Drug-drug interaction prediction method based on deep learning
Through deep learning technology, the drug is characterized by encoding and feature extraction, combined with the multi-head graph attention mechanism and multi-layer graph neural network, the shortcomings of drug-drug interaction prediction in the existing technology are solved, and more accurate and efficient drug interaction prediction is achieved.
Patent Information
- Application Number
- CN202411689877.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-06-03
AI Technical Summary
The lack of effective deep learning methods for predicting drug-drug interactions in the prior art, making it difficult to accurately evaluate and predict drug combination interactions when treating complex diseases.
Using a deep learning-based method, the feature vectors of drug combinations are obtained by encoding the drug feature information, and the independent and related features of drugs are extracted using multi-head graph attention mechanism and multi-layer graph neural network, and finally predicting drug-drug interactions through feature fusion and MLP prediction models.
Improves the accuracy and performance of drug-drug interaction predictions, enabling more efficient identification of potential drug interactions, thereby helping to develop safer and more effective treatment options.
Smart Images
Figure CN120089232A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of prediction of drug-drug interactions, and particularly to a method for predicting drug-drug interactions based on deep learning. Background Art
[0002] Drug-drug interactions (DDIs) refer to the pharmacological and clinical responses of drug combinations, which are different from the known modes of action of the two drugs when used alone. When treating complex diseases, a single drug may not provide sufficient efficacy. Therefore, identifying potential DDIs and using multiple drugs in combination is crucial for medical treatment plans. Compared with single drugs, drug combinations often exhibit new pharmacological and clinical responses, which can not only enhance the therapeutic effect but also bring new efficacy and side effects.
[0003] The treatment strategy of drug combinations aims to enhance efficacy and reduce the development of drug resistance through the synergistic effects of drugs with different mechanisms or targets. However, the complexity of drug combinations also brings challenges of drug interactions, which may lead to unexpected adverse reactions. To maximize the advantages of drug combinations while reducing risks, it is necessary to accurately predict and evaluate drug-drug interactions.
[0004] In the existing exploration of the field of drug-drug interactions (DDIs), there is a lack of practical and effective methods for applying deep learning technology in the field of DDIs. Summary of the Invention
[0005] To solve the existing problems, the present invention provides a method for predicting drug-drug interactions based on deep learning, and the specific scheme is as follows:
[0006] A method for predicting drug-drug interactions based on deep learning, comprising the following steps:
[0007] S1, encoding the characteristic information of the drugs to obtain the characteristic vectors of the drug combination;
[0008] S2, inputting the characteristic vectors of the drug combination into a multi-head graph attention mechanism to respectively extract the independent drug molecule characteristics of the two drugs in the drug combination;
[0009] S3, constructing an association network of all drugs, and using a multi-layer graph neural network to extract and obtain drug association characteristics;
[0010] S4, fusing and reclassifying the independent drug molecule characteristics and drug association characteristics to predict drug-drug interactions.
[0011] Preferably, the specific steps of encoding the characteristic information of the drugs in step S1 include:
[0012] S11. Convert the drug SMILES string into a molecular graph using RDKit. The nodes of the molecular graph represent atoms, and the edges of the molecular graph represent chemical bonds. The formula is expressed as: G(A) = (V(A), E(A));
[0013] Among them, A represents drug A, where V(A) is the set of nodes, and each node represents an atom in the drug molecular graph; E(A) is the set of edges in the molecular graph, and each edge represents the chemical bond between atoms in the molecular graph;
[0014] S12. Use Node2vec to obtain the substructure feature encoding information of the topological sequence on the molecular graph;
[0015] S13. Obtain the substructure information through the PubMed database.
[0016] Preferably, step S2 extracts the independent drug molecular features of the two drugs in the drug combination, specifically including the following steps:
[0017] S21. Use the multi-head graph attention network multi-head GAT to perform iterative calculations on drug A, and perform feature update and extraction through the i-layer multi-head GAT. The formula is expressed as:
[0018]
[0019] Among them, || represents the feature concatenation operation, M represents the maximum number of multi-heads of GAT, and W represents the weight matrix;
[0020] represents the attention coefficient, representing the importance of the feature of the i-th node of the drug to the feature of the j-th node. The calculation formula is expressed as:
[0021]
[0022] Among them, a T ∈ R C' is a learnable weight vector;
[0023] S22. Use the global addition pooling gap to process the drug to obtain the drug information feature vector of each drug. The formula is expressed as: Z m = gap(v 1 , v 2 ,..., v n ).
[0024] Preferably, in step S3, the drug association features are obtained by training with drug information, including the steps:
[0025] S31. Encode the PubMed sequence substructure information of the drug into substructure features using a soft threshold function, which is expressed by the formula:
[0026] x kout = sign(x kin ) × ReLU(||x kin || - θ)
[0027] where xkin and xkout are the input and output of neuron k in the layer, and θ is the threshold.
[0028] S32. Extract features from the encoded feature vector of the drug through the drug association network, which is expressed by the formula:
[0029]
[0030] where is the adjacent matrix of the drug in the drug association network.
[0031] Preferably, step S4 specifically includes the following steps:
[0032] S41. Perform adaptive fusion on the drug independent features and drug association features, which is expressed by the formula:
[0033] Z = λ m × Z m + λ d × Z d
[0034] where Z m is the drug independent feature, Z d is the drug association feature, and λ is the fusion parameter.
[0035] The calculation formula for the λ fusion parameter is expressed as:
[0036] where f(·) is the softmax function.
[0037] S42. Construct an MLP prediction model and input the drug combination features for prediction, which is expressed by the formula:
[0038] Y pout = MLP(Concat(Z A , Z B ))
[0039] where Y pout is the predicted label, Z A is the feature of drug A in the drug combination after extraction, and Z B is the feature of drug B in the drug combination after extraction.
[0040] Preferably, the present invention also discloses a computer-readable storage medium with a computer program stored thereon. After the computer program runs, it executes the method described in any one of the above.
[0041] Preferably, the present invention also discloses a computer system, including a processor and a storage medium. The storage medium stores a computer program, and the processor reads and runs the computer program from the storage medium to execute the method described in any one of the above.
[0042] The beneficial effects of the present invention are as follows:
[0043] The present invention proposes a deep learning framework for predicting potential drug-drug interactions, which is a dynamic fusion network for feature extraction based on drug-independent information and drug-associated information. From the perspective of independent molecules, atomic properties and molecular graphs are considered to reflect the chemical and topological characteristics of drug molecules. From the perspective of drug association, a soft-threshold dimensionality reduction network is designed to retain key features. In the feature fusion process, two adaptive parameters are designed to connect these multi-perspective features, improving the prediction performance. Description of the Drawings
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0045] Figure 1 It is a schematic diagram of the principle flow of the method of the present invention;
[0046] Figure 2 It is a schematic diagram of the module structure of the embodiment of the method of the present invention;
[0047] Figure 3 It is a schematic diagram of obtaining drug-associated features of the present invention;
[0048] Figure 4 It is a schematic diagram of the structure of the electronic device of the present invention. Detailed Embodiments
[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.
[0050] For example of implementation Figures 1 to 4 。
[0051] The present invention provides a method for analyzing and predicting drug-drug interactions using deep learning, such as Figure 1 , which is a schematic flowchart of the method for analyzing and predicting drug-drug interactions through deep learning provided by an embodiment of the present invention. The method includes the steps of:
[0052] S1. Encode the characteristic information of the drug to obtain the characteristic vector of the drug combination.
[0053] The specific steps of encoding the characteristic information of the drug include:
[0054] S11. Use RDKit to convert the drug SMILES string into a molecular graph, where the molecular graph nodes represent atoms and the molecular graph edges represent chemical bonds, and the formula is expressed as: G(A)=(V(A), E(A));
[0055] Among them, A represents drug A, where V(A) is the set of nodes, and each node represents an atom in the drug molecular graph; E(A) is the set of edges in the molecular graph, and each edge represents the chemical bond between atoms in the molecular graph;
[0056] S12. Use Node2vec to obtain the substructure feature encoding information of the topological sequence on the molecular graph;
[0057] S13. Obtain the substructure information through the PubMed database.
[0058] Specifically, in order to obtain the characteristic vector of the drug combination, first construct a molecular graph for each small molecule drug, and design a feature extraction module to collect chemical information and identify the molecular graph feature embedding. Small molecule drugs are composed of multiple atoms connected by chemical bonds. Specifically, they can be described by the Simplified Molecular Input Line Entry Specification (SMILES) sequence. Using the RDKit toolkit based on DeepChem, construct a two-dimensional molecular graph for each drug, where the feature matrix Xm∈Rn×d is predefined and contains a 78-dimensional attribute vector for each atom. Each attribute vector contains 5 types of atomic attributes, including atomic symbols, the number of neighboring atoms, the total number of connected hydrogen atoms, implicit valence, and aromaticity and other chemical properties.
[0059] To explore the local structural features around each atom in more depth, a random-walk-based method was adopted to generate substructure sequences starting from each atom. Specifically, the random-walk algorithm starts from an atomic node in the molecular graph. At each step, the current atom is used as the starting point, and the next adjacent atom is randomly selected along its chemical bonds. This process can be repeated multiple times to generate a series of different paths or substructure sequences. The advantage of this method is that it can cover a large number of substructures in the molecular graph through traversal and sampling, capturing the local structural information around each atom. Since these random-walk paths are diverse, it is possible to obtain atom-based diverse substructure sequences globally, thus better describing the overall molecular structure of the drug.
[0060] To further enrich the expression of drug features, the structural fingerprints of drugs were also obtained from the PubChem database, which provides another perspective for drug description. The PubChem substructure fingerprint (PubChemFP) is a high-dimensional encoding method composed of the encodings of 881 key substructures. These substructures are defined by systematically analyzing the data of a large number of compounds and can represent the common chemical substructures that may appear in molecules. Each drug can be represented as an 881-dimensional vector x = (x 1 , x 2 ,..., x 881 ), where each element x i corresponds to the presence or absence of a specific substructure in the drug molecule. This high-dimensional fingerprint vector provides a standardized representation for each drug, facilitating subsequent feature comparison and model training.
[0061] The sparsity and high-dimensional features of this fingerprint vector make it particularly useful in dealing with large-scale drug data. However, since each drug usually contains only a few substructures, this means that in practical applications, the 881-dimensional fingerprint vector is often highly sparse, with most positions having zero information. Although this sparsity brings computational efficiency, it may also pose certain challenges to model performance. Therefore, further optimization and processing are required in feature engineering. This optimization process helps to extract the most representative and predictive features in drug molecules, thus achieving better performance in the drug-drug interaction prediction model.
[0062] S2, input the feature vector of the drug combination into the multi-head graph attention mechanism to separately extract the independent drug molecular features of the two drugs in the drug combination.
[0063] After the above steps, a molecular graph representation was generated for each drug, including atoms, chemical bonds, and predefined atomic properties. Then, the random walk algorithm was used to extract high-dimensional features containing substructure information from each molecular graph. In this algorithm, each atom obtained a new representation. In addition, since the chemical bonds in each molecule are different, their importance also needs to be emphasized. For this reason, a multi-head graph attention network (GAT) was proposed, introducing an attention mechanism to reflect the importance of different chemical bonds.
[0064] Among them, extracting the independent drug molecular features of two drugs in the drug combination specifically includes the following steps:
[0065] S21. Use the multi-head graph attention network multi-head GAT to perform iterative calculations on drug A, and perform feature update and extraction through the i-layer multi-head GAT. In GAT, the node embedding definition formula is expressed as:
[0066]
[0067] Among them, || represents the feature concatenation operation, connecting the output results of multiple attention mechanisms, M represents the maximum multi-head number of GAT, and W represents the weight matrix; the multi-head graph attention neural network we used in the first layer, and the graph attention network in the second layer. The multi-head graph attention network (Multi-Head GAT) can more comprehensively capture information at multiple levels and different relationships in the graph by parallel learning multiple sets of attention weights, improving the feature learning ability of the graph neural network.
[0068] The attention coefficient α between each input node i and its first-order neighbors in the graph i,j is calculated as follows:
[0069]
[0070] where a T ∈R C' is a learnable weight vector, T is the corresponding transpose, and elu is a non-linear activation function. When x is negative, y is equal to 0. Then, the'softmax' function is introduced to normalize all neighbor nodes j of node i for simple calculation and comparison. On the drug molecular graph, GAT has multiple advantages. First, it can flexibly learn the relationship weights between nodes, thus better capturing the complex structure in the drug molecule.
[0071] S22. Use the global sum pooling gap to process the drug to obtain the drug information feature vector of each drug, and the formula is expressed as: Z m =gap(v 1 ,v 2 ,...,vn )。
[0072] S3. Construct the association network of all drugs, and use the multi-layer graph neural network to extract and obtain drug association features.
[0073] Through the study of the PubChemFP sequence information, it is found that the information at most positions in the sequence is 0, indicating that the information of the sequence is relatively sparse. The original input of this feature may have an adverse impact on the performance of our model, so we choose to perform dimensionality reduction processing on the sequence information.
[0074] Among them, training with drug information to obtain drug association features includes the steps of:
[0075] S31. When performing dimensionality reduction encoding on the original sequence, we introduce a soft-threshold operation function to ensure that key feature information is retained after dimensionality reduction processing of the sequence. Encoding the substructure information of the drug PubMed sequence into substructure features using the soft-threshold function, which is expressed by the formula:
[0076] x kout = sign(x kin ) × ReLU(|x kin |- θ)
[0077] where x kin and x kout are the input and output of neuron k in the layer, and θ is the threshold.
[0078] The soft-threshold function performs an element-wise operation on each element in the input vector x kin , and linearly compresses the elements greater than the set threshold, thereby obtaining a relatively key output vector.
[0079] S32. Extract features from the drug-encoded feature vector through the drug association network, which is expressed by the formula:
[0080]
[0081] where is the adjacent matrix of the drug in the drug association network.
[0082] S4. Fuse and reclassify the independent drug molecule features and drug association features to predict drug-drug interactions.
[0083] Step S4 specifically includes the following steps:
[0084] S41. For the features Z a and Z b learned from the two perspectives of intra-molecular and inter-molecular of the drug obtained respectively,, an adaptive-weight feature fusion method is used to fuse two features of a single drug. To effectively combine the internal structural features of the drug and the associative features with other molecules, a learned weight vector is used to perform a weighted combination of the two features, and the embedding is converted into feature weights through a non-linear transformation, as follows:
[0085]
[0086] where W is the weight matrix and β is the bias matrix, and the adaptive weight coefficients λ m and λ d can be obtained respectively from the self-calculation of the features Zm and Zd. For the combination of the two features of each drug, the final drug feature is obtained as follows:
[0087] Z = λ m ×Z m +λ d ×Z d
[0088] S42. Construct an MLP prediction model to determine whether two given drugs interact. The MLP model includes three hidden layers. After each hidden layer, the ReLU function is applied for non-linear transformation, and dropout regularization is used to prevent overfitting. Finally, the output layer performs the final linear transformation to obtain the classification prediction result, denoted as out. The formula is expressed as:
[0089] Y pout = MLP(Concat(Z A , Z B ))
[0090] where Y pout is the predicted label, Z A is the feature of drug A in the drug combination after extraction, and Z B is the feature of drug B in the drug combination after extraction.
[0091] The probability (classification label) of drug interaction is calculated using the Softmax function, and its input is the final output of the fully connected (FC) layer. The calculation process is as follows:
[0092] p t = softmax(W out ·(Z drugA ||Z drugB ) + b out )
[0093] where p t represents the probability of class t, and W out and b outThey are the weight matrix and bias vector of the output layer, representing the embedded features of the two drugs in the drug combinations learned from the previous layers.
[0094] For a given set of combinations with labels, we use cross-entropy as the loss function to train the model, aiming to minimize the loss during the training process. The expression is as follows:
[0095]
[0096] where Θ represents the set of all trainable weight and bias parameters involved in the model, N is the total number of samples in the training dataset, t i is the i-th label, and λ is the L 2 regularization hyperparameter.
[0097] Furthermore, the present invention adopts the AdamW optimizer to optimize the training process. The training process includes alternating training between single drugs and drug combinations, promoting the model to improve drug feature learning, and allowing the evaluation of drug combinations in the validation set after training.
[0098] Furthermore, the performance of the drug combination synergistic effect screening and prediction method of the present invention can be evaluated by the following indicators:
[0099] Recall(Rec) = TP / (TP + FN);
[0100] Precision(Pre) = TP / (TP + FP);
[0101] F1 = 2·(Rec·Pre) / (Rec + Pre);
[0102] Accuracy(Acc) = (TP + TN) / (TP + FN + FP + TN);
[0103] where TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, FN represents the number of false negative samples, AUC is synonymous with AUROC, representing the area under the ROC curve. AUPR is the area under the precision-recall (PR) curve.
[0104] The following examples are given to illustrate the effect of the method of the present invention with experimental data:
[0105] Experimental results: The present invention compares with the methods of predicting drug targets of the current four leading deep learning methods. To comprehensively evaluate the performance, all the comparison methods were run 5 times with 5-fold cross-validation on all drug-drug datasets, and the average experimental results are shown in Table 1:
[0106] Table 1: Comparison methods of advanced models on different datasets
[0107]
[0108] The present invention conducts a comprehensive ablation experiment to analyze the importance of each key point in the model. By sequentially removing the key components of the model, the specific impact on the screening prediction performance is analyzed. These variants include:
[0109] 1) The present invention cancels the original adaptive feature fusion strategy and instead uses a simple feature concatenation method. (noadaptive fusion)
[0110] 2) The present invention removes the soft-threshold feature dimensionality reduction processing in the drug association feature module. (no soft-threshold)
[0111] 3) The present invention removes the atomic sequence feature from the independent drug molecular feature module. (no substructuresequence)
[0112] 4) Under this condition, only the molecular features of the drug are used, and the extraction of drug-drug interaction (DDI) graph features is not considered. (no DDI graph feature extraction)
[0113] 5) In this experimental setting, only drug association features are used, and the extraction of drug molecular features is completely ignored. (no molecular feature extraction)
[0114] Table 2 Experiments of different modules of the model on the stnn dataset
[0115]
[0116] Case Study: This case study aims to evaluate the practical ability of the model in identifying unknown drug-drug interactions (DDIs). When conducting the case study, we first used all the known DDIs in the benchmark dataset as positive samples, and these sample data were collected from DrugBank 5.0.3. These known DDIs were used to form the training dataset for training our model. To generate negative samples, we constructed them by randomly selecting drug pairs that did not show interactions in the positive samples. Specifically, we randomly paired the drug pairs and ensured that these drug pairs had no recorded interactions in the actual drug interaction data. In this way, we were able to ensure that the negative samples included in the training dataset could effectively simulate the non-interaction situation between drug pairs in the real world. In the prediction phase, we set a threshold of 0.5 to determine the predicted DDIs. If the prediction score of the model for a drug pair is greater than 0.5, then the drug pair is predicted to have an interaction; otherwise, it is predicted to have no interaction. This method ensures that our model can provide clear prediction results when outputting. We first predicted the potential top 20 DDIs from all unknown drug pairs. Next, we searched for the actual evidence of these prediction results in relevant databases and literature to verify the prediction accuracy of the model. All the results are summarized in Table 3. In this way, we can intuitively evaluate the actual performance of the model in predicting interactions between unknown drug pairs and verify its effectiveness and reliability in real scenarios. Table 3. The top 20 results predicted by the model
[0117]
[0118]
[0119] In addition, to deeply study the potential drug-drug interactions (DDIs) that may be related to specific diseases, we conducted a targeted analysis. We collected two groups of drugs, with each group of drugs targeting a specific disease. These diseases are in the areas we are interested in, and the drug combinations were carefully selected to ensure their relevance to the diseases. For each group of drugs, we used the model to predict the potential DDIs related to them. These prediction results helped us identify the possible drug interactions that may exist in the treatment of specific diseases, thus providing guidance for further experimental research and clinical applications. All the prediction results are shown in detail in Tables 4 and 5 to facilitate the analysis and comparison of the potential DDIs of each group of drugs. In this way, we can not only verify the prediction ability of the model in different disease areas but also explore the important possible interactions between drugs, thus providing valuable insights for the treatment of diseases and drug development. Table 4. The top 10 predicted DDIs of drugs related to breast cancer.
[0120]
[0121]
[0122] Table 5. The top 10 predicted DDIs of drugs related to parasites
[0123]
[0124]
[0125] The present invention also provides a computer system, as Figure 4 shown, comprising a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory complete communication with each other through the communication bus. The processor can call the logical instructions in the memory to execute the following methods:
[0126] S1. Encode the characteristic information of the drugs to obtain the characteristic vectors of the drug combinations;
[0127] S2. Input the drug characteristics into the multi-head graph attention mechanism to separately extract the independent drug molecule characteristics of the two drugs in the drug combination;
[0128] S3. Construct the association network of all drugs, and use the multi-layer graph neural network to extract and obtain the drug association characteristics;
[0129] S4. Fuse and reclassify the drug independent characteristics and the drug association characteristics to predict drug-drug interactions.
[0130] When the above-mentioned logical instructions in the memory are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.
[0131] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is configured to execute the methods provided in the above embodiments, including:
[0132] S1. Encode the characteristic information of the drug to obtain the characteristic vector of the drug combination;
[0133] S2. Input the drug characteristics into the multi-head graph attention mechanism to separately extract the independent drug molecule characteristics of the two drugs in the drug combination;
[0134] S3. Construct the association network of all drugs, and use the multi-layer graph neural network to extract and obtain the drug association characteristics;
[0135] S4. Fuse and reclassify the drug independent characteristics and the drug association characteristics to predict drug-drug interactions.
[0136] The device embodiments described above are merely illustrative. 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 may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.
[0137] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, also by hardware. Based on this understanding, the above technical solutions, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., including several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.
[0138] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications 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.
Claims
1. A method for predicting drug-drug interactions based on deep learning, characterized in that: The following steps are involved: S1, encode the characteristic information of the drugs and obtain the characteristic vector of the drug combination; S2, inputs the feature vector of the drug combination into the multi-head graph attention mechanism to extract the independent drug molecular features of the two drugs in the drug combination respectively; S3, constructs the association network of all drugs and uses a multi-layer graph neural network to extract and obtain drug association features; S4, independent drug molecular features and drug association features are integrated and reclassified to predict drug-drug interactions.
2. The method according to claim 1, characterized in that The specific steps of encoding the characteristic information of the drug in step S1 include: S11. Use RDKit to convert the drug SMILES string into a molecular graph. The nodes of the molecular graph represent atoms, and the edges of the molecular graph represent chemical bonds. The formula is: G(A) = (V(A), E(A)); Where A represents drug A, V(A) is a node set, each node represents an atom in the drug molecule graph; E(A) is a set of edges in the molecule graph, each edge represents a chemical bond between atoms in the molecule graph; S12, using Node2vec to obtain substructure feature encoding information of topological sequences on the molecular graph; S13. Obtain substructure information through PubMed database.
3. The method according to claim 1, characterized in that Step S2 extracts the independent drug molecular features of the two drugs in the drug combination, specifically including the following steps: S21. Use the multi-head GAT to iteratively calculate drug A, and update and extract features through i-layer multi-head GAT. The formula is expressed as: Among them, || represents the feature concatenation operation, M represents the maximum number of multi-heads of GAT, and W represents the weight matrix; Represents the attention coefficient, which represents the importance of the characteristics of the drug's i node to the characteristics of the j node. The calculation formula is expressed as: where a T ∈R C ' is the learnable weight vector; S22. Use global additive pooling gap to process the drugs and obtain the drug information feature vector of each drug. The formula is: Z m =gap(v1,v2,...,v n ).
4. The method according to claim 1, characterized in that: In S3, drug information training is used to obtain drug association features, including the following steps: S31. Encode the substructure information of the drug PubMed sequence into substructure features using a soft threshold function. The formula is: x kout =sign(x kin )×ReLU(Px kin P-θ) Where xkin and xkout are the input and output of neuron k in the layer, and θ is the threshold; S32, extracting the drug coding feature vector through the drug association network, and the formula is expressed as: in It is the adjacency matrix of drugs in the drug association network.
5. The method according to claim 1, characterized in that: Step S4 The specific steps include: S41. Adaptively fuse drug-independent features and drug-associated features. The formula is: Z=λ m ×Z m +λ d ×Z d Among them, Z m Drug-independent characteristics, Z d is the drug association feature, λ is the fusion parameter; The calculation formula of the λ fusion parameter is expressed as: Where f(·) is the softmax function; S42. Construct an MLP prediction model and input the drug combination features for prediction. The formula is: Y pout =MLP(Concat(Z A ,Z B )) Among them, Y pout is the predicted label, Z A is the extracted feature of drug A in the drug combination, Z B It is the characteristic of drug B in the drug combination after extraction.
6. A computer-readable storage medium, characterized in that: A computer program is stored on the medium, and after the computer program is run, the method according to any one of claims 1 to 5 is executed.
7. A computer system, characterized in that: The method comprises a processor and a storage medium, wherein a computer program is stored in the storage medium, and the processor reads and runs the computer program from the storage medium to execute the method as claimed in any one of claims 1 to 5.