Optimizing Protein Characterization using Machine Learning-Guided Mass Spectrometry

Abstract

Mass spectrometry-based proteomics excels at high-throughput identification of proteins expressed in complex biological samples. However, the technology struggles to identify low abundance proteins due to large amounts of redundant data acquired for high abundance proteins with little collected for low abundance proteins. To improve the identification sensitivity of these proteins, I designed a machine learning classifier that assesses protein identification confidence on-the-fly, during mass spectrometry analysis. Proteins deemed confidently identified are excluded from further analysis, saving mass spectrometry resources for lower abundance proteins. Simulating data from a HEK293 cell lysate mass spectrometry analysis, our algorithm uses 16.2% - 66.2% fewer mass spectrometry resources with a 2.6% - 39.5% drop in protein identifications. When applied to live mass spectrometry experiments, these saved resources will likely improve the overall protein identification sensitivity of the experiment, particularly for lower abundance proteins, and will therefore provide a better understanding of the cell’s biology.