A training method and a prediction method of a chemical genetic toxicity prediction model

Abstract

The present disclosure provides a training method of a chemical genetic toxicity prediction model, which can be applied to the technical field of chemical environmental health risk assessment. The training method of the chemical genetic toxicity prediction model comprises: obtaining original data about chemicals with known genetic toxicity, wherein the original data comprises chemical codes of the chemicals, differential gene expression data of the chemicals, and in vitro high-throughput test data of the chemicals; generating a training sample data set according to the original data, wherein the training sample data set comprises an atomic feature matrix, a connection relationship matrix, a differential gene expression matrix, an in vitro high-throughput test matrix, and label information of the chemicals; and training an initial model using the training sample data set to obtain a chemical genetic toxicity prediction model. The present disclosure also provides a prediction method of chemical genetic toxicity.

Description

Technical Field

[0001] The present disclosure relates to the technical field of chemical environmental health risk assessment, and more specifically to a training method and a prediction method for a chemical genotoxicity prediction model. Background Art

[0002] It is reported that 90-95% of cancers are attributed to environmental exposures and lifestyle factors, and that genotoxicity of pollutants, which damages genetic material at different levels, is one of the main factors inducing cancer. Therefore, in order to protect humans from such potential risks, it is very important to assess the genotoxic potential of chemicals to which humans are exposed.

[0003] The genotoxicity of chemicals has complex mechanisms, encompassing not only mutagenicity at genetic loci with varying modes of action, but also non-mutagenicity, which involves chromosomal aberrations rather than direct gene action. Currently, in vivo animal testing is widely used to assess chemical genotoxicity, but this method is time-consuming, costly, and inefficient. Summary of the Invention

[0004] In view of the above problems, the present disclosure provides a training method and a prediction method for a chemical genotoxicity prediction model.

[0005] According to a first aspect of the present disclosure, a method for training a chemical genotoxicity prediction model is provided, comprising:

[0006] Obtaining raw data on chemicals with known genotoxicity, wherein the raw data includes chemical codes of the chemicals, differential gene expression data of the chemicals, and in vitro high-throughput testing data of the chemicals;

[0007] Generating a training sample data set based on the raw data, wherein the training sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, an in vitro high-throughput test matrix, and label information of the chemicals;

[0008] The initial model was trained using the above training sample data set to obtain a chemical genotoxicity prediction model.

[0009] According to an embodiment of the present disclosure, generating a training sample dataset based on the original data includes:

[0010] Converting the chemical code into a two-dimensional graphic structure representing the spatial structure of the chemical, wherein atoms in the two-dimensional graphic structure are represented by points and chemical bonds are represented by lines;

[0011] Determining the number of atoms, atomic feature vectors, and connection relationships between atoms corresponding to the chemical code according to the two-dimensional graphic structure;

[0012] Constructing the atomic characteristic matrix according to the number of atoms and the atomic characteristic vector;

[0013] The connection relationship matrix is ​​constructed according to the number of atoms and the connection relationship between the atoms.

[0014] According to an embodiment of the present disclosure, the above-mentioned atomic feature vector is represented by a digital vector of length n, where n is the number of atomic features, 50≤n≤75.

[0015] According to an embodiment of the present disclosure, the connection relationship between the above atoms includes using a value "1" to indicate that a chemical bond exists between the two atoms, and a value "0" to indicate that no chemical bond exists between the two atoms.

[0016] According to an embodiment of the present disclosure, the initial model includes a chemical structure input layer, a differential gene expression input layer, an in vitro high-throughput test data input layer, and a two-dimensional convolutional neural network layer;

[0017] The above-mentioned training sample data set is used to train the initial model, and the chemical genotoxicity prediction model obtained includes:

[0018] Input the atomic feature matrix and the connection relationship matrix in the training sample data set into the chemical structure input layer, and output a first vector;

[0019] Inputting the differential gene expression matrix in the training sample data set into the differential gene expression input layer, and outputting a second vector;

[0020] Inputting the in vitro high-throughput test matrix in the training sample data set into the in vitro high-throughput test data input layer, and outputting a third vector;

[0021] Concatenate the first vector, the second vector, and the third vector to form a fourth vector;

[0022] Inputting the fourth vector into the two-dimensional convolutional neural network layer, and outputting a predicted value of genotoxicity corresponding to the chemical;

[0023] The initial model is optimized according to the predicted values ​​and the label information of the chemical to obtain a genetic toxicity prediction model for the chemical.

[0024] According to an embodiment of the present disclosure, the above-mentioned training sample data set is used to train the initial model to obtain a chemical genotoxicity prediction model, which includes:

[0025] The initial model is trained using the training sample data set and the preset m sets of hyperparameters to obtain m genotoxicity prediction models for chemicals to be verified, where m ≥ 1;

[0026] The area under the receiver operating characteristic curve is used to evaluate the m genetic toxicity prediction models of the chemicals to be verified, and m evaluation values ​​are obtained;

[0027] Determine the maximum evaluation value among the above m evaluation values;

[0028] The hyperparameter set corresponding to the maximum evaluation value is determined as the final hyperparameter of the chemical genotoxicity prediction model.

[0029] According to an embodiment of the present disclosure, the above training method further includes:

[0030] Before training the initial model using the training sample data set and the preset m sets of hyperparameters, for each of the multiple hyperparameters in the initial model, determining z candidate values ​​corresponding to the hyperparameter, where z ≥ 1;

[0031] The z candidate values ​​corresponding to the above hyperparameters are combined with the z candidate values ​​corresponding to other hyperparameters to obtain m sets of hyperparameter sets.

[0032] According to an embodiment of the present disclosure, the initial model is trained using the training sample data set and the preset m sets of hyperparameters to obtain m genotoxicity prediction models for chemicals to be verified, including:

[0033] For each of the m groups of hyperparameters, the initial model is trained using k-fold cross validation to obtain k genotoxicity prediction models for chemicals to be verified, where k ≥ 1;

[0034] The m genotoxicity prediction models of the chemicals to be verified are evaluated using the area under the receiver operating characteristic curve, and the m evaluation values ​​obtained include:

[0035] The k genotoxicity prediction models of the chemicals to be verified are evaluated using the area under the receiver operating characteristic curve to obtain k initial evaluation values;

[0036] Based on the above k initial evaluation values, an evaluation mean of the above k initial evaluation values ​​is determined, wherein the above evaluation mean is the evaluation value of the genetic toxicity prediction model for the chemical to be verified corresponding to the above hyperparameters.

[0037] Another aspect of the present disclosure provides a method for predicting the genotoxicity of a chemical, comprising:

[0038] Obtaining data on the chemical to be predicted, wherein the data on the chemical to be predicted includes chemical coding, differential gene expression, and in vitro high-throughput testing data;

[0039] Generating a test sample data set based on the data of the chemical to be predicted, wherein the test sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, and an in vitro high-throughput test matrix;

[0040] Inputting the test sample data set into a chemical genotoxicity prediction model and outputting a predicted value, wherein the chemical genotoxicity prediction model is trained by the training method for the chemical genotoxicity prediction model;

[0041] The above predicted values ​​are analyzed to obtain the analysis results.

[0042] According to an embodiment of the present disclosure, the above prediction method further includes:

[0043] If the above analysis results show that the above predicted value is greater than or equal to the preset threshold, it is determined that the above chemical to be predicted has genotoxicity;

[0044] If the above analysis results show that the above predicted value is less than the above preset threshold, it is determined that the above chemical to be predicted does not have genotoxicity;

[0045] The method for determining the preset threshold includes:

[0046] Inputting a sample data set generated from sample data of a plurality of chemicals with known genotoxicity into the chemical genotoxicity prediction model to obtain a sample prediction value corresponding to each of the chemicals with known genotoxicity;

[0047] Calculate the true positive rate and false positive rate based on the sample prediction values ​​and the active labels corresponding to the chemicals with known genotoxicity;

[0048] With the false positive rate on the x-axis and the true positive rate on the y-axis, the receiver operating characteristic curve was obtained;

[0049] The point where the true positive rate relative to the false positive rate changes the most in the receiver operating characteristic curve is determined, and the predicted value corresponding to the point where the change rate is the most is used as the preset threshold.

[0050] Another aspect of the present disclosure provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the training method or prediction method.

[0051] Another aspect of the present disclosure further provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to perform the above-mentioned training method or prediction method.

[0052] Another aspect of the present disclosure further provides a computer program product, including a computer program, which implements the above-mentioned training method or prediction method when executed by a processor.

[0053] According to the embodiments of the present disclosure, this solution obtains the chemical coding, differential gene expression and in vitro high-throughput test data of chemicals with known genotoxicity, and generates a training sample data set including an atomic feature matrix, a connection relationship matrix, a differential gene expression matrix, an in vitro high-throughput test matrix and label information of the above chemicals based on the above data; then the initial model is trained using the training sample data set to obtain a chemical genotoxicity prediction model. Therefore, the chemical genotoxicity prediction model constructed by this solution can directly read the atomic feature matrix and connection relationship matrix calculated from the chemical chemical structure, and automatically capture the chemical structure by considering the interaction between atoms in the molecule. It does not require artificially defined features such as molecular fingerprints or descriptors, saving time for molecular descriptor calculation and selection, and improving the accuracy of the model. At the same time, the chemical genotoxicity prediction model provided by the embodiments of the present disclosure can simultaneously capture the chemical structural characteristics, differential gene expression characteristics and the interactive relationship of in vitro high-throughput testing, overcoming the problem that chemical genotoxicity prediction cannot accurately predict screening by relying solely on chemical structure.

[0054] According to the embodiments of the present disclosure, the chemical genotoxicity prediction model provided is suitable for large-scale chemical genotoxicity screening. The method is fast, simple, and efficient, and has broad application prospects in fields such as chemical risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above contents and other objects, features and advantages of the present disclosure will become more apparent through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0056] Figure 1 A flowchart of a method for training a chemical genotoxicity prediction model according to an embodiment of the present disclosure is schematically shown;

[0057] Figure 2 A flowchart schematically illustrates a method for determining final hyperparameters of a chemical genotoxicity prediction model according to an embodiment of the present disclosure;

[0058] Figure 3 The flowchart of the method for predicting the genotoxicity of chemicals according to an embodiment of the present disclosure is schematically shown;

[0059] Figure 4 A flowchart of a method for predicting the genotoxicity of chemicals according to another embodiment of the present disclosure is schematically shown;

[0060] Figure 5 Schematic diagram of a hybrid convolutional neural network model is shown;

[0061] Figure 6 The figure schematically shows chemical differential gene expression data and in vitro high-throughput test data according to another embodiment of the present disclosure. DETAILED DESCRIPTION

[0062] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the detailed description below, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0063] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise," "include," etc. used herein indicate the presence of the features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0064] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0065] When expressions such as "at least one of A, B and C, etc." are used, they should generally be interpreted in accordance with the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0066] At present, there are many quantitative structure-activity relationship (QSAR) models based on the molecular fingerprints or molecular descriptors of compounds, and they have shown excellent predictive ability in chemical toxicity assessment. The use of this method has improved the efficiency of chemical genotoxicity screening to a certain extent and has become an important tool for chemical risk assessment. However, due to the complexity of the structure of genotoxic substances, small changes in chemical structure may bring about huge changes in toxicity. Models based only on structural features have shown problems such as weak generalization ability in practical applications. Molecular descriptors or molecular fingerprints are sparse for chemical representation. If only relying on artificially defined molecular descriptors or molecular fingerprints, it may cause the loss of molecular structure information, and redundant features may also affect the predictive performance of the model.

[0067] With the development of high-throughput screening technology, the rapid accumulation of large-scale biological data can provide more information for the risk assessment of chemicals. The characteristic changes of high-throughput biological data have been used as indicators to reflect the underlying mechanisms of diseases and are related to chemical perturbations of the entire biological system. Therefore, although many studies have used high-throughput biological data for toxicity prediction and risk assessment, most of them focus on the analysis of a single type of high-throughput data and do not explore the synergy and interaction between biological data and structural features. In fact, structural features and biological data have complementary effects. In toxicity prediction and risk assessment, the comprehensive analysis method of coupling chemical structure features and biological data can provide more effective information.

[0068] In the process of implementing the present disclosure, it was found that machine learning prediction models can be used to evaluate genetic toxicity caused by chemical exposure. Although machine learning prediction models can improve prediction efficiency, their prediction accuracy is poor because they are based only on chemical structure or single high-throughput biological data.

[0069] In view of this, the present disclosure addresses the above technical problems and provides a chemical genotoxicity prediction model based on chemical structure characteristics, differential gene expression characteristics and in vitro high-throughput testing. The model can simultaneously capture the interactive relationship between chemical structure characteristics, differential gene expression characteristics and in vitro high-throughput testing. Compared with other models, the prediction model provided by the present disclosure has relatively excellent prediction performance.

[0070] Specifically, an embodiment of the present disclosure provides a method for training a chemical genotoxicity prediction model, comprising:

[0071] Obtaining raw data on chemicals with known genotoxicity, wherein the raw data includes chemical codes of the chemicals, differential gene expression data of the chemicals, and in vitro high-throughput test data of the chemicals; generating a training sample data set based on the raw data, wherein the training sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, an in vitro high-throughput test matrix, and label information of the chemicals; and training an initial model using the training sample data set to obtain a chemical genotoxicity prediction model.

[0072] According to the embodiments of the present disclosure, the basic principle can be understood as inputting the atomic feature matrix representing the atomic features obtained by calculating the chemical and the connection relationship matrix representing the connection relationship between atoms into the chemical structure input layer (the chemical structure input layer can be a one-dimensional graph convolution layer) to obtain a first vector representing the chemical structure features; inputting the differential gene matrix into the differential gene input layer (the differential gene input layer can be a fully connected layer) to obtain a second vector representing the differential gene expression features; inputting the in vitro high-throughput test matrix into the in vitro high-throughput test data input layer (the in vitro high-throughput test data input layer can be a fully connected layer) to obtain a third vector representing the in vitro high-throughput test features; then, the first vector, the second vector and the third vector are spliced ​​and input into a two-dimensional convolutional neural network to establish a mathematical function relationship between them and the molecular genetic toxicity, thereby realizing the prediction of chemical genetic toxicity.

[0073] This disclosure proposes for the first time a method for predicting chemical genotoxicity based on chemical structure, differentially expressed genes, and high-throughput in vitro testing data. Prior to this disclosure, there had been no reports of using hybrid convolutional neural network models for genotoxicity prediction using chemical structure and multiple biological data as input.

[0074] Figure 1 The flowchart of the method for training a chemical genotoxicity prediction model according to an embodiment of the present disclosure is schematically shown.

[0075] like Figure 1 As shown, the training method of the chemical genotoxicity prediction model of this embodiment includes operations S101 to S103.

[0076] In operation S101 , raw data on chemicals with known genotoxicity are acquired, wherein the raw data include chemical codes of the chemicals, differential gene expression data of the chemicals, and in vitro high-throughput test data of the chemicals.

[0077] According to an embodiment of the present disclosure, the raw data of chemicals may include, for example, raw data of multiple chemicals.

[0078] According to an embodiment of the present disclosure, the original data of chemicals with known genotoxicity may be obtained from literature or public databases, for example.

[0079] According to an embodiment of the present disclosure, the chemical code of the chemical may include the SMILES code of the chemical.

[0080] According to an embodiment of the present disclosure, the differential gene expression of a chemical refers to the amount of differential gene expression change in tens of thousands of genes in a cell line treated with the chemical. For example, in another embodiment of the present disclosure, Figure 6 As shown, only the expression changes of 295 marker genes related to genotoxicity of chemicals in human hepatocellular carcinoma (HepG2) cell lines were retained and expressed as z scores.

[0081] According to an embodiment of the present disclosure, the in vitro high-throughput test data of a chemical refers to the activity value or response data of the chemical in thousands of biological test endpoints in the ToxCast project and the Tox21 project. For example, in another embodiment of the present disclosure, Figure 6 As shown in Figure 2, we only retained the response data of chemicals in 434 biological test endpoints related to genotoxicity. The response data were expressed as binary variables, where "1" indicated that the chemical responded in the biological test endpoint and "0" indicated that the chemical did not respond in the biological test endpoint.

[0082] In operation S102 , a training sample data set is generated based on the original data, wherein the training sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, an in vitro high-throughput test matrix, and label information of the chemicals.

[0083] According to an embodiment of the present disclosure, the atomic feature matrix may include, for example, an M×N digital matrix constructed from N feature vectors of M atoms of a chemical.

[0084] According to an embodiment of the present disclosure, an atomic feature vector is represented by a digital vector of length n, where 50≤n≤75. The atomic feature represents chemical and topological properties such as atomic type and hybridization degree.

[0085] According to an embodiment of the present disclosure, the connection relationship matrix may include, for example, an M×M digital matrix consisting of connection relationships between M atoms of a chemical.

[0086] According to an embodiment of the present disclosure, the connectivity relationships between atoms in the connectivity matrix include using a value of "1" to indicate the presence of a chemical bond between two atoms, and a value of "0" to indicate the absence of a chemical bond between two atoms. If all the values ​​on the diagonal of the connectivity matrix are "1," it means that each atom is connected to itself, forming a closed loop.

[0087] According to an embodiment of the present disclosure, the label information of a chemical may include, for example, the genotoxicity category of the chemical, wherein the genotoxicity category is represented by a binary category, where "1" indicates genotoxicity and "0" indicates non-genotoxicity.

[0088] In operation S103 , the initial model is trained using the training sample data set to obtain a chemical genotoxicity prediction model.

[0089] According to the embodiments of the present disclosure, this solution obtains the chemical coding, differential gene expression data and in vitro high-throughput test data of chemicals with known genotoxicity, and generates a training sample data set based on the above data. The training sample data set includes an atomic feature matrix, a connection relationship matrix, a differential gene expression matrix, an in vitro high-throughput test matrix and label information of the above chemicals; and then uses the training sample data set to train the initial model to obtain a chemical genotoxicity prediction model. Therefore, the chemical genotoxicity prediction model constructed by this solution can directly read the atomic feature matrix and connection relationship matrix calculated from the chemical chemical structure, and automatically capture the chemical structure by considering the interaction between atoms in the molecule. It does not require artificially defined features such as molecular fingerprints or descriptors, saving time for molecular descriptor calculation and selection, and improving the accuracy of the model. At the same time, the chemical genotoxicity prediction model provided by the embodiments of the present disclosure can simultaneously capture the chemical structural characteristics, differential gene expression characteristics and the interactive relationship of in vitro high-throughput testing, overcoming the problem that chemical genotoxicity prediction cannot accurately predict screening by relying solely on chemical structure.

[0090] According to an embodiment of the present disclosure, the above-mentioned generation of a training sample data set based on the above-mentioned original data includes: converting the above-mentioned chemical code into a two-dimensional graphic structure representing the spatial structure of the above-mentioned chemical, wherein atoms in the above-mentioned two-dimensional graphic structure are represented by points and chemical bonds are represented by lines; according to the above-mentioned two-dimensional graphic structure, determining the number of atoms, atomic feature vectors and connection relationships between atoms corresponding to the above-mentioned chemical code; constructing the above-mentioned atomic feature matrix based on the above-mentioned atomic number and the above-mentioned atomic feature vector; and constructing the above-mentioned connection relationship matrix based on the above-mentioned atomic number and the connection relationship between the above-mentioned atoms.

[0091] According to an embodiment of the present disclosure, for chemical A, the number of atoms of chemical A is determined to be M, the number of atomic features is N, and the atomic feature vector is represented by a digital vector of length n, where n is the number of atomic features, 50≤n≤75; then an M×N atomic feature matrix is ​​constructed based on the number of atoms M, the number of atomic features N, and the atomic feature vector n; and an M×M connection relationship matrix is ​​constructed based on the number of atoms M and the connection relationship between atoms.

[0092] According to an embodiment of the present disclosure, the above-mentioned initial model includes a chemical structure input layer, a differential gene expression input layer, an in vitro high-throughput test data input layer and a two-dimensional convolutional neural network layer.

[0093] According to an embodiment of the present disclosure, the initial model is trained using the above-mentioned training sample data set to obtain a chemical genetic toxicity prediction model, including: inputting the above-mentioned atomic feature matrix and the above-mentioned connection relationship matrix in the above-mentioned training sample data set into the above-mentioned chemical structure input layer, and outputting a first vector; inputting the above-mentioned differential gene expression matrix in the above-mentioned training sample data set into the above-mentioned differential gene expression input layer, and outputting a second vector; inputting the above-mentioned in vitro high-throughput test matrix in the above-mentioned training sample data set into the above-mentioned in vitro high-throughput test data input layer, and outputting a third vector; splicing the above-mentioned first vector, the above-mentioned second vector and the above-mentioned third vector to form a fourth vector; inputting the above-mentioned fourth vector into the above-mentioned two-dimensional convolutional neural network layer, and outputting a predicted value of the genetic toxicity corresponding to the above-mentioned chemical; optimizing the above-mentioned initial model according to the above-mentioned predicted value and the label information of the above-mentioned chemical to obtain the above-mentioned chemical genetic toxicity prediction model.

[0094] According to an embodiment of the present disclosure, the chemical structure input layer in the initial model is followed by three one-dimensional graph convolution layers, and the i-th graph convolution layer contains unit i nodes, the convolution step length is stride j ; After the graph convolution layer, the maximum pooling layer is connected. After the atomic feature matrix and connection relationship matrix in the training sample data set are input into the chemical structure input layer, the first vector is output after passing through three layers of one-dimensional graph convolution layers and maximum pooling layers in sequence. The first vector is a one-dimensional vector of length u.

[0095] According to an embodiment of the present disclosure, two fully connected layers are connected after the differential gene expression input layer, and after the differential gene expression matrix in the above-mentioned training sample data set is input into the above-mentioned differential gene expression input layer, a second vector is output after passing through two fully connected layers, wherein the second vector is a one-dimensional vector with a length of q.

[0096] According to an embodiment of the present disclosure, two fully connected layers are connected after the in vitro high-throughput test data input layer, and after the above-mentioned in vitro high-throughput test matrix in the above-mentioned training sample data set is input into the above-mentioned in vitro high-throughput test data input layer, a third vector is output after passing through two fully connected layers, wherein the third vector is a one-dimensional vector with a length of p.

[0097] According to an embodiment of the present disclosure, the above-mentioned two-dimensional convolutional neural network layer includes three two-dimensional convolutional layers, three two-dimensional maximum pooling layers and two fully connected layers. The first fully connected layer in the two fully connected layers includes unit lnodes, serving as the input and output of the two-dimensional convolutional neural network layer, where the output layer outputs the prediction results of the chemical's genetic toxicity.

[0098] According to an embodiment of the present disclosure, the above-mentioned training sample data set is used to train the initial model to obtain the chemical genotoxicity prediction model, which includes: using the above-mentioned training sample data set and a preset m group of hyperparameter sets to train the above-mentioned initial model to obtain m chemical genotoxicity prediction models to be verified, wherein the above-mentioned m ≥ 1; using the area under the receiver operating characteristic curve to evaluate the above-mentioned m chemical genotoxicity prediction models to be verified to obtain m evaluation values; determining the maximum evaluation value among the above-mentioned m evaluation values; and determining the hyperparameter set corresponding to the above-mentioned maximum evaluation value as the final hyperparameter of the above-mentioned chemical genotoxicity prediction model.

[0099] According to an embodiment of the present disclosure, the area under the receiver operating characteristic curve (AUC-ROC) is used as a model evaluation indicator, and the hyperparameter corresponding to the maximum AUC-ROC value is used as the final hyperparameter of the chemical genotoxicity prediction model.

[0100] According to an embodiment of the present disclosure, each set of hyperparameters includes lr, dropout, λ, batchsize, where lr is the learning rate, dropout is the dropout coefficient, λ is the L2 weight decay regularization term parameter, and batchsize is the batch size, where dropout refers to randomly "temporarily discarding" a part of neuron nodes with a certain probability during the training of the deep network.

[0101] According to an embodiment of the present disclosure, the above-mentioned training method also includes: before training the above-mentioned initial model using the above-mentioned training sample data set and the preset m groups of hyperparameter sets, for each of the multiple hyperparameters in the above-mentioned initial model, determining z alternative values ​​corresponding to the above-mentioned hyperparameter, where z≥1; and combining the z alternative values ​​corresponding to the above-mentioned hyperparameters with the z alternative values ​​corresponding to other hyperparameters respectively to obtain m groups of hyperparameter sets.

[0102] According to an embodiment of the present disclosure, multiple alternative values ​​are set for each of the hyperparameters lr, dropout, λ, and batchsize, and different alternative values ​​of each hyperparameter are combined to obtain multiple sets of hyperparameter sets.

[0103] Figure 2 The flowchart of the method for determining the final hyperparameters of the chemical genotoxicity prediction model according to an embodiment of the present disclosure is schematically shown.

[0104] like Figure 2Said method includes operations S201 to S206.

[0105] In operation S201 , for each hyperparameter in a plurality of hyperparameter sets in an initial model, z candidate values ​​corresponding to the hyperparameter are determined, where z≧1.

[0106] In operation S202 , the z candidate values ​​corresponding to the hyperparameters are respectively combined with the z candidate values ​​corresponding to other hyperparameters to obtain m sets of hyperparameters.

[0107] In operation S203 , an initial model is trained using the training sample data set and m sets of hyperparameter sets to obtain m genotoxicity prediction models of chemicals to be verified, where m≧1.

[0108] In operation S204 , the m genetic toxicity prediction models of the chemicals to be verified are evaluated using the area under the receiver operating characteristic curve to obtain m evaluation values.

[0109] In operation S205 , a maximum evaluation value among the m evaluation values ​​is determined.

[0110] In operation S206 , the hyperparameter set corresponding to the maximum evaluation value is determined as the final hyperparameter of the chemical genotoxicity prediction model.

[0111] According to an embodiment of the present disclosure, the above-mentioned training of the above-mentioned initial model using the above-mentioned training sample data set and the preset m groups of hyperparameter sets to obtain m genetic toxicity prediction models for chemicals to be verified includes: for each group of hyperparameters in the m groups of hyperparameter sets, using k-fold cross-validation to train the above-mentioned initial model to obtain k genetic toxicity prediction models for chemicals to be verified, wherein k≥1; the above-mentioned evaluation of the above-mentioned m genetic toxicity prediction models for chemicals to be verified using the area under the receiver operating characteristic curve to obtain m evaluation values ​​includes: evaluating the above-mentioned k genetic toxicity prediction models for chemicals to be verified using the area under the receiver operating characteristic curve to obtain k initial evaluation values; and determining the evaluation mean of the above-mentioned k initial evaluation values ​​based on the above-mentioned k initial evaluation values, wherein the above-mentioned evaluation mean is the evaluation value of the genetic toxicity prediction model for chemicals to be verified corresponding to the above-mentioned hyperparameters.

[0112] According to an embodiment of the present disclosure, for example, the above-mentioned initial model is trained using five-fold cross-validation, including: for each set of hyperparameters in the m sets of hyperparameter sets, the training sample data set is randomly divided into five parts, one of the data is taken as a validation set, and the remaining four data are taken as training sets for training the initial model, and finally five genotoxicity prediction models of chemicals to be verified are obtained for each set of hyperparameters; the above-mentioned five genotoxicity prediction models of chemicals to be verified are evaluated using the area under the above-mentioned receiver operating characteristic curve to obtain five initial evaluation values; based on the above-mentioned five initial evaluation values, the evaluation mean of the above-mentioned five initial evaluation values ​​is determined, wherein the above-mentioned evaluation mean is the evaluation value of the genotoxicity prediction model of the chemical to be verified corresponding to the hyperparameter, thereby determining the evaluation value corresponding to one of the sets of hyperparameters; repeating the above-mentioned operations until m evaluation values ​​corresponding to the m sets of hyperparameter sets are determined; determining the largest evaluation value among the m evaluation values; and determining the hyperparameter set corresponding to the largest evaluation value as the final hyperparameter of the chemical genotoxicity prediction model.

[0113] According to an embodiment of the present disclosure, the area under the receiver operating characteristic curve is used to evaluate the above-mentioned five genotoxicity prediction models of chemicals to be verified, and the method for obtaining five initial evaluation values ​​includes: inputting the verification set into the genotoxicity prediction model of the chemical to be verified, outputting the predicted value and the genotoxicity label of the chemical (1 or 0), calculating the true positive (True Positive, TP), true negative (True Negative, TN), false positive (False Positive, FP), false negative (False Negative, FN), true positive rate and false positive rate, making a receiver operating characteristic curve and calculating the AUC-ROC value, and using the AUC-ROC value as the initial evaluation value.

[0114] in,

[0115] TPR = TP / (TP + FN);

[0116] FPR = FP / (FP+TN);

[0117] TP: indicates the number of samples in the validation set that were predicted to be genotoxicants and were actually genotoxicants;

[0118] FP: indicates the number of samples in the validation set that were predicted to be genotoxicants but were actually non-genotoxicants;

[0119] FN: indicates the number of samples in the validation set that were predicted to be non-genotoxic but were actually genotoxic;

[0120] TN: indicates the number of samples in the validation set that were predicted to be non-genotoxic and were actually non-genotoxic.

[0121] Figure 3The flowchart of the method for predicting the genotoxicity of chemicals according to an embodiment of the present disclosure is schematically shown.

[0122] like Figure 3 As shown, the method includes operations S301 to S304.

[0123] In operation S301 , data of a chemical to be predicted is obtained, wherein the data of the chemical to be predicted includes chemical coding, differential gene expression data, and in vitro high-throughput test data.

[0124] In operation S302 , a test sample data set is generated based on the data of the chemical to be predicted, wherein the test sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, and an in vitro high-throughput test matrix.

[0125] In operation S303 , the sample data set to be tested is input into a chemical genotoxicity prediction model, and a prediction value is output, wherein the chemical genotoxicity prediction model is trained by the training method for the chemical genotoxicity prediction model.

[0126] In operation S304, the predicted value is analyzed to obtain an analysis result.

[0127] According to an embodiment of the present disclosure, the above-mentioned prediction method also includes: when the above-mentioned analysis results show that the above-mentioned prediction value is greater than or equal to the preset threshold, determining that the above-mentioned chemical to be predicted has genotoxicity; when the above-mentioned analysis results show that the above-mentioned prediction value is less than the above-mentioned preset threshold, determining that the above-mentioned chemical to be predicted does not have genotoxicity; wherein, the method for determining the above-mentioned preset threshold includes: inputting a sample data set generated by sample data of multiple chemicals with known genotoxicity into the above-mentioned chemical genotoxicity prediction model to obtain a sample prediction value corresponding to each of the above-mentioned chemicals with known genotoxicity; calculating the true positive rate and the false positive rate based on the above-mentioned sample prediction values ​​and the active labels corresponding to the above-mentioned chemicals with known genotoxicity; obtaining a receiver operating characteristic curve with the false positive rate as the x-axis and the true positive rate as the y-axis; determining the point with the maximum change rate of the true positive rate relative to the false positive rate in the above-mentioned receiver operating characteristic curve, and taking the prediction value corresponding to the above-mentioned maximum change rate point as the above-mentioned preset threshold.

[0128] According to an embodiment of the present disclosure, calculating the true positive rate and the false positive rate based on the above-mentioned sample prediction values ​​and the activity labels corresponding to the above-mentioned known genotoxic chemicals includes: sorting the sample prediction values ​​from large to small, and calculating the true positive rate and the false positive rate based on the prediction value of each known chemical and the toxicity label corresponding to the chemical (1 indicates genotoxicity and 0 indicates non-genotoxicity).

[0129] Figure 4The flowchart of the method for predicting the genotoxicity of chemicals according to another embodiment of the present disclosure is schematically shown.

[0130] like Figure 4 As shown, the method includes operations S401 to S412.

[0131] In operation S401 , raw data of chemicals with known genotoxicity are obtained, wherein the raw data include chemical codes of the chemicals, differential gene expression of the chemicals, and in vitro high-throughput test data of the chemicals.

[0132] In operation S402 , a training sample data set is generated according to the original data, wherein the training sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, an in vitro high-throughput test matrix, and label information of chemicals.

[0133] In operation S403 , the training sample data set is divided into a training set and a test set.

[0134] In operation S404 , the initial model is trained using the training set using a five-fold crossover method to obtain a genotoxicity prediction model for the chemical to be verified.

[0135] In operation S405 , the chemical genotoxicity prediction model to be verified is verified using the test set to obtain the chemical genotoxicity prediction model.

[0136] In operation S406 , data of the chemical to be predicted is obtained, wherein the data of the chemical to be predicted includes chemical codes, differential gene expression data, and in vitro high-throughput test data.

[0137] In operation S407 , a test sample data set is generated according to the data of the chemical to be predicted, wherein the test sample data set includes an atomic feature matrix, a connection relationship matrix, a differential gene expression matrix, and an in vitro high-throughput test matrix.

[0138] In operation S408 , the sample data set to be tested is input into the chemical genotoxicity prediction model.

[0139] In operation S409 , calculations are performed on the sample data to be tested to determine a predicted value.

[0140] In operation S410, it is determined whether the predicted value is greater than or equal to a preset threshold. If the predicted value is greater than or equal to the preset threshold, operation S411 is performed; if the predicted value is less than the preset threshold, operation S412 is performed.

[0141] In operation S411 , a prediction result that the chemical to be predicted has genotoxicity is output.

[0142] In operation S412 , a prediction result indicating that the chemical to be predicted does not have genotoxicity is output.

[0143] According to the embodiments of the present disclosure, the chemical genotoxicity prediction model can directly read molecular features calculated from the chemical's chemical structure, automatically capturing the chemical's structure by considering the interactions between atoms within the molecule. It also simultaneously captures the interactions between chemical structural features, differential gene expression characteristics, and in vitro high-throughput testing, overcoming the difficulty in accurately predicting and screening chemical genotoxicity simply by relying on chemical structure. Furthermore, the disclosed prediction model is suitable for large-scale chemical genotoxicity screening, offering a rapid, simple, and efficient method with broad application prospects in fields such as chemical risk assessment.

[0144] The technical solution of the present disclosure is further illustrated below through specific embodiments.

[0145] Example

[0146] (1) Acquisition and preprocessing of chemical data

[0147] The original data of 261 chemicals with genotoxic binary labels in the CRCGN dataset were downloaded. The original data included the compound structures and their differential gene expression data in human hepatocellular carcinoma cell lines, and in vitro high-throughput test data matching the 261 chemicals were obtained. Finally, 244 chemicals and their SMILES codes, differential gene expression data and in vitro high-throughput test data were obtained, of which 92 chemicals were genotoxic and 152 were not genotoxic.

[0148] (2) Chemical structure conversion into atomic feature matrix and connection relationship matrix

[0149] The atomic feature matrix representing the atomic features obtained by chemical structure calculation is combined with the connection relationship matrix representing the connection relationship between atoms, wherein the length n of the atomic feature vector is set to 75.

[0150] (3) Training and hyperparameter determination of hybrid convolutional neural network models

[0151] The raw data of chemicals were randomly divided into training set and test set in a ratio of 4:1. The training set was used to determine the hyperparameters and establish the model of the hybrid convolutional neural network model, and the test set was used to evaluate the model's predictive ability.

[0152] Figure 5 Schematic diagram of a hybrid convolutional neural network model is shown schematically.

[0153] like Figure 5 As shown in FIG, the hybrid convolutional neural network model includes a chemical structure input layer, a differential gene expression data input layer, and an in vitro high-throughput test data input layer.

[0154] After the chemical structure input layer, three one-dimensional graph convolution layers are connected. After each graph convolution layer, the linear rectifier function (ReLU) is used as the activation function to convert the linear features in the neural network into nonlinear features, and batch normalization is performed to keep the input of each layer of the neural network in the same distribution. Dropout is then used to prevent overfitting. The i-th graph convolution layer contains unit i nodes, where unit i is [256, 256, 256], and the convolution step size is stride j , where stride j is [1, 1, 1]; then the maximum pooling layer is connected to obtain the digital vector U representing the structural characteristics of the chemical.

[0155] A fully connected layer with 256 nodes is connected to the input layer of the differential gene expression data and the in vitro high-throughput test data. Then, the bitangent sine curve function (tanh) is used as the activation function. After using dropout, the fully connected layer with 100 nodes and the activation function of the linear rectifier function (ReLU) outputs the digital vector Q representing the differential gene expression characteristics and the digital vector P representing the in vitro high-throughput test characteristics.

[0156] The above vectors U, Q and P are concatenated and input into a two-dimensional convolutional neural network to output the predicted value of the genotoxicity corresponding to the chemical.

[0157] The convolutional neural network is divided into 8 layers. The first layer of the convolutional neural network is a two-dimensional convolution layer, which contains 30 convolution kernels with a size of 1×150, a convolution step of 1, and uses the ReLU function as the activation function; followed by the first layer of the two-dimensional maximum pooling layer, with a pooling kernel size of 1×2.

[0158] The second and third layers of the convolutional neural network are similar to the first layer, but with slight differences: the number of convolution kernels in the second and third layers is 10 and 5, respectively, and the size is 1×5. The second and third 2D max pooling layers are also similar to the first layer, placed after the corresponding convolution layers, except that the pooling kernel size is 1×3. The final 2D max pooling layer is followed by dropout to prevent overfitting.

[0159] The first fully connected layer in the fully connected layer contains unit l nodes, where unit l is [300, 100], and the activation functions are bitangent sine function (tanh) and sigmoid function, which serve as the input and output of the two-dimensional convolutional neural network, respectively. The output layer outputs the predicted value s of the chemical's genotoxicity, which ranges from 0 to 1.

[0160] To prevent the model from overfitting, an L2 weight decay regularization term is added to each layer of the two-dimensional convolutional layer, and the parameter is set to λ.

[0161] The adaptive moment estimation optimizer (Adam) method is used as the trainer during training, the learning rate is set to lr, 100 iterations are performed each time, and the model parameters of each iteration are saved.

[0162] In the training set, when using the grid screening method to obtain the optimal hyperparameters, candidate values ​​for the learning rate lr were [0.01, 0.001, 0.0001], the batch size batchsize ...

[0163] (4) Determination of the preset threshold of the optimal hybrid convolutional neural model

[0164] Based on the final hyperparameters obtained by the method described in step (3), the optimal hybrid convolutional neural prediction model is trained in the training set, and the test set is predicted to obtain the predicted activity values ​​of the chemicals in the test set. The receiver operating characteristic curve (ROC curve) is drawn based on the predicted activity values ​​combined with their true genotoxicity labels. The predicted value corresponding to the maximum point t of the TPR relative to FPR change rate in the ROC curve is used as the preset threshold s for determining genotoxicity t , this embodiment presets a threshold value s t It is 0.740.

[0165] (5) Comparison of prediction performance with other existing machine learning models

[0166] In order to better reflect the excellent performance of the hybrid convolutional neural network model described in the present embodiment in the genotoxicity prediction method, the prediction model of the present embodiment is compared with the performance of similar machine learning models of the support vector machine based on PubChem molecular fingerprint and random forest based on MACCS molecular fingerprint. The prediction results of the test set are calculated to evaluate the predictive ability of the model by calculating true positive (TP), true negative (TN), false positive (FP), false negative (FN), true positive rate (TPR), precision (Precision), F1 score (F1-Score), Matthews correlation coefficient (MCC) and AUC-ROC comprehensive evaluation model.

[0167] in,

[0168] TPR=TP / (TP+FN)

[0169] Precision = TP / (TP+FP)

[0170] F1 Score=2*(Precision*TPR) / (Precision+TPR)

[0171]

[0172] Machine learning models were established on the same dataset, and the prediction performance evaluation indicators of each model are shown in Table 1.

[0173] Table 1

[0174]

[0175] As shown in Table 1, the AUC-ROC, F1-Score and MCC of the prediction model provided in this embodiment are 0.889, 0.811 and 0.697 respectively, which shows that the hybrid convolutional neural network model proposed in the embodiment of the present disclosure has excellent predictive performance in genotoxicity. The AUC-ROC, F1-Score and MCC of the method provided in the embodiment of the present disclosure on the test set are much higher than those of the other two similar methods, which shows that the method provided in the embodiment of the present disclosure has a stronger ability to distinguish between genotoxic substances and non-genotoxic substances than the other two similar methods, and has better generalization ability. At the same time, the method provided in the embodiment of the present disclosure introduces the differential gene expression data of chemicals and in vitro high-throughput test data compared with the other two similar methods, indicating that there is a complementary relationship between the chemical structure of chemicals, differential genes and in vitro high-throughput test data in the prediction of genotoxicity, and the prediction method coupling chemical structure with multiple biological data has better performance. In short, the prediction model provided in the embodiment of the present disclosure has superior predictive ability for chemical genotoxicity.

[0176] In summary, the chemical genotoxicity prediction and screening model based on a hybrid convolutional neural network established in the embodiments of the present disclosure captures the interactive relationship between chemical structural characteristics, differential gene expression characteristics, and in vitro high-throughput testing, and can more accurately predict the genotoxicity of chemicals.

[0177] Those skilled in the art will appreciate that the features described in the various embodiments and / or claims of this disclosure may be combined and / or coupled in various ways, even if such combinations and / or couplings are not explicitly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this disclosure may be combined and / or coupled in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or couplings are intended to fall within the scope of this disclosure.

[0178] The embodiments of the present disclosure are described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present disclosure. Although each embodiment has been described separately above, this does not mean that the measures in each embodiment cannot be used in combination to advantage. The scope of the present disclosure is defined by the appended claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present disclosure.

Claims

1. A method for training a chemical genotoxicity prediction model, comprising: Acquiring raw data on chemicals with known genotoxicity, wherein the raw data includes chemical codes of the chemicals, differential gene expression data of the chemicals, and in vitro high-throughput testing data of the chemicals; Generating a training sample data set according to the original data, wherein the training sample data set includes an atomic feature matrix, a connectivity matrix, a differential gene expression matrix, an in vitro high-throughput test matrix, and label information of the chemical; Using the training sample data set to train an initial model to obtain a chemical genotoxicity prediction model; The initial model includes a chemical structure input layer, a differential gene input layer, an in vitro high-throughput test data input layer, and a two-dimensional convolutional neural network layer; The method of using the training sample data set to train the initial model to obtain a chemical genotoxicity prediction model includes: Inputting the atomic feature matrix and the connection relationship matrix in the training sample data set into the chemical structure input layer, and outputting a first vector; Inputting the differential gene expression matrix in the training sample data set into the differential gene expression input layer, and outputting a second vector; Inputting the in vitro high-throughput test matrix in the training sample data set into the in vitro high-throughput test data input layer, and outputting a third vector; concatenate the first vector, the second vector, and the third vector to form a fourth vector; Inputting the fourth vector into the two-dimensional convolutional neural network layer, and outputting a predicted value of genetic toxicity corresponding to the chemical; The initial model is optimized according to the predicted value and the label information of the chemical to obtain the chemical genotoxicity prediction model.

2. The method according to claim 1, wherein Generating a training sample data set according to the original data includes: Converting the chemical code into a two-dimensional graphic structure representing the spatial structure of the chemical, wherein atoms in the two-dimensional graphic structure are represented by points and chemical bonds are represented by lines; Determining, based on the two-dimensional graphic structure, the number of atoms, atomic feature vectors, and connection relationships between atoms corresponding to the chemical code; Constructing the atomic feature matrix according to the number of atoms and the atomic feature vectors; The connection relationship matrix is ​​constructed according to the number of atoms and the connection relationship between the atoms.

3. The method according to claim 2, wherein: The atomic feature vector is represented by a digital vector of length n, where n is the number of atomic features, and 50≤n≤75.

4. The method according to claim 3, wherein: The connection relationship between the atoms includes using a value "1" to indicate that there is a chemical bond between the two atoms, and a value "0" to indicate that there is no chemical bond between the two atoms.

5. The method according to claim 1, wherein The method of using the training sample data set to train the initial model to obtain a chemical genotoxicity prediction model includes: The initial model is trained using the training sample data set and the preset m sets of hyperparameters to obtain m genotoxicity prediction models of chemicals to be verified, wherein m ≥ 1; Evaluating the m genetic toxicity prediction models of the chemicals to be verified using the area under the receiver operating characteristic curve to obtain m evaluation values; Determining the maximum evaluation value among the m evaluation values; The hyperparameter set corresponding to the maximum evaluation value is determined as the final hyperparameter of the chemical genotoxicity prediction model.

6. The method according to claim 5, further comprising: Before training the initial model using the training sample data set and the preset m sets of hyperparameters, determining, for each hyperparameter of the multiple hyperparameter sets in the initial model, z candidate values ​​corresponding to the hyperparameter, wherein z ≥ 1; The z candidate values ​​corresponding to the hyperparameters are respectively combined with the z candidate values ​​corresponding to other hyperparameters to obtain m sets of hyperparameter sets.

7. The method according to claim 6, wherein: The method of training the initial model using the training sample data set and the preset m sets of hyperparameters to obtain m genotoxicity prediction models for chemicals to be verified includes: For each set of hyperparameters in the m sets of hyperparameters, the initial model is trained using k-fold cross validation to obtain k genotoxicity prediction models of chemicals to be verified, where k ≥ 1; The m evaluation values ​​obtained by evaluating the m genetic toxicity prediction models of the chemicals to be verified using the area under the receiver operating characteristic curve include: Using the area under the receiver operating characteristic curve, the k genetic toxicity prediction models of the chemicals to be verified are evaluated to obtain k initial evaluation values; Based on the k initial evaluation values, an evaluation mean of the k initial evaluation values ​​is determined, wherein the evaluation mean is the evaluation value of the genetic toxicity prediction model for the chemical to be verified corresponding to the hyperparameter.

8. A method for predicting the genetic toxicity of a chemical, comprising: Acquiring data on the chemical to be predicted, wherein the data on the chemical to be predicted includes chemical codes, differentially expressed genes, and in vitro high-throughput testing data; Generating a test sample data set according to the data of the chemical to be predicted, wherein the test sample data set includes an atomic feature matrix, a connectivity relationship matrix, a differential gene matrix, and an in vitro high-throughput test matrix; Inputting the test sample data set into a chemical genotoxicity prediction model and outputting a predicted value, wherein the chemical genotoxicity prediction model is trained by the chemical genotoxicity prediction model training method according to any one of claims 1 to 7; The predicted value is analyzed to obtain an analysis result.

9. The method according to claim 8, further comprising: If the analysis result shows that the predicted value is greater than or equal to a preset threshold, determining that the chemical to be predicted has genotoxicity; If the analysis result shows that the predicted value is less than the preset threshold, determining that the chemical to be predicted does not have genotoxicity; The method for determining the preset threshold includes: Inputting a sample data set generated from sample data of a plurality of chemicals with known genotoxicity into the chemical genotoxicity prediction model to obtain a sample prediction value corresponding to each of the chemicals with known genotoxicity; Calculating a true positive rate and a false positive rate based on the sample prediction value and the active signature corresponding to the known genotoxic chemical; With the false positive rate on the x-axis and the true positive rate on the y-axis, the receiver operating characteristic curve was obtained; The point where the true positive rate has the largest change rate relative to the false positive rate in the receiver operating characteristic curve is determined, and the predicted value corresponding to the point where the change rate has the largest change rate is used as the preset threshold.

Images