Unsupervised Structural Classification of Dissolved Organic Matter Based on Fragmentation Pathways

Abstract

Dissolved organic matter (DOM) is considered an essential component of the Earth’s ecological and biogeochemical processes. Structural information of DOM components at the molecular level remains one of the most extraordinary analytical challenges. Advances in determination of chemical formulas from the molecular studies of DOM have provided limited indications on structural signatures and potential reaction pathways. In this work, we extend the structural characterization of a wetland DOM sample using precursor and fragment molecular ions obtained by a sequential electrospray ionization–Fourier transform–ion cyclotron resonance tandem mass spectrometry (ESI-FT-ICR CASI-CID MS/MS) approach. The DOM chemical complexity resulted in near 900 precursors (P) and 24 000 fragment (F) molecular ions over a small m/z 261–477 range. The DOM structural content was dissected into families of structurally connected precursors based on neutral mass loss patterns (P_n−1 + F_1:n + C) across the two-dimensional (2D) MS/MS space. This workflow identified over 1900 structural families of DOM compounds based on a precursor and neutral loss (H₂O, CH₄O, and CO₂). The inspection of structural families showed a high degree of isomeric content (numerous identical fragmentation pathways), not discriminable with sole precursor ion analysis. The connectivity map of structural families allows for the visualization of potential biogeochemical processes that DOM undergoes throughout its lifetime. This study illustrates that integrating effective computational tools on a comprehensive high-resolution mass fragmentation strategy further enables the DOM structural characterization.

Graphical Abstract

1. INTRODUCTION

Decoding the chemical structure of dissolved organic matter remains not only one of the most interesting but also challenging analytical tasks. Although the molecular features of DOM have been the focus of a multitude of studies over the last decades,^1–7 the elucidation of its compositional structures and a clear view of DOM isomeric complexity persist as one of the most demandingly difficult analytical problems.^8–12

Nuclear magnetic resonance (NMR)^2,13–18 and hyphenated ultra-high-resolution mass spectrometry (UHRMS)^{9–11,19–22} have been the leading approaches in the structural characterization of DOM. Although most advanced NMR techniques have provided valuable multidimensional information on DOM structural characteristics, the extraordinary molecular complexity of this material is still overcoming the NMR capabilities to resolve discrete molecular structures.^1,13,23 On the other hand, analytical approaches integrating ultra-high-resolution mass spectrometry, gas/liquid separation techniques, and tandem mass spectrometry strategies have provided much of the existing information on the chemical diversity of DOM.^3,9,12,24,25 With the progressive increase in computational power and the high demand in the analysis of complex data, the characterization of DOM at the molecular level has been addressed by molecular dynamics and machine learning approaches as complementary tools to experimental workflows.^26–29

In general, the study of DOM structural complexity using UHRMS has commonly focused on strategies that analyze regular patterns solely based on molecular ions.^{6,7,15,30–35} The van Krevelen-type diagram has been the preferred approach to map UHRMS data from complex samples.^2,36,37 By plotting O/C vs H/C ratios from the molecular composition, it is possible to visualize clusters of compounds that exhibit similar structural characteristics. Despite that lipid, protein, carbohydrate, tannin, lignin, and carboxylic-rich alicyclic (CRAM)-type compound classes have been routinely identified in DOM,^38,39 a structural assignment cannot be accurately provided solely on the basis of chemical formulas.^1,40 Kim et al.³⁶ additionally explored the combination of van Krevelen plots with Kendrick mass defect^37,41,42 to provide structural information based on reaction pathways. For instance, the replacement of 2H by an oxygen atom found along a diagonal of two parallel CH₂ series was suggested as an oxidation pathway of a primary alcohol to a carboxylic acid.

Several parameters derived from chemical formulas are also commonly utilized to predict structural signatures and compositional trends of DOM molecular species. For instance, double bond equivalents (DBE) and the aromaticity index are used to estimate the degree of structure unsaturation and identify aromatic/condensed species in DOM components, respectively.^43,44 Furthermore, the occurrence of several regular patterns in DOM and their potential correlation with families of structurally related compounds has also been reported.¹ However, an explanation on the origin of these regularities and the structural correlation among the compounds belonging to the homologous families has not been yet provided. Reports based on the tandem mass spectrometry of selected molecular ions have shown promise for the identification of DOM structural features.^{4,20,22,45–47}

A Fourier transform–ion cyclotron resonance mass spectrometry (FT-ICR MS/MS) study⁴⁷ of solid-phase-extracted (SPE)-atmospheric organic matter has suggested that structural analogies could exist among the members of a CH₂ homologous series since they share identical neutral losses during collision-induced dissociation (CID). Similarly, several regular patterns in DOM chemical formulas such as the H₂ and CH₂ series and a replacement of CH₄ by an oxygen atom have been found using molecular-level analyses by FT-ICR MS.^23,45 Although the CH₄ vs O pattern has not been clearly explained from a structural perspective,⁴⁸ it has been attributed to potential interchanges of functionalities (e.g., C₂H₅ vs CHO). Moreover, in a different contribution, the same authors⁴⁹ utilized known degradation pathways observed for lignin (a possible component of DOM) to structurally explain newly found repeating patterns in DOM chemical formulas. Interestingly, O₂ and CH₄ vs O₂ regularities found in DOM components were correlated with aromatic ring openings (+O₂) and a combination of aromatic ring openings after one demethylation (–CH₂) and one side-chain oxidation (–H₂), respectively.

New structural insights into the H₂ and CH₂ homologous series from low-molecular-weight compounds of Suwannee River fulvic acid standard using size exclusion chromatography-electrospray ionization-time-of-flight (TOF) tandem mass spectrometry have been reported by These et al.⁴⁸ The similarity found in the fragmentation patterns of homologous isolated precursors (fragments exhibiting the same H₂ or CH₂ difference as their corresponding precursors) suggested that structural dissimilarities among family members presumably lied on their corresponding core structures.

The structural complexity of marine DOM using ultra-high-resolution tandem mass analysis based on an orbitrap MS/MS workflow was explored by Cortés-Francisco et al. ²⁰ Although this study was not oriented to the analysis of structural regularities found in DOM, the potential fragmentation pathways proposed for one of the precursor ions attributed to a lipidlike compound showed the utility of integrating MS/MS data with van Krevelen information to provide new structural understandings of DOM components.

The advantages of DOM analysis using complementary trapped ion mobility spectrometry-FT-ICR MS/MS with correlated harmonic excitation field have been shown,¹² while this work demonstrated the isolation by mobility and tandem MS/MS at the level of chemical formula, its routine application is unviable due to the large number of isomers and isobars present in DOM samples. There is a need for simplified strategies capable of establishing structural patterns based on MS1 and tandem MS/MS information using shorter experimental and processing time scales. In this report, we propose a systematic nominal mass UHRMS/MS followed by a computational model capable of correlating structural features (or families) based on the fragmentation pathways of precursor molecules.

Data-independent acquisition (DIA) is an acquisition strategy in mass spectrometry based on a parallel collection of MS/MS spectra and has recently been utilized to improve the signal-to-noise ratio, reproducibility, and ultimate analyte coverage.^50–52 Recent advances in computing power and electronics have enabled two-dimensional (2D) FT-ICR MS as an emerging DIA tool to analyze complex mixtures.^53–55 The application of an RF pulse sequence to manipulate the ion’s cyclotron radii in the ICR cell,^56,57 along with no ion isolation and ion-neutral collisions (infrared multiphoton dissociation and electron capture dissociation are mostly used), led to the correlation of precursor and fragment ion signals with enhanced resolution and sensitivity. Nevertheless, the presence of abundant scintillation noise⁵⁸ and difficulties associated with data processing are still important limitations that need to be addressed to obtain comprehensive MS/MS data.

The introduction of continuous accumulation of selected ions (CASI) in FT-ICR MS instrument by Senko et al.⁵⁹ provided a way to increase the sensitivity and dynamic range while reducing space charge effects by sequentially transmitting smaller m/z segments. More recently, this strategy has also been implemented in top-down mass spectrometry⁶⁰ and protein imaging,⁶¹ respectively. In the case of DOM analysis, CASI has allowed for the detection of a larger number of chemical formulas when compared to traditional broadband acquisitions.^23,62 Despite the increase in the number of chemical formulas, there is a need for further CASI implementations combined with sequential fragmentation (CASI MS/MS). In the case of complex mixtures analysis, CASI MS/MS workflows can greatly benefit from new computational algorithms for MS/MS data processing and structural correlations.

In this work, we extend the structural characterization of a wetland DOM sample from Pantanal, Brazil, using precursor and fragment molecular ions obtained with electrospray ionization–Fourier transform–ion cyclotron resonance tandem mass spectrometry (ESI-FT-ICR CASI-CID MS/MS). The families of structurally related DOM compounds are identified based on characteristic mass loss patterns across heteroatom classes. We propose a novel graphical analysis of interconnected structural families as a potential tool that helps to understand DOM biogeochemical processes.

2. MATERIALS AND METHODS

2.1. Sample Preparation.

The DOM sample was obtained by the SPE of surface water collected from wetlands located at Pantanal National Park, Brazil. Details on sampling, sample treatment, and the SPE procedure are described by Hertkorn et al.² and Dittmar et al.⁶³ Briefly, 2 L of surface water was collected using HCl precleaned brown plastic bottles. Samples were kept refrigerated on ice and filtered using GFF precombusted glass fiber filters (0.7 μm nominal pore size) within 6 h after collection. Filtered samples acidified to a pH 2 with concentrated HCl were loaded by gravity onto a 1g-Varian Bond Elut PPL cartridge using Teflon tubing. The PPL cartridge was preconditioned with methanol followed by pH 2 Milli-Q water. The loaded cartridge was then rinsed with pH 2 Milli-Q water and dried in a N₂ gas flow for 5 min prior to the elution of DOM molecules with 20 mL of methanol. SPE-DOM extracts were stored in precombusted glass vials at −20 °C until further analysis. The choice of the sample comes from its recent UHRMS and IMS-UHRMS characterization (recent papers^11,12 and a 2016 report by Hertkorn et al.²). The SPE-DOM sample was diluted 10 times by dissolving it in 1 mL of denatured ethanol. All solvents used were of Optima LC-MS grade or better, obtained from Fisher Scientific (Pittsburgh, PA).

2.2. ESI-FT-ICR MS.

A SolariX 9T ESI-FT-ICR MS spectrometer (Bruker Daltonics, MA), equipped with an infinity ICR cell, was optimized for high transmission in the 100–1200 m/z range. Samples were ionized using an electrospray ionization source (Apollo II ESI design, Bruker Daltonics, Inc., MA) in a negative ion mode at 200 μL/h injection. Typical operating conditions were 3700–4200 V capillary voltage, 2 L/min dry gas flow rate, 2.0 bar nebulizer gas pressure, and a dry gas temperature of 200 °C. Operational parameters were as follows: funnel rf amplitude 160 peak-to-peak voltage (Vpp), capillary exit −150 V, deflector plate −140 V, skimmer 1–20 V, transfer line RF 350 Vpp, octupole RF amplitude 350 Vpp, and collision cell RF 1100 Vpp. An arginine cluster ion series (173–1740 Da) was used during the instrument tuning and control optimization. The broadband MS₁ spectrum (first MS dimension) of 115 coadded scans was collected at 4 MW data acquisition size (mass resolution of 4 M at 400 m/z).

2.3. ESI-FT-ICR CASI-CID MS/MS.

For CASI-CID experiments, ions at odd nominal masses were sequentially isolated (1 Da window) in the quadrupole (m/z range 261–477), accumulated for 5–7 s in the collision cell, and subject to CID prior to the analysis in the ICR cell. Multiple CID collision voltages (15–27 V) tailored to the precursor nominal m/z were utilized for a better coverage across low and high m/z fragments. The same ion optic parameters used in the broadband analysis were utilized during the MS/MS experiments. Up to 100 scans were coadded for each tandem mass spectrum (MS₂) in the segmental acquisition mode. Eight predefined segments were acquired and stitched for each experiment using the serial run mode.

2.4. ESI-FT-ICR CHEF SORI MS/MS.

Differences between nominal mass and chemical formula-based MS/MS were evaluated for the case of the 267.087412 m/z ion (C₁₃H₁₅O₆) using correlated harmonic excitation field (CHEF),^12,46,64,65 shots ejection of isobaric ions (~0.002% power and 0.04 pulse length), and sustained off-resonance irradiation (SORI)-CID (1.4% SORI power, 0.1 s pulse length, and −500 Hz frequency offset). A sweep excitation was applied, and 600 MS/MS scans were collected at a 2 MW data size.

2.5. Data Processing.

Data was processed using Data Analysis (v. 5.2, Bruker Daltonics, CA), and all other plots were created using OriginPro 2016 (Originlab Co., MA). The assignment of chemical formulas was conducted using Composer software (version 1.0.6, Sierra Analytics, CA) and confirmed with Data Analysis (version 5.2, Bruker Daltonics). The assignment of formulas was based on the lowest formula errors, the presence of isotopologue signals, and the removal of isolated assignments (deassignment of peaks belonging to classes with only a few sparsely scattered members). The theoretical formula constraints of C_4–50H_4–100N_0–3O_0–25S_0–2, S/N > 3, m/z range 100–900, error <1 ppm and 0 < O/C ≤ 2, 0.3 ≤ H/C ≤ 2.5, and DBE-O ≤ 10 were considered.⁶⁶ The internal walking calibration performed in Composer software using oxygen homologous series (O₄–O₂₀) resulted in an average error <80 ppb for the mass range 229–890 Da. Both odd and even electron configurations were allowed in Data Analysis software. The MS/MS spectra were internally calibrated using a list of exact masses of fragment ions obtained from commonly occurring neutral losses in DOM and their combinations.^8,22 A four-column excel file containing (1) the accurate mass of assigned peaks from both MS2 and MS1 (odd masses m/z 261–477), (2) the isolated nominal mass, (3) the intensity, and (4) the chemical formulas was created as an input file for further data processing using Graph-DOM, an in-house code written in Python 3.7.3.

Ordered fragmentation pathways were computed based on the following equation

$$\mathrm{P}=\left[{\mathrm{N}\mathrm{L}}_1+{\mathrm{N}\mathrm{L}}_2+{\mathrm{N}\mathrm{L}}_3+\dots {\mathrm{N}\mathrm{L}}_n\right]+C$$ (1)

where P corresponds to the chemical formula of the isolated precursor at nominal mass and NL is the neutral species lost during the fragmentation of the precursor and fragment ions. In this study, CH₄, O, H₂O, CO, CH₂O, CH₄O, and CO₂ were considered as potential neutral losses.^20,46,67,68 The sequence $\left[{\mathrm{N}\mathrm{L}}_1-{\mathrm{N}\mathrm{L}}_2-{\mathrm{N}\mathrm{L}}_3-\dots {\mathrm{N}\mathrm{L}}_n\right]$ in eq 1 is an ordered array of neutral losses (NL) generated by an approach similar to the one recently described by Simon et al.⁶⁷ Differently from Simon’s approach, we sequentially match the exact mass of the theoretical NL with the mass difference of two consecutive assigned peaks with 1 mDa tolerance error. The core fragment (C) was defined as the lowest mass assigned fragment in a given pathway. Note that this approach also considers the search of multiple NL if the mass difference between two peaks does not match the accurate mass of a single NL. Due to the large amount of fragment data collected in this study, we set the NL multiple at 2. Nevertheless, the Graph-DOM code allows the user to define both the type of NL and its multiples.

The families of structurally related compounds were identified using a conceptual model (P_n−1 + F_1:n + C), defined in the Graph-DOM code, based on de novo matching of fragmentation pathways. Briefly, a precursor chemical formula along with the full fragmentation pathway is searched across all computed pathways in ascending order of mass. Note that the n – 1 subscript in the model indicates the presence of the P_n−1 precursor as the first fragment of the P_n’s fragmentation pathway (see Figure 3 panel B). The term F_1:n defines the full match condition for all fragments in the pathway to consider a precursor in a family. Cytoscape v.3.8.2⁶⁹ was used to visualize the complexity of DOM in the form of structural networks formed by a neutral-loss-based interconnection of family members. A list of the precursors found in the structural families was imported into Cytoscape and defined as nodes. Structural functionalities based on neutral loss differences among precursors in a family were imported as edges.

Figure 3.

Conceptual models designed to compute ordered fragmentation pathways (panel A) and find structural families in DOM based on sequential matching of fragmentation pathways (Panel B). Note that for the precursor P1 to be considered in a family, its ion mass should match (1 mDa tolerance) the mass of the first fragment in P₂’s fragmentation pathway.

3. RESULTS AND DISCUSSION

The broadband ESI-FT-ICR MS spectrum of the SPE-DOM sample showed a typical distribution of [M – H]⁻ ion signals with a maximum of around 400 m/z (Figure 1A). A section of the spectrum (4 Da mass range) depicts the characteristic DOM pattern of most abundant signals located at every other odd m/z and lower intensity peaks at even m/z (see the inset, Figure 1A). The van Krevelen plot (Figure 1B) obtained after assigning near 4000 molecular formulas showed a dominance of CHO (green) and CHON (orange) heteroatom classes in the region 0.3 < O/C < 0.8–0.4 < H/C < 1.8 attributed to lignin and tannin-type molecules followed by less abundant CHOS (blue) compound classes associated with sulfonated carboxylic-rich alicyclic components.^2,36,39,70

Figure 1.

ESI-FT-ICR MS broadband spectrum of the SPE-DOM sample and expanded view of the m/z range 406–410 shown in the inset (A). Van Krevelen plot obtained after the chemical formula assignment of mass signals with black arrows describing DOM reaction pathways previously suggested by Kim et al.³⁶ CHO, CHOS, CHON, and CHONS compound classes are represented in green, orange, blue, and gray colors, respectively (B). The section of an MS/MS spectrum showing [M – H]⁻ precursor ions isolated at nominal mass 313. Assigned molecular formulas are displayed with heteroatoms indicated with the color code (C). Typical MS/MS spectrum of the precursors isolated at nominal m/z 313 with annotated common neutral losses observed in DOM (D). Note that single peaks showed at nominal masses may comprise an envelope of multiple mass signals. For instance, nine peaks resulting from the CO₂ loss of precursors fragmented at m/z 313 are shown at m/z 269 (Panel D, inset).

A closer view of the nominal mass 313 (Figure 1C) shows the characteristic isobaric complexity of the sample, where up to 14 precursor ions of the CHO, CHON, CHOS, and CHONS classes were coisolated and fragmented. Similar patterns resulting in an average of 10 precursor ions per MS₂ spectrum (total 110 fragment spectra collected) are found across the studied mass range (m/z 261–477). Typical fragmentation patterns showing common DOM neutral losses of CH₄, H₂O, CO, CH₄O, and CO₂ were observed across the fragmentation data set (see the MS₂ profile of precursors isolated at 313 m/z in Figure 1D). Few other less abundant neutral losses associated with sulfur (SO₃) and nitrogen (NH₂OH and HNO₃) species were also observed.

The analysis of potential reaction pathways previously reported for DOM³⁶ and described by black arrows in Figure 1B suggests that compounds found along a pathway (e.g., Redox) in the van Krevelen space are part of a structural family with a potential common backbone. Since structural questions are difficult, if not impossible, to answer solely based on chemical composition obtained from UHRMS, here we explored a fragmentation strategy that will provide new information about the structural complexity of DOM as a complementary tool to the traditional van Krevelen plot.

The application of the ESI-FT-ICR CASI-CID MS/MS workflow resulted in more than 24 000 total assigned chemical formulas (~900 precursors). The CHO constituted the most abundant compound class (80% of all of the precursors assigned), followed by the CHOS (~17%) and CHON (<3%) classes (Figure S1). 2D MS/MS plots generated using all identified molecular formulas (A) and the filtered m/z signals assigned to the CHO, CHON, and CHOS compound classes, respectively (B–D), are shown in Figure 2. A closer view to the panels B–D in Figure 2 confirmed the clear dominance of the CHO compounds during fragmentation (>23 000 chemical formulas) over the less abundant CHON and CHOS compound classes. Consequently, the O-heteroatom class will constitute the main focus of this study.

Figure 2.

2D MSMS plots generated after the chemical formula assignment of ion signals obtained from the FT-ICR CASI-CID MS/MS experiments.

Similar to the 2D mass spectrum described by van Agthoven,⁵⁴ in our 2D MS/MS plot of fragment m/z vs precursor m/z, typical straight lines can be observed. The examination of Figure 2A denotes that data points are aligned over diagonal lines described by eq 2

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{s}\mathrm{o}\mathrm{r}\left(\frac{m}{z}\right)=\mathrm{F}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\left(\frac{m}{z}\right)+\mathrm{n}\mathrm{e}\mathrm{u}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{o}\mathrm{s}\mathrm{s}$$ (2)

The first diagonal line observed (right toward the left) represents the precursor line, and it contains precursor ions. Since the chemical formula assignment was based on accurate mass (error <1 ppm), precursors and fragments can be directly correlated (data points horizontally aligned in the 2D MS/MS domain).^8,20,22 Neutral loss lines are parallel to the precursor line, and the NL mass (relative to the precursor line) can be determined by the intercept of eq 2. For instance, the characteristic line of one H₂O loss (first line from the precursor line in Figure 2B) can be described using the equation Precursor $\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{s}\mathrm{o}\mathrm{r}\left(\frac{m}{z}\right)=\mathrm{F}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\left(\frac{m}{z}\right)+18$. Other typical NLs observed (e.g., CO, CH₄O, CO₂, etc.) and their corresponding multiples can be visualized in the form of their characteristic lines.

The alignment of the MS/MS data along unique NL lines observed in Figure 2A–C evidenced the similarity of DOM fragmentation pathways regardless of the precursor chemical composition. These structural patterns are in good agreement with previous findings obtained from fragmentation experiments of a few selected nominal masses.^8,46,47,71 The systematic occurrence of NL-line patterns in the 2D MS/MS space suggests that DOM molecules are clustered by the families of compounds that could likely share common backbone structures.

The analysis of the complex fragmentation data generated from the FT-ICR CASI-CID MS/MS experiments was performed by designing an efficient data mining approach implemented in the Graph-DOM (Figure 3). The first step consisted of computing all possible ordered fragmentation pathways for the assigned precursors (Figure 3, panel A) using eq 1. For instance, a fragmentation pathway for the precursor C₁₆H₁₉O₉ is described as C₁₆H₁₉O₉ = [H₂O + CO₂ + CH₄O + CO₂] + C₁₃H₁₃O₃. Since NLs are directly correlated with structural functionalities, H₂O, CH₄O, and CO₂ chemical units could be interpreted as CID fragments associated with hydroxyl, methoxy, and carboxylic moieties, respectively. Since a precursor formula comprises a variety of isomeric species, multiple fragmentation pathways (with the same or different core fragments) can be associated with the same precursor.¹¹ Note that the core fragment chemical formula could be interpreted as the backbone of the precursor structure and can also hold isomeric diversity. In the examples shown, the core fragment is limited by the lower m/z experimentally observed (in this instrument and settings, m/z below 100 are not detected).

Over 10⁷ ordered fragmentation pathways were computed for the CHO compound class following the workflow described in Figure 3A. Precursor compounds within the mass range 395–477 exhibited the highest number of fragmentation pathways (>100 000), in agreement with the extensive amount of product ions detected (see Figure S1). An average of 7 million pathways was found for precursor molecules containing 11–13 oxygens in their composition. On the other hand, less oxygenated DOM compounds (8–12 oxygens) generated a larger number of core fragments (Figure S2). The relatively high abundance of fragmentation pathways and core fragments for O-rich molecules suggests that the degree of oxygenation plays a key role in DOM structural diversity.

Assuming that the fragmentation pathway of a precursor C_xH_yO_z is fully matched to the pathway of another precursor C_x+aH_y+bO_z+c, we could presume that they are structurally related. Consequently, the compositional difference between these two precursors will be the chemical unit C_aH_bO_c. Since many of the fragments assigned in the MS/MS spectra are also observed in the MS₁ domain, other precursors will likely show the same behavior as both C_xH_yO_z and C_x+aH_y+bO_z+c. Therefore, they can be grouped into families characterized by an NL-based sequence resulting from the difference in chemical units among precursors.

The computation of structural families of DOM was conducted by implementing the conceptual model P_n−1 + F_1:n + C graphically described in Figure 3B. An overlapping strategy of fragmentation pathways in the form of P = [F₁ + F₂ + ··· + F_n] + C was utilized. The overlap step consisted of matching both the initial lowest mass precursor P₁ and its fragmentation pathway in the database generated from the previous step (Figure 3, panel A) in ascending order of mass. The initial precursor P₁ is further grouped into the family [P₁, P_n−1] with the newly matched precursor P_n−1, and the chemical unit difference NL_Pn−1→P1 is stored as the structural difference between P₁ and P_n−1. The resulting pathway P_n−1 = P₁ + [F₁ + F₂ + ··· + Fn] + C₁ is searched again for a new match, and the loop is repeated until no further match is found. Finally, the family ([P₁, P₂,…, P_n−1, P_n]) is created as an array of the precursors sharing the same fragmentation pathways. The chemical unit difference identified as a neutral loss among precursors within a family represents the functionality that is being added/subtracted to/from the family members. This array of neutral-loss-based moieties illustrates the potential biogeochemical transformation processes experienced by DOM molecules. Once a family is retrieved, a new precursor higher in mass than P₁ is reset as an initial lowest mass precursor and the pathway matching algorithm is repeated until all potential families are computed. Note that since various fragmentation pathways might be common to different precursors, multiple identical structural families will be expected. We define these sequences of analogous precursors as isomeric families, and they are an important indication of confidence during the computation of the families.

The model performance to retrieve CHO structural families (coverage of precursors, intermediate fragments, and core fragments) is described in Figure 4. Although the coverage of precursors in the families was 60%, over 1900 DOM structural families were identified. A higher coverage was found for both the intermediate (~90%) and core fragments (>80%). Note that since the same core and intermediate fragments might be found at different nominal masses, those fragments were counted in the families every time they were linked with a different precursor. These results suggest that there are potential structural families that remain undetected under the current conceptual model. While our workflow provides higher confidence, in grouping structurally related DOM compounds, than previous approaches, there are still limitations associated with the considerations of the proposed model.

Figure 4.

Number of covered precursors, core and intermediate fragments by the model P_n−1 + F_1:n + C (A), distribution of the number of families per family size (B), and families per oxygen class of the uppermost precursor (compound with the highest oxygen content in the family) (C), respectively, for the CHO class.

Structural families containing two-to-four precursor members of the CHO class were the most abundant (261–477 m/z range). A decrease in the number of families was also observed as the family size (number of precursors in a family) increased from four to six members (Figure 4B). Up to five precursors were found in over 300 structural families and the lowest abundant family (<100) contained six DOM compounds. The relatively high number of two-member families (>400) could be attributed to the limited mass range analyzed in the current study, preventing the match of fragmentation pathways from precursors with higher mass (>477).

The number of families per oxygen class of the uppermost precursor within a family depicts a Gaussian-type distribution centered in the O-class 10 (Figure 4C). This pattern is in good agreement with the distribution of pathways and core fragments per O-class found for the CHO compound class (Figure S2). Nevertheless, a closer view of Figure 4C evidenced a shift of the distribution toward less oxygenated family parents and an increase in the number of these uppermost precursors with eight to nine oxygens.

Overall, the families retrieved from the FT-ICR CASI-CID MS/MS data collected in the studied mass range showed that the structural transformation of CHO components in DOM depends on oxygenation/deoxygenation processes driven by both single and mixed addition or subtraction of H₂O, CH₄O, and CO₂ chemical units (Figure S3). This finding suggests that the structural alteration of DOM involves complex mechanisms compared to the uniform trends (e.g., hydration and carboxylation) previously observed from broadband FT-MS data.^36,45 Although our findings are constrained to O-compounds negatively ionized, the proposed approach allows the structural analysis of other molecular classes (e.g., CHOS and CHON) upon the availability of substantial fragmentation data.

A closer view of the compositional relationship among members within a family revealed that oxygenation (increase in the O/C ratio) through the addition of carboxylic (CO₂) moieties increases the unsaturation degree of the resulting species (+1 DBE). Conversely, hydroxyl (H₂O) and methoxy (CH₄O) additions are accompanied by a decrease in one DBE unit of the subsequent molecule.

A 2D MS/MS representation of the molecular transformations exhibited by the structural family [C₁₄H₁₃O₅–C15H13O7–C16H17O8–C17H17O10–C17H19O11] identified in the SPE-DOM sample is shown in Figure 5A. The double arrows placed between family members indicate that the potential biogeochemical transformations of DOM can be viewed from a bidirectional perspective. For instance, a sequential addition (synthesis-like) of carboxylic, hydroxyl, and methoxy moieties (–CO₂, –CH₄O, –CO₂, and –OH) starting from C₁₄H₁₃O₅ up to the family parent C₁₇H₁₉O₁₁ is described in Figure 5A. This successive functionalization of O-depleted low-molecular-weight compounds resulting in high-molecular-weight O-rich molecules could be explained as aging processes. Evidence of an increase in oxidized species observed in relatively old DOM from deep ocean water^19,23 compared to younger freshwater DOM has been previously reported. Similar findings of fresh (¹⁴C dating) DOM exhibiting less oxygenated and lighter molecules compared to older terrestrial DOM species have also been reported by Benk et al.⁷² However, the consistent decrease in unsaturation of oxygenated high-molecular-weight DOM components observed in the previous study contrasts with our results of alternating unsaturation patterns along an ascending-order structural family (e.g., DBE change 8–7-8–7-8 from C₁₄H₁₃O₅ to C₁₇H₁₉O₁₁). A notable increase in O-rich molecules at the expense of the consumption of a poor oxygenated species was also reported in biodegradation experiments conducted on DOM from landfill leachate⁷³ and from the surface of glaciers and ice sheets.⁷⁴ Although the impact of biodegradation on DOM structural transformation was not investigated in this study, the increasing oxygenation trend reported in both contributions is in good agreement with the O-based functionalization found in our structural families. Other abiotic processes such as a photo or chemical oxidation have also been indicated as responsible for the presence of highly oxygenated and CRAM species in DOM,^19,75,76 yet supporting our findings observed along a structural family in the ascending order.

Figure 5.

2D MS/MS visualization of a characteristic DOM family of six precursors (Panel A). Chemical unity (H₂O, CH₄O, and CO₂) differences among precursors are shown using a color code. Fragmentation pathways described as neutral losses are also shown as colored bars. Van Krevelen plot (B) of CHO class compounds obtained from the MS₁ experiment highlighting the compositional nature of the structural family.

The analysis of structural families in the reverse direction (top to bottom) suggests that DOM molecules could also undergo mineralization-like transformations resulting in low-molecular-weight reduced species. For instance, the oxygen-rich family parent C₁₇H₁₉O₁₁ (Figure 5A) experiences defunctionalization processes characterized by consecutive eliminations of H₂O, CO₂, CH₄O, and CO₂, resulting in the poorly oxygenated low-molecular-weight compound C₁₄H₁₃O₅. Interestingly, it has been suggested that highly oxygenated compounds⁷⁷ and aromatic oxidized species⁷⁸ from terrestrial DOM, determined by broadband FT-ICR MS, are preferentially removed by biodegradation, resulting in low-molecular-weight components. Similarly, a significant decrease in aromatic content and oxygen functionalities was observed by Ward et al.⁷⁶ during the photodegradation experiments of soil DOM compared to dark controls. Moreover, Hawkes et al.⁷⁹ have found that hydrothermal environments such as the ones observed in ocean deep hydrothermal vents could induce potential defunctionalization processes (e.g., decarboxylation and dehydration) of O-rich high-molecular-weight DOM species, resulting in less O-functionalized low-molecular-weight components. The results described in Figure 5 illustrate that our model provides useful information that could help to elucidate the complex DOM transformational mechanisms at the structural level.

A representation of the DOM structural family (Figure 5A) superimposed on the van Krevelen space generated for the sample’s CHO class is shown in Figure 5B. The discontinuous line patterns described by different directional vectors representing neutral-loss-based functionalities contrast with the traditional straight lines utilized in the van Krevelen plot to describe chemical transformations and reaction pathways of DOM components deduced from elemental composition obtained from UHRMS data (Figure 1B).^1,36 Therefore, our results suggest that DOM biogeochemical transformation mechanisms are more complex than traditionally described, based upon the heterogeneous nature of the structural information obtained from neutral mass loss patterns observed in this study. For example, DOM molecules assigned from an MS₁ analysis describing a regular addition/subtraction of H₂O chemical units are conventionally interpreted as a family characterized by a hydration/condensation process. Similarly, chemical formulas differing in exactly CO₂ have also been placed into a homologous series resulting from carboxylation/decarboxylation pathways.⁴⁹ However, our findings indicate that CHO compounds in this DOM sample form more complex families characterized by multiple heterogeneous combinations of neutral-loss-based structural moieties (e.g., H₂O, CH₄O, and CO₂) such as the one described in Figure 5. These results illustrate that the integration of efficient computational tools with comprehensive UHRMS fragmentation workflows allows the identification of valuable structural information of DOM components that cannot be accurately predicted by traditional FT-MS workflows.

The visualization of the computed families using Cytoscape confirms the notion that DOM forms a complex assembly of interconnected molecules (Figure 6). Similar results using broadband FT-ICR MS data of DOM from both surface and deep sea²⁶ water samples and from secondary organic aerosols⁸⁰ have been reported.

Figure 6.

View of the three main clusters observed in the network of neutral-loss-based structurally connected DOM precursors for the CHO class. Precursor molecules are described by nodes and the family indexes are shown as edges. An expanded view of 14 interconnected DOM families is shown in the inset. A comprehensive web-based network can be found at https://github.com/Usman095/Graph-DOM.

A closer view of the structural network in Figure 6 revealed three main clusters of related DOM components (red dots) connected by neutral-loss-based structural functionalities (edges). A more detailed analysis of a specific region of the network described in the inset of Figure 6 illustrates that several precursors are common to multiple families. This result, not previously observed at the precursor level, shows the crucial role that structural isomers play in the interconnection of DOM compounds and confirms that isomeric diversity is a fundamental component of DOM molecular complexity. The level of complexity observed in this network suggests that previous elemental-based composition interpretations cannot accurately describe structural patterns in DOM.

In this model, the intersection of structural families relates to the isomeric content of DOM. However, it should be noted that the model may overestimate the number of fragmentation pathways due to the nominal mass CASI CID data collection. The analysis of fragmentation pathways determined by the nominal mass and chemical formula-based MS/MS for the case of the 267.087412 m/z ion (C₁₃H₁₅O₆) showed that all nine fragmentation pathways determined by chemical formula-based MS/MS are also observed in the nominal mass analysis (Table S1). This is an expected result and speaks to the effective processing of the computational code. The nominal mass MS/MS processing resulted in 13 additional fragmentation channels. While some of the additional fragmentation channels (overestimation) can be derived from differences in the fragmentation mechanism (CASI CID vs SORI CID), the application of the model to nominal mass CASI CID MS/MS will inherently carry potential overestimations.

The analytical power of this workflow is based on the fast acquisition of nominal mass CASI-CID data sets from complex DOM samples. The model applied to nominal mass CASI CID MS/MS effectively reports all of the “real” fragmentation pathways. One alternative to reduce the workflow overestimation is to utilize chemical formula-based MS/MS, but this approach is unpractical for routine DOM analysis. A more viable alternative is the implementation of complementary artificial intelligence and machine learning approaches trained with small subsets of chemical formula-based MS/MS data from DOM samples.