Novel Missing Data Imputation Approaches Enhance Quantitative Trait Loci Discovery in Multi-Omics Analysis

Abstract

AbstractHandling missing values in multi-omics datasets is essential for a broad range of analyses. While several benchmarks for multi-omics data imputation methods have recommended certain approaches for practical applications, these recommendations are not widely adopted in real-world data analyses. Consequently, the practical reliability of these methods remains unclear. Furthermore, no existing benchmark has assessed the impact of missing data and imputation on molecular quantitative trait loci (xQTL) discoveries. To establish the best practice for xQTL analysis amidst missing values in multi-omics data, we have thoroughly benchmarked 16 imputation methods. This includes methods previously recommended and in use in the field, as well as two new approaches we developed by extending existing methods. Our analysis indicates that no established method consistently excels across all benchmarks; some can even result in significant false positives in xQTL analysis. However, our extension to a recent Bayesian matrix factorization method,FLASH, exhibits superior performance in multi-omics data imputation across various scenarios. Notably, it is both powerful and well-calibrated for xQTL discovery compared to all the other methods. To support researchers in practically implementing our approach, we have integrated our extension toFLASHinto the R package flashier, accessible athttps://github.com/willwerscheid/flashier. Additionally, we provide a bioinformatics pipeline that implementsFLASHand other methods compatible with xQTL discovery workflows based on tensorQTL, available athttps://cumc.github.io/xqtl-pipeline/code/data_preprocessing/phenotype/phenotype_imputation.html.