Eight-gene signature predicts recurrence in lung adenocarcinoma

Abstract

BACKGROUND:

Recurrence significantly influences the survival in patients with lung adenocarcinoma (LUAD). However, there are less gene signatures that predict recurrence risk of LUAD.

OBJECTIVE:

We performed this study to construct a model to predict risk of recurrence in LUAD.

METHODS:

RNA-seq data from 426 patients with LUAD were downloaded from The Cancer Genome Atlas (TCGA) and were randomly assigned into the training ( $n=$ 213) and validation set ( $n=$ 213). Differentially expressed genes (DEGs) between recurrent and non-recurrent tumors in the training set were identified. Recurrence-associated DEGs were selected using multivariate Cox regression analysis. The recurrence risk model that identifies patients at low and high risk for recurrence was constructed, followed by the validation of its performance in the validation set and a microarray dataset.

RESULTS:

In total, 378 DEGs, including 20 recurrence-associated DEGs, were identified between the recurrent and non-recurrent tumors in the training set. The signatures of 8 genes (including AZGP1, INPP5J, MYBPH, SPIB, GUCA2A, HTR1B, SLC15A1 and TNFSF11) were used to construct the prognostic model to assess the risk of recurrence. This model indicated that patients with high risk scores had shorter recurrence-free survival time compared with patients with low risk scores. ROC curve analysis of this model showed it had high predictive accuracy (AUC $>$ 0.8) to predict LUAD recurrence in the TCGA cohort (the training and validation sets) and GSE50081 dataset. This prognostic model showed high predictive power and performance in predicting recurrence in LUAD.

CONCLUSION:

We concluded that this model might be of great value for evaluating the risk of recurrence of LUAD in clinics.

Keywords

8-gene signature lung adenocarcinoma prognostic model recurrence The Cancer Genome Atlas

1. Introduction

Lung cancer, the most commonly diagnosed cancer worldwide, accounts for 11.6% of the total cancer cases and 18.4% of the total cancer-related deaths in 2018 [1]. Prognosis of lung cancer remains poor and the 5-year survival rate is less than 20% [2]. The 5-year postoperative recurrence in non-small cell lung cancer (NSCLC) ranges from 19.3% to 50% [3, 4, 5]. Adenocarcinoma is the predominant histologic type of lung cancer and is a leading cause of cancer death worldwide [6]. The 5-year postoperative recurrence in LUAD is 17% $\sim$ 40% [7, 8, 9]. The poor prognosis and high death rate of this disease, on a great extent, depend on the postoperative recurrence and aggressive metastasis.

NSCLC recurrence has been identified to be independently influenced by the clinical characteristics like age, histologic type, invasion and tumor size [3]. However, histopathology and clinical indicators are far from sufficient to predict clinical outcome and risk of recurrence in lung cancer.

Molecular markers are incorporated into recurrence identification, prediction and decision-making in LUAD in clinics [10, 11, 12]. Global changes in gene expression represent the common cause of tumorigenesis as well as predict the postoperative prognosis and recurrence in lung adenocarcinoma (LUAD) [6, 13, 14]. The overexpression of metastasis associated with the colon cancer 1 gene (MACC1) was reported to be associated with postoperative recurrence in LUAD [15]. Larsen et al. identified a 54-gene signature with potential to predict the risk of recurrence in LUAD [13]. The cell cycle progression score (CCP score) that consists of 31 cell cycle genes has been used for predicting disease response, mortality, survival and recurrence in LUAD [16, 17]. In addition, the combination of CCP score with molecular prognostic score (mPS) had higher performance in predicating distant and local recurrence in LUAD compared with CCP score alone [17]. However, this strategy includes too many genes to be widely implemented in preclinical or clinical practices. Signature including genes’ function in single molecular process might be insufficient to identify cancer development and metastasis. A signature with moderate number of genes that relate to multiple functions might be of great interest for predicting recurrence in human cancers.

We performed this study to identify a gene signature with prognostic power for predicting the risk of recurrence in LUAD following resection. Differentially expressed genes (DEGs) that associate with the recurrence in LUAD would be identified in a LUAD cohort in the The Cancer Genome Atlas (TCGA). An 8-gene model with high predictive power for the risk of recurrence in LUAD was identified and validated. The potential of using the prognostic model for the prediction of postoperative recurrence in LUAD would be discussed and defined.

2. Materials and methods

2.1 Study population and assignment

The available RNA-seq data and clinical information of 426 evaluable patients with LUAD were downloaded from TCGA (https://gdc-portal.nci.nih.gov/) in March 2019. Patients ( $n=$ 426) were randomly divided into two sets: the training set ( $n=$ 213) and the validation set ( $n=$ 213).

2.2 DEGs identification

The DEGs between the samples with ( $n=$ 76) and without ( $n=$ 137) recurrence in the training set were identified using the Limma package (version 3.34.7; https://bioconductor.org/packages/release/bioc/html/li mma.html) [18]. DEGs were considered statistically significant if its false discovery rate (FDR) $<$ 0.05 and fold-change (FC) $\geqslant$ 1.2 ( $|$ log ${}_{2}$ FC $|\geqslant$ 0.263). The expression profiles of all DEGs were presented using a heatmap analysis (pheatmap package, version 1.0.8; https://cran.r-project.org/web/packages/pheatmap/index. html) [19].

2.3 Identification of recurrence-associated DEGs

The recurrence-associated DEGs were identified using the univariate and multivariate Cox regression analyses (survival package, version 2.41-1; http://bioconduc tor.org/packages/survivalr/) [20]. Kaplan-Meier log-rank test $P$ -value $<$ 0.05 was considered as the threshold.

2.4 Selection of the optimal combination of DEGs

The DEGs that might be used for the construction of model were selected from recurrence-associated DEGs using the LASSO Cox regression model in penalized package (version 0.9.50; https://cran.r-project.org/web/ packages/penalized/index.html) [21]. The ‘lambda’ parameter (Cox-PH model) of the optimal combination of genes was obtained using the cross-validation likelihood algorithm (1000 times). Regression coefficients ( $\beta$ ) of all recurrence-associated DEGs with recurrence were obtained.

2.5 Establishment of the recurrence risk model

The expression levels of the recurrence-associated DEGs in the aforementioned optimal combination were used for the analysis of optimal cut-off values using the X-Tile Bio-Informatics tool (version 2.41.3; https://medicine.yale.edu/lab/rimm/research/software. aspx) [22]. Monte Carlo methods $P$ value $<$ 0.05 was used as the threshold. The expression status of each recurrence-associated DEG was defined as 1 when the expression level was higher than the cut-off value, or otherwise 0. The recurrence risk score of each sample was defined as the combination of the weighted expression status by the LASSO Cox $\beta$ value, and calculated as: risk score $=$ $\Sigma\beta_{\text{DEGn}}$ $\times$ Status ${}_{\text{DEGn}}$ (0/1). Samples with risk scores higher than the median value were stratified into the high-risk group, or otherwise the low-risk group.

Table 1
Baseline characteristics of patients with LUAD

Clinical characteristics	Training set ( $N=$ 213)	Validation set ( $N=$ 213)	Entire set ( $N=$ 426)
Age (mean $\pm$ SD)	65.12 $\pm$ 10.07	65.30 $\pm$ 10.24	65.21 $\pm$ 10.15
Gender (Male/Female)	105/108	89/124	194/232
Pathologic M (M0/M1/NA)	139/13/61	135/4/74	274/17/135
Pathologic N (N0/N1/N2N3/NA)	144/33/31/0/5	137/45/24/2/5	281/78/55/2/10
Pathologic T (T1/T2/T3/T4/NA)	68/113/23/7/2	82/113/12/5/1	150/226/35/12/3
Pathologic stage (I/II/III/IV/-)	109/52/37/13/2	125/51/27/5/5	234/103/64/18/7
Radiation therapy (Yes/No/NA)	28/175/10	28/178/7	56/353/17
Targeted molecular therapy (Yes/No/NA)	69/133/11	69/135/9	138/268/20
Smoking history (Current/Reform/Never/NA)	23/59/9/122	17/48/13/135	40/107/22/257
Recurrence (Yes/No)	76/137	75/138	151/275
Recurrence free survival time (months $\pm$ SD)	26.21 $\pm$ 26.64	26.35 $\pm$ 27.47	26.78 $\pm$ 27.03
Death (Dead/Alive)	69/144	70/143	139/287
Overall survival time (months $\pm$ SD)	29.92 $\pm$ 28.08	30.16 $\pm$ 28.42	30.04 $\pm$ 28.22
Follow up (months, median and range)	18.93 (0.34 $\sim$ 224.40)	18.93 (0.13 $\sim$ 235.40)	18.93 (0.13 $\sim$ 235.40)

SD, standard deviation. NA, not applicable.

2.6 Evaluation and assessment of the prognostic model

Kaplan-Meier log-rank test was used to evaluate the survival difference between the high- and low-risk groups. $P$ value $<$ 0.05 was defined as significantly different. Additionally, the areas under the ROC curve (AUC) in the training and validation sets were calculated for evaluating the performance of this model. The sensitivity and specificity were calculated. The third microarray dataset GSE50081 that consists of samples from 124 LUAD patients was downloaded from the NCBI’s Gene Expression Omnibus (GEO) and used to validate the predictive performance of this model.

2.7 Identification of characteristics that associate with LUAD recurrence

The clinical characteristics that associate with recurrence in LUAD were identified using the univariate and multivariate Cox regression analyses (survival package). Kaplan-Meier log-rank test $P$ -value $<$ 0.05 was considered as the threshold. The 95% confidence interval (CI) value of each clinical characteristic was evaluated. Stratification analyses were performed for the training and validation sets.

2.8 Identification and enrichment of the DEGs between patients with high and low risk of recurrence in LUAD

To investigate the differences between patients with high and low risk recurrence scores, the DEGs between the two groups were identified using Limma package (version 3.34.7) with the criteria of FDR $<$ 0.05 and $|$ log ${}_{2}$ FC $|\geqslant$ 0.263. The Gene Ontology biological processes and pathways associated with the DEGs were identified using clusterProfiler (Version 3.6.0; http:// bioconductor.org/packages/release/bioc/html/clusterPro filer.html) [23]. $P<$ 0.05 was considered statistically significant.

2.9 Statistical analysis

Data statistical analysis was performed using the SPSS 22.0 software. The differences in continuous variables (patient age and survival time) were analyzed using $t$ -test, and those in categorical variables were analyzes using $\chi^{2}$ test, respectively. $P<$ 0.05 was regarded as statistically significant.

3. Results

3.1 Identification of LUAD recurrence-associated DEGs

The baseline characteristics of the 426 patients with LUAD are shown in Table 1. The mean age was 65.21 $\pm$ 10.15 years, and the male-to-female ratio was 1: 1.19. This study set consists of 234, 103, 64, 18, and 7 cases with stage I–IV and unreported stage tumors. In the training set ( $n=$ 213), a total of 378 DEGs (including 279 down-regulated and 99 up-regulated genes) were identified in patients with recurrent LUAD ( $n=$ 76) compared with patients with non-recurrent LUAD ( $n=$ 137; Fig. 1A). The distinct expression profiles of these DEGs were presented by a heatmap (Fig. 1B).

Figure 1.

Identification and statistics of the LUAD recurrence-associated differentially expressed genes (DEGs) in the training set ( $n=$ 213). (A) the volcano plot of the DEGs between patient with and without recurrent LUAD in the training set ( $n=$ 213). Green notes the significantly up-regulated and down-regulated genes, respectively. DEGs were defined with a false discovery rate $<$ 0.05 and $|$ log ${}_{2}$ (fold-change) $|$ $\geqslant$ 0.263. (B) the heatmap depicting the two-way hierarchical clustering of the DEGs in the training set. The color scale indicates the up-regulation (red) or down-regulation (green) of expression. Rec, recurrent LUAD samples ( $n=$ 76); Non-rec, LUAD samples without recurrence.

The prognosis and recurrence characteristics of the 213 patients in the training set were used for the identification of recurrence-associated DEGs. Univariate Cox regression analysis identified that 155 DEGs were associated with the recurrence in LUAD, and the subsequent multivariate Cox regression analysis screened out 20 DEGs.

3.2 Construction and assessment of the recurrence risk model

The optimal combination of DEGs ( $n=$ 8) was identified in the training set using the LASSO Cox regression model (Table 2). We observed that the expression of 4 DEGs, including alpha-2-glycoprotein 1, zinc-binding protein (AZGP1), inositol polyphosphate 5-phosphatase PIPP (INPP5J), myosin binding protein H (MYBPH), and SPIB encoding gene, were negatively correlated with the recurrence in LUAD, whereas the other 4 genes including guanylin (GUCA2A), serotonin receptor 1B (HTR1B), peptide transporter 1 (SLC15A1) encoding gene and cytokine receptor activator of NF- $\kappa$ B ligand (RANKL) encoding gene TNFSF11 were positively correlated with the recurrence.

Table 2
The 8 LUAD prognosis-associated differentially expressed genes in the optimal group by LASSO Cox regression model

ID	$\beta$	$P$ value	Hazard ratio	95% CI	Cutoff
AZGP1	$-$ 0.08	3.292 $\times$ 10 ${}^{-02}$	0.45	0.03–0.74	2.21
GUCA2A	0.24	3.430 $\times$ 10 ${}^{-05}$	5.82	2.53–13.39	0.65
HTR1B	0.03	3.503 $\times$ 10 ${}^{-03}$	4.96	1.69–6.51	0.45
INPP5J	$-$ 0.18	4.759 $\times$ 10 ${}^{-02}$	0.42	0.11–0.52	0.95
MYBPH	$-$ 0.12	4.784 $\times$ 10 ${}^{-02}$	0.71	0.41–0.89	0.21
SLC15A1	0.02	2.117 $\times$ 10 ${}^{-03}$	1.29	1.03–2.64	1.36
SPIB	$-$ 0.13	2.682 $\times$ 10 ${}^{-02}$	0.35	0.24–0.64	1.08
TNFSF11	0.03	2.341 $\times$ 10 ${}^{-03}$	1.16	1.05–3.52	1.07

HR, hazard ratio; CI, confidence interval. $\beta$ , regression coefficients.

Table 3

The differences in the baseline characteristics of patients with high and low risk scores of LUAD in the dataset GSE50081

Clinical characteristics	High ( $N=$ 62)	Low ( $N=$ 62)	$P$ value
Age (mean $\pm$ SD)	69.91 $\pm$ 8.93	67.36 $\pm$ 10.48	0.147 ${}^{\text{a}}$
Gender (Male/Female)	35/27	29/33	0.28 ${}^{\text{b}}$
Pathologic M (M0/M1/NA)	0/62/0	0/62/0	NA
Pathologic N (N0/N1/N2N3/NA)	43/19/0/0/0	49/13/0/0/0	0.305 ${}^{\text{b}}$
Pathologic T (T1/T2/T3/T4/NA)	17/44/1/0/0	24/37/1/0/0	0.252 ${}^{\text{b}}$
Pathologic stage (I/II/III/IV/-)	41/21/0/0/0	48/14/0/0/0	0.231 ${}^{\text{b}}$
Smoking history (Current/Reform/Never/NA)	18/28/9/7	17/27/14/4	0.841 ${}^{\text{b}}$
Recurrence (Yes/No)	22/40	13/49	0.110 ${}^{\text{b}}$
Death (Dead/Alive)	33/29	16/46	0.003 ${}^{\text{b}}$
Overall survival time (months $\pm$ SD)	39.12 $\pm$ 25.88	47.69 $\pm$ 28.00	0.079 ${}^{\text{a}}$

${}^{\text{a}}$ and ${}^{\text{b}}$ notes the difference is analyzed using $t$ -test and $\chi^{2}$ test, respectively. NA, not available.

Figure 2.

The construction, validation and accuracy of the recurrence risk model in the training set (A), validation set (B), and entire cohort (C). The distribution of sample risk score (left), Kaplan-Meier survival analysis (middle) and the predictive accuracy (ROC curve; right) in the three sets of patients with LUAD. HR, hazard ratio. In the training set, the area under the ROC curve (AUC) for 3-year and 5-year survival was 0.97 and 0.94, with the sensitivity of 83.96% and 77.20% and the specificity of 72.86% and 88.79%, respectively. In the validation set, 3-year and 5-year AUC was 0.93 and 0.86, with the sensitivity of 82.72% and 92.02% and the specificity of 67.85% and 64.77%, respectively. In the entire set, 3-year and 5-year AUC was 0.96 and 0.92, with the sensitivity of 84.41% and 88.90% and the specificity of 70.67% and 79.23%, respectively.

Figure 3.

The validation for the prognostic power of the recurrence risk model. The GSE50081 dataset consists of 124 patients with LUAD was used for the validation. The area under the ROC curve (AUC) for 3-year and 5-year survival was 0.75 and 0.62, with the sensitivity of 69.35% and 79.03% and the specificity of 80.65% and 48.39%, respectively.

Figure 4.

Survival analysis for the pathologic stage of LUAD patients in the training set (A), validation set (B) and the entire cohort (C). Higher pathologic stage developed to shorter recurrence-free survival time.

The recurrence risk model for predicting the recurrence in LUAD was constructed using the linear combination of 8 DEGs, as risk score $=$ ( $-$ 0.07960) $\times$ Status ${}_{\text{AZGP1}}$ $+$ (0.23581) $\times$ Status ${}_{\text{GUCA2A}}$ $+$ (0.02453) $\times$ Status ${}_{\text{HTR1B}}$ $+$ ( $-$ 0.17848) $\times$ Status ${}_{\text{INPP5J}}$ $+$ ( $-$ 0.11699) $\times$ Status ${}_{\text{MYBPH}}$ $+$ (0.01756) $\times$ Status ${}_{\text{SLC15A1}}$ $+$ ( $-$ 0.13153) $\times$ Status ${}_{\text{SPIB}}$ $+$ (0.02540) $\times$ Status ${}_{\text{TNFSF11}}$ . The risk score of each sample was calculated, and 106 and 107 samples in the training set were divided into the low- and high-risk group according to the median score, respectively (Fig. 2A). Survival analysis showed that patients in the high-risk group had shorter recurrence-free survival (RFS) time compared with patients in the low-risk group ( $p=$ 8.422 $\times$ 10 ${}^{-7}$ , HR $=$ 3.30, 95% CI 1.99–5.45). The accuracy evaluation using ROC curve analysis showed the 3-year and 5-year AUC of the 8-gene recurrence risk model was 0.97 and 0.94 with the the sensitivity of 83.96% and 77.20% and the specificity of 72.86% and 88.79%, respectively. In the validation set ( $n=$ 213; Fig. 2B) and entire cohort ( $n=$ 426, Fig. 2C), patients had higher risk scores showed shorter RFS time compared with patients had lower risk scores (validation set: $p=$ 0.038, HR $=$ 1.63, 95% CI 1.02–2.60; and entire cohort: $p=$ 6.932 $\times$ 10 ${}^{-7}$ , HR $=$ 2.31, 95% CI 1.64 $\sim$ 3.24; Fig. 2), respectively. The ROC curve analyses showed the high accuracy of the recurrence risk model in predicting the recurrence in LUAD in the validation set and entire cohort (3-year AUC $=$ 0.93 and 0.96, 5-year AUC $=$ 0.86 and 0.92, respectively). Besides, the performance of this model in predicting recurrence in LUAD was validated in a third dataset GSD50081 (Fig. 3). The basic characteristics of subjects in GSD50081 are presented in Table 3. There was significant difference in death rates between patient had high and low risk scores ( $p=$ 0.003, Table 3). No difference was found in the recurrence rate and overall survival time. Patients had higher scores showed obvious shorter RFS time compared with patients had low risk scores ( $p=$ 0.044, HR $=$ 1.97, 95% CI 1.01–3.83), with a 3-year and 5-year AUC of 0.75 and 0.62, respectively (sensitivity: 69.35% and 79.03%; specificity: 80.65% and 48.39%, respectively). These results suggested the high predictive accuracy of the recurrence risk model for predicting the risk of recurrence in LUAD.

Table 4

The predictive values of the clinical characteristics for LUAD recurrence in the training set, validation set and entire cohort

Clinical characteristics	Univariable cox		Multivariable cox
	HR (95% CI)	$P$ value	HR (95% CI)	$P$ value
Training set ( $N=$ 213)
Age (mean $\pm$ sd)	1.01 (0.99–1.03)	0.501	–	–
Gender (Male/Female)	1.20 (0.77–1.89)	0.425	–	–
Pathologic M (M0/M1)	1.46 (0.57–3.69)	0.426	–	–
Pathologic N (N0/N1/N2N3)	1.32 (0.99–1.76)	0.068	–	–
Pathologic T (T1/T2/T3/T4)	1.57 (1.17–2.11)	2.571 $\times$ 10 ${}^{-3}$	1.17 (0.84–1.63)	0.343
Pathologic stage (I/II/III/IV)	1.43 (1.15–1.79)	2.314 $\times$ 10 ${}^{-3}$	1.27 (1.10–1.67)	0.020
Radiation therapy (Yes/No)	1.84 (1.05–3.22)	0.030	1.22 (0.66–2.27)	0.526
Targeted molecular therapy (Yes/No)	1.16 (0.71–1.90)	0.553	–	–
Smoking history (Current/Reform/Never)	1.34 (0.80–2.24)	0.261	–	–
Risk score (High/Low)	3.30 (1.99–5.45)	8.422 $\times$ 10 ${}^{-7}$	2.90 (1.69–4.99)	1.120 $\times$ 10 ${}^{-4}$
Validation set ( $N=$ 213)
Age (mean $\pm$ sd)	1.01 (0.99–1.04)	0.261	–	–
Gender (Male/Female)	0.65 (0.40–1.05)	0.761	–	–
Pathologic M (M0/M1)	1.39 (0.34–5.75)	0.663	–	–
Pathologic N (N0/N1/N2N3)	1.37 (1.04–1.82)	0.054	–	–
Pathologic T (T1/T2/T3/T4)	1.60 (1.15–2.23)	5.254 $\times$ 10 ${}^{-3}$	1.49 (0.93–2.16)	0.352
Pathologic stage (I/II/III/IV)	1.32 (1.02–1.71)	0.035	1.98 (1.09–3.60)	0.024
Radiation therapy (Yes/No)	2.12 (1.21–3.71)	7.182 $\times$ 10 ${}^{-3}$	1.17 (0.87–1.57)	0.292
Targeted molecular therapy (Yes/No)	0.95 (0.58–1.59)	0.856	–	–
Smoking history (Current/Reform/Never)	0.83 (0.51–1.37)	0.476	–	–
Risk score (High/Low)	1.63 (1.02–2.60)	0.038	1.31 (1.19–2.16)	3.207 $\times$ 10 ${}^{-03}$
Entire cohort ( $N=$ 426)
Age (mean $\pm$ sd)	1.01 (0.99–1.03)	0.223	–	–
Gender (Male/Female)	0.90 (0.65–1.24)	0.519	–	–
Pathologic M (M0/M1)	1.41 (0.65–3.05)	0.361	–	–
Pathologic N (N0/N1/N2N3)	1.34 (1.10–1.63)	3.741 $\times$ 10 ${}^{-3}$	0.95 (0.70–1.28)	0.724
Pathologic T (T1/T2/T3/T4)	1.58 (1.27–1.96)	3.917 $\times$ 10 ${}^{-5}$	1.26 (0.91–1.62)	0.544
Pathologic stage (I/II/III/IV)	1.38 (1.17–1.63)	1.396 $\times$ 10 ${}^{-4}$	1.59 (1.04–2.43)	0.033
Radiation therapy (Yes/No)	1.97 (1.33–2.93)	5.776 $\times$ 10 ${}^{-4}$	1.23 (0.93–1.62)	0.140
Targeted molecular therapy (Yes/No)	1.06 (0.75–1.50)	0.725	–	–
Smoking history (Current/Reform/Never)	1.06 (0.747–1.512)	0.736	–	–
Risk score (High/Low)	2.31 (1.642–3.237)	6.932 $\times$ 10 ${}^{-7}$	1.98 (1.38–2.85)	2.230 $\times$ 10 ${}^{-4}$

Figure 5.

Identification and statistics of the differentially expressed genes (DEGs) between patients with high and low risk of recurrence in LUAD. (A) The volcano plot of the DEGs between patient with low and high LUAD recurrence risk in the entire set ( $n=$ 426). Blue indicates the significantly up-regulated and down-regulated genes, respectively. DEGs were defined with a false discovery rate $<$ 0.05 and $|$ log ${}_{2}$ (fold-change) $|$ $\geqslant$ 0.263. (B) Venn diagram of the common DEGs between comparisons of recurrent vs. non-recurrent samples and high-risk group vs. low-risk group. (C) The two-way hierarchical clustering of the DEGs. Samples are ordered from low to high risk scores. The color scale indicates the up-regulation (red) or down-regulation (green) of expression.

3.3 Independent predictors of recurrence risk

The stepwise Cox regression analysis showed that pathologic stage was the independent risk factor for LUAD recurrence in the three sets (Table 4). Survival analysis showed patients with higher pathologic stage (III–IV) had shorter RFS time compared with patients with lower pathologic stage (I–II, Kaplan-Meier log-rank test $P$ -value $<$ 0.05, Fig. 4A–C).

Table 5
The pathways associated with the differentially expressed genes (DEGs) between patients with high and low recurrence risk

Term	Count	$P$ Value	Genes
hsa04080: Neuroactive ligand-receptor interaction	28	1.702 $\times^{-09}$	DRD2, HTR1B, NPFFR2, NPFFR1, HTR1D, NTSR1…
hsa00980: Metabolism of xenobiotics by cytochrome P450	9	5.128 $\times^{-05}$	GSTA3, ADH4, UGT2B11, ADH7, AKR1C1…
hsa04610: Complement and coagulation cascades	8	5.609 $\times^{-04}$	F11, FGB, SERPINA5, F2, SERPIND1…
hsa04060: Cytokine-cytokine receptor interaction	14	6.766 $\times^{-03}$	IL11, CCR6, CCL14, TNFSF11, CXCR5, IL22RA2…
hsa00830: Retinol metabolism	5	8.887 $\times^{-03}$	CYP2A13, ADH4, UGT2B11, UGT2B4, ADH7
hsa00380: Tryptophan metabolism	4	1.299 $\times^{-02}$	KYNU, OGDHL, ACMSD, IDO2
hsa00500: Starch and sucrose metabolism	4	1.448 $\times^{-02}$	ENPP3, UGT2B11, UGT2B4, AMY2B
hsa04020: Calcium signaling pathway	8	3.174 $\times^{-02}$	ERBB4, DRD5, GRIN2A, ADRA1A, PTGFR, NTSR1…
hsa04514: Cell adhesion molecules (CAMs)	6	4.022 $\times^{-02}$	CADM3, NRXN2, CD40LG, CLDN6, NLGN4X, CD22
hsa04664: Fc epsilon RI signaling pathway	4	4.471 $\times^{-02}$	FCER1A, LAT, MS4A2, PLA2G3

Figure 6.

The Gene Ontology biological processes associated with the differentially expressed genes (DEGs) between patients with high and low recurrence risk.

3.4 Identification of pathways related to the risk of recurrence

To investigate the molecular mechanisms that may relevant to the recurrence in LUAD, the 426 patients were assigned into the high-risk and low-risk groups using the prognostic model. Subsequently, 616 DEGs were selected between the two groups (Fig. 5A), including 378 DEGs that overlapped between DEGs by the comparison of LUAD tumor with and without recurrence (Fig. 5B and C). Enrichment analysis showed these DEGs were associated with 16 Gene Ontology biological processes including GO:0051969 $\sim$ regulation of transmission of nerve impulse (including HTR1B), GO:0031644 $\sim$ regulation of neurological system process (including HTR1B), GO:0007267 $\sim$ cell-cell signaling (including HTR1B) and GO:0006955 $\sim$ immune response (including TNFSF11; Fig. 6A). Pathway enrichment analysis showed HTR1B was associated with hsa04080: Neuroactive ligand-receptor interaction and TNFSF11 was involved in hsa04060: Cytokine-cytokine receptor interaction pathway, respectively (Table 5).

4. Discussion

This study identified prognostic model of a combination of 8-gene signature with high predictive power for recurrence in LUAD. The 8 genes include 4 recurrence “negative” DEGs (including AZGP1, INPP5J, MYBPH and SPIB) and 4 “positive” DEGs (including GUCA2A, HTR1B, SLC15A1 and TNFSF11). We divided patients into two groups according to the median risk scores according to the 8-gene signature, and confirmed the significant differences in RFS and recurrence rate between patients in the high-risk and low-risk groups, and the high accuracy (AUC $>$ 0.8). This high performance of this model in predicting the postoperative recurrence in LUAD highlighted the crucial roles of the 8 genes in LUAD recurrence.

Gene expression profiles have important effect on the local and distant recurrence in tumors apart from the histopathology stage [3, 13, 15]. A 21-gene recurrence score was used for the prediction of local and distant recurrence in most commonly diagnosed breast cancer in women [11, 12]. The 21-gene model showed notable performance in N0 and N $+$ hormone receptor – positive patients [11], as well as in hormone-receptor – positive, human epidermal growth factor receptor type 2 (HER2)-negative and axillary node – negative breast cancers [12]. A CCP scoring system based on 31 proliferation-related genes, including cyclin A2 (CCNA2), CCNE2, cell division cycle 25A (CDC25A), kinase inhibitor 1A (CDKN1A) and CDNK1B, showed high performance in predicting and stratifying early-stage LUAD at risk of recurrence [17, 24]. These clinical reports demonstrated the efficacy and performance of multiple gene signature in predicting the recurrence in cancer.

Our present study identified a recurrence risk model with high performance for predicting the recurrence in LUAD. The recurrence risk model consists of the signatures of 8 genes with multiple functions. Genes including HTR1B [25] and GUCA2A [26], peptide absorption and transport like SLC15A1 [27] take important roles in metabolism, TNFSF11 relates to bone formation [28], AZGP1 involves in tumor growth and recurrence [29, 30], and MYBPH participates in cell motility and metastasis [31]. However, none of these genes has been identified to be related with the recurrence in cancers including LUAD. Here in our present study we identified the association of them with the recurrence in LUAD. This finding may show the crucial roles of them in LUAD development and progression.

Of these 8 genes, MYBPH and AZGP1 have been reported to be related to LUAD [31, 32]. MYBPH is a myosin binding protein with largely unknown function [33]. It shares sequence and structural similarity with cardiac MYBPC which has significant impact on contractility [34, 35, 36]. MYBPC3 mutation accounts for $\sim$ 25% hypertrophic cardiomyopathy [35]. Hosono et al. [31] showed that MYBPH inhibited the activation of Rho kinase 1 (ROCK1) via directly interacting with ROCK1, and negatively regulated the organization of actomyosin and thus reducing cell motility, invasion and LUAD metastasis. They also revealed that MYBPH was a transcriptional target of thyroid transcription factor 1 (TTF-1/NKX2-1), a lung development regulator that regulates lung epithelial morphogenesis [37]. TTF-1 expression was positive in most LUAD and was positively associated with good overall survival in patients with advanced LUAD [38, 39]. Kyuichi et al. [39] suggested that the recurrence rate in TTF-1-negative LUAD was significantly higher than that in TTF-1-positive patients. These results suggested the important role of TTF-1 expression in LUAD recurrence prediction. In addition, the express of another largely unknown gene AZGP1 was downregulated in recurrent LUAD samples and the content of serum AZGP1 autoantibody was identified to be positively correlated with the 5-year overall survival in LUAD [32]. AZGP1 was reported negatively correlated with prostate cancer recurrence [29]. The inclusion of AZGP1 and TTF-1-targeted MYBPH in the recurrence risk model suggested the potent clinical value of using this model for predicting LUAD recurrence.

Other genes, including INPP5J and GUCA2A were reported to be directly or indirectly correlated with cancer cell proliferation, tumor size and survival [40, 41]. Kondo et al. had shown the high INPP5J expression in breast cancer and its positive correlation with disease-free survival [41]. GUCA2A is a component of guanylin-GUCY2C tumor suppressor axis [42]. Guanylin stimulates lipolysis in human visceral adipocytes [43]. The silencing of guanylin – GUCY2C axis promoted obesity-induced colorectal cancer [44]. The inclusion of INPP5J and GUCA2A as well as the other genes like SPIB, HTR1B, SLC15A1 and TNFSF11 suggested that complex mechanism underlying LUAD development and recurrence.

5. Conclusions

Our study presented a novel recurrence risk model of 8-gene signature with high performance in predicting the recurrence in LUAD. These 8 genes, including AZGP1, INPP5J, MYBPH, SPIB, GUCA2A, HTR1B, SLC15A1 and TNFSF11 had complex molecular functions, and most of them had not been reported to be correlated with LUAD before. Our present study highlighted their correlation with LUAD recurrence and development. However, clinical validation using a large cohort should be performed to validate the performance of this model which might be of great value for predicting postoperative recurrence in LUAD and making timely prevention.

Footnotes

Conflict of interest

The authors declare that they have no competing interests.

References

Bray

Ferlay

Soerjomataram

Siegel

R.L.

Torre

L.A.

and Jemal

, Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA: A Cancer Journal for Clinicians 68 (2018), 394–424.

Suresh

Ali

Ahmad

Philip

P.A.

and Sarkar

F.H.

, The role of cancer stem cells in recurrent and drug-resistant lung cancer, Oxygen Transport to Tissue 890 (2016), 57–74.

Dziedzic

D.A.

Rudzinski

Langfort

and Orlowski

, Risk factors for local and distant recurrence after surgical treatment in patients with non-small-cell lung cancer, Clinical Lung Cancer 17 (2016), e157–e167.

Taylor

M.D.

Nagji

A.S.

Bhamidipati

C.M.

Theodosakis

Kozower

B.D.

Lau

C.L.

and Jones

D.R.

, Tumor recurrence after complete resection for non-small cell lung cancer, Annals of Thoracic Surgery 93 (2012), 1813–1821.

Hans-Stefan

Gesine

Günther

Christiane

Andreas

Rolf-Edgar

and Stefan

, Matrix metalloproteinase-12 expression correlates with local recurrence and metastatic disease in non-small cell lung cancer patients, Clinical Cancer Research 11 (2005), 1086–1092.

Lander

E.S.

, Comprehensive molecular profiling of lung adenocarcinoma, Nature 511 (2014), 543–550.

Mattonen

S.A.

Palma

D.A.

Johnson

Louie

A.V.

Landis

Rodrigues

Chan

Etemad-Rezai

Yeung

T.P.C.

and Senan

, Detection of local cancer recurrence after stereotactic ablative radiation therapy (SABR) for lung cancer: physician performance versus radiomic assessment, International Journal of Radiation Oncology Biology Physics 94 (2016), 1121–1128.

Hideki

Kyuichi

Chaft

J.E.

Daniel

Sima

C.S.

Ming-Ching

James

Travis

W.D.

Rizk

N.P.

and Rudin

C.M.

, Solid predominant histologic subtype in resected stage i lung adenocarcinoma is an independent predictor of early, extrathoracic, multisite recurrence and of poor postrecurrence survival, Journal of Clinical Oncology 33 (2015), 2877–2884.

Kelsey

C.R.

Marks

L.B.

Donna

Hubbs

J.L.

Ready

N.E.

D’Amico

T.A.

and Boyd

J.A.

, Local recurrence after surgery for early stage lung cancer: an 11-year experience with 975 patients, Cancer 115 (2010), 5218–5227.

10.

Isaksson

Jönsson

Monsef

Brunnström

Bendahl

P.O.

Jönsson

Staaf

and Planck

, CA 19-9 and CA 125 as potential predictors of disease recurrence in resectable lung adenocarcinoma, Plos One 12 (2017), e0186284.

11.

Mitch

Jack

Christopher

John

Mallon

E.A.

Janine

Emma

Anita

Michael

and Aman

, Prediction of risk of distant recurrence using the 21-gene recurrence score in node-negative and node-positive postmenopausal patients with breast cancer treated with anastrozole or tamoxifen: a TransATAC study, Journal of Clinical Oncology 28 (2010), 1829–1834.

12.

Sparano

J.A.

Gray

R.J.

Makower

D.F.

Pritchard

K.I.

Albain

K.S.

Hayes

D.F.

Geyer

C.E.

Dees

E.C.

Perez

E.A.

and Olson

J.A.

, Prospective Validation of a 21-Gene Expression Assay in Breast Cancer, New England Journal of Medicine 373 (2015), 2005–2014.

13.

Larsen

J.E.

Pavey

S.J.

Passmore

L.H.

Bowman

R.V.

Hayward

N.K.

and Fong

K.M.

, Gene expression signature predicts recurrence in lung adenocarcinoma, Clinical Cancer Research 13 (2007), 2946–2954.

14.

Jie

, Gene-expression profiles predict survival of patients with lung adenocarcinoma, Nature Medicine 8 (2002), 816–824.

15.

Hidehiko

Hidetaka

Takamitsu

Takeshi

Tsunehiro

and Kosei

, Overexpression of MACC1 mRNA in lung adenocarcinoma is associated with postoperative recurrence, Journal of Thoracic & Cardiovascular Surgery 141 (2011), 895–898.

16.

Eguchi

Kadota

Chaft

Evans

Kidd

Tan

K.S.

Dycoco

Kolquist

Davis

and Hamilton

S.A.

, Cell cycle progression score is a marker for five-year lung cancer-specific mortality risk in patients with resected stage I lung adenocarcinoma, Oncotarget 7 (2016), 35241–35256.

17.

Aramini

Casali

Stefani

Bettelli

Wagner

Sangale

Hughes

Lanchbury

J.S.

Maiorana

and Morandi

, Prediction of distant recurrence in resected stage I and II lung adenocarcinoma, Lung Cancer 101 (2016), 82–87.

18.

Ritchie

M.E.

Belinda

Yifang

Law

C.W.

Wei

and Smyth

G.K.

, limma powers differential expression analyses for RNA-sequencing and microarray studies, Nucleic Acids Research 43 (2015), e47.

19.

Wang

Cao

Zeng

Wang

Cheng

Zhu

and Hai

, RNA-seq analyses of multiple meristems of soybean: novel and alternative transcripts, evolutionary and functional implications, BMC Plant Biology, 14, 1 (2014-06-17) 14 (2014), 169.

20.

Wang

Hang

Zou

and Mao

J.H.

, A novel gene expression-based prognostic scoring system to predict survival in gastric cancer, Oncotarget 7 (2016), 55343–55351.

21.

Tibshirani

, The lasso method for variable selection in the Cox model, Statistics in Medicine 16 (1997), 385–395.

22.

Camp

R.L.

Marisa

D.F.

and Rimm

D.L.

, X-tile: a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization, Clinical Cancer Research 10 (2004), 7252–7259.

23.

Huang

Zhao

Xia

Wei

Chen

and Shi

, FGF2 and FAM201A affect the development of osteonecrosis of the femoral head after femoral neck fracture, Gene 652 (2018), 39–47.

24.

Raphael

Elisha

Susanne

Gutin

A.S.

Lanchbury

J.S.

Yifan

Archer

M.A.

Corinne

Jones

J.T.

and Kristen

, Validation of a molecular and pathological model for five-year mortality risk in patients with early stage lung adenocarcinoma, Journal of Thoracic Oncology 10 (2015), 67–73.

25.

Hernández

Camarena

González

Caballero

Flores

and Aguilar

, A family-based association study of the HTR1B gene in eating disorders, Revista Brasileira De Psiquiatria 38 (2016), 239–345.

26.

Rodríguez

Gómez-Ambrosi

Catalán

Ezquerro

Méndez-Giménez

Becerril

Ibáñez

Vila

Margall

M.A.

and Moncada

, Guanylin and uroguanylin stimulate lipolysis in human visceral adipocytes, International Journal of Obesity 40 (2016), 1405–1415.

27.

Zhang

Cheng

Y.C.

Lee

W.M.

and Hu

, Slc15a1 is involved in the transport of synthetic F5-peptide into the seminiferous epithelium in adult rat testes, Scientific Reports 5 (2015), 16271.

28.

Mencej

Preželj

Kocijančič

Ostanek

and Marc

, Association of TNFSF11 gene promoter polymorphisms with bone mineral density in postmenopausal women, Maturitas 55 (2006), 219–226.

29.

Henshall

S.M.

Horvath

L.G.

Quinn

D.I.

Eggleton

S.A.

Grygiel

J.J.

Stricker

P.D.

Biankin

A.V.

Kench

J.G.

and Sutherland

R.L.

, Zinc-alpha2-glycoprotein expression as a predictor of metastatic prostate cancer following radical prostatectomy, Journal of the National Cancer Institute 98 (2006), 1420.

30.

Sidaway

, Prostate cancer: AZGP1 expression predicts favourable outcomes, Nature Reviews Urology 14 (2017), 391.

31.

Hosono

Yamaguchi

Eri

Kiyoshi

Chinatsu

Shuta

Yukako

Michiyo

Seiichi

and Kohei

, MYBPH, a transcriptional target of TTF-1, inhibits ROCK1, and reduces cell motility and metastasis, Embo Journal 31 (2014), 481–493.

32.

Albertus

D.L.

Seder

C.W.

Chen

Wang

Hartojo

Lin

Silvers

Thomas

D.G.

Giordano

T.J.

and Chang

A.C.

, AZGP1 autoantibody predicts survival and histone deacetylase inhibitors increase expression in lung adenocarcinoma, Journal of Thoracic Oncology 3 (2008), 1236–1244.

33.

Mouton

Loos

Moolman-Smook

J.C.

and Kinnear

C.J.

, Ascribing novel functions to the sarcomeric protein, myosin binding protein H (MyBPH) in cardiac sarcomere contraction, Experimental Cell Research 331 (2015), 338–351.

34.

Kellermayer

Sziklai

Papp

Decker

Lakatos

and Martonfalvi

, Topology of interactions between titin molecules and myosin thick filaments, Biophysical Journal 114 (2018), 646a.

35.

Ribeiro

M.C.

Birket

M.J.

Tertoolen

Monshouwer

, Kloots Mummery

C.L.

and Passier

, Hypertrophic cardiomyopathy – impaired sarcomeric motion and contractile dysfunction caused by defective MYBPC3 in hPSC derived cardiomyocytes, From Fetus Towards Adult 123 (2016).

36.

Tschöpe

Kherad

Klein

Lipp

Blaschke

Gutterman

Burkhoff

Hamdani

Spillmann

and Van

L.S.

, Cardiac contractility modulation: mechanisms of action in heart failure with reduced ejection fraction and beyond, European Journal of Heart Failure 21 (2019), 14–22.

37.

Minoo

Hamdan

Warburton

Stepanik

and Delemos

, TTF-1 regulates lung epithelial morphogenesis, Developmental Biology 172 (1995), 694–698.

38.

Schilsky

J.B.

Ahn

Datta

Travis

W.D.

Kris

M.G.

Chaft

J.E.

Rekhtman

and Hellmann

M.D.

, Prognostic impact of TTF-1 expression in patients with stage IV lung adenocarcinomas, Lung Cancer 108 (2017), 205–211.

39.

Kyuichi

Jun-Ichi

Sarkaria

I.S.

Sima

C.S.

Xiaoyu

Akihiko

Rusch

V.W.

Travis

W.D.

and Adusumilli

P.S.

, Thyroid transcription factor-1 expression is an independent predictor of recurrence and correlates with the IASLC/ATS/ERS histologic classification in patients with stage I lung adenocarcinoma, Cancer 119 (2013), 931–938.

40.

Zhu

Yuan

Wang

Gong

Xie

and Li

, MiR-661 contributed to cell proliferation of human ovarian cancer cells by repressing INPP5J expression, Biomedicine & Pharmacotherapy 75 (2015), 123–128.

41.

Kondo

Kim

T.S.

Wanifuchi

Hato

Hisada

Nishimoto

Nishikawa

and Toyama

, Abstract P6-07-34: The prognostic impact of inositol polyphosphate 5-phosphatase PIPP (INPP5J) expression in breast cancer tissue, San Antonio Breast Cancer (2017). 10.1158/1538-7445.

42.

Blomain

E.S.

Lin

J.E.

Colóngonzález

Kim

G.W.

Hyslop

Zhan

Snook

A.E.

and Waldman

S.A.

, Abstract 2882: calorie-induced silencing of the tumor suppressive guanylin-GUCY2C paracrine axis underlies colorectal cancer in obesity, Cancer Research 75 (2015), 2882–2882.

43.

Rodríguez

Gómezambrosi

Catalán

Ezquerro

Méndezgiménez

Becerril

Ibáñez

Vila

Margall

M.A.

and Moncada

, Guanylin and uroguanylin stimulate lipolysis in human visceral adipocytes, International Journal of Obesity 40 (2016), 1405–1415.

44.

Lin

J.E.

Colon-Gonzalez

Blomain

Kim

G.W.

Aing

Stoecker

Rock

Snook

A.E.

Zhan

and Hyslop

T.M.

, Obesity-induced colorectal cancer is driven by caloric silencing of the guanylin-GUCY2C paracrine signaling axis, Cancer Research 76 (2016), 339–346.

Eight-gene signature predicts recurrence in lung adenocarcinoma

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

2. Materials and methods

2.1 Study population and assignment

2.2 DEGs identification

2.3 Identification of recurrence-associated DEGs

2.4 Selection of the optimal combination of DEGs

2.5 Establishment of the recurrence risk model

Table 1 Baseline characteristics of patients with LUAD

2.7 Identification of characteristics that associate with LUAD recurrence

2.8 Identification and enrichment of the DEGs between patients with high and low risk of recurrence in LUAD

2.9 Statistical analysis

3. Results

3.1 Identification of LUAD recurrence-associated DEGs

Table 2 The 8 LUAD prognosis-associated differentially expressed genes in the optimal group by LASSO Cox regression model

Table 5 The pathways associated with the differentially expressed genes (DEGs) between patients with high and low recurrence risk

4. Discussion

5. Conclusions

Footnotes

Conflict of interest

References

Table 1
Baseline characteristics of patients with LUAD

Table 2
The 8 LUAD prognosis-associated differentially expressed genes in the optimal group by LASSO Cox regression model

Table 5
The pathways associated with the differentially expressed genes (DEGs) between patients with high and low recurrence risk