Knowledge-driven machine learning method for screening reproductive toxicity of chemical substance zebrafish

Abstract

The invention relates to a method for screening reproductive toxicity of chemical substance zebrafish by knowledge-driven machine learning, which belongs to the technical field of machine learning and computational toxicology, and is characterized by comprising the following steps: S1, constructing a reproductive toxicity prediction model; s1.1, constructing an in-vivo test original data set and an in-vitro test original data set of zebrafish reproductive toxic compounds to form a compound original data set; s1.2, calculating molecular characteristics of the compound original data set to form a compound characteristic data set; the molecular features comprise functional connectivity fingerprints and molecular descriptors; s1.3, inputting the compound characteristic data set into a deep neural network to obtain a reproductive toxicity prediction model; and S2, calculating molecular characteristics of a to-be-detected sample, and inputting the molecular characteristics into the reproductive toxicity prediction model to obtain a toxicity prediction result. According to the method, the limitation of traditional dependence on low-flux animal experiments is overcome, the compound screening efficiency is remarkably improved, and meanwhile, the evaluation accuracy is kept.

Description

Technical Field

[0001] The present application relates to the technical fields of machine learning and computational toxicology, and in particular to a method for screening the reproductive toxicity of chemicals in zebrafish using knowledge-driven machine learning. Background Art

[0002] Traditional methods for testing reproductive toxicity of chemicals usually rely on low-throughput animal experiments, which are not only costly and time-consuming, but also involve the use of a large number of animals. For example, in the Endocrine Disruptor Screening Program (EDSP) implemented by the U.S. Environmental Protection Agency (EPA), zebrafish and other animals are usually used for biological experiments. Although such experiments can accurately assess the effects of chemicals on the reproductive system to a certain extent, with the sharp increase in the number of compounds, this method faces the challenge of being difficult to handle high-throughput chemical screening. Especially when thousands of compounds need to be screened quickly, traditional animal experiments cannot complete these tasks under reasonable time and resource conditions, resulting in extremely low screening efficiency.

[0003] With the development of modern computer technology, quantitative structure-activity relationship (QSAR) models based on machine learning have become a potential alternative to animal experiments. These models construct toxicity prediction models through the structural feature information of chemicals (such as molecular fingerprints and molecular descriptors). However, traditional QSAR models mainly rely on chemical structure information and lack a deep understanding of the toxic mechanism of chemicals. Although this model can improve the screening speed, it also has many shortcomings. First, due to the lack of in-depth consideration of biological mechanisms (such as receptor binding, gene expression regulation, etc.), the model has poor interpretability and it is difficult to provide researchers with effective mechanism analysis. Secondly, traditional QSAR models may have overfitting problems due to excessive reliance on chemical structure information. Especially when dealing with complex biological data, this overfitting may cause the model to lose accuracy in predicting unknown compounds. Summary of the invention

[0004] Based on the needs in the above-mentioned fields, the present invention constructs a machine learning model for screening the reproductive toxicity of zebrafish chemicals and predicting the toxicity mechanism, and provides a method for screening the reproductive toxicity of chemicals in zebrafish based on the model. This method effectively integrates multidimensional data from in vivo biological experiments and in vitro high-throughput biological tests, overcomes the limitations of traditional reliance on low-throughput animal experiments, and significantly improves the efficiency of screening compounds. The introduction of in vitro data sets greatly shortens the screening time and reduces experimental costs while maintaining the accuracy of the evaluation, making this method suitable for rapid screening of large-scale compounds.

[0005] The present invention provides a method for screening zebrafish reproductive toxicity and mechanism analysis by knowledge-driven machine learning, which is characterized by comprising the following steps: S1 Construction of reproductive toxicity prediction model; S1.1 Construct the original in vivo and in vitro test data sets of zebrafish reproductive toxicity compounds to form the original data set of compounds; S1.2 Calculate the molecular features of the original compound data set, obtain the in vivo test feature data set and the in vitro test feature data set, and merge them into a compound feature data set; The molecular features include: functional connectivity fingerprints and molecular descriptors; S1.3 inputting the compound characteristic data set into a deep neural network to obtain a reproductive toxicity prediction model; S2 calculates the molecular characteristics of the sample to be tested and inputs them into the reproductive toxicity prediction model to obtain the toxicity prediction results.

[0006] In the above method, the original in vivo test data set is reproductive toxicity endpoint data mediated by androgen receptor, aryl hydrocarbon receptor, estrogen receptor, peroxisome proliferator-activated receptor and thyroid hormone receptor respectively; Compounds are labeled based on whether they are reproductive toxic.

[0007] In the above method, the in vitro test original data set is generated by in vitro biological testing, and the reproductive toxicity reaction data mediated by androgen receptor, aryl hydrocarbon receptor, estrogen receptor, peroxisome proliferator-activated receptor and thyroid hormone receptor are screened out according to the nuclear receptor priority principle; The in vitro biological test includes: a test on the compound binding to a receptor, a test on protein stabilization, a test on gene expression regulation, a test on transcription factor activity regulation and a test on cell proliferation.

[0008] In the above method, step S1.2 further includes: S1.21 construct a quantitative structure-activity relationship model (QSAR model) based on the compound feature data set and the training set, and use the QSAR model to predict and fill in the missing values ​​in the in vitro test feature data set; The training set includes: CASRN, name and in vitro biological test results of the compound; The missing values ​​include completely random missing, random missing associated with observed data, and non-random missing associated with unobserved data themselves; Filling methods include: model prediction based on sample characteristics, or supplementation based on inference models of known data.

[0009] In the above method, it is characterized in that the reproductive toxicity prediction model is constructed based on the harmful outcome pathway framework, and the in vitro test feature data set is input into the middle layer neurons of the neural network for processing; Each neuron represents the process and result of an in vitro biological test in the training set.

[0010] In the above method, the reproductive toxicity prediction model includes: An input layer for receiving a feature dataset of the compound; At least one intermediate layer to simulate the process and results of biological testing; An output layer for outputting the reproductive toxicity endpoint data of the compound.

[0011] In the above method, step S1.3 further includes: S1.31 The reproductive toxicity prediction model was trained using a stochastic gradient descent algorithm, and the predictive performance of the model was evaluated by the leave-one-out cross-validation method to ensure the accuracy and generalization ability of the model.

[0012] In the above method, step S3 further includes: S1.32 Using the reproductive toxicity prediction model, the weight distribution and activation state of each neuron in the biological test process were calculated to construct a virtual toxicity pathway.

[0013] In the above method, the middle layer is 5 layers, which are as follows: Neural network layer 2: 11 neurons in total. In this layer, the model captures the receptor binding process and the starting point of the signal transduction pathway. Neural network layer 3: 12 neurons in total, representing the protein stabilization process; the model captures the effect of compounds on protein stability; The fourth layer of the neural network: a total of 7 neurons, representing the gene expression regulation process. In this layer, the model simulates the effect of compounds on gene expression in cells; Neural network layer 5: a total of 25 neurons, representing the transcription factor activity regulation process. In this layer, the model captures the effects of compounds on the activation, inhibition or modification of transcription factors; Neural network layer 6: a total of 9 neurons, representing the cell proliferation process. In this layer, the model captures the effect of compounds on cell proliferation. The 5 middle layers are densely connected, with each of the 64 neurons sharing a weighted connection with every neuron in the layer before and after it In the above method, an in vitro test characteristic data set of the sample to be tested is extracted, and a quantitative structure-activity relationship model is used to fill in the missing values.

[0014] This application has at least the following beneficial technical effects: The present invention combines the disclosed in vivo and in vitro test data sets to construct a machine learning model for the screening of reproductive toxicity of zebrafish chemicals and the prediction of toxicity mechanisms. This method effectively integrates multidimensional data from in vivo biological experiments and in vitro high-throughput biological tests, overcomes the limitations of traditional reliance on low-throughput animal experiments, and significantly improves the efficiency of screening compounds. The introduction of in vitro data sets greatly shortens the screening time and reduces experimental costs while maintaining the accuracy of the evaluation, making this method suitable for rapid screening of large-scale compounds.

[0015] In the method of the present invention, the reproductive toxicity prediction model constructed is a deep neural network model, which not only screens reproductive toxicity based on the molecular structure information of chemicals but also combines the results of in vitro biological tests, and can simulate the complex interactions between chemicals and biological systems. This multi-level simulation improves the interpretability of the model, avoids the limitations of traditional chemical structure-based QSAR models, and makes the model more accurate and in-depth in analyzing the toxicity mechanism of chemicals.

[0016] The model constructed in the method of the present invention adopts a multi-level biological testing process to simulate multiple biological reaction steps from receptor binding, protein stabilization to gene expression regulation and cell proliferation. Through this hierarchical neural network design, the model can accurately capture the effects of chemicals at different stages in the body, ensuring that each biological process is fully evaluated. This not only improves the accuracy of the prediction, but also provides an important scientific basis for the analysis of toxicity mechanisms.

[0017] The method of the present invention uses a quantitative structure-activity relationship (QSAR) model to handle missing data problems, and can effectively fill in missing values ​​in in vitro test data sets. With the help of model prediction, the integrity of the data set is ensured, while the accuracy of data analysis is improved, and data deviations caused by missing values ​​are avoided. This enables the present invention to maintain efficient analysis capabilities even when the data is incomplete.

[0018] The model constructed in the method of the present invention effectively achieves a comprehensive analysis of the toxicity mechanism of chemicals by accurately simulating the multi-level reactions of chemicals in organisms, especially accurately simulating the mechanisms of action in different biological processes (such as receptor binding, gene expression regulation, etc.). Unlike traditional methods, this model can not only predict reproductive toxicity, but also reveal the specific pathways and mechanisms of toxicity formation, thereby providing a more in-depth scientific basis for toxicity research. This innovation enables researchers to more accurately understand the impact of chemicals on biological systems, which will help future toxicity mechanism research and risk assessment. Through similar data integration and machine learning algorithms, the method of the present invention can be applied to the screening and prediction of other toxicities. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for screening zebrafish reproductive toxicity and mechanism analysis using knowledge-driven machine learning of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the implementation of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below in conjunction with the drawings in the embodiments of this application.

[0021] The following is combined with Figure 1 And further explain in detail the method of knowledge-driven machine learning screening of zebrafish reproductive toxicity and mechanism analysis in this application.

[0022] Example 1. Construction of a reproductive toxicity prediction model The method provided in this application for knowledge-driven machine learning screening of zebrafish reproductive toxicity and mechanism analysis includes the following steps.

[0023] S1. Construction of a reproductive toxicity prediction model S1.1 Construct the original data sets of in vivo and in vitro tests containing zebrafish reproductive toxicity compounds respectively.

[0024] The original in vivo test data set was derived from the Comparative Toxicogenomics (CTD) database and published scientific literature, and collected data on zebrafish reproductive toxicity endpoint compounds mediated by five nuclear receptors: androgen receptor (AR), aryl hydrocarbon receptor (AhR), estrogen receptor (ER), peroxisome proliferator-activated receptor (PPAR), and thyroid hormone receptor (TR). The data set contains 177 compounds, of which 113 have zebrafish reproductive toxicity and 64 have no zebrafish reproductive toxicity. Compounds with reproductive toxicity are marked as 1, and compounds without reproductive toxicity are marked as 0.

[0025] The original in vitro test data set is based on the in vitro biological test data of 177 chemicals extracted from the U.S. Environmental Protection Agency ToxCast / Tox21 in vitro toxicity screening database v4.1 (EPA ToxCast / Tox21 invitroDBv4.1), totaling 1439 biological tests. According to the biological process targets defined by ToxCast (including receptor binding, protein stabilization, gene expression regulation, transcription factor activity regulation and cell proliferation) and the nuclear receptor priority principle, 64 biological tests were finally screened as the in vitro test data set (see Table 2). That is, this data set was generated through in vitro biological tests, covering reproductive toxicity reaction data mediated by androgen receptor (AR), aryl hydrocarbon receptor (AhR), estrogen receptor (ER), peroxisome proliferator-activated receptor (PPAR) and thyroid hormone receptor (TR). Specific biological tests include: compound and receptor binding test, compound and protein stabilization test, compound and gene expression regulation test, compound and transcription factor activity regulation test and compound and cell proliferation test.

[0026] Finally, the above in vivo test datasets were combined with the in vitro test datasets to construct a dataset containing 177 compounds (Table 1), which had in vitro test data from 64 biological tests (Table 2) and in vivo test data of zebrafish fertility reduction.

[0027] Table 1 The original data set of the in vivo test

[0028] Table 2 In vitro test original data set: ToxCast / Tox21 in vitro biological test data set selected according to the neural network model

[0029] S1.2 Calculate the molecular characteristics of zebrafish reproductive toxic compounds and construct a characteristic dataset.

[0030] Based on the SMILES code of the compound in step S1, the functional connectivity fingerprint (FCFP) of the in vivo test compound was calculated using the RDKit package to obtain the in vivo test feature data set; the molecular descriptors of the in vitro biological test compound were calculated using the RDKit package to obtain the in vitro test feature data set; the two were combined to form the compound feature data set.

[0031] S1.21. Construction of quantitative structure-activity relationship model (QSAR) and filling of missing values The present invention constructs a quantitative structure-activity relationship (QSAR) model based on the compound feature data set, and uses the model to predict and fill in missing values ​​in the in vitro test feature data set. Missing values ​​include completely random missing, random missing associated with observed data, and non-random missing associated with unobserved data itself. The filling method includes: model prediction based on sample characteristics, or supplementation based on an inference model of known data.

[0032] Specifically, in order to fill in the missing values, the experimenters used QSAR modeling for all compounds in the EPA ToxCast / Tox21 database for 64 biological tests of the 177 compounds shown in Table 1 (as shown in Table 2) to fill in the missing data. In the QSAR modeling process, the training set data includes the American Chemical Abstracts Registry Number (CASRN), name and active / inactive test value (i.e., the results of in vitro biological tests) of the compound.

[0033] The CASRN of the compound was used to obtain the SMILES of the compound from PubChem. The molecular descriptors calculated by RDKit (https: / / rdkit.org), the molecular access control system (MACCS) fingerprints, and the functional connectivity fingerprint (FCFP) with a bond radius of 3 were tested in QSAR modeling as feature information to build the model to fill the data gaps caused by uncertain experimental results or untested compounds.

[0034] The entire modeling process was completed in PyCaret v3.2.0 (Python v3.9.9). A total of 16 algorithms were tested for each biological test. To solve the problem of data imbalance, an oversampling technique (SMOTE) was used. Finally, the best performing algorithm in each biological test was selected as the QSAR model of the biological test to fill the missing data.

[0035] Finally, three different molecular feature data were compared, and the QSAR model constructed using the RDKit descriptor was finally selected as the best model. The average performance of the QSAR models of different targets constructed using the RDKit descriptor was the best, with an average AUC of 0.82.

[0036] By this method, the activity test values ​​of the above 177 compounds in the above 64 bioassays were extracted from the EPA ToxCast / Tox 21 database. A total of 627 missing in vitro test activity data were filled, so that the final in vitro test feature data set contained 3754 active data and 7574 inactive data.

[0037] The following code demonstrates how to extract the molecular features of compounds (functional connectivity fingerprints, MACCS fingerprints, and RDKit molecular descriptors) through RDKit, and use the PyCaret library to build, optimize, and evaluate QSAR models for activity prediction.

[0038] # Install required libraries # !pip install rdkit pycaret pandas # Import necessary libraries from rdkit import Chem from rdkit.Chem import AllChem, Descriptors, MACCSkeys from pycaret.classification import * import pandas as pd # Read the raw data set of compound numbers # Assume that the dataset has a 'smiles' column (SMILES representation of the compound) and an 'activity' column (Activity data: 1 means active, 0 means inactive) data = pd.read_csv('compound_data.csv') # Define functions to calculate functional connectivity fingerprint (FCFP), MACCS fingerprint and RDKit descriptor def calculate_features(smiles): mol = Chem.MolFromSmiles(smiles) if mol: # Calculate functional connectivity fingerprint (FCFP) fcfp = AllChem.GetMorganFingerprintAsBitVect(mol, radius=3, nBits= 1024) fcfp_bits = list(map(int, fcfp.ToBitString())) # Calculate MACCS fingerprint maccs = MACCSkeys.GenMACCSKeys(mol) maccs_bits = list(map(int, maccs.ToBitString())) # Calculate RDKit descriptor descriptors = [Descriptors.MolWt(mol), Descriptors.MolLogP(mol), Descriptors.NumRotatableBonds(mol)] # Returns the combined features (FCFP, MACCS, RDKit descriptor) return fcfp_bits + maccs_bits + descriptors else: # If the molecule parsing fails, return a null value return [None] * (1024 + 166 + 3) # Calculate molecular features for all compounds features = data['smiles'].apply(calculate_features) features_df = pd.DataFrame(features.tolist(), columns=[f"fcfp_{i}" for i in range(1024)] + [f"maccs_{i}" for i in range(166)]+ ['MolWt', ' MolLogP', 'NumRotBonds']) # Merge feature data with activity data data = pd.concat([features_df, data['activity']], axis=1) # Use PyCaret to build the model clf = setup(data, target='activity', train_size = 0.8, fold = 5, fix_ imbalance = True, session_id=123) # Train multiple models and select the best model best_model = compare_models() # Fine-tune the best model tuned_model = tune_model(best_model) # Use the final model to make predictions predictions = predict_model(tuned_model) # Output the model's prediction results print(predictions) # Save the final model save_model(tuned_model, 'best_qsar_model') S1.3. The compound feature dataset is constructed through a deep neural network algorithm to obtain a reproductive toxicity prediction model.

[0039] The reproductive toxicity prediction model is built based on the adverse outcome pathway (AOP) framework, and the in vitro test feature data set is input into the intermediate layer neurons of the neural network for processing.

[0040] The reproductive toxicity prediction model constructed in the present invention is as follows: An input layer (layer 1 of the neural network) is used to receive the feature dataset of the compound, specifically the 1024-bit FCFP fingerprint.

[0041] At least one intermediate layer for processing the in vitro test characteristic data set to simulate the process and results of biological testing.

[0042] In this embodiment, the middle layer of the model contains 5 layers with a total of 64 neurons, each of which represents the process and result of a biological test in the training set. According to the biological process target representing the biological test, the 64 neurons are allocated to these 5 layers, and the specific structure of the middle layer is as follows: The second layer (middle layer) of the neural network, with a total of 11 neurons, represents the receptor binding process and is the starting point of the signal transduction pathway. In this process, chemicals bind to receptors on the cell surface, initiating a series of intracellular signal transduction reactions. Receptor binding is the first step for organisms to perceive external chemicals. The successful activation of this step determines the start of the subsequent signal transduction chain, thereby triggering various physiological and biochemical reactions in the cell.

[0043] The third layer (middle layer) of the neural network, with a total of 12 neurons, represents the protein stabilization process. In this layer, the effect of compounds on protein stability is reflected, mainly including how compounds affect the conformational stability, unfolding tendency, and interaction with auxiliary proteins of specific proteins. Protein stabilization is an important step in maintaining normal cell function. Abnormal stability may lead to protein dysfunction and induce toxic reactions.

[0044] The 4th layer (middle layer) of the neural network, with a total of 7 neurons, represents the gene expression regulation process. In this layer, the model simulates the effect of compounds on gene expression in cells. This involves the activation or inhibition of transcription factors, which in turn affects RNA synthesis. The gene expression regulation process is crucial because it determines which genes are expressed or inhibited at a specific time point, thereby affecting the response and function of the cell.

[0045] The fifth layer (middle layer) of the neural network, with a total of 25 neurons, represents the regulation of transcription factor activity. This layer mainly involves the effects of compounds on the activation, inhibition or modification of transcription factors. Transcription factors play a key role in regulating gene expression. They can precisely control the transcription of downstream genes. Therefore, the impact on transcription factors will directly change multiple physiological and biochemical pathways and affect the overall function of cells.

[0046] The sixth layer (middle layer) of the neural network, with a total of 9 neurons, represents the cell proliferation process. In this layer, the model captures the effect of compounds on the rate of cell proliferation. Cell proliferation is a fundamental process for the growth, repair, and maintenance of normal functions of organisms. The effect of compounds on cell proliferation may be achieved through a variety of pathways, including affecting cell cycle regulatory factors, leading to accelerated or inhibited cell growth, thus having a profound impact on the overall physiological process.

[0047] The five intermediate layers above are densely connected, with each of the 64 neurons sharing weighted connections with every neuron in the layer before and after it.

[0048] An output layer (layer 7 of the neural network) is used to output the reproductive toxicity endpoint data of the compound.

[0049] 1.31 Model training and performance prediction After the above model is built, it is trained. During the training process, the Stochastic Gradient Descent (SGD) algorithm is used to optimize the weight parameters of the network. The initial learning rate is set to 0.01 and the momentum parameter is 0.9 to accelerate convergence and reduce oscillation. The model is trained and evaluated using the binary cross entropy loss function, and the training ends after 200 iterations. In order to prevent overfitting and improve the generalization ability of the model, the dropout technology is also introduced during the training process to randomly remove some neurons in each layer.

[0050] After the model training is completed, the performance prediction is performed using the leave-one-out cross-validation method. The prediction effect of the final model of the present invention is: Training set: AUC( train) = 0.84, ACC( train ) = 0.89, F1( train ) = 0.78, MCC( train ) =0.66; Leave-one-out cross validation: AUC( test ) = 0.88, ACC( test ) = 0.81, F1( test ) = 0.75, MCC( test ) = 0.60.

[0051] Code example: #Import necessary libraries import os import numpy as np import pandas as pd import tensorflow as tf from arch_functions import create_architecture, layer_divide from generate_input import load_data from datetime import datetime # Define input parameters assay_file = 'path / to / assay_file.csv' dataset = 'path / to / dataset' endpoint = 'in_vivo_outcome' directory = 'data' fingerprints = 'FCFP' identifier = 'compound_id' num_nodes = [11,12,7, 25, 9] # Load in vivo and in vitro test data sets (in vitro and in vivo) profiles, in_vivo = load_data(dataset, directory, assay_file, endpoint, fingerprints, identifier=identifier) # Define the number of compounds to be used for training num_chemicals = profiles.shape[0] num_fragments = profiles.shape[1] - sum(num_nodes) # Stratify the data split_data = layer_divide(profiles, num_fragments, num_nodes) print(split_data[0].shape, split_data[1].shape) split_data.append(in_vivo.values.reshape((profiles.shape[0], 1))) feed_dict, output, weights = create_architecture(num_chemicals, split_data) y_true = list(feed_dict.items())[-1][0] # Define training parameters cost = tf.reduce_mean(tf.keras.backend.binary_crossentropy(y_true, output)) regularization_losses = tf.get_collection (tf.GraphKeys.REGULARIZATION_LOSSES) accuracy = tf.reduce_mean(tf.keras.metrics.binary_accuracy(y_true, output) + 0.001 * sum(regularization_losses)) lr_placeholder = tf.placeholder(tf.float32, [], name='learning_rate') cost_summary = tf.summary.scalar('cost', cost) accuracy_summary = tf.summary.scalar('accuracy', accuracy) merged = tf.summary.merge_all() # Set the SGD algorithm optimizer = tf.train.MomentumOptimizer(lr_placeholder, 0.9, use_ nesterov=False).minimize(cost) # Set TensorFlow logs now = datetime.utcnow().strftime('%Y%m%d%H%M%S') root_logdir = 'tf_logs' logdir = f'{root_logdir} / run-{now}' # Initialize variables for monitoring and saving model training lr_wait = 0 stop_wait = 0 best_cost = np.inf lr = 0.01 init = tf.global_variables_initializer() file_writer = tf.summary.FileWriter(logdir, tf.get_default_graph()) train = True epoch = 0 saver = tf.train.Saver() # Start training the model with tf.Session() as sess: sess.run(init) while train: feed_dict[lr_placeholder] = lr sess.run(optimizer, feed_dict=feed_dict) current_cost = sess.run(cost, feed_dict=feed_dict) if epoch % 50 == 0: print('Epoch:', epoch) print('Cost:', current_cost) print('Accuracy:', sess.run(accuracy, feed_dict=feed_dict)) summary, _ = sess.run([merged, optimizer], feed_dict=feed_dict) file_writer.add_summary(summary, epoch) epoch += 1 if epoch == 100001: train = False if np.less(current_cost, best_cost - 0.0001): best_cost = current_cost lr_wait = 0 stop_wait = 0 else: lr_wait += 1 stop_wait += 1 if lr_wait > 50: if lr > 0.00001: lr = max(lr * 0.9, 0) print(f'Reducing learning rate to {lr}') lr_wait = 0 if stop_wait >= 200: stopped_epoch = epoch print(f'Early stopping on epoch {epoch}') train = False # Save model training information saver.save(sess, os.path.join('.', f'{dataset}_{endpoint}_{assay_ file}_{fingerprints}')) file_writer.close() # Save the edge weights of each layer of the model for weight, counter in zip(weights[:-1], range(len(split_data) - 1)): w = pd.DataFrame(sess.run(weight), index=split_data[counter].columns, columns=split_data[counter + 1].columns) w.to_csv(os.path.join(directory, f'{dataset}_{endpoint}_{assay_file}_ {fingerprints}_{weight.name[:-2]}.csv')) # Save the output edge weights output_W = pd.DataFrame(sess.run(weights[-1]), index=split_data[-2] .columns) output_W.to_csv(os.path.join(directory, f'{dataset}_{endpoint}_ {assay_file}_{fingerprints}_output_W.csv')) Through the reproductive toxicity screening model, the toxic action process of chemicals in organisms can be simulated and their potential toxic pathways can be predicted. First, the model activates different neurons in the deep neural network layer by layer according to the molecular characteristics of the input chemicals. Each neuron represents a specific biological reaction in the biological test process, such as receptor binding, gene expression regulation, protein stabilization, etc.

[0052] In this process, the model records each activated neuron and its connection with other neurons, generating a "virtual toxic pathway" of the chemical. This pathway shows the interaction between chemicals and biological systems, from initial receptor binding to the regulation of gene expression, ultimately affecting cell function and biological processes, simulating the toxic action path of chemicals on organisms.

[0053] Through this virtual toxicity pathway, researchers can understand in detail how chemicals cause toxic reactions in organisms step by step. This pathway starts from the initial combination of chemical-biological targets and gradually affects the genes, proteins, and cells, thereby constructing the dynamic process of the entire toxicity mechanism. With the output of this model, scientists can identify important targets activated in the toxicity pathway, help further design more targeted experimental verification, and conduct in-depth analysis of the toxicity mechanism of chemicals. In addition, these toxicity pathways can also provide theoretical support for the safety assessment and environmental risk prediction of chemicals, and help formulate reasonable prevention and control strategies.

[0054] Application Experiment Example 1. Application of Reproductive Toxicity Prediction Model I Given a chemical, bis(2-ethylhexyl) tetrabromophthalate (TBPH, CAS 26040-51-7), its reproductive toxicity and toxicity pathway are predicted.

[0055] Firstly, the SMILES code of TBPH was obtained, and its FCFP molecular fingerprint and molecular descriptor were calculated using the RDKit software package.

[0056] Subsequently, the activity test values ​​of 64 biological tests involved in the method principle were extracted from the ToxCast / Tox21 project, and after checking, 10 missing values ​​were found, and the constructed QSAR model was used to fill the detected missing values.

[0057] Next, the calculated FCFP molecular fingerprint and in vivo test data were input into the reproductive toxicity prediction model for analysis. The predicted result was an activity value of 1, indicating that TBPH has reproductive toxicity, which is consistent with the existing test results.

[0058] Finally, the matrix multiplication result between the input vector from the previous layer and the optimized edge weights after the neurons corresponding to the inactive biological test results are dropped is output, and one or more neurons with the highest value in the output vector of each layer are selected as the neurons activated by the compound to form a toxic pathway.

[0059] The specific predicted pathways are: First, the AR-binding neuron NVS_NR_hAR, the PPARα receptor-binding neuron NVS_NR_hPPARa, and the PPARγ receptor-binding neuron NVS_NR_hPPARg (layer 2 of the neural network) were activated; The second step activated the protein immobilization neurons OT_AR_ARSRC1_0960, OT_ER_ERbERb_0480, and OT_ER_ERbERb_1440 (layer 3 of the neural network); The third step activated the DNA-binding neurons OT_ERa_EREGFP_0120 and OT_ERa_GFPERaERE_0480 (layer 4 of the neural network); The fourth step activated the transcriptional neurons ATG_chAR_XSP2, TOX21_PPARd_BLA_Antagonist_ratio, and TOX21_PPARg_BLA_Antagonist_ratio (layer 5 of the neural network); Finally, the cell proliferation neuron ACEA_AR_agonist_80hr (layer 6 of the neural network) was activated; this is consistent with the results of existing in vivo animal experiments.

[0060] Application Experimental Example 2. Application of Reproductive Toxicity Prediction Model II Given a chemical, decabromodiphenylethane (DBDPE, CAS 84852-53-9), we need to predict whether it is reproductive toxic and predict its toxicity pathway.

[0061] Firstly, the SMILES code of DBDPE was obtained, and the FCFP molecular fingerprint and RDKit molecular descriptor were calculated using the RDKit software package.

[0062] Subsequently, the activity test values ​​of the aforementioned 64 biological tests were extracted from the ToxCast / Tox 21 project, and 15 missing values ​​were checked, and the missing values ​​were filled using the QSAR model constructed in the present invention.

[0063] Next, the FCFP molecular fingerprint and in vivo test data were input into the model constructed in Example 1 for prediction. The predicted result was that the activity value was 1, that is, DBDPE had reproductive toxicity, which was consistent with the test result.

[0064] Finally, the matrix multiplication result between the input vector from the previous layer and the optimized edge weights after the neurons corresponding to the inactive biological test results are dropped is output. One or more neurons with the highest value in the output vector of each layer are selected as neurons activated by the compound to form a toxic pathway.

[0065] The specific predicted pathways are: First, AR receptor-binding neurons NVS_NR_hAR and ERα receptor-binding neurons NVS_NR_hER (layer 2 of the neural network) were activated; The second step activated the protein immobilization neurons OT_AR_ARSRC1_0960, OT_ER_ERbERb_0480, and OT_ER_ERbERb_1440 (layer 3 of the neural network); The third step activated the DNA-binding neurons OT_ERa_EREGFP_0120 and OT_ERa_GFPERaERE_0480 (layer 4 of the neural network); The fourth step activated the transcription neurons ATG_ERa_TRANS, TOX21_ERb_BLA_Antagonist_ratio, and TOX21_AR_BLA_Agonist_ratio (layer 5 of the neural network); Finally, the cell proliferation neurons ACEA_ER_80hr (layer 6 of the neural network) were activated, which is consistent with the results of existing animal experiments.

[0066] Application Experiment Example 3. Application of Reproductive Toxicity Prediction Model III Given the chemical 2,4-dibromophenol (CAS 615-58-7), we need to predict whether it is reproductive toxic and predict its toxicity pathway.

[0067] Firstly, the SMILES code of 2,4-dibromophenol was obtained, and the FCFP molecular fingerprint and RDKit molecular descriptor were calculated using the RDKit software package.

[0068] Subsequently, the activity test values ​​of 64 biological tests were extracted from the ToxCast / Tox 21 project, and 9 missing values ​​were checked, and the missing values ​​were filled using the QSAR model constructed in the present invention.

[0069] Next, the FCFP molecular fingerprint and in vivo test data were input into the model constructed by the present invention for prediction. The prediction result was that the activity value was 0, that is, 2,4-dibromophenol had no reproductive toxicity, which was consistent with the existing experimental test results.

[0070] The predicted pathway showed that shedding occurred at the second layer of neurons, resulting in interruption of the activation pathway, so it was predicted that 2,4-dibromophenol would not be reproductively toxic.

[0071] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for screening the reproductive toxicity of chemical substances in zebrafish by using knowledge-driven machine learning, characterized in that: The following steps are involved: S1 Construction of reproductive toxicity prediction model; S1.1 Construct the original in vivo and in vitro test data sets of zebrafish reproductive toxicity compounds to form the original data set of compounds; S1.2 Calculate the molecular features of the original compound data set, obtain the in vivo test feature data set and the in vitro test feature data set, and merge them into a compound feature data set; The molecular features include: functional connectivity fingerprints and molecular descriptors; S1.3 inputting the compound characteristic data set into a deep neural network to obtain a reproductive toxicity prediction model; S2 calculates the molecular characteristics of the sample to be tested and inputs them into the reproductive toxicity prediction model to obtain the toxicity prediction results.

2. The method according to claim 1, characterized in that The original data set of the in vivo test is reproductive toxicity endpoint data mediated by androgen receptor, aryl hydrocarbon receptor, estrogen receptor, peroxisome proliferator-activated receptor and thyroid hormone receptor, respectively; Compounds are labeled based on whether they are reproductive toxic.

3. The method according to claim 2, characterized in that The original in vitro test data set is generated through in vitro biological testing, and the reproductive toxicity reaction data mediated by androgen receptor, aryl hydrocarbon receptor, estrogen receptor, peroxisome proliferator-activated receptor and thyroid hormone receptor are screened out according to the nuclear receptor priority principle; The in vitro biological test includes: a test on the compound binding to a receptor, a test on protein stabilization, a test on gene expression regulation, a test on transcription factor activity regulation and a test on cell proliferation.

4. The method according to claim 3, characterized in that The step S1.2 further includes: S1.21 Build a quantitative structure-activity relationship model based on the compound feature dataset and training set, and use the model to predict and fill in missing values ​​in the in vitro test feature dataset; The training set includes: CASRN, name and in vitro biological test results of the compound; The missing values ​​include completely random missing, random missing associated with observed data, and non-random missing associated with unobserved data themselves; Filling methods include: model prediction based on sample characteristics, or supplementation based on inference models of known data.

5. The method according to claim 4, characterized in that The reproductive toxicity prediction model is constructed based on the adverse outcome pathway framework, and the in vitro test feature data set is input into the middle layer neurons of the neural network for processing; Each neuron represents the process and result of an in vitro biological test in the training set.

6. The method according to claim 5, characterized in that The reproductive toxicity prediction model comprises: An input layer for receiving a feature dataset of the compound; At least one intermediate layer to simulate the process and results of biological testing; An output layer for outputting the reproductive toxicity endpoint data of the compound.

7. The method according to claim 6, characterized in that The step S1.3 also includes: S1.31 The reproductive toxicity prediction model was trained using a stochastic gradient descent algorithm, and the predictive performance of the model was evaluated by the leave-one-out cross-validation method to ensure the accuracy and generalization ability of the model.

8. The method according to claim 7, characterized in that The step S1.3 also includes: S1.32 Using the reproductive toxicity prediction model, the weight distribution and activation state of each neuron in the biological test process were calculated to construct a virtual toxicity pathway.

9. The method according to claim 6, characterized in that The middle layer has five layers, which are as follows: Neural network layer 2: 11 neurons in total. In this layer, the model captures the receptor binding process and the starting point of the signal transduction pathway. Neural network layer 3: 12 neurons in total, representing the protein stabilization process; The model captures the effects of compounds on protein stability; The fourth layer of the neural network: a total of 7 neurons, representing the gene expression regulation process. In this layer, the model simulates the effect of compounds on gene expression in cells; Neural network layer 5: a total of 25 neurons, representing the transcription factor activity regulation process. In this layer, the model captures the effects of compounds on the activation, inhibition or modification of transcription factors; Neural network layer 6: a total of 9 neurons, representing the cell proliferation process. In this layer, the model captures the effect of compounds on cell proliferation. The five intermediate layers are densely connected, with each of the 64 neurons sharing weighted connections with every neuron in the layer before and after it.

10. The method according to claim 9, characterized in that Step 2 includes extracting the in vitro test characteristic data set of the samples to be tested and filling in the missing values ​​using the quantitative structure-activity relationship model.

Images