Causal inference and mechanism for unraveling the removal of four pesticides from lettuce (Lactuca sativa L.) via ultrasonic processing and various immersion solutions

Highlights

• •Reduction of these pesticides mainly affected by the residuals and immersion time.
• •Efficiency of ultrasonic processing mainly depends on pesticides types and treating-time.
• •Molecular docking was used to elucidate the interaction between pesticide lettuce leaves.
• •Machine leaning was aided causal inference for carbamates and pyrethroid reducing.

Abstract

This study explores the reduction of carbamates (CAs) and pyrethroids (PYs) − commonly used pesticides − in lettuce using various immersion solutions and ultrasonic processing. It also examines the role of machine learning and molecular docking in understanding the mechanisms of pesticide reduction. The results revealed that the highest reduction of both CAs and PYs exceeded 80 % on lettuce leaves. In most samples, the reduction increased with the power of ultrasonic processing and processing time. The results of machine learning models (XGBoost and SHAP) showed that during the immersion cleaning of CAs and PYs, as well as during both immersion cleaning and ultrasonic processing of CAs + PYs, the reduction was most influenced by the initial pesticide levels and immersion time. Gas Chromatography-Mass Spectrometry (GC–MS) analysis of lettuce’s wax layer identified 24 compounds, including fatty alcohols, fatty acids, fatty acid esters, and triterpenoids. Despite the absence of active sites, the lipophilic nature of long-chain aliphatic compounds aids in pesticide binding, while triterpenoids form strong hydrogen bonds with pesticides, indicating a robust adsorption on the lettuce surface. This study aims to offer insights into the efficient removal of chemical pesticide residues from fruits and vegetables, addressing critical concerns for food safety and human health.

1. Introduction

Fresh vegetables are a vital source of essential nutrients such as vitamins, phenolics, carotenoids, steroids, flavonoids, anthocyanins, and minerals, contributing significantly to human health [1], [2]. However, the presence of chemical pesticide residues in vegetables has become a growing concern due to the need for high-quality produce and the rising awareness of potential health risks associated with pesticide exposure [3]. Meanwhile, the escalating costs of pesticide control, the emergence of pesticide resistance, and heightened consumer awareness of the risks associated with pesticide residues have amplified the demand for effective removal methods [4]. The increased use of pyrethroids (PYs) and carbamates (CAs), following restrictions on organophosphate and organochlorine pesticides, has exacerbated this issue. The residues of PYs and CAs have been widely detected in various vegetables [5], [6]. In Tanzania, the highest residues of cypermethrin and bendiocarb were 42.4 mg/kg and 15.6 mg/kg in raw vegetables, which were almost 606 and 312 times of maximum residual limit (MRL), respectively [7]. The ingestion of vegetables containing pesticide residues can lead to health problems, including immune system suppression, hormonal disorders, reproductive defects, and cancer [8]. Unintentional pesticide poisonings have also been reported, underlining the urgency of mitigating this risk [9], [10]. Therefore, effective methods to eliminate PYs and CAs from vegetables are essential.

Vegetables are typically consumed raw or undergo various processing methods such as peeling, grinding, cooking, and sterilization [11]. It is known that these processes can result in a reduction of essential nutrients; however, consuming raw vegetables is considered the best way to preserve these bioactive compounds [12], [13]. Unfortunately, conventional water-based washing methods are insufficient for pesticide removal [1], [9] due to the propensity of lipophilic pesticides to adhere to the vegetable surface [11]. To address this challenge, several methods such as electrolytic oxidized water, ozone-water, microbubbles, ozone-microbubble treatment, and X-ray diffraction spectroscopy have been explored for pesticide removal [14], [15], [16], [17]. Some success also has been reported, as exemplified by Parmar, Korat, Shah and Singh [18], who achieved a substantial reduction in pyrethroid concentration by washing okra with a 2 % brine solution. However, the efficiency, causative factors, and molecular mechanisms involved in pesticide removal by different washing methods remain poorly understood.

Lettuce (Lactuca sativa L.), a widely consumed leafy vegetable known for its nutritional value, holds global significance, with China producing the majority of the world's lettuce [19], [20]. In recent times, machine learning (ML) models have been employed to aid in causal inference [21]. In this study, we investigated the concentrations of PYs (Fenpropathrin and λ-Cyhalothrin) and CAs (Fenobucarb and Indoxacarb) in lettuce leaves following washing treatments involving acetic acid, sodium chloride, sodium bicarbonate, and ultrasonic processing, three machine-learning models (K-nearest neighbors Regressor (KNNR), Random Forest (RF) and extreme gradient boosting (XGBoost)) were employed to unravel the removal of these pesticides from lettuce (Lactuca sativa L.) by the different washing methods, molecular docking were further used to analyze the mechanism based on the composition of wax layer of lettuce leaves. The primary objective of this study is to shed light on the efficacy of various methods for cleaning pyrethroids and carbamates from vegetable leaves.

2. Materials and methods

2.1. Materials

Fenobucarb (FNC), Indoxacarb (IDC), fenpropathrin (FNP), and λ-cyhalothrin (LCT) standard products (≥98.5 %) were procured from Tan Mo Quality Inspection Standard Material Center. We obtained chromatographic pure acetonitrile and methanol from Thermo Fisher Technology (China) Co., LTD. Primary Secondary Amine (PSA) and C18 were provided by NanoChrom Technologies (Suzhou) Co., Ltd. All other analytical pure chemicals were sourced from Chengdu Kelon Chemical Co., Ltd. Emulsions of Fenobucarb (FNC), Indoxacarb (IDC), fenpropathrin (FNP), and λ-cyhalothrin (LCT) were acquired from different suppliers: Ganzhou Yutian Chemical Co., LTD, Suzhou Fumeishi Plant Protection Agent Co., LTD, Guangdong Liwei Chemical Co., Ltd., and Zhejiang Wilda Chemical Co., Ltd., respectively. Organic fresh lettuce (Lactuca sativa L.) was purchased from Entuji Supermarket on the Chenglong campus of Sichuan Normal University, Chengdu, Sichuan Province, China.

2.2. Preparation of samples

To prepare the samples, we followed the recommended usage doses of FNC, IDC, FNP, and LCT emulsions, which were 2.5 g/L for FNC, 0.06 g/L for IDC, 0.26 g/L for FNP, and 0.065 g/L for LCT. We created two sets of pesticide mixtures: one at the recommended dose and another at twice the recommended dose, each totaling 1 L. Next, we took 500 g of lettuce and immersed it in the pesticide mixture solution for 1 min. Afterward, we allowed the lettuce to air-dry indoors, and then we collected 10-gram samples. For further analysis, these 10-gram lettuce samples were soaked in 500 mL of aqueous solutions containing sodium chloride (at concentrations of 0.1 %, 0.5 %, 1.0 % w/v), sodium bicarbonate (at concentrations of 0.1 %, 0.5 %, 1.0 % w/v), and acetic acid (at concentrations of 0.1 %, 0.5 %, 1.0 % v/v). We then subjected these samples to ultrasonication, applying two power levels (120w and 240w) for 5, 10, and 15 min. Finally, we analyzed the samples using ultra-high-performance liquid chromatography (HPLC).

2.3. Determination of FNC, IDC, FNP, and LCT

The quantification of FNC, IDC, FNP, and LCT concentrations was carried out based on a previously reported method with certain adaptations [22]. To elaborate, a 5 g sample of SCP was subjected to homogenization and then transferred to a 50 ml centrifuge tube. Following this, 25 ml of acetonitrile was introduced to the tube. The mixture underwent vortexing for 1 min, followed by a 30-minute ultrasonic extraction. Subsequently, 3 g of NaCl was added, and another minute of vortexing was performed. Of the resulting supernatant, 10 ml was collected and transferred to a volumetric flask after stratification between the acetonitrile and aqueous phases. Rotary evaporation was then employed to achieve dryness, after which the residue was dissolved using 2 ml acetonitrile.

The prepared supernatant was combined with an extraction containing 100 mg PSA, 50 mg C18, and 300 mg magnesium sulfate in a centrifuge tube. After shaking for 30 s, the mixture underwent centrifugation. The resultant supernatant was filtered through a 0.22 μm filter membrane. Finally, the concentration of FNC, IDC, FNP, and LCT in the filtrate was ascertained using Ultra-High Performance Liquid Chromatography (UPLC, ACQUITY UPLC H-Class) equipped with an LC-20AT pump (Shimadzu), a CTO-20A column oven (Shimadzu), an UPLC column (ACQUITY BEH C18, 2.1 × 50 mm, 1.7 μm), and a Photodiode Array Detector.

In the UPLC analysis, the column oven temperature was maintained at 40 °C, and the detection wavelength was set at 250 nm. The mobile phases consisted of (A) acetonitrile and (B) 0.1 % formic acid in water. The linear gradient program (time in minutes and percentage of A) was as follows: t = 0–40 %; t = 4–70 %; t = 7–80 %; t = 9–85 %; t = 9.5–40 %; t = 12–40 %. The flow rate was set at 0.3 mL/min, and the injection volume was 10 μL. Throughout the analysis, the detection wavelength remained at 250 nm. The standard curves for the pesticides FNC, IDC, FNP, and LCT were established with high correlation coefficients, as follows: for FNC, y = 200.84x + 4705 (R² = 0.9998); for IDC, y = 2990.9x + 1411.5 (R² = 0.9927); for FNP, y = 965.4x + 220 (R² = 0.9925); and for LCT, y = 1035.7x − 276.11 (R2 = 0.9927). In these equations, 'x' represents the pesticide concentration, and 'y' denotes the corresponding peak area. The analysis demonstrated strong precision, with relative standard deviations being less than 5 %. Furthermore, the spiked recoveries of the pesticides in this method were found to be consistently above 95 %.

2.4. Machine learning analysis

Details of the immersion cleaning methods, cleaning solution concentrations, types of pesticides, initial immersion concentrations, removal rates, and their standard deviations can be found in the Supplementary materials. For machine learning (ML) analysis, we employed Python or R and utilized the caret (version 4.3.2), Random Forest (version 4.3.2) and the XGBoost (version 1.7.1) packages to run the K-nearest neighbors Regressor (KNNR), Random Forest (RF), and XGBoost regression and SHAP explainer models, respectively. The source code for this analysis is available in the Supplementary materials.

2.5. Composition of lettuce wax layer

The compositions of lettuce wax layer measured by the methods reported by Klavins and Klavins [23]. In detail, lettuce (500 g) was placed in a 1.0 L beaker, to which 500 mL of a chloroform, n-hexane, and methanol mixture (in a 3:1:1 ratio) was added. The beaker was then placed in an ultrasonic cleaner for extraction by sonication at 25 °C, operating at 40 kHz and 100 % power, for 1 min. This procedure was independently repeated three times for each sample. The extract was subsequently evaporated under reduced pressure to 4 mL at 35 °C using a rotary evaporator and transferred to a 25 mL test tube. Following this, 15 mL of the concentrated extraction solution was used to further dissolve the wax residue in the rotary evaporation flask. The extracts were combined, dried under a nitrogen stream, and then treated with 0.4 mL of BSTFA reagent. The samples were immediately sealed, made waterproof, and heated in an oven at 70 °C for 40 min to facilitate the derivatization reaction. After completion of the reaction, the BSTFA reagent was dried off under nitrogen, and the sample was allowed to cool to room temperature. Subsequently, 1 mL of n-hexane was added to dissolve the derivatized waxes. To this, 20 μL of n-hexadecane internal standard solution was added, and the mixture was thoroughly vortexed. The sample was then filtered through a 0.22 μm membrane for gas chromatography-mass spectrometry (GC–MS) analysis. The GC–MS analysis was conducted using an Agilent 6890–5975 system with a DB-5MS (30 mm × 25 mm × 25 μm) quartz capillary column. The inlet port temperature was maintained at 240 ℃, and nitrogen served as the carrier gas at a flow rate of 400 mL/min. The oven temperature was programmed to remain at 50 ℃ for 1 min, then increased at 40 ℃/min to 200 ℃ for 1 min, followed by a 3 ℃/min increase to 300 ℃, held for 15 min. The detector temperature was set at 300 ℃. The system operated in a full scan mode, covering a scanning range of m/z 50–800.

2.6. Molecular docking

The molecular docking was processed according to the method reported by Li et al., (2021) [16]. In detail, ChemDraw software were employed to drawn the molecular structures of FNC, IDC, FNP, and LCT and matrices, and saved in MOL format. And Sybyl-X2.0 software was used to process the three dimensional structure conversion and force field optimization of the molecules, then the active sites on the matrix and pesticide molecules were obtained.

2.7. Data processing and statistical analysis

The experiments in this study were conducted in triplicate. Differences were assessed for significance (p < 0.05) using one-way ANOVA with Duncan's multiple test and Student's t-test by using R software (4.3.1).

3. Results

3.1. Removal of CAs residue by immersion cleaning

The reduction in Fenobucarb (FNC) and Indoxacarb (IDC) concentrations on lettuce leaves after immersion cleaning with various concentrations of acetic acid (AA), sodium chloride (SC), or sodium bicarbonate (SB) for different durations is shown in Fig. 1. As expected, all these solutions proved effective in reducing CAs residues. The most substantial reduction, at 83.4 %, was achieved when 1.0 % SC was used for 15 min to clean lettuce leaves treated with 60 mg/L IDC. In contrast, the least reduction, only 13.8 %, was observed when leaves were cleaned with 0.1 % AA for 5 min after treatment with 5000 mg/L FNC. In general, higher concentrations of AA, SC, and SB, along with longer cleaning durations, were more effective in removing FNC and IDC. Reduction in the concentrations of FNC and IDC was inversely related to their initial immersion concentrations. Interestingly, the removal efficiency of different concentrations of AA, SC, and SB at various durations showed no significant differences when the immersion concentration of IDC was 120 mg/L. Furthermore, the reduction of IDC in samples treated with 1.0 % SB and SC for 15 min was significantly higher than that in samples treated with 0.1 % SB and SC for 5 min (p < 0.05).

Fig. 1.

Effects of various methods of immersion cleaning on reduction of Fenobucarb (FNC), Indoxacarb (IDC). AA indicated acetic acid; SC indicted sodium chioride; SB indicated sodium bicarbonate. Different letter indicated significant different (p < 0.05).

3.2. Removal of PYs residue by immersion cleaning

Fig. 2 presents the reduction in Fenpropathrin (FNP) and Cyhalothrin (LCT) concentrations on lettuce leaves after immersion cleaning with various concentrations of acetic acid (AA), sodium chloride (SC), or sodium bicarbonate (SB) for different durations. It's important to note that, except for samples immersed in 130 mg/L LCT, both the concentration of the cleaning solution and immersion time influenced the reduction process. While all these solutions contributed to the reduction of FNP and LCT, their efficiencies varied. When the immersion concentration of FNP was 260 mg/L, the highest reduction (82.9 %) in FNP occurred with 1.0 % AA for 15 min, whereas the least reduction (23.0 %) was observed with 0.1 % SC for 5 min (Fig. 2a–f). In the case of LCT at an immersion concentration of 65 mg/L (Fig. 2g–l), the highest reduction (73.0 %) was achieved with 0.5 % SC for 15 min, while the lowest reduction (6.0 %) was noted with 0.1 % AA for 5 min. Interestingly, AA concentrations and immersion time had no significant impact on LCT reduction. However, samples cleaned with 0.5 % and 1.0 % SB for 15 min showed significantly higher reductions compared to other treatments. Similarly, samples cleaned with 0.5 % and 1.0 % SC for 15 min exhibited significantly higher reductions than those cleaned with 0.5 % SC for 5 min.

Fig. 2.

Effects of various methods of immersion cleaning on reduction of Fenpropathrin (FNP) and Cyhalothrin (LCT). AA indicated acetic acid; SC indicted sodium chioride; SB indicated sodium bicarbonate. Different letter indicated significant different (p < 0.05).

3.3. Removal of CAs and PYs residue by ultrasonic processing

Fig. 3 illustrates the reduction in Fenobucarb (FNC), Fenpropathrin (FNP), Indoxacarb (IDC), and Cyhalothrin (LCT) concentrations on lettuce leaves after ultrasonic processing (UP) at varying power and duration levels. Surprisingly, the processing time, immersion concentrations of FNC, and the power of UP had minimal effects on removal efficiency (Fig. 3a and b). However, it's noteworthy that the reduction in samples immersed in 2500 mg/L FNC and treated under 240 W for 15 min exhibited a significant increase (p < 0.05). For FNP, when the immersion concentration was 520 mg/L, the reduction in samples treated with 240 W for 15 min was significantly higher than that in samples treated with 120 W for 5 or 10 min (Fig. 3d, p < 0.05). Conversely, the lowest efficiency was observed when samples were treated with 120 W for 5 min (Fig. 3f). For samples immersed in 260 mg/L FNP, 60 mg/L IDC, 65 mg/L LCT, and 130 mg/L LCT, most samples exhibited increased reduction with higher UP power and longer processing times (Fig. 3c, e, g, and h).

Fig. 3.

Effects of ultrasonic processing on reduction of Fenobucarb (FNC), Fenpropathrin (FNP), Indoxacarb (IDC), and Cyhalothrin (LCT). Different letter indicated significant different (p < 0.05).

3.4. Pesticide reduction prediction with machine learning

We compared the performance of KNNR, RF, LGBoost and XGBoost models under two scenarios: removal of CAs and PYs residue by immersion cleaning and ultrasonic processing (Table S1), suggesting that the prediction of XGBoost model was more efficient than other models. The resulting trained model, denoted as xgbReg, was subsequently applied to predict the reduction of both CAs and PYs achieved through immersion cleaning and ultrasonic processing. Table 1 presents the model parameters, including RMSE, MSE, MAE, MAPE, accuracy, and R², demonstrating a strong alignment between predictions and experimental reduction data. To quantify the factors affecting reduction, the SHAP explainer was utilized to interpret XGBoost's prediction processes (Fig. 4). It generated SHAP values that elucidate the extent to which various factors influence pesticide reduction. These factors encompassed immersion cleaning methods (Methods), concentrations of immersion solution (Con), power of ultrasonic processing (UP), immersion time (Time), initial immersion concentration of pesticides (IN), pesticides concentrations in the control group (CK), and pesticide type (Pesticide). The analysis revealed distinct vulnerabilities in reduction factors. For immersion cleaning of CAs, the presence of initial residuals had the most significant impact (mean |SV| = 6.7), while the choice of cleaning method exhibited the least influence (mean |SV| = 1.2) (Fig. 4a). Similarly, among immersion cleaning of PYs, the choice of method was the least influential (mean |SV| = 1.9), whereas immersion time had the greatest impact (mean |SV| = 6.1) (Fig. 4b). In cases involving both immersion cleaning and ultrasonic processing of CAs + PYs, reduction was most susceptible to the presence of initial residuals and immersion time (mean |SV| = 7.3 and 5.4, respectively). Conversely, the choice of method and power of ultrasonic processing had the least influence (mean |SV| = 0.2 and 1.9, respectively) (Fig. 4c and d). The effectiveness order of immersion cleaning in the CAs group was CK > IN > Time > Con > Methods. In the PYs group, it was Time > IN > CK > Con > Methods (Fig. 5). For the CAs + PYs group, it was CK > Time > IN > Con > Methods > Pesticide. In the ultrasonic processing group, the effectiveness order was Time > Pesticide > IN > CK > Con (Fig. 5). SHAP feature relevance of the model is visualized in Fig. 6. Features that positively influence the forecast are highlighted in red, while those with a negative impact are shown in blue. Notably, immersion time (Time) and the concentration of the immersion solution (Con) exhibit a positive correlation with the output value in these four groups. A more in-depth SHAP analysis underscores the significant role of these features in distinguishing between positive and negative regions.

Table 1.

Parameters of the XGBoost model training and testing dataset of cleaning the pesticides by immersion and UP.

Cleaning method	pesticides	RMSE	MSE	MAE	MAPE	accuracy	R2
Immersion cleaning	CAs	5.9298	35.1620	4.2310	0.1118	0.8882	0.8331
Immersion cleaning	PYs	6.1543	37.8756	4.5506	0.1391	0.8609	0.8364
Immersion cleaning	CAs + PYs	1.3619	1.8547	0.6610	0.1057	0.9843	0.9923
ultrasonic processing	CAs + PYs	1.7583	3.0915	1.3745	0.1644	0.9356	0.9910

Fig. 4.

Correlations between experimentally measured and XGBoost-predicted reduction of CAs (a), PYs (b) and CAs + PYs (c) by immersion cleaning (IC), and CAs + PYs (d) by ultrasonic processing (UP).

Fig. 5.

Mean absolute SHAP values show the immersion cleaning methods (methods), concentrations of immersion solution (con), power of ultrasonic processing (UP), immersion time (time), initial immersion concentration of pesticides (IN), pesticides concentrations in the control group (ck), and pesticide type (pesticide) contribute to the XGBoost’s prediction of the reduction of CAs (a), PYs (b) and CAs + PYs (c) by immersion cleaning (IC), and CAs + PYs (d) by ultrasonic processing (UP).

Fig. 6.

SHAP values displaying the effect of the immersion cleaning methods (methods), concentrations of immersion solution (con), power of ultrasonic processing (UP), immersion time (time), initial immersion concentration of pesticides (IN), pesticides concentrations in the control group (ck), and pesticide type (pesticide) contribute to the XGBoost’s prediction of the reduction of CAs (a), PYs (b) and CAs + PYs (c) by immersion cleaning (IC), and CAs + PYs (d) by ultrasonic processing (UP).

3.5. Interaction between pesticides and wax layer

The compositions of waxy layer of lettuce leaves were investigated by GC–MS (Table S2 and Fig. S1). And 24 compounds were identified, including 11 fatty alcohols, 5 fatty acids, 3 fatty acid esters, and 5 triterpenoids. The molecules of these compounds underwent Tripos force field optimization to achieve their best conformation. Subsequent superimposition using SYBYL software and exploration through the GASP pharmacophore identification technique (Table S3 and S4) revealed active sites common to triterpenoids. These sites, as illustrated in the Fig. 7a, can act as both acceptors for pesticide molecule binding (donor site DS) and hydrogen bond donors (acceptor site AA). When pesticide molecules adsorb onto triterpenoid surfaces, the triterpenoids function as either proton donors or acceptors, involving the hydrogen atoms in their hydroxyl groups in hydrogen bond formation.

Fig. 7.

Molecular properties of pesticides (a: acceptors for pesticide molecule binding (donor site DS) and hydrogen bond donors (acceptor site AA); b and c: active sites of λ-cyhalothrin (LCT); d and e: lipophilic site of Fenobucarb (FNC); f and g: active site of Indoxacarb (IDC); h and i: lipophilic site of fenpropathrin (FNP)).

Four pesticides were analyzed for their surface properties. Cyfluthrin's oxygen atoms form hydrogen bonds efficiently (Fig. 7b, red area), with its chlorine and hydrogen atoms creating potential lipophilic sites (Fig. 7c, yellow area). The nitrogen atom of sec-butylcarbamate also forms hydrogen bonds readily (Fig. 7d, red region), and its hydrogen atoms contribute to its lipophilic sites (Fig. 7e, yellow region). Indoxacarb's oxygen atom, possessing strong electronegativity, easily forms hydrogen bonds (Fig. 7f, red region), while its chlorine atoms create potential lipophilic regions (Fig. 7g, yellow region). Lastly, the nitrogen atom of alpha-cypermethrin is electronegative and forms hydrogen bonds efficiently (Fig. 7h, red region), with its chlorine atoms contributing to lipophilic sites (Fig. 7i, yellow area). Hydrophilic or hydroxyl-rich matrices, such as those in lettuce waxes containing hydroxyl or oxygen atoms, demonstrate a strong ability to form hydrogen bonds. This suggests a high likelihood of pesticides forming hydrogen bonds with these groups, indicating a strong adsorption capacity on the lettuce surface. The LCA compounds did not exhibit active sites after overlapping, likely due to their simpler molecular structures hindering the formation of binding sites for proton donors or acceptors. Despite lacking active sites, the lipophilic nature of LCA compounds enables them to bind to pesticide molecules, albeit with a lower binding capacity compared to that of triterpenoids.

4. Discussion

Residues of carbamate (CA) and pyrethroid (PY) pesticides in fruits and vegetables have garnered increasing attention, and addressing the elimination or reduction of these pesticide residues has become a pressing concern [7], [11]. It is widely recognized that FNC, IDC, FNP, and LCT are hydrophobic compounds, which, when absorbed onto the waxy surfaces of fruits and vegetables, become challenging to remove [4], [16]. The efficiency of pesticide residue removal hinges on the physical and chemical properties of the pesticides, as well as the washing method employed, encompassing factors such as the washing solution, temperature, and duration [11], [24]. In this study, we investigated the removal efficiency of the most commonly used pesticides, including CAs and PYs, following cleaning with acetic acid (AA), sodium chloride (SC), sodium bicarbonate (SB), or ultrasonic processing (UP). It's worth noting that Zhang et al. (2007) reported a significant reduction in pesticide residues, including chlorpyrifos, p, p-DDT, cypermethrin, and chlorothalonil, through washing with acetic acid solutions (at 10.0 % concentration for 20 min). In our study, all of these methods effectively reduced the concentrations of CAs and PYs in lettuce; however, the removal efficiencies varied. And molecular docking was employed to analysis the reason for difficult removal. Since the molecular structure of LCA is relatively simple, it is difficult to form binding sites that can be used as donors or receptors. They contains a large number of highly stable C-C single bonds and C-H single bonds. The electronegativity difference between hydrogen atoms and carbon atoms is very small, so C-H bond is very non-polar and belongs to a strong non-polar chemical bond[16]. The adsorption between them and pesticide molecules is mainly a weak interaction such as van der Waals force[16]. Therefore, the binding ability of LCA is lower than that of triterpenoids and pesticides. When pesticide molecules adsorb onto the surface of the triterpenes, the triterpenes can function either as proton donors or receptors. This might be the participation of oxygen and hydrogen atoms in the hydroxyl groups in forming hydrogen bonds[25], [26].

Machine learning have been widely applied to analysis the causal inference [27]. Some popular models including KNNR, RF, LGBoost and XGBoost[28] were used in this study. Extreme Gradient Boosting (XGBoost), a robust implementation of gradient boosting decision trees [29], is the most efficient model to assess the factors influencing pesticide removal efficiency. In this study, it highlighted that pesticide concentrations, pesticide residues, and the duration of immersion cleaning were the most crucial factors affecting pesticide removal efficiency. It reported that the removal efficiency can vary depending on the type of pesticides and vegetables when washing pesticide residues[30]. Our results underscored that immersion time and solution concentrations were more influential factors in removal efficiency than the immersion cleaning method itself [11], [17]. The attachment mode of pesticide residues to the matrix involves both the binding of pesticide formulation solution residue onto the matrix surface due to natural air drying and the binding of pesticide molecules on the matrix surface [4], [31]. Despite variations in the solubility order of these pesticides (FNC > IDC > FNP > LCT), the type of pesticide was found to be of lesser importance in the reduction process. However, we observed that LCT was the most challenging pesticide to remove, primarily due to differences in immersion concentration and initial pesticide residues.

It reported that ultrasonic processing (UP) significantly reduced the residues of pyrazophos, chlorothalonil, and carbendazim on pakchoi leaves when compared to traditional water immersion [32]. Ultrasound proved to be a more efficient method for removing pesticide residues. Our findings also align with this, demonstrating the effectiveness of UP in decreasing the concentration of FNC, IDC, FNP, and LCT. The cavitation phenomenon generated by ultrasound is likely the primary reason for the efficient removal of pesticide residues during ultrasonic cleaning. During the collapse of cavitation bubbles, pesticide molecules break down along with water molecules [30], [33]. Additionally, some pesticide molecules may undergo oxidation due to the presence of highly reactive OH radicals produced during the breakdown of water molecules [34]. Given that both the duration of UP and its power can influence the collapse of cavitation bubbles and the generation of reactive OH radicals, these factors affect the reduction of these pesticides. Moreover, variations in chemical bonds and residual concentrations within pesticides can lead to differences in reduction efficiency. Our results are consistent with these theoretical predictions.

5. Conclusion

Residual pesticides in fruits and vegetables pose a threat to human health, with carbamates (CAs) and pyrethroids (PYs) being among the most widely used pesticides worldwide. Our study demonstrated that different immersion solutions and ultrasonic processing (UP) can effectively reduce the concentrations of CAs, including FNC and IDC, and PYs, including FNP and LCT, in lettuce. The lipophilic nature of long-chain aliphatic compounds (LCA) compounds enables them to bind to pesticide molecules, pesticides forming hydrogen bonds with triterpenoids which is a strong adsorption capacity on the lettuce surface. We employed the ML model to predict and analyze the factors influencing the removal efficiency of these pesticides, and XGBoost was the best predictor. The variations in reduction rates can be attributed to the diverse chemical properties of these pesticides, with the immersion methods being of lesser importance. However, the order of effectiveness in pesticide removal differs from that of immersion cleaning, with processing time and pesticide types emerging as the two most critical factors affecting efficiency. These findings provide insights into reducing CAs and PYs residues primarily through ultrasonic processing and various immersion solutions.

Funding

This work was financially supported by the Sichuan Science and Technology Program (2022NSFSC0113), National Natural Science Foundation of China (31801644), Sichuan Normal University (SYJS2022016), and Innovation Training for University Students in Sichuan Province (202310636050).

CRediT authorship contribution statement

Sijia Zhao: Writing – original draft, Methodology, Investigation, Formal analysis. Xinyi Huang: Visualization, Validation, Software, Investigation. Guanyu Chen: Visualization, Software, Formal analysis. Haixiong Qin: Methodology, Investigation. Bowen Xu: Methodology, Investigation. Yu Luo: Methodology, Investigation. Ying Liao: Methodology, Investigation. Shufang Wang: Methodology, Investigation, Formal analysis. Shen Yan: Software, Methodology, Investigation. Jiayuan Zhao: Writing – review & editing, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: