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

Abstract

Searching spectral libraries in 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 that 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.

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.

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).

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).

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.

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.