Drug synergistic effect prediction method, system and equipment based on heterograph tensor decomposition and medium

Abstract

The invention discloses a drug synergistic effect prediction method, system and device based on heterograph tensor decomposition and a medium, and the prediction method comprises the steps: obtaining an SMILES sequence of a drug, extracting the molecular structure characteristics of the drug, and obtaining the molecular structure characteristic representation of the drug; constructing a drug pair heterograph in each cell line, and obtaining local interaction characteristics of drugs through a heterograph conversion network; performing Tucker decomposition on the heterogeneous graph relation tensor of the three channels, splicing a decomposition result with local interaction features of the medicine, and extracting a global interaction feature vector of the medicine; and according to the molecular structure feature representation of the drug and the global interaction feature vector of the drug, predicting the collaborative score of the current drug-drug combination in the cell line. According to the method, the SMILES sequence of the medicine and gene expression information of a cell line where the medicine is located are integrated, the SMILES sequence is converted into the medicine molecular structure diagram by utilizing the diagram convolutional network, so that the molecular structure characteristics of the medicine are extracted, and due to the design, TensoGraph can more accurately reflect the complexity and diversity of a medicine interaction network in the real world; and the prediction performance is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of drug reaction prediction, and particularly relates to a drug synergistic effect prediction method, system, device and medium based on heterogeneous graph tensor decomposition. BACKGROUND

[0002] With the continuous development of treatment methods for major diseases such as cancer, single drug treatment is facing clinical bottlenecks such as strong toxic side effects and high drug resistance. Multi-drug combination therapy can not only improve the treatment effect, but also reduce drug toxicity and drug resistance, so it has gradually become an important trend in cancer treatment. In order to effectively design drug combination schemes, the prediction of drug synergistic effect has become a key link in precision medicine and drug research and development.

[0003] Traditionally, the evaluation of drug synergistic effect relies on experimental methods such as high-throughput screening (HTS) technology. However, such experimental methods face significant limitations such as high cost, low efficiency, and long experimental period when facing the exponentially growing drug combination space and cancer heterogeneity. In recent years, with the rapid development of computing technology, researchers have begun to explore machine learning, deep learning, and even graph neural networks to model drug synergistic effect.

[0004] In the prior art, the 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 machine (SVM), random forest (RF), and gradient boosting decision tree (GBDT) can handle multi-modal features, but have limited ability to capture complex nonlinear relationships and feature interactions; deep learning methods such as feedforward neural network (FNN), convolutional neural network (CNN), and Transformer structure are more suitable for modeling high-dimensional complex relationships, and to some extent, improve the prediction accuracy, but often lack interpretability and integration of biological priori; graph neural network (GNN) is good at handling structured data and can model complex interactions between drugs and cell lines, so it has become an important direction for drug synergistic prediction in recent years.

[0005] Although graph neural networks provide a new solution paradigm for this problem, the current technology still has the following key deficiencies:

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

[0007] (2) Insufficient modeling ability of interaction relationships: There are complex nonlinear and multi-level interaction relationships between drugs-drugs and drugs-cell lines, and traditional feature concatenation or simple weighting cannot accurately model these relationships, affecting the prediction performance of the model.

[0008] (3) Model interpretability is poor: In clinical applications, the model output not only needs to be accurate, but also needs to explain its prediction mechanism. Many current models are black box structures, and lack of interpretable interaction mechanism design.

[0009] (4) Poor scalability and generalization ability: Some models rely on the coverage of drug combinations in the training set. Once encountering unseen combinations or new cell lines, the generalization ability is significantly reduced, limiting the application range of the model in practice. Therefore, how to design a drug synergy prediction model with stronger heterogeneous information fusion ability, more complex interaction modeling mechanism, while considering the interpretability and generalization performance, is still a difficulty and hotspot in current research. SUMMARY

[0010] In order to solve the technical problem that the existing technology often does not consider the influence of drugs on the cell line environment, and cannot depict the multi-relationship heterogeneity interaction among "drug-drug-cell line", resulting in inaccurate prediction of drug interaction, the purpose of the present application is to provide a drug synergy prediction method, system, device and medium based on heterogeneous graph tensor decomposition.

[0011] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0012] A drug synergy prediction method based on heterogeneous graph tensor decomposition, comprising the following steps:

[0013] Obtaining the SMILES sequence of the drug, and extracting the molecular structure features of the drug according to the SMILES sequence of the drug to obtain the molecular structure feature representation of the drug;

[0014] Constructing a drug pair heterogeneous graph in each cell line, obtaining a three-channel heterogeneous graph relationship tensor according to the drug pair heterogeneous graph in each cell line, and obtaining the local interaction features of the drug through a heterogeneous graph conversion network;

[0015] Tucker decomposition is performed on the three-channel heterogeneous graph relationship tensor, and the decomposition result is spliced with the local interaction features of the drug to extract the global interaction feature vector of the drug;

[0016] According to the molecular structure feature representation of the drug and the global interaction feature vector of the drug, a fusion feature vector mechanism is constructed through drug combination to predict the synergy score of the current drug-drug combination in the cell line.

[0017] Further, the SMILES sequence of each drug is converted into a drug molecular structure graph, and a L-layer graph convolution network is used to update the node representation by aggregating the neighbor information of each atom node;

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

[0019] Further, the L-layer graph convolution network is used to update the node representation by aggregating the neighbor information of each atom node, which is performed by the following formula:

[0020]

[0021] wherein, is the meta-path tensor including the information of the node itself, is the degree matrix of H i (l) is the feature matrix of the l-th layer node, i (l-1) is the feature matrix of the l-1-th layer node, i (0) is the initial representation, and ReLU is the activation function, i (l-1) is the learnable weight parameter matrix of the linear transformation.

[0022] Further, the molecular structure feature representation m i is calculated by the following formula:

[0023]

[0024] wherein, i is the molecular structure feature representation of the i-th drug, and n is the total number of drugs.

[0025] Further, the local interaction feature of the drug is obtained by the following process: integrating all drug data and constructing a drug combination heterogeneous graph in each cell line, performing convolutional transformation processing on the three-channel heterogeneous graph relationship tensor by the heterogeneous graph transformation network, extracting meta-path information, and fusing the meta-path information by the semantic-level attention mechanism to obtain the local interaction feature of the drug.

[0026] Further, the local interaction feature of the drug is obtained by the following process: distinguishing the relationships of all drug combinations under the r-th cell line according to the three biological effect types of synergistic effect, additive effect and antagonistic effect, respectively constructing three edge sets, and converting the three edge sets into synergistic adjacency matrix, additive adjacency matrix and antagonistic adjacency matrix respectively to form a three-channel heterogeneous graph relationship tensor.

[0027] The three-channel heterogeneous graph relationship tensor is processed by the heterogeneous graph transformation network to obtain an intermediate adjacency relationship matrix.

[0028] The intermediate adjacency relationship matrix is continuously calculated to obtain a 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 spliced to obtain an embedding representation;

[0029] According to the embedding representation, the weight of each channel of the three-channel heterogeneous graph relationship tensor is learned through an attention mechanism;

[0030] According to the weight of each channel, the channel embeddings are aggregated according to the attention score to obtain the interaction feature of the drug.

[0031] Further, the weight m r is calculated by the following formula:

[0032]

[0033] wherein ‖ c is a function of channel splicing, Z r|c|i is the representation of the ith drug in the cth channel in the cell line r, q is a query vector in the attention mechanism, is a learnable weight matrix, b r is a learnable bias vector, and tanh is an activation function.

[0034] The interaction feature Z r of the drug is calculated by the following formula:

[0035]

[0036] wherein Z r|c is the cth channel of the interaction feature of the drug in the cell line, m r|i is the ith value in the weight vector of each channel, and m r|j is the jth value in the weight vector of each channel.

[0037] A drug synergistic effect prediction system based on heterogeneous graph tensor decomposition, comprising:

[0038] A feature extraction module is configured to obtain a SMILES sequence of a drug, and extract a molecular structure feature of the drug according to the SMILES sequence of the drug to obtain a molecular structure feature representation of the drug.

[0039] A conversion module is configured to construct a drug pair heterogeneous graph in each cell line, obtain a three-channel heterogeneous graph relationship tensor according to the drug pair heterogeneous graph in each cell line, and obtain a local interaction feature of the drug through a heterogeneous graph conversion network.

[0040] a decomposition module, configured to perform Tucker decomposition on the three-channel heterogeneous graph relationship tensor, and splice the decomposition result with the local interaction feature of the drug to extract a global interaction feature vector of the drug;

[0041] a prediction module, configured to construct a fusion feature vector mechanism through drug combination according to the molecular structure feature representation of the drug and the global interaction feature vector of the drug, and predict a synergy score of the current drug-drug combination in the cell line.

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

[0043] A computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the drug synergy prediction method based on heterogeneous graph tensor decomposition.

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

[0045] The drug synergy prediction method based on heterogeneous graph tensor decomposition of the present application constructs a heterogeneous graph in a specific cell line from the perspective of a heterogeneous graph, and combines a graph transformation network (GTN) with tensor decomposition (Tucker decomposition) to capture the characteristics of drugs in a cell environment. The present application integrates the SMILES sequence of the drug and the gene expression information of the cell line where the drug is located, and also uses a graph convolution network (GCN) to convert the SMILES sequence into a drug molecular structure graph, thereby extracting the drug molecular structure features. Such 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.

[0046] The core of the present application is to first couple a 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. Introducing the cell line microenvironment. The present application explicitly models the cell line as the third dimension, constructs the relationship tensor of each drug in each cell line, models the heterogeneous graph relationship tensor of "drug-drug-cell line", and makes the synergy / additive / antagonistic three relationships change dynamically with the cell environment, which first solves the prediction distortion problem caused by "environmental independence", and overcomes the problem of ignoring the regulation of the cell line on the drug efficacy in the prior art, which regards the "drug-drug" (drug combination) relationship as a single and static binary variable.

[0048] 2. Global high-order tensor decomposition. The present application 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 capture fine-grained local structures.

[0049] 3. Local heterogeneous meta-path learning. The present application introduces a heterogeneous graph transformation network (GTN) to capture local topological differences in cell line-specific graphs through an automatic meta-path discovery mechanism, overcoming the problem of insufficient representation caused by "fixed paths" and the difficulty of existing graph neural networks to adapt to multiple relationship scenarios due to their reliance on manual meta-paths.

[0050] In summary, the present application solves the two major pain points of environmental heterogeneity and representation integrity in a unified framework through the system-level innovation of "explicit modeling of cell line dimension + global tensor decomposition + local meta-path learning + multi-modal fusion", significantly outperforms existing baselines in experimental accuracy, and has biological interpretability, thereby substantially overcoming the defects of inaccurate prediction, environmental irrelevance, and single representation in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition of the present application;

[0052] Figure 2 A flowchart of the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition of the present application;

[0053] Figure 3 A schematic diagram of the drug synergistic effect prediction system based on heterogeneous graph tensor decomposition of the present application. DETAILED DESCRIPTION

[0054] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0055] Referring to Figure 1 and Figure 2 , the present application provides a drug synergistic effect prediction method based on heterogeneous graph tensor decomposition (TensoGraph), which effectively solves the information heterogeneity problem in existing graph neural network methods, specifically comprising the following steps:

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

[0057] Firstly, the SMILES sequence of each drug is converted into a drug molecular structure graph, which is completed by means of the open source chemical information tool RDKit. Next, taking the ith drug as an example, the process of extracting the molecular structure features of the drug and obtaining the molecular structure feature representation of the drug according to the SMILES sequence of the drug is described in detail.

[0058] For the ith drug molecular structure graph can be constructed by , wherein is the set of all atomic nodes of the ith drug, is the set of all chemical bonds in the ith drug.

[0059] Each atomic node corresponds to a chemical atom (such as C, H, O, …). Each atom has its corresponding atomic node feature vector, which contains various attributes of the atom, such as element type, number of implicit hydrogens, number of valence electrons, bonding situation, charge state, and hybridization type. Therefore, the atomic node feature vector corresponding to each atom can be used as the initial feature of the drug. The initial feature matrix of the ith drug is composed of element type, number of implicit hydrogens, number of valence electrons, bonding situation, charge state, and hybridization type, wherein denotes an n i ×c real number matrix, n i is the number of atoms of the ith drug, and c is the dimension of the feature. Each row of the initial feature matrix of the ith drug represents a feature vector of an atom.

[0060] Each edge represents the chemical bond connection between two atoms, and the adjacency matrix is used to describe the chemical bond connection relationship, denotes an n i ×n i real number matrix. For example, if there is a chemical bond between the jth atom and the kth atom in the ith drug, then the adjacency matrix A 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] When the L-layer graph convolution network (GCN) is used to extract features of the drug molecular structure graph , the node representation is updated by aggregating the neighbor information of each atomic node. The update formula in the lth layer GCN is as follows:

[0062]

[0063] where, The adjacency matrix is augmented by adding the identity matrix I (self-loop) to incorporate the information of the nodes, i.e. The meta-path tensor including the information of the nodes themselves, is the degree matrix of , representing the sum of connections of each node (including self-loop). H i (l) represents the feature matrix of the l-th layer nodes, H i (l-1) represents the feature matrix of the l-1-th layer nodes, which is initially represented as H i (0) = X i . ReLU is an activation function, W i (l-1) is the learnable weight parameter matrix of the linear transformation of the layer.

[0064] Through the stacking operation of the L-layer GCN, the node features are aggregated from further distance neighbor information layer by layer, so as to effectively integrate the information of the whole molecular graph structure. After the completion of the graph convolution, in order to obtain the unified representation of the whole molecule, the feature matrix of the nodes in the last layer, i.e. the L-th layer of the GCN is aggregated by the global average pooling operation, and the molecular structure feature representation m i of the drug is obtained, and the calculation formula is:

[0065]

[0066] wherein m i is the molecular structure feature representation of the i-th drug, which comprehensively represents the molecular structure information of the drug, and n is the total number of drugs.

[0067] S2: Construct a drug combination heterogeneous graph in each cell line, obtain the local interaction feature of the drug through the heterogeneous graph transformation network, and then capture the local topological difference in the cell line specific graph.

[0068] Specifically, the present application integrates all drug data, and constructs a drug combination heterogeneous graph in each cell line, considers the heterogeneity of the information in the graph, extracts meta-path information through a heterogeneous graph transformation network (GTN), further fuses the meta-path information by a semantic level attention mechanism, and obtains a local interaction feature, so that the representation deficiency caused by "fixed path" can be overcome.

[0069] More specifically, taking the r-th cell line as an example, the relationships of the drug combination data of all drug combinations in the r-th cell line are distinguished according to three biological effect types of synergistic effect, additive effect and antagonistic effect, and three types of edge sets are constructed, which are denoted as a 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] where, and are the first, second, L-th composite meta-path tensor in cell line r, respectively. is the first intermediate adjacency matrix of the first layer GTN network is the c-th channel of is the second intermediate adjacency matrix of the first layer GTN network is the c-th channel of is the product of and is the normalized degree matrix of the product. is the c-th channel of the output of the second layer GTN network is the c-th channel of the first composite meta-path tensor is the normalized degree matrix of the c-th channel of the first composite meta-path tensor is the c-th channel of the first composite meta-path tensor is the c-th channel of the L-th layer GTN network output is the c-th channel of the L-1-st composite meta-path tensor is the c-th channel of the L-1-st composite meta-path tensor is the c-th channel of the L-1-st composite meta-path tensor is the normalized degree matrix of the c-th channel of the L-1-st composite meta-path tensor is the c-th channel of the L-1-st composite meta-path tensor is the c-th channel of the L-1-st composite meta-path tensor is the c-th channel of the L-1-st composite meta-path tensor

[0076] Finally, a GCN is applied to each channel of all meta-path tensors and the representations of multiple channels are concatenated, and the final embedding representation is:

[0077]

[0078] where, c represents the function of concatenating along the channel, the adjacency matrix is augmented by adding an identity matrix I (self-loop) to incorporate the information of the node itself is the normalized degree matrix of , i.e. is the meta-path tensor including the information of the node itself, W r ∈ R m×d 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.) is used to decompose the global interaction features across cell lines while preserving the high-order structure of the tensor, avoiding information collapse, and significantly improving the cross-domain generalization capability. The specific process is to decompose the three-channel heterogeneous graph relationship tensor Tucker decomposition is performed as follows:

[0086]

[0087] wherein, is the core tensor, and is the compressed low-rank tensor, capturing the potential high-order correlation between all modes, capturing the multi-dimensional potential mode of drug interaction (such as pathway combination). is the low-dimensional embedding of the drug as the initiator of the interaction, which is the global structure or functional feature of the drug. is the low-dimensional embedding of the drug as the receiver of the interaction, which reflects the response characteristics of the drug. is the low-dimensional feature of the interaction type (synergistic / additive / antagonistic). r1, r2, r3 represent the ranks in the above three modes (U r , V r , J r ) respectively, and × n represents the tensor-matrix product of mode n.

[0088] To predict the synergistic effect of drug combination, attention should be paid to the initiative action of the drug (i.e., which drugs tend to initiate synergistic effect). The global structure or functional feature of the drug encodes the characteristics of the drug as the initiator, which is directly related to the dominant role of the drug on the combination, so it is used as the global interaction feature of the drug. The mode matrix related to the first mode, i.e., the global structure or functional feature of the drug U r and the interaction feature of the drug in the cell line obtained in step S2 Z r are spliced together as the final interaction embedding of the drug in the cell line r, which is represented as follows:

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

[0090] wherein, H r is the global interaction feature vector of the drug, and the i-th row in the global interaction feature vector H r of the drug represents the interaction embedding H i,r of the drug i in the cell line r.

[0091] S4: obtaining a molecular structure feature representation m of the drug according to step S1 i and a global interaction feature vector H of the drug obtained in step S3 r , a fusion feature vector mechanism is constructed for drug combinations to predict the synergistic score of the current "drug-drug" combination in the cell line.

[0092] The present application fuses the global interaction feature vector, the local interaction feature of the drug, the molecular structure feature of the drug and the fingerprint feature of the drug end to end, forms an interpretable joint representation, and still maintains robust prediction under small sample and high noise conditions through multi-modal unified representation, and provides mechanism support for personalized combination therapy.

[0093] The object of the present application is to find the interaction relationship of "drug-drug" in the cell line, which can be written as "drug-drug-cell line" combination relationship and combination score. Therefore, the following takes the "drug-drug-cell line" combination of the i th drug and the j th drug in the cell line r as an example to predict their interaction.

[0094] In order to further improve the accuracy of drug combination synergistic effect prediction, a multi-input deep neural network prediction module (which uses four different MLP networks) is designed. The multi-input deep neural network prediction module adopts a modular extraction, dimension reduction and fusion of multiple features related to drug combinations, and finally unifies the prediction. The multi-input deep neural network prediction module receives four types of input features, including: the molecular structure feature representation (m i , m j ) of the i th drug and the j th drug obtained in step S1; the fingerprint feature (X i , X j ) of the i th drug and the j th drug, x i is the i th row in the fingerprint feature matrix X of all drugs, X j is the j th row in the fingerprint feature matrix X of all drugs; the interaction feature (H i,r , H j,r ) of the i th drug and the j th drug in the cell line r obtained in step S3; and the cell line r feature (C r ), C r is the gene expression matrix of the cell. Each type of feature is dimensionally reduced by a two-layer deep neural network prediction module with batch normalization, and all the dimensionally reduced features are spliced to predict the synergistic score of the drug combination

[0095]

[0096] Wherein, m′ i= MLP1(m i ), m' j = MLP1(m j ), m' i is the i-th drug molecule structure feature m i is the new representation after passing through the first MLP network, m' j is the j-th drug molecule structure feature m j is the new representation after passing through the first MLP network. X' i = MLP2(X i ), X' j = MLP2(X j ), X' i is the i-th drug fingerprint feature X i is the new representation after passing through the second MLP network, X' j is the j-th drug fingerprint feature X j is the new representation after passing through the second MLP network. H' i,r = MLP3(H i,r ), H' j,r = MLP3(H j,r ), H' i,r is the i-th drug-drug-cell line interaction feature H i,r is the new representation after passing through the third MLP network, H' j,r is the j-th drug-drug-cell line interaction feature H h,r is the new representation after passing through the third MLP network. C' r = MLP4(C r ), C' t is the gene expression matrix of cell r C r is the new representation after passing through the fourth MLP network.

[0097] To evaluate and optimize the deviation between the predicted value and the true value, the multi-input deep neural network prediction module adopts the mean square error as the loss function is defined as follows:

[0098]

[0099] where y t is the true score of the t-th "drug-drug-cell line" group, is the score predicted by the model corresponding to the "drug-drug-cell line" combination, denotes the total number of drug combinations, and r is the total number of cell lines, is a permutation combination calculation, and t represents the t-th "drug-drug-cell line" combination.

[0100] The heterogenous graph tensor decomposition method will be described in detail below through specific embodiments.

[0101] The specific embodiments of the present application use the O'neil dataset on the Loewe score to verify the drug synergy prediction method based on the heterogenous graph tensor decomposition described in the present application. The O'neil dataset includes comprehensive data of 39 cell lines, 38 drugs and 22737 drug combinations, of which there are 1973 drug combination data with synergistic effect, 12087 drug combination data with additive effect and 8677 drug combination data with antagonistic effect. The specific embodiments of the present application divide the dataset into 10 parts, take 8 parts as the training set, 1 part as the validation set and 1 part as the test set, so as to perform ten-fold cross-validation. Based on the above data, the network is trained using the Adam optimizer with a learning rate of 1*10-4, and the synergistic score of the drug combination in a specific cell line is predicted. According to the synergistic score, it is determined whether the drug has synergistic effect or antagonistic effect.

[0102] The specific embodiments of the present application 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 removes the part of the GCN extracting the molecular structure features of the drug on the basis of the TensoGraph of the present application, and only retains the drug interaction features, drug fingerprint features and cell line features for drug prediction.

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

[0105] Specifically, the TensoGraph-GCN-GTN model removes the part of the GCN extracting the molecular structure features of the drug and removes the local interaction features generated by the GTN on the basis of the TensoGraph of the present application, and only uses the global interaction features of the drug, the Infomax fingerprint features and the cell line features for prediction.

[0106] Specifically, the TensoGraph-GCN-Tucker-GTN model removes all interaction features obtained by GTN and Tucker decomposition on the basis of the TensoGraph of the present application, and only retains drug molecular structure features, Infomax fingerprint features and cell line features for prediction.

[0107] Specifically, the TensoGraph-GCN-Tucker-GTN model removes all interaction features obtained by GTN and Tucker decomposition on the basis of the TensoGraph of the present application, and only retains drug molecular structure features, Infomax fingerprint features and cell line features for prediction.

[0108] Further, 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 can be regarded as ablation experiments of the TensoGraph model.

[0109] In a specific embodiment of the present application, 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 on the basis of the O'neil data set.

[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 experiment of O'Neil on Loewe score

[0112]

[0113]

[0114] From the experimental results of Table 1, it can be seen that the TensoGraph model improves 2.6% in the MSE index and 1.3% in the RMSE index compared with the TensoGraph-GCN model removing the molecular structure features of the drug; improves 2.4% in the MSE index and 1.1% in the RMSE index compared with the TensoGraph-GCN-Tucker model removing the molecular structure features of the drug and global interaction features; improves 4.3% in the MSE index and 2.2% in the RMSE index compared with the TensoGraph-GCN-GTN model removing the molecular structure features of the drug and local interaction features; improves 7.7% in the MSE index and 3.9% in the RMSE index compared with the TensoGraph-Tucker-GTN model removing the interaction features between drugs; improves 10.2% in the MSE index and 5.2% in the RMSE index compared with the TensoGraph-GCN-Tucker-GTN model removing the molecular structure features of the drug and the interaction features between drugs. The 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] In view of the heterogeneity of information in the drug interaction network, the application proposes a drug synergy model based on heterogeneous graph tensor decomposition, which integrates the SMILES sequence of the drug, the Infomax fingerprint feature and the gene expression data. First, the SMILES sequence of the drug is converted into a drug molecular graph, and a graph neural network is used to extract the molecular structure features of the drug. Then, the heterogeneous graph constructed in the cell line is transformed through a graph transformation network and Tucker decomposition to extract the drug interaction features. Finally, a deep neural network is used to minimize the prediction error to predict the synergy score, thereby effectively fusing the relationship between drug combinations and cell lines, improving the prediction accuracy of drug synergy, and accelerating the process of drug research and development, which lays a solid foundation for its wide application in the field of drug synergy prediction.

[0116] Referring to Figure 3 The drug synergy prediction system based on heterogeneous graph tensor decomposition provided by the application comprises:

[0117] The feature extraction module is configured to obtain the SMILES sequence of the drug, and extract the molecular structure features of the drug according to the SMILES sequence of the drug, to obtain the molecular structure feature representation of the drug.

[0118] The conversion module is configured to construct a drug pair heterogeneous graph in each cell line, obtain a three-channel heterogeneous graph relationship tensor according to the drug pair heterogeneous graph in each cell line, and obtain the local interaction features of the drug through a heterogeneous graph conversion network.

[0119] a decomposition module configured to perform Tucker decomposition on the three-channel heterogeneous graph relational tensor, and splice the decomposition result with the local interaction feature of the drug to extract a global interaction feature vector of the drug;

[0120] a prediction module configured to construct a fusion feature vector mechanism through drug combination according to the molecular structure feature representation of the drug and the global interaction feature vector of the drug, and predict a synergistic score of the current drug-drug combination in the cell line.

[0121] The foregoing embodiments of the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition involve all relevant contents of the steps, which can be cited as the functional description of the corresponding functional modules of the drug synergistic effect prediction system based on heterogeneous graph tensor decomposition in the embodiments of the present application, and will not be repeated here.

[0122] In an embodiment of the present application, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition.

[0123] In an embodiment of the present application, a computer readable storage medium is provided, specifically a computer readable storage medium (Memory), which is a memory device in a computer device and is used to store programs and data. It can be understood that the computer readable storage medium herein can include an internal storage medium in a computer device, and of course can also include an extended storage medium supported by the computer device. The computer readable storage medium provides a storage space, which stores an operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer readable storage medium to implement the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition in the above embodiments.

[0124] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0125] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart 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, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0126] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0128] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing the technical solutions of the present application, but not for limiting it. Although the present application is described in detail with reference to the above embodiments, those skilled in the field should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the claims of the present application.

Claims

1. A method for predicting drug synergy based on heterogeneous graph tensor decomposition, characterized in that, The method comprises the following steps: Obtaining the SMILES sequence of the drug, and extracting the molecular structure features of the drug according to the SMILES sequence of the drug to obtain the molecular structure feature representation of the drug; A drug pair heterograph is constructed in each cell line, and a three-channel heterograph relationship tensor is obtained according to the drug pair heterograph in each cell line, and a local interaction feature of the drug is obtained through a heterograph conversion network; Tucker decomposition is performed on the three-channel heterograph relationship tensor, and the decomposition result is spliced with the local interaction feature of the drug to extract a global interaction feature vector of the drug; According to the molecular structure feature representation of the drug and the global interaction feature vector of the drug, 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. 2.The drug synergy prediction method based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, Each SMILES sequence of the drug is converted into a drug molecular structure graph, and a L-layer graph convolution network is used to update the node representation by aggregating the neighbor information of each atomic node; Then, through the stacking operation of the L-layer heterograph transformation network, the feature matrix of the last layer node of the heterograph transformation network is aggregated through the global average pooling operation to obtain the molecular structure feature representation of the drug. 3.The drug synergy prediction method based on heterogeneous graph tensor decomposition according to claim 2, characterized in that, The L-layer graph convolution network is used to update the node representation by aggregating the neighbor information of each atomic node, which is performed by the following formula: wherein, is an ego-path tensor including information of the node itself, is a degree matrix of i (l) is a feature matrix of the l-th layer node, i (l-1) is a feature matrix of the l-1-th layer node, i (0) is an initial representation, ReLU is an activation function, i (l -1) is a learnable weight parameter matrix of linear transformation.​ 4.The drug synergy prediction method based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, molecular structure feature representation of a drug i calculated by the formula: where m i is the molecular structure feature representation of the ith drug, and n is the total number of drugs. 5.The drug synergy prediction method based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, The local interaction feature of the drug is obtained by the following process: integrating all drug data and constructing a drug combination heterograph in each cell line, performing convolutional transformation processing on the three-channel heterograph relationship tensor through a heterograph conversion network, fusing the meta-path information through a semantic-level attention mechanism, and obtaining the local interaction feature of the drug. 6.The drug synergy prediction method based on heterogeneous graph tensor decomposition according to claim 1, characterized in that, The local interaction feature of the drug is obtained by the following process: distinguishing the relationships of all drug combinations under the rth cell line according to three biological effect types of synergistic effect, additive effect and antagonistic effect, respectively constructing three types of edge sets, and converting the three types of edge sets into synergistic adjacency matrix, additive adjacency matrix and antagonistic adjacency matrix to form a three-channel heterograph relationship tensor; The three-channel heterograph relationship tensor is processed through convolutional transformation by a heterograph transformation network to obtain an intermediate adjacency relationship matrix; The intermediate adjacency relationship matrix is further calculated to obtain a meta-path tensor; the heterograph transformation network is applied to each channel of all meta-path tensors, and the representations of multiple channels are spliced to obtain an embedding representation; According to the embedding representation, the three-channel heterograph relationship tensor is learned through an attention mechanism to obtain the weight of each channel; According to the weight of each channel, the channel embeddings are aggregated according to the attention score to obtain the interaction feature of the drug. 7.The drug synergy prediction method based on heterogeneous graph tensor decomposition according to claim 6, characterized in that, the weight m of each channel r is calculated by the formula: where ‖ c is a function of channel concatenation, Z r|c|i is the representation of the c-th channel of the i-th drug in the cell line r, q is the query vector in the attention mechanism, is a learnable weight matrix, b r is a learnable bias vector, and tanh is the activation function; Drug interaction profile Z r is calculated by the formula: where Z r|c is the cth channel of the interaction profile of the drug in the cell line, m r|i is the ith value in the weight vector for each channel, m r|j is the jth value in the weight vector for each channel. 8.A drug synergy prediction system based on heterogeneous graph tensor decomposition, characterized in that, It comprises: A feature extraction module is used to obtain the SMILES sequence of the drug, and extract the molecular structure features of the drug according to the SMILES sequence of the drug to obtain the molecular structure feature representation of the drug; A conversion module is used to construct a drug pair heterograph in each cell line, obtain a three-channel heterograph relationship tensor according to the drug pair heterograph in each cell line, and obtain a local interaction feature of the drug through a heterograph conversion network; The decomposition module is used for performing Tucker decomposition on the three-channel heterogeneous graph relation tensor, and splicing the decomposition result with the local interaction feature of the drug to extract a global interaction feature vector of the drug; The prediction module is used for constructing a fusion feature vector mechanism through drug combination according to the molecular structure feature representation of the drug and the global interaction feature vector of the drug, and predicting a synergistic score of the current drug-drug combination in the cell line.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the drug synergistic effect prediction method based on heterogeneous graph tensor decomposition as claimed in any one of claims 1 to 7.

Images