A Bayesian Approach to Protein Inference Problem in Shotgun Proteomics

Abstract

The protein inference problem represents a major challenge in shotgun proteomics. Here we describe a novel Bayesian approach to address this challenge by incorporating the predicted peptide detectabilities as the prior probabilities of peptide identification. We propose a rigorious probabilistic model for protein inference, and provide practical algoritmic solutions to this problem. We used a complex synthetic protein mixture to test our method and obtained promising results.

SIR: Deterministic protein inference from peptides assigned to MS data

“…The strict division in ProteinProphet of the result into correct and incorrect identifications has fostered the development of other statistical models that give low but non-zero probability values for proteins that do not have conclusive evidence of their presence in the sample [6,7]. Bayesian approaches that attempt to incorporate peptide detect ability to improve protein inference have also been proposed [8]. Clearly the final result from minimal list approach depends on the underlining assumption in the statistical model.…”

“…Previous protein inference algorithms can be divided into two main groups those that provide a minimal list that explains the observed data [1,[6][7][8] and those that aim at representing the full peptide and protein set organized into groups to provide a simplified view of the result that can be further explored [9][10][11]. Basically all protein inference algorithms discussed in the literature only consider the degeneration at the peptide-protein level [1,[9][10][11], but as it can be seen from Fig.…”

Identifying Abundant Immunotherapy and Other Targets in Solid Tumors

“…Proteins are declared identified using either a formal protein-inference algorithm 18,21 or by thoughtful heuristics which commonly requires two confident peptides per protein. Almost all proteins have peptides that map to only a small number of its peptides, and some peptides will match to multiple different proteins, especially those that derive from members of homologous gene families.…”

Solution to Statistical Challenges in Proteomics Is More Statistics, Not Less

“…In proteomics, a variety of methods have been proposed for estimating the protein identification FDR: by using protein-level targetdecoy search 6 , by adjusting protein-level target-decoy search results to account for the random 2 scattering of false-positive peptides 7 , by analyzing decoy-free peptide search results and using unique peptides to assign protein identification FDRs 8 , or by using generative parametric models to compute probabilities that each protein is present given the spectral evidence of shared and unique peptides [9][10][11][12][13] . Creating empirical generative models in proteomics-i.e., modeling the process by which proteins become peptides and peptides become spectra, and spectra are sampled-is not very difficult, and so these approaches are primarily concerned with balancing accuracy of the model with the ability to perform efficient Bayesian inference.…”