GEO data

Probe re-annotation pipeline

The GPL18109 probe set sequences for the Agilent-038314 CBC Homo sapiens lncRNA + mRNA microarray V2.0 were downloaded from the Agilent website (https://www.agilent.com/). PCG and lncRNA expression data from the Agilent-based expression profile of ESCC cohorts (GSE53624 and GSE53622) was obtained by re-annotating microarray probes according to the sequences of the probe sets and the annotations of PCG and lncRNA records in GENCODE (GRCh38, release 21, http://www.gencodegenes.org/). We used BLASTn (ftp://ftp.ncbi.nlm.nih.gov/blast/executables/LATEST/) to align the probe sequences to those of noncoding and coding transcript sequences from GENCODE.

The alignments were filtered as follows: (i) only probes perfectly matched to a transcript were retained, resulting in two sets of probes targeting protein-coding and -non-coding transcripts, respectively; (ii) probes targeting noncoding transcripts that perfectly matched cDNA coding sequences were removed; (iii) all transcripts corresponding to the retained probes were mapped to the genome and annotated as PGCs or lncRNAs.

Sample collection and preparation

Besides above 179 ESCC patients from GEO database, we collected 105 ESCC patients as the experimental set from the Chaoshan District of Guangdong Province, which has a high prevalence of ESCC [26]. The experimental set was used for external validation. Samples were collected from the Department of Oncological Surgery of the Central Hospital of Shantou City, P.R. China between February 2012 and December 2013. Tumor and paired nontumor tissues were collected from patients who underwent surgical resection. After examination by a pathologist, tissues were immediately frozen in liquid nitrogen and stored at − 80 °C. Partial tissue samples were used for hematoxylin and eosin staining to confirm the diagnosis and analysis of pathological grade, metastasis, and tumor cell content. Tumor samples contained > 80% tumor tissue free of necrosis were selected. Only those died of ESCC were included in the study. The follow-up for patients after esophageal resection continued until death, and it extended to March 2016. We excluded patients suffering from severe postoperative complications, other tumors, or those who died of other causes. The clinical data were available in Table 1. Cases were classified according to the TNM classification of the International Union Against Cancer, 7th edition. Evaluation of tumor differentiation was based on the guidelines of the World Health Organization (WHO) Pathological Classification of Tumors. Overall survival (OS) was defined as the interval between surgery and death from tumors or the last observation of surviving patients. The study was approved by the Ethics Committees of the Central Hospital of Shantou City and Shantou University Medical College. Written informed consent to use resected samples for research purposes was obtained from all patients.

Reverse transcription (RT) and real-time PCR

Total RNA was extracted using TRIzol (15596-018, Life Technologies, Grand Island, NY, USA) and purified using a PureLink RNA Mini Kit (12183018A; Life Technologies) according to the manufacturer’s protocol. The purity and concentration of RNA were determined according to the ratio of absorbance of 260 nm/280 nm light using a NanoDrop ND-2000 spectrophotometer (ND-2000, Thermo Fisher Scientific, Waltham, MA, USA). The cDNA synthesis was performed by reverse transcription using random hexamer primers (Takara, Dalian, Liaoning, China). Real-time PCR was performed using a SYBR Premix Ex Taq kit (DRR037A, DRR081A; Takara). Briefly, reverse transcription was performed at 37 °C for 15 min and at 85 °C for 5 s. Real-time PCR was performed using an ABI 7500 real-time PCR system (Life Technologies) as follows: 95 °C, 30 s; 95 °C, 5 s; 60 °C, 34 s (30 cycles). Relative quantification of mRNA expression was calculated using the 2^−ΔΔCt method. Quantitative RT-PCR (qRT-PCR) was performed in triplicate and repeated at least three times, as described previously [34, 35]. All methods were performed in accordance with guidelines and regulations of the ethics committees identified above.

Statistical analysis

All analyses were performed using the R program (www.r-project.org), including packages named pROC, rpart, and survival downloaded from Bio-conductor (http://www.bioconductor.org/). Univariate Cox regression analyses were used to identify common PCGs and lncRNAs that associated with OS. P < 0.05 indicates a statistically significant difference. All insignificant PCGs and lncRNAs with P >0.05 were excluded. We performed principal component analysis (PCA) [36–38] to reduce expression data and to capture innate patient characteristics.

We used the R NbClust package (Euclidean distance, complete linkage), and it provides 30 indices for determining the number of clusters. The output of the software proposes the best clustering scheme from the results obtained by varying combinations of the clusters, distance measures, and clustering methods. These operations identify the optimal number of clusters by calculating several cluster indices. The clusters were validated using the Calinski and Harabasz index [39, 40].

Kaplan–Meier survival analysis was performed to test the equality of survival distributions of different groups for each ESCC cohort, and statistical significance was assessed using the two-sided log-rank test. Further, time-dependent receiver operating characteristic (ROC) curves were used to compare the sensitivity and specificity of the survival prediction, and the area under the curve (AUC) value was calculated from the ROC curve. Multiclass ROC curves were used to evaluate the effectiveness of multi classification predictions, which were calculated using the R packages called pROC [41, 42]. Classification and regression tree (CART) analysis was performed using a Recursive Partitioning and Regression Trees (RPART) library in R to develop a risk-staging model and to determine predictors of survival from the set of survival-related PCGs and lncRNAs [43, 44]. This is a nonparametric statistical method that uses a series of dichotomous splits to create a decision tree.

To begin the CART analysis, patients identified from the GSE53624 dataset were randomly split into the training and testing groups using the function “sample” of the R program [45]. CART was applied first to the training group and then to the test sample to assess the model’s generalizability and to evaluate the overfitting of the model. When the classification tree was generated, error tests and pruning were performed to construct the final tree of parameters with the best size, lowest misclassification rate, and lowest complexity. The selection process of the prognostic model is shown in Fig. 1.

The associations between coexpressed lncRNA and PGCs in the molecular staging model were computed using Pearson correlation coefficients visualized with Cytoscape. We next performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses of the coexpressed protein-coding genes to predict biological functions using the ClueGo plugin of Cytoscape (version 3.2.3) [46], which is a commonly used functional annotation tool that can assess over-representation of functional categories among a gene set of interest. Enrichment analysis, which was performed using the functional annotation chart and functional annotation clustering options, was limited to GO terms and KEGG pathways in the “Biological Process” categories. Functional annotations with P < 0.05 were considered significant [47].

Selection of ESCC microarray datasets and acquisition of PCG and lncRNA expression values

According to the dataset screening criteria described in Methods, 179 samples (119 from the GSE53624 dataset and 60 from the GSE53622 dataset) were selected. All these ESCC samples contained tumor tissues and adjacent normal tissues. Probe reannotation of the Agilent-038314 CBC Homo sapiens lncRNA + mRNA microarray V2.0 identified 17,434 PCGs and 6253 lncRNAs. We retained probes that were uniquely mapped to the genomic coordinates of PCGs or lncRNAs derived from GENCODE. Multiple probes (or probe sets) mapping to the same PCGs or lncRNAs were integrated using the arithmetic mean of the values of multiple probes (or probe sets) and were used to generate new PCG and lncRNA expression profile values from the GSE53624 and GSE53622 datasets. Further, we used fold-change values (cancer/normal) on a log2 scale in the next analysis.

Selection of prognostic PCGs and lncRNAs from the GSE53624 dataset

The ESCC patient cohort from GSE53624 (n = 119) was selected to explore the association of OS with PCGs or lncRNAs. We conducted univariate Cox proportional hazards regression analysis of the PCG and lncRNA expression profiling data, with OS as the dependent variable, and identified 2404 PCGs and 635 lncRNAs that significantly associated with OS (P < 0.05) (Fig. 2a), which were therefore considered potential prognostic markers.

Identification of three distinct clusters of patients with ESCC

To identify genes that are more important for staging and to reduce the dimensionality of the profile with2404 PCGs and 635 lncRNAs, PCA was performed. This analysis demonstrated that the survival-related PCG-lncRNA set in GSE63624 was reduced to three independent principal components PC1, PC2, and PC3, accounting for 99% of the variance of the component space (Fig. 2b). Using the three PC scores as variables, we applied the NbClust clustering procedure to cluster the patients with ESCC identified from the GSE53624 dataset and highlighted a three-cluster partition as the best, with the largest cluster index number (Fig. 2c, d).

Association of patient groups with survival

Kaplan–Meier analysis revealed that the prognostic PCGs and lncRNAs had the potential to classify the 119 patients into three groups with different OS estimates. Patients in the high-risk group had shorter OS compared with those in the middle- and low-risk groups (median OS: 16.5 months vs. 26.5 months and 50.9 months, P < 0.05) (Fig. 3a). The 5-year OS rate was approximately 20% in the high-risk group, approximately 30% in the middle-risk group, and > 50% in the low-risk group.

Comparison of the PCG-lncRNA grouping versus TNM staging in survival prediction

According to Kaplan–Meier analysis, the PCG-lncRNA grouping yielded a better classification of patients compared with that of the TNM staging system (Fig. 3b). To compare the sensitivity and specificity in survival prediction between TNM staging and PCG-lncRNA grouping, we performed ROC analysis. In the GSE53624 dataset, the predictive ability of the PCG-lncRNA grouping was better than that of TNM staging (AUC: 0.69 vs. 0.65, P < 0.05) (Fig. 3c).

Construction of the LINC01800-SEMA3A-BEX2 (LSB) staging model

Kaplan–Meier and ROC analyses showed that the PCG-lncRNA grouping improved the classification of patients, indicating its value as a novel, efficient staging plan. We next pursued identifying markers to classify the three groups as follows. We randomly divided the GSE53624 dataset into a training set (n = 59) and a testing set (n = 60) for internal validation. Next, we selected the first 100 genes with the highest absolute loading, e.g., the 100 genes with the highest positive or negative correlation corresponding to each principal component. The clinical attributes of patients, such as age, sex, tobacco use, alcohol use, tumor location, tumor grade, T stage, N stage, and TNM stage, were used as variables to perform the CART routines.

Subsequently, CART analysis of the training set generated the final tree composed by LINC01800, semaphorin 3A (SEMA3A), and brain-expressed X-linked 2 (BEX2) identified from the survival-related PCGs and lncRNAs. Moreover, there were no clinical attributes of patients with lowest error rate remaining in the classification tree produced using RPART (Fig. 4a, b; Table 2). Higher expression levels of BEX2 and LINC01800 were associated with longer OS (univariable Cox regression coefficient < 0). Higher expression levels of SEMA3A were associated with shorter OS (univariable Cox regression coefficient > 0). Multiclass ROC was used to evaluate the predictive ability of the LSB (stands for the first letter of each of the three genes) staging model, and the AUC was 0.89 (P < 0.05) in the training set.

Ensemble ID	Gene symbol	Gene name	Chromosome location	Coefficienta	P valuea	Gene expression level association with prognosis
ENSG00000075213	SEMA3A	Semaphorin 3A	Chromosome 7: 83955777–84492724 (−)	0.17	0.01	High
ENSG00000133134	BEX2	Brain-expressed X-linked protein 2	Chromosome X: 103309346–103311046 (−)	− 0.22	0.01	Low
ENSG00000234572	LINC01800		Chromosome 2: 64846130–64863626 (−)	− 0.20	0.00	Low

PCG Protein-coding gene, lncRNA long non-coding RNA

^aDerived from univariable Cox regression analysis of the GSE53624 dataset

The LSB staging model was used to stratify patients in the test set (n = 60) and entire GSE53624 dataset (n = 119), and the AUC values were 0.79 and 0.83 (P < 0.05), indicating that this model efficiently stratified patients into different prognostic groups (Fig. 4c).

To validate its power for staging efficiency, the LSB staging model was evaluated using an independent dataset (GSE53622, n = 60). Kaplan–Meier survival curves for patients with LSB I, II, and III ESCC, which were classified according to the respective cutoff values of the three molecular markers, are shown in Fig. 5a (median OS: 56.7 vs. 39.2 vs. 24.5 months, P = 0.01). The 5-year OS rates were of was 63.4% in patients with LSB I ESCC, 39.6% in patients with LSB II ESCC, and 21.2% in patients with LSB III ESCC. The percentages of patients with LSB I, LSB II, and LSB III ESCC were 33.3%, 38.3% and 28.4%, respectively, similar to those in the training group (LSBs I-III: 32.0%, 34.0%, and 34.0%, respectively). Moreover, the AUC of the LSB staging model was 0.68, greater than 0.66 for the TNM staging (P < 0.05) (Fig. 5b).

Validation of the LSB staging model using an experimental dataset

Functional classifications of the LSB genes

To further investigate the potential biological roles of the three markers, a coexpression network comprising SEMA3A, BEX2, and LINC01800 was constructed by computing Pearson correlation coefficients of the GSE53624 and GSE53622 datasets (Fig. 6a). GO and KEGG analysis of the PCGs which were coexpressed with LINC01800, SEMA3A, and BEX2 revealed that in the two datasets, the coexpressed PCGs were significantly enriched in 28 GO terms and 3 KEGG pathways (P < 0.05). These findings suggest that SEMA3A, BEX2, and LINC01800 may be involved in tumorigenesis through interacting with those coexpressed PCGs that influence biological processes such as angiogenesis, cell migration, cell differentiation, and cell adhesion (Fig. 6b).