A knowledge graph-based method for predicting cancer drug response

Abstract

This invention discloses a knowledge graph-based method for predicting cancer drug response, belonging to the field of bioinformatics. To address the problems of low accuracy and lack of interpretability in existing cancer drug response prediction methods, this invention constructs a knowledge graph based on heterogeneous data on the relationships between cell lines, drugs, tissues, and proteins. Multiple sub-knowledge graphs are then generated using meta-paths. A hierarchical propagation strategy based on multi-level attention is employed to extract and fuse high-order neighbor information on the sub-knowledge graphs to update the embedding representations of drugs and cell lines. Finally, after introducing drug-cell line similarity information through contrastive learning, the learned cell line-drug embeddings are used to predict the sensitivity of cancer cell lines to drugs. This invention is used for cancer drug response prediction.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method for predicting cancer drug response. Background Technology

[0002] Cancer is one of the world's major health challenges, posing a serious threat to human health. The heterogeneity of tumors means that the same drug may have varying effects on different patients. Designing personalized treatment plans based on a patient's genomic data can improve treatment outcomes and reduce costs. However, ethical and operational limitations hinder widespread clinical drug response testing. Therefore, Cancer Drug Response (CDR) prediction has become a crucial element in achieving precision medicine.

[0003] In recent years, researchers have proposed various CDR prediction methods, among which graph neural network-based methods have shown good performance. For example, the DeepCDR method uses a Uniform Graph Convolution Neural Network (UGCN) to extract drug features and a Convolutional Neural Network (CNN) to combine and synthesize cell line omics data for CDR prediction. DRPreter divides a pre-constructed cell line graph into multiple subgraphs based on biological pathways, and then uses a Graph Neural Network (GCN) method to learn cell line and drug information for CDR prediction. GraphCDR utilizes GCN to aggregate features of neighboring nodes on the cell line-drug heterogeneous graph and introduces a contrastive learning task as a regularizer, improving the generalization ability of the CDR prediction model.

[0004] However, these CDR prediction methods have some shortcomings. For example, graph neural network-based methods often ignore key biomarkers, tissue types, and biological relationships between cancer cell lines and drugs, limiting the understanding of complex biological semantics and thus the interpretability of the model. Furthermore, GCN-based methods do not fully consider the differences in nodes and edges when extracting information from relevant heterogeneous graphs, thus limiting the accuracy of the model. Finally, graph neural network-based methods ignore the similarity between drug cell line pairs (DCPs). Summary of the Invention

[0005] The purpose of this invention is to address the problems of low accuracy and lack of interpretability in predicting the response to existing cancer drugs.

[0006] A knowledge graph-based method for predicting cancer drug response includes:

[0007] S100: Acquire cell line-drug response data, drug-protein association data, cell line-tissue association data, protein-protein association data, drug molecular fingerprint data, and cell line gene mutation (MUT), copy number variation (CNV), and gene expression (EXP) omics data.

[0008] S200. Calculate the Jaccard similarity of drug molecule fingerprints. The Jaccard similarity constitutes a drug similarity network.

[0009] The correlation measure between EXP, MUT and CNV data of cell lines is calculated, and the correlation measure between EXP, MUT and CNV data is fused to obtain the similarity between cell lines. The similarity between cell lines constitutes a cell line similarity network.

[0010] The drug similarity network and cell line similarity network were reconstructed using KNN, and the reconstructed drug similarity network and cell line similarity network were used as similarity information between cell lines and drugs.

[0011] S300. Using the obtained cell line-drug response data, drug-protein association data, cell line-protein association data, cell line-tissue association data, protein-protein association data, cell line similarity information, and drug similarity information, construct a knowledge graph (KG).

[0012] Based on cell line-drug response data, drug-protein association data, cell line-protein association data, cell line-tissue association data, protein-protein association data, cell line similarity information, and drug similarity information, multiple relationship chains between cells, drugs, proteins, and tissues are identified as meta-paths. The knowledge graph KG is sampled using meta-paths, and each meta-path yields a corresponding sub-knowledge graph.

[0013] S400. Initialize the embedding of entities and relations in each sub-knowledge graph; obtain multi-layer neighbor information of cell lines or drugs through iterative hierarchical propagation in each sub-knowledge graph, and use multi-head attention mechanism to assign weights to different types of entities and relations in each layer, thereby aggregating the neighbor entity information of each layer in the sub-knowledge graph.

[0014] In each sub-knowledge graph, the cell line or drug entity embeddings of each layer are aggregated using a hierarchical attention mechanism. Then, the aggregated cell line and drug embedding representations are concatenated to obtain the cell line and drug embedding representations under each sub-knowledge graph.

[0015] Finally, a meta-path-level attention mechanism is used to weighted aggregate the cell line drug response relationship embeddings under each sub-knowledge graph, resulting in the final cell line drug response relationship embedding e. c,d :

[0016] S500, DCP based on cell line drug response, also known as e c,d The probability of drug sensitivity in cell lines was predicted using machine learning.

[0017] Furthermore, the calculation of Jaccard similarity of drug molecule fingerprints is used to construct drug similarity:

[0018]

[0019] in, and These are the molecular fingerprints of drug d1 and drug d2, respectively.

[0020] Furthermore, the process of calculating the correlation measure between EXP, MUT, and CNV data of cell lines and fusing the correlation measure between EXP, MUT, and CNV data to obtain the similarity between cell lines includes:

[0021] The correlation metrics among EXP, MUT, and CNV data of cell lines are calculated as follows:

[0022]

[0023] in, and ρ represents the gene expression values ​​of gene g1 and gene g2, respectively, and ρ is the standardization function.

[0024]

[0025] in, and These are the copy number variation values ​​for genes g1 and g2, respectively.

[0026]

[0027] in, and ρ represents the binary gene mutation values ​​of genes g1 and g2, respectively, and ρ is the standardization function.

[0028] The similarities between the cell lines obtained by fusion are as follows:

[0029]

[0030] Among them, α1 and α2 are adjustable hyperparameters used to balance the weights of the three omics similarities.

[0031] Furthermore, the drug similarity network and cell line similarity network reconstructed using KNN are represented as follows:

[0032]

[0033] Among them, S drug N represents the reconstructed drug similarity network. t Let S represent the set of neighboring nodes of drug t, and u represent a node in the set of neighboring nodes of drug t; cell N represents the reconstructed cell line similarity network. a Let b represent the set of neighboring nodes of cell line a, and let b represent the nodes in the set of neighboring nodes of cell line a.

[0034] Furthermore, the process of identifying multiple relationships between cell, drug representation, protein, and tissue expression as metapaths in S300 includes:

[0035] Cell lines are represented as Cell, drugs as Drug, proteins as Protein, and tissues as Tissue. Using the obtained cell line-drug response data, drug-protein association data, cell line-protein association data, cell line-tissue association data, protein-protein association data, cell line similarity information, and drug similarity information, denoted as relations r1-relation r7 respectively, meta-paths related to cancer drug response prediction are determined, including meta-paths M1-M5:

[0036]

[0037] Metapaths are used to generate corresponding sub-knowledge graphs on a knowledge graph.

[0038] Furthermore, S400 utilizes a multi-head attention mechanism to assign weights to different types of entities and relationships in each layer, thereby aggregating neighbor entity information from each layer of the sub-knowledge graph. This process includes:

[0039] A multi-head attention mechanism is employed to integrate the neighbor information of entity h by processing multiple information subspaces in parallel.

[0040]

[0041] Among them, W O It is a weight matrix that maps the connection headers back to the original embedding space. Embed the entity as a neighboring entity in the i-th sub-knowledge graph at the l-th layer; head j It is the neighbor entity aggregation embedding under the j-th head in multi-head attention, e t Embedding of the tail entity; Let α be the set of neighbors of entity h under the i-th sub-knowledge graph; j(h,r,t) represents the weight score of the triple (h,r,t) under the j-th attention head in the subgraph. It controls how much information propagates from t to h among the j attention heads, and is expressed as:

[0042]

[0043] α j (h,r,t) is α j ′(h,r,t) is obtained by normalization using the softmax function;

[0044]

[0045] Where W1, W2, and W3 are weight matrices, b1, b2, and b3 are bias terms, || represents the concatenation operation, and ReLU is the non-linear activation function. and Let h, r, and t be the embedding representations of entity h, relation r, and entity t in the l-th layer of the i-th sub-knowledge graph, respectively. Use the LeakyReLU activation function;

[0046] The neighboring entities of entity h from the previous step are embedded in the l-th layer of each sub-knowledge graph. The embedding representation of entity h at layer l is obtained by aggregating the entity's embedding representation at layer l-1 using the Bi-Interaction method:

[0047]

[0048] Where ⊙ represents the dot product operation, W4 and W5 are weight matrices, and LeakyReLU is the activation function. and They represent the embedding of entity h in the (l-1)th layer of the i-th sub-knowledge graph and the embedding of neighboring entities, respectively.

[0049] Furthermore, in S400, the process of aggregating the cell line or drug entity embeddings of each layer in each sub-knowledge graph using a hierarchical attention mechanism, and then concatenating the aggregated cell line and drug embedding representations to obtain the cell line drug embedding representation under each sub-knowledge graph includes:

[0050] In each sub-knowledge graph, cell line or drug entities at each layer are embedded and aggregated using a hierarchical attention mechanism:

[0051]

[0052] in, It is a transpose operation, σ(·) is the tanh activation function, and q (i,l) w (i,l) and b(i,l) Let ω be the attention vector, weight matrix, and bias vector of the embedding representation of the i-th sub-knowledge graph at layer l. (i,l) Let H2 be the weights before softmax normalization of the embeddings of the l-th layer of the i sub-knowledge graphs, H2 be the number of heads in the hierarchical multi-head attention mechanism, and L be the maximum depth of hierarchical propagation; e (i,l) Indicates the cell line or drug corresponding to e (i,0) and e (i,L) Embedding representations of the 0th and Lth layers of cell lines or drug entities, γ (i,0) and γ (i,L) The weights representing the embedding representations of cell lines or drug entities at layers 0 and L;

[0053] The cell lines and drug embeddings obtained from each sub-knowledge graph are spliced ​​together separately:

[0054]

[0055] Here, || represents the concatenation operation. and These are the embedding representation vectors for each layer of cell line and drug aggregation, respectively, which are the e-vectors corresponding to each cell line and drug. i ; The embedding of the DCP obtained in the i-th sub-knowledge graph.

[0056] Furthermore, the final cell line drug response relationship embedding representation in S400 in, and β i Let and be the cell line drug response relationship embeddings and corresponding weights generated in the i-th sub-knowledge graph, respectively; H3 is the number of heads in the meta-path level multi-head attention mechanism; LeakyReLU is the non-linear activation function; m is the total number of meta-paths; e c,d This is for the final cell line drug embedding.

[0057] Furthermore, in S500, before using machine learning to predict the probability of drug sensitivity in cell lines, it is necessary to first use contrastive learning to analyze the e... c,d Optimize, for the optimized e c,d Predict using machine learning; utilize contrastive learning to analyze e. c,d The optimization process includes:

[0058] Based on all obtained cell line drug response relationships (DCP), i.e., e c,dA contrastive learning model is used for optimization. Training samples are divided into positive and negative groups based on the similarity between DCP vectors. The optimization objective is to maximize the similarity of positive sample DCPs and minimize the similarity of negative sample DCPs. The loss function for contrastive learning is expressed as:

[0059]

[0060] Among them, e c,d It is the DCP embedding representation obtained from S4004. This is the DCP embedding representation in the positive sample pair. It is the DCP embedding representation in negative sample pairs. For the embedding transpose operation, τ is an adjustable temperature coefficient, K is the number of negative samples, E is the expected value, and * is used to calculate the cosine similarity between two DCP embedding representations.

[0061] Furthermore, in the process of predicting the probability of drug sensitivity in cell lines using machine learning, S500 utilizes MLP and the sigmoid function to predict the drug sensitivity of cancer cell lines:

[0062]

[0063] In this context, MLP stands for Multilayer Perceptron, and Sigmoid is the activation function. To predict the probability of drug sensitivity in the obtained cell lines.

[0064] The beneficial effects of this invention are:

[0065] This invention first integrates diverse biological data to construct a heterogeneous dataset encompassing associations between cell lines, drugs, tissues, and proteins. The knowledge graph (KG) built upon this dataset effectively alleviates the data sparsity problem. Secondly, based on biological knowledge, we define multiple meta-paths closely related to cancer drug response to perform multi-faceted analysis of the KG. Next, this application applies a hierarchical propagation strategy based on a multi-level attention mechanism to each sub-knowledge graph, effectively aggregating high-order neighborhood information of drugs or cell lines. Furthermore, this invention employs a multi-head self-attention mechanism at the meta-path level to adaptively assign weights to the cell line-drug pair embeddings learned by each meta-path sub-knowledge graph, improving the model's interpretability. Finally, this invention further introduces DCP similarity information through a self-supervised learning-based contrastive learning module, enhancing the model's generalization ability. This invention can effectively improve the accuracy of cancer drug response prediction. It outperforms existing state-of-the-art CDR prediction methods under various experimental conditions, demonstrating its leading predictive performance and practical value. Attached Figure Description

[0066] Figure 1 This is a flowchart for predicting cancer drug response based on knowledge graphs.

[0067] Figures 2-4 ROC curves for five-fold cross-validation experiments (hot start, cell line cold start, drug cold start) on the constructed dataset and other methods.

[0068] Figures 5-7 To generate PR curves for three types of five-fold cross-validation experiments (hot start, cell line cold start, drug cold start) on the constructed dataset and other methods.

[0069] Figure 8 To predict performance graphs for independent testing experiments with other methods on the constructed dataset.

[0070] Figure 9 This is a graph showing the impact of different embedding dimensions on prediction performance.

[0071] Figures 10-15 The graphs are divided into AUC, AUPR, F1, ACC, Recall, and Specificity plots, which show the effect of different numbers of entity-level attention heads and meta-path-level attention heads on the model's prediction performance on the constructed dataset.

[0072] Figure 16 This is a diagram showing the results of tissue-specific experiments.

[0073] Figure 17 This is a plot showing the tissue distribution and proportion of cell lines in the independent test set.

[0074] Figure 18 This is a graph showing the dynamic weight changes of each metapath under the Warm start experimental setup.

[0075] Figure 19 This is a graph showing the dynamic weight changes of each metapath under the Cell line cold start experimental setup.

[0076] Figure 20 This is a graph showing the changes in the weights of each metapath during the training phase of five-fold cross-validation for drug cold start. Detailed Implementation

[0077] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0078] This embodiment is a cancer drug response prediction method based on knowledge graphs, including:

[0079] Step A, Data Acquisition:

[0080] Step a1: Obtain drug and cell line response data, i.e., cell line-drug response data, from the Genomics of Drug Sensitivity in Cancer (GDSC) database.

[0081] Step a2: Collect drug target protein data, i.e. drug-protein association data, from the GDSC, Cancer Cell Line Encyclopedia (CCLE), and DrugBank databases.

[0082] Step a3: Select the top 200 key genes in each cell line given in CCLE, i.e., cell line-protein association data.

[0083] Step a4: Collect the tissue information of the above cell lines in the CCLE database, i.e., cell line-tissue association data;

[0084] Step a5: Collect the above genes and input them into the STRING database. Filter for protein-protein interaction data with a confidence level greater than 0.7, i.e., PPI (Protein Protein Interaction) data. PPI data is used as protein-protein association data.

[0085] Step a6: Collect the drug identification number in GDSC, then obtain the Simplified Molecular InputLine Entry System (SMILES) data of the corresponding drug from the public drug database PubChem based on the drug identification number, and then convert the SMILES into the form of molecular fingerprints (Morgan Fingerprints).

[0086] Step a7: Collect gene mutation (MUT), copy number variation (CNV), and gene expression (EXP) omics data for each cell line in the CCLE database;

[0087] Step B, Similarity Calculation:

[0088] Step b1: Calculate the Jaccard similarity of drug molecule fingerprints, and the Jaccard similarity constitutes a drug similarity network;

[0089] The calculation of Jaccard similarity of drug molecule fingerprints is used to construct drug similarity:

[0090]

[0091] in, and These are the molecular fingerprints of drug d1 and drug d2, respectively.

[0092] Step b2: Calculate the correlation measure between EXP, MUT and CNV data of cell lines, fuse the correlation measures between EXP, MUT and CNV data to obtain the similarity between cell lines, and the similarity between cell lines constitutes the cell line similarity network.

[0093] The correlation metrics between the EXP, MUT, and CNV data of the cell lines calculated in step b2 are as follows:

[0094]

[0095] in, and ρ represents the gene expression values ​​of genes g1 and g2, respectively, and ρ is the standardization function.

[0096]

[0097] in, and These are the copy number variation values ​​of genes g1 and g2, respectively.

[0098]

[0099] in, and ρ represents the binary gene mutation values ​​of genes g1 and g2, respectively, and ρ is the standardization function.

[0100] The similarities between cell lines are as follows:

[0101]

[0102] Among them, α1 and α2 are adjustable hyperparameters used to balance the weights of the three omics similarities.

[0103] Step b3: Reconstruct the drug similarity network and cell line similarity network using KNN, and use the reconstructed drug similarity network and cell line similarity network as the similarity information between cell lines and drugs;

[0104] The drug similarity network and cell line similarity network reconstructed using KNN are represented as follows:

[0105]

[0106] Among them, S drugN represents the reconstructed drug similarity network. t Let S represent the set of neighboring nodes of drug t, and u represent a node in the set of neighboring nodes of drug t; cell N represents the reconstructed cell line similarity network. a Let b represent the set of neighboring nodes of cell line a, and let b represent the nodes in the set of neighboring nodes of cell line a.

[0107] Step C: Knowledge Graph and Meta-Path Construction:

[0108] Step c1: Using the obtained cell line-drug response (relationship r1), drug-protein association (relationship r2), cell line-protein association (relationship r3), cell line-tissue association (relationship r4), protein-protein association (relationship r5), cell line similarity information (relationship r6), and drug similarity information (relationship r7), construct a knowledge graph (KG);

[0109] Step c2: Represent cell lines as Cell, drugs as Drug, proteins as Protein, and tissues as Tissue, and define meta-pathways related to cancer drug response prediction, including meta-pathways M1-M5:

[0110]

[0111] Metapath M1 indicates that a drug affects the expression or activity of related proteins by acting on specific cell lines, which may ultimately affect the effects of the same drug or other drugs, thus contributing to cancer drug response prediction research.

[0112]

[0113] The metapath M2 reveals how the tissue environment from which a cell line originates can influence drug response, helping us understand and predict the sensitivity of cell lines from different tissue sources to drugs. This is of great significance for guiding personalized treatment selection.

[0114]

[0115] Metapath M3 indicates that drugs interact with target proteins, thereby influencing related PPI networks and ultimately acting on cell lines. This helps to elucidate the mechanism of drug action and predict the effects of drugs on different cell lines.

[0116]

[0117] Metapath M4 reveals how similar cell lines influence the effect of drugs on specific proteins, which is important when comparing differences in the response of different cell lines to the same drug.

[0118]

[0119] This metapath M5 explores how drug similarity affects proteins through its influence on cell lines, providing guidance for predicting drug responses.

[0120] Metapaths are used to generate corresponding sub-knowledge graphs on a knowledge graph.

[0121] Step c3: Use the meta-paths defined in c2 to obtain sub-knowledge graphs from the knowledge graph constructed in c1;

[0122] In this process, meta-paths are used to sample the knowledge graph, and each meta-path yields its corresponding sub-knowledge graph as follows:

[0123]

[0124] in, For the metapath M i The generated sub-knowledge graph, where m is the number of meta-paths.

[0125] Step D: Hierarchical Propagation Based on Multi-Level Attention Mechanisms:

[0126] Step d1: Initialize the embedding of entities and relations in each sub-knowledge graph;

[0127] The specific steps for initializing the entity and relation embeddings in each sub-knowledge graph in step d1 are as follows:

[0128] The nn.Embedding method is used to generate initial embeddings for cell lines, drugs, proteins, tissues, and relations r1-r6; specifically, this method accepts the indices of entities and relations as input and constructs two matrices. and Let N be the embedding matrix for entities and the embedding matrix for relations, respectively. e N represents the number of entities. r d represents the number of relations. e and d r These represent the dimensions of entity and relation embeddings, respectively. More specifically, each row in matrices E and R represents the embedding representation of an entity and a relation, respectively. For a given entity h and relation r, their embedding representations are eh and rr, respectively. h and e r .

[0129] Step d2: In each sub-knowledge graph, obtain multi-layer neighbor information of cell lines or drugs through iterative hierarchical propagation, and use multi-head attention mechanism to assign weights to different types of entities and relationships in each layer, thereby aggregating the neighbor entity information of each layer in the sub-knowledge graph;

[0130] In step d2, the information about the cell line or drug at each level obtained through hierarchical propagation in each sub-knowledge graph is represented as follows:

[0131]

[0132] Where l is the depth of the meta-path propagation. For metapath M i The generated sub-knowledge graph is defined by (h, r, t) as a set of triples, where h is the head entity, t is the tail entity, r represents the relationship between entities h and t, m is the number of meta-paths, and L is the maximum propagation depth. Let h be the set of neighboring entities of entity h at the l-th layer of the i-th sub-knowledge graph.

[0133] In step d2, the multi-head attention mechanism is used to assign weights to different types of entities and relations in each layer, thereby aggregating the neighbor entity information of each layer in the sub-knowledge graph. The specific steps are as follows:

[0134] Step d21: Employ a multi-head attention mechanism to integrate the neighbor information of entity h by processing multiple information subspaces in parallel.

[0135]

[0136] Among them, W O It is a weight matrix that maps the connection headers back to the original embedding space. Embed the entity as a neighboring entity in the i-th sub-knowledge graph at the l-th layer; head j It is the neighbor entity aggregation embedding under the j-th head in multi-head attention, e t Embedding of the tail entity; Let α be the set of neighbors of entity h under the i-th sub-knowledge graph; j (h,r,t) represents the weight score of the triple (h,r,t) under the j-th attention head in the subgraph. It controls how much information propagates from t to h among the j attention heads, and is expressed as:

[0137]

[0138] α j (h,r,t) is α j ′(h,r,t) is obtained by normalization using the softmax function;

[0139]

[0140] Where W1, W2, and W3 are weight matrices, b1, b2, and b3 are bias terms, || represents the concatenation operation, and ReLU is the non-linear activation function. and Let h, r, and t be the embedding representations of entity h, relation r, and entity t in the l-th layer of the i-th sub-knowledge graph, respectively. This is the LeakyReLU activation function.

[0141] Step d22: Embed the neighbor entities of entity h from the previous step into the l-th layer of each sub-knowledge graph. The embedding representation of entity h at layer l is obtained by aggregating the entity's embedding representation at layer l with the entity's embedding representation at layer l-1 using the Bi-Interaction method. The specific formula is as follows:

[0142]

[0143] Where ⊙ represents the dot product operation, W4 and W5 are weight matrices, and LeakyReLU is the activation function. and They represent the embedding of entity h in the (l-1)th layer of the i-th sub-knowledge graph and the embedding of neighboring entities, respectively.

[0144] Step d3: In each sub-knowledge graph, the cell line or drug entity embeddings of each layer are aggregated using a hierarchical attention mechanism, and then the aggregated cell line and drug embedding representations are spliced ​​together to obtain the cell line drug embedding representation under each sub-knowledge graph.

[0145] The specific steps in step d3 to obtain the cell line drug embedding representation under each sub-knowledge graph are as follows:

[0146] Step d31: Embed the cell line or drug entity of each layer in each sub-knowledge graph and aggregate them using a hierarchical attention mechanism:

[0147]

[0148] in, It is a transpose operation, σ(·) is the tanh activation function, and q (i,l) w (i,l) and b (i,l) Let ω be the attention vector, weight matrix, and bias vector of the embedding representation of the i-th sub-knowledge graph at layer l. (i,l) Let H2 be the weights before softmax normalization of the embeddings of the l-th layer of the i sub-knowledge graphs, H2 be the number of heads in the hierarchical multi-head attention mechanism, and L be the maximum depth of hierarchical propagation; e (i,l) Indicates the cell line or drug corresponding to e (i,0) and e (i,L) Embedding representations of the 0th and Lth layers of cell lines or drug entities, γ (i,0) and γ (i,L)The weights represent the embedding representations of cell lines or drug entities at layers 0 and L.

[0149] Step d32: Perform a splicing operation on the cell line and drug embeddings obtained from each sub-knowledge graph in step d31:

[0150]

[0151] Here, || represents the concatenation operation. and These are the embedding representation vectors for each layer of cell line and drug aggregation, respectively, which are the e-vectors corresponding to each cell line and drug. i . The embedding of the DCP obtained in the i-th sub-knowledge graph.

[0152] Step d4: Employ a meta-path-level attention mechanism to weighted aggregate the cell line drug response relationship embeddings under each sub-knowledge graph, resulting in the final cell line drug response relationship embeddings:

[0153]

[0154] in, and β i Let and be the cell line drug response relationship embeddings and corresponding weights generated in the i-th sub-knowledge graph, respectively. H3 is the number of heads in the meta-path-level multi-head attention mechanism. LeakyReLU is a non-linear activation function, m is the total number of meta-paths, and e is the number of heads. c,d This is for the final cell line drug embedding.

[0155] Step E, Reaction Sensitivity:

[0156] Step e1: Perform drug response relationships (DCPs) on all obtained cell lines, i.e., e c,d We optimized this model using contrastive learning, specifically training a contrastive learning model to obtain more discriminative DCP vector representations. First, the training samples were divided into positive and negative groups based on the similarity between DCP vectors. Specifically, cell line-drug pairs with high vector cosine similarity belonged to the positive group, while randomly selected cell line-drug pairs with low similarity were assigned to the negative group. The optimization objective was to maximize the similarity of positive sample DCPs and minimize the similarity of negative sample DCPs; this unsupervised method aimed to improve the ability to distinguish between positive and negative DCP features without predefined labels. In our study, positive sample pairs were DCPs identified by high scores (high similarity) in cosine similarity, while negative sample pairs were randomly selected from all DCPs. The loss optimization function for contrastive learning can be expressed as:

[0157]

[0158] Among them, e c,d It is the DCP embedding representation obtained in step d4. This is the DCP embedding representation in the positive sample pair. It is the DCP embedding representation in negative sample pairs. For the embedding transpose operation, τ is an adjustable temperature coefficient, K is the number of negative samples, E is the expected value, and * is used to calculate the cosine similarity between two DCP embedding representations.

[0159] Step e2: Optimize the DCP feature representation e from step e1. c,d As a basis for predicting the response sensitivity of cancer cell lines to drugs, a multilayer perceptron (MLP) and a sigmoid function were used to predict the response sensitivity of cancer cell lines to drugs.

[0160] The sensitivity of cancer cell lines to drug response was expressed using MLP and the sigmoid function as follows:

[0161]

[0162] In this context, MLP stands for Multilayer Perceptron, and Sigmoid is the activation function. To predict the probability of drug sensitivity in the obtained cell lines.

[0163] Example:

[0164] On the GDSC dataset, various experiments were conducted to test and evaluate the model's performance. For example... Figures 2 to 8 As shown.

[0165] To test and evaluate the performance of this application, it was evaluated on a dataset, using AUC (area under the ROC curve), AUPR (area under the PR curve), precision, recall, and F1 score as evaluation metrics. AUROC is the area under the ROC (receiver operating features) curve, with FP on the x-axis and TP on the y-axis. AUPR is the area under the PR curve, with precision on the x-axis and recall on the y-axis. The specific calculation formulas are as follows:

[0166]

[0167] Wherein, TP represents the number of correctly predicted drug-cell line response relationships, FP represents the number of incorrectly predicted drug-cell line response relationships, TN represents the number of correctly predicted non-drug-cell line response relationships, and FN represents the number of incorrectly predicted non-drug-cell line response relationships.

[0168] To evaluate the impact of the number of heads in multi-level attention, this embodiment examines the effects of different numbers of heads (1 to 6) in meta-path-level attention and different numbers of heads (2 to 8) in entity-level attention on prediction performance. Furthermore, this invention calculates and performs detailed comparative analyses of AUC, AUPR, F1 score, Precision, Recall, ACC, and Specificity values ​​for each depth. Figure 7 As shown, when the number of entity-level attention heads increases to 6 and the number of meta-path-level attention heads increases to 4, the model shows performance improvement on the aforementioned evaluation metrics. This phenomenon may be because the increased number of heads allows the model to extract feature information from more dimensions. When the number of heads is moderate, it can effectively capture key information, thereby improving the overall prediction performance.

[0169] To verify the superiority of this invention, we conducted a series of comparative experiments on the constructed dataset, comparing the cancer drug response prediction method of this invention with other state-of-the-art methods (including GraphCDR, TGSA, DeepTTA, DRPreter, DeepDSC, SubCDR, GraphDRP, and RF). These comparative experiments included four experimental settings: warm-start five-fold cross-validation, cell line cold-start five-fold cross-validation, drug cold-start five-fold cross-validation, and independent testing. Detailed experimental results are shown in Tables 1-6. Figures 2-7 Tables 7 and 8 present the results of the case analysis. Figure 8 For independent testing of experimental results, Figure 9 For the embedding dimension sensitivity experiment, Figures 10-15 The results show the sensitivity experiments for entity-level attention head number H1 and meta-path-level attention head number H3, respectively. Figure 16 For the AUC and AUPR values ​​of each organization, Figure 17 To independently test the tissues to which the cell lines belong and their proportions. Figures 18-20 This represents the changes in the weights of each metapath during the training phase in five-fold cross-validation for hot start, five-fold cross-validation for cell line cold start, and five-fold cross-validation for drug cold start.

[0170] The results demonstrate that this invention effectively captures the complex biological semantic information between cell lines and drugs by integrating rich biological data and applying a multi-level attention mechanism, thereby significantly improving predictive performance. Furthermore, the invention exhibited superior performance in four different experimental settings, further demonstrating its advancement in understanding complex drug-cell line response relationships and providing a solid scientific foundation for personalized cancer treatment.

[0171] Table 1 shows the performance comparison results of warm starts on the constructed dataset.

[0172]

[0173] Table 2 shows the comparison results of cold start performance of cell lines on the constructed dataset.

[0174]

[0175] Table 3 shows the comparison results of drug cold start performance on the constructed dataset.

[0176]

[0177] Table 4. AUC values ​​for all methods under each experimental setting.

[0178]

[0179] Table 5. AUPR values ​​for all methods under each experimental setting.

[0180]

[0181] Table 6 shows the performance comparison results of independent tests on the constructed dataset.

[0182]

[0183] Table 7. Top 20 cell lines and their drug-prediction scores

[0184]

[0185] Table 8. The 10 cell lines most sensitive to dasatinib and GSK690693

[0186]

[0187]

[0188] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A cancer drug response prediction method based on knowledge graph, characterized in that, include: S100: Acquire cell line-drug response data, drug-protein association data, cell line-tissue association data, protein-protein association data, drug molecular fingerprint data, and cell line gene mutation (MUT), copy number variation (CNV), and gene expression (EXP) omics data. S200. Calculate the Jaccard similarity of drug molecule fingerprints. The Jaccard similarity constitutes a drug similarity network. The correlation measure between EXP, MUT and CNV data of cell lines is calculated, and the correlation measure between EXP, MUT and CNV data is fused to obtain the similarity between cell lines. The similarity between cell lines constitutes a cell line similarity network. The drug similarity network and cell line similarity network were reconstructed using KNN, and the reconstructed drug similarity network and cell line similarity network were used as similarity information between cell lines and drugs. S300. Using the obtained cell line-drug response data, drug-protein association data, cell line-protein association data, cell line-tissue association data, protein-protein association data, cell line similarity information, and drug similarity information, construct a knowledge graph (KG). Based on cell line-drug response data, drug-protein association data, cell line-protein association data, cell line-tissue association data, protein-protein association data, cell line similarity information, and drug similarity information, multiple relationships between cells, drugs, proteins, and tissues are identified as metapaths. The knowledge graph (KG) is sampled using meta-paths, and each meta-path yields a corresponding sub-knowledge graph; S400. Initialize the embedding of entities and relations in each sub-knowledge graph; obtain multi-layer neighbor information of cell lines or drugs through iterative hierarchical propagation in each sub-knowledge graph, and use multi-head attention mechanism to assign weights to different types of entities and relations in each layer, thereby aggregating the neighbor entity information of each layer in the sub-knowledge graph. In each sub-knowledge graph, the cell line or drug entity embeddings of each layer are aggregated using a hierarchical attention mechanism. Then, the aggregated cell line and drug embedding representations are concatenated to obtain the cell line and drug embedding representations under each sub-knowledge graph. Finally, a meta-path-level attention mechanism is used to weighted aggregate the cell line drug response relationship embeddings under each sub-knowledge graph, resulting in the final cell line drug response relationship embedding e. c,d : S500, DCP based on cell line drug response, also known as e c,d The probability of drug sensitivity in cell lines was predicted using machine learning.

2. The cancer drug response prediction method based on knowledge graph according to claim 1, characterized in that, The calculation of Jaccard similarity of drug molecule fingerprints is used to construct drug similarity: in, and These are the molecular fingerprints of drug d1 and drug d2, respectively.

3. The cancer drug response prediction method based on knowledge graph according to claim 1, characterized in that, The process of calculating the correlation measure between EXP, MUT, and CNV data of cell lines and fusing the correlation measure between EXP, MUT, and CNV data to obtain the similarity between cell lines includes: The correlation measure between EXP, MUT, and CNV data of cell lines is calculated as follows: in, and ρ represents the gene expression values ​​of gene g1 and gene g2, respectively, and ρ is the standardization function. in, and These are the copy number variation values ​​for genes g1 and g2, respectively. in, and ρ represents the binary gene mutation values ​​of genes g1 and g2, respectively, and ρ is the standardization function. The similarities between the cell lines obtained by fusion are as follows: Among them, α1 and α2 are adjustable hyperparameters used to balance the weights of the three omics similarities.

4. The cancer drug response prediction method based on knowledge graph according to claim 3, characterized in that, The drug similarity network and cell line similarity network reconstructed using KNN are represented as follows: Among them, S drug N represents the reconstructed drug similarity network. t Let S represent the set of neighboring nodes of drug t, and u represent a node in the set of neighboring nodes of drug t; cell N represents the reconstructed cell line similarity network. a Let b represent the set of neighboring nodes of cell line a, and let b represent the nodes in the set of neighboring nodes of cell line a.

5. The cancer drug response prediction method based on knowledge graph according to claim 1, characterized in that, The process of identifying multiple relationships between cell, drug representation, protein, and tissue expression as metapaths in S300 includes: Cell lines are represented as Cell, drugs as Drug, proteins as Protein, and tissues as Tissue. Using the obtained cell line-drug response data, drug-protein association data, cell line-protein association data, cell line-tissue association data, protein-protein association data, cell line similarity information, and drug similarity information, denoted as relations r1-relation r7 respectively, meta-paths related to cancer drug response prediction are determined, including meta-paths M1-M5: Metapaths are used to generate corresponding sub-knowledge graphs on a knowledge graph.

6. A cancer drug response prediction method based on knowledge graphs according to any one of claims 1 to 5, characterized in that, S400 uses a multi-head attention mechanism to assign weights to different types of entities and relationships in each layer, and then aggregates the neighbor entity information of each layer in the sub-knowledge graph. The process includes: A multi-head attention mechanism is employed to integrate the neighbor information of entity h by processing multiple information subspaces in parallel. Among them, W O It is a weight matrix that maps the connection headers back to the original embedding space. Embed the entity as a neighboring entity in the i-th sub-knowledge graph at the l-th layer; head j It is the neighbor entity aggregation embedding under the j-th head in multi-head attention, e t Embedding of the tail entity; Let α be the set of neighbors of entity h under the i-th sub-knowledge graph; j (h,r,t) represents the weight score of the triple (h,r,t) under the j-th attention head in the subgraph. It controls how much information propagates from t to h among the j attention heads, and is expressed as: α j (h,r,t) is α j ′(h,r,t) is obtained by normalization using the softmax function; Where W1, W2, and W3 are weight matrices, b1, b2, and b3 are bias terms, || represents the concatenation operation, and ReLU is the non-linear activation function. and Let θ represent the embedding representations of entity h, relation r, and entity t in the l-th layer of the i-th sub-knowledge graph, respectively, and let θ be the LeakyReLU activation function. The neighboring entities of entity h from the previous step are embedded in the l-th layer of each sub-knowledge graph. The embedding representation of entity h at layer l is obtained by aggregating the entity's embedding representation at layer l-1 using the Bi-Interaction method: Where ⊙ represents the dot product operation, W4 and W5 are weight matrices, and LeakyReLU is the activation function. and They represent the embedding of entity h in the (l-1)th layer of the i-th sub-knowledge graph and the embedding of neighboring entities, respectively.

7. The cancer drug response prediction method based on knowledge graph according to claim 6, characterized in that, In S400, the process of aggregating the cell line or drug entity embeddings of each layer in each sub-knowledge graph using a hierarchical attention mechanism, and then concatenating the aggregated cell line and drug embedding representations to obtain the cell line and drug embedding representations under each sub-knowledge graph includes: In each sub-knowledge graph, cell line or drug entities at each layer are embedded and aggregated using a hierarchical attention mechanism: in, It is a transpose operation, σ(·) is the tanh activation function, and q (i,l) w (i,l) and b (i,l) Let ω be the attention vector, weight matrix, and bias vector of the embedding representation of the i-th sub-knowledge graph at layer l. (i,l) Let H2 be the weights before softmax normalization of the embeddings of the l-th layer of the i sub-knowledge graphs, H2 be the number of heads in the hierarchical multi-head attention mechanism, and L be the maximum depth of hierarchical propagation; e (i,l) Indicates the cell line or drug corresponding to e (i,0) and e (i,L) Embedding representations of the 0th and Lth layers of cell lines or drug entities, γ (i,0) and γ (i,L) The weights representing the embedding representations of cell lines or drug entities at layers 0 and L; The cell lines and drug embeddings obtained from each sub-knowledge graph are spliced ​​together separately: Here, || represents the concatenation operation. and These are the embedding representation vectors for each layer of cell line and drug aggregation, respectively, which are the e-vectors corresponding to each cell line and drug. i ; The embedding of the DCP obtained in the i-th sub-knowledge graph.

8. The cancer drug response prediction method based on knowledge graph according to claim 7, characterized in that, The final cell line drug response relationship embedding representation in S400 in, and β i Let and be the cell line drug response relationship embeddings and corresponding weights generated in the i-th sub-knowledge graph, respectively; H3 is the number of heads in the meta-path level multi-head attention mechanism; LeakyReLU is the non-linear activation function; m is the total number of meta-paths; e c,d This is for the final cell line drug embedding.

9. The cancer drug response prediction method based on knowledge graph according to claim 8, characterized in that, Before using machine learning to predict the probability of drug sensitivity in cell lines in S500, contrastive learning is needed to first analyze the e c,d Optimize, for the optimized e c,d Predict using machine learning; Using contrastive learning to study e c,d The optimization process includes: Based on all obtained cell line drug response relationships (DCP), i.e., e c,d A contrastive learning model is used for optimization. Training samples are divided into positive and negative groups based on the similarity between DCP vectors. The optimization objective is to maximize the similarity of positive sample DCPs and minimize the similarity of negative sample DCPs. The loss function for contrastive learning is expressed as: Among them, e c,d It is the DCP embedding representation obtained from S4004. This is the DCP embedding representation in the positive sample pair. It is the DCP embedding representation in negative sample pairs. For the embedding transpose operation, τ is an adjustable temperature coefficient, K is the number of negative samples, E is the expected value, and * is used to calculate the cosine similarity between two DCP embedding representations.

10. A cancer drug response prediction method based on knowledge graphs according to claim 9, characterized in that, In the process of predicting the probability of drug sensitivity in cell lines using machine learning, S500 utilizes MLP and the sigmoid function to predict the drug sensitivity of cancer cell lines: In this context, MLP stands for Multilayer Perceptron, and Sigmoid is the activation function. To predict the probability of drug sensitivity in the obtained cell lines.

Images