Missing Value Imputation Approach for Mass Spectrometry-based Metabolomics Data

Abstract

Missing values exist widely in mass-spectrometry (MS) based metabolomics data. Various methods have been applied for handling missing values, but the selection can significantly affect following data analyses. Typically, there are three types of missing values, missing not at random (MNAR), missing at random (MAR), and missing completely at random (MCAR). Our study comprehensively compared eight imputation methods (zero, half minimum (HM), mean, median, random forest (RF), singular value decomposition (SVD), k-nearest neighbors (kNN), and quantile regression imputation of left-censored data (QRILC)) for different types of missing values using four metabolomics datasets. Normalized root mean squared error (NRMSE) and NRMSE-based sum of ranks (SOR) were applied to evaluate imputation accuracy. Principal component analysis (PCA)/partial least squares (PLS)-Procrustes analysis were used to evaluate the overall sample distribution. Student’s t-test followed by correlation analysis was conducted to evaluate the effects on univariate statistics. Our findings demonstrated that RF performed the best for MCAR/MAR and QRILC was the favored one for left-censored MNAR. Finally, we proposed a comprehensive strategy and developed a public-accessible web-tool for the application of missing value imputation in metabolomics (https://metabolomics.cc.hawaii.edu/software/MetImp/).

Nutrimetabolomics: An Integrative Action for Metabolomic Analyses in Human Nutritional Studies

“…A typical metabolomic dataset may contain 20% of missing data in up to 80% of all variables . Missingness can occur from analytical, computational, or biological causes, but, regardless of origin, a principled strategy to deal with them is required. Features showing a high degree of missingness can, for example, be removed from the data matrix whereas remaining missing values can be imputed, i.e.…”

NS-kNN: a modified k-nearest neighbors approach for imputing metabolomics data

“…To simulate measurements missing due to instrument limits of detection, a percentage of the lowest abundance values for each metabolite were identified for potential removal, similar to cutoffs used in previous studies (Shah et al, 2017; Wei et al, 2018b). Of the values identified, the lowest 80% abundance values were removed (representing values below a hypothetical limit of detection).…”

“…Importantly, a sample can be a neighbor only if it has a measured value for the metabolite missing in the target sample. (Alternatively, nearest neighbor metabolites can be found instead of samples (Armitage et al, 2015; Gromski et al, 2014; Wei et al, 2018b)). A weighted combination of the corresponding values for the missing metabolite in the nearest neighbors is used as the imputed value.…”

“…Although there have been a few studies in the literature examining the differences between imputation methods on metabolomics data (Armitage et al, 2015; Di Guida et al, 2016; Gromski et al, 2014; Wei et al, 2018b), the creation of realistic datasets for use in such studies has gone relatively unexplored. This is a critical step because the ideal assessment of imputations would be a comparison to the actual missing data, but these data are (by definition) not available for real experimental datasets.…”

“…While kNN and random forest methods have been previously shown to perform better than simplistic replacement methods, these algorithms inherently assume data to be MCAR (Wei et al, 2018b). Because we expect more realistic metabolomics data to contain both types of missing data, and often a majority that are MNAR, it is critical that imputation algorithms account for MNAR values in order to reduce imputation errors.…”