Endocrine disrupter identification method and system

Abstract

The invention discloses an endocrine disrupter identification method and system, and the method comprises the steps: carrying out the multi-dimensional feature extraction of a to-be-identified target compound, and obtaining the features of the target compound; respectively inputting the target compound characteristics into an activity prediction model corresponding to each prediction end point type to obtain an activity probability prediction result corresponding to each prediction end point; generating a recognition result of the target compound corresponding to each activity probability prediction result; according to the endocrine disrupter identification method, the identification efficiency of the endocrine disrupter is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for identifying endocrine disruptors. Background Technology

[0002] Endocrine-disrupting chemicals (EDCs) are a class of pollutants that are widely present in the environment and pose a significant threat to global public health. They can enter organisms through air, water, soil, and food, and specifically bind to targets of nuclear receptors (NRs, such as estrogen receptor ER, androgen receptor AR, peroxisome proliferator-activated receptor PPAR, etc.) or non-nuclear receptors (nonNRs, such as oxidative stress-related receptors, DNA damage repair-related proteins, etc.), interfering with cell signal transduction pathways and normal cell fate regulation, thereby causing endocrine system disorders.

[0003] Currently, there are only a few thousand known EDC species, while the number of potential EDC species in the environment is as high as tens of thousands or even hundreds of thousands. Moreover, their mechanisms of action are highly diverse, making the efficient identification, mechanism analysis, and multi-system toxicity assessment of EDCs a core challenge in the field of environmental toxicology.

[0004] In existing technologies, the screening and toxicity assessment of endocrine disruptors (EDCs) are primarily based on in vitro cell experiments, determining EDC properties by detecting the effects of compounds on the activity of specific receptors. While such methods can directly reflect the endocrine-disrupting effects of compounds, a single experiment can only evaluate a few compounds, which cannot meet the screening needs of the massive number of potential EDCs. Summary of the Invention

[0005] This invention provides a method and system for identifying endocrine disruptors, addressing the technical problem of how to improve existing methods for identifying endocrine disruptors and thereby increasing the efficiency of identifying a large number of potential endocrine disruptors.

[0006] To address the aforementioned technical problems, the present invention provides a method for identifying endocrine disruptors, comprising: In response to the compound recognition signal, multidimensional feature extraction is performed on the target compound to be identified to obtain the target compound features; The target compound features are input into the activity prediction model corresponding to each prediction endpoint type to obtain the activity probability prediction results corresponding to each prediction endpoint. Generate identification results for the target compound corresponding to each of the predicted activity probabilities; The training process of the activity prediction model includes: The acquired data related to adverse outcome pathways of endocrine disruptors were analyzed and processed, and a global AOP network was constructed based on the analysis results; Enrichment analysis is performed on the node gene set of the global AOP network to obtain key node data and determine the predicted endpoint type corresponding to the key node data; wherein, the enrichment analysis is used to screen key node data associated with key target genes of endocrine disruptors. Several sets of interaction data between compounds and their corresponding toxicity endpoints are obtained from a preset database. The interaction data are then standardized based on the predicted endpoint type to obtain multiple sets of correlation data between standard compounds and their corresponding predicted endpoints. Each set of correlation data corresponds to a type of predicted endpoint. The associated data is subjected to multidimensional feature extraction processing to obtain multiple sets of multidimensional molecular feature data; The pre-constructed initial activity prediction model is trained based on the multidimensional molecular feature data of each group to obtain the activity prediction model corresponding to each prediction endpoint type.

[0007] As one preferred embodiment, the step of parsing and processing the acquired data related to adverse outcome pathways of endocrine disruptors, and constructing a global AOP network based on the parsing results, includes: Obtain adverse outcome path-related data of endocrine disruptors from the AOP-Wiki database, and parse the adverse outcome path-related data to obtain initial parsing data, wherein the initial parsing data includes at least key events and key event relationships; The initial parsed data is structurally transformed using the key events as nodes and the key event relationships as edges to obtain an AOP directed graph. The AOP directed graph is then processed to construct a network topology to obtain an initial global AOP network. Calculate the node distances in the initial global AOP network, and optimize the initial global AOP network based on the node distances to obtain the global AOP network. The optimization process is used to treat nodes with a distance of 1 as direct associations and nodes with a distance of 2 as indirect associations.

[0008] As one preferred embodiment, the enrichment analysis of the node gene set of the global AOP network to obtain key node data includes: Gene set matching is performed on each key event node in the global AOP network to obtain the node gene set corresponding to each key event node. Identifying key target genes associated with endocrine disruptor perturbations based on protein-protein interactions; Enrichment analysis was performed on each of the node gene sets, and key nodes associated with the target gene were screened based on the enrichment analysis results to obtain key node data. The key node data includes nuclear receptor pathway nodes directly associated with the mechanism of action of nuclear receptor-mediated endocrine disruptors and non-nuclear receptor pathway nodes directly associated with the mechanism of action of non-nuclear receptor-mediated endocrine disruptors.

[0009] As one preferred embodiment, the identification of key target genes related to endocrine disruptor disturbances based on protein-protein interactions includes: Obtain initial endocrine disruptor data from a preset multi-source database, perform deduplication on the initial endocrine disruptor data, and generate an endocrine disruptor list; Gene association data corresponding to the endocrine disruptor list were extracted from the CTD database to construct an endocrine disruptor-associated gene set. The endocrine disruptor-associated gene set was mapped to a protein interaction network, and key target genes related to endocrine disruptor perturbation were screened based on the interaction output results.

[0010] As one preferred embodiment, the predicted endpoint types include nuclear receptor-mediated molecular initiation event predicted endpoints, non-nuclear receptor-mediated molecular initiation event predicted endpoints, key toxicity mechanism nodes of endocrine disruptors predicted endpoints, and multi-endpoint toxicity node predicted endpoints. Determining the prediction endpoint type corresponding to the key node data includes: A specific AOP network is constructed based on key node data, and nuclear receptor-mediated molecular initiation event nodes and non-nuclear receptor-mediated molecular initiation event nodes are extracted from the specific AOP network. Topology analysis was performed on the specific AOP network, and key toxicity mechanism nodes of endocrine disruptors were screened based on the topology analysis results. Extract the set of toxic outcome nodes from the specific AOP network, and cluster the set of toxic outcome nodes based on a preset toxicity system to obtain multi-endpoint toxic nodes; The nuclear receptor-mediated molecule initiation event nodes, non-nuclear receptor-mediated molecule initiation event nodes, key toxicity mechanism nodes of endocrine disruptors, and multi-endpoint toxicity nodes were labeled to obtain the types of predicted endpoints.

[0011] As one preferred embodiment, the construction of a specific AOP network based on key node data includes: The key node data is sequentially subjected to hierarchical limitation and redundancy merging to obtain the first key node data; Extract the first key event relationship data from the global AOP network that has a direct or indirect relationship with the first key node data, and remove the self-loops in the first key event relationship data to obtain the second key event relationship data; The adverse outcome nodes in the first key node data are categorized and replaced to obtain the second key node data. Construct a special node subnetwork based on the second key node data and the second key event relationship data; The specific node subnetwork is initialized as a directed graph, and the directed graph is filtered for effective paths to obtain a specific AOP network, wherein the specific AOP network includes a nuclear receptor pathway specific AOP network and a non-nuclear receptor pathway specific AOP network.

[0012] As one preferred embodiment, the standardization of the interaction data based on the predicted endpoint type yields multiple sets of associated data for standard compounds and their corresponding predicted endpoints, including: The interaction correlation data between compounds and their corresponding toxic endpoints in a preset multi-source database is obtained, and the interaction correlation data is integrated and processed to obtain an initial interaction dataset. The initial interactive dataset is sequentially subjected to unified identification processing, contradiction removal processing, and label normalization processing to obtain a structured standardized dataset. Based on the predicted endpoint type, the structure-normalized dataset is matched and associated to obtain multiple sets of associated data; wherein, the matching and association process is designed to correspond one-to-one between compound data and the activity tags of the corresponding predicted endpoint, and each set of associated data corresponds to a predicted endpoint type.

[0013] As one preferred embodiment, the multidimensional feature extraction process for the associated data includes: For each set of associated data, 1D hybrid fingerprint feature extraction, 2D molecular graph feature extraction and 3D molecular space feature extraction are performed respectively to obtain 1D molecular fingerprint feature data, 2D molecular graph feature data and 3D molecular space feature data corresponding to each set of associated data; The 1D molecular fingerprint feature data, the 2D molecular graph feature data, and the 3D molecular spatial feature data corresponding to each set of associated data are integrated and processed to generate multiple sets of multidimensional molecular feature data, wherein each set of multidimensional molecular feature data corresponds to an associated data of a prediction endpoint type.

[0014] As one preferred embodiment, the generation of identification results for the target compounds corresponding to each of the activity probability prediction results specifically includes: Based on the activity probability prediction results corresponding to each prediction endpoint and the preset judgment rules, the target compound is subjected to endocrine disruptor category identification processing to obtain endocrine disruptor category identification results. Based on the endocrine disruptor category identification results and their corresponding activity probability prediction results, the importance of the target compound is calculated to obtain the endocrine disruptor importance score. The activity probability prediction results corresponding to the predicted endpoint, the endocrine disruptor category identification results, and the endocrine disruptor importance scores are integrated to generate the identification results of the target compound.

[0015] Another aspect of the present invention provides an endocrine disruptor identification system, comprising: The acquisition module is used to extract multi-dimensional features of the target compound to be identified in response to the compound recognition signal, so as to obtain the features of the target compound. The prediction module is used to input the characteristics of the target compound into the activity prediction model corresponding to each prediction endpoint type to obtain the activity probability prediction result corresponding to each prediction endpoint. An identification module is used to generate identification results for the target compound corresponding to each of the activity probability prediction results; The prediction module further includes: The parsing unit is used to parse and process the acquired data related to adverse outcome pathways of endocrine disruptors, and to construct a global AOP network based on the parsing results; An analysis unit is used to perform enrichment analysis on the node gene set of the global AOP network to obtain key node data and determine the predicted endpoint type corresponding to the key node data; wherein, the enrichment analysis is used to screen key node data associated with key target genes of endocrine disruptors. A standardization unit is used to acquire several sets of interaction data between compounds and their corresponding toxicity endpoints in a preset database, and to standardize the interaction data based on the predicted endpoint type to obtain multiple sets of association data between standard compounds and their corresponding predicted endpoints; wherein each set of association data corresponds to a type of predicted endpoint. The extraction unit is used to perform multidimensional feature extraction processing on the associated data to obtain multiple sets of multidimensional molecular feature data; The training unit is used to train the pre-constructed initial activity prediction model based on the multidimensional molecular feature data of each group, so as to obtain the activity prediction model corresponding to each prediction endpoint type.

[0016] Compared with the prior art, the present invention has the following advantages: 1) This invention takes the adverse outcome pathway (AOP) of endocrine disruptors as its core. Through global AOP network analysis and node gene set enrichment analysis, it accurately screens key node data associated with key target genes, ensuring that the predicted endpoint type is highly consistent with the biological mechanism of endocrine disruption. At the same time, it combines compound-toxic endpoint interaction data from a pre-set database, and after standardization and multi-dimensional feature extraction, it provides a comprehensive and standardized training data foundation for the model, effectively reducing the impact of data heterogeneity on prediction accuracy.

[0017] 2) This invention effectively solves the problems of low efficiency, high cost, and limited coverage in traditional endocrine disruptor identification methods. Traditional methods are difficult to meet the needs of large-scale compound screening, while this invention, by constructing an efficient activity prediction model, can directly and quickly output the activity probability results of multiple prediction endpoints based on the multidimensional features of compounds, significantly shortening the identification cycle and reducing experimental costs. It is especially suitable for the preliminary screening of unknown compounds or a large number of candidate compounds in the environment. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the endocrine disruptor identification method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the MF3DGNet structure and molecular characterization in one embodiment of the present invention. Figure 3 This is a schematic diagram of an endocrine disruptor identification system in one embodiment of the present invention; Figure label: Among them, 11 is the acquisition module; 12 is the prediction module; and 13 is the recognition module. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0023] One embodiment of the present invention provides a method for identifying endocrine disruptors. For details, please refer to [link to specific documentation]. Figure 1 , Figure 1 The diagram shown is a flowchart of an endocrine disruptor identification method according to one embodiment of the present invention, which includes steps S1-S3: S1: In response to the compound recognition signal, perform multidimensional feature extraction on the target compound to be identified to obtain the target compound features; In this embodiment, the compound identification signal is actively triggered by the user through the EDC-MF3DNet interactive computing platform. There are two specific input methods: First, the molecular structure of the target compound is directly drawn through the JSME chemical editor integrated in the platform front end, and the editor converts the structure into standard SMILES (CanonicalSMILES) in real time; Second, the standard SMILES text of the target compound is directly uploaded in batches (the number of batch uploads is <50 to avoid computational overload), and the platform back end verifies the validity of the SMILES format through the RDKit tool.

[0024] Based on the validated standard SMILES, features were extracted in three steps using tools such as RDKit and PyBioMed, and finally integrated into the target compound features.

[0025] Specifically, the invention first performs 1D hybrid fingerprint feature extraction. Two one-dimensional (1D) molecular fingerprints are used to characterize prior molecular features: the ECFP fingerprint is used to capture information such as atomic connectivity, while the PubChem fingerprint focuses on identifying information such as functional groups. Furthermore, the molecular structure is modeled as a graph, with atoms in the molecule considered as nodes and bonds as edges. in, Indicates atoms in a molecule. Indicates the bonds in a molecule. This represents the embedding vector dimension. Node features consist of 16 atomic attributes (such as atom type, number of free radical electrons, etc.), with some atomic attributes using one-hot encoding. Edge features are represented by four bond attributes (bond type, 3D features, conjugated bonds, and ring states). Node and edge features are integrated into a 2D molecular graph feature matrix. Furthermore, RDKit is used to calculate the three-dimensional coordinates of each atom. As a vector feature , representing the 3D characteristics of molecules.

[0026] The 1D hybrid fingerprint feature vector, 2D molecular graph feature matrix, and 3D spatial feature vector are dimensionally aligned to form a multi-dimensional fused target compound feature. Detailed molecular features are shown in Table 1.

[0027] Table 1 S2: Input the target compound features into the activity prediction model corresponding to each prediction endpoint type to obtain the activity probability prediction results corresponding to each prediction endpoint; In one embodiment of the present invention, the predicted endpoint types are divided into four categories, and each category of predicted endpoints corresponds to one independently trained active prediction model (not a single model covering all endpoints). The specific matching relationship is as follows: Nuclear receptor-mediated molecular initiation events (NR-MIE) were used to predict the effects of compounds on 20 NRs (such as ER). AR, PPAR The three modes of action (agonist, antagonist, and binding) of the compound correspond to 20 × 3 = 60 models; non-nuclear receptor-mediated molecular initiation events (nonNR-MIE) are used to predict the activation activity of the compound for "increased oxidative stress" and "increased DNA damage", corresponding to 2 models; key toxicity mechanism nodes (KE) of the EDC are used to predict the activation activity of the compound for "increased apoptosis" and "activation of the p53 signaling pathway", corresponding to 2 models; and multi-endpoint toxicity (AO) is used to predict the eight types of organ toxicity of the compound (such as hepatotoxicity, carcinogenicity, and reproductive toxicity), corresponding to 8 models.

[0028] All activity prediction models are based on the MF3DNet (Multi-Feature Fusion 3Draph Neural Network) architecture and are trained independently for datasets with different prediction endpoints. The core computation process is as follows: The 1D hybrid fingerprint features of the target compound are further learned through independent linear layers to strengthen the association between functional groups and activity. The 3D spatial feature vector and the 2D molecular graph feature matrix are input into the graph attention convolution (ATConv) layer, which automatically assigns weights to high-contribution atoms through the attention mechanism to highlight the structural fragments that are key to the prediction endpoint. The fused 2D+3D features are output as molecular-level feature vectors through graph max pooling and average pooling, and then concatenated with the 1D feature vectors to form full-dimensional fused features.

[0029] The concatenated feature input is processed by a four-layer fully connected network that learns nonlinear associations step by step using the ReLU activation function. The last layer maps the output to probability values ​​in the [0,1] interval using the Softmax function, which is the activity probability prediction result. A probability ≥ 0.5 indicates activity for the predicted endpoint, while a probability < 0.5 indicates inactivity. Each predicted endpoint outputs one activity probability value, and the probability results of all endpoints are summarized into a probability matrix.

[0030] S3: Generate the identification results of the target compounds corresponding to each activity probability prediction result; Preferably, in one embodiment of the present invention, generating identification results of the target compound corresponding to each activity probability prediction result specifically includes: Based on the predicted activity probability results corresponding to each predicted endpoint and the preset judgment rules, the target compound is subjected to endocrine disruptor category identification processing to obtain the endocrine disruptor category identification results; specifically, based on the activity probability matrix obtained in step S2, the target compound is divided into 3 categories of EDC according to the rules, and the specific judgment conditions are as follows: Potential EDC (pEDC): Only meets the "endocrine disruption toxicity prediction endpoint activity probability ≥ 0.5" in the multi-endpoint toxicity test, and has no NR-MIE / nonNR-MIE activity; Potential nuclear receptor-mediated EDC (NR-pEDC): simultaneously satisfying "at least one NR-MIE predictive endpoint activity probability ≥ 0.5 (any mode of agonism / antagonism / binding)" + "endocrine disruption toxicity probability ≥ 0.5"; Potential non-nuclear receptor-mediated EDC (nonNR-pEDC): simultaneously satisfying "at least one nonNR-MIE predictive endpoint activity probability ≥0.5 (either increased oxidative stress or increased DNA damage)" + "endocrine disruption toxicity probability ≥0.5".

[0031] The importance of the target compound is calculated based on the endocrine disruptor category identification results and their corresponding activity probability prediction results, resulting in an endocrine disruptor importance score. Specifically, the risk priority of the target compound is quantified by calculating the EDC ranking score (S_EDC), where the EDC ranking score calculation formula for potential nuclear receptor-mediated EDC (NR-pEDC) is expressed as follows: in, To predict the total number of NR-MIEs interacting with the compound, For the compound to the first The predicted probability of individual phase-dependent NR-MIE.

[0032] The formula for calculating the EDC ranking score of potential non-nuclear receptor-mediated EDC (nonNR-pEDC) is expressed as follows: in, To predict the total number of NR-MIEs interacting with the compound, For the compound to the first The predicted probability of individual phase-dependent NR-MIE.

[0033] The formula for calculating the EDC ranking score of potential EDCs (pEDCs) is expressed as follows: in, This is the rating coefficient. For the probability weight of endocrine-disrupting toxic AO, This is the predicted probability of endocrine disruption toxicity AO. The weights for the NR-MIE portion, For nonNR-MIE weights, and All are greater than 0.5.

[0034] The activity probability prediction results corresponding to the predicted endpoint, the endocrine disruptor category identification results, and the endocrine disruptor importance scores are integrated to generate the target compound identification results.

[0035] The identification results of the target compounds are presented in the form of structured reports and visual charts, specifically including: Basic information: Target compound name, PubChemCID, standard SMILES; Classification results: EDC category (pEDC / NR-pEDC / nonNR-pEDC) and determination criteria (e.g., "NR-pEDC, because ER..."). (Binding activity 0.85 + endocrine toxicity 0.72”). Probability details: Activity probability of each predicted endpoint (e.g., ER in NR-MIE) (Binding 0.85, AR antagonism 0.32; oxidative stress in nonNR-MIE 0.41; hepatotoxicity in AO 0.62). Sorting information: S_EDC score and global ranking (e.g., S_EDC=3.83, ranked 1st); Visualization charts: Molecular structure heatmaps are generated using the MolDraw2DCairo engine in RDKit (atomic colors change from blue to red according to weight, with darker red indicating a higher contribution to EDC activity; for example, the carboxyl group in MEHP is dark red).

[0036] The training process of the activity prediction model includes: Step S21: Analyze and process the obtained data related to adverse outcome pathways of endocrine disruptors, and construct a global AOP network based on the analysis results; Preferably, in one embodiment of the present invention, the acquired data related to adverse outcome pathways of endocrine disruptors are parsed and processed, and a global AOP network is constructed based on the parsing results, including: Obtain adverse outcome path-related data of endocrine disruptors from the AOP-Wiki database, and parse the adverse outcome path-related data to obtain initial parsing data, which includes at least key events and key event relationships.

[0037] It should be noted that the data in this invention comes from the AOP-Wiki database, a database dedicated to adverse outcome paths. The data is obtained in XML format and contains all AOP entries related to endocrine disruptors (EDCs). In this embodiment, the xml2R package is used to perform structured parsing of the XML file to obtain overall AOP attribute information and detailed information on key events (KEs) and key event relationships (KERs).

[0038] The overall AOP attribute information includes core identifiers (AOP ID and title) and confidence status (OECD status, reflecting the scientific consensus of AOP, and SAAOP status, reflecting the effectiveness of AOP). Key event information includes the ID, title, abbreviation, and biological organizational level of each KE; it also needs to clarify the type of KE, classifying them as "Molecular Initiation Event (MIE, the starting node of AOP)," "Intermediate Key Event (KE, the transmission node of AOP)," and "Adverse Outcome (AO, the final toxicity node of AOP)." Key event relationship (KER) information includes the causal transmission direction between KEs, obtained by extracting the upstream and downstream KE-IDs of each KER.

[0039] After parsing the database, a structured dataset is formed, with each row corresponding to one AOP entry. The columns include "AOP attributes (ID / title / status)", "KE list (including the ID / type / level of each KE)" and "KER list (including upstream and downstream KE relationships)".

[0040] Furthermore, the initial parsed data is structurally transformed using key events as nodes and key event relationships as edges to obtain an AOP directed graph. This AOP directed graph is then processed for network topology construction to obtain the initial global AOP network. It should be noted that before constructing the AOP directed graph, the AOPs in the initial parsed data must be quality-screened to ensure the confidence of the subsequent network. Specific screening rules include removing invalid AOPs and removing missing AOPs. Removing invalid AOPs includes deleting AOPs with an SAAOP status of "archived" (invalid archiving), which are usually marked by the database due to insufficient scientific evidence or outdated conclusions. Removing missing AOPs includes removing AOPs lacking any of the core elements "MIE, KE, or AO" (e.g., incomplete paths containing only MIE and AO but without an intermediate KE), as missing AOPs cannot form a complete poison mechanism chain and will interfere with the network topology.

[0041] Based on the filtered high-confidence AOP, the structure is transformed according to the rule of "KE as nodes and KER as directed edges". Specifically, each KE is treated as a network node, and the node attributes carry its ID, type (MIE / KE / AO), and biological level. Each KER is treated as a directed edge, and the direction of the edge is consistent with the causal propagation direction of upstream KE to downstream KE. The relationship between the above nodes and edges is transformed into a directed graph structure through the iraphR package, forming an AOP directed graph.

[0042] The AOP directed graph is processed for network topology, and an initial global AOP network is generated using the iraphR package. The initial global AOP network needs to be visualized and verified using a biological network visualization tool (Cytoscape software). The association logic between nodes and edges is manually verified to ensure that the network structure is consistent with the actual toxicity mechanism chain of AOP.

[0043] Furthermore, the node distances in the initial global AOP network are calculated, and the initial global AOP network is optimized based on these node distances to obtain the global AOP network. The optimization process treats nodes with a distance of 1 as directly related and nodes with a distance of 2 as indirectly related. The node distance refers to the shortest path length between two KE nodes in the global AOP network (the path is determined by the number of KE nodes connected by the KER). The specific calculation rules are as follows: If two KEs are directly connected by a KER (e.g., KE1→KE2), then the node distance is 1; if two KEs are indirectly connected by an intermediate KE (e.g., KE1→KE3→KE2), then the node distance is 2; and so on.

[0044] The initial global AOP network only contains direct keyers (distance = 1). However, the poison mechanism chain of EDC contains many indirect causal relationships. To improve the comprehensiveness of keyer coverage, nodes with a distance of 2 are considered as indirect relationships and added to the initial network as potential keyers. The network topology is then updated again using iraphR packets to form the final global AOP network. The core characteristics of the global AOP network are: Covers 1080 KE nodes (including MIE, intermediate KE, and AO); Including directly associated KERs (distance=1) + indirectly associated KERs (distance=2), the KER coverage is improved by approximately 30% compared to the initial network; The network can be visualized using Cytoscape, clearly showing the direct and indirect toxicity transfer pathways from MIE to AO in EDC, providing a complete network structure foundation for subsequent screening of key nodes based on gene enrichment analysis.

[0045] S22: Enrichment analysis is performed on the node gene set of the global AOP network to obtain key node data and determine the prediction endpoint type corresponding to the key node data; among which, enrichment analysis is used to screen key node data associated with key target genes of endocrine disruptors. Preferably, in one embodiment of the present invention, enrichment analysis is performed on the node gene set of the global AOP network to obtain key node data, including: Gene set matching is performed on each key event node in the global AOP network to obtain the node gene set corresponding to each key event node. It should be noted that the KE in the global AOP network only contains biological event descriptions. Gene set matching is needed to establish the association between KE and functional genes in order to screen EDC-related nodes through gene enrichment analysis.

[0046] In this embodiment, a triple association rule is used to achieve gene set matching, ensuring that the gene set of each KE is complete and reliable. First, experimentally verified KE-gene correspondences can be directly extracted from a verified KE-gene association database, as this type of gene set has the highest reliability and requires no additional verification.

[0047] Secondly, for the KEs for which no genes were matched in the first layer, the genes were supplemented using the KE-O association file provided in the AOP-Wiki database, following the procedure below: extract the "biological process O term" corresponding to the current KE from the KE-O association file; download the ene2o.z file from the NCBI database and obtain all human genes corresponding to the O term through reverse matching of the O term; manually review and confirm the functional association between the gene and the KE to form the second gene set.

[0048] Finally, for KE terms where no genes were matched in the first two layers, supplementary genes were calculated based on the semantic similarity between the KE title and OBP terms. The specific process included: preprocessing the names of KE and OBP terms, including unifying semantic vocabulary, deleting meaningless common words, converting abbreviations to full names, and uniformly converting them to uppercase; segmenting and removing punctuation from the preprocessed KE titles and OBP terms, calculating the Jaccard similarity coefficient (number of similar words / total number of words), and setting a threshold of ≥0.3, which was verified to balance sensitivity and specificity. For OBP terms with a similarity coefficient ≥0.3, after manual confirmation of their functional consistency with KE, the corresponding genes were obtained through the ene2o.z file, forming the third gene set.

[0049] After removing duplicates from the genes obtained from the triple association, they are integrated into the complete node gene set of the current KE.

[0050] Based on protein-protein interactions, key target genes associated with endocrine disruptor perturbation can be identified. EDC induces toxic effects by perturbing specific target genes. Therefore, it is necessary to first identify the key target genes of EDC perturbation before enrichment analysis can be used to screen out KE nodes (i.e., EDC-related nodes) associated with these target genes.

[0051] Preferably, in one embodiment of the present invention, identifying key target genes associated with endocrine disruptor perturbation based on protein-protein interactions includes: Initial endocrine disruptor data is obtained from a pre-defined multi-source database. This initial data is then deduplicated to generate an endocrine disruptor list. Specifically, in this embodiment, EDC data is obtained from three authoritative sources, and the reliability of the EDCs is ensured through intersection deduplication: Source 1, the DEDuCT database; Source 2, the EU endocrine disruptor list; Source 3, relevant literature. The EDC names / CAS numbers from the three sources are matched, their intersection is taken, and duplicates are removed, ultimately yielding a number of high-confidence EDCs.

[0052] Gene association data corresponding to the endocrine disruptor list in the CTD database were extracted to construct an endocrine disruptor-associated gene set. Specifically, based on the aforementioned EDC list, associated genes were extracted from the CTD_chem_gene_ixns file in the CTD database (Comparative Toxicology Genomics Database) according to the following rules: only association data for Homosapiens (humans) were retained; only associations with ≥2 pieces of literature evidence were retained; if the same EDC-gene pair exhibited opposing action patterns such as "agonist / antagonist" or "activator / inhibitor," the action pattern with higher frequency in the literature was retained. The final endocrine disruptor-associated gene set was obtained.

[0053] The endocrine disruptor (EDC) associated gene set was mapped to a protein-protein interaction (PPI) network, and key target genes related to EDC perturbation were screened based on the interaction output. It should be noted that the EDC associated gene set includes a large number of indirectly associated genes, requiring screening using a PPI network to identify key target genes that play a central role in EDC toxicity. Specifically, the EDC associated genes were mapped to the PPI network of the STRING database (protein-protein interaction database), and the overall interaction score of each gene with other genes was obtained. A screening threshold was set, retaining genes with an overall score >400. In the STRING database, a score >400 indicates a high degree of confidence in the protein-protein interaction, excluding weak associations. Finally, a number of key target genes for EDC perturbation were identified.

[0054] Furthermore, enrichment analysis is performed on the gene sets of each node. Based on the enrichment analysis results, key nodes associated with target genes in the gene sets of each node are screened to obtain key node data. This key node data includes nuclear receptor pathway nodes directly associated with the mechanism of action of nuclear receptor-mediated endocrine disruptors and non-nuclear receptor pathway nodes directly associated with the mechanism of action of non-nuclear receptor-mediated endocrine disruptors. In this embodiment, gene set enrichment analysis (GSEA) is used to determine whether the gene set of each KE in the global AOP network is significantly associated with key target genes of EDC, thereby screening out key nodes related to EDC and distinguishing between NR / nonNR pathway nodes.

[0055] Specifically, the R package clusterProfiler was used to perform gene set enrichment analysis, using the key target gene set of EDC perturbation as the foreground gene set and the node gene set of each KE as the background gene set. The "enrichment score", "normalized enrichment score (NES)" and "original p-value" of each KE were calculated, where NES>0 indicates that the KE is in an activated state under EDC perturbation, and NES<0 indicates an inhibited state. This invention focuses on the toxic pathways activated by EDC, so only KEs with NES>0 are retained. The original p-values ​​were corrected using the Benjamini-Hochberg method to obtain the corrected q-values ​​(used to control the false discovery rate and avoid false screening).

[0056] Nodes significantly associated with key EDC target genes were screened according to the following rules to distinguish between NR and nonNR pathways: The NR pathway node screening rules include: the KE gene set must be related to the nuclear receptor (NR) signaling pathway; NES>0; q value of MIE / KE node<0.01; q value of AO node<0.25. The screening rules for nonNR pathway nodes include: the gene set of KE is not related to the nuclear receptor signaling pathway, but is related to nonNR pathways such as oxidative stress, DNA damage, and apoptosis; NES>0, MIE / KEq<0.01, and AOq<0.25.

[0057] Furthermore, to ensure that the selected nodes meet the requirements for EDC toxicity mechanism research, a secondary screening was conducted according to biological tissue levels to eliminate nodes at irrelevant levels. Specifically, for MIE nodes, only the molecular level was retained, since the molecular initiation events of EDC all occur at the molecular level; for KE nodes, the four levels of molecular, cellular, organ, and tissue were retained to focus on the microscopic toxicity transmission of EDC; and for AO nodes, the molecular level was excluded.

[0058] Redundancy was merged among the KEs describing the same biological process but with different names in the screened nodes, and the key node data obtained were: nuclear receptor (NR) pathway nodes: containing 20 NR-MIEs, 12 NR-related hub KEs, and 6 NR-related AOs; and non-nuclear receptor (nonNR) pathway nodes: containing 2 nonNR-MIEs, 8 nonNR-related hub KEs, and 2 nonNR-related AOs.

[0059] Preferably, in one embodiment of the present invention, constructing a specific AOP network based on key node data includes: The key node data is sequentially subjected to hierarchical limitation and redundancy merging to obtain the first key node data. Hierarchical limitation means that the node data after target gene screening is screened according to the biological tissue level. Redundancy merging means that KEs describing the same biological process but with different names are merged in the screened nodes.

[0060] Furthermore, the first key event relationship data with direct or indirect correlation to the first key node data in the global AOP network is extracted, and self-loops in the first key event relationship data are removed to obtain the second key event relationship data. Specifically, based on the KER association logic of the global AOP network, two types of KERs are extracted between the first key node data, namely direct association KERs (node ​​distance = 1) and indirect association KERs (node ​​distance = 2). Invalid KERs that are related to themselves by a node are removed, and finally the second key event relationship data is obtained.

[0061] The adverse outcome nodes in the first critical node data are categorized and replaced to obtain the second critical node data. Specifically, to facilitate subsequent multi-endpoint toxicity prediction, the AO nodes in the first critical node data need to be categorized and replaced according to a predefined toxicity system. For example, specific AO descriptions are mapped to corresponding toxicity categories, such as: Livercell necrosis → hepatotoxicity; Uterine adenocarcinoma → carcinogenicity; Neuronal damage → neurotoxicity. It is ensured that the replaced AO categories are consistent with the toxicological effects of the original AOs. The replaced AO nodes, together with other nodes (MIE / KE), constitute the second critical node data.

[0062] A unique node subnetwork is constructed based on the second key node data and the second key event relationship data. The second key node data (nodes) and the second key event relationship data (edges) are input into the igraphR package to construct two types of unique node subnetworks: NR path unique node subnetworks and nonNR path unique node subnetworks. The NR path unique node subnetwork contains only nodes and KERs along the NR path, forming a poison chain of "NR-MIE→hub KE→AO". The nonNR path unique node subnetwork contains only nodes and KERs along the nonNR path, forming a poison chain of "nonNR-MIE→hub KE→AO". The subnetworks are visualized using Cytoscape software, and the association logic between nodes and edges is manually verified to ensure that it conforms to the EDC poison mechanism.

[0063] Furthermore, the specific node subnetwork is initialized as a directed graph, and the directed graph is subjected to effective path filtering to obtain a specific AOP network, which includes a nuclear receptor pathway specific AOP network and a non-nuclear receptor pathway specific AOP network.

[0064] Specifically, the two types of special node sub-networks are initialized as directed graphs, where nodes = second key nodes, edges = second key nodes, and the direction of the edges is consistent with the direction of poison propagation. Valid paths starting from the MIE, propagating through the KE, and ending at the AO are filtered, while invalid paths are excluded. Specific filtering rules include: Each path has 4-15 nodes (if the number of nodes is less than 4, the mechanism chain is incomplete; if it is greater than 15, noisy nodes are easily introduced); loops in the path (such as KE1→KE2→KE1) are removed to avoid poison propagation logic loops; valid paths that meet the conditions are merged to finally obtain two types of specific AOP networks.

[0065] Preferably, in one embodiment of the present invention, determining the prediction endpoint type corresponding to the key node data includes: A specific AOP network was constructed based on key node data, and nuclear receptor-mediated molecular initiation event nodes and non-nuclear receptor-mediated molecular initiation event nodes were extracted from the specific AOP network. Specifically, from the NR pathway-specific AOP network, MIE nodes with a biological level of Molecular, directly associated with nuclear receptors, and a q value < 0.01 were screened; and NR-MIEs not covered by AOP-Wiki were supplemented by searching relevant literature, all of which served as endpoints for predicting nuclear receptor-mediated molecular initiation events. Similarly, from the nonNR pathway-specific AOP network, MIE nodes with a biological level of Molecular, unrelated to nuclear receptors, and a q value < 0.05 were screened as endpoints for predicting non-nuclear receptor-mediated molecular initiation events.

[0066] Furthermore, topology analysis was performed on the specific AOP network to screen key toxicity mechanism nodes of endocrine disruptors (EDCs). Specifically, igraph v1.0.1 was used to perform topology analysis on KE nodes in the NR / nonNR pathway-specific AOP network, calculating core metrics including in-degree (number of upstream nodes of the KE), out-degree (number of downstream nodes of the KE), and total degree (in-degree + out-degree). KE nodes with a total degree ≥ 3 were selected as key toxicity mechanism nodes of endocrine disruptors, and their gene sets were combined to ensure that the nodes were directly related to the EDC toxicity mechanism.

[0067] Toxicity outcome node sets are extracted from specific AOP networks. Based on a predefined toxicity system (including carcinogenicity, developmental toxicity, respiratory toxicity, immune system toxicity, endocrine system toxicity, nervous system toxicity, reproductive system toxicity, digestive system toxicity, hepatotoxicity, nephrotoxicity, cardiotoxicity, skin toxicity, and other toxicities), these toxicity outcome node sets are clustered to obtain multi-endpoint toxicity nodes. Specifically, from the NR / nonNR pathway-specific AOP networks, AO nodes with a biological level of Organ / Tissue and a q value < 0.25 are screened. Based on a predefined toxicity system, the AO node sets are classified, and each clustered AO node is uniformly labeled as the corresponding multi-endpoint toxicity prediction endpoint.

[0068] Nodes representing nuclear receptor-mediated molecular initiation events, non-nuclear receptor-mediated molecular initiation events, key toxicity mechanisms of endocrine disruptors, and multi-endpoint toxicity were labeled to obtain the types of predicted endpoints. These predicted endpoint types include predicted endpoints for nuclear receptor-mediated molecular initiation events, non-nuclear receptor-mediated molecular initiation events, key toxicity mechanisms of endocrine disruptors, and multi-endpoint toxicity.

[0069] S23: Obtain several sets of interaction data between compounds and their corresponding toxicity endpoints from a preset database, and standardize the interaction data based on the predicted endpoint type to obtain multiple sets of association data between standard compounds and their corresponding predicted endpoints; wherein, each set of association data corresponds to a type of predicted endpoint. Preferably, in one embodiment of the present invention, the interaction data is standardized based on the predicted endpoint type to obtain multiple sets of associated data of standard compounds and their corresponding predicted endpoints, including: Interaction data between compounds and their corresponding toxicity endpoints were obtained from a pre-defined multi-source database. This interaction data was then integrated to obtain an initial interaction dataset. In this embodiment, the multi-source databases include T3DB, TTD, NPASS, BindingDB, CTD, BioGRID, DrugCentral, ToxCast, and TOXRIC. The heterogeneous data from these nine databases were mapped to a unified field system. Core fields include: compound information (name, CASID, SMILES, PubChemCID, InChIKey); and toxicity endpoint information (target type, target name, mode of action, activity value, and number of literature evidence). All data was converted to CSV format using Python's pandas library, with consistent date and concentration units. Completely duplicate records were removed based on the key combination CASID + target name + mode of action to obtain the initial interaction dataset. Furthermore, the initial interactive dataset is subjected to unified identification processing, contradiction removal processing, and label normalization processing in sequence to obtain a structured standardized dataset; Specifically, the unified identification process includes using the PubChemPy library to convert the compound CAS IDs and InChIKeys in the initial dataset into PubChem CIDs; for compounds without CAS IDs / InChIKeys, obtaining the CIDs through fuzzy matching in PubChem using SMILES, and setting the matching threshold to SMILES similarity ≥ 0.95; supplementing the IUPAC names and standard SMILES of compounds through the PubChem API; deleting records missing PubChem CIDs or standard SMILES, and finally obtaining a standardized dataset of compound information.

[0070] The contradiction elimination process includes contradiction elimination within a single dataset and contradiction elimination after merging multiple data sources. The rules for contradiction elimination within a single dataset include that within the same dataset, for the same compound of the same nuclear receptor (NR), agonist and antagonist activities should not coexist; for the same compound-target association, positive (active) and negative (inactive) labels should not be marked simultaneously; if both IC50 and Ki values exist in the same record and the determination results are contradictory according to the criteria that IC50 < 10 μM is active and Ki > 100 μM is inactive, delete the record. The rules for contradiction elimination after merging multiple data sources include that for the same compound-target-mode of action association in different databases, if the mode of action conflicts or the determination results of the activity values are contradictory, delete the association; if only the numerical values of the activity values are different but the determination results are consistent, retain the record with the smallest activity value.

[0071] The label normalization process includes the following rules: Normalization of NR-related toxicity endpoint labels: Retain four basic labels: Agonist, Antagonist, Binding, and Inactive; if the sample size of a certain mode of action of a certain NR < 500, merge Agonist, Antagonist, and Binding into a unified Active label, and only retain two labels: Active and Inactive; Normalization of non-NR-related toxicity endpoint labels: Only retain two labels: Active and Inactive; based on the activity value or literature evidence, IC50 < 10 μM or literature evidence ≥ 2 articles are determined as Active, IC50 > 100 μM or no literature evidence is determined as Inactive, and records with 10 μM < IC50 < 100 μM are excluded.

[0072] Finally, the dataset after the above three rounds of processing is subjected to final structure optimization using the RDKit tool. Specifically, it includes separating salts / mixtures and removing inorganic substances, and finally obtaining a structurally standardized dataset. Each record contains PubChem CID, standard SMILES, target type, target name, normalized activity label, activity value, and the number of literature evidence.

[0073] Furthermore, the structure-normalized dataset is bound one-to-one with four types of predicted endpoints, ensuring that each set of associated data corresponds to only one type of predicted endpoint, thus providing targeted data for model training. Based on the predicted endpoint type, the structure-normalized dataset is subjected to matching and association processing to obtain multiple sets of associated data; the matching and association processing is designed to map compound data one-to-one with the activity labels of the corresponding predicted endpoints, with each set of associated data corresponding to one type of predicted endpoint.

[0074] Specifically, a predicted endpoint type index column is added to the structure-normalized dataset; data is extracted by grouping according to the index. For example, all records indexed as non-nuclear receptor-mediated molecular initiation events (increased oxidative stress) are extracted to form the associated data for that endpoint; all records indexed as multi-endpoint toxicity predicted endpoints (hepatotoxicity) are extracted to form the associated data for hepatotoxicity endpoints; each group of associated data is encapsulated in a compound feature-activity tag format. Ultimately, multiple groups of associated data are obtained, where each group corresponds to only one predicted endpoint type, with no data overlap.

[0075] S24: Perform multidimensional feature extraction processing on the associated data to obtain multiple sets of multidimensional molecular feature data; Preferably, in one embodiment of the present invention, multi-dimensional feature extraction processing is performed on the associated data, including: For each set of associated data, 1D hybrid fingerprint feature extraction, 2D molecular graph feature extraction and 3D molecular space feature extraction are performed respectively to obtain 1D molecular fingerprint feature data, 2D molecular graph feature data and 3D molecular space feature data corresponding to each set of associated data; Specifically, the invention first performs 1D hybrid fingerprint feature extraction. Two one-dimensional (1D) molecular fingerprints are used to characterize prior molecular features: the ECFP fingerprint is used to capture information such as atomic connectivity, while the PubChem fingerprint focuses on identifying information such as functional groups. Furthermore, the molecular structure is modeled as a graph, with atoms in the molecule considered as nodes and bonds as edges. in, Indicates atoms in a molecule. Indicates the bonds in a molecule. This represents the embedding vector dimension. Node features consist of 16 atomic attributes (such as atom type, number of free radical electrons, etc.), with some atomic attributes using one-hot encoding. Edge features are represented by four bond attributes (bond type, 3D features, conjugated bonds, and ring states). Node and edge features are integrated into a 2D molecular graph feature matrix. Furthermore, RDKit is used to calculate the three-dimensional coordinates of each atom. As a vector feature , representing the 3D characteristics of molecules.

[0076] The 1D molecular fingerprint feature data, 2D molecular graph feature data and 3D molecular spatial feature data corresponding to each set of associated data are integrated and processed to generate multiple sets of multidimensional molecular feature data. Each set of multidimensional molecular feature data corresponds to a type of associated data for predicting endpoints.

[0077] S25: Train the pre-constructed initial activity prediction model based on the multidimensional molecular feature data of each group to obtain the activity prediction model corresponding to each prediction endpoint type.

[0078] First, independent linear layers are used to learn feature representations of ECFP and PubChem fingerprints separately. The output features of both are then merged and input into a module containing linear layers and Dropout layers to further learn the fused molecular fingerprint features. GNNs are then used to learn the fingerprint features for a given molecular graph. Molecular characterization is performed. The topological structure of the molecular graph is represented by an adjacency matrix. During message passing, the feature vector of each atom is iteratively updated by aggregating information from its neighboring atoms connected by chemical bonds. A Geometric Vector Perceptron (GVP) layer is used to learn the 3D spatial conformation features of the input molecule. Subsequently, Graph Attention Convolution (GATConv) is used to further learn the output features of the GVP layer and the 2D graph features of the molecule. Graph max pooling and graph average pooling are applied in the readout layer to output a molecular graph feature vector that incorporates 3D conformation information. Finally, the fused fingerprint features are concatenated with the molecular graph features, passed through a normalization layer and a fully connected network containing four linear layers and two dropout layers for binary classification prediction, and the probability distribution is output using the Softmax function. This network architecture is named MF3DGNet (Multi-Feature Fusion 3D Graph Neural Network), and its detailed structure can be found in [link to detailed structure]. Figure 2 As shown, Figure 2 This is a schematic diagram of the structure and molecular characterization of MF3DGNet provided in an embodiment of the present invention.

[0079] By utilizing PyTorch's tensor indexing and dynamic batch alignment methods, the indices and corresponding scores of high-contribution atoms are extracted from the attention weights. The weights are then normalized and mapped to a continuous blue-red gradient color spectrum. Combined with the MolDraw2DCairo vector rendering engine in the RDKit tool, the weight information is embedded into the molecular topology in an atom highlighting mode to generate a molecular structure heatmap for identifying warning structures. The deeper the red, the higher the contribution of the atom to the structure predicted by the model.

[0080] During model training, the dataset obtained in step S24 is randomly divided into training, validation, and test sets in an 8:1:1 ratio. The model is trained on the training set, hyperparameters are adjusted and the model is validated on the validation set, and performance is finally evaluated on the test set. He initialization of model weights is used to optimize training stability. L2 regularization, dropout, and early stopping mechanisms are employed to avoid overfitting and underfitting during training. Training is terminated if there is no improvement after 30 consecutive epochs to avoid learning noise in the model. Furthermore, to reduce the impact of randomness in sample partitioning on the results, the above training and evaluation process is repeated using five different random seeds, and the mean and standard deviation (SD) of the test set performance metrics are reported. Class-balanced sampling and a weighted FocalLoss function are used to address the class imbalance problem in the training set. The Adam optimizer is used for training, with a maximum training epoch of 200. An adaptive learning rate scheduler is used to dynamically adjust the learning rate by monitoring changes in validation metrics to improve convergence speed. The anneal algorithm in the hyperopt tool is used to search for key hyperparameters in the model to optimize model performance. Model performance was evaluated using metrics including Accuracy (ACC), Recall, Precision, F1 Score, Balanced Accuracy (bACC), Receiver Operating Characteristic-Area Under Curve (ROC-AUC), and Precision-Recall-Area Under Curve (PR-AUC). To compare the performance of MF3DGNet with other models, we constructed RF, SVM, LR, kNN, and XGBoost prediction models as baselines for each toxicity endpoint based on ECFP fingerprints. Data was partitioned into training, validation, and test sets in an 8:1:1 ratio, and class imbalance was addressed using the SMOTE algorithm. Parameter tuning and model evaluation were performed using 10-fold cross-validation. The mean ± standard deviation of bACC, F1 score, and ROC-AUC metrics for MF3DGNet and traditional algorithms on the test set were compared. Furthermore, performance was compared with published GNN models on the same dataset. Five models were repeatedly trained using MF3DGNet with different random seeds based on their training data, and different evaluation metrics were obtained on the test set.

[0081] One embodiment of the present invention provides an endocrine disruptor identification system. For details, please refer to [link to specific documentation]. Figure 3 , Figure 3 The diagram shown is a flowchart illustrating an endocrine disruptor identification method according to one embodiment of the present invention, which includes: The acquisition module 11 is used to extract multi-dimensional features of the target compound to be identified in response to the compound recognition signal, so as to obtain the features of the target compound. Prediction module 12 is used to input the characteristics of the target compound into the activity prediction model corresponding to each prediction endpoint type to obtain the activity probability prediction result corresponding to each prediction endpoint. The identification module 13 is used to generate identification results of the target compound corresponding to each activity probability prediction result; The prediction module also includes: The parsing unit is used to parse and process the acquired data related to adverse outcome pathways of endocrine disruptors, and to construct a global AOP network based on the parsing results; The analysis unit is used to perform enrichment analysis on the node gene set of the global AOP network to obtain key node data and determine the prediction endpoint type corresponding to the key node data; among them, the enrichment analysis is used to screen key node data associated with key target genes of endocrine disruptors. The standardization unit is used to acquire several sets of interaction data between compounds and their corresponding toxicity endpoints in a preset database. Based on the predicted endpoint type, the interaction data is standardized to obtain multiple sets of association data between standard compounds and their corresponding predicted endpoints. Each set of association data corresponds to a type of predicted endpoint. The extraction unit is used to perform multidimensional feature extraction processing on the associated data to obtain multiple sets of multidimensional molecular feature data; The training unit is used to train the pre-constructed initial activity prediction model based on each group of multidimensional molecular feature data to obtain the activity prediction model corresponding to each prediction endpoint type.

[0082] This invention constructs an interactive prediction platform based on the R Shiny framework, integrating Shiny's reactive programming logic to achieve real-time display of prediction algorithm results. The front-end integrates the JSME chemistry editor to draw compound structures and generate SMILES in real time, or upload compound SMILES. The back-end relies on RDKit and PyBioMed to calculate molecular fingerprints and descriptors (1D / 2D / 3D features), and calls PyTorch and PyTorch Geometric to load pre-trained models for EDC prediction. Matplotlib is used for result visualization, and the rdMolDraw2D module of RDKit is called to generate molecular structure diagrams.

[0083] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: 1) This invention establishes an EDC prediction graph neural network model MF3DGNet that integrates multidimensional molecular features (1D molecular fingerprint, 2D molecular graph and 3D atomic coordinate features) and attention mechanism to identify warning structures. Compared with the single feature prediction of traditional machine learning methods, this model can comprehensively and efficiently learn the relationship between molecular structure features and prediction endpoints, and improve the accuracy of prediction.

[0084] 2) This invention, by combining events in an AOP network, enables EDC prediction tasks using multiple criteria such as NR, nonNR, and AO mechanisms. Compared to previous methods for predicting EDC based on a single criterion, this not only improves the comprehensiveness of predictions but also provides clues for mechanism analysis. Furthermore, compared to single EDC screening and prediction, the prediction of key toxicity mechanism nodes and multi-terminal toxicity prediction based on an AOP network further deepens EDC prediction research.

[0085] 3) In potential EDC prediction tasks, this invention can calculate SEDC based on the prediction results of multiple prediction endpoints to achieve EDC importance ranking, which helps researchers to conduct subsequent EDC screening research.

[0086] 4) The algorithm of this invention can screen out a large number of potential EDCs, predict possible mechanisms and conduct toxicity assessments in a short time. Compared with traditional experimental methods, this invention significantly reduces research costs and improves screening efficiency through computational methods.

[0087] 5) This invention develops the EDC-MF3DGNet online platform, which supports user-defined data input and result download, provides a convenient EDC prediction tool, and integrates multiple R libraries, providing rich visualization functions to facilitate data analysis and result display for users.

[0088] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for identifying endocrine disruptors, characterized in that, include: In response to the compound recognition signal, multidimensional feature extraction is performed on the target compound to be identified to obtain the target compound features; The target compound features are input into the activity prediction model corresponding to each prediction endpoint type to obtain the activity probability prediction results corresponding to each prediction endpoint. Generate identification results for the target compound corresponding to each of the predicted activity probabilities; The training process of the activity prediction model includes: The acquired data related to adverse outcome pathways of endocrine disruptors were analyzed and processed, and a global AOP network was constructed based on the analysis results; Enrichment analysis is performed on the node gene set of the global AOP network to obtain key node data and determine the predicted endpoint type corresponding to the key node data; wherein, the enrichment analysis is used to screen key node data associated with key target genes of endocrine disruptors. Several sets of interaction data between compounds and their corresponding toxicity endpoints are obtained from a preset database. The interaction data are then standardized based on the predicted endpoint type to obtain multiple sets of correlation data between standard compounds and their corresponding predicted endpoints. Each set of correlation data corresponds to a type of predicted endpoint. The associated data is subjected to multidimensional feature extraction processing to obtain multiple sets of multidimensional molecular feature data; The pre-constructed initial activity prediction model is trained based on the multidimensional molecular feature data of each group to obtain the activity prediction model corresponding to each prediction endpoint type.

2. The endocrine disruptor identification method as described in claim 1, characterized in that, The process of parsing and processing the acquired data related to adverse outcome pathways of endocrine disruptors, and constructing a global AOP network based on the parsing results, includes: Obtain adverse outcome path-related data of endocrine disruptors from the AOP-Wiki database, and parse the adverse outcome path-related data to obtain initial parsing data, wherein the initial parsing data includes at least key events and key event relationships; The initial parsed data is structurally transformed using the key events as nodes and the key event relationships as edges to obtain an AOP directed graph. The AOP directed graph is then processed to construct a network topology to obtain an initial global AOP network. Calculate the node distances in the initial global AOP network, and optimize the initial global AOP network based on the node distances to obtain the global AOP network. The optimization process is used to treat nodes with a distance of 1 as direct associations and nodes with a distance of 2 as indirect associations.

3. The endocrine disruptor identification method as described in claim 1, characterized in that, The enrichment analysis of the node gene set of the global AOP network to obtain key node data includes: Gene set matching is performed on each key event node in the global AOP network to obtain the node gene set corresponding to each key event node. Identifying key target genes associated with endocrine disruptor perturbations based on protein-protein interactions; Enrichment analysis was performed on each of the node gene sets, and key nodes associated with the target gene were screened based on the enrichment analysis results to obtain key node data. The key node data includes nuclear receptor pathway nodes directly associated with the mechanism of action of nuclear receptor-mediated endocrine disruptors and non-nuclear receptor pathway nodes directly associated with the mechanism of action of non-nuclear receptor-mediated endocrine disruptors.

4. The endocrine disruptor identification method as described in claim 3, characterized in that, The identification of key target genes associated with endocrine disruptor disturbances based on protein-protein interactions includes: Obtain initial endocrine disruptor data from a preset multi-source database, perform deduplication on the initial endocrine disruptor data, and generate an endocrine disruptor list; Gene association data corresponding to the endocrine disruptor list were extracted from the CTD database to construct an endocrine disruptor-associated gene set. The endocrine disruptor-associated gene set was mapped to a protein interaction network, and key target genes related to endocrine disruptor perturbation were screened based on the interaction output results.

5. The endocrine disruptor identification method as described in claim 1, characterized in that, The predicted endpoint types include nuclear receptor-mediated molecular initiation event predicted endpoints, non-nuclear receptor-mediated molecular initiation event predicted endpoints, key toxicity mechanism nodes of endocrine disruptors predicted endpoints, and multi-endpoint toxicity node predicted endpoints. Determining the prediction endpoint type corresponding to the key node data includes: A specific AOP network is constructed based on key node data, and nuclear receptor-mediated molecular initiation event nodes and non-nuclear receptor-mediated molecular initiation event nodes are extracted from the specific AOP network. Topology analysis was performed on the specific AOP network, and key toxicity mechanism nodes of endocrine disruptors were screened based on the topology analysis results. Extract the set of toxic outcome nodes from the specific AOP network, and cluster the set of toxic outcome nodes based on a preset toxicity system to obtain multi-endpoint toxic nodes; The nuclear receptor-mediated molecule initiation event nodes, non-nuclear receptor-mediated molecule initiation event nodes, key toxicity mechanism nodes of endocrine disruptors, and multi-endpoint toxicity nodes were labeled to obtain the types of predicted endpoints.

6. The endocrine disruptor identification method as described in claim 5, characterized in that, The construction of a specific AOP network based on key node data includes: The key node data is sequentially subjected to hierarchical limitation and redundancy merging to obtain the first key node data; Extract the first key event relationship data from the global AOP network that has a direct or indirect relationship with the first key node data, and remove the self-loops in the first key event relationship data to obtain the second key event relationship data; The adverse outcome nodes in the first key node data are categorized and replaced to obtain the second key node data. Construct a special node subnetwork based on the second key node data and the second key event relationship data; The specific node subnetwork is initialized as a directed graph, and the directed graph is filtered for effective paths to obtain a specific AOP network, wherein the specific AOP network includes a nuclear receptor pathway specific AOP network and a non-nuclear receptor pathway specific AOP network.

7. The endocrine disruptor identification method as described in claim 1, characterized in that, The interaction data is standardized based on the predicted endpoint type to obtain multiple sets of associated data for standard compounds and their corresponding predicted endpoints, including: The interaction correlation data between compounds and their corresponding toxic endpoints in a preset multi-source database is obtained, and the interaction correlation data is integrated and processed to obtain an initial interaction dataset. The initial interactive dataset is sequentially subjected to unified identification processing, contradiction removal processing, and label normalization processing to obtain a structured standardized dataset. Based on the predicted endpoint type, the structure-normalized dataset is matched and associated to obtain multiple sets of associated data; wherein, the matching and association process is designed to correspond one-to-one between compound data and the activity tags of the corresponding predicted endpoint, and each set of associated data corresponds to a predicted endpoint type.

8. The method for identifying endocrine disruptors as described in claim 1, characterized in that, The multidimensional feature extraction process for the associated data includes: For each set of associated data, 1D hybrid fingerprint feature extraction, 2D molecular graph feature extraction and 3D molecular space feature extraction are performed respectively to obtain 1D molecular fingerprint feature data, 2D molecular graph feature data and 3D molecular space feature data corresponding to each set of associated data; The 1D molecular fingerprint feature data, the 2D molecular graph feature data, and the 3D molecular spatial feature data corresponding to each set of associated data are integrated and processed to generate multiple sets of multidimensional molecular feature data, wherein each set of multidimensional molecular feature data corresponds to an associated data of a prediction endpoint type.

9. The method for identifying endocrine disruptors as described in claim 1, characterized in that, The generation of identification results for the target compounds corresponding to each of the predicted activity probabilities specifically includes: Based on the activity probability prediction results corresponding to each prediction endpoint and the preset judgment rules, the target compound is subjected to endocrine disruptor category identification processing to obtain endocrine disruptor category identification results. Based on the endocrine disruptor category identification results and their corresponding activity probability prediction results, the importance of the target compound is calculated to obtain the endocrine disruptor importance score. The activity probability prediction results corresponding to the predicted endpoint, the endocrine disruptor category identification results, and the endocrine disruptor importance scores are integrated to generate the identification results of the target compound.

10. An endocrine disruptor identification system, characterized in that, include: The acquisition module is used to extract multi-dimensional features of the target compound to be identified in response to the compound recognition signal, so as to obtain the features of the target compound. The prediction module is used to input the characteristics of the target compound into the activity prediction model corresponding to each prediction endpoint type to obtain the activity probability prediction result corresponding to each prediction endpoint. An identification module is used to generate identification results for the target compound corresponding to each of the activity probability prediction results; The prediction module further includes: The parsing unit is used to parse and process the acquired data related to adverse outcome pathways of endocrine disruptors, and to construct a global AOP network based on the parsing results; An analysis unit is used to perform enrichment analysis on the node gene set of the global AOP network to obtain key node data and determine the predicted endpoint type corresponding to the key node data; wherein, the enrichment analysis is used to screen key node data associated with key target genes of endocrine disruptors. A standardization unit is used to acquire several sets of interaction data between compounds and their corresponding toxicity endpoints in a preset database, and to standardize the interaction data based on the predicted endpoint type to obtain multiple sets of association data between standard compounds and their corresponding predicted endpoints; wherein each set of association data corresponds to a type of predicted endpoint. The extraction unit is used to perform multidimensional feature extraction processing on the associated data to obtain multiple sets of multidimensional molecular feature data; The training unit is used to train the pre-constructed initial activity prediction model based on the multidimensional molecular feature data of each group, so as to obtain the activity prediction model corresponding to each prediction endpoint type.

Images