A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins

By constructing non-target biological protein plasmids and surface plasmon resonance technology, combined with the XGBoost model, the problem of predicting non-target biological toxicity in the early stage of pesticide research and development was solved, efficient and accurate toxicity assessment was achieved, resource consumption and cost were reduced, and more Reliable ecological security support.

CN118777590BActive Publication Date: 2024-11-26INST OF PLANT PROTECTION CHINESE ACAD OF AGRI SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411237372.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-05
Publication Date
2024-11-26
Estimated Expiration
2044-09-05

AI Technical Summary

Technical Problem

The existing technology is difficult to effectively predict the toxicity of pesticides to non-target organisms in the early stages of pesticide research and development, resulting in resource waste and ecological risks. The prediction results of traditional methods are inaccurate and costly.

Method used

By constructing non-target biological protein plasmids, combining surface plasmon resonance technology (SPR) and protein chips, the interaction parameters of pesticides and non-target biological proteins were detected using Biacore T200 biomolecular interaction analyzer, and a regression model based on XGBoost was constructed for toxicity prediction. .

Benefits of technology

It realizes efficient and accurate assessment of pesticide toxicity, reduces experimental costs and resource consumption, provides more reliable ecological security data support, can accurately predict LD50 values, and is suitable for different types and dosage forms of pesticides.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118777590B_ABST
    Figure CN118777590B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting the toxicity of pesticides to non-target organisms by using non-target biological proteins, which belongs to a method for predicting the toxicity of pesticides by using non-target biological proteins. The method realizes efficient evaluation of the toxicity of pesticides by inducing expression and purification of non-target biological proteins, combining advanced technologies such as surface plasmon resonance technology (SPR) and protein chips. The present invention successfully replaces traditional non-target biological toxicology experiments by innovatively combining surface plasmon resonance technology, protein chip technology and artificial intelligence algorithms, and realizes a more animal-friendly and ethical toxicity prediction method. With efficient sample utilization, accurate toxicity assessment and simplified experimental procedures, the present invention significantly improves the efficiency and accuracy of toxicity detection, reduces the consumption of experimental resources and the burden on operators. At the same time, it reduces experimental costs, improves high-throughput detection capabilities, and makes the overall process of pesticide toxicity assessment more convenient and economical.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to a method for predicting the toxicity of pesticides by using non-target biological proteins, and particularly relates to a method for predicting the toxicity of pesticides to non-target organisms by using non-target biological proteins. Background Art

[0002] With the continuous development of agricultural production, the use of pesticides has become increasingly widespread, especially the rapid popularization of aerial plant protection drone spraying technology in recent years, which will inevitably bring about the risk of drift to non-target organisms and environmental toxicological hazards. Taking bees as an example, the large-scale use of neonicotinoid insecticides has seriously affected the normal pollination and survival of bees, leading to the occurrence of colony collapse disorder. In order to protect bees and other pollinators, the European Union banned the outdoor use of several major neonicotinoid insecticides, including thiamethoxam, in 2013 and 2018 respectively. Due to the widespread use of neonicotinoid insecticides, the bee population has dropped sharply and ecological problems have become increasingly serious.

[0003] The research and development costs of pesticides are extremely high. In the traditional development process, a large number of compounds are usually synthesized in the later stages before toxicity bioassays for bees can be carried out. However, these animal toxicity assays usually require dozens to hundreds of grams of samples to be tested, especially for some low-toxic compounds, which is usually difficult to achieve in the early stages of compound discovery. This high demand greatly increases experimental costs and resource consumption.

[0004] Although some compounds meet the requirements for pesticide registration in indoor bioassays, they only show toxicity to bees and ecology in field experiments or even after they are widely used in the market. This delayed discovery eventually led to these compounds being banned in some countries or organizations and forced to withdraw from the market, resulting in a huge waste of time and R&D funds.

[0005] In addition, many pesticides not only bind to target proteins, but also interact with proteins of other non-target organisms. For example, neonicotinoid insecticides bind to the olfactory receptor proteins of bees. This interaction makes it difficult for the traditional acetylcholinesterase-based evaluation system and QSAR model-based toxicity prediction methods to effectively deal with the ecotoxicological problems caused by pesticides to non-target organisms.

[0006] Therefore, it is particularly important to predict toxicity in the early stages of pesticide research and development. Currently, there are two main methods for pesticide toxicity testing:

[0007] (1) Quantitative structure-activity relationship (QSAR) model combined with machine learning

[0008] Kwon S, Bae H, Jo J, et al. Comprehensive ensemble in QSAR prediction for drug discovery. BMC bioinformatics, 2019, 20: 1-12. discloses that QSAR models use the relationship between chemical structure and biological activity, combined with machine learning algorithms (such as random forests, support vector machines, and neural networks) to predict the toxicity of chemical substances. These models usually rely on the molecular descriptors of the compounds and predict the toxicity of new compounds through training data sets. The shortcomings are:

[0009] First, there is a large difference between the predicted results and the actual experimental data, which reduces the credibility of the model; second, only a classification model is constructed without regression analysis, which can only provide toxicity levels (high, medium, and low) and cannot accurately predict the toxicity LD 50 Numerical values, limited prediction range;

[0010] (2) Spectroscopy combined with machine learning

[0011] Spectroscopy combined with machine learning is used to analyze the spectral characteristics of chemical substances to predict their toxicity. After the spectral data is processed, it is modeled through a machine learning algorithm to predict the toxicity of new samples. Its shortcomings are: first, spectral data is easily affected by environmental factors, resulting in unstable experimental results and large errors; second, the model is easily affected by changes in experimental conditions, resulting in unreliable prediction results and low accuracy;

[0012] Based on the above problems, it is of vital importance to evaluate and predict the potential environmental toxicological risks of pesticides to non-target organisms in the early stages of pesticide development. Through early prediction, compounds that may pose risks to non-target organisms can be identified and eliminated in a timely manner. This can not only significantly reduce the high cost of pesticide research and development, but also effectively reduce unnecessary ecological risks and promote safer and sustainable pesticide creation and development strategies. Through this process, researchers can more efficiently screen out pesticides that are friendly to the environment and ecosystem, thereby improving research and development efficiency and avoiding waste of resources due to later toxicity issues. Summary of the invention

[0013] The present invention aims to overcome the deficiencies of the prior art and provides a method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins.

[0014] A method of the present invention for predicting the toxicity of pesticides to non-target organisms by using non-target biological proteins comprises the following steps:

[0015] (1) Plasmid transformation of protein target gene

[0016] Constructing a non-target biological protein plasmid and performing plasmid transformation on the target gene of the non-target biological protein plasmid;

[0017] (2) Induction of protein target genes

[0018] Inducing the target gene treated in step (1) and collecting the supernatant;

[0019] (3) Protein purification

[0020] Purify the supernatant collected in step (2) to obtain a final protein concentration of no less than 1 mg / ml;

[0021] (4) Protein chip preparation

[0022] a. Prepare 1×PBS-P+protein coupling running buffer;

[0023] b. Dilute the protein purified in step (3) to a final concentration of 80 μg / mL using a sodium acetate solution at pH 1.0-4.0;

[0024] c. Ultrasonic exhaust of pH 1.0-4.0 protein sodium acetate solution at -4°C for 1-3 minutes;

[0025] d. Couple the protein sodium acetate treated in step c to the chip in a running buffer with a coupling amount of 8000-12000 RU, and fix it on the chip surface for later use;

[0026] Wherein, the coupling amount is according to the formula: get , Among them, Rmax is the maximum binding capacity of the chip surface, which is substituted into 100RU in protein test; Sm is the stoichiometric ratio, which is selected as 1 when unknown; RL is the ligand coupling level, analyte MW is the molecular weight of the bee olfactory protein, and ligand MW is the molecular weight of the pesticide compound to be tested;

[0027] (5) Pesticide configuration

[0028] The prepared pesticide sample stock solution was diluted downward into 11 concentration gradients, namely 3000 μmol / L, 1000 μmol / L, 333.3 μmol / L, 111.1 μmol / L, 37 μmol / L, 12.3 μmol / L, 4.1 μmol / L, 1.37 μmol / L, 0.46 μmol / L, 0.15 μmol / L, 0.05 μmol / L; set aside;

[0029] (6) Detection of animal protein and pesticide dynamics

[0030] The temperature of the Biacore T200 biomolecular interaction analyzer was set to 21°C, 28°C, and 35°C, respectively. The chip packaged in step (4) was placed in the chip placement slot. The instrument parameters were set as follows: the Hertz number was set to 10 Hz, the loading speed was 10 μl / min, the binding time was 120 s, the dissociation time was 130 s, and the chip regeneration condition was to regenerate the chip with a pH 1.8, 10 mmol / L glycine hydrochloride solution at a flow rate of 30 μl / min for 30 s;

[0031] (7) Data processing

[0032] a. Open the Biacore T200 Evaluation Software (data analysis software) for the Biacore T200 biomolecular interaction analyzer in step (6) and click Open in File to find the file.

[0033] b. Open the software front-end interface and click the "Data Processing" button;

[0034] c. The software is based on the Sikuli library. It uses image recognition to find and control the buttons on the interface. Click SolventCorrection to perform solvent correction analysis and identify solvent correction.

[0035] d. In the Kinetics / Affinity-Select Curves interface, under Select evaluation mode, identify and select Single mode; in the Sample drop-down list in the Curve column, select the sample to be analyzed; identify and select the pesticide samples with different concentrations configured in step (5) for testing. If there are abnormal value results, uncheck the box in front of this concentration to delete this concentration;

[0036] e. Select kinetics for fitting, select 1:1 Binding for Model, select fit local for Rmax, and then click Fit to fit the data; after the kinetics fitting is completed, perform thermodynamic fitting to obtain the binding constant K between the test compound and the protein a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G ;

[0037] Plot of steady-state binding level versus analyte concentration Calculate the equilibrium dissociation constant K for a 1:1 interaction D The equations include the following:

[0038] ;

[0039] Where: K D is the equilibrium dissociation constant; Rmax is the surface analyte binding capacity; offset is the bulk refractive index contribution in the sample;

[0040] ;

[0041] The component of the mass transfer constant that is independent of flow rate is calculated using the following equation:

[0042] ;

[0043] Where D is the diffusion coefficient of the analyte, f is the volume flow rate of the solution through the flow cell, and h, w, and l are the dimensions of the flow cell;

[0044] ;

[0045] where G is the conversion factor from surface concentration to RU, and for proteins on sensor chip CM5, G is approximately 10 9 ;

[0046] ;

[0047] For equilibrium thermodynamics, the van't Hoff equation states:

[0048] ;

[0049] where Δ G is the standard Gibbs free energy change, R is the universal gas constant, T is the absolute temperature, and KD is the equilibrium dissociation constant;

[0050] (8) Model construction

[0051] a. Introduce the data obtained from the non-target biological toxicity database, the literature, and the toxicity test data of some similar compounds on bees into the data set;

[0052] b. Substitute the binding constant K obtained in step e of (7) a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G As a data set, the data set is divided into a training set and a test set at a ratio of 8:2;

[0053] c. Globally optimize hyperparameters based on the Bayesian optimization algorithm and build a regression model for bee toxicity prediction based on XGBoost;

[0054] The described model architecture:

[0055] Define the following feature variables:

[0056] ;

[0057] Target variable:

[0058] ;

[0059] The squared error loss function is used for regression tasks, defined as , where y is the actual value, is the predicted value. XGBoost uses the gradient boosting framework to improve the model performance by gradually adding weak learners. At the tth iteration, the model is as follows:

[0060] ;

[0061] in is the t-th tree, and the goal is to minimize the loss function:

[0062] ;

[0063] Bayesian optimization is used to select the best hyperparameters for the XGBoost model. Bayesian optimization finds the optimal hyperparameters by maximizing the expected improvement:

[0064] ;

[0065] in is the prediction of the current model, is the current optimal solution;

[0066] The optimal hyperparameters were found through Bayesian optimization , and used to train the XGBoost model:

[0067] ;

[0068] Where T is the total number of trees, It is the tth tree. When predicting new data, the feature variable X is input into the trained model to obtain the predicted value:

[0069] ;

[0070] e. Substitute the test set into the model for verification and optimization;

[0071] f. When predicting, the binding constant K of the ligand pesticide to be tested detected by the Biacore T200 biomolecular interaction analyzer a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time tc , free energy Δ G By inputting the numerical value into the model, the LD of the new compound to non-target organisms can be obtained. 50 value, obtain the toxicity level and complete the prediction of its toxicity.

[0072] As a further improvement of the present invention, the plasmid transformation of the protein target gene in step (1) is achieved by the following steps:

[0073] (1) Construction of plasmids for non-target biological proteins;

[0074] (2) Add 1 µL of plasmid to 50-200 µL of BL21 competent cells, swirl gently to mix, and place on ice for 10-30 minutes;

[0075] (3) Heat shock in a 42°C water bath for 70-90 seconds, then quickly transfer the tube to an ice bath for 1-3 minutes.

[0076] (4) Add 300 µL LB medium, mix well, and culture at 37 °C, 200-500 rpm with shaking for 1-2 h;

[0077] (5) In a clean bench, evenly spread 100-300 μL of bacterial solution on an LB plate containing 100-300 μg / mL kanamycin using a disposable coating rod and place at room temperature until the liquid is absorbed;

[0078] (6) Invert the plate and transfer it to a 37 °C biochemical incubator for 24 h in the dark.

[0079] As a further improvement of the present invention, the induction of the protein target gene in step (2) is achieved by the following steps:

[0080] (1) Pick two single colonies and inoculate them into 5 mL LB liquid medium containing 0.1-0.3% kanamycin at 37°C and 220 rpm / min for about 6-8 h.

[0081] (2) Inoculate 200-400 mL LB liquid culture medium at a ratio of 1%, add 0.1% kanamycin, incubate at 37°C and 220 rpm / min, and induce for 1-2 h;

[0082] (3) Then, add 150 μL IPTG to a final concentration of 0.5 mmol / L, incubate at 16 °C, 150 rpm / min, and induce for 12-16 h;

[0083] (4) After induction, centrifuge at 8,000 × g for 10 min (4 °C), discard the supernatant, collect the sediment, and resuspend the collected bacterial pellet in 15-30 mL binding buffer;

[0084] (5) Ultrasonic disruption of cells for 8-12 min;

[0085] (6) Centrifuge at 8,000 rpm at 4°C for 5-15 min and collect the unpurified protein supernatant.

[0086] As a further improvement of the present invention, the protein purification in step (3) is achieved by the following steps:

[0087] (1) Preparation of gravity purification column: Add 2 1.5 mL affinity columns containing nickel to a 10 mL purification column, add 20 mL ultrapure water to remove the original ethanol in the column material, and add 10 mL of 50 mmol / L binding buffer to equilibrate the column material;

[0088] (2) Slowly load the collected protein supernatant and collect the flow-through sample;

[0089] (3) The flow-through sample is eluted again, and the elution is repeated three times to obtain a preliminarily purified protein solution;

[0090] (4) Desalt the purified protein solution through a desalting column connected to the ÄKTA protein rapid purification system to obtain a protein solution with a final protein concentration of no less than 1 mg / ml.

[0091] As a further improvement of the present invention, in step (8) c, the optimized hyperparameters include: n_estimators, max_depth, learning_rate, min_child_weight, subsample, colsample_bytree, and gamma.

[0092] As a further improvement of the present invention, the non-target biological protein is any one of olfactory receptor protein, human serum albumin, bovine serum albumin, hemoglobin, globulin, hemoglobin, myosin, collagen, enzyme protein, bee protein, zebrafish protein, human liver cells, and zebrafish cells.

[0093] The method of the present invention for predicting the toxicity of pesticides to non-target organisms by using non-target biological proteins has the following advantages:

[0094] 1. Replace traditional non-target biological toxicology experiments, animal-friendly and ethical

[0095] This application achieves efficient evaluation of pesticide toxicity by inducing expression and purification of non-target biological proteins, combined with advanced technologies such as surface plasmon resonance (SPR) and protein chips. This process does not involve the use of any living organisms, fundamentally eliminating harm to organisms, and also avoiding animal ethical issues caused by experiments;

[0096] 2. Accurately predict the toxicity of pesticides to non-target organisms

[0097] The present invention uses the interaction parameters between pesticides and key proteins of non-target organisms as input variables, and generates a prediction model that can accurately evaluate the toxicity of pesticides to non-target organisms through training and optimization of the XGBoost model. This method significantly improves the accuracy of pesticide toxicity assessment and provides more reliable data support for the environmental safety of non-target organisms. In addition, the model also has a strong generalization ability and can be applied to the prediction of pesticide toxicity of different types and dosage forms, helping researchers and regulatory agencies to more effectively evaluate and manage the ecological risks of pesticides;

[0098] 3. Comprehensively consider the impact of two non-target biological proteins

[0099] The present invention further improves the accuracy of pesticide toxicity prediction by comprehensively considering the effects of two non-target biological proteins. Specifically, the present invention not only focuses on the interaction between a single non-target biological protein and pesticides, but also incorporates the interaction parameters of two non-target biological proteins with different functions into the model to construct a more complex and comprehensive XGBoost prediction model. This comprehensive consideration of dual protein parameters can not only greatly avoid the biggest problem of false positives when surface plasmon resonance technology is used to detect the interaction between pesticides and proteins; it can also enable the model to more accurately reflect the potential toxic effects of pesticides in different organisms, thereby providing more comprehensive data support for toxicity assessment;

[0100] 4. Save protein and compound samples

[0101] The present invention utilizes protein chip technology to achieve extremely high sample utilization efficiency in the toxicity prediction process. Only about 100 microliters of protein sample and about 2 milligrams of the compound to be tested are required to complete an accurate toxicity assessment. This technology is particularly critical in the early stages of new pesticide research and development, especially when the amount of compound synthesis is limited. It can significantly save precious sample resources, thereby improving research efficiency, reducing experimental costs, and reducing the number of animals used in experiments, meeting environmental protection and ethical requirements. At the same time, the high-throughput characteristics of protein chip technology enable it to process multiple samples simultaneously, greatly shortening the time for toxicity assessment, and providing faster and more economical technical support for pesticide research and development;

[0102] 5. Advantages of efficient artificial intelligence toxicity prediction

[0103] The present invention proposes an efficient artificial intelligence toxicity prediction method, which uses a deep learning model, especially the XGBoost algorithm, to achieve accurate prediction of pesticide toxicity. With its powerful processing power and high flexibility, the XGBoost algorithm can automatically extract key features from a large amount of experimental data and optimize the model, thereby significantly improving the accuracy and reliability of toxicity prediction. Through this algorithm, the present invention can not only quickly identify compounds that may be highly toxic, but also reduce the impact of human intervention and reduce errors in the toxicity prediction process. The high efficiency of the model is reflected in its ability to process complex data. It can complete the analysis of a large amount of data in a short time and automatically adjust parameters to optimize the prediction results. In addition, when processing data, the XGBoost algorithm can adaptively weigh multiple factors to ensure the generalization ability of the model on different data sets. This enables the toxicity prediction method of the present invention to maintain a high level of accuracy in different application scenarios, providing strong technical support for the creation and safety assessment of new pesticides;

[0104] 6. Simplify the experimental process and realize intelligence

[0105] The present invention greatly simplifies the traditional toxicity detection experimental process by combining surface plasmon resonance technology and protein chips. Compared with traditional methods, the experimental operation of this technology is simpler and easier to control, which significantly reduces the complex steps in the experimental process. By introducing intelligent data processing and toxicity prediction systems, experimental results can be automatically generated after the experiment, greatly improving work efficiency. This intelligent system not only speeds up data analysis, but also ensures the accuracy and consistency of the results, thereby reducing errors caused by human operation. In addition, due to the high degree of automation of the entire experimental process, the workload of operators is significantly reduced, allowing them to focus more on high-level tasks such as experimental design and data interpretation;

[0106] 7. Low experimental cost and high efficiency

[0107] The present invention has significant advantages in production cost and energy consumption. Compared with traditional toxicity detection methods, its production cost is low and its efficiency is high. By using surface plasmon resonance technology and protein chip technology, the chip can be reused, which greatly reduces the experimental cost. In addition, the present invention can also perform high-throughput detection, thereby significantly improving the detection efficiency of compound toxicity.

[0108] In summary, the present invention has successfully replaced traditional non-target biological toxicology experiments by innovatively combining surface plasmon resonance technology, protein chip technology and artificial intelligence algorithms, and achieved a more animal-friendly and ethical toxicity prediction method. With efficient sample utilization, accurate toxicity assessment and simplified experimental procedures, the present invention significantly improves the efficiency and accuracy of toxicity detection, reduces the consumption of experimental resources and the burden on operators. At the same time, the intelligent processing system of the present invention further reduces experimental costs, improves high-throughput detection capabilities, and makes the overall process of pesticide toxicity assessment more convenient and economical. These advantages not only promote the sustainable development of the field of pesticide research and development, but also provide strong technical support for pesticide safety assessment, and have broad application prospects and important practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 This is the SDS-PAGE electrophoresis diagram of the purified honey bee CSP3 protein, M, protein molecular weight standard; 1, CSP3 protein;

[0110] Figure 2 This is the SDS-PAGE electrophoresis diagram of the purified honey bee OBP14 protein, M, protein molecular weight standard; 1, OBP14 protein;

[0111] Figure 3 A schematic diagram of the construction process of a rapid prediction method for bee toxicity based on surface plasmon resonance technology and CSP3-OBP14 protein chip based on XGBoost algorithm;

[0112] Figure 4 Flow chart for prediction of toxicity of pesticide compounds to be tested;

[0113] Figure 5 Sensorgram of the experiment of detecting CSP3 and the pesticide compound to be tested by SPR;

[0114] Figure 6 Sensorgram of the experiment of detecting OBP14 and the tested pesticide compounds by SPR;

[0115] Figure 7 SPR affinity saturation curve of the tested pesticide and CSP3 1:1 binding fitting;

[0116] Figure 8 SPR affinity saturation curve of the tested pesticide and OBP14 1:1 binding fitting;

[0117] Fig. 9 The front-end interface of the bee toxicity data processing and prediction system based on the XGBoost algorithm. DETAILED DESCRIPTION

[0118] Example 1

[0119] A method of the present invention for predicting the toxicity of pesticides to non-target organisms by using non-target biological proteins comprises the following steps:

[0120] (1) Plasmid transformation of protein target gene

[0121] Constructing a non-target biological protein plasmid and performing plasmid transformation on the target gene of the non-target biological protein plasmid;

[0122] (2) Induction of protein target genes

[0123] Inducing the target gene treated in step (1) and collecting the supernatant;

[0124] (3) Protein purification

[0125] Purify the supernatant collected in step (2) to obtain a final protein concentration of no less than 1 mg / ml;

[0126] (4) Protein chip preparation

[0127] a. Prepare 1×PBS-P+protein coupling running buffer;

[0128] b. Dilute the protein purified in step (3) to a final concentration of 80 μg / mL using a sodium acetate solution at pH 1.0-4.0;

[0129] c. Ultrasonic exhaust of pH 1.0-4.0 protein sodium acetate solution at -4°C for 1-3 minutes;

[0130] d. Couple the protein sodium acetate treated in step c to the chip in a running buffer with a coupling amount of 8000-12000 RU, and fix it on the chip surface for later use;

[0131] Wherein, the coupling amount is according to the formula: get , Among them, Rmax is the maximum binding capacity of the chip surface, which is substituted into 100RU in protein test; Sm is the stoichiometric ratio, which is selected as 1 when unknown; RL is the ligand coupling level;

[0132] (5) Pesticide configuration

[0133] The prepared pesticide sample stock solution was diluted downward into 11 concentration gradients, namely 3000 μmol / L, 1000 μmol / L, 333.3 μmol / L, 111.1 μmol / L, 37 μmol / L, 12.3 μmol / L, 4.1 μmol / L, 1.37 μmol / L, 0.46 μmol / L, 0.15 μmol / L, 0.05 μmol / L; set aside;

[0134] (6) Detection of animal protein and pesticide dynamics

[0135] The temperature of the Biacore T200 biomolecular interaction analyzer was set to 21°C, 28°C, and 35°C, respectively. The chip packaged in step (4) was placed in the chip placement slot. The instrument parameters were set as follows: the Hertz number was set to 10 Hz, the loading speed was 10 μl / min, the binding time was 120 s, the dissociation time was 130 s, and the chip regeneration condition was to regenerate the chip with a pH 1.8, 10 mmol / L glycine hydrochloride solution at a flow rate of 30 μl / min for 30 s;

[0136] (7) Data processing

[0137] a. Open the Biacore T200 Evaluation Software (data analysis software) for the Biacore T200 biomolecular interaction analyzer in step (6) and click Open in File to find the file.

[0138] b. Open the software front-end interface and click the "Data Processing" button;

[0139] c. The software is based on the Sikuli library. It uses image recognition to find and control the buttons on the interface. Click SolventCorrection to perform solvent correction analysis and identify solvent correction.

[0140] d. In the Kinetics / Affinity-Select Curves interface, under Select evaluation mode, identify and select Single mode; in the Sample drop-down list in the Curve column, select the sample to be analyzed; identify and select the pesticide samples with different concentrations configured in step (5) for testing. If there are abnormal value results, uncheck the box in front of this concentration to delete this concentration;

[0141] e. Select kinetics for fitting, select 1:1 Binding for Model, select fitlocal for Rmax, and then click Fit to fit the data; after the kinetic fitting is completed, perform thermodynamic fitting to obtain the binding constant K between the test compound and the protein a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G ;

[0142] Plot of steady-state binding level (Req) versus analyte concentration (C) to calculate the equilibrium dissociation constant K for a 1:1 interaction D The equations include the following:

[0143] ;

[0144] Where: K D is the equilibrium dissociation constant (M); Rmax is the surface analyte binding capacity (RU); offset is the bulk refractive index contribution in the sample;

[0145] ;

[0146] The component of the mass transfer constant that is independent of flow rate is calculated using the following equation:

[0147] ;

[0148] where D is the diffusion coefficient of the analyte (m 2 ·s -1 ), f is the volume flow rate of the solution through the flow cell (m 3 ·s -1 ), h, w, l are the dimensions of the flow cell (height, width, length, in m);

[0149] ;

[0150] where G is the conversion factor from surface concentration to RU, and for proteins on sensor chip CM5, G is approximately 10 9 ;

[0151] ;

[0152] For equilibrium thermodynamics, the van't Hoff equation states:

[0153] ;

[0154] where Δ G is the standard Gibbs free energy change, R is the universal gas constant, T is the absolute temperature (K), and KD is the equilibrium dissociation constant;

[0155] (8) Model construction

[0156] a. Introduce the data obtained from the non-target biological toxicity database, the literature, and the toxicity test data of some similar compounds on bees into the data set;

[0157] b. Substitute the binding constant K obtained in step e of (7) a , dissociation constant K d , equilibrium dissociation constant KD , relaxation time t c , free energy Δ G As a data set, the data set is divided into a training set and a test set at a ratio of 8:2;

[0158] c. Globally optimize hyperparameters based on the Bayesian optimization algorithm and build a regression model for bee toxicity prediction based on XGBoost;

[0159] The described model architecture:

[0160] Define the following feature variables:

[0161] ;

[0162] Target variable:

[0163] ;

[0164] The squared error loss function is used for regression tasks, defined as , where y is the actual value, is the predicted value. XGBoost uses the gradient boosting framework to improve the model performance by gradually adding weak learners (decision trees). At the tth iteration, the model is as follows:

[0165] ;

[0166] in is the t-th tree, and the goal is to minimize the loss function:

[0167] ;

[0168] Bayesian optimization is used to select the best hyperparameters for the XGBoost model. Bayesian optimization finds the optimal hyperparameters by maximizing the expected improvement (El):

[0169] ;

[0170] in is the prediction of the current model, is the current optimal solution;

[0171] The optimal hyperparameters were found through Bayesian optimization , and used to train the XGBoost model:

[0172] ;

[0173] Where T is the total number of trees, It is the tth tree. When predicting new data, the feature variable X is input into the trained model to obtain the predicted value:

[0174] ;

[0175] e. Substitute the test set into the model for verification and optimization;

[0176] f. When predicting, the binding constant K of the ligand pesticide to be tested detected by the Biacore T200 biomolecular interaction analyzer a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G By inputting the numerical value into the model, the LD of the new compound to non-target organisms can be obtained. 50 value, obtain the toxicity level and complete the prediction of its toxicity.

[0177] Example 2

[0178] The non-target biological protein described in the present invention is an olfactory receptor protein, any one of human serum albumin, bovine serum albumin, hemoglobin, globulin, hemoglobin, myosin, collagen, enzyme protein, honey bee protein, zebrafish protein, human liver cells, and zebrafish cells. Replacing different proteins, cells, or tissues can predict different toxicities. For example, human liver cells / proteins are used for SPR experiments to assist in replacing and predicting the metabolic toxicity of pesticides to human liver; chip-encapsulated bird cells / protein cells are used for SPR experiments to predict the toxicity of pesticides to birds; chip-encapsulated zebrafish cells / proteins are used for SPR experiments to predict the toxicity of pesticides to aquatic organisms; chip-encapsulated silkworm cells / proteins are used for SPR experiments to predict and measure the ecological toxicity of pesticides; chip-encapsulated rat nerve cells / proteins are used for SPR experiments to predict the toxicity of pesticides to the mammalian nervous system; chip-encapsulated rabbit eye cells / proteins are used for SPR experiments. Predict the irritation and damage of pesticides to the eyes; conduct SPR experiments on chip-encapsulated plant cells / proteins to evaluate and predict the growth inhibition and potential phytotoxicity of pesticides on plants; conduct SPR experiments on chip-encapsulated zebrafish embryonic cells / proteins to predict the developmental toxicity of pesticides on aquatic organisms; conduct SPR experiments on chip-encapsulated pig liver cells / proteins to simulate and predict the metabolic toxicity of pesticides in farm animals; conduct SPR experiments on chip-encapsulated mouse cardiomyocytes / proteins to evaluate and predict the potential toxicity of pesticides to the heart; conduct SPR experiments on chip-encapsulated sea urchin embryonic cells / proteins to predict the toxicity of pesticides to marine invertebrates, etc.; including but not limited to the above.

[0179] Research has found that allelopathic protein 3 (CSP3) plays a vital role in the olfactory system of Italian honey bees, helping to detect pheromones and odorants required for foraging, navigation, and colony cohesion. CSP3 can bind and transport various ligands, so its interaction with neonicotinoid compounds is important for the sense of smell of bees. At the same time, CSP3 acts as a pheromone carrier in larvae, and its expression level in larvae is more than ten times that of adult bees. Therefore, compared with adult bees, bee larvae are at a much higher risk of exposure to sublethal concentrations of neonicotinoid pesticides.

[0180] The olfactory binding protein (OBP14) of honey bees plays an important role in the olfactory system of honey bees, identifying and responding to chemical signals in the environment by binding and transporting odor molecules. Neonicotinoids are a widely used class of pesticides that mainly act on the central nervous system of insects, causing neural overexcitation and death. However, this type of pesticide also has significant toxic effects on the OBP14 of honey bees. Neonicotinoid pesticides may compete with OBP14 for binding sites, interfere with normal odor molecule transmission, induce conformational changes in OBP14, and affect its ability to bind and transport odor molecules. In addition, bees exposed to neonicotinoid pesticides may have decreased expression levels of OBP14, resulting in impaired foraging and navigation abilities, which in turn affects their survival and reproduction.

[0181] It can be seen that allelopathic protein 3 (CSP3) and olfactory binding protein (OBP14) both play an important role in the olfactory system of bees, helping bees detect pheromones and odorants required for foraging, navigation, and colony cohesion. Both CSP3 and OBP14 can bind and transport various ligands, so their interactions with neonicotinoid compounds are of great significance to the olfactory function of bees. It is worth noting that due to the important role played by CSP3 and OBP14 in the olfactory system of bees and the controversy over their sensitivity to neonicotinoid pesticides, it is very important to predict the toxicity of compounds to bees based on these two proteins at the same time. This not only enables a more comprehensive assessment of the effects of pesticides on bees, but also greatly avoids the false positive problem that occurs when surface plasmon resonance technology detects the interaction between pesticides and a single protein. By comprehensively considering the interaction between CSP3 and OBP14, its toxicity can be predicted more accurately.

[0182] The present invention is further described below by taking honey bee olfactory receptor proteins (CSP3 and OBP14) as an example:

[0183] (1) Plasmid transformation of target genes of honey bee chemosensory protein CSP3 and honey bee odorant binding protein OBP14

[0184] a. Construct CSP3 and OBP14 plasmids respectively;

[0185] b. Take 1 µL of CSP3 and OBP14 plasmids respectively and add them to 50-200 µL BL21(DE3) competent cells; swirl gently to mix, and place on ice for 10-30 min;

[0186] c. Heat shock at 42°C in a water bath for 80 seconds, then quickly transfer the tube to an ice bath for 1-3 minutes;

[0187] d. Add 300 µL LB medium, mix well, and culture at 37 °C and 400 rpm for 1.5 h.

[0188] e. In a clean bench, spread the bacterial solution (200 μL) evenly on an LB plate containing kanamycin (200 μg / mL) using a disposable coating rod and place at room temperature until the liquid is absorbed;

[0189] f. Invert the plate and transfer it to a 37 °C biochemical incubator for 24 h.

[0190] (2) Induced expression of CSP3 and OBP14 target genes in Escherichia coli

[0191] a. Pick two single colonies and inoculate them into 5 mL LB liquid medium containing 0.2% kanamycin at 37°C and 220 rpm / min for about 8 h.

[0192] b. Inoculate 1% of each strain into 300 mL LB liquid medium, add 0.1% kanamycin, and induce at 37°C, 220 rpm / min, for 1-2 h.

[0193] c. Then, 150 μL IPTG was added to a final concentration of 0.5 mmol / L, and the cells were induced at 16 °C and 150 rpm / min for 12-16 h.

[0194] d. After induction, centrifuge at 8,000×g for 10 min (4°C), discard the supernatant, collect the sediment, and suspend the collected bacterial precipitate in 15-30 mL binding buffer; (the binding buffer is: 50 mmol / L imidazole, 20 mmol / L Tris-HC, 500 mmol / L NaCl, adjusted to pH 8.0);

[0195] e. Ultrasonic disruption of cells for 10 min (ultrasonic disruptor: power: 40%; ultrasound: 3 s, pause: 5 s);

[0196] f. Centrifuge at 8,000 rpm for 12 min at 4°C and collect the unpurified protein supernatant.

[0197] (3) Purification of CSP3 and OBP14 proteins

[0198] a. Preparation of gravity purification column: add 2 1.5 mL of affinity column material containing nickel (Ni) to 2 10 mL purification columns respectively, add 20 mL of ultrapure water to remove the original ethanol in the column material, and add 10 mL of 50 mmol / L binding buffer to balance the column material; (Binding buffer is: 50 mmol / L imidazole, 20 mmol / L Tris-HCl, 500 mmol / L NaCl, adjusted to pH 8.0);

[0199] b. Slowly load the supernatants of the two proteins collected respectively and collect the flow-through samples;

[0200] c. Elute the flow-through samples again and repeat the elution three times;

[0201] d. Add 50 mL of 50 mmol / L imidazole binding buffer to elute the impurities (Binding buffer: 50 mmol / L imidazole, 20 mmol / L Tris-HCl, 500 mmol / L NaCl, adjusted to pH 8.0);

[0202] e. Add 5 mL elution buffer to elute the target protein and collect it (Elution buffer: 250 mmol / L imidazole, 20 mmol / L Tris-HCl, 500 mmol / L NaCl, adjusted to pH 8.0);

[0203] f. Elute and collect the protein solution eluted with elution buffer again;

[0204] g. Desalting the purified protein solutions through desalting columns connected to the ÄKTA protein rapid purification system;

[0205] h. Take the desalted CSP3 protein solution and OBP14 protein solution for SDS-PAGE identification (see Appendix Figure 1 , 2);

[0206] j. Quantify the final concentration of recombinant CSP3 protein and OBP14 protein by BCA method, which should not be less than 1 mg / ml. If the concentration is not reached, ultrafiltration is performed using a 10 KD ultrafiltration tube.

[0207] (4) Preparation of CSP3-OBP14 protein chip

[0208] a. Prepare 1×PBS-P+protein coupling running buffer;

[0209] b. Use pH 4.0 sodium acetate solution to dilute CSP3 protein and OBP14 protein to a final concentration of 80 μg / mL respectively;

[0210] c. Ultrasonic exhaust was performed on the sodium acetate solutions of CSP3 protein and OBP14 protein at pH 4.0 at -4°C for 1-3 min respectively;

[0211] d. In order to meet the testing requirements of more test compounds, the two proteins (CSP3 and OBP14) were manually coupled to the CM5 chip (coupling amount was 8000-12000 RU) and fixed on the chip surface.

[0212] The coupling amount is based on the formula: , Among them, Rmax is the maximum binding capacity of the chip surface, which is usually substituted with 100RU in protein testing; Sm is the stoichiometric ratio, which is selected as 1 when unknown; RL is the ligand coupling level; the preferred coupling amount is 1.5 times RL; analyte MW is the molecular weight of the honey bee olfactory protein (CSP3: 19.6 kDa and OBP14: 19.1 kDa), and ligand MW is the molecular weight of the pesticide compound to be tested (100-1500);

[0213] The manual coupling steps are as follows: put the CM5 chip into the Biacore T200 biomolecular interaction analysis system, mix EDC and NHS in a ratio of 1:1 and pass them through channels 2 and 4 to activate the chip surface. The loading speed is 10 μl / min and the time is set to 420 s; that is, pass the CSP3 protein prepared with pH 4.0 sodium acetate solution into channel 2 at a loading speed of 30 μl / min for 300 s each time until the coupling amount is greater than 10,000 RU. Pass the ethanolamine solution at a loading speed of 10 μl / min for 300 s each time until the coupling amount is greater than 10,000 RU. Pass the ethanolamine solution at a loading speed of 10 μl / min for 420 s; pass the OBP14 protein prepared with pH 4.0 sodium acetate solution into channel 4 at a loading speed of 30 μl / min for 300 s each time until the coupling amount is greater than 10,000 RU. Pass the ethanolamine solution at a loading speed of 10 μl / min for 420 s;

[0214] If analyte MW / ligand MW>100, it is necessary to replace it with CM7 chip.

[0215] (5) Pesticide configuration

[0216] Select finished neonicotinoid pesticides (imidacloprid, nitenpyram, acetamiprid, chlorothiazide, thiamethoxam, thiacloprid, clothianidin, bispyribac, flupyralid, sulfoxaflor, cyclohexidine, and some similar compounds, etc.), carbamates, triazoles, organophosphorus and other pesticide technicals, including but not limited to the above pesticide technicals; also include different pesticide formulations, such as solid preparations (powders, granules, pellets, tablets, strips), dispersible solid preparations, solution preparations, solution preparations, etc.;

[0217] Prepare running buffer: Take 105mL of 10 × PBS-P+, add 0.22 μm filtered deionized water (895mL) to dilute 1000mL to prepare 1.05 × PBS-P+. Prepare 1.05 × PBS-P+10% DMSO running buffer and solvent correction stock solution according to Table 1, and prepare solvent correction point solution according to Table 2;

[0218] ;

[0219] Prepare pesticide solution: Dissolve the ligand pesticide to be tested in 100% DMSO to a stock solution of 0.06 M. Then dilute the ligand pesticide to be tested with the prepared Running Buffer containing 10% DMSO to 11 concentration gradients, namely 3000 μM, 1000 μM, 333.3 μM, 111.1 μM, 37 μM, 12.3 μM, 4.1 μM, 1.37 μM, 0.46 μM, 0.15 μM, 0.05 μM. Set a repeat concentration for injection and add a 0 concentration.

[0220] (6) Detection of the kinetics of CSP3 and OBP14 with neonicotinoid pesticides

[0221] The temperature of the Biacore T200 biomolecular interaction analysis system was set to 21°C, 28°C, and 35°C, respectively. The instrument parameters of the packaged CM5 chip were set as follows: the Hertz number was set to 10Hz, the loading speed was 10μl / min, the binding time was 120s, and the dissociation time was 130s. Filtered PBS-P+ buffer containing 10% DMSO was used as the running buffer. The neonicotinoid compounds were diluted into 11 concentration gradients (3000 μM, 1000 μM, 333.3μM, 111.1 μM, 37 μM, 12.3 μM, 4.1 μM, 1.37 μM, 0.46 μM, 0.15 μM, 0.05 μM) using the same running buffer. These compounds were then injected into channels Fc1 - Fc4 at a flow rate of 30 μL / min for a 120-second binding phase and a 130-second dissociation phase. The chip was regenerated using a 10 mmol / L glycine hydrochloride solution at pH 1.8 at a flow rate of 30 μl / min for 30 s.

[0222] Preferably, in the Kinetics / Affinity interface, set the Flow path points to 2-1, 4-3.

[0223] (7) Data processing

[0224] a. Open the Biacore T200 Evaluation Software (data analysis software) for the Biacore T200 biomolecular interaction analyzer in step (6) and click Open in File to find the file.

[0225] b. Open the software front-end interface and click the "Data Processing" button, as shown in the attached Fig. 9 As shown;

[0226] c. The software is based on the Sikuli library. It uses image recognition to find and control the buttons on the interface. Click SolventCorrection to perform solvent correction analysis and identify solvent correction.

[0227] d. In the Kinetics / Affinity-Select Curves interface, under Select evaluation mode, identify and select Single mode; in the Sample drop-down list in the Curve column, select the sample to be analyzed; identify and select the pesticide samples with different concentrations configured in step (5) for testing. If there are abnormal value results, uncheck the box in front of this concentration to delete this concentration, such as Figure 7 , Figure 8 As shown;

[0228] e. Select kinetics for fitting, select 1:1 Binding for Model, select fitlocal for Rmax, and then click Fit to fit the data; after the kinetic fitting is completed, perform thermodynamic fitting to obtain the binding constant K between the test compound and the protein a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G ;

[0229] Plot of steady-state binding level Req versus analyte concentration C Calculate the equilibrium dissociation constant K for a 1:1 interaction D The equations include the following:

[0230] ;

[0231] ;

[0232] Where: K D is the equilibrium dissociation constant (M); Rmax is the surface analyte binding capacity (RU); offset is the bulk refractive index contribution in the sample;

[0233] The component of the mass transfer constant that is independent of flow rate is calculated using the following equation:

[0234] ;

[0235] where D is the diffusion coefficient of the analyte (m 2 ·s -1 ), f is the volume flow rate of the solution through the flow cell (m 3 ·s -1 ), h, w, l are the dimensions of the flow cell (height, width, length, in m);

[0236] ;

[0237] where G is the conversion factor from surface concentration to RU, and for proteins on sensor chip CM5, G is approximately 10 9 ;

[0238] ;

[0239] For equilibrium thermodynamics, the van't Hoff equation states:

[0240] ;

[0241] where Δ Gis the standard Gibbs free energy change, R is the universal gas constant, T is the absolute temperature (K), and KD is the equilibrium dissociation constant.

[0242] (8) Model construction

[0243] a. Introduce the data obtained from the non-target biological toxicity database, the literature, and the toxicity test data of some similar compounds on bees into the data set;

[0244] b. Substitute the binding constant K obtained in step e of (7) a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G As a data set, the data set is divided into a training set and a test set at a ratio of 8:2;

[0245] c. Globally optimize hyperparameters based on the Bayesian optimization algorithm and build a regression model for bee toxicity prediction based on XGBoost;

[0246] The described model architecture:

[0247] Define the following feature variables:

[0248] ;

[0249] Target variable:

[0250] ;

[0251] The squared error loss function is used for regression tasks, defined as , where y is the actual value, is the predicted value. XGBoost uses the gradient boosting framework to improve the model performance by gradually adding weak learners (decision trees). At the tth iteration, the model is as follows:

[0252] ;

[0253] in is the t-th tree, and the goal is to minimize the loss function:

[0254] ;

[0255] Bayesian optimization is used to select the best hyperparameters for the XGBoost model. Bayesian optimization finds the optimal hyperparameters by maximizing the expected improvement (El):

[0256] ;

[0257] in is the prediction of the current model, is the current optimal solution;

[0258] The optimal hyperparameters were found through Bayesian optimization , and used to train the XGBoost model:

[0259] ;

[0260] Where T is the total number of trees, It is the tth tree. When predicting new data, the feature variable X is input into the trained model to obtain the predicted value:

[0261] ;

[0262] The code is as follows:

[0263] from sikuli import *

[0264] import pandas as pd

[0265] import xgboost as xgb

[0266] from sklearn.model_selection import train_test_split

[0267] from sklearn.metrics import mean_squared_error

[0268] from bayes_opt import BayesianOptimization

[0269] import numpy as np

[0270] # Open the model front-end interface

[0271] App.open("path_to_your_application_executable")

[0272] # Wait for the interface to load

[0273] wait("main_interface_image.png")

[0274] # (2) Click the "Data Processing" button

[0275] click("data_processing_button_image.png")

[0276] # (3) Use image recognition to find and click "solvent correction" to perform solvent correction analysis

[0277] wait("solvent_correction_button_image.png")

[0278] click("solvent_correction_button_image.png")

[0279] # (5) Select Single mode in the Kinetics / Affinity-Select Curves interface

[0280] wait("kinetics_affinity_select_curves_interface_image.png")

[0281] click("select_evaluation_mode_dropdown_image.png")

[0282] click("single_mode_option_image.png")

[0283] # Select the sample to be analyzed in the Sample drop-down list

[0284] click("sample_dropdown_image.png")

[0285] click("desired_sample_option_image.png")

[0286] # Detect inappropriate concentration results and remove the check mark

[0287] for concentration_checkbox in findAll("concentration_checkbox_image.png"):

[0288] click(concentration_checkbox)

[0289] # (6) Select kinetics for fitting

[0290] click("kinetics_tab_image.png")

[0291] click("model_dropdown_image.png")

[0292] click("1_1_binding_option_image.png")

[0293] click("rmax_dropdown_image.png")

[0294] click("fit_local_option_image.png")

[0295] click("fit_button_image.png")

[0296] # Wait for the fitting to complete

[0297] wait("fitting_complete_image.png")

[0298] # Perform thermodynamic fitting

[0299] click("thermodynamics_tab_image.png")

[0300] # Wait for the thermodynamic fit to complete and grab the results

[0301] wait("thermodynamics_fitting_complete_image.png")

[0302] results = {

[0303] "Ka1": find("ka1_value_image.png").text(),

[0304] "Ka2": find("ka2_value_image.png").text(),

[0305] "Kd1": find("kd1_value_image.png").text(),

[0306] "Kd2": find("kd2_value_image.png").text(),

[0307] "KD1": find("kd1_value_image.png").text(),

[0308] "KD2": find("kd2_value_image.png").text(),

[0309] "tc1": find("tau_c1_value_image.png").text(),

[0310] "tc2": find("tau_c2_value_image.png").text(),

[0311] "ΔG1": find("delta_h1_value_image.png").text(),

[0312] "ΔG2": find("delta_h2_value_image.png").text(),

[0313] }

[0314] # Print or store the results

[0315] print(results)

[0316] # Close the application

[0317] App.close("path_to_your_application_executable")

[0318] # Convert the captured results into a DataFrame

[0319] data = pd.DataFrame([results])

[0320] # Read data from Excel file

[0321] file_path = 'path_to_your_excel_file.xlsx'

[0322] df = pd.read_excel(file_path)

[0323] # Print the first few lines of data to check if it was read correctly

[0324] print(df.head())

[0325] # Define features and labels

[0326] X = df.drop("LD50", axis=1)

[0327] y = df["LD50"]

[0328] # Dataset division

[0329] X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

[0330] # Define the hyperparameter space of the XGBoost model

[0331] def xgb_evaluate(max_depth, gamma, colsample_bytree):

[0332] params = {

[0333] 'max_depth': int(max_depth),

[0334] 'gamma': gamma,

[0335] 'colsample_bytree': colsample_bytree,

[0336] 'eval_metric': 'rmse',

[0337] 'objective': 'reg:squarederror',

[0338] 'n_estimators': 100,

[0339] 'learning_rate': 0.1

[0340] }

[0341] model = xgb.XGBRegressor(**params)

[0342] model.fit(X_train, y_train)

[0343] predictions = model.predict(X_test)

[0344] return -mean_squared_error(y_test, predictions)

[0345] Bayesian Optimization

[0346] xgb_bo = BayesianOptimization(

[0347] xgb_evaluate,

[0348] {

[0349] 'max_depth': (3, 10),

[0350] 'gamma': (0, 1),

[0351] 'colsample_bytree': (0.3, 1)

[0352] },

[0353] random_state=42 )

[0355] xgb_bo.maximize(init_points=5, n_iter=25)

[0356] # Get the best parameters

[0357] best_params = xgb_bo.max['params']

[0358] best_params['max_depth'] = int(best_params['max_depth'])

[0359] # Train the XGBoost model using the optimal parameters

[0360] model = xgb.XGBRegressor(

[0361] max_depth=best_params['max_depth'],

[0362] gamma = best_params['gamma'],

[0363] colsample_bytree=best_params['colsample_bytree'],

[0364] eval_metric='rmse',

[0365] objective='reg:squarederror',

[0366] n_estimators=100,

[0367] learning_rate=0.1 )

[0369] model.fit(X_train, y_train)

[0370] # Evaluate the model

[0371] predictions = model.predict(X_test)

[0372] mse = mean_squared_error(y_test, predictions)

[0373] print("Mean Squared Error:", mse)

[0374] # Use the captured data to make predictions

[0375] new_data = data[['Ka1', 'Ka2', 'Kd1', 'Kd2', 'KD1', 'KD2', 'tc1', 'tc2', 'ΔG1', 'ΔG2']]

[0376] predicted_ld50 = model.predict(new_data)

[0377] print("Predicted LD50 value:", predicted_ld50)

[0378] The above algorithms are not limited to the XGBoost algorithm, but also include the following commonly used algorithms: Linear Regression, Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naive Bayes, Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), LightGBM, CatBoost, Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), GRU (Gated Recurrent Unit), Bayesian Optimization, Genetic Algorithm, Particle Swarm Optimization, Simulated Annealing, etc. Annealing、Reinforcement Learning、Principal Component Analysis (PCA), Factor Analysis、Independent Component Analysis (ICA), T-SNE (T-distributed Stochastic Neighbor Embedding), Autoencoder、Variational Autoencoder (VAE), Generative Adversarial Networks,GANs、K-means Clustering、Hierarchical Clustering、DBSCAN(Density-Based Spatial Clustering of Applications with Noise)、Fuzzy C-Means、Gaussian Mixture Model(GMM)、Bayesian Networks、Markov Chains、Hidden Markov Models(HMM)、Seq2Seq Models、Active Learning、Multi-task Learning、Transfer Learning、Zero-shot Learning、One-shot Learning、Semi-supervised Learning、Weakly-supervised Learning、Transformer Models、BERT(Bidirectional Encoder Representations from Transformers)、GPT(Generative Pre-trained Transformer), Ensemble Learning Methods, Boosted Trees, and Relational Networks, etc.

[0379] e. Substitute the test set into the model for verification and optimization;

[0380] f. When predicting, the binding constant K of the ligand pesticide to be tested detected by the Biacore T200 biomolecular interaction analyzer a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G By inputting the numerical value into the model, the LD of the new compound to non-target organisms can be obtained. 50 value, obtain the toxicity level and complete the prediction of its toxicity.

[0381] Next, we selected neonicotinoid pesticides, including imidacloprid, nitenpyram, acetamiprid, thiamethoxam, thiacloprid, clothianidin, flupyrazone, sulfoxaflor, and some similar compounds, such as IPP217B16, IPP251A19, IPP217D25, IPP250C3, and IPP250S10. The software automatically captured the test data (see Table 3), substituted it into the global optimization hyperparameters based on the Bayesian optimization algorithm, and constructed a regression model for bee toxicity prediction based on XGBoost (R 2 =0.91).

[0382] The neonicotinoid compounds I-6 and PMDD-04 were selected and tested using this method to obtain the binding constant K a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , and the free energy Δ G Then, the actual LD ​​was obtained through the bee live toxicity test. 50 The predicted LD 50 The results show that the LD predicted by the model 50 The value is very close to the actual measured value, the error is extremely small, and the toxicity level prediction is completely accurate.

[0383] Table 3 Some biomolecular interaction datasets and LD used to build models 50 Dataset

[0384]

[0385] Table 4 Biomolecular interaction information of two compounds used to validate the model

[0386]

[0387] Table 5 Prediction results of XGBoost model

[0388]

[0389] SSE = predicted LD50 - actual LD50.

[0390] It can be seen that the above Example 2 constructs an XGBoost model based on the interaction parameters between the tested neonicotinoid pesticide and the bee olfactory protein (CSP3 and OBP14), realizes the accurate prediction of the toxicity of neonicotinoid pesticides to bees, and can more effectively evaluate its harm to bees. Comprehensively considering that neonicotinoid drugs affect bees mainly through two olfactory proteins, chemosensory protein (CSP3) and odor binding protein (OBP14), by analyzing their interaction information, accurately predicting the toxicity to bees, and predicting the toxicity of compounds to bees based on two proteins, this can not only more comprehensively evaluate the impact of compounds on bees, but also greatly avoid the biggest problem of false positives when surface plasmon resonance technology is used to detect the interaction between pesticides and proteins. This method does not require traditional bee toxicology experiments, avoids the use of a large number of bees, and complies with the principle of animal-friendly. Through surface plasmon resonance technology and protein chips, the toxicity of neonicotinoid pesticides can be evaluated without harming bees, which has important ethical and environmental significance. Through the developed deep learning model, especially the XGBoost algorithm, the present invention can efficiently and accurately predict the toxicity of neonicotinoid compounds. The algorithm can make full use of a large amount of experimental data, automatically perform feature selection and model optimization, and significantly improve the accuracy and reliability of toxicity prediction.

Claims

1. A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins, characterized in that The following steps are involved: (1) Plasmid transformation of protein target gene Constructing a non-target biological protein plasmid and performing plasmid transformation on the target gene of the non-target biological protein plasmid; (2) Induction of protein target genes Inducing the target gene treated in step (1) and collecting the supernatant; (3) Protein purification Purify the supernatant collected in step (2) to obtain a final protein concentration of no less than 1 mg / ml; (4) Protein chip preparation a. Prepare 1×PBS-P+protein coupling running buffer; b. Dilute the protein purified in step (3) to a final concentration of 80 μg / mL using a sodium acetate solution at pH 1.0-4.0; c. Ultrasonic exhaust of pH 1.0-4.0 protein sodium acetate solution at -4°C for 1-3 minutes; d. Couple the protein sodium acetate treated in step c to the chip in a running buffer with a coupling amount of 8000-12000 RU, and fix it on the chip surface for later use; Wherein, the coupling amount is according to the formula: get , Among them, Rmax is the maximum binding capacity of the chip surface, which is substituted into 100RU in protein test; Sm is the stoichiometric ratio, which is 1 when unknown; RL is the ligand coupling level; analyte MW is the molecular weight of the bee olfactory protein, and ligand MW is the molecular weight of the pesticide sample to be tested; (5) Pesticide configuration The prepared pesticide sample stock solution was diluted downward into 11 concentration gradients, namely 3000 μmol / L, 1000 μmol / L, 333.3 μmol / L, 111.1 μmol / L, 37 μmol / L, 12.3 μmol / L, 4.1 μmol / L, 1.37 μmol / L, 0.46 μmol / L, 0.15 μmol / L, 0.05 μmol / L; set aside; (6) Detection of animal protein and pesticide dynamics The temperature of the Biacore T200 biomolecular interaction analyzer was set to 21°C, 28°C, and 35°C, respectively. The chip packaged in step (4) was placed in the chip placement slot. The instrument parameters were set as follows: the Hertz number was set to 10 Hz, the loading speed was 10 μl / min, the binding time was 120 s, the dissociation time was 130 s, and the chip regeneration condition was to regenerate the chip with a pH 1.8, 10 mmol / L glycine hydrochloride solution at a flow rate of 30 μl / min for 30 s; (7) Data processing a. Open the Biacore T200 Evaluation Software (data analysis software) for the Biacore T200 biomolecular interaction analyzer in step (6) and click Open in File to find the file. b. Open the software front-end interface and click the "Data Processing" button; c. The software is based on the Sikuli library, and uses image recognition to find and control the buttons on the interface, and automatically clicks on solventcorrection to perform solvent correction analysis and identify solvent correction; d. In the Kinetics / Affinity-Select Curves interface, under Select evaluation mode, identify and select Single mode; in the Sample drop-down list in the Curve column, select the sample to be analyzed; identify and select the pesticide samples with different concentrations configured in step (5) for testing, automatically determine the abnormal value results, and uncheck the box in front of this concentration to delete this concentration; e. Select kinetics for fitting, select 1:1 Binding for Model, select fit local for Rmax, and then click Fit to fit the data; after the kinetics fitting is completed, perform thermodynamic fitting to obtain the binding constant K between the test compound and the protein a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy ΔG; Plot of steady-state binding level Req versus analyte concentration C Calculate the equilibrium dissociation constant K for a 1:1 interaction D The equations include the following: ; ; Where: K D is the equilibrium dissociation constant M; Rmax is the surface analyte binding capacity RU; offset is the bulk refractive index contribution in the sample; The component of the mass transfer constant that is independent of flow rate is calculated using the following equation: ; Where D is the diffusion coefficient of the analyte, f is the volume flow rate of the solution through the flow cell, and h, w, and l are the dimensions of the flow cell; ; where G is the conversion factor from surface concentration to RU, and for proteins on sensor chip CM5, G is approximately 10 9 ; ; For equilibrium thermodynamics, the van't Hoff equation states: ; Where ΔG° is the standard Gibbs free energy change, R is the universal gas constant, T is the absolute temperature, and KD is the equilibrium dissociation constant; (8) Model construction a. Introduce the data obtained from the non-target biological toxicity database, the literature, and some similar compounds in the bee toxicity test experiment into the data set; b. Substitute the binding constant K obtained in step e of (7) a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G As a data set, the data set is divided into a training set and a test set at a ratio of 8:2; c. Globally optimize hyperparameters based on the Bayesian optimization algorithm and build a regression model for bee toxicity prediction based on XGBoost; The described model architecture: Define the following feature variables: ; Target variable: ; The squared error loss function is used for regression tasks, defined as , where y is the actual value, is the predicted value. XGBoost uses the gradient boosting framework to improve the model performance by gradually adding weak learners. At the tth iteration, the model is as follows: ; in is the t-th tree, and the goal is to minimize the loss function: ; Bayesian optimization is used to select the best hyperparameters for the XGBoost model. Bayesian optimization finds the optimal hyperparameters by maximizing the expected improvement: ; in is the prediction of the current model, is the current optimal solution; The optimal hyperparameters were found through Bayesian optimization , and used to train the XGBoost model: ; Where T is the total number of trees, It is the tth tree. When predicting new data, the feature variable X is input into the trained model to obtain the predicted value: ; e. Substitute the test set into the model for verification and optimization; f. When predicting, the binding constant K of the pesticide sample to be tested will be detected by the Biacore T200 biomolecular interaction analyzer. a , dissociation constant K d , equilibrium dissociation constant K D , relaxation time t c , free energy Δ G By inputting the numerical value into the model, the LD of the new compound to non-target organisms can be obtained. 50 value, obtain the toxicity level and complete the prediction of its toxicity.

2. A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins as claimed in claim 1, characterized in that The plasmid transformation of the protein target gene in step (1) is achieved by the following steps: (1) Construction of plasmids for non-target biological proteins; (2) Add 1 µL of plasmid to 50-200 µL of BL21 competent cells, swirl gently to mix, and place on ice for 10-30 min. (3) Heat shock in a 42°C water bath for 70-90 seconds, then quickly transfer the tube to an ice bath for 1-3 minutes. (4) Add 300 µL LB medium, mix well, and culture at 37 °C, 200-500 rpm with shaking for 1-2 h; (5) In a clean bench, evenly spread 100-300 μL of bacterial solution on an LB plate containing 100-300 μg / mL kanamycin using a disposable coating rod and place at room temperature until the liquid is absorbed; (6) Invert the plate and transfer it to a 37 °C biochemical incubator for 24 h in the dark.

3. A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins as claimed in claim 1, characterized in that Induction of the protein target gene in step (2) is achieved by the following steps: (1) Pick two single colonies and inoculate them into 5 mL LB liquid medium containing 0.1-0.3% kanamycin at 37°C and 220 rpm / min for about 6-8 h. (2) Inoculate 200-400 mL LB liquid culture medium at a ratio of 1%, add 0.1% kanamycin, incubate at 37°C, 220 rpm / min, and induce for 1-2 h; (3) Then, add 150 μL IPTG to a final concentration of 0.5 mmol / L, incubate at 16 °C, 150 rpm / min, and induce for 12-16 h; (4) After induction, centrifuge at 8,000 × g for 10 min (4 °C), discard the supernatant, collect the precipitated cells, and resuspend the collected precipitated cells in 15-30 mL binding buffer; (5) Ultrasonic disruption of cells for 8-12 min; (6) Centrifuge at 8,000 rpm at 4°C for 5-15 min and collect the unpurified protein supernatant.

4. A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins as claimed in claim 1, characterized in that Step (3) Protein purification is achieved by the following steps: (1) Preparation of gravity purification column: Add 2 1.5 mL affinity columns containing nickel to a 10 mL purification column, add 20 mL ultrapure water to remove the original ethanol in the column material, and add 10 mL of 50 mmol / L binding buffer to equilibrate the column material; (2) Slowly load the collected protein supernatant and collect the flow-through sample; (3) Elute the flow-through sample again, repeat the elution three times, collect the eluate, and obtain a preliminarily purified protein solution; (4) The protein solution after preliminary purification is desalted by a desalting column connected to the ÄKTA protein rapid purification system to obtain a protein solution with a final protein concentration of not less than 1 mg / ml.

5. A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins as claimed in claim 1, characterized in that The non-target biological protein is any one of olfactory receptor protein, human serum albumin, bovine serum albumin, hemoglobin, globulin, myosin, collagen, enzyme protein, honey bee protein, zebrafish protein, human liver cells, and zebrafish cells.

6. A method for predicting the toxicity of pesticides to non-target organisms using non-target biological proteins as claimed in claim 1, characterized in that In step (8) c, the optimized hyperparameters include: n_estimators, max_depth, learning_rate, min_child_weight, subsample, colsample_bytree, gamma.