Skip to main content
NCBI home page
ACS Environmental Au logoLink to ACS Environmental Au
. 2025 Aug 5;5(6):573–582. doi: 10.1021/acsenvironau.5c00062

Machine Learning-Assisted Recognition of Environmental Sulfur-Containing Chemicals in Nontargeted Mass Spectrometry Analysis of Inadequate Mass Resolution

Brian Low †, Tingting Zhao †, Xingfang Li ‡,*, Tao Huan †,*
PMCID: PMC12635934  PMID: 41277996

Abstract

Sulfur (S)-containing compounds can be unambiguously identified by their distinctive isotope patterns in mass spectrometry (MS) when the instrument has a mass resolution exceeding 500,000. However, many environmental research laboratories that perform nontargeted analysis rely on high-resolution mass spectrometry (HRMS) instruments, such as quadrupole time-of-flight mass spectrometry (QTOF MS). These HRMS instruments typically operate at a mass resolution of less than 50,000. At such limited resolution, confidently recognizing sulfur isotope patterns is challenging. This work develops a machine learning (ML) strategy for recognizing and predicting the number of S present using HRMS at a mass resolution as low as 25,000. We benchmarked our ML strategy on experimental data, where 200 S-containing standard compounds were mixed into complex environmental samples. In positive electrospray ionization (ESI) mode, our ML strategy achieved accuracies ranging from 87.4 to 95.0% for S recognition and accuracies ranging from 86.3 to 96.6% for S number prediction. Notably, the ML method performed similarly well in negative ESI mode. Our ML strategy was further evaluated on an external experimental water dataset where it correctly recognized the presence of S for all 24 previously reported 2-mercaptobenzothiazole disinfection byproducts (DBPs). The developed ML strategy was implemented into SulfurFinder, an R program, to facilitate automated data cleaning, S recognition, and S number prediction in HRMS data. SulfurFinder combined with HPLC-HRMS analysis of a wastewater sample tentatively identified 169 potential S-containing features. Of these, three were confirmed as S-containing pharmaceuticals. An additional S-containing drug was also putatively annotated using molecular networking. The development of SulfurFinder significantly boosts the capability of conventional HRMS to address the challenge of S recognition in the era of exposomics, supporting a wide range of environmental applications.

Keywords: high-resolution mass spectrometry, inadequate mass resolution, isotope patterns, machine learning, nontargeted analysis, sulfur-containing chemicals, exposome

graphic file with name vg5c00062_0009.jpg

graphic file with name vg5c00062_0007.jpg

1. Introduction

Sulfur (S)-containing organic compounds are frequently detected in environmental samples, such as water and soil. , They come from heavy usage in the industrial and pharmaceutical sectors. These S-containing compounds play critical roles as additives in the gas, oil, and rubber industries and have been found to leach into surface waters. Although known S-containing metabolites are only a small fraction of most chemical databases, they are almost 23% of the DrugBank database, highlighting their bioactive roles. In particular, the presence of S-containing contaminants in source water has led to the growing concern of human exposure to these chemicals through drinking water. For example, S-containing insecticides, such as fipronil, are frequently detected in drinking water and have been linked to the disruption of human thyroid development. Harmful S-containing compounds in drinking water may also be a result of disinfection byproducts (DBPs) formed from water treatment. − Although a few are regulated, the primary driver of cytotoxicity in drinking water is from unregulated DBPs. In contrast to known contaminants, these emerging DBPs are “unknown unknowns” that underscore the urgent need for innovative analytical strategies.

Fourier-transform ion cyclotron resonance mass spectrometry (FT-ICR MS) has been the method of choice for the detection of S-containing compounds in complex samples. It has been widely used to study S-containing compounds in biological and environmental applications. ,− This is owing to its ultrahigh mass resolving power, sub-ppm mass accuracy, and low signal fluctuations. , With these instrumental conditions, it is possible to annotate S-containing molecules based on accurate mass and isotope patterns. In particular, the characteristic mass difference (+1.9958 Da) and relative intensity (∼4.5%) between 34S/32S isotopic peaks can be accurately measured in FT-ICR MS for unambiguous S recognition. However, the high costs and demanding maintenance requirements of FT-ICR MS have limited their adoption in routine analyses. Liquid chromatography (LC) coupled with high-resolution mass spectrometry, such as quadrupole time-of-flight (QTOF), are more commonly used platforms for the nontargeted analysis of environmental samples. However, these mass analyzers are typically operated at a mass resolution of less than 50,000. Adjacent isotopologues cannot be resolved from each other at such resolution, convoluting the isotope pattern needed for S recognition. The non-negligible signal fluctuations in both the mass and intensity dimensions further complicate the annotation of S-containing compounds.

To overcome the above-mentioned limitations of HRMS systems with inadequate mass resolution, we propose a machine learning (ML) strategy to recognize S in small compounds (m/z ≤500) using their isotope patterns. This idea was inspired by previous work where ML was used to recognize uncommon elements based on their unique isotope patterns in MS data. − However, using isotope patterns for S recognition is more challenging than recognizing halogenated compounds. This is because the abundance of 34S is significantly lower than 37Cl and 81Br, resulting in more subtle changes to the isotope pattern. Thus, we addressed this issue in this work by developing a sequential ML strategy with two layers. In the first layer, a series of ML models will recognize whether a compound contains S. If S is present, the second layer will apply another set of ML models to determine how many S are in the compound. Using simulated isotope patterns, we first trained and tested the ML models. We then further validated our strategy on experimental datasets where S-containing standards were spiked into complex matrices including wastewater and plasma. In addition, an external dataset which monitored S-containing photodegradation products was evaluated to determine the robustness of our ML models. Notably, we benchmarked our strategy on datasets in both positive and negative electrospray ionization (ESI) modes. Additionally, we present SulfurFinder, a streamlined R program to aid the discovery of S-containing compounds in environmental samples. SulfurFinder automates (1) data cleaning with confident annotation of natural isotopes, adducts, and in-source fragments, (2) S recognition, and (3) S number prediction. As a proof-of-principle application of SulfurFinder, we analyzed a wastewater sample where three annotated S-containing pharmaceutical drugs were confirmed using authentic standards.

2. Materials and Methods

2.1. Reagents and Chemicals

LC-MS grade methanol (MeOH), water (H2O), acetonitrile (ACN), formic acid (FA), ammonium acetate, and sodium formate were purchased from Fisher Scientific (Waltham, MA, USA). Ammonium hydroxide was purchased from MilliporeSigma (Burlington, MA, USA). Oasis HLB cartridges (6 mL, 200 mg) were obtained from Waters (Milford, MA, USA). Polyvinylidene difluoride (PVDF) syringe filters (0.45 μm) were obtained from Sigma-Aldrich (St. Louis, MO, USA). S-containing standards were purchased from MedChemExpress (Monmouth Junction, NJ, USA).

2.2. Evaluation of the Filtering Method to Extract S-Containing Compounds

To evaluate how well mass and intensity-based filtering can identify S-containing compounds, we extracted the molecular formulas (MFs) from the NORMAN Suspect List (downloaded on March 24, 2024). The MFs were then cleaned using four criteria: (1) a maximum monoisotopic mass of 500 Da; (2) organic molecules that contain at least one C and one H atom; (3) an elemental composition within the following: C, H, N, O, P, S, and Cl; and (4) a maximum of one S and two Cl. A total of 16,326 and 5,323 unique non-S- and S-containing MFs were retained after the cleaning for downstream analysis, respectively.

To simulate the MS spectrum for each MF, we used the isopattern­() function from the R package “enviPat” (ver. 2.6) to generate the monoisotopic mass and theoretical isotope pattern. Due to the limited mass resolution of QTOF MS, isotopologues were grouped at a mass resolution of 25,000 to generate the M, M+1, M+2, M+3, and M+4 isotopic peaks (Text S1). To mimic experimental conditions, a random mass error in ppm that followed a normal distribution with a mean of 0 and standard deviation of 4 was applied. In addition, intensity fluctuations were added to the isotopic peaks using normal distributions with a mean of 1 and standard deviations of 11.8% for the M+1 peak and 15.8% for the M+2, M+3, and M+4 isotopic peaks. Details on parameter selection for the simulated isotope patterns can be found in Text S2.

2.3. Model Training and Evaluation

A sequential ML approach was implemented; S is recognized in the first layer, and then the number of S is predicted in the second layer. To train a series of ML models to accomplish these tasks, we extracted unique MFs from the PubChem database. We then used several criteria to clean the MFs (detailed in Text S3). Using the strategy mentioned in the previous section, isotope patterns were simulated for model training. To further enhance the robustness, we added noise to the simulated MS spectra. Random mass error in ppm was simulated using a normal distribution with a mean of 0 and a standard deviation of 4. Random intensity fluctuation was simulated using a normal distribution with a mean of 1 and a standard deviation of 10%.

In the first layer where S recognition occurs, 163,528 and 156,072 S- and non-S-containing MFs were used for model training, respectively. The MFs were divided into a 70:30 ratio for training and testing. Random forest models were trained using the R package “ranger” (ver. 0.14.1). Three random forest ML models (i.e., the M+2, M+3, and M+4 models) were trained for the first layer; each model predicts outcomes based on spectral information extending up to their corresponding isotopic peak. Up to 25 ML features were engineered depending on the number of isotopic peaks detected. These features include the m/z of the M peak, mass differences between isotopic peaks, and intensity differences between isotopic peaks. Information on the ML features included in each model is detailed in Text S4. Hyperparameters were tuned on the training dataset using grid search. In brief, the set of hyperparameters that resulted in the lowest out-of-bag errors were chosen. The trained models were then evaluated on the testing datasets. To train the random forest ML models for S number prediction in the second layer, 79,719 and 83,809 1 and ≥2 S-containing MFs were used, respectively. Using the same strategy of S recognition for the first layer, we trained, tuned, and tested three models (i.e., the M+2, M+3, and M+4 models) using the 25 features previously described (Text S4). Detailed information on hyperparameter tuning, including the final parameters, feature importance, and model testing can be found in Text S5.

2.4. Model Benchmarking on Experimental Data

A total of 200 S-containing standards were pooled to a final concentration of 1 μg/mL in ACN/H2O (1:1, v/v). A list of the standards can be found in Table S1. We mixed this standard pool with wastewater at five ratios (standard to wastewater, v/v): 5:1, 3:1, 1:1, 1:3, and 1:5. The standard wastewater mixtures were analyzed on a Bruker Impact II ultrahigh-resolution Qq-time-of-flight mass spectrometer (Bremen, Germany) coupled to a Thermo Vanquish UHPLC system (Waltham, MA, USA). Separation was achieved on a Waters reversed-phase UPLC Acquity BEH C18 column (1.7 μm, 1.0 mm × 100 mm, 130 Å) (Milford, MA, USA). The data was collected in positive and negative ESI modes with data-dependent acquisition (DDA). Further details on sample preparation and experimental conditions can be found in Text S6.

2.5. Model Validation on Experimental Water Dataset

To further validate the performance of our models, we reprocessed an external dataset from a previous study on 2-mercaptobenzothiazole (2-MBT). In this work, the DBPs of 2-MBT were monitored at different UV fluence levels using the HPLC-QTOF-MS platform operated in positive and negative ESI modes. Details on dataset preprocessing can be found in Text S7. We validated our ML strategy by evaluating whether the models could accurately recognize and predict the number of S for 13 and 11 previously reported compounds found in positive and negative ESI modes, respectively.

2.6. Analysis of Wastewater Sample

As a proof-of-principle application for SulfurFinder, we analyzed a wastewater sample. Briefly, wastewater was collected from the Gold Bar Wastewater Treatment Plant in Edmonton, AB, Canada. The sample was concentrated using solid-phase extraction (SPE) with Waters Oasis HLB cartridges (Milford, MA, USA). The sample was analyzed in positive ESI mode using the same platform as described in the model benchmarking experiments. Further details on sample preparation and experimental conditions can be found in Text S6. The raw files were first converted to mzML files using MSConvert. MS-DIAL (ver. 4.92) was then used for feature extraction and alignment. Feature-based molecular networking was performed within the GNPS web platform. The parameters used in MS-DIAL and GNPS are listed in Text S8. Molecular networking results were exported and visualized in Cytoscape (ver. 3.10.2). PubChemLite (downloaded on Feb 04, 2025) was used to putatively annotate S-containing compounds.

3. Results and Discussion

3.1. Challenges in Recognizing S-Containing Compounds

Currently, the detection of S-containing compounds in complex samples mainly relies on FT-ICR MS. , Notably, S-containing compounds can be identified using filtering approaches for mass differences and relative intensities between the M+2 and M isotopic peaks. However, since FT-ICR MS is not accessible to most analytical and environmental laboratories, we wanted to determine whether the more standard QTOF MS can be used to recognize S-containing compounds in HRMS-based nontargeted analysis. To accomplish this task, we investigated the MFs from the NORMAN Suspect List, a database of emerging environmental contaminants that are potentially harmful to human health. For simplicity, we only considered compounds that contained ≤1 S. After simulating the isotope patterns for these 21,649 chemicals in a manner that is consistent with QTOF MS spectra of 25,000 mass resolution, we first checked the m/z differences between M+2 and M isotopic peaks. As shown in Figure A, there is a significant overlap of m/z differences between non-S- and S-containing compounds. A major contributor to this overlap is the compounds that contain elements that contribute to M+2 isotopic peaks, including Cl, C, and O. QTOF instruments cannot always resolve 34S from 37Cl, 13C2, and 18O. Thus, these non-S-containing compounds will have m/z differences similar to those of S-containing compounds. This problem is further exacerbated by mass errors in the isotope pattern. Besides the mass difference overlap, Figure B also shows a clear overlap of M+2/M isotopic peak intensity ratios between S- and non-S-containing compounds.

1.

1

Open in a new tab

Demonstration of a filtering approach for recognizing sulfur. (A) Distribution of m/z differences between the M+2 and M isotopic peaks for non-S (n = 16,326) and S-containing (n = 5,323) compounds in the NORMAN database. The dashed line represents the m/z difference between 34S and 32S. (B) Distribution of M+2 relative intensities for non-S and S-containing compounds. The dashed line represents the relative intensity of 34S. (C) TPR and FPR of a filtering strategy to recognize S-containing compounds.

Although m/z or intensity differences alone are insufficient to distinguish non-S- and S-containing compounds, we examined whether combining these two variables could be enough for S recognition. Thus, using a filtering approach similar to FT-ICR MS studies, , we used a mass difference cutoff of 1.9958 ± 0.01 Da between the M+2 and M isotopic peaks and an M+2 intensity cutoff of 4.5 ± 2.0%. For evaluation, we calculated the true positive rate (TPR) and false positive rate (FPR). These two metrics are critical indicators of data quality. A low TPR indicates that many S-containing molecules cannot be annotated. This is problematic in data interpretation as these compounds might be informative about a given environmental question. Alternatively, a high FPR is problematic because many non-S-containing compounds will be falsely annotated as containing S, and resources would be wasted on interpreting misleading data. As shown in Figure C, this filtering approach resulted in a low TPR of 0.39 and a noticeable FPR of 0.18. This performance is clearly not suitable for real-world applications as many S-containing compounds will be missed, and non-S-containing compounds will be falsely annotated to contain S. Notably, this filtering strategy would perform even worse with multiple S atoms as their M+2 intensities fall outside the given intensity threshold. In short, a more nuanced strategy is needed to accurately distinguish S- and non-S-containing compounds.

3.2. Using Machine Learning to Recognize S-Containing Compounds

To improve the filtering approach for recognizing S-containing compounds in QTOF MS analysis, we proposed a sequential ML strategy with two layers. In the first layer, the ML-based S recognition model predicts whether the chemical represented in the MS data contains S. To account for variability in isotopic peak detection due to experimental conditions, particularly for low-abundance signals, the strategy includes three distinct ML models: the M+2, M+3, and M+4 models, which uses spectral information up to the M+2, M+3, and M+4 isotopic peaks, respectively (Figure ). During model application, a suitable model is automatically picked based on the number of isotopic peaks observed in the experimental data. For example, if only the M, M+1, and M+2 isotopic peaks are detected, then the M+2 model will be used. This design addresses a critical technical challenge by leveraging data from all detected isotopic peaks, maximizing the information used. It also maintains flexibility and adaptability to account for real-world experimental data, which often lack some potential isotopic peaks, especially the M+3 and M+4 isotopic peaks due to their low abundances. Once the first layer determines the compound contains S, the second layer will proceed to predict the number of S atoms. Similar to the first layer, three ML models (i.e., M+2, M+3, and M+4 models) were trained to accomplish this task.

2.

2

Open in a new tab

Schematic overview of the ML strategy for the recognition and annotation of the number of S.

After the ML models were trained, they were evaluated on the testing dataset from the 70:30 training-test split using classification accuracies and receiver operating characteristic (ROC) curves. As shown in Figure A,B, the models performed well as the area under the ROC curves (AUCs) for all were above 0.96. For S recognition, the classification accuracies were 90.5, 91.9, and 98.7%, respectively for the M+2, M+3, and M+4 models. Improved performance is expected going from the M+2 to the M+4 model, as the M+4 model leverages more information from the isotope pattern. A similar trend was observed for S number prediction as the classification accuracies were 89.5, 91.1, and 98.1%, respectively for the M+2, M+3, and M+4 models. We further validated our ML models using cross validation and y-scrambling as detailed in Text S9. Finally, we reanalyzed the simulated QTOF isotope patterns of the NORMAN database compounds as described in the previous section. For the purposes of this benchmarking test, we removed MFs present in the NORMAN database from the PubChem training data to avoid data leakage for an unbiased evaluation. As shown in Figure C, our sequential ML strategy achieved a TPR of 0.99 and an FPR of 0.00 on compounds that the models did not see during its training. These results are significantly better than the filtering approach with a TPR of 0.39 and an FPR of 0.18 (Figure C).

3.

3

Open in a new tab

ROC curves on the testing datasets from the 70:30 training–test split. Results of using isotopic peaks for (A) S recognition and (B) S number prediction. (C) Evaluation of the ML strategy for S recognition on the unique compounds in the NORMAN database using TPR and FPR.

3.3. Benchmarking on Experimental Datasets

Next, we proceeded to benchmark our ML strategy on real-world experimental data. For this task, we prepared a standard pool containing a variety of 200 S-containing compounds, including endogenous metabolites, pesticides, and pharmaceuticals, which represent the possible S-containing compounds in the environment. Notably, we included S-containing compounds containing single and multiple sulfur atoms. Specifically, 135 compounds have a single S atom and 65 compounds have ≥2 S atoms. We mixed this standard pool with wastewater at varying ratios to mimic real environmental water sample conditions. We analyzed the standard wastewater mixtures in both positive and negative ESI modes to evaluate the robustness of our ML strategy. For benchmarking, the isotope pattern of each S-containing compound was extracted from the raw data using a series of criteria. Briefly, an isotopic peak was recognized based on three aspects: (1) retention time; (2) mass difference; and (3) chromatogram profiles (peak–peak Pearson correlation). Notably, to minimize the influence of signal fluctuation, MS spectra were averaged across the chromatographic peak. Detailed information on isotopic peak extraction can be found in Text S10.

We first benchmarked the first layer of our ML strategy on the standard wastewater mixtures in positive ESI mode. For the compounds that had detectable M+2 isotopic peaks, the S recognition ML models performed well, with accuracies ranging from 87.4 to 95.0% (Figure A). We then benchmarked the second layer of our ML strategy using the compounds that were recognized to contain S in the first layer. As shown in Figure B, the number of S can also be predicted with high accuracies ranging from 86.3 to 96.6%. In particular, with our ML strategy, it is possible to recognize S in compounds that contain Cl, which was demonstrated as impossible with intensity or mass difference cutoffs. In fact, for small molecules with m/z ≤500, a mass resolution of approximately 400,000 is needed to theoretically distinguish between the 34S- and 37Cl-containing isotopologues (Figure S1). Although modern high-end Orbitrap MS may be able to achieve such high resolving power, the scan speed would be compromised. Our strategy recognized compounds that contain S and Cl with accuracies ranging from 66.7 to 86.2% (Figure S2). This suggests that our ML strategy can be very useful in the screening of S-containing compounds in environmental studies using QTOF and Orbitrap MS.

4.

4

Open in a new tab

ML strategy was benchmarked by mixing S-containing standards in wastewater at varying ratios. (A) Results for S recognition in standards that had a detected M+2 isotopic peak in positive ESI mode. The number above each bar represent the percentage (%) of standards correctly recognized to contain S. (B) Results for predicting the number of S for compounds recognized to contain S in positive ESI mode. The number above each bar represents the percentage (%) of correct predictions. The same samples were also analyzed in negative ESI mode. (C) Results for S recognition in standards that had a detected M+2 isotopic peak in negative ESI mode. (D) Results for predicting the number of S for compounds recognized to contain S in negative ESI mode.

Similar to the positive ESI mode results, our ML strategy had good results in negative ESI mode. For S recognition, we achieved accuracies ranging from 80.0 to 87.7% (Figure C). We then predicted the number of S for compounds that were recognized to contain S. Again, we obtained good performance with accuracies ranging from 81.0 to 86.9% (Figure D). We note that the model performance in negative ESI mode was slightly poorer compared to positive ESI mode. The reasoning was that S-containing compounds were detected at lower intensities in negative ESI mode compared to positive ESI mode (Figure S3A). To investigate whether the compounds in negative ESI mode had larger intensity fluctuations in their experimental isotope patterns, we calculated an isotope similarity score. This scoring method is based on previous work and compares the experimental isotope pattern to theoretical (detailed in Text S11). As shown in Figure S3B, S-containing compounds had lower isotope similarity scores and thus larger intensity fluctuations in negative ESI mode, resulting in poorer prediction outcomes. Furthermore, we used the 1:5 standard to wastewater (v/v) mixtures as representative examples to explore the relationship between intensity and isotope similarity score (i.e., intensity fluctuation). In positive and negative ESI modes, the Spearman’s ρ between these two variables were 0.43 and 0.51 (Figure S4A,B), respectively. These results show that intensity is moderately correlated with isotope similarity score. Thus, S-containing compounds should be sufficiently abundant so that intensity fluctuations are minimized in the isotope pattern, thereby improving S-recognition.

Compared to the simulated isotope patterns, S recognition is more challenging for real experimental data, such as the standard wastewater mixtures. Real experimental data is affected by coeluting ions and ion suppression. In the standard wastewater mixtures in positive ESI mode, 10.2–13.0% of S-containing compounds did not have a detectable M+2 isotopic peak (Figure S5A). As a representative example, we further analyzed the 1:5 standard to wastewater (v/v) mixture collected in positive ESI mode to investigate why the M+2 isotopic peaks were missing. Among the 18 S-containing compounds that did not have an M+2 isotopic peak, the main reason was their low intensities (Figure S5B). However, we observed that for 4 out of 18 compounds that had high intensities (i.e., >10,000), the M+2 isotopic peaks were still not extracted due to poor peak shapes. The peak–peak correlations between the M and M+2 isotopic peaks were lower than the threshold needed for identifying isotopic peaks in our program, and thus those compounds were not recognized. The results suggest that sample preparation to concentrate S-containing compounds and better LC separation methods to mitigate poor peak shapes are all critical for comprehensive S-recognition.

To further demonstrate the robustness of our ML strategy, we designed a similar benchmarking experiment to the standard wastewater mixtures. Notably, we mixed the 200 S-containing compounds with plasma instead of wastewater. The detailed sample preparation and experimental conditions can be found in Text S12. Even with a significantly different sample matrix, we were able to obtain good results for S recognition and S number prediction in both positive and negative ESI modes (Figure S6). We note that in negative ESI mode, prediction outcomes for S recognition were poorer. For example, for the standard plasma mixture with the lowest concentration of S-containing compounds (i.e., 1:5 standard to plasma (v/v) mixture), the accuracy for S recognition was only 57.4%. This was a significant drop in prediction performance compared to the analogous standard wastewater mixture (i.e., 1:5 standard to wastewater (v/v) mixture) where the S recognition accuracy was 81.3%. S recognition was poorer in the standard plasma mixture than the standard wastewater mixture due to plasma having a significantly more complex matrix than wastewater. This difference in matrix complexity can be seen in the chromatograms of wastewater and plasma where many more intense peaks were detected in plasma (Figure S7). Thus, S recognition was more challenging in plasma due to the increased background interference from ion suppression and coelution in the isotope patterns.

Next, we benchmarked our ML strategy by reprocessing an external water dataset where the photodegradation products of 2-MBT were monitored to investigate DBPs. In brief, 13 and 11 compounds were reported in positive and negative ESI modes, respectively. These reported compounds were a mixture of confirmed and putative annotations. Notably, these compounds were chemically diverse, ranging from compounds that contained no sulfur atoms to three sulfur atoms. For evaluation, we wanted to determine whether our ML models could accurately recognize and predict the correct number of S for these compounds. For both ionization modes, we were able to correctly recognize the presence of S for all reported compounds (Table S2). In addition, except for one reported S-containing compound in negative ESI mode, we correctly predicted whether each chemical contained 1 or ≥2 S. We note that although the external dataset was relatively small, the class of chemicals evaluated (i.e., photodegradation byproducts) were significantly different from the S-containing compounds in the previously benchmarked standard mixtures. Furthermore, the data was collected using an MS from a different vendor. Despite these differences, we still achieved consistent good performance, demonstrating the robustness of our ML strategy. In addition to validating our ML strategy, we wanted to leverage it to see whether we could discover novel S-containing DBPs that were not reported in the original study. In negative ESI mode, we highlight feature #756 as a potential photodegradation product that was predicted to contain ≥2 S. As shown in Figure A, this feature formed substantially after being irradiated with 500 mJ/cm2 but eventually photodegraded after 4,000 mJ/cm2. We tentatively proposed the MF to be C7H5NO6S3 and compared its experimental isotope pattern to its theoretical counterpart. We obtained a high isotope similarity score of 0.98, suggesting that C7H5NO6S3 was a likely MF candidate (Figure B). Using MS/MS data, we propose the structure of the feature to be related to benzothiazole-2-sulfonic acid (Figure C). Feature #756 and the reference spectrum had a modified cosine score of 0.78 and were separated by a mass shift corresponding to a loss of SO3. Altogether, the correct recognition of reported S-containing DBPs and the discovery of a potentially new DBP demonstrate that our ML strategy can be critical for advancing DBP research.

5.

5

Open in a new tab

Feature #756 is a potential photodegradation product of 2-MBT. (A) Intensities of the feature as a function of UV fluence. (B) Comparison of the experimental isotope pattern with its theoretical counterpart. (C) Comparison of the MS/MS data of feature #756 with reference data for benzothiazole-2-sulfonic acid. The two compounds are structurally related with a modified cosine score of 0.78.

3.4. SulfurFinder

To streamline the recognition of S-containing compounds in nontargeted LC-HRMS data, we employed our ML strategy to develop SulfurFinder for automated data cleaning, S recognition, and S number prediction. The input of the program only requires the raw LC-HRMS data and a feature table that contains feature ID, m/z, retention time, and intensities for each sample. Thus, SulfurFinder can be easily integrated into existing MS data analysis workflows as the required information is all easily obtained from the user’s preferred peak picking algorithm.

Notably, a series of unique features were implemented into SulfurFinder to improve S recognition. To address the issue that peaks with low abundance and/or poor peak shapes may be missed from peak picking, SulfurFinder directly extracts the isotope patterns for each feature from the raw LC-HRMS data. The identification of isotopic peaks was based on retention time, mass difference, and chromatogram profiles (peak–peak Pearson correlation). Notably, MS spectra were averaged across the chromatographic peak to minimize signal fluctuations, thereby improving prediction outcomes. We further applied intensity cutoffs to remove coeluting ions in the isotope pattern mistakenly identified as isotopic peaks. Further details on isotopic peak extraction can be found in Text S10. SulfurFinder will then use the averaged MS spectra for S recognition. The program will automatically select the best ML model depending on the number of isotopic peaks detected. SulfurFinder will use a different set of models in the second layer to predict the number of S for compounds recognized to contain S in the first layer. Most importantly, SulfurFinder was trained on representative QTOF data which has significantly lower mass resolution compared to conventional S recognition methods involving FT-ICR MS. Thus, SulfurFinder makes S recognition possible for more analytical and environmental laboratories. In addition, data cleaning can be conducted to annotate natural isotopes, adducts, and in-source fragments to improve the confidence of S annotation results. Detailed information on each step of SulfurFinder can be found in Text S13. To ensure that SulfurFinder was not mistakenly annotating too many compounds as S-containing, we tested the program on a method blank sample (Text S14). As expected of a clean sample, only 4.4% of the features were recognized as S-containing, demonstrating the high selectivity of SulfurFinder. SulfurFinder is freely accessible on GitHub: https://github.com/HuanLab/SulfurFinder.

3.5. Proof-of-Principle Application

As a proof-of-principle application, SulfurFinder was tested on a wastewater sample (Text S6). After peak picking and alignment, we obtained 5,738 features (Figure A). We first removed 2,199 features that did not elute during the analytical gradient or had similar intensities in the method blank. From the remaining 3,539 features, 88 natural isotopes, 564 adducts, and 114 in-source fragments were identified and removed. A total of 169 potential S-containing features were finally recognized. Notably, SulfurFinder only took 8 min for S recognition, S number prediction, and data cleaning when using a standard desktop computer (Intel i9–12900K CPU@3.2 GHz with 16 cores and 32 GB memory), demonstrating the efficiency of the program (Figure S8).

6.

6

Open in a new tab

Application of SulfurFinder on a wastewater sample in positive ESI mode. (A) 169 S-containing features were annotated. (B) Classification of S-containing features based on the number of S. (C) SulfurFinder and molecular networking were leveraged to putatively annotate feature #3482 as pirlimycin, which is structurally related to clindamycin with a modified cosine score of 0.83.

Of the 169 annotated S-containing features by SulfurFinder, 81.7% had experimental MS/MS but no match from the MS-DIAL and NIST20 MS/MS library search. These features would conventionally be considered as “unknown unknowns” and their de novo annotation would be extremely hard. However, with the use of SulfurFinder, we can at least confidently recognize S in these molecules. Thus, a certain level of chemical information is provided and guides downstream compound confirmation. Notably, of the compounds annotated through MS/MS library search, only nine were S-containing. This finding suggests that many S-containing environmental contaminants may not be in MS/MS libraries, demonstrating the need for SulfurFinder to recognize them. On the other side, since the LC-MS data was collected in DDA mode, low-abundance features may not trigger an MS/MS event. Notably, 29.6% of the compounds recognized to contain S did not have MS/MS data. However, since the annotation was based on MS1 isotope patterns only, SulfurFinder was still able to gain some insight into these features by recognizing them as S-containing. These results highlight the potential of SulfurFinder in annotating S-containing compounds with minimal spectral requirements.

Figure B shows the distribution of 1 S and ≥2 S-containing features in the dataset. To putatively annotate the 169 potential S-containing features, we used the PubChemLite database, a subset of PubChem containing chemicals relevant to environmental and exposomics applications. Briefly, for each S-containing feature in our experiment, we queried its experimental m/z and isotope pattern against PubChemLite for potential MF matches. Detailed information on MF matching can be found in Text S15. Of the 169 potential S-containing features, 116 (68.6%) of them matched literature-reported S-containing chemicals in the database. The unmatched features could be due to the limited size of the database. When leveraging SulfurFinder’s S number prediction results to further refine the database search, we found that 86 (50.9%) of the S-containing features matched S-containing chemicals with the number of S predicted. These putative MF annotations thus offer good starting points to confirm emerging S-containing chemicals in environmental samples.

Finally, among the annotated S features, we confirmed three S-containing drugs with authentic standards: clindamycin, gliclazide, and clopidogrel (Text S16). Notably, SulfurFinder correctly predicted the number of S for clindamycin and gliclazide as well. Finally, in addition to the SulfurFinder annotations, we leveraged molecular networking with MS/MS to putatively annotate additional S-containing compounds (Figure C). In particular, we focus on feature #3482 which did not have a match from the MS/MS library search but was annotated as having 1 S from SulfurFinder. After molecular networking, feature #3482 was linked to clindamycin (modified cosine score: 0.83) with a mass shift corresponding to the loss of CH2. Thus, we putatively annotated this feature as pirlimycin, another S-containing drug molecule and analog of clindamycin. These results suggest that S-containing pharmaceuticals are an important source of environmental contaminants due to their incomplete removal in wastewater treatment. −

3.6. Limitations and Future Directions

The present study highlights SulfurFinder, a streamlined ML-based tool that leverages isotope patterns for precise S recognition and S number prediction. However, several limitations should be acknowledged. First, low-abundant S-containing compounds that do not have sufficient isotopic peaks cannot be recognized by SulfurFinder. In addition, due to the inadequate mass resolution of QTOF and Orbitrap MS instruments, predictions for compounds containing both S and Cl may be poor. To address these concerns, future work could focus on optimizing sample enrichment procedures, such as solid phase extraction (SPE), to concentrate S-containing chemicals. This will allow the full isotope pattern to be detected, improving the confidence of SulfurFinder predictions. Furthermore, future work could explore whether ML models using MS/MS data can improve S recognition in cases where isotopic peaks are missing and/or convoluted due to insufficient mass resolution.

4. Conclusions

This work proposed a ML strategy for the nontargeted analysis of S-containing chemicals in complex environmental samples. Notably, we were able to annotate S-containing features with high accuracy using MS instruments with inadequate mass resolution. This ML strategy was integrated into an R program, SulfurFinder, to automate data cleaning, S recognition, and S number prediction in LC-HRMS data. As a proof-of-principle application, we tested the program on a wastewater sample and annotated >100 potential S-containing chemicals, including three confirmed drug molecules. Although this work focused on environmental samples, SulfurFinder can also be easily applied to biological samples where S plays important biological roles.

Supplementary Material

vg5c00062_si_001.pdf (1.2MB, pdf)
vg5c00062_si_002.xlsx (27KB, xlsx)

Acknowledgments

The project was supported by grants from the Canada Foundation for Innovation, Natural Sciences and Engineering Research Council of Canada, Genome Canada, Genome British Columbia, Alberta Innovates, and the Canada Research Chairs Programs. We thank Alisa Hui for proofreading this manuscript.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsenvironau.5c00062.

  • Approximate mass resolution required to resolve the 34S-containing and 37Cl-containing isotopologues (Figure S1); comparison of S-recognition accuracies for compounds containing only S and compounds containing both S and Cl in positive ESI mode (Figure S2); comparison of compounds benchmarked in the 1:5 standard to wastewater (v/v) mixture in positive and negative ESI modes (Figure S3); relationship between intensity and isotope similarity score (Figure S4); S cannot be recognized in chemicals with no detectable M+2 isotopic peak (Figure S5); ML strategy was benchmarked by mixing S-containing standards in plasma at varying ratios (Figure S6); plasma had a more complicated sample matrix than wastewater (Figure S7); processing times of SulfurFinder on the wastewater sample with 3,539 metabolic features after RT and blank filtering (Figure S8); grouping adjacent isotopologues (Text S1); parameter selection for isotope pattern simulations (Text S2); preparation of the PubChem database (Text S3); machine learning features used in the ML models (Text S4); tuning, feature importance, and evaluation of ML models (Text S5); sample preparation and experimental conditions for the standard wastewater mixtures (Text S6); data preprocessing for the 2-MBT photodegradation dataset (Text S7); parameters used in MS-DIAL and GNPS for wastewater samples (Text S8); further model validation on ML models (Text S9); details on isotopic peak extraction (Text S10); isotope similarity scoring (Text S11); sample preparation and experimental conditions for the standard plasma mixtures (Text S12); details of SulfurFinder (Text S13); evaluation of SulfurFinder on a method blank sample (Text S14); MF annotation using PubChemLite (Text S15); comparison of experimental and reference data for S-containing drug molecules (Text S16); and benchmarking results for 2-MBT DBPs dataset (Table S2) (PDF)

  • S-containing standards detected in the 5:1 standard to wastewater (v/v) mixture (positive and negative ESI modes) (Table S1) (XLSX)

The authors declare no competing financial interest.

Published as part of ACS Environmental Au special issue “2025 Rising Stars in Environmental Research”.

References

  1. Chou C.-L.. Sulfur in Coals: A Review of Geochemistry and Origins. Int. J. Coal Geol. 2012;100:1–13. doi: 10.1016/j.coal.2012.05.009. [DOI] [Google Scholar]
  2. Schwarzenbach, R. P. ; Gschwend, P. M. ; Imboden, D. M. . Environmental Organic Chemistry; John Wiley & Sons, 2016. [Google Scholar]
  3. Nakabayashi R., Saito K.. Ultrahigh Resolution Metabolomics for S-Containing Metabolites. Curr. Opin. Biotechnol. 2017;43:8–16. doi: 10.1016/j.copbio.2016.07.003. [DOI] [PubMed] [Google Scholar]
  4. Zhuang J., Jiang Z., Chen D., Li J., Crabbe M. J. C., Qiu M., Zheng Y., Qu W.. Thyroid-Disrupting Effects of Exposure to Fipronil and Its Metabolites from Drinking Water Based on Human Thyroid Follicular Epithelial Nthy-Ori 3–1 Cell Lines. Environ. Sci. Technol. 2023;57(15):6072–6084. doi: 10.1021/acs.est.2c08627. [DOI] [PubMed] [Google Scholar]
  5. Liberatore H. K., Westerman D. C., Allen J. M., Plewa M. J., Wagner E. D., McKenna A. M., Weisbrod C. R., McCord J. P., Liberatore R. J., Burnett D. B., Cizmas L. H., Richardson S. D.. High-Resolution Mass Spectrometry Identification of Novel Surfactant-Derived Sulfur-Containing Disinfection Byproducts from Gas Extraction Wastewater. Environ. Sci. Technol. 2020;54(15):9374–9386. doi: 10.1021/acs.est.0c01997. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Wang M., Helbling D. E.. A Non-Target Approach to Identify Disinfection Byproducts of Structurally Similar Sulfonamide Antibiotics. Water Res. 2016;102:241–251. doi: 10.1016/j.watres.2016.06.042. [DOI] [PubMed] [Google Scholar]
  7. Jiang P., Qiu J., Gao Y., Stefan M. I., Li X.-F.. Nontargeted Identification and Predicted Toxicity of New Byproducts Generated from UV Treatment of Water Containing Micropollutant 2-Mercaptobenzothiazole. Water Res. 2021;188:116542. doi: 10.1016/j.watres.2020.116542. [DOI] [PubMed] [Google Scholar]
  8. Allen J. M., Plewa M. J., Wagner E. D., Wei X., Bokenkamp K., Hur K., Jia A., Liberatore H. K., Lee C.-F. T., Shirkhani R., Krasner S. W., Richardson S. D.. Drivers of Disinfection Byproduct Cytotoxicity in U.S. Drinking Water: Should Other DBPs Be Considered for Regulation? Environ. Sci. Technol. 2022;56(1):392–402. doi: 10.1021/acs.est.1c07998. [DOI] [PubMed] [Google Scholar]
  9. Richardson, S. D. ; Postigo, C. . Chapter 11 - Discovery of New Emerging DBPs by High-Resolution Mass Spectrometry. In Comprehensive Analytical Chemistry; Pérez, S. ; Eichhorn, P. ; Barceló, D. , Eds.; Applications of Time-of-Flight and Orbitrap Mass Spectrometry in Environmental, Food, Doping, and Forensic Analysis; Elsevier, 2016; Vol. 71, pp. 335–356. [Google Scholar]
  10. Nakabayashi R., Tsugawa H., Mori T., Saito K.. Automation of Chemical Assignment for Identifying Molecular Formula of S-Containing Metabolites by Combining Metabolomics and Chemoinformatics with 34S Labeling. Metabolomics. 2016;12(11):168. doi: 10.1007/s11306-016-1115-5. [DOI] [Google Scholar]
  11. Miura D., Tsuji Y., Takahashi K., Wariishi H., Saito K.. A Strategy for the Determination of the Elemental Composition by Fourier Transform Ion Cyclotron Resonance Mass Spectrometry Based on Isotopic Peak Ratios. Anal. Chem. 2010;82(13):5887–5891. doi: 10.1021/ac902931x. [DOI] [PubMed] [Google Scholar]
  12. Nakabayashi R., Sawada Y., Yamada Y., Suzuki M., Hirai M. Y., Sakurai T., Saito K.. Combination of Liquid Chromatography-Fourier Transform Ion Cyclotron Resonance-Mass Spectrometry with 13C-Labeling for Chemical Assignment of Sulfur-Containing Metabolites in Onion Bulbs. Anal. Chem. 2013;85(3):1310–1315. doi: 10.1021/ac302733c. [DOI] [PubMed] [Google Scholar]
  13. Nakabayashi R., Sawada Y., Aoyagi M., Yamada Y., Hirai M. Y., Sakurai T., Kamoi T., Rowan D. D., Saito K.. Chemical Assignment of Structural Isomers of Sulfur-Containing Metabolites in Garlic by Liquid Chromatography–Fourier Transform Ion Cyclotron Resonance–Mass Spectrometry1234. J. Nutr. 2016;146(2):397S–402S. doi: 10.3945/jn.114.202317. [DOI] [PubMed] [Google Scholar]
  14. Smith D. F., Podgorski D. C., Rodgers R. P., Blakney G. T., Hendrickson C. L.. 21 T FT-ICR Mass Spectrometer for Ultrahigh-Resolution Analysis of Complex Organic Mixtures. Anal. Chem. 2018;90(3):2041–2047. doi: 10.1021/acs.analchem.7b04159. [DOI] [PubMed] [Google Scholar]
  15. Bowman A. P., Blakney G. T., Hendrickson C. L., Ellis S. R., Heeren R. M. A., Smith D. F.. Ultra-High Mass Resolving Power, Mass Accuracy, and Dynamic Range MALDI Mass Spectrometry Imaging by 21-T FT-ICR MS. Anal. Chem. 2020;92(4):3133–3142. doi: 10.1021/acs.analchem.9b04768. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Richardson S. D.. Water Analysis: Emerging Contaminants and Current Issues. Anal. Chem. 2009;81(12):4645–4677. doi: 10.1021/ac9008012. [DOI] [PubMed] [Google Scholar]
  17. Meusel M., Hufsky F., Panter F., Krug D., Müller R., Böcker S.. Predicting the Presence of Uncommon Elements in Unknown Biomolecules from Isotope Patterns. Anal. Chem. 2016;88(15):7556–7566. doi: 10.1021/acs.analchem.6b01015. [DOI] [PubMed] [Google Scholar]
  18. Zhao T., Wawryk N. J. P., Xing S., Low B., Li G., Yu H., Wang Y., Shen Q., Li X.-F., Huan T.. ChloroDBPFinder: Machine Learning-Guided Recognition of Chlorinated Disinfection Byproducts from Nontargeted LC-HRMS Analysis. Anal. Chem. 2024;96(6):2590–2598. doi: 10.1021/acs.analchem.3c05124. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Zhao T., Low B., Shen Q., Wang Y., Hidalgo Delgado D., Chau K. N. M., Pang Z., Li X., Xia J., Li X.-F., Huan T.. Exposome-Scale Investigation of Cl-/Br-Containing Chemicals Using High-Resolution Mass Spectrometry, Multi-Stage Machine Learning, and Cloud Computing. Anal. Chem. 2025;97:11099. doi: 10.1021/acs.analchem.5c00503. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Mohammed Taha H., Aalizadeh R., Alygizakis N., Antignac J.-P., Arp H. P. H., Bade R., Baker N., Belova L., Bijlsma L., Bolton E. E., Brack W., Celma A., Chen W.-L., Cheng T., Chirsir P., Čirka Ĺ., D’Agostino L. A., Djoumbou Feunang Y., Dulio V., Fischer S., Gago-Ferrero P., Galani A., Geueke B., Głowacka N., Glüge J., Groh K., Grosse S., Haglund P., Hakkinen P. J., Hale S. E., Hernandez F., Janssen E. M.-L., Jonkers T., Kiefer K., Kirchner M., Koschorreck J., Krauss M., Krier J., Lamoree M. H., Letzel M., Letzel T., Li Q., Little J., Liu Y., Lunderberg D. M., Martin J. W., McEachran A. D., McLean J. A., Meier C., Meijer J., Menger F., Merino C., Muncke J., Muschket M., Neumann M., Neveu V., Ng K., Oberacher H., O’Brien J., Oswald P., Oswaldova M., Picache J. A., Postigo C., Ramirez N., Reemtsma T., Renaud J., Rostkowski P., Rüdel H., Salek R. M., Samanipour S., Scheringer M., Schliebner I., Schulz W., Schulze T., Sengl M., Shoemaker B. A., Sims K., Singer H., Singh R. R., Sumarah M., Thiessen P. A., Thomas K. V., Torres S., Trier X., van Wezel A. P., Vermeulen R. C. H., Vlaanderen J. J., von der Ohe P. C., Wang Z., Williams A. J., Willighagen E. L., Wishart D. S., Zhang J., Thomaidis N. S., Hollender J., Slobodnik J., Schymanski E. L.. The NORMAN Suspect List Exchange (NORMAN-SLE): Facilitating European and Worldwide Collaboration on Suspect Screening in High Resolution Mass Spectrometry. Environ. Sci. Eur. 2022;34(1):104. doi: 10.1186/s12302-022-00680-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Chambers M. C., Maclean B., Burke R., Amodei D., Ruderman D. L., Neumann S., Gatto L., Fischer B., Pratt B., Egertson J., Hoff K., Kessner D., Tasman N., Shulman N., Frewen B., Baker T. A., Brusniak M.-Y., Paulse C., Creasy D., Flashner L., Kani K., Moulding C., Seymour S. L., Nuwaysir L. M., Lefebvre B., Kuhlmann F., Roark J., Rainer P., Detlev S., Hemenway T., Huhmer A., Langridge J., Connolly B., Chadick T., Holly K., Eckels J., Deutsch E. W., Moritz R. L., Katz J. E., Agus D. B., MacCoss M., Tabb D. L., Mallick P.. A Cross-Platform Toolkit for Mass Spectrometry and Proteomics. Nat. Biotechnol. 2012;30(10):918–920. doi: 10.1038/nbt.2377. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Tsugawa H., Cajka T., Kind T., Ma Y., Higgins B., Ikeda K., Kanazawa M., VanderGheynst J., Fiehn O., Arita M.. MS-DIAL: Data-Independent MS/MS Deconvolution for Comprehensive Metabolome Analysis. Nat. Methods. 2015;12(6):523–526. doi: 10.1038/nmeth.3393. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Nothias L.-F., Petras D., Schmid R., Dührkop K., Rainer J., Sarvepalli A., Protsyuk I., Ernst M., Tsugawa H., Fleischauer M., Aicheler F., Aksenov A. A., Alka O., Allard P.-M., Barsch A., Cachet X., Caraballo-Rodriguez A. M., Da Silva R. R., Dang T., Garg N., Gauglitz J. M., Gurevich A., Isaac G., Jarmusch A. K., Kameník Z., Kang K. B., Kessler N., Koester I., Korf A., Le Gouellec A., Ludwig M., Martin H C., McCall L.-I., McSayles J., Meyer S. W., Mohimani H., Morsy M., Moyne O., Neumann S., Neuweger H., Nguyen N. H., Nothias-Esposito M., Paolini J., Phelan V. V., Pluskal T., Quinn R. A., Rogers S., Shrestha B., Tripathi A., van der Hooft J. J. J., Vargas F., Weldon K. C., Witting M., Yang H., Zhang Z., Zubeil F., Kohlbacher O., Böcker S., Alexandrov T., Bandeira N., Wang M., Dorrestein P. C.. Feature-Based Molecular Networking in the GNPS Analysis Environment. Nat. Methods. 2020;17(9):905–908. doi: 10.1038/s41592-020-0933-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Shannon P., Markiel A., Ozier O., Baliga N. S., Wang J. T., Ramage D., Amin N., Schwikowski B., Ideker T.. Cytoscape: A Software Environment for Integrated Models of Biomolecular Interaction Networks. Genome Res. 2003;13(11):2498–2504. doi: 10.1101/gr.1239303. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Elapavalore A., Ross D. H., Grouès V., Aurich D., Krinsky A. M., Kim S., Thiessen P. A., Zhang J., Dodds J. N., Baker E. S., Bolton E. E., Xu L., Schymanski E. L.. PubChemLite Plus Collision Cross Section (CCS) Values for Enhanced Interpretation of Nontarget Environmental Data. Environ. Sci. Technol. Lett. 2025;12:166. doi: 10.1021/acs.estlett.4c01003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Xing S., Shen S., Xu B., Li X., Huan T.. BUDDY: Molecular Formula Discovery via Bottom-up MS/MS Interrogation. Nat. Methods. 2023;20:881–890. doi: 10.1038/s41592-023-01850-x. [DOI] [PubMed] [Google Scholar]
  27. Mahieu N. G., Patti G. J.. Systems-Level Annotation of a Metabolomics Data Set Reduces 25 000 Features to Fewer than 1000 Unique Metabolites. Anal. Chem. 2017;89(19):10397–10406. doi: 10.1021/acs.analchem.7b02380. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Guo J., Huan T.. Comparison of Full-Scan, Data-Dependent, and Data-Independent Acquisition Modes in Liquid Chromatography-Mass Spectrometry Based Untargeted Metabolomics. Anal. Chem. 2020;92(12):8072–8080. doi: 10.1021/acs.analchem.9b05135. [DOI] [PubMed] [Google Scholar]
  29. Batt A. L., Bruce I. B., Aga D. S.. Evaluating the Vulnerability of Surface Waters to Antibiotic Contamination from Varying Wastewater Treatment Plant Discharges. Environ. Pollut. 2006;142(2):295–302. doi: 10.1016/j.envpol.2005.10.010. [DOI] [PubMed] [Google Scholar]
  30. Silva B. F. da, Jelic A., López-Serna R., Mozeto A. A., Petrovic M., Barceló D.. Occurrence and Distribution of Pharmaceuticals in Surface Water, Suspended Solids and Sediments of the Ebro River Basin, Spain. Chemosphere. 2011;85(8):1331–1339. doi: 10.1016/j.chemosphere.2011.07.051. [DOI] [PubMed] [Google Scholar]
  31. Fabregat-Safont D., Gracia-Marín E., Ibáñez M., Pitarch E., Hernández F.. Analytical Key Issues and Challenges in the LC-MS/MS Determination of Antibiotics in Wastewater. Anal. Chim. Acta. 2023;1239:340739. doi: 10.1016/j.aca.2022.340739. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

vg5c00062_si_001.pdf (1.2MB, pdf)
vg5c00062_si_002.xlsx (27KB, xlsx)

Articles from ACS Environmental Au are provided here courtesy of American Chemical Society

RESOURCES