Generalized peakgroup scoring boosts identification rates and accuracy in mass spectrometry based discovery proteomics

The statistical validation of peptide and protein identifications in mass spectrometry proteomics is a critical step in the analytical workflow. This is particularly important in discovery experiments to ensure only confident identifications are accumulated for downstream analysis and biomarker consideration. However, the inherent nature of discovery proteomics experiments leads to scenarios where the search space will inflate substantially due to the increased number of potential proteins that are being queried in each sample. In these cases, issues will begin to arise when the machine learning algorithms that are trained on an experiment specific basis cannot accurately distinguish between correct and incorrect identifications and will struggle to accurately control the false discovery rate. Here, we propose an alternative validation algorithm trained on a curated external data set of 2.8 million extracted peakgroups that leverages advanced machine learning techniques to create a generalizable peakgroup scoring (GPS) method for data independent acquisition (DIA) mass spectrometry. By breaking the reliance on the experimental data at hand and instead training on a curated external dataset, GPS can confidently control the false discovery rate while increasing the number of identifications and providing more accurate quantification in different search space scenarios. To first test the performance of GPS in a standard experimental environment and to provide a benchmark against other methods, a novel spike-in data set with known varying concentrations was analyzed. When compared to existing methods GPS increased the nunmber of identifications by 5-18% and was able to provide more accurate quantification by increasing the number of ratio validated identifications by 24-74%. To evaluate GPS in a larger search space, a novel data set of 141 blood plasma samples from patients developing acute kidney injury after sepsis was searched with a human tissue spectral library (10000+ proteins). Using GPS, we were able to provide a 207-377% increase in the number of candidate differentially abundant proteins compared to the existing methods while maintaining competitive numbers of global identifications. Finally, using an optimized human tissue library and workflow we were able to identify 1205 proteins from the 141 plasma samples and increase the number of candidate differentially abundant proteins by 70.87%. With the addition of machine learning aided differential expression, we were able to identify potential new biomarkers for stratifying subphenotypes of acute kidney injury in sepsis. These findings suggest that by using a generalized model such as GPS in tandem with a massive scale spectral library it is possible to expand the boundaries of discovery experiments in DIA proteomics. GPS is open source and freely available on github at (https://github.com/InfectionMedicineProteomics/gscore).