The protein inference problem, i.e., determining which proteins are present in a biological sample, is key to understanding the roles of proteins and, more broadly, many biological processes. Protein identification is typically achieved by first cleaving proteins into smaller sequences called peptides. Peptides are then identified using tandem mass spectrometry, a process that produces mass spectra, and in which peptide identification consists of associating, via dedicated tools, a mass spectrum to a peptide sequence. Protein inference consists of identifying, from a list of identified peptides, the proteins that most likely produced them, and were therefore present in the original sample. Usually, peptide identification and protein inference are two separate steps, which are sequentially achieved. However, by proceeding in such a way, a significant amount of potentially useful information contained in the spectra may be discarded in the second step. Moreover, AI-based tools can now predict the likelihood of a peptide’s identification when its parent protein is present in the sample. In this paper, we present the Global Spectrum Interpretation (GSI) model, a protein inference model that integrates all this information to produce more accurate protein identifications. We show that GSI is NP-hard and provide a Mixed Integer Linear Program (MILP) formulation for it. This MILP is then benchmarked against state-of-the-art protein inference models on several datasets. Our results show that GSI’s promising and original approach achieves performance comparable to current models and outperforms other widely used ones, while being more explainable.

BACKGROUND: In proteomics, the interpretation of mass spectra representing peptides carrying multiple complex modifications remains challenging, as it is difficult to strike a balance between reasonable execution time, a limited number of false positives, and a huge search space allowing any number of modifications without a priori. The scientific community needs new developments in this area to aid in the discovery of novel post-translational modifications that may play important roles in disease. RESULTS: To make progress on this issue, we implemented SpecGlobX (SpecGlob eXTended to eXperimental spectra), a standalone Java application that quickly determines the best spectral alignments of a (possibly very large) list of Peptide-to-Spectrum Matches (PSMs) provided by any open modification search method, or generated by the user. As input, SpecGlobX reads a file containing spectra in MGF or mzML format and a semicolon-delimited spreadsheet describing the PSMs. SpecGlobX returns the best alignment for each PSM as output, splitting the mass difference between the spectrum and the peptide into one or more shifts while considering the possibility of non-aligned masses (a phenomenon resulting from many situations including neutral losses). SpecGlobX is fast, able to align one million PSMs in about 1.5 min on a standard desktop. Firstly, we remind the foundations of the algorithm and detail how we adapted SpecGlob (the method we previously developed following the same aim, but limited to the interpretation of perfect simulated spectra) to the interpretation of imperfect experimental spectra. Then, we highlight the interest of SpecGlobX as a complementary tool downstream to three open modification search methods on a large simulated spectra dataset. Finally, we ran SpecGlobX on a proteome-wide dataset downloaded from PRIDE to demonstrate that SpecGlobX functions just as well on simulated and experimental spectra. We then carefully analyzed a limited set of interpretations. CONCLUSIONS: SpecGlobX is helpful as a decision support tool, providing keys to interpret peptides carrying complex modifications still poorly considered by current open modification search software. Better alignment of PSMs enhances confidence in the identification of spectra provided by open modification search methods and should improve the interpretation rate of spectra.