Despite considerable efforts, the concordance between the lipidome composition and quantities of individual species obtained by different analytical approaches from similar samples remains poor. Let us consider, for example, the composition of triacylglycerols (TAG) in human blood plasma – one of the most abundant and easily detectable lipid class in one of the best characterized body fluids. Conflicting numbers of TAG species were reported: 31 [8]; 18 [9] and 24 [10] with significantly different molecular compositions. Graessler et al and Oresic et al reported, respectively, 6 and 4 species comprising fatty acid moieties with odd numbers of carbon atoms, while Quehenberger et al reported none. Quehenberger et al pointed to TAG 566 and Graessler et al to TAG 568 as the group of isobaric molecules with the highest number of carbon atoms and double bonds, while Pietilainen et al [11] identified considerably larger species with 58 carbon atoms and up to 9 double bonds.

Although in population-wide studies the sample origin is an important factor, we argue that automated interpretation of spectra by software remains one of the major sources of lipidome compositional discordances [12]. Consider that lipids constitute a structurally diverse class of biomolecules that are differently ionized and fragmented in MS and MS/MS experiments. Major cleavage pathways of lipid molecular ions are often instrument-dependent (reviewed in [4], [13], [14]). Also, depending on the mass spectrometer design, lipid ions are detected with different mass resolution and accuracy. Therefore, the spectra interpretation software is usually tailored for a certain type of the mass spectrometer and/or spectra acquisition method; it may be underperforming if applied to spectra acquired at alternative instrument platform(s). Additionally, the software usually targets a selection of lipid classes that reflects scientific interests of the developer's team and may not recognize uncommon lipid classes or species - for example, those comprising some unconventional fatty acid moieties, or may be performing sub-optimally.

We also argue that efforts to develop software relying on a single “gold standard” interpretation algorithm may be counterproductive – unless users are able to incorporate their own interpretation algorithms into the same software framework. Along these lines, we developed the software LipidXplorer [12]. LipidXplorer implements the molecular fragmentation query language (MFQL) that can describe any user-defined fragmentation pathway of any lipid class and implement it into a customized spectra interpretation method. LipidXplorer design supports any shotgun lipidomics routine, from high throughout top-down screening for molecular diagnostics and biomarker discovery to the targeted absolute quantification of low abundant lipid species [6], [8], [15]–[20].

While processing spectra for the MasterScan, LipidXplorer imposes two experiment-specific constraints. The first constraint defines the minimal intensity of precursor and fragments peaks. The second constraint, termed the occupancy threshold, is related to the frequency at which a particular peak was observed in all acquired spectra; for example, setting the threshold to 100% implies that the MasterScan will only encompass peaks observed in each and every spectrum. These constraints help to adjust processing settings to the actual level of chemical noise and harmonize the confidence of lipid species identification between independent experiments. We underscore that spectra are processed independently of a mass spectrometer type: they are imported in generic machine-independent formats and, while building a MasterScan, only peak attributes reflecting basic instrument- dependent spectra features, are considered.

The MasterScan database holds all MS and MS/MS spectra acquired from all samples in the experiment and can be further interrogated by queries written in the molecular fragmentation query language (MFQL). The MFQL interpreter is written using PLY (Python Lex-Yacc) (http://www.dabeaz.com/ply/), a lexer/parser generator based on Lex and Yacc. Basically, each query defines what structure-specific “signature” ions and/or their boolean combinations should be recognized in MS and MS/MS spectra within the MasterScan. Both precursor and fragment ions could serve as “signatures” leading to unequivocal identification of species in complex lipid extracts [12], [22] (Figure 1). Typically, each query targets one lipid class, while many queries can be successively executed by the LipidXplorer. Lipid species identified by all executed queries are reported in a single results file. A collection of basic MFQL scripts covering major lipid classes is included into the distributed version of LipidXplorer and is available at its wiki site: https://wiki.mpi-cbg.de/wiki/lipidx/index.php/Main_Page.

Here we demonstrate how low mass resolution precursor ion spectra and neutral loss spectra can be interpreted by LipidXplorer and provide evidence that these interpretations are consistent with alternative analyses by DDA-driven acquisition of full MS/MS spectra. Let us explain the interpretation algorithm using a precursor ion scan spectrum as an example. A precursor ion scanning spectrum plots the abundance of one specific fragment ion produced from a precursor isolated within the resolution-dependent mass window that is moving with a certain small increment, usually 0.05 to 0.2 Th, along m/z range (Figure 2). If several precursor ion spectra are acquired, they can be aligned by precursor masses and then transposed to the format: [precursor mass]: [frag1, abundance], [frag2, abundance], …, [fragn, abundance], where the precursor masses are all masses within the considered mass range. This operation corresponds to the transformation of spectra in panel B to spectra in panel A, however the latter only comprises a restricted set of fragment masses. The spectra alignment algorithm employed by LipidXplorer [12] effectively creates a “virtual” MS/MS spectrum for each precursor mass observed in individual precursor ion spectra: the only difference with shotgun datasets acquired by DDA is that they comprise no survey MS spectra. The algorithm is also adjustable to the actual mass resolution and accuracy. Therefore, upon building a MasterScan from transposed spectra, lipid identification could proceed with the same MFQL queries in the usual way.

To validate cross-platform interpretation capabilities of LipidXplorer, we analyzed a commercial sample of the total lipid extract from E.coli (Avanti Polar Lipids, Alabaster, AL) by shotgun experiments performed in different analytical modes and on different mass spectrometers. Upon collision induced fragmentation, molecular anions of phosphatidylethanolamines (PE) and phosphatidylglycerols (PG) – the major constituents of E.coli lipidome produce abundant acyl anion fragments of their fatty acid moieties that, together with precursor masses, unequivocally identify their molecular species [5], [29], [31], [32] (Figure 1). We first acquired MS/MS spectra from all PE and PG precursors on a quadrupole time-of-flight instrument QSTAR Pulsar i and on LTQ Orbitrap Velos using data-dependent acquisition in negative ion mode. Molecular anions were selected with the unit mass resolution to prevent co-fragmenting neighboring precursors. Collision energies were optimized as described in [20], [32] and, in the course of analyses, either ramped with precursor masses within the range of 44 to 56.5 eV (m/z 600 to 850) (QSTAR) or applied as a normalized collision energy nCE=45% [20] (LTQ Orbitrap Velos). For better consistency, spectra were acquired with approximately the same mass resolution of 7500 (full width at half maximum, FWHM) for both LTQ Orbitrap Velos and QSTAR; the impact of mass resolution on lipid identification accuracy was examined in [12]. Experiments were performed in 4 replicas; a shotgun dataset comprising 31 MS and 321 MS/MS spectra was processed and individual species were quantified (Figure 3). Note that E.coli does not produce ether glycerophospholipids. If analyzed by precursor scanning for acyl anions of fatty acid moieties, identification of ether lipids would rely on matching a single acyl anion to the precursor mass since complementary alkoxide fragment is usually ca 20-fold less abundant [32]. In positive mode plasmenyl species of PE could be distinguished from plasmanyl species by specific fragments [33] accountable via boolean scans [23].

In parallel, the same extract was infused into a triple quadrupole mass spectrometer TSQ Vantage (Thermo Fisher Scientific) and 72 precursor ion scan spectra were successively acquired for masses of acyl anion fragments of common fatty acids, including all fatty acids recognized in the experiment above. Spectra were also acquired under unit mass resolution consistently with the experiment settings applied on the QSTAR and LTQ Orbitrap; collision energy (CE) was 50 eV; collision gas pressure 1.5 mTorr. The MasterScan was composed from the aligned and transformed spectra and interpreted by the same MFQL queries identifying species of PE and PG lipid classes [12]. To quantify the species, isotopic correction of precursors and fragment intensities was applied [12], [20]. We found that quantitative profiles obtained in three independent experiments on hybrid tandem machines QSTAR and LTQ Orbitrap and on a triple quadrupole Vantage instrument, were consistent (Figure 3).

Using the same E.coli lipid extract, we further tested if LipidXplorer could consistently interpret neutral loss scanning spectra. Upon collisional fragmentation in positive ion mode, molecular cations of PE and ammonium adducts of PG undergo facile neutral losses of their head groups (Δ m/z 141.02 and Δ m/z 189.04, respectively), which are conventionally used for their shotgun profiling [23], [34]. We performed shotgun neutral loss experiments on the TSQ Vantage under basic instrument settings similar to described above; however spectra were acquired in positive mode under CE=22 eV. Spectra were acquired in three replicas, processed using MFQL queries accounting for the head group neutral losses (available at the LipidXplorer wiki site) and normalized abundances of species compared (Figure 4). Note that in neutral loss scanning spectra identified lipids may only be annotated by the total number of carbons and double bonds in both fatty acid moieties. For consistent comparison with profiles deduced from negative mode precursor ion spectra (Figure 3), the abundances of isobaric species of the same lipid class were combined. For each lipid class, we observed good correlation between the profiles, suggesting that LipidXplorer consistently interpreted both precursor ion scanning and neutral loss scanning spectra.

We demonstrated that, because of its innovative design concept, LipidXplorer consistently interprets shotgun spectra acquired in any mode (MS, data-dependent MS/MS, precursor and neutral loss scanning) by any tandem mass spectrometer (quadrupole time-of-flight, linear ion trap Orbitrap or triple quadrupole). LipidXplorer imports and aligns MS and MS/MS spectra and identifies lipids under resolution, noise thresholds, tolerance and identification settings that can be customized for any mass spectrometer, acquisition mode or dataset. Since spectra interpretation does not rely on a reference database, the software could encompass species of any lipid class that were ionized and fragmented during a shotgun experiment. The same MFQL-based approach could be applied to identify lipids in LC-MS/MS spectra. However, species quantification would require a separate module for time-integrating their chromatographic peaks.

While it has become possible to consistently interpret shotgun spectra, the question remains: how accurate are those identifications for a given dataset and experiment settings? No statistical framework has yet been developed to either estimate a global false discovery rate, or compute the local probability that a particular assignment is correct. In some instances it should be possible to revert to manual inspection of acquired spectra, yet this approach is biased and hardly applicable to large scale lipidomics efforts.