Multi-view-based prediction method and model for learning anti-echinococcosis drug combination

Abstract

The invention discloses a multi-view-based prediction method and module for learning an echinococcosis-resistant drug combination, and the method comprises the steps: integrating the pathway information of a disease and the similarity information of the disease, extracting the individual input of the disease, the molecular fingerprint information of a drug used by the drug, and the similarity information of the drug, and employing a multi-view-angle feature extraction expert module to achieve the prediction of the echinococcosis-resistant drug combination. And constructing high-order representation from different angles, and connecting representation of a drug combination synergistic effect and a drug-drug interaction task through a gating mechanism by utilizing a self-adaptive task fusion module. According to the method, an echinococcosis-resistant drug combination prediction model based on multi-view multi-task learning is designed, the model pays attention to safe drug use while paying attention to synergistic drug combination, adverse reactions among drugs are avoided, and therefore intelligent prediction of echinococcosis-resistant drug combination is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer chemistry, and in particular to a method and model for predicting anti-hydatid disease drug combinations based on multi-view learning. Background Art

[0002] Echinococcosis, commonly known as hydatid disease, is a chronic zoonotic parasitic disease caused by the larvae of the tapeworm Echinococcus in the host, which seriously affects human health and the development of animal husbandry and is a global public health problem. For patients who do not meet the surgical indications and are intolerant to surgery, drug treatment is the main treatment method. Compared with single-drug treatment, drug synergy can enhance the treatment effect and make it easier for patients to recover. This is because different drug combinations produce synergistic effects on the target biological pathways, making the treatment of hydatid disease more effective. At the same time, through drug synergy, the dose of each drug can be reduced, thereby reducing the potential side effects and drug tolerance that patients may face. Currently, screening existing drugs through in vitro and in vivo experiments of Echinococcus is still the mainstream of drug research and development. However, only a small number of drugs are studied through clinical trials. On the one hand, they are time-consuming, expensive, and may subject patients to unnecessary treatments. On the other hand, due to the huge space of drug combinations, it is difficult to explore all drug combination spaces only by in vitro and in vivo biological experimental methods. That is, the inefficiency and extensive drug trial and error in biological experiments cannot be ignored, and it is necessary to explore and establish an efficient experimental evaluation method.

[0003] In recent years, with the rapid development of computer technology, various computational methods have been proposed for drug combination prediction, which provides the possibility for effectively exploring the large synergistic drug combination space. Among them, deep learning technology, as a powerful artificial intelligence technology, has been widely applied to the medical field. Such as for cancer tumor discovery, drug development, personalized medicine, etc. However, the related research on echinococcosis is relatively less, and there are certain limitations in the methods. On the one hand, the successful training of current deep learning models depends on large-scale and high-quality data sets, which should contain a large number of drug synergistic combinations. However, the effective synergistic drug combination data for hydatid disease is extremely scarce, which significantly limits the training and prediction accuracy of deep learning methods. On the other hand, existing computational synergy prediction methods cannot detect potential drug-drug adverse interactions that may increase the risk of combination therapy, resulting in the difficulty of clinically treating the predicted drug combinations. Summary of the Invention

[0004] To solve the technical problems existing in the background art, the present invention proposes a multi - perspective learning - based anti - hydatid disease drug combination prediction method and model, which integrates the pathway information of the disease and the disease similarity information to extract the individual input of the disease, and uses the molecular fingerprint information of the drug and the drug similarity information. A multi - perspective feature extraction expert module is adopted to construct high - order representations from different angles, and an adaptive task fusion module is used to connect the representations of the drug combination synergy and the drug - drug interaction tasks through a gating mechanism. A multi - perspective learning - based anti - hydatid disease drug combination prediction method model is designed. While paying attention to the synergistic drug combinations, this model also pays attention to safe drug use and avoids adverse reactions between drugs, thus realizing the intelligent prediction of the combination use of anti - hydatid disease drugs.

[0005] A multi - perspective learning - based anti - hydatid disease drug combination prediction method proposed by the present invention includes the following steps:

[0006] S1. Construct an input data set of drug data and disease data, which includes positive samples and negative samples; the positive samples include drug and drug combination information for treating hydatid disease, and also include drug and drug combination information for parasitic diseases similar to hydatid disease. Specifically, the data set information is sourced from relevant drugs and drug combinations regarding the quality of hydatid disease in CNKI and PubMed databases. Specifically, the drugs and drug combinations related to parasitic diseases similar to hydatid disease include 263 single drugs and 283 drug combinations;

[0007] The negative samples of the input data set are drug combinations that clearly have no synergy and are not suitable for use together;

[0008] S2. Take two drugs and a disease C as inputs and convert the input features into latent vectors h v1 、h v2 、h v3 ;

[0009] h v1 is the basic feature corresponding to each drug data and disease data; h v2 is the feature of each drug treating the disease alone; h v3 is the feature in the case of two drugs jointly treating the disease;

[0010] S3. The gating unit learns the weight Wk of each latent vector through the features of the drug combination and the disease, and calculates the fused feature Ok of task k:

[0011] I num refers to the latent vectors output by the num expert networks in step S2, and M is the total number of latent vectors output by the expert networks: I numis the latent vector output by num expert networks in step S2, is I num the weight of the corresponding latent vector;

[0012] S4: Input the fused features into the prediction module to calculate the prediction probabilities of positive and negative samples respectively, and select the category with the highest probability as the classification label.

[0013] Preferably, in step S2, the expert network includes an initial feature extraction expert network for extracting h v1 a single-drug treatment feature expert network for extracting h v2 and a combination treatment feature expert network for extracting h v3

[0014] Preferably: Define two drugs as drug A and drug B, and define one disease as disease C, then d A and d B respectively represent the molecular fingerprint information of drug A and drug B after being transformed by the rdkit tool, and dis represents the initial feature representation of the disease, then:

[0015] h v1 = MLP(d A ) || MLP(d B ) || MLP(dis);

[0016] where, || represents concatenation; MLP(d A ) represents the output of d A through the multi-layer perceptron model; MLP(d B ) represents the output of d B through the multi-layer perceptron model; MLP(dis) represents the output of dis through the multi-layer perceptron model;

[0017] h v2 = MLP(d A || dis) MLP(d B || dis);

[0018] where: || represents concatenation; MLP(d A || dis) represents the output of d A after being combined with dis through the multi-layer perceptron model; MLP(d B || dis) represents the output of d B after being combined with dis through the multi-layer perceptron model;

[0019] h v3 = MLP(d A || d B || dis)

[0020] ​where: MLP(d A ||d B ||dis) represents the output of the multi - layer perceptron model after combining d B 、d A and dis;

[0021] Specifically, dis = dis pathway ||dis sim , dis pathway is the pathway information of the disease, and dis sim is the similarity information of the disease.

[0022] Preferably, in step S3, W k = softmax(gate k (CAT(d A , d B , dis)));

[0023] gate k (x) = Layer k (Layer k (x));

[0024] In the formula: softmax(gate k (CAT(d A , d B , dis))) represents performing the softmax activation function on gate k (CAT(d A , d B , dis)); gate k (CAT(d A , d B , dis)) represents performing the gate control unit processing on CAT(d A , d B , dis), and CAT(d A , d B , dis) represents concatenating d A 、d B and dis; Layer k (Layer k (x)) represents inputting Layer k (x) into the neural network for processing; Laye k r(x) represents inputting x into the neural network for processing, and x is CAT(d A , d B , dis).

[0025] A multi - perspective learning - based anti - echinococcosis drug combination prediction model, comprising:

[0026] Initial feature extraction expert network module, which is used to extract the basic features h of drugs and diseases v1 , so as to realize the independence of drug and disease features;

[0027] Monotherapy feature expert network module: extract the effects of each drug and disease on the disease respectively, and obtain the output h v2 ;

[0028] Combined therapy feature expert network module, which is used to extract the combined features h of the combined action of two drugs on the disease v3 ; h v3 Reflect the characteristic effect in the case of drug combination therapy for diseases;

[0029] Adaptive task fusion module, including two independent gating units, automatically learns to select relevant information under the corresponding task from multiple latent vectors, and obtains the fusion feature OK of task k;

[0030] Prediction module, calculates the prediction probabilities of positive and negative samples through the fusion feature, and selects the category with the higher probability as the classification label.

[0031] Preferably, the adaptive task fusion module includes opposing gating units, and the gating units learn the weights of each latent vector through the features of drug combinations and diseases.

[0032] Preferably, the initial feature extraction expert network module, the monotherapy feature expert network module, and the combined therapy feature expert network module adopt different neural network structures. It realizes different interaction information extraction modes for the combination of drug x and drug y, and at the same time reduces the information redundancy brought by the same interaction mode. The number of each type of expert (including the initial feature extraction expert network module, the monotherapy feature expert network module, and the combined therapy feature expert network module) can be single or multiple. Each expert converts the input features into d-dimensional latent vectors.

[0033] In the present invention, the proposed multi-perspective learning anti-echinococcosis drug combination prediction method and model extract input features into multiple high-order vectors from different angles, the adaptive task fusion module learns the weights of specific tasks, and sums the weighted vectors of these two tasks; finally, each summation vector is input into the corresponding prediction module to obtain the prediction output of the specified task, avoiding adverse reactions between drugs, thereby realizing more accurate and effective intelligent prediction of the combination of anti-echinococcosis drugs.

[0034] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be understood through the practice of the present invention. Description of the Drawings

[0035] Figure 1 Schematic diagram of the structure of the present invention;

[0036] Figure 2 Schematic diagram of the connection structure between the air knife and the frame of the present invention;

[0037] In the figure: DrugA is drug A; DrugB is drug B; DiseaseC is disease C; Experts1-2 is: single-drug treatment feature expert network module; Experts3-4 is single-drug treatment feature expert network module; Experts5-6 is combination treatment feature expert network module; GateNetwork is adaptive task fusion module; DDI is the interaction between drugs and drugs, and DDS is the synergistic treatment effect of drugs and drugs on diseases. Detailed implementation manners

[0038] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar symbols represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation to the present invention.

[0039] As Figure 1 - Figure 2 shown, a multi-perspective learning-based method for predicting anti-echinococcosis drug combinations includes the following steps:

[0040] S1. Construct an input data set of drug data and disease data, and the input data set includes positive samples and negative samples; the positive samples include drug and drug combination information for treating echinococcosis, and also include drug and drug combination information for parasitic diseases similar to echinococcosis. Specifically, the drugs and drug combinations related to parasitic diseases similar to echinococcosis include 263 single drugs and 283 drug combinations;

[0041] The negative samples of the input data set are drug combinations that include obviously no synergistic effect and are not suitable for use together. The negative samples and positive samples have the following specific sample information:

[0042] Multi-task drug combination effect data set

[0043]

[0044] Parasite drug combination synergy data set

[0045]

[0046] S2. Take two drugs and a disease C as inputs, and convert the input features into latent vectors h v1 、h v2 、h v3 ;

[0047] h v1 is the basic feature corresponding to each drug data and disease data; h v2 is the feature of each drug treating the disease alone; h v3 is the feature in the case of two drugs treating the disease in combination;

[0048] The expert network includes an initial feature extraction expert network for extracting h v1 a single-drug treatment feature expert network for extracting h v2 and a combination treatment feature expert network for extracting h v3 ;

[0049] S3. Learn the weight W of each latent vector through the features of the drug combination and the disease k , and calculate the fusion feature Ok of task k through the weight W k : k is 1 or 2, where task 1 is the synergistic effect of the drug combination and task 2 is the drug-drug interaction. Since there are two tasks, there are two different gating units to calculate the outputs of different tasks;

[0050] I num refers to the latent vector output by the num expert network in step S2. M is the total number of latent vectors output by the expert network. In this embodiment, M is 6; I num is a latent vector output by the expert network in step S2, is I num the weight of the corresponding latent vector;

[0051] As Figure 1 shown, in this embodiment, there are three types of expert networks. Each type of expert adopts a different neural network structure to implement different interaction information extraction modes for the combination of drug A and drug B, reducing the information redundancy caused by the same interaction mode. The number of each type of expert can be single or multiple. Each expert converts the input features into d-dimensional latent vectors. In this article, the following three types of experts are mainly adopted. Respectively: initial feature extraction expert network; single-drug treatment feature expert network; combination treatment feature expert network.

[0052] As Figure 2 shown, the num expert network is one of the six expert networks. The first expert network is the 1 expert network, the second expert network is the 2 expert network... the sixth expert network is the 6 expert network, I 1 is the output of the first expert network, I 2 is the output of the second expert network.......I 6 is the output of the sixth expert network;

[0053] As Figure 2 shown, after intensive connection between drug A and disease C, the first output (output-1) is obtained; after intensive connection between drug B and disease C, the second output (output-2) is obtained; after intensive connection between drug A and drug B, the first output (output-3) is obtained; after intensive connection of disease C, the fourth output (output-4) is obtained; after intensive connection of drug A, drug B and disease C, the fifth output (output-5) is obtained;

[0054] Among them, (output-1) and (output-2) are the characteristic outputs of each drug treating the disease alone, and (output-5) is the basic feature corresponding to the drug data and disease data; (output-3) and (output-4) are the features in the case of two drugs treating the disease jointly;

[0055] S4: Input the fused features into the prediction module to calculate the prediction probabilities of positive and negative samples respectively, and select the class with the highest probability as the classification label;

[0056] The prediction module includes a prediction layer, which consists of two prediction towers. Using the fused representation Ok as the input, the prediction tower of task k predicts y of certain tasks k , since both of our two tasks are classification tasks, after the two-dimensional vector output obtained in step S3, the vector is sent to the softmax layer to calculate the prediction probabilities of positive and negative samples respectively. Select the class with the highest probability as the classification label.

[0057] Preferably: Define two drugs as drug A and drug B, and define one disease as disease C, then d A and d B respectively represent the molecular fingerprint information of drug A and drug B after being transformed by the rdkit tool, and dis represents the initial feature representation of the disease, then:

[0058] h v1 = MLP(d A ) || MLP(d B );

[0059] Among them, || represents concatenation; MLP(d A ) represents the output of d A through the multi-layer perceptron model; MLP(d B ) represents the output of d B through the multi-layer perceptron model; MLP(dis) represents the output of dis through the multi-layer perceptron model;

[0060] h v2 = MLP(dA ||dis)MLP(d B ||dis);

[0061] Where: || represents concatenation; MLP(d A ||dis) represents the output of d A after being combined with dis through a multi-layer perceptron model; MLP(d B ||dis) represents the output of d B after being combined with dis through a multi-layer perceptron model;

[0062] h v3 = MLP(d A ||d B ||dis)

[0063] Where: MLP(d A ||d B ||dis) represents the output of d B 、d A after being combined with dis through a multi-layer perceptron model;

[0064] Specifically, dis = dis pathway ||dis sim , dis pathway is the pathway information of the disease, dis sim is the similarity information of the disease.

[0065] Preferably, in step S3, W k = softmax(gate k (CAT(d A , d B , dis)));

[0066] gate k (x) = Layer k (Layer k (x));

[0067] In the formula: softmax(gate k (CAT(d A , d B , dis))) represents applying the softmax activation function to gate k (CAT(d A , d B , dis)); gate k (CAT(d A , d B , dis)) represents CAT(d A , d B, perform gating unit processing on (dis), CAT(d A , d B , (dis) indicates concatenation processing on d A 、d B 、dis; Layer k (Layer k (x)) indicates inputting Layer k (x) into the neural network for processing; Laye k r(x) indicates inputting x into the neural network for processing, where x is CAT(d A , d B , dis).

[0068] It should be noted that represents the weights of the num expert networks of the k gating units, and its calculation method is also as shown in the above formula.

[0069] A multi-view learning-based anti-echinococcosis drug combination prediction model, including:

[0070] An initial feature extraction expert network module, which is used to extract the basic features h of drugs and diseases v1 to achieve the independence of drug and disease features;

[0071] A single-drug treatment feature expert network module: extracting the effects of each drug and disease on the disease respectively to obtain the output h v2 ;

[0072] A combination treatment feature expert network module, which is used to extract the combined features h of the combined action of two drugs on the disease v3 ; h v3 reflects the characteristic effect in the case of drug combination treatment of diseases;

[0073] An adaptive task fusion module, which automatically learns to select relevant information under the corresponding task from multiple latent vectors to obtain the fusion feature OK of task k;

[0074] A prediction module, which calculates the prediction probabilities of positive and negative samples through the fusion feature and selects the category with the higher probability as the classification label.

[0075] Preferably, the adaptive task fusion module includes opposing gating units, and the gating units learn the weights of each latent vector through the features of drug combinations and diseases.

[0076] Preferably, the initial feature extraction expert network module, the monotherapy feature expert network module, and the combination therapy feature expert network module adopt different neural network structures. Different interaction information extraction modes for the combination of drug x and drug y are realized, while reducing the information redundancy brought by the same interaction mode. The number of each type of expert (including the initial feature extraction expert network module, the monotherapy feature expert network module, and the combination therapy feature expert network module) can be single or multiple. Each expert converts the input features into d-dimensional latent vectors.

[0077] To evaluate the performance of the above methods and models, the predictive ability of the model is evaluated through five-fold cross-validation, in which the training samples are randomly divided into five approximately equal subsets. In each iteration, one subset is reserved as the test set, and the remaining subsets are used as the training set. The average prediction accuracy of the five-fold cross-validation is used as the final metric for performance evaluation. The performance evaluation metrics include the area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUPR), recall rate (Rec), precision (Prec), F1 score, and accuracy (ACC) to comprehensively reflect the performance of the model in various aspects. The experimental results are shown in the table. The formulas for each evaluation metric are as follows:

[0078]

[0079]

[0080]

[0081]

[0082] AUC AUPR Rec Prec F1 ACC XGBoost 0.9116 0.8891 0.8187 0.8239 0.8519 0.8668 DeepSynergy 0.9018 0.9128 0.9108 0.8498 0.8618 0.8401 Transynergy 0.9213 0.9219 0.9003 0.8618 0.8612 0.8737 GAECDS 0.9019 0.9323 0.9023 0.8712 0.8532 0.8312 Attensyn 0.9402 0.9419 0.8219 0.8693 0.8617 0.8519 MMLSyn 0.9432 0.9586 0.9013 0.9081 0.9018 0.9173

[0083] The results show that the model of the present invention performs excellently in predicting the effects of drug combinations against Echinococcus granulosus. Specifically, this method obtained an AUC of 0.9561 and an AUPR of 0.9622 in the prediction, exceeding the performance metrics of most comparative methods in the Echinococcus granulosus drug combination dataset, fully demonstrating the powerful predictive ability of the model.

[0084] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A method for predicting anti-echinococcosis drug combinations based on multi-perspective learning, characterized in that: The steps include: S1. Construct an input data set of drug data and disease data, the input data set includes positive samples and negative samples; the positive samples include information on drugs and drug combinations for treating echinococcosis, and also include information on drugs and drug combinations for parasitic diseases similar to echinococcosis; the negative samples of the input data set include drug combinations that are obviously not synergistic and are not suitable for use together; S2, taking two drugs and a disease C as input, converting the input features into a latent vector h through an expert network v1 、h v2 、h v3 ; h v1 is the basic feature corresponding to each drug data and disease data; h v2 Characterize the disease treated by each drug individually; v3 Characterizes when two drugs are combined to treat a disease; S3. Learn the weight W of each latent vector by the characteristics of drug combination and disease k , through the weight W k Calculate the fusion feature Ok of task k: M is the total number of potential vectors output by the expert network: I num is the potential vector output by the num expert network in step S2, For I num The weight of the corresponding latent vector; S4: Input the fused features into the prediction module to calculate the prediction probabilities of positive and negative samples respectively, and select the category with the highest probability as the classification label.

2. The method for predicting anti-echinococcosis drug combinations based on multi-perspective learning according to claim 1, characterized in that: The expert network in step S2 includes a v1 The initial feature extraction expert network is used to extract h v2 Expert network for extracting h v3 A network of experts on combination therapy characteristics.

3. The method for predicting anti-echinococcosis drug combinations based on multi-perspective learning according to claim 1, characterized in that: Define two drugs as drug A and drug B, and define a disease as disease C, then d A and d B They represent the molecular fingerprint information of drug A and drug B after conversion by the rdkit tool, and dis represents the initial characteristic representation of the disease: h v1 =MLP(d A )||MLP(d B )||MLP(dis); Among them, || represents connection; MLP(d A ) means d A Through the output of the multi-layer perceptron model; MLP (d B ) means d B The output of the multilayer perceptron model; MLP(dis) represents the output of dis through the multilayer perceptron model; h v2 =MLP(d A ||dis)MLP(d B ||dis); Among them: || represents connection; MLP(d A ||dis) means d A After being combined with dis, the output of the multi-layer perceptron model; MLP (d B ||dis) means d B The output of the multilayer perceptron model after being combined with dis; h v3 =MLP(d A ||d B ||dis) Among them: MLP (d A ||d B ||dis) means d B ,d A The output of the multilayer perceptron model after being combined with dis.

4. The method for predicting anti-echinococcosis drug combinations based on multi-perspective learning according to claim 3 is characterized in that: dis=dis pathway ||dis sim ,dis pathway is the pathway information of the disease, sim It is the similarity information of the disease.

5. The method for predicting anti-echinococcosis drug combinations based on multi-perspective learning according to claim 1, characterized in that: In step S3: W k =softmax(gate k (CAT(d A ,d B ,dis))); gate k (x)=Layer k (Layer k (x)); Where: softmax(gate k (CAT(d A ,d B ,dis))) indicates the gate k (CAT(d A ,d B ,dis)) performs softmax activation function processing; gate k (CAT(d A ,d B ,dis)) means CAT(d A ,d B ,dis) for door control unit processing, CAT(d A ,d B ,dis) indicates that d A d B , dis for splicing; Layer k (Layer k (x)) indicates the Layer k (x) is input to the neural network for processing; Layer k r(x) represents the processing of x input into the neural network, x is CAT(d A ,d B ,dis).

6. A multi-perspective learning anti-echinococcosis drug combination prediction model, characterized in that: include: Initial feature extraction expert network module, the initial feature extraction expert network module is used to extract the basic features of drugs and diseases v1 ; Single drug treatment feature expert network module: extract the effect of each drug and disease on the disease and get the output h v2 ; Combined treatment feature expert network module, used to extract the joint features of two drugs acting on the disease v3 ; The adaptive task fusion module automatically learns to select relevant information under the corresponding task from multiple potential vectors and obtains the fusion feature OK of task k; The prediction module calculates the prediction probability of positive and negative samples by fusing features and selects the category with high probability as the classification label.

7. The multi-perspective learning anti-echinococcosis drug combination prediction model according to claim 6, characterized in that: The adaptive task fusion module includes an adversarial gating unit that learns the weight of each latent vector by the characteristics of drug combinations and diseases.

8. The multi-perspective learning anti-echinococcosis drug combination prediction model according to claim 6, characterized in that: The initial feature extraction expert network module, the monotherapy feature expert network module and the combination therapy feature expert network module adopt different neural network structures.

9. The multi-perspective learning anti-echinococcosis drug combination prediction model according to claim 6, characterized in that: The number of the initial feature extraction expert network module, the monotherapy feature expert network module and the combination therapy feature expert network module can be single or multiple.

Images