Integrating Computational Chemistry and Ion Mobility Spectrometry into Non-Targeted Analysis Workflows for the Structural Elucidation of Emerging PFAS

Abstract

Per- and polyfluoroalkyl substances (PFAS) are an evolving class of synthetic chemicals that are pervasive in the environment due to widespread manufacturing and consumer use, such that PFAS contamination is of great concern. Although thousands of PFAS structures have been reported to date, this number increases daily with the identification of new PFAS from advances in non-targeted analysis (NTA) workflows. Ion mobility spectrometry in combination with mass spectrometry (IMS-MS) has recently been incorporated into PFAS NTA studies. Although MS provides essential precursor and fragmentation data, IMS offers complementary structural information via the measurement of the ion-neutral collision cross section (CCS) values. Experimental CCS values can then be compared with those calculated in silico from candidate 3D structures to meet confidence-level criteria in analyte assignments within NTA workflows. Although this approach has been applied since the late 1990s, PFAS often exhibited poorer agreement between the calculated and experimental CCS values. To address this limitation, we propose an optimized computational workflow to calculate Boltzmann-weighted CCS values for PFAS structures generated via quantumchemical calculations. This workflow was assessed with experimental CCS values from 56 known PFAS structures across seven classes and resulted in an average percent error of 2.0%. Moreover, 11 new PFAS structures were proposed from NTA, with average errors of 1.3%. The combination of this new computational workflow and experimental IMS-MS measurements therefore establishes a workflow for structural elucidation of emerging PFAS.

Graphical Abstract

1. INTRODUCTION

PFAS comprise a diverse group of persistent environmental contaminants with wide-ranging chemical structures.^1,2 As the use of PFAS in industrial and commercial products has increased since the 1950s, the number of unique PFAS produced in manufacturing processes has grown exponentially.³ The Organisation for Economic Co-operation and Development (OECD) broadly defines PFAS as a species containing one fully fluorinated methyl (CF₃) or methylene (CF₂) carbon atom, which translates to over seven million compounds reported in PubChem.^4,5 The term “legacy” PFAS is commonly used to describe long-chain perfluoroalkyl acids (PFAAs), such as perfluorosulfonic acids (PFSAs) and perfluorocarboxylic acids (PFCAs).⁶ Perfluorooctanesulfonic acid (PFOS) is one of the most well-characterized legacy PFAS, composed of a hydrophobic 8-carbon perfluoroalkyl chain attached to a sulfonic acid headgroup. Although voluntarily phased out of manufacturing for commercial products, PFOS contamination is still a matter of concern due to persistence in the environment and detection in the blood of >96% of Americans tested at a concentration of ~4 parts per billion (ppb) on average.^7,8 “Emerging” PFAS is also a common term linked to short-chain PFAAs and alternative polyfluorinated PFAS, such as fluorotelomers (FTs) and perfluoroalkyl phosphates (PFPs), which have been implemented as replacements for legacy PFAS. Emerging PFAS make up over 99% of the 14,000+ unique PFAS structures reported in the U.S. Environmental Protection Agency’s CompTox Chemicals Dashboard.⁹ Because this number continues to grow as manufacturing companies regularly develop alternative PFAS, it is imperative to optimize and develop better NTA workflows to characterize novel PFAS in complex matrices.^9–12

Ion mobility spectrometry in combination with mass spectrometry (IMS-MS) has been demonstrated as a useful method for the detection and identification of PFAS.^10–14 Specifically, MS provides fragmentation and precursor mass information for each detected feature while IMS enables structural analyses with the capability of separating constitutional and conformational isomers.^15–19 In an IMS experiment, such as drift tube IMS (DTIMS), ions are separated in an inert buffer gas, such as nitrogen, under the influence of a weak electric field.²⁰ The mobility of the ion is then determined by the time it takes for the ion to migrate through the buffer gas, which, if conditions within the IMS are within the so-called “low-field limit,” correlates to the ion-neutral collision cross section (CCS) via the Mason–Schamp equation.²⁰ DTIMS, traveling wave IMS (TWIMS), and trapped IMS (TIMS) operate within this limit where ion motion can be described by the Mason–Schamp equation, meaning the measured drift time translates to a CCS that reflects the gas-phase structure of an equilibrated ensemble of analyte conformations.²⁰ Although acquisition of drift time measurements varies across instrumental platforms, such as with TWIMS, where CCS values are determined based on calibration against known standards, CCS values are considered highly reproducible under standardized experimental conditions.²¹

Many research groups calculate and report experimental CCS values so they may be used by others in targeted and suspect screening analyses for known analytes.^22–25 To date, experimental CCS values for various PFAS species have been reported by using commercially available standards to explore the use of IMS for PFAS detection and identification.^13,23,26 IMS experiments have been valuable in separating halogenated ions from nonhalogenated isobars, due to differences in density and shape.²⁶ This has been demonstrated with CCS vs m/z trendlines, which show a region enriched in halogenated species that is distinct from the nonhalogenated chemical space.²⁶ These trendlines can then be leveraged to detect and identify novel PFAS and other halogenated molecules in NTA. To date, the validation of structures from NTA is extremely challenging (Figure 1), but IMS is showing promise for addressing this limitation. Specifically, in targeted and suspect screening analyses, the features of interest are termed “Known Knowns” or “Known Unknowns,” and most have available isotopically labeled standards that can be analyzed directly in the instrumental platform for validation.^13,23,27 However, NTA differs in that detected features are classified as “Unknown Unknowns,” many of which have no analytical standards or known structures.^10–12 Therefore, the CCS values from IMS experiments provide structural information on the “Unknown Unknowns” by enabling comparison with CCS values calculated for candidate geometries consistent with a given molecular formula and compound class, where agreement within 3% has historically been considered supportive of a potential structural match.^28,29

Figure 1.

Using IMS experimental and theoretical collision cross section values in non-targeted analyses. Experimental features from NTA lack standards, but if experimental cross sections exist, they can be compared to CCS values from theoretical structures. In the computational workflow, CCS values calculated from theoretical output structures can be compared to experimental CCS values to determine potential structures for the molecular formulas proposed from NTA feature selection.

To further aid in structural identifications for NTA workflows when standards are unavailable, researchers have incorporated machine learning (ML) algorithms to predict CCS values of emerging contaminant species.^26,30,31 However, current ML approaches are often unable to distinguish between diastereomers or structural isomers, and their predictive performance is generally strongest when the analyte class is well represented within its training set. Because most IMS data sets are dominated by “typical” organic molecules (composed of C, H, O, N, P, S),^28,32–35 computational workflows remain important for generating CCS predictions across chemically distinct PFAS subclasses while explicitly accounting for stereochemistry. Accordingly, CCS values can be calculated from theoretical candidate structures generated by mapping the potential energy surface (PES) of the analyte and optimizing representative conformations using appropriate levels of quantum-chemical theory (i.e., density functional theory (DFT), coupled cluster theory). From these structures, a Boltzmann-weighted (BW) CCS value, which assumes a thermodynamic equilibrium at 298 K, can then be computed for the thermally accessible conformational ensemble populated under low-field IMS conditions, providing better agreement between experimental and calculated CCS ^33–35 values.

In this work, we adopted computational methods from the literature and refined the applied levels of theory to ensure compatibility with highly fluorinated species.³⁴ Because fluorinated molecules are underrepresented in computational analyses, this workflow is very important for NTA studies with PFAS and potentially other halogenated chemicals. Although quantum-chemical methods account for the unique electronic properties of fluorine, few studies have probed the conformational landscape of PFAS in the gas phase.^36–40 Existing IMS-related studies have been limited to PFCAs and focused on unique intra- and intermolecular interactions such as hydrogen bonding and dimerization.^37,38 Additionally, these studies have not employed automated conformer search methodologies, which have been deemed advantageous when performing BW-CCS analyses in order to achieve comprehensive representation of ion populations.^34,41,42

The use of quantum-chemical calculations paired with dynamic modeling of ion-neutral collision events has achieved accurate comparisons between experimental CCS values and those calculated from theoretical structures. However, these methods have been most commonly applied to organic cations, and several difficulties arise with PFAS.³³ Specifically, in addition to high degrees of halogenation, the majority of PFAS are analyzed in negative ionization mode; thus, computational approaches must be compatible with anionic, electron-dense species. In these cases, it is important to use basis sets that include diffuse functions, which add flexibility to describe electron densities that occupy relatively large volumes. This is particularly critical for anions, where the extra electron density is less tightly bound and extends further from the molecule. Standard basis sets without diffuse functions can underestimate the molecular size and distort predicted structures and properties. Here, we propose an optimized computational workflow to calculate BW-CCS values for PFAS from various classes as the structural diversity of emerging PFAS continues to expand while remaining vastly understudied. This approach optimizes conformational searching, ensemble reduction, structural optimizations, and energetic refinement to provide a new way to assign structures to emerging environmental contaminants from NTA.

2. METHODS

2.1.. Experimental CCS Values

All experimental CCS values in this manuscript were referenced from pre-existing in-house liquid chromatography (LC)-IMS-MS libraries for PFAS, pharmaceuticals, and pesticides, also noted in Table S1. All library CCS values were obtained from commercially available reference standards characterized using the Agilent 6560 IM-QTOF platform, as previously described.^13,14,23 Briefly, following direct infusion, ionization was performed with electrospray. The ions were transmitted into a high-pressure ion funnel and then a trapping funnel, where the ion packet was stored and released into a 78 cm drift tube filled with nitrogen buffer gas at a typical pressure of ~4 Torr. The ions were then separated under the influence of a uniform electric field (17.3 V/cm) in the drift tube, based on their size, shape, and charge state, with their drift times recorded. All CCS values were obtained using the single-field calibration method, which allows experimental CCS values to be derived from drift time measurements relative to Agilent Tune Mix reference ions.²¹ Furthermore, all IMS measurements were performed in triplicate to obtain average CCS values and standard deviations for all standards.

2.2.. Theoretical CCS Values

2.2.1. Conformer Search.

To calculate theoretical CCS values for the PFAS studied here, a computational workflow for generating a refined conformer ensemble was employed. The first step in the computational workflow was to perform a conformer search by mapping the PES of the ion of interest and identifying what structures make up the conformational ensemble that surrounds the global minimum structure, utilizing the Global Geometry Optimization and Ensemble Generator (GOAT) algorithm available in ORCA (v. 6.1.0).^42–45 Each GOAT calculation employed the GFN2-xTB semiempirical tight-binding model, parametrized with temperature bins of 500, 750, 2000, and 3000 K for comprehensive coverage of the PES.^34,46,47 The GOAT output file provided the assignment of a global minimum structure and a list of additional structures that make up the conformational ensemble and their relative energies. The ensemble reported in the GOAT output file was limited by a maximum energy threshold of either 5 kcal/mol (~21 kJ/mol) or 15 kcal/mol (~63 kcal/mol) relative to the global minimum structure, depending on PFAS structure, as detailed below. To improve energetic order ahead of ensemble reduction, all GOAT-generated structures were screened via single-point energy at the r2SCAN-3c level of theory in CENSO.^48,49

2.2.2. Ensemble Reduction.

Depending on the size and flexibility of the ion of interest, tens to thousands of configurations may be generated during the conformer search. It is neither feasible nor necessary to perform full geometry optimizations for every structure, and therefore, only unique structures that are low in energy relative to the global minimum should be taken into consideration for BW analysis. If more than ~35 structures were generated by GOAT, then the ensemble was reduced in size by cosine similarity sorting, whereby all structures were assigned a vector representation of the mass-weighted distance between each atom and the molecular center of mass.⁵⁰ Using a custom Python script implementing hierarchical clustering based on cosine similarity with the Ward criteria for dendrogram links, a similarity cutoff was applied to retain only geometrically unique structures.⁵¹ The molecule-specific thresholds are reported in Table S2.

2.2.3. Structural Optimization and Energetic Refinement.

After the ensemble was reduced to a manageable number of unique configurations (10–40 structures), DFT calculations were used to optimize the molecular geometry, electronic energy, and thermochemistry of each structure.^52–55 Full optimizations were performed at the B3LYP/6–311+G(d,p) level of theory, along with the calculation of thermochemical corrections and electrostatic potential derived partial charges, using Gaussian v. 16.01.^56–58 The partial charges were calculated according to the Merz–Singh–Kollman scheme and were constrained to reproduce the dipole moment of the molecule.^59,60 Gaussian thermochemical data was extracted using the Featherstone Laboratories Suite GUI.⁶¹ Additional structural optimization and thermochemical correction calculations were performed at the B3LYP-D3/def2-TZVPPD level of theory with the RIJCOSX approximation and def2/J auxiliary basis set using ORCA v. 6.1.0.^34,62–71 The partial charges for these optimizations were calculated using the CHELPG partition scheme with a 0.1 Å grid, in which each grid point was constrained to be within 3.0 Å of any atom in the system.^34,72 ORCA thermochemical data were extracted using the PodPals GUI.⁷³ After optimization, only unique structures with relative energies within 15 kJ/mol of the global minimum were retained. To improve calculation accuracy, electronic energies for geometry-optimized structures were refined at the DLPNO–CCSD-(T)/def2-TZVPPD level of theory employing the def2/C auxiliary basis set.^74–79 All coupled cluster energies were extracted using the PodPals GUI.⁷³ Proposed global minimum input/output files are available via the ioChem-BD database associated with this manuscript (10.19061/iochem-bd-6-679).

2.2.4. CCS Calculations.

Once the final ensemble of low-energy structures had been optimized and refined as described above, the ion-neutral CCS value for each conformer was calculated using the trajectory method in MobCal-MPI using nitrogen as the buffer gas.^33,35,80 Calculated CCS (^BWCCS_N2) values were calculated under zero-field parameters (temperature: 298 K, buffer gas: N₂, number of cycles: 10, number of relative velocities: 104, number of collisions: 512).^{33,35,52,80,81} Finally, the DFT thermochemistry data, coupled cluster energies, and calculated CCS values for each conformer structure were compiled for BW-CCS analysis using the PodPals GUI.^34,52,73 BW-CCS values were calculated using eq 1

$${\text{CCS}}_{\text{BW}}\text{}=\sum _i{p}_i\times {\text{CCS}}_{\text{i}}$$ (1)

in which the calculated CCS value of each conformer (CCS_i) is multiplied by its corresponding Boltzmann population (p_i) at 298 K, to determine the BW average across all conformers.^33–35,81 The ^BWCCS_N2 values were then compared with experimental CCS values to determine which theoretical molecular structures had CCS values within 3% of the experimental value.

3. RESULTS AND DISCUSSION

The first objective of this study was to benchmark the performance of the BW-CCS workflow using established methods from the literature with analytes from the same chemical class of interest.³³ This enables assessment of agreement between experimental CCS values and candidate geometries generated from quantum-chemical calculations through percent error (%E). To exclude the unknown variable of excess electron density, initially, species with only low numbers of fluorine atoms were assessed. These included one pharmaceutical, fluoxetine, and two pesticides, tau fluvalinate and lufenuron (Figure 2). Each step of the computational workflow was then evaluated for the analytes, including conformational searching, ensemble reduction, structural optimizations, energetic refinement, and BW-CCS calculations. For the structural optimization step, DFT calculations were performed without diffuse functions. Here, low %E was observed between the computed and experimental CCS values (Figure 2). Although all three of these compounds are classified as PFAS by the OECD, these structures are not highly fluorinated as are most PFAS of interest.^5,82 For example, when comparing the [M-H]⁻ species of lufenuron and PFOS, the atomic composition of lufenuron is ~20% fluorine, whereas PFOS is ~60%. However, the initial results at low fluorination showed promise for applying the computational workflow to ions with higher fluorine compositions.

Figure 2.

Theoretical Boltzmann-weighted collision cross section values for minimally fluorinated compounds. The proposed global minimum structures for fluoxetine, tau fluvalinate, and lufenuron are shown with their ^BWCCS_N2 values (178.1 ± 1.0 Ų, 210.7 ± 1.2 Ų, 205.1 ± 1.0 Ų, respectively), which were <1% error of experimental CCS values (177.2 ± 0.2 Ų, 209.5 ± 0.2 Ų, and 204.0 ± 0.1 Ų, respectively).

To further validate that the structural optimization methods within the computational workflow were compatible with more highly fluorinated species, DFT calculations were performed at the same level of theory for PFOS, specifically ωB97X-D3/def2-TZVPP. In this case, the experimental ^DTCCS_N2 value for PFOS was 168.3 ± 0.2 Ų, and the ^BWCCS_N2 value was 161.2 ± 1.1 Ų, resulting in 4.2% error and a structure that was too compact (Figure 3A). Because the acceptable threshold is commonly considered less than 3%, this ^BWCCS_N2 value for PFOS was not satisfactory, especially when compared to the minimally fluorinated compounds.^28,29 To assess whether %E for PFOS was a result of the selected DFT method, analogous optimizations were performed with the addition of diffuse functions at the B3LYP/6–311+G(d,p) level of theory. With this modification, the ^BWCCS_N2 value was 167.6 ± 1.2 Ų, which deviates from the experimental ^DTCCS_N2 value by 0.4%. Looking at the global minimum structures obtained from both methods, it is apparent that the added diffuse functions produced an optimized structure that is more linear and extended, which agrees better with the experimental CCS value for PFOS (Figure 3A). Further investigation of different combinations of the ωB97X and B3LYP functionals paired with the def2-TZVP and 6–311G basis sets was performed to probe the effects of applying additional diffuse functions, polarization functions, and dispersion corrections. Although beyond the focus of this study, it was concluded that the addition of diffuse functions was critical to achieve accurate geometries for low-energy structures.

Figure 3.

Method optimization for highly fluorinated compounds. (A) Input structure of PFOS and proposed global minimum structures following application of the two DFT methods, without and with diffuse functions. The DFT calculations with diffuse functions produced a more extended structure with a ^BWCCS_N2 value (167.6 ± 1.2 Ų) that aligns with the experimental reference value (168.3 ± 0.2 Ų). The relative energy and structure plots for (B) no diffuse functions (ωB97X-D3/def2-TZVPP) and (C) added diffuse functions (B3LYP/6–311+G(d,p)) demonstrate that the former produced smaller CCS values than those determined experimentally. The “original ensemble” consists of all of the conformer structures generated from the GOAT calculation at the beginning of the workflow. The “final ensemble” consists of the reduced number of structures, which were retained at the end of the workflow and included in the BW-CCS analysis.

To ensure the difference in experimental and theoretical collision cross sections was related to the level of theory applied for the optimization calculations, rather than a downstream effect of the ensemble reduction, all 70 PFOS structures obtained from the conformer search were optimized with both DFT methods. Figure 3B,C represent the relative energy (kJ/mol) of each conformer versus their calculated CCS value (Ų). Although both plots generally follow a similar distribution in which relative energy decreases as CCS increases, the structures optimized without diffuse functions had smaller CCS values than the structures optimized with the added parameter. The differences between these two methods are greatest with respect to the proposed global minimum structure, which ultimately contributes the most to the BW population.^33,52 Without diffuse functions, the global minimum structure was calculated to have a theoretical CCS value of 161.9 ± 1.3 Ų, which contributes to 87.9% of the relative population as calculated by 298 K Gibbs energy. However, the global minimum structure for the calculations with added diffuse functions had a theoretical CCS value of 169.0 ± 1.4 Ų and makes up 86.1% of the relative population. While both structures contribute almost equally to their respective ^BWCCS_N2 values, the smaller structures generated without diffuse functions do not agree as well with the experimental data relative to the structures optimized with diffuse functions. While the addition of diffuse functions does not seem to be necessary for ions with low fluorine content, it is recommended to include them when modeling highly fluorinated species. Thus, the remainder of this work was conducted by executing the structural optimizations with diffuse functions included in the DFT parameters.

The computational workflow was then applied across multiple classes of PFAS to assess how structural diversity affects %E. The workflow was initially applied to 47 individual species representing seven PFAS classes, including PFSAs, PFCAs, FTs, perfluoro ether sulfonic acids (PFESAs), perfluoro ether carboxylic acids (PFECAs), perfluoroalkyl sulfonamides (PFASAs), and PFPs. Of these 47 compounds, 44 were [M-H]⁻ species, two [M-H–CO₂]⁻, and one [M+H]⁺. All experimental ^DTCCS_N2 values are published in the Baker Lab LC-IMS-MS PFAS library and included in Table S1 with the ion type modeled.²³ For six out of the seven PFAS classes (PFSAs, FTs, PFESAs, PFECAs, PFASAs, and PFPs), the average percent errors were determined to be less than 3% for 33 of the 47 species (Figure 4A). This workflow, however, did not work well with the PFCAs, where the average percent error was 5.6%, and 7 out of 8 PFCAs exceeded the acceptable threshold of 3%. PFCAs are an interesting class of PFAS that have been well-documented to undergo in-source fragmentation during MS analyses. Thus, while the deprotonated version of some PFCAs is observed at the detector, many decarboxylate to the [M-H–CO₂]⁻ form before detection.⁸³ For example, [M-H–CO₂]⁻ ions make up ~70% of the experimental CCS values for monomer PFCAs published in the Baker Lab LC-IMS-MS PFAS library.²³ Thus, when validating this workflow, accurately calculating theoretical CCS values for decarboxylated species is critical, but the %E was large for these molecules in comparison with the other PFAS. Therefore, we decided to investigate an alternative route for optimization. For example, when the computational workflow was applied directly to the [M-H–CO₂]⁻ ion of perfluorononanoic acid (PFNA), a ^BWCCS_N2 value of 153.6 ± 0.8 Ų resulted, which deviates by 4.5% from the experimental CCS value of 147.0 ± 0.2 Ų (Figure 5). However, when the computational workflow was first applied to the [M-H]⁻ form of PFNA, and then it was decarboxylated and a second structural optimization was performed, the resulting ^BWCCS_N2 was within 1.7% of the experimental CCS value (Figure 5). Differences in ^BWCCS_N2 between these various optimizations seemed to be a result of partial charges assigned to the carbon atoms that differed greatly across the two approaches (Table S3). For example, the fourth carbon in the alkyl chain was assigned a slightly positive partial charge (0.057) in the direct optimization of [M-H–CO₂]⁻, but was assigned a negative partial charge (−0.220) following sequential [M-H]⁻ and [M-H–CO₂]⁻ optimizations. Thus, these assignments have a considerable effect on the results of the trajectory method calculations, which are influenced by charge density.^33,52

Figure 4.

Percent error between theoretical and experimental CCS values for seven PFAS classes. (A) Initially, the average percent error for 33 out of 47 unique PFAS and the 6 PFAS classes was <3%. (B) Following modification of the workflow for [M-H-CO₂]⁻ PFCA species, the percent error for 44 out of 51 PFAS across all 7 classes was <3%. Average error is denoted with a *.

Figure 5.

Input structure of PFNA and ^BWCCS_N2 results from two workflows to model [M-H-CO₂]⁻ species. Initially, decarboxylation of the input structure without prior optimization was assessed, but yields a theoretical ^BWCCS_N2 value with 4.5% error (153.6 ± 0.8 Ų) when compared with the experimental CCS of 147.0 ± 0.2 Ų (left). Applying the workflow first to the deprotonated species and then reoptimizing the decarboxylated structures produces a theoretical ^BWCCS_N2 value that aligns within 1.7% (149.4 ± 0.9 Ų) of the experimental value (right).

Although the proposed modification of first performing [M-H]⁻ optimization and then [M-H–CO₂]⁻ in the computational workflow achieves better agreement for decarboxylated species, the deprotonated PFCAs remained challenged by an ~6% %E between experimental CCS values and theoretical structures (Table S4). This error seems to be the result of an unlikely interaction between one of the carbonyl oxygen atoms and the middle of the carbon backbone, resulting in calculated CCS values smaller than the experimental CCS values. This observation is consistent across long-chain PFCAs, including perfluorodecanoic acid (C10), perfluorododecanoic acid (C12), and perfluorotetradecanoic acid (C14). However, once this workflow was altered to better model [M-H–CO₂]⁻ ions through the [M-H]⁻ and then [M-H–CO₂]⁻ optimization steps, the average %E for the PFCA class was reduced to 2.1%. This modification also allowed for additional short-chain and branched PFCA compounds to be assessed because their only published experimental CCS values are for [M-H–CO₂]⁻ species. Therefore, the workflow was validated with 51 PFAS across the seven previously defined classes, including 38 [M-H]⁻, 12 [M-H–CO₂]⁻, and one [M+H]⁺ species. All 7 PFAS classes and 44 out of the 51 PFAS had %E less than 3% (Figure 4B), demonstrating the workflow is robust against the structural diversity of PFAS. The average %E and standard deviations for each PFAS class are reported in Table S5.

To address the challenge of calculating ^BWCCS_N2 values for deprotonated PFCA species, the original ensemble of structures generated in the conformer search section of this workflow was further investigated. As demonstrated by PFOS (Figure 3A), and consistent with the majority of PFAS analogs investigated in this study, it was concluded that energetically favored conformations tend to be elongated, linear structures. When the original ensemble for the [M-H]⁻ form of PFNA was visualized, it was observed that none of the structures included within the 5 kcal/mol cutoff retained any semblance of a linear conformation. Thus, in order to access the more elongated structures and attain a complete original ensemble, the energy cutoff in the conformer search calculation was increased to 15 kcal/mol. Visualization of this larger ensemble confirmed that the linear structures are found at higher energy levels at this step in the workflow. This is a consequence of the lower level of theory being utilized to generate hundreds to thousands of structures, which is necessary to ensure that the conformer search calculations remain efficient in time and resources. Once understood, this strategy was applied to the 51 PFAS that were assessed in the original data set, with the addition of the five deprotonated PFCAs that could now be properly modeled. The average %E for the 56 PFAS species, encompassing 43 [M-H]⁻, 12 [M-H–CO₂]⁻, and one [M+H]⁺ ion types, was determined to be 2.0%. Using the refined workflow with a conformer search cutoff of 15 kcal/mol, the average %E and standard deviations for each PFAS class are reported in Table S6. The errors across the PFCA, FT, PFESA, PFECA, and PFP classes were determined to be normally distributed, but both the PFSA and PFASA classes are influenced by an extreme outlier, 8Cl-PFOS and EtFOSE, respectively. Although this modification greatly improved the ensemble coverage for deprotonated PFCAs, reducing the average %E of the five analogues assessed in this study from 6.5% to 1.5%, not all PFAS require this high energy threshold. Within the original data set of 51 PFAS examined using the 5 kcal/mol cutoff, no extreme outliers were identified. In addition, the 15 kcal/mol cutoff generated 10-fold more structures than the original 5 kcal/mol cutoff on average. This leads to the need for more computational time and resources to generate and process a substantial number of structures, which is not necessary unless working with linear PFCA-like compounds. Thus, it is recommended to use personal discretion when deciding whether a high energy cutoff is needed for structures proposed by NTA.

The final goal of this study was to assess how this workflow performed for PFAS structures proposed from experimental NTA methods. In the work published by Kirkwood et al., aquatic passive samplers were deployed in the Cape Fear River downstream of a fluorochemical manufacturing site in North Carolina.¹² Following NTA, 11 PFAS structures were proposed for 11 unknown features, and confidence levels were assigned respective to the PFAS-specific confidence criteria provided by Charbonnet et al.^12,84 The 11 proposed structures include per- and polyfluorinated ether sulfonic acids and multiheaded perfluorinated ethers, where confidence levels ranged from level 2 (probable structure) to level 3 (tentative candidate) based on updated guidance for PFAS identification with IMS.^12,84,85 Of the 11 features of interest, the structures for compounds 8 and 11 were proposed with the lowest confidence (level 3) and were specifically defined as fragmentation-based candidate structures (Figure 6A,6B).¹² However, upon application of our computational workflow, the ^BWCCS_N2 values of the global minimum structures identified for compounds 8 and 11 demonstrate good agreement (~1.4%) with the experimental CCS values reported. Application of the workflow to all 11 proposed NTA structures also illustrated an average %E of 1.3%. Thus, the CCS values calculated from the theoretical structures using the proposed workflow aligned with experimental CCS values obtained from NTA features. Specifically, four out of 11 compounds fell below a 1% percent error threshold, five were between 1 and 2%, zero fell between 2 and 3%, and two were under 4%. Although CCS values calculated from theoretical structures alone will not increase the confidence level assignment for a given structure, current reporting criteria require that library or CCS values from theoretical structures must support all candidates.⁸⁵ Thus, as the gap continues to widen between the number of accessible analytical standards and features detected by NTA, the need for computational tools that accurately generate theoretical CCS values is essential for maintaining confidence in identifications. Therefore, the present study has demonstrated that our computational workflow can be applied across multiple classes of PFAS and used for commonly observed ion types. Furthermore, this workflow can be applied to eliminate candidate structures that may align with retention time and m/z information, but do not fall within 3% of the expected CCS value.⁸⁵

Figure 6.

Computational workflow applied to PFAS structures proposed from NTA. Global minimum structures of Compound 8 and Compound 11 with respective ^BWCCS_N2 values (163.4 ± 0.6 Ų) and (164.9 ± 1.3 Ų), which align with experimental values, 161.5 Ų and 167.5 Ų, respectively.

4. CONCLUSION

The present work showcases the use of a computational workflow to accurately model gas-phase conformer ensembles and calculate CCS values from the theoretical structures to complement IMS experiments. Although this workflow is akin to other BW-CCS calculation schemes with organic molecules, this work demonstrated that DFT optimizations performed with the addition of diffuse functions produced more extended structures that aligned well with experimental measurements for highly fluorinated compounds.³⁴ It should be noted that this observation was challenged by linear PFCA compounds. To retain accuracy for the decarboxylated PFCAs, a 2-step approach optimizing the [M-H]⁻ version and then the [M-H–CO₂]⁻ ion led to enhanced agreement with experimental values. Additionally, it was determined that this challenge can be overcome for deprotonated PFCAs as well by increasing the energy threshold for the conformer search from 5 to 15 kcal/mol. Overall, this workflow was validated across 56 unique PFAS, and the average %E was 2.0 ± 2.1%. Additionally, the performance of this workflow was assessed for 11 proposed PFAS structures identified via NTA from environmental samples. Given the remarkable agreement observed between the experimental CCS values and CCS values calculated for the theoretical structures, it was concluded that this workflow has the potential to supplement future NTA PFAS studies utilizing IMS. However, as analytical separations and measurements continue to evolve and enhance the detection and identification of emerging environmental contaminants, computational resources must also progress to characterize features of interest. Thus, this is just the beginning of many needed developments in integrating NTA with computational chemistry and would benefit from the leverage of artificial intelligence methods in the future.

Supplementary Material

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jasms.6c00201.

Experimental and theoretical CCS values with percent error analysis for all species (Table S1), the cosine similarity parameters used for ensemble reduction (Table S2), the partial charges assigned to [M-H–CO₂]⁻ PFNA before and after workflow modification (Table 3), experimental and theoretical CCS values with percent error analysis for PFCA compounds before workflow modification (Table S4), the summary statistics for individual PFAS classes with 5 and 15 kcal/mol energy cutoffs, respectively (Tables S5 and S6) (XLSX)

Data Availability Statement

Input/Output files for DFT and DLPNO-CCSD(T) calculations for the proposed global minimum structure of each species, ioCHEM-bd database (doi:10.19061/iochem-bd-6-679)