methyLImp2: faster missing value estimation for DNA methylation data

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• DNA methylation data imputation is crucial for analyzing incomplete epigenomic datasets.
• Existing general-purpose imputation methods, like methyLImp, show competitive performance but suffer from long running times.
• The computational burden limits the application of imputation methods to large-scale DNA methylation datasets.

Purpose of the Study:

• To significantly reduce the running time of the methyLImp DNA methylation imputation method.
• To enable the analysis of large DNA methylation datasets that were previously computationally infeasible.
• To maintain the prediction performance of the original methyLImp method while improving efficiency.

Main Methods:

• Implemented a chromosome-wise parallel version of methyLImp to leverage multi-core processing.
• Introduced a mini-batch approach to process subsets of samples, reducing memory requirements and computation time.
• Evaluated the performance and runtime of methyLImp2 on large-scale DNA methylation datasets.

Main Results:

• Achieved a several 10-fold reduction in runtime through chromosome-wise parallelization.
• Further reduced runtime from days to hours or minutes using the mini-batch approach for large datasets.
• Demonstrated that methyLImp2 maintains high prediction accuracy comparable to the original methyLImp method.

Conclusions:

• methyLImp2 offers a computationally efficient solution for DNA methylation data imputation.
• The parallel and mini-batch strategies make methyLImp2 suitable for large-scale epigenomic studies.
• The R package methyLImp2 is available on Github and under review for Bioconductor.