Drug interaction prediction method and device based on multi-modal data fusion
By constructing a multimodal data set and performing feature vector fusion, the problem of low accuracy in drug interaction prediction in the prior art is solved, and a more efficient drug interaction prediction effect is achieved.
Patent Information
- Application Number
- CN202411109787.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-26
AI Technical Summary
Existing drug interaction prediction algorithms fail to effectively consider the potential correlation between the multimodal characteristics of drugs and DDI events, resulting in low accuracy of prediction results.
By constructing a drug data set containing molecular maps, SMILES sequences, drug-drug information networks and drug feature similarity data, various feature vectors are calculated, and feature vectors are fused, and finally a pre-trained drug interaction prediction model is input for prediction.
It improves the accuracy of drug interaction prediction, can effectively mine key information between the data of each modality, and improves the quality of DDI prediction results.
Smart Images

Figure CN119028486A_ABST
Abstract
Description
Technical Field
[0001] Embodiments of the present invention relate to the field of neural network technology. More specifically, embodiments of the present invention relate to a drug interaction prediction method and device based on multimodal data fusion. Background Art
[0002] The prediction of drug-drug interaction (DDI) is one of the hot topics in biomedical research. When a drug is used together with one or more other drugs, the interaction between drugs may produce better effects, but it may also have adverse consequences. Therefore, it is crucial to accurately predict potential DDIs in order to maximize the synergistic benefits between drugs.
[0003] Currently, DDI can be predicted through a model composed of a three-dimensional graph neural network and a pre-trained text attention mechanism. However, existing algorithms usually make predictions based on similarity, network, or drug structure, and rarely consider the potential correlation between other multimodal characteristics of drugs and DDI events. They cannot effectively mine the key information between the data of each modality, resulting in low accuracy of DDI prediction results. Summary of the invention
[0004] In this context, embodiments of the present invention are intended to provide a method and apparatus for predicting drug interactions based on multimodal data fusion, which can improve the accuracy of DDI prediction results.
[0005] In a first aspect of an embodiment of the present invention, a method for predicting drug interactions based on multimodal data fusion is provided, comprising:
[0006] Constructing a drug data set based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data sets at least include a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data;
[0007] Based on the drug data set, determine a feature vector set of each drug to be predicted; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector;
[0008] Performing feature vector fusion on each of the feature vector sets of the drug to be predicted to obtain a fused feature vector of each of the drug to be predicted;
[0009] The fused feature vector is input into a pre-trained drug interaction prediction model to output a drug interaction prediction result.
[0010] In an example of this implementation, determining a molecular graph feature vector of a target drug to be predicted based on the drug dataset includes:
[0011] Acquire a target molecule graph dataset corresponding to the target drug to be predicted from the drug dataset;
[0012] Input the target molecular graph dataset into a pre-built graph convolutional network, and output an embedding vector corresponding to the target molecular graph dataset;
[0013] The embedding vector is calculated based on the maximum function and the average function to obtain the molecular graph feature vector of the target drug to be predicted.
[0014] In an example of this implementation, based on the drug dataset, determining a SMILES sequence feature vector of a target drug to be predicted includes:
[0015] Acquire target SMILES sequence data corresponding to the target drug to be predicted from the drug data set; wherein the characters contained in the target SMILES sequence data correspond one-to-one to the target nodes contained in the target drug to be predicted;
[0016] Input the target SMILES sequence data into a pre-built Transformer model, and output a query matrix, a key matrix, and a value matrix;
[0017] Determine the attention weight of each character in the target SMILES sequence data based on the query matrix, the key matrix and the value matrix;
[0018] Concatenating the attention weights of each character to obtain a concatenated attention weight;
[0019] The concatenated attention weights are calculated through a feedforward network and a maximum pooling layer to obtain a SMILES sequence feature vector of the target drug to be predicted.
[0020] In an example of this implementation, determining a drug information feature vector of a target drug to be predicted based on the drug dataset includes:
[0021] Acquire a target drug-drug information network corresponding to the target drug to be predicted from the drug data set;
[0022] Calculating the target drug-drug information network based on the initial attention coefficient of each node of the target drug to be predicted to obtain a final attention coefficient;
[0023] Get the preset weight matrix;
[0024] Based on the weight matrix and the final attention coefficient, a drug information feature vector of the target drug to be predicted is calculated.
[0025] In one embodiment of this implementation, the target drug-drug information network is calculated based on the initial attention coefficient of each node of the target drug to be predicted to obtain a final attention coefficient β ij The calculation formula is:
[0026]
[0027] Among them, W is the preset parameter matrix, LeakyReLU is the activation function, is the preset weight vector, XD i is the feature matrix of node i, m is the number of drug attributes, XD j is the feature matrix of node j, where node j is the neighboring node of node i, N(i) represents the set of neighboring nodes of node i, XD k is the feature matrix of node k in the set N(i) of neighboring nodes of node i, T represents transposition, ∥ represents the connection operation, r ij is the initial attention coefficient of node i.
[0028] In an example of this implementation, the drug to be predicted includes a first drug to be predicted and a second drug to be predicted, and determining the drug feature similarity feature vector of the first drug to be predicted based on the drug data set includes:
[0029] Acquire first drug feature similarity data corresponding to the first drug to be predicted from the drug data set; wherein the first drug feature similarity data at least includes a first drug-enzyme feature matrix and a first drug-target feature matrix;
[0030] Acquire second drug feature similarity data corresponding to the second drug to be predicted from the drug data set; wherein the second drug feature similarity data at least includes a second drug-enzyme feature matrix and a second drug-target feature matrix;
[0031] Calculating cosine similarity based on the first drug-enzyme feature matrix and the second drug-enzyme feature matrix;
[0032] Calculating the Jaccard similarity based on the first drug-target feature matrix and the second drug-target feature matrix;
[0033] The cosine similarity and the Jaccard similarity are concatenated and dimensionally reduced to obtain a drug feature similarity feature vector of the first drug to be predicted.
[0034] In one embodiment of this implementation, the feature vector set of each drug to be predicted is respectively subjected to feature vector fusion to obtain a fused feature vector of each drug to be predicted, including:
[0035] Based on the feature vector set of each drug to be predicted, a set of attention values of each node of each drug to be predicted is calculated; wherein the set of attention values at least includes a molecular graph attention value, a SMILES sequence attention value, a drug-drug information network attention value, and a drug feature similarity attention value;
[0036] Regularizing the attention value set of each node by a softmax function to obtain an attention coefficient set of each node; wherein the attention coefficient set includes at least a molecular graph attention coefficient, a SMILES sequence attention coefficient, a drug-drug information network attention coefficient, and a drug feature similarity attention coefficient;
[0037] Based on the attention coefficient set of each node, feature vector fusion is performed on the feature vector set of each drug to be predicted to obtain a fused feature vector of each drug to be predicted.
[0038] In a second aspect of the embodiments of the present invention, a drug interaction prediction device based on multimodal data fusion is provided, comprising:
[0039] A construction unit, used to construct a drug data set based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data sets at least include a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data;
[0040] A determination unit, used to determine a feature vector set of each of the drugs to be predicted based on the drug data set; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector;
[0041] A fusion unit, used for fusing the feature vectors of each of the feature vector sets of the drug to be predicted to obtain a fused feature vector of each of the drug to be predicted;
[0042] The output unit is used to input the fused feature vector into a pre-trained drug interaction prediction model and output the drug interaction prediction result.
[0043] In a third aspect of an embodiment of the present invention, a computing device is provided, comprising: at least one processor, a memory, and an input-output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute any one of the methods described in the first aspect.
[0044] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, which includes instructions, and when the instructions are executed on a computer, the computer executes the method described in any one of the first aspects.
[0045] According to the drug interaction prediction method and device based on multimodal data fusion according to the embodiment of the present invention, it is possible to analyze a drug data set constructed based on the SMILES sequence of the drug to be predicted, obtain a feature vector set for each drug to be predicted, and the feature vector set includes multimodal data such as molecular graph feature vectors, SMILES sequence feature vectors, drug information feature vectors, and drug feature similarity feature vectors. Based on the feature vector set including the multimodal data, the drug interaction is predicted to obtain a drug interaction prediction result; the method can take into account the potential correlation between other multimodal features of the drug and the DDI event, and can effectively mine the key information between each modality data, thereby improving the accuracy of the DDI prediction result. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:
[0047] Figure 1 A schematic diagram of a process flow of a drug interaction prediction method based on multimodal data fusion provided by an embodiment of the present invention;
[0048] Figure 2 A schematic diagram of updating local features of a molecular graph of a drug provided by an embodiment of the present invention;
[0049] Figure 3 A flowchart of SMILES sequence data construction provided by an embodiment of the present invention;
[0050] Figure 4 An architectural diagram of a drug interaction prediction method based on multimodal data fusion provided by an embodiment of the present invention;
[0051] Figure 5 A schematic diagram of the structure of a drug interaction prediction device based on multimodal data fusion provided by an embodiment of the present invention;
[0052] Figure 6 A schematic diagram of the structure of a medium according to an embodiment of the present invention is schematically shown;
[0053] Figure 7 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is schematically shown.
[0054] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION
[0055] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.
[0056] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.
[0057] According to an embodiment of the present invention, a drug interaction prediction method and device based on multimodal data fusion are proposed.
[0058] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.
[0059] The principle and spirit of the present invention are explained in detail below with reference to several representative embodiments of the present invention.
[0060] Exemplary Methods
[0061] Reference below Figure 1 , Figure 1 This is a flow chart of a drug interaction prediction method based on multimodal data fusion provided by an embodiment of the present invention. It should be noted that the embodiments of the present invention can be applied to any applicable scenario.
[0062] Figure 1 The process of the drug interaction prediction method based on multimodal data fusion provided by an embodiment of the present invention shown includes:
[0063] Step S101, constructing a drug dataset based on the SMILES sequence of the drug to be predicted.
[0064] In an embodiment of the present invention, the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data sets include at least a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data.
[0065] In the embodiment of the present invention, the molecular graph dataset can be constructed by using RDKit to obtain the molecular graph corresponding to the drug. Taking a drug i as an example, it corresponds to a molecular graph composed of atoms and chemical bonds. A molecular graph can be used as a network, and each atom is used as a node v i , connecting two atoms v i and v j The chemical bond is edge e ij , construct the initial features G of atoms in the molecular graph based on atomic properties such as atomic symbol, formal charge, and chirality i v={G i v1, ..., G i v i , ..., G i v t}, G i v i is node v in the molecular graph of drug i i The characteristic matrix of , t is the number of atoms contained in a molecular graph, and in addition, the node adjacency matrix A is constructed i g={A i g1,...,A i g i , ..., A i g t}, A i g i is node v in drug i i The adjacency matrix of the molecular graph is: if there is an edge between two nodes in the molecular graph, it is 1, otherwise it is 0. The initial characteristics G of the chemical bonds in the molecular graph are constructed based on the bond type, conjugation, ring membership and stereochemical properties. i e. Accordingly, the initial features of the atoms in the molecular graph corresponding to the entire molecular graph dataset are Gv = {G1v, ..., G i v, ..., G n v}, the adjacency matrix is Ag = {A1g, ..., A i g,…,A n g}, containing the initial characteristics of the bond Ge, and n is the number of drugs.
[0066] Please also read Figure 2 , Figure 2 A schematic diagram of updating the local features of the molecular graph of a drug provided by an embodiment of the present invention; wherein the dark nodes are v i, v i The edge between the nodes is e ij , feature updates are performed by combining its own features with the edge features between other connected nodes.
[0067] The construction method of SMILES sequence data can be: SMILES string encoding uses simple integer encoding, and RDKit is used to obtain the atom list List contained in each SMILES = {List1, ..., List i , ..., List n}, calculate the length ml of the list containing atoms for each SMILES i , i∈{1,...,n} and put the characters (atoms) that have not appeared in the dictionary into the corresponding integers in the order in which they are put. After the characters are put into the dictionary, one-hot encoding is performed on the characters put into the dictionary. Get the longest length of the atom list in all SMILES strings ML=max(ml1,...,ml n ), then take a drug i as an example, and look up List from the dictionary i The one-hot encoding of the atoms contained in List i If the length is shorter than ML, zero padding is used to ensure that each Smile has the same length, that is, contains t characters. This constitutes the embedded sequence XS corresponding to the SMILES formula of the drug i ={xs1, ..., xs i , ..., xs t}, the dimension is one-hot encoding dimension d, xS i is the i-th character in the SMILES string. The process is as follows Figure 3 As shown, Figure 3 Flow chart of constructing SMILES sequence data provided by an embodiment of the present invention. Other SMILES corresponding codes are the same as above, and the SMILES sequence data XS={XS1, ..., XS i , ..., XS n}.
[0068] The drug-drug information network can be constructed by collecting drug information from a database, and constructing an adjacency matrix and a feature matrix based on the drug information, wherein the behavior of the feature matrix is a drug node, and the column is a drug attribute node (drug similarity, oral bioavailability, blood-brain barrier, etc.). If the drug has this attribute, it is set to 1, otherwise it is 0, so as to obtain a drug-drug information network XD = {XD1, ..., XD i ,...,XD n}, In addition, an adjacency matrix AD is constructed, where the value is 1 if two drugs have an edge, otherwise it is 0.
[0069] The drug feature similarity data can be constructed by collecting biochemical features that may affect drug interactions from the database, including targets and enzymes, and constructing a drug-enzyme feature matrix XE = {XE1, ..., XE i , ..., XE n} and drug-target feature matrix XT = {XT1, ..., XT i , ..., XT n Each feature (enzyme, target) corresponds to a set of descriptors, and a drug can be represented as a binary feature vector, where each entry represents the presence or absence of the corresponding descriptor.
[0070] Step S102: determining a feature vector set of each drug to be predicted based on the drug dataset.
[0071] In the embodiment of the present invention, the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector.
[0072] Optionally, based on the drug dataset, a method for determining a molecular graph feature vector of a target drug to be predicted may include the following steps:
[0073] Obtaining a target molecule graph dataset corresponding to the target drug to be predicted from the drug dataset:
[0074] Input the target molecular graph dataset into a pre-built graph convolutional network, and output an embedding vector corresponding to the target molecular graph dataset;
[0075] The embedding vector is calculated based on the maximum function and the average function to obtain the molecular graph feature vector of the target drug to be predicted.
[0076] Among them, by implementing this implementation method, the target analysis graph dataset can be processed through a pre-built graph convolutional network to obtain an embedding vector corresponding to the target molecular graph dataset, and then the embedding vector can be calculated using the maximum function and the average function to obtain the molecular graph feature vector, thereby accurately obtaining the feature vector of the target drug to be predicted in terms of the molecular graph.
[0077] Specifically, taking a molecular graph as an example, in order to fully learn the atomic and chemical bond features, the atomic information in the molecular graph is first updated using the atomic (node) and chemical bond (edge) features, and the features are updated by traversing the atoms in the molecular graph. The formula is as follows:
[0078] G i v i =Relu(G i v i ||∑j∈N(i )G i e ij );
[0079] Where N(i) is the node v i Neighbor nodes, G i v i is node v in molecular graph i i The characteristic information of G i e ij For connecting node v i Its neighbor node v j The edge ij II is the feature information of , II is the splicing function, and the atomic features of other molecular graphs are updated in the same way as above.
[0080] After updating the atomic features in all molecular graphs, the GCN model is constructed. Taking a molecular graph i as an example, GCN (Graph Convolutional Network) is used to achieve effective embedding extraction. The principle is as follows. The feature vector of the output of the l+1 layer can be expressed as:
[0081]
[0082] Where σ is the activation function Relu, D is the adjacency matrix A i g’s degree matrix, I represents the identity matrix, Zg (l) Represents the hidden vector of the lth layer. When 1=0, that is, Zg 0 Then it is the feature matrix G i v,W (l) is the weight matrix. When iterating from 1=0 to 1=L-1, we get the embedding vector Zg of the molecular graph i. L .
[0083] ZGar L Each row in the graph represents the information of a node in the molecular graph. The number of nodes involved in a molecular graph is N, and each row The mean function and the max function are used to integrate the key information of nodes in the molecular graph. Specifically, each row of node feature vectors is concatenated with the maximum value of each row of node feature vectors. The concatenated feature vector has N rows. The average is taken to finally obtain a feature vector of the molecular graph i corresponding to drug i, as shown below:
[0084]
[0085] The steps for extracting the feature vectors of the molecular graphs corresponding to other drugs are the same as above, and finally the molecular graph feature vector ZG of the target drug to be predicted is obtained. i , ..., ZGn}.
[0086] And, based on the drug dataset, a method for determining a SMILES sequence feature vector of a target drug to be predicted may include the following steps:
[0087] Acquire target SMILES sequence data corresponding to the target drug to be predicted from the drug data set; wherein the characters contained in the target SMILES sequence data correspond one-to-one to the target nodes contained in the target drug to be predicted;
[0088] Input the target SMILES sequence data into a pre-built Transformer model, and output a query matrix, a key matrix, and a value matrix;
[0089] Determine the attention weight of each character in the target SMILES sequence data based on the query matrix, the key matrix and the value matrix;
[0090] Concatenating the attention weights of each character to obtain a concatenated attention weight;
[0091] The concatenated attention weights are calculated through a feedforward network and a maximum pooling layer to obtain a SMILES sequence feature vector of the target drug to be predicted.
[0092] Among them, by implementing this implementation method, the target SMILES sequence data can be processed through a pre-built Transformer model to obtain the query matrix, key matrix and value matrix corresponding to the target SMILES sequence data, and then the query matrix, key matrix and value matrix are processed to obtain the splicing attention weight; finally, based on the splicing attention weight, an accurate feature vector of the target drug to be predicted in the SMILES sequence can be obtained.
[0093] Specifically, taking a target SMILES sequence data XS i For example, through the three linear layers of the Transformer model (a deep learning model based on the self-attention mechanism), three independent matrices are generated, namely the query matrix (Query), the key matrix (Key) and the value matrix (Value). The three matrices are used to calculate the attention score. Since the multi-head attention mechanism is used in this model, Q (h) , K (h) and V (h) Used to calculate multiple groups of attention scores, where Q (h) =XS i W Q (h) , K (h) =XS i WK (h) 、V (h) =XS i W V (h) ,in, is the projection matrix, d Q d K and d V They are three projection matrices W respectively Q , W K and W V The dimension of , h is the number of heads, h∈{1,...,H}.
[0094] The attention scores are normalized using the sofimax function to obtain the attention weights. The weight of each head represents the sequence XS i Each character in the XS i Characters with higher attention weights are considered more critical to the task being performed. This is shown below:
[0095]
[0096] Concatenate the multi-head attention weights, W head is the weight matrix used for training. As shown below:
[0097] MS=concat(head (1) , ..., head (h) , ..., head (H) )W head ;
[0098] The concatenated multi-head attention weights are passed through the feedforward network and the maximum pooling layer to obtain the target SMILES sequence data XS of the target drug i to be predicted i The SMILES sequence feature vector ZS i , where WS1, Wb1, WS2 and Wb2 are learning parameters, and finally traverse XS to obtain the target SMILES sequence data XS i The SMILES sequence feature vector ZS={ZS1,...,ZS i , ..., ZS n}.
[0099] ZS i =maxpool(Relu(MSWS1+Wb1)WS2+Wb2);
[0100] In addition, based on the drug data set, a method for determining a drug information feature vector of a target drug to be predicted may include the following steps:
[0101] Acquire a target drug-drug information network corresponding to the target drug to be predicted from the drug data set;
[0102] Calculating the target drug-drug information network based on the initial attention coefficient of each node of the target drug to be predicted to obtain a final attention coefficient;
[0103] Get the preset weight matrix;
[0104] Based on the weight matrix and the final attention coefficient, a drug information feature vector of the target drug to be predicted is calculated.
[0105] Among them, by implementing this implementation method, the target drug-drug information network can be processed through a shared self-attention mechanism to obtain a drug information feature vector of the target drug to be predicted; thereby improving the accuracy of the feature vector of the obtained drug-drug information network.
[0106] Specifically, a shared self-attention mechanism is introduced with an attention coefficient of r ij =t(WXD i ,WXD j ), which is used to set the influence of neighboring node j on drug node i, where W is the parameter matrix to be learned, t represents the self-attention mechanism, which is normalized to compare the sizes of various attention coefficients, N(i) is the set of neighboring nodes of drug node i, j∈N(i), as shown below.
[0107]
[0108] Furthermore, the target drug-drug information network is calculated based on the initial attention coefficients of each node of the target drug to be predicted to obtain the final attention coefficient β ij The calculation formula is:
[0109]
[0110] Among them, W is the preset parameter matrix, LeakyReLU is the activation function, is the preset weight vector, XD i is the feature matrix of node i, m is the number of drug attributes, XD j is the feature matrix of node j, where node j is the neighboring node of node i, N(i) represents the set of neighboring nodes of node i, XD k is the feature matrix of node k in the set N(i) of neighboring nodes of node i, T represents transposition, ∥ represents the connection operation, r ijis the initial attention coefficient of node i. This formula can be used to get the final attention coefficient more accurately.
[0111] In order to make the learning process of the self-attention mechanism more stable, the attention coefficient is calculated using multi-head attention, and the weighted feature vectors of the neighboring nodes of drug node i are aggregated to obtain the feature vector of node i, where σ is the activation function RELU, is the normalized attention coefficient of the k-th head attention mechanism, W k is the weight matrix used for feature transformation.
[0112]
[0113] Then, the average strategy is used to update the feature i of the drug node to obtain ZD i , and further generate the final drug information feature vector ZD = {ZD1,...,ZD i ,...,ZD n}:
[0114]
[0115] As an optional implementation, the drug to be predicted includes a first drug to be predicted and a second drug to be predicted, and the drug to be predicted includes the first drug to be predicted and the second drug to be predicted. The method of determining the drug feature similarity feature vector of the first drug to be predicted based on the drug data set may include the following steps:
[0116] Acquire first drug feature similarity data corresponding to the first drug to be predicted from the drug data set; wherein the first drug feature similarity data at least includes a first drug-enzyme feature matrix and a first drug-target feature matrix;
[0117] Acquire second drug feature similarity data corresponding to the second drug to be predicted from the drug data set; wherein the second drug feature similarity data at least includes a second drug-enzyme feature matrix and a second drug-target feature matrix;
[0118] Calculating cosine similarity based on the first drug-enzyme feature matrix and the second drug-enzyme feature matrix;
[0119] Calculating the Jaccard similarity based on the first drug-target feature matrix and the second drug-target feature matrix;
[0120] The cosine similarity and the Jaccard similarity are concatenated and dimensionally reduced to obtain a drug feature similarity feature vector of the first drug to be predicted.
[0121] Among them, by implementing this implementation method, the cosine similarity and Jaccard similarity can be calculated for the drug-enzyme feature matrix and the second drug-target feature matrix of the first drug to be predicted and the second drug to be predicted, respectively, and then the cosine similarity and Jaccard similarity can be concatenated and dimensionally reduced to make the obtained drug feature similarity feature vector of the first drug to be predicted more accurate.
[0122] Specifically, based on the cosine similarity and the Jaccard similarity, the cosine similarity CE of the first drug-enzyme feature matrix and the second drug-enzyme feature matrix are calculated respectively. i ,...,CE n}, and calculate the Jaccard similarity JT = {JT1,...,JT i ,...,JT n} as shown below.
[0123]
[0124] whereXE i and XE j Respectively represent the drug-enzyme feature matrix of the first drug to be predicted i and the second drug to be predicted j in the drug-enzyme feature matrix, XT i and XT j Respectively represent the drug-target feature matrix of the first drug i to be predicted and the second drug j to be predicted in the drug-target feature matrix, |XT i ∩XT j | is the intersection of the two, |XT i ∪XT j | is the union of the two.
[0125] Further, CE i and JT i The concatenation and dimensionality reduction are performed as the drug feature similarity feature vector of the first drug i to be predicted, where: is the concatenation operation, DR() is the dimension reduction function, and the final feature vector ZF={ZF1,...,ZF i ,...,ZF n}:
[0126]
[0127] Step S103 , performing feature vector fusion on each feature vector set of the drug to be predicted to obtain a fused feature vector of each drug to be predicted.
[0128] As an optional implementation, step S103 performs feature vector fusion on each feature vector set of the drug to be predicted, and the method of obtaining the fused feature vector of each drug to be predicted may include the following steps:
[0129] Based on the feature vector set of each drug to be predicted, a set of attention values of each node of each drug to be predicted is calculated; wherein the set of attention values at least includes a molecular graph attention value, a SMILES sequence attention value, a drug-drug information network attention value, and a drug feature similarity attention value;
[0130] Regularizing the attention value set of each node by a softmax function to obtain an attention coefficient set of each node; wherein the attention coefficient set includes at least a molecular graph attention coefficient, a SMILES sequence attention coefficient, a drug-drug information network attention coefficient, and a drug feature similarity attention coefficient;
[0131] Based on the attention coefficient set of each node, feature vector fusion is performed on the feature vector set of each drug to be predicted to obtain a fused feature vector of each drug to be predicted.
[0132] By implementing this implementation, a set of attention coefficients of each node of the drug to be predicted during feature vector fusion can be calculated, thereby accurately calculating the contribution of each node feature to the feature vector fusion.
[0133] Specifically, the self-attention mechanism is used to realize multimodal feature fusion, including the molecular graph feature vector ZG of the target drug to be predicted, the SMILES sequence feature vector ZS, the drug information feature vector ZD, and the drug feature similarity feature vector ZF. Taking the drug node i in ZG as an example, a nonlinear transformation is applied and multiplied with the shared attention mechanism to obtain the attention value set of node i (the attention value set at least includes the molecular graph attention value WG i 、SMILES sequence attention value WS i , Drug-drug information network attention value WD i And the drug feature similarity attention value WF i ), which is used to calculate the contribution of each node feature to feature fusion, as shown below:
[0134] WG i =q T gtanh(wGg(ZG i ) T +bG);
[0135] Similarly, the attention values of drug node i in ZS, ZD, and ZF are as follows:
[0136] WS i =q T gtanh(wSg(ZS i ) T +bS);
[0137] WD i =q T gtanh(wDg(ZD i ) T +bD);
[0138] WF i =q T gtanh(wFg(ZF i ) T +bF);
[0139] Among them, wG, wS, wD and wF are weight matrices, bG, bS, bD and bF are bias vectors, and q is the shared attention vector.
[0140] The attention value is regularized using the softmax function to obtain the attention coefficient of node i. These coefficients are used to finally determine the weight of each feature during fusion, as shown below:
[0141]
[0142] Similarly, αS i =softmax(WS i ), αD i =softmax(WD i ), αF i =softmax(WF i ).
[0143] The attention coefficients corresponding to the final molecular graph feature vector ZG, SMILES sequence feature vector ZS, drug information feature vector ZD, and drug feature similarity feature vector ZF are:
[0144] αG={αG1,...,αG i , ..., αG n};
[0145] αS={αS1,...,αS i , ..., αS n};
[0146] αD={αD1,...,αD i , ..., αD n};
[0147] αF={αF1,...,αFi , ..., αF n}.
[0148] Finally, the fused feature vector after fusing the multimodal data using this weight is as follows:
[0149] Emb=αGgZG+αGgZS+αGgZD+αGgZF;
[0150] Step S104, inputting the fused feature vector into a pre-trained drug interaction prediction model, and outputting a drug interaction prediction result.
[0151] In the embodiment of the present invention, the drug interaction prediction model obtained by pre-training can measure the drug interaction prediction score by cosine similarity, which is expressed as:
[0152]
[0153] Among them, Emb i and Emb j are the fused feature vectors of drug node i and drug node j that need to predict interactions in the training set, The drug interaction prediction result between the two drugs is predicted, and the drug interaction prediction result may be a drug similarity score.
[0154] Please also read Figure 4 , Figure 4 An architectural diagram of a drug interaction prediction method based on multimodal data fusion provided by an embodiment of the present invention; Figure 4 It shows the specific processing process of the drug data set from step S102 to step S104.
[0155] In addition, the drug interaction prediction model can be trained first. Before training, a training data set can be constructed first. Specifically, the data set can be divided according to the drug interaction list D={t, j, l}, l∈{0, 1}, where i and j are two drugs, and l indicates whether the two drugs have an interaction. In this way, a training set D_train={i, j, l} and a test set D_test={i, j, l} are divided, and the model is trained using the training set.
[0156] Implementing the above steps during the training process can get the prediction results From the training set, we can get the prediction results The corresponding true label Y ij , in order to minimize the model's prediction results and the true label Y ijTo optimize the prediction results, the binary cross entropy can be used as the loss function to train the model, and finally the test set is used to evaluate the model. The loss function expression is as follows:
[0157]
[0158] Y ij is the true label of drug pair i and j that needs to predict interaction, where Y ij ∈{0, 1}.
[0159] The present invention can take into account the potential correlation between other multimodal features of drugs and DDI events, and can effectively mine key information between each modal data, thereby improving the accuracy of DDI prediction results. In addition, the present invention can also accurately obtain the feature vector of the target drug to be predicted in terms of the molecular graph. In addition, the present invention can also obtain accurate feature vectors of the target drug to be predicted in terms of the SMILES sequence. In addition, the present invention can also improve the accuracy of the feature vector obtained in terms of the drug-drug information network. In addition, the present invention can also more accurately obtain the final attention coefficient. In addition, the present invention can also make the drug feature similarity feature vector of the first drug to be predicted more accurate. In addition, the present invention can also accurately calculate the contribution of each node feature to the feature vector fusion.
[0160] Exemplary Devices
[0161] After introducing the method of the exemplary embodiment of the present invention, next, refer to Figure 5 A drug interaction prediction device based on multimodal data fusion according to an exemplary embodiment of the present invention is described, and the device comprises:
[0162] A construction unit 501 is used to construct a drug data set based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data sets at least include a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data;
[0163] A determination unit 502 is used to determine a feature vector set of each of the drugs to be predicted based on the drug data set; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector;
[0164] A fusion unit 503 is used to perform feature vector fusion on each feature vector set of the drug to be predicted, so as to obtain a fused feature vector of each drug to be predicted;
[0165] The output unit 504 is used to input the fused feature vector into a pre-trained drug interaction prediction model and output a drug interaction prediction result.
[0166] The present invention can take into account the potential correlation between other multimodal features of drugs and DDI events, and can effectively mine key information between each modal data, thereby improving the accuracy of DDI prediction results. In addition, the present invention can also accurately obtain the feature vector of the target drug to be predicted in terms of the molecular graph. In addition, the present invention can also obtain accurate feature vectors of the target drug to be predicted in terms of the SMILES sequence. In addition, the present invention can also improve the accuracy of the feature vector obtained in terms of the drug-drug information network. In addition, the present invention can also more accurately obtain the final attention coefficient. In addition, the present invention can also make the drug feature similarity feature vector of the first drug to be predicted more accurate. In addition, the present invention can also accurately calculate the contribution of each node feature to the feature vector fusion.
[0167] Exemplary Media
[0168] After introducing the method and apparatus of the exemplary embodiment of the present invention, next, reference is made to Figure 6 For a description of a computer-readable storage medium according to an exemplary embodiment of the present invention, please refer to Figure 6 , the computer-readable storage medium shown is a CD 60, on which a computer program (i.e., a program product) is stored. When the computer program is executed by the processor, the steps described in the above method implementation are implemented, for example, a drug data set is constructed based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data set includes at least a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data; based on the drug data set, a feature vector set of each drug to be predicted is determined; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector; the feature vector sets of each drug to be predicted are respectively fused to obtain a fused feature vector of each drug to be predicted; the fused feature vector is input into a pre-trained drug interaction prediction model, and a drug interaction prediction result is output; the specific implementation method of each step is not repeated here.
[0169] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical or magnetic storage media, which are not listed here one by one.
[0170] Exemplary Computing Devices
[0171] After introducing the method, apparatus and medium of the exemplary embodiments of the present invention, next, reference is made to Figure 7 A computing device for drug interaction prediction based on multimodal data fusion according to an exemplary embodiment of the present invention.
[0172] Figure 7 A block diagram of an exemplary computing device 70, which may be a computer system or a server, is shown that is suitable for use in implementing embodiments of the present invention. Figure 7 The computing device 70 shown is merely an example and should not bring any limitation to the functionality and scope of use of embodiments of the present invention.
[0173] like Figure 7 As shown, the components of the computing device 70 may include, but are not limited to: one or more processors or processing units 701, a system memory 702, and a bus 703 connecting different system components (including the system memory 702 and the processing unit 701).
[0174] The computing device 70 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computing device 70, including volatile and non-volatile media, removable and non-removable media.
[0175] The system memory 702 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 7021 and / or cache memory 7022. The computing device 70 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the ROM 7023 may be used to read and write non-removable, non-volatile magnetic media ( Figure 7 is not shown in the Figure 7As shown in FIG. 7 , a disk drive for reading and writing a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing a removable non-volatile optical disk (e.g., a CD-ROM, a DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 703 via one or more data medium interfaces. System memory 702 may include at least one program product having a set (e.g., at least one) of program modules that are configured to perform the functions of various embodiments of the present invention.
[0176] A program / utility 7025 having a set (at least one) of program modules 7024 may be stored, for example, in system memory 702, and such program modules 7024 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment. Program modules 7024 generally perform the functions and / or methods of the embodiments described herein.
[0177] The computing device 70 may also communicate with one or more external devices 704 (e.g., a keyboard, a pointing device, a display, etc.). Such communication may be performed via an input / output (I / O) interface 705. In addition, the computing device 70 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 706. Figure 7 As shown, the network adapter 706 communicates with other modules (such as the processing unit 701, etc.) of the computing device 70 via the bus 703. It should be understood that although Figure 7 Not shown, other hardware and / or software modules may be used in conjunction with computing device 70 .
[0178] The processing unit 701 executes various functional applications and data processing by running the program stored in the system memory 702, for example, constructing a drug data set based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data set includes at least a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data; based on the drug data set, determining a feature vector set of each drug to be predicted; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector; performing feature vector fusion on the feature vector set of each drug to be predicted, respectively, to obtain a fused feature vector of each drug to be predicted; inputting the fused feature vector into the pre-trained drug interaction prediction model, and outputting the drug interaction prediction result. The specific implementation of each step will not be repeated here. It should be noted that although several units / modules or sub-units / sub-modules of the drug interaction prediction device based on multimodal data fusion are mentioned in the above detailed description, this division is only exemplary and not mandatory. In fact, according to the embodiments of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.
[0179] In the description of the present invention, it should be noted that the terms “first”, “second” and “third” are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0180] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0181] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.
[0182] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0183] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0184] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0185] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in 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.
[0186] In addition, although the operations of the method of the present invention are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.
Claims
1. A drug interaction prediction method based on multimodal data fusion, comprising: Constructing a drug data set based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data sets at least include a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data; Based on the drug data set, determine a feature vector set of each drug to be predicted; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector; Performing feature vector fusion on each of the feature vector sets of the drug to be predicted to obtain a fused feature vector of each of the drug to be predicted; The fused feature vector is input into a pre-trained drug interaction prediction model to output a drug interaction prediction result.
2. According to the drug interaction prediction method based on multimodal data fusion according to claim 1, the step of determining a molecular graph feature vector of a target drug to be predicted based on the drug dataset comprises: Acquire a target molecule graph dataset corresponding to the target drug to be predicted from the drug dataset; Input the target molecular graph dataset into a pre-built graph convolutional network, and output an embedding vector corresponding to the target molecular graph dataset; The embedding vector is calculated based on the maximum function and the average function to obtain the molecular graph feature vector of the target drug to be predicted.
3. The drug interaction prediction method based on multimodal data fusion according to claim 1, determining a SMILES sequence feature vector of a target drug to be predicted based on the drug dataset, comprising: Acquire target SMILES sequence data corresponding to the target drug to be predicted from the drug data set; wherein the characters contained in the target SMILES sequence data correspond one-to-one to the target nodes contained in the target drug to be predicted; Input the target SMILES sequence data into a pre-built Transformer model, and output a query matrix, a key matrix, and a value matrix; Determine the attention weight of each character in the target SMILES sequence data based on the query matrix, the key matrix and the value matrix; Concatenating the attention weights of each character to obtain a concatenated attention weight; The concatenated attention weights are calculated through a feedforward network and a maximum pooling layer to obtain a SMILES sequence feature vector of the target drug to be predicted.
4. The drug interaction prediction method based on multimodal data fusion according to claim 1, wherein determining a drug information feature vector of a target drug to be predicted based on the drug dataset comprises: Acquire a target drug-drug information network corresponding to the target drug to be predicted from the drug data set; Calculating the target drug-drug information network based on the initial attention coefficient of each node of the target drug to be predicted to obtain a final attention coefficient; Get the preset weight matrix; Based on the weight matrix and the final attention coefficient, a drug information feature vector of the target drug to be predicted is calculated.
5. According to the drug interaction prediction method based on multimodal data fusion according to claim 4, the target drug-drug information network is calculated based on the initial attention coefficient of each node of the target drug to be predicted to obtain the final attention coefficient β ij The calculation formula is: in, W is the preset parameter matrix, LeakyReLU is the activation function, is the preset weight vector, XD i is the feature matrix of node i, m is the number of drug attributes, XD j is the feature matrix of node j, where node j is the neighboring node of node i, N(i) represents the set of neighboring nodes of node i, XD k is the feature matrix of node k in the set N(i) of neighboring nodes of node i, T represents transposition, ∥ represents the connection operation, r ij is the initial attention coefficient of node i.
6. The drug interaction prediction method based on multimodal data fusion according to claim 1, wherein the drug to be predicted includes a first drug to be predicted and a second drug to be predicted, and the step of determining the drug feature similarity feature vector of the first drug to be predicted based on the drug data set comprises: Acquire first drug feature similarity data corresponding to the first drug to be predicted from the drug data set; wherein the first drug feature similarity data at least includes a first drug-enzyme feature matrix and a first drug-target feature matrix; Acquire second drug feature similarity data corresponding to the second drug to be predicted from the drug data set; wherein the second drug feature similarity data at least includes a second drug-enzyme feature matrix and a second drug-target feature matrix; Calculating cosine similarity based on the first drug-enzyme feature matrix and the second drug-enzyme feature matrix; Calculating the Jaccard similarity based on the first drug-target feature matrix and the second drug-target feature matrix; The cosine similarity and the Jaccard similarity are concatenated and dimensionally reduced to obtain a drug feature similarity feature vector of the first drug to be predicted.
7. The drug interaction prediction method based on multimodal data fusion according to claim 1, wherein the feature vector set of each drug to be predicted is subjected to feature vector fusion to obtain a fused feature vector of each drug to be predicted, comprising: Based on the feature vector set of each drug to be predicted, a set of attention values of each node of each drug to be predicted is calculated; wherein the set of attention values at least includes a molecular graph attention value, a SMILES sequence attention value, a drug-drug information network attention value, and a drug feature similarity attention value; Regularizing the attention value set of each node by a softmax function to obtain an attention coefficient set of each node; wherein the attention coefficient set includes at least a molecular graph attention coefficient, a SMILES sequence attention coefficient, a drug-drug information network attention coefficient, and a drug feature similarity attention coefficient; Based on the attention coefficient set of each node, feature vector fusion is performed on the feature vector set of each drug to be predicted to obtain a fused feature vector of each drug to be predicted.
8. A drug interaction prediction device based on multimodal data fusion, comprising: A construction unit, used to construct a drug data set based on the SMILES sequence of the drug to be predicted; wherein the drug data set includes drug sub-data sets of each drug to be predicted, and the drug sub-data sets at least include a molecular graph data set, a SMILES sequence data, a drug-drug information network, and drug feature similarity data; A determination unit, used to determine a feature vector set of each of the drugs to be predicted based on the drug data set; wherein the feature vector set includes at least a molecular graph feature vector, a SMILES sequence feature vector, a drug information feature vector, and a drug feature similarity feature vector; A fusion unit, used for performing feature vector fusion on each feature vector set of the drug to be predicted, to obtain a fused feature vector of each drug to be predicted; The output unit is used to input the fused feature vector into a pre-trained drug interaction prediction model and output the drug interaction prediction result.
9. A computing device, comprising: at least one processor, memory, and input-output unit; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the method according to any one of claims 1 to 7.
10. A computer-readable storage medium comprising instructions, which, when executed on a computer, causes the computer to execute the method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Drug interaction prediction method based on multi-modal network and width learning
CN115713966A
Molecular pre-training representation method and system fusing SMILES sequence and molecular map
CN115762659A
Drug-target interaction prediction method fusing multi-dimensional features
CN116206775A
Multi-category drug interaction prediction method, device, equipment and medium
CN116936126A
Drug-target interaction prediction method and system based on multi-view comparative learning
CN117766019A
