In Silico Collision Cross Section Calculations to Aid Metabolite Annotation

Abstract

The interpretation of ion mobility coupled to mass spectrometry (IM-MS) data to predict unknown structures is challenging and depends on accurate theoretical estimates of the molecular ion collision cross section (CCS) against a buffer gas in a low or atmospheric pressure drift chamber. The sensitivity and reliability of computational prediction of CCS values depend on accurately modeling the molecular state over accessible conformations. In this work, we developed an efficient CCS computational workflow using a machine learning model in conjunction with standard DFT methods and CCS calculations. Furthermore, we have performed Traveling Wave IM-MS (TWIMS) experiments to validate the extant experimental values and assess uncertainties in experimentally measured CCS values. The developed workflow yielded accurate structural predictions and provides unique insights into the likely preferred conformation analyzed using IM-MS experiments. The complete workflow makes the computation of CCS values tractable for a large number of conformationally flexible metabolites with complex molecular structures.

Graphical Abstract

INTRODUCTION

The metabolome is the total collection of biologically active small molecules with molecular weights lower than about ~1.5 kDa.^1,2 This includes endogenous molecules that are biosynthesized by metabolic networks in "primary metabolism", specialized "secondary metabolite" signaling, or defense molecules, molecules derived from diet or environmental exposures (the exposome), and molecules derived from the biosynthetic interactions with associated microbes (the microbiome). Metabolites, both known and unknown, have structures that span a wide chemical structure space.^3,4 The chemical space of metabolites of 100 atoms or less is on the order of ~10⁶⁰ unique molecules, which makes it a grand challenge to explore and characterize metabolomes using experimental and theoretical tools.^5–7 Metabolomics utilizes sophisticated experimental and computational technologies for the identification of biologically relevant metabolites.^8,9 Identification of metabolites is important in order to understand metabolic diseases, develop precision medicine, and elucidate the pathways altered in complex metabolic networks.^10–12 The experimental annotation of metabolites is performed by matching chemical features against databases or, if available, with appropriate chemical standards.^13,14 Currently, this last approach is nontrivial as the vast majority of metabolites do not have such standards. For example, the HMDB database contains a small percentage of the total number of metabolites across multiple organisms, and most of the of molecules in the HMDB do not have authentic chemical standards.^8,15 Further, ~1% of available compounds in the U.S. Environmental Protection Agency (EPA) Distributed Structure-Searchable Toxicity (DSSTox) Database, PubChem, ChemSpider, and the American Chemical Society’s CAS databases, which in aggregate contain millions of chemicals,^16,17 have known chemical structures. Even the most powerful analytical technologies such as nuclear magnetic resonance spectroscopy (NMR)^18–21 and mass spectrometry (MS)^22–24 can have difficulties in confidently identifying unknown compounds for a number of technical reasons. Therefore, it is tremendously challenging to build a complete experimental map of any given metabolome, considering the large gaps in our knowledge of metabolite structures.

Ion mobility coupled to mass spectrometry (IM-MS) has emerged as an effective tool to study the structures of unknown compounds while providing high selectivity and fast multidimensional separations.^25–27 Ion mobility spectrometry (IMS) separates gas-phase ions based on differences in their rotationally averaged collisional cross-section (CCS) values. The advantage of calculating the molecular surface area or CCS as compared to other properties such as MS/MS spectra or chromatographic retention time is that CCS can be measured to a relative standard deviation (RSD) ranging from 0.25 to 6.0% depending on the instrument employed, thereby providing a unique physiochemical property for structure elucidation of target metabolites.^28,29 Isomeric metabolites that commonly exist in biological samples can, in many cases, be accurately distinguished by CCS values. More importantly, CCS values are reproducible and relatively insensitive toward instrumental resources and laboratory environments. McLean and co-workers^30,31 and Baker and co-workers³² have used standard chemicals to construct experimental CCS databases with >1000 CCS values.

Data curation efforts of the MS community have contributed positively to the sharing of mass spectra, an approach that could be mimicked for IM data. One example is the MassBank database, with a wide user base and contributors from many different countries.³³ The MassBank of North America and the European MassBank³⁴ have further accelerated the sharing of mass spectral data for annotated metabolites. These data servers include autocuration of spectra and chemical structure information (InChI keys). On the other hand, the GNPS³⁵ spectral database utilizes a crowd-sourcing approach to annotate unknown compounds. However, these resources are still reported in different formats, and at present, there are no standardized procedures for data collection, collation, standardization, and sharing.^30,32

Over the past decade, IM-MS³⁶ has emerged as an advanced technique for the analysis of metabolites.^37–40 IM-MS experiments enable measurement of both mass and structural features within milliseconds. Four main types of IM-MS technologies are used widely: traveling wave (TWIMS),^41,42 drift tube (DTIM),^38,43 trapped IM (TIMS),⁴⁴ and differential mobility MS (DMS).⁴⁵ IM-MS uses different types of ionization techniques which include matrix assisted laser desorption ionization (MALDI), electrospray ionization (ESI), and atmospheric pressure chemical ionization (APCI). Liquid-phase separations such as capillary electrophoresis, gas chromatography, and supercritical fluid chromatography can also be integrated into IM-MS systems. Although different experimental methods are used to measure CCS, the typical experimental uncertainty and differences between platforms amount to ~3%, which is the accepted threshold for comparing predicted and experimental CCS values.

In silico technologies have shown promise for the comprehensive identification of compounds and as a way by which to expand database content through the introduction of accurate theoretical information.^46–49 High throughput and comprehensive identification of metabolites can afford a deeper understanding of the role that small molecules play in a given biological system. Importantly, in silico approaches are "standard free" and, in principle, can provide high quality predictions for a range of properties. Many advances in in silico approaches for structure determination have been reported, including NMR chemical shifts^50,51 and spin coupling, chromatographic retention times,^52,53 ion mobility CCS,^49,54 and tandem mass spectra.^55,56 Renslow et al. have developed a quantum chemistry-based pipeline (ISiCLE) for CCS calculations that produces CCS predictions in good agreement with experimental values.⁴⁹ To identify the protonated and deprotonated state of a molecule, ISiCLE considers the lowest energy conformation of the molecule. More recently, Grimme et al. have shown that the lowest energy protonated state does not necessarily give the best agreement with experimental CCS values.⁵⁷ Other high energy states were found to also be important in order to obtain accurate CCS predictions. This finding suggests that in order to obtain the most accurate CCS predictions all of the protonated and deprotonated states should be correctly modeled. Furthermore, ISiCLE uses simulated annealing molecular dynamics (MD) methods to generate a number of temporally correlated structures, which are then subjected to QM optimization, making this a computationally expensive approach.⁵⁸ MD trajectories use molecular mechanics for the Hamiltonian. Moreover, because of errors in small-molecule force fields there is no guarantee that the global minimum conformation will be identified.⁵⁹ Investing computational resources on the DFT refinement of MD-generated initial geometries is not necessarily optimal. Hopkins et al. performed CCS calculation using Mobcal-MPI with 25 different computational methods, showing an accuracy of CCS predictions of ~3.0% relative to experimental values.⁶⁰ However, this approach does not extensively map the conformational space of a given molecule, which is known to be important for CCS prediction.

Two basic approaches currently exist for predicting IM CCS values. Appropriately constructed machine learning approaches greatly accelerate CCS calculation and have modest CPU needs. However, training set dependencies limit the application of this approach and the quality of results. On the other hand, standard quantum mechanics (QM) methods are very reliable but computationally intensive. In this work, we have developed a CCS calculation workflow that combines the best features of both machine learning and standard QM methods. We have compared our calculated CCS values computed using our in silico workflow with the available database results to confirm their accuracy. Twenty metabolites and their possible molecular states (protonated/deprotonated/neutral) were explored in this benchmark. Furthermore, we performed our own set of TWIMS experiments to measure CCS values for the 20 metabolites considered in this study. The CCS values calculated using our workflow were in good agreement with available experimental values in the literature (±3%), with the values determined herein (±3%), and with the average or consensus values between the two experimental values (±3%). This high-throughput workflow was automated via a series of Python scripts and has the potential to compute CCS values for large numbers of metabolites to enhance structural analysis in metabolomics.

METHODS

The developed workflow consists of several distinct steps (see Figure 1). Initially, we determine the molecular protonation state of the corresponding gas-phase ions (typically [M + H]⁺ or [M − H]⁻) and then generate the conformations for these using the RDKit toolkit.^61–63 All generated conformations undergo geometry optimization using a QM-based ML model called ASE_ANI^64–66 followed by an unsupervised clustering step using in-house unsupervised clustering code (viz. AutoGraph)⁶⁷ to obtain structurally distinct conformations. Standard DFT geometry optimization and atomic charge calculation are then performed on a representative conformation from each cluster at the B3LYP/6–31+G(d,p) and B3LYP/6–311++G(d,p) level of theory, respectively, using the Gaussian 16 software package.^68–71 The input file for CCS calculation is prepared by extracting the geometry and atomic charges from the DFT computations. Finally, the roomtemperature N₂-based trajectory method is used to calculate the average collisional cross-sectional areas using the HPCCS code developed by Zanotto et al.^72,73 and predict the structure of the target metabolites by comparing the computed results using a Boltzmann-weighted average over multiple conformers with experimental CCS values. We describe details of each step of the workflow in the Supporting Information. The TWIMS CCS experimental values are also reported in the Supporting Information, together with instrumental parameters (Tables S1 and S2). The workflow is integrated in our freely available Web server, www.pomics.org.

Figure 1.

In silico collisional cross section calculation workflow.

RESULTS AND DISCUSSION

The goal of this study was to develop an in silico workflow to accurately calculate the IM CCS value of an unknown metabolite in order to assign its 3D structure and filter out false positives more easily. This workflow involves an eightstep process starting from protonation-state determination to Boltzmann-weighted structure assignment, as illustrated in Figure 1. To validate the proposed workflow, we computed the CCS values for 20 metabolites (see Figure S1) with atom counts ranging from 10 to 40 and compared the computed results to literature CCS values and to CCS values measured in this work. Comparisons against literature values and the TWIMS experimental data are summarized in Table S3.

The number of molecular states for each metabolite are given in Table 1. As an example, for carnosine we have generated five states including one neutral, three monoprotonated ions, and one doubly protonated ion. As our aim was structure elucidation, we performed a rigorous study covering a range of likely molecular states. Apart from the pK_a determination, chemical intuition can also play an important role in predicting likely protonation states of a particular metabolite. Each of the molecular states was then converted into a 1D SMILES string and subjected to conformation generation using the RDKit conformation generation tool. We only requested 1000 conformers, but this number could be increased or decreased depending on the application or on the desired coverage of the conformational space. The total number of conformations generated for each metabolite are depicted in Table 1, and it can be seen that for conformationally rigid species few conformers are generated, while for more flexible molecules many more conformations are produced. The AutoGraph clustering method greatly reduces the conformational complexity by lumping related conformations into distinct clusters. The five predicted molecular states for carnosine (viz. models 1–5) generated 990, 994, 994, 994, and 996 conformers, respectively, which after ANI-1ccx minimization and clustering represent 16, 13, 15, 12, and 14 unique conformations, respectively. A representative cluster of the carnosine molecule is depicted in Figure 2. The results for the remaining 19 metabolites are summarized in Table 1. Optimization of large number of generated conformation using a ML model followed by structural similarity based clustering helps to identify relevant conformations in a large conformational space and greatly reduces the number of QM geometry optimization yielding a computationally efficient workflow without compromising accuracy.

Table 1.

Compound Name, Molecular State, Molecular Charge, Number of Atoms, Number of Conformations, Number of Clusters, Boltzmann-Weighted CCS Values, and the Percent Error between the In Silico Predicted CCS Value As Compared to the Corresponding Experimental Consensus CCS Value^a

no.	compd name	molecular state	molecular charge	no. of atoms	conformation no.	cluster no.	Boltzmann-weighted CCS	error (%)
1	carnosine	model 1	0	30	990	16	137.15	9.40
		model 2	1	31	994	13	166.54	9.91
		model 3	1	31	994	15	163.57	8.27
		model 4	1	31	994	12	150.21	0.11
		model 5	2	32	996	14	217.5	31.02
2	L-anserine	model 1	0	33	998	8	141.8	8.49
		model 2	1	34	996	9	159.74	3.70
		model 3	1	34	996	10	163.96	6.18
3	abscisic acid	model 1	0	39	1000	13	154.38	5.44
		model 2	1	40	1000	9	162.86	0.05
		model 3	1	40	1000	10	166.43	2.19
4	O-succinyl-l-homoserine	model 1	0	28	1000	21	134.4	7.80
		model 2	1	29	1000	17	145.45	0.39
5	L-tyrosine	model 1	0	24	961	9	118.68	19.97
		model 2	−1	23	902	10	148.53	4.14
6	L-citrulline	model 1	0	25	1000	14	122.85	10.18
		model 2	−1	24	1000	17	141.86	4.59
		model 3	−1	26	999	17	141.96	4.65
7	quinolinic acid	model 1	0	17	87	7	105.66	27.81
		model 2	1	18	66	5	142.14	5.00
8	nicotinic acid	model 1	0	14	21	3	90.78	40.42
		model 2	1	15	18	3	132.17	3.55
9	guanidinoacetic acid	model 1	0	15	989	10	90.15	40.92
		model 2	1	16	792	6	131.49	3.38
		model 3	1	16	997	6	131.1	3.10
		model 4	1	16	998	10	130.21	2.43
10	citramalic acid	model 1	0	18	998	11	98.98	22.53
		model 2	−1	17	999	10	122.52	1.01
		model 3	−1	17	995	9	119.62	1.39
		model 4	−1	17	997	9	121.58	0.24
11	N-methyl-l-gIutam.ate	model 1	0	22	1000	14	105.34	25.17
		model 2	1	23	999	17	129.3	1.98
12	serotonin	model 1	0	25	987	7	131.36	0.38
		model 2	1	26	914	8	165.04	20.11
		model 3	1	26	901	4	161.6	18.41
13	L-mimosine	model 1	0	24	935	8	127.93	12.00
		model 2	1	25	919	10	150.9	5.05
		model 3	1	25	958	9	145.36	1.43
14	L-tryptophan	model 1	0	27	934	9	136.84	9.62
		model 2	1	28	919	11	161.55	7.14
		model 3	1	28	924	10	159.69	6.06
15	L-ornithine	model 1	0	21	999	14	110.09	16.81
		model 2	1	22	1000	12	127.39	0.95
		model 3	1	22	997	11	126.63	1.55
16	N,N-dimethylglycine	model 1	0	16	999	7	100.1	25.42
		model 2	1	17	1000	6	118.19	6.23
17	kynurenine	model 1	0	27	993	10	137.15	7.64
		model 2	1	28	978	8	146.94	0.47
18	L-asparagine	model 1	0	17	997	10	111.86	15.03
		model 2	1	18	991	7	124.49	3.36
		model 3	1	18	999	8	128.93	0.20
19	L-2-aminoadipic acid	model 1	0	22	999	16	112.99	16.44
		model 2	1	23	1000	14	129.56	1.55
20	glutamine	model 1	0	20	1000	12	115.5	13.14
		model 2	1	21	1000	9	129.92	0.58

^aNitrogen gas was used as the drift gas.

Figure 2.

Structure clusters for carnosine (Model 4) as an example. A total of 994 ANI optimized conformers yielded 12 clusters.

To assign the structure correctly and to understand the dependency of the CCS values on the molecular state, we compared the calculated CCS values with literature value, and the CCS values measured in this work and report the quality of the results in the form of the percent error. The results are tabulated in Table 2 for the 20 molecules in each of their molecular states (neutral and charged). For carnosine, the consensus CCS value for the [M + H]⁺ ionic species is 150.1 Ų. Therefore, it is expected from the in silico modeling that one of the three singly protonated species should best match experiment as this is the species being detected experimentally. For carnosine model 4, where the terminal amine group is protonated (Figure 3), a computed CCS value of 150.21 Ų was calculated using our workflow. The CCS error was 0.11% relative to the consensus experimental value indicating that the prediction in this case is well within the experimental error estimated as ±3% in Table 2.²⁹ The other two singly protonated species (models 2 and 3, 166.54 and 163.57 Ų, respectively) have 9.9% and 8.2% errors relative to experiment. We were further interested in calculating the CCS values of the neutral species (model 1) and the doubly protonated species (model 5) to gain more insight into the effect structure plays in the predicted CCS values, despite these not being experimentally detectable. When compared to the CCS value of monoprotonated carnosine the neutral and doubly protonated models exhibited percent errors of 9.4% and 31.0%, respectively. From this analysis we concluded that the [M + H]⁺ ion corresponds to model 4. For the remaining systems we only investigated the singly charged protonated or deprotonated species (be it positive or negative) and the neutral species. In the absence of experimental electrospray data indicating multiple charges, we did not explore other higher charge states for the remaining 19 metabolites.

Table 2.

Experimental Uncertainties of CCS Values for the Studied Metabolites (Ų)^a

no.	compound name	experimental CCS (Literature)	experimental CCS (this work)	error (%)b	consensus
1	camosine	152.2	147.9	2.89	150.1
2	L-anserine	156.0	151.7	2.81	153.9
3	abscisic acid	160.6	165.0	2.69	162.8
4	O-succinyl-l-homoserine	147.5	142.3	3.63	144.9
5	L-tyrosine	145.8	139.0	4.86	142.4
6	L-citrulline	139.5	131.2	6.33	135.4
7	quinolinic acid	139.0	131.1	6.01	135.1
8	nicotinic acid	128.4	126.6	1.38	127.5
9	guanidinoacetic acid	126.9	127.2	0.25	127.1
10	citramalic acid	124.9	117.7	6.09	121.3
11	N-methyl-l-glutam.ate	133.7	130.0	2.85	131.9
12	serotonin	151.9	144.2	5.34	148.1
13	L-mimosine	143.5	143.1	0.26	143.3
14	L-tryptophan	143.5	145.8	1.60	144.7
15	L-ornithine	129.8	127.4	1.88	128.6
16	N,N-dimethylglycine	123.9	127.2	2.59	125.6
17	kynurenine	151.1	144.2	4.76	147.7
18	L-asparagine	131.5	125.8	4.56	128.7
19	L-2-aminoadipic Acid	131.5	131.6	0.05	131.6
20	glutamine	133.5	127.9	4.34	130.7

^aNitrogen gas was used as the drift gas for CCS measurement. The average of two experimental data, i.e., consensus CCS values, is also reported.

^bAverage % of error is 3.3 ± 2

Figure 3.

Protonation state is modeled by pK_a prediction and chemical intuition. Only cationic species were generated to match the electrospray ionization mode. The pK_a is calculated using the EPIK module implemented in the Schrödinger software suite.

TWIMS-MS experiments exhibit excellent between-lab and between-run reproducibility but require calibration to a set of compounds of known CCS values in order to calculate CCS experimental values.⁷⁴ These known values are most typically obtained using drift tube ion mobility (DTIM) and the Mason–Schamp equation,²⁵ though DTIM measured CCS values can vary largely between laboratories and instruments. For example, Hines et al. measured the CCS of the polyalanine hexamer to be 190.8 Ų, while Picache et al. measured 194.0 Ų, a 1.6% difference.^75,76 Much of the error calculated for our experimental TWIMS CCS values may be attributed to di?erences between the DTIM system used to measure CCS values for our polyalanine calibration⁷⁵ and the DTIM system used to generate the database used for our comparison with select metabolites.³² For this reason, it is essential to calibrate TWIMS measurements using CCS values from a single DTIM database if the goal is to match unknown compounds to CCS measurements against that same database.

To further validate the results obtained from this workflow, we performed the CCS computation for additional metabolites with different molecular states. Depending on the electrospray ionization mode used for IM-MS experiments with each analyte, either the [M + H]⁺ or [M − H]⁻ molecular states were generated. The neutral state was always considered for each metabolite essentially as an internal reference in that the neutral molecule should not match experiment well at all. A total of 52 states were generated for the 19 metabolites (excluding the five for carnosine), and we executed the workflow to obtain the computed CCS values. The final step of this workflow is to Boltzmann average the CCS values of the multiple conformers.

The Boltzmann weighting step takes the total energies and uses them to weight each individual conformer to determine how much it contributes to the observed CCS value. To illustrate this, we use model 4 of carnosine as an exemplar (see Table 3). A similar analysis for the other molecular states of carnosine and the 19 other metabolites is reported in the Supporting Information (see Tables S4–S57). It is observed from Table 3 that the most stable conformation of model 4 is conformation number 3, which contributes 99% to the Boltzmann-weighted CCS value. Conformation number 11 is 2.55 kcal mol⁻¹ higher in energy than conformation 3, and this high energy conformer contributes only 1% to the Boltzmann-weighted CCS value of 150.21 Ų. The remaining high energy conformers negligibly contribute to the computed CCS value. We observe that, in many cases, molecules can exist in multiple conformations having an effect in CCS calculations. Therefore, in our experience, it is always better to consider Boltzmann average properties rather than the lowest energy state. For example, model 3 of L-tryptophan (see Table S43) has two conformations (4 and 9) with relative populations of 54% and 45% in the gas phase, respectively. Conformations 4 and 9 have CCS values of 166.44 and 156.03 Ų, and the experimentally reported value is 154.22 Ų. The Boltzmann average CCS value is 159.69 Ų. Therefore, the Boltzmann average value gives a significantly better prediction of the CCS value relative to the global minimum. In our previous work on NMR chemical shift calculation⁵⁰ for metabolites, we observed that high-energy conformations had a subtle impact on the computed NMR chemical shifts, and in order to obtain high-resolution predictions, the entire ensemble was essential. Beyond giving accurate CCS values, the ensembles themselves give molecularlevel insights into the conformational space of molecules, which in many cases can be quite informative in their own right.

Table 3.

Conformation Number, Relative Energy, Mole Fraction, and the CCS Value for Carnosine (Model 4, Protonated Species)

conformation no.	relative energy (kcal/mol)	mol fraction	CCS value (Å2)
1	25.25	0.00	169.05
2	28.35	0.00	172.56
3	0.00	0.99	150.21
4	28.76	0.00	170.15
5	25.05	0.00	170.87
6	22.69	0.00	169.19
7	8.86	0.00	160.17
8	14.49	0.00	160.97
9	12.31	0.00	160.86
10	22.97	0.00	167.64
11	2.55	0.01	150.15
12	13.56	0.00	163.02

To further understand the conformational energy surface, we have further investigated the AutoGraph clustering of the ANI-1ccx optimized conformations. Figure 4 shows the clustering of carnosine (model 4) based on weighted degree and energy values. This figure confirms that both conformations 3 and 11 are in the low energy basin. Cluster regions with the QM energies are shown in Figure S3, which further confirms that conformations 3 and 11 are low energy conformations and, hence, the most probable structures of carnosine. From this analysis we predict that conformation 3 is the most likely structure of the carnosine as it traverses the IM device along with a small fraction of conformation 11. This suggests that our protocol not only can return a CCS value but can provide molecular level insights into the metabolite under study.

Figure 4.

Unsupervised clustering of carnosine (model 4) by AutoGraph. Twelve different colors represent the 12 cluster regions (Left). Low to high energies (ANI-1ccx potential) are represented by blue to red, respectively, and the relative energy value is in kcal mol⁻¹ (right). Figure generated with the Gephi software.

Using the best matching CCS value with experiment for all 20 metabolites yields an average error of 2.2% (within experimental uncertainties), suggesting that our computed CCS values are highly reliable and will lead to highly accurate CCS and structure predictions for unknown metabolites.

CONCLUSIONS

This article introduces a robust in silico workflow to accurately predict the CCS values of small molecules. Moreover, we also evaluated the experimental CCS values and compared them with available literature values to establish that the experimental variation between laboratories is ~±3%, which along with other analyses^28,29 establishes the reproducibility of experimental values. The computational workflow is an eightstep process, and the pipeline utilizes the best aspects of force fields, machine learning QM and QM methods to achieve highly accurate and reliable results. Accurate CCS prediction primarily depends on the molecular state (protonated/deprotonated) of a metabolite. Hence, thorough characterization of relevant molecular states yields CCS predictions within 3% of experimental values. The unsupervised clustering method included in this workflow reduces the possibility of human bias and error in cluster selection. The QM-ML model and clustering technique makes this protocol more computationally efficient. All of the steps can be processed in an automated way using series of Python-based scripts, making this protocol even more useful in building an in silico library of predicted CCS values and can be used to assign the structure of an unknown metabolite. Our newly developed workflow maintains an excellent balance between accuracy and computational cost, and we anticipate that this advanced protocol will be a useful tool for structure prediction for the metabolomics community and other communities studying small molecular species. It should be mentioned here that structure elucidation of any unknown compound, including metabolites, is challenging and only one data point is very unlikely to identify an unknown. CCS prediction is only one part of a larger workflow using multiple techniques that include retention time matching, MS/MS databases, isotopic cluster assignments, NMR, and even MSⁿ. Hence, accurate structure elucidation involves determining a range of properties both experimentally and computationally. CCS calculations, however, can reduce the molecular space and can be used as a filter to remove false positives. When coupled with other techniques, it can lead to the more reliable annotation of unknown metabolites.