A drug synergistic effect prediction method, system, device and medium based on heterogeneous graph tensor decomposition

Abstract

The application discloses a drug synergistic effect prediction method, system, device and medium based on heterogeneous graph tensor decomposition, and the prediction method comprises the following steps: obtaining a SMILES sequence of a drug, extracting a molecular structure feature of the drug, and obtaining a molecular structure feature representation of the drug; constructing a drug pair heterogeneous graph in each cell line, and obtaining a local interaction feature of the drug through a heterogeneous graph conversion network; performing Tucker decomposition on a three-channel heterogeneous graph relationship tensor, splicing the decomposition result with the local interaction feature of the drug, and extracting a global interaction feature vector of the drug; and predicting a synergistic score of a current drug-drug combination in a cell line according to the molecular structure feature representation of the drug and the global interaction feature vector of the drug. The application integrates the SMILES sequence of the drug and gene expression information of the drug pair in the cell line, and also converts the SMILES sequence into a drug molecular structure graph by using a graph convolution network, so as to extract the molecular structure feature of the drug. Such a design enables TensoGraph to more accurately reflect the complexity and diversity of the drug interaction network in the real world, thereby improving the prediction performance.

Description

Technical Field

[0001] This invention belongs to the field of drug response prediction technology, specifically relating to a method, system, device and medium for predicting drug synergy based on heterogeneous graph tensor decomposition. Background Technology

[0002] With the continuous development of treatment methods for major diseases such as cancer, single-drug therapy faces clinical bottlenecks such as strong toxic side effects and high drug resistance. Multi-drug combination therapy, by acting on multiple targets, can not only improve treatment efficacy but also reduce drug toxicity and drug resistance, thus gradually becoming an important trend in cancer treatment. To effectively design drug combination regimens, predictive research on drug synergy has become a crucial link in precision medicine and drug development.

[0003] Traditionally, the assessment of drug synergies has relied on experimental methods, such as high-throughput screening (HTS). However, these methods face significant limitations, including high cost, low efficiency, and long experimental cycles, when dealing with the exponentially growing space of drug combinations and cancer heterogeneity. In recent years, thanks to the rapid development of computing technology, researchers have begun to explore machine learning, deep learning, and even graph neural networks to predict and model drug synergies.

[0004] In existing technologies, mainstream methods can be mainly divided into three categories: traditional machine learning methods, deep neural network methods, and graph neural network methods. Traditional machine learning methods, such as Support Vector Machines (SVM), Random Forests (RF), and Gradient Boosting Trees (GBDT), can handle multimodal features, but have limited ability to capture complex nonlinear relationships and feature interactions. Deep learning methods, such as Feedforward Neural Networks (FNN), Convolutional Neural Networks (CNN), and Transformer structures, are more suitable for modeling high-dimensional complex relationships and improve prediction accuracy to some extent, but often lack interpretability and integration of biological priors. Graph Neural Networks (GNN), due to their ability to process structured data and model complex interactions between drugs and between drugs and cell lines, have become an important direction for drug synergistic prediction in recent years.

[0005] Although graph neural networks offer a new paradigm for solving this problem, current technology still has several key shortcomings:

[0006] (1) Limited ability to handle heterogeneous information: Drug synergy is affected by various types of data, such as chemical structure, target, gene expression, etc. Existing models often have difficulty integrating and modeling these heterogeneous information efficiently at the same time.

[0007] (2) Insufficient ability to model interaction relationships: There are complex nonlinear and multi-level interaction relationships between drugs and between drugs and cell lines. Traditional feature splicing or simple weighting is difficult to accurately model these relationships, affecting the model's prediction performance.

[0008] (3) Poor model interpretability: In clinical applications, model outputs not only need to be accurate, but also need to explain their prediction mechanisms. Currently, many models are black-box structures and lack interpretable interaction mechanism designs.

[0009] (4) Poor scalability and generalization ability: Some models rely on the coverage of drug combinations in the training set. Once they encounter unseen combinations or new cell lines, their generalization ability drops significantly, limiting the applicability of the models in practice. Therefore, how to design a drug synergy prediction model with stronger heterogeneous information fusion ability, more complex interactive modeling mechanism, and at the same time take into account interpretability and generalization performance remains a difficult and hot topic in current research. Summary of the Invention

[0010] To address the technical problem that existing technologies often fail to consider the impact of drugs on the cell line environment and cannot characterize the heterogeneous interactions among "drug-drug-cell line," resulting in inaccurate predictions of drug interactions, the present invention aims to provide a method, system, device, and medium for predicting drug synergy based on heterogeneous graph tensor decomposition.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] A method for predicting drug synergy based on heterogeneous graph tensor decomposition includes the following steps:

[0013] Obtain the SMILES sequence of the drug, and extract the molecular structural features of the drug based on the SMILES sequence to obtain the molecular structural feature representation of the drug.

[0014] Construct a heterogeneous graph of drug pairs in each cell line. Based on the heterogeneous graph of drug pairs in each cell line, obtain a three-channel heterogeneous graph relation tensor. Then, obtain the local interaction features of drugs through a heterogeneous graph transformation network.

[0015] Tucker decomposition was performed on the heterogeneous graph relation tensor of the three channels, and the decomposition results were concatenated with the local interaction features of the drug to extract the global interaction feature vector of the drug.

[0016] Based on the molecular structural features of drugs and their global interaction feature vectors, a fusion feature vector mechanism is constructed through drug combinations to predict the synergistic score of the current drug-drug combination in the cell line.

[0017] Furthermore, the SMILES sequence of each drug is transformed into a drug molecule structure diagram, and an L-layer graph convolutional network is used to update the node representation by aggregating the neighbor information of each atom node.

[0018] Then, through the stacking operation of the L-layer heterogeneous graph transformation network, the feature matrix of the last layer node of the heterogeneous graph transformation network is aggregated through the global average pooling operation to obtain the molecular structure feature representation of the drug.

[0019] Furthermore, an L-layer graph convolutional network is used to update the node representation of the drug molecule structure diagram by aggregating the neighbor information of each atom node, as shown in the following formula:

[0020]

[0021] in, This is a meta-path tensor that includes information about the nodes themselves. For is The degree matrix, H i (l) H is the feature matrix of the l-th layer node. i (l-1) H is the feature matrix of the (l-1)th layer node. i (0) Let W be the initial representation, ReLU be the activation function, and W be the value of W. i (l-1) is the learnable weight parameter matrix for linear transformation.

[0022] Furthermore, the molecular structural characteristics of the drug represent m i Calculated using the following formula:

[0023]

[0024] Where, m i Let n represent the molecular structural features of the i-th drug, where n is the total number of drugs.

[0025] Furthermore, the local interaction features of the drugs are obtained through the following process: integrating all drug data and constructing a heterogeneous graph of drug combinations in each cell line; performing convolutional transformation on the three-channel heterogeneous graph relation tensor through a heterogeneous graph transformation network to extract metapath information; and fusing the metapath information using a semantic-level attention mechanism to obtain the local interaction features of the drugs.

[0026] Furthermore, the local interaction characteristics of drugs are obtained through the following process: the relationships of all drug combinations in the r-th cell line are distinguished according to three biological effect types: synergistic, additive, and antagonistic. Three types of edge sets are constructed respectively, and these three types of edge sets are transformed into synergistic adjacency matrices, additive adjacency matrices, and antagonistic adjacency matrices, respectively, to form a three-channel heterogeneous graph relation tensor.

[0027] A heterogeneous graph transformation network is used to perform convolution transformation on the three-channel heterogeneous graph relation tensor to obtain the intermediate adjacency relation matrix;

[0028] The intermediate adjacency relation matrices are further computed to obtain the meta-path tensor; the heterogeneous graph transformation network is applied to each channel of all meta-path tensors, and the representations of multiple channels are concatenated to obtain the embedded representation;

[0029] Based on the embedding representation, the weights of each channel are obtained by learning the heterogeneous graph relation tensor of the three channels through an attention mechanism.

[0030] Based on the weight of each channel, the embeddings of each channel are aggregated by attention score weighting to obtain the interaction features of the drug.

[0031] Furthermore, the weight m of each channel r Calculated using the following formula:

[0032]

[0033] in,‖ c It is a function spliced ​​along the channel, Z r|c|i is the representation of the c-th channel of the i-th drug in cell line r, and q is the query vector in the attention mechanism. It is a learnable weight matrix, b r It is a learnable bias vector, and tanh is the activation function;

[0034] Drug interaction characteristics Z r Calculated using the following formula:

[0035]

[0036] Among them, Z r|c It is the c-th channel of drug interaction characteristics in cell lines, m r|i For each channel, the i-th value in the weight vector, m r|j The j-th value in the weight vector for each channel.

[0037] A drug synergy prediction system based on heterogeneous graph tensor decomposition includes:

[0038] The feature extraction module is used to obtain the SMILES sequence of the drug and extract the molecular structure features of the drug based on the SMILES sequence to obtain the molecular structure feature representation of the drug.

[0039] The transformation module is used to construct the heterogeneity map of drug pairs in each cell line. Based on the heterogeneity map of drug pairs in each cell line, a three-channel heterogeneity map relation tensor is obtained, and the local interaction features of drugs are obtained through the heterogeneity map transformation network.

[0040] The decomposition module is used to perform Tucker decomposition on the heterogeneous graph relation tensor of the three channels, and concatenate the decomposition results with the local interaction features of the drug to extract the global interaction feature vector of the drug.

[0041] The prediction module is used to predict the synergistic score of the current drug-drug combination in the cell line by constructing a fusion feature vector mechanism through drug combination based on the molecular structural feature representation and global interaction feature vector of the drug.

[0042] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the drug synergy prediction method based on heterogeneous graph tensor decomposition.

[0043] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the drug synergy prediction method based on heterogeneous graph tensor decomposition.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention presents a drug synergy prediction method based on heterogeneous graph tensor decomposition. From the perspective of heterogeneous graphs, it constructs a heterogeneous graph within a specific cell line and combines a heterogeneous graph transformation network (GTN) with tensor decomposition (Tucker decomposition) to capture drug characteristics in the cellular environment. This invention integrates the drug's SMILES sequence and the drug's gene expression information in its cell line. Furthermore, it utilizes a graph convolutional network (GCN) to convert the SMILES sequence into a drug molecular structure graph, thereby extracting drug molecular structural features. This design allows TensoGraph to more accurately reflect the complexity and diversity of drug interaction networks in the real world, thus improving prediction performance.

[0046] The core of this invention lies in the first-time coupling of the three-dimensional heterogeneous tensor of "cell line-drug-relationship" with graph learning, forming a "global-local" dual-channel modeling mechanism, which has the following advantages:

[0047] 1. Introduction of cell line microenvironment. This invention explicitly models cell lines as a third dimension, constructs a relationship tensor for each drug pair in each cell line, and models a heterogeneous graph relationship tensor of "drug-drug-cell line". This allows the three relationships of synergy, additive, and antagonism to change dynamically with the cell environment. For the first time, this invention solves the prediction distortion problem caused by "environment independence" and overcomes the problem in existing technologies that treat "drug-drug" (drug combination) relationships as a single, static binary variable, ignoring the regulation of drug efficacy by cell lines.

[0048] 2. Global High-Order Tensor Decomposition. This invention performs Tucker decomposition on the heterogeneous graph relation tensor to obtain global interaction features across cell lines, while preserving the high-order structure of the tensor, avoiding information collapse, significantly improving cross-domain generalization ability, and overcoming the problem that traditional tensor methods only extract global low-dimensional embeddings and cannot characterize fine-grained local structures.

[0049] 3. Local Heterogeneous Meta-Path Learning. This invention introduces a Heterogeneous Graph Transformation Network (GTN), which captures local topological differences in cell line-specific graphs through an automatic meta-path discovery mechanism. This overcomes the insufficient representation caused by "fixed paths" and the problem that existing graph neural networks rely on hand-crafted meta-paths and are difficult to adapt to multi-relationship scenarios.

[0050] In summary, this invention, through a system-level innovation of "explicit modeling at the cell line level + global tensor decomposition + local meta-path learning + multimodal fusion", solves the two major pain points of environmental heterogeneity and representation integrity for the first time within a unified framework. The experimental accuracy is significantly better than existing baselines, and it also has biological interpretability, thus substantially overcoming the shortcomings of existing technologies such as inaccurate prediction, environmental irrelevance, and single representation. Attached Figure Description

[0051] Figure 1 This is a flowchart of the drug synergy prediction method based on heterogeneous graph tensor decomposition of the present invention;

[0052] Figure 2 This is a flowchart of the drug synergy prediction method based on heterogeneous graph tensor decomposition of the present invention.

[0053] Figure 3 A schematic diagram of the drug synergy prediction system based on heterogeneous graph tensor decomposition of the present invention. Detailed Implementation

[0054] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0055] See Figure 1 and Figure 2 This invention provides a drug synergy prediction method based on heterogeneous graph tensor decomposition (TensoGraph), which effectively solves the information heterogeneity problem in existing graph neural network-based methods. Specifically, it includes the following steps:

[0056] S1: Obtain the SMILES sequence of the drug, and extract the molecular structural features of the drug based on the SMILES sequence to obtain the molecular structural feature representation of the drug.

[0057] First, the SMILES sequence of each drug is converted into a drug molecular structure diagram, a process accomplished using the open-source chemical information tool RDKit. Below, using the i-th drug as an example, we will detail the process of extracting the molecular structural features of the drug based on its SMILES sequence to obtain its molecular structural feature representation.

[0058] For the drug molecule structure diagram of the i-th drug It can be done To construct, in which, Let be the set of all atomic nodes of the i-th drug. Let be the set of all chemical bonds in the i-th drug.

[0059] Each atomic node Each atom corresponds to a chemical atom (e.g., C, H, O, etc.). Each atom has a corresponding atomic node feature vector, containing various attributes of that atom, such as element type, number of hidden hydrogen atoms, number of valence electrons, bonding configuration, charge state, and hybridization type. Therefore, the atomic node feature vector corresponding to each atom can serve as the initial features of the drug. The initial feature matrix of the i-th drug... It is composed of attributes such as element type, number of hidden hydrogen atoms, number of valence electrons, bonding situation, charge state, and hybridization type. Representing n i A real field matrix of size ×c, n i Let c be the number of atoms in the i-th drug, and c be the dimension of the feature. Each row of the initial feature matrix of the i-th drug represents the feature vector of one atom.

[0060] Each edge Adjacency matrix represents the chemical bond between two atoms. It is used to describe these chemical bond connections. Representing n i ×n i A real-field matrix. For example, if there is a chemical bond between the j-th atom and the k-th atom in the i-th drug, then the adjacency matrix A is... i (j,k)=1, otherwise the adjacency matrix A i (j,k)=0, where j≠k, and j and k are any two different atoms in the drug.

[0061] This invention uses an L-layer graph convolutional network (GCN) to visualize drug molecule structures. During feature extraction, the node representation is updated by aggregating the neighbor information of each atomic node. The update formula for this process in the l-th layer GCN is as follows:

[0062]

[0063] in, The adjacency matrix is ​​expanded to incorporate node information by adding an identity matrix I (self-loop). This is a meta-path tensor that includes information about the nodes themselves. yes The degree matrix represents the total number of connections (including self-loops) for each node. H i (l) H represents the feature matrix of the l-th layer node. i (l-1) This represents the feature matrix of the (l-1)th layer nodes, initially represented by H. i (0) =X i ReLU is the activation function, W i (l-1) It is the learnable weight parameter matrix of the linear transformation of this layer.

[0064] Through the stacking operation of L layers of GCN, node features are aggregated layer by layer from information from more distant neighbors, thereby effectively integrating the information of the entire molecular graph structure. After graph convolution, in order to obtain a unified representation of the entire molecule, the feature matrix of the nodes in the last layer of GCN, i.e., the Lth layer, is... Aggregation is performed using global average pooling to obtain the molecular structural feature representation m of the drug. i The calculation formula is as follows:

[0065]

[0066] Where, m i This represents the molecular structure characteristics of the i-th drug, comprehensively characterizing the molecular structure information of the drug, where n is the total number of drugs.

[0067] S2: Construct a heterogeneous graph of drug pairs in each cell line, obtain the local interaction features of drugs through a heterogeneous graph transformation network, and then capture local topological differences in a cell line-specific graph.

[0068] Specifically, this invention integrates all drug data and constructs a heterogeneous graph of drug combinations in each cell line, taking into account the heterogeneity of information in the graph. Metapath information is extracted through a heterogeneous graph transformation network (GTN), and these metapath information are further fused using a semantic-level attention mechanism to obtain local interaction features. Therefore, it can overcome the insufficient representation caused by "fixed paths".

[0069] More specifically, taking the r-th cell line as an example, this invention distinguishes the relationships between all drug combinations in the r-th cell line according to three types of biological effects: synergistic, additive, and antagonistic. Three types of edge sets are constructed for each type, denoted as the synergistic edge set. Additive edge set Antagonistic Edge Set Furthermore, these three types of edge sets are transformed into cooperative adjacency matrices A respectively. r|1 Additive adjacency matrix A r|2 and antagonistic adjacency matrix A r|3 This constitutes a three-channel heterogeneous graph relation tensor. This represents a tensor of form n×n×3, where n is the total number of drugs. This heterogeneous graph relation tensor... It is a heterogeneous graph in the r-th cell line because the cell line is explicitly modeled as a third dimension, thereby enabling research in the three dimensions of "drug-drug-cell line" and not limited to "drug-drug" (drug combination) research.

[0070] This invention employs a heterogeneous graph transform network (GTN) to perform convolutional transformation on the three-channel heterogeneous graph relation tensor. Specifically, an L-layer GTN network is used to automatically learn multiple drug semantic paths. To obtain different types of meta-path information, this invention introduces L+1 multi-channel learnable parameters into the L-layer GTN. In each layer, for multi-channel learnable parameters The softmax function is applied, where 1×1 represents a convolution operation, 3 represents the three types of edges, and c is the number of channels. The specific formula is as follows:

[0071]

[0072] in, This represents the intermediate adjacency relation matrix of all GTN networks. These matrices serve as temporary intermediate products of the GTN network and are used to subsequently compute the GTN network's output. An L-layer GTN will generate L+1 intermediate adjacency relation matrices. Specifically, the first layer has two intermediate adjacency relation matrices, namely the first intermediate adjacency relation matrix. Second intermediate adjacency matrix Each of the remaining l-th layers has only one intermediate adjacency matrix. This represents the l-th convolutional layer network. It corresponds to the l-th layer network The learnable parameter matrix.

[0073] After stacking L layers of GTN networks, the aforementioned intermediate adjacency matrix is ​​obtained. These intermediate adjacency matrices are then calculated to obtain the final output of each GTN layer, which is the metapath tensor. A large number of meta-path tensors implies richer path information, and the specific calculation process is as follows:

[0074]

[0075] in, and These are the first, second, and Lth composite element path tensors in cell line r, respectively. The first intermediate adjacency matrix of the first layer GTN network The c-th channel, The second intermediate adjacency matrix of the first-layer GTN network The c-th channel, for and The normalization degree matrix of the product of . For the output of the second-layer GTN network The c-th channel, For the first composite element path tensor The normalization matrix of the c-th channel, It is the first composite element path tensor The c-th channel. The output of the Lth layer GTN network The c-th channel, It is the path tensor of the (L-1)th composite element. The normalization matrix of the c-th channel, It is the path tensor of the (L-1)th composite element. The c-th channel.

[0076] Finally, GCN is applied to all metapath tensors. Each channel is processed, and the representations of multiple channels are concatenated to obtain the final embedded representation. for:

[0077]

[0078] in,‖ c The function representing splicing along the channel. The adjacency matrix is ​​expanded by adding an identity matrix I (self-loop) to incorporate information about the nodes themselves. yes The normalized degree matrix, i.e. W is a meta-path tensor that includes information about the nodes themselves. r ∈R m×d It is a trainable weight matrix shared across channels. X is the fingerprint feature matrix of all drugs, where n is the total number of drugs, m is the fingerprint feature dimension, and σ is the activation function. The fingerprint feature matrix X of all drugs can be obtained using the existing tool Deep Graph Infomax (DGI). Embedded representation This refers to the embedding representation of all drugs in cell line r.

[0079] According to the embedding representation Semantic attention mechanisms are used to compute the final drug interaction features in cell lines. First, the importance of each channel, i.e., the weight of each channel, needs to be learned through the attention mechanism.

[0080]

[0081] in,‖ c Z represents the function for splicing along the channel. r|c|i This represents the c-th channel of the i-th drug in cell line r. It is the query vector in the attention mechanism. It is a learnable weight matrix, meaning in Chinese. It is a learnable bias vector, and tanh is the activation function. Based on the weight m of each channel... r Embed each channel into Z r|c Attention score-weighted aggregation yields the final drug interaction characteristics Z in the cell line. r .

[0082]

[0083] Among them, Z r|c It is Z r The c-th channel, m r|i Weight m for each channel r The i-th value in the vector, m r|j Weight m for each channel r The j-th value in the vector, It is the interaction characteristic of the final drug in the cell line.

[0084] S3: Based on the three-channel heterogeneous graph relation tensor constructed in step S2 The global interaction feature vector of the drug is extracted using Tucker decomposition.

[0085] The Tucker decomposition method (Some mathematical notes on three-mode factor analysis. Psychometrika, 1966, 31(3):279-311.) was used to decompose the global interaction features across cell lines, while preserving the higher-order structure of the tensor, avoiding information collapse, and significantly improving cross-domain generalization ability. Specifically, the three-channel heterogeneous graph relation tensor was decomposed... Tucker decomposition is performed as follows:

[0086]

[0087] in, It is the core tensor, a compressed low-rank tensor that captures the potential high-order correlations between all patterns and captures the multidimensional potential patterns of drug interactions (such as pathway combinations). It is a low-dimensional embedding of the drug as the initiator of the interaction, and it is the global structural or functional feature of the drug. It is a low-dimensional embedding of the drug as an interaction receiver, reflecting the drug's response characteristics. These are low-dimensional features representing the interaction types (cooperative / additive / antagonistic). r1, r2, and r3 represent the three modes mentioned above (U... r V r J r Rank under ) × n This represents the tensor-matrix product of pattern n.

[0088] Predicting the synergistic effect of drug combinations requires focusing on the active roles of the drugs (i.e., which drugs tend to trigger synergistic effects). The global structural or functional characteristics of the drugs encode the characteristics of the drugs as initiators, directly relating to their dominant role in the combination; therefore, these are considered as the global interaction characteristics of the drugs. The pattern matrix associated with the first model, i.e., the global structural or functional characteristics U of the drugs, is... r The interaction characteristics Z of the drug in the cell line obtained in step S2 r When pieced together, the final interactive embedding of the drug in cell line r is represented as follows:

[0089] H r =[U r Z r ],

[0090] Among them, H r Let H be the global interaction feature vector of the drug. r The i-th row represents the interactive embedding H of drug i in cell line r. i,r .

[0091] S4: Based on the molecular structural characteristics of the drug obtained in step S1, m i The global interaction feature vector H of the drug obtained in step S3 r A fusion feature vector mechanism was constructed for drug combinations to predict the synergistic score of the current "drug-drug" combination in cell lines.

[0092] This invention integrates global interaction feature vectors, local interaction features of drugs, molecular structure features of drugs, and fingerprint features of drugs end-to-end to form an interpretable joint characterization. Through multimodal unified characterization, it maintains robust prediction under small sample and high noise conditions, providing mechanistic support for personalized combination therapy.

[0093] The objective of this invention is to find the drug-drug interactions in cell lines, which can be expressed as drug-drug-cell line combinations and combination fractions. Therefore, the following example uses the drug-drug-cell line combination of the i-th drug and the j-th drug in cell line r to predict their interactions.

[0094] To further improve the accuracy of predicting drug combination synergistic effects, a multi-input deep neural network prediction module was designed (this multi-input deep neural network prediction module uses four different MLP networks). The multi-input deep neural network prediction module extracts, reduces, and fuses multiple features related to the drug combination in separate modules, and then performs a unified prediction. This multi-input deep neural network prediction module receives four types of input features, including: the molecular structure feature representations of the i-th and j-th drugs obtained in step S1 (m... i m j ); fingerprint features (X) of the i-th drug and the j-th drug i X j ), x i For the i-th row of the fingerprint feature matrix X of all drugs, X j The j-th row of the fingerprint feature matrix X for all drugs; the interaction features (H) between the i-th and j-th drugs in cell line r obtained in step S3. i,r H j,r ); and cell line r characteristics (C r ), C r This is the gene expression matrix of the cell. Each type of feature is dimensionality-reduced using a two-layer deep neural network prediction module with batch normalization. All dimensionality-reduced features are then concatenated to predict the synergistic score of the drug combination.

[0095]

[0096] Where, m′ i=MLP1(m i ), m′ j =MLP1(m j ), m′ i It is the i-th drug molecule structural feature m i The new representation obtained after passing through the first MLP network, m′ j It is the j-th drug molecule structural feature m j The new representation obtained after passing through the first MLP network. X′ i =MLP2(X i ), X′ j =MLP2(X j ), X′ i X is the i-th drug fingerprint feature. i The new representation obtained after passing through the second MLP network, X′ j X is the j-th drug fingerprint feature. j The new representation obtained after passing through the second MLP network. H′ i,r =MLP3(H i,r ), H′ j,r =MLP3(H j,r ), H′ i,r H is the interaction feature of the i-th drug in cell line r. i,r The new representation obtained after passing through the third MLP network, H′ j,r H is the interaction feature of the j-th drug in cell line r. h,r The new representation obtained after passing through the third MLP network. C′ r =MLP4(C r ), C′ t The gene expression matrix C of cell r r The new representation obtained after passing through the fourth MLP network.

[0097] To evaluate and optimize the deviation between the predicted and actual values, the multi-input deep neural network prediction module uses mean squared error as the loss function. The definition is as follows:

[0098]

[0099] Among them, y t This is the true score for group t, "drug-drug-cell line". This is the score predicted by the model corresponding to the "drug-drug-cell line" combination. This represents the total number of drug combinations, where r is the total number of cell lines. For permutation and combination calculations, t represents the t-th group of "drug-drug-cell line" combinations.

[0100] The following detailed examples illustrate the heterogeneous graph tensor decomposition method.

[0101] A specific embodiment of this invention uses the O'neil dataset on the Loewe score to verify the drug synergy prediction method based on heterogeneous graph tensor decomposition described in this application. The aforementioned O'neil dataset includes comprehensive data on 39 cell lines, 38 drugs, and 22,737 drug combinations. Among these, 1,973 drug combinations exhibit synergistic effects, 12,087 exhibit additive effects, and 8,677 exhibit antagonistic effects. In this specific embodiment, the dataset is divided into 10 parts: 8 parts are used as the training set, 1 part as the validation set, and 1 part as the test set, for 10-fold cross-validation. Training is then performed based on this data using the Adam optimizer with a learning rate of 1*10⁻⁴ to predict the synergistic score of drug combinations in specific cell lines. The synergistic score is then used to determine whether the drug combination exhibits synergistic or antagonistic effects.

[0102] The specific embodiments of the present invention are compared with the TensoGraph-GCN model, the TensoGraph-GCN-Tucker model, the TensoGraph-GCN-GTN model, the TensoGraph-Tucker-GTN model, and the TensoGraph-GCN-Tucker-GTN model.

[0103] Specifically, the TensoGraph-GCN model, based on the TensoGraph of this invention, removes the part of GCN that extracts drug molecular structure features, and only retains drug interaction features, drug fingerprint features, and cell line features for drug prediction.

[0104] Specifically, the TensoGraph-GCN-Tucker model, based on the TensoGraph of this invention, not only removes the function of GCN in extracting drug molecular structure features, but also removes the global interaction features obtained by Tucker decomposition, and only uses the local interaction features of the drug, Infomax fingerprint features and cell line features for prediction.

[0105] Specifically, the TensoGraph-GCN-GTN model, based on the TensoGraph of this invention, removes the drug molecule structure feature extraction part from GCN and removes the local interaction features generated by GTN, and only uses the global interaction features of the drug, Infomax fingerprint features and cell line features for prediction.

[0106] Specifically, the TensoGraph-Tucker-GTN model, based on the TensoGraph of this invention, removes all interactive features obtained by GTN and Tucker decomposition, and retains only drug molecule structure features, Infomax fingerprint features and cell line features for prediction.

[0107] Specifically, the TensoGraph-GCN-Tucker-GTN model, based on the TensoGraph of this invention, removes the drug molecule structure features extracted by GCN and all interaction features obtained by GTN and Tucker decomposition, and only uses Infomax fingerprint features and cell line features for prediction.

[0108] Furthermore, the TensoGraph-GCN model, TensoGraph-GCN-Tucker model, TensoGraph-GCN-GTN model, TensoGraph-Tucker-GTN model, and TensoGraph-GCN-Tucker-GTN model can be regarded as ablation experiments of the TensoGraph model.

[0109] In a specific embodiment of the present invention, based on the O'neil dataset, the TensoGraph model is compared with the TensoGraph-GCN model, the TensoGraph-GCN-Tucker model, the TensoGraph-GCN-GTN model, the TensoGraph-Tucker-GTN model, and the TensoGraph-GCN-Tucker-GTN model.

[0110] Specifically, when the TensoGraph model is compared with the comparison method, the final experimental results are shown in Table 1.

[0111] Table 1. Ablation experiments on O'Neil's Loewe scores

[0112]

[0113]

[0114] As shown in Table 1, the TensoGraph model improved the MSE by 2.6% and the RMSE by 1.3% compared to the TensoGraph-GCN model (which removes drug molecule structure features); compared to the TensoGraph-GCN-Tucker model (which removes both drug molecule structure features and global interaction features), the MSE improved by 2.4% and the RMSE by 1.1%; compared to the TensoGraph-GCN-GTN model (which removes both drug molecule structure features and local interaction features), the MSE improved by 4.3% and the RMSE by 2.2%; compared to the TensoGraph-Tucker-GTN model (which removes drug-to-drug interaction features), the MSE improved by 7.7% and the RMSE by 3.9%; and compared to the TensoGraph-GCN-Tucker-GTN model (which removes both drug molecule structure features and drug-to-drug interaction features), the MSE improved by 10.2% and the RMSE by 5.2%. Experimental results show that each component in the model is crucial, and the TensoGraph model helps to extract and fuse different features to obtain richer features, thereby making more accurate predictions.

[0115] To address the heterogeneity of information in drug interaction networks, this invention proposes a drug synergy model based on heterogeneous graph tensor decomposition. This model integrates drug SMILES sequences, Infomax fingerprint features, and gene expression data. First, the drug SMILES sequences are transformed into a drug molecular graph, and a graph neural network is used to extract drug molecular structural features. Then, the heterogeneous graph constructed in the cell line is decomposed using a graph transformation network and Tucker decomposition to extract drug interaction features. Finally, a deep neural network is used to minimize the prediction error to predict the synergy score, thereby achieving an effective fusion of drug combination and cell line relationship, improving the accuracy of drug synergy prediction, accelerating the drug development process, and laying a solid foundation for its widespread application in the field of drug synergy prediction.

[0116] See Figure 3 The present invention provides a drug synergy prediction system based on heterogeneous graph tensor decomposition, comprising:

[0117] The feature extraction module is used to obtain the SMILES sequence of the drug and extract the molecular structure features of the drug based on the SMILES sequence to obtain the molecular structure feature representation of the drug.

[0118] The transformation module is used to construct the heterogeneity map of drug pairs in each cell line. Based on the heterogeneity map of drug pairs in each cell line, a three-channel heterogeneity map relation tensor is obtained, and the local interaction features of drugs are obtained through the heterogeneity map transformation network.

[0119] The decomposition module is used to perform Tucker decomposition on the heterogeneous graph relation tensor of the three channels, and concatenate the decomposition results with the local interaction features of the drug to extract the global interaction feature vector of the drug.

[0120] The prediction module is used to predict the synergistic score of the current drug-drug combination in the cell line by constructing a fusion feature vector mechanism through drug combination based on the molecular structural feature representation and global interaction feature vector of the drug.

[0121] All relevant content of each step involved in the aforementioned embodiments of the drug synergy prediction method based on heterogeneous graph tensor decomposition can be referenced to the functional description of the corresponding functional module of the drug synergy prediction system based on heterogeneous graph tensor decomposition in the embodiments of the present invention, and will not be repeated here.

[0122] In one embodiment of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the drug synergy prediction method based on heterogeneous graph tensor decomposition.

[0123] In one embodiment of the present invention, a computer-readable storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the drug synergy prediction method based on heterogeneous graph tensor decomposition described in the above embodiment.

[0124] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting drug synergy based on heterogeneous graph tensor decomposition, characterized in that, Includes the following steps: Obtain the SMILES sequence of the drug, and extract the molecular structural features of the drug based on the SMILES sequence to obtain the molecular structural feature representation of the drug. Construct a heterogeneous graph of drug pairs in each cell line. Based on the heterogeneous graph of drug pairs in each cell line, obtain a three-channel heterogeneous graph relation tensor. Then, obtain the local interaction features of drugs through a heterogeneous graph transformation network. Tucker decomposition was performed on the heterogeneous graph relation tensor of the three channels, and the decomposition results were concatenated with the local interaction features of the drug to extract the global interaction feature vector of the drug. Based on the molecular structure features of drugs and the global interaction feature vectors of drugs, a fusion feature vector mechanism is constructed through drug combination to predict the synergistic score of the current drug-drug combination in the cell line. The local interaction features of drugs are obtained through the following process: all drug data are integrated and a heterogeneous graph of drug combinations in each cell line is constructed. The heterogeneous graph relationship tensor of the three channels is processed by convolution transformation through a heterogeneous graph transformation network to extract metapath information. The metapath information is fused using a semantic-level attention mechanism to obtain the local interaction features of drugs. Specifically, the drug combination data will be used in the first... The relationships between all drug combinations in a cell line are distinguished according to three types of biological effects: synergistic, additive, and antagonistic. Three types of edge sets are constructed respectively, and these three types of edge sets are transformed into synergistic adjacency matrices, additive adjacency matrices, and antagonistic adjacency matrices, forming a three-channel heterogeneous graph relation tensor. A heterogeneous graph transformation network is used to perform convolution transformation on the three-channel heterogeneous graph relation tensor to obtain the intermediate adjacency relation matrix; The intermediate adjacency relation matrices are further computed to obtain the meta-path tensor; the heterogeneous graph transformation network is applied to each channel of all meta-path tensors, and the representations of multiple channels are concatenated to obtain the embedded representation; Based on the embedding representation, the weights of each channel are obtained by learning the heterogeneous graph relation tensor of the three channels through an attention mechanism. Based on the weight of each channel, the embeddings of each channel are aggregated by attention score weighting to obtain the interaction features of the drug.

2. The method for predicting drug synergy based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, The SMILES sequence of each drug is converted into a drug molecular structure diagram, and the drug molecular structure diagram is then used... Layered graph convolutional networks update node representations by aggregating the neighbor information of each atomic node; Then through The stacking operation of the heterogeneous graph transformation network is performed, and then the feature matrix of the last layer node of the heterogeneous graph transformation network is aggregated through global average pooling to obtain the molecular structure feature representation of the drug.

3. The method for predicting drug synergy based on heterogeneous graph tensor decomposition according to claim 2, characterized in that, Use of drug molecular structure diagrams Layered graph convolutional networks update node representations by aggregating the neighbor information of each atomic node, using the following formula: , in, This is a meta-path tensor that includes information about the nodes themselves. For is The degree matrix, For the first Feature matrix of layer nodes For the first Feature matrix of layer nodes For the initial representation, For activation function, is the learnable weight parameter matrix for linear transformation.

4. The method for predicting drug synergy based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, Drug molecular structure characteristics representation Calculated using the following formula: ， in, For the first The molecular structural characteristics of a drug are represented. This represents the total number of drugs.

5. The method for predicting drug synergy based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, Weight of each channel Calculated using the following formula: in, It is a function that splices along the channel. In cell lines The Middle The first drug The representation of each channel, It is the query vector in the attention mechanism. It is a learnable weight matrix. It is a learnable bias vector. For activation functions; Drug interaction characteristics Calculated using the following formula: in, The first characteristic of drug interaction in cell lines One channel, In the weight vector of each channel, the th One value, In the weight vector of each channel, the th Values.

6. A drug synergy prediction system based on heterogeneous graph tensor decomposition, characterized in that, include: The feature extraction module is used to obtain the SMILES sequence of the drug and extract the molecular structure features of the drug based on the SMILES sequence to obtain the molecular structure feature representation of the drug. The transformation module is used to construct the heterogeneity map of drug pairs in each cell line. Based on the heterogeneity map of drug pairs in each cell line, a three-channel heterogeneity map relation tensor is obtained, and the local interaction features of drugs are obtained through the heterogeneity map transformation network. The decomposition module is used to perform Tucker decomposition on the heterogeneous graph relation tensor of the three channels, and concatenate the decomposition results with the local interaction features of the drug to extract the global interaction feature vector of the drug. The prediction module is used to predict the synergistic score of the current drug-drug combination in the cell line by constructing a fusion feature vector mechanism through drug combination based on the molecular structure feature representation and global interaction feature vector of the drug. The local interaction features of drugs are obtained through the following process: all drug data are integrated and a heterogeneous graph of drug combinations in each cell line is constructed. The heterogeneous graph relationship tensor of the three channels is processed by convolution transformation through a heterogeneous graph transformation network to extract metapath information. The metapath information is fused using a semantic-level attention mechanism to obtain the local interaction features of drugs. Specifically, the drug combination data will be used in the first... The relationships between all drug combinations in a cell line are distinguished according to three types of biological effects: synergistic, additive, and antagonistic. Three types of edge sets are constructed respectively, and these three types of edge sets are transformed into synergistic adjacency matrices, additive adjacency matrices, and antagonistic adjacency matrices, forming a three-channel heterogeneous graph relation tensor. A heterogeneous graph transformation network is used to perform convolution transformation on the three-channel heterogeneous graph relation tensor to obtain the intermediate adjacency relation matrix; The intermediate adjacency relation matrices are further computed to obtain the meta-path tensor; the heterogeneous graph transformation network is applied to each channel of all meta-path tensors, and the representations of multiple channels are concatenated to obtain the embedded representation; Based on the embedding representation, the weights of each channel are obtained by learning the heterogeneous graph relation tensor of the three channels through an attention mechanism. Based on the weight of each channel, the embeddings of each channel are aggregated by attention score weighting to obtain the interaction features of the drug.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the drug synergy prediction method based on heterogeneous graph tensor decomposition as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the drug synergy prediction method based on heterogeneous graph tensor decomposition as described in any one of claims 1 to 5.

Images