Prognostic score model based on six m6A‐related autophagy genes for predicting survival in esophageal squamous cell carcinoma

Abstract Background Prognostic signatures based on autophagy genes have been proposed for esophageal squamous cell carcinoma (ESCC). Autophagy genes are closely associated with m6A genes. Our purpose is to identify m6A‐related autophagy genes in ESCC and develop a survival prediction model. Methods Differential expression analyses for m6A genes and autophagy genes were performed based on TCGA and HADd databases followed by constructing a co‐expression network. Uni‐variable Cox regression analysis was performed for m6A‐related autophagy genes. Using the optimal combination of feature genes by LASSO Cox regression model, a prognostic score (PS) model was developed and subsequently validated in an independent dataset. Results The differential expression of 13 m6A genes and 107 autophagy genes was observed between ESCC and normal samples. The co‐expression network contained 13 m6A genes and 96 autophagy genes. Of the 12 m6A‐related autophagy genes that were significantly related to survival, DAPK2, DIRAS3, EIF2AK3, ITPR1, MAP1LC3C, and TP53 were used to construct a PS model, which split the training set into two risk groups with significant different survival ratios (p = 0.015, 1‐year, 3‐year, and 5‐year AUC = 0.873, 0.840, and 0.829). Consistent results of GSE53625 dataset confirmed predictive ability of the model (p = 0.024, 1‐year, 3‐year, and 5‐year AUC = 0.793, 0.751, and 0.744). The six‐gene PS score was an independent prognostic factor from clinical factors (HR, 2.362; 95% CI, 1.390–7.064; p‐value = 0.012). Conclusion Our study recommends 6 m6A‐related autophagy genes as promising prognostic biomarkers and develops a PS model to predict survival in ESCC.


| INTRODUC TI ON
Esophageal cancer ranks the eighth for incidence and sixth for cancer-related mortality worldwide and is among the most aggressive human malignancies. 1 Esophageal squamous cell carcinoma (ESCC), the dominant histological subtype in East Asian, has an extremely low five-year survival rate and a high incidence of recurrence and metastasis. 2 Despite recent advancements in systemic therapies of ESCC, there is a lack of approved targeted therapeutics. 3 Discovery of prognostic biomarkers and prediction models for risk stratification holds great promise for future progress in improving patient outcomes and advancing individualized therapies in ESCC.
Autophagy, a lysosome-dependent self-degradation process, suppresses tumor initiation but advances tumor progression, playing opposing roles in the biology of cancer. 4,5 Past studies shows that autophagy plays a positive or negative role in regulating esophageal cell survival in a context-dependent manner and may have important implications for patients outcomes as a promising therapeutic target. 6,7 Efficient and useful prognostic signatures based on autophagy-related genes have been reported for predicting survival in esophageal cancer. 8,9 N6-methyladenosine (m6A) methylation is a commonly seen modification in eukaryotic messenger RNA (mRNA) and m6A regulators are primarily composed of methyltransferases (writers), demethylases (erasers), and RNA binding proteins (readers). 10 m6A regulators have potential as prognostic biomarkers and a strong prognostic signature based on m6A regulators has been proposed for ESCC. 11 m6A methylation plays a crucial role in various tumorigenesis-related biological processes, including autophagy. 12 There is evidence that FTO, a well-known m6A demethylase, is involved in modulating autophagosome formation in autophagy. 13 METTL3, a primary m 6 A methyltransferase, activates autophagyrelated pathways under hypoxia in ESCC. 14 Moreover, two recent studies consistently show that METTL3-mediated m6A methylation negatively modulates autophagy. 15,16 In light of the close relationships between m6A regulators and autophagy genes, we speculated that m6A-related autophagy genes may have greater prognostic significance in ESCC.
In the present study, a comprehensive research into autophagy genes, m6A genes and their correlations in ESCC was performed.
Based on the m6A-related autophagy genes identified, a prognostic score (PS) model for survival prediction in ESCC was developed and validated. Our study might shed light on the roles of m6A-related autophagy genes into the crucial molecular mechanisms associated with ESCC prognosis.

| Identification and function annotation of m6A-related autophagy genes
Total 232 autophagy genes were firstly downloaded from HADb

| Development and validation of PS model based on feature autophagy genes
The uni-variate Cox regression analysis was performed for the data in training set to select the autophagy genes that were significantly related to overall survival (OS) time from the autophagy genes in the co-expression network using survival package (version 2.41-1, http://bioco nduct or.org/packa ges/survi valr/) in R. p-Value<0.05 was regarded as a significance threshold.
Least absolute shrinkage and selection operator (LASSO) Cox regression model 21 along with 1000-fold cross-validation likelihood was employed to identify the optimal combination of feature autophagy genes using penalized package (https://cran.r-proje ct.org/ web/packa ges/penal ized/index.html) in R. 22 We put data of the significant survival-related genes in LASSO Cox regression model to achieve variable selection and shrinkage. Cross-validation (CV) procedure was iterated for 1000 times to determine the optimal penalty regularization parameter λ.
We further explored associations of each feature autophagy gene with OS of patients in the training set using survival package.
According to median expression level of each feature autophagy gene, all samples in the training set were divided into high expression group (expression value was higher than the median expression level) and low expression group (expression value was lower than the median expression level), separately.
Based on expression of feature autophagy genes weighted by LASSO Cox regression coefficients, PS was calculated for each sample as follows: Wherein β genes denotes LASSO Cox regression coefficient, and Exp genes denotes gene expression level in the training set.
According to median PS value in the training set or the validation set, samples were split into a high-risk subgroup (PS was higher than the median PS) and low-risk subgroup (PS was lower than the median PS). Survival analysis was carried out using survival package with plotting Kaplan-Meier curves. OS time of different groups was compared using log-rank test. The area under the ROC curve (AUC) is used to assess predictive performance of the PS model. Uni-and multi-variable Cox regression analyses were performed to identify independent prognostic factors using survival package. Log-rank p < 0.05 suggested significance. Based on LASSO Cox regression coefficients and expression levels of the 6 feature autophagy genes, PS was calculated for each sample using the following formula:

| Principal component analysis
Using median PS as cutoff, the training set was separated into a high-risk group and a low-risk group with significantly different OS time (p = 0.015, Figure 5A). Besides, 1-, 3-, and 5-year AUC were 0.873, 0.840, and 0.829, separately ( Figure 5B).

| Successful validation of the PS model in GSE53625 dataset
Prognostic score model based on the 6 feature autophagy genes was further tested on an independent validation set of GSE53625 to verify its prognostic capability in ESCC. The high-risk group had significantly lower survival ratio compared with the low-risk group (p-value = 0.024, Figure 5A), similar to the results of the training set. Figure 5B shows that 1-year, 3-year, and 5-year AUC were 0.793, 0.751, and 0.744. These results proved that PS based on the six feature autophagy genes was efficient and accurate in discriminating high-risk patients from low-risk patients in ESCC. Table 1   Abbreviations: CI, confidence interval; HR, hazard ratio; PS, prognostic score; SD, standard deviation.

| PCA analysis for high-risk and low-risk samples based on m6A genes
relationships between m6A genes and autophagy genes as well as prognostic implications of m6A-related autophagy genes in ESCC.
One important innovation of our study is an m6A-autophagy

| CON CLUS ION
Combining data from TCGA and HADb databases with comprehensive bioinformatics analysis, we unraveled a list of m6A-related autophagy genes in ESCC. We identified 6 promising prognostic genes, developed, and validated a PS model based on them, which performed well in distinguishing high-risk patients from low-risk patients in ESCC. This study provides a more detailed portrait of m6A genes and autophagy genes in the biology of ESCC and facilities individualized outcome prediction for patients.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
Gene expression data analyzed in this study are available from The