Evaluation of different approaches for missing data imputation on features associated to genomic data

Area of Science:

• Genomics
• Bioinformatics
• Data Science

Background:

• Missing data is prevalent in genomics, potentially biasing analyses.
• Data imputation methods are crucial for handling missing values in genomic datasets.
• Understanding missing data types (MCAR, MAR, MNAR) is key for effective imputation.

Purpose of the Study:

• To evaluate six imputation approaches for genomic data.
• To identify optimal imputation methods for non-coding variants.
• To assess the performance of imputation algorithms using ClinVar data.

Main Methods:

• Utilized 31,245 variants from ClinVar, annotated with 13 genome-wide features.
• Applied and compared six distinct imputation algorithms.
• Evaluated imputation performance based on data characteristics and algorithm efficiency.

Main Results:

• Random Forest and kNN algorithms demonstrated superior performance in imputation tasks.
• Conservation scores (phyloP7, phyloP20) showed robust imputation across methods.
• PhasCons scores exhibited poor imputation performance irrespective of the algorithm used.

Conclusions:

• Random Forest and kNN are recommended as top imputation methods for genomic data, including non-coding variants.
• kNN provides a computationally feasible alternative to Random Forest for imputation.
• This imputation methodology can enhance variant prioritization in genomic screening.