Qualitative Analysis of Real Drug Evidence using DART-MS and the Inverted Library Search Algorithm

Abstract

Chromatographic-less mass spectrometry techniques like direct analysis in real time- mass spectrometry (DART-MS) are steadily being employed as seized drug screening tools. However, these newer analytical platforms require new computational methods to best make-use of the collected data. The inverted library search algorithm (ILSA) is a recently developed method designed specifically for working with mass spectra of mixtures collected with DART-MS, and has been implemented as a function in the NIST/NIJ DART-MS Data Interpretation Tool (DIT). This paper demonstrates how DART-MS and the ILSA/DIT can be used to analyze seized drug evidence, while discussing insights gathered during the evaluation of 92 adjudicated case samples. The evaluation verified that the combination of DART-MS and the ILSA/DIT can be used as an informative tool to help analysts screen seized drug evidence, but also revealed several factors, such as the influence of incorporating multiple in-source fragmentation spectra and the effect of scoring thresholds, an analyst must consider while employing these methods. Use cases demonstrating the benefit of the non-scoring metrics provided by the ILSA/DIT and demonstrating how the ILSA/DIT can be used to identify novel substances are also presented. A summary of considerations for using the ILSA/DIT for drug screening concludes this paper.

1. Introduction

The prevalence of new emerging drugs and opioids has led to increased caseloads and longer turnaround times for forensic laboratories [1, 2]. In response, many laboratories are looking for new approaches that can improve analysis efficiency. Typically, seized drug evidence is analyzed with a screening technique (e.g., color tests or gas chromatography-flame ionization detection) followed by a confirmation method (e.g., gas chromatography-mass spectrometry). One potential approach to improve the efficiency of seized drug analysis is to replace traditional screening techniques with ambient ionization mass spectrometry (AI-MS) platforms that can rapidly produce informative mass spectra [3].

A commonly used AI-MS platform in seized drug analysis is the combination of a direct analysis in real time (DART) ionization source coupled with a single quadrupole or time-of-flight mass spectrometer (MS) capable of collecting mass spectra at multiple in-source collision induced dissociation (is-CID) voltages [4]. To improve data interpretation, forensic laboratories will often collect mass spectra at multiple is-CID energies—one where there is typically minimal to no fragmentation (producing predominantly a protonated or deprotonated molecule), and additional ones where fragment ions can be observed. The resulting is-CID mass spectra can be difficult to interpret, as all ions are indiscriminately fragmented – unlike traditional tandem mass spectrometry where a specific ion is isolated and fragmented – leading to fragment ions that are not directly attributable to a parent ion.

With any MS platform, data interpretation can be greatly simplified with libraries of reference mass spectra and library-search algorithms [5]. We recently updated the NIST DART-MS Forensics Database (Reference Library) [6, 7] and developed the inverted library search algorithm (ILSA) [8, 9]. The updated Reference Library contains is-CID mass spectra of each compound measured at three fragmentation levels, to be consistent with most current practices in seized drug analysis [10]. The ILSA searches input query mass spectra for partial patterns that are similar to reference pure compound mass spectra contained in the Reference Library, an approach that is particularly useful when the query spectrum contains signatures of multiple compounds. The ILSA also simultaneously compares multiple is-CID mass spectra, collected at varying is-CID energies, thus providing a more comprehensive assessment of similarity between a set of query and library spectra. The ILSA has been implemented in the interactive NIST/NIJ DART-MS Data Interpretation Tool [11], colloquially known as the DIT.

In this paper, we detail how the ILSA/DIT can be used to interpret is-CID mass spectra of real seized drug evidence through demonstrative examples and discuss several general observations and relevant insights about the use of DART-MS and the ILSA/DIT for qualitative analysis of seized drug evidence. All is-CID mass spectra and source code used to generate the results presented in this paper are available in the Supplemental Information.

2. Methods

2.1. Materials and instrumental analysis

We obtained 92 adjudicated seized drug case sample extracts from the Maryland State Police Forensic Sciences Division. The samples contained between zero and six controlled substances in a mixture with as many as six cutting agents or diluents. Most of the samples were comprised of one or two controlled substances and up to one additional compound. Many of the samples contained opioids—as was expected given the high prevalence of these compounds in the region—though the sample set also contained at least one instance of amphetamines, arylcyclohexylamines, benzodiazepines, cocaine, lysergamides, opiates, steroids, synthetic cathinones, synthetic cannabinoids, and tryptamines. A complete list of the sample contents can be found in Supplemental Table 1.

Samples were provided as methanolic solutions and/or suspensions, with approximate overall concentrations of 1 mg/mL. We were also provided with gas chromatography-mass spectrometry (GC-MS) datafiles for each sample. To confirm the sample compositions we ran all raw datafiles through the Automated Mass Spectral Deconvolution and Identification System (AMDIS) [12] using a Simple analysis, the SWGDRUG mass spectral library (version 3.9), a component width of 32 arbitrary units (a.u.), medium resolution, high sensitivity, and medium shape requirements while allowing for multiple identifications per compound. This was completed to attempt to identify the maximum number of compounds in a sample and may not be representative of common reporting practices in the field.

For each case sample extract, we collected is-CID mass spectra at three orifice 1 voltages—low (+30 V), mid, (+60 V), and high (+90 V)—using a JEOL AccuTOF LC-4G mass spectrometer (JEOL USA, Peabody, MA) coupled with an IonSense (Saugus, MA) DART-SVP source. The three is-CID mass spectra were collected using the parameter switching option with a cycle time of 0.2 s cycle⁻¹. The MS was operated in positive ionization mode, with an orifice temperature of 120 °C, an orifice 2 voltage of +5 V, a ring lens voltage of +5 V, an RF ion guide voltage of +700 V, and a mass scan range of m/z 80 to m/z 800. The DART was operated in positive ionization mode, using high purity helium as the source gas with a source gas temperature of 400 °C, and an exit grid voltage of 150 V. A methanolic solution of the synthetic cannabinoid AB-FUBINACA, with an approximate concentration of 0.25 mg mL⁻¹, was used as a positive control and polyethylene glycol (PEG-600) was used for mass calibration. Samples were introduced to the DART gas stream using glass microcapillary rods (Corning, NY) that were previously heated in a GC oven at 300 °C overnight to remove plasticizers and other contaminants. Data for each sample was collected in a 1 min analysis with an introduction method of blank capillary, AB-FUBINACA, Sample Replicate #1, Sample Replicate #2, Sample Replicate #3, and AB-FUBINACA. A multi-point mass drift compensation was completed within the JEOL instrument software (msAxel) using the base peak of AB-FUBINACA (m/z 352.1456), and then a single averaged and background subtracted mass spectrum of the sample from the three replicates was extracted for each is-CID voltage.

2.2. Data analysis

The is-CID mass spectra of all case samples were analyzed using the ILSA [9] and version 5 of the NIST DART-MS Forensics Database (Earthworm) as the Reference Library. A first-pass automated analysis was done with a command line implementation of the ILSA, intended to help us generate general observations. Interesting examples from the first pass analysis were further analyzed interactively with the DIT (version 2.0).

The ILSA can be summarized as a three-step search that (i) identifies target m/z values with a relative intensity greater than a user-defined targeting threshold in the low-fragmentation is-CID mass spectrum that could represent potential compounds in the mixture, (ii) searches the library for compounds that could be matches to the target m/z given a user-defined mass tolerance, and (iii) compares the library spectra of the potential matches to the query spectra. Step 3 produces several scores and metrics that can be used to guide an analyst in deciding whether potential library matches are components of the query mixture. For all searches discussed herein, we used a mass tolerance of ±0.005 Da and a targeting threshold of 1% relative intensity.

2.2.1. Command line ILSA for automated evaluation

To conduct a rapid first-pass evaluation of how well the ILSA performs at identifying known compounds in seized evidence, we wrote a simple command line application that searched all 92 case samples using specified sets of is-CID mass spectra: (i) only the low-fragmentation mass spectrum, (ii) all available is-CID mass spectra, (iii) the low- and high-fragmentation mass spectra, and (iv) the low- and mid-fragmentation mass spectra. Each search produces an Excel workbook containing the search results. Each workbook contains a separate worksheet for every identified target m/z value. If the target has potential library matches, the worksheet contains the library compounds’ name, generic class (e.g., Fentanyls), the spectral similarity scoring metrics at each is-CID level searched, the average similarity scores, and other metrics described in [9]. The spectral similarity scoring metrics are average FPIE (fraction of peak intensity explained) and average RevMF (reverse match factor) as described in [8, 9]. The results also include the Δm/z, which is the difference between the theoretical and observed m/z values, and the IRD (isotope ratio difference) which measures, for each compound, the similarity of the ratio of the observed protonated molecule and its major isotope to the theoretical ratio. If the target does not have any potential matches in the library, the worksheet states this plainly.

To objectively decide whether the results for any given target would provide beneficial or misleading information about the query mixture, we developed the assessment scheme shown in Figure 1. For target m/z values that had potential matches in the library, the target result was categorized as either Ideal, Desirable, Acceptable, Non-Descriptive, or Undesirable, based on the known identity of the compounds in the mixture, as determined by GC-MS, the number of potential library matches as well as their class, and the average FPIE or average RevMF scores of the potential library matches (see section 2.1). For example, if one of the target m/z values has a single potential library match with a score greater than the decision-making threshold, but neither that compound nor another compound of the same class was determined, by GC-MS, to be in the mixture, that result is deemed Undesirable. Here we only report result categories determined with an arbitrary 0.7 a.u. decision-making threshold. An Excel Pivot Table containing complete results for a variety of decision-making thresholds is provided in the Supplemental Information. Example results to better explain each result category are provided in the Supplemental Information.

Figure 1:

Target result assessment scheme. The number in parenthesis shows the desirability order of these result categories (1 being the most desirable).

2.2.2. NIST/NIJ DART-MS Data Interpretation Tool (DIT) for interactive evaluation

The DIT is a recently developed application for interacting with a is-CID Reference Library and searching query mass spectra using the ILSA. It was developed in collaboration with several local, state, and federal forensic laboratories to ensure it contained functions and features of utility to practicing forensic chemists [11].

An example search-result generated by the DIT is shown in Figure 2. The DIT requires users to upload at least a low-fragmentation is-CID mass spectrum, but allows for up to three (low-, mid-, and high-fragmentation) is-CID mass spectra. The DIT implementation of the ILSA reports the average FPIE and average RevMF for all potential compounds identified for each target. The results also include the Δm/z, the IRD, and the Match Type, which explains how the Target m/z can be explained by the library compound or its low-fragmentation is-CID mass spectrum. In the DIT version employed in this paper (version 2.0), the potential Match Types are protonated molecule, base peak in library spectrum, major isotope of the protonated molecule, major isotope of the base peak, and major fragment ion in library spectrum. We should note that the protonated molecule and the library spectrum base peak are identical for most drugs in the library.

Figure 2:

Snapshot of the results of Sample 2 provided by the DART-MS Data Interpretation Tool (version 2.0).

3. Results and discussion

The complete results for all performed ILSA searches (92 case samples with four is-CID configurations), as well as the automated interpretations of the search results using the classification scheme (Figure 1) with several decision-making thresholds are provided in the Supplemental Information.

3.1. Example searches

In this section, we describe a few interesting examples that demonstrate how the results from the ILSA/DIT can be used to interpret the is-CID mass spectra of case samples and illustrate a few known limitations of the current implementation of the algorithm.

3.1.1 –. Sample 13: Complex mixture containing fentanyl and fluorofentanyl – the role of multiple is-CID spectra

Consider the low-fragmentation (+30 V) is-CID mass spectrum of a case sample extract (Sample 13) shown in Figure 3. The colored peaks (blue and red) indicate the m/z values with a relative intensity of at least 1%, identified as potential targets in step 1 of the ILSA. There are 34 targets identified, 17 of which have potential matches in the library (peaks marked in red in Figure 3). In Table 1, we present the 17 targets with select potential matches along with the category of the results (based on Figure 1), assuming a decision-making threshold of 0.7 a.u., when using either the low-fragmentation spectrum or all three is-CID spectra. Supplemental Tables S2 and S3 contain condensed lists of all potential matches from the DART-MS Library.

Figure 3:

Low-fragmentation is-CID mass spectrum of sample 13. Thirty-four target peaks (colored) were identified using a relative intensity threshold of 1%. Targets marked in red have potential matches in the library. Targets in blue do not have potential matches in the library. Inset is a section of the spectrum containing the top-10 targets by relative intensity, all with potential matches in the library.

Table 1:

Summary results for Sample 13 using only the low-fragmentation is-CID mass spectrum and all three is-CID mass spectra. Only targets that had a match in the database are listed. For targets where there are more than one possible compound, only the compound that was present in the mixture is shown. Categorization of results was completed using a score threshold of 0.7 a.u. Note that this extract contained tetracaine as an internal standard. MF indicates major fragment ion. Supplemental Tables S1 and S2 list the complete search results.

DIT Target #	Target m/z	Relative Intensity (%)	Peak Identification	FPIE Result (low)	RevMF Result (low)	FPIE Result (low, mid, high)	ReMF Result (low, mid, high)
1	337.2285	100.0	Fentanyl	Desirable	Desirable	Desirable	Desirable
2	265.1917	98.1	Tetracaine	Ideal	Ideal	Ideal	Ideal
3	114.0658	28.8	Creatinine	Ideal	Ideal	Non-Descriptive 2	Non-Descriptive 2
4	338.2318	26.2	Isotope match of Target 1	Desirable	Desirable	Desirable	Desirable
5	195.0872	22.0	Caffeine	Ideal	Ideal	Non-Descriptive 2	Ideal
6	266.1948	17.6	Isotope match of Target 2	Ideal	Ideal	Ideal	Ideal
7	343.0792	12.2	Etizolam	Ideal	Ideal	Ideal	Ideal
8	235.1801	10.2	Lidocaine	Ideal	Ideal	Non-Descriptive 2	Ideal
9	183.0860	8.1	Mannitol	Non-Descriptive 2	Ideal	Non-Descriptive 2	Non-Descriptive 2
10	165.0758	5.1	MF Ion (Mannitol)	Non-Descriptive 2	Ideal	Non-Descriptive 2	Non-Descriptive 2
12	345.0765	4.6	Isotope match of Target 7	Ideal	Ideal	Ideal	Ideal
14	176.1066	3.9	MF Ion (Tetracaine)	Ideal	Ideal	Ideal	Ideal
16	281.2004	3.7	4-ANPP	Undesirable†	Non-Descriptive 1	Non-Descriptive 1	Non-Descriptive 1
21	196.0898	2.3	Isotope match of Target 5	Ideal	Ideal	Non-Descriptive 2	Ideal
25	236.1831	1.6	Isotope match of Target 8	Ideal	Ideal	Non-Descriptive 2	Ideal
29	115.0673	3.3	Isotope match of Target 3	Ideal	Ideal	Non-Descriptive 2	Non-Descriptive 2
34	355.2195	1.1	ortho-fluorofentanyl	Acceptable	Desirable	Desirable	Desirable

^*Phenacetin was also in the mixture but was present below 1% relative intensity.

^†Impramine was identified as the only potential target with a score above 0.7 a.u. for this Target m/z.

Based on GC-MS analysis, the components of Sample 13 were fentanyl, etizolam, 4—ANPP, ortho-fluorofentanyl, creatinine, phenacetin, mannitol, caffeine, and lidocaine. One of the 9 compounds (phenacetin) presented a protonated molecule that was below our targeting and “noise” threshold of 1% relative intensity, leading to 8 identifiable mixture components. Using FPIE scores as a decision-making metric and considering only the low-fragmentation mass spectrum (Table 1, FPIE Result (low)) led to five of the eight identifiable mixture components being categorized as an Ideal or Desirable result using the ILSA. It is important to note that although multiple targets can have the same potential library matches (e.g., Target 1 and Target 4 both correlate to fentanyl), scores are computed for a compound and so will be identical across targets. Of the other three compounds, the protonated molecules of mannitol, 4-ANPP, and ortho-fluorofentanyl all had computed FPIE scores below the decision-making threshold of 0.7 a.u. In the case of mannitol (Targets 9 and 10), the search result was considered a Non-Descriptive 2 result (no compounds with scores greater than the threshold). Low scores for mannitol are often observed and can likely be attributed to a combination of low ionization efficiency and significant fragmentation, even under the low-setting. With ortho-fluorofentanyl (Target 34), the result was considered Acceptable as other fluoro-containing fentanyl analogs appeared with scores greater than the threshold and, when used for screening, would have alerted a drug chemist to the fact that a fentanyl analog was present. The target result with 4-ANPP (Target 16) was considered undesirable as the FPIE score for imipramine—a compound with a different class designation in the Reference Library—was greater than 0.7 a.u. Using RevMF scores as a decision-making metric (Table 1, RevMF Result (low)) led to seven of the eight identifiable mixture components being categorized as Ideal or Desirable. With this scoring metric, the result for 4-ANPP (Target 16) was Non-Descriptive 1 as both 4-ANPP and imipramine had RevMF scores greater than 0.7 a.u.

If all three is-CID spectra were incorporated into the ILSA search, fentanyl, etizolam, 4—ANPP, and ortho-fluorofentanyl are categorized as Ideal or Desirable results using either FPIE or RevMF scores. Interestingly, none of the cutting agents (mannitol, lidocaine, creatine, and caffeine) had FPIE scores greater than 0.7 a.u, and only caffeine had a RevMF greater than 0.7 a.u. Using all three is-CID spectra, there were no undesirable results.

From this example, we can see the effect of using the is-CID spectra at different fragmentation levels on computed scores and subsequent decision making. The added spectra removed any false positive identifications, but also dropped the computed scores for most cutting agents below the decision-making threshold of 0.7 a.u., leading to five false negatives (as opposed to one false negative using only the low-fragmentation spectrum and RevMF). Since DART-MS is traditionally used as a screening tool to compliment confirmatory GC-MS analysis, false negatives for cutting agents and diluents may be more acceptable than false positives for controlled substances. It should be noted that the decision-making threshold of 0.7 a.u. was selected arbitrarily. If we had used the three is-CID spectra configuration and a lower decision-making threshold, we may have still been able to remove the undesirable result (false identification) while not missing true mixture components. Similarly, it is conceivable that we could have prevented undesirable results while searching only the low-fragmentation spectra by adjusting the decision-making threshold.

3.1.2 –. Sample 59: A novel cathinone – the limitations of library searching

As a second example, consider the low fragmentation is-CID spectrum of another case sample extract (sample 59) shown in Figure 4. The primary target identified in this spectrum is m/z 250.1442, which matched well with several of the cathinones (molecular formula of C₁₄H₁₉NO₃) available in the database. Based on the GC-MS results, the true match for this peak was N-ethyl pentylone—a compound not in the version of the database used for this analysis. Accordingly, the automated assessment scheme deemed this result as undesirable—a novel compound cannot be identified and will not be associated with a general class of compounds. In practice, however, achieving this result with a novel drug might be deemed desirable, if not ideal. The results of searching the is-CID mass spectra of this sample are shown in Table 2. Using only the low-fragmentation spectrum, one might surmise that a synthetic cathinone with a number of isomers is present in the sample. This result would provide a good starting point for further exploration. By including all three is-CID spectra in the search, the results (i) further support the assumption that a synthetic cathinone is present in the sample and (ii) begin to limit the possible structure of that cathinone based on poor FPIE or RevMF scores. For example, the low scores for 3’,4’-Methylenedioxy-N-tert-butylcathinone may suggest that a tert-butyl group is not present on the unknown cathinone. While “not in library” will remain a common limitation of mass spectral library-searching of new compounds, a comprehensive library can still be used to make a reasonable inference on the structure of the unknown. It should be noted that identification of the novel cathinone was possible only because other cathinones with the same protonated molecule were in the Reference Library. Given the targeted approach of the ILSA, identification of new compounds at m/z values not in the Reference Library is not possible.

Figure 4:

Section of low-fragmentation is-CID mass spectrum of Sample-59. Inset is a section of the spectrum containing the top-2 targets by relative intensity, all with potential matches in the library.

Table 2:

Search results for Sample 59 using the low-fragmentation mass spectrum and all available is-CID mass spectra. Since the actual compound was not in the library these results were categorized as undesirable.

DIT Target #	Target m/z	Relative Intensity (%)	Peak Identification	FPIE Score (low)	RevMF Score (low)	FPIE Score (low, mid, high)	ReMF Score(low, mid, high)
1	250.1442	100.0	3’,4’-Methylenedioxy-N,N-diethylcathinone	0.892	0.991	0.673	0.655
			3’,4’-Methylenedioxy-N-tert-butylcathinone	0.752	0.959	0.649	0.563
			3’,4’-Methylenedioxy-α-dimethylamino-Isovalerophenone	0.795	0.984	0.810	0.721
			3,4-Methylenedioxy-α-methylaminohexanophenone	0.840	0.981	0.845	0.851
			3,4-Methylenedioxy-α-methylaminoisohexanophenone	0.841	0.985	0.935	0.713
			3’4’-methylenedioxy-α-ethylamino-Isovalerophenone	0.812	0.993	0.743	0.853
			N,N-Dimethylpentylone	0.859	0.993	0.862	0.950
2	251.1464	27.8	Isotope Match of Target 1	--	--	--	--

3.1.3. Sample 46 – A fentanyl and tramadol mixture – the role of other ILSA metrics and the impact of poor fragmentation

As a final example, we consider Sample 46 which contains fentanyl, tramadol, caffeine, levamisole, mannitol, phenylpropamide, and procaine. The low-fragmentation mass spectrum of the sample (Figure 5) contains 35 targets at the 1% relative intensity threshold, with 15 of the targets having potential matches in the library (Table 3). Using only the low-fragmentation data, procaine and caffeine produced ideal results, and fentanyl produced a Desirable result. In this example, the observed accurate mass of the isotope peak for fentanyl (nominal m/z 338, Target 5) was within mass tolerance of the protonated molecule of metodesnitazene. The scores for metodesnitazene were high which may lead to an analyst to believe this compound is also present in the mixture, and this conjecture would be further supported by the matching of the metodesnitazene isotope at nominal m/z 339 (Target 18). The inclusion of all is-CID mass spectra did not change overall result category as scores for fentanyl and metodesnitazene remained above 0.7 a.u.

Figure 5:

Section of low-fragmentation spectrum of sample 46.

Table 3:

Search results for Sample 46 using only the low-fragmentation is-CID mass spectrum (top row of results) and all three is-CID mass spectra (bottom, italicized row of results). Only compounds that were present in the mixture are shown, along with the result category for the FPIE and RevMD results, using a score threshold of 0.7 a.u. The Δm/z and isotope ratio difference (IRD) are also provided.

DIT Target #	Target m/z	Relative Intensity (%)	Peak Identification	Δm/z (Da)	IRD	FPIE Result	ReMF Result
1	237.1597	100.0	Procaine	−0.0006	0.060	Ideal	Ideal
						Ideal	Ideal
2	337.2250	71.0	Fentanyl	−0.0030	0.034	Desirable	Desirable
						Desirable	Desirable
3	205.0780	47.4	Levamisole	−0.0020	0.023	Ideal	Ideal
						Ideal	Ideal
4	238.1617	20.1	Isotope of Target 1	−0.0020	0.060	Ideal	Ideal
						Ideal	Ideal
5	338.2282	19.3	Isotope of Target 2 -or- Not in Mixture	−0.0032 0.0049	0.034 −0.097	Non-Descriptive 1	Non-Descriptive 1
						Non-Descriptive 1	Non-Descriptive 1
6	249.1584	13.4	Not in Mixture	−0.0019	0.011	Undesirable‡	Undesirable‡
						Non-Descriptive 2	Non-Descriptive 2
7	264.1941	10.0	Tramadol	−0.0023	0.024	Non-Descriptive 2	Ideal
						Non-Descriptive 2	Non-Descriptive 2
8	150.0894	8.0	Not in Mixture	−0.0025	−0.097	Non-Descriptive 2	Undesirableϟ
						Non-Descriptive 2	Non-Descriptive 2
9	206.0808	6.7	Isotope of Target 3	−0.0025	0.023	Ideal	Ideal
						Ideal	Ideal
16	195.0857	3.1	Caffeine	−0.0025	−0.087	Non-Descriptive 2	Ideal
						Non-Descriptive 2	Ideal
18	339.2312	2.5	Not in Mixture (Isotope of Target 5)	0.0046	−0.097	Undesirable†	Undesirable†
						Undesirable †	Undesirable †
21	118.1196	2.3	MF ion of Procaine	−0.0036	0.060	Ideal	Ideal
						Ideal	Ideal
23	250.1619	2.2	Isotope of Target 6	−0.0018	0.011	Undesirable‡	Undesirable‡
						Non-Descriptive 2	Non-Descriptive 2
24	265.1969	2.0	Isotope of Target 7	−0.0028	0.024	Non-Descriptive 2	Ideal
						Non-Descriptive 2	Non-Descriptive 2
28	136.0742	1.9	Not in Mixture	−0.0021	0.034	Undesirable	Undesirable
						Non-Descriptive 2	Undesirable

^†Metodesnitazene was identified as another potential target, unrelated to fentanyl, with a score above 0.7 a.u. for this search.

^‡Pindolol was identified as the only potential compound for Target 6 with a score above 0.7 a.u. for this search.

^ϟCathinone was identified as the only potential compound for Target 8 with a score above 0.7 a.u. for this search.

This example highlights the potential complication of relying solely on similarity scores to draw conclusions and the value of the additional metrics reported by the ILSA/DIT. For every potential library match in the search results, the ILSA/DIT reports a Δm/z value and an isotope ratio difference (IRD). The reported Δm/z is the difference between the theoretical accurate mass of the library compound and the observed m/z value of the target. The IRD describes how different the observed isotopic envelope is from the theoretical envelope for the library compound in the low fragmentation mass spectra. All of the compounds that were present in the mixture had Δm/z values ranging from −0.006 Da to −0.0036 Da, indicating that the mass calibration was off slightly in the negative direction for that measurement. In the case of metodesnitazene, the reported Δm/z value was +0.0049 Da, just within the specified ±0.005 Da mass tolerance. For high resolution measurements such as this, a Δm/z that is in the opposite direction of prominent matches to other targets is a good indication that the potential library match is not in the mixture, and so we can surmise that metodesnitazene is not in the sample.

This assertion is further supported by considering the IRD for fentanyl (Target 2). IRD values range from −9.99 to +9.99 where 0 indicates an exact match between the theoretical isotopic distribution and the observed isotopic distribution. IRD values that skew positive would indicate the intensity of the major isotope is higher than expected (and therefore may be influenced by the presence of another compound or fragment ion with the same or near identical m/z to the major isotope) while values that skew negative would indicate the relative intensity of the major isotopic peak is lower than expected (and therefore may be influenced by the presence of another compound or fragment ion approximately 1 Da less than the compound of interest). In this example, the IRD for fentanyl is close to zero (0.034), which indicates that the major isotope at nominal m/z 338 is not likely to be influenced by the presence of a protonated molecule of a second compound.

A final consideration is that the m/z 339 (Target 18) could be explained as an M+2 isotope of fentanyl, but this functionality is currently not implemented in the ILSA/DIT. Therefore, even though the FPIE and RevMF scores for metodesnitazene seem reasonable, the additional metrics and an aggregate evaluation of all mixture targets can help rule out that presumptive identification.

Another noteworthy occurrence in this example is the change in classification of tramadol (Target 7) when including additional fragmentation spectra; the result is deemed Ideal using only the low fragmentation mass spectrum and RevMF but becomes a Non-Descriptive 2 result when all fragmentation spectra are considered. While changes in result classification as a function of the configuration of is-CID spectra considered in the ILSA are not uncommon (see cutting agents in example 3.1.1.), this example points to another limitation of the ILSA as currently implemented. The low-fragmentation mass spectrum of tramadol has a prominent protonated molecule (nominal m/z 264) and another major fragment ion with nominal m/z 58. The higher-fragmentation mass spectra of tramadol are dominated by the m/z 58 peak. However, the ILSA only considers peaks greater than m/z 80 as a lower limit where computing similarity scores. Accordingly, scores computed using higher fragmentation mass spectra of tramadol are being computed on noise peaks. This is an important limitation of scoring incorporated in the ILSA and should be considered when matching spectra that have major fragment ions less than m/z 80.

3.2. General observations from automated result interpretations

As noted earlier, we conducted a batch search of all 92 case samples with 4 input spectral configurations and performed an objective analysis of the search results using ‘ground truth’ data from GC-MS and the interpretation scheme shown in Figure 1. Given that our case sample dataset is small and fairly specific, it is difficult to draw any general conclusions about the performance of the ILSA/DIT for all drug screening applications. However, there were some insights gathered through the automated evaluation study that may prove generalizable with additional testing.

3.2.1. Effect of number of is-CID spectra

The number of is-CID spectra to include in an ILSA search can affect the interpretability of the result. The low-fragmentation spectrum is typically dominated by protonated molecule and will generally score well with most potential library matches within mass tolerance of the target. By including higher fragmentation is-CID spectra, the scores consider additional information which can lead to better discrimination between potential library matches for a given target. However, the inclusion of additional fragmentation spectra can also complicate interpretation if not considered with care due to low relative abundances, lack of specific fragmentation, major fragment ions being present below the mass scan range, or other phenomena that occur during the measurement of complex mixtures.

To better understand the effect of including additional is-CID mass spectra, the classification of all targets in the 92 mixtures (excluding those labeled as not in mixture, not in library, and tetracaine) were compared when the ILSA search was completed using just the low-fragmentation spectrum and when it was completed using the low-fragmentation spectrum plus one or more higher fragmentation spectra. Depending on the configuration, between 66% and 83% of all objective target classifications remained unchanged when additional is-CID spectra were included. This was likely largely driven by targets that correlated to compounds with a unique database match (where an Ideal result remained Ideal) or targets with a database match consisting of only compounds from the appropriate class (where a Desirable result remained Desirable). Out of the target identifications (487 per configuration and score combination) from these 92 case samples, the target classification did not change for any instance of arylcyclohexylamines, benzodiazepines, lysergamides, opioids (non-fentanyl), and steroids in any combination of configuration and score. While this was the observation in this set of samples, this result may not be universally applicable to these classes especially since they often represent a single compound or small number of compounds in each class. Also, this result only applies to the objective, automated classification, and does not consider the additional data that would be provided to a drug chemist making a subjective identification.

Depending on the configuration between 10% and 21% of target classifications improved with the inclusion of additional fragmentation data and between 7% and 14% of target classifications worsened as a result. The opiate class saw the highest number of improvements – dominated by Non-Descriptive 1 results being reclassified as Ideal due to the additional information provided by one or more is-CID mass spectra. The objective classification of cocaine often improved (from Non-Descriptive 1 to Ideal) with the addition of higher fragmentation spectra that allowed for differentiation from Scopolamine. Synthetic cannabinoids and tryptamines did not show a regression in classification, but the additional data did help elevate some Non-Descriptive results to either Desirable or Ideal. Table 4 shows the number, and types, of improvements and regressions seen when comparing the low-fragmentation classification to the classification obtained when using all three is-CID spectra.

Table 4:

Number of total changes in classification when comparing the classification obtained for targets using only the low-fragmentation mass spectrum and when using all the is-CID mass spectra. Values above the diagonal indicate an improvement in the classification while values below the diagonal (bolded and italicized) indicate a drop in classification. In each cell, the left column indicates the number of changes when using FPIE as the scoring metric (along with a 0.7 a.u. scoring threshold) and the right column indicates the number of changes when using RevMF as the scoring metric.

		Low-fragmentation only classification
		Ideal		Desirable		Acceptable		Non-Descriptive 1		Non-Descriptive 2		Undesirable
All is-CID spectra classification	Ideal			10	9	0	3	58	66	0	1	0	1
	Desirable	2	0			3	0	6	13	01	0	0	0
	Acceptable	0	0	4	5			0	0	0	0	0	0
	Non-Descriptive 1	12	0	0	0	0	0			0	2	1	0
	Non-Descriptive 2	29	45	1	0	0	1	4	11			4	5
	Undesirable	0	0	0	0	0	0	4	4	0	0

Drops in the objective classification from Ideal to Non-Descriptive 2 were the most common decrease. This was largely due to the score of the actual target dropping below the scoring threshold, leading to targets with no potential compounds above the threshold. This most frequently occurred with cutting agents and/or diluents, and scores typically dipped just below the 0.7 a.u. threshold. There were also four instances where the addition of all is-CID mass spectra turned a Non-Descriptive 1 result into to an undesirable result when using the RevMF score.

3.2.2. Understanding scores

A common question that arises in all mass spectral library search applications is “What is a good score?” and there is no correct general answer to that question. And as noted in the previous section, the choice of input is-CID configuration can affect the computed scores and interpretations. A summary of all computed scores for compounds known to be in the data set (via GC-MS) is provided in Table 5. The compounds are only attributed to the first target with which they match in a search result. For example, if fentanyl matches as a protonated molecule for Target 1 (relative intensity 100%) in the mixture spectrum and as a major fragment for Target 5 (relative intensity 15%), we are only including its score with the match to Target 1 in Table 5.

Table 5:

Summary of average spectral similarity scores for known compounds in mixtures across different is-CID configurations, disaggregated by the relative intensity of its target peak in the low fragmentation mixture mass spectrum.

Target Intensity (%)	Number of occurrences	Average FPIE				Average RevMF
		Low	Low Mid	Low High	Low Mid High	Low	Low Mid	Low High	Low Mid High
Base Peak [100]	78	0.890	0.886	0.892	0.888	0.965	0.943	0.943	0.936
High [50,100)	39	0.870	0.868	0.877	0.873	0.970	0.960	0.938	0.943
Mid [10,50)	73	0.807	0.771	0.767	0.756	0.928	0.888	0.825	0.832
Low [5,10)	28	0.758	0.698	0.668	0.658	0.801	0.761	0.648	0.672
Ultra-low [1,5)	31	0.648	0.592	0.599	0.578	0.765	0.730	0.644	0.661

There are two general trends we can note. First, computed scores are generally higher for higher intensity targets. This is expected given that the mixture spectrum of higher intensity targets is likely to present more of the ions necessary to match well with the pure compound library spectra. Second, for similar reasons, the effect of the input is-CID configuration to the ILSA is more pronounced with the lower intensity target matches. The higher fragmentation mixture mass spectra of these lower intensity targets are less likely to present ions that match well with the higher fragmentation mass spectra of the pure compounds in the library.

3.2.3. Targeting thresholds

A final observation worth mentioning is that the majority of compounds that were identified with GC-MS had target intensities greater than 10% (see Table 5) in the low-fragmentation DART-MS spectra. While this mostly reflects the nature of dataset, where mixtures usually only had 2 to 3 compounds, it is an interesting parameter to explore when using the ILSA/DIT. Selecting a lower targeting threshold will minimize the opportunity to miss on a mixture compound (false negative) but will increase the opportunity of matching compounds that are not in the mixture (false positive). It is worth noting that there were 30 instances when a known compound in the mixture would have not been identified if the targeting threshold was selected as greater than 5%.

3.2.4. Key takeaways

For the convenience of readers interested in using the ILSA/DIT to interpret is-CID mass spectra of seized drugs, we use this section of the paper to share our most important observations from this evaluation. Most of these considerations are not necessarily specific to the ILSA/DIT and may be applicable to other types of mass spectral library searching and chromatographic-less mass spectrometry.

• The ILSA has been successfully employed to analyze both simple (pure compound) and complex mixture spectra of actual seized drug samples.

• As with any mass spectral library-search based analysis, outcomes are highly dependent on the quality of the Reference Library. If a compound is not included in the library, it cannot be directly identified. However, a comprehensive library with spectra of structurally similar compounds can still guide inferences about structures of unknowns.

• For mixture components that produce ions with low relative intensities in the is-CID mass spectra, the FPIE score will, generally, be more forgiving than the RevMF as it does not consider the relative abundance of peaks.

• Making decisions strictly based on numerical values of scoring metrics is not advisable. The “optimal” decision-making threshold will likely vary as a function of the (i) scoring metric, (ii) input spectral configuration (i.e., searching low-fragmentation mass spectra only versus searching all is-CID mass spectra), and (iii) classes of compounds. If automated screening is required, the choice of a threshold should be based acceptability of false positive identifications (using a lower threshold) or false negative identifications (using a higher threshold) in the analysis. The same considerations should be made when choosing the targeting relative intensity threshold. For example, a 1% relative intensity threshold will likely have a higher possibility of false positive identifications than a 5% relative intensity threshold.

• The Δm/z value may be useful in ruling out potential compounds in a mixture when using high-resolution mass measurements. The Δm/z values will, generally, trend in one direction across the m/z range with true mixture components. Library matches with Δm/z values that deviate from that trend are unlikely to be true mixture components.

• The IRD value may be useful in ruling out potential compounds in a mixture when a target has multiple potential library matches, where some are explained as protonated molecules or base peaks and others are major isotope matches from targets with higher relative intensities. IRD values close to zero suggest that the potential library match explained as a major isotope is accurate since the isotopic distribution observed in mixture mass spectrum follows the theoretical distribution of the library compound. An IRD value significantly greater than zero suggests that an additional compound, matching the target as a protonated molecule or base peak, is present in the mixture.

• Compounds that do not fragment into any observable ions may have much lower FPIE or RevMF scores when higher level is-CID spectra are included in the search.

• Competitive ionization should be considered when trying to reconcile differences between DART-MS and GC-MS results. Compounds present at low concentrations in the mixture and/or with low proton affinities (e.g., heroin and noscapine) may not be detectable by DART-MS when in the presence of high proton affinity compounds (e.g., fentanyl, cocaine, and xylazine).

• The inclusion of major isotopes, base peaks, and major fragment ions in the ILSA and DIT increases the explainability of peaks in mixture spectra.

• Using GC-MS data as ground truth may not be ideal as GC-MS is, generally, less sensitive and may not be able to reliably identify low concentration components that are detectable by DART-MS. However, when dealing with real seized drug samples, ground truth may never be known.

• At present, we highly recommend interacting with the search results as the aggregate evaluation of all reported metrics across all targets in the mixture can often provide significant insights as shown in the examples—automated evaluation with decision-making thresholds remains unreliable.

4. Conclusions

The use of the ILSA, and the DIT, for qualitative seized drug screening was demonstrated through example analyses and automated evaluations of real case samples. The ILSA/DIT was largely successful in identifying the components of these samples, though there are some known limitations that have been documented. To summarize, using the ILSA/DIT with DART-MS can be a useful approach for identifying mixture components even with complex composition like seized drug evidence. There is still work to be done in determining preferable parameters and configuration, as well as continuing to build upon the library, but the underlying method has been shown to identify most major mixture components and may be a reliable screening technique in seized drug analysis.