Introduction

The appearance of quantitative and robust high-resolution (HR) MS instruments while recording HR-full scan, allow analysts to go further beyond pre-acquisition choices and targeted analysis [1–9]. With the same HRMS instrument, targeted and untargeted data, routine and research, quantitative or qualitative analyses can be performed and can reveal the presence, level, or fate of exogenous and endogenous compounds [10]. In routine labs using HRMS, global and untargeted data are easily acquired but data are usually treated in a targeted manner to quantify expected compounds [1–11]. However, when treated as untargeted, this large data can possibly reveal unpredicted compounds of interest in particular samples [10, 12–17].

Typically, global and untargeted (LC-)HRMS acquisition allows analysts to reveal biomarkers and toxic/illicit compounds in, respectively, metabolomics and screening labs (typically, labs involved in food safety, waste water, environmental, security, antido**, forensic, and toxicology analyses). Even further, in clinical labs, untargeted analysis could also be applied to patients’ blood samples in order to reveal unexpected compounds or unexpected levels for preliminary diagnosis attempt [18–21].

Metabolomics labs are used for untargeted workflow (see online resources, Figure S-1) [22]. Untargeted data treatment usually starts with the detection of all ions, called features. A feature is a m/z accurate value and its retention time. Frequently, metabolomics has to cope with unknown-unknown features that are compounds with no entries in databases. In this case, the determination of chemical composition with substructure identification (neutral losses and fragmentation trees) is the first step [23–26]. Here, the identifications of unknown-unknowns will not be considered. However, the degree of confidence in known-unknown identification will be treated carefully because it is the fundamental step prior biological interpretations.

Screening labs focus essentially on known-unknowns that are compounds with entries in databases such as Chemspider, HMDB, MetLin, mzcloud (https://www.mzcloud.org), etc. [23, 27–35]. In screening labs, reference (authentic, pure) standards are generally available and (semi)-targeted analysis are performed. But untargeted approach has already shown its potential for the discovery of unexpected xenobiotics [10–14] or to show the effects of toxins on endogenous metabolites (toxicometabolomics) [36]. From LC-HR-full scan acquisition, data mining can be time-consuming [12]. But, using a very stringent selection of compounds (e.g., P-values, fold differences in comparison to controls, etc.), a rapid and easy data treatment is possible (e.g., ≤4 h) when sample number is limited (e.g., ≤25/d) and if a limited number of unknown molecules are considered (e.g., ≤50).

In the past, total ion current (TIC) or base peak chromatograms could be used to reveal unexpected but intense ions. Today, dedicated software allows treating HRMS full scan data more in depth and comparing efficiently between- or within-samples or subjects (Figure 1 and Supplementary Materials, Figure S-1). Today, user-friendly softwares allow using such untargeted workflow routinely (Figure 1a) by lab heads or experienced technicians. This untargeted workflow would consist of a preliminary diagnosis attempt with the identification of unexpected metabolites or unexpected levels (relative quantification) and could be followed by further targeted, quantitative, and confirmative analysis.

Figure 1
figure 1

Data treatment and representation in untargeted analysis with HR-full scan acquisition. (a) Untargeted workflow in metabolomics/screening laboratories. (b) 2D-gel representation of the global HR-full scan (Progenesis QI software, Newcastle, UK); the software data treatment allows regrou** related adducts (right-hand box) for further processing. Principal component analysis showing differences between two groups and (c) main compounds involved in the group discrimination (d)

Full size image

In untargeted diagnostic analysis, the number of uncovered compounds that would have to be potentially identified are >1000 (endogenous compounds and xenobiotics). If all reference standards have to be systematically injected in the same analytical sequence, identification would be too time-consuming. In addition, some reference standards would probably be unavailable or too expensive to purchase.

In the identification of features/compounds, there are two main “visions” coming from metabolomics and screening lab communities. Whereas various recent articles can be taken into account, both visions are based on two central reference documents that have been edited in 2002 and 2007 by the European Communities, EC [37] and the Metabolomics Standard Initiative, MSI [38], respectively.

For LC-HRMS analysis, the document from the European Communities (2002/657/EC) [37] asks for (1) the relative retention time (RRT) of the analyte(s), set with a tolerance of ≤ ±2.5%, (2) relative product ion intensities (ideally, 4 MS2 ion ratios), set with a tolerance of ±20 to ±50% with the reference MS2 spectra, and (3) ≥4 identification points (IPs): 2 IPs for the precursor m/z and 2.5 IPs for a product ion determined accurately.

In the MSI document [38], there are four levels of metabolite identification: the first level (1) is a definitive identification performed with two orthogonal analyses compared with the reference standard(s) [38–43]. The second level (2) is the putative annotation (identification) of compounds and refers to the correspondence with MS data from databases or literature. The third and fourth levels (3 and 4) stand for putatively characterized compound classes and unknown compounds, respectively.

In light of HRMS data, key information and confidence level for known-unknown identification should be renewed as it has been mentioned already by some authors [5, 39, 40, 44]. In this article, we will focus on the degree of confidence in known-unknown identification using HRMS instruments [2, 45]. The vision of metabolomics and screening labs, sustained by the references mentioned above [37, 38], will be considered and merged. An identification confidence scale and an ID score will be proposed in an attempt to build a bridge between the metabolomics and screening lab communities.

The need to identify known-unknowns should grow with the increasing availability of HRMS in various routine labs, and the further implementation of global and untargeted approaches as preliminary diagnosis prior to possible further targeted analyses (Figure 1a). The aim of the proposed ID scale and score is to establish a base and a degree of confidence in known-unknown identification with an inter-laboratory and inter-community understanding. Confident identification is crucial for biological interpretations in order to stratify diseases, reveal known biomarkers, understand disease mechanisms, monitor therapeutic outcomes, and uncover acute or chronic intoxications, etc. Finally, using LC-HRMS data from plasma extracts, concrete examples of identification with ID score determination will be discussed.

Parameters to Consider for the Identification of Unknown Compounds Using HRMS Analysis

Before proposing a scale of confidence for the identification of non-peptidic known-unknown compounds (see next chapter), this is crucial to evaluate the probability of misidentification. There are various parameters influencing false positive identifications, defined here as a wrong match between a feature (m/z-RT item) and an attributed hit (entry) from a database. These parameters depend mainly on compound composition and isomeric possibilities, MW and mass defect, matrix and its clean-up, chromatography, and detection specificity including mass accuracy (MA) and mass resolution (R; R ≥ 20,000 at full-width-half-maximum) (see on-line resources, Figures S-2a to S-2g).

In the Supplementary Materials, Figure S-2e depicts the compound entries in ChemSpider per MW ranges [27]. The Gaussian bell curve shows that the majority of compounds are between 300 and 500 Da. Figure 2a shows, with a log scale, the number of hits found in Chemspider, with Xcalibur and in HMDB against MA tolerance (in ppm) for four m/z monoisotopic values, corresponding to four drugs from 150 to 500 Da. Xcalibur software (Xcalibur ver. 3.0.63; Thermo Fisher Scientific, San Diego, CA, USA), used as a chemical composition generator, was set with the following possibilities: C0-30, H0-60, O0-15, N0-10, P0-3, S0-3 with or without Cl0-3, F0-3, Br0-3.

Figure 2
figure 2

(a) Number of hits (entries) for 4 m/z (corresponding to four drugs) found in Chemspider, Xcalibur used as a chemical composition generator (here set with C0-30, H0-60, O0-15, N0-10, P0-3, S0-3, Cl0-3, F0-3, and Br0-3) and HMDB against mass accuracy tolerance (in ppm). (b) Retention time difference (+/–; in min) for unknown compounds after retention time alignments (αmeas) with the reference analysis (αref) and for the three chromatography classes, A, B, and C, corresponding to three relative retention time deviations. Chromatography Class D has no chromatographic references. (◊), (*) and (**) are RRT tolerance values according to [37] and [51] with and without isotopic internal standards, respectively. RT of unknown compound should be >2 void volumes (2xT0)

Full size image

In general, there is a log difference in the number of hits found in the HMDB, a chemical composition generator, and Chemspider (Figure 2a). With the chemical composition generator, the number of compositions significantly increased with (1) higher MA tolerance, (2) higher m/z monoisotopic values, and (3) when Br, Cl, and F atoms were taken into account (e.g., synthetic xenobiotics) (Supplementary Materials, Figure S-2a and b).

Overall, these figures shows that the risk of false positive identification depends on the MW (m/z), MA, and the database chosen, especially if all possible isomers are considered. This risk has to consider two extremes: (1) “small size” databases (e.g., HMDB, Metlin, KEGG, etc. [S-1) and the already ≈ 5000 commercially available compounds (food additives, drugs, endogenous or plant metabolites, and phytochemicals) [23]. The crucial question is to define the degree of confidence and this is why an identification scale and ID score considering HRMS data among other information can be useful.

Confidence Scale for Known-Unknown Compound Identification Using HRMS Analysis

In their review article [49], Lehotay and colleagues criticized the identification points system and defined it as “non-scientific.” They argued that identification requires “careful deductive thinking” rather than “arbitrary requirements that need to be met by an unthinking analyst.” One can agree to some extent to this point of view. Indeed, identification depends on many non-proportional parameters and is a case-by-case work. Thus, it appears simplistic to identify compounds with points attributed or not (1 or 0). However, a “perfect” confidence scale taking into account all parameters in small increments would be unworkable. Nevertheless, the great advantage of a confidence scale, ID score, and IPs, is the common basis given to the wide community of analysts. This is why similar work about identification confidence has already been done in-depth for NMR experiments [50].

The proposed scale tries to be simple and opens a space for deductive comments with free attribution of some identification points (IPs) by the analyst. Today, it is all the more important to give some flexibility to analysts because unknown identification can depend on specific HRMS system characteristics or software packages [23, 51]. On the other hand, a common base appears necessary to reduce possible confusion originating from specific MS manufacturers’ possibilities [23]. The proposed ID scale and score allow building a common basis in known-unknown identification for scientists in screening and metabolomics labs. Whereas the scale is built on the quantitative collection of IPs, the degree of flexibility given for the attribution of IPs, is rather qualitative. This flexibility opens ID scores to inter-bioanalyst variability but ID score difference should remain moderate.

Peak shape symmetry, blank injection, carry-over, background subtraction, chromatography performance and retention time alignment, signal intensity, mass accuracy calibration, signal/noise ratio, and limit of identification/detection/quantification will not be considered but can influence somehow the identification confidence. Eventually, putative identification relies on the analyst’s responsibility and competence to guarantee the above parameters and the traceability of the data [49].

The proposed scale is based on the following three criteria corresponding to the first three columns in Table 1: Criterion I, Criterion II, and Criterion III.

Table 1 Confidence Scale for Unknown Compound Identification and ID Score. The first 3 Columns Allow Determining the ID Score as [a Number, a Letter, and a Number] (e.g., 2D3) Corresponding, Respectively, to a I. General Identification Category (1 to 4), II. Chromatography Class (A to D; see Table 2) and III. Identification Point (IPs) Level (1 to 3). IP-Level is Attributed by Collecting a Certain Number of IPs (from 5 to 30 IPs; see Table 3)
Full size table
  1. I.

    General Identification Category: a number from confirmed (#1) to unknown (#4) (four categories: 1, 2, 3, and 4).

  2. II.

    Chromatography Class: a letter from the narrowest chromatographic tolerance (A) to no chromatographic references (D) (four classes: A, B, C, and D)

  3. III.

    Identification Point Level: a number from the maximum (#1; >15 IPs) to the minimum level depending on the number of IP attributed (#3; ≥5 IPs) (three levels: 1, 2, and 3).

For each of the three criteria, a grade is given allowing to determine an Identification Score (ID score) as [a number, a letter, and a number]. For instance, an unknown compound, identified in one LC-HRMS analysis, with no chromatographic reference but with 5 IPs attributed, would have an ID score defined as 2D3. For each identification score, Table 1 associates a qualitative word expression [46]. Below is the complementary information for each criterion.

I. General Identification Category

In Table 1, the first criterion/column (I) is based on the denomination given by the Metabolomics Standards Initiative (MSI) where four general levels of identification confidence have been proposed [38]. Here, these four levels are denoted as General Identification Category. According to the MSI and other references [38–43], these four categories stand for 1, confirmed compounds (definitively identified), 2, putatively identified compounds, 3, putatively identified classes or annotated compounds, and 4, unknown.

Category #1, confirmed compounds, is based on NMR analysis [50] or two orthogonal analyses with the reference standard(s) injected in the same analytical sequence and, in the proposed scale, with at least one analysis using HRMS (e.g., LC-, CE-, SFC- or GC-). Category #2, the large box of putatively identified compounds, is based on HRMS analysis with or without reference standards. Finally, Categories #3 and #4 concern putatively characterized compound classes or annotated compounds (possibly identified compounds but with a too low confidence), and unknown compounds, respectively.

Low to high identification confidence is established according to a Chromatography Class and a number of IPs attributed (column #2 and #3 in Table 1, respectively). Below are the four Chromatography Classes (II) and three IP-levels (III).

II. Chromatography Class

The second criterion is based on relative retention time (RRT) and has four classes of confidence, A to D (Table 2 and Figure 2b). Classes A, B, and C stand for identical, very similar, or close hydro/lipophilicity, respectively, whereas for Class D, there are no chromatographic references. Even if chromatographic references are not always mandatory for good identifications, it is usually inevitable when different isomers have to be discriminated. This point has to be documented by the analysts.

Table 2 Chromatography Classes, A to D, According to Relative Retention Time (RRT) Tolerance Values
Full size table

RRT is determined in the analytical run containing the unknown molecule(s) and in the reference run containing the reference standard(s). RRT definition is given in Figure 2b (box in the right-hand side) and determination of RRT deviation is shown in Supplementary Materials, Figure S-3. The two runs: #1, unknown(s) and chromatographic internal standard(s) (IS) and #2, authentic reference standard(s) and IS, are compared. The Class A corresponds to RRT deviation with the smallest tolerance and is based on the 2002/657/EC and WADA TD/2015/IDCR documents [37, 52]. According to 2002/657/EC, RRT tolerance should be ≤ ±2.5%, whereas for the WADA document [52], RRT tolerance is ≤ ±1% but ≤ ±0.5% when isotopic internal standards (IS) are used. The Chromatography Class A “imposes” reference standards to be injected in the same chromatographic conditions and generally within a few days. According to various documents [37, 52, 53], RT of the unknown compounds should be at least 2× the RT of the void volume, t0 (LC system and column). When a second analysis is performed for confirmation, it should be recommended to have RT of the unknown ≥ 10× void volume RT.

Classes B and C correspond to RRT deviations (tolerance) ≤ ±10% and ≤ ±25%, respectively, which allow using previous analytical runs, in silico calculated RRT [54, 55], and chromatography databases. The lowest, Class, D, is attributed when there are no chromatographic references or RRT deviations > ±25%. Figure 2b depicts for Classes A, B, and C, the typical delta of time (min) that are accepted after alignments, with RRT tolerance ≤ ±0.5, 1, and 2.5% (A), ≤ ±10% (B), and ≤ ±25% (C) (Figure 2b and Supplementary Materials, Figure S-3a).

Beside the fact that an accurate m/z or a specific chemical formula can meet many stereochemical possibilities, there is a probability associated with each isomer for its detection in a specific environment (occurrence in matrix, ubiquitous molecule or not, etc.). For instance, Chemspider can rank the isomers with their links to various sources, including PubMed. This allows giving an occurrence probability for a compound in a real situation. This information should be provided especially for Chromatography Classes B, C, and D. Moreover, this occurrence probability can be used by the analyst to allocate IPs.

III. Identification Point Levels and IPs attribution

There are three IP-levels from 1 to 3 for very high, high, and normal level, respectively. IP-levels are attributed based on the number of IPs collected for the unknown compound: 5 to 9.5, 10 to 15, and >15 IPs, respectively. Below 5 IPs, the unknown should be degraded to Category #3 as annotated compounds.

Table 3 lists the six different groups (from i to vi) where IPs can be collected from HRMS data. The six groups take into account (i) the precursor ion determination with MA, R at m/z (R m/z ), and if the adduct has been identified (e.g., [M + H+]+), (ii) the fragment ion(s) from different MSn acquisitions, (iii) the relative ion intensity in MSn spectra (ion ratio), (iv) the relative isotopic abundance (RIA), (v) the fine isotopic distribution (FID), and (vi) additional information. MSn stands for any kind of induced dissociation/fragmentation taking place after the MS entrance (e.g., MS2, MS3, MSE, MSALL, MSAIF, SWATH-MS, vDIA, etc.). Additional information is given in Table 3 and its legend.

Table 3 Attribution of Identification Points (IPs) from 6 Groups (i to vi) and IP-Levels (1 to 3) According to the Number of IPs Collected. There is a Maximum Number of IPs per Group and a Total of 30 IPs can be Collected
Full size table

Each group has a maximum number of IPs that can be collected (column #3, Table 3) whereas the number of IPs per group could exceed these maxima (column #4, Table 3). For instance, the maximum number of IPs for the group (ii) [fragment ion(s)], is 10, whereas 5 MS2 ions measured with a MA <5 ppm in a product scan acquisition (MS2 with precursor isolation width ≤5 Da), could give 12.5 IPs. Figure 3 plots the maximum number of IPs per group and the number of IPs needed to obtain one of the 3 IP-levels. IPs have to be collected from at least two groups.

Figure 3
figure 3

Number of identification points (IPs) for each of the six different groups where IPs can be collected in order to establish the IP-level. IP-level is 1, 2, or 3 (◊) for very high, high, and normal level corresponding to 5 to 9.5, 10 to 15, and >15 IPs collected, respectively. Each group has a maximum of IPs (*) and a maximum of 30 IPs (∆) can be collected. (□) Corresponds to the minimum number of IPs collected for identification, according to 2002/657/EC recommendations (8.5 = 2 IPs for precursor m/z, ±2.5 IPs for one MS2 fragment m/z, and 4 IPs for relative ion intensities of 4 m/z)

Full size image

Details on the attribution of IPs are given in the 4th column of Table 3. Whereas the attribution of IPs is, to some extent, arbitrary, this is based on real data. For instance, in HMDB, there are significantly more hits corresponding to accurate m/z values when all adducts are considered rather than [M + H+]+ adduct only (mean = 8× in the positive mode; Supplementary Materials, Figure S-2g). In HMDB, there are 33 and 15 adducts listed for positive and negative mode, respectively. Thus, when the adduct is identified (e.g., [M + H+]+) thanks to other adducts; see Figure 1b, [43]), MA ≤5 ppm and Rm/z ≥50,000, an additional IP can be allocated.

In most HRMS analyses, the determination of precursor ion MA is a prerequisite (group i) and, if possible, identification of the adduct type by other observed adducts should be investigated (Figure 1b). Group ii can collect up to 10 IPs that is consistent with the fact that MSn fragments (unless unspecific neutral losses) from a precursor give strong evidence in unknown identification [42]. Additional IPs must be collected from another group than ii. MSn fragments should be compared with measured fragments found with reference standard fragmentation(s) (≠ in silico). IPs can be collected from in silico proposed MSn fragments but in group vi (see below). MSn fragments can be produced in (1) MS2, MS3, or MS4, (2) SWATH-MS or vDIA [56, 57], and (3) MSE, MSALL, MSAIF[57, 58], which are all different denominations corresponding to, respectively, narrow (≤5 Da), medium (usually between 20 and 150 Da), or large (>200 Da) precursor isolation in the quadrupole. Thus, different IP numbers are attributed depending on precursor isolation width (≤5, ≤25, and >25 Da; Table 3, column #4) [37, 54]. The attribution of 2.5 IPs for a MS2 product ion (precursor isolation width ≤5 Da) corresponds to 2002/657/EC recommendations [37].

Group iii consider MSn ion ratios that are key information in GC-MS spectra with electron-impact fragmentation. In the 2002/657/EC document, tolerance (relative) for ion ratios is between ≤ ±10% and ≤ ±50%, depending on fragment ion intensities [37]. Unfortunately, in LC-HRMS, the MSn ion ratio is very dependent on mobile phase, ion source design ESI parameters, collision cell design, fragmentation energy, etc., making of the tolerance proposed in 2002/657/EC unrealistic unless reference standards are injected with the exact same LC-HRMS conditions. More globally, and according to Mol et al. [59], ion ratio tolerance of ±45% is an excellent fit-for-purpose compromise in (LC-)HRMS analysis. Even wider tolerance can be useful to reduce false positive identification as far as additional MSn ions are used [42]. This is why two wide tolerances in MSn ion ratios between unknown and reference spectra have been proposed to collect IPs in (LC-)HRMS analysis: 1 IP and 0.5 IP for a ≤ ±45% and ≤ ±90% tolerance (relative). For instance, with a least/most ion ratio = 0.58, tolerance values of ±45% and ±90% would give ratio margins from 0.406 to 0.754 and 0.058 to 1.102, respectively.

Searching for the identity of a known-unknown compound, the m/z value is frequently used to match entries in databases. The number of hits proposed can be reduced if the chemical composition is established or at least if some chemical possibilities can be discarded. In this respect, relative isotopic abundance (RIA; group iv) as well as the fine isotopic distribution (FID; group v) give key information [47]. RIA and FID take into account M, M + 1, M + 2 etc. intensities and resolved isotopes (e.g., 15N,33S, 34S, 18O and 13C), respectively, and can be determined theoretically with appropriate software. Various RIA calculations have been proposed and results reliability can be dependent on ion intensities, HRMS technology used, Rm/z, and the number of isotopes considered [60–66]. The use of FID is straightforward (even if isotopic abundance can vary slightly, (< ~4% [67, 68]) but relies on Rm/z and the degree of overlap between isotopes. Therefore, in groups iv and v (RIA and FID, respectively), the number of IPs is attributed according to the analyst’s documented interpretation in the capability to reduce the number of chemical compositions with these two parameters.

Figures S-3b and S-3c present typical information that can be obtained with RIA and FID. The RIA tolerance can be adapted to the number of carbons. Indeed, the discrimination between chemical compositions is more easily obtained for small molecules with <15 [C] than for compounds containing >30 [C] (Supplementary Materials, Figure S-3b). The interest of FID can be seen with the discriminations of 13C, 2H, 15N, 18O, 33S, and 34S isotopes at M + 1 and M + 2 of sulfamethoxazole (C10H11N3O3S, Supplementary Materials, Figure S-3c).

Some parameters have not been taken into account. For instance, as found in mzcloud database, the number of entries increases exponentially when MSn m/z is <100 (Supplementary Materials, Figure S-2f). Similar results were found for GC-MS [49]. It suggests that identifications of unknown using product m/z ≤100 are less specific in comparison to m/z >100. No bonus/penalty system has been considered for the ID scale as some authors have proposed [5], but the analyst should also report contradicting information.

Finally, additional information (group vi, Table 3) can be used to collect up to 7 IPs. This maximum represents the second value behind group ii (up to 10 IPs attributed for MSn fragment ions; Figure 3). For instance, non-analytical information from a policeman, judge, witness, clinician, or patient testimony could be considered to corroborate the analytical data. In source fragment ions [69], MS1 adducts and their ratios, collision cross-section (measured from a drift time in ion mobility MS) [70–72], (bio)transformation products from the unknown [73], correlation between the levels of known and unknown metabolite(s) [48], and specific effects of the unknown compounds can be considered to collect IPs. In group vi, up to 3 IPs can be attributed in a subdivision denominated “other, to be documented” that gives some degree of flexibility to the analysts for using specific HRMS instrument characteristics, other chemical information, free or commercial tools, software and databases, etc. (e.g., collision cross-section library, in silico fragments, etc.).

From these six groups (i–vi, Table 3), a maximum of 30 IPs can be collected and three IP-levels can be determined as normal (3), high (2), and very high (1) for, respectively, 5 to 9.5, 10 to 15, and >15 IPs attributed. Below five IPs collected, compound identification should be considered as low and belong to the Category #3 (annotated compounds). Noteworthy, the use of intra-lab rather than inter-lab databases should be more efficient to obtain higher ID scores. Indeed, RRT, fragment ions, etc. are usually dependent on the LC-HRMS systems and conditions used.

In this proposed identification confidence scale, the lowest ID score for the identification of a known-unknown compound considering the 2002/657/EC document [37], would be 2A3 (strong confidence, Table 1). This relates to I, one LC-/GC-HRMS analysis (not confirmed by a second orthogonal analysis), II, Chromatography Class A for the narrowest RRT tolerance possible, and III, 8.5 IPs for the accurate m/z determinations of the precursor (2 IPs) and one MS2 fragment ion (2 IPs) and for four relative ion MS2 intensities (4 IPs).

Today, good or strong identification confidence of known-unknown compounds can be obtained from HRMS data without the need to have RRT ≤ ±2.5% and reference standard(s) injected in the same sequence (Chromatography Class A). Indeed, many IPs can be attributed allowing to accept a larger tolerance in chromatography constraints. The following Concrete examples will demonstrate this point.

Concrete Examples and Discussion

Theoretical m/z values, RIA, and FID have been calculated with Xcalibur software (Thermo Fisher Scientific). The first example is the putative identification of imatinib, an anticancer drug, in a human plasma extract by LC-HRMS analysis recording HR-full scan (no MSn). This unique analysis (Criterion I, Category #2 or further) was compared with a previous analysis performed 2 years earlier with roughly the similar LC column (C18) and conditions (gradient with MeCN:H2O with formic acid). The previous analysis was considered as the reference because pure imatinib standard was spiked and quantified in plasma.

RRT values (Criterion II) were calculated using an endogenous metabolite known to be abundant in plasma, LysoPC(18:2(9Z,12Z), which showed the same profile in both chromatograms (Figure 4a). RRT difference was –20.4% allowing to allocate the Chromatography Class C in the ID score. In this first example, 7 IPs were attributed (Criterion III, IP-level 3; Figure 4a-e). Table 3 and Figure 3 show how IP-level is attributed in relation to the number of IPs collected. In group i, 3 IPs were collected for the precursor ion identified as the [M + H+]+ because other co-eluting adducts were detected (Na+, K+, and 2H+; same RT) and because MA was ≤3 ppm and R ≥50,000 (Figure 4a). In group iii, 1 IP was obtained for RIA at M + 1 and M + 2 that matched imatinib RIAtheor with –8% and –17%, respectively (Figure 4b). In group vi, 3 IPs were collected as follows: 2 IPs for the detection of 2 known MS2 fragment ions but detected in the MS1 scan (in source fragmentation; MA <5 ppm, same RT as imatinib, Figure 4d) and 1 IP for S, Cl, and Si removals in proposed chemical formulae given by Chemspider (FID incompatibility) (Figure 4c) combined with the much higher number of references in PubMed for imatinib than further hits (according to Chemspider) (Figure 4e). Then an ID score of 2C3 (7 IPs) was established (good evidence in Table 1). More stringent acceptance on RIA errors could have been applied: RIA at A + 1 ≤ 5% (according to Knolhoff et al. [66]). This would have removed the attribution of 1 IP based on RIA even if the ID score had not changed. This underlines that the flexibility given to the analyst could possibly change the number of IP collected and the ID score but differences should remain moderate.

Figure 4
figure 4

Putative identification of imatinib found in a human plasma extract by LC-HRMS analysis. Seven IPs were collected (IP-level 3; see main text) establishing the ID score for imatinib identification at 2C3 (good evidence). (a) After the comparison with a similar analysis performed 4 y before and containing imatinib pure standard, the chromatography class was determined to be C (RRT with reference sample was ≤25%; IS for RRT calculation was lysoPC(18:2(9Z,12Z), m/z theor = 520.33977 extracted with ±5 ppm extraction window). Imatinib m/z was observed with H+, Na+, K+, and 2H+ adducts with MA ≤3 ppm and at R ≥50,000 (3 IPs). (b) Relative isotopic abundance (RIA; 1 IP) and (c) fine isotopic distribution (FID) allow discarding a few chemical composition possibilities (e.g., C ≈ 30, no Si, no Cl, no Br, no S). (d) Extracted ion chromatograms from a LC-HR full scan acquisition, depict imatinib and two known imatinib MS2 ions that were detected as in-source fragments (2 IPs). (e) In combination with FID, only C29H31N7O corresponding to imatinib is the most credible hit in large database (1 IP) because Chemspider shows >7000 links with PubMed whereas further possibilities show <40 hits

Full size image

The second example shows the putative identification of triclosan, an antibacterial and antifungal agent, found in a human plasma extract. Triclosan was detected in only one individual, whereas 50 anonymous patients were analyzed by LC-HRMS performing two simultaneous acquisitions: HR-full scan and MSALL (MS/MS with no precursor selection). No additional analyses were performed (Criterion I, Category #2 or further) and no RRT comparison with a reference sample was possible (Criterion II, Chromatography Class D). From the HRMS data, 10 IPs were collected allowing determination of an ID score of 2D2 (good evidence in Table 1).

Figure 5 depicts this identification and the number of IPs collected. In Group i (see Table 3), 2 IPs were collected for the precursor ion (not formally identified as [M – H+]- by other adducts). Whereas no IPs were attributed for MSn fragments (no fragment ions found in the MSALL scans), RIA allowed the collection of 2 IPs (Group ii). In fact, the chemical composition of triclosan, C12H7Cl3O2, shows a specific isotopic pattern (see spectra and XIC in Figure S-3d), whereas the low levels detected gave relatively poor RIA difference (up to 37% for lowest abundance ions) between measured and theoretical RIA (Rm/z = 54,000). In Group vi, 6 IPs were obtained in the additional information as follows. Two IPs were collected from the detection of two well-known plasma metabolites of triclosan: triclosan-glucuronide, [C18H15Cl3O8-H+]–, and triclosan-sulfate, [C12H7Cl3O5S-H+]–. One IP was given for a MSALL fragment ion corresponding to triclosan m/z but detected at the RT of the two triclosan metabolites (see Figure 5a), 2 IP were attributed for the Cl3 specific RIA of triclosan and its two metabolites (up to M + 7 for triclosan-sulfate, see Figure 5b and Supplementary Materials, Figure S-3d) and 1 IP was collected for the very low number of other credible hit(s) found in Chemspider and HMDB database. Searching with the unknown accurate m/z, only triclosan was returned from HMDB (all adducts in negative mode, MA ≤5 ppm) and considering C12H7Cl3O2, only triclosan shows references in PubMed, according to Chemspider (Figure 5c). This is in agreement with the ubiquitous use of triclosan in many consumer products in contrast to the other known isomers.

Figure 5
figure 5

Putative identification of triclosan in a human plasma extract by LC-HRMS analysis. Ten IPs were collected and corresponding to IP-level 2 (see main text) establishing the ID score for triclosan identification at 2D2 (good evidence). (a) Extracted ion chromatograms (HR full scan) depict triclosan (two IPs) and two known triclosan metabolites (triclosan-glucuronide and triclosan-sulfate with 50× more peak intensities (2x1 IPs, [M – H+]–; MA <5 ppm; not found in other plasma extracts). In the bottom chromatogram, a MSALL fragment ion corresponding to triclosan m/z can be observed at the two metabolite RT, 1 IP). (b) Measured relative isotopic abundance (RIA) of triclosan-sulfate, detected here up to M + 7 (2 IPs; see spectra and XIC in Figure S-3d). MA and RIA differences between theoretical and measured values (Rm/z = 54,000) are depicted and allow establishing the chemical composition, C12H7Cl3O5S; see spectrum in Figure S-3. Similar Cl3-specific RIA patterns were found for triclosan-sulfate and triclosan-glucuronide and, therefore, two additional IPs were attributed. (c) In Chemspider, triclosan is the most credible compound for C12H7Cl3O2 because it has a lot more links to various Sources than the other hits (1 IP)

Full size image

The third and last example shows the putative identification of caffeine in a human plasma extract. The ID score was 2D2 (good evidence in Table 1), which relates to one LC-HRMS analysis, performed with HR-full scan and MSALL acquisitions (Criterion I, Category #2), no chromatographic references (Criterion II, Chromatography Class D), and 10 IPs attributed (Criterion III, IP-levels 2). Two IPs were obtained from the precursor ion (MA <5 ppm; Figure 6a), 4 IPs were attributed from 4 known fragment ions of caffeine (see https://www.mzcloud.org or [74]), (1 IP per ion detected in MSALL scans; Figure 6b), 1 IP was attributed from RIA and 2 IPs were given by FID (Figure 6a and b). The last IP was attributed from the following analyst’s deduction: from the spectra at M + 1 and M + 2 (FID and RIA), a high probability that the unknown chemical composition contains N4, C8, and S0 was established (Figure 6). From Xcalibur used as formula generator, the only realistic hit for the accurate m/z measured was C8H11O2N4 +. In Chemspider, although more than 600 isomers have this formula, only caffeine and enprofylline have a significant number of links to PubMed, respectively, 19,779 and 239. The presence of the unknown compound in ≈ 30% of analyzed plasma adds evidence for the identification of caffeine rather than enprofylline or other isomers.

Figure 6
figure 6

Putative identification of caffeine ([C8H11O2N4]+) in a human plasma extract by LC-HRMS analysis. Ten IPs were collected (IP-level 2; see main text) establishing the ID score for caffeine identification at 2D2 (good evidence). (a) Extracted ion chromatograms (m/z theor ±5 ppm) depict the unknown compound (2 IPs; adduct not formally identified with other adducts) and three of its isotopes, Display (b) depicts 4 MSALL fragments known to be MS2 ions of caffeine (4x1 IPs; see https://www.mzcloud.org or [72]). Display (c) shows MA (ppm) and RIA difference (%; 1 IP) of caffeine isotopes between measured and theoretical values (Rm/z = 70,000). Spectra in (d) show the FID ([2] for consistency IPs) and the interest to resolve fine isotopes for chemical composition determination. Here, [13CC7H11O2N4]+ and [C8H11O2 15NN3]+, corresponding to M + 1 isotopes, are resolved, whereas, at M + 2, no 34S such as [C8H19O32S34S]+ at m/z = 197.0829 is observed. The fine discrimination of m/z isotopes together with the ion intensity (RIA) allow establishing that the unknown composition has N4, C8, and S0 and is eventually C8H10O2N4. (e) In Chemspider, caffeine is the most credible hit because the unknown was found in many plasma extracts and has >80× more links to PubMed in comparison to further hits (2 IPs)

Full size image

Concluding Remarks

With the increasing availability of HRMS in routine labs, new investigations will be realized using full scan or data-independent acquisition. The untargeted data treatment will reveal unexpected compounds of interest (pollutants, drugs, etc.) or unexpected concentrations. These revealed compounds will have to be identified for pertinent biological interpretations [preliminary diagnosis] and possible further targeted analyses. The injection of reference standards in the same analytical sequence and possibly in two orthogonal analyses can be mandatory for identifications with utmost certainty. But this is too time-consuming to inject 100s to 1000s of reference standards in the same analytical sequence for known-unknown identification. This is why an ID scale and score has been proposed to establish a base and a degree of confidence in known-unknown identification. Indeed, fair, good, or strong confidence can be established without the injection of reference standards when various identification points (IPs) can be collected from HRMS data. The three presented examples of identification have shown this point.

Based on two key references for screening and metabolomics labs [37, 38], a confidence scale and an ID score have been proposed for the rating of (non-peptidic) known-unknown identification in (LC-)HRMS analysis (Tables 1, 2 and 3) and are based on three criteria: I, a general identification category (1 to 4); II, a chromatography class (A to D); and III, an identification point level (1 to 3) attributed from the number of IPs collected. Therefore, the ID score is based on these three criteria and given as [a number, a letter, and a number].

This scale and ID score are new proposals and logically could/should be modified and improved. The purpose of these proposals is also an attempt to build a bridge between the metabolomics and screening lab communities. It is a step forward in the need to rate the confidence of putative identifications of known-unknown compounds with or without the availability of reference standards. The need to identify known-unknowns should grow with the increasing availability of HRMS in various routine labs and the further implementation of global and untargeted approaches to reveal biomarkers, monitor therapeutic outcomes, uncover acute or chronic intoxications, etc. The proposed ID scale, score, and confidence provide a base to these untargeted analyses.