Open search unveils modification patterns in formalin-fixed, paraffin-embedded thermo HCD and SCIEX TripleTOF shotgun proteomes

Highlights

• We offer a concrete list of five mass shifts to include when identifying proteins from FFPE tissues by shotgun proteomics.

• We offer a variety of methods that may be used to detect excessive false discovery rates in open search algorithms.

• We provide workarounds for problems likely to vex SCIEX TripleTOF users when attempting to use open-source software.

Abstract

The application of database search algorithms with very wide precursor mass tolerances for the “Open Search” paradigm has brought new efforts at post-translational modification discovery in shotgun proteomes. This approach has motivated the acceleration of database search tools by incorporating fragment indexing features. In this report, we compare open searches and sequence tag searches of high-resolution tandem mass spectra to seek a common “palette” of modifications when analyzing multiple formalin-fixed, paraffin-embedded (FFPE) tissues from Thermo Q-Exactive and SCIEX TripleTOF instruments. While open search in MSFragger produced some gains in identified spectra, careful FDR control limited the best result to 24% more spectra than narrow search (worst result: a loss of 9%). Open pFind produced high apparent sensitivity for PSMs, but entrapment sequences hinted that the actual error rate may be higher than reported by the software. Combining sequence tagging, open search, and chemical knowledge, we converged on this set of PTMs for our four FFPE sets: mono- and di-methylation (nTerm and Lys), single and double oxidation (Met and Pro), and variable carbamidomethylation (nTerm and Cys).

Introduction

A confluence of technologies in the early 2000s made formalin-fixed, paraffin-embedded (FFPE) tissues a viable resource for biomarker discovery. Rapidly scanning LTQs (linear ion traps) became available in 2003 [1]. LC-MS/MS for protein identification matured considerably as a discipline as Molecular and Cellular Proteomics produced its initial publication guidelines in 2004 [2]. Fractionation strategies for digging deeply into complex mixtures became commonplace as methods such as GeLC-MS (2003) [3] and MudPIT (1999) [4] became widespread. Greater cell type specificity became possible through laser capture microdissection (1999) [5]. Bernard Metz and coworkers began untangling the complex chemistry of FFPE through chemical treatment of synthetic peptides in 2004 [6]. In 2005, Brian Hood and colleagues produced some of the first LC-MS/MS inventories of FFPE tissues [7].

Hood’s early experiment combined laser capture microdissection and gas-phase fractionation with a Thermo Scientific LTQ to identify approximately 2000 distinct peptides from each sample. Their work revealed key themes that would dominate FFPE proteomics for years to come. First, many peptides released from FFPE samples contained “missed cleavages,” internal basic residues that increased their precursor charges; high resolution of precursor ions and eventually fragment ions would make these peptides more accessible in years to come. Second, FFPE peptides less frequently featured lysine residues at their C-termini, hinting at the role of primary amines in FFPE modification chemistry. Hood et al. gamely identified formyl adducts (+12 Da) of lysines and noted that more than half of observed Met residues were oxidized in some of their LC-MS/MS experiments. The broader palette of modifications, however, awaited more powerful informatic tools.

Identifying post-translational modifications under the 1995 Sequest model [8] requires a foreknowledge of all PTM masses along with the residues where they may be observed. Because the algorithm must consider an exponentially growing number of decorated peptides as multiple potential modification sites are found in each peptide sequence, PTM searches have gained a reputation for prohibitive slowness as well as a penchant for false positives [9]. Bioinformatics researchers have produced a variety of alternative strategies for the discovery of PTM masses, particularly sequence tagging and spectral alignment. The first of these techniques attempts to infer partial sequences directly from fragment ions and then allow for arbitrary mass shifts to be introduced to either side of this “tag” to reconcile the spectrum to a database peptide sequence [10]. Spectral alignment approaches such as QuickMod [11] or spectral networks [12] compare experimental MS/MS scans from different precursor masses to determine if fragment ions align with or without adding the precursor mass difference as a PTM. The approaches were generally recommended as a strategy for determining which PTMs should be included in a final database search rather than being seen as authoritative identifications of themselves [13]. In 2015, Ying Zhang and colleagues used spectral alignment for PTM discovery in FFPE to discover the prominence of lysine methylation in FFPE LC-MS/MS data [14].

Joel Chick of the Gygi Laboratory published the open search strategy in 2015 [15] offering the attractive promise of identifying 28% more spectra at similar FDRs by simply relaxing the precursor mass filter in Sequest (invoking a “∼10-fold” time penalty). Established proteome informatics teams at the University of Michigan and the Beijing Chinese Academy of Sciences created new search engines (MSFragger [16]) or modified existing ones (Open pFind [17]) to greatly accelerate the database search algorithm to enable routine open searching. A key enabling development was the use of indexing to determine which experimental spectra matched a calculated fragment ion mass, irrespective of precursor mass, a strategy first published in 2007 for Marshall Bern’s ByOnic algorithm [18]. Kong et al. achieved similar gains in identified spectra to those of the Gygi team through MSFragger, reporting 33.6% more spectra identified with a much lower time penalty. Notably, the MSFragger paper incorporated a single paragraph relating the use of the software to detect methylol adducts (+30 Da) in 442 LC-MS/MS experiments from FFPE breast cancer data (Kong et al. remained mute, however, concerning the frequency of lysine methylation in these data.). When Hao Chi and colleagues adapted pFind for open search by adding fragment indexing, precursor analysis, sequence tagging database reduction, and machine learning discrimination, they were able to profoundly boost identification rates over their earlier “narrow” searches in pFind, gaining between 61.2% and 88.2% PSMs [17].

To mainstream proteomics users, such gains in identified PSMs might have seemed the answer to long-standing frustrations surrounding the low fraction of MS/MS scans that we identify from experiments (50% of MS/MS scans measured at high resolution is currently thought to be a good result). Long experience in bioinformatics, however, caused our team to feel some skepticism at these claims, and we designed this study to evaluate the gains achievable through open search algorithms. We chose to emphasize FFPE experiments because of their complex and long-studied modification chemistry, and we included both Thermo high-resolution “HCD” [19] RAW files and SCIEX high-resolution “beam-type CID” WIFF files from the TripleTOF [20], since the latter have seen only minimal attention in open search articles to date. Would open search substantially improve the fraction of spectra we identified? Would its discovered palette of PTMs expand greatly upon those visible through sequence tagging?