Cooperative drug prediction method and system based on heterogeneous hypergraph

Abstract

The invention provides a collaborative drug prediction method and system based on a heterogeneous hypergraph, and relates to the field of collaborative drug prediction.The collaborative drug prediction method comprises the steps that information of various cells and structures of various drugs are obtained, and corresponding cell fusion features and drug structure features are obtained; constructing a heterogeneous hypergraph based on the drug structure features and the cell fusion features, performing convolution processing to obtain drug node features, cell node features and hyperedge features, and training a neural network by using the extracted features to obtain a prediction model; and obtaining a to-be-detected drug, and inputting the to-be-detected drug into the prediction model to obtain a prediction result. According to the method, the heterogeneous hypergraph is constructed based on the collaborative information of the known drug combination in the cells, and the feature information of the drug nodes, the cell nodes and the hyperedges is extracted by adopting the difference of the multilayer hypergraph convolutional layers, so that the defects of excessive smoothness and redundancy of the information during feature extraction of the heterogeneous network by a common graph neural network are overcome; and a complex relation mode is difficult to capture.

Description

Technical Field

[0001] The present invention relates to the field of collaborative drug prediction, and in particular, to a collaborative drug prediction method and system based on heterogeneous hypergraphs. Background Art

[0002] At present, drug development mainly follows the traditional concept of "one drug, one target". However, for some complex diseases such as cancer and Alzheimer's disease, which involve multiple target genes, it is difficult to achieve the best therapeutic effect with only one drug. In this case, it may be more appropriate to design a combination therapy of drugs for complex diseases. So far, some computational models for predicting collaborative drugs have been proposed, which can be divided into two categories. The first type of method directly uses the characteristic information such as the structure, target, enzyme, and pathway of the drug itself to predict collaborative drugs. The second type of method comprehensively considers the characteristic information of both drugs and cells to predict on which specific cells the drug combination can exhibit a synergistic effect. However, after extracting the drug characteristics and cell characteristics, the existing methods directly use the extracted drug characteristics and cell characteristics to predict collaborative drugs, without considering the interaction between drug characteristics and cell characteristics, which may lead to information loss and misleading conclusions, and the prediction accuracy is relatively low. Summary of the Invention

[0003] The purpose of the present invention is to provide a collaborative drug prediction method and system based on heterogeneous hypergraphs to improve the above problems. To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0004] In a first aspect, the present application provides a collaborative drug prediction method based on heterogeneous hypergraphs, including:

[0005] Obtaining information of multiple cells and the structures of multiple drugs;

[0006] Obtaining corresponding cell fusion features according to the information of multiple cells, and obtaining corresponding drug structure features according to the structures of multiple drugs;

[0007] Constructing a heterogeneous hypergraph based on the drug structure features and cell fusion features, and performing convolutional processing to obtain drug node features, cell node features, and hyperedge features, and training a neural network with the extracted features to obtain a prediction model;

[0008] Obtaining a drug to be tested, and inputting the drug to be tested into the prediction model to obtain a prediction result.

[0009] In a second aspect, the present application further provides a collaborative drug prediction system based on heterogeneous hypergraphs, including:

[0010] A first module for obtaining information of multiple cells and the structures of multiple drugs;

[0011] The second module is used to obtain corresponding cell fusion features according to the information of multiple cells and obtain corresponding drug structure features according to the structures of multiple drugs;

[0012] The third module is used to construct a heterogeneous hypergraph based on the drug structure features and cell fusion features, perform convolutional processing to obtain drug node features, cell node features and hyperedge features, and train a neural network with the extracted features to obtain a prediction model;

[0013] The fourth module is used to obtain a drug to be tested, input the drug to be tested into the prediction model, and obtain a prediction result.

[0014] The beneficial effects of the present invention are as follows:

[0015] Based on the synergistic information of known drug combinations in cells, the present invention constructs a heterogeneous hypergraph, and uses a multi-layer hypergraph convolutional layer to differentially extract the feature information of drug nodes, cell nodes and hyperedges, and constructs the connection between drugs, cells and synergistic effects, overcoming the problems of over-smoothing and redundancy of information existing in the feature extraction of heterogeneous networks by ordinary graph neural networks, and it is difficult to capture complex relationship patterns.

[0016] Other features and advantages of the present invention will be described in the subsequent specification, and part of them will become obvious from the specification, or be understood by implementing the embodiments of the present invention. Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a flowchart of a synergistic drug prediction method based on a heterogeneous hypergraph according to an embodiment of the present application;

[0019] Figure 2 It is a structural diagram of a synergistic drug prediction system based on a heterogeneous hypergraph according to an embodiment of the present application.

[0020] Markings in the figure: 100 - the first module; 200 - the second module; 210 - the first unit; 220 - the second unit; 221 - the first subunit; 222 - the second subunit; 223 - the second subunit; 223 - the third subunit; 230 - the third unit; 231 - the fourth subunit; 232 - the fifth subunit; 233 - the sixth subunit; 234 - the seventh subunit; 235 - the eighth subunit; 300 - the third module; 400 - the fourth module. Detailed Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and illustrated in the accompanying drawings herein can be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0022] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not require further definition and explanation in subsequent drawings.

[0023] Existing methods often use graph neural networks to extract the structural features of drugs from the two-dimensional molecular graphs of drugs when extracting drug features. The feature transfer between nodes depends on the edges connecting the nodes (i.e., chemical bonds in the original molecule). However, in drug molecular graphs, some nodes may be connected to each other through multiple paths, which means that the features of a single node may be repeatedly represented on other nodes after being aggregated multiple times. This multiple-path problem can lead to the repeated transmission and superposition of the same information at different levels and paths, resulting in redundancy of drug structural features, thereby affecting the capture of key structural features of drugs by the model.

[0024] After existing methods extract the multi-source feature information of cells, they often use methods such as splicing or average weighting to fuse the multi-source cell information. However, such an approach does not consider the correlation relationships between the multi-source information. Taking the three types of information of gene expression, copy number aberration, and somatic mutation of cells as an example, they not only reflect different states and functions of cells respectively, but also have complex mutual influences among them. For example, the gene expression level may be affected by somatic mutations, and copy number aberrations may in turn affect gene expression. The correlation relationships between these multi-source information are extremely complex, and direct splicing or average weighting obviously cannot capture these deep-level interaction patterns.

[0025] Existing methods generally construct a heterogeneous network based on the synergistic information of known drug combinations in cells and use graph neural networks to extract features from the heterogeneous network. However, in a heterogeneous network, the types of nodes are diverse, each type of node may have different features, and different types of node pairs may also have different potential associations. Since graph neural networks assume that all nodes and edges have the same feature space and relationship type, this may lead to problems when graph neural networks handle this heterogeneity. In addition, there may be multiple paths in the heterogeneous network, and some nodes may be connected to each other through multiple types of edges. When graph neural networks handle this multiple-path problem, it may lead to over-smoothing and redundancy of information, making it difficult to capture complex relationship patterns.

[0026] The discovery that similar drug pairs have similar inhibitory effects on similar cell lines is an important breakthrough in the field of drug research, which reveals the close relationship between drugs and cells. After extracting drug features and cell features, existing methods directly use the extracted drug features and cell features to predict synergistic drugs, without considering the interaction between drug features and cell line features. However, the interaction between drugs and cell lines is complex and diverse. If these features are not considered simultaneously but only independent feature extraction is performed, it may lead to loss of information and misleading conclusions, and it is also impossible to comprehensively understand the mechanism of drug action and accurately predict the drug effect.

[0027] Example 1

[0028] See Figure 1 , this example provides a method for predicting synergistic drugs based on a heterogeneous hypergraph, including steps S100, S200, S300, and S400;

[0029] S100. Obtain information on multiple cells and the structures of multiple drugs

[0030] The information on the cells includes the gene expression profiles, copy number abnormalities, and somatic mutation data of the cells; the structures of the drugs include the SMILES structures of the drugs.

[0031] S200. Obtain corresponding cell fusion features based on the information of multiple cells and corresponding drug structure features based on the structures of multiple drugs, specifically including:

[0032] S210. Construct molecular graphs respectively according to the SMILES of each drug, process the molecular graphs using a message passing neural network based on a fusion attention mechanism, and update the node features using edge features to obtain all sub-structure features in each molecular graph;

[0033] S211. Generate atomic initial vectors according to the atoms in the drugs and edge vectors according to the chemical bonds:

[0034] First, construct a drug molecular graph. Using the SMILES of the drug as input, RDKit is used to convert the SMILES into a two-dimensional undirected graph G=(V, E) of the drug molecule. Here, V is the node set of graph G, that is, the atoms in the drug molecule, and E is the edge set of graph G, representing the chemical bonds in the drug molecule.

[0035] Generate the initial atomic vectors according to the atoms in the drug and generate edge vectors according to the chemical bonds;

[0036] For the initial vectors of each atom in graph G, it is necessary to use RDKit to extract the atomic symbol (dimension 118, using one-hot encoding), atomic degree (dimension 10, using one-hot encoding), implicit valence electrons (dimension 7, using one-hot encoding), formal charge (dimension 1, using numerical encoding), number of radical electrons (dimension 1, using numerical encoding), hybridization state (dimension 5, using one-hot encoding), and aromaticity (dimension 1, using numerical encoding) of each atom in the drug molecule. Concatenate the above seven atomic features to obtain an atomic vector x with a dimension of 143 i ; For the vectors of each edge in graph G, directly perform one-hot encoding according to the type of chemical bond. Since there are only four types of chemical bonds, namely single bond, double bond, triple bond, and aromatic bond, the dimension of the initial feature vector of the edge is 4. For the i-th atom v i and the j-th atom v j in graph G, denote the edge between them as e ij , and its feature vector is x ij . Also, because the drug molecular graph is an undirected graph, e ij and e ji are actually the same.

[0037] In this step, the feature vectors of the atoms can be initialized first. For the feature vector of the i-th atom, the present invention considers performing a linear transformation on the original atomic vector x i using a multi-layer perceptron (MLP) to extract better node representations, as follows:

[0038] h i = MLP(x i )

[0039] where MLP(·) represents the MLP processing layer, and h i is the processed result, which is used as the final initial atomic vector.

[0040] S212. Process the initial atomic vectors of two connected atoms and the edge vector between them through a multi-layer perceptron to obtain a directed edge vector:

[0041] Assume that node v i is the core node of a sub-structure. Then it can be considered that starting from the remaining nodes v j in the current sub-structure, the core node v i will be reached eventually. Regarding v i as the end of the path formed during the process of starting from node v j and reaching node v i , with the starting point of this path being v j and the ending point being v i , then this path is a directed path. Considering different nodes in the same sub-structure as core nodes and converting the molecular graph into a directed graph, the edges e ij and e ji become two independent directed edges. To more clearly highlight the difference between them, rename them as e i→j (representing the edge from node v i to node v j ) and e j→i (representing the edge from node v j to node v i ). Therefore, the v i on the path to v j can be weighted by the weight value of the edge e i connected to v j→i along the path. If there are multiple paths from v j to v i , then each path is considered separately. According to the method of weighting node features using edge weights proposed here, sub-structures of different sizes and shapes will be generated eventually (because when the weight value of the edge is approximately equal to 0, the nodes on the rest of the path will be cut off, and thus, the sub-structure will be divided into an irregular shape). Considering each node in graph G as the core node of a certain sub-structure in turn, a learnable weight is assigned to each directed edge, restricted within the range of [0,1], so as to extract as many sub-structures as possible from graph G.

[0042] For the initial directed edge vector j→i of edge e[[ID= fifty-two]] is also initialized using a multi-layer perceptron as follows:

[0043] where x ji represents the initial feature vector of edge e j→i ; τ is a constant to avoid gradient flow saturation when using the Sigmoid function; σ(·) represents the Sigmoid function; || represents the concatenation operation. h i is the atom v iThe initial vector P (of the original xji atom) MLP (h||vector hj,)) h j is the atom v j The initial vector of the atom;

[0044] It should be noted here that since the molecular graph is converted into a directed graph, and Need to be processed by the MLP processing layer respectively.

[0045] S213. Input the atomic initial vector and the directed edge vector into the message passing neural network to obtain the directed edge feature;

[0046] Traditional MPNNs use three parts: message passing, message aggregation, and message update to extract node features in graph-structured data. In each iteration of the MPNN, a node will receive messages from its neighbor nodes and update its own features accordingly. This means that after l iterations, a node will aggregate the features of all nodes within a path length of l. This is similar to the idea of aggregating node features along the path during the substructure extraction process described earlier in the present invention. However, this approach of MPNN will result in feature redundancy in the structural features extracted from the two-dimensional undirected graph of drug molecules during the process of aggregating node features. To avoid feature redundancy, the present invention does not update the features of the nodes, but updates the features of the edges between the nodes, which can well avoid this problem because a node will only appear once in a path of the directed graph. Therefore, the present invention considers adopting an edge feature update strategy on the basis of the traditional MPNN. Correspondingly, the improved MPNN also consists of three parts: message passing, message aggregation, and message update, as follows:

[0047]

[0048] Among them, h j and h k respectively represent the atomic initial vectors of node v j and node v k ; represents the message passed from edge e k→j to edge e j→i in the l-th layer; represents the directed edge vector of edge e j→i in the (l - 1)-th layer; represents the feature aggregated from the adjacent edges of e j with node v j→i as the intermediate node in the l-th layer; represents the updated edge e j→iDirected edge vector; Aggravation (l) (·) represents the message aggregation function; N(i) represents the set of neighbor nodes of node vi; Update (l) (·) represents the message update function; to avoid feature redundancy caused by loop paths, the present invention will remove node v from the set of neighbor nodes N(j) of node v j 。 i 。

[0049] Take the output of the last layer (L layer) As the directed edge feature.

[0050] S214. Process the directed edge features using the attention mechanism to obtain the attention scores between the directed edge features;

[0051] Use the multi-head attention mechanism to update the node features. First, calculate the query vector q , key vector k j→i , and value vector v j→i for each directed edge feature, and the calculation method is as follows:

[0052]

[0053] Among them, W Q , W K , and W V all represent linear transformation matrices, which are randomly initialized and updated through the loss function. Calculate the attention scores between each directed edge feature. Here, taking the directed edge features j→i of edge e u→j and edge e (j→i)(u→j) as an example, the attention score is calculated as follows:

[0054] socre j→i =q u→j ·k j→i

[0055] Among them, q u→j represents the query vector of the directed edge feature ; k j→i represents the key vector of the directed edge feature . After calculating the attention scores between each feature vector, it is necessary to normalize the attention scores, and the calculation is as follows:

[0056]

[0057] <00 industrie000320>Among them, softmax(·) represents the softmax function; exp(·) represents the inverse function of the natural logarithm function. < industrie000320>

[0058] S215. Update the directed edge features based on the attention scores to obtain updated edge features, and update the atomic initial vectors based on the updated edge features to obtain substructure features.

[0059] According to the normalization result, the updated edge features are obtained as:

[0060]

[0061] where w′ j→i is the updated edge feature, and v u→j is the value vector of.

[0062] Finally, after updating the edge features, update the atomic initial vectors of the nodes:

[0063]

[0064] where n i is the substructure feature, which can be regarded as the feature vector extracted with the atom v i as the core node; h i is the atomic initial vector of the atom v i .

[0065] S220. Cluster the substructure features in all molecular graphs to obtain the centroid information of the clusters; according to the substructure features and the centroid information of the clusters where they are located, obtain the drug structure features, specifically as follows:

[0066] S221. Randomly select multiple substructure features as the initial centroids, calculate the distance between each substructure feature and each initial centroid, assign it to the cluster where the nearest centroid is located, and update the initial centroids until the centroid positions remain unchanged to obtain the final centroids;

[0067] Based on the description of the structure-activity relationship, drugs with similar structures often have similar chemical properties, and the chemical properties of drugs are largely determined by their substructures. Therefore, the present invention considers performing k-means clustering operations on all substructure feature vectors of all drug molecules, and setting the number of clusters to K * ;

[0068] First, randomly select K * data points as the initial centroids For each substructure feature n i , calculate its distance from each centroid, and assign it to the cluster where the nearest centroid is located, and then update the centroid until the centroid position no longer changes significantly. Let the K * final centroids after clustering be Each final centroid is related to the substructure feature vector ni For vectors with the same dimension, the final centroid contains the common features of all substructure features in the cluster.

[0069] S222. Weight the final centroid according to the number of substructures of the current drug in each cluster, and sum the weighted centroids of all clusters to obtain the key feature vector.

[0070] Assume that the number of atoms of a drug is N atom , then after being processed by the improved MPNN, it can obtain N atom substructure features. Now, since we know the cluster to which each substructure feature belongs, therefore, we can distinguish the importance of different substructure features according to the clusters to which these N atom substructure features belong. For example, for the N atom substructure features of the current drug, among which m1 belong to the first cluster, m2 belong to the second cluster, and so on, there are m U belonging to the U-th cluster. Then the key feature vector of the current drug can be expressed as:

[0071]

[0072] where K * is the number of clusters, N atom is the number of atoms, m U is the number of substructure features in the U-th cluster; μ′ U is the feature representation of the final centroid of the U-th cluster.

[0073] S223. Concatenate the key feature vector and the substructure features of the drug and input them into a multi-layer perceptron for processing to obtain the drug structure features.

[0074]

[0075] where n1, n2,..., n ωI represent the substructure features of the current I-th drug, ω I represents the number of substructure features of the I-th drug; d I is the drug structure feature of the I-th drug.

[0076] S230. Construct an enrichment matrix of features according to the cell information and perform feature fusion to obtain the cell fusion feature, specifically as follows:

[0077] S231. Obtain at least one Gaussian regularization vector of the cell based on the cell information corresponding to the cell:

[0078] Cell information includes the gene expression profile, copy number abnormality, and somatic mutation data of cells. The Gaussian regularization vectors of the gene expression profile, copy number abnormality, and somatic mutation of cells are calculated as follows:

[0079]

[0080]

[0081] Among them, gene J 、copy J and var J respectively represent the column vectors of the expression values of the gene expression profile, copy number abnormality, and somatic mutation of the Jth cell among all cells. and and and respectively represent the average value and standard deviation of the expression abundances of the gene expression profile, copy number abnormality, and somatic mutation of the Jth cell among all cells (expression abundance refers to the expression level of a specific gene or protein in cells, tissues, or biological samples in biology, and the expression abundance reflects the degree of activity of a specific gene or protein in an organism).

[0082] S232. Perform dimensional normalization on the Gaussian regularization vectors to obtain the corresponding cell feature vectors; in this embodiment, there are three Gaussian regularization vectors, so three cell feature vectors with the same dimension are obtained, as follows:

[0083] Use MLP to perform feature dimensional normalization respectively to obtain cell feature vectors with a dimension of F and

[0084] S233. Stack the feature vectors of each cell feature vector to obtain the enrichment square matrix corresponding to the cell feature vector;

[0085] Stack the feature vectors of F gene′ J 、F copy′ J and F var J ' respectively to construct the enrichment square matrix S gene 、S copy and S var ;

[0086] S234. Use the matrix iteration fusion method to perform feature fusion on each enrichment square matrix, and then perform weighted fusion on the feature fusion results of all enrichment square matrices to obtain the final fusion result; as follows:

[0087]

[0088] Among them, λ gene , λ copy and λ var are weight decay parameters, all set to 0.25; D gene , D copy and D var are respectively the degree matrices of S gene , S copy and S var ; E′ is the identity matrix; are respectively the feature fusion results of the three enrichment square matrices;

[0089] Perform weighted fusion as follows:

[0090]

[0091] Among them, A is a diagonal matrix, and the q-th diagonal element of A is the q-th diagonal element of. Since the diagonal elements of are not necessarily all 1, so A -1 is used to normalize it.

[0092] S235. Take the average of the row vectors in the final fusion result and process it to obtain the cell fusion feature of this cell;

[0093] That is, take the average of each row vector of the matrix to obtain the fusion feature vector c J of the J-th cell.

[0094] S300. Construct a heterogeneous hypergraph based on the drug structure features and cell fusion features, and perform convolution processing to obtain drug node features, cell node features, and hyperedge features. Use the extracted features to train a neural network to obtain a prediction model;

[0095] Drug co-occurrence samples are represented by quadruples of drug-drug-cell-synergy labels, which describe whether two drugs exhibit a synergistic effect on a cell. A traditional graph consists of nodes and edges, where an edge connects two nodes, representing the relationship between the nodes. In contrast, a hypergraph introduces the concept of a hyperedge, which can connect multiple nodes, representing the relationship between multiple nodes. A heterogeneous hypergraph is an extension of a conventional hypergraph, where nodes and hyperedges can have different types or labels. In a conventional hypergraph, nodes are connected by hyperedges, and each hyperedge can connect any number of nodes. However, in a heterogeneous hypergraph, nodes and hyperedges can have additional attributes or labels, thus being able to represent more complex relationships. A heterogeneous hypergraph can be defined as G H ={V H , E H}, where, V H={v1, v2,..., v M} represents the set of nodes of the heterogeneous hypergraph, and E H ={e1, e2,..., e N} represents the set of hyperedges of the heterogeneous hypergraph.

[0096] S310. Obtain the hyperedge weight matrix weighted by drug nodes and cell nodes, and obtain the incidence matrix of the hypergraph;

[0097] The hypergraph can be represented by the incidence matrix H, and its representation is as follows:

[0098]

[0099] Among them, represents the incidence relationship between the * ith node and the kth hyperedge e K . From the perspective of drug and cell nodes, the hypergraph is represented as H ∈ R represents that the node belongs to the hyperedge e K . From the perspective of drugs and cells, the hypergraph is represented as H ∈ R M×N The hyperedges are weighted by the hyperedge weight matrices Ψ d ∈ R N×N and Ψ c ∈ R N×N . The matrix Ψ d ∈ R N×N and Ψ c ∈ R N×N are diagonal matrices; the diagonal elements in the diagonal matrix represent the degree of the hyperedge (i.e., the number of vertices included), and the non-diagonal elements are zero. Similarly, from the perspective of hyperedges, the hypergraph is represented as H T ∈ R N×M The drug and cell nodes are weighted by the positive values stored in the diagonal matrices Φ d ∈ R M×M and Φ c ∈ R M×M respectively. The diagonal matrix is obtained according to the known incidence information of nodes and hyperedges; when there is no prior information about the importance of nodes or hyperedges, the diagonal values in the diagonal matrix can be simply set to 1, and this is the method adopted in the present invention. Ψ d ∈ R N×N is the hyperedge weight matrix weighted by drug nodes, and Ψ c ∈ R N×N is the hyperedge weight matrix weighted by cell nodes;

[0100] S320. Based on the drug structure characteristics, cell fusion characteristics, hyperedge weight matrix and incidence matrix, perform convolution on drug nodes and cell nodes respectively to obtain drug node features and cell node features;

[0101]

[0102] Among them, represents the structural feature vector of the I-th drug at the l-th * layer, initialized to d I ; represents the fusion feature vector of the J-th cell, initialized to c J ; and represent the trainable weight matrix between the l-th * layer and the (l + 1)-th * layer, and are randomly initialized.

[0103] Although the feature dimensions of d I and c J can be changed, due to the spectral radius limitation, stacking multiple convolutional layers in the above convolution operation formula may increase the possibility of vanishing gradients. Therefore, to maintain the scale invariance of d I and c J during the processing, a symmetric normalization operation is applied:

[0104]

[0105] where D d and D c are the diagonal matrices of drug nodes and cell nodes respectively, and the calculation methods of the two node diagonal matrices are and B d and B c represent the hyperedge diagonal matrices, and their calculation methods are and It can be seen that B d and B c are actually calculated in the same way. Here, it is mainly to distinguish the two formulas. Because through calculation, it can be known that and are both positive semi-definite matrices (E′ is the identity matrix). Therefore, it can be determined that and the maximum eigenvalues of are both less than 1. Therefore, by performing multiple convolution operations, the drug node features of the I-th drug node and the cell node features of the J-th cell node can be extracted. * The number of convolutional layers used here is also L

[0106] Next, the hyperedge information needs to be encoded as follows:

[0107] S330. Concatenate any two drug structure features and one cell fusion feature, and input them into a multi-layer perceptron to obtain the initial hyper-edge features;

[0108] Initial hyper-edge features Initialized as MLP(d I ||d P ||c J ), where d I , d P and c J represent the feature vectors of the I-th drug, the P-th drug, and the J-th cell, respectively;

[0109] S340. Convolve the hyper-edges based on the initial hyper-edge features and the adjacency matrix to obtain the hyper-edge features.

[0110]

[0111] Among them, represents the feature vector of the K-th hyper-edge at the l * -th layer, initialized as MLP(d I ||d P ||c J ), || represents the concatenation operation, and MLP(·) represents the processing using the MLP processing layer; and represent the trainable weight matrix between the l * -th layer and the l * +1-th layer, and are randomly initialized.

[0112] Symmetric normalization is also applied to maintain the scale invariance:

[0113]

[0114] Among them, L d and L c represent the hyper-edge pair angle matrices, and the calculation methods of the two hyper-edge pair angle matrices are and J d and J c represent the node pair angle matrices, and their calculation methods are and By performing multiple convolution operations, the potential feature vector of the associated hyper-edge between the I-th drug, the P-th drug, and the J-th cell is extracted The number of convolution operation layers used here is also L * .

[0115] After encoding the drug nodes, cell nodes, and hyperedges, a decoder can be used to predict the synergy labels of different drug combinations under different cells. Specifically, the decoder layer is trained based on the latent data representations of the nodes and hyperedges. When predicting the synergy effect of the I-th drug and the P-th drug under the J-th cell, the prediction result is as follows:

[0116]

[0117] where N represents the total number of known hyperedges; p(·) represents the decoder layer; represents the synergy label (0 or 1); and represent the associations of the I-th drug and the P-th drug with the J-th cell in the reconstructed hypergraph, respectively. If and are both 1, it means there is a synergy effect between the I-th drug and the P-th drug under the J-th cell. At this time, is 1, otherwise there is no synergy effect, is 0.

[0118] In this embodiment, the calculation of the loss function of the model includes two parts: the heterogeneous hypergraph reconstruction loss and the binary cross-entropy loss for sample prediction, as follows:

[0119] S350. Reconstruct the heterogeneous hypergraph based on the predicted values of the prediction model, and obtain the first loss function L reconstruct ;

[0120] [[ID=#33]]

[0121] where represents the association relationship between the I * -th node and the K-th hyperedge before reconstruction; represents the association relationship between the I * -th node and the K-th hyperedge after reconstruction;

[0122] S360. Construct a binary cross-entropy loss function L bce based on the predicted values and true values of the prediction model;

[0123]

[0124] where D represents the total number of drugs; C represents the total number of cells; is the predicted value of the s-th sample; y s is the true value of the s-th sample.

[0125] S370. Perform weighted fusion on the first loss function and the binary cross-entropy loss function to obtain the target loss function Lrb ; as follows:

[0126] L rb = βL reconstruct +(1 - β)L bce

[0127] where β is the fusion weight.

[0128] S400. Obtain the drug to be tested, input the drug to be tested into the prediction model, and obtain the prediction result.

[0129] The present invention first introduces a drug structure feature extraction module, which adopts the method of first updating the edge features and then using the edge features to update the node features, combines the attention mechanism to optimize the processing process of the traditional MPNN, removes duplicate paths and useless paths by updating the edge features, reduces the number of nodes that need to be processed in the feature aggregation operation and improves the processing speed, so as to overcome the problems of repeated extraction of node features and slow extraction speed caused by the traditional MPNN when facing the multiple path problem; after extracting several sub - structure features of each drug, perform k - means clustering on the sub - structure features of all drugs to mine the common features of various sub - structures, and then use the common features to enhance the drug structure features, so as to ensure that the drug structure features incorporate stable common information while retaining specific information. Secondly, introduce a cell fusion feature extraction module, construct a feature enrichment matrix by stacking cell feature vectors, and adopt a matrix fusion strategy to fuse multi - source cell information, ensuring that each feature information can incorporate heterologous features to strengthen itself on the basis of retaining its own features, and overcoming the problem that feature fusion methods such as splicing or average weighting do not consider the correlation relationship between multi - source information. Thirdly, introduce a heterogeneous hypergraph feature extraction module, construct a heterogeneous hypergraph based on the synergistic information of known drug combinations in cells, and design a multi - layer hypergraph convolutional network to perform feature learning on the nodes and hyper - edges in the heterogeneous network, overcoming the disadvantages of over - smoothing and redundancy of information and difficulty in capturing complex relationship patterns when ordinary graph neural networks perform feature extraction on heterogeneous networks. Finally, introduce a multi - layer hypergraph convolutional network, distinguish and consider the contribution differences of different types of nodes to hyper - edges, effectively extract node features and hyper - edge features, and use the node features and hyper - edge features to predict synergistic drugs, improving the stability of the model. The present invention combines the binary cross - entropy loss of drug synergy sample prediction and the heterogeneous hypergraph reconstruction loss, comprehensively considers the model loss using two types of tasks, and improves the generalization ability of the model.

[0130] Example 2

[0131] Refer to Figure 2 , this embodiment provides a synergistic drug prediction system based on a heterogeneous hypergraph, including:

[0132] The first module 100 is used to obtain information of various cells and structures of various drugs;

[0133] The second module 200 is used to obtain corresponding cell fusion characteristics based on information of multiple cells and obtain corresponding drug structure characteristics based on structures of multiple drugs;

[0134] The third module 300 is used to construct a heterogeneous hypergraph based on the drug structure characteristics and cell fusion characteristics, and perform convolution processing to obtain drug node characteristics, cell node characteristics and hyperedge characteristics. The extracted features are used to train a neural network to obtain a prediction model;

[0135] The fourth module 400 is used to obtain a drug to be tested, input the drug to be tested into the prediction model, and obtain a prediction result.

[0136] As an optional implementation manner, the second module includes:

[0137] The first unit 210 is configured to construct a molecular graph based on the SMILES of each drug, process the molecular graph based on a message passing neural network fused with an attention mechanism, and update the node features using the edge features to obtain all substructure features in each molecular graph;

[0138] The second unit 220 is used to cluster the substructure features in all molecular graphs to obtain the centroid information of the cluster; and obtain the drug structure features based on the substructure features and the centroid information of the cluster to which they belong;

[0139] The third unit 230 is used to construct a feature enrichment matrix based on the cell information and perform feature fusion to obtain a cell fusion feature.

[0140] As an optional implementation manner, the second unit 220 includes:

[0141] The first subunit 221 is used to randomly select multiple substructure features as initial centroids, calculate the distance between each substructure feature and each initial centroid, assign it to the cluster where the nearest centroid is located, and update the initial centroids until the centroid position remains unchanged, thereby obtaining the final centroid;

[0142] The second subunit 222 is configured to weight the final centroid according to the number of substructures of the current drug in each cluster, and sum the weighted centroids of all clusters to obtain a key feature vector;

[0143] The third subunit 223 is used to splice the key feature vector and the substructure feature of the drug and input the concatenated data into a multi-layer perceptron for processing to obtain the drug structure feature.

[0144] As an optional implementation manner, the third unit 230 includes:

[0145] The fourth sub-unit 231 is configured to obtain at least one Gaussian regularization vector of the cell based on the cell information corresponding to the cell;

[0146] The fifth sub-unit 232 is configured to perform dimension normalization processing on the Gaussian regularization vector to obtain a corresponding cell feature vector;

[0147] The sixth sub-unit 233 is configured to perform feature vector stacking on each cell feature vector to obtain an enrichment square matrix corresponding to the cell feature vector;

[0148] The seventh sub-unit 234 is configured to perform feature fusion on each enrichment square matrix by using a matrix iterative fusion method, and then perform weighted fusion on the feature fusion results of all enrichment square matrices to obtain a final fusion result;

[0149] The eighth sub-unit 235 is configured to perform averaging processing on the row vectors in the final fusion result to obtain the cell fusion feature of the cell.

[0150] As described above, the above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A collaborative drug prediction method based on heterogeneous hypergraphs, characterized in that, Including: Obtaining information of multiple cells and structures of multiple drugs; Obtaining corresponding cell fusion features according to the information of multiple cells, and obtaining corresponding drug structure features according to the structures of multiple drugs; Constructing a heterogeneous hypergraph based on the drug structure features and cell fusion features, and performing convolutional processing to obtain drug node features, cell node features and hyperedge features, training a neural network with the extracted features to obtain a prediction model; Obtaining a drug to be tested, inputting the drug to be tested into the prediction model to obtain a prediction result.

2. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 1, wherein Obtaining corresponding cell fusion features according to the information of multiple cells, and obtaining corresponding drug structure features according to the structures of multiple drugs, including: Constructing a molecular graph respectively according to the SMILES of each drug, processing the molecular graph by a message passing neural network based on a fusion attention mechanism, and updating the node features by using edge features to obtain all sub-structure features in each molecular graph; Clustering the sub-structure features in all molecular graphs to obtain centroid information of clusters; obtaining drug structure features according to the sub-structure features and the centroid information of the clusters where they are located; Constructing an enrichment matrix of features according to cell information and performing feature fusion to obtain cell fusion features.

3. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 2, wherein Clustering the sub-structure features in all molecular graphs to obtain centroid information of clusters; obtaining drug structure features according to the sub-structure features and the centroid information of the clusters where they are located, including: Randomly selecting multiple sub-structure features as initial centroids, calculating the distance between each sub-structure feature and each initial centroid, assigning it to the cluster where the nearest centroid is located, and updating the initial centroid until the centroid position remains unchanged to obtain the final centroid; Weighting the final centroid according to the number of sub-structures of the current drug in each cluster, and summing the weighted centroids of all clusters to obtain a key feature vector; Inputting the concatenation of the key feature vector and the sub-structure features of the drug into a multi-layer perceptron for processing to obtain drug structure features.

4. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 2, wherein Constructing an enrichment matrix of features according to cell information and performing feature fusion to obtain cell fusion features, including: Obtaining at least one Gaussian regularization vector of a cell based on the cell information corresponding to the cell; Performing dimensional normalization processing on the Gaussian regularization vector to obtain a corresponding cell feature vector; Stacking feature vectors for each cell feature vector to obtain an enrichment matrix corresponding to the cell feature vector; Performing feature fusion on each enrichment matrix by using a matrix iterative fusion method, and then performing weighted fusion on the feature fusion results of all enrichment matrices to obtain a final fusion result; Performing averaging processing on the row vectors in the final fusion result to obtain the cell fusion feature of the cell.

5. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 1, wherein Constructing a heterogeneous hypergraph based on the drug structure features and cell fusion features, and performing convolutional processing to extract drug node features, cell node features and hyperedge features, including: Obtaining a hyperedge weight matrix weighted by drug nodes and cell nodes, and obtaining an incidence matrix of the hypergraph; Performing convolution on drug nodes and cell nodes respectively based on the drug structure features, cell fusion features, hyperedge weight matrix and incidence matrix to obtain drug node features and cell node features; Perform feature splicing on any two drug structure features and one cell fusion feature, and input them into a multi-layer perceptron to obtain the initial hyperedge features; Perform convolution on the hyperedges based on the initial hyperedge features and the adjacency matrix to obtain hyperedge features.

6. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 5, wherein Train a neural network using the extracted features to obtain a prediction model, including: Reconstruct the heterogeneous hypergraph according to the predicted values of the prediction model, and obtain the first loss function based on the differences between the heterogeneous hypergraphs before and after reconstruction; Construct a binary cross-entropy loss function based on the predicted values and the true values of the prediction model; Perform weighted fusion on the first loss function and the binary cross-entropy loss function to obtain the target loss function; Adjust the parameters of the neural network based on the target loss function to obtain the prediction model.

7. A collaborative drug prediction system based on heterogeneous hypergraphs, characterized in that, Including: The first module is used to obtain information on multiple cells and the structures of multiple drugs; The second module is used to obtain the corresponding cell fusion features according to the information of multiple cells, and obtain the corresponding drug structure features according to the structures of multiple drugs; The third module is used to construct a heterogeneous hypergraph based on the drug structure features and the cell fusion features, and perform convolution processing to obtain drug node features, cell node features, and hyperedge features. Train a neural network using the extracted features to obtain a prediction model; The fourth module is used to obtain the drug to be tested, input the drug to be tested into the prediction model, and obtain the prediction result.

8. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 7, wherein The second module includes: The first unit is used to respectively construct a molecular graph according to the SMILES of each drug, process the molecular graph using a message passing neural network based on the fusion attention mechanism, and update the node features using the edge features to obtain all sub-structure features in each molecular graph; The second unit is used to cluster the sub-structure features in all molecular graphs to obtain the centroid information of the clusters; according to the sub-structure features and the centroid information of the clusters where they are located, obtain the drug structure features; The third unit is used to construct an enrichment matrix of features according to the cell information and perform feature fusion to obtain the cell fusion features.

9. The collaborative drug prediction method based on heterogeneous hypergraph according to claim 8, wherein The second unit includes: The first sub-unit is used to randomly select multiple sub-structure features as the initial centroids, calculate the distances between each sub-structure feature and each initial centroid, assign them to the clusters where the nearest centroids are located, and update the initial centroids until the centroid positions remain unchanged to obtain the final centroids; The second sub-unit is used to weight the final centroids according to the number of sub-structures of the current drug in each cluster, sum the weighted centroids of all clusters to obtain the key feature vector; The third sub-unit is used to splice the key feature vector and the sub-structure features of the drug and input them into a multi-layer perceptron for processing to obtain the drug structure features.

10. The collaborative drug prediction method based on a heterogeneous hypergraph according to claim 8, wherein, The third unit includes: The fourth sub-unit is used to obtain at least one Gaussian regularization vector corresponding to the cell based on the cell information corresponding to the cell; The fifth sub-unit is used to perform dimensional normalization processing on the Gaussian regularization vector to obtain the corresponding cell feature vector; The sixth sub-unit is used to stack the feature vectors of each cell feature vector to obtain the enrichment matrix corresponding to the cell feature vector; The seventh sub-unit is used to perform feature fusion on each enrichment matrix using the matrix iterative fusion method, and then perform weighted fusion on the feature fusion results of all enrichment matrices to obtain the final fusion result; The eighth sub-unit is used to average and process the row vectors in the final fusion result to obtain the cell fusion feature of the cell.

Images