We assessed extraction efficiency by measuring DOC concentrations in the bulk seawater sample and permeate of each replicate. Detailed results are shown in Figure S2 and Table S1. The input seawater had a concentration of approximately 70 μmol L⁻¹. For the 0.2 g cartridges, the highest extraction efficiency calculated based on the permeate concentration was 54.4 ± 1% (mean ± SD) and was observed for the 100 mL samples, followed by 47.4 ± 3% and 42.3 ± 1% for 500 and 1,000 mL respectively. The 1.0 g cartridge extracted more DOC from 1,000 mL (53.1 ± 2.3%) than the 0.2 g cartridge, as would be expected. The data also suggested that the reactive sites on the 0.2 g bed mass cartridge became saturated at seawater volumes ≥500 mL. Extraction efficiencies calculated using the concentration of DOC that was eluted from the resin exhibited the same trend but were lower than those calculated from permeate concentrations. This is likely due to the fact that not all compounds extracted onto PPL resin could be eluted with methanol, but could also result from sample processing following the elution (e.g., drying). The approximate amount of injected carbon ranged from an average of 0.19 μmol for the 100 mL samples to 0.83 μmol for 500 mL to 1.48 μmol for the 1,000 mL samples extracted with 0.2 g PPL and 1.86 μmol per injection for the 1,000 mL samples extracted with 1 g PPL. The reported accuracy of injection concentrations is linked to the concentration of DOC in permeates of each SPE extraction and the dilution volume.

The simplest assessment of DOM MS data was performed by comparing overall intensities of total ion currents (TIC) (Figure 2). The featureless TIC, regardless of injected concentration, highlighted the overall complexity of the PPL-extracted DOM sample, and thus the depth of the analytical challenge faced by DOM analysts. As expected, the two 1,000 mL samples, which had a higher concentration, gave the highest response. The intensity of their TICs only differed in the early retention time range, where the 1,000 mL sample extracted onto 0.2 g of PPL resin showed a lower response. This difference indicated that polar compounds were less efficiently retained as the resin became saturated and is a response that is consistent with the hydrophobic nature of PPL.

The overall number of features (defined as a single peak in an extracted ion chromatogram; XIC) in each sample was identified based on thresholds and deconvolution settings in the software tool used (MZmine2) and the strict requirement that features must appear in the XIC of 4 of 5 replicates. After blank subtraction, it was found that the number of features increased with injection concentration and that more features were identified in positive mode than negative mode. The overall sum of features observed in positive and negative mode across all groups was 13,987 and 7,328. The number of positive (negative) features (Figure S3) increased from 1,235 (465) in the lowest concentration samples (100 mL extraction volume) to 5,653 (2,715) of the medium concentrated samples (500 mL extraction volume) to 7,766 (3,408) and 8,167 (4,217) features in the highest concentrated samples (1,000 mL/0.2 g and 1,000 mL/1 g) respectively. These data highlight the concentration dependency of the mass spectrometer, which needs to be carefully controlled when performing comparative analyses. The feature tables are available in the Supporting Information of this article.

The two samples with the highest extraction efficiency (100 mL/0.2 g and 1,000 mL/1 g) also had the highest proportion of unique MS features in positive mode (45 and 41%, respectively, compared with 22 and 33% for the other two samples, absolute values shown in Table S2). The trend was similar for features identified in negative mode as well but all samples exhibited a higher proportion of unique features in negative mode. The observed similarity of the two samples with high extraction efficiencies may have been a coincidence, but is not unexpected given that polar compounds, for example, were extracted more efficiently if the PPL column was not saturated. Besides differences in sample composition, different injection concentrations affect signal/noise ratios and can result in concentration dependent changes in peak shape, which can lead to retention time misalignment between different sample groups and thus more unique features. More relaxed retention time tolerances or stricter filtering toward minimum repetition of features will likely alleviate this problem. If samples are likely to be low in DOC concentration then injecting a short dilution series will be useful.

From a global dataset perspective, the number of features with assigned formulas within a 5 ppm mass error tolerance increased with injected sample concentration. This finding is consistent with the hypothesis that low S/N ratios will be omitted from the molecular formula assignment calculation at lower injection concentrations. For instance, the number of features with assigned formulas within a 5 ppm mass error tolerance is slightly lower for the smaller extraction volumes (shown in Figure 3B). This can be explained by the assumption that some features were either noise or had a molecular composition which was not considered through our calculation parameters (e.g., contained elements other than C, H, O, N, and S, Na, Cl).

To increase the confidence of molecular formula annotation, we made our data analysis more stringent through the alignment of XICs from both positive and negative ionization modes. This alignment groups different ion species from the same molecule e.g., adducts (M+H+, M+Na+, M-H+, and M+Cl−). If two or more matching ion species were aligned, a consensus molecular formula (the highest ranked common formula) was created. The overall number of consensus features of all volume groups was 3,060 and resulted in 2,600 molecular formulas (shown in the Supporting Information). To display the chemical space of the molecular formulas observed in the different groups, we created Van Krevelen diagrams, displaying the H/C vs. O/C ratios of the molecular formulas, differentiated between positive mode, negative mode and consensus formulas (Figure 3A).

In order to loosely categorize “likely” and “unlikely” H/C and O/C ratios in DOM, we also mapped out the distributions of previously characterized metabolite formulas in Van Krevelen space using structures < 500 Da from the Supernatural database (Banerjee et al., 2015), which includes molecular formulas from more than 230,000 natural products. The Supernatural Van Krevelen plot and density plots of H/C and O/C ratios are shown in Figures 3C,D. The highest density of molecules in the Supernatural database is constrained by H/C ratios between 0.5 and 2.5 and O/C ratios between 0 and 0.5. There is a striking resemblance between this area and the hotspot region of H/C vs. O/C ratios linked to consensus (or shared) features between ionization modes and between DOM sample groups (Figures S7, S8). As such, we have higher confidence in molecular formula annotations for DOM features that appear in these Van Krevelen regions. Many masses resulted in multiple possible molecular formulas within our mass tolerance constraints, and our analysis displays only the best match (highest ranked consensus formula for various adducts). Therefore, the elemental ratio information provided by our high confidence region is more appropriate for describing a statistically significant global sample composition. However, such an approach may not be suitable for confident annotation of particular ion species. Furthermore, post-production modification of metabolites in seawater will likely drive some H/C and O/C ratios away from those predicted by a metabolite database (e.g., loss of unsaturation, photo-oxidation; Arakawa et al., 2017), and so, while the comparison with the Supernatural database is useful for constraining our dataset, we cannot exclude the remaining molecular formulas without further scrutiny of the data.

Statistical comparison to known compounds from large scale libraries can help to increase the confidence of molecular formula assignments. The frequency of individual molecular formulas in the supernatural database provides an empirical basis to demonstrate that many molecules often share the same molecular formula, but not the same chemical structures. The frequency of molecular formulas in this database ranges from many unique to several 100 redundant formulas, with an average of 5.5 structures per molecular formula. Chromatographic retention times (e.g., Figures S4, S5) could be used to differentiate between individual molecules with identical formulas. However, absolute and relative retention times often vary between studies due to differences in LC gradients, HPLC column type, and instrument conditions. Thus, relying on MS1 features, even when molecular formula assignments are robust, is not a suitable characteristic for compound library searches, unless the standardized methods are used for library generation and sample analysis.

For this reason, we examined the efficacy of tandem MS in data driven acquisition (DDA) mode, to provide molecular level information for DOM in an untargeted context. Unlike MS1 feature detection and molecular formula assignment, MS/MS spectra do not depend on chromatographic reproducibility and allow comparison between different instrument platforms if the same fragmentation methods and similar fragmentation energies are used. Tandem mass spectra matching to library entries are considered level two annotations according to the 2007 metabolomics standards initiative (Sumner et al., 2007) and probably the most effective approach for high-throughput annotation of small molecules in complex samples. In our workflow, we acquired MS/MS spectra in top 5 DDA mode, in which up to five precursor ions are submitted for MS/MS after each MS1 survey scan. With a medium resolution (R_{m/z 200} = 17.000), the iterative scan cycles result in several thousand spectra per sample. To reduce redundancy and computation time we clustered identical spectra to consensus spectra (Frank et al., 2008) and searched them against several spectral libraries, including GNPS (Wang et al., 2016), MassBank (Horai et al., 2010), ReSpect (Sawada et al., 2012) HMDB (Forsythe and Wishart, 2009), and NIST 2014 (Stein, 2014). Hereby a strict precursor mass tolerance of m/z 0.01 was used first, followed by a maximum allowed precursor delta mass of up to m/z 100, with the intention to annotate putative analogs. Besides dereplication of known compounds from the libraries, an additional multiple spectral alignment of all clustered spectra was performed. Thereby, all spectra are compared to each other using cosine similarity scoring (Stein and Scott, 1994) and spectrum-spectrum matches, with a cosine score higher or equal to 0.7 are connected as “nodes” in a network (Watrous et al., 2012; Wang et al., 2016). If nodes yielded a database hit during a spectrum library search, they can be labeled as specific compounds and nodes around the “hit” can be assumed to have similar chemical scaffolds. If the delta masses, fragmentation patterns and chemical formulas of the database hit are carefully interpreted, putative structures of previously unknown spectra can be proposed with a certain confidence.

The spectral network of all samples acquired is shown in Figure 4A. Nodes that contained MS/MS spectra from blanks were subtracted from all samples. Depending on the origin of clustered spectra, nodes were labeled gray (100 mL, 0.2 g PPL), orange (500 mL, 0.2 g PPL), green (1,000 mL, 0.2 g PPL), and red (1,000 mL, 1 g PPL) if spectra came exclusively from this group. If a node contained spectra from more than one group, it was considered as shared feature and was labeled blue. The size of the nodes was increased proportionally to the number of samples which contained a particular MS/MS spectrum. Nodes with a library hit were shown with a diamond shape with a thick border if the precursor matched within 10 ppm, and as a thin line for hits observed with higher precursor delta masses (from analog search). Overall, we observed 6,272 nodes. Of these only 32 had an associated positive library hit. We also evaluated MS/MS information in the context of extraction volume and total injected carbon, by examining the total number of nodes contributing from each concentration/extraction group, as well as by tallying the number of unique nodes per group (Figure S6 as well as in a Table S2). The highest number of both total nodes and unique nodes were identified in the lowest concentration sample group (0.19 μmol injected C). The remaining samples had a smaller number of nodes and unique nodes, but the 1,000 mL/1 g PPL group had an elevated number of both types of nodes when compared with the other two concentration groups. This trend with respect to injection concentration was different from that observed for feature number and unique feature numbers as determined from MS1 (Figure S6, right). In that case, the number of both node types increased with injection concentration. In order to generate a metric for assessing the quality of the data resulting from differences in injection concentration, we investigated spectral annotation rates (Figure S6). Spectral annotation rates were low across all groups but the lowest annotation rate was found for the lowest injection concentration (100 mL extraction volume). Thus, while the number of nodes (total and unique) was highest for the lowest concentration injection, these nodes were annotated with lower efficiency. In addition, this sample group also showed a high number of self-looped nodes (i.e., many spectra were not connected to other nodes). The highest concentration injection groups (1.48 and 1.86 μmol injected C) had fewer self-looped nodes. Together, these observations suggest that the quality of MS/MS spectra generated from the lowest concentration group was so poor (low S/N) that even identical spectra could not be confidently identified. Thus, we propose the use of spectral annotation rate and number of unique nodes (Figure S6, left) as a quality control check, to determine whether certain samples in a dataset are likely to be poorly represented because they were injected at too low of a concentration.

The overall library annotation rate of the dataset was at 0.5%, and for the two 1,000 mL (1.48 and 1.86 μmol injected C) groups around 1%. The low annotation efficiency is likely a result of a combination of factors, all of which need to be addressed in future work.

One important reason for the low annotation efficiency is linked to the fact that more than one molecule is isolated prior to fragmentation in the collision cell, which results often in chimeric spectra (i.e., DOM MS/MS spectra are often a combination of fragments from multiple molecules with very similar masses that could not be separated by the unit resolution of the quadrupole, and naturally, yield lower matching scores to library MS/MS spectra). The general field of metabolomics has been grappling with the issue of chimeric spectra. Besides technical improvements in chromatographic or gas phase precursor separation (multi-dimensional chromatography or ion mobility), repetitive or large scale analysis of different samples could help to bypass this problem. Here, a possible solution could be an alignment of numerous chimeric spectra and searching for consensus fragments. However, the bioinformatic tools for the detection (Lawson et al., 2017) and separation of chimeric spectra are still in development. In addition, low annotation rates also result from the low number of available library spectra that match compounds in DOM. Nevertheless, the reported annotation rate is in a similar range as other non-targeted metabolomic datasets (Bouslimani et al., 2015; Petras et al., 2016; Floros et al., 2017) that are typically less complex than DOM.

This still poses a universal barrier to untargeted MS/MS analyses. This problem is independent from the data acquisition parameters and can be solved by expanding the chemical space in spectral libraries. Community driven databases such as GNPS (Wang et al., 2016) provide a good platform for researchers to contribute to the growth of spectral library knowledge, much like how nucleotide or protein databases were established in recent years (Bateman et al., 2004; Sayers et al., 2012).

Besides looking for perfect matches, spectral libraries were searched for spectral similarity/analog hits (the same way sequence libraries are used to search for homolog hits with BLAST Altschul et al., 1997). The analog search was set to allow precursors differences of up to m/z 100. Taking the resulting putative analogs into account, we could increase the overall annotation rate to 13.7% and to more than 20% for the two highest injection concentration groups. As a caution, we know that removing the stringent precursor mass filter can increase the rate of false positive hits. Therefore, analog hits should always be considered as putative annotations requiring subsequent statistical or ideally, manual validation. Assuming that the spectral networking approach connects similar spectra to subnetworks, then true library hits in a network should be surrounded by analog hits.

Following this logic, we could observe a subnetwork of amino-sugars (Figure 4B) containing one node with a library hit to N-acetyl-glucosamine, which is surrounded by several other analog hits to amino sugars or glycans. The inclusion of the library hit increases our confidence in the analog annotations in this subnetwork of an amino-sugar molecular family. The finding is in line with evidence from other studies showing that amino sugars are highly abundant in marine DOM (Ogawa et al., 2001; Aluwihare et al., 2005; Davis et al., 2009). In other subnetworks acetamido-oxohexanoic acid could be identified and a node connected to it was annotated as well as acetamido-oxohexanoic acid, but with a delta mass of m/z 48.036. Furthermore, we could identify subnetworks with several hits to the marine natural product B27A19 from a marine sponge library, and another molecular family of acyl-carnitines, both of which contained analog hits to the same compounds in the subnetworks. All spectrum library mirror plots of the above mentioned database hits are shown Figures S9–S11.

To further evaluate the effect of concentration on spectral quality, we examined the distribution of MS/MS library matches across the different injection concentrations. Acetamido-oxohexanoic acid, for example, was found in at least 4 of 5 replicates across all samples except for the lowest concentration injections. N-Acetyl-glucosamine and B27A19 were present in the three highest concentration samples as well, but only occurred in some of the replicates. The underlying reason could be the automatic triggering of MS/MS acquisition through data dependent acquisition (DDA). If a precursor ion is not among the most abundant ions in a survey scan then it will not be selected for subsequent MS/MS. By using a dynamic exclusion list for precursors that had already been submitted, we assumed that DDA would consider most of the ions, depending on the complexity of a given time point in the LC run, and the scan speed of the mass spectrometer. Nevertheless, for medium and low abundant compounds, the machine might not have had enough time to acquire MS/MS scans, and for some compounds, small shifts in chromatographic profiles may have changed the order of MS/MS selection, which triggered MS/MS acquisition in some but not all samples. Repeated measurements can alleviate this bias and increase annotation rates.

To test the initial assumption of intensity dependency, we plotted the sample frequency (number of samples contributing to one consensus spectrum) of all network nodes, shown in Figure 5. Only around 110 consensus spectra where found in all four groups (20 samples, grouped into five replicates of four different injection concentrations). The precursor intensity, plotted on the second y-axis, shows that these ions are among the most abundant compounds in DOM extracts (Figure 5). Looking at some of the chemical IDs, we observed that the fatty acid glycerin ester mono-laurine and the terpene dihydroactinidiolide, which shares structural similarity with the head group of beta-carotene, were present in all samples. In this context, it is very interesting that oxidation products of beta-carotene have been recently described as abundant in refractory DOM (Arakawa et al., 2017).

If we inspect the nodes that occur at lower frequencies, we can see that the average precursor intensity decreases accordingly, and that only around 10% of the nodes are shared by 10 or more samples (mainly from the 1,000 mL groups with 1.46 and 1.86 μmol injected C). Of the low abundance compounds, more than 50% of all nodes are only found in one sample. Pantothenic acid, a cofactor involved in fatty acid and secondary metabolite biosynthesis, provides an example of such a compound. Next to the above described problem of precursor selection, chimeric spectra, as noted previously, may be another reason for the high number of unique nodes. Chimeric spectra can result in an artificial diversification of MS/MS patters, which then appear as unique nodes. The repetitive analysis of the same sample can help to bypass this problem, for example through an alignment of chimeric spectra and search of consensus fragments. However, the bioinformatic tools for the detection (Lawson et al., 2017) and separation of chimeric spectra are still being developed in our data analysis workflow. Another missing piece of the software aided data analysis is the connection of MS1 and MS/MS features. Due to lower sensitivity of MS/MS and the overall higher numbers of MS1 features, our results indicate that for a comprehensive analysis of DOM, we need to merge MS1 feature tables and the networking database dereplication output. Currently, this is typically done either manually or with simple search functions in R or excel. The matching often results in difficulties, especially if precursor masses differ slightly from averaged MS1 masses from XICs or if the MS/MS event was triggered too far off the apex of the chromatographic peak. A more streamlined approach for large scale data analysis would likely involve implementing an MS/MS scan number assignment to XICs during MS1 feature extraction. Once established, this link could be used downstream during data analysis to link network nodes back to MS1 features and to reduce redundancy in the network, providing the most comprehensive molecular characterization of DOM. Most recently, such an approach has been described for MZmine2 (Olivon et al., 2017), but the software implementation and usage of the workflow is not yet straightforward. Recently, we tested a more user-friendly graphical software module directly implemented in MZmine2, which is now available within the latest version (MZmine2.28). We caution that the evaluation of this tool is still ongoing.

Surface water from the Scripps Pier (La Jolla, USA) was collected with bucket and transported in a 20 L PTFE-carboy on February 2nd 2017 (10:05:39 (PST): temperature 14.58°C, chlorophyll a 0.47 μg/L; salinity 33.18 g/kg). The seawater sample was filtered through a membrane filter (0.2 μm pore size; Acropak, Pall) and adjusted to pH 2 with hydrochloric acid (38% p.a., LCMS trace metal grade, J.T. Baker, USA). The well mixed filtrate was poured into combusted glass bottles. Five replicates each of 100, 500, and 1,000 mL were extracted through 0.2 g bed mass PPL cartridges as well as five replicates of 1,000 mL through 1 g PPL cartridges.

Before use, the cartridges were rinsed and activated with one cartridge volume of methanol (LC-MS grade, Fisher Chemical, Belgium) and refilled with methanol for conditioning overnight (see Figure S1). Afterwards, the cartridges were rinsed with two cartridge volumes of water (LCMS grade, J.T. Baker, USA), two cartridge volumes of methanol and two cartridge volumes of water at pH 2 (acidified with HCl). For extraction, the filtered and acidified seawater [100 ml (0.2 g), 500 ml (0.2 g), 1,000 ml (0.2 g), 1,000 ml (1.0 g)] was gravity passed through each PPL cartridge with a flow rate below 5 ml min⁻¹. Subsequently, remaining salt was removed with three cartridge volumes of pH 2 water. After drying with inert pure nitrogen gas, DOM was eluted with 3 ml (for 0.2 g PPL) and 6 ml (for 1.0 g PPL) of methanol into combusted glass vials. After taking an aliquot of extract for determining extraction efficiencies by direct concentration measurements (see below) the extracts were dried down with a vacuum centrifuge (Centrivap, labconco) and transferred in an amber glass GC vial. The extract was dried down again and stored at −20°C until later dissolved in 100 μL methanol for LC-MS/MS analysis. All glassware used during sampling and sample treatment was pre-combusted for 4 h at 450°C. All other materials were cleaned with acidified ultrapure water and rinsed with sample water before use. For verification of possible contamination, procedural blanks for each volume and cartridge size were performed by repeating the sample preparation with water (LC-MS grade, J.T. Baker, USA).

DOC concentrations were analyzed by high-temperature catalytic combustion using a TOC-VCPH/CPN Total Organic Carbon Analyzer equipped with an ASI-V autosampler (Shimadzu, Japan). Standard solutions ranging from 10 to 100 μmol C L⁻¹ were used for calibration and Deep Atlantic Seawater reference material (DSR, D. A. Hansell, University of Miami, Florida, USA) as well as Deep Pacific Seawater (CCE P1604) and Scripps Pier Water reference material were measured to control for instrumental precision (1 μmol L⁻¹) and accuracy (1.5 μmol L⁻¹). Aliquots of the acidified filtrate (pre-extraction) were sampled for quantification of DOC. To calculate DOC concentrations of extracts, 250 μl of the methanol extracts were isolated based on weight and evaporated overnight at 50°C before re-dissolving in 15 ml ultrapure water at pH 2 for DOC analysis. Additionally, the last 40 mL of permeate from each SPE extraction was taken to determine the DOC concentration of the PPL flow through (post-extraction). Extraction efficiency was then calculated by subtracting this permeate DOC concentration from the pre-filtered DOC concentration in the seawater entering the SPE column.

DOM samples were re-dissolved in 100 μL methanol and 1% formic acid of which 10 μL were injected into a ultra-high performance liquid chromatography (UPLC) system coupled to a Q-Exactive orbitrap mass spectrometer (Thermo Fisher Scientific, Bremen, Germany) in three independent runs, first in high resolution positive mode, then in high resolution positive DDA MS/MS mode and finally in UHR negative mode. For the chromatographic separation, a C18 core-shell column (Kinetex, 100 × 2 mm, 1.8 um particle size, 100 A pore size, Phenomenex, Torrance, USA) with a flowrate of 0.5 mL/min (Solvent A: H2O + 0.1% formic acid (FA), Solvent B: Acetonitrile (ACN) + 0.1% FA) was used. After injection, the samples were eluted during a linear gradient from 0 to 0.5 min, 5% B, 0.5 to 8 min 5 to 50% B, 8 to 10 min 50 to 99% B, followed by a 2 min washout phase at 99% B and a 3 min re-equilibration phase at 5% B. For positive mode measurements, the electrospray ionization (ESI) parameters were set to 52 L/min sheath gas flow, 14 L/min auxiliary gas flow, 0 L/min sweep gas flow and 400°C auxiliary gas temperature. The spray voltage was set to 3.5 kV and the inlet capillary to 320°C. 50 V S-lens level was applied. MS scan range was set to 150–1,500 m/z with a resolution at m/z 200 (R_{m/z 200}) of 140,000 with one micro-scan in positive mode. The maximum ion injection time was set to 100 ms with automated gain control (AGC) of 1.0E6. MS/MS spectra were recorded in data dependent acquisition (DDA) mode. Both MS1 survey scans (150–1,500 m/z) and up to 5 MS/MS scans of the most abundant ions per duty cycle were measured with R_{m/z 200} of 17,500 with one micro-scan in positive mode. The maximum ion injection time was set to 100 ms with automated gain control (AGC) targets set to 1.0E6 for survey scans and 3.0E5 for MS/MS with minimum 10% C-trap filling. MS/MS precursor selection windows were set to m/z 1. Normalized collision energy was set to a stepwise increase from 20 to 30 to 40% with z = 1 as default charge state. MS/MS experiments were triggered at the apex of peaks within 2–15 s from their first occurrence. Dynamic exclusion was set to 5 s. Ions with unassigned charge states were excluded from DDA as well as isotope peaks. For negative mode measurements, the electrospray ionization (ESI) parameters were identical to the positive mode measurements besides the polarity switching and an adjusted spray voltage. MS scan range was also set to 150–1,500 m/z with R _{m/z 200} of 140,000, one micro-scan and maximum ion injection time of 100 ms with an AGC target of 1.0E6.

Thermo.raw datasets were converted to.mzXML in centroid mode using MSConvert (Chambers et al., 2012). LC-MS/MS.raw and.mzXML data can be found on the Mass spectrometry Interactive Virtual Environment (http://massive.ucsd.edu/) with the accession number MSV000080562.

As the first step of data analysis MS1 feature extraction was performed with MZmine2 (Pluskal et al., 2010). For both positive and negative mode data, mass detection was performed with a signal threshold of 3.0e4 and 0.6 s minimum peak width. Extracted ion chromatograms were built with a minimum peak height of 9.0e4 and a relative mass tolerance of 5 ppm. Chromatographic deconvolution was performed with the baseline cutoff algorithm with a baseline level of 3.0e4 and a minimum peak height of 9.0e4. The maximum peak length was set to 2 min. For isotope peak grouping, mass and retention time tolerances were set to 5 ppm and 0.1 min respectively. For alignment of extracted ion chromatograms (XICs) between samples, the same mass and retention time tolerances as above were used. After alignment, only XIC which contained at least 2 isotope peaks and which occurred at least 4 times out of 5 replicates were further considered. The aligned peak list was further filtered with a duplicate peak filter in which duplicate XIC within 5 ppm mass windows and 0.1 min retention time windows. All peak areas of XICs as well as binary information of XIC (present/not present) can be found in the Data Sheet 2, positive-mode, negative-mode and consensus XICs.

Molecular formulas of MS1 features (XIC) < 500 m/z were calculated with an in-house R script applying the Rdisop Bioconductor package (https://bioconductor.org/packages/release/bioc/html/Rdisop.html), implementing the method developed by (Böcker et al., 2009). The generated formulas were filtered by the following rules: N rule, O/C ratio ≤ 1, H/C ratio ≤ 4 and >0, element counts: C ≤ 100, H ≤ 200, O ≤ 80, N ≤ 10, S ≤ 2 modified from (Cortés-Francisco et al., 2014) and our inspection of ~233.000 molecular formulas obtained from the supernatural structure database (Banerjee et al., 2015), and maximal mass error of 5 ppm. Resulting molecular formulas were scored by the method described by (Böcker et al., 2009) and the top1 hit was used for downstream statistical analysis. The code to perform the above formula calculations and filtering is available as a jupyter notebook at https://github.com/DorresteinLaboratory/adductMatchAndFormPrediction.

Following the filtering process presented above, molecular formulas were subsequently filtered by presence of matching ions detected in both positive and negative modes. For that, the most common ESI ion species (Huang et al., 1999) ([M+H]⁺, [M+Na]⁺ for positive mode and [M-H]⁺ and [M+Cl]⁻ for negative mode) were searched in each ionization mode, matched, and tagged in the data. Calculated molecular formulas can be found in the Data Sheet 2.

MS/MS spectra were analyzed with GNPS (Wang et al., 2016). Therefore, the data was filtered by removing all MS/MS peaks within a 17 Da window of the precursor m/z and MS/MS spectra were filtered by choosing only the top 6 peaks in 50 Da windows. The data was then clustered with MS-Cluster (Frank et al., 2008), with a precursor mass tolerance of 0.01 Da and a MS/MS fragment ion tolerance of 0.01 Da. A spectral network was then created with a minimum spectral similarity of cosine 0.7 and more than 4 matched peaks. Only the top 10 edges connecting one node were kept in the network. Consensus spectra were searched against the GNPS spectral library as well as Massbank, ReSpect, HMDB, and NIST14 (Forsythe and Wishart, 2009; Horai et al., 2010; Sawada et al., 2012; Stein, 2014; Wang et al., 2016) with a precursor mass tolerance of 0.01 Da and a MS/MS fragment ion tolerance of 0.01 Da as well as in analog search mode with as maximum precursor delta mass of m/z 100. Library hits discussed in this article were inspected manually and mirror plots of spectrum library matches are shown in the Supplemental Information. Finally, spectral networks were the visualized in Cytoscape 3.4 (Shannon et al., 2003).

All LC-MS/MS data can be found on the Mass spectrometry Interactive Virtual Environment (MassIVE) at https://massive.ucsd.edu/ with the accession number: MSV000080562

Molecular Networking Data and all results of the Spectra Library Comparison can be found at the Global Natural Product Social Molecular Networking (GNPS) website with the links:

http://gnps.ucsd.edu/ProteoSAFe/status.jsp?task=e89fe14eddbb43b5baa2b3d65c257661,

http://gnps.ucsd.edu/ProteoSAFe/status.jsp?task=81b58b4ec67145b490cb0490f39165aa,

The code to perform ion species matching and molecular formula calculation is available at:

https://github.com/DorresteinLaboratory/adductMatchAndFormPrediction.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.