Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network

Abstract

The invention provides a Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network, which comprises the following steps: constructing a Parkinson's disease domain knowledge graph containing drugs, targets, metabolic enzymes, adverse reactions and drug interaction relationships, entities and relationships are extracted from authoritative medical guidelines, literatures and databases through a natural language processing technology; based on the knowledge graph, using a knowledge graph embedding model to carry out vectorization expression on drugs and related entities thereof, and combining molecular fingerprint information and protein sequence features of the drugs to construct node features; inputting the node features into a graph neural network model, analyzing a potential action path between drugs through a multi-hop relationship aggregation mechanism, and predicting the occurrence probability of a drug combination and adverse reactions or drug-drug interaction; and outputting the prediction probability, and displaying a knowledge graph path having key influence on a prediction result in a visual mode to assist a doctor in identifying potential drug risks which are not widely reported.

Description

A Method for Predicting Drug Risk in Parkinson's Disease Based on Knowledge Graph Embedding and Graph Neural Networks Technical Field

[0001] This invention relates to the field of Parkinson's disease drug risk prediction technology, and in particular to a Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network. Background Technology

[0002] Parkinson's disease (PD) is a common neurodegenerative disease in middle-aged and elderly people, with highly heterogeneous clinical manifestations and disease progression. Currently, the diagnosis and treatment of PD is gradually evolving towards data-driven, knowledge-guided, and intelligent assistance. The deep integration of artificial intelligence and medical informatics provides a technological foundation for precision medicine. Existing research has attempted to utilize the collaborative operation of knowledge graphs, natural language processing (NLP), and graph neural networks (GNNs) to construct structured medical knowledge and intelligent decision support systems. This system covers the entire process from medical knowledge extraction, patient data parsing, multimodal feature fusion to personalized treatment recommendations, specifically including: knowledge structuring of medical guidelines, standardization of clinical data, semantic modeling of drug-target-metabolic pathways, and key aspects such as treatment strategy optimization based on reinforcement learning and rule engines.

[0003] However, existing drug risk prediction methods directly use static databases (such as SIDER and FAERS) to match known adverse reactions with drug interactions, without modeling potential risks arising from multi-hop relationships, metabolic pathway conflicts, or structural similarities between drugs. This may lead to insufficient ability to identify novel or hidden ADRs / DDIs, or misjudgments in multi-drug combination scenarios, thus affecting the safety and individualization of treatment plans. Furthermore, traditional methods typically rely on a single data source or rule system, making it difficult to integrate multimodal information such as patient genotype, imaging characteristics, and dynamic symptom changes, limiting the depth and breadth of risk prediction. In clinical practice, physicians lack systematic analytical tools for the potential risks of complex drug combinations, resulting in high trial-and-error costs and safety risks in treatment decisions. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a method for predicting the risk of Parkinson's disease medication based on knowledge graph embedding and graph neural networks.

[0006] The second objective of this invention is to propose a Parkinson's disease drug risk prediction device based on knowledge graph embedding and graph neural networks.

[0007] The third objective of this invention is to provide an electronic device.

[0008] The fourth objective of this invention is to provide a computer-readable storage medium.

[0009] The fifth objective of this invention is to provide a computer program product.

[0010] To achieve the above objectives, a first aspect of this invention proposes a method for predicting Parkinson's disease drug risks based on knowledge graph embedding and graph neural networks, comprising: S1, constructing a knowledge graph in the Parkinson's disease domain containing drugs, targets, metabolic enzymes, adverse reactions, and drug interactions, and extracting entities and relationships from authoritative medical guidelines, literature, and databases using natural language processing technology; S2, based on the knowledge graph, using a knowledge graph embedding model to vectorize drugs and their related entities, and constructing node features by combining the molecular fingerprint information and protein sequence features of the drugs; S3, inputting the node features into a graph neural network model, analyzing potential action paths between drugs through a multi-hop relationship aggregation mechanism, and predicting the probability of drug combinations and adverse reactions or drug-drug interactions; S4, outputting the predicted probabilities, and visually displaying the knowledge graph paths that have a key impact on the prediction results to assist doctors in identifying potential drug risks that have not been widely reported.

[0011] In one embodiment of the present invention, the construction of a Parkinson's disease domain knowledge graph containing drugs, targets, metabolic enzymes, adverse reactions, and drug interactions, and the extraction of entities and relationships from authoritative medical guidelines, literature, and databases using natural language processing technology, further includes: S11, the knowledge graph containing "metabolic pathway" relationships between drugs and metabolic enzymes, "transport pathway" relationships between drugs and transport proteins, and "potentially causing" relationships between drugs and adverse reactions; S12, the natural language processing technology including a BERT-based named entity recognition model and a relation extraction model, used to extract entities such as drugs, targets, metabolic enzymes, and adverse reactions and their semantic relationships from unstructured text.

[0012] In one embodiment of the present invention, the step of using the knowledge graph embedding model to vectorize the drug and its related entities based on the knowledge graph, and constructing node features by combining the molecular fingerprint information of the drug with the protein sequence features, further includes: S21, the knowledge graph embedding model is a RotatE or ComplEx model, used to map entities such as drugs, targets, metabolic enzymes, and adverse reactions into dense vector representations of 100-500 dimensions; S22, the molecular fingerprint information includes extended connectivity fingerprint (ECFP) or MACCS fingerprint, and the protein sequence features are embedded using a pre-trained protein language model such as ESM.

[0013] In one embodiment of the present invention, the step of inputting the node features into a graph neural network model and analyzing the potential action paths between drugs through a multi-hop relationship aggregation mechanism to predict the probability of drug combinations and adverse reactions or drug-drug interactions further includes: S31, the graph neural network model is a relational graph convolutional network (RelationalGCN) or a graph attention network (GAT) used to handle multi-hop relationships in heterogeneous graph structures; S32, the model generates training data through a negative sampling strategy, wherein negative samples are generated by randomly replacing adverse reaction nodes or drug pairs, and ensuring that the replaced triples do not exist in the knowledge graph or are clinically unreasonable.

[0014] In one embodiment of the present invention, the method further includes: generating a personalized drug dosage adjustment strategy based on a reinforcement learning algorithm, and cross-validating the strategy with drug contraindications and dosage restrictions information in a knowledge graph to ensure that the recommended dosage adjustment scheme complies with clinical safety guidelines.

[0015] To achieve the above objectives, a second aspect of the present invention proposes a Parkinson's disease drug risk prediction device based on knowledge graph embedding and graph neural networks, comprising: a knowledge graph construction module for constructing a Parkinson's disease domain knowledge graph containing drugs, targets, metabolic enzymes, adverse reactions, and drug interactions, and extracting entities and relationships from authoritative medical guidelines, literature, and databases using natural language processing technology; a node feature generation module for vectorizing drugs and related entities based on the knowledge graph using a knowledge graph embedding model, and constructing node features by combining the molecular fingerprint information and protein sequence features of the drugs; a drug risk prediction module for inputting the node features into a graph neural network model, analyzing potential action paths between drugs through a multi-hop relationship aggregation mechanism, and predicting the probability of drug combinations and adverse reactions or drug-drug interactions; and a result display module for outputting the predicted probabilities and visually displaying the knowledge graph paths that have a key impact on the prediction results, to assist doctors in identifying potential drug risks that have not been widely reported.

[0016] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;

[0017] The memory stores computer-executed instructions;

[0018] The processor executes computer execution instructions stored in the memory to implement the method as described in any one of the first aspects.

[0019] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of the first aspects.

[0020] To achieve the above objectives, a fifth aspect of the present invention provides a computer program product that, when executed by a processor, implements the method described in any one of the first aspects.

[0021] The technical solutions provided by the embodiments of the present invention bring at least the following beneficial effects: they can prospectively predict potential adverse drug reactions and drug-drug interaction risks in multi-drug combination therapy for Parkinson's disease, and improve the safety and personalization of treatment plans.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0024] Figure 1 is a flowchart of a Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network according to an embodiment of the present invention.

[0025] Figure 2 is a schematic diagram of the structure of a Parkinson's disease drug risk prediction device based on knowledge graph embedding and graph neural network according to an embodiment of the present invention. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0027] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0028] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0029] Figure 1 is a flowchart of a Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network according to an embodiment of the present invention.

[0030] As shown in Figure 1, the Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network includes the following steps:

[0031] S1 constructs a knowledge graph for the Parkinson's disease domain that includes drugs, targets, metabolic enzymes, adverse reactions, and drug interactions, and extracts entities and relationships from authoritative medical guidelines, literature, and databases using natural language processing technology.

[0032] Specifically, this step aims to construct a knowledge graph in the Parkinson's disease (PD) domain, encompassing drugs, targets, metabolic enzymes, adverse reactions, and drug interactions. It then extracts entities and relationships from authoritative medical guidelines, literature, and databases using Natural Language Processing (NLP) technology. This step is the core of the entire system's knowledge foundation, providing structured and semantic medical knowledge support for subsequent intelligent decision-making and risk prediction.

[0033] At the technical implementation level, the construction of the knowledge graph first relies on the integration of multi-source heterogeneous data, including authoritative medical resources such as the "Chinese Guidelines for the Treatment of Parkinson's Disease (Fourth Edition)," PubMed, DrugBank, KEGG, SIDER, and FAERS. For unstructured text data, a BERT-based Named Entity Recognition (NER) model is used for entity extraction, covering key medical entities such as drugs (e.g., levodopa, pramipexole), targets (e.g., D2 receptor, COMT enzyme), metabolic enzymes (e.g., CYP2D6, CYP3A4), and adverse reactions (e.g., dyskinesia, hallucinations). Relation extraction is achieved through a BERT-based sequence labeling model or a Joint Entity and Relation Extraction (JERE) model, identifying semantic relationships between entities, such as "drug-acts on-target," "drug-causes-adverse reaction," and "drug-interacts with drug."

[0034] At the parameter level, entity recognition models typically use BERT-Base or BERT-Large as the base model, with a maximum input text length of 512 tokens. The number of entity categories is defined based on domain knowledge, generally between 10 and 20. Relation extraction models use a multi-layer Transformer structure, with 20-50 relation types typically covering pharmacology, toxicology, and metabolism. During model training, cross-entropy loss is used, the optimizer is AdamW, the learning rate is 2e-5, the batch size is 16-32, and the training epochs are 3-10. Specific parameters are adjusted based on the dataset size and model convergence.

[0035] In application scenarios, this knowledge graph can serve as the semantic foundation for intelligent decision-making engines, supporting drug recommendation, dosage optimization, adverse reaction warning, and drug interaction analysis. For example, in multi-drug combination scenarios, the system can infer potential metabolic interference risks based on the relationship between "drug A - metabolism via - enzyme X" and "drug B - inhibition of - enzyme X" in the graph.

[0036] The technical value of this step lies in the fact that by constructing structured medical knowledge, it significantly improves the system's ability to understand complex drug relationships, providing high-quality knowledge input for subsequent graph neural network (GNN) and reinforcement learning (RL) algorithms, thereby enhancing the accuracy and interpretability of predictions.

[0037] Furthermore, S1 includes:

[0038] S11, the knowledge graph includes the "metabolic pathway" relationship between drugs and metabolic enzymes, the "transport pathway" relationship between drugs and transport proteins, and the "potential cause" relationship between drugs and adverse reactions.

[0039] Specifically, in this invention, the knowledge graph contains the "metabolic pathway" relationship between drugs and metabolic enzymes, the "transport pathway" relationship between drugs and transport proteins, and the "potential cause" relationship between drugs and adverse reactions. It forms the core knowledge foundation for drug risk prediction and personalized treatment recommendations. The technical implementation of this step is based on medical ontology modeling and relation extraction technology, aiming to model and store the pharmacokinetic characteristics and clinical risk factors of drugs in a structured form.

[0040] In some implementations, this step first uses Natural Language Processing (NLP) technology to extract drug entities and their related biomolecular entities (such as metabolic enzymes, transport proteins, and adverse reactions) from authoritative medical literature, drug databases (such as DrugBank, KEGG, and SIDER), and clinical guidelines. The relation extraction model can be fine-tuned based on a pre-trained language model (such as BioBERT or ClinicalBERT) and uses sequence labeling or Joint Entity and Relation Extraction (JERE) methods to identify semantic relations in the form of triplets, such as (drug A, metabolized via, CYP2D6), (drug B, transported via, P-glycoprotein), (drug C, may cause, hallucinations), etc.

[0041] At the parameter level, relation extraction models are typically trained on labeled datasets. Their performance can be evaluated using F1-score, precision, and recall. A target F1-score of 0.85 or higher is expected to ensure the reliability of the knowledge graph. Relation types within the knowledge graph must adhere to medical ontology standards (such as SNOMED CT, RxNorm, and DrugBank ID) and employ unified naming conventions and semantic encoding to ensure consistency and retrievability across data sources.

[0042] In application scenarios, the relational network constructed in this step provides crucial topological input for subsequent graph neural network (GNN) models. For example, in adverse drug reaction (ADR) prediction tasks, GNNs infer potential ADR risks by aggregating neighborhood information of drug nodes (such as metabolic enzyme inhibition, transporter protein expression levels, and known adverse reaction pathways). In drug-drug interaction (DDI) prediction, the metabolic and transport pathway relationships between drugs can reveal potential enzyme competition or transport inhibition mechanisms, thereby improving the accuracy and clinical interpretability of predictions.

[0043] In terms of technical effectiveness, this step significantly enhances the system's understanding of drug metabolism mechanisms and adverse reaction pathways by structurally modeling key pharmacokinetic relationships between drugs and biomolecules. The constructed triplet relationship network provides rich contextual information for the GNN model, helping to identify potential risks not explicitly recorded in traditional databases, thereby improving the safety and personalization of Parkinson's disease drug treatment plans.

[0044] S12, the natural language processing technology includes a BERT-based named entity recognition model and a relation extraction model, used to extract entities such as drugs, targets, metabolic enzymes, adverse reactions and their semantic relationships from unstructured text.

[0045] Specifically, the Natural Language Processing (NLP) technology employed in this invention includes a BERT-based Named Entity Recognition (NER) model and a Relation Extraction (RE) model, used to extract key entities and their semantic relationships from unstructured medical text. In some implementations, the NER model identifies entities such as drug names, targets, metabolic enzymes, and adverse reactions in the text by fine-tuning a pre-trained BERT model (such as BERT-base, BERT-large, or domain-adapted BioBERT). The model input is medical text that has undergone word segmentation and stemming, and the output is entity type labeling results, such as "drug-levodopa", "target-D2 receptor", "metabolic enzyme-CYP2D6", "adverse reaction-dyskinesia", etc. During entity recognition, a CRF (Conditional Random Field) or BiLSTM-CRF structure can be optionally used as a sequence labeling layer to improve the accuracy of entity boundary recognition. Furthermore, the relation extraction model, building upon entity recognition, leverages BERT's context-aware capabilities and combines them with relation classification tasks to identify semantic relationships between entities, such as "drug - acts on -> target," "drug - metabolizes via -> enzyme," and "drug - causes -> adverse reactions." The model typically employs joint embedding of entity pairs, outputting the relation type probability distribution by appending a fully connected classifier to the BERT output layer. For training data, annotated medical text corpora (such as BioNLP, MedMentions, and the RxNorm extended dataset) can be optionally used, and the model is evaluated using metrics such as F1-score, accuracy, and recall. In terms of parameter settings, the hidden layer dimension of the BERT model is typically 768 or 1024, the maximum sequence length is set to 512, the learning rate is set to 2e-5 to 5e-5, the batch size is 16 to 32, and the number of training epochs is 3 to 10. This step plays a crucial role in the entire system, providing a structured knowledge input foundation for the knowledge graph construction module. It ensures that the system can efficiently and accurately extract medical entities and relationships from unstructured texts such as authoritative guidelines and medical literature, thereby supporting subsequent intelligent decision-making and risk prediction functions.

[0046] S2, Based on the knowledge graph, a knowledge graph embedding model is used to vectorize the drug and its related entities, and node features are constructed by combining the drug's molecular fingerprint information and protein sequence features.

[0047] Specifically, this step involves vectorizing the drug and its related entities based on a knowledge graph embedding model, and constructing node features by combining the drug's molecular fingerprint information and protein sequence features. This is a key preprocessing step in the graph neural network (GNN) drug risk prediction algorithm within the intelligent processing and decision engine module of this invention. Its technical implementation principle is based on a knowledge graph embedding (KGE) and multimodal feature fusion strategy, aiming to provide the GNN model with node representations rich in semantic and biochemical information, thereby improving the accuracy and interpretability of adverse drug reactions (ADRs) and drug-drug interactions (DDIs) predictions.

[0048] In some implementations, knowledge graph embedding models can employ advanced algorithms such as RotatE, ComplEx, or Relational GCN to map entities (e.g., drugs, targets, adverse reactions, metabolic enzymes) and relationships (e.g., "drug-acts on-target", "drug-causes-adverse reaction") in the knowledge graph to a low-dimensional dense vector space (e.g., 100-500 dimensions). The embedding process typically models the semantic relationships between entities by maximizing the scores of known triples while minimizing the scores of negative samples. For example, in RotatE, entity embedding simulates the semantic directionality of relationships through rotation operations in complex space, thereby enhancing the ability to model heterogeneous relationships.

[0049] Furthermore, the feature construction of drug nodes not only relies on KGE embedding vectors but also needs to fuse their molecular fingerprint information. Molecular fingerprints (such as ECFP and MACCS) are binary or sparse vectors based on drug molecule structures, used to characterize the chemical similarity of drugs. In this invention, ECFP_4 (extended connected fingerprint with a diameter of 4) or MACCS key fingerprints can be used, with a length typically of 1024 or 2048 bits. After hashing or normalization, these are concatenated with KGE vectors to form a hybrid feature vector. For protein nodes, pre-trained protein language models (such as ESM-1b and ProtBERT) can be used to extract their sequence embedding vectors, typically 1280 dimensions or higher, to capture the structural and functional features of proteins.

[0050] Optionally, node features can also be weighted and fused using an attention mechanism. For example, multi-head attention can be used to select features from KGE, molecular fingerprints, and protein embeddings to enhance the model's focus on key biochemical properties. The fused node features will then serve as input to GNN models (such as GraphSAGE and GAT) for subsequent link prediction tasks, such as the identification of potential ADRs and DDIs.

[0051] The technical value of this step lies in its significant improvement in the representation ability of drug nodes in graph structures through the fusion of multi-source heterogeneous features. This enables GNN models to more accurately capture the complex relationships between drugs and adverse reactions, as well as between drugs themselves. In practical applications, this method can be widely used for risk prediction in multi-drug combination scenarios for Parkinson's disease, especially when dealing with novel drug combinations or rare adverse reactions, demonstrating strong generalization ability and prediction accuracy.

[0052] Furthermore, S2 includes:

[0053] S21, the knowledge graph embedding model is a RotatE or ComplEx model, used to map entities such as drugs, targets, metabolic enzymes, and adverse reactions into dense vector representations of 100-500 dimensions.

[0054] Specifically, in this invention, the Knowledge Graph Embedding (KGE) model employs RotatE or ComplEx models to map entities such as drugs, targets, metabolic enzymes, and adverse reactions into dense vector representations of 100-500 dimensions, thereby enabling the modeling and reasoning of complex semantic relationships between entities. This step is a crucial component in constructing an intelligent drug risk prediction model, providing high-quality input features for subsequent predictions of adverse drug reactions (ADRs) and drug-drug interactions (DDIs) using graph neural networks (GNNs).

[0055] At the technical implementation level, the RotatE model is based on entity-relation-entity triples in complex space. Its core idea is to treat relations as rotation operations between entity vectors; that is, by rotating the head entity vector in complex space, an approximate representation of the tail entity is obtained. Its loss function typically employs a negative sampling strategy, combined with the scoring function of the triples (such as Euclidean distance or inner product) for optimization. The ComplEx model, on the other hand, extends the ability of TransE-like models to handle asymmetric relations by introducing complex vectors and a bilinear interaction mechanism.

[0056] Regarding parameter settings, the embedding dimension of entities and relationships is typically between 100 and 500 dimensions, with the specific dimension size adjusted based on the size of the knowledge graph and the complexity of the task. For example, in this system, considering the multi-hop relationships and semantic complexity between Parkinson's disease drugs and adverse reactions, the embedding dimension can be set to 300 dimensions to balance the model's expressive power and computational efficiency. During training, the learning rate is generally set to 10. -3 Up to 10 -4The batch size is 128-256, the optimizer can be Adam or AdamW, and the number of training epochs is usually 100-300 to ensure that the model converges and has good generalization ability.

[0057] In application scenarios, this embedding model is used to construct vector representations of entities such as drugs, targets, metabolic enzymes, and adverse reactions, serving as input node features for graph neural networks (such as GAT and R-GCN). By mapping entities to a unified semantic space, the system can capture the potential associations between drugs and their targets, metabolic pathways, and adverse reactions, thereby improving the ability to predict unknown or hidden drug risks.

[0058] The technical value of this step lies in the fact that by structuring heterogeneous medical knowledge into low-dimensional dense vectors, it not only enhances the model's ability to understand the relationships between entities, but also provides highly interpretable and semantically rich input representations for subsequent graph neural network reasoning, which is an important foundation for realizing intelligent and personalized drug risk prediction.

[0059] S22, the molecular fingerprint information includes extended connectivity fingerprint (ECFP) or MACCS fingerprint, and the protein sequence features are used to generate an embedded representation through a pre-trained protein language model such as ESM.

[0060] Specifically, in the drug risk prediction technology of this invention, the extraction of molecular fingerprint information and protein sequence features are key input steps for constructing drug-adverse reaction (ADR) and drug-drug interaction (DDI) prediction models. Specifically, molecular fingerprint information is characterized using Extended-Connectivity Fingerprints (ECFP) or Molecular Access System Fingerprints (MACCS), while protein sequence features are embedded using a pre-trained protein language model (such as ESM, Evolutionary Scale Modeling).

[0061] In some implementations, ECFP fingerprints are generated based on the topological structure of drug molecules, through iterative expansion of atomic neighborhoods. Typically, ECFP4 or ECFP6 versions are used, with bit widths set to 1024 or 2048 bits to capture local structural features of the molecule. The ECFP generation process includes: first, initial feature encoding of each atom (such as atom type, charge, hybridization state, etc.); then, iterative expansion of its neighborhood to generate substructural features of different radii; and finally, a fixed-length binary fingerprint is formed through hash mapping and bit vector compression. MACCS fingerprints, on the other hand, are based on predefined 166 seed structure rules, directly determining whether a molecule contains specific chemical groups, and are suitable for quickly matching the relationship between known structural features and drug effects.

[0062] Furthermore, protein sequence feature extraction employs pre-trained protein language models (such as ESM-1b, ESM-2, or the more advanced ESM-3). These models are trained on large-scale protein sequence data and can capture the evolutionary conservation, domain features, and functional information of proteins. In this invention, after the protein sequence is input into the model, a context-aware embedding representation is generated using a Transformer architecture. The output is a continuous vector of length 640-1280 dimensions, the specific dimension depending on the selected model version. This embedding vector can be used as feature input for protein nodes such as drug targets, metabolic enzymes, or transport proteins, for node feature construction in graph neural networks (GNNs).

[0063] This step plays a crucial role in the overall technical solution, as its output features directly influence the GNN model's predictive ability for drug-adverse reactions and drug-drug interactions. By fusing molecular fingerprints with protein embedding features, the system can more comprehensively model the complex relationship between the chemical structure and biological mechanism of action of drugs, thereby improving the accuracy and generalization ability in identifying potential risks. For example, when predicting whether a drug-drug interaction (DDI) exists between drug A and drug B, the GNN will simultaneously consider the ECFP features of both drugs (reflecting chemical similarity) and the ESM embeddings of their targets (reflecting biological mechanism similarity), using multi-hop relationship reasoning to uncover potential pharmacological conflict paths.

[0064] In terms of technical specifications, ECFP fingerprint generation typically employs iterative expansion with a radius of 2-3, and a bit depth of 1024 or 2048 to ensure sufficient coverage of molecular structural information. The embedding dimension of ESM models is generally 640-1280 dimensions, which can be adjusted according to the model version (e.g., 640 dimensions for ESM-1b, 1280 dimensions for ESM-2). Furthermore, protein sequence input needs to be standardized, such as removing non-standard amino acid symbols, standardizing sequence length, or using truncation / padding strategies to adapt to model input requirements.

[0065] In summary, this step, by structuring drug molecule and protein sequence information, provides high-quality node feature input for subsequent graph neural network models, and is one of the core foundations for achieving in-depth drug risk prediction and personalized treatment recommendations.

[0066] S3. Input the node features into the graph neural network model, analyze the potential action paths between drugs through a multi-hop relationship aggregation mechanism, and predict the probability of drug combinations and adverse reactions or drug-drug interactions.

[0067] Specifically, inputting the node features into the graph neural network model and analyzing the potential action paths between drugs through a multi-hop relationship aggregation mechanism to predict the probability of drug combinations and adverse reactions or drug-drug interactions (ADRs / DDIs) is one of the core functions of the intelligent processing and decision engine module in this invention. Its technical implementation is based on the synergistic effect of Knowledge Graph Embedding (KGE) and Graph Neural Network (GNN).

[0068] In the specific implementation, node features are generated by knowledge graph embedding models (such as RotatE, ComplEx, or RelationalGCN). Each drug node is mapped to a low-dimensional (e.g., 100-500 dimensional) dense vector, which integrates multi-dimensional semantic information such as the drug's chemical structure (e.g., ECFP molecular fingerprint), target (e.g., dopamine receptor subtype), metabolic enzymes (e.g., CYP2D6, CYP3A4), and transport proteins (e.g., P-glycoprotein). The node feature vector is aggregated through multi-hop relationships using graph neural network models (e.g., GAT, GraphSAGE, or R-GCN). That is, the relationship information between the drug node and its neighboring nodes (e.g., targets, enzymes, adverse reactions) is aggregated layer by layer through a message-passing mechanism, thereby constructing a context-aware representation of the drug in a complex biological network.

[0069] In multi-hop aggregation, the model distinguishes the semantics of different relationships by edge type (such as "acts on", "metabolic via", "causes", etc.) and uses relation-aware aggregation functions (such as Relational Attention or Relation-specific Message Aggregation) to weight and fuse different relationship paths. For example, when predicting whether there is a DDI between drug A and drug B, the model aggregates the relationship between drug A and metabolic enzyme C, the relationship between drug B and the same metabolic enzyme C, and the inhibitory / inducing properties of enzyme C in drug metabolism, thereby inferring the potential conflict between the two in the metabolic pathway.

[0070] The model's output layer employs a probabilistic regression or classification structure, combined with a Sigmoid or Softmax activation function, to output the probability of drug combinations occurring with specific adverse reactions or interactions. During training, known "drug-adverse reaction" or "drug-drug-interaction" triples from the knowledge graph are used as positive samples, while negative samples are generated through negative sampling strategies (such as random tail entity replacement or adversarial negative sampling), thus constructing a supervised learning task. Model evaluation metrics include AUC-ROC, F1-score, and Precision@K to ensure the clinical reliability of the prediction results.

[0071] This step is of critical value in the system, especially in multi-drug combination scenarios. It can identify potential risks not covered by conventional databases, providing doctors with forward-looking and interpretable drug risk warnings, thereby improving the safety and personalization of Parkinson's disease treatment plans.

[0072] Furthermore, S3 includes:

[0073] S31, the graph neural network model is a relational graph convolutional network (Relational GCN) or a graph attention network (GAT), used to handle multi-hop relationships in heterogeneous graph structures.

[0074] Specifically, in this invention, the step "the graph neural network model is a relational graph convolutional network (RelationalGCN) or a graph attention network (GAT) used to handle multi-hop relationships in heterogeneous graph structures" is one of the core implementation methods of the potential ADR and DDI prediction algorithm based on knowledge graph embedding and graph neural networks (GNN) in the intelligent processing and decision engine module. This step aims to perform in-depth prediction of potential adverse reactions (ADRs) and drug-drug interactions (DDIs) in Parkinson's disease drug treatment regimens through graph neural networks, with particular attention to complex risks that may be caused by multi-hop relationships but are not explicitly recorded in traditional databases.

[0075] In some implementations, the system employs Relational GCN or GAT as the graph neural network model to handle the heterogeneous nodes and multi-type edge relationships contained in the knowledge graph. Relational GCN introduces a relation-specific weight matrix to perform independent message passing and aggregation for each relation type (such as "drug-acts on-target", "drug-metabolism via-enzyme", etc.), thereby effectively modeling the complex semantic associations between drugs and their multi-hop context entities. GAT, on the other hand, uses an attention mechanism to assign dynamic weights to different neighbor nodes, enhancing the model's sensitivity to critical paths, and is particularly suitable for graph structures with noise or sparse connections.

[0076] At the parameter level, models are typically trained in an embedding space of 100-500 dimensions. Node features are formed by concatenating entity vectors generated by knowledge graph embedding (KGE) models (such as RotatE and ComplEx) with drug molecule fingerprints (such as ECFP and MACCS). In multi-hop relationship modeling, the number of model layers is generally set to 2-4 layers to balance computational complexity and context awareness. During training, positive samples use known triples from the knowledge graph, such as "drug-cause -> adverse reaction" or "drug A-interacts with drug B," while negative samples are generated by randomly replacing tail entities or through adversarial sampling to ensure the diversity and rationality of the training data.

[0077] In practical applications, this step is used to process the graph structure representation of a patient's current medication combination. The GNN model aggregates neighborhood information (such as targets, metabolic enzymes, known adverse reactions, etc.) of drug nodes and calculates the probability of association between them and potential adverse reaction or interacting nodes. For example, when the system detects that a patient is simultaneously using levodopa and a certain CYP2D6 inhibitor, the GNN model can identify the potential risk of elevated blood drug concentrations caused by the drug pair through multi-hop pathways such as "metabolic enzyme-inhibitor-drug" and output a corresponding risk score.

[0078] In terms of technical effectiveness, this step significantly enhances the depth and breadth of drug risk prediction. It not only identifies known risks but also uncovers potential, unreported adverse drug reactions (ADRs) and drug-induced drug defects (DDIs) through implicit relationships within the graph structure. Its multi-hop relationship-based modeling capability enables the system to possess stronger contextual understanding and reasoning abilities when facing complex multi-drug combination scenarios. This provides clinicians with more forward-looking and interpretive evidence for drug safety assessment, enhancing the system's clinical applicability and decision-making credibility.

[0079] S32, the model generates training data through a negative sampling strategy, wherein negative samples are generated by randomly replacing adverse reaction nodes or drug pairs, and it is ensured that the replaced triples do not exist in the knowledge graph or are clinically unreasonable.

[0080] Specifically, in the drug risk prediction technology of this invention, the model generates training data through a negative sampling strategy, which is a key step in improving the generalization ability and discrimination performance of graph neural networks (GNNs) in adverse drug reaction (ADR) and drug-drug interaction (DDI) prediction tasks. This strategy is mainly used to construct a link prediction task dataset for training the GNN model. By introducing negative samples, the model can learn to distinguish between real drug-adverse reaction or drug-drug interaction relationships and non-existent or unreasonable spurious relationships.

[0081] In practice, negative samples are generated by randomly replacing adverse reaction nodes or drug nodes to construct new triples (drug A, relation R, adverse reaction X) or (drug A, relation I, drug B). The replacement operation must meet two core conditions: first, the replaced triple must not exist in the knowledge graph; second, it must be logically unreasonable in clinical practice to avoid introducing misleading data. For example, based on a known positive sample of "drug A - causing -> adverse reaction X", "adverse reaction X" can be randomly replaced with another adverse reaction Y that does not have a relationship with drug A in the knowledge graph. Then, it can be checked whether this adverse reaction Y has a potential association with drug A in the clinical context (e.g., through pharmacological mechanisms, metabolic pathways, etc.).

[0082] Furthermore, to ensure that the distribution of negative samples closely approximates the real data, a negative sampling method based on embedding distance can be adopted. This involves selecting nodes that are farthest from the positive samples in the embedding space as negative samples, thereby enhancing the model's sensitivity to semantic differences. Additionally, an adversarial negative sampling mechanism can be introduced. By jointly training a generative adversarial network (GAN) or a negative sampler with the model, more challenging negative samples can be dynamically generated, improving the model's robustness.

[0083] Regarding parameter settings, the ratio of negative samples to positive samples is typically set between 1:1 and 1:5, with the specific ratio adjustable based on the sparsity of the dataset. During sampling, a maximum number of attempts (e.g., 10) can be set to ensure that the generated negative samples meet the above two conditions. Simultaneously, a clinical rationality verification module can be incorporated, introducing medical expert rules or semantic constraints based on medical knowledge to filter out negative samples that are clearly unreasonable in pharmacology or clinical practice.

[0084] This step is of significant value in the overall technical solution, particularly in predicting potential risks in multi-drug combination scenarios. By introducing appropriate negative samples, the model can more accurately identify unrecorded but clinically significant adverse drug reactions (ADRs) and drug-induced drug reactions (DDIs), thereby improving the sensitivity and specificity of predictions. Furthermore, a well-designed negative sampling strategy helps alleviate the long-tail distribution problem in knowledge graphs, enhancing the model's ability to identify rare but high-risk events and providing clinicians with more forward-looking and practical drug risk warning information.

[0085] S4, output the predicted probability, and visualize the knowledge graph path that has a key impact on the prediction result to help doctors identify potential drug risks that have not been widely reported.

[0086] Specifically, this step involves outputting predicted probabilities and visually displaying the knowledge graph paths that have a key impact on the prediction results to assist doctors in identifying potential drug risks that are not widely reported. This step is one of the core functions of the auxiliary diagnosis and treatment suggestion output module (1.4) in the intelligent auxiliary diagnosis and treatment system of this invention, aiming to improve the interpretability and clinical applicability of the system's recommendation results.

[0087] At the technical implementation level, the system first structurally integrates the predicted probabilities output by the potential ADR and DDI prediction algorithms based on graph neural networks (GNN) in the intelligent processing and decision engine module (1.3). The predicted probabilities are typically represented as floating-point numbers, ranging from 0 to 1, representing the likelihood of adverse drug reactions or drug-drug interactions occurring. For example, the probability of a serious DDI occurring between drug A and drug B is 0.83. The system maps this value to the corresponding risk level (e.g., high, medium, low) and combines it with clinical severity standards (e.g., WHO Adverse Drug Reaction Classification Standard CTCAE v5.0) for risk classification.

[0088] Furthermore, the system extracts the knowledge graph paths that contribute most to the prediction results from the GNN model through the interpretable AI (XAI) component. Specifically, it uses GNNExplainer or attention-based visualization methods to identify entities and relationships in the graph that are highly relevant to the prediction results. For example, when predicting that drug A may cause abnormal liver function, the system highlights key pathways such as "drug A → metabolism via → CYP2D6 → inhibition → metabolic pathway of drug B", revealing its potential pharmacokinetic conflict mechanism.

[0089] In its visualization implementation, the system employs interactive graph display tools (such as D3.js and Cytoscape.js) to present entity nodes and relational edges in the knowledge graph in a topological structure. The node color, size, edge thickness, and color variations are all correlated with the predicted probability and risk level; for example, high-risk paths are highlighted in red, medium-risk paths in orange, and low-risk paths in gray. Doctors can click on nodes to view their attribute information (such as the SMILES structure of drugs, gene polymorphism information of metabolic enzymes, etc.) and understand the basis for predictions through path tracing functionality.

[0090] This step is of significant value in real-world clinical settings, particularly when using multiple drugs in combination or novel drug combinations, as it helps physicians identify potential risks not covered by standard databases. For example, when recommending the combined use of dopamine receptor agonists and antidepressants for elderly patients with Parkinson's disease (PD), the system can reveal potential conflicts between the two in the CYP450 metabolic pathway, thereby avoiding the possible accumulation of adverse reactions.

[0091] Through this step, the system not only outputs quantitative risk prediction results, but also provides a visual explanation path based on medical knowledge, which significantly enhances the transparency and credibility of AI decision-making and provides clinicians with scientific and intuitive decision support.

[0092] The Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural networks in this invention can effectively integrate medical knowledge and multimodal patient data. Through knowledge graph embedding and graph neural networks, it achieves prospective prediction of adverse drug reactions and drug-drug interactions in Parkinson's disease, improving the safety and personalization of treatment plans.

[0093] In addition, embodiments of the present invention also generate personalized drug dosage adjustment strategies based on reinforcement learning algorithms, and cross-validate these strategies with drug contraindications and dosage restrictions in a knowledge graph to ensure that the recommended dosage adjustment schemes comply with clinical safety guidelines.

[0094] Specifically, this step involves generating personalized drug dosage adjustment strategies based on reinforcement learning (RL) algorithms and cross-validating these strategies with drug contraindications and dosage restrictions in a knowledge graph to ensure that the recommended treatment complies with clinical safety guidelines. In some implementations, this process first constructs an offline reinforcement learning framework, utilizing historical clinical data (such as electronic medical records (EMR) and clinical trial data) for policy learning, avoiding potential treatment risks associated with online exploration. The RL model can employ offline variants of algorithms such as Deep Q-Networks (DQN) or Soft Actor-Critic (SAC) to adapt to the characteristics of limited data and the need for highly safe decision-making in medical scenarios.

[0095] State space (State, S) t The action space contains multidimensional clinical characteristics of the patient, such as UPDRS score, drug response, comorbidities, metabolic enzyme information, and drug dosage history. It is usually represented in vector form and standardized or normalized. t This is defined as an operation involving adjustments to drug dosage, including increasing, decreasing, maintaining, or adding / discontinuing a specific drug. All actions must be filtered through dosage restrictions, contraindications, and drug interaction rules in the knowledge graph to ensure clinical feasibility. The reward function (Reward, R) t+1 Taking into account indicators such as symptom improvement, occurrence of adverse reactions, aggravation of exercise complications and changes in quality of life, a weighted linear combination was adopted, and the weight coefficients (w1, w2, w3, w4) were optimized and determined based on feedback from clinical experts and historical data.

[0096] During strategy learning, the system optimizes the dosage adjustment strategy by maximizing the cumulative discount reward (γ∈[0.9,0.99]). The generated policy network outputs the optimal action suggestion for the current state. Subsequently, this suggestion is cross-validated with entity relationships in the knowledge graph, such as checking for conflicts between "patient - has -> liver dysfunction" and "drug - contraindicated in -> liver dysfunction", or verifying whether the recommended dose exceeds the "drug - maximum recommended dose" threshold. If a conflict exists, the system will automatically correct or reject the action to ensure that the recommended plan is within the clinical safety boundary. This step plays a crucial role in the intelligent processing and decision engine module, providing a scientific, safe, and interpretable dosage optimization path for individualized drug treatment of Parkinson's disease patients, significantly improving the accuracy and safety of treatment.

[0097] The Parkinson's disease drug risk prediction method based on knowledge graph embedding and graph neural network in this invention introduces reinforcement learning algorithm to generate personalized drug dosage adjustment strategy, and combines contraindication and dosage restriction information in knowledge graph for cross-validation, which further improves the clinical safety and dynamic adaptability of drug recommendation plan, and realizes more accurate Parkinson's disease drug management that is more in line with individualized treatment needs.

[0098] To achieve the above embodiments, the present invention also proposes a Parkinson's disease drug risk prediction device based on knowledge graph embedding and graph neural networks. Figure 2 is a schematic diagram of the structure of a Parkinson's disease drug risk prediction device based on knowledge graph embedding and graph neural networks provided in an embodiment of the present invention. As shown in Figure 2, the device includes:

[0099] The knowledge graph construction module 100 is used to construct a knowledge graph in the field of Parkinson's disease that includes drugs, targets, metabolic enzymes, adverse reactions and drug interactions, and to extract entities and relationships from authoritative medical guidelines, literature and databases through natural language processing technology.

[0100] The node feature generation module 200 is used to vectorize the drug and its related entities based on the knowledge graph using the knowledge graph embedding model, and to construct node features by combining the molecular fingerprint information of the drug with protein sequence features.

[0101] The drug risk prediction module 300 is used to input the node features into a graph neural network model, analyze the potential action paths between drugs through a multi-hop relationship aggregation mechanism, and predict the probability of drug combinations and adverse reactions or drug-drug interactions.

[0102] The results display module 400 is used to output the predicted probability and visualize the knowledge graph path that has a key impact on the prediction results to help doctors identify potential drug risks that have not been widely reported.

[0103] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0104] To implement the above embodiments, the present invention also proposes an electronic device, comprising: a processor, and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0105] To implement the above embodiments, the present invention also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0106] To implement the above embodiments, the present invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0107] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0108] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.

[0109] This invention is intended to provide implementation schemes for users to selectively prevent the use or access to personal information data. That is, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information can be de-identified to protect user privacy.

[0110] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0111] Furthermore, the terms "first" and "second" 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0112] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0113] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0114] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0115] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0116] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0117] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0118] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0119] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting the risk of Parkinson's disease medications based on knowledge graph embedding and graph neural networks, characterized in that, include: S1. Construct a knowledge graph for the Parkinson's disease domain, including drugs, targets, metabolic enzymes, adverse reactions, and drug interactions. Extract entities and relationships from authoritative medical guidelines, literature, and databases using natural language processing technology. S2. Based on the knowledge graph, use a knowledge graph embedding model to vectorize drugs and related entities, and construct node features by combining drug molecular fingerprint information and protein sequence characteristics. S3. Input the node features into a graph neural network model, analyze potential action paths between drugs through a multi-hop relationship aggregation mechanism, and predict the probability of drug combinations and adverse reactions or drug-drug interactions. S4. Output the predicted probabilities and visualize the knowledge graph paths that have a key impact on the prediction results to assist doctors in identifying potential drug risks that are not widely reported.

2. The method as described in claim 1, characterized in that, The construction of a Parkinson's disease knowledge graph encompassing drugs, targets, metabolic enzymes, adverse reactions, and drug interactions, and the extraction of entities and relationships from authoritative medical guidelines, literature, and databases using natural language processing (NLP) techniques, further includes: the knowledge graph containing metabolic pathway relationships between drugs and metabolic enzymes, transport pathway relationships between drugs and transport proteins, and potential causative relationships between drugs and adverse reactions; the NLP techniques include BERT-based named entity recognition and relation extraction models, used to extract entities such as drugs, targets, metabolic enzymes, and adverse reactions, along with their semantic relationships, from unstructured text.

3. The method as described in claim 1, characterized in that, The method of vectorizing drugs and related entities based on the knowledge graph using a knowledge graph embedding model and constructing node features by combining the molecular fingerprint information of the drug with protein sequence features also includes: the knowledge graph embedding model is a RotatE or ComplEx model, used to map entities such as drugs, targets, metabolic enzymes, and adverse reactions into dense vector representations of 100-500 dimensions; the molecular fingerprint information includes extended connectivity fingerprints ECFP or MACCS fingerprints; and the protein sequence features are embedded using a pre-trained protein language model such as ESM.

4. The method as described in claim 1, characterized in that, The step of inputting the node features into a graph neural network model and analyzing the potential action paths between drugs through a multi-hop relationship aggregation mechanism to predict the probability of drug combinations and adverse reactions or drug-drug interactions further includes: S31, the graph neural network model is a relational graph convolutional network or a graph attention network, used to handle multi-hop relationships in heterogeneous graph structures; S32, the model generates training data through a negative sampling strategy, wherein negative samples are generated by randomly replacing adverse reaction nodes or drug pairs, and ensuring that the replaced triples do not exist in the knowledge graph or are clinically unreasonable.

5. The method as described in claim 1, characterized in that, Also includes: Personalized drug dosage adjustment strategies are generated using reinforcement learning algorithms, and these strategies are cross-validated with drug contraindications and dosage restrictions in a knowledge graph to ensure that the recommended dosage adjustment schemes comply with clinical safety guidelines.

6. A Parkinson's disease drug risk prediction device based on knowledge graph embedding and graph neural networks, characterized in that, include: The knowledge graph construction module is used to build a knowledge graph in the field of Parkinson's disease that includes drugs, targets, metabolic enzymes, adverse reactions and drug interactions, and extracts entities and relationships from authoritative medical guidelines, literature and databases through natural language processing technology. The node feature generation module is used to vectorize the drug and its related entities based on the knowledge graph using the knowledge graph embedding model, and to construct node features by combining the drug's molecular fingerprint information and protein sequence features. The drug risk prediction module is used to input the node features into the graph neural network model, analyze the potential action paths between drugs through a multi-hop relationship aggregation mechanism, and predict the probability of drug combinations and adverse reactions or drug-drug interactions. The results display module is used to output the predicted probability and visualize the knowledge graph path that has a key impact on the prediction results to help doctors identify potential drug risks that have not been widely reported.

7. The apparatus as claimed in claim 6, characterized in that, Also includes: The dosage adjustment strategy generation module is used to generate personalized drug dosage adjustment strategies based on reinforcement learning algorithms, and cross-validate the strategies with drug contraindications and dosage restrictions in the knowledge graph to ensure that the recommended dosage adjustment plan complies with clinical safety guidelines.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method as described in any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-5.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1-5.

Images