DeepMethyGene: a deep-learning model to predict gene expression using DNA methylations.

Area of Science:

• Epigenetics
• Computational Biology
• Genomics

Background:

• Gene expression dictates cellular functions.
• DNA methylation is a key epigenetic regulator of gene expression.
• Predicting gene expression from DNA methylation is crucial for understanding biological processes.

Purpose of the Study:

• To develop an advanced deep learning model, DeepMethyGene, for predicting gene expression using DNA methylation data.
• To improve upon existing state-of-the-art models in predictive performance.
• To investigate the relationship between DNA methylation patterns and gene expression levels.

Main Methods:

• Developed DeepMethyGene, an adaptive recursive convolutional neural network based on ResNet.
• Transformed methylation Beta values to M values for data optimization.
• Implemented residual blocks to address gradient vanishing in deep networks.

Main Results:

• DeepMethyGene achieved superior predictive performance (R² = 0.640) compared to the geneEXPLORE model (R² = 0.449).
• Prediction accuracy was significantly influenced by the number of methylation sites and their distance to transcription start sites (TSS).

Conclusions:

• DeepMethyGene demonstrates high efficacy in predicting gene expression from DNA methylation.
• Understanding the methylation-gene expression interplay provides a foundation for disease prediction and clinical strategies.
• The study provides accessible code and data for further research.