Metabolite signal identification in accurate mass metabolomics data with MZedDB, an interactive m/z annotation tool utilising predicted ionisation behaviour 'rules'

Changes in the overall metabolite composition of living cells (metabolome) reflect a key end point in gene expression and make a major contribution to organism phenotype [1]. Although, no single analytical platform can offer a fully comprehensive survey of the chemical diversity representing the metabolome, continuous improvements in mass spectrometry (MS) instrumentation have allowed development of relatively standardised metabolite profiling or fingerprinting procedures [2]. A fundamental principle of mass spectrometry is the representation of metabolite features in any biological matrix by measurement of the spectrum of signals reflecting the mass to charge ratios (m/z) of their ionisation products. One advantage of MS over alternative spectroscopic methods such as Nuclear Magnetic Resonance (NMR) and Fourier Transform Infrared (FT-IR) is the opportunity to putatively annotate directly a spectral component by virtue of its atomic mass. In the context of a metabolomics experiment these 'first pass' annotations may be used to develop hypotheses relating to metabolite identity which are then tested by subsequent, more targeted, analytical chemistry methods. Traditional hyphenated MS profiling approaches provide simultaneous detection and quantification of discrete metabolite-derived peaks after chromatographic separation. In gas chromatography MS (GC-MS) a couple of hundred well resolved metabolite peaks are identified where possible by matching their positively charged ion spectrum (following fragmentation by electron impact in the gas phase) and column retention time to those of known standards [3]. GC-MS methods are well established and robust but are limited to analysis only of volatile metabolites. Although derivatisation can increase the volatility of a wide range of metabolite classes such chemical modification further increases the complexity of any annotation process based on atomic mass.

Alternative profiling methods utilising liquid chromatography coupled to mass spectrometry (LC-MS) provide sensitive tools for the analysis of a wider range of metabolites with higher polarity, lower volatility and much larger mass range without a need for derivatisation [4]. Resolved metabolite peaks can only be efficiently ionised when outside of the liquid phase under atmospheric pressure. Typical approaches include atmospheric pressure chemical ionisation (APCI) and electrospray ionisation (ESI). In contrast to electron impact, both methods allow 'soft' ionisation with little fragmentation in which a major product may be a pseudo-molecular ion comprising the protonated (+ ve ion data) or de-protonated (-ve ion data) parent molecule.

Peak detection and particularly spectral deconvolution in GC-MS and LC-MS are both technically challenging, time-consuming and very difficult to automate; LC-MS particularly has been hampered by poor analyte peak resolution and retention time variability which confound metabolite peak alignment [5, 6]. As a consequence, recent approaches to LC-MS profiling have concentrated on the non-targeted quantification of all detected ionisation products above a pre-set noise threshold which are compiled subsequently as several thousand un-annotated variables [7–9]. The variables in this high dimensional data consist simply of an ion mass and its retention time. Such data tables will contain much redundancy as the data acquisition rate of most instruments will be sufficient to generate at least 3–5 scans across each eluting peak. In addition, even with the extra resolving power of ultra high pressure liquid chromatography (UHPLC) a large number of peaks will still overlap or even co-elute in many regions of the chromatogram which demands the use of powerful signal alignment software for data pre-processing.

An alternative approach to LC-MS profiling is to generate a metabolite fingerprint representation without using the chromatographic dimension in which data variables are simply the detected mass values [10]. In such approaches [2, 10–16] the sample may be dissolved in an appropriate solvent and injected directly into the ion source (Direct Injection Mass Spectrometry; DIMS) or infused as a 'plug' flow using a HPLC system without a chromatography column (Flow Injection Electrospray Ionisation Mass Spectrometry; FIE-MS). Data representations may take two forms depending mainly on the accuracy and subsequent resolving power (mass/mass accuracy) of the system mass analyser. In nominal mass fingerprinting methods the intensities of all masses from all scans are integrated within pre-defined ranges (mass bins). Linear ion traps with quadrupole (Q) detectors are extremely robust and offer rapid scanning over large m/z ranges. However, as full-scan mass resolution is often less than 1000 all signals are binned to the nearest nominal mass value from around m/z 50 up to as high as m/z 2000, depending on instrument; under these circumstances metabolites are considered to be mono-isotopic and each mass bin could obviously contain signals derived from several metabolites. Despite these limitations nominal mass fingerprinting methods have attracted considerable interest as a first pass investigation tool as data pre-processing is quick and trivial with minimal likelihood of error and with a short cycle time (typically less than 5 minutes) they are suitable for experimental designs which require a high throughput of samples [10]. Time-of-flight (TOF) and hybrid Q-TOF mass analysers exhibit an average resolving power as high as 10,000 which, with internal calibration, may be sufficient to allow binning of signals in finer mass ranges (0.1 amu). This level of resolution however is not sufficient to distinguish between a large proportion of metabolites of different elemental composition which remain effectively isobaric at the maximum mass resolution achievable. With the advent of ultra-high accuracy mass analyzers such as the Orbitrap (resolution ~100,000) and Fourier Transform Ion Cyclotron Resonance (FT-ICR) analyzer (resolution potentially in excess of 1,000,000) this latter problem is greatly reduced [17–19]. Particularly using FT-ICR-MS instruments there is the opportunity to generate an accurate mass fingerprint (often referred to as a profile as variables are now considered effectively discrete) of an infused sample extract [20–22]. In such data representations mass 'peaks' are detected in the raw data and the centroid mass intensity calculated. The resolving power and sensitivity of these ultra-high accuracy mass analyzers is such that metabolite signals from molecules containing naturally abundant elemental isotopes (e.g. ¹³C, ⁴¹K, ¹⁵N, ¹⁷O,³⁴S and ³⁷Cl) are visible in the data [17, 18].

The past few years has witnessed a rapid increase in the number of publications describing experiments using particularly electrospray ionisation mass spectrometry (ESI-MS) profiling or fingerprinting at a range of mass accuracies. This activity has been accompanied by the recognition that the majority of m/z variables in most complex biological matrices cannot be matched unambiguously to metabolites of identical elemental composition found in publicly accessible databases [9, 19–27]. With the ultra-high mass accuracy achievable using FT-ICR-MS methodology [17, 18] this might seem surprising for metabolites with a mass < 500 Da for which the number of isobaric molecules based on calculated elemental composition should be very limited. Several factors contribute to this problem. In the simplest situation, given that many natural metabolites (particularly from plant sources) remain to be structurally identified, the relevant molecules may simply not be present in current databases. A more complex issue stems from the fact that all metabolites have a chance of containing one or more natural isotopes [19, 23, 24] of constituent atoms (e.g. ¹³C, ⁴¹K, ¹⁵N, ¹⁷O,³⁴S and ³⁷Cl); metabolite signals thus are rarely present as a single mono-isotopic peak and the spectral isotope distributions often results in measured masses of significant intensity that by chance are undistinguishable (ie not resolvable) from completely unrelated chemistry [18, 24]. A final factor relates to the diversity of potential ionisation products generated in the ion source in addition to parental pseudo-molecular ions (i.e. [M+H]¹⁺ and [M - H]^1-). Thus, in both fingerprinting and profiling approaches it is common to find adducts with either sample matrix components (e.g. in positive ion data: [M + Na]¹⁺, [M + K]¹⁺ and [M + 2Na - H]¹⁺ and negative ion data: [M + Cl]^1- and [M + Na - 2H]^1-) or mobile phase solvents (e.g. [M + NH4]¹⁺ in positive ion data). Additionally, although ESI is traditionally considered a 'soft' ionisation method a number of more sensitive metabolites will rearrange and fragment with neutral loss of common moieties such as water from a hydroxyl group [M - H2O + H]¹⁺ or ammonia from an amine group [M - NH2 + H]¹⁺ or formate from a carboxyl group [M - CO2H + H]¹⁺. Clearly, it is possible also that a range of potential ionisation products will contain various combinations of isotopes, adducts and neutral losses. Finally, ionisation products also can carry more than one charge (e.g. [M + 2H]²⁺) or can, through non-covalent interactions, produce stable molecular clusters, for example dimers, in which only one component is charged (e.g. [2M + H]¹⁺). The potential to form different ionisation products will be dependent on the structure and physical properties of individual metabolites.

It can be concluded that a key aspect of any variable annotation strategy is to assess the relationships between signals and to base putative annotations on models that best fit all of the data. The presence of isotopic atoms and either loss, or gain, of selected chemical moieties all afford specific changes in accurate mass which mean that the exact mathematical relationships between detected ions are informative. Natural isotope abundances in metabolites reflect the percentages at which isotopes of an element occur naturally in the environment. Thus analysis of isotope relative abundance patterns provide a useful approach to 'filter' unlikely annotations based on molecular formulae calculated from accurate mass measurements [19, 23]. Annotation strategies have been reported recently incorporating algorithms interrogating some, but not all, of these atomic mass mathematical relationships [19, 22, 23, 25, 26]. In most instances the adopted strategy centred on an analysis of the full spectrum of aligned peaks resolved at ultra-high mass accuracy in the biological matrices under comparison. A less computationally intensive strategy more suitable for high throughput experiments is to focus the annotation exercise on m/z signals which make significant contributions to models explaining the metabolome differences between two biologically different sample classes [10, 14–16, 22, 28]. In nominal mass metabolite fingerprint data [10, 14–16, 28] such experiments make the simple assumption that 'explanatory' metabolites will be represented by several mathematically-related ionisation products (e.g. pseudo-molecular parental ion, ¹³C isotope, salt adducts) which additionally exhibit tight co-variance in a simple correlation analysis; where LC-MS profiling is used ionisation products related to the same metabolite should additionally be found at the same retention time [14, 29]. Similar approaches are clearly possible using accurate mass data [7, 8, 24, 27].

It is time consuming and often computationally intensive to pre-process LC-MS profile/fingerprint data to a level where the researcher is absolutely confident of understanding the origin of any ionisation product (i.e. whether signals of interest are parental pseudo-ions, salt adducts, solvent adducts, neutral losses, clusters). This is a particular problem when high-throughput is required or for exploratory analyses of a new biological matrix. An alternative approach described here is the use of a database strategy that will allow searching based on all potential ionisation products that may be predicted to form during electrospray ionisation.

Searching for likely annotation candidates based on accurate mass information in publicly accessible databases is in itself time consuming as individual database coverage of natural chemistry varies and so a comprehensive search requires query of several. Unfortunately, with few exceptions, databases with appropriate metabolite mass information can contain much redundancy, resources for curation are often limited (consequently it is not uncommon to find mistakes relating to mass values, molecular formulae and structure) and some of the entries relate to ionic states, often from interactions with salts. To overcome some of these problems we have developed MZedDB [30] a database which uses an archive in a common format of all metabolite 'structures' derived from several widely used and publicly accessible databases. The value of such an approach to avoid ontology problems was recognised recently by a consortium of yeast researchers [31]. Using a set of 'rules' derived from structural information and physical properties (such as number of H-bond acceptor/donors, number of OH/COOH/NH2 groups, number of acidic H or basic O in molecule) MZedDB will generate a list of potential adducts and neutral loss fragments that are likely to form for each structure and calculates on the fly the accurate mass of every potential ionisation product which provide targets for searches based on accurate mass. Starting with a list of m/z signals MZedDB supports a range of manual or semi-automated annotation strategies based on either m/z mass or predicted molecular formulae at a range of mass resolutions. In the present article we describe the development of MZedDB and illustrate several typical applications relating to annotation of ESI-MS data which take into account both predicted and measured ionisation behaviour.

In high throughput LC-MS fingerprinting/profiling techniques utilising electrospray ionisation (ESI-MS) it is a common precaution to use analytical instruments at lower than maximal mass resolution to avoid problems with data alignment [21, 24]. As a result, m/z annotation suggestions may often include a large number of structurally diverse but effectively isobaric candidates when only parental pseudo-ions (i.e. [M + H]⁺ or [M - H]^- m/z signals) are considered as ionisation products. Although recent papers have stressed the importance of recognising isotopes to avoid miss-annotation of m/z signals in accurate mass LC-MS data [9, 18, 21, 24–26] the present paper shows clearly that ionisation products other than parental pseudo-ions in reality can account for upwards of 40% of the signals in many accurate mass spectra. Against this finding it is unsurprising that many accurate mass m/z signals lack annotation 'hits' if structures other than parental pseudo-ions are not included in the searchable databases. Thus we conclude that annotation decisions can be usefully guided by both determining from the outset whether any signal in question is actually an adduct, isotope or neutral loss fragment or indeed any combination of all three. Takahashi et al. [27] report the ability to modify a search of the species-specific KNApSAcK database dependent upon the presence of sodium, potassium or ammonium adduct ions derived from the solvent used for sample preparation. Other researchers have downloaded the general metabolite database KEGG LIGAND and calculated the exact mass of 7 possible adduct ions for each entry to provide query targets in a customised database [22]. MZedDB consolidates and expands on this strategy by harvesting SMILES and calculating the accurate mass of a much more comprehensive list of metabolites from a range of databases, as well as including hyperlinks to rapidly access detailed information from entries in the original databases. By restricting MZedDB searches to entries derived from only a selected number of the original databases m/z annotations can be refined in a species specific context. An additional advantage of MZedDB centres on the ability to extend annotation searches to include a much larger range of adducts, neutral loss fragments and clusters by implementation (using on the fly calculations) of a simple set of Potential Ionisation Product 'rules' which effectively filter searches to exclude unlikely candidates based on SMILES information and which can be customised in line with any prior knowledge concerning the matrix under analysis (e.g. salt content of tissue, extraction or HPLC solvent). Ultimately, metabolite identity assignment is crucial to derive biological meaning from metabolome modelling experiments and thus data annotation tools such as MZedDB which allow more intelligent searching need to be expanded to cover more comprehensively a wider range of natural chemistry and take into account the characteristics of new biological matrices. For example, to improve the analysis of human urine samples it is anticipated that future implementations of MZedDB will be extended to include any common biotransformations of parent metabolites (e.g. glucuronidation, sulphation, and de-methylation) in attempts to link ionisation products to original metabolites.