A deep learning-based echinococcosis drug combination synergy prediction method, device, equipment and medium

Abstract

The application discloses a kind of based on deep learning's echinococcosis drug combination synergistic prediction method, device, equipment and medium, wherein method includes: using ResGAT network obtains drug information feature vector, constructs echinococcosis target protein PPI network, using node2vec network obtains target protein feature vector, aggregation is used as echinococcosis information feature vector;Training and constructing single drug MLP prediction model;Single drug MLP prediction model and Bliss are used to form drug combination synergistic prediction model, and drug combination synergistic prediction result is obtained.The application sufficiently utilizes drug combination chemistry molecular graph and echinococcosis single drug data information, improves the accuracy of drug combination joint screening prediction, effectively solves the problem of insufficient training data of drug combination, improves the accuracy of sample data limited echinococcosis drug combination screening prediction.

Description

Technical Field

[0001] This invention relates to the field of drug combination synergistic prediction technology for echinococcosis, and particularly to a method, apparatus, equipment and medium for echinococcosis drug combination synergistic prediction based on deep learning. Background Technology

[0002] Hydatid disease is a zoonotic parasitic disease caused by the larvae of Echinococcus tapeworms. It is divided into cystic hydatid disease (CE) caused by Echinococcus granulosus and alveolar hydatid disease (AE) caused by Echinococcus multilocularis. The disease manifests as the formation of cysts in organs (mainly the liver and lungs), leading to decreased organ function, swelling, and pain. Drugs play a crucial role in the treatment of hydatid disease; antiparasitic drugs, antitumor drugs, and certain plant extracts are widely used, especially benzimidazole due to its widespread application. However, long-term use of these drugs may lead to adverse reactions in the skin and mucous membrane systems and potential teratogenicity. Furthermore, hydatid disease is prevalent in remote pastoral areas, and most high-risk populations live in poverty, highlighting the urgent need for cost-effective and efficient drugs.

[0003] Current medications for echinococcosis are mainly divided into single-agent drugs and drug combinations. Considering the limitations of single-agent drugs, including drug tolerance and resistance, there is increasing interest in drug combinations. This approach aims to enhance therapeutic efficacy and slow the development of drug resistance by using drugs that act on different mechanisms or targets. However, not all drug combinations produce the expected synergistic effects. Existing drug combinations still exhibit side effects or poor efficacy, and traditional drug development processes require significant time and resources to screen candidate compounds. Drug screening prediction, using computational methods, can rapidly assess the potential activity and toxicity of a large number of molecules. Therefore, in-depth research into drug synergies can not only deepen our understanding of drug interactions but also provide a wider range of options for clinical drug screening. It can help identify better drug combinations from existing FDA-approved drugs to mitigate the side effects of traditional benzimidazole and shorten clinical trial time, which is crucial for improving the treatment efficacy and safety of echinococcosis.

[0004] Patent document CN112927766B discloses a method for screening combination drugs for a disease. The scheme determines the final score of a drug combination by calculating the correlation coefficient between the gene target sets corresponding to any two candidate drug combinations from N candidate drugs and the network distance score of the drugs; then, it selects the candidate drug combinations with the highest final scores as the combination drugs for the target disease. However, this scheme has the following drawbacks: 1. It is not suitable for drug combination screening and prediction for echinococcosis with limited sample data, and cannot improve the accuracy of drug combination screening and prediction for echinococcosis. 2. It does not effectively obtain the potential feature vectors of the drugs, resulting in poor capture of the interaction relationships between drug molecules. 3. It does not effectively utilize the target protein PPI network, and uses node2vec and attention weight aggregation methods to derive disease features, resulting in poor disease feature expression. Summary of the Invention

[0005] The purpose of this invention is to provide a method, device, equipment, and medium for synergistic prediction of drug combinations for echinococcosis based on deep learning, which solves the problems of long time consumption, high resource consumption, and difficulty in shortening clinical trial time in existing synergistic prediction of drug combinations for echinococcosis.

[0006] This invention provides a method, apparatus, device, and medium for synergistic prediction of drug combinations for echinococcosis based on deep learning.

[0007] The first aspect: a deep learning-based collaborative prediction method for echinococcosis drug combinations, including:

[0008] S1. Establish positive and negative samples for single drugs and drug combinations to obtain SMILES information for single drugs and drug combinations for echinococcosis.

[0009] S2. Obtain target protein information for echinococcosis and construct a target protein PPI network;

[0010] S3. Use the ResGAT network to process information on single drugs and drug combinations SMILES to obtain drug information feature vectors;

[0011] S4. For the target protein in the PPI network, use the node2vec network with biased random walk to obtain the feature vector of the target protein.

[0012] S5. Use a method incorporating attention concepts to weight and aggregate target protein feature vectors as echinococcosis information feature vectors;

[0013] S6. Use drug information feature vectors and echinococcosis information feature vectors to train and construct a single-drug MLP prediction model;

[0014] S7. Use the single-drug MLP prediction model and Bliss to form a drug combination synergistic prediction model to obtain the drug combination synergistic prediction results.

[0015] Furthermore, in step S3, the ResGAT network is used to process information on single drugs and drug combinations (SMILES) to obtain drug information feature vectors, including the following steps:

[0016] S31. Convert the drug SMILES string into a molecular graph, where nodes represent atoms and edges represent chemical bonds. The formula is as follows:

[0017] (1)

[0018] Where A represents drug A, This is a set of nodes, where each node represents an atom in the drug molecule diagram. It is the set of edges in a molecular diagram, where each edge represents a chemical bond between atoms in the molecular diagram.

[0019] S32. Iteratively calculate drug A using a ResGAT network enhanced with residual structure. After each layer iteration, add initial encoded features, expressed by the formula:

[0020] (2)

[0021] Where 1≤i≤n represents the number of atoms in the molecule diagram of drug A, l=0, 1, 2, ..., N represents the number of ResGAT network layers. This represents the characteristic of atom i output from the l-th layer of the drug A molecule diagram. This represents the initial coding features of atom i;

[0022] S33. The function F(∙) represents the operation of the ResGAT network, which aggregates features of all neighboring nodes in the molecular graph through a message-passing mechanism, performing a weighted sum based on interest to obtain the updated node feature vector. The formula is expressed as:

[0023] (3)

[0024] in, Represents atoms The set of adjacent nodes;

[0025] The attention coefficient represents the drug. middle Node feature pairs The importance of node features; The formula for representing is:

[0026] (4)

[0027] in, represents the shared attention mechanism vector, and || represents the feature concatenation operation;

[0028] S34. Process drug A using global max pooling (GMP) to obtain the drug information feature vector of drug A, expressed by the formula:

[0029] (5)

[0030] Where n represents the number of atoms in drug A molecule, Let be the drug information feature vector of drug A.

[0031] Furthermore, in step S5, the target protein feature vector is weighted and aggregated using an attention-based approach as the echinococcosis information feature vector, including the following steps:

[0032] S51. Based on the importance of target proteins to echinococcosis, an attention mechanism is used to assign weights to the importance of target proteins, expressed by the following formula:

[0033] (6)

[0034] in, This represents the dot product operation. This represents the initial characteristics of unweighted echinococcosis. The feature vector representing the target protein. Target protein feature vector The weights;

[0035] S52. In a PPI network, all proteins in each layer are linearly combined to form the feature vector of the target protein in that layer, expressed by the formula:

[0036] (7)

[0037] Where n is the number of target protein layers, 0 ≤ i ≤ n, The normalized value representing the weight of the target protein associated with echinococcosis in the i-th layer is expressed by the formula:

[0038] (8)

[0039] S53. Integrate the target protein features of all layers to obtain the echinococcosis information feature vector of echinococcosis 𝑤, expressed by the formula:

[0040] (9)

[0041] in, and These represent the aggregate weight and the bias, respectively.

[0042] Further, in step S6, a single-drug MLP prediction model is trained and constructed using drug information feature vectors and echinococcosis information feature vectors. This includes obtaining the anti-echinococcosis activity score of a single drug using the single-drug MLP prediction model, expressed by the formula:

[0043] (10)

[0044] in, Drug A fused feature vector, , This represents the MLP prediction model. Let A be the feature vector of drug A. Using the feature vector of echinococcosis information, the MLP prediction model is trained through cross-entropy to predict the anti-echinococcosis activity of a single drug. The higher the value, the more effective drug A is in treating echinococcosis.

[0045] Furthermore, in S7, a drug combination synergistic prediction model is formed by using a single-drug MLP prediction model and Bliss, including:

[0046] S71, Obtaining Drug Combinations fused feature vector The formula is expressed as:

[0047] (11)

[0048] in, Drug A fused feature vector, , Drug B fusion feature vector, ;

[0049] S72, Obtaining Drug Combinations Anti-echinococcosis activity fraction The formula is expressed as:

[0050] (12)

[0051] in, Indicates drug combination The fused feature vector, This represents the MLP prediction model.

[0052] S73, Defining Drug Combinations Expected activity fraction against echinococcosis The formula is expressed as:

[0053] (13)

[0054] in, This represents the activity fraction of drug A alone. This represents the activity fraction of drug B alone. drug combination Expected activity fraction;

[0055] S74, Calculating Drug Combinations Synergistic effect score The formula is expressed as:

[0056] (14)

[0057] Among them, if drug combination Activity fraction against echinococcosis Higher than drug combination Expected activity fraction This indicates a combination It has a synergistic effect.

[0058] The second aspect: a deep learning-based echinococcosis drug combination collaborative prediction device, comprising:

[0059] The drug feature extraction module is used to obtain drug information feature vectors;

[0060] The disease feature extraction module is used to obtain feature vectors of echinococcosis information;

[0061] The single-drug MLP prediction model module obtains the drug activity score based on the drug information feature vector and the echinococcosis information feature vector.

[0062] The Bliss collaborative screening model module uses drug activity scores to collaboratively screen and obtain collaborative prediction results for drug combinations.

[0063] Third aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0064] Fourth aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method as provided in the first aspect.

[0065] The beneficial effects of this invention are:

[0066] 1. This invention improves the accuracy of drug combination screening prediction by making full use of drug combination chemical molecular maps and single drug data information for echinococcosis, effectively solving the problem of insufficient training data for drug combinations and improving the accuracy of drug combination screening prediction for echinococcosis with limited sample data.

[0067] 2. This invention uses a residual attention mechanism network (resGAT) to obtain the latent feature vector of a drug. The residual attention mechanism network (resGAT) provides a better understanding of drug features and interactions. Residual connections help alleviate the gradient vanishing problem common in neural networks. Residual connections allow gradients to flow directly through the network, thus preventing performance loss when training deeper models. The attention mechanism with residual connections more effectively captures the interactions between drug molecules and uses the chemical bond relationship information in the molecular graph to more accurately represent drug features.

[0068] 3. This invention constructs a PPI network for echinococcosis targets and uses node2vec and attention weight aggregation to derive disease features. This method conducts in-depth research on PPI networks related to echinococcosis, learns low-dimensional feature representations of nodes through node2vec, and aggregates these features using an attention mechanism. This not only enhances the expressive power of the features but also captures important nodes and relationships in the network, providing deeper biological insights for the study of echinococcosis. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating a deep learning-based synergistic prediction method for echinococcosis drug combinations according to the present invention.

[0070] Figure 2 This is a schematic diagram of the structure of a deep learning-based hydatid disease drug combination synergistic prediction device according to the present invention.

[0071] Figure 3 This is a schematic diagram of the structure of the drug feature extraction module of the present invention;

[0072] Figure 4 This is a schematic diagram of the disease feature extraction module of the present invention;

[0073] Figure 5 This is a schematic diagram of the process for constructing the single-drug MLP prediction method of this invention;

[0074] Figure 6 This is a schematic diagram of the process for constructing the drug combination synergistic prediction method of the present invention;

[0075] Figure 7 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0076] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0077] Existing drug combinations for echinococcosis still have some side effects or poor efficacy. The traditional drug development process requires a lot of time and resources to screen candidate compounds, which is not conducive to shortening the clinical trial time and is not conducive to the treatment effect and safety of echinococcosis.

[0078] To address the above problems, this invention provides a deep learning-based method for synergistic prediction of echinococcosis drug combinations. Figure 1 This is a flowchart illustrating a deep learning-based synergistic prediction method for echinococcosis drug combinations provided in an embodiment of the present invention. The method includes:

[0079] S1. Establish positive and negative samples for single drugs and drug combinations to obtain SMILES information for single drugs and drug combinations for echinococcosis.

[0080] Information on the efficacy of single-drug and drug combination treatments for echinococcosis can be obtained through literature mining. This involves retrieving relevant literature on echinococcosis and extracting data on single-drug and drug combination treatments that clearly describe "effective treatment results." For example, this invention, by searching literature websites such as PubMed and CNKI, identified 63 drugs for treating cystic echinococcosis (CE) and 54 drugs for treating alveolar echinococcosis (AE), establishing positive samples for single-drug treatments. For drug combinations, there were 24 drugs for CE and 26 drugs for AE, establishing positive samples for drug combinations.

[0081] In addition to positive sample drugs, FDA-approved drugs were used as single-drug negative samples. Positive and negative samples were used to train the single-drug anti-echinococcosis activity model. In drug combinations, drug combinations that were clearly marked as effective in the literature were considered positive samples. Combinations of single drugs involved in positive samples of drug combinations that were not clearly marked in the literature were considered negative samples. Detailed data information is shown in Table 1.

[0082] Table 1 Drug Information

[0083]

[0084] The SMILES information corresponding to the drug is obtained by querying in PubChem and used as the drug input information of this invention.

[0085] S2. Obtain target protein information for echinococcosis and construct a target protein PPI network.

[0086] When predicting and screening drug combinations for echinococcosis, considering the underlying biological mechanisms and the specificity of echinococcosis is crucial for accurately characterizing the disease, which helps in accurately screening effective drug combinations.

[0087] A PPI network for echinococcosis was constructed based on target proteins associated with echinococcosis.

[0088] First, target proteins associated with echinococcosis can be collected from the Harmonizome database, including 35 direct targets of cystic echinococcosis (CE) and 61 direct targets of alveolar echinococcosis (AE). Harmonizome is a comprehensive bioinformatics database that includes information on target proteins associated with echinococcosis derived from abstracts of biomedical publications.

[0089] To identify proteins that are directly related to echinococcosis, human protein maps were used to collect neighboring target proteins that interact with the target.

[0090] Based on their interrelationships, target proteins are stratified into direct target proteins, second-order target proteins, etc. Second-order target proteins indicate that the protein is a sequence neighbor of the target protein of echinococcosis.

[0091] For the 35 direct target proteins of cystic echinococcosis, a total of 276 interaction relationships were collected, of which 128 were second-order target proteins. Detailed data are shown in Table 2.

[0092] Table 2 Target protein information

[0093]

[0094] For the 61 direct target proteins of alveolar echinococcosis, a total of 457 interaction relationships were collected, of which 424 were second-order target proteins.

[0095] S3. Use the ResGAT network to process information on single drugs and drug combinations SMILES to obtain drug information feature vectors.

[0096] like Figure 3 As shown, the open-source cheminformatics software RDKit can be used to convert the drug SMILES string into a molecular graph. Molecular graph nodes represent atoms, and molecular graph edges represent chemical bonds. The formula is as follows:

[0097] (1)

[0098] Where A represents drug A, This is a set of nodes, where each node represents an atom in the drug molecule diagram. It is the set of edges in a molecular diagram, where each edge represents a chemical bond between atoms in the molecular diagram.

[0099] The ResGAT network with residual structure enhancement is used to iteratively compute drug A. After each layer iteration, initial encoded features are added, as expressed by the formula:

[0100] (2)

[0101] Where 1≤i≤n represents the number of atoms in the molecular diagram of drug A, and l=0, 1, 2, …, N represents the number of ResGAT network layers. This represents the characteristic of atom i output from the l-th layer of the drug A molecule diagram. This represents the initial coding characteristics of atom i.

[0102] Using residual structures enhances the representational power of the GAT network, facilitating the capture of complex features and patterns. The initial encoded features combined with the features obtained from N layers of ResGAT can better capture subtle features and patterns in the data, thereby enhancing the network's expressive power and generalization performance.

[0103] The function F(∙) represents the operation of the ResGAT network. Nodes in the molecular graph are aggregated through a message-passing mechanism. The features of all neighboring nodes are summed based on interest to obtain the updated node feature vector, as expressed by the formula:

[0104] (3)

[0105] in Represents atoms The set of adjacent nodes.

[0106] The attention coefficient represents the drug. middle Node feature pairs The importance of node features; The formula for representing is:

[0107] (4)

[0108] in, Let || represent the shared attention mechanism vector, and || represent the feature concatenation operation; after N layers of operations, the feature representation of each atom in drug A is obtained.

[0109] Finally, global max pooling (GMP) is used to process drug A to obtain the drug information feature vector of drug A, expressed by the formula:

[0110] (5)

[0111] Where n represents the number of atoms in drug A molecule, Let be the drug information feature vector of drug A.

[0112] S4. For the target protein in the PPI network, use a biased random walk node2vec network to obtain the feature vector of the target protein.

[0113] like Figure 4 As shown, the node2vec network is used to perform a random walk on the PPI network associated with echinococcosis, generating the topological relationship of target protein node sequences, and the feature vector of each target protein in the PPI network is derived from the node sequence topological relationship.

[0114] For each target protein node, multiple random walks are performed to obtain a random walk sequence. During each walk, the transition probability function is used to determine the sequence. The probability of transitioning from the current node to the next node is calculated, and sampling is performed based on these transition probabilities to obtain the next node to be sampled, ultimately generating a node sequence. Finally, the SkipGram model is used to transform these random walk sequences into node embedding vectors. Specifically:

[0115] For each target protein node v, the probability of visiting the next node x is:

[0116] (15)

[0117] in, Z represents the transition probability between two nodes, and Z represents the normalization constant.

[0118] in, The calculation is as follows:

[0119]

[0120] in, This represents the weighting coefficient between node v and node x. The representation is as follows:

[0121] (16)

[0122] Where (t, x) represents the edges traversed by a node during a random walk. This represents the shortest path distance between node t and node x. p and q are hyperparameters used in the walk to control the node direction during the random walk. When the value of p is high, the random walk tends to return to previously visited nodes; when the value of q is high, the random walk tends to perform a breadth-first search.

[0123] After obtaining the sequences generated by the random walk, these sequences are used as input to the SkipGram part. The objective function of the SkipGram model is to maximize the co-occurrence probability between node pairs, that is, given a node, predict its context nodes to obtain the feature vector of the node.

[0124] Through this objective function, the node2vec network can simultaneously learn the probability distribution of nodes and the similarity relationships between nodes, thereby generating meaningful node embeddings. The specific objective function is as follows:

[0125] (17)

[0126] Where V represents the set of all nodes in the graph; Let u be the set of its neighboring nodes, which are determined by the transition probabilities in a random walk. Nodes in the generated sequence; This represents the learned feature vector of node u.

[0127] The objective function aims to maximize the sum of the log probabilities of all nodes. This ensures that the learned node feature vectors reflect the structural information of the graph to the greatest extent possible, making similar nodes as close as possible in the feature vector space. The optimization process uses gradient descent to continuously adjust the node feature vectors until the objective function reaches its maximum value, enabling Node2Vec to learn the low-dimensional feature vector of each node.

[0128] For each target protein node in the PPI network associated with echinococcosis, the above method was used to... Optimization is performed, resulting in the final optimized node features. As the feature vector of the target protein, for ease of understanding, we will use vector p to represent the feature vector of the target protein obtained after training and optimization.

[0129] S5. Use an attention-based approach to weight and aggregate target protein feature vectors as echinococcosis information feature vectors.

[0130] Each target protein has a different importance in echinococcosis. Based on the importance of the target proteins in echinococcosis, an attention-based approach is used to assign weights to the importance of the target proteins. The formula is as follows:

[0131] (6)

[0132] in, This represents the dot product operation. This represents the initial characteristics of unweighted echinococcosis. The feature vector representing the target protein. target protein The weights of the feature vectors;

[0133] After obtaining the importance weight of each target protein through dot product, the target proteins in each layer are linearly combined, as expressed by the formula:

[0134] (7)

[0135] Where n is the number of target protein layers, 0 ≤ i ≤ n, The normalized value representing the weight of the target protein associated with echinococcosis in the i-th layer is expressed by the formula:

[0136] (8)

[0137] Then, the target protein features of all layers are integrated to obtain the echinococcosis information feature vector of echinococcosis w, expressed by the formula:

[0138] (9)

[0139] in, and These represent the aggregate weight and the bias, respectively.

[0140] S6. Use drug information feature vectors and echinococcosis information feature vectors to train and construct a single-drug MLP prediction model;

[0141] like Figure 5 As shown, the anti-echinococcosis activity score of a single drug is obtained using a single-drug MLP prediction model, expressed by the formula:

[0142] (10)

[0143] in, Drug A fused feature vector, , This represents the MLP prediction model. Let A be the feature vector of drug A. Using the feature vector of echinococcosis information, the MLP prediction model is trained through cross-entropy to predict the anti-echinococcosis activity of a single drug. The higher the value, the more effective drug A is in treating echinococcosis.

[0144] S7. Use drug information feature vectors and echinococcosis information feature vectors to train and construct a single-drug MLP prediction model, such as... Figure 6 As shown, a drug combination synergistic prediction model is constructed using a single-drug MLP prediction model and Bliss, including the following steps:

[0145] S71, Obtaining Drug Combinations fusion feature vector The formula is expressed as:

[0146] (11)

[0147] in, Drug A fused feature vector, , Drug B fusion feature vector, ;

[0148] S72, Obtaining Drug Combinations Anti-echinococcosis activity fraction The formula is expressed as:

[0149] (12)

[0150] in, Indicates drug combination The fused feature vector, This represents the MLP prediction model.

[0151] S73, Defining Drug Combinations Expected activity fraction against echinococcosis The formula is expressed as:

[0152] (13)

[0153] in, This represents the activity fraction of drug A alone. This represents the activity fraction of drug B alone. drug combination Expected activity fraction;

[0154] S74, Calculating Drug Combinations Synergistic effect score The formula is expressed as:

[0155] (14)

[0156] Among them, if drug combination Activity fraction against echinococcosis Higher than drug combination Expected activity fraction This indicates a combination It has a synergistic effect.

[0157] Furthermore, this invention minimizes the weighted sum of the two-part loss function during training, expressed as: ,in Indicates the loss of a single drug task. The loss representing the drug combination is calculated using the binary cross-entropy loss function, as shown below:

[0158] (18)

[0159] Where N is the total number of training samples. The actual label value of the sample. These are the predicted values ​​for the training samples.

[0160] Furthermore, this invention employs the AdamW optimizer to optimize the training process, which includes alternating training between single drugs and drug combinations, promoting model improvement in drug feature learning, and allowing evaluation of drug combinations on a post-training validation set.

[0161] Furthermore, the performance of the drug combination synergy screening and prediction method of the present invention can be evaluated using the following formula:

[0162] (19)

[0163] (20)

[0164] (twenty one)

[0165] (twenty two)

[0166] 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 and represents the area under the ROC curve, and AUPR is the area under the precision-recall (PR) curve.

[0167] The following experimental data illustrates the effectiveness of the method of the present invention:

[0168] This invention evaluates the performance of five leading drug synergy combination screening and prediction methods based on graph neural networks. All comparison algorithms were trained and validated only on drug combinations for screening echinococcosis. The prediction results of the five drug synergy combination screening and prediction algorithms on the CE and AE datasets are shown in Tables 3 and 4.

[0169] Table 3 Comparison results of different algorithms on the CE dataset

[0170]

[0171] Table 4 Comparison results of different algorithms on the AE dataset

[0172]

[0173] To explore the contributions of ResGAT, the initial coding feature representation of echinococcosis, and the single-drug prediction module to this invention, an ablation experiment was employed. The specific impact on screening prediction performance was analyzed by sequentially removing key components of the model. These variations included:

[0174] 1) No attention mechanism (no_resGAT): Drug features are obtained directly through RdKit processing without using an attention network with residual connections to integrate neighbor node information in the drug molecule graph;

[0175] 2) The target protein-target protein interaction PPI network (no_ppi) is not considered, that is, only drug feature information is considered, including single drug and drug combination information, and ResGAT is used to process all input drugs;

[0176] 3) No single-drug supervised training (no_single), only drug combination data, combined with PPI network features of echinococcosis, and drug features are processed by ResGAT.

[0177] The algorithm proposed in this invention and its variants were comprehensively evaluated on the CE and AE datasets using multiple evaluation metrics such as AUC, AUPR, ACC, Recall, Precision, and F1 Score. The results are shown in Tables 5 and 6.

[0178] Table 5 Ablation experiments on the CE dataset

[0179]

[0180] Table 6 Ablation experiments on the AE dataset

[0181]

[0182] Experimental data analysis involved randomly selecting five positive samples from the dataset as an independent test set. The remaining samples were used to train the model, and the trained model was saved. The saved model was then used to screen and predict these five independent test samples, yielding corresponding scores, as shown in Tables 7 and 8. Analysis indicates that the scores of these five positive samples were all high, demonstrating that this invention can relatively accurately screen and predict effective drug combinations for echinococcosis among approved drugs.

[0183] Table 7. Predictive performance of independent test samples for screening cystic echinococcosis (CE).

[0184]

[0185] Table 8. Predictive performance of screening independent test samples for alveolar echinococcosis (AE).

[0186]

[0187] This invention also provides a deep learning-based synergistic prediction device for echinococcosis drug combinations, such as... Figure 2 As shown, the device includes four modules: a drug feature extraction module, a disease feature extraction module, a single-drug MLP prediction model module, and a Bliss collaborative screening model module.

[0188] The Drug Extract module is used to obtain drug information feature vectors.

[0189] The disease feature extraction module (Diease Extract) is used to obtain feature vectors of echinococcosis information;

[0190] The single-drug MLP prediction model module obtains the drug activity score based on the drug information feature vector and the echinococcosis information feature vector.

[0191] The Bliss collaborative screening model module uses drug activity scores to collaboratively screen and obtain collaborative prediction results for drug combinations.

[0192] The device of this invention uses the resGAT network to obtain drug information feature vectors, uses the node2Vec and attention weight aggregation method to extract echinococcosis information feature vectors, learns the anti-parasitic activity scores of each drug in the drug combination through a multilayer perceptual MLP prediction model, and finally combines the two drugs to calculate the synergistic score of the drug combination using Bliss scores, and determines whether the drug combination has a synergistic effect.

[0193] The present invention also provides an electronic device, Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 7 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute the following methods:

[0194] S1. Establish positive and negative samples for single drugs and drug combinations to obtain SMILES information for single drugs and drug combinations for echinococcosis.

[0195] S2. Obtain target protein information for echinococcosis and construct a target protein PPI network;

[0196] S3. Use the ResGAT network to process information on single drugs and drug combinations SMILES to obtain drug information feature vectors;

[0197] S4. For the target protein in the PPI network, use the node2vec network with biased random walk to obtain the feature vector of the target protein.

[0198] S5. Use a method incorporating attention concepts to weight and aggregate target protein feature vectors as echinococcosis information feature vectors;

[0199] S6. Use drug information feature vectors and echinococcosis information feature vectors to train and construct a single-drug MLP prediction model;

[0200] S7. Use the single-drug MLP prediction model and Bliss to form a drug combination synergistic prediction model to obtain the drug combination synergistic prediction results.

[0201] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0202] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0203] S1. Establish positive and negative samples for single drugs and drug combinations to obtain SMILES information for single drugs and drug combinations for echinococcosis.

[0204] S2. Obtain target protein information for echinococcosis and construct a target protein PPI network;

[0205] S3. Use the ResGAT network to process information on single drugs and drug combinations SMILES to obtain drug information feature vectors;

[0206] S4. For the target protein in the PPI network, use the node2vec network with biased random walk to obtain the feature vector of the target protein.

[0207] S5. Use a method incorporating attention concepts to weight and aggregate target protein feature vectors as echinococcosis information feature vectors;

[0208] S6. Use drug information feature vectors and echinococcosis information feature vectors to train and construct a single-drug MLP prediction model;

[0209] S7. Use the single-drug MLP prediction model and Bliss to form a drug combination synergistic prediction model to obtain the drug combination synergistic prediction results.

[0210] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0211] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented 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. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for synergistic prediction of drug combinations for echinococcosis, characterized in that, include: S1. Establish positive and negative samples for single drugs and drug combinations to obtain SMILES information for single drugs and drug combinations for echinococcosis. S2. Obtain target protein information for echinococcosis and construct a target protein PPI network; S3. Use the ResGAT network to process information on single drugs and drug combinations SMILES to obtain drug information feature vectors; S4. For the target protein in the PPI network, use the node2vec network with biased random walk to obtain the feature vector of the target protein. S5. Use a method incorporating attention concepts to weight and aggregate target protein feature vectors as echinococcosis information feature vectors; S6. Use drug information feature vectors and echinococcosis information feature vectors to train and construct a single-drug MLP prediction model; S7. Use the single-drug MLP prediction model and Bliss to form a drug combination synergistic prediction model to obtain the drug combination synergistic prediction results. S3 uses a ResGAT network to process information on single drugs and drug combinations (SMILES) to obtain drug information feature vectors, including the following steps: S31. Convert the drug SMILES string into a molecular graph, where nodes represent atoms and edges represent chemical bonds. The formula is as follows: (1) Where A represents drug A, This is a set of nodes, where each node represents an atom in the drug molecule diagram. It is the set of edges in a molecular diagram, where each edge represents a chemical bond between atoms in the molecular diagram; S32. Iteratively calculate drug A using a ResGAT network enhanced with residual structure. After each iteration, add initial encoded features, expressed by the formula: (2) Where 1≤i≤n represents the number of atoms in the molecule diagram of drug A, l=0, 1, 2, ..., N represents the number of ResGAT network layers. This represents the characteristic of atom i output from the l-th layer of the drug A molecule diagram. This represents the initial coding features of atom i; S33. The function F(∙) represents the operation of the ResGAT network, which aggregates features of all neighboring nodes in the molecular graph through a message-passing mechanism, performing a weighted sum based on interest to obtain the updated node feature vector. The formula is expressed as: (3) in, Represents atoms The set of adjacent nodes; This represents the attention coefficient, indicating the drug's... middle Node feature pairs The importance of node features; The formula for representing is: (4) in, represents the shared attention mechanism vector, and || represents the feature concatenation operation; S34. Process drug A using global max pooling (GMP) to obtain the drug information feature vector of drug A, expressed by the formula: (5) Where n represents the number of atoms in drug A molecule, Let A be the drug information feature vector of drug A; In step S5, a method incorporating attention principles is used to weight and aggregate target protein feature vectors as echinococcosis information feature vectors. This includes the following steps: S51. Based on the importance of target proteins to echinococcosis, an attention mechanism is used to assign weights to the importance of target proteins, expressed by the following formula: (6) in, This represents the dot product operation. This represents the initial characteristics of unweighted echinococcosis. The feature vector representing the target protein. Target protein feature vector The weights; S52. In a PPI network, all proteins in each layer are linearly combined to form the feature vector of the target protein in that layer, expressed by the formula: (7) Where n is the number of target protein layers, 0 ≤ i ≤ n, The normalized value representing the weight of the target protein associated with echinococcosis in the i-th layer is expressed by the formula: (8) S53. Integrate the target protein features of all layers to obtain the echinococcosis information feature vector of echinococcosis 𝑤, expressed by the formula: (9) in, and These represent the aggregate weight and bias, respectively.

2. The method for synergistic prediction of echinococcosis drug combinations based on deep learning according to claim 1, characterized in that, In step S6, a single-drug MLP prediction model is trained and constructed using drug information feature vectors and echinococcosis information feature vectors. This includes obtaining the anti-echinococcosis activity score of a single drug using the single-drug MLP prediction model, expressed by the formula: (10) in, Drug A fused feature vector, , This represents the MLP prediction model. Let A be the feature vector of drug A. Using the feature vector of echinococcosis information, the MLP prediction model is trained through cross-entropy to predict the anti-echinococcosis activity of a single drug. The higher the value, the more effective drug A is in treating echinococcosis.

3. The method for synergistic prediction of echinococcosis drug combinations based on deep learning according to claim 2, characterized in that, The S7 method utilizes a single-drug MLP prediction model and Bliss to form a drug combination synergistic prediction model, including the following steps: S71, Obtaining Drug Combinations fused feature vector The formula is expressed as: (11) in, Drug A fused feature vector, , Drug B fusion feature vector, ; S72, Obtaining Drug Combinations Anti-echinococcosis activity fraction The formula is expressed as: (12) in, Indicates drug combination The fused feature vector, This represents the MLP prediction model. S73, Defining Drug Combinations Expected activity fraction against echinococcosis The formula is expressed as: (13) in, This represents the activity fraction of drug A alone. This represents the activity fraction of drug B alone. drug combination Expected activity fraction; S74, Calculating Drug Combinations Synergistic effect score The formula is expressed as: （14） Among them, if drug combination Activity fraction against echinococcosis Higher than drug combination Expected activity fraction This indicates a combination It has a synergistic effect.

4. A drug combination synergistic prediction device for assisting in the screening of echinococcosis, applied to the method described in any one of claims 1-3, characterized in that, The device includes: The drug feature extraction module is used to obtain drug information feature vectors; The disease feature extraction module is used to obtain feature vectors of echinococcosis information; The single-drug MLP prediction model module obtains the drug activity score based on the drug information feature vector and the echinococcosis information feature vector. The Bliss collaborative screening model module uses drug activity scores to collaboratively screen and obtain collaborative prediction results for drug combinations.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the deep learning-based hydatid disease drug combination synergistic prediction method as described in any one of claims 1 to 3.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of a deep learning-based synergistic prediction method for echinococcosis drug combinations as described in any one of claims 1 to 3.

Images