A prognosis-related molecular subtype for early-stage non-small lung cell carcinoma by multi-omics integration analysis

RNA-Seq data, DNA methylation data and clinical information of NSCLC patients were downloaded from the UCSC Xena website (http://​xena.​ucsc.​edu/​). The RNA-seq data were log₂ transformed RSEM normalized counts and mapped to HUGO gene symbols. The DNA methylation levels were represented by β-values (from 0 to 1). Methylation sites were filtered by the following criteria: 1) probes located in the X or Y chromosome; 2) SNP present within the assay of probe; 3) probes did not annotate with any reference genes; 4) probes located in the shelves and oversea regions of CpG island. Genes and methylation sites with missing value in more than 20% of patients were excluded, and patients without mRNA data or methylation data were also removed from further analysis. Data were centralized and standardized before analysis.

We randomly divided the data set into training set and test set, in which training set contained 60% patients (Table 1). We selected biomarkers and built prediction models for molecular subtypes in training set, and validated the models in test set. In the biomarker selection phase, univariate Wilcoxon test was firstly used to selected differentially expressed genes and methylated sites (P < 0.001) compared with other subtypes in the training dataset. Then, a multivariate partial least square (PLS) model was established, and 10 variables with largest variable important projection (VIP) values were selected as biomarkers for each subtype. Random forest (RF) and support vector machine (SVM) models were constructed with 10 selected biomarkers in training dataset. The model prediction ability was evaluated in training and test datasets, respectively.

The prediction model performance was evaluated by sensitivity (SE), specificity (SP) and area under the ROC curve (AUC). SE and SP were defined by: where.

True Positive (TP): the patient belongs to a subtype, and the prediction model predicts the patient as this subtype;

False Positive (FP): the patient does not belong to a subtype, but the prediction model predicts the patient as this subtype;

True Negative (TN): the patient does not belong to a subtype, and the prediction model does not predict the patient as this subtype;

False Negative (FN): the patient belongs to a subtype, but the prediction model does not predict the patient as this subtype.

AUC were defined by: where n₀ and n₁ are the number of patients who belong to and not belong to a subtype respectively, and r_i is the rank of i_th patient of a subtype in the ranked list.

An overview workflow of this study was shown in Fig. 1. In total, 303 LUSC and 351 LUAD patients with coupled mRNA and methylation data were analyzed in this study. After screening according to the criteria in the “method” section, 16249 mRNAs and 162,926 methylation sites were included. In the univariate Cox regression analysis, 14 mRNAs and 362 methylation sites in LUSC, and 143 mRNAs and 458 methylation sites in LUAD were associated with overall survival at the 0.001 level.

In the identification of prognosis-related molecular subtypes, the genes and methylation sites related with overall survival were used in the PAM clustering analysis, with cluster number K ranging from 2 to 5 (Supplementary Fig. 1). In the Kaplan-Meier (K-M) survival curves of LUSC and LUAD patients for different cluster number K, optimum survival curves were identified by 4 clusters for LUSC and 2 clusters for LUAD, respectively (Fig. 2). These subtypes showed different gene expression and DNA methylation patterns (Fig. 3). There were highly expressed and hypermethylated gene regions for LUSC-C1 and LUAD-C2, highly expressed region for LUAD-C1, and hypermethylated regions for LUSC-C3 and LUSC-C4.

Functional enrichment was conducted for genes and methylated genes related with overall survival in LUSC and LUAD. Firstly, 362 and 458 methylation sites in LUSC and LUAD were located in 267 and 339 genes, respectively. The distribution of methylation sites was showed in Supplementary Fig. 2. The methylated genes in LUSC were significantly enriched in 8 GO biological process terms, 4 GO cellular component terms, and 7 GO molecular function terms (Supplementary Table 1). The 14 differentially expressed genes were also enriched in 3 GO cellular component terms at 0.2 level.

For LUAD patients, the methylated genes were significantly enriched in 11 biological process terms, 9 cellular component terms, 7 molecular function terms, and 2 KEGG pathways (Supplementary Table 2). The differentially expressed genes were significantly enriched in 21 biological process terms, 12 cellular component terms, 11 molecular function terms, and 7 KEGG pathways.

We next sought to select specific biomarkers and build prediction models for molecular subtypes. There were only 4 patients and obvious hypermethylated region in LUSC-C4 subtype, so we selected biomarkers and built prediction model for LUSC-C1, LUSC-C3 and LUAD. In the training dataset which included 60% patients, univariate Wilcoxon test was firstly used, and 114, 182, 303 biomarkers with P < 0.001 were selected for LUSC-C1, LUSC-C3 and LUAD. Then, these biomarkers were further used to build a multivariate PLS model and 10 variables with VIP values were selected as biomarkers for each subtype, respectively (Table 2). Ten methylation sites were selected as biomarkers for prediction of LUSC-C1 and LUSC-C3, respectively. Seven biomarkers (cg00894870, cg03041700, cg04738309, cg08356572, cg09844983, cg11416447 and cg22627950) were down-regulated, and 3 biomarkers (cg15987088, cg22627950 and cg24599434) were up-regulated in LUSC-C1 (Supplementary Fig. 3). Two biomarkers (cg02074191 and cg12065562) were down-regulated, and 8 biomarkers (cg00431236, cg00894870, cg02590972, cg03041700, cg04417954, cg08356572, cg11416447 and cg22627950) were up-regulated in LUSC-C3 (Supplementary Fig. 4). Nine genes and 1 methylation site were selected as biomarkers for LUAD subtype prediction. These biomarkers were all down-regulated in LUAD-C1 (Supplementary Fig. 5).

To evaluate the predictive performance of biomarkers, RF and SVM models for subtype prediction were built, and SE, SP and AUC were calculated in training and test datasets (Table 3). As results, we got a good prediction performance in the both training and test datasets. AUCs for LUSC-C1, LUSC-C3 and LUAD in RF model were 0.84, 0.77 and 0.83, whereas 0.85, 0.85 and 0.83 in SVM model, respectively. The prediction model built by SVM model was superior to the one built by RF model. These subtypes were significantly associated with overall survival, except for LUAD subtypes in test set, which also tended to be significant (Supplementary Fig. 6).