Imputation of label-free quantitative mass spectrometry-based proteomics data using self-supervised deep learning

Abstract

Imputation techniques provide means to replace missing measurements with a value and are used in almost all downstream analysis of mass spectrometry (MS) based proteomics data using label-free quantification (LFQ). Some methods only impute assuming the limit of detection (LOD) was not passed and therefore impute missing values with too low or too high intensities, potentially leading to biased results in downstream statistical analysis. Here we test how self supervised deep learning models can impute missing values in the context of LFQ at different levels: precursors, aggregated peptides or protein groups. We demonstrate how collaborative filtering, denoising autoencoders, and variational autoencoders can be used to reconstruct missing values and can make more relevant features available for downstream analysis compared to current approaches. Additionally, we show that deep learning approaches can model data in its entirety for imputation and offer an approach for controlled evaluation of imputation approaches. We applied our method, proteomics imputation modeling mass spectrometry (PIMMS), to an alcohol-related liver disease (ALD) cohort with blood plasma proteomics data available for 358 individuals. We identified 49 additional proteins (+52.7%) that are significantly differentially abundant across disease stages compared to traditional methods and found that some of these were predictive of ALD progression in machine learning models. We, therefore, suggest the use of deep learning approaches for imputing missing values in MS-based proteomics and provide workflows for these.

Evaluating proteomics imputation methods with improved criteria

“…Deep neural networks have proven highly generalizable in other contexts. Recent "deep" impute methods may be a step in the right direction [20], though much work remains to be done. In the future, data-driven imputation methods will likely be broadly adopted as part of general signal processing workflows in proteomics.…”

“…On the basis of the resulting citation counts (Figure 1), we selected four of the most popular imputation methods: k-nearest neighbor (kNN) [3], MissForest [11], Gaussian sampling [9], and low value replacement (Figure 1). We also include a non-negative matrix factorization (NMF) imputation method, which has recently been proposed for proteomics [18][19][20]. By focusing only on the most commonly used imputation methods, our aim is to provide a practical comparison that will be beneficial to experimental proteomicists.…”

Mass spectrometry-based proteomics data from thousands of HeLa control samples

“…We hypothesized that this data could be used to explore self-supervised deep learning models for imputing MS-based proteomics data on the basis of ionised peptides generated from denatured proteins. Therefore in a companion paper we adapted three deep learning architectures for imputation of proteomics data, de ned a work ow for model comparison and analysed the effect of imputation using different methods on a downstream analysis task 3 .…”

“…From the MaxQuant search of each le, we dumped the tab separated les to PRIDE along the associated raw le. In a companion paper we used the "evidence.txt" for precursor quanti cations, "peptides.txt" for aggregated peptides and "proteinGroups.txt" for protein groups referencing them by their gene group 3 .…”

Evaluating Proteomics Imputation Methods with Improved Criteria

“…On the basis of the resulting citation counts (Figure ), we selected four of the most popular imputation methods: k -nearest neighbor (kNN), MissForest, Gaussian sampling, and low value replacement. We also include a non-negative matrix factorization (NMF) imputation method, which has recently been proposed for proteomics. − By focusing on only the most commonly used imputation methods, our aim is to provide a practical comparison that will be beneficial to experimental proteomicists. For this reason, seldom used R packages (e.g., imp4p, impSeqRob, and QRLIC) have been omitted from our analysis.…”