BioProEV: A Bioinformatics Pipeline for Biologically-Relevant Handling of Missing Values in the Analysis of Extracellular Vesicles by Mass Spectrometry

While proteomics is being increasingly applied to investigate extracellular vesicles (EVs), there remains no consensus on how to address the inevitable missing values within proteomics datasets. Here, we devised a three-step approach that prioritized retaining biologically-relevant information in EV samples using a dataset containing two populations of EVs analysed by liquid chromatography electrospray ionization tandem mass spectrometry (LC-ESI-MS/MS). Firstly, to avoid overfitting, we excluded proteins in which more than half of the values were missing. Next, we statistically tested for low protein abundance in a single population: when the number of potential "missing not at random" (MNAR) values was significantly enriched, these values were replaced with the lowest possible value of "1". Finally, all remaining missing values were then imputed using the Random Forest machine learning algorithm. Our final dataset included 49.9 % of proteins that originally contained at least one missing value, from which 84.7 % were listed in the ExoCarta database, significantly more than would be expected by chance, strongly indicating biologically relevant imputation. To enable other EV researchers to analyse proteomics data in a robust, easy-to-use and peer-reviewed manner, we provide BioProEV, a bioinformatics pipeline to impute missing values with biological relevance in EV datasets.