Extending the coverage of spectral libraries: a neighbor-based approach to predicting intensities of peptide fragmentation spectra

Abstract

Searching spectral libraries in tandem mass spectrometry (MS/MS) is an important new approach to improving the quality of peptide and protein identification. The idea relies on the observation that ion intensities in an MS/MS spectrum of a given peptide are generally reproducible across experiments, and thus, matching between spectra from an experiment and the spectra of previously identified peptides stored in a spectral library can lead to better peptide identification compared to the traditional database search. However, the use of libraries is greatly limited by their coverage of peptide sequences: even for well-studied organisms a large fraction of peptides have not been previously identified. To address this issue, we propose to expand spectral libraries by predicting the MS/MS spectra of peptides based on the spectra of peptides with similar sequences. We first demonstrate that the intensity patterns of dominant fragment ions between similar peptides tend to be similar. In accordance with this observation, we develop a neighbor-based approach which first selects peptides that are likely to have spectra similar to the target peptide and then combines their spectra using a weighted K-nearest neighbor method to accurately predict fragment ion intensities corresponding to the target peptide. This approach has the potential to predict spectra for every peptide in the proteome. When rigorous quality criteria are applied, we estimate that the method increases the coverage of spectral libraries available from the National Institute of Standards and Technology by 20–60%, although the values vary with peptide length and charge state. We find that the overall best search performance is achieved when spectral libraries are supplemented by the high quality predicted spectra.

1. Introduction

Liquid chromatography (LC) coupled with tandem mass spectrometry (MS/MS) is a widely used platform for high-throughput identification and quantification of proteins in biological samples [1, 2]. In addition to experimental steps in the pipeline, computational and statistical procedures play important roles in determining the content of the complex proteome sample. However, even with the best analytical platforms and modern software, only about 10–30% of spectra are identified in a typical experiment [3, 4]. In most situations this results in a large fraction of identified proteins (more than 50%) being covered by a single identified peptide, which weakens the confidence of protein identification [5–7].

There are several reasons for low identification coverage, involving biological, analytical and also computational factors. Biological samples may contain genomic variants, contaminants or enzymes with incompletely understood specificity and cleavage patterns. They may also contain a number of post-translationally modified proteins, some with yet unknown modifications. Analytical platform variations, e.g. the differences in sample preparation or the inherently stochastic nature of peptide elution, ion current variation in ionization sources, or the ability of a peptide to ionize and fragment well, are another notable source of variation. Analytical techniques can also introduce chemical modifications of several residues, e.g. oxidation of methionine and histidine, or cysteine alkylation. Similarly, a large number of peptides may be truncated before even reaching the mass spectrometer, resulting in the spectra of unexpectedly short peptides [8]. Most of these sources are aggregated via the concept of peptide detectability [9, 10] or the proteotypic property of peptides [11]. Finally, development of computational approaches for accurate identification of tandem mass spectra, estimating the false discovery rates of peptide-spectrum matches [12, 13], or inferring the identity and quantity of proteins [14], is still an open challenge. Among these, improving the quality and confidence of peptide-spectrum matching has the potential to directly impact biological discoveries.

There are two main strategies used for assigning peptide sequences to experimental MS/MS spectra. Depending on whether the reference protein sequences of the organism(s) under study are already known, these approaches are referred to as database searching or de novo sequencing [15]. Database searching strategies match experimental spectra with theoretical spectra corresponding to the peptides available in the database (for sequenced organisms), whereas de novo algorithms attempt to infer peptide sequences directly from experimental spectra; implicitly assuming that the search space contains all possible peptides within the mass/charge tolerance of the experimental spectra. In database search, it is expected that allowing for different fragment ion intensities in a theoretical spectrum will result in improved search outcomes. These intensities may be determined ad-hoc, as in SEQUEST [16], or using computational approaches that predict experimental spectra directly from peptide sequence [17–20]. One recently introduced approach, extending the ideas of small molecule identification, is that of spectral libraries [21–25], where the spectra of all identified peptides from previous experiments are aggregated into a library of consensus spectra corresponding to the identified peptides. Several approaches indicated that searching spectral libraries has the potential to identify more peptide sequences given the same database size. These studies resulted in hybrid algorithms that utilize spectral libraries for previously identified peptides and conventional database searches on the remaining spectra [26–28]. Recently, new algorithms have been developed to exploit differential ion intensities in matching experimental and theoretical spectra [29].

Since spectral libraries cover a relatively small fraction of peptide sequences, an alternative solution has been to in silico predict the fragment spectra. Several algorithms have been developed ranging from the kinetic models of peptide fragmentation [17, 18] to fully data-driven algorithms [19, 20, 30–32]. Recently, Li et al. investigated the variability of tandem mass spectra of identical peptides and showed that spectral reproducibility in the same experiment (i.e. in the same run) is significantly higher than over different experiments (i.e. over different runs and different labs), even for the same analytical procedure [20]. They also concluded that algorithms such as MassAnalyzer [17, 18] and PeptideART [20] provide predictions that consistently exceed 70% of the correlation level achieved by the spectral reproducibility of identified peptides. Venable and Yates have also shown that the variance of peptide-spectrum match scores is dependent on both the peptide sequence and its quantity [33].

In this work, we study ion intensity patterns in CID spectra from similar peptides. We show that the intensities of fragment ions can be accurately predicted using a weighted average of the spectra from peptides of the same length and similar (neighboring) sequences whose consensus spectra were already collected in spectral libraries. We then compare this neighbor-based approach with both database search and spectral library methods that rely on predicted intensities of fragment ions [17, 18] and demonstrate that our approach provides an improvement in peptide identification.

2. Methods

2.1. Tandem mass spectrometry data

We used peptide tandem mass spectral libraries provided by the National Institute of Standards and Technology (NIST; http://peptide.nist.gov/). These libraries are comprised of consensus spectra that were obtained by processing individual spectra from multiple samples and experiments. NIST currently provides spectral libraries for nine species (H. sapiens, M. musculus, D. melanogaster, C. elegans, S. cerevesiae, E. coli, D. radiodurans, M. smegmatis and R. norvegicus) from ion trap platforms. The matched spectra were annotated by labeling conventional product ions, along with the immonium ions, internal ions and common neutral loss ions. In cases of ambiguous peak assignments (NIST allows up to two different ion types per peak), we picked the top annotation. The library also contains post-translationally modified peptides; however, in this work only unmodified peptides were considered. Peptides with cysteine carbamidomethylation were excluded as only a subset of cysteine-containing peptides were modified. Basic statistics of the NIST Library are summarized in Table 1.

Table 1.

NIST spectral library statistics.

Species	Number of spectra	Spectra +2	Spectra +3
H. sapiens	159,664	113,323	46,341
M. musculus	75,856	56,463	19,393
D. melanogaster	55,009	39,458	15,551
C. elegans	48,432	34,029	14,403
S. cerevesiae	41,533	31,165	10,368
E. coli	31,953	23,197	8,756
D. radiodurans	5,106	3,988	1,118
M. smegamatis	3,107	1,908	1,199
R. norvegicus	12,582	8,839	3,743
	Total spectra (unique peptides)	312,370 (283,122)	120,872 (112,564)

We converted the original consensus spectra stored in the .msp files into vectors of annotated ion intensities, referred to here as standard spectra. Given any peptide of length l and a particular charge state, a standard spectrum corresponds to a fixed-length vector representation of the intensities of the particular types of fragment ions. A standard spectrum for a doubly charged precursor ion contains the peaks of the following product ions: b, y, b–18, y–18, b–17, y–17, b–35, y–35, b+18, a, b⁺⁺, [b–18]⁺⁺, [b–17]⁺⁺, [b+18]⁺⁺, y⁺⁺, [y–18]⁺⁺, [y–17]⁺⁺, b–43, b–44, b–45, b–46, [b–43]⁺⁺, [b–44]⁺⁺, [b–45]⁺⁺, [b–46]⁺⁺, y–43, y–44, y–45, y–46, [y–43]⁺⁺, [y–44]⁺⁺, [y–45]⁺⁺, [y–46]⁺⁺ for each fragmentation site and prec–18, prec–17, prec–35, prec–36, prec–45, prec–46 for the neutral loss ions derived from the precursor ion. Ions that were absent from the original spectra were assigned zero intensity in the standard spectrum. For a triply charged precursor, fragment ions b⁺⁺⁺, y⁺⁺⁺, [b–18]⁺⁺⁺, [y–18]⁺⁺⁺, [b–17]⁺⁺⁺, [b–43]⁺⁺⁺, [b–44]⁺⁺⁺, [b–45]⁺⁺⁺, [b–46]⁺⁺⁺ were also considered. Therefore, for charge +2, the dimension of the standard spectrum vector was 33·(l – 1) + 6; for charge +3, the dimension of the vector was 42·(l –1) + 6, where l is the peptide length. If the same peptide was found in multiple species/libraries, the consensus spectrum with the most peaks was used. In this work, l ∈ {7, 8, …, 20} for +2 precursors and l ∈ {12, 13, …, 25} for +3 precursors.

Mosquito data set

The Aedes albopictus C6/36 cell line was cultured under appropriate conditions (1xMEM supplemented with 10% fetal bovine serum, non-essential amino acids, L-glutamine and antibiotic/antimycotic solution in the presence of 5% CO₂) prior to subcellular fractionation into nuclear, cytoplasmic and membranous fractions via mechanical disruption and centrifugation. Membrane fractions, normalized to cell number, were resolved using SDS-PAGE preceding gel excision. Gel slices were treated with Solution I (25mM ammonium bicarbonate prepared in 50% acetonitrile) prior to desiccation and rehydration with Solution II (25mM ammonium bicarbonate) and in gel trypsin digestion overnight at 37°C. Peptide fragments were eluted using Solution III (50% acetonitrile and 5% formic acid) via sonication prior to sample concentration via vacuum drying.

Digested samples were diluted with 20 microliters of LC solvent A (98% water, 2% acetonitrile, 0.1% formic acid). Four microliters of each sample was analyzed by nanoLC-MS/MS on an Eksigent nanoLC-2D coupled to a Thermo LTQ-orbitrap XL. A 60-minute gradient from 95% solvent A to 60% solvent B (acetonitrile with 0.1% formic acid) was used that provided separation of the peptides. The mass spectrometer was set up to capture one MS scan followed by MS/MS spectra for the top five precursor ions. Dynamic exclusion was employed such that if the same precursor m/z is selected twice within 15 seconds it is excluded from selection for 30 seconds. Overall, the data set contained 85,016 doubly and 67,568 triply charged MS/MS spectra.

2.2. Problem formulation and classification model

Given a spectral library = {(p_i, s_i)}, where pairs (p_i, s_i) represent peptides with their standard consensus spectra (peptide-spectrum pairs; PSPs), our goal is to predict the spectrum s that corresponds to any previously unseen peptide p. We seek to express the predicted spectrum ŝ of p as a weighted average of a subset of spectra from that are expected to be most similar to s. More specifically, the prediction is made using a weighted K-nearest neighbor formulation as

$$\widehat{s}={\sum }_{k=1}^K{w}_k·s_k,$$ (1)

where K is a positive integer, {w_k} is a set of weights, and {s_k} is a set of K spectra in that are most similar to s. Because s is unknown, it is necessary to find {s_k} and determine weights {w_k} based on the sequence of peptide p and a set of peptides from that have the same length as p.

To determine the parameters from Eq. (1), we first define a measure of similarity between standard spectra. The spectral similarity ρ between spectra s_i and s_j corresponding to some equal-length peptides is measured using the cosine function. That is,

$$\rho (s_i,s_j)=\frac{\langle s_i,s_j\rangle }{\Vert s_i\Vert ·\Vert s_j\Vert },$$

where 〈 u,v 〉 is a dot product between two vectors and $\Vert u\Vert =\sqrt{\langle u,u\rangle }$ is a Euclidean norm of vector u. Because all elements of a standard spectrum are non-negative, it holds that 0 ≤ ρ(s_i,s _j ) ≤ 1.

Generally, each weight w_k in Eq. (1) should be proportional to ρ(s,s_k ). To achieve this, we constructed a classification model to approximate the spectral similarity function from peptide sequences only. This procedure enabled us to select the K spectra that were expected to be most similar to s as well as the set of weights {w_k}. Formally, given two equal-length peptides p_i and p_j, a classification model was trained to provide a similarity score ρ( p_i, p _j ) between spectra s_i and s_j corresponding to p_i and p_j, respectively.

The scoring function σ was learned using a training set {(x_i, y_i )}, where x_i was a vector encoding a pair of peptide sequences $p_i^1$ and $p_i^2$. The class label was set to y = +1 if $\rho (s_i^1,s_i^2)\ge {\rho }_{\mathit{high}}$ and y_i = −1 if $\rho (s_i^1,s_i^2)<{\rho }_{\mathit{low}}$. That is, the positive set and negative set were comprised of pairs of peptides with very similar and very dissimilar spectra. Setting ρ_high = ρ_low would utilize all data points (pairs of peptides of the same length). Vector x_i was constructed using peptide sequences only. Considering that both peptides were of length l, another string of length l could be constructed over an extended alphabet corresponding to the pairs of amino acid symbols that were found at each position (there are 210 such symbols, e.g. AA, AC, AD, etc.). Vector x_i represented a binary encoding of this string as well as the Hamming distance (normalized by l) between $p_i^1$ and $p_i^2$; it results in a sparse vector of dimension 210·l +1 (except for the normalized Hamming distance, only l elements in each data point x_i were set to 1, while the remaining elements were 0). Note that the order of the two amino acids is not relevant; that is, AD and DA correspond to the same symbol in the extended alphabet.

We trained Support Vector Machines (SVMs) using the SVM^light software [34]. We used default SVM^light parameters on a balanced data set of positive and negative data points (note that a separate classifier was trained for each peptide length l). To minimize the chance of overfitting, we used the linear kernel function. The output of a trained SVM was further mapped to a 0–1 interval using the sigmoid function [35]. Thus, the similarity function ρ was approximated using the soft outputs of a classification model.

The weight w_k associated with s_k was determined by the rank r_k of the SVM score σ (p, p_k ) among the top K scoring PSPs (those with the highest scores σ). Here we used an exponentially decaying function

$$w_k=e^{1-\sqrt{r_k}},$$

where r_k ∈{1,2,…K}. We also define a confidence score reflecting the reliability of the prediction ŝ as the average of the SVM scores of the nearest neighbors of peptide p. That is,

$$c(p)=\frac{1}{K}{\sum }_{k=1}^K\sigma (p,p_k),$$ (2)

where {p_k} is a set of K peptides in with highest SVM scores σ (p, p_k ) relative to the target peptide. This confidence score is expected to be higher for more similar peptides and was used to determine the quality of the predicted spectrum.

2.3. Performance evaluation

The performance of the classification model and spectrum prediction was evaluated using 10-fold cross-validation. Specifically, the set of PSPs = {(p_i, s_i)} was split into ten disjoint partitions. Then, in each step of the process the accuracy of the spectrum prediction was evaluated on PSPs in one (test) partition using the scoring function trained on pairs of PSPs from the other nine partitions (training). However, the neighbors for constructing a spectrum were selected from the entire spectral library. The classifier performance was estimated using the area under the ROC curve. On the other hand, the performance of the spectrum prediction was evaluated using the spectral similarity between the observed spectrum s and the predicted spectrumŝ, i.e. ρ(s,ŝ), as well as spectral library search.

To assess the performance of the predicted spectra in an actual proteomics search, we created two duplicate NIST libraries where all peptide sequences were kept but the original consensus spectra were replaced by the predicted spectra. The first duplicate library (NIST^KNN) was constructed using the weighted K-nearest neighbor approach using the training and test sets (to avoid overfitting) from the 10-fold cross-validation described above. The second duplicate library (NIST^MA) was generated using MassAnalyzer [17, 18] with the default parameters, and was used to evaluate whether the neighbor-based approach provides advantages compared to the kinetic model for prediction of peak intensities. All libraries were separated by precursor charge and contained 283,122 PSPs for charge +2 and 112,564 for charge +3. Spectral searches were then performed on a human cell line data set [3] to evaluate the performance of NIST, NIST^KNN, and NIST^MA libraries. For this purpose, we used the SpectraST software [24] with the default parameters.

For each library, an equal-size decoy library was generated using SpectraST’s create mode, which first annotates each peak in each peptide-spectrum pair, and then shuffles the corresponding peptide sequence and repositions the peaks according to the shuffled sequence [36]. The query spectra were searched against a combined target/decoy library. The false discovery rate (FDR) was estimated using the decoy count method; that is,

$$\mathit{FDR}=\frac{2\times FP}{TP+FP},$$

where FP is the number of hits to decoy and TP is the number of hits to the target library.

As an additional comparison, we performed a sequence database search using InsPecT [37] on the same query spectra. Two sequence databases in the FASTA format were created for charge +2 (length 7 to 20) and +3 (length 12 to 25) separately, in which each peptide sequence from the corresponding target spectral library was represented as an entry. A corresponding equal-size decoy database was created by shuffling each peptide in the target database, but keeping K, R and P in their original positions to maintain the number of tryptic termini. To ensure each peptide sequence was considered as a single candidate we set the protease option to trypsin and RequireTermini option to 2. Default parameters were used except that the precursor mass tolerance (PMTolerance) was set to 3 and 4.5 Da for charges +2 and +3, respectively, to match SpectraST’s 1.5 Da/e precursor m/z tolerance. Note that we removed all peptides with missed cleavage sites from all libraries/databases to ensure that the search space between spectral and sequence search were equal. This resulted in 223,392 doubly and 62,600 triply charged spectra in the final data set.

3. Results

3.1. Peptides with similar sequence have similar standard spectra

Because fragmentation of peptides is highly dependent on amino acid sequence [38], peptides with similar sequences can be expected to have similar standard spectra. To test this hypothesis we examined all pairs of equal-length peptides and calculated the spectral similarity between their standard spectra as a function of the number of amino acid mismatches (Hamming distance) between peptides. In Figure 1A we show that the average spectral similarity monotonically decreases with the Hamming distance between peptides, suggesting that sequence similarity is indeed predictive of spectral similarity. This result also justifies the inclusion of multiple sequence neighbors in the candidate set in increasing order of the Hamming distance between two peptides (similarity of spectra for peptides with one mismatch has been recently studied by Hu et al. [39]). Figure 1B shows that values of spectral similarity follow an extreme value-like distribution, with the majority of peptide pairs having moderately low (0.3 to 0.4) similarity values. We note that Figure 1 corresponds to the doubly charged peptides of length 20 (8,934 peptides); however, similar trends were observed for all other lengths and charge states (Figure S1; Supplementary Materials).

Figure 1.

The spectral similarity of standard spectra corresponding to all pairs of doubly charged peptides of length 20. (A) Spectral similarity between standard spectra, averaged over pairs with Hamming distance d, where d ∈ {1, 2, …, 20}; (B) Histogram of all spectral similarity values.

3.2. Evaluation of the spectral similarity scoring function

As illustrated in Figure 1, simple sequence-based features such as the Hamming distance are predictive of spectral similarity between peptides. Other studies have also shown that the frequency of a peptide bond’s fragmentation in CID depends on the identity of amino acids N- or C-terminal to it [30, 40]. A standard example is the preference of proline (P) on the C-terminal side of the fragmentation site, which suggests that the presence of proline induces an intense peak corresponding to the peptide bond fragmentation N-terminal from the proline. Consequently, a strong peak may be lost due to a P-to-non-P substitution. Another observation, illustrated in Figure S2 (Supplementary Materials), is that the most influential amino acids on the similarity between two spectra are generally located in the middle of the sequence for longer peptides, and towards N-terminus for shorter peptides. This may be a consequence of the fact that the ions of particular lengths account for most of the overall intensity (Figure S3; Supplementary Materials).

Although these relatively simple empirical rules offer clues as to how the fragmentation spectrum is influenced by the composition of amino acids and their location along the peptide sequence, they are difficult to combine in the presence of multiple changes. Rather than manually selecting features that potentially are predictive of spectral similarity (e.g. the number of substitutions, the presence/absence of a P-to-non-P substitution), a similarity function was learned. The sparse binary data representation was selected to incorporate both amino acid differences and their locations in the pairs of peptides.

We assessed the performance of the scoring function and spectrum prediction using a cross-validation approach as described in Section 2.3. The parameters ρ_high and ρ_low were set so that each model used top 5% and bottom 5% of the available data. The areas under the ROC curves (AUCs) for all models are shown in Table S1 (Supplementary Materials). Various properties of the predictors in terms of spectral similarity were also analyzed and presented in Figure 2.

Figure 2.

For each target peptide p (with standard spectrum s) of length 20, the similarity scores σ (p, p_k ) between p and a top neighbors p_k (with standard spectra s_k) were predicted; k ∈ {1, 2, …, K}. We consider K = 11 and precursor ion charge of +2. (A) The Hamming distance between p and p_k, averaged over all p’s, as a function of the rank of p_k ’s predicted similarity. Note that because the majority of peptides do not have close neighbors, the average Hamming distance is generally high for all top-scoring peptides; (B) The number of times p and its nearest neighbor (with largest predicted similarity) have an identical amino acid at position i (match count) where i ∈ {1,2, …, 19}; position 20 is ignored because it is either K or R. Match count is summed over all p’s ; (C) The spectral similarity between s and s_k, averaged over all p’s, as a function of the rank of p_k ’s predicted similarity (solid line), or as a function of the rank of p_k ’s Hamming distance in ascending order (dashed line).

As shown in Figure 2A, neighbors with higher SVM scores have on average smaller Hamming distance relative to their target. Figure 2B shows that the distribution of influential amino acids along the alignment between each target peptide and its highest scoring neighbor (i.e. peptide with the largest predicted SVM score) has the same trend as observed in Figure S2. In addition, as depicted in Figure 2C, the spectral similarity between a target’s and a neighbor’s standard spectrum is on average higher for neighbors with higher SVM scores, suggesting that the learned scoring function provides an effective estimate of the spectral similarity between two peptides. Although similar trends are observed by using only the Hamming distance as a feature, the learned scoring function is more predictive of the spectral similarity than the Hamming distance alone.

We further studied how well the predicted SVM score σ ( p₁, p₂) agrees with the actual spectral similarity ρ(s₁,s₂ ). A heat map depicting these two quantities is shown in Figure S4 (Supplementary Materials), indicating that the predicted similarity score positively correlates with the actual similarity (Pearson correlation coefficient: 0.43).

3.3 Evaluation of spectrum prediction

We calculated the average spectral similarity between predicted and true standard spectra for peptides with different lengths and charges (Table S1; Supplementary Materials). The accuracy of predicting the spectrum of a target peptide depends not only on the discriminating power of the classification model, but also on the neighbors that can be found in a spectral library and used to synthesize a target spectrum. In particular, the target sequences for which a better set of neighbors can be found should be expected to have better predictions of their standard spectra. To verify this, we calculated the confidence score for each target peptide as described in Methods, in addition to evaluating the prediction accuracy. The relationship between these two quantities is shown in Figure S5 (Supplementary Materials), indicating that prediction accuracy can be inferred from the confidence score. For example, the average spectral similarity for the subsets of predictions with confidence scores greater than 10^th, 50^th, and 90^th percentile of the distribution were 0.751, 0.782, and 0.814, respectively. We believe that the low spectral prediction quality for some target peptides is due to the lack of good quality neighbors in spectral libraries (e.g. for target sequences not similar to any sequence in a library), which suggests that the neighbor-based approach may not be appropriate for predicting spectra of these peptides.

We also investigated the influence of parameter K, which controls the number of neighbors used in the construction of the predicted spectrum and estimating the confidence of prediction. The average spectral similarity between predicted and observed spectra as a function of K (Figure S6; Supplementary Materials), however, did not show significant variation. Thus, we determined that K = 11 provided a good balance between speed and accuracy of prediction for all lengths and charges.

3.4. Assessment of predicted spectra in spectral library searches

The usefulness of predicted spectra was evaluated in two different spectral library and database searches. We first used a large set of spectra derived from a human cell line and searched it against three spectral libraries (NIST, NIST^KNN, and NIST^MA) and a sequence database. The sample comes from an extract of the human erythroleukemia cell line K562, where trypsin-digested peptides were eluted by a multistage gradient (Agilent 1100 Series HPLC) into electrospray ionization source of an LCQ ion trap mass spectrometer [3]. This set of identified spectra had already been included in the NIST spectral library; thus, the NIST library had an unfair advantage over its counterparts that used predicted spectra. However, it remained useful for estimating the empirical upper limits of identification.

Figure 3A–B shows that both predicted spectral libraries outperformed the sequence database search. This is probably because SpectraST’s scoring function [24] has utilized information of well-predicted peak intensities to improve the scores for the correct PSPs. We also compared the unique peptide identifications between predicted and real libraries at 1% FDR estimated by using a target-decoy search approach. The Venn diagrams in Figure 3C–D show that the predicted and real libraries have a large number of overlapping unique peptide identifications, while the real (NIST) library still has most unique identifications.

Figure 3.

Comparing sensitivity of spectral library search (NIST, NIST^KNN, and NIST^MA) and sequence database search (InsPecT). The number of positive identifications plotted as a function of FDR for charge +2 (A) and +3 (B). Venn diagrams corresponding to the unique peptide identifications for charge +2 (C) and +3 (D).

We then investigated the nature of the set of unique peptides that were identified in the NIST library but missed by the NIST^KNN library. As shown in Table S2 (Supplementary Materials), the average spectral similarity between predicted and real standard spectra was higher for peptides identified in both NIST^KNN and NIST than for peptides identified only in NIST. In addition, the average confidence scores for the peptides identified only by NIST were comparatively lower (Table S2), which suggests that confidence scores can be used to select high quality predicted spectra. To evaluate this, we searched the same query spectra against smaller spectral libraries consisting of subsets of predicted spectra with high confidence scores (we used the 50%, 40%, 30%, 20% and 10% of spectra with the highest confidence scores). As shown in Figure 4, spectral search against these smaller NIST^KNN libraries resulted in noticeably more identifications than against NIST^MA libraries with the same search space.

Figure 4.

Comparing sensitivity of spectral search on NIST^KNN and NIST^MA with equal search space corresponding to the 20% of the most confident predictions in NIST^KNN for (A) charge +2 and (B) charge +3 precursor ions. This resulted in 56,642 out of 283,122 peptides ions for charge +2 and 22,528 out of 112,564 peptide ions for charge +3. The same set of peptide ions were then selected from the NIST^MA for the parallel spectral search. To minimize the impact of decoy library (randomly shuffled peptide sequences) on small target libraries, we generated 50 decoy libraries using SpectraST and performed 50 independent searches in which the target library was appended with each decoy library. The numbers of positive IDs at each FDR cutoff were averaged over all 50 runs.

The increase in coverage of three well-studied proteomes for the situations when high-confidence predicted libraries can be created is shown in Table 2. These results indicate that the approach is practically useful for improving peptide identification.

Table 2.

The number of peptides in various proteomes additionally covered with high-confidence spectral predictions of the neighbor-based approach. Three thresholds for the confidence level were considered: top 10%, top 20% and top 30% of predicted spectra.

Table 2A
Human
+2	Length	Covered by NIST	Covered by the neighbor approach
			Top 10%	Top 20%	Top 30%
	7–20	113,323	23,176	50,479	80,117
	7–10	37,939	9,289	21,953	36,654
	11–15	50,857	9,478	18,974	28,324
	16–20	24,527	4,409	9,552	15,139
+3	Length	Covered by NIST	Covered by the neighbor approach
			Top 10%	Top 20%	Top 30%
	12–25	46,341	10,629	23,118	37,618
	12–15	15,666	3,974	8,094	13,374
	16–20	18,878	5,329	11,805	18,469
	21–25	11,797	1,326	3,219	5,775

Table 2B
Fly
+2	Length	Covered by NIST	Covered by the neighbor approach
			Top 10%	Top 20%	Top 30%
	7–20	39,458	3,200	8,436	20,216
	7–10	11,490	2,069	5,158	8,820
	11–15	18,580	2,500	4,954	7,384
	16–20	9,388	1,175	2,570	4,012
+3	Length	Covered by NIST	Covered by the neighbor approach
			Top 10%	Top 20%	Top 30%
	12–25	15,551	2,450	5,423	8,817
	12–15	4,256	822	1,713	2,903
	16–20	6,594	1308	2,959	4,579
	21–25	4,701	320	751	1,335

Table 2C
Yeast
+2	Length	Covered by NIST	Covered by the neighbor approach
			Top 10%	Top 20%	Top 30%
	7–20	31,165	4,185	10,736	26,552
	7–10	10,893	2,849	7,190	12,098
	11–15	13,866	3,142	6,360	9,476
	16–20	6,406	1,492	3,166	4,978
+3	Length	Covered by NIST	Covered by the neighbor approach
			Top 10%	Top 20%	Top 30%
	12–25	10,368	2,590	6,111	10,181
	12–15	3,100	875	1,928	3,336
	16–20	4,299	1,361	3,258	5,183
	21–25	2,969	354	925	1,662

Finally, to evaluate spectral identification in less well studied organisms and using unequal search spaces we performed a spectral library search on the set of experimental spectra from an in-house mosquito sample (Section 2.1). As shown in Figure 5, at equal (but small) search space the NIST library resulted in slightly more peptide identifications than its predicted counterpart NIST^KNN. However, the predicted libraries Mosquito^KNN and Mosquito^KNN-80% led to a significantly larger number of identifications than NIST even though they contain only predicted spectra. This suggests that a substantial expansion in search space compensates for the decrease in quality of predicted spectra. In particular, the 85-fold increase in the search space of the Mosquito^KNN library resulted in three times more identifications at 1% FDR compared to NIST, whereas Mosquito^KNN-80% almost doubled the number of identifications with only a 14-fold increase in search space. This is possibly because Mosquito^KNN-80% contains a higher quality predicted spectra than Mosquito^KNN and thus it achieves a better trade-off between the search space size and spectral quality. We also observe that the search of a hybrid library, which contained real spectra of NIST-covered peptides and predicted spectra of the remaining mosquito peptides, resulted in even larger number of identifications than the search of Mosquito^KNN.

Figure 5.

Comparison of spectral library searches with unequal search space. The numbers of positive identifications are plotted as a function of FDR for each of the seven libraries; the same types of lines indicate the same search space between groups of libraries. NIST and NIST^KNN are spectral libraries described in Section 2.3. Mosquito^KNN contained predicted spectra of all tryptic peptides (length 7–20 for charge +2 and 12–25 for +3) from the set of A. aegypti proteins in Swiss-Prot, and Mosquito^KNN-80% was a subset of Mosquito^KNN in which the predicted spectra had confidence scores greater than the 80^th percentile threshold. Mosquito^MA and Mosquito^MA-80% were counterparts of Mosquito^KNN and Mosquito^KNN-80%, respectively, but the spectra were generated using MassAnalyzer. Hybrid was a library in which NIST spectral library was combined with Mosquito^KNN; if a peptide sequence was present in both libraries, the spectrum from NIST was retained. In total, the set of A. aegypti peptides contained 395,213 tryptic peptides, of which 4,685 were present in NIST. Mosquito^KNN-80% and Mosquito^MA-80% each contained 65,346 peptides.

4. Discussion

In this study we present evidence that similar peptides generally produce fragmentation spectra with similar ion intensities of the corresponding product ions. We then exploit this observation to develop a novel neighbor-based method for predicting peak intensities of the product ions in a CID fragmentation spectrum. Finally, we demonstrate that the spectral libraries built from the predicted spectra can be used to improve peptide identification over a standard sequence database search.

The analysis of similarities of standard spectra corresponding to peptides of various lengths revealed some interesting findings. For example, we noticed that most short peptides (l ∈ {7, 8}) have very similar standard spectra regardless of the peptide sequence. Such peptides usually have higher peaks corresponding to the b₂ and y_l−2 ions for +2 and ${\text{y}}_{l-2}^{++}$ ions for +3 precursors (Figure S3). This effect, however, diminishes for longer peptides (l ≥ 12 for +2 and l ≥ 23 for +3 precursor ions; Figure S3).

Current approaches addressing the prediction of ion intensities in MS/MS spectra are based on two strategies: modeling of the basic chemistry of peptide fragmentation and data-driven strategies in which machine learning is used to train predictors from a large collection of peptide-spectrum pairs. While accurate, both of these approaches encounter problems, stemming from either incomplete understanding of chemistry of peptide ionization and fragmentation or because they rely on de novo learning of the rules of peptide fragmentation in a data-driven manner. The neighbor-based approach presented here exploits experimental spectra of previously identified peptides and, importantly, has the ability to estimate the accuracy of prediction based on the types of neighbors it utilized for prediction. Thus, our method extends spectral library-based search algorithms beyond previously identified spectra.

Previous studies provided evidence that spectral library searching was superior to a traditional database search and that the use of the kinetic model for predicting fragmentation spectra provided only minor advantages over the traditional database search [41, 42]. The results of our study, however, suggest that the prediction of intensities of fragment ions does improve peptide identification at equal search space (Figure 3), although we note that the methods for spectral library search and database search use different methods for estimating FDR.

While real and predicted spectral libraries provide better performance than conventional database search at equal search space, it is of greater interest to compare their performance in a more realistic application scenario. Our results suggest that for well-studied organisms with high spectral coverage of all detectable peptides, spectral library searches should be expected to outperform conventional database searches (Figure 3). On the other hand, for organisms with low coverage, the peptide identification process can significantly benefit if a high quality predicted library can be generated (Figure 5). In such situations, our results show that the best performance is achieved using hybrid libraries, i.e. libraries where a real spectral library available for an organism is complemented with high quality predicted spectra for the remaining peptides. We also suggest that the estimated quality of predicted spectra may be useful for developing more powerful peptide-spectrum matching algorithms. Generally, such approaches should lead to improved peptide and protein identification with impacts on studies ranging from fundamental biology to human disease intervention.