Integrative analysis and machine learning application to TCGA DNA methylation, RNA-sequencing, and variant dataset in endometrial cancer for predicting recurrence

초록

Objectives According to the cancer genome atlas (TCGA), endometrial cancer (EC) is divided into 4 molecular subtypes with distinctive prognostic features. However, the prognosis within each molecular subtype can vary depending on histologic and molecular factors. Recently, multi-omics studies evaluating diagnostic and prognostic markers in EC are increasing, but no study exists that tried to perform multi-omics analysis to find prognostic factors within each molecular subtype. By merging omics datasets and machine learning analyses, we attempted to identify biomarkers related to the recurrence of EC within each molecular subtype. Methods From TCGA multi-omics datasets, 116 EC samples with DNA methylation, RNA-sequencing, and common variants, as well as recurrence data, were used for analysis. Differentially expressed genes (DEGs) and differentially methylated regions (DMRs) were retrieved by t-test between the recurrence and non-recurrence groups. DEGs and DMRs were visualized using volcano plots and heat maps. The machine learning approaches, namely decision tree (DT) and random forest (RF) models, were used to find the molecular factors to best discriminate the following 4 groups: copy number-high (CN-high) with recurrence, CN-high without recurrence, copy number-low (CN-low) with recurrence, and CN-low without recurrence. Critical gene expressions or DNA methylation levels in each molecular subtype were presented as a boxplot. Results Three omics datasets were merged with 378,278 CpG sites, 53,409 normalized gene expression levels, 118,555 genomic region variants, and 18,603 gene variants. Through DT and RF, PARD6G-AS1, CSMD1, TESC, and CD44 were selected. Boxplot revealed that hypomethylation of PARD6G-AS1, hypermethylation of CSMD1, and increased TESC expression were significantly associated with an increased risk of recurrence in the CN-high group. Also, overexpression of CD44 was significantly associated with recurrence in the CN-low group. After validation of these 4 parameters using TCGA raw data, PARD6G-AS1 and CD44 maintained their statistical significance (P = 0.006 and P = 0.020, respectively). Hypomethylation of PARD6G-AS1 in CN-high and CD44 overexpression in CN-low demonstrated significant association with both advanced stage and lymph node metastasis. Conclusions Hypomethylation of PARD6G-AS1 in CN-high and CD44 overexpression in CN-low could be predictive markers in endometrial cancer recurrence.