A QSAR–ICE–SSD model prediction of the PNECs for alkylphenol substances and application in ecological risk assessment for rivers of a megacity

Abstract

Alkylphenols (APs) are ubiquitous and generally present in higher residue levels in the environment. The present work focuses on the development of a set of in silico models to predict the aquatic toxicity of APs with incomplete/unknown toxicity data in aquatic environments. To achieve this, a QSAR-ICE-SSD model was constructed for aquatic organisms by combining quantitative structure–activity relationship (QSAR), interspecies correlation estimation (ICE), and species sensitivity distribution (SSD) models in order to obtain the hazardous concentrations (HCs) of selected APs. The research indicated that the keywords “alkylphenol” and “nonylphenol” were most commonly studied. The selected ICE models were robust (R²: 0.70–0.99; p-value < 0.01). All models had a high reliability cross- validation success rates (>75%), and the HC₅ predicted with the QSAR-ICE-SSD model was 2-fold than that derived with measured experimental data. The HC₅ values demonstrated nearly linear decreasing trend from 2-MP to 4-HTP, while the decreasing trend from 4-HTP to 4-DP became shallower, indicates that the toxicity of APs to aquatic organisms increases with the addition of alkyl carbon chain lengths. The ecological risks assessment (ERA) of APs revealed that aquatic organisms were at risk from exposure to 4-NP at most river stations (the highest risk quotient (RQ) = 1.51), with the highest relative risk associated with 2.9% of 4-NP detected in 82.9% of the sampling sites. The targeted APs posed potential ecological risks in the Yongding and Beiyun River according to the mixture ERA. The potential application of QSAR-ICE-SSD models could satisfy the immediate needs for HC₅ derivations without the need for additional in vivo testing.

1. Introduction

Alkylphenols (APs) have attracted worldwide attention as an important group of emerging contaminants, as exposure to even ng/L levels may affect the structure or function of endocrine and reproductive systems through disruption of hormone biosynthesis and metabolism (Ashfaq et al., 2019; Jin et al., 2014; Lee et al., 2022; McLaren and Rawlins, 2022; Nowak et al., 2019; Ying et al., 2002). APs exert endocrine disrupting effects by binding to estrogen receptors in organisms, leading to disruptions in reproductive function, secretion, development, and immunity (Li et al., 2019; Liu et al., 2022a). They can enter humans and ecological receptors through various pathways, including (but not limited to) ingestion through the food chain, dermal absorption, and inhalation, and may contribute to many human diseases such as diabetes, obesity, reproductive disorders, cardiovascular disease, birth defects, and breast cancer (Qiu et al., 2021). As a group, APs are ubiquitous in the environment and generally have higher residue levels compared to other endocrine disrupting chemicals (EDCs), such as natural and synthetic estrogens (Urbatzka et al., 2012; Wang et al., 2016). APs in aquatic environments are mainly derived from the degradation of AP ethoxylates (Bergé et al., 2014). APs have been frequently detected in water and sediments of both freshwater and marine environments (Cheng et al., 2018; Graca et al., 2016; Koh et al., 2006; Liu et al., 2016a). Generally, APs are detected in water bodies and tissue samples at concentration ranging from ng/L to μg/L (Gu et al., 2016; Yang et al., 2014), and their remaining concentrations in surface water can cause notable estrogenic potency in aquatic wildlife (Wang et al., 2018).

Beijing is one of the largest megacities in the world. The city is scattered with man-made, semi-natural and natural lakes and land systems (Zhang et al., 2022). The Yongding River and Beiyun River, the northern branch of the Haihe River, flow through most of Beijing and are important water sources and ecological corridors. The dense population and intensive industrial and agricultural activities inevitably lead to the continuous discharge of organic pollutants into the rivers (Lei et al., 2019; Pernet-Coudrier et al., 2012; Zhang et al., 2014). In a recent study, 4-t-octylphenol and nonylphenol were detected in 100% of water samples collected from Beijing–Tianjin–Hebei River agglomeration, with concentrations ranging from 23 to 255 ng/L (Lei et al., 2021). Therefore, understanding the toxicity of APs to aquatic organisms is essential to protect the environment and human health (Lu et al., 2015).

The toxicity of a compound provides the basis for deriving a hazardous concentration (HCs) that is used in ecological risk assessment (ERA) to assess potential effects of a compound on aquatic organisms (Fu et al., 2021; Guo et al., 2021; Zhang et al., 2020; Feng et al., 2019; Feng et al., 2012; Macleod et al., 2004). A reliable evaluation of the potential effects of a chemical on aquatic ecosystems requires ecotoxicological data for a diversity of species, which is typically not available in practice (Douziech et al., 2020; Johnson et al., 2020). In addition, increasing the number of species for which chemical effects are measured experimentally is often impractical and discouraged by international efforts to reduce animal testing (Fentem et al., 2021). Alternatively, the paucity of data can lead to the non-commissioning of these chemicals due to the lack of risk evaluation. Therefore, it is necessary to apply toxicity extrapolation and other in silico approaches that use available toxicity data for surrogate species to derive hazard levels protective of diverse species (Wu et al., 2021; Zhong et al., 2021).

The use of models to extrapolate and predict toxicity data is widely applied internationally (Bejarano et al., 2017; Brill et al., 2016; Fan et al., 2019; Fojut et al., 2012a; Fojut et al., 2012b; Khan et al., 2019; Roy et al., 2015; Wang et al., 2020; Wu et al., 2015; Wu et al., 2016; Zhang et al., 2017b). Quantitative Structure Activity Relationships (QSAR) models are based on a variety of mathematical approaches to predict activities and properties of untested chemicals on the basis of their molecular structure (Chen et al., 2015; Mu et al., 2014; Wu et al., 2013; Huang et al., 2022). QSAR model are recognized worldwide as alternatives to animal testing to estimate chemical effects on cells and tissues and untested vertebrates, invertebrates, and microorganisms (Galimberti et al., 2020). Typically, QSAR models predict acute toxicity representative for a fish, an invertebrate, and an algal species. Interspecies Correlation Estimation (ICE) models are log-linear least squares regressions of acute toxicity developed from measured toxicity of chemicals previously tested in two taxa. In application, ICE models use available toxicity data of tested species (i.e., surrogate species) as input to predict sensitivity of untested taxa (i.e., predicted species, genus, family) (Bejarano et al., 2017; Feng et al., 2013a; Feng et al., 2013b; Raimondo et al., 2010; Shen et al., 2022; Willming et al., 2016). ICE models have been widely used to predict the acute toxicity of chemical substances and to derive protective toxicity thresholds with good predictive results (Fan et al., 2019; Feng et al., 2013b; Raimondo et al., 2010; Wang et al., 2020; Willming et al., 2016) and use of QSAR-estimated values for fish and invertebrates as input into ICE models has been shown to accurately predict acute toxicity to a diversity of species (Bejarano et al., 2017; Douziech et al., 2020). Since Web-ICE provides a range of ICE models of varying robustness, model selection criteria recommended by Willming et al. (2016) were applied in the present study. To reduce uncertainty associated with ICE predictions, we only used models that met the following criteria: 1) model mean square error (MSE) ≤ 0.95; 2) correlation coefficient, R² > 0.6; 3) slope > 0.6. Species Sensitivity Distributions (SSD) are cumulative probability distributions of chemical toxicity on diverse species, and the 5th percentile of the SSD is most often used as the HC₅ (hazardous concentration for 5% of species) applied in many international risk evaluations (Belanger et al., 2017). SSDs composed of ICE-estimated acute toxicity has been demonstrated to be comparable to SSDs developed from measured data (Dyer et al., 2008; Wu et al., 2015; Wu et al. 2016; Zhang et al., 2017b).

The combination of QSAR and ICE models can greatly expand the ability to predict toxicity of untested chemicals, estimate potential effects on untested species, and inform SSDs (Barron et al., 2012; Douziech et al., 2020; Raimondo et al., 2022; Zhang et al., 2021). Several applied studies have demonstrated the comparability of HC₅s derived from combined QSAR-ICE-SSD models to SSDs developed from measured toxicity (Bejarano et al., 2017; Douziech et al., 2020; Golsteijn et al., 2012; Gredelj et al., 2018; Zhang et al., 2017b). Nevertheless, QSAR and ICE models both have limitations that may be compounded by linking models if uncertainties are not properly considered. QSAR models are only able to predict chemical toxicity for certain species or taxa types (e. g., “Daphnia”, “fish”). The majority of available ICE models are based on North American species and can predict to species, genus, and family levels. A number of studies have shown good comparison between HC₅s generated from measured data with those from QSAR-ICE-SSDs models (e.g., Douziech et al., 2020; Zhang et al., 2017). Conversely, Barron et al. (2012) did not observe strong agreement of HC₅s derived using QSA-RICE-SSD models with those based on measured data; however, comprehensive uncertainty analyses were not performed in Barron et al. (2012) to identify the source of the disagreement. Similarly, He (He et al., 2017) recently presented combined QSAR-ICE models to derive water quality criteria but did not discuss how to appropriately characterize the uncertainty associated with combining the two models. Therefore, there is a need to better evaluate and understand the sources of uncertainty for HC₅s based on in silico methods only. Moreover, such evaluations should be extended to include diverse chemical groups. Finally, there is a need to validate the science and accuracy of the predictions using QSAR-ICE-SSD models to increase the reliability of the results.

The objective of the present study was to apply a set of in silico models to predict the aquatic toxicity of APs with incomplete/unknown toxicity data in the aquatic environment and demonstrate how the models can be used in an ERA. To achieve this, we 1) analyze the current research status of APs using bibliometric software, 2) acquire toxicity data of several APs with structures similar to nonylphenols from the ECOTOX database, 3) applied the QSAR-ICE-SSD models to predict the toxicity data of APs and derive HC₅ values for APs with lack of sufficient toxicity data, and 4) conduct an ERA for APs in two rivers in China.

2. Materials and methods

2.1. Data collection and analysis for bibliometric

The Web of Science (WOS) Core Collection SCI-expand database (https://www.webofscience.com/wos/alldb/basic-search, last accessed on October 15, 2021) was used to assess the availability of previously published studies for APs. The WOS Core Collection was chosen since it is considered the most comprehensive database, embodying the most influential and relevant journals. An advanced search was performed on the literature, and the specific requirements were as follows: 1) the search topic (TS) = (“alkylphenol*” OR “alkyl-phenol*” OR “alkyl phenol*”) 2) the type of the document was set to “article”, 3) the language of the article was set to English, and 4) the search data was set from 1975 to October 15, 2021. A second round of manual screening was conducted based research hotspot analysis (keyword co-occurrence and clustering) using CiteSpace software (Kolahi et al., 2017; Niu et al., 2021). In the network of co-occurrence and clustering analysis, nodes represent specific key terms, such as country, institution, or keyword, where the larger the node size, the higher the frequency of occurrence within the literature.

2.2. Aps selected for this study and toxicity data collection

Data were available for APs that were structurally similar to nonylphenol and were: 2-methylphenol (2-MP), 4-ethylphenol (4-EP), 2-propylphenol (2-PPP), 2-butylphenol (2-BP), 4-pentylphenol (4-PTP), 4-hexylphenol (4-HXP), 4-heptylphenol (4-HTP), 4-octylphenol (4-OP), 4-nonylphenol (4-NP), 4-decylphenol (4-DP). The basic physicochemical property parameters of these APs were obtained from website searches (https://pubchem.ncbi.nlm.nih.gov, last accessed on October 30, 2021; https://www.chemspider.com, last accessed on October 30, 2021) and are presented in Supplementary Information (SI) (Table S1). Acute and chronic toxicity data were obtained from the United States Environmental Protection Agency (US EPA) ECOTOX Database (https://cfpub.epa.gov/ecotox/, last accessed on October 30, 2021). The available toxicity data for APs are provided shown in the Supplemental Information (Table S2–S13). The estimated and actual toxicity values of 4-OP, 4-NP, 4-DP were exceeded their water solubility. As long as the toxicity concentration is not higher than 10 times of solubility, it is generally considered to be reliable data (Mensink et al., 2008). In addition, Solvents may be used to prepare stock solutions. It is, however, usually not advised that solvents be added directly to the test vessels to enhance solubility. Solvents should not be toxic to the tested species at the tested concentrations (ECA, 2008). According to several OECD guidelines, the concentration of solvents should not exceed 0.01%. The highest solvent concentration used should be reflected in a solvent control (Liu et al., 2016b; Moermond et al., 2016). In addition, in the derivation of the PNEC value of pollutants, the toxicity data used are generally at the level of μg/L or mg/L, it is only a statistical value, does not mean that this value must be below the water solubility of pollutants, because in some actual environments, pollutants are likely to occur enrichment, such as enrichment in particulate matter or some algae, or joint reactions with other pollutants, and in the derivation of the PNEC value is generally taken HC₅, for the toxicity data is particularly high is not very influential, thus, for scientific research needs, we put this part of the data together to discuss.

In the present study, two separate datasets were used to develop SSDs: (1) measured APs toxicity data obtained from the literature as provided in Table S2–S13 and (2) data derived from QSAR-ICE model. The ICE models used to predict acute toxicity from available surrogate species were from the publicly available Web-ICE application (www3.epa.gov/webice, version 3.3, last accessed on December 15, 2021). Screening and evaluation of measured data followed the principles of accuracy, relevance, and reliability presented in Feng et al. (2013b) and Liu et al. (2020b). Data were required to have been obtained from tests conducted in accordance with standard test methods (e.g., the methods of the Organization for Economic Cooperation and Development). Briefly, aquatic acute toxicity tests found within the database or literature included the following characteristics: 96-h and 48-h 50% of the lethal concentration (LC₅₀) or 50% of the effect concentration (EC₅₀) for fish and invertebrates, respectively, and the toxicological endpoints were death or a terminal effect such as immobility or respiratory inhibition. For chronic toxicity data obtained from durations > 96-h, the No Observable Effect Concentration (NOEC) was the priority endpoint; if unavailable, the Lowest Observable Effect Concentration (LOEC), Maximum Acceptable Toxic Concentration (MATC), or EC_X (X = 10 or 20) were used. In situations where different endpoints were available for the same species, the most sensitive endpoint was selected. When multiple toxicity values were available for the same species and the same endpoint, the geometric mean was taken as the mean toxicity value for the species.

The derivation of acute and chronic HC₅s for the APs followed the steps in Fig. 1. The toxicity data were separated into acute and chronic toxicity data. If the data were acceptable for criteria derivation (Feng et al., 2019), then they were used directly in the SSD to derive an HC₅. Data for a given chemical were considered insufficient for the derivation of HC₅ if they did not: (1) Cover at least 3 different trophic levels including producers; (2) contain at least 10 species including the following taxa: 1 species of carp of the family Scleractiniaceae, 1 species of non-carp of the family Scleractiniaceae, 1 species of zooplankton, one non-fish benthic (e.g., shellfish, benthic crustaceans, etc.), one amphibian or other aquatic animal belonging to a different phylum from the above, one phytoplankton or 1 amphibian or other aquatic animal belonging to a different phylum from the above, 1 phytoplankton or aquatic vascular plant. The QSAR-ICE model was used to predict toxicity to diverse taxa and derive the acute HC₅ when the above listed requirements were met by measured data. All available acute and chronic measured toxicity data available for the same AP compound were used to calculate the mean acute-to-chronic ratio (ACR) of HC₅ values. The chronic HC₅ values for other APs that did not have sufficient measured chronic data were obtained by dividing the corresponding acute HC₅ values by ACR.

Fig. 1.

Step-by-step diagram of the HC₅ derivation for APs from measured data as reported in the ECOTOX database.

Available measured data were only sufficient to develop measured SSDs for the 2-MP acute data, 4-OP chronic data, 4-NP acute data, and 4-NP chronic data (Table S2–S13). QSAR-ICE models were used to predict toxicity data for other APs as the basis for SSD to derive acute HC₅ values, as described below.

2.3. Toxicity data prediction using QSAR-ICE model

The Toxicity Estimation Software Tool (TEST, version 5.1, 2020) is a QSAR modeling application developed by US EPA to estimate toxicities and physical properties from molecular structure. TEST estimates toxicity values for several endpoints: 1) 96-h Fathead minnow (Pimephales promelas) LC₅₀, 2) 48-h Daphnia magna LC₅₀, 3) 48-h Tetrahymena pyriformis (T. pyriformis) 50% of the growth inhibition concentration (IGC₅₀) (Martin et al., 2008). Web-ICE was then used to extrapolate acute values for untested species using the sensitivities of the Fathead minnow and Daphnia magna as predicted by TEST. Because there is little correlation between algal and animal taxa, our work focused only on invertebrate and fish species.

The combined QSAR-ICE approach was developed from models based on surrogate species using the following steps: As noted above, the first step used QSAR models to predict toxicity to the Fathead minnow and Daphnia magna using the TEST models for fish and invertebrates, respectively. Second, all available ICE models for these two surrogate species were identified. Third, QSAR estimates of surrogate species sensitivity were entered into the appropriate ICE models, resulting in the combined QSAR-ICE model. The QSAR models are defined as equation (1):

$${\text{Log}}_{10}(\text{Surrogate}\text{Toxicity})=a_0+a_1{f}_1+a_2{f}_2+\cdots +a_n{f}_n$$ (1)

Where f₁, f₂ and f_n are the independent descriptors, such as molecular weight, the octanol–water partition coefficient, heat of formation, total energy, electronic energy, gradient norm, gradient norm per atom, ionization potential, etc., and a₀, a₁, and a_n are fitted parameters. The ICE models are defined as equation (2):

$${\text{Log}}_{10}(\text{Predicted}\text{Toxicity})=a+\text{b}\times {\text{Log}}_{10}(\text{Surrogate}\text{Toxicity})$$ (2)

Where a is the intercept and b is the slope of the regression. In this case, the output of equation (1) is used as the input of equation (2). The above conditions were required to be met by ICE model for inclusion in the present study.

2.4. The HC₅ derivation and reliability verification

To develop SSDs for acute measured values, acute predicted values, and chronic measured values, sensitivity of each species within each data set were ranked and assigned percentiles. The data were fitted with log–logistic distribution to construct SSDs. The acute HC₅ for SSDs was determined by the Origin 9.0 statistical software (Liu et al., 2020a). Species ranks from the two distributions were visually examined to determine the similarities of species placement in the SSD curves. The most sensitive species were located in the lowest percentiles of the SSD, such that the species in the first quartile of the SSD curve was the most sensitive species in the dataset for that chemical. To verify the reliability of the QSAR-ICE-SSD model, the predicted HC₅ values were compared to those derived from the measured acute toxicity data for 2-MP and 4-NP.

2.5. Ecological risks assessment of APs to aquatic organisms in river

2.5.1. Sample collection, pretreatment and analysis of APs

Exposure scenarios were derived from water samples collected from two rivers, Yongding River (115.71°−116.26°E, 39.58°−40.1°N) and Beiyun River (116.64°−116.90°E, 39.76°−39.93°N), which flow through Beijing from North to South. A total of 35 river water samples (27 of the Yongding River and 8 of the Beiyun River) were collected in January 2022. These samples were collected in pre-cleaned amber glass bottles. After sealing without headspace, the samples were transported to the laboratory within 2 h for pretreatment. In total, four field blank samples were collected during the sampling period. Sample pretreatment was conducted according to previously published methods (Gu et al., 2016; Li et al., 2019) with some modifications. First, the water samples were filtered through 0.7 μm glass microfiber filters to remove suspended particles. Then, 500 mL of filtered water sample was spiked with 100 ng/L of three internal standards (200 μL) and extracted on Oasis HLB cartridges (500 mg, 6 mL, Waters Corporation, Milford, MA, USA). The cartridges were conditioned with 6 mL methyl tert-butyl ether (MTBE), followed by 6 mL HPLC-grade methanol (MeOH) and 12 mL Milli-Q (MQ) water. The water samples were passed through the cartridges at a flow rate of 5–10 mL/min. After being loaded with water samples, the cartridges were washed with 12 mL of MQ water and then dried under vacuum for 30 min. Finally, the analytes were eluted with 6 mL mixture of MTBE and MeOH (v/v 1:1). The eluents were concentrated under a stream of nitrogen and redissolved in 200 μL MeOH for further instrument analysis. More detailed information on sample pretreatment and instrumental conditions were described in the Supplemental Information (Text S1).

2.5.2. Rqs of individual compound

The risk quotient (RQ) is internationally adopted approach used to assess potential ecotoxicological risk of compounds in aquatic ecosystems (Gros et al., 2010). The chronic and sublethal risks of APs in two rivers of Beijing were assessed using the RQ, which is the ratio of an estimated or measured environmental concentration of a chemical divided by the predicted no effect concentration (PNEC), such that (Eq. (3)):

$$\text{RQ}=\frac{\text{MEC}}{\text{PNEC}}$$ (3)

where MEC is the measured environmental concentration of a single compound at each site and the PNEC concentration at which no adverse effects are likely to occur. The PNEC was obtained using Equation (4):

$$\text{PNEC}=\frac{{\text{HC}}_5}{\text{SF}}$$ (4)

where the HC₅ was obtained from the constructed SSD curve of chronic data. PNECs were calculated as the derived HC₅ divided by a safety factor (SF), which was determined by the amount of supporting toxicity data (Hou et al., 2022). In the present study, the SF was set to 2 when the number of species for which data was available exceeded 15 (ECB, 2002). For each chemical, the risk assessment ranking criterion were classified as no risk if RQ < 0.01; low risk if 0.01 ≤ RQ < 0.1; moderate risk if 0.1 ≤ RQ < 1, and high risk if RQ ≥ 1(EMA, 2006; Kosma et al., 2014).

2.5.3. RQs of mixture compounds

Aquatic organisms are typically exposed to complex mixtures of pollutants in their natural environment, and synergistic effects may occur when various pollutants co-occur (Riva et al., 2019). Therefore, considering one compound at a time is not sufficient to predict the synergistic effects of chemical mixtures in aquatic ecosystems. The cumulative risk of mixtures is predicted according to the framework proposed by Backhaus and Faust (2012). The mixture risk quotient (RQ_MIX) was calculated as (Eq. (5)):

$${\text{RQ}}_{\text{MIX}}={\sum }_{\text{i =1}}^{\text{n}}\frac{{\text{MEC}}_{\text{i}}}{{\text{PNEC}}_{\text{i}}}$$ (5)

where i represents individual compounds and n is the total number of compounds in the mixture. If the RQ_MIX exceeded 1, the potential for ecological risk posed by the mixture could not be excluded.

3. Results and discussion

3.1. Bibliometric results and analysis

The visualization of the keywords with the highest co-occurrence and frequency of major keywords was summarized and displayed in Fig. 2. The node colors range from gray to red representing the years 1975 to 2021. The highest frequencies were for “alkylphenol” (620, 77.79%) and “nonylphenol” (616, 77.29%), which confirms that these two terms were the primary topics published in the APs research field (Gu et al., 2021; Lei et al., 2021). Not surprisingly, the high frequency of keywords “water”, “waste water” and “aquatic environment” highlights the high occurrence of APs in water bodies. Other frequently detected keywords related to chemistry and analysis, including “chemical” and “mass spectrometry”, indicate that APs are still difficult to study in terms of analytical detection. This is likely due to the complex structures of APs and their many congeners and isomers, which often need to be extracted and enriched by complex pretreatment process. Furthermore, the literature review established a lack of definitive methods of the detection for APs (Lei et al., 2021; Lu et al., 2021; Wang et al., 2016). It is worth noting that the frequency of the term “Degradation” is also high, indicating that treatment measures are also needed to account for degraded APs in the water environment.

Fig. 2.

Visual clusters analysis of keywords co-occurrence.

3.2. Predicting the toxicity data using QSAR-ICE models

3.2.1. Development and verification of the models

Developing a QSAR model requires at least 10 or more compounds with similar chemical structures with the same effect endpoint for the same tested species and toxicity values obtained from the same laboratory. There were only 10 target compounds identified in the present study and the data did not meet the conditions for constructing QSARs models. Therefore, TEST was used to predict the toxicity values of 2-MP, 4-EP, 2-PPP, 2-BP, 4-PTP, 4-HXP, 4-HTP, 4-OP, 4-NP and 4-DP to 96-h LC₅₀ of Fathead minnow (Pimephales promelas), 48-h LC₅₀ of Daphnia magna, and 48-h IGC₅₀ of T. pyriformis (Table 1). The number of chemicals represent the number of chemicals that were used in the development of the QSAR model in TEST. All predictions satisfy r² > 0.6, q² > 0.6, thus, QSAR model is considered to have acceptable predictive power.

Table 1.

Predicted the toxicity values of APs using QSAR models for three species.

APs	Endpoint	Predicted value (mg/L)	r2	q2	#Chemicals
4-EP	Fathead minnow LC50 (96-h)	15.01	0.918	0.89	124
	Daphnia magna LC50 (48-h)	19.62	0.808	0.734	174
	T. pyriformis IGC50 (48-h)	80.06	0.855	0.745	57
2-PPP	Fathead minnow LC50 (96-h)	11.11	0.912	0.86	37
	Daphnia magna LC50 (48-h)	13.42	0.808	0.734	174
	T. pyriformis IGC50 (48-h)	60.09	0.946	0.915	26
2-BP	Fathead minnow LC50 (96-h)	6.59	0.912	0.86	37
	Daphnia magna LC50 (48-h)	8.2	0.808	0.734	174
	T. pyriformis IGC50 (48-h)	54.7	0.928	0.891	41
4-PTP	Fathead minnow LC50 (96-h)	2.5	0.901	0.772	35
	Daphnia magna LC50 (48-h)	1.37	0.858	0.744	19
	T. pyriformis IGC50 (48-h)	4.6	0.922	0.893	98
4-HXP	Fathead minnow LC50 (96-h)	1.37	0.986	0.978	16
	Daphnia magna LC50 (48-h)	0.85	0.82	0.752	143
	T. pyriformis IGC50 (48-h)	2.01	0.973	0.938	10
4-HTP	Fathead minnow LC50 (96-h)	0.72	0.986	0.978	16
	Daphnia magna LC50 (48-h)	0.16	0.987	0.976	11
	T. pyriformis IGC50 (48-h)	1.23	0.97	0.937	37
4-OP	Fathead minnow LC50 (96-h)	0.4	0.986	0.978	16
	Daphnia magna LC50 (48-h)	0.18	0.987	0.976	11
	T. pyriformis IGC50 (48-h)	0.64	0.97	0.937	37
4-DP	Fathead minnow LC50 (96-h)	0.14	0.986	0.978	16
	Daphnia magna LC50 (48-h)	0.21	0.987	0.976	11
	T. pyriformis IGC50 (48-h)	0.18	0.97	0.937	37

The Web-ICE models were used to predict acute toxicity values for other species from the QSAR values predicted in Table 1. When using ICE models for toxicity prediction, the robustness of the models needs to be evaluated using the statistical parameters listed in the methods to ensure the highest confidence in prediction accuracy. A description of ICE model performance and accuracy is provided by Raimondo et al. (2010) and Willming et al. (2016). These studies found that ICE models produce highly accurate predictions when the models have: 1) relatively low mean square error (MSE ≤ 0.95); 2) high square correlation coefficient (R²) value (R² > 0.6); and 3) slope > 0.6. Raimondo et al. (2010) also note the taxonomic relatedness of the surrogate and predicted species also influences model accuracy, with prediction accuracy associated with more closely related species. Thus, we used the QSAR Daphnia magna toxicity values to predict to invertebrates and the fathead minnow QSAR value as a surrogate for vertebrates (Raimondo et al., 2010; Willming et al., 2016).

Since T. pyriformis was not available as a surrogate in the Web-ICE database, only Daphnia magna and Pimephales promelas were used as surrogate species for toxicity prediction in this study. All toxicity values predicted from the 2 surrogate species were pooled for a total of 23 predicted species. The statistical parameters of the ICE models for predicting to untested species using Daphnia magna and Pimephales promelas as surrogate species are shown in Table 2.

Table 2.

Statistical parameters of the ICE model using Daphnia magna and Pimephales promelas as surrogate species. R²: correlation coefficient, MSE: mean square error, CVSR: cross- validation success rates, TD: taxonomic distance.

Surrogate species	Predicted species	Intercept	Slope	R2	MSE	CVSR (%)	TD
Daphnia magna	Thamnocephalus platyurus	0.22	0.92	0.9881	0.060	91	4
	Ceriodaphnia dubia	− 0.19	1.00	0.9596	0.266	81	2
	Daphnia pulex	− 0.14	1.01	0.9713	0.124	90	1
	Simocephalus serrulatus	− 0.04	1.00	0.8832	0.217	87	2
	Paratanytarsus parthenogeneticus	0.74	0.94	0.9893	0.044	100	5
	Utterbackia imbecillis	0.16	0.90	0.9694	0.115	100	6
	Lymnaea stagnalis	− 0.01	1.01	0.9620	0.191	78	6
	Physa gyrina	− 0.02	0.99	0.9694	0.147	89	6
	Amblema plicata	− 0.09	0.88	0.9500	0.188	90	6
	Branchinecta lynchi	0.31	0.90	0.9804	0.093	100	4
	Megalonaias nervosa	− 0.03	0.93	0.9625	0.163	91	6
	Margaritifera falcata	0.41	0.86	0.9588	0.147	90	6
Pimephales promelas	Lepomis macrochirus	− 0.02	0.93	0.7531	0.575	77	4
	Ictalurus punctatus	0.07	0.97	0.8465	0.401	82	4
	Oncorhynchus kisutch	− 0.09	0.82	0.7518	0.478	75	4
	Cyprinus carpio	− 0.35	1.04	0.9173	0.197	80	2
	Oncorhynchus clarkii	− 0.41	0.94	0.7943	0.392	81	4
	Jordanella floridae	0.06	0.91	0.8643	0.380	90	4
	Carassius auratus	0.07	0.98	0.9672	0.109	95	2
	Poecilia reticulata	0.29	0.86	0.8375	0.276	78	4
	Oryzias latipes	0.50	0.92	0.9291	0.235	78	4
	Oncorhynchus mykiss	− 0.28	0.97	0.8384	0.368	84	4
	Cyprinodon variegatus	0.99	0.70	0.7437	0.432	77	4

The correlation coefficient (R²) of these models ranges from 0.7437 to 0.9893, and the slope ranges from 0.70 to 1.04. The distribution of model parameters showed that 91.3% of all intercepts ranged from −0.5 to 0.5. The model pair of Pimephales promelas- Cyprinodon variegatus had the smallest slope = 0.7 and the smallest R² = 0.7437 used here.

To further evaluate the consistency and predictive ability of the selected ICE models, the MSE cutoffs linked to cross- validation success rates of 70% and 60% were identified and were assumed to have high reliability and moderate reliability, respectively (Fan et al., 2021). All MSE values were<0.95, ranged from 0.044 to 0.575. All models had a cross- validation success rates > 75%, with 74% (17 models) of the models having a cross- validation success rates > 80%, indicating that all of models with the high reliability (Fan et al., 2019; Wang et al., 2020).

All of the models were selected to evaluate the influence of taxonomic distance on the model reliability (Fig. 3). The taxonomic distance of all the models ranged from 1 (shared genus) to 6 (shared kingdom), with 4 being the most represented at 47.8% of all models. Regardless of taxonomic distance, the R² and the cross- validation success rates were>0.7 and 75%, respectively, showing robust models exist with higher taxonomic distance when the MSE (≤0.95), R² (≥0.6) and slope (≥0.6) criteria are met (Willming et al., 2016) and that taxonomic distance alone should not be used to qualify or disqualify a model for use.

Fig. 3.

The relationship between taxonomic distance and model reliability.

3.2.2. Predicting the toxicity data of APs

The LC₅₀ or EC₅₀ values of 26 species for 10 APs were predicted by QSAR-ICE models, as shown in Table 3.

Table 3.

The LC₅₀ or EC₅₀ values of 26 species for 10 APs were predicted by QSAR-ICE models (mg/L).

Species	2-MP	4-EP	2-PPP	2-BP	4-PTP	4-HXP	4-HTP	4-OP	4-NP	4-DP
Daphnia magna	9.4	19.6	13.4	8.2	1.4	0.9	0.2	0.2	0.2	0.2
Thamnocephalus platyurus	7.1	14.1	9.9	6.3	1.2	0.8	0.2	0.2	0.2	0.2
Ceriodaphnia dubia	6.5	13.7	9.4	5.7	0.9	0.6	0.1	0.1	0.1	0.1
Daphnia pulex	7.0	17.0	11.6	7.0	1.1	0.7	0.1	0.1	0.2	0.2
Simocephalus serrulatus	9.4	19.9	13.5	8.2	1.4	0.8	0.2	0.2	0.2	0.2
Paratanytarsus parthenogeneticus	29.4	58.9	41.3	26.0	4.9	3.0	0.7	0.7	0.8	0.8
Utterbackia imbecillis	5.5	10.8	7.7	4.9	1.0	0.6	0.1	0.2	0.2	0.2
Lymnaea stagnalis	10.3	21.8	14.8	9.0	1.5	0.9	0.2	0.2	0.2	0.2
Physa gyrina	8.2	17.1	11.8	7.2	1.2	0.8	0.2	0.2	0.2	0.2
Amblema plicata	2.5	4.8	3.4	2.2	0.5	0.3	0.1	0.1	0.1	0.1
Branchinecta lynchi	7.7	15.1	10.7	6.9	1.4	0.9	0.2	0.2	0.2	0.3
Megalonaias nervosa	4.4	8.8	6.2	3.9	0.8	0.5	0.1	0.1	0.1	0.1
Margaritifera falcata	6.8	12.8	9.2	6.0	1.3	0.9	0.2	0.2	0.2	0.3
Pimephales promelas	20.7	15.0	11.1	6.6	2.5	1.4	0.7	0.4	0.3	0.1
Lepomis macrochirus	9.7	7.2	5.4	3.3	1.4	0.8	0.4	0.3	0.2	0.1
Ictalurus punctatus	17.8	13.0	9.7	5.9	2.3	1.3	0.7	0.4	0.3	0.1
Oncorhynchus kisutch	2.8	2.2	1.7	1.1	0.5	0.3	0.2	0.1	0.1	0.0
Cyprinus carpio	14.1	10.0	7.3	4.3	1.5	0.8	0.4	0.2	0.1	0.1
Oncorhynchus clarkii	4.5	3.3	2.5	1.5	0.6	0.3	0.2	0.1	0.1	0.0
Jordanella floridae	10.1	7.5	5.7	3.5	1.5	0.8	0.5	0.3	0.2	0.1
Carassius auratus	19.4	14.1	10.5	6.3	2.4	1.4	0.7	0.4	0.3	0.1
Poecilia reticulata	9.8	7.5	5.8	3.7	1.6	1.0	0.6	0.3	0.2	0.1
Oryzias latipes	29.1	21.6	16.4	10.2	4.2	2.4	1.3	0.8	0.5	0.3
Oncorhynchus mykiss	8.0	5.8	4.4	2.6	1.0	0.6	0.3	0.2	0.1	0.1
Cyprinodon variegatus	10.1	8.1	6.5	4.5	2.3	1.5	1.0	0.6	0.5	0.3
Tetrahymena pyriformis	126.5	80.1	60.1	54.7	4.6	2.0	1.2	0.6	0.3	0.2

3.2.3. Derivation of the acute HC₅ values of APs

The acute SSDs were developed from the acute toxicity data pre dicted from QSAR-ICE models (Table 3) and the HC₅ values were then derived using Origin 9.0 software (Fig. 4). The acute HC₅ of the APs were listed in Table 4.

Fig. 4.

Species sensitivity distributions (SSDs) for APs based on acute toxicity data.

Table 4.

The chronic HC₅ of APs based acute HC₅ and MACR.

APs	Predicted HC5 (μg/L)	Acute HC5 (μg/L)	Chronic HC5 (μg/L)
2-MP	n/a	1362.7a	30.28
4-EP	3437.16	1718.58	38.19
2-PPP	2581.07	1290.54	28.68
2-BP	1538.15	769.08	17.09
4-PTP	530.27	265.14	5.89
4-HXP	337.29	168.65	3.75
4-HTP	72.98	36.49	0.81
4-OP	82.19	41.10	0.91
4-NP	n/a	37.62a	0.77b
4-DP	53.99	27.00	0.60

^aSSD developed from measured acute values.

^bSSD developed from measured chronic values.

3.2.4. Reliability verification of the HC₅

Since enough measured acute toxicity data were available for 2-MP and 4-NP to develop SSDs with measured data for these compounds, we compared the HC₅ derived from the measured acute toxicity data and the QSAR-ICE modeled toxicity data for these compounds. The accuracy of the QSAR-ICE modeling approach was verified by comparing the measured-HC₅ and predicted-HC₅ values, as shown in Fig. 5. The measured-HC₅ and predicted-HC₅ of 2-MP were 1362.7 and 2649.72 μg/L, respectively, the ratio of predicted-HC₅ value to measured-HC₅ value was 1.95. The measured-HC₅ and predicted-HC₅ of 4-NP were 37.62 and 77.8 μg/L, respectively, the ratio of predicted-HC₅ value to measured-HC₅ value was 2.1. And the mean ratio of predicted-HC₅ value to measured-HC₅ value for these two compounds was 2. This ratio is well within the range of interlaboratory variability for acute toxicity tests (Hrovat et al., 2009; Fairbrother 2008; Raimondo et al., 2010) and demonstrate high accuracy of the QSAR-ICE models for these two chemicals. For the remaining APs, the predicted HC₅s values were adjusted by a safety factor of two to obtain uncertainty-adjusted HC₅s values to calculate RQs (Table 4).

Fig. 5.

Comparison of SSDs for measured acute and QSAR-ICE-based predicted species toxicity values.

3.3. Derivation of the chronic HC₅ values of APs

Chronic toxicity data for alkylphenols were rarely sufficient to derive chronic HC₅, and neither QSAR derived from TEST nor ICE models predict chronic toxicity. In this study, the available acute and chronic toxicity data (in Table S2–S13) were used to calculate the mean acute and chronic ratios (MACR), and the acute HC₅ was divided by the ACR to obtain the corresponding chronic HC₅. The MACR was calculated to be 45, which was close to the acute HC₅ to chronic HC₅ ratio of 4-NP (37.62/0.77 = 49), indicating that the mean ACR obtained by this method was reliable for APs. The chronic HC₅ of APs were shown in Table 4. Predicted HC₅s were divided by a correction factor of 2 to obtain the Acute HC₅, which was divided by a mean ACR of 45 to obtain the chronic HC₅. Since measure values were used for acute SSDs for 2-MP and 4-MP, predicted values were not applied for these two compounds.

Both acute and chronic HC₅ values of APs (Table 4) for aquatic organisms decreased from 2-MP to 4-DP (acute HC₅ from 1362.7 to 27 μg/L, and chronic HC₅ from 29.44 to 0.60 μg/L, Table 4), indicating that the toxicity of APs to aquatic organisms increases with the addition of alkyl carbon chain lengths. The HC₅ values demonstrated nearly linear decreasing trend from 2-MP to 4-HTP, while the decreasing trend from 4-HTP to 4-DP became shallower. For example, the HC₅ of 2-MP was 37.3 times higher than that of 4-HTP, while the HC₅ of 4-HTP was only 1.35 times higher than that of 4-DP.

Several countries or agencies, such as the United States (US EPA, 2005), the European Union (ECB, 2002), Canada (CCME, 2002), the World Health Organization (Bontje et al., 2004), and Japan (Ministry of the Environment Japan, 2010), have established the water quality criteria of NP for the protection of aquatic organisms. The chronic HC₅ value was 0.77 μg/L and the final PNEC value was 0.38 μg/L of nonylphenol in this study. The PNEC was approximately 18 times lower than the criteria continuous concentration value of 6.59 μg/L derived by US EPA, and slightly higher than the water quality criteria value of 0.33 μg/L formulated by EU but in the same order of magnitude (Hong et al., 2022). In addition, some researchers have also studied water quality criteria or risk assessment of NP based different toxicity effect endpoints (Gao et al., 2015; Jin et al., 2014; Lei et al., 2012; Zhang et al., 2017a), which was close to the PNEC value obtained in this study. The Ministry of Ecology and Environment of the People’s Republic of China issued the “Technical Guideline for Deriving Water Quality Criteria for Freshwater Organisms” in 2022 (https://www.mee.gov.cn/xxgk2018/xxgk/xxgk01/202203/t20220314_971352.html). However, it emphasized that the guideline was unsuitable in deriving the water quality criteria of EDCs and highly enriched organics. Therefore, no water quality criteria or standard for NP has been developed in China. Apart from NP, few studies have been conducted on water quality criteria for other APs. This is consistent with the bibliometric results, show that studies on NP are the most numerous, and very few studies on the toxicity of APs other than NP.

3.4. Ecological risk assessment

The study area and locations of sampling sites were shown in Fig. 6 (A). Nine target APs (4-DP not detected) were detected in water samples collected from the Yongding and Beiyun River, demonstrating their ubiquitous presence in these locations (Fig. 6(B)). The concentrations of the individual chemicals were measured at the ng/L level. In the Yongding River, the concentration of APs ranged from ND to 580.8 ng/L, the highest concentration was 4-NP, and the detection rate of 4-OP and 4-NP was 88.9% and 100%, respectively. In the Beiyun River, the concentration of APs ranged from ND to 306.3 ng/L, and 4-NP were the highest, and the detection rates of 2-MP, 2-PPP, 2-t-BP, 4-PTP and 4-NP were 100%, 100%, 100%, 100% and 96.3%, respectively. It could be discerned that the degree of AP pollution in Yongding River was greater than that in Beiyun River, with the levels of 4-NP in both rivers presented at the highest levels. The highest concentrations of 4-NP measured in the current study was similar to that found by Lei et al. (2021).

Fig. 6.

Study area and locations of sampling sites (A), the concentrations of APs in rivers (B), heat plot of the ecotoxicological risk, represented by chronic RQs, of each APs (y-axis) for aquatic organism in the Yongding and Beiyun River sampling sites (x-axis) (C).

The chronic RQs of the 9 APs are represented by a heat plot (Fig. 6 (C)). Of the 9 APs, 4-NP had medium or high risk to aquatic organisms at most river stations (highest RQ = 1.51), although five stations had low or no risk associated with this compound. For 4-NP, high risk was associated with 2.9% of sampling sites and medium risk was determined at 82.9% of sites. The compounds 4-OP and 4-t-BP did not pose a high risk at any station but had medium or low risk at some stations with the highest RQs calculated as 0.22 and 0.04, respectively. Approximately 5.7% of 4-OP detected in the sampling sites had medium risk. Many studies demonstrated frequent detection of 4-t-BP, 4-OP and 4-NP in water and sediments of both freshwater and marine environments (Gu et al., 2021; Lei et al., 2021; Li et al., 2019). The other APs generally posed low or no risk. The RQ values of APs at all sampling sites were not very high (the highest RQ was 1.51), indicating that with the exception of the high risks noted above, the selected APs do not pose a very high risk to aquatic organisms in both the Yongding River and the Beiyun River. Therefore, 4-NP presented the highest risks to aquatic organisms of all the APs detected in these two rivers, consistent with the findings of Lei et al. (Lei et al., 2021).

To assess the potential influence of AP mixtures, the synergic effects of the 9 APs in the two rivers water were predicted by RQ_MIX (Fig. 6(C)). The RQ_MIX in 1 sampling sites exceeded 1, 33 sampling sites exceeded 0.1, all sampling sites with RQ_MIX exceeded 0.01, which suggests the potential additive or synergistic ecological risks of AP mixtures.

The methods used in the present study calculated RQs based on the best available data to ensure a conservative approach and adjust for uncertainty around limited datasets. Because the science of SSDs is a blend of statistical theory, ecotoxicological testing, research reliability, and biodiversity, the utility of SSDs has been well reviewed, and they are considered advanced assessment tools in ERA and other disciplines (Belanger and Carr, 2020). Furthermore, the HC₅ is considered a practical option that takes advantage of summarizing large data sets, places importance on obtaining additional data, and recognizes that the input data are specific to unobserved effects, which implies additional conservatism (Belanger and Carr, 2019; Carr and Belanger, 2019; Sorgog and Kamo, 2019). Although this study used models to predict AP toxicity, which contain inherent uncertainties, they improve the accuracy of risk assessment compared to assessment factors (AF) used in risk assessment when data are insufficient. The use of AFs is common within ERA practice and is an arbitrary adjustment factor intended to account for various types of uncertainty where data are limited. Despite decades of criticism for their arbitrary nature, AFs continue to be applied because the method is simple and convenient. AFs are applied in cases where available toxicity data are insufficient; however, despite often referred to as “uncertainty factors”, their application is confounded by unquantifiable uncertainty. Ideally, the QSAR-ICE-SSD method can reduce the reliance on AFs and is more suitable for deriving PNECs for risk evaluation (Wheeler et al., 2002). The risk of APs to aquatic organisms in the Yongding River and Beiyun River is lower based on SSD approach than AF method, which can reduce the occurrence of overprotection.

The ubiquitous occurrence of APs in aquatic environments has led to an emerging concern for their synergistic effect (Crini et al., 2022; Meijer et al., 2021). The adverse impacts of APs on endocrine- disrupting effects, especially 4-OP and 4-NP had been reported in the previous studies (Hong et al., 2020; Hong et al., 2022), and the potencies were relatively low compared to those of natural and synthetic estrogens (Barber et al., 2022; Li et al., 2019). The potential risks of APs in natural water bodies are complex and should be comprehensively assessed.

3.5. Uncertainty analysis

Model predictions as used here for QSAR-ICE minimizes the use of experimental animals in toxicity experiments. The use of in silico methods to derive water quality criteria of pollutants is the future of water quality criteria development (Bejarano et al., 2017; Dyer et al., 2008; Fan et al., 2021; Fan et al., 2019; Gredelj et al., 2018; Van den Berg et al., 2019). The application of QSAR-ICE models has recently expanded to include toxicity data for species native to China (about 60 native species of amphibians and fish), which can broaden application of QSAR- ICE models to be used to derive HC₅ for China (He et al., 2017).

Uncertainty analysis is important to evaluate and properly communicate results of model predictions. Douziech et al. (2020) determined that the influence of QSAR uncertainty on the HC₅ is small compared to the uncertainty of ICE estimates. Accounting for QSAR and ICE uncertainty influences the interpretation of the HC₅ more than the HC₅₀ due to SSD uncertainty being greatest at the lowest range of the distribution (Douziech et al., 2020). It was also found that statistical uncertainty in HC₅ was significantly correlated with interspecific variation (P < 0.05), and that an increase in interspecific variation was associated with an increase in statistical uncertainty. In addition, an increase in measured ecotoxicity values tended to decrease the uncertainty of HC₅ values (Douziech et al., 2020). It is important to account for the uncertainty in QSAR and ICE models predictions. For example, if the chemical toxicity to surrogate species obtained from structural descriptors are outside the range of data used to build the QSAR or ICE models, the predicted values should be evaluated against model robustness parameters and risk assessors may limit their weight in the assessment or use them qualitatively. QSAR and ICE models should be statistically validated using approaches such as using a base training and external validation datasets or leave-one-out cross-validation to ensure the accuracy and precision of the model. Other considerations for QSAR models include the selection of the most important molecular descriptors, evaluation of sample size requirements and assessment of the impact of taxonomic distance. Currently, both QSAR and ICE models can only predict the acute toxicity and cannot predict chronic and reproductive toxicity. APs contaminants have been found to affect reproductive fitness at concentrations less than those based on survival and growth (Jin et al., 2014), and more thorough evaluations of APs would require the development of additional models, perhaps as those employing machine learning methods (Cordero et al., 2021; Fan et al., 2021; Hou et al., 2020; Liu et al., 2022b; Wu et al., 2021; Zhang et al., 2021; Zhong et al., 2021).

Additionally, there are several distributions available to derive the SSD for ERA (Wang et al., 2015), such as log-normal, log-logistic, triangular, or Burr Type III (Jin et al., 2011; Wu et al., 2014; Wu et al., 2013). The linearized log-normal distribution has become the most adopted approach by most researchers. In addition to these parametric methods, the standard nonparametric bootstrap method and the bootstrap regression method based on the log-logistic regression model have also been adopted by some researchers (Liu et al., 2016a; Wang et al., 2015). There are always uncertainties in risk assessments associated with the stochastic nature of the environment and the systematic errors introduced by the chosen assessment processes and models. The results of this assessment suggest that the choice of toxicity data has a greater impact on HC₅ than the statistical methods used to extract thresholds from the data (Liu et al., 2016c).

4. Conclusions

In silico methods have been proven helpful in the assessment of potential risks of chemicals with incomplete/unknown toxicity data to aquatic communities. In this study, a QSAR-ICE-SSD methodology was used to predict the HC₅ value of APs to aquatic species. The ICE models used in the analyses were robust (R²: 0.70–0.99; p-value < 0.01) with high cross- validation success rates (>75% predicted within 5-fold of measured value). The acute HC₅ predicted with the QSAR-ICE-SSD model was within 2-fold of the HC₅s derived with measured experimental data. The HC₅ values demonstrated nearly linear decreasing trend from 2-MP to 4-HTP, while the decreasing trend from 4-HTP to 4-DP became shallower, and the toxicity of APs to aquatic organisms increases with the addition of alkyl carbon chain lengths.

The ERA of APs revealed that 4-NP had medium or high risk to aquatic organisms at most river stations (highest RQ = 1.51), five stations had low or no risk associated with this compound. High risk was associated with 2.9% of 4-NP detected in the sampling sites with 82.9% of sites having RQs indicative of medium risk. And the target APs posed potential ecotoxicological risks in the Yongding and Beiyun River according to the mixture ERA. APs pollution, particularly 4-NP, should be further investigated to protect the structure and function of aquatic ecosystems and ensure the sustainable use of river water resources.