CN113948156A

CN113948156A - Multitask neural network method for predicting degradation half-life of chemicals in four environment media

Info

Publication number: CN113948156A
Application number: CN202111388088.5A
Authority: CN
Inventors: 陈景文; 许喆; 傅志强
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2021-10-20
Filing date: 2021-11-22
Publication date: 2022-01-18
Anticipated expiration: 2041-11-22
Also published as: CN113948156B

Abstract

The invention belongs to the technical field of high-throughput prediction for chemical risk management, and discloses a multitask neural network method for predicting degradation half-life of chemicals in four environment media. On the basis of the known molecular structure of the chemical, the degradation half-life of the chemical in four media can be obtained by calculating the molecular fingerprint and applying the constructed method. The method is simple, convenient and efficient, has low cost, and can save the resource investment of experimental tests. The construction process of the method is as follows: collecting degradation half-reduction period data; molecular PubChem fingerprint calculation; training a multitask neural network model; evaluating the performance of the model by selecting indexes such as a measurement speculative value-predicted value decision coefficient; characterizing a model application domain by referring to an OECD guide rule; the prediction model established by the invention has good fitting capability, robustness and prediction capability, can effectively predict the degradation half-life of chemicals in an application domain in four environment media, provides a necessary tool for sound management of the chemicals, and has important significance.

Description

Multitask neural network method for predicting degradation half-life of chemicals in four environment media

Technical Field

The invention belongs to the technical field of high-throughput screening for chemical risk management, and discloses a method for predicting degradation half-life of chemicals in four environment media (atmosphere, water body, soil and sediment) based on a quantitative structure-activity relationship (QSAR) model.

Background

The environmental persistence evaluation of chemicals is one of the core contents of chemical risk management. Environmental persistence refers to the ability of a chemical to degrade and transform with difficulty and to remain unchanged for a long period of time in the environment. The most common indicator for assessing the environmental persistence of a chemical is its degradation half-life (t) in an environmental medium_1/2) I.e., the time required for degradation of the chemical from the ambient medium to remove half of the initial amount. t is t_1/2Is an important index for determining the environmental fate of chemicals and is also a key parameter for evaluating and controlling the durability, the bioaccumulation and the toxic (PBT) chemicals by global regulations.

T of chemical in ambient medium_1/2Is determined by various degradation reactions (such as biodegradation, hydrolysis, photolysis, atmospheric oxidative degradation, etc.). The economic cooperation and development Organization (OECD) evaluates the environmental persistence of chemicals mainly based on biodegradability tests, issuing test guidelines for biodegradability of chemicals in surface water (OECD guideline 309), sediment (OECD guideline 308) and soil (OECD guideline 307). Testing each of the obtained media organismsThe degradation half-life is compared to persistence standards in the relevant regulations (e.g., REACH regulations) to determine whether a chemical is non-persistent (nP), persistent (P), or very persistent (vP).

Experimental testing to obtain t of chemical in environmental medium_1/2Low efficiency, long time consumption, high cost, and the need to develop efficient (high throughput, low cost) analog prediction techniques. A quantitative structure-activity relationship (QSAR) -based computational simulation technology can effectively predict t of chemicals in an environmental medium by establishing the correlation between the molecular structure characteristics of the chemicals and the environmental behavior parameters thereof_1/2. With the development of a machine learning algorithm, the QSAR based on the machine learning algorithm has strong advantages in the aspect of mining the internal relation between the predicted end point and the molecular characteristics, wherein the multi-task learning technology can learn the related information between different predicted end points through a characteristic and parameter sharing mechanism, so that the QSAR effectively improves the prediction performance of the model on the basis of simultaneously predicting different end points, and is expected to be applied to chemicals t_1/2Plays an important role in the aspect of prediction and is helpful for screening out the preferentially controlled persistent chemicals.

At present, some researches have been carried out to construct t of chemicals in environmental media_1/2QSAR predictive model of (1). The literature 'Water Res, 2019,157, 181-190' constructs a multiple linear regression model of biodegradation half-reduction period of aromatic hydrocarbon and derivatives thereof in Water; the document J.Chemnformatics, 2018,10 and 10' constructs a K nearest neighbor regression model of biodegradation half-reduction period of hydrocarbon compounds in water; the literature, "ecotox. environ. safe, 2016,129,10-15," constructs a support vector machine model that predicts half-life of herbicide biodegradation in soil. The existing models have a small application domain and are only suitable for predicting a single endpoint, i.e. a chemical in a single medium t_1/2The correlation between different end points is neglected, so that the model prediction effect is difficult to further improve and the like; while also lacking t in the deposit_1/2The predictive model of (1).

Based on the reasons, t of 250 chemicals in four media of atmosphere, water body, soil and sediment is comprehensively obtained by gathering documents_1/2Data, data set covering organic acidsEster, ether, ketone, alcohol, phenol, aniline, polycyclic aromatic hydrocarbon, heterocyclic compound, halogenated hydrocarbon and other chemicals, and a multilayer feedforward neural network combined multitask learning technology is used for constructing a method capable of simultaneously predicting t of the chemicals in four media_1/2The multi-task neural network quantitative model is characterized in the application domain of the model, and the application range of the model is determined.

Disclosure of Invention

The invention constructs a simple and efficient method for predicting t of chemicals in four media_1/2The method can predict the degradation half-life of the chemicals in four media simultaneously according to SMILES codes of the chemicals, and provides a basic tool for screening the PBT chemicals; in the modeling process, the QSAR model is constructed and used by referring to the OECD, and the robustness and the prediction capability of the model are examined through internal and external verification.

The technical scheme of the invention is as follows:

a multitask neural network method for predicting degradation half-life of chemicals in four environmental media comprises the following steps:

(1) data gathering

250 chemicals were collected from the literature in four media t_1/2(ii) a Generating SMILES codes corresponding to chemicals by using an RDkit package in Python 3.8.8 software;

(2) calculating molecular fingerprints of chemicals

Adopting Open Babel 2.3.2.2 software to convert the CSV format file recorded with the chemical SMILES code into an SDF format file; inputting the SDF file into PaDEL-Descriptor 2.21 software, and calculating PubChem molecular fingerprints of 250 chemicals;

(3) model training

Molecular fingerprinting and logt of PubChem of chemicals_1/2Merging data; randomly splitting a data set into a training set and a verification set according to a ratio of 4:1, and using chemicals to logt in four media_1/2(t_1/2Unit: h) taking PubChem fingerprints of chemicals as independent variables, and training a multi-task model by adopting a multi-layer feedforward neural network and combining a multi-task learning technology; to avoid overfitting, a batch process is usedPhysical and Dropout methods; determining the optimal hyper-parameter of the algorithm by a grid search method; constructing a model based on the optimal hyper-parameters and verifying logt of the compounds_1/2And (4) predicting data to represent the external prediction performance of the model.

The model optimal hyperparameters are as follows: the neuron comprises a first hidden layer and a second hidden layer, wherein the first hidden layer and the second hidden layer respectively comprise 100 neurons and 10 neurons; a Dropout layer is arranged behind the first hidden layer, and the Dropout rate is set to be 20%; both hidden layers adopt a linear rectification function (ReLU) as an activation function; training 16 chemicals in each batch, wherein the total number of iterations is 300, namely, batchsize is 16, and epoch is 300; the loss function index is Mean Square Error (MSE); the optimizer selects an adaptive moment estimate (Adam); the optimizer step size is set to 0.005, i.e. leaningrate is 0.005; setting the same weight factors for the four tasks during optimization;

(4) model evaluation

Estimation-prediction decision coefficient (R) using training set measurements²) Root Mean Square Error (RMSE), Mean Absolute Error (MAE) characterizes model goodness of fit; r of verification set²RMSE, MAE characterize model prediction capability; ten-fold cross-validation coefficient (Q) using training set² ₁₀) The model robustness is characterized.

The prediction effect of the final model is as follows:

chemicals log in atmosphere_1/2(t_1/2Unit: h) the predicted effect of (2): r² _train＝0.988，RMSE_train＝0.094，MAE_train＝0.070，Q² ₁₀＝0.889，R² _test＝0.713，RMSE_test＝0.348，MAE_test＝0.244；

Chemical logt in water_1/2(t_1/2Unit: h) the predicted effect of (2): r² _train＝0.976，RMSE_train＝0.121，MAE_train＝0.087，Q² ₁₀＝0.895，R² _test＝0.802，RMSE_test＝0.305，MAE_test＝0.205；

Chemical log in soil_1/2(t_1/2Unit: h) predicted effect ofAnd (4) fruit: r² _train＝0.981，RMSE_train＝0.112，MAE_train＝0.084，Q² ₁₀＝0.941，R² _test＝0.883，RMSE_test＝0.261，MAE_test＝0.204；

Chemical logt in deposits_1/2(t_1/2Unit: h) the predicted effect of (2): r² _train＝0.979，RMSE_train＝0.107，MAE_train＝0.079，Q² ₁₀＝0.924，R² _test＝0.870，RMSE_test＝0.261，MAE_test＝0.199；

(5) Application domain characterization

Generating MACCS molecular fingerprints of the chemicals by using an RDkit software package, and calculating the similarity (Tanimoto similarity) between the chemical molecules A of the verification set and the chemical molecules B of the training set according to the following calculation formula:

wherein S is_ABIs the degree of similarity of the valleys of molecules A and B, X_jAIs the jth fingerprint feature of molecule A, X_jBIs the jth feature of the molecule B and n is the number of feature bits of the fingerprint.

Defining a similarity threshold (S)_cutoff) And the minimum number of similar molecules (N)_min) Defining an application domain, i.e. if the similarity between the training set and the target molecule is greater than S_cutoffChemical number of (2) exceeds N_minThen the molecule is determined to be within the application domain. The application domains of the invention are: s_cutoff＝0.6，N_min＝5。

The invention has the advantages that: the established model can simultaneously predict the logt of the chemicals in four media_1/2Because the correlation information between different end points is considered, the prediction performance of the model is greatly improved, and the model has a clearly characterized application domain. The method is simple, convenient, efficient and low in cost, is expected to play a role in high-throughput prediction of chemical degradation half-life data, and is used for healthy chemical managementProvides a basic tool, and meets the national important requirements of chemical risk management and control and new pollutant treatment.

Drawings

FIG. 1 shows the overall process flow.

FIG. 2 shows the chemical logt in four media_1/2(t_1/2The unit is: h) measuring a linear fitting graph of the speculative value and the predicted value, wherein the chemicals in the training set and the verification set are respectively 200 and 50; (a) is in the atmosphere; (b) is in a body of water; (c) is in the soil; (d) is in the deposit.

Detailed Description

The following further describes a specific embodiment of the present invention with reference to the drawings and technical solutions.

Example 1

Given a chemical 2,2' -dichlorobiphenyl (CAS number: 13029-08-8), logt is to be predicted in four media_1/2(unit: h). Firstly, according to SMILES code of 2,2' -dichlorobiphenyl, utilizing RDkit software package to calculate MACCS molecular fingerprint, then calculating the similarity of the MACCS molecular fingerprint and grain of chemical molecules in a training set, and calculating to obtain that the similarity of the chemical molecules in the training set and the grain is more than 0.6 (S)_cutoff) Has 6 molecules (more than N)_min5), 2' -dichlorobiphenyl is within the domain of model applications. Further, the PubChem molecular fingerprint is calculated by using the PaDEL-Descriptor software, and the multi-task neural network model constructed by the method is used for prediction. The results were obtained:

logt_{1/2 (atmosphere)}＝2.31，logt_{1/2 (Water body)}＝3.79，logt_{1/2 (soil)}＝4.15，logt_{1/2 (sediment)}4.25, the corresponding measurement estimate is: logt (r)_{1/2 (atmosphere)}＝2.23，logt_{1/2 (Water body)}＝3.74，logt_{1/2 (soil)}＝4.23，logt_{1/2 (sediment)}The predicted value and the measurement estimate agree very well with 4.23.

Example 2

Given a chemical cyclohexanol (CAS number: 108-93-0), it is predicted to be logt in four media_1/2(unit: h). Firstly, according to SMILES code of cyclohexanol, using RDkit software package to calculate MACCS molecular fingerprintThen calculating the similarity of the particles with the chemical molecules in the training set, wherein the similarity of the molecules in the training set with the chemical molecules is calculated to be more than 0.6 (S)_cutoff) Has 8 molecules (more than N)_min5), cyclohexanol is within the model application domain. Further, the PubChem molecular fingerprint is calculated by using the PaDEL-Descriptor software, and the multi-task neural network model constructed by the method is used for prediction. The results were obtained:

logt_{1/2 (atmosphere)}＝1.70，logt_{1/2 (Water body)}＝1.77，logt_{1/2 (soil)}＝1.60，logt_{1/2 (sediment)}2.25, the corresponding measurement estimate is: logt (r)_{1/2 (atmosphere)}＝1.74，logt_{1/2 (Water body)}＝1.74，logt_{1/2 (soil)}＝1.74，logt_{1/2 (sediment)}The predicted value and the measurement estimate agree very well with 2.23.

Example 3

Given a chemical n-pentane (CAS number: 109-66-0), it is predicted to be logt in four media_1/2(unit: h). Firstly, according to SMILES code of n-pentane, utilizing RDkit software package to calculate MACCS molecular fingerprint, then calculating the similarity of the MACCS molecular fingerprint and the grain of chemical molecules in a training set, and calculating to obtain that the similarity of the chemical molecules in the training set and the grain of the chemical molecules in the training set is more than 0.6 (S)_cutoff) Has 15 molecules (more than N)_min5) so n-pentane is within the model application domain. Further, the PubChem molecular fingerprint is calculated by using the PaDEL-Descriptor software, and the multi-task neural network model constructed by the method is used for prediction. The results were obtained:

logt_{1/2 (atmosphere)}＝1.30，logt_{1/2 (Water body)}＝2.75，logt_{1/2 (soil)}＝3.24，logt_{1/2 (sediment)}3.74, the corresponding measurement estimate is: logt (r)_{1/2 (atmosphere)}＝1.23，logt_{1/2 (Water body)}＝2.74，logt_{1/2 (soil)}＝3.23，logt_{1/2 (sediment)}The predicted value and the measurement estimate agree very well with 3.74.

Claims

1. A multitasking neural network method for predicting degradation half-life of chemicals in four environmental media, comprising the steps of:

(1) data gathering

250 chemicals were collected from the literature for degradation log half-life in four media_1/2And generating a SMILES code corresponding to the chemical;

(2) calculating molecular fingerprints of chemicals

Converting the CSV format file recording the SMILES code of the chemical into an SDF format file; calculating PubChem molecular fingerprints of 250 chemicals according to the SDF format file;

(3) model training

Molecular fingerprinting and logt of PubChem of chemicals_1/2Merging data; randomly splitting a data set into a training set and a verification set according to a ratio of 4:1, and using the logt of chemicals in four media_1/2Taking PubChem molecular fingerprints of chemicals as independent variables, and training a multi-task model by adopting a multi-layer feedforward neural network and combining a multi-task learning technology; determining the optimal hyper-parameter of the algorithm by a grid search method; constructing a model based on the optimal hyper-parameters, predicting degradation half-life data of the chemicals in the verification set, and representing the external prediction performance of the model;

the model optimal hyperparameters are as follows: the neuron comprises a first hidden layer and a second hidden layer, wherein the first hidden layer and the second hidden layer respectively comprise 100 neurons and 10 neurons; a Dropout layer is arranged behind the first hidden layer, and the Dropout rate is 20%; the two hidden layers adopt a linear rectification function as an activation function; training 16 chemicals in each batch, wherein the total number of iterations is 300, namely, batchsize is 16, and epoch is 300; the loss function index is Mean Square Error (MSE); the optimizer selects an adaptive moment estimate; the optimizer step size is set to 0.005, i.e. leaningrate is 0.005; setting the same weight factors for the four tasks during optimization;

(4) model performance assessment

Coefficient R is determined using training set measurement guess-predictor²Root mean square error RMSE, mean absolute error MAE characterize model goodness of fit; r of verification set²RMSE, MAE characterize model prediction capability; ten-fold cross-validation coefficient Q using training set² ₁₀Characterizing model robustness;

(5) application domain characterization

Generating MACCS molecular fingerprints of the chemicals, and calculating the similarity of the grains between the chemical molecules A in the verification set and the chemical molecules B in the training set, wherein the calculation formula is as follows:

wherein S is_ABIs the degree of similarity of the valleys of molecules A and B, X_jAIs the jth fingerprint feature of molecule A, X_jBIs the jth feature of the molecule B, and n is the number of feature bits of the fingerprint;

by a self-defined similarity threshold S_cutoffAnd the minimum number of similar molecules N_minTo define the application domain, i.e. if the similarity of the targeted molecule to the trough in the training set is greater than S_cutoffChemical number of (2) exceeds N_minThen the molecule is determined to be within the application domain.

2. The method of claim 1, wherein the defined application domains are: s_cutoff＝0.6，N_min＝5。