Target-decoy search strategy for mass spectrometry-based proteomics

Abstract

Accurate and precise methods for estimating incorrect peptide and protein identifications are crucial for effective large-scale proteome analyses by tandem mass spectrometry. The target-decoy search strategy has emerged as a simple, effective tool for generating such estimations. This strategy is based on the premise that obvious, necessarily incorrect "decoy" sequences added to the search space will correspond with incorrect search results that might otherwise be deemed to be correct. With this knowledge, it is possible not only to estimate how many incorrect results are in a final data set but also to use decoy hits to guide the design of filtering criteria that sensitively partition a data set into correct and incorrect identifications.

Fig. 1

Decoy PSMs indicate incorrect target PSMs, depending on the underlying proportion of target and decoy sequences. Under the reversed-decoy model, the proportion of target and decoy peptides considered are approximately equal (5th-ranked, reversed-decoy). Thus, the proportion of decoy PSMs observed in the presence of correct identifications equals the proportion of target PSMs that are incorrect (Top-ranked, reversed-decoy). When the underlying proportion of target and decoy sequences are not equal, as is usually the case with randomly created protein sequence lists, one must first measure this proportion (5th-ranked, random-decoy), and then apply it to the condition containing correct identifications (top-ranked, random-decoy). See ref. for further details

Fig. 2

Venn diagram of basic measurements related to estimated false positive identifications. The total number of identifications are contained within the rectangle. All correct identifications are contained within the white circle. All identifications passing a given set of selection criteria (positive identifications) are contained within the black circle. The overlap between these circles are true positives (TP). False positive identifications (FP) are the remaining positive identifications, and false negative identifications (FN) are the remaining correct identifications that do not meet the selection criteria. True negatives (TN) are the incorrect identifications that are correctly classified as such by the selection criteria. This Venn diagram scheme is elaborated in Fig. 3

Fig. 3

Considering multiple selection criteria enhances accuracy. Selection criteria applied to score distributions (left) determine the form of the Venn diagrams (right). Venn diagram shapes and colors correspond with those in Fig. 2. (a) Distribution of FP and TP hits sorted by an arbitrary score. When no score criteria are applied, all selected correct identifications are denoted in grey circle, and all selected incorrect identifications are denoted in black rectangle. (b) Application of a single score threshold, which excludes most incorrect identifications (lighter region), can yield an acceptable precision rate, but yields sub-optimal sensitivity. (c) Considering two scores allows for greater separation between correct and incorrect identifications. The distribution of incorrect identifications is indicated by the distribution of decoy hits. Application of global criteria that excludes most decoy hits in two score dimensions (lighter region) provides greater sensitivity than one score alone. (d) Designing selection criteria that take into account numerous peptide measurements, such as mass accuracy, charge, enzymatic specificity, and peptides per protein, can yield far greater sensitivity while maintaining acceptable precision