Clinical test result prediction method and system based on hierarchical interaction graph neural network

Abstract

The invention provides a clinical test result prediction method and system based on a hierarchical interaction graph neural network. The method comprises the following steps: coding clinical test information; pre-training a drug embedding vector and a disease embedding vector; constructing an interaction graph, wherein the interaction graph comprises four layers of nodes; taking the interaction graph as input, defining a label of the interaction graph according to a test result, enhancing node embedding by utilizing an updating rule of an improved graph convolutional network, and introducing a learnable layer-independent attention matrix to calculate an importance weight of an edge so as to adaptively adjust a transfer weight of information between nodes according to the importance weight; and constructing a loss function of the improved graph convolutional network, and completing training of the improved graph convolutional network according to the loss function. According to the method, a more accurate and reliable solution can be provided for prediction of clinical test results, and the method has important application value.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical trial prediction, and in particular to a clinical trial result prediction method and system based on a hierarchical interactive graph neural network. Background Art

[0002] Clinical trials are a core part of new drug development, but their success rate is extremely low (less than 10% on average) and they are costly. Traditional prediction methods have the following limitations: 1. Insufficient data integration: Existing methods usually rely on single-modality data (such as drug molecular structure or disease description) and fail to fully integrate the complex relationships between multi-source data (drugs, diseases, trial inclusion and exclusion criteria, etc.).

[0003] 2. Lack of dynamic interaction modeling: The success of clinical trials is dynamically influenced by multiple factors such as drug metabolomics (ADMET) properties, disease risk, and trial design. Existing models find it difficult to capture these hierarchical cross-modal interactions.

[0004] 3. Poor interpretability: Existing black-box models cannot reveal the contribution of key influencing factors (such as drug toxicity or disease risk) to prediction results, limiting their application in decision support.

[0005] 4. Data missing problem: Clinical trial data often have some modal missing (such as molecular structure not disclosed due to commercial secrets), and traditional methods lack an effective filling mechanism. Summary of the Invention

[0006] The purpose of the present invention is to provide a clinical trial result prediction method and system based on a hierarchical interactive graph neural network to at least solve one problem existing in the background technology.

[0007] In a first aspect, the present invention provides a method for predicting clinical trial results based on a hierarchical interactive graph neural network, the method comprising: Acquire clinical trial information for multiple patients, each clinical trial information including drug information, disease information, trial inclusion and exclusion criteria, and trial results, and encode the drug information, disease information, and trial inclusion and exclusion criteria into an embedding vector of uniform dimension; Pre-training drug embedding vectors based on ADMET properties and obtaining historical disease trial risks for each disease to pre-train disease embedding vectors based on the historical disease trial risks; An interaction graph is constructed using drug embedding vectors, disease embedding vectors, and trial inclusion / exclusion criteria embedding vectors as input nodes, and ADMET attribute embedding vectors and disease risk embedding vectors as external knowledge nodes. Taking the interaction graph as input and defining the labels of the interaction graph based on experimental results, the method enhances node embedding using the updated rule of the improved graph convolutional network, introduces a learnable layer-independent attention matrix to calculate the importance weights of edges, and adaptively adjusts the transmission weight of information between nodes according to the importance weights; Construct a loss function for the improved graph convolutional network, and complete training of the improved graph convolutional network according to the loss function.

[0008] Furthermore, the step of encoding drug information, disease information, and trial inclusion / exclusion criteria into an embedding vector of uniform dimension includes: Medication information is encoded according to the following formula:

[0009] in, is the drug embedding vector, is the total number of drug molecules contained in the mth drug, is the jth drug molecule contained in the mth drug, is the molecular encoding function, for -dimensional real vector; Disease information is encoded according to the following formula:

[0010]

[0011]

[0012] in, is the disease embedding vector, is the graph attention model, For the patient's diseases, is the total number of diseases suffered by the patient, For diseases embedded using GRAM, is the attention weight, is a feed-forward network with a single hidden layer, Represents the set of all ancestor nodes of i. The ancestors of a node represent the higher-level categories of the current node. is the jth randomly initialized learnable embedding vector, is the i-th randomly initialized learnable embedding vector, T is the transpose, is the current disease node itself; The trial inclusion and exclusion criteria were coded according to the following formula:

[0013] in, For the test, the standard embedding vector is The inclusion criteria for the test were Encode the function for the eligibility criteria.

[0014] Furthermore, the step of pre-training the drug embedding vector according to the ADMET attributes and obtaining the historical disease trial risk of each disease to pre-train the disease embedding vector according to the historical disease trial risk includes: Taking the drug embedding vector as input and the predicted score of ADMET property as the target, a multi-layer high-speed network is used to generate the ADMET property embedding vector:

[0015] The ADMET attribute embedding vectors are input into a single-layer fully connected neural network, and the output of the fully connected layer is processed by the Sigmoid function to obtain the binary prediction value:

[0016] The loss function of the single-layer fully connected neural network is set to:

[0017] in, Embedding vectors for ADMET properties, including absorption properties , distribution attribute embedding vector , metabolic attribute embedding vector , excretion attribute embedding vector , toxicity attribute embedding vector , For multi-layer high-speed network, is a single-layer fully connected neural network. is the binary prediction value, , is the true binary label, , is the first loss value.

[0018] Furthermore, the step of pre-training the drug embedding vector according to the ADMET attributes and obtaining the historical disease trial risk of each disease to pre-train the disease embedding vector according to the historical disease trial risk also includes: Taking the disease embedding vector as input and the disease risk score as the target, a two-layer high-speed network is used to generate the disease risk embedding vector:

[0019] The disease risk embedding vector is input into a single-layer fully connected neural network, and the output of the fully connected layer is processed by the Sigmoid function to obtain the disease risk prediction value:

[0020] The loss function of the single-layer fully connected neural network is set to:

[0021] in, is the disease risk embedding vector, It is a double-layer high-speed network. is the disease risk prediction value, , is a single-layer fully connected neural network. is the second loss value, is a binary label with trial success or failure as a function of disease information, .

[0022] Furthermore, the step of constructing an interaction graph using the drug embedding vector, the disease embedding vector, and the trial inclusion / exclusion criteria embedding vector as input nodes, and the ADMET attribute embedding vector and the disease risk embedding vector as external knowledge nodes includes: The interaction graph also includes an aggregation node and a prediction node; The aggregation node includes: 、 and Interaction nodes ,connect 、 、 、 、 Pharmacokinetic nodes , through the disease risk embedding vector Enhanced Interaction Nodes Strong interaction nodes ; The prediction node To connect pharmacokinetic nodes and enhanced interactive nodes Node.

[0023] Furthermore, the steps of taking the interaction graph as input, defining the labels of the interaction graph according to the experimental results, enhancing node embedding using the updated rule of the improved graph convolutional network, introducing a learnable layer-independent attention matrix to calculate the importance weights of edges, and adaptively adjusting the transmission weights of information between nodes according to the importance weights include: Define A as the adjacency matrix of the interaction graph and initialize all node embeddings:

[0024] in, is the initial node embedding matrix, K=13, for dimensional real vector space; The node embedding is enhanced according to the following formula:

[0025] in, The updated node embedding matrix for the improved graph convolutional network at layer l, is the bias parameter of the lth layer, is a learnable layer-independent attention matrix, , is the weight matrix of the l-th layer transformation embedding, is, L is the depth of the improved graph convolutional network, is the unit multiplication, is a linear rectification function; A learnable layer-independent attention matrix is ​​constructed according to the following formula:

[0026] in, is the element in the i-th row and j-th column of the matrix V, which is used to represent the importance weight of the edge between node i and node j. is a two-layer fully connected neural network. is the embedding vector of node i, is the embedding vector of node j, , , yes The transpose of the i-th row of yes The transpose of the jth row of For splicing operation.

[0027] Furthermore, the step of constructing a loss function of the improved graph convolutional network and completing the training of the improved graph convolutional network according to the loss function includes: If the test result is successful, the label is defined as 1; if the test result is failed, the label is defined as 0; The output of the Lth layer of the improved graph convolutional network is used as the success probability of the experiment:

[0028] in, is the probability of success of the experiment, is the prediction node representation of the Lth layer; The loss function of the improved graph convolutional network is constructed according to the following formula:

[0029] in, is the third loss value, is the true binary label, is the probability of success of the experiment, , .

[0030] Furthermore, the method further comprises: The improved graph convolutional network also includes a filling module. When the drug information is missing, the filling module generates a drug replacement embedding vector based on the disease embedding vector associated with the drug information and the trial inclusion and exclusion criteria embedding vector, and replaces the actual drug embedding vector with the drug replacement embedding vector to construct an interaction graph:

[0031] The loss function of the padding module is constructed according to the following formula, and the padding module is trained to minimize the fourth loss value:

[0032] in, is the drug replacement embedding vector, To fill the network, is the fourth loss value.

[0033] In a second aspect, the present invention provides a clinical trial result prediction system based on a hierarchical interactive graph neural network, the system comprising: A historical trial information acquisition module is used to obtain clinical trial information of multiple patients, each of which includes drug information, disease information, trial inclusion and exclusion criteria, and trial results, and encode the drug information, disease information, and trial inclusion and exclusion criteria into an embedded vector of uniform dimension; A vector pre-training module, configured to pre-train drug embedding vectors based on ADMET properties and obtain historical disease trial risks for each disease to pre-train disease embedding vectors based on the historical disease trial risks; An interaction graph construction module is used to construct an interaction graph using drug embedding vectors, disease embedding vectors, and trial inclusion / exclusion criteria embedding vectors as input nodes, and ADMET attribute embedding vectors and disease risk embedding vectors as external knowledge nodes; A node enhancement module is configured to take the interaction graph as input, define labels for the interaction graph based on experimental results, enhance node embeddings using an improved graph convolutional network update rule, and introduce a learnable layer-independent attention matrix to calculate edge importance weights, so as to adaptively adjust the transfer weight of information between nodes based on the importance weights; The model training module is used to construct a loss function of the improved graph convolutional network and complete the training of the improved graph convolutional network according to the loss function.

[0034] In a third aspect, the present invention provides a readable storage medium, which stores one or more programs, which, when executed by a processor, implement the above-mentioned clinical trial result prediction method based on hierarchical interactive graph neural network.

[0035] In a fourth aspect, the present invention provides an electronic device, comprising a memory and a processor, wherein: The memory is used to store computer programs; When the processor is used to execute the computer program stored in the memory, it implements the above-mentioned clinical trial result prediction method based on hierarchical interactive graph neural network.

[0036] Compared with the prior art, the present invention has the following advantages: This invention constructs an interaction graph by encoding drug information, disease information, and trial inclusion / exclusion criteria into embedding vectors of uniform dimensionality. These embedding vectors serve as input nodes, while also introducing ADMET attribute embedding vectors and disease risk embedding vectors as external knowledge nodes. This processing approach overcomes the limitations of single-modality data and enables the effective integration of multi-source data. Furthermore, because the constructed interaction graph incorporates multimodal information such as drug, disease, trial inclusion / exclusion criteria, ADMET attributes, and disease risk, processing it through an improved graph convolutional network (GCN) can capture the complex hierarchical interactions between nodes. A learnable layer-independent attention matrix dynamically identifies the importance of information transfer between nodes, enabling the model to adapt to the dynamic influence of multiple factors, such as pharmacokinetic (ADMET) attributes, disease risk, and trial design, accurately simulating the dynamic interactions between these cross-modal factors and addressing the issue of missing dynamic interaction modeling. Furthermore, when data from some modalities is missing, other relevant nodes and edges in the interaction graph can still provide information. The improved GCN, through information transfer between nodes, leverages existing data to compensate for and infer the missing modal data to a certain extent. For example, even if the molecular structure of a drug is not disclosed due to commercial confidentiality, the model can still predict clinical trial results by combining disease information, trial inclusion and exclusion criteria, ADMET properties and other relevant information with the node relationships in the interaction graph, thereby solving the problem of missing data and improving the model's adaptability to incomplete data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flowchart of a clinical trial result prediction method based on a hierarchical interactive graph neural network proposed in one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an improved convolutional graph neural network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a clinical trial results system based on a hierarchical interactive graph neural network proposed in one embodiment of the present invention.

[0038] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be the common meanings understood by people with ordinary skills in the field to which the invention belongs. The words "including" and similar words used in this article mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0040] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a clinical trial result prediction method based on a hierarchical interactive graph neural network, the method comprising steps S101 to S105, wherein: Step S101: Acquire clinical trial information of multiple patients, each clinical trial information including drug information, disease information, trial inclusion and exclusion criteria, and trial results, and encode the drug information, disease information, and trial inclusion and exclusion criteria into an embedded vector of uniform dimension; It should be noted that the clinical trial information is historically known information. Each patient has a corresponding clinical trial information. Specifically, the drug information is coded according to the following formula:

[0041] in, is the drug embedding vector, is the total number of drug molecules contained in the mth drug, is the jth drug molecule contained in the mth drug, is the molecular encoding function, for -dimensional real vector.

[0042] Drug molecule embedding is crucial for modeling multimodal associations in clinical trial outcome prediction. However, traditional methods (such as Morgan fingerprints or SMILES encoding) have significant drawbacks: their single-modality nature fails to capture the spatial complexity of molecular graphs (e.g., chemical bond relationships between atoms), leading to loss of efficacy or toxicity features. Furthermore, in multidrug combination scenarios, concatenation or weighting strategies can easily introduce noise or redundancy. To address this, this paper proposes a drug molecule embedding method based on average pooling. Specifically, as shown above, a molecular graph message passing neural network (MPNN) encodes atomic and chemical bond features (three hidden layers, 100 dimensions) and aggregates multidrug embeddings using average pooling to avoid feature shifts caused by quantitative differences and interference from anomalous molecules. This design balances molecular structure modeling with dynamic aggregation of multiple drugs. Experimental results show that its AUC improves by approximately 5% compared to sum pooling, significantly enhancing the ability to characterize pharmacokinetic properties (e.g., absorption rate and toxicity) while ensuring embedding stability in multidrug scenarios. This provides a robust and highly discriminative feature foundation for subsequent cross-modal interactions, ultimately supporting the model's generalization performance in complex prediction tasks.

[0043] In addition, the disease information is encoded according to the following formula:

[0044]

[0045]

[0046] in, is the disease embedding vector, is the graph attention model, For the patient's diseases, is the total number of diseases suffered by the patient, For diseases embedded using GRAM, is the attention weight, is a feed-forward network with a single hidden layer, Represents the set of all ancestor nodes of i. The ancestors of a node represent the higher-level categories of the current node. is the jth randomly initialized learnable embedding vector, is the i-th randomly initialized learnable embedding vector, T is the transpose, is the current disease node itself.

[0047] For example, in ICD 10 codes, D41 (tumor of the urinary organs) and D41.2 (tumor of the ureter) are ancestors of D41.20 (tumor of the right ureter). C34 (malignant neoplasm of bronchus and lung) and C34.9 (malignant neoplasm of bronchus or lung, unspecified) are ancestors of C34.91 (malignant neoplasm of right bronchus or lung, unspecified).

[0048] In clinical trial prediction, disease embedding requires accurate representation of disease semantics and hierarchical associations. However, traditional methods (such as flat text encoding or static ontology encoding) have a key flaw: they cannot dynamically integrate the hierarchical structure of medical ontologies (such as the ICD) (e.g., the relationship between disease parent and child categories). This results in the ineffective utilization of semantic associations between diseases (e.g., the subordination between "lung cancer" and "malignant tumor"), limiting the accuracy of cross-modal interactions. To this end, this paper proposes a disease embedding method based on the Graph Attention Model (GRAM). As shown in the above formula, an embedding is generated by convexly combining the current node and its ancestor nodes in the disease ontology. The calculation (input is the concatenation of the embeddings of the disease and its ancestor nodes) dynamically learns the contributions of key nodes in the hierarchical path (e.g., prioritizing "malignant tumor" over specific subtypes). This method explicitly models the tree structure of the disease ontology, allowing the embedding to retain specific disease characteristics while inheriting high-level semantic commonalities (such as disease risk categories). Experiments have shown that compared with baseline models (such as pure BERT encoding), its AUC in disease risk prediction tasks is improved by 8%, and its interpretability is enhanced (for example, visualizing attention weights can locate high-risk hierarchical nodes), providing a disease representation foundation that is both fine-grained and generalized for subsequent drug-disease interactions.

[0049] In addition, in some embodiments, Clinical-BERT is used to extract sentence-level embeddings of the trial inclusion criteria. Multi-scale one-dimensional convolutions (kernel sizes 1, 3, 5, and 7) are used to capture semantic features of different granularities. Finally, the embedding vectors are generated after concatenation through a fully connected layer. Specifically, the trial inclusion criteria are encoded according to the following formula:

[0050] in, For the test, the standard embedding vector is The inclusion criteria for the test were Encode the function for the eligibility criteria.

[0051] In clinical trial prediction, embedding inclusion and exclusion criteria requires accurate analysis of the semantic logic of complex medical text (e.g., the opposition between inclusion and exclusion criteria). However, traditional methods (such as general-purpose BERT or single-scale convolution) have significant shortcomings: general-purpose models have poor domain adaptability for specialized terms (e.g., "abnormal liver function threshold") and fail to distinguish the semantic differences between inclusion and exclusion criteria, resulting in ineffective modeling of the interactive relationships between key conditions (e.g., "age > 18 years" and "contraindicated during pregnancy"). To address this issue, this paper proposes a dynamic embedding method based on Clinical-BERT and multi-scale convolution. Specifically, as shown in the above formula, sentence-level semantic vectors are first extracted using domain-pretrained Clinical-BERT. Then, four sets of one-dimensional convolutions with different kernel sizes (1, 3, 5, and 7) are applied to the inclusion / exclusion criteria, capturing both short-range rules (e.g., numerical range) and long-range semantics (e.g., complex logical descriptions). Finally, these vectors are concatenated and passed through a fully connected layer to generate a unified embedding. This method significantly improves the ability to parse complex medical texts through multi-granularity feature fusion and conditional type separation. Ablation experiments show that its AUC is 12% higher than that of a single convolution kernel model, and visualization shows that the model can distinguish key conditions (such as the strong contribution of "tumor stage III" to prediction), providing a highly discriminative structured text representation for cross-modal interaction, supporting the model's accurate reasoning under the dynamic influence of inclusion and exclusion criteria.

[0052] Step S102: pre-training the drug embedding vector according to the ADMET attributes, and obtaining the historical disease trial risk of each disease to pre-train the disease embedding vector according to the historical disease trial risk; It is important to note that in clinical trial prediction, ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties are core indicators of drug safety and efficacy. However, traditional approaches (such as independent prediction models or manual feature engineering) have key drawbacks: their disjunctive modeling lacks deep interaction between pharmacokinetic properties and the primary task (e.g., trial outcome prediction), and single-task models struggle to capture cross-attribute correlations (e.g., the effect of toxicity on metabolic rate). To address this, this paper proposes an ADMET pre-training framework based on a multi-layer Highway Network. This framework generates attribute-specific embeddings from drug molecules and outputs binary predictions using a sigmoid algorithm. This design mitigates gradient vanishing through cross-layer connections within the Highway Network while sharing underlying molecular features to capture implicit correlations between attributes (e.g., toxic molecules are often associated with low absorption rates). The pre-trained ADMET embeddings are then injected into the primary model as external knowledge. Experimental results show that this approach improves the prediction AUC of HINT by 10%. Furthermore, the attention mechanism can pinpoint key attributes (e.g., the high contribution of toxicity to failed trials), significantly enhancing the model's ability to quantify drug safety margins and providing high-information prior knowledge for subsequent cross-modal fusion of pharmacokinetic nodes.

[0053] Specifically, in some embodiments, a drug embedding vector is used as input and the prediction score of the ADMET property is used as the target, and a multi-layer high-speed network is used to generate the ADMET property embedding vector:

[0054] The ADMET attribute embedding vectors are then input into a single-layer fully connected neural network, and the output of the fully connected layer is processed by the Sigmoid function to obtain the binary prediction value:

[0055] And set the loss function of the single-layer fully connected neural network to:

[0056] in, Embedding vectors for ADMET properties, including absorption properties , distribution attribute embedding vector , metabolic attribute embedding vector , excretion attribute embedding vector , toxicity attribute embedding vector , For multi-layer high-speed network, is a single-layer fully connected neural network. is the binary prediction value, , is the true binary label, , is the first loss value.

[0057] In addition, the input and output dimensions of multi-layer highway networks are the same. If the input embedding is larger than the desired dimension (for example, multiple embeddings are concatenated as input), the FC layer should be applied first to reduce the input dimension, and then the highway network should be used to maintain the desired size.

[0058] Pre-trained predictive models for absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties are used in drug discovery to gain insights into drug interactions throughout the body. ADMET properties include: Absorption: The absorption model describes how a drug is absorbed by the body and reaches its site of action. Drugs that are poorly absorbed are generally less than ideal.

[0059] Distribution: Drug distribution models measure the ability of a molecule to move through the bloodstream to different parts of the body. Stronger distribution motions are generally preferred.

[0060] Metabolism: The rate at which a drug is metabolized determines the duration of its effectiveness.

[0061] Excretion: The excretion rate of a drug measures how much of the toxic components of the drug are eliminated from the body.

[0062] Toxicity: Drug toxicity measures the damage a drug can affect the body.

[0063] Furthermore, it should be noted that, in addition to drug characteristics, the applicant also considers knowledge extracted from historical trials for the target disease. Disease information comes from multiple sources: 1) disease descriptions and disease ontologies, and 2) historical trial success rates for the disease. Because historical trials have captured the success rates of each disease at different trial stages, the applicant will use these disease-specific success rates as supervision for training the trial risk prediction model. More specifically, trial success statistics are sourced from ClinicalTrials.gov. Disease risk embedding vectors are generated using a two-layer highway network.

[0064] Specifically, in some embodiments, a disease risk embedding vector is generated using a two-layer high-speed network with the disease embedding vector as input and the disease risk score as the target:

[0065] The disease risk embedding vector is input into a single-layer fully connected neural network, and the output of the fully connected layer is processed by the Sigmoid function to obtain the disease risk prediction value:

[0066] The loss function of the single-layer fully connected neural network is set to:

[0067] in, is the disease risk embedding vector, It is a double-layer high-speed network. is the disease risk prediction value, , is a single-layer fully connected neural network. is the second loss value, is a binary label with trial success or failure as a function of disease information, , for example, 0 means the test failed, and 1 means the test succeeded.

[0068] In summary, disease risk pre-training requires integrating the inherent risk of a disease with historical experimental patterns. However, traditional approaches (such as relying solely on disease descriptions or static ontology encodings) are limited. They fail to leverage critical prior knowledge about historical experimental success rates (e.g., the high failure rate of tumors), making it difficult for the model to quantify the potential impact of the disease on experimental outcomes. To address this, this paper proposes a disease risk pre-training framework based on a two-layer high-speed network. As shown above, the input is a disease embedding (integrating ICD hierarchy and description information). Training is supervised by historical experimental success rate labels to generate risk embeddings and predictions. This design dynamically couples disease ontology semantics (e.g., the association between "lung cancer" and its ICD parent category) with cross-stage success rates (e.g., the decay of success rates from stage II to stage III). Experimental results show that injecting the risk embedding into the main model improves the prediction area under the predictive curve by 7%. Furthermore, the attention weights account for the disease risk hierarchy (e.g., the high weight of the "late-stage malignant tumor" node). Ultimately, this model enables accurate identification of high-risk experimental scenarios across disease generalization, providing a quantitative basis for the multi-dimensional interaction between drug, disease, and risk.

[0069] Step S103: constructing an interaction graph using the drug embedding vector, disease embedding vector, and trial inclusion / exclusion criteria embedding vector as input nodes, and the ADMET attribute embedding vector and disease risk embedding vector as external knowledge nodes; It should be noted that the interaction graph constructed in this step is used to connect all input data sources and important factors affecting the success of clinical trials. It consists of four layers of nodes, which are interconnected between two layers. That is, the first layer nodes are input nodes, and the second layer nodes are external knowledge nodes. In addition, the interaction graph also includes aggregation nodes and prediction nodes, that is, the third layer nodes are aggregation nodes, and the fourth layer nodes are prediction nodes. Among them, the aggregation nodes include: 、 and Interaction nodes ,connect 、 、 、 、 Pharmacokinetic nodes , through the disease risk embedding vector Enhanced Interaction Nodes Strong interaction nodes ; The prediction node To connect pharmacokinetic nodes and enhanced interactive nodes Node.

[0070] In some embodiments, the interaction node is obtained according to the following formula:

[0071] The pharmacokinetic node is obtained according to the following formula:

[0072] The strong interaction nodes are obtained according to the following formula:

[0073] The prediction node is obtained according to the following formula:

[0074] in, 、 、 、 Each is a single-layer fully connected network, where 、 The input dimension is 2 * d, and the output dimension is d, The input dimension is 3 * d, and the output dimension is d. The input dimension is 5 * d, and the output dimension is d.

[0075] In summary, while constructing hierarchical interaction graphs for clinical trial prediction aims to model the hierarchical dynamic relationships of multimodal data, traditional approaches (such as single graph structures or linear feature concatenation) suffer from significant shortcomings. These approaches fail to distinguish the interaction patterns between components at different levels, such as drug properties (e.g., ADMETs), disease risk, and trial design (e.g., the indirect impact of pharmacokinetics on efficacy), leading to the ineffective capture of key cross-modal relationships (e.g., the synergistic effects of toxic molecules and high-risk diseases). To address this, the present invention designs a four-layer interaction graph (input nodes, external knowledge nodes, aggregation nodes, and prediction nodes). This graph integrates ADMET properties, disease risk, and multimodal interaction features layer by layer through formulated aggregation rules (e.g., pharmacokinetic nodes), ultimately outputting the results through the prediction nodes. This architecture explicitly distinguishes local attribute interactions from global decision dependencies through a hierarchical information flow (input → knowledge → aggregation → prediction), enabling interpretable modeling of complex factors (e.g., the drug-disease-risk triangle), providing a structured reasoning framework that combines logic and generalization for end-to-end prediction.

[0076] Step S104: using the interaction graph as input and defining the labels of the interaction graph based on the experimental results, enhancing node embedding using the updated rule of the improved graph convolutional network, introducing a learnable layer-independent attention matrix to calculate the importance weights of edges, and adaptively adjusting the transfer weights of information between nodes based on the importance weights; It is important to note that trial embeddings provide an initial representation of different trial components and their interactions through a graph. To further improve prediction, the applicants designed a dynamic attention graph neural network that leverages this interaction graph to model influential trial components and help improve predictions.

[0077] Specifically, define A as the adjacency matrix of the interaction graph and initialize all node embeddings:

[0078] in, is the initial node embedding matrix, K=13, for dimensional real vector space; The node embedding is enhanced according to the following formula:

[0079] in, The updated node embedding matrix for the improved graph convolutional network at layer l, is the bias parameter of the lth layer, is a learnable layer-independent attention matrix, , is the weight matrix of the l-th layer transformation embedding, is, L is the depth of the improved graph convolutional network, is the unit multiplication, is a linear rectification function; A learnable layer-independent attention matrix is ​​constructed according to the following formula:

[0080] in, is the element in the i-th row and j-th column of the matrix V, which is used to represent the importance weight of the edge between node i and node j. is a two-layer fully connected neural network. is the embedding vector of node i, is the embedding vector of node j, , , yes The transpose of the i-th row of yes The transpose of the jth row of For splicing operation.

[0081] It should be noted that there are ReLU and Sigmoid activation functions in the hidden layer and output layer respectively. The matrix V is element-wise multiplied by the adjacency matrix A, so that the information of the edge with higher prediction score will be given higher weight to propagate.

[0082] In summary, dynamic attention graph neural networks (DANs) aim to optimize the interaction weight distribution of cross-modal nodes in clinical trial prediction. However, traditional graph networks (such as GCNs) have a key bottleneck. Their fixed adjacency matrix cannot dynamically adjust the strength of inter-node associations based on context (e.g., the differential impact of drug toxicity on specific disease risks). This results in the homogenization of key paths (e.g., the synergistic effect of highly toxic drugs and low-success-rate diseases), weakening the model's ability to capture complex causal relationships. To this end, this paper proposes a dynamic graph network based on a learnable attention matrix V. This calculates the attention scores of node embeddings and concatenations, achieving dynamic weight distribution (e.g., assigning higher weights to toxicity nodes and disease risk nodes). By combining graph structure priors (A) with dynamic attention (V), this design adaptively focuses on key paths while preserving hierarchical interaction logic, providing a flexible and interpretable inference engine for quantitative analysis and decision support of complex drug-disease-risk relationships.

[0083] Step S105: constructing a loss function of the improved graph convolutional network, and completing the training of the improved graph convolutional network according to the loss function.

[0084] It should be noted that, in some embodiments, if the test result is successful, the label is defined as 1, and if the test result is failed, the label is defined as 0, thereby obtaining a true binary label.

[0085] Then the output of the Lth layer of the improved graph convolutional network is used as the success probability of the experiment:

[0086] in, is the probability of success of the experiment, is the prediction node representation of the Lth layer; The loss function of the improved graph convolutional network is constructed according to the following formula:

[0087] in, is the third loss value, is the true binary label, is the probability of success of the experiment, , .

[0088] It should be pointed out that the first loss value, the second loss value, and the third loss value mentioned above are all minimized to achieve the training of the corresponding model or network.

[0089] In addition, in some embodiments, the improved graph convolutional network includes a filling module, a dynamic attention graph neural network, a knowledge embedding module, an input embedding module, and a hierarchical interaction graph construction module, wherein the input embedding module is used to encode each experimental information, the hierarchical interaction graph construction module is used to construct an interaction graph, the knowledge embedding module is used to pre-train each embedding vector, the dynamic attention graph neural network is used for prediction, and the filling module is used to fill in missing data.

[0090] Specifically, when the drug information is missing, the filling module generates a drug replacement embedding vector based on the disease embedding vector associated with the drug information and the trial inclusion and exclusion criteria embedding vector, and replaces the actual drug embedding vector with the drug replacement embedding vector to construct an interaction graph:

[0091] The loss function of the padding module is constructed according to the following formula, and the padding module is trained to minimize the fourth loss value:

[0092] in, is the drug replacement embedding vector, To fill the network, is the fourth loss value.

[0093] During the training of the filling module, it is necessary to actively delete the drug information in the complete historical test information, and then use the deleted historical test information to train it, and continuously update the filling module by minimizing the fourth loss value.

[0094] Furthermore, after the improved convolutional graph neural network is trained, the test information is obtained. This test information includes at least disease information, test inclusion and exclusion criteria, and may also include drug information. This test information is input into the trained improved convolutional graph neural network to obtain the test success probability.

[0095] In summary, in clinical trial predictions, the missing data filling module needs to solve the problem of missing drug molecular information due to commercial secrets, but traditional methods (such as mean interpolation or direct elimination of missing samples) have serious defects: they ignore the deep association between drugs and diseases, and inclusion and exclusion criteria (such as specific diseases often correspond to specific drug categories), resulting in the filling results deviating from the true distribution or losing key features. To this end, the present invention proposes a neural network filling model based on disease embedding and inclusion and exclusion criteria embedding, generates missing drug embeddings through formulas, and supervises training with mean square error loss, and reconstructs molecular features using cross-modal correlations (such as the implicit association between "liver disease tests" and "metabolic drugs"). This module realizes dynamic feature completion by jointly optimizing the filling network and main model parameters, significantly enhancing the robustness of the model in scenarios with incomplete data, and providing a high-fidelity solution for the common privacy data missing in practical applications.

[0096] Compared with the prior art, the present invention has the following advantages: 1. Modeling the hierarchical interaction between multimodal data through four layers of nodes (input, knowledge, aggregation, prediction), dynamically integrating pharmacokinetic properties and disease risk into the prediction framework for the first time.

[0097] 2. By introducing a learnable attention matrix VV, the information transmission weight between nodes is adaptively adjusted to enhance the model's sensitivity to critical paths (such as drug toxicity).

[0098] 3. By combining medical ontology (ICD), molecular graph structure and natural language processing (Clinical-BERT), and through ADMET and disease risk pre-training, the embedding quality is significantly improved.

[0099] 4. Through the joint optimization of the missing data filling module and the graph neural network, the stability of the model is ensured in scenarios with incomplete data.

[0100] like Figure 3 As shown, an embodiment of the present invention further proposes a clinical trial result prediction system based on a hierarchical interactive graph neural network, the system comprising: A historical trial information acquisition module 10 is configured to acquire clinical trial information of multiple patients, each of which includes drug information, disease information, trial inclusion and exclusion criteria, and trial results, and encode the drug information, disease information, and trial inclusion and exclusion criteria into an embedded vector of uniform dimension; a vector pre-training module 20 for pre-training drug embedding vectors according to ADMET properties and obtaining historical disease trial risks for each disease to pre-train disease embedding vectors according to the historical disease trial risks; An interaction graph construction module 30 is configured to construct an interaction graph using the drug embedding vector, the disease embedding vector, and the trial inclusion / exclusion criteria embedding vector as input nodes, and the ADMET attribute embedding vector and the disease risk embedding vector as external knowledge nodes; a node enhancement module 40 for taking the interaction graph as input, defining labels for the interaction graph based on experimental results, enhancing node embeddings using an improved graph convolutional network update rule, and introducing a learnable layer-independent attention matrix to calculate edge importance weights, so as to adaptively adjust the transfer weight of information between nodes based on the importance weights; The model training module 50 is used to construct a loss function of the improved graph convolutional network and complete the training of the improved graph convolutional network according to the loss function.

[0101] On the other hand, the present invention also proposes a readable storage medium having one or more programs stored thereon, which, when executed by a processor, implements the above-mentioned clinical trial result prediction method based on hierarchical interactive graph neural network.

[0102] On the other hand, the present invention also proposes an electronic device, including a memory and a processor, wherein the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the above-mentioned clinical trial result prediction method based on hierarchical interactive graph neural network.

[0103] Those skilled in the art will appreciate that the logic and / or steps represented in the flowcharts 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 an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device), or in conjunction with such instruction execution system, apparatus, or device. For purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system, apparatus, or device, or in conjunction with such instruction execution system, apparatus, or device.

[0104] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting, or processing it in another suitable manner as necessary, and then storing it in a computer memory.

[0105] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the aforementioned embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or a combination of the following technologies known in the art may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0106] While the embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations of these embodiments are possible. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as set forth in the claims. Furthermore, the invention described herein is susceptible to other embodiments and may be practiced or implemented in a variety of ways.

Claims

1. A clinical trial results prediction method based on a hierarchical interactive graph neural network, characterized in that: The method comprises: Acquire clinical trial information for multiple patients, each clinical trial information including drug information, disease information, trial inclusion and exclusion criteria, and trial results, and encode the drug information, disease information, and trial inclusion and exclusion criteria into an embedding vector of uniform dimension; Pre-training drug embedding vectors based on ADMET properties and obtaining historical disease trial risks for each disease to pre-train disease embedding vectors based on the historical disease trial risks; An interaction graph is constructed using drug embedding vectors, disease embedding vectors, and trial inclusion / exclusion criteria embedding vectors as input nodes, and ADMET attribute embedding vectors and disease risk embedding vectors as external knowledge nodes. The interaction graph is used as input, and labels of the interaction graph are defined based on experimental results. Node embedding is enhanced using an improved graph convolutional network update rule. A learnable layer-independent attention matrix is ​​introduced to calculate edge importance weights, so as to adaptively adjust the information transfer weights between nodes based on the importance weights. Construct a loss function for the improved graph convolutional network, and complete training of the improved graph convolutional network according to the loss function.

2. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 1, characterized in that: The step of encoding drug information, disease information, and trial inclusion / exclusion criteria into an embedding vector of uniform dimension includes: Medication information is encoded according to the following formula: ； in, is the drug embedding vector, is the total number of drug molecules contained in the mth drug, is the jth drug molecule contained in the mth drug, is the molecular encoding function, represents a d-dimensional real vector; Disease information is encoded according to the following formula: ； ； ； in, is the disease embedding vector, is the graph attention model, For the patient's diseases, is the total number of diseases suffered by the patient, For diseases embedded using GRAM, is the attention weight, is a feed-forward network with a single hidden layer, Represents the set of all ancestor nodes of i. The ancestors of a node represent the higher-level categories of the current node. is the jth randomly initialized learnable embedding vector, is the i-th randomly initialized learnable embedding vector, T is the transpose, is the current disease node itself; The trial inclusion and exclusion criteria were coded according to the following formula: ； in, For the test, the standard embedding vector is The inclusion criteria for the test were Encode the function for the eligibility criteria.

3. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 2, characterized in that: The steps of pre-training the drug embedding vector according to the ADMET attributes and obtaining the historical disease trial risk of each disease to pre-train the disease embedding vector according to the historical disease trial risk include: Taking the drug embedding vector as input and the predicted score of ADMET property as the target, a multi-layer high-speed network is used to generate the ADMET property embedding vector: ； The ADMET attribute embedding vectors are input into a single-layer fully connected neural network, and the output of the fully connected layer is processed by the Sigmoid function to obtain the binary prediction value: ； The loss function of the single-layer fully connected neural network is set to: ； in, Embedding vectors for ADMET properties, including absorption properties , distribution attribute embedding vector , metabolic attribute embedding vector , excretion attribute embedding vector , toxicity attribute embedding vector , For multi-layer high-speed network, is a single-layer fully connected neural network. is the binary prediction value, , is the true binary label, , is the first loss value.

4. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 2, characterized in that: The step of pre-training the drug embedding vector according to the ADMET attributes and obtaining the historical disease trial risk of each disease to pre-train the disease embedding vector according to the historical disease trial risk also includes: Taking the disease embedding vector as input and the disease risk score as the target, a two-layer high-speed network is used to generate the disease risk embedding vector: ； The disease risk embedding vector is input into a single-layer fully connected neural network, and the output of the fully connected layer is processed by the Sigmoid function to obtain the disease risk prediction value: ； The loss function of the single-layer fully connected neural network is set to: ； in, is the disease risk embedding vector, It is a double-layer high-speed network. is the disease risk prediction value, , is a single-layer fully connected neural network. is the second loss value, is a binary label with trial success or failure as a function of disease information, .

5. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 3, characterized in that: The step of constructing an interaction graph using the drug embedding vector, the disease embedding vector, and the trial inclusion and exclusion criteria embedding vector as input nodes and the ADMET attribute embedding vector and the disease risk embedding vector as external knowledge nodes includes: The interaction graph also includes an aggregation node and a prediction node; The aggregation node includes: 、 and Interaction nodes ,connect 、 、 、 、 Pharmacokinetic nodes , through the disease risk embedding vector Enhanced Interaction Nodes Strong interaction nodes ; The prediction node To connect pharmacokinetic nodes and enhanced interactive nodes Node.

6. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 5, characterized in that: The steps of taking the interaction graph as input, defining the labels of the interaction graph according to the experimental results, enhancing node embedding using the updated rule of the improved graph convolutional network, introducing a learnable layer-independent attention matrix to calculate the importance weight of the edge, and adaptively adjusting the transfer weight of the information between nodes according to the importance weight include: Define A as the adjacency matrix of the interaction graph and initialize all node embeddings: ； in, is the initial node embedding matrix, K=13, for dimensional real vector space; The node embedding is enhanced according to the following formula: ； in, The updated node embedding matrix for the improved graph convolutional network at layer l, is the bias parameter of the lth layer, is a learnable layer-independent attention matrix, , is the weight matrix of the l-th layer transformation embedding, is, L is the depth of the improved graph convolutional network, is the unit multiplication, is a linear rectification function; A learnable layer-independent attention matrix is ​​constructed according to the following formula: ； in, is the element in the i-th row and j-th column of the matrix V, which is used to represent the importance weight of the edge between node i and node j. is a two-layer fully connected neural network. is the embedding vector of node i, is the embedding vector of node j, , , yes The transpose of the i-th row of yes The transpose of the jth row of For splicing operation.

7. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 6, characterized in that: The steps of constructing the loss function of the improved graph convolutional network and completing the training of the improved graph convolutional network according to the loss function include: If the test result is successful, the label is defined as 1; if the test result is failed, the label is defined as 0; The output of the Lth layer of the improved graph convolutional network is used as the success probability of the experiment: ； in, is the probability of success of the experiment, is the prediction node representation of the Lth layer; The loss function of the improved graph convolutional network is constructed according to the following formula: ； in, is the third loss value, is the true binary label, is the probability of success of the experiment, , .

8. The clinical trial result prediction method based on hierarchical interactive graph neural network according to claim 2, characterized in that: The method further comprises: The improved graph convolutional network also includes a filling module. When the drug information is missing, the filling module generates a drug replacement embedding vector based on the disease embedding vector associated with the drug information and the trial inclusion and exclusion criteria embedding vector, and replaces the actual drug embedding vector with the drug replacement embedding vector to construct an interaction graph: ； The loss function of the padding module is constructed according to the following formula, and the padding module is trained to minimize the fourth loss value: ； in, is the drug replacement embedding vector, To fill the network, is the fourth loss value.

9. A clinical trial results prediction system based on a hierarchical interactive graph neural network, characterized in that: The system comprises: A historical trial information acquisition module is used to obtain clinical trial information of multiple patients, each of which includes drug information, disease information, trial inclusion and exclusion criteria, and trial results, and encode the drug information, disease information, and trial inclusion and exclusion criteria into an embedded vector of uniform dimension; A vector pre-training module, configured to pre-train drug embedding vectors based on ADMET properties and obtain historical disease trial risks for each disease to pre-train disease embedding vectors based on the historical disease trial risks; An interaction graph construction module is used to construct an interaction graph using drug embedding vectors, disease embedding vectors, and trial inclusion / exclusion criteria embedding vectors as input nodes, and ADMET attribute embedding vectors and disease risk embedding vectors as external knowledge nodes; A node enhancement module is configured to take the interaction graph as input, define labels for the interaction graph based on experimental results, enhance node embeddings using an improved graph convolutional network update rule, and introduce a learnable layer-independent attention matrix to calculate edge importance weights, so as to adaptively adjust the transfer weight of information between nodes based on the importance weights; The model training module is used to construct a loss function of the improved graph convolutional network and complete the training of the improved graph convolutional network according to the loss function.

10. A readable storage medium storing one or more programs, which, when executed by a processor, implement the clinical trial result prediction method based on a hierarchical interactive graph neural network as described in any one of claims 1 to 8.

Images