For cancer diagnosis, many DNA methylation markers have been identified. However, few studies have tried to identify DNA methylation markers to diagnose diverse cancer types simultaneously, i.e., pan-cancers. In this study, we tried to identify DNA methylation markers to differentiate cancer samples from the respective normal samples in pan-cancers. We collected whole genome methylation data of 27 cancer types containing 10,140 cancer samples and 3386 normal samples, and divided all samples into five data sets, including one training data set, one validation data set and three test data sets. We applied machine learning to identify DNA methylation markers, and specifically, we constructed diagnostic prediction models by deep learning. We identified two categories of markers: 12 CpG markers and 13 promoter markers. Three of 12 CpG markers and four of 13 promoter markers locate at cancer-related genes. With the CpG markers, our model achieved an average sensitivity and specificity on test data sets as 92.8% and 90.1%, respectively. For promoter markers, the average sensitivity and specificity on test data sets were 89.8% and 81.1%, respectively. Furthermore, in cell-free DNA methylation data of 163 prostate cancer samples, the CpG markers achieved the sensitivity as 100%, and the promoter markers achieved 92%. For both marker types, the specificity of normal whole blood was 100%. To conclude, we identified methylation markers to diagnose pan-cancers, which might be applied to liquid biopsy of cancers.

Whole-genome bisulfite sequencing generates a comprehensive profiling of the gene methylation levels, but is limited by a high cost. Recent studies have partitioned the genes into landmark genes and target genes and suggested that the landmark gene expression levels capture adequate information to reconstruct the target gene expression levels. This inspired us to propose that the methylation level of the promoters in landmark genes might be adequate to reconstruct the promoter methylation level of target genes, which would eventually reduce the cost of promoter methylation profiling. Here, we propose a deep learning model called Deep-Gene Promoter Methylation (D-GPM) to predict the whole-genome promoter methylation level based on the promoter methylation profile of the landmark genes from The Cancer Genome Atlas (TCGA). D-GPM-15%-7000 × 5, the optimal architecture of D-GPM, acquires the least overall mean absolute error (MAE) and the highest overall Pearson correlation coefficient (PCC), with values of 0.0329 and 0.8186, respectively, when testing data. Additionally, the D-GPM outperforms the regression tree (RT), linear regression (LR), and the support vector machine (SVM) in 95.66%, 92.65%, and 85.49% of the target genes by virtue of its relatively lower MAE and in 98.25%, 91.00%, and 81.56% of the target genes based on its relatively higher PCC, respectively. More importantly, the D-GPM predominates in predicting 79.86% and 78.34% of the target genes according to the model distribution of the least MAE and the highest PCC, respectively.

Background: Gene promoter methylation plays a critical role in a wide range of biological processes, such as transcriptional expression, gene imprinting, X chromosome inactivation, etc . Whole-genome bisulfite sequencing generates a comprehensive profiling of the gene methylation levels but is limited by a high cost. Recent studies have partitioned the genes into landmark genes and target genes and suggested that the landmark gene expression levels capture adequate information to reconstruct the target gene expression levels. Moreover, the methylation level of the promoter is usually negatively correlated with its corresponding gene expression. This result inspired us to propose that the methylation level of the promoters might be adequate to reconstruct the promoter methylation level of target genes, which would eventually reduce the cost of promoter methylation profiling. Results: Here, we developed a deep learning model (D-GPM) to predict the whole-genome promoter methylation level based on the methylation profile of the landmark genes. We benchmarked D-GPM against three machine learning methods, namely, linear regression (LR), regression tree (RT) and support vector machine (SVM), based on two criteria: the mean absolute deviation (MAE) and the Pearson correlation coefficient (PCC). After profiling the methylation beta value (MBV) dataset from the TCGA, with respect to MAE and PCC, we found that D-GPM outperforms LR by 9.59% and 4.34%, RT by 27.58% and 22.96% and SVM by 6.14% and 3.07% on average, respectively. For the number of better-predicted genes, D-GPM outperforms LR in 92.65% and 91.00%, RT in 95.66% and 98.25% and SVM in 85.49% and 81.56% of the target genes. Conclusions: D-GPM acquires the least overall MAE and the highest overall PCC on MBV-te compared to LR, RT, and SVM. For a genewise comparative analysis, D-GPM outperforms LR, RT, and SVM in an overwhelming majority of the target genes, with respect to the MAE and PCC. Most importantly, D-GPM predominates among the other models in predicting a majority of the target genes according to the model distribution of the least MAE and the highest PCC for the target genes.