MS-GF+ makes progress towards a universal database search tool for proteomics

Abstract

Mass spectrometry (MS) instruments and experimental protocols are rapidly advancing, but the software tools to analyse tandem mass spectra are lagging behind. We present a database search tool MS-GF+ that is sensitive (it identifies more peptides than most other database search tools) and universal (it works well for diverse types of spectra, different configurations of MS instruments and different experimental protocols). We benchmark MS-GF+ using diverse spectral data sets: (i) spectra of varying fragmentation methods; (ii) spectra of multiple enzyme digests; (iii) spectra of phosphorylated peptides; and (iv) spectra of peptides with unusual fragmentation propensities produced by a novel alpha-lytic protease. For all these data sets, MS-GF+ significantly increases the number of identified peptides compared with commonly used methods for peptide identifications. We emphasize that although MS-GF+ is not specifically designed for any particular experimental set-up, it improves on the performance of tools specifically designed for these applications (for example, specialized tools for phosphoproteomics).

Figure 1

Various spectral types. Spectral types are represented as paths in the graph representing possible choices of the fragment method (Fragmentation), the instrument measuring product ion m/z (Instrument), the protocol used to prepare a sample (Protocol), and the enzyme used to digest proteins (Enzyme). ‘Low’ in Instrument indicates low-resolution instruments (e.g. linear ion-trap), ‘High’ indicates high-resolution instruments (e.g. Orbitrap), and ‘TOF’ indicates time-of-flight instruments. ‘Phosphorylation’ and ‘Ubiquitination’ in Protocol indicate that spectra are generated from phosphopeptides and ubiquitinated peptides, respectively. A path in the graph represents a spectral type. For example, the green path (CID, Low, Phosphorylation, Trypsin) represents low-precision CID spectra of trypsin digests generated from a sample enriched for phosphopeptides. The blue, red, green, and magenta paths represent spectral types of the datasets used in recent studies by Frese et al. [20], Swaney et al. [1], Huttlin et al. [21], and Starita et al. [22], respectively. Different combinations of analysis tools were used for different studies. Frese et al. used an in-house tool for peak filtering, de-isotoping, and charge deconvolution, Mascot for database search, Percolator for re-scoring, and RockerBox [58] for peptide-level FDR control. Swaney et al. used an in-house tool for peak filtering, OMSSA [27] for database search, and an in-house tool for both peptide- and protein-level FDR control. Huttlin et al. used an in-house tool for re-calibrating peak masses, SEQUEST for database search, an in-house tool for re-scoring, and peptide- and protein-level FDR control. Starita et al. used the Trans-Proteomics Pipeline [45] along with SEQUEST for database search. The same datasets were analyzed by MS-GF+ without using any additional tool with scoring parameters trained separately for different spectral types.

Figure 2

Benchmarking MS-GF+ against Mascot+Percolator. Percent increases in the number of identified PSMs for MS-GF+ compared to Mascot+Percolator for all 19 datasets. Each bar represents a spectral dataset of a specified spectral type. For (CID, Low, Standard, Trypsin) and (ETD, Low, Standard, Trypsin), there are two corresponding datasets, one from human and the other from yeast. We distinguish them by adding ‘*’ to the yeast datasets. For the (CID, Low, Phosphorylation, Trypsin) and (CID, Low, Ubiquitination, Trypsin) datasets, the number of phosphorylated and ubiquitinated PSMs were counted instead of the number of all identified PSMs. For the (ETD,Low,Standard,αLP) dataset, Mascot+Percolator identified no PSM.

Figure 3

Comparison of MS-GF+ and other tools for diverse spectral types. The numbers of identified PSMs (a–c) or peptides (d) at 1% FDR are shown. Numbers above bars represent the percentages of increase in the number of identifications for MS-GF+ compared to other tools. (a) Results for the human datasets with varying fragmentations and instruments. MS-GF+, Mascot+Percolator, and Mascot results are shown along with the results in [20]. Percolator greatly increased the number of identifications as compared to Mascot, but MS-GF+ outperformed Mascot+Percolator for all the datasets. (b) Increase in the number of identifications due to the availability of high-precision product ion peaks. For the three human datasets representing HH spectra, MS-GF+, Mascot+Percolator, and Mascot were run using search parameters for HL spectra. The results of these searches (denoted by HL) are compared with the numbers of identifications for the regular searches (denoted by HH). HH searches identified more PSMs than HL searches for every tool and every dataset. The difference was larger for CID and HCD than ETD spectra. (c) Results for the yeast datasets with varying fragmentations and enzymes. MS-GF+ and Mascot+Percolator results are shown. MS-GF+ outperformed Mascot+Percolator for all these datasets. (d) Comparison of MS-GF+ and the results in [1] that used OMSSA along with in-house post-processing tools for the yeast datasets. The numbers of (unique) peptides at the peptide-level 1% are shown. In [1], only the number of identified peptides matched to proteins identified at 1% protein-level FDR was counted while for MS-GF+, the number of identified peptides was counted regardless of their matched proteins.

Figure 3

Comparison of MS-GF+ and other tools for diverse spectral types. The numbers of identified PSMs (a–c) or peptides (d) at 1% FDR are shown. Numbers above bars represent the percentages of increase in the number of identifications for MS-GF+ compared to other tools. (a) Results for the human datasets with varying fragmentations and instruments. MS-GF+, Mascot+Percolator, and Mascot results are shown along with the results in [20]. Percolator greatly increased the number of identifications as compared to Mascot, but MS-GF+ outperformed Mascot+Percolator for all the datasets. (b) Increase in the number of identifications due to the availability of high-precision product ion peaks. For the three human datasets representing HH spectra, MS-GF+, Mascot+Percolator, and Mascot were run using search parameters for HL spectra. The results of these searches (denoted by HL) are compared with the numbers of identifications for the regular searches (denoted by HH). HH searches identified more PSMs than HL searches for every tool and every dataset. The difference was larger for CID and HCD than ETD spectra. (c) Results for the yeast datasets with varying fragmentations and enzymes. MS-GF+ and Mascot+Percolator results are shown. MS-GF+ outperformed Mascot+Percolator for all these datasets. (d) Comparison of MS-GF+ and the results in [1] that used OMSSA along with in-house post-processing tools for the yeast datasets. The numbers of (unique) peptides at the peptide-level 1% are shown. In [1], only the number of identified peptides matched to proteins identified at 1% protein-level FDR was counted while for MS-GF+, the number of identified peptides was counted regardless of their matched proteins.

Figure 3

Comparison of MS-GF+ and other tools for diverse spectral types. The numbers of identified PSMs (a–c) or peptides (d) at 1% FDR are shown. Numbers above bars represent the percentages of increase in the number of identifications for MS-GF+ compared to other tools. (a) Results for the human datasets with varying fragmentations and instruments. MS-GF+, Mascot+Percolator, and Mascot results are shown along with the results in [20]. Percolator greatly increased the number of identifications as compared to Mascot, but MS-GF+ outperformed Mascot+Percolator for all the datasets. (b) Increase in the number of identifications due to the availability of high-precision product ion peaks. For the three human datasets representing HH spectra, MS-GF+, Mascot+Percolator, and Mascot were run using search parameters for HL spectra. The results of these searches (denoted by HL) are compared with the numbers of identifications for the regular searches (denoted by HH). HH searches identified more PSMs than HL searches for every tool and every dataset. The difference was larger for CID and HCD than ETD spectra. (c) Results for the yeast datasets with varying fragmentations and enzymes. MS-GF+ and Mascot+Percolator results are shown. MS-GF+ outperformed Mascot+Percolator for all these datasets. (d) Comparison of MS-GF+ and the results in [1] that used OMSSA along with in-house post-processing tools for the yeast datasets. The numbers of (unique) peptides at the peptide-level 1% are shown. In [1], only the number of identified peptides matched to proteins identified at 1% protein-level FDR was counted while for MS-GF+, the number of identified peptides was counted regardless of their matched proteins.

Figure 3

Comparison of MS-GF+ and other tools for diverse spectral types. The numbers of identified PSMs (a–c) or peptides (d) at 1% FDR are shown. Numbers above bars represent the percentages of increase in the number of identifications for MS-GF+ compared to other tools. (a) Results for the human datasets with varying fragmentations and instruments. MS-GF+, Mascot+Percolator, and Mascot results are shown along with the results in [20]. Percolator greatly increased the number of identifications as compared to Mascot, but MS-GF+ outperformed Mascot+Percolator for all the datasets. (b) Increase in the number of identifications due to the availability of high-precision product ion peaks. For the three human datasets representing HH spectra, MS-GF+, Mascot+Percolator, and Mascot were run using search parameters for HL spectra. The results of these searches (denoted by HL) are compared with the numbers of identifications for the regular searches (denoted by HH). HH searches identified more PSMs than HL searches for every tool and every dataset. The difference was larger for CID and HCD than ETD spectra. (c) Results for the yeast datasets with varying fragmentations and enzymes. MS-GF+ and Mascot+Percolator results are shown. MS-GF+ outperformed Mascot+Percolator for all these datasets. (d) Comparison of MS-GF+ and the results in [1] that used OMSSA along with in-house post-processing tools for the yeast datasets. The numbers of (unique) peptides at the peptide-level 1% are shown. In [1], only the number of identified peptides matched to proteins identified at 1% protein-level FDR was counted while for MS-GF+, the number of identified peptides was counted regardless of their matched proteins.

Figure 4

Constructing a Directed Acyclic Graph (DAG) in the case of two “amino acids” with real masses 2.012 and 2.996. Assume that only singly-charged b-ion with a real of f set 1.008 contributes to the scoring. The spectrum S is converted into S^′ by shifting each peak by 1.008 to the left. Each arrowed line in S^′ represents a pair of peaks separated approximately by 2 Da (blue) or 3 Da (red) that form a duo (solid) or does not form a duo (dashed) for a fragment mass tolerance 0.01 Da. A DAG G is constructed from S^′. The number in the vertex represents its label. The color of the edge represents its label (0 for dashed grey and 1 for solid black).

Figure 5

Illustration of the MS-GF+ Directed Acyclic Graph (DAG) scoring. The peptide ABAA is converted into its Boolean string P = 010010101 and the spectrum S is converted into a labeled DAG G as described in the text. The number in the vertex represents its label. The color of the edge represents its label (0 for grey and 1 for black). The vertex i is colored depending on the peptide character i (white for 0 and black for 1). We also color vertex 0 as black. The procedure to compute Score(P, G) is illustrated. All edges are partitioned into 8 classes depending on s_i,j, p_i, and p_j. For example, there are 5 edges with s_i,j = p_i = p_j = 0.