False discovery rates of protein identifications: a strike against the two-peptide rule

Abstract

Most proteomics studies attempt to maximize the number of peptide identifications and subsequently infer proteins containing two or more peptides as reliable protein identifications. In this study, we evaluate the effect of this "two-peptide" rule on protein identifications, using multiple search tools and data sets. Contrary to the intuition, the "two-peptide" rule reduces the number of protein identifications in the target database more significantly than in the decoy database and results in increased false discovery rates, compared to the case when single-hit proteins are not discarded. We therefore recommend that the "two-peptide" rule should be abandoned, and instead, protein identifications should be subject to the estimation of error rates, as is the case with peptide identifications. We further extend the generating function approach (originally proposed for evaluating matches between a peptide and a single spectrum) to evaluating matches between a protein and an entire spectral data set.

Figure 1

Identification of peptides in the Shewanella data set using different approaches and scoring functions. Each point in the curves is generated by varying the scoring threshold and computing the number of hits in the target and the decoy database exceeding the threshold.

Figure 2

(a) Identification of proteins in the human data set using different approaches and scoring functions. (b) Similar plot as in panel (a) for Shewanella data set.

Figure 3

(a) Protein identification in the human data set using X!Tandem search results with different scoring approaches at the protein level. (b) Similar plot as in panel (a) for an arbitrarily selected subset of Shewanella data set containing 1.25 million spectra.

Figure 4

Identification of proteins, using the unique peptides only (peptides that are not shared between multiple proteins), in the human data set using InsPecT search results with different approaches.

Figure 5

(a) Identification of proteins in the human data set using MS-GF scores, without and with length correction. (b) Similar plot as in panel (a) for Shewanella data set.