Predicting lung adenocarcinoma disease progression using methylation-correlated blocks and ensemble machine learning classifiers

Abstract

Applying the knowledge that methyltransferases and demethylases can modify adjacent cytosine-phosphorothioate-guanine (CpG) sites in the same DNA strand, we found that combining multiple CpGs into a single block may improve cancer diagnosis. However, survival prediction remains a challenge. In this study, we developed a pipeline named “stacked ensemble of machine learning models for methylation-correlated blocks” (EnMCB) that combined Cox regression, support vector regression (SVR), and elastic-net models to construct signatures based on DNA methylation-correlated blocks for lung adenocarcinoma (LUAD) survival prediction. We used methylation profiles from the Cancer Genome Atlas (TCGA) as the training set, and profiles from the Gene Expression Omnibus (GEO) as validation and testing sets. First, we partitioned the genome into blocks of tightly co-methylated CpG sites, which we termed methylation-correlated blocks (MCBs). After partitioning and feature selection, we observed different diagnostic capacities for predicting patient survival across the models. We combined the multiple models into a single stacking ensemble model. The stacking ensemble model based on the top-ranked block had the area under the receiver operating characteristic curve of 0.622 in the TCGA training set, 0.773 in the validation set, and 0.698 in the testing set. When stratified by clinicopathological risk factors, the risk score predicted by the top-ranked MCB was an independent prognostic factor. Our results showed that our pipeline was a reliable tool that may facilitate MCB selection and survival prediction.

Machine learning in the identification of prognostic DNA methylation biomarkers among patients with cancer: a systematic review of epigenome-wide studies

“…Similar to clustering-based workflows, an initial screening was generally the first step for the 30 studies using supervised feature selection-based workflows (Figure 4.2) (6-9, 24-28, 30, 31, 36, 40-42, 48, 51, 59, 60, 63, 66-68, 71-74, 81, 84, 85). In the next step, nearly all studies used one to three types of supervised feature selection techniques to pre-select prognostically relevant DNA methylation biomarkers prior to training one or more prognostic models.…”

“…In the stage of model training, further feature selection was performed by more than half of the 30 studies (8, 9, 24, 25, 28, 31, 36, 40-42, 48, 51, 66, 68, 72-74, 81, 84, 85). Sample-splitting was the most commonly used internal model validation method ( N = 12), and external validation was performed in 15 studies (7, 8, 24-26, 31, 42, 48, 59, 66-68, 72, 74, 81).…”

“…Lastly, we found that 18 identified studies first constructed a Cox prognostic model based on single selected CpG sites (i.e., methylation score), which was subsequently used as a single predictor, together with other clinical variables, to construct a second Cox prognostic model (6, 31, 36, 39-42, 55, 57, 66-69, 73, 78, 80, 81). It might be an inappropriate approach to construct the first “methylation score”, in which the relative weight of individual CpGs were consequently fixed for the second prognostic model.…”

“…As to studies using more than one supervised ML model for feature selection, four studies intersected CpGs selected by different ML techniques (9,31,48,84), whereas two studies only selected the biomarkers retained in one model with the best performance (72,85). Notably, one study (81) We identified a total of seven studies that applied deep learning-based feature transformation to reduce dimensionality (Figure 4.3). In the first step, five studies used autoencoder (44,49,52,58,76), and one used bidirectional deep neural networks (65), to transform original input features into a reduced number of new features (i.e., hidden nodes), and then they performed univariable Cox regression analysis to select hidden nodes significantly associated with prognosis.…”

“…As to studies using more than one supervised ML model for feature selection, four studies intersected CpGs selected by different ML techniques (9,31,48,84), whereas two studies only selected the biomarkers retained in one model with the best performance (72,85). Notably, one study (81) first identified methylationcorrelated blocks (MCBs, i.e., blocks of closely co-methylated CpGs) and then selected MCBs associated with survival by LASSO. Next, elastic-net model, support vector regression, and Cox regression were separately trained and used for ranking We identified a total of seven studies that applied deep learning-based feature transformation to reduce dimensionality (Figure 4.3).…”