values (p < 0.05, FDR < 0.05): pathways in cancer (hsa05200) and ECM-receptor interaction (hsa04512) (Table 2). The top ten enriched signaling pathways in upregulated and downregulated DEGs were illustrated, respectively (Fig. 4). Of note, no significant pathway was identified in upregulated set, and only one, the pathways in cancer (hsa5200), was identified as significant in downregulated set.

Next, the PPI networks were initially obtained by the STRING database and visualized by Cytoscape with degrees of each nodes ≥ 5. A total of 291 nodes 1883 edges were included in the PPI networks (Fig. 5). The top 20 hub genes with highest degrees were determined, including CD44 molecule (CD44), erb-b2 receptor tyrosine kinase 2 (HER2), cadherin 1 (CDH1), 2′-5′-oligoadenylate synthetase 1–3 (OAS1–3), 2′-5′-oligoadenylate synthetase-like (OASL), ISG15 ubiquitin-like modifier (ISG15), bone morphogenetic protein 4 (BMP4), signal transducer and activator of transcription 1 (STAT1), early growth response 1 (EGR1), cyclin D1 (CCND1), vimentin (VIM), Wnt family member 5A (WNT5A), KIT proto-oncogene receptor tyrosine kinase (KIT), bone morphogenetic protein 2 (BMP2), interferon regulatory factor 9 (IRF9), MX dynamin-like GTPase 1 (MX1), FYN proto-oncogene, Src family tyrosine kinase (FYN), and HECT and RLD domain containing E3 ubiquitin protein ligase family member 6 (HERC6) (Fig. 5, Table 3). In addition, the top scored three modules were determined by MCODE in Cytoscape, with KEGG enrichment results (Fig. 6). Furthermore, the siRNAs of the hub genes were summarized (Additional file 1: Table S1) [33‐51].

The prognostic values of the hub genes were assessed by the KM plotter in GC. High HER2, CDH1, OAS1, OAS3, ISG15, BMP4, CCND1, and WNT5A expression levels were associated with poor OS, whereas high CD44, STAT1, EGR1, VIM, KIT, and FYN expression levels were associated with favorable OS. OAS2, OASL, BMP2, IRF9, MX1, and HERC6 were not significantly associated with OS (Fig. 7). The mRNAs expression of CD44, HER2, CDH1, OAS1, OAS2, OAS3, OASL, ISG15, STAT1, CCND1, and WNT5A were significantly upregulated in tumor while only KIT was significantly downregulated in tumor (TCGA) compared to normal (TCGA normal + GTEx normal) (Fig. 8a). Next, we further compared the mRNA expression of the hub genes between tumor (TCGA) and normal (TCGA) by the data retrieved from the Xena system. In fact, the results from the Xena (TCGA tumor vs TCGA normal) were different from GEPIA (TCGA tumor vs TCGA normal + GTEx normal). Only five hub genes (STAT1, OAS3, OAS2, CDH1, ISG15) significantly exhibited upregulation and two hub genes (KIT and EGR1) exhibited downregulation according to the thresholds (adj.p value < 0.05 and |logFC| > 1) (Additional file 2: Table S2). Interestingly, the gene with the most significant logFC value is KIT (logFC = − 2.11514), whereas the gene with the most significant adj.p value is STAT1 (adj.p value = 8.99E−12).

Furthermore, the mRNA expression of the hub genes (IRF9 was not available) was externally validated in GSE13861 (Additional file 3: Figure S1). Consistently, CD44, OAS3, ISG15, STAT1, and WNT5A were significantly upregulated whereas KIT was significantly downregulated in tumor compared to normal in GSE13861. Moreover, BMP4 was significantly upregulated in tumor in GSE13861. OASL, EGR1, and BMP2 were significantly downregulated in tumor in GSE13861 (Additional file 3: Figure S1). Moreover, the mRNA expression of all the hub genes in specific clinic stages had been analyzed. In fact, only CD44 (p = 0.0146), VIM (p = 1.07e−05) and KIT (0.00759) exhibited significant stage-specific expression (Additional file 4: Figure S2).

To further elucidate the underlying mechanism between the DEGs, the STAD of TCGA data was employed based on GEPIA platform. Of note, 87.8% (65/74) gene-gene correlations were positive. What is more, OAS1, 3, and CDH1 featured high degrees and strong correlations with other hub genes. Additionally, VIM was negatively correlated with CCND1, HER2, and CDH1, respectively. KIT was negatively correlated with HER2, ISG15, and OAS1, respectively (Fig. 8b). Meanwhile, to investigate the potential roles of the hub genes in other target therapies, GSE19043 and GSE95414 were retrieved for external investigation (Additional file 5: Table S3). In GSE19043, none of the hub genes exhibited differential expression between gefitinib group and control, whereas in GSE95414, only six of the hub genes, including VIM, BMP2, CD44, OAS3, KIT, and WNT5A, showed slight fold change values > 1 between T-DM1-resistant cell lines and control (Additional file 6: Figure S3). In summary, the hub genes identified in this study may not be directly involved in gefitinib (EGFR inhibition, GSE95414) and T-DM1 (GSE19043) (Additional file 6: Figure S3).