Intelligent passive sampling and non-targeted analytical method for monitoring emerging contaminants in water

Abstract

The present application belongs to the technical field of environmental monitoring, and relates to a method for collecting and quantifying new pollutants in water, and relates to an intelligent passive sampling and non-target analysis method for monitoring new pollutants in water. The method predicts the diffusion coefficient and ionization efficiency of the compound by constructing a machine learning model, overcomes the quantitative problem caused by the lack of key parameters in the coupling process of DGT and NTA technology, thereby fully utilizing the advantages of DGT in obtaining representative water samples and time-weighted concentration, and combining the high screening ability of NTA, realizing high-throughput quantitative analysis of new pollutants. The model constructed by the present application covers a wider range of chemical structures and experimental conditions. Combined with the predicted key parameters, reliable high-throughput concentration estimation of the identified new pollutants in the water body can be carried out without standard samples, providing strong technical support for new pollutant monitoring and risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring technology, and relates to methods for collecting and quantifying new pollutants in water, and to an intelligent passive sampling and non-targeted analysis method for monitoring new pollutants in water. Background Technology

[0002] Emerging pollutants refer to toxic and hazardous chemical substances released into the environment that possess characteristics such as biotoxicity, environmental persistence, and bioaccumulation, posing significant risks to the ecological environment or human health, but which are not yet included in management or where existing management measures are insufficient. The accumulation of emerging pollutants in water bodies poses a potential threat to aquatic ecosystems and human health. Revealing the types and concentration levels of emerging pollutants in the aquatic environment is of great significance for their remediation. Traditional environmental monitoring methods typically target compounds for which existing analytical methods and standards exist. However, the difficulty in obtaining and the high cost of emerging pollutant standards result in a limited number of monitorable emerging pollutants, making it difficult to accurately and efficiently reveal the types and concentrations of emerging pollutants in the aquatic environment. Furthermore, the low concentrations and diverse sources of emerging pollutants present challenges for representative sample collection, pretreatment, and analysis.

[0003] Passive sampling techniques such as diffraction gradient diffusion (DGT) can pre-enrich pollutants, providing an average weighted concentration of pollutants during sampling, resulting in more representative samples. The pretreatment method is simple, making it suitable for collecting novel pollutants in water. However, in DGT sampling, the quantitative accumulation (M) of compounds in the binding layer depends on the availability of the diffusion coefficient (D) of each pollutant in the diffusion layer. Experimentally determining the D value of each pollutant is inefficient, costly, and lacks sufficient standards, limiting DGT passive sampling to targeted monitoring of novel pollutants with available standards. On the other hand, non-targeted high-resolution mass spectrometry (NTA) can rapidly identify hundreds of compounds in a sample without the need for standards. However, in the absence of standards, quantification in non-targeted analysis relies on calibration curves of response factors (RF) or similar structural substitutes, leading to significant uncertainty and limited applicability. Combining DGT and NTA technologies may leverage their advantages in sample representativeness and high-throughput identification, but high-throughput monitoring remains constrained by quantification issues.

[0004] Existing studies have attempted to predict diffusion coefficients (D) using machine learning models (Sci. Total Environ., 2020, 706, 135691), but their applicability remains limited due to insufficient training data coverage. Some studies have predicted RF based on compound structure (Water Res. 2023, 233, 119735), however, RF is easily affected by various factors such as testing methods, mass spectrometry type, and detector status, resulting in poor generalization and applicability of the model. In contrast, ionization efficiency (IE) is also closely related to compound structure, but is less affected by the instrument and has been shown to have good transferability between different instruments and laboratories (Anal. Chem. 2020, 92(7), 4691-4699). Existing studies have attempted to achieve semi-quantitative non-targeted analysis based on IE prediction models, but the consideration of relevant influencing factors is still insufficient, limiting the prediction accuracy. Therefore, it is necessary to develop D and IE machine learning prediction models with broader coverage and higher accuracy to address the lack of quantitative parameters, thereby promoting the effective coupling of the two technologies and achieving high-throughput monitoring. Summary of the Invention

[0005] This invention provides a method for predicting key quantitative parameters (D and IE) of DGT passive sampling and NTA technology using machine learning models, which solves the problem of difficulty in quantification in the coupled application of DGT and NTA, and enables high-throughput monitoring of new pollutants in the aquatic environment.

[0006] The specific technical solution of this invention is as follows:

[0007] A smart passive sampling and non-targeted analysis method for monitoring emerging pollutants in water includes the following steps:

[0008] Step (1) Data collection and preprocessing

[0009] The diffusion coefficients (D values) of novel contaminants such as antibiotics, perfluorinated compounds, and endocrine disruptors in agarose gel diffusion layers were collected and uniformly corrected to values ​​at 25°C. Then, all D values ​​were converted to a negative logarithmic form (-logD) with base 10.

[0010] Collect ionization efficiency data (logIE) of the aforementioned new pollutants under positive and negative ion modes of the electrospray ionization source. pos and logIE neg These are used to construct prediction models for the ionization efficiency of new pollutants under different modes.

[0011] Step (2) Model Building

[0012] To construct a diffusion coefficient-logD prediction model, firstly, molecular descriptors and molecular fingerprints are calculated based on the structure of the pollutants in step (1) as molecular structural features. The dataset is divided into training and testing sets. The Extreme Gradient Boosting Tree (XGBoost) algorithm is used to arrange and combine all molecular descriptors and molecular fingerprints to form n sets of feature combinations. The n sets of feature combinations are used as inputs and -logD is used as the output to construct n -logD models.

[0013] For logIE pos and logIE neg In constructing the prediction model, in addition to molecular structure features, descriptors such as mobile phase composition, gradient, instrument type, and ion source technology are additionally introduced into the input features. One-hot encoding is used to convert non-numerical information into binary features. The dataset is divided into training and testing sets, and the XGBoost algorithm is used to construct the model. In the input features, the descriptors such as mobile phase composition, gradient, and instrument type remain constant, while the molecular structure features are constructed using combinations of different molecular descriptors and molecular fingerprints. That is, all molecular descriptors and molecular fingerprints are permuted and combined to form n feature combinations, constructing n logIEs. pos and logIE neg Model.

[0014] Model performance was evaluated using cross-validation and an external test set. Root mean square error (RMSE) was used. RMS Using the coefficient of determination (R²) as evaluation indicators, the optimal values ​​were selected as log-logD and logIE. pos and logIE neg Predictive model. The prediction model uses a cosine similarity score threshold (Si) between the structure to be predicted and the structures in the training set. cutoff ), and the minimum number of compounds in the training set that satisfy the threshold (N) min Define the application domain of the model to ensure the reliability of the model's predictions.

[0015] Step (3) DGT Sampling and Analysis

[0016] The DGT sampler was deployed in the water body to be tested, and continuous sampling was performed for a certain period of time. After the sampling was completed, the bound layer was eluted and concentrated. High-performance liquid chromatography-high-resolution mass spectrometry (HPLC-MS / MS) was used to perform non-targeted analysis of the samples. Electrospray ionization was used to detect compounds in the samples in both positive and negative modes. Non-targeted data were acquired using a data-dependent mode. Peak alignment, noise reduction, and screening were performed on the detected peaks. The structures of new contaminants in the samples were identified by matching with a secondary mass spectrometry database using unknown substance screening software.

[0017] Step (4) Application of semi-quantitative strategies

[0018] Using two sets of standards that can be ionized in positive and negative ionization modes respectively, a standard curve is established, the RF (i.e., the slope of the standard curve) is obtained, and the logIE is used. pos and logIE neg The predictive model obtains the logIE of the standard product under the corresponding mode. pos and logIE neg Predicted values. LogRF and logIE are fitted under positive and negative ionization modes respectively. pos and logIE neg Obtain the formula for the linear relationship between the predicted values.

[0019] Based on the structure of the newly identified pollutants, -logD and logIE were used. pos and logIE neg Predictive models predict -logD and logIE for new pollutants. pos and logIE neg value.

[0020] Using the logRF and logIE obtained above pre Based on the relationship, the RF is derived, and combined with the response area of ​​the new pollutant, the concentration of the new pollutant in the sample (C) is calculated using the following formula. S ):

[0021] C S = Area / RF

[0022] Here, Area represents the peak area of ​​the newly identified pollutant.

[0023] Further calculations of the cumulative amount of new pollutants in DGT allow for the derivation of the time-weighted average concentration of these pollutants in the environment, enabling high-throughput quantitative prediction of new pollutants in water without the need for standards. The specific formula is as follows:

[0024]

[0025] in, M is the concentration of new contaminants in the water measured by DGT; M is the mass of new contaminants accumulated on the binding layer, which is equal to the volume of the eluent and the concentration of new contaminants in the eluent. S The product of Δg and Δg; Δg is the thickness of the diffusion layer; D is the diffusion coefficient of the new pollutant in the diffusion layer, obtained by the -logD prediction model; A is the area of ​​the DGT exposure window; t is the deployment time of the DGT sampler in the water.

[0026] Effects and benefits of the present invention:

[0027] (1) The model constructed in this invention covers a wider range of pollutant chemical structures and testing conditions, and has a wide range of applications and strong scalability.

[0028] (2) The method constructed in this invention provides reliable results for predicting the concentration of compounds within the application domain.

[0029] (3) This invention uses machine learning models to predict key parameters, and can simultaneously perform quantitative analysis on dozens to hundreds of pollutants, realize concentration estimation without the need for standards, break through the limitations of standards, and realize high-throughput monitoring. Attached Figure Description

[0030] Figure 1 The measured and predicted values ​​of the diffusion coefficients of the compounds in the examples are shown.

[0031] Figure 2 The measured and predicted values ​​of the compound ionization efficiency under the positive ionization mode in the examples are shown.

[0032] Figure 3 The measured and predicted values ​​of the compound ionization efficiency under negative ionization mode in the examples are shown.

[0033] Figure 4 This example compares the predicted and measured concentrations of 15 compounds. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0035] Step (1) Construction of diffusion coefficient prediction model

[0036] New pollutant D values ​​were collected from the literature. Only data measured using agarose gel as the diffusion layer were retained. The diffusion coefficients of all compounds measured at different temperatures were uniformly converted to values ​​at 25 °C using formula (1), in centimeters per second (cm² / s).

[0037] (1)

[0038] Where T is the temperature set for the experiment; D T (cm² / s) is the diffusion coefficient of chemical pollutants in DGT at temperature T; D 25 (cm² / s) is the diffusion coefficient of chemical pollutants in DGT at a temperature of 25 °C.

[0039] Finally, all D values ​​were converted to negative logarithmic form. Data for 187 novel pollutants were collected through this process for model building. Before modeling, the SMILES strings of compounds were converted into numerical features, including Molecular Access System Bonds (MACCS), Avalon Molecular Fingerprints (AFP), Extended Connectivity Fingerprints (ECFP), Counted Morgan Fingerprints (C_ECFP), Mordred Molecular Descriptors (Mordred), and RDKit Molecular Descriptors (RDKit). Based on the permutations and combinations of different features, a predictive model was developed using the XGBoost algorithm.

[0040] The dataset was randomly divided into a training set (80%) and a test set (20%). Hyperparameter optimization was performed on the training set using random search and five-fold cross-validation to select the optimal hyperparameter combination (Table 1). Model performance was measured using the root mean square error (RMSE). RMS ) and coefficient of determination (R) 2 Evaluation. Using S cutoff With N min The application domain of the model is defined. Results show that the model based on the combination of MACCS, ECFP, C_ECFP, and RDKit performs best. Its performance on the test set is R... 2 test = 0.733, E RMS-test =0.084( Figure 1 This indicates that the model has good predictive and generalization capabilities.

[0041] Table 1. Hyperparameters of the best prediction model

[0042]

[0043] To strike a balance between model performance and application domain coverage, a relatively loose application domain boundary was adopted for the model. The threshold for the -logD model was set to S. cutoff = 0.15 and N min = 1. This set of thresholds is used to determine whether a new pollutant falls within the application domain of the model.

[0044] Step (2) Construction of ionization efficiency prediction model

[0045] LogIE data (logIE) for ESI positive and negative ionization modes were collected from the literature. pos With logIE neg Information such as instrument type, mobile phase composition, pH, and solvent properties. pos The dataset contains 335 unique chemical substances, totaling 3319 data points; logIE negThe dataset contains 132 unique chemical substances, totaling 1319 data points. Similar to the diffusion coefficient model, molecular structure features of the same type are generated based on the compound's SMILES string. Non-numerical information such as instrument type, mobile phase composition, pH, and solvent properties are converted into descriptors containing 0s or 1s using one-heat encoding. For logIE... pos The prediction model yielded 32 relevant descriptors. (For logIE...) neg The prediction model yielded 16 relevant descriptors. The model was developed using the XGBoost algorithm, and hyperparameters were optimized through random search and five-fold cross-validation (Table 1). The modeling performance of different feature combinations was examined.

[0046] For logIE pos The optimal feature combination for the model is RDKit and Mordred descriptors. The best model's performance on the test set is R0. 2 test = 0.917, E RMS_test = 0.402 ( Figure 2 This indicates that the model has good predictive and generalization abilities. (logIE) neg The optimal combination of models is C_ECFP and Mordred. The performance of the optimal model on the test set is R. 2 test =0.894, E RMS_test = 0.361( Figure 3 The optimal model hyperparameters are shown in Table 1. (logIE) pos The application domain threshold of the model is S cutoff =0.55 and N min = 3; logIE neg The threshold of the model is S cutoff = 0.35 and N min = 1.

[0047] Step (3) Fitting the relationship between response factor and ionization efficiency

[0048] Using 51 standards (Table 2), standard curves were constructed under both positive and negative ESI ionization modes to obtain the radiometric emission (RF) for each compound, which was then converted to logarithmic form (logRF). Based on logIE... pos and logIE neg The predictive model obtains the logIE of the standard product under the corresponding mode. pos and logIE negPredicted values. Linear relationships between the predicted values ​​of logRF and logIE were fitted under different modes, and their regression equations were obtained. The results show that 26 and 25 compounds exhibited effective ionization responses in positive and negative ionization modes, respectively. The obtained regression equations are as follows:

[0049] logRF pos = 0.669 logIE pos + 4.228

[0050] logRF neg = 0.666 logIE neg + 5.125

[0051] The coefficient of determination R of the regression equation 2 The values ​​are 0.605 and 0.768 respectively, indicating that logRF and logIE... pre It exhibits a good linear relationship.

[0052] Table 2 Information on 51 Standard Products

[0053]

[0054]

[0055]

[0056] Step (4) Validation and application of the semi-quantitative strategy

[0057] A diffusion gel sampler (DGT) was prepared and deployed for sampling novel contaminants. The DGT sampler consisted of a plastic shell, a filter membrane, a diffusion gel layer, and a binding layer. The diffusion gel layer was prepared using agarose, and the binding layer was prepared using a non-polar macroporous adsorption resin (XDA-1) and a weak anion exchange resin (WAX) suitable for sampling different types of novel contaminants. The DGT preparation method was based on previous literature (Sci. Total Environ. 2018, 618, 1605-1612).

[0058] The assembled sampler was stored in a 0.01 mol / L sodium chloride solution at 4 °C. During sampling, the DGT sampler was deployed at a wastewater treatment plant in Dalian City for 7 consecutive days. Water temperature was recorded in real time using a thermometer to correct for the diffusion coefficient. The WAX ​​binding layer was eluted with 10 mL of 5‰ ammonia-methanol solution for 24 hours, and the XDA-1 binding layer was eluted with 10 mL of methanol by sonication for 20 min. The eluent was dried under nitrogen, redissolved in 100 μL of methanol, and filtered through a 0.22 μm PTFE membrane. Analysis was performed using a UHPLC-Q-Exactive Orbitrap mass spectrometer. The instrument was operated in both positive and negative ionization modes, and data were acquired using full scan and data-dependent secondary mass spectrometry modes.

[0059] The mass spectrometry instrumentation methods are shown in Table 3. Chromatographic separation employed a gradient elution program with a mobile phase flow rate of 0.3 mL / min. For analysis in positive ionization mode, the initial ratio of mobile phase A (0.1% formic acid aqueous solution) to B (methanol) was 95:5, maintained for 0.5 min; then, at 11 min, the gradient transitioned to 1:99 and was maintained for 14 min; finally, at 17 min, the ratio was restored to the initial 95:5. For analysis in negative ionization mode, the initial ratio of mobile phase A (10 mM ammonium acetate aqueous solution) to B (methanol) was 97:3, transitioning to 50:50 at 4 min; then, at 17 min, it was further adjusted to 2:98 and maintained for 22 min; finally, at 22.1 min, the ratio was restored to 97:3 and maintained for 25 min.

[0060] Table 3 Mass Spectrometry Instrument Methods

[0061]

[0062] Non-targeted analysis was performed using Compound Discoverer 3.3, retaining peaks with a height greater than 10,000, a signal-to-noise ratio greater than 3, and a response area greater than 3 times that detected in blank samples. The mzCloud mass spectrometry database was used for matching to identify structures. For candidate structures with a mass deviation ≤ 5 ppm, a mass spectrometry similarity match score ≥ 80, and a matching fragment number ≥ 3, the structure with the highest match score was retained. The logIE of the identified compounds was predicted using a model. pre The results were then converted to a specific RF for the current instrument using a regression equation. The cumulative amount of compounds in the sample was calculated by combining the response peak area and sample volume, and the concentration of compounds in the environment (C0) was further calculated by combining the diffusion coefficient and sampling time. DGT The reliability of the method was evaluated by performing targeted quantification of some pollutants.

[0063] Targeted quantification was performed using 15 standards (Table 4 and ). Figure 4The results showed that the average multiple error between the measured concentration and the predicted concentration was 2.2 times. For the 11 compounds in the application domain, the average multiple error was 1.9 times, indicating that the accuracy of this method is acceptable.

[0064] Table 4 Information on 15 Standard Samples

[0065]

[0066] In addition to the 15 structures identified by targeted analysis, 85 compounds were identified in this embodiment through non-targeted analysis, of which 43 were detected in positive ion mode and 42 in negative ion mode. The identified compounds can be divided into eight categories: pharmaceuticals (23.53%), functional additives (18.82%), transformation products (15.29%), natural products (15.29%), industrial materials (9.41%), other categories (8.24%), pesticides (5.88%), and personal care products (3.54%).

[0067] Using the established prediction model, a semi-quantitative analysis was performed on the above compounds (Table 5), with predicted concentrations ranging from 2.1 to 11009.9 ng / L. Pharmaceutical compounds showed the highest detection rate, with predicted concentrations between 2.1 and 1000.8 ng / L. Some antihypertensive drugs and arthritis medications were detected. The highest concentration was found in estriol (1000.8 ng / L). The predicted concentrations of functional additives ranged from 13.1 to 2852.1 ng / L, mainly including plasticizers, preservatives, and antibacterial agents such as butylparaben and triclocarban, with N-butylbenzenesulfonamide showing the highest concentration (2852.1 ng / L). In addition, various transformation products were detected, including drug metabolites and plasticizer monoester transformation products. The highest concentration of galazone, a transformation product of synthetic musk, was 11009.9 ng / L. The results show that this method can achieve effective identification and semi-quantitative analysis of compounds in the absence of standards, and the predicted concentration is comparable to that reported in the literature, verifying the reliability of the method and its application potential in high-throughput screening of water pollutants.

[0068] Table 5. Compounds identified by non-targeted analysis and semi-quantitative results

[0069]

[0070]

[0071]

[0072] Note: The predicted D value is at 19.3℃ (water temperature), which is corrected from the D value of the compound predicted by the model at 25℃.

Claims

1. A smart passive sampling and non-targeted analysis method for monitoring emerging pollutants in water, characterized in that, Includes the following steps: Step (1) Data collection and preprocessing Collect the diffusion coefficient D values ​​of new pollutants in the agarose gel diffusion layer and uniformly correct them to values ​​at 25°C; then convert all D values ​​to the negative logarithmic form -logD with base 10. Collect ionization efficiency data of new pollutants under positive and negative ion modes of electrospray ionization source (logIE) pos and logIE neg These are used to construct prediction models for the ionization efficiency of new pollutants under different modes; Step (2) Model Building To construct a diffusion coefficient −logD prediction model, firstly, molecular descriptors and molecular fingerprints are calculated based on the structure of pollutants in step (1) as molecular structural features; the dataset is divided into training set and test set, and the Extreme Gradient Boosting Tree (XGBoost) algorithm is used to arrange and combine all molecular descriptors and molecular fingerprints to form n sets of feature combinations. The n sets of feature combinations are used as input and −logD is used as output to construct n −logD models. For logIE pos and logIE neg In constructing the prediction model, in addition to molecular structure features, the input features include mobile phase composition, gradient, instrument type, and ion source technology descriptor; a one-hot encoding method is used to convert non-numerical information into binary features; The dataset was divided into training and testing sets, and the XGBoost algorithm was used to build the model. In the input features, the mobile phase composition, gradient, and instrument type descriptors remained constant, while the molecular structure features were constructed using combinations of different molecular descriptors and molecular fingerprints. Specifically, all molecular descriptors and molecular fingerprints were permuted and combined to form n feature combinations, thus constructing n logIEs. pos and logIE neg Model; Step (3) DGT Sampling and Analysis The DGT sampler was deployed in the water body to be tested and sampled continuously for a certain period of time. After the sampling was completed, the binding layer was eluted and concentrated. High performance liquid chromatography-high resolution mass spectrometry was used to perform non-targeted analysis on the samples. Electrospray ionization mode was used to detect compounds in the samples in both positive and negative modes. Non-targeted data were collected in a data-dependent mode. The detected peaks were aligned, denoised, and screened. The structure of new pollutants in the samples was identified by matching the secondary mass spectrometry database using unknown substance screening software. Step (4) Application of semi-quantitative strategies Using two sets of standards that can be ionized in positive and negative ionization modes respectively, a standard curve is established, and the slope RF of the standard curve is obtained. The logIE is then used to... pos and logIE neg The predictive model obtains the logIE of the standard product under the corresponding mode. pos and logIE neg Predicted values; fitting logRF and logIE under positive and negative ionization modes respectively. pos and logIE neg Find the linear relationship between the predicted values ​​and obtain the formula. Based on the structure of the newly identified pollutants, −logD and logIE were used. pos and logIE neg Predictive models predict -logD and logIE for new pollutants. pos and logIE neg value; Using the logRF and logIE obtained above pre Based on the relationship, the RF is derived, and combined with the response area of ​​the new pollutant, the concentration C of the new pollutant in the sample is calculated using the following formula. S : C S = Area / RF Where Area represents the peak area of ​​the newly identified pollutant; Further calculation of the cumulative amount of new pollutants in DGT allows for the derivation of the time-weighted average concentration of new pollutants in the environment, enabling high-throughput quantitative prediction of new pollutants in water without the need for standards. The specific formula is as follows: ； in, M is the concentration of new contaminants in the water measured by DGT; M is the mass of new contaminants accumulated on the binding layer, which is equal to the volume of the eluent and the concentration of new contaminants in the eluent. S The product of Δg and Δg; Δg is the thickness of the diffusion layer; D is the diffusion coefficient of the new pollutant in the diffusion layer, obtained by the -logD prediction model; A is the area of ​​the DGT exposure window; t is the deployment time of the DGT sampler in the water.

2. The intelligent passive sampling and non-targeted analysis method for monitoring new pollutants in water according to claim 1, characterized in that, In step (1), the new pollutants are antibiotics, perfluorinated compounds, and endocrine disruptors.

3. The intelligent passive sampling and non-targeted analysis method for monitoring new pollutants in water according to claim 1, characterized in that, In step (2), the model performance is evaluated through cross-validation and an external test set; the root mean square error Em is used. RMS Using the coefficient of determination R² as the evaluation index, the optimal values ​​are -logD and logIE. pos and logIE neg The predictive model is defined by a cosine similarity score threshold between the structure to be predicted and the structures in the training set, and the minimum number of compounds in the training set that satisfy the threshold, thus ensuring the reliability of the model's predictions.

Images