Statistical and machine learning methods to analyze large-scale mass spectrometry data

Modern biology is faced with vast amounts of data that contain valuable information yet to be extracted. Proteomics, the study of proteins, has repositories with thousands of mass spectrometry experiments. These data gold mines could further our knowledge of proteins as the main actors in cell processes and signaling. Here, we explore methods to extract more information from this data using statistical and machine learning methods. First, we present advances for studies that aggregate hundreds of runs. We introduce MaRaCluster, which clusters mass spectra for large-scale datasets using statistical methods to assess similarity of spectra. It identified up to 40% more peptides than the state-of-the-art method, MS-Cluster. Further, we accommodated large-scale data analysis in Percolator, a popular post-processing tool for mass spectrometry data. This reduced the runtime for a draft human proteome study from a full day to 10 minutes. Second, we clarify and promote the contentious topic of protein false discovery rates (FDRs). Often, studies report lists of proteins but fail to report protein FDRs. We provide a framework to systematically discuss protein FDRs and take away hesitance. We also added protein FDRs to Percolator, opting for the best-peptide approach which proved superior in a benchmark of scalable protein inference methods. Third, we tackle the low sensitivity of protein quantification methods. Current methods lack proper control of error sources and propagation. To remedy this, we developed Triqler, which controls the protein quantification FDR through a Bayesian framework. We also introduce MaRaQuant, which proposes a quantification-first approach that applies clustering prior to identification. This reduced the number of spectra to be searched and allowed us to spot unidentified analytes of interest. Combining these tools outperformed the state-of-the-art method, MaxQuant/Perseus, and found enriched functional terms for datasets that had none before.QC 20181001