Endocrine disrupter prediction method based on harmful outcome path network

Abstract

The invention discloses an endocrine disrupter prediction method based on a harmful outcome path network, and belongs to the field of endocrine, and the method comprises the steps: obtaining experimental activity data containing qualitative data and quantitative data according to a pre-constructed harmful outcome path network; acquiring endocrine disrupter list data as a second data set; constructing a target spot qualitative prediction model according to the qualitative data; performing target activity data prediction on the second data set by using a target qualitative prediction model to obtain a target activity data set; constructing an endocrine disruption effect qualitative prediction model by using the target activity data set; constructing a target spot quantitative prediction model according to the quantitative data; predicting whether the compound to be detected has the endocrine disruption effect or not by utilizing the endocrine disruption effect qualitative prediction model; and for the compound with the endocrine interference effect, predicting the corresponding target quantitative activity by using the target quantitative prediction model to obtain a quantitative harmful outcome path. According to the method, the endocrine disrupter EDCs is qualitatively identified and quantitatively classified.

Description

Technical Field

[0001] The present application relates to the field of endocrinology, and more specifically, to a method for predicting endocrine disruptors based on a harmful outcome pathway network. Background Art

[0002] With the acceleration of industrialization, humans are exposed to a large number of exogenous chemicals, namely endocrine disruptors (EDCs). These chemicals can inadvertently disrupt the complex endocrine system and cause adverse health effects. Exposure to EDCs has been repeatedly associated with serious diseases and developmental disorders (such as reproductive disorders and cognitive defects). This grim reality has led government agencies and the scientific community to take a series of necessary actions to identify and classify EDCs. On March 31, 2023, the European Union (EU) adopted a revision of the Classification, Labelling and Packaging Regulation, requiring endocrine disruptor labels to be affixed to substances and mixtures by 2028. Therefore, evaluating the endocrine disrupting properties of more than 100,000 commercial chemicals is a major unresolved challenge.

[0003] The EU has set three criteria for considering a substance as an EDC: (a) it shows an adverse effect (e.g., impaired function); (b) it has an endocrine disrupting mechanism (e.g., alters endocrine system function); and (c) there is a biologically plausible link between the adverse effect and the endocrine disrupting mechanism (i.e., the adverse effect is the result of the endocrine disrupting mechanism). With the advancement of experimental technology, the United States Environmental Protection Agency (US EPA), the Organization for Economic Cooperation and Development (OECD), and the EU have established a wide range of in vitro and in vivo testing technologies (e.g., the EDSP program and the ToxCastTM / Tox21 program) to monitor the endocrine disrupting mechanisms and adverse effects of chemicals. The field of artificial intelligence has worked hard to develop new machine learning algorithms and deep learning architectures, providing new opportunities for high-throughput virtual screening of EDCs. Many platforms (e.g., the VEGA platform and Endocrine Disruptome) or models (e.g., CoMPARA, CERAPP, and NRMEA) have been developed and are freely available. However, all experimental techniques and virtual screening methods target only one specific effect, either an endocrine disrupting mechanism or a physiologically measurable adverse effect. Until now, biologically plausible links between evidence have required manual judgment based on expert experience, which has made it difficult for tools that meet all three criteria simultaneously.

[0004] Adverse outcome pathway networks (AOP networks) can help establish the above causal links. It can condense multi-level toxicological information such as molecular initiating events (MIEs), key events (KEs), and adverse outcomes (AOs) after chemical exposure observed in various laboratory experiments and population epidemiology. Despite recent advances in experimental and computational techniques, it is still challenging to develop and apply new technologies to identify EDCs because there is a lack of biologically plausible links between chemical effects detected at the macromolecular level of biological tissues and the adverse effects and top results traditionally used in EDC hazard identification and classification. Faced with this challenge, a new technical solution is urgently needed that can combine advanced artificial intelligence algorithms and hierarchical biological knowledge in AOPnetwork to build a predictive model with biological information. By encoding hierarchical biological knowledge (including toxicological mechanisms, adverse outcomes, and biological causal relationships), high-throughput screening, qualitative hazard identification, and quantitative hazard classification of EDCs can be achieved, meeting the three standards of government agencies, improving prediction performance, and simultaneously revealing the endocrine disruption patterns of EDCs.

[0005] The Chinese patent application, application number CN201811597767.1, published on March 29, 2019, discloses a method for high-throughput screening of endocrine disruptors based on hierarchical warning structures, using substructure frequency analysis and substructure proportion analysis to extract the primary warning structure of active data compounds; using SARpy software to extract the secondary warning structure of compounds that meet the primary warning structure; using SARpy software to extract the tertiary warning structure; combining the primary warning structure and the secondary warning structure to form an activity prediction module, first screening out compounds with characteristic structures in the prediction module, and then screening out warning compounds with potential endocrine disrupting effects based on the secondary warning structure; using the tertiary warning structure as an interference activity prediction module, and then screening interference activity based on the interference activity prediction module. However, the extraction of the warning structure in this scheme is based only on molecular fingerprints and substructure analysis, and lacks consideration of the mechanism of action and harmful consequences of endocrine disruptors. Summary of the invention

[0006] 1. Technical problems to be solved

[0007] In view of the problem that the prior art lacks both qualitative identification and quantitative classification of endocrine disruptors EDCs identified by the EU, the present application provides an endocrine disruptor prediction method based on an adverse outcome pathway network, which obtains experimental activity data and endocrine disruptor list data, and pre-processes the experimental activity data to construct a target qualitative prediction model and an endocrine disrupting effect qualitative prediction model to predict whether the test compound has an endocrine disrupting effect; at the same time, a target quantitative prediction model is constructed to predict the target quantitative activity of compounds with endocrine disrupting effects, obtain a quantitative adverse outcome pathway, and identify sensitive adverse outcomes and sensitive quantitative adverse outcome pathways based on the quantitative adverse outcome pathways; finally, the endocrine disruptors are classified according to the sensitive harmful outcomes and the sensitive quantitative harmful outcome pathways.

[0008] 2. Technical solution

[0009] The purpose of this application is achieved through the following technical solutions.

[0010] One aspect of the present application provides an endocrine disruptor prediction method based on a harmful outcome pathway network, comprising: obtaining experimental activity data as a first data set according to a pre-constructed harmful outcome pathway network, and preprocessing the first data set; wherein the first data set includes qualitative data and quantitative data; obtaining endocrine disruptor inventory data as a second data set; constructing a target qualitative prediction model based on the qualitative data in the pre-processed first data set using a machine learning algorithm; predicting target activity data for the second data set using the target qualitative prediction model to obtain a target activity data set; constructing an endocrine disruptor inventory using a machine learning algorithm using the target activity data set Effect qualitative prediction model; based on the quantitative data in the first data set after preprocessing, a machine learning algorithm is used to construct a target quantitative prediction model; using the endocrine disruption effect qualitative prediction model to predict whether the test compound has endocrine disruption effect; for compounds with endocrine disruption effect, the target quantitative prediction model is used to predict the corresponding target quantitative activity and obtain the quantitative harmful outcome path; based on the quantitative dose-effect relationship of each event in the quantitative harmful outcome path, sensitive harmful outcomes and sensitive quantitative harmful outcome paths are obtained; based on the sensitive harmful outcomes and the sensitive quantitative harmful outcome paths, endocrine disruptors are graded for hazard classification to obtain the prediction results of endocrine disruptors.

[0011] Furthermore, the experimental activity data include data related to molecular initiation events, key events and harmful outcome events; molecular initiation events represent starting events in the harmful outcome pathway network; key events represent event nodes in the harmful outcome pathway network that reflect changes in chemical substances and ultimately lead to harmful outcomes; harmful outcome events represent endpoint events in the harmful outcome pathway network.

[0012] Furthermore, the first data set is preprocessed, including: performing quality assessment on the experimental activity data; screening the experimental activity data according to the quality assessment result; and using the screened experimental activity data as the preprocessed first data set.

[0013] Furthermore, a machine learning algorithm is used to construct a target quantitative prediction model or a target qualitative prediction model, including: using a molecular descriptor-based modeling method or a molecular graph-based modeling method to construct a target quantitative prediction model or a target qualitative prediction model.

[0014] Furthermore, a molecular descriptor-based modeling method is used to construct a target quantitative prediction model or a target qualitative prediction model, including: converting the chemical structure SMILES of the compound into a numerical molecular descriptor or a molecular fingerprint; based on the molecular descriptor or the molecular fingerprint, a machine learning algorithm is used to construct a target quantitative prediction model or a target qualitative prediction model.

[0015] Furthermore, a molecular graph-based modeling method is sampled to construct a target quantitative prediction model or a target qualitative prediction model, including: converting the chemical structure SMILES of the compound into a molecular graph structure, in which atoms are represented as nodes and bonds are represented as edges; according to the molecular graph structure, a graph neural network GNN algorithm is used to construct a target quantitative prediction model or a target qualitative prediction model.

[0016] Furthermore, the target activity dataset was used and the graph neural network-based machine learning algorithm GCNConvEdge was adopted to construct a qualitative prediction model for endocrine disruption effects, including: taking the harmful outcome pathway network as a directed acyclic graph, the target activity data as node input information, the causal relationship between key events as edge input information, and whether there is an endocrine disruption effect as output information.

[0017] Furthermore, according to the quantitative dose-effect relationship of each event in the quantitative harmful outcome pathway, sensitive harmful outcomes and sensitive quantitative harmful outcome pathways are obtained, including: according to the quantitative dose-effect relationship of each event in the quantitative harmful outcome pathway, the harmful outcome with the lowest highest no-effect concentration NOAEL is obtained as the sensitive harmful outcome; based on the sensitive harmful outcome, the corresponding quantitative harmful outcome pathway is determined to be the sensitive quantitative harmful outcome pathway, and the sensitive quantitative harmful outcome pathway includes: sensitive molecule initiating events, sensitive key events and sensitive harmful outcomes; the NOAEL values ​​of sensitive molecule initiating events, sensitive key events and sensitive harmful outcomes in the sensitive quantitative harmful outcome pathway are compared to see whether they meet the following requirements: the NOAEL value of the sensitive molecule initiating event is less than the NOAEL value of the sensitive key event is less than the NOAEL value of the sensitive harmful outcome; if so, the corresponding pathway is taken as the sensitive quantitative harmful outcome pathway.

[0018] Furthermore, according to the sensitive harmful outcomes and the sensitive quantitative harmful outcomes pathway, the endocrine disruptors are graded for hazard, and the prediction results of endocrine disruptors are obtained, including: obtaining the highest no-effect concentration (NOAEL) of each key event in the sensitive quantitative harmful outcomes pathway; using the uncertainty factor method, extrapolating the highest no-effect concentration (NOAEL) to the sensitive harmful outcomes to obtain the highest no-effect concentration (NOAEL) of the sensitive harmful outcomes based on each key event; obtaining the minimum value of the highest no-effect concentration (NOAEL) of the sensitive harmful outcomes of each key event as the hazard grading threshold of the corresponding compound; and grading the test compound according to the hazard grading threshold.

[0019] Furthermore, the target activity dataset was used to construct a qualitative prediction model for endocrine disruption effects using a machine learning algorithm: the machine learning algorithm used was GCNConvEdge, a machine learning algorithm based on a graph neural network.

[0020] Another aspect of the present application also provides a system for predicting endocrine disruptors, including: a data acquisition module, which acquires experimental activity data as a first data set, acquires endocrine disruptor inventory data as a second data set, and preprocesses the first data set; wherein the first data set includes qualitative data and quantitative data; a first data set processing module, which uses a machine learning algorithm to construct a target qualitative prediction model based on the qualitative data in the preprocessed first data set; a second data set processing module, which uses the target qualitative prediction model to predict target activity data of the second data set to obtain a target activity data set; a third data set processing module, which uses the target activity data set to construct an endocrine disruptor effect prediction model using a machine learning algorithm A qualitative prediction model is used; the fourth data set processing module uses a machine learning algorithm to build a target quantitative prediction model based on the quantitative data in the first data set after preprocessing; the prediction module uses the endocrine disruption effect qualitative prediction model to predict whether the test compound has endocrine disruption effect; for compounds with endocrine disruption effect, the target quantitative prediction model is used to predict the corresponding target quantitative activity to obtain a quantitative harmful outcome path; based on the quantitative dose-effect relationship of each event in the quantitative harmful outcome path, sensitive harmful outcomes and sensitive quantitative harmful outcome paths are obtained; endocrine disruptors are graded according to the sensitive harmful outcomes and sensitive quantitative harmful outcome paths to obtain the prediction results of endocrine disruptors.

[0021] 3. Beneficial effects

[0022] Compared with the prior art, the advantages of this application are:

[0023] (1) This application uses machine learning and machine learning algorithms, combined with experimental activity data and endocrine disruptor inventory data, to construct a target qualitative prediction model, an endocrine disruption effect qualitative prediction model, and a target quantitative prediction model, which can effectively establish the biological connection between the toxicological mechanism and harmful outcomes of EDCs, thereby overcoming the "black box" defect of traditional data-driven models;

[0024] (2) This application provides a visual interactive form that can directly encode hierarchical prior knowledge into a directed acyclic graph language and convert this hierarchical knowledge into a graph neural network (GNN). By adopting the machine learning algorithm GCNConvEdge based on a graph neural network, the qualitative prediction model of endocrine disruption effects constructed in this application can transparently display the complete process from compound-protein interaction to individual phenotypic toxicity, providing comprehensive knowledge information for prediction and mechanism elucidation in the context of environmental toxicology;

[0025] (3) Compared with traditional data-driven models, this application has achieved excellent performance in qualitative and quantitative hazard identification and hazard classification tasks by organizing and learning complete hierarchical biological knowledge of key event (KE) information (including toxicological mechanisms and harmful outcomes) and key event relationship (KER) information (biologically credible connections) about the harmful outcome pathway network. Specifically, in qualitative prediction tasks, the method of this application can achieve a ROC-AUC value of 0.90; in quantitative prediction tasks, the method of this application can control the mean square error (MSE) at a low level of 0.08. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is an exemplary flow chart of an endocrine disruptor prediction method based on a harmful outcome pathway network of the present application;

[0027] Figure 2 The EDC screening test method collected for this application;

[0028] Figure 3 Qualitative data collected for this application according to the standardized screening assay (top) and the ratio (%) of the amount of data used in this study to the amount of data collected (bottom);

[0029] Figure 4 Quantitative data collected for this application according to the standardized screening assay (top) and the ratio (%) of the amount of data used in this study to the amount of data collected (bottom);

[0030] Figure 5 A qualitative prediction model of 74 optimal targets was constructed for this application;

[0031] Figure 6A method for constructing a qualitative prediction model of endocrine disruption effects based on a harmful outcome pathway network according to the present invention;

[0032] Figure 7 Obtain the best prediction performance for the novel GCNConvEdge architecture of this application;

[0033] Figure 8 A quantitative prediction model of 52 optimal targets was constructed for this application. DETAILED DESCRIPTION

[0034] The present application is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0035] Example 1

[0036] like Figure 1 As shown, a method for predicting endocrine disruptors based on a harmful outcome pathway network includes: obtaining experimental activity data as a first data set according to a pre-constructed harmful outcome pathway network, and preprocessing the first data set; wherein the first data set includes qualitative data and quantitative data; obtaining endocrine disruptor inventory data as a second data set; constructing a target qualitative prediction model based on the qualitative data in the pre-processed first data set using a machine learning algorithm; predicting target activity data for the second data set using the target qualitative prediction model to obtain a target activity data set; constructing a qualitative prediction model of endocrine disrupting effects using a machine learning algorithm using the target activity data set. Prediction model; based on the quantitative data in the first data set after preprocessing, a machine learning algorithm is used to construct a target quantitative prediction model; using the endocrine disruption effect qualitative prediction model, it is predicted whether the test compound has endocrine disruption effect; for compounds with endocrine disruption effect, the target quantitative prediction model is used to predict the corresponding target quantitative activity and obtain the quantitative harmful outcome path; based on the quantitative dose-effect relationship of each event in the quantitative harmful outcome path, sensitive harmful outcomes and sensitive quantitative harmful outcome paths are obtained; based on the sensitive harmful outcomes and the sensitive quantitative harmful outcome paths, endocrine disruptors are graded for hazard classification and the prediction results of endocrine disruptors are obtained.

[0037] Specifically, this example selected EDCs mediated by estrogen (E), androgen (A), thyroid (T) and steroidogenesis (S) as research objects, and after collecting the harmful outcome pathway network of EATS-mediated EDCs, a model was constructed.

[0038] According to the pre-constructed adverse outcome pathway (AOP) network, experimental activity data are obtained as the first data set: using the established AOP network, experimental activity data related to key events in the AOP network are extracted from relevant databases, literature and other sources. Experimental activity data include qualitative data (such as positive / negative results) and quantitative data (such as half effective concentration EC50, half inhibitory concentration IC50, etc.). The collected experimental activity data are organized into a structured first data set. Preprocess the first data set: clean the experimental activity data in the first data set to remove missing values, outliers and other noise data. Standardize the quantitative data, such as unifying the concentration unit and normalizing the numerical range. Encode the qualitative data, such as converting positive / negative results to 1 / 0, etc.

[0039] Obtain the endocrine disrupting chemicals (EDCs) inventory data as the second data set: Obtain the compound information of known or suspected EDCs from the EDCs list published by authoritative organizations, relevant databases and other channels. Sort out the EDCs inventory data and extract key information such as compound name, molecular structure, and physical and chemical properties. Use the sorted EDCs inventory data as the second data set.

[0040] Collect standard test methods for EDCs: Based on three official guidance documents, namely the OECD document "Revised Guidance Document No. 150 on Standardized Test Guidelines for the Evaluation of Chemical Endocrine Disruption", the ECHA document "Guidelines for the Identification of Endocrine Disruptors in Accordance with Regulation (EU) No. 528 / 2012 and Regulation (EC) No. 1107 / 2009", and the EPA document "Guidelines for Testing in Endocrine Disruptor Screening Programs", standard test methods for EDCs were collected. According to the guidance documents, standardized tests with non-mammals (such as poultry), non-EATS-related tests, and repeated test methods were deleted. The test endpoints of the screened standardized test methods were collected and a list of test endpoints was compiled. The test endpoints were matched one by one with the events in the AOP network: the correspondence between the test endpoints and the molecular initiation events, key events, and harmful outcomes in the AOP network was analyzed. A one-to-one mapping relationship was established between each test endpoint and a specific event in the AOP network.

[0041] like Figure 2 As shown, 26 detection methods were finally screened out: Based on the three official guidance documents and the correspondence between the test endpoints and the AOP network events, 26 detection methods were finally selected, including 7 in vitro tests and 19 in vivo tests.

[0042] like Figure 3As shown, the collection of qualitative data of EDCs: According to the selected detection method, the collection of qualitative data of EDCs was carried out. The data of the selected standardized tests came from five public databases, including eChemPortal, DrugBank, EADB, EndocrineDisruptorKnowledgeBas, and CompTox Chemicals Dashboard. We collected a total of 21,700 experimental results from the five databases. Considering the importance of data quality to model building, the collected experimental result data were screened: the reliability of each experimental result with / without restrictions was evaluated. Finally, less than half of the experimental result data were selected to ensure the high quality and reliability of the data.

[0043] The toxic endpoints of 26 detection methods were mapped to the key events in the adverse outcome pathway (AOP) network: the key events in the AOP network include molecular initiating events (MIE), key events (KE) and adverse outcomes (AO). The correspondence between the toxic endpoints of each detection method and specific key events in the AOP network was analyzed. A mapping table between the toxic endpoints of the detection method and the key events of the AOP network was established. According to the mapping relationship between the toxic endpoints of the detection method and the key events of the AOP network, the screened experimental result data were classified into the corresponding key events: using the mapping table, the AOP network key events corresponding to each experimental result were determined. The experimental results were classified and sorted according to the corresponding key events. Statistical mapping and classification results: A total of 74 key events in the AOP network were matched to the toxic endpoints of the detection method. 132,947 toxicology records were successfully classified into different key events.

[0044] EDCs quantitative data collection, such as Figure 4 As shown, the highest no-effect concentration (NOAEL) was selected as the indicator of quantitative data: NOAEL is an important toxicological parameter for evaluating the safety of compounds, indicating the highest dose or exposure level that does not cause significant adverse effects under experimental conditions. It is generally believed that compound exposure doses below NOAEL are safe for humans. NOAEL quantitative data of EDCs were collected based on the same five databases as qualitative data collection: eChemPortal, DrugBank, EADB, Endocrine Disruptor Knowledge Base, and CompTox Chemicals Dashboard; from the above databases, a total of 10,200 experimental result data of 18 detection methods were collected. Note that there are no available NOAEL data for the other 8 detection methods. The collected NOAEL experimental result data were quality assessed and screened: the quality of the toxicology record of each experimental result was evaluated. Finally, 6,713 high-quality toxicology records were selected, accounting for 65.18% of the total data.

[0045] The toxic endpoints of 18 detection methods were mapped to key events in the adverse outcome pathway (AOP) network: The correspondence between the toxic endpoints of each detection method and specific key events in the AOP network was analyzed. A mapping table between the toxic endpoints of the detection methods and the key events of the AOP network was established. According to the mapping relationship between the toxic endpoints of the detection methods and the key events of the AOP network, the screened NOAEL experimental result data were classified into the corresponding key events: The mapping table was used to determine the AOP network key events corresponding to each NOAEL experimental result. The NOAEL experimental results were classified and sorted according to the corresponding key events. Statistical mapping and classification results: The toxic endpoints of 18 detection methods were mapped to 52 key events in the AOP network. 54,200 toxicology records were successfully classified into different key events. NOAEL quantitative experimental data of EDCs were widely collected from five authoritative databases, and the data were screened, mapped and classified in combination with the AOP network, and finally a high-quality, structured quantitative data set related to key events was obtained. These NOAEL data can be used for subsequent quantitative prediction model construction

[0046] Construct a qualitative prediction model of targets based on key events. Data set division: The qualitative data in the first data set after preprocessing are randomly divided into a training set and a test set. 80% of the compounds are selected as the training set for QSAR modeling, and the remaining 20% ​​of the compounds are used as the test set to evaluate the model performance. Traditional QSAR modeling method: The chemical structure is filtered and standardized, and the QSAR-ready structure is generated through the standardized workflow in KNIME. The standardized chemical structure is converted into a molecular descriptor and a molecular fingerprint (2756 bits) using PaDEL-descriptor software. The converted data is strictly cleaned to remove blank values, outliers, and erroneous values. The features are normalized using Z scores so that the features of the training set follow a normal distribution with a mean of 0 and a variance of 1. Features with small variance (var=0) are removed, and univariate feature selection methods (such as the mutual information method) are used to reduce variables and reduce information noise.

[0047] Select machine learning algorithms: Select five representative machine learning algorithms, including random forest (RF), support vector machine (SVM), decision tree (DT), naive Bayes (NB) and k-nearest neighbor (kNN) algorithms. Hyperparameter optimization: Use learning curve and grid search methods to optimize the hyperparameters of each machine learning algorithm.

[0048] The optimized hyperparameters include: RF: n_estimators (number of trees) and max_depth (maximum depth of the tree); SVM: C (regularization parameter) and kernel (kernel type used in the algorithm); DT: max_depth (maximum depth of the tree); NB: feature_prior (feature likelihood, such as Gaussian, polynomial, complementary, Bernoulli); kNN: n_neighbors (number of neighbors). Imbalanced class problem handling: The synthetic minority oversampling technique (SMOTE) is used to solve the imbalanced class problem that may exist in the dataset. SMOTE balances the dataset by synthesizing new minority class samples to improve the model's prediction performance for the minority class. Model training and evaluation: Using the optimized hyperparameters and the balanced training set, the models of the five machine learning algorithms were trained separately. The performance of each model was evaluated on the test set, and the model with the best performance was selected as the final target qualitative prediction model. Using the preprocessed qualitative data, the traditional QSAR modeling method and multiple machine learning algorithms were used to construct a target qualitative prediction model based on key events. These models can be used to predict whether a compound has biological activity associated with specific key events.

[0049] Construct a molecular graph-based target qualitative prediction model, chemical structure representation: Convert the SMILES string of each compound into a molecular graph. Use the open source software package RDKit to determine the set of atoms and chemical bonds in the molecular graph. Eigenvector calculation and initialization: Calculate and initialize the eigenvectors in the molecular graph. Atomic features include 8 attributes, such as atomic number and formal charge. Chemical bond features include 4 attributes, such as chemical bond type and conjugation.

[0050] Select the graph neural network (GNN) algorithm: Use the directed message passing neural network (DMPNN) as the graph neural network algorithm. DMPNN can learn to predict chemical toxicity directly from the graph structure of chemicals. Information passing and feature learning: DMPNN summarizes information from neighboring atoms and chemical bonds through a series of message passing steps. In each message passing step, DMPNN updates the feature vector of the current atom based on the feature vectors of neighboring atoms and chemical bonds. Through multiple message passing, DMPNN gradually builds an understanding of local chemical properties.

[0051] Hyperparameter optimization: Hyperparameter optimization was performed using the Bayesian optimization method using the Hyperopt python package. The optimized hyperparameters include: depth: number of information transfer steps; dropout probability: regularization technique to prevent overfitting; hidden size: size of key information vector; number of feedforward network layers: number of fully connected layers used for feature transformation and prediction. Imbalanced class problem handling: SMOTE upsampling method was used to balance the ratio of active and inactive compounds. SMOTE balances the dataset by synthesizing new minority class (active) samples to improve the model's prediction performance for active compounds. Model training and evaluation: The DMPNN model was trained using the optimized hyperparameters and balanced training set. The performance of the model was evaluated on the test set and compared with the model built by the traditional QSAR method. The molecular graph was used to represent the chemical structure, and the directed information transfer neural network (DMPNN) algorithm was used to construct a molecular graph-based target qualitative prediction model. Compared with the traditional QSAR method, this method can directly learn molecular features from the graph representation of the chemical structure without manually designing molecular descriptors, and has stronger expressive power and flexibility.

[0052] like Figure 5As shown, this embodiment establishes six binary QSAR classification models for each event and selects the best model. Specifically, for each key event, the two QSAR modeling methods mentioned above (traditional method and molecular graph-based method) are used to establish the model respectively. For each modeling method, three different data partitioning methods (such as random partitioning, time series partitioning, and chemical structure similarity partitioning) are used to establish the model. Therefore, for each key event, a total of 6 binary QSAR classification models (2 modeling methods × 3 data partitioning methods) are established. The performance of each QSAR classification model is evaluated using three indicators: F1-score, ROC-AUC value, and accuracy. F1-score is the harmonic mean of precision and recall, which is suitable for the evaluation of unbalanced data sets. The ROC-AUC value represents the area under the receiver operating characteristic curve, reflecting the overall performance of the classifier at different thresholds. The accuracy rate represents the proportion of correctly predicted samples to the total number of samples. For each key event, the QSAR classification model with the best performance is selected as the best target qualitative prediction model according to the F1-score indicator. A total of 74 best target qualitative prediction models corresponding to 74 key events were selected. The performance indicators of the 74 best target qualitative prediction models were statistically analyzed. The mean and standard deviation of F1-score, ROC-AUC value and accuracy were calculated. The results showed that the average F1-score was 0.42 and the standard deviation was 0.19. The average ROC-AUC value was 0.71 and the standard deviation was 0.12. The average accuracy was 0.88 and the standard deviation was 0.11. The statistical and visual analysis results of the performance indicators of these best models showed that the constructed QSAR model had good predictive ability as a whole and could be used to predict the effects of compounds on specific key events.

[0053] The target qualitative prediction model is used to make predictions and construct a qualitative prediction model for endocrine disruption effects, including: for the compounds in the second data set, the best target qualitative prediction model constructed in step S2 is used to make predictions. For each compound, the corresponding best model is used to predict its effect on each key event (i.e., target activity). The prediction results are integrated to obtain a complete target activity data set, including original data and predicted data. Known EDCs are collected as positive samples. Due to the lack of clear non-EDC data, it is necessary to construct assumed non-EDCs as negative samples. 81 chemicals clearly identified as EDCs by ECHA are selected as positive samples. 810 assumed non-EDCs are generated as negative samples using the DUD-E method. These assumed non-EDCs are similar to EDCs in physical characteristics but different from EDCs in topological structure to minimize the possibility that they are actually EDCs. The target activity data of positive samples (EDCs) and negative samples (assumed non-EDCs) are preprocessed.

[0054] For the missing qualitative data of target activity, the best performing model was used to fill in the missing qualitative data of target activity. A total of 64,336 missing qualitative data of target activity of 891 chemicals (81 EDCs + 810 assumed non-EDCs) were filled in. Feature engineering was performed on the preprocessed target activity data to extract meaningful feature representations. Features can be generated using methods such as molecular descriptors, molecular fingerprints or graph representations of compounds. Necessary preprocessing operations such as normalization and standardization were performed on the features. A machine learning algorithm suitable for processing target activity data was selected, such as feedforward neural networks, convolutional neural networks or graph neural networks. According to the characteristics of the data and the requirements of the task, a suitable network architecture and hyperparameters were designed. The preprocessed target activity dataset was divided into training set, validation set and test set. The qualitative prediction model of endocrine disruption effects was trained using the training set, and hyperparameters were tuned on the validation set. The performance of the model was evaluated on the test set, using appropriate evaluation indicators such as accuracy, precision, recall, F1 value, etc.

[0055] like Figure 6 As shown, different architectures are used to construct qualitative prediction models for endocrine disruption effects and compare performance. The new GCNConvEdge architecture developed in this application is used to construct a qualitative prediction model for endocrine disruption effects based on the adverse outcome path (AOP) network. The GCNConvEdge architecture is a variant of a graph convolutional neural network (GCN), which can simultaneously consider node features and edge features to better capture information in the AOP network. The AOP network is represented as a graph, in which nodes represent key events and edges represent causal relationships between key events. The GCNConvEdge layer is used to extract features and transfer information on the AOP network to learn the representation of nodes and edges. After the GCNConvEdge layer, a fully connected layer and a softmax activation function are used for classification to predict whether a compound has an endocrine disruption effect. A method based on the traditional GNN architecture is used to construct a qualitative prediction model for endocrine disruption effects. Traditional GNN architectures, such as graph convolutional networks (GCNs) and graph attention networks (GATs), only consider node features without considering edge features. The AOP network is represented as a graph, in which nodes represent key events and edges represent connections between key events. The traditional GNN layer is used to extract features and transfer information to the AOP network to learn the representation of nodes. After the GNN layer, a fully connected layer and softmax activation function are used for classification to predict whether the compound has an endocrine disrupting effect.

[0056] A qualitative prediction model for endocrine disrupting effects was constructed using an architecture based on machine learning (ML) ensemble methods. Ensemble methods such as random forests and gradient boosted decision trees improve performance by combining the prediction results of multiple base models. For each key event in the AOP network, the best target qualitative prediction model was used to predict the effect of the compound. The prediction results of all key events were used as features to train the ML ensemble model for classification and predict whether the compound has an endocrine disrupting effect. Models with three different architectures were trained and evaluated using the same training, validation, and test sets. The performance of different models was compared using appropriate evaluation metrics such as accuracy, precision, recall, and F1 values. Qualitative prediction models for endocrine disrupting effects were constructed using the novel GCNConvEdge architecture, the traditional GNN architecture, and the ML ensemble architecture, and their performance was compared. This comparison can reveal the key role of the AOP network in EDC identification and demonstrate the advantages of the model based on the GCNConvEdge architecture in capturing AOP network information.

[0057] In short, in the GCNConvEdge architecture, this example uses the harmful outcome pathway network as a typical directed acyclic graph, embeds the target activity as a node, and embeds the causal relationship (KER, i.e., biologically plausible connection) between events as an edge into the three-layer GCNConvEdge architecture, and finally makes qualitative predictions of EDC based on the fully connected neural network layer. Since GCNConvEdge can integrate the hierarchical biological knowledge of the harmful outcome pathway network into a unified representation, this example believes that it can provide the most comprehensive biological knowledge for EDC prediction. Unlike GCNConvEdge, the traditional GNN-based architecture (the second architecture) ignores the causal relationship between events and uses the harmful outcome pathway network as an undirected acyclic graph. This architecture learns node embeddings from the harmful outcome pathway network, ignores the causal relationship between events, and uses four different GNN algorithms for EDC prediction. In the ML-based ensemble architecture (the third architecture), the KER information is completely removed, only the event information is retained, and it is converted into fixed tabular data. The third architecture only learns node information from the harmful outcome pathway network for EDC prediction.

[0058] Three different architectures were used to process adverse outcome pathway (AOP) networks and predict endocrine disruptors (EDCs): Based on the GCNConvEdge architecture: The AOP network was represented as a directed acyclic graph (DAG), where nodes represent key events and edges represent causal relationships (KERs, i.e., biologically plausible connections) between events. Target activity was embedded as nodes and KERs as edges, and this information was input into the GCNConvEdge architecture. The GCNConvEdge architecture consists of three GCNConvEdge layers, which can process node embedding and edge embedding simultaneously to capture the hierarchical biological knowledge in the AOP network. After the GCNConvEdge layer, a fully connected neural network layer was used for qualitative prediction of EDCs. By integrating the hierarchical biological knowledge of the AOP network into a unified representation, the GCNConvEdge architecture can provide the most comprehensive biological knowledge for EDC prediction.

[0059] Based on the traditional GNN architecture: The AOP network is represented as an undirected acyclic graph, where nodes represent key events and edges represent connections between events (ignoring causal relationships). Target activity is embedded as a node, and this information is input into the traditional GNN architecture. Traditional GNN architectures such as graph convolutional networks (GCNs) and graph attention networks (GATs) only consider node embeddings and ignore edge embeddings (i.e., causal relationships between events). Four different GNN algorithms (such as GCN, GAT, GraphSAGE, and GIN) are used to perform feature extraction and information transfer on the AOP network to learn the representation of nodes. After the GNN layer, a fully connected neural network layer is used for qualitative prediction of EDC. Due to the ignoring of causal relationships between events, the traditional GNN architecture may not be able to fully utilize the biological knowledge in the AOP network.

[0060] Based on ML ensemble architecture: extract key event information from AOP network, convert it into fixed tabular data, and completely delete KER information. Use the best target qualitative prediction model to predict the target activity corresponding to each key event. Use the predicted target activity as a feature to build a model based on machine learning (ML) ensemble methods, such as random forest, gradient boosting decision tree, etc. Use ML ensemble model for EDC qualitative prediction, and improve performance by combining the prediction results of multiple basic models. Since KER information is completely deleted, ML ensemble architecture may not be able to capture the causal relationship and biological knowledge in AOP network. By comparing the above three architectures, it can be seen that GCNConvEdge architecture has obvious advantages in utilizing the hierarchical biological knowledge of AOP network. It can simultaneously consider node information (key events) and edge information (KER), capture the causal relationship between events, and provide more comprehensive biological knowledge for EDC prediction. In contrast, the traditional GNN architecture ignores the causal relationship between events, while the ML ensemble architecture completely deletes KER information and may not be able to fully utilize the biological knowledge in AOP network. This comparison highlights the advantages and potential of GCNConvEdge architecture in endocrine disruptor identification.

[0061] Performance comparison and results of different architectures on the qualitative prediction task of endocrine disruption effects are shown in Figure 2. Figure 7As shown in the figure, the superior performance of the GCNConvEdge architecture: the model based on the GCNConvEdge architecture outperforms other architectures in all evaluation metrics (ROC-AUC, accuracy, and F1-score). The superior performance of the GCNConvEdge architecture is attributed to its ability to exploit the complete hierarchical biological knowledge in the adverse outcome pathway (AOP) network, including key event information and causal relationships (KER) between events. Specifically, the model of the GCNConvEdge architecture achieved 0.90, 0.96, and 0.78 in the ROC-AUC, accuracy, and F1-score indicators, respectively. Compared with the GCNConvEdge architecture, other architectures (traditional GNN architectures and ML integrated architectures) performed poorly in various indicators. The values ​​of other architectures ranged from 0.86-0.89 in the ROC-AUC indicator, 0.92-0.95 in the accuracy indicator, and 0.65-0.75 in the F1-score indicator. The performance differences of these architectures may stem from their limitations in processing AOP network information, such as ignoring the causal relationship between events or completely removing KER information. The indispensability of each component of the AOP network: The superior performance of the GCNConvEdge architecture on all indicators highlights the indispensability of each component (key events and KERs) in the AOP network. By simultaneously considering key event information and causal relationships between events, the GCNConvEdge architecture is able to more comprehensively utilize the hierarchical biological knowledge of the AOP network, thereby obtaining more accurate prediction results. Importance of event information: When the integrated model only learns half of the event information, its prediction performance is significantly reduced compared to the GCNConvEdge architecture. Compared with the GCNConvEdge architecture, the integrated model decreased by 0.14, 0.10, and 0.16 in the ROC-AUC, accuracy, and F1-score indicators, respectively. This result shows that event information is crucial for accurately predicting endocrine disruption effects, and removing part of the event information significantly affects the performance of the model. The model based on the GCNConvEdge architecture performed best on the qualitative prediction task of endocrine disruption effects, thanks to its ability to fully utilize the complete hierarchical biological knowledge in the AOP network.

[0062] Construct a quantitative prediction model for targets, such as Figure 8As shown, the quantitative data in the first data set after pretreatment are randomly divided into a training set and a test set. 80% of the compounds are selected as the training set for QSAR modeling and model training. The remaining 20% ​​of the compounds are used as the test set to evaluate the performance of the developed model. Five classic machine learning (ML) algorithms and one deep learning (DL) algorithm are selected to construct a target quantitative prediction model. Classical ML algorithms include random forest (RF), decision tree (DT), Xgboost, naive Bayes (NB) and k-nearest neighbor (kNN). The DL algorithm is a fully connected neural network (FCNN). Six quantitative QSAR models are constructed for each key event, including five models based on ML algorithms and one model based on DL algorithms. A total of 312 regression models (52 key events × 6 algorithms) were constructed. For each key event, the same training set and test set were used to train and evaluate the six models. The R2 indicator (coefficient of determination) was used to evaluate the performance of the six models for each key event. For each key event, the model with the highest R2 indicator was selected as the best prediction model. The performance indicators of each best model were recorded, including the mean absolute error (MAE), mean square error (MSE), and R2 indicators. The performance of the best prediction models for 52 key events was analyzed. The mean and standard deviation of the MAE, MSE, and R2 indicators of the 52 models were calculated. The results showed that the average MAE of the 52 models was 0.18±0.08, the average MSE was 0.08±0.05, and the average R2 was 0.58±0.20, indicating that these models had high predictive performance. Further analysis found that 50 (96.15%) of the 52 best models were built using tree-based algorithms (such as RF, DT, and Xgboost). Six quantitative QSAR models were constructed for each key event using machine learning algorithms and machine learning algorithms, and the model with the best performance was selected as the final target quantitative prediction model. The results showed that these models had high predictive performance, among which the models built using tree-based algorithms accounted for the highest proportion.

[0063] Use the qualitative prediction model of endocrine disruption effect to predict whether the test compound has endocrine disruption effect; input the test compound into the constructed qualitative prediction model of endocrine disruption effect. The model outputs the prediction result of whether the test compound has endocrine disruption effect. Input the compound qualitatively predicted to have endocrine disruption effect into the constructed target quantitative prediction model. The model outputs the quantitative activity value of each key event to form a quantitative adverse outcome pathway (qAOP). Analyze the quantitative dose-effect relationship of each event in the qAOP. Determine the adverse outcome (AO) and qAOP that are most likely to be activated by EDCs at the lowest environmental concentration as sensitive AO and sensitive qAOP.

[0064] Hazard classification of endocrine disruptors based on sensitive AO and sensitive qAOP: Hazard classification of EDCs based on the type and severity of sensitive AO. Determine the lowest no-observable-adverse-effect level (NOAEL) of EDCs based on the quantitative dose-effect relationship in the sensitive qAOP. EDCs are classified into high hazard (NOAEL≤1mg / kg bw / day), medium hazard (NOAEL>1mg / kg bw / day) and other levels based on NOAEL. Comprehensively consider the qualitative prediction results of EDCs, sensitive AO, sensitive qAOP and hazard classification results. Output the endocrine disrupting effect prediction results of each compound to be tested, including whether it has endocrine disrupting effect, main sensitive AO, sensitive qAOP and hazard level.

[0065] This application predicts potential EDCs from the global existing chemical inventory: three chemical lists from China, the United States and the European Union were collected and sorted, with a total of 225,814 chemical substances with unique structures. A qualitative prediction model for endocrine disruption effects based on the harmful outcome pathway network was used to predict whether each chemical substance had an endocrine disruption effect. The results showed that a total of 6,461 (2.86%) chemicals were characterized as potential EDCs.

[0066] Prioritization and hazard grading of potential EDCs: Three quantitative rules were used to prioritize 6461 EDCs. The results showed that a total of 293 qAOPs were activated by at least one EDC, of ​​which 73 qAOPs could be activated by more than half of the EDCs. EDCs tended to activate 40 sensitive qAOPs, which in turn produced 40 sensitive AOs, with NOAELs ranging from 430 mg / kg bw / day (medium hazard) to 1 mg / kg bw / day (high hazard). Reproductive effects (3986 EDCs, 61.69%) and developmental effects (765 EDCs, 11.84%) were the most sensitive AOs.

[0067] Analysis of the gender-specific toxic effects of EDCs on the reproductive system: 2535 EDCs mainly had harmful effects on the male reproductive system (such as the testicles and seminal vesicles). 1435 EDCs mainly had harmful effects on the female reproductive system (such as the ovaries and uterus). Only 99 EDCs had harmful effects on both the male and female reproductive systems (such as impaired fertility). The remaining 1710 EDCs (26.47%) mainly affected the adrenal and pituitary glands, which play a key role in the HPG and HPA axes and are essential for endocrine balance.

[0068] The qualitative prediction model of endocrine disrupting effects and the quantitative prediction model of targets were used to predict the endocrine disrupting effects and classify the hazards of the tested compounds. The case analysis results showed that this method can effectively screen out potential EDCs from the global chemical inventory and determine their sensitive AO, sensitive qAOP and hazard level. At the same time, the analysis also revealed the gender-differentiated toxic effects of EDCs on the reproductive system, providing an important basis for the risk assessment and management of endocrine disruptors.

[0069] The invention of the present application and its implementation methods are described schematically above. The description is not restrictive. Without departing from the spirit or basic features of the present application, the present application can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention of the present application. The actual structure is not limited to this. Any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the present invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all fall within the scope of protection of this patent. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A method for predicting endocrine disruptors based on a harmful outcome pathway network, characterized in that: include: According to the pre-constructed harmful outcome pathway network, experimental activity data is obtained as a first data set, and the first data set is pre-processed; wherein the first data set includes qualitative data and quantitative data; Obtaining endocrine disruptor inventory data as a second data set; Based on the qualitative data in the preprocessed first data set, a machine learning algorithm is used to construct a qualitative prediction model for the target; Using the target qualitative prediction model, predicting target activity data of the second data set to obtain a target activity data set; Using the target activity dataset and machine learning algorithms, a qualitative prediction model for endocrine disruption effects was constructed; According to the quantitative data in the preprocessed first data set, a machine learning algorithm is used to construct a target quantitative prediction model; Use the qualitative prediction model of endocrine disrupting effect to predict whether the tested compound has endocrine disrupting effect; For compounds with endocrine disrupting effects, the target quantitative prediction model is used to predict the corresponding target quantitative activity and obtain the quantitative harmful outcome pathway; According to the quantitative dose-effect relationship of each event in the quantitative harmful outcome pathway, sensitive harmful outcomes and sensitive quantitative harmful outcome pathways are obtained; According to the sensitive harmful outcomes and sensitive quantitative harmful outcome pathways, the hazards of endocrine disruptors are graded and the prediction results of endocrine disruptors are obtained.

2. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 1, characterized in that: Experimental activity data include data related to molecular initiation events, key events, and adverse outcome events; Molecular initiation events represent the starting events in the network of harmful outcome pathways; Key events represent event nodes in the harmful outcome pathway network that reflect changes in chemical substances and ultimately lead to harmful outcomes; The adverse outcome event represents the endpoint event in the adverse outcome pathway network.

3. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 2, characterized in that: The first data set is preprocessed, including: Conduct quality assessment of experimental activity data; Based on the quality assessment results, the experimental activity data were screened; The screened experimental activity data is used as the first data set after preprocessing.

4. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 1, characterized in that: Use machine learning algorithms to build a quantitative prediction model or a qualitative prediction model for targets, including: A molecular descriptor-based modeling method or a molecular graph-based modeling method is used to construct a target quantitative prediction model or a target qualitative prediction model.

5. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 4, characterized in that: A molecular descriptor-based modeling approach is used to construct a quantitative target prediction model or a qualitative target prediction model, including: Convert the chemical structure of a compound into a numerical molecular descriptor or molecular fingerprint; Based on molecular descriptors or molecular fingerprints, a machine learning algorithm is used to construct a target quantitative prediction model or a target qualitative prediction model.

6. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 4, characterized in that: Sampling molecular graph-based modeling methods to build target quantitative prediction models or target qualitative prediction models, including: Convert the chemical structure of the compound into a molecular graph structure, where atoms are represented as nodes and bonds are represented as edges; According to the molecular graph structure, the graph neural network (GNN) algorithm is used to construct a target quantitative prediction model or a target qualitative prediction model.

7. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 1, characterized in that: Using the target activity dataset and machine learning algorithms, a qualitative prediction model for endocrine disruption effects was constructed, including: The harmful outcome pathway network is regarded as a directed acyclic graph, the target activity data is used as the node input information, the causal relationship between key events is used as the edge input information, and whether there is an endocrine disruption effect is used as the output information.

8. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 1, characterized in that: Obtain sensitive adverse outcomes and sensitive quantitative adverse outcome pathways, including: According to the quantitative dose-effect relationship of each event in the quantitative adverse outcome pathway, the adverse outcome with the lowest highest no-effect concentration is obtained as the sensitive adverse outcome; Based on the sensitive harmful outcomes, the corresponding quantitative harmful outcome pathway is determined as the sensitive quantitative harmful outcome pathway, and the sensitive quantitative harmful outcome pathway includes: sensitive molecular initiating events, sensitive key events and sensitive harmful outcomes; Compare whether the highest no-effect concentration values ​​of sensitive molecular initiating events, sensitive key events and sensitive adverse outcomes in the sensitive quantitative adverse outcome pathway meet the following requirements: The highest no-effect concentration value of the sensitive molecule initiating event is less than the highest no-effect concentration value of the sensitive key event, which is less than the highest no-effect concentration value of the sensitive adverse outcome; if these conditions are met, the corresponding pathway will be taken as the sensitive quantitative adverse outcome pathway.

9. The method for predicting endocrine disruptors based on harmful outcome pathway network according to claim 1, characterized in that: The prediction results of endocrine disruptors are obtained, including: Obtain the highest no-effect concentration for each key event in the sensitive quantitative adverse outcome pathway; Using the uncertainty factor method, the highest no-effect concentration was extrapolated to the sensitive and harmful outcomes to obtain the highest no-effect concentration of the sensitive and harmful outcomes based on each key event; Obtain the minimum value of the highest no-effect concentration of the sensitive harmful outcomes of each key event as the hazard classification threshold of the corresponding compound; The tested compounds were classified according to the hazard classification thresholds.

10. A system for predicting endocrine disruptors, characterized in that: include: A data acquisition module, which acquires experimental activity data as a first data set, acquires endocrine disruptor inventory data as a second data set, and preprocesses the first data set; wherein the first data set includes qualitative data and quantitative data; A first data set processing module, which uses a machine learning algorithm to construct a target qualitative prediction model based on the qualitative data in the preprocessed first data set; A second data set processing module, using a target qualitative prediction model to perform target activity data prediction on the second data set to obtain a target activity data set; The third data set processing module uses the target activity data set and a machine learning algorithm to build a qualitative prediction model for endocrine disruption effects; A fourth data set processing module, based on the quantitative data in the preprocessed first data set, uses a machine learning algorithm to construct a target quantitative prediction model; The prediction module uses the qualitative prediction model of endocrine disrupting effect to predict whether the test compound has endocrine disrupting effect; for compounds with endocrine disrupting effect, the target quantitative prediction model is used to predict the corresponding target quantitative activity to obtain the quantitative harmful outcome path; based on the quantitative dose-effect relationship of each event in the quantitative harmful outcome path, sensitive harmful outcomes and sensitive quantitative harmful outcome paths are obtained; endocrine disruptors are classified according to the sensitive harmful outcomes and sensitive quantitative harmful outcome paths to obtain the prediction results of endocrine disruptors.

Images