A Machine Learning Approach to Decipher Protein-Protein Interactions in Human Plasma to Facilitate the Characterization of Metabolic Pathways

Abstract

Immunoprecipitation coupled to mass spectrometry (IP-MS) methods are often used to identify protein-protein interactions (PPIs) in biological samples. While these approaches are prone to false-positive identifications through contamination and antibody non-specific binding, their results can be filtered by combining the use of negative controls and computational modelling. However, such filtering does not effectively detect false-positive interactions when IP-MS is performed on human plasma samples, given a higher propensity for non-specific interactions. Therein, proteins cannot be overexpressed or inhibited, and existing modelling algorithms are not adapted for execution without such controls. Hence, we introduce MAGPIE, a novel machine learning-based approach for identifying PPIs in human plasma using IP-MS, which leverages negative controls that include antibodies targeting proteins not known to be present in human plasma. Unsupervised learning algorithms are first applied to label-free MS quantification data to identify a set of high-quality negative controls that can be used for false- positive interaction modelling. MAGPIE then uses a logistic regression classifier to assess the reliability of PPIs detected in IP-MS experiments using antibodies targeting known plasma proteins. When applied to five IP-MS experiments, our algorithm identified 68 PPIs with an FDR of 20%. MAGPIE significantly outperformed a state-of-the-art PPI discovery tool, detecting three times more interactions at half the FDR. PPIs identified by MAGPIE are further supported by known or predicted interactions in the STRING PPI repository. Finally, our approach provides an unprecedented ability to detect human plasma PPIs, enabling a better understanding of biological processes in plasma.