A drug response prediction method based on interpretable multi-relational graph neural network

Abstract

The present invention is applicable to the field of artificial intelligence, specifically a drug response prediction method of an interpretable multi-relational graph neural network, comprising the following steps: constructing a directed heterogeneous graph containing cell line and drug nodes, as well as sensitivity, resistance, drug similarity, and cell similarity relationship types; using the R-GCN model to process the directed heterogeneous graph, updating the node representation through node feature aggregation, generating positive and negative samples for training, and predicting the potential link relationship between cell lines and drugs; combining the original graph with the prediction results to generate a calculation graph, extracting the edge neighborhood graph to be explained, using an optimizable mask and edge type weighting mechanism, and optimizing the mutual information and structural score to obtain an explanation subgraph, and defining the explanation process from the two aspects of mutual information and structural score. The present invention uses the R-GCN model with an attention mechanism to perform link prediction, and uses a mask and adaptive weighting mechanism to obtain an explanation subgraph, which clearly explains the biological mechanism of the prediction results from many aspects.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a drug reaction prediction method that can explain a multi-relationship graph neural network. Background Art

[0002] In the field of precision medicine, the prediction of cancer drug response (CDR) is the key to achieving personalized treatment plans.

[0003] In recent years, graph neural networks (GNNs) have shown potential in CDR prediction tasks in the field of bioinformatics. Although GNN-based prediction models such as MOFGCN and DualGCN have achieved certain results in CDR prediction, they have not paid enough attention to interpretability. Like a "black box" operation, it is difficult for doctors to understand the mechanism of drug action on patients based on their predictions and to formulate accurate and personalized treatment plans. Patients may also question treatment plans and reduce compliance because they do not understand the basis of the predictions. This lack of interpretability has hindered the effective application of the model in clinical practice and limited its ability to provide strong support for precision medicine.

[0004] Therefore, in the prediction of cancer drug response CDR in existing technologies, it is impractical to conduct extensive drug testing on each patient due to practical and economic factors. Therefore, the development of computational models that can accurately predict CDR has become an urgent need in clinical practice. Existing prediction methods have significant defects, including the black box characteristics of the model and insufficient understanding of complex biological network relationships. In addition, existing methods fail to fully utilize the importance of specific types of edges in bioinformatics, and lack clear benchmark GT in actual data sets, making it difficult to evaluate the effectiveness of explanations. Summary of the invention

[0005] The purpose of the embodiments of the present invention is to provide a drug response prediction method that can interpret multi-relationship graph neural networks, aiming to solve the technical problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions.

[0007] The embodiment of the present invention provides a drug response prediction method that can interpret a multi-relationship graph neural network, comprising the following steps:

[0008] Obtain cell line multi-omics data, drug data, and drug response triplet data;

[0009] DNN was used to process cytological multi-omics data to obtain cell line feature representation, and GNN was used to process drug data to obtain drug feature representation. Similarity triplets were generated by setting cosine similarity thresholds to obtain cell line similarity triplets and drug similarity triplets respectively. The half-maximum inhibitory concentration of the drug was converted into a sensitivity and resistance relationship to obtain the drug response relationship.

[0010] The obtained similarity data are integrated with the drug response relationship to construct a directed heterogeneous graph containing cell line and drug nodes, as well as relationship types such as sensitivity, resistance, drug similarity, and cell similarity;

[0011] The R-GCN model is used to process directed heterogeneous graphs, update node representations through node feature aggregation, generate positive and negative samples for training, and predict potential links between cell lines and drugs;

[0012] The original graph and the predicted results are combined to generate a computational graph, and the neighborhood graph of the edge to be explained is extracted. The explained subgraph is obtained by maximizing the mutual information and structural score optimization using the optimizable mask and edge type weighting mechanism, and the explanation process is defined from the two aspects of mutual information and structural score.

[0013] The real benchmark GT is constructed using network relationships and known triples. The accuracy of the explanation process is evaluated by calculating the precision, recall and F1 score. The stability of the explanation result is evaluated by calculating the explanation stability index, and the average evaluation index after five-fold cross-validation is output.

[0014] Furthermore, the step of using DNN to process cytological multi-omics data to obtain cell line feature representation includes: the cell line multi-omics data is , and learn its features through the DNN layer to obtain the cell line representation , where cell num represents the number of cell lines and dim represents the feature dimension;

[0015] The step of using GNN to process drug data to obtain drug feature representation includes: using the GNN layer to extract drug representation from drug data, setting the input representation of the GNN after the drug representation data is processed as , the drug representation is obtained after calculation by the GNN layer , where drug num represents the number of drugs and dim represents the feature dimension;

[0016] The generation process of the cell line similarity triplet includes: calculating the cosine similarity between cell line i and cell line j: ; Among them, H c [i] represents the characterization vector of cell line i; H c [j] represents the characterization vector of cell line j; when simij When ≥0.96, cell lines i and j are considered similar, forming a cell line similarity triplet (i, C sim , j);

[0017] The generation process of the drug similarity triplet includes: calculating the cosine similarity between drug m and drug n: ; Among them, H d [m] represents the characterization vector of drug m, H d [n] represents the characterization vector of drug n; when sim mn When ≥0.85, drug m is considered similar to drug n, forming a drug similarity triplet (m, D sim , n);

[0018] The step of converting the half-maximal inhibitory concentration of the drug into a sensitivity-resistance relationship to obtain a drug-response relationship includes: converting the half-maximal inhibitory concentration IC50 value of the drug into a sensitivity-resistance relationship, and for cell line c and drug d, if the IC50 value of drug d in cell line c is lower than the sensitivity threshold, it is determined to be a sensitive relationship (c, Sen, d); otherwise, it is a resistance relationship (c, Res, d).

[0019] Furthermore, the steps of constructing a directed heterogeneous graph include:

[0020] Cell line characterization c and drug characterization d As a node, the cell line node set V c Each node in V corresponds to a cell line representation, and the drug node set V d Each node in corresponds to a drug representation, and the overall node set V is expressed as: V = V c ∪V d ;

[0021] The cell line similarity triple (i, C sim ,j) in the relationship C sim , drug similarity triples (m, D sim , n) in the relationship D sim And the sensitivity relationship (c, Sen, d) and resistance relationship (c, Res, d) obtained by converting the IC50 value are used as edges to form an edge set E;

[0022] Construct a directed heterogeneous graph G=(V,E,R), where R represents a set of relations of different types, R={Sen, Res, D sim , C sim}, for each cell line node i∈V c and drug node j∈V dIf there is a sensitivity relationship (i, Sen, j), a directed edge from cell line node i to drug node j is added to the graph, indicating that cell line i is sensitive to drug j; if there is a resistance relationship (i, Res, j), a directed edge from cell line node i to drug node j is added, indicating that cell line i is resistant to drug j; for cell line similarity triples (i, C sim , j), add a directed edge from cell line node i to cell line node j, indicating that cell lines i and j have similarities; for drug similarity triples (m, D sim , n), add a directed edge from drug node m to drug node n, indicating that drugs m and drug n have similarities.

[0023] Furthermore, the step of predicting the potential link relationship between the cell line and the drug includes:

[0024] In the R-GCN framework, node parameter update is divided into two steps: First, for each node V i ∈V at level l, in addition to aggregating nodes V from itself and its neighbor nodes connected to it through relations r∈R j In addition to the features of ∈V, the attention mechanism is introduced to calculate the attention weight and the attention coefficient , used to measure the importance of node j to the node through the relationship; set and are the hidden states of node i and node j in layer l, respectively, through a shared linear transformation W a And the activation function LeakyReLU to calculate the attention score :

[0025] ;

[0026] Among them, || represents the splicing operation, and are the hidden states of node i and node j in layer l respectively; The linear transformation weight matrix representing the node’s own characteristics is used to transform the hidden state of node i in layer l Perform linear transformation; Represents the linear transformation weight matrix associated with relation r, used for the hidden state of node j in layer l Perform a linear transformation related to the relation r;

[0027] Then, the attention score is normalized by the Softmax function to obtain the attention coefficient :

[0028] ;

[0029] in, represents the set of neighbor nodes connected to node i through relationship r; represents the attention score between node i and its neighbor node k connected by a relationship;

[0030] Based on the calculated attention coefficient , perform node feature aggregation, and the aggregation operation formula is:

[0031] ;

[0032] Among them, σ is a nonlinear activation function; c i,r is a problem-specific normalization constant; W r is the parameter of relation r; represents the attention coefficient; and are the hidden states of node i and node j in layer l respectively; The linear transformation weight matrix representing the node’s own characteristics is used to transform the hidden state of node i in layer l Perform linear transformation to participate in the calculation of node feature aggregation to consider the influence of node characteristics; Represents the linear transformation weight matrix associated with relation r, which is used to transform the hidden state of node j in layer l Perform linear transformation related to the relationship r so that when aggregating node features, neighbor node features can be processed specifically according to different relationships, thereby better integrating neighbor node information;

[0033] The aggregated information undergoes a nonlinear transformation before being passed to the next layer l+1, and finally the embedded features of the node are obtained;

[0034] The existing triples (i, r, j) in the directed heterogeneous graph are regarded as positive samples, and negative samples are generated by randomly perturbing the subject node i or the object node j in the positive samples;

[0035] Through parameter updating, feature aggregation, and generation and training of positive and negative samples, the R-GCN model is applied to perform link prediction in directed heterogeneous networks to identify potential new link relationships between cell lines and drugs.

[0036] Furthermore, the steps of generating the computation graph and extracting the neighborhood graph include: combining the original directed graph G 0 And the corresponding prediction result Y 0 , generate the computational graph G c , integrating the predicted existence sensitivity relationship into G c , for the edge to be explained, from G c Extract the k-hop neighborhood graph G from n, the k-hop neighborhood graph contains the nodes and edges that are within k-step connection range with the edge to be explained.

[0037] Furthermore, the step of obtaining the explanation subgraph by maximizing mutual information and structural score optimization by utilizing the optimizable mask and edge type weighting mechanism includes: n Apply the optimizable mask and adaptive edge type weighting mechanism to obtain the initial interpretation subgraph G s , which can optimize the mask screening G n The information in the , by adjusting the mask value to determine the nodes and edges to focus on; set the edge type set to R = {r 1 ,r 2 ,...,r m}, m represents the number of edge types, and edge type r in G n The frequency of occurrence in is f r , then the weight of the edge type W r It can be calculated by the following formula: ,f i Represents the edge type set R={r 1 ,r 2 ,...,r m}; by maximizing G s With G 0 The mutual information between them is iteratively optimized. Given the predicted link (i, r, j), the relevant node cluster V of cell line i is determined. i and the node cluster V related to drug j j , where V i Contains cell line nodes that have similar or indirect similarity to i, V j Contains drug nodes that are similar to j; in the graph G composed of these node clusters c For node V i and V j , find the subgraph G that is most critical to the prediction result (i, r, j) s , the mutual information MI calculation formula is: ; Among them, H(Y) is the entropy of the predicted result Y, H(Y|G=G s） is the conditional entropy, indicating that the computational graph is restricted to G s The uncertainty of Y at time , is calculated as: , where P Φ Indicates the prediction result Y in a given subgraph G s The probability distribution under the condition, Φ is a fixed model parameter; In a given subgraph G s The expectation of the predicted result Y under the condition is the expectation of Y under G sThe probability distribution under the condition is weighted averaged; the mask is optimized through multiple iterations to maximize MI and obtain the final explanation subgraph G s .

[0038] Furthermore, the steps of defining the explanation process from the perspectives of mutual information and structural scoring include: in terms of mutual information, by calculating and optimizing mutual information, determining the subgraph structure that is highly correlated with the prediction result from the perspective of information theory; mutual information measures the G s With G 0 The degree of dependence between them, the larger the value, the greater the subgraph G s The stronger the ability to explain the prediction results; In terms of structural scoring: define the structural score function, expressed as:

[0039] ;

[0040] Based on G s , where N is the total number of edges, r is the type of relationship in the subgraph, and N r is the number of nodes participating in the r relationship, W r is the adaptive weight of the relationship type edge, r i is the node participating in the relationship, coefficient Penalty (N) = αN + βN 2 , α and β are positive hyperparameters used to control the size of the graph; the final optimization goal is obtained by synthesizing the structure score and the mutual information formula .

[0041] Furthermore, the step of constructing a true benchmark GT using network relations and known triples includes:

[0042] GT is constructed using resistance edges, sensitivity edges, drug similarity edges, cell line similarity edges and known triplets in directed heterogeneous networks. First, one-hop similar cell line neighbors and one-hop similar drug neighbors are identified. For cell line i, the cosine similarity between it and other cell lines j is calculated. If the similarity meets the set threshold, j is the one-hop similar cell line neighbor of i. Similarly, the one-hop similar neighbors of the drug are determined. Then, the existing sensitive or resistance relationships are searched between these similar drugs and cell lines to construct GT.

[0043] Further, in the step of evaluating the accuracy of the explanation process by calculating precision, recall and F1 score:

[0044] Precision is used to measure the accuracy of the model explanation. It calculates the proportion of correct explanations among the first k explanations given by the model. The calculation formula is: ; Precision@k represents the precision, TP@k represents the number of correct explanations of the model in the first k predictions, and FP@k represents the number of incorrect explanations in the first k predictions;

[0045] The recall rate is used to evaluate the coverage of all correct explanations by the model explanation. It calculates the proportion of the number of correct explanations in the first k explanations to the number of all correct explanations of the model. The calculation formula is: ; TP@k is the number of correct explanations in the first k predictions, FN@N is the number of ground truth cases that are not correctly explained among the total N explanations of the model; Recall@k is the recall rate;

[0046] The F1 score is used to balance the evaluation indicators and is calculated as follows:

[0047] .

[0048] Furthermore, the step of evaluating the stability of the interpretation results includes: introducing an interpretation stability index, evaluating the influence of the model interpretation on the convergence of the prediction model, setting Indicates that when the prediction model is in the training period e k When N E represents the total number of edges to be explained, then the stability is expressed as: .

[0049] Compared with the prior art, the drug response prediction method of the present invention that can explain the multi-relationship graph neural network has the following beneficial effects:

[0050] First, the present invention constructs a directed heterogeneous network, uses the R-GCN model with an attention mechanism to perform link prediction, and uses a mask and adaptive weighting mechanism to obtain an explanation subgraph, which clearly explains the biological mechanism of the prediction results from multiple aspects and enhances the understanding and trust of clinical decision-making;

[0051] Second, the present invention uses network relationships to construct GT and adopts five-fold cross-validation, and quantitatively evaluates the interpretation quality through multiple indicators, overcoming the problem of lack of clear GT in actual data sets. It can better handle high-dimensional complex data, fully consider the significance of biological specific edge types, and show more outstanding performance and practical value in the field of cancer drug response prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention.

[0053] Figure 1 This is an implementation diagram of a drug response prediction method that can interpret a multi-relationship graph neural network according to the present invention;

[0054] Figure 2A logic diagram of a drug reaction prediction method that can explain a multi-relationship graph neural network according to the present invention;

[0055] Figure 3 A schematic diagram of a logical flow of predicting link relationships provided by the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0057] At present, although GNNs-based prediction models such as MOFGCN and DualGCN have achieved certain results in CDR prediction, they have not paid enough attention to interpretability. They operate like "black boxes", and it is difficult for doctors to understand the mechanism of drug action on patients based on their predictions, and it is impossible to formulate accurate and personalized treatment plans. Patients may also question the treatment plan and reduce compliance because they do not understand the basis of the prediction. This lack of interpretability hinders the effective application of the model in clinical practice and limits its ability to provide strong support for precision medicine. In addition, algorithms such as Explaine, GNNExplainer and PGExplainer face the problem of high dimensionality and high complexity of data when processing CDR data. These data integrate information such as cell line multi-omics and drug molecular fingerprints. It is difficult for the algorithm to effectively extract feature information and mine potential relationships from them, resulting in limited accuracy and effectiveness of the interpretation, and unable to provide a reliable basis for the prediction results; in bioinformatics, specific edge types such as sensitivity, resistance, drug similarity and cell similarity between cell lines and drugs are extremely critical to understanding drug response mechanisms. However, when existing explanatory algorithms generate explanatory subgraphs, they often fail to fully consider the importance of these edge types, and only focus on the smallest subgraph that affects the prediction, ignoring the broader meaningful structure, making the explanation lack biological meaning and unable to truly reflect the drug reaction process; actual data sets lack clear GTs, and existing technologies make it difficult to construct GT evaluation methods based on biological knowledge. This makes it impossible to quantitatively measure the effectiveness of explanatory algorithms, and it is difficult for researchers to compare the pros and cons of algorithms, which hinders algorithm improvement. For example, when comparing different algorithms, there is no objective standard, and it is not known which algorithm explanation is more accurate and meaningful, which is not conducive to promoting the development of algorithms in this field.

[0058] The present invention aims to solve the key technical problems of CDR prediction in the field of precision medicine. On the one hand, in response to the problem of insufficient interpretability of existing GNNs-based prediction models, the present invention provides a method that can clearly explain the biological mechanism of the prediction results, so as to enhance doctors' and patients' understanding and trust in model decisions, thereby supporting clinical decision-making and the formulation of personalized treatment plans. On the other hand, considering the shortcomings of existing methods in dealing with high-dimensional complex data sets, the present invention is committed to building more effective models to deeply understand the complex biological relationships in cancer drug responses. In addition, in order to solve the problem of lack of clear GT in actual data sets, the present invention will explore methods for constructing GT and propose evaluation indicators to promote the improvement and development of explanatory algorithms.

[0059] Specifically, the present invention collects cell line multi-omics data, drug SMILES data and drug response triple data from data sources such as the Cancer Cell Line Encyclopedia (CCLE), PubChem, and Genomics of Cancer Drug Sensitivity (GDSCv2), and obtains cell line and drug representations through deep neural network (DNN) layer and graph neural network (GNN) layer processing, generates similarity triples according to specific cosine similarity and converts the drug half-maximum inhibitory concentration (IC50) value into sensitivity and resistance relationship, thereby constructing a directed heterogeneous graph G. Then, the relational graph convolutional network (R-GCN) model with attention mechanism is used for link prediction. The general R-GCN model processes the neighbor node information in a relatively average way, and after adding the attention mechanism, the model can dynamically assign weights according to different relationships and the importance of nodes to the target node, and then aggregate features according to its node update rules and generate positive and negative samples. Then, the original graph and the predicted results are combined to generate a correlation graph, and the mask and adaptive weighting mechanism are applied to optimize the mutual information and structural scores to obtain the explained subgraph. Finally, the GT was constructed using the network relationship, and a five-fold cross-validation was used to evaluate the explanation quality through precision (Precision@k), recall (Recall@k), F1 score (F1@k) and explanation stability indicators to achieve cancer drug response prediction and explanation evaluation.

[0060] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0061] Please refer to Figure 1 and Figure 2 In an embodiment of the present invention, a drug response prediction method that can interpret a multi-relational graph neural network includes the following steps:

[0062] S101, obtaining cell line multi-omics data, drug data, and drug response triplet data;

[0063] Among them, in step S101 of the present invention, multi-omics data of cell lines are collected from CCLE, including genomic data (such as gene sequence, mutation information, etc.), transcriptomic data (gene expression profile) and proteomic data (protein expression and modification information). These data provide multi-dimensional information for in-depth research on cell line characteristics; SMILES data of drugs are obtained from PubChem, and the molecular structure of drugs is described in the form of strings to prepare for subsequent feature extraction based on graph neural networks; drug response triple data are obtained from GDSCv2. This data resource compiles the sensitivity response data of tumor cell lines to various drugs. It is currently the largest public repository of tumor cell drug sensitivity and provides key data support for studying the reaction relationship between drugs and cell lines.

[0064] S102, using DNN to process cytology multi-omics data to obtain cell line feature representation, using GNN to process drug data to obtain drug feature representation, generating similarity triplets by setting cosine similarity thresholds, and obtaining cell line similarity triplets and drug similarity triplets respectively; and converting the half-maximum inhibitory concentration of the drug into a sensitivity and resistance relationship to obtain a drug response relationship;

[0065] S103, integrating the obtained similarity data with the drug response relationship, and constructing a directed heterogeneous graph including cell line and drug nodes, as well as relationship types such as sensitivity, resistance, drug similarity, and cell similarity;

[0066] S104. Use the R-GCN model to process directed heterogeneous graphs, update node representations through node feature aggregation, generate positive and negative samples for training, and predict potential links between cell lines and drugs;

[0067] S105, combining the original graph and the prediction result to generate a computational graph, extracting the edge neighborhood graph to be explained, using the optimizable mask and edge type weighting mechanism, and optimizing the interpretation subgraph by maximizing the mutual information and structural score, and defining the interpretation process from the two aspects of mutual information and structural score;

[0068] S106. Use network relationships and known triples to build a true benchmark GT. Evaluate the accuracy of the interpretation process by calculating precision, recall, and F1 score. Evaluate the stability of the interpretation result by calculating the interpretation stability index. Output the average evaluation index after five-fold cross-validation.

[0069] Furthermore, the step of using DNN to process cytological multi-omics data to obtain cell line feature representation includes: the cell line multi-omics data is , and learn its features through the DNN layer to obtain the cell line representation , where cell num represents the number of cell lines and dim represents the feature dimension;

[0070] The step of using GNN to process drug data to obtain drug feature representation includes: using the GNN layer to extract drug representation from drug data, setting the input representation of the GNN after the drug representation data is processed as , the drug representation is obtained after calculation by the GNN layer , where drug num represents the number of drugs and dim represents the feature dimension;

[0071] The generation process of the cell line similarity triplet includes: calculating the cosine similarity between cell line i and cell line j: ; Among them, H c [i] represents the characterization vector of cell line i; H c [j] represents the characterization vector of cell line j; when sim ij When ≥0.96, cell lines i and j are considered similar, forming a cell line similarity triplet (i, C sim , j);

[0072] The generation process of the drug similarity triplet includes: calculating the cosine similarity between drug m and drug n: ; Among them, H d [m] represents the characterization vector of drug m, H d [n] represents the characterization vector of drug n; when sim mn When ≥0.85, drug m is considered similar to drug n, forming a drug similarity triplet (m, D sim , n);

[0073] The step of converting the half-maximal inhibitory concentration of the drug into a sensitivity-resistance relationship to obtain a drug-response relationship includes: converting the half-maximal inhibitory concentration IC50 value of the drug into a sensitivity-resistance relationship, and for cell line c and drug d, if the IC50 value of drug d in cell line c is lower than the sensitivity threshold, it is determined to be a sensitive relationship (c, Sen, d); otherwise, it is a resistance relationship (c, Res, d).

[0074] Through the above data processing steps, the present invention converts the original collected data into a form suitable for constructing a directed heterogeneous network, including node representation and different types of directed edge relationships, laying a foundation for subsequent link prediction and interpretation work.

[0075] Furthermore, in step S103 of the present invention, the step of constructing a directed heterogeneous graph includes:

[0076] Cell line characterization c and drug characterization d As a node, the cell line node set V cEach node in V corresponds to a cell line representation, and the drug node set V d Each node in corresponds to a drug representation, and the overall node set V is expressed as: V = V c ∪V d ;

[0077] The cell line similarity triple (i, C sim ,j) in the relationship C sim , drug similarity triples (m, D sim , n) in the relationship D sim And the sensitivity relationship (c, Sen, d) and resistance relationship (c, Res, d) obtained by converting the IC50 value are used as edges to form an edge set E;

[0078] Construct a directed heterogeneous graph G=(V,E,R), where R represents a set of relations of different types, R={Sen, Res, D sim , C sim}, for each cell line node i∈V c and drug node j∈V d If there is a sensitivity relationship (i, Sen, j), a directed edge from cell line node i to drug node j is added to the graph, indicating that cell line i is sensitive to drug j; if there is a resistance relationship (i, Res, j), a directed edge from cell line node i to drug node j is added, indicating that cell line i is resistant to drug j; for cell line similarity triples (i, C sim , j), add a directed edge from cell line node i to cell line node j, indicating that cell lines i and j have similarities; for drug similarity triples (m, D sim , n), add a directed edge from drug node m to drug node n, indicating that drugs m and drug n have similarities.

[0079] The present invention integrates cell lines, drugs and various relationships between them into a directed heterogeneous network by constructing a directed heterogeneous graph, providing a structured data basis for subsequent link prediction and interpretation work, making it possible to analyze the complex relationships between cell lines and drugs and potential drug response mechanisms within this network framework.

[0080] Further, such as Figure 2 and Figure 3As shown, in step S104 of the present invention, for the link prediction part, after successfully constructing the directed heterogeneous network, in order to deeply explore the potential relationship between cell lines and drugs and predict new links, we introduce the relational graph convolutional network (R-GCN) combined with the attention mechanism to perform link prediction. This can not only utilize the ability of R-GCN to process multi-relational directed heterogeneous graphs, but also use the attention mechanism to adaptively focus on important information, thereby more accurately obtaining new potential links.

[0081] The R-GCN model was chosen because it has shown effectiveness in dealing with downstream tasks based on multi-relational directed heterogeneous graphs. On this basis, the attention mechanism is integrated to enable the model to dynamically assign weights according to the importance of different relations and nodes to the target node, so as to better capture the key information in the graph structure.

[0082] Specifically, the step of predicting the potential link relationship between the cell line and the drug includes:

[0083] In the R-GCN framework, node parameter update is divided into two steps:

[0084] First, for each node V i ∈V at level l, in addition to aggregating nodes V from itself and its neighbor nodes connected to it through relations r∈R j In addition to the features of ∈V, the attention mechanism is introduced to calculate the attention weight and the attention coefficient , used to measure the importance of node j to the node through the relationship; set and are the hidden states of node i and node j in layer l, respectively, through a shared linear transformation W a And the activation function LeakyReLU to calculate the attention score :

[0085] ;

[0086] Among them, || represents the splicing operation, and are the hidden states of node i and node j in layer l respectively; The linear transformation weight matrix representing the node’s own characteristics is used to transform the hidden state of node i in layer l Perform linear transformation; Represents the linear transformation weight matrix associated with relation r, used for the hidden state of node j in layer l Perform a linear transformation related to the relation r;

[0087] Then, the attention score is normalized by the Softmax function to obtain the attention coefficient :

[0088] ;

[0089] in, represents the set of neighbor nodes connected to node i through relationship r; represents the attention score between node i and its neighbor node k connected by a relationship;

[0090] Based on the calculated attention coefficient , perform node feature aggregation, and the aggregation operation formula is:

[0091] ;

[0092] Among them, σ is a nonlinear activation function; c i,r is a problem-specific normalization constant; W r is the parameter of relation r; represents the attention coefficient; and are the hidden states of node i and node j in layer l respectively; The linear transformation weight matrix representing the node’s own characteristics is used to transform the hidden state of node i in layer l Perform linear transformation to participate in the calculation of node feature aggregation to consider the influence of node characteristics; Represents the linear transformation weight matrix associated with relation r, which is used to transform the hidden state of node j in layer l Perform linear transformation related to the relationship r so that when aggregating node features, neighbor node features can be processed specifically according to different relationships, thereby better integrating neighbor node information;

[0093] Before being passed to the next layer l+1, the aggregated information undergoes a nonlinear transformation to obtain the embedded features of the nodes. This step helps the model learn more complex node representations to adapt to the complex relationship structure in the directed heterogeneous graph.

[0094] The existing triples (i, r, j) in the directed heterogeneous graph are regarded as positive samples, and negative samples are generated by randomly perturbing the subject node i or the object node j in the positive samples. This method of generating positive and negative samples enables the model to learn the ability to distinguish between real links and false links, thereby making more accurate predictions in link prediction tasks. Through parameter updating, feature aggregation, and the generation and training of positive and negative samples, the R-GCN model is used to perform link prediction in directed heterogeneous networks, identifying potential new link relationships between cell lines and drugs, which provides a basis for the subsequent interpretation of drug response prediction results.

[0095] After completing the previous link prediction, in order to deeply understand the basis and internal mechanism of the model's prediction, we need to reasonably explain the prediction results. In step S105 of the present invention, the steps of generating a computational graph and extracting a neighborhood graph include: combining the original directed graph G 0 And the corresponding prediction result Y 0 , generate the computational graph G c , integrating the predicted existence sensitivity relationship into G c , for the edge to be explained, from G c Extract the k-hop neighborhood graph G from n The k-hop neighborhood graph contains the nodes and edges that are within the k-step connection range of the edge to be explained. These elements are considered to be related to the prediction result of the edge to be explained; for example, if k=2, it will include all nodes and edges connecting them that are no more than 2 hops away from the nodes at both ends of the edge to be explained.

[0096] Furthermore, the step of obtaining the explanation subgraph by maximizing mutual information and structural score optimization by utilizing the optimizable mask and edge type weighting mechanism includes: n Apply the optimizable mask and adaptive edge type weighting mechanism to obtain the initial interpretation subgraph G s , which can optimize the mask screening G n The information in the ,is used to determine the nodes and edges to focus on by adjusting the mask value;

[0097] For the adaptive edge type weighting mechanism, we dynamically determine the weights according to the distribution of edge types (e.g., sensitivity, resistance, drug similarity, cell similarity) in the data;

[0098] Set the edge type set to R={r 1 ,r 2 ,...,r m}, m represents the number of edge types, and edge type r in G n The frequency of occurrence in is f r , then the weight of the edge type W r It can be calculated by the following formula: ,f i Represents the edge type set R={r 1 ,r 2 ,...,r m}; thus, in different neighborhood graphs G n In the , the weight of the edge type will be adaptively adjusted according to its actual distribution in the subgraph. For example, if the frequency of occurrence of a sensitive edge in a certain subgraph is high, its weight will be increased accordingly, so that the model pays more attention to such key edge types, thereby preliminarily screening out potentially important subgraph structures;

[0099] By maximizing G sWith G 0 The mutual information between them is iteratively optimized. Given the predicted link (i, r, j), the relevant node cluster V of cell line i is determined. i and the node cluster V related to drug j j , where V i Contains cell line nodes that have similar or indirect similarity to i, V j Contains drug nodes that are similar to j; in the graph G composed of these node clusters c For node V i and V j , find the subgraph G that is most critical to the prediction result (i, r, j) s , the mutual information MI calculation formula is: ; Among them, H(Y) is the entropy of the predicted result Y, H(Y|G=G s） is the conditional entropy, indicating that the computational graph is restricted to G s The uncertainty of Y at time , is calculated as: , where P Φ Indicates the prediction result Y in a given subgraph G s The probability distribution under the condition, Φ is a fixed model parameter; In a given subgraph G s The expectation of the predicted result Y under the condition is the expectation of Y under G s The probability distribution under the condition is weighted averaged; the mask is optimized through multiple iterations to maximize MI and obtain the final explanation subgraph G s .

[0100] Furthermore, the steps of defining the explanation process from the perspectives of mutual information and structural scoring include: in terms of mutual information, by calculating and optimizing mutual information, determining the subgraph structure that is highly correlated with the prediction result from the perspective of information theory; mutual information measures the G s With G 0 The degree of dependence between them, the larger the value, the greater the subgraph G s The stronger the explanatory power of the prediction results; for example, if the MI value is high, it means that G s The combination of nodes and edges in is closely related to the prediction results, thus providing a strong basis for the explanation;

[0101] Structural scoring: Define the structural score function, expressed as:

[0102] ;

[0103] Based on G s , where N is the total number of edges, r is the type of relationship in the subgraph, and N r is the number of nodes participating in the r relationship, W ris the adaptive weight of the relationship type edge, r i is the node participating in the relationship, coefficient Penalty (N) = αN + βN 2 , α and β are positive hyperparameters used to control the size of the graph; the final optimization goal is obtained by synthesizing the structure score and the mutual information formula ;

[0104] This enables the model to take into account both mutual information and subgraph structural characteristics during optimization, select key explanatory subgraphs more accurately, improve explanatory capabilities, and provide more reasonable and biologically meaningful explanations for drug response predictions.

[0105] In order to evaluate the quality of explanation and five-fold cross-validation, after obtaining the model's explanation of the prediction results, evaluating the quality of these explanations becomes a key step. At the same time, the use of five-fold cross-validation can more comprehensively and robustly evaluate the performance of the model on different data subsets, avoiding the deviation of the evaluation results due to the randomness of data division, thereby obtaining more representative and convincing model performance evaluation conclusions.

[0106] Furthermore, the step of constructing a true benchmark GT using network relations and known triples includes:

[0107] GT is constructed using resistance edges, sensitivity edges, drug similarity edges, cell line similarity edges and known triplets in directed heterogeneous networks. First, one-hop similar cell line neighbors and one-hop similar drug neighbors are identified. For cell line i, the cosine similarity between it and other cell lines j is calculated. If the similarity meets the set threshold, j is the one-hop similar cell line neighbor of i. Similarly, the one-hop similar neighbors of the drug are determined. Then, the existing sensitive or resistance relationships are searched between these similar drugs and cell lines to construct GT.

[0108] The final constructed GT can be divided into three cases:

[0109] Case 1: If cell line i is sensitive to drug J, and drug J is similar to drug j, then cell line i is considered to be sensitive to drug j, forming the corresponding GT triplet (i, Sen, j);

[0110] Case 2: If cell line I is sensitive to drug J, and cell line I is similar to cell line i, then cell line i is considered to be sensitive to drug J, and a GT triplet (i, Sen, J) is constructed;

[0111] Case 3: If cell line I is sensitive to drug J, cell line I is similar to cell line i, and drug J is similar to drug j, then cell line i is considered to be sensitive to drug j, and a GT triplet (i, Sen, j) is generated;

[0112] In this way, the relational information in the network is fully utilized to construct GT, providing a reliable reference standard for the subsequent accurate evaluation of the quality of model interpretation. At the same time, in order to perform five-fold cross-validation, the drug response triples and similarity triples are combined and randomly shuffled, and then divided into five subsets on average. Each subset will be used as a test set in turn, and the remaining four subsets will be used as training sets for subsequent model performance evaluation. Throughout the process, the ratio of positive and negative samples in the drug response prediction model in the model is set to 1:2 to ensure the balance and effectiveness of the training data.

[0113] Further, in the step of evaluating the accuracy of the explanation process by calculating precision, recall and F1 score:

[0114] Precision is used to measure the accuracy of the model explanation. It calculates the proportion of correct explanations among the first k explanations given by the model. The calculation formula is: ; Precision@k represents the precision, TP@k represents the number of correct explanations of the model in the first k predictions, and FP@k represents the number of incorrect explanations in the first k predictions;

[0115] The recall rate is used to evaluate the coverage of all correct explanations by the model explanation. It calculates the proportion of the number of correct explanations in the first k explanations to the number of all correct explanations of the model. The calculation formula is:

[0116] ; TP@k is the number of correct explanations in the first k predictions, FN@N is the number of ground truth cases that are not correctly explained among the total N explanations of the model; Recall@k is the recall rate;

[0117] Taking into account the precision and recall rate, the F1 score is used to balance the evaluation indicators. The calculation formula is:

[0118] , which can more comprehensively reflect the quality of model explanation and avoid the one-sided evaluation that may occur when using precision or recall alone;

[0119] Furthermore, the step of evaluating the stability of the interpretation results includes: introducing an interpretation stability index, evaluating the influence of the model interpretation on the convergence of the prediction model, setting Indicates that when the prediction model is in the training period e k When N E represents the total number of edges to be explained, then the stability is expressed as: , which measures the stability of the model by calculating the overlap of explanations provided by the explanatory models at different training epochs.

[0120] Finally, the final evaluation results are output: After completing the five-fold cross-validation, the evaluation indicators obtained in each fold validation process are averaged. By calculating these average indicators, an evaluation result that comprehensively reflects the performance of the model on the entire data set can be obtained, avoiding the randomness and bias that may be caused by a single evaluation.

[0121] In summary, the beneficial effects of the drug response prediction method of the explainable multi-relationship graph neural network of the present invention are as follows: the present invention constructs a directed heterogeneous network, uses the R-GCN model with an attention mechanism to perform link prediction, and uses a mask and adaptive weighting mechanism to obtain an explanation subgraph, which clearly explains the biological mechanism of the prediction result from multiple aspects and enhances the understanding and trust of clinical decision-making; at the same time, the present invention uses network relationships to construct GT and adopts five-fold cross-validation, and quantitatively evaluates the explanation quality through multiple indicators, which overcomes the problem of lack of clear GT in actual data sets, can better process high-dimensional complex data, and fully considers the significance of biological specific edge types, showing more excellent performance and practical value in the field of cancer drug response prediction.

[0122] This invention integrates the attention mechanism into the relational graph convolutional network (R-GCN). Based on R-GCN's processing of multi-relational directed heterogeneous graphs, the attention mechanism uses a unique calculation method to enable the model to dynamically assign weights based on the importance of different relationships and nodes to the target node. From calculating the attention score using shared linear transformation and activation function, to obtaining the attention coefficient through normalization by the Softmax function, and then to aggregating node features based on this, this series of closely connected calculation steps constitutes a complete attention mechanism system. This innovation not only improves the model's ability to capture key information of the graph structure, but also optimizes the information fusion process and significantly enhances the model performance.

[0123] The present invention introduces an adaptive weight mechanism. Different from traditional fixed weights, this mechanism dynamically determines the weights through a specific formula based on the distribution of edge types in the data. Specifically, according to the frequency of occurrence of edge types in a specific subgraph, the corresponding weights are calculated using a formula so that the weights can be adjusted as the local characteristics of the data change. When generating the initial explanatory subgraph, the adaptive weights work together with the optimizable mask to screen and weight different edge types based on the weights, thereby affecting the generation of the subgraph structure and improving the model's ability to explain the prediction results.

[0124] The present invention lays the foundation for building GT by using specific neighbor identification methods, such as identifying one-hop similar cell line neighbors and one-hop similar drug neighbors based on specific similarity calculation and threshold setting. On top of this, according to the carefully designed three-case classification rules, GT triples are constructed based on similar neighbor relationships, so that GT can accurately reflect the complex relationships in the data and provide a reliable reference for model interpretation quality assessment.

[0125] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A drug response prediction method based on an interpretable multi-relational graph neural network, characterized in that: The following steps are involved: Obtain cell line multi-omics data, drug data, and drug response triplet data; DNN was used to process cytological multi-omics data to obtain cell line feature representation, and GNN was used to process drug data to obtain drug feature representation. Similarity triplets were generated by setting cosine similarity thresholds to obtain cell line similarity triplets and drug similarity triplets respectively. The half-maximum inhibitory concentration of the drug was converted into a sensitivity and resistance relationship to obtain the drug response relationship. The obtained similarity data are integrated with the drug response relationship to construct a directed heterogeneous graph containing cell line and drug nodes, as well as sensitivity, resistance, drug similarity, and cell similarity relationship types; The R-GCN model is used to process directed heterogeneous graphs, update node representations through node feature aggregation, generate positive and negative samples for training, and predict potential links between cell lines and drugs; The original graph and the predicted results are combined to generate a computational graph, and the neighborhood graph of the edge to be explained is extracted. The explained subgraph is obtained by maximizing the mutual information and structural score optimization using the optimizable mask and edge type weighting mechanism, and the explanation process is defined from the two aspects of mutual information and structural score. The real benchmark GT is constructed using network relationships and known triples. The accuracy of the explanation process is evaluated by calculating the precision, recall and F1 score. The stability of the explanation result is evaluated by calculating the explanation stability index, and the average evaluation index after five-fold cross-validation is output.

2. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 1, characterized in that: The step of using DNN to process cytological multi-omics data to obtain cell line feature representation includes: the cell line multi-omics data is , and learn its features through the DNN layer to obtain the cell line representation , where cellnum represents the number of cell lines and dim represents the feature dimension; The step of using GNN to process drug data to obtain drug feature representation includes: using the GNN layer to extract drug representation from drug data, setting the input representation of the GNN after the drug representation data is processed as , the drug representation is obtained after calculation by the GNN layer , where drug num represents the number of drugs and dim represents the feature dimension; The generation process of the cell line similarity triplet includes: calculating the cosine similarity between cell line i and cell line j: ; Among them, H c [i] represents the characterization vector of cell line i; H c [j] represents the characterization vector of cell line j; when sim ij When ≥0.96, cell lines i and j are similar, forming a cell line similarity triplet (i, C sim , j); The generation process of the drug similarity triplet includes: calculating the cosine similarity between drug m and drug n: ; Among them, H d [m] represents the characterization vector of drug m, H d [n] represents the characterization vector of drug n; when sim mn When ≥0.85, drug m is similar to drug n, forming a drug similarity triplet (m, D sim , n); The step of converting the half-maximal inhibitory concentration of the drug into a sensitivity-resistance relationship to obtain a drug-response relationship includes: converting the half-maximal inhibitory concentration IC50 value of the drug into a sensitivity-resistance relationship, and for cell line c and drug d, if the IC50 value of drug d in cell line c is lower than the sensitivity threshold, it is determined to be a sensitive relationship (c, Sen, d); otherwise, it is a resistance relationship (c, Res, d).

3. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 2, characterized in that: The steps to construct a directed heterogeneous graph include: Cell line characterization c and drug characterization d As a node, the cell line node set V c Each node in V corresponds to a cell line representation, and the drug node set V d Each node in corresponds to a drug representation, and the overall node set V is expressed as: V = V c ∪V d ; The cell line similarity triple (i, C sim ,j) in the relationship C sim , drug similarity triples (m, D sim , n) in the relationship D sim And the sensitivity relationship (c, Sen, d) and resistance relationship (c, Res, d) obtained by converting the IC50 value are used as edges to form an edge set E; Construct a directed heterogeneous graph G=(V,E,R), where R represents a set of relations of different types, R={Sen, Res, D sim , C sim }, for each cell line node i∈V c and drug node j∈V d If there is a sensitivity relationship (i, Sen, j), a directed edge from cell line node i to drug node j is added to the graph, indicating that cell line i is sensitive to drug j; if there is a resistance relationship (i, Res, j), a directed edge from cell line node i to drug node j is added, indicating that cell line i is resistant to drug j; for cell line similarity triples (i, C sim , j), add a directed edge from cell line node i to cell line node j, indicating that cell lines i and j have similarities; for drug similarity triples (m, D sim , n), add a directed edge from drug node m to drug node n, indicating that drugs m and drug n have similarities.

4. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 3, characterized in that: The step of predicting the potential link relationship between the cell line and the drug comprises: In the R-GCN framework, node parameter update is divided into two steps: First, for each node V i ∈V at level l, in addition to aggregating nodes V from itself and its neighbor nodes connected to it through relations r∈R j In addition to the features of ∈V, the attention mechanism is introduced to calculate the attention weight and the attention coefficient , used to measure the importance of node j to the node through the relationship; set and are the hidden states of node i and node j in layer l, respectively, through a shared linear transformation W a And the activation function LeakyReLU to calculate the attention score : ； Among them, || represents the splicing operation, and are the hidden states of node i and node j in layer l respectively; The linear transformation weight matrix representing the node’s own characteristics is used to transform the hidden state of node i in layer l Perform linear transformation; Represents the linear transformation weight matrix associated with relation r, used for the hidden state of node j in layer l Perform a linear transformation related to the relation r; Then, the attention score is normalized by the Softmax function to obtain the attention coefficient : ； in, represents the set of neighbor nodes connected to node i through relationship r; represents the attention score between node i and its neighbor node k connected by a relationship; Based on the calculated attention coefficient , perform node feature aggregation, and the aggregation operation formula is: ； Among them, σ is a nonlinear activation function; c i,r is a problem-specific normalization constant; W r is the parameter of relation r; represents the attention coefficient; and are the hidden states of node i and node j in layer l respectively; The linear transformation weight matrix representing the node’s own characteristics is used to transform the hidden state of node i in layer l Perform linear transformation; Represents the linear transformation weight matrix associated with relation r, which is used to transform the hidden state of node j in layer l Perform a linear transformation related to the relation r; The aggregated information undergoes a nonlinear transformation before being passed to the next layer l+1, and finally the embedded features of the node are obtained; The existing triples (i, r, j) in the directed heterogeneous graph are regarded as positive samples, and negative samples are generated by randomly perturbing the subject node i or the object node j in the positive samples; Through parameter updating, feature aggregation, and generation and training of positive and negative samples, the R-GCN model is applied to perform link prediction in directed heterogeneous networks to identify potential new link relationships between cell lines and drugs.

5. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 4, characterized in that: The steps of generating a computation graph and extracting a neighborhood graph include: combining the original directed graph G0 with the corresponding prediction result Y0 to generate a computation graph G c , integrating the predicted existence sensitivity relationship into G c , for the edge to be explained, from G c Extract the k-hop neighborhood graph G from n , the k-hop neighborhood graph contains the nodes and edges that are within k-step connection range with the edge to be explained.

6. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 5, characterized in that: The step of obtaining the explanation subgraph by maximizing mutual information and structural score optimization by utilizing the optimizable mask and edge type weighting mechanism includes: For G n Apply the optimizable mask and adaptive edge type weighting mechanism to obtain the initial interpretation subgraph G s , which can optimize the mask screening G n The information in the ,is used to determine the nodes and edges to focus on by adjusting the mask value; Set the edge type set to R={r1,r2,...,r m }, m represents the number of edge types, and edge type r in G n The frequency of occurrence in is f r , then the weight of the edge type W r It can be calculated by the following formula: ； Among them, f i Represents the edge type set R={r1,r2,...,r m }The frequency of occurrence of the i-th edge type in; By maximizing G s Iteratively optimize the mutual information mask with G0, and determine the relevant node cluster V of cell line i given the predicted link (i, r, j). i and the node cluster V related to drug j j , where V i Contains cell line nodes that have similar or indirect similarity to i, V j Contains drug nodes that are similar to j; In the graph G composed of node clusters c For node V i and V j , find the subgraph G that is most critical to the prediction result (i, r, j) s , the mutual information MI calculation formula is: ； Among them, H(Y) is the entropy of the predicted result Y, H(Y|G=G s） is the conditional entropy, indicating that the computational graph is restricted to G s The uncertainty of Y at time , is calculated as: , where P Φ Indicates the prediction result Y in a given subgraph G s The probability distribution under the condition, Φ is a fixed model parameter; In a given subgraph G s The expectation of the predicted result Y under the condition is the expectation of Y under G s The probability distribution under the conditions is weighted averaged; By optimizing the mask through multiple iterations, MI is maximized to obtain the final explanation subgraph G s .

7. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 6, characterized in that: The steps of defining the explanation process from the aspects of mutual information and structural scoring include: In terms of mutual information, by calculating and optimizing mutual information, the subgraph structure that is highly correlated with the prediction results is determined from the perspective of information theory; mutual information measures the s The degree of dependence between G0 and G0, the larger the value, the smaller the subgraph G s The stronger the ability to explain the prediction results; Structural scoring: Define the structural score function, expressed as: ; Based on G s , where N is the total number of edges, r is the type of relationship in the subgraph, and N r is the number of nodes participating in the r relationship, W r is the adaptive weight of the relationship type edge, r i is the node participating in the relationship, coefficient Penalty (N) = αN + βN 2 , α and β are positive hyperparameters used to control the size of the graph; the final optimization goal is obtained by synthesizing the structure score and the mutual information formula .

8. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 7, characterized in that: The step of constructing a true benchmark GT using network relations and known triples includes: GT is constructed using resistance edges, sensitivity edges, drug similarity edges, cell line similarity edges and known triplets in directed heterogeneous networks. First, one-hop similar cell line neighbors and one-hop similar drug neighbors are identified. For cell line i, the cosine similarity between it and other cell lines j is calculated. If the similarity meets the set threshold, j is the one-hop similar cell line neighbor of i. Similarly, the one-hop similar neighbors of the drug are determined. Then, the existing sensitive or resistance relationships are searched between these similar drugs and cell lines to construct GT.

9. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 8, characterized in that: In the step of evaluating the accuracy of the explanation process by calculating precision, recall, and F1 score: Precision is used to measure the accuracy of the model explanation. It calculates the proportion of correct explanations among the first k explanations given by the model. The calculation formula is: ； Precision@k represents the precision, TP@k represents the number of correct explanations of the model in the first k predictions, and FP@k represents the number of incorrect explanations of the first k predictions; The recall rate is used to evaluate the coverage of all correct explanations by the model explanation. It calculates the proportion of the number of correct explanations in the first k explanations to the number of all correct explanations of the model. The calculation formula is: ； TP@k is the number of correct explanations among the first k predictions, FN@N is the number of ground truth cases that are not correctly explained among the total N explanations of the model; Recall@k is the recall rate; The F1 score is used to balance the evaluation indicators and is calculated as follows: 。 10. The drug response prediction method based on an interpretable multi-relational graph neural network according to claim 9, characterized in that: The step of evaluating the stability of the interpretation results comprises: Introducing an explanation stability index to evaluate the impact of model explanation on the convergence of the prediction model; Stability is expressed as: ； in, Indicates that when the prediction model is in the training period e k When N E Indicates the total number of edges to be explained.

Images