Completing Single-Cell DNA Methylome Profiles via Transfer Learning Together With KL-Divergence

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Whole-genome bisulfite sequencing (WGBS) yields sparse methylome data with low biological material.
• This sparsity limits studies on systems with limited sample availability, such as mammalian preimplantation embryos.
• Existing imputation methods struggle with minimal data coverage or lack of complementary data.

Purpose of the Study:

• To develop and evaluate computational methods for imputing sparse methylome profiles using transfer learning and Kullback-Leibler (KL) divergence.
• To enhance the utility of WGBS data from low-input samples, particularly in early embryonic development.
• To improve data coverage for downstream methylome analyses.

Main Methods:

• Explored transfer learning to leverage less sparse methylome profiles from different tissues of the same species.
• Utilized KL divergence to maximize information extraction from input data.
• Employed a deep neural network to capture DNA sequence and local methylation patterns for imputation.
• Trained predictive models for methylome completion with very low coverage (below 2%).

Main Results:

• Transfer learning individually improved prediction performance by 29.98%.
• KL divergence individually improved prediction performance by 29.43%.
• The combined use of transfer learning and KL divergence resulted in a 38.70% increase in prediction performance.
• Imputation drastically increased data coverage from 0.06-1.47% to 43.80-73.6%.

Conclusions:

• Transfer learning and KL divergence are effective for imputing sparse methylome profiles, especially for low-coverage data.
• The developed approach significantly enhances data coverage, enabling previously infeasible downstream analyses.
• This method holds promise for studying epigenetics in systems with limited biological material, like early embryos.