A System and Operation Method for Estrogen Receptor Target Identification and Toxicity Prediction Based on Graph Neural Networks and Molecular Simulation Technology

Abstract

This invention discloses an estrogen receptor target identification and toxicity prediction system and its operation method based on graph neural networks and molecular simulation technology. The system converts compound structures into molecular graphs, constructs atomic node features and chemical bond edge features / types, inputs these into a graph neural network for characterization learning, and outputs the binding probabilities of estrogen receptor subtypes such as ESR1, ESR2, and GPER1, as well as predicted toxicity endpoints such as logAC50 / AC50. Based on the predicted target, the system automatically selects the receptor structure and adaptively generates docking box parameters. It then calls AutoDock Vina to complete molecular docking to obtain binding energies and optimal conformations. The system also displays the three-dimensional complex, interaction statistics, public database annotation information, and toxicological explanations based on a large language model in an interactive interface. This system achieves a high-throughput, interpretable, integrated "prediction-docking-annotation" analysis workflow for ER targets.

Description

Technical Field

[0001] This invention relates to the field of computational pharmaceutical chemistry and toxicology informatics technology, specifically to an estrogen receptor target identification and toxicity prediction system and operating method based on graph neural networks and molecular simulation technology. Background Technology

[0002] Estrogen receptors (ERs) are important targets in drug development and environmental chemical risk assessment. Currently, the prediction of bioactivity and toxicity assessment of ER-related compounds typically rely on a cascaded workflow of multiple independent tools. Typical approaches include: 1) A workflow combining high-throughput toxicity screening platforms such as EPA ToxCast with traditional QSAR scripts: Platforms like ToxCast provide a unified data warehouse and hardware infrastructure for collecting experimental toxicity data such as AC50 and EC50; researchers then use publicly available data to write their own QSAR training scripts using programming libraries such as RDKit and scikit-learn to achieve quantitative structure-activity relationship modeling of specific toxicity indicators. 2) A structural biology workflow centered on software such as AutoDockVina, MGLTools, and PyMOL / DiscoveryStudio: Researchers prepare receptor structures and ligand libraries in advance, run scripts such as prepare_receptor and prepare_ligand in the command-line environment, and call AutoDockVina to complete batch molecular docking; the generated pdbqt / pdb files are then visualized and analyzed using software such as PyMOL or DiscoveryStudio. 3) Workflow combining online toxicity annotation services such as PubChem and EPACompTox with generalized large language model question-and-answer: Users query the synonyms, GHS descriptions and related literature information of compounds based on SMILES on web pages or APIs, and copy the results to the generalized conversational large language model to obtain explanations or questions and answers in natural language form.

[0003] While the aforementioned "multi-tool cascade" approach is relatively mature in its individual stages, its overall structure exhibits a clear "loose coupling" characteristic, specifically manifested in the following ways: 1) Lack of a unified data structure and end-to-end management capabilities: Existing solutions require frequent file export and import between multiple tools, easily leading to formatting errors or data loss, and making it difficult to uniformly manage and track compound states, prediction results, and docking products. 2) Lack of automated error detection and process guidance mechanisms: Traditional workflows generally rely on manual checks of model loading status and input data integrity. If software crashes or kernel restarts occur during calculations, it is often difficult to detect and locate them in a timely manner, resulting in insufficient reliability of the results. 3) Lack of integrated visualization and toxicological interpretation capabilities: Existing tools struggle to simultaneously display 3D complex structures, key interactions, public database annotation information, and text interpretations specific to compounds and targets within a single interface, requiring users to repeatedly switch between multiple software programs and web pages. 4) Lack of support for batch processing of multiple targets / compounds: Many existing products are mainly for the analysis of single targets or single compounds, and lack unified scheduling and batch processing capabilities for large-scale compound libraries and multi-target scenarios, making it difficult to meet the high-throughput screening needs of environmental chemicals and candidate drugs.

[0004] In summary, existing "multi-tool cascaded" technical systems still have shortcomings in areas such as integrated data management, automated quality control, comprehensive visualization, and batch screening, and cannot yet meet the integrated requirements for high-throughput, end-to-end, and interactive analysis in estrogen receptor-related toxicity assessment. Furthermore, existing ER target prediction and toxicity QSAR models largely rely on fixed-length molecular fingerprints and artificial physicochemical descriptors as input features. These features struggle to explicitly characterize structural information such as atom-bond connection topology, local functional group combinations, and bond type differences, resulting in limited generalization ability for novel skeletons or structurally complex compounds, and difficulty in providing explanations traceable to specific atoms / bonds. Graph neural networks can perform message passing or attention aggregation on molecular graphs, automatically learn substructure representations related to target selectivity and toxicity endpoints, and output interpretable information such as atom contribution. Therefore, it is necessary to combine graph neural networks with molecular simulation methods such as molecular docking to construct a unified "prediction-docking-interpretation" integrated technical system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an integrated computing platform for ER targets, which addresses the problems of "disjointed links between target prediction, molecular docking and toxicological interpretation, lack of automation and interactivity" in the existing "multi-tool cascade" scheme for estrogen receptor-related computational processes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An estrogen receptor target identification and toxicity prediction system based on graph neural networks and molecular simulation technology, the system being an integrated computing platform comprising the following six core modules: (I) Data Preparation and Model Training Module Composition: Includes a data import submodule, a molecular structure standardization submodule, a molecular graph construction and feature encoding submodule, a graph neural network model training submodule (target classification and toxicity regression), a model evaluation and saving submodule, and an optional fingerprint / descriptor generation and fusion submodule.

[0007] Function: This function loads training data (e.g., ER_Ligands.xlsx, ER_Toxicity.xlsx, etc.), performs structural normalization on training sample compounds, and constructs a molecular graph G=(V,E), where V is the set of atomic nodes and E is the set of chemical bond edges. It generates node feature vectors for nodes and edge features / types for edges. The molecular graph is then input into a graph neural network for representation learning, yielding a molecular-level embedding representation and outputting target probability distributions such as ESR1, ESR2, and GPER1, as well as predicted values ​​for toxicity endpoints such as logAC50 / AC50. Optionally, a 2048-bit Morgan fingerprint and physicochemical property descriptors are further generated and spliced / gated with the molecular-level embedding to improve prediction robustness.

[0008] Connection relationships: After the model training is completed, the graph neural network model parameters, node / edge feature encoding dictionary, feature standardization parameters and related data objects are registered to the global namespace or a unified model management structure for subsequent batch inference, docking and visualization modules to call.

[0009] (ii) Model status check module (pre-inspection module) Components: Includes controls for triggering checks (such as buttons), model overload controls, and status output display panels, etc.

[0010] Function: Used to detect whether key objects such as graph neural network prediction models, node / edge feature encoding dictionaries, toxicity regression / fusion related parameters, and receptor structure libraries have been correctly loaded; when missing or abnormal objects are detected, it can prompt the user to return to the training unit to re-execute, or prompt the user to restart the computing kernel and run the relevant units in sequence.

[0011] Connection relationship: As a pre-screening checkpoint before batch compound processing and docking analysis, it avoids hidden errors such as "missing model or data" caused by kernel restart or interruption of execution.

[0012] (III) Batch Compound Input Module Components include a file upload control, a SMILES text input control, a processing button, a result output area, and a batch_session data structure for maintaining the batch context.

[0013] Function: Used to receive structural files and multi-line SMILES strings in formats such as .sdf, .mol, .mol2, .pdb, .smi, and .txt, enabling batch import of compounds and automatically performing hydrogen supplementation, three-dimensional embedding, and MMFF force field optimization to generate standardized SDF / PDB / SMI files.

[0014] Structural relationship: An independent workspace directory is automatically created each time a batch process runs, for example, Output / BatchCompound / Batch_ <timestamp>The system is structured into subdirectories such as input, predictions, docking, visualization, and toxicity. Simultaneously, the structural information and corresponding file path of each compound are registered in batch_session['compounds'] to achieve batch management and status tracking.

[0015] (iv) Batch Target and Toxicity Prediction Module Components: Includes a prediction trigger control, a result output area, a molecular graph construction and batch processing submodule, a graph neural network inference submodule, and an optional baseline model / fusion submodule.

[0016] Function: This function standardizes the structures of each compound imported from the batch input module and constructs a molecular graph. It batch-processes and generates a tensor input graph neural network prediction model with node features, edge types, and node masks. It outputs the probability distributions of ESR1, ESR2, and GPER1 and determines recommended targets. It also outputs predicted values ​​of toxicity endpoints such as logAC50 / AC50. Optionally, it uses similarity-weighted k-NN and other methods to output control toxicity estimates and fuses them with the graph neural network output. It can also optionally output the importance of atoms / bonds to explain key substructures.

[0017] Structural relationships: The prediction results are written to the batch_session['compounds'][name]['predictions'] field and exported as a predictions / batch_predictions.csv file for subsequent analysis and storage.

[0018] (v) Batch Integration Module Composition: Includes docking trigger control, result output area, automatic docking box parameter calculation function, AutoDockVina scheduling logic, and Receptor structure directory, etc.

[0019] Function: Based on the recommended target points obtained from the batch prediction module, it automatically selects the corresponding acceptor structure for each compound, automatically calculates the center and side length of the docking box based on the three-dimensional coordinate range of the acceptor (for example, extending outward by about 20 Å based on the acceptor coordinate range), calls the prepare_receptor and prepare_ligand scripts to generate PDBQT files, runs AutoDockVina (for example, setting parameters such as --num_modes9 and --exhaustiveness8) to complete molecular docking, and uses OpenBabel to export the PDB file of the energy-optimal conformation.

[0020] Structural relationships: Docking results (including optimal docking energy, conformation file path, and acceptor file path, etc.) are stored in the batch_session['compounds'][name]['docking'] field and summarized in the docking / batch_docking_summary.csv file.

[0021] (vi) Interactive visualization and toxicity annotation module Components include a compound selection dropdown menu, a refresh control, interaction display options, a py3Dmol 3D view component, a toxicity information HTML panel, and a chat control based on a large language model interface.

[0022] Function: Reads the calculation results of the aforementioned modules in real time, presents information such as the three-dimensional protein-ligand complex structure, hydrophobic / polar interactions and metal coordination, and outputs the interaction data in tabular form; at the same time, it displays two-dimensional structure diagrams, AC50 predicted values ​​and target information cards, synonyms from PubChem, GHS hazard descriptions and publicly available toxicity information, and provides toxicity explanations and interactive Q&A through a large language model.

[0023] Structural relationships: The chat_state structure stores the context information of the current compound (such as name, AC50, target, GHS tag, toxicity index, etc.) for multi-round question and answer calls; the HTML panel uniformly references data obtained from the batch prediction module, batch docking module and online database to ensure consistency of the displayed content.

[0024] The second objective of this invention is to provide an operational method for an estrogen receptor target identification and toxicity prediction system based on graph neural networks and molecular simulation technology, comprising the following stages: 1) Training Phase The user runs the training section of the platform, which completes the import of training data, structure standardization, molecular graph construction and feature encoding, graph neural network model training and evaluation, and registers the relevant models and data objects as global variables or writes them into a unified model management structure. 2) Pre-inspection stage: Users perform a model status check to confirm that the resources required for the graph neural network prediction model, feature encoding / normalization parameters, receptor structure library, and (optionally) fusion model have been correctly loaded, ensuring the availability of subsequent batch processing and docking processes; 3) Batch processing stage: The process sequentially executes three stages: batch compound input, batch target and toxicity prediction, and batch docking. Multiple compounds are uniformly distinguished and scheduled to form batch prediction and docking results with a unified data structure. 4) Analysis phase: In the interactive visualization and toxicity annotation module, users can select any docked compound and conduct a comprehensive analysis of its three-dimensional complex structure, key interactions, PubChem annotation information, and toxicological interpretation based on a large language model. Users can also conduct contextualized multi-round follow-up questions and verifications through the chat panel. 5) Single-compound expansion stage: When it is necessary to move away from batch processing scenarios, users can call the analysis unit for single compounds in the platform, and directly perform prediction, docking and annotation operations on a single compound based on the structure and intermediate results in the existing catalog, so as to achieve refined analysis.

[0025] The working principle and key technical features of this invention platform mainly include: 1. Unified coupling of molecular graph representation and graph neural network: Compounds are uniformly represented as molecular graphs, and node / edge features are encoded. Graph neural networks perform message passing or attention aggregation on the molecular graphs to learn molecular-level embedding representations related to ER subtype selectivity and toxicity endpoints, which are used to simultaneously output target probability distributions and toxicity indicators such as logAC50 / AC50. Graph embeddings can optionally be fused with global features such as Morgan fingerprints and physicochemical descriptors to improve the robustness and generalization ability of the model. 2. Automatic docking box derivation and batch management mechanism: The docking box parameters are automatically derived by using the three-dimensional structural coordinates of the receptor, avoiding subjective bias caused by manually setting the center and size of the box. At the same time, a workspace directory distinguished by timestamps is used to hierarchically manage the input of each batch of compounds, prediction results, docking products and visualization files, which facilitates result traceability and process reproducibility. 3. Multimodal output and unified interface presentation: The py3Dmol 3D structural view, tabular interaction information, HTML-formatted toxicity cards, and text explanations generated by a large language model are integrated into the same interactive interface to achieve multimodal coupled display of structural information, numerical results, and textual knowledge, thereby improving the efficiency of result interpretation. 4. Interactive intelligent question-and-answer closed loop: Based on large language model interfaces such as DeepSeek, and combined with the compound context information recorded in chat_state, professional explanations related to specific compounds and targets are generated. This allows users to continuously ask follow-up questions about prediction results, docking conformations, and toxicity risks, thus forming a closed-loop analysis process of "prediction-verification-interpretation". 5. Linkage between interpretability and structural verification: By using graph neural networks to output atom / bond importance scores through methods such as attention weights, gradient attribution, or masking perturbations, key substructures driving target discrimination and toxicity estimation are located and used together with key interacting residues obtained from molecular docking for structural rationality verification and subsequent molecular optimization decisions.

[0026] Compared with the aforementioned existing solutions of the "multi-tool cascade" type, the present invention has significant differences and beneficial effects in terms of system structure and functional implementation, mainly reflected in: 1. End-to-end integrated design: This invention encapsulates target classification, toxicity prediction, molecular docking, 3D visualization, PubChem data acquisition, and toxicological annotation based on a large language model into a single platform. It completes the entire chain of "target prediction - QSAR - docking - visualization - toxicological interpretation - AI" question answering in a consistent data context and user interface, avoiding frequent manual data migration between multiple software programs. 2. Automated verification and batch processing capabilities: This invention achieves automatic verification and status management for multi-compound and multi-target tasks through a model state checking module and a batch compound processing and docking mechanism, reducing the need for manual intervention and the operational threshold, and ensuring the robustness and repeatability of the calculation process. 3. Synchronous display of structural data and text explanation: This invention displays three-dimensional protein-ligand complexes, interaction information, toxicity prediction data, and toxicological explanations based on a large language model in a single interactive interface. Users can complete structure verification, result interpretation, and risk assessment in a unified view. Existing platforms typically only cover one or a few aspects, making it difficult to achieve collaborative analysis of multimodal information. 4. Interactive AI-assisted analysis targeting specific targets: The large language model module of this invention is not called in isolation, but performs contextual reasoning based on the contextual data such as compounds, targets, docking and toxicity that have been aggregated within the platform, thereby outputting toxicological interpretations that are more in line with the specific ER target and compound context; in contrast, traditional general question answering models often lack this docking context and automatic data splicing capabilities.

[0027] In summary, this invention achieves integrated analysis of the entire chain of high-throughput prediction, molecular docking, and toxicological interpretation for estrogen receptor targets through systematic module division, unified data structure, and multimodal interactive output. It is significantly different from existing loosely coupled, multi-tool combination schemes in terms of structural form and functional integration. Attached Figure Description

[0028] Figure 1 This is a screenshot of the interface of an online platform for an integrated system for estrogen receptor target identification and toxicity analysis based on graph neural networks and molecular simulation technology, according to the present invention.

[0029] Figure 2 This is a flowchart of the six modules of an integrated system for estrogen receptor target identification and toxicity analysis based on graph neural networks and molecular simulation technology according to the present invention.

[0030] Figure 3 This is a graph showing the change of the loss function of the graph neural network model during the training process with the number of training rounds in an embodiment of the present invention.

[0031] Figure 4 The figures show the ROC and PR curves of the model on the validation set in this embodiment of the invention.

[0032] Figure 5 This is a scatter plot comparing the predicted and measured values ​​of the quantitative structure-activity relationship (QSAR) model in this embodiment of the invention.

[0033] Figure 6 This is a histogram showing the distribution of ER-related ligand AC50 (represented by log10 nM) in an embodiment of the present invention. Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0036] This application makes the following... Figure 1 and 2 This paper presents an integrated system for estrogen receptor target identification and toxicity analysis based on graph neural networks and molecular simulation technology. Under a unified data structure and interactive interface, the platform integrates functional modules such as molecular feature calculation, machine learning target prediction, quantitative structure-activity relationship (QSAR) toxicity inference, AutoDock Vina molecular docking, py3Dmol 3D visualization, PubChem data acquisition, and toxicological knowledge annotation based on large language models (e.g., DeepSeek). It supports an integrated workflow for batch processing of multiple compounds and in-depth analysis of single compounds, and can be used for high-throughput screening and risk assessment of estrogen receptor-related compounds.

[0037] Example 1 (Construction and Training of a Graph Neural Network Target Recognition and Toxicity Prediction Model) This embodiment illustrates how the graph neural network in the present invention is used for estrogen receptor target identification and toxicity prediction.

[0038] (1) Data preparation Collect compound data with target tags (at least ESR1, ESR2, GPER1) and / or toxicity endpoint values ​​(e.g., AC50 or logAC50), perform deduplication, normalization, and tag normalization on SMILES and structural files, and divide them into training and validation sets.

[0039] (2) Molecular map construction and feature encoding Each compound is represented as a molecular graph G=(V,E), where V is the set of atomic nodes and E is the set of chemical bond edges. A node feature vector is generated for each node, and an edge feature or edge type is generated for each edge. Preferably, node features include one or more of the following: atomic type, atomic degree, valence state / formal charge, aromaticity, whether it is in a ring, hybridization type, and hydrogen count; edge features include one or more of the following: bond type, whether it is conjugated, whether it is in a ring, and bond orientation / stereoscopic information.

[0040] (3) Graph Neural Network Structure and Output Node features and edge features are input into a graph neural network encoder and subjected to multi-layer message passing or multi-head attention aggregation to obtain node embeddings. The node embeddings are then aggregated through a readout layer to obtain molecular-level embedding vectors. Based on these molecular-level embedding vectors, a target classification output is set up to output the probability distributions of ESR1, ESR2, and GPER1. A toxicity regression output is set up to output predicted toxicity endpoints such as logAC50 / AC50. Preferably, the graph neural network is a graph Transformer network that includes edge-type embeddings.

[0041] (4) Training and preservation The target classification uses the cross-entropy loss function, and the toxicity regression uses the mean squared error loss function or the Huber loss function, which are then jointly optimized by weighted summation. After training, the model parameters, node / edge encoding dictionary, and feature standardization parameters are saved for online inference.

[0042] (5) Reasoning and interpretability The system performs structure normalization and molecular graph construction consistent with the training on the compound to be predicted, inputs it into the trained graph neural network, and outputs target probability and toxicity endpoint values. Optionally, it generates atom / bond importance scores based on attention weights, gradient attribution, or masking perturbations to explain the key substructures driving the prediction, and verifies them in conjunction with molecular docking results.

[0043] Example 2: Taking 4-Hydroxytamoxifen (hereinafter referred to as "compound 1") as an example, target prediction, GHS hazard information retrieval, and molecular docking evaluation were performed. The molecular weight (MW) of compound 1 is 387.22.

[0044] Compound 1 was predicted using a target prediction method, and the results showed that its main target was estrogen receptor β (ESR2), with a predicted AC50 of 0.4449 μM. The target prediction probability distribution of compound 1 is shown in Table 1.

[0045] Table 1. Target prediction probability distribution for Compound 1 A search of public databases yielded GHS hazard classification information for Compound 1, with the signal word "Warning" and hazard warnings related to acute and reproductive toxicity. These hazard statements included: H302 (harmful if swallowed), H312 (harmful if in contact with skin), H332 (harmful if inhaled), and H361 (suspected impairment of fertility or harm to the fetus).

[0046] Furthermore, using ESR2 as the receptor protein, compound 1 was docked using a molecular docking method, and a docking score of -7.021 kcal / mol was obtained, indicating that compound 1 can form a relatively stable binding conformation in the ESR2 ligand binding pocket.

[0047] The interaction types after compound 1 docked with ESR2 included hydrophobic interactions and hydrogen bonding, with hydrophobic interactions being the main contributor (81 / 84, approximately 96.4%) and hydrogen bonding being a secondary orientation effect (3 / 84, approximately 3.6%). The interaction statistics are shown in Table 2.

[0048] Table 2. Statistical analysis of docking interactions between compound 1 and ESR2 Note: The hydrophobic interaction distances in Table 2 range from 3.12 to 4.49 Å; the hydrogen bond distances range from 3.03 to 3.36 Å. The hydrogen bonds are mainly formed by the ligand atom O16 and the acceptor residues GLU276, TYR397, and LYS401, with the shortest hydrogen bond distance being 3.03 Å.

[0049] As shown in Table 2, ARG346 and HIS279 are high-frequency contact residues, followed by GLU276. This indicates that compound 1 forms the main binding stability in the ESR2 pocket through continuous hydrophobic / van der Waals contacts, and achieves anchoring and spatial orientation constraint through a small number of key hydrogen bonds. This result is consistent with the high probability prediction of ESR2 and docking score results shown in Table 1.

[0050] Given the important role of ESR2 (ERβ) in reproductive system development and homeostasis, immune and inflammatory regulation, nervous system protection, cardiovascular reactivity regulation, and metabolic processes, compound 1, as a potential ESR2 ligand, may affect these physiological processes. Its specific manifestation depends on its functional properties as an agonist, antagonist, or selective regulator, and further confirmation is needed by combining transcriptional activity and downstream pathway experiments.

[0051] Example 3 uses Diethylstilbestrol (hereinafter referred to as "Compound 2") as an example to perform target prediction, GHS hazard and toxicity information retrieval, and molecular docking evaluation. The molecular weight (MW) of Compound 2 is 268.15.

[0052] Compound 2 was predicted using a target prediction method, and the results showed that its main target was estrogen receptor β (ESR2), with a predicted AC50 of 0.0139 μM. The target prediction probability distribution of compound 2 is shown in Table 3.

[0053] Table 3. Target prediction probability distribution of compound 2 By searching public databases, GHS hazard classification information for compound 2 can be obtained, with the signal word "Danger". The hazard claims include at least: H315 (causes skin irritation), H319 (causes severe eye irritation), H335 (may cause respiratory irritation), H341 (suspected to cause genetic defects), H350 (may be carcinogenic), H360 (may impair fertility or cause harm to the fetus), H371 / H372 / H373 (may cause organ damage, single or long-term / repeated exposure), as well as H400 / H401 (toxic to aquatic life) and H410 / H411 (long-term harm to aquatic life), etc.

[0054] Meanwhile, the toxicity data section shows the presence of toxicological data sources and entries, indicating that its risk characterization information is relatively complete. In actual management, it should be handled and disposed of as a highly hazardous chemical to avoid inhalation, ingestion, and skin / eye contact, and to prevent it from entering environmental water bodies.

[0055] Furthermore, using ESR2 as the receptor protein, compound 2 was docked using a molecular docking method, and a docking score of -10.241 kcal / mol was obtained, indicating that compound 2 can form a relatively stable binding conformation in the ESR2 ligand binding pocket; this result is consistent with its predicted trend of strong activity at the nanomolar level (0.0139 μM).

[0056] Given that ESR2 (ERβ) plays an important role in reproductive system homeostasis, immune inflammation regulation, and neuroprotection, compound 2, as a potential ligand of ESR2, may affect these processes. Its specific pharmacological properties need to be confirmed through functional experiments. In addition, considering its carcinogenicity, reproductive toxicity, and ecotoxicity, priority should be given to conducting safety evaluations related to genotoxicity, reproductive and developmental toxicity, and long-term exposure.

[0057] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.< / timestamp>

Claims

1. A system for estrogen receptor target identification and toxicity prediction based on graph neural networks and molecular simulation technology, characterized in that, The system is an integrated computing platform, comprising the following six core modules: I. Data Preparation and Model Training Module: Used to load the training dataset, perform structural standardization on the compounds in the training data and construct a molecular graph, generating node features and edge features / edge types; Based on the molecular graph, a graph neural network prediction model is trained. The graph neural network prediction model is used to output the target probability distributions of ESR1, ESR2, and GPER1, as well as toxicity endpoint values ​​such as logAC50 / AC50. It can also optionally generate Morgan fingerprints and physicochemical property descriptors, which are used to fuse with the graph embedding representation of the graph neural network or as a control model. II. Model Status Check Module: As a pre-inspection checkpoint, it automatically verifies whether key model objects and data objects have been correctly loaded into memory before task execution. If missing or abnormal objects are detected, a prompt is issued to ensure the stability of subsequent processes. III. Batch Compound Input Module: It provides file upload and text input interfaces, supports batch import of various molecular file formats, and automatically performs molecular structure standardization, protonation, three-dimensional conformation generation and optimization. It creates an independent workspace directory for each batch to manage input files and intermediate data. IV. Batch Target and Toxicity Prediction Module: The system performs molecular graph construction and batch processing on the input compounds, calls the loaded graph neural network prediction model to output the ER subtype binding probability prediction results and logAC50 / AC50 toxicity index prediction results for each compound, and optionally outputs the atom or bond importance based on attention / gradient to explain the key substructures driving target discrimination and toxicity estimation; wherein the toxicity index prediction results can be obtained from the graph neural network regression output or fused with the k-nearest neighbor prediction results based on similarity weighting. V. Batch Integration Module: Based on the predicted recommended target, the corresponding three-dimensional structure of the receptor is automatically selected for each compound. Based on the receptor coordinates, the molecular docking parameters are adaptively calculated. The external molecular docking engine AutoDock Vina is scheduled to complete the batch docking task and extract the optimal conformation and binding energy. VI. Interactive Visualization and Toxicity Annotation Module: It reads the calculation results of the aforementioned modules in real time, provides a unified graphical user interface, and integrates three-dimensional structure visualization components, two-dimensional molecular structure display, toxicity prediction result cards, public database information query panel, and context-aware question-and-answer unit based on a large language model to achieve multimodal linkage display of structure, data and text interpretation.

2. An operational method for an estrogen receptor target identification and toxicity prediction system based on graph neural networks and molecular simulation technology, characterized in that, Includes the following steps: S1, Training Phase: The user runs the training part of the platform, completes the import of training data, standardization of compound structure, construction of molecular graph and encoding of node / edge features, trains the graph neural network prediction model for outputting the probability of ER subtype target and the toxicity endpoint value, and registers the graph neural network model parameters, feature encoding dictionary and related data objects as global variables or writes them into a unified model management structure. S2, Pre-inspection phase: Users perform model status checks to confirm that the required resources have been loaded correctly, ensuring the availability of subsequent batch processing and integration processes; S3, Batch Processing Stage: The process sequentially executes three stages: batch compound input, batch target and toxicity prediction, and batch docking. In the batch target and toxicity prediction stage, a molecular graph is constructed for each compound and a graph neural network prediction model is called to obtain the target probability and toxicity endpoint prediction values. Multiple compounds are uniformly distinguished and scheduled to form batch prediction and docking results with a unified data structure; S4, Analysis Phase: In the interactive visualization and toxicity annotation module, users can select any docked compound and conduct a comprehensive analysis of its three-dimensional complex structure, key interactions, PubChem annotation information, and toxicological interpretation based on a large language model. Users can also conduct contextualized multi-round follow-up questions and verifications through the chat panel. S5, Single-compound expansion stage: When it is necessary to move away from batch processing scenarios, users can call the analysis unit for single compounds in the platform, and directly perform prediction, docking and annotation operations on a single compound based on the structure and intermediate results in the existing catalog, so as to achieve refined analysis.

Images