Improved Protein Identification in Shotgun Proteomics with a Group-Level Extension of the LPGF Model.

Abstract

Shotgun proteomics relies on robust statistical methods to increase the confidence of the protein identifications. Previously, we introduced LPGF, a protein probability model based on unique peptides. In this study, we extend LPGF to protein groups that share peptides by considering only peptides that are unique to each group. This strategy preserves the principle of unique evidence while enabling the probability estimation at the protein-group level. Applying our method to three tissues from the Human Proteome Map (HPM), we evaluated the gain in protein identifications obtained with group-unique peptides compared to those obtained using only protein-unique peptides and different scores. The accuracy of false discovery rate (FDR) estimation was further validated using a standard data set for protein inference based on human Protein Epitope Signature Tags (PrESTs), as well as through an entrapment experiment using the substantially larger HPM data set. To facilitate adoption, we developed an R package (b10prot) that integrates the full workflow, including protein-level and group-level LPGF score computation and protein grouping via an R port of our previous tool PAnalyzer. Our results demonstrate that extending LPGF to protein groups increases sensitivity while maintaining robust FDR control, thus providing the proteomics community with a practical framework for more comprehensive protein identification.