D-GPM: A Deep Learning Method for Gene Promoter Methylation Inference.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Whole-genome bisulfite sequencing provides comprehensive gene methylation profiling but is costly.
• Landmark gene expression can reconstruct target gene expression, suggesting a similar principle for methylation.
• Developing cost-effective promoter methylation profiling methods is crucial for large-scale studies.

Purpose of the Study:

• To propose a deep learning model (D-GPM) for predicting whole-genome promoter methylation levels.
• To leverage promoter methylation profiles of landmark genes to infer those of target genes.
• To reduce the overall cost of promoter methylation profiling.

Main Methods:

• Development of a deep learning model named Deep-Gene Promoter Methylation (D-GPM).
• Utilizing promoter methylation data from landmark genes in The Cancer Genome Atlas (TCGA).
• Comparison of D-GPM performance against regression tree, linear regression, and support vector machine models.

Main Results:

• The optimal D-GPM architecture (D-GPM-15%-7000 × 5) achieved a low Mean Absolute Error (MAE) of 0.0329 and a high Pearson Correlation Coefficient (PCC) of 0.8186 on testing data.
• D-GPM significantly outperformed regression tree, linear regression, and support vector machine models in terms of MAE and PCC across a majority of target genes.
• D-GPM demonstrated superior predictive performance, dominating in the least MAE and highest PCC for a substantial percentage of target genes.

Conclusions:

• The D-GPM model effectively predicts whole-genome promoter methylation levels using landmark gene methylation profiles.
• This deep learning approach offers a cost-effective alternative to traditional whole-genome bisulfite sequencing for promoter methylation analysis.
• D-GPM shows significant potential for advancing large-scale genomic studies by enabling efficient and accurate methylation profiling.