A deep learning framework for imputing missing values in genomic data

MotivationThe presence of missing values is a frequent problem encountered in genomic data analysis. Lost data can be an obstacle to downstream analyses that require complete data matrices. State-of-the-art imputation techniques including Singular Value Decomposition (SVD) and K-Nearest Neighbors (KNN) based methods usually achieve good performances, but are computationally expensive especially for large datasets such as those involved in pan-cancer analysis.\n\nResultsThis study describes a new method: a denoising autoencoder with partial loss (DAPL) as a deep learning based alternative for data imputation. Results on pan-cancer gene expression data and DNA methylation data from over 11,000 samples demonstrate significant improvement over standard denoising autoencoder for both data missing-at-random cases with a range of missing percentages, and missing-not-at-random cases based on expression level and GC-content. We discuss the advantages of DAPL over traditional imputation methods and show that it achieves comparable or better performance with less computational burden.\n\nAvailabilityhttps://github.com/gevaertlab/DAPL\n\nContactogevaert@stanford.edu