Imputing missing RNA-sequencing data from DNA methylation by using a transfer learning-based neural network.

Area of Science:

• Genomics
• Epigenetics
• Bioinformatics

Background:

• Gene expression links DNA to phenotype but RNA-seq data is often missing.
• DNA methylation regulates gene expression and can predict RNA-seq data.
• Existing imputation methods often lack pan-cancer applicability.

Purpose of the Study:

• To develop a novel method for imputing missing RNA-seq data from DNA methylation data.
• To leverage large pan-cancer datasets for improved imputation accuracy.
• To validate the method's utility in downstream cancer research.

Main Methods:

• Developed TDimpute, a transfer learning-based neural network.
• Trained a general model on The Cancer Genome Atlas (TCGA) pan-cancer data.
• Fine-tuned the model on specific cancer datasets for imputation.

Main Results:

• TDimpute significantly outperforms existing methods, improving imputation accuracy by 7-11%.
• Demonstrated utility in downstream analyses: identifying methylation-driving and prognosis-related genes, clustering, and survival analysis.
• Validated on an independent Wilms tumor dataset from the TARGET project.

Conclusions:

• TDimpute provides an effective solution for RNA-seq imputation.
• The method is valuable even with limited training samples.
• Enables robust genomic analyses using imputed gene expression data.