Mass spectrometry offers numerous ways to analyze the composition, function, and interactions of complex proteomes. Unfortunately, it suffers from the variation introduced by technological artifacts, which reduces the reproducibility and reliability of the results. In order to detect and minimize deviations from optimal performance, researchers monitor standard mixtures of proteins or peptides and various associated metrics using statistical summaries. Although this approach to monitoring multiple analytes and metrics is often beneficial, most of these methods do not scale well to multivariate situations. In this paper, we present MSstatsQC-ML, a machine learning approach to quality control that optimizes decision-making from standard mixtures with many analytes and metrics. MSstatsQC-ML combines machine learning classifiers with experimental design strategies to simulate possible suboptimal MS runs that have not yet been observed. For training the classifiers, the proposed approach incorporates informative features from metrics. Analysis of longitudinal values of each feature allows us to interpret the root causes of suboptimal performance and helps to design preventive actions. In evaluations on quality control data from discovery and targeted proteomic experiments, MSstatsQC-ML reduced error rates of detecting suboptimal performance and outperformed traditional approaches. MSstatsQC-ML is available as part of the open-source MSstatsQC R/Bioconductor package.
Doğu et al. (Mon,) studied this question.