Tackling Missing Values in Mass Spectrometry-based Proteomics Data

Abstract

In the development of therapeutics, analysis of differentially abundant proteins (DAPs) using mass spectrometry (MS) is essential. However, MS-based data suffers from high rates of missing values that severely complicate downstream analyses. Various imputation methods have been proposed to deal with the missing data, but there is no standard protocol for selecting a method. Here we have comprehensively evaluated common methods, to develop a best practice for imputation to inform downstream statistical analyses of MS proteomics data. We compared the performance of five imputation methods in their application to values missing completely at random and missing not at random introduced into data from the Cancer Cell Line Encyclopedia, and data simulated from a multivariate mixed-effects model respectively. Performance was measured in true positive rate (TPR) and false positive (FPR) of detected DAPs (%adj 0 05, est. log2 fold-change ¡1, and an accuracy metric [&] 103). The FPR was below 5% for all methods under all conditions tested. If less than 10% of the data was missing, imputation did not increase the TPR compared to removing missing values. For 30% missingness irrespective of data or missingness type, the TPR was below 80%; and for 50% missingness the TPR was 25- 75% depending on imputation method. Since the FPR was controlled, no artefacts were introduced by any methods under any circumstances. For large proportions of missingness (50%), we recommend imputation with Principal Component Analysis imputation if the sample size is large (= ¡ 50). With small sample sizes (= = 10) or small proportions of missingness (10%), imputation is advised against.