Method and system for predicting interaction between drug and target protein

Abstract

The invention belongs to the technical field of intelligent medical treatment, and particularly relates to a method and system for predicting interaction between a drug and target protein. The method comprises the following steps: 1, acquiring structural modal data of a drug, structural data of a target protein and transcriptome expression data of the drug and the target protein in breast cancer, non-small cell lung cancer, melanoma and colon cancer; 2, preprocessing the obtained structural modal data and heterogeneous network modal data of the drug and the target protein; 3, constructing a model, wherein the model comprises a single-mode feature extraction network, a multi-mode feature fusion network based on an attention mechanism and a DTI classification network; 4, inputting the data preprocessed in the step 2 into the model constructed in the step 3; when the DTI is identified, the data relationship between different modes can be concerned at the same time, and the data relationship in the same mode can be fully learned. The method aims to solve the problem of utilization of multi-modal data, the problem of collinearity among multi-modal features and the problem of learning of a data relationship in a single modal in the prior art.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent medical technology, and specifically relates to a method and system for predicting the interaction between a drug and a target protein. Background Art

[0002] The development of a new small molecule drug is a complex and lengthy process. The entire process takes about 10 to 20 years and costs about $2.6 billion. Drug development is divided into four stages: target selection and verification, compound screening and optimization of lead compounds, preclinical research, and clinical trials. In the target selection and verification stage, it is necessary to determine the target associated with a specific disease. Then, the lead compound that can inhibit or activate the target is searched in the molecular database, and the physicochemical properties of the lead compound are optimized. Next, the lead compound becomes a candidate drug and enters the preclinical stage, where pharmacokinetic studies and toxicity tests are conducted in animals. Finally, the candidate drug is applied to patients in clinical trials. Only if the safety and efficacy of the candidate drug pass all clinical trials can the candidate drug be approved for marketing. If DTI (drug-target interaction) is accurately identified in the early stages of drug development, the screening range of the compound can be greatly narrowed, the optimization of the lead compound can be accelerated, the drug development cycle can be shortened, and the drug development efficiency can be improved. Therefore, identifying DTI is a core link in the development of small molecule drugs.

[0003] It is estimated that among the 20,244 proteins in humans, about 6,000 to 8,000 proteins can become targets for drug therapy, but to date, only about 2,700 proteins are drug targets, and a large number of potential drug targets remain to be verified. Among the tens of millions of compounds in the small molecule database, only about 20,000 of them can interact with the corresponding targets. Large-scale small molecule and target protein data show that there are still a large number of potential DTIs that have not yet been discovered. With the continuous accumulation of chemical and biomedical data, drugs and target proteins have data in multiple modalities such as structure, heterogeneous networks, and transcriptome expression. The rapid growth of chemical and biomedical data has promoted the intelligent analysis of DTI. Compared with traditional in vivo and in vitro experiments, computer-assisted intelligent methods can effectively identify large-scale DTIs, reduce the time and cost of in vitro cell experiments, and help people better understand the complex interactions between drugs and targets, accelerate drug development, and promote the progress of human medicine.

[0004] At present, there are two main types of intelligent analysis methods for identifying DTI at home and abroad. One is the unimodal method, which refers to the analysis of only one data type (unimodal data) of drugs and target proteins, using a classifier to learn the data relationship between drugs and target proteins, and determine whether there is an interaction between them. The other is the multimodal method, which refers to the analysis of multiple data types (multimodal data) of drugs and target proteins at the same time, extracting drug and target features from different modalities, and comprehensively analyzing the different modal features of drugs and targets, so as to confirm whether there is an action relationship between drugs and targets. Multimodal data provides more information than unimodal data. Multimodal methods use multiple modalities of drugs and target proteins as input. Compared with unimodal methods, they can analyze the interaction between drugs and targets more comprehensively from different perspectives.

[0005] Existing multimodal methods have some limitations. Existing multimodal methods only integrate the structure and heterogeneous network modal data of drugs and target proteins when identifying DTI, without considering their transcriptome expression modal data, which can provide an unbiased data-driven mechanism to help better understand the mechanism of action of drugs. In addition, existing multimodal methods cannot efficiently integrate multimodal features, and the collinearity between multimodal features is difficult to eliminate, which can easily lead to model overfitting, causing the model to tend to remember the relationship between different modal data and fail to fully learn the relationship between data within the same modality. Summary of the invention

[0006] The present invention provides a method for predicting the interaction between a drug and a target protein, which aims to solve the problems of utilizing multimodal data, the problem of collinearity between multimodal features and the problem of learning data relationships within a single modality in the prior art.

[0007] The present invention provides a system for predicting the interaction between a drug and a target protein, which is used to implement a method for identifying the interaction between a drug and a target protein based on a semi-supervised neural network.

[0008] The present invention is achieved through the following technical solutions:

[0009] A method for predicting the interaction between a drug and a target protein, the method comprising the following steps:

[0010] Step 1: Obtain the structural modal data of the drug, the structural data of the target protein, and the transcriptome expression data of the drug and target protein in breast cancer, non-small cell lung cancer, melanoma, and colon cancer;

[0011] Step 2: Preprocess the acquired structural modal data and heterogeneous network modal data of drugs and target proteins;

[0012] Step 3: Build a model, which includes a unimodal feature extraction network, a multimodal feature fusion network based on an attention mechanism, and a DTI classification network;

[0013] Step 4: Input the data preprocessed in step 2 into the model constructed in step 3; this allows the model to simultaneously focus on the data relationship between different modalities and fully learn the data relationship within the same modality when identifying DTI.

[0014] Furthermore, the step 2 preprocesses the structural modal data of the drug and the target protein by:

[0015] The chemical structure of small molecule drugs is converted into graph structure data, where nodes represent atoms in molecules, edges represent chemical bonds between atoms, and each amino acid in the target protein is encoded as a feature vector with a dimension of 100. Therefore, the target protein can be represented by a two-dimensional feature matrix, where the rows of the matrix represent amino acids and the columns represent feature values.

[0016] The step 1 of preprocessing the heterogeneous network modal data is specifically as follows:

[0017] Heterogeneous networks are constructed using multiple interactions of drug-protein, drug-drug, and drug-disease, and are represented as graph structures, where nodes represent a biological entity in a drug, protein, or disease, and edges represent the interactions between them.

[0018] Furthermore, the step 3 is specifically as follows: the structural modal data and heterogeneous network modal data of the drug and target protein preprocessed in step 2 and the transcriptome expression modal data without preprocessing are used as input, and the single modal feature extraction is performed through the single modal feature extraction network through the structural, heterogeneous network, and transcriptome expression modal feature extraction networks respectively;

[0019] The multimodal feature fusion network based on the attention mechanism controls the information expressed by each modal feature, reduces the collinearity between multimodal features, and retains each single modal feature when fusing multimodal features, thereby effectively fusing multimodal features;

[0020] The fused features are input into the classification network of DTI to obtain the probability of interaction between the drug and the target protein.

[0021] Furthermore, the unimodal feature extraction network includes a transcriptome expression modality learning module, a structural modality learning module, and a heterogeneous network modality learning module.

[0022] The transcriptome expression mode learning module specifically uses the gene transcription expression profile perturbed by small molecule drugs and gene expression profiles perturbed by gene expression changes Two PPI networks were constructed, and two SGCNs were used to extract the transcriptome expression features of drugs and target proteins from the two PPI network graph structures respectively; the PPI network is represented by a graph structure, in which each node represents a protein, and the attribute of the node is the expression value of the corresponding gene in the expression spectrum that guides the synthesis of the protein. The entire PPI network contains 978 nodes, corresponding to 978 landmark genes in the expression spectrum; the edges connecting the nodes represent the interactions between proteins.

[0023] Furthermore, the structural modality learning module uses the graph neural network AttentiveFP with an attention mechanism to learn from the drug molecule graph structure The structural features of the drug are extracted by using three parallel convolutional neural networks to construct Extract the structural features of the target protein; Attentive FP finally outputs the drug structural modal features Then, three CNNs with different convolution kernel sizes are used to learn the target protein features, and finally the three features are concatenated as the final target protein structural features. The structural features of the drug and the target protein are spliced ​​to obtain the output of the structural modality learning module

[0024] Furthermore, the heterogeneous network modality learning module first uses a preset meta-path to guide the random walk process in the heterogeneous network until the entire heterogeneous network graph structure is traversed; and then uses the Skip-Gram model to learn the heterogeneous network modality characteristics of drugs and target proteins from the graph structure.

[0025] Furthermore, the multimodal feature fusion network based on the attention mechanism in step 2 includes a feature fusion module based on the attention mechanism, which fuses multiple modal features output by the transcriptome expression modality learning module, the structural modality learning module, and the heterogeneous network modality learning module into a feature vector;

[0026] The Kronecker product is used to map the three unimodal features into a common representation space to learn the data relationship between different modalities. When fusing multimodal features, an additional constant dimension with a value of 1 is appended to each unimodal feature representation in order to retain the unique characteristics of each unimodality while learning the data relationship between modalities.

[0027] Furthermore, the DTI classification network of step 3 includes a classification prediction module, which uses a fully connected layer for classification prediction. The fully connected layer contains 1 neuron for fusion feature z fusion To classify, firstly fusion Perform a linear transformation, then calculate the category probability, and finally make a classification prediction based on the probability.

[0028] A system for predicting the interaction between a drug and a target protein, the system using the above-mentioned method for identifying the interaction between a drug and a target protein based on a semi-supervised neural network, the system comprising:

[0029] Acquisition module: Acquisition of drug structural modal data, target protein structural data, and transcriptome expression data of drugs and target proteins in breast cancer, non-small cell lung cancer, melanoma, and colon cancer;

[0030] Preprocessing module: preprocessing the structural modal data and heterogeneous network modal data of drugs and target proteins;

[0031] Model building module: build a model, which includes a unimodal feature extraction network, a multimodal feature fusion network based on attention mechanism, and a DTI classification network;

[0032] Calculation module: Input the preprocessed data into the model constructed in step 3 to obtain the probability of interaction between the drug and the target protein; when identifying DTI, it can not only pay attention to the data relationship between different modalities, but also fully learn the data relationship within the same modality.

[0033] A computer-readable storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0034] The beneficial effects of the present invention are:

[0035] The attention mechanism-based feature fusion module of the present invention uses the attention mechanism to control the information expressed by each modal feature, thereby reducing the collinearity between multimodal features.

[0036] The feature fusion module based on the attention mechanism of the present invention retains each single modal feature when fusing multimodal features. When identifying DTI, it can not only pay attention to the data relationship between different modalities, but also fully learn the data relationship within the same modality. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the method of the present invention.

[0038] Figure 2 Flow chart of the present invention for identifying DTI by a multimodal feature fusion network (AttMFFN) based on an attention mechanism; wherein A is a transcriptome expression modality learning module; B is a structural modality learning module; C is a heterogeneous network modality learning module; D is a feature fusion module based on an attention mechanism; and E is a classification prediction module;

[0039] Figure 3It is the prediction performance comparison result of AttMFFN and existing multimodal DTI recognition methods;

[0040] Figure 4 It is the generalization performance comparison result of AttMFFN and existing multimodal DTI recognition methods;

[0041] Figure 5 It is a PCA visualization display of single modal features in AttMFFN using different fusion methods; A is the structural modal feature in AttMFFN using the feature fusion module based on the attention mechanism; B is the heterogeneous network modal feature in AttMFFN using the feature fusion module based on the attention mechanism; C is the transcription spectrum modal feature in AttMFFN using the feature fusion module based on the attention mechanism; D is the structural modal feature in AttMFFN using the splicing operation; E is the heterogeneous network modal feature in AttMFFN using the splicing operation; F is the transcription spectrum modal feature in AttMFFN using the splicing operation;

[0042] Figure 6 To compare the performance of AttMFFN using different fusion methods;

[0043] Figure 7 A is a visualization of molecular docking; A is a visualization of the co-crystal structure of the breast cancer therapeutic target AKT1 combined with tadalafil; B is a visualization of the co-crystal structure of the breast cancer therapeutic target CDK4 combined with flubendazole; C is a visualization of the co-crystal structure of the non-small cell lung cancer therapeutic target EGFR combined with quinacrine. DETAILED DESCRIPTION

[0044] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0045] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0046] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly indicates otherwise.

[0047] The following is attached to this application specification Figure 1-7 , the technical solutions in the embodiments of the present application are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present application, but the present application may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0049] Implementation Method 1

[0050] This embodiment provides a method for predicting the interaction between a drug and a target protein, the method comprising the following steps:

[0051] Step 1: Obtain the structural modal data of the drug, the structural data of the target protein, and the transcriptome expression data of the drug and target protein in breast cancer, non-small cell lung cancer, melanoma, and colon cancer;

[0052] Step 2: Preprocess the acquired structural modal data and heterogeneous network modal data of drugs and target proteins;

[0053] Step 3: Build a model, which includes a unimodal feature extraction network, a multimodal feature fusion network based on an attention mechanism, and a DTI classification network;

[0054] Step 4: Input the data preprocessed in step 2 into the model constructed in step 3; this enables the model to focus on the data relationship between different modalities while fully learning the data relationship within the same modality when identifying DTI, significantly increasing the efficiency of identifying DTI.

[0055] Furthermore, the step 2 preprocesses the structural modal data of the drug and the target protein by:

[0056] The chemical structure of small molecule drugs is converted into graph structure data, where nodes represent atoms in molecules, edges represent chemical bonds between atoms, and each amino acid in the target protein is encoded as a feature vector with a dimension of 100. Therefore, the target protein can be represented by a two-dimensional feature matrix, where the rows of the matrix represent amino acids and the columns represent feature values.

[0057] The step 2 of preprocessing the heterogeneous network modal data is specifically as follows:

[0058] A heterogeneous network is constructed using multiple interactions such as drug-protein, drug-drug, and drug-disease, and the heterogeneous network is represented as a graph structure, in which nodes represent a biological entity in a drug, protein, or disease, and edges represent interactions between them; the original data of transcriptome expression profiles of drugs and target proteins are used as transcriptome expression modal data without preprocessing; the transcriptome expression modal data represents the expression of specific marker genes in cell lines after being disturbed by small molecule drugs or gene expression changes (such as knockdown or overexpression). In the transcriptome expression data of drugs, rows represent genes corresponding to drugs or target proteins, and columns represent the expression values ​​of marker genes.

[0059] Furthermore, the step 3 specifically comprises taking the structural modal data and heterogeneous network modal data of the drug and target protein pre-processed in step 2 and the transcriptome expression modal data without pre-processing as input, and performing single modal feature extraction through the structural, heterogeneous network, and transcriptome expression modal feature extraction networks respectively;

[0060] The multimodal feature fusion network based on the attention mechanism controls the information expressed by each modal feature, reduces the collinearity between multimodal features, and retains each single modal feature when fusing multimodal features, thereby effectively fusing multimodal features;

[0061] The fused features are input into the classification network of DTI to obtain the probability of interaction between the drug and the target protein.

[0062] Furthermore, the unimodal feature extraction network includes a transcriptome expression modality learning module, a structural modality learning module, and a heterogeneous network modality learning module.

[0063] The transcriptome expression modality learning module is specifically as follows: Figure 2 (A) shows the gene transcription profile perturbed by small molecule drugs. and gene expression changes (knockdown or overexpression) perturbed gene transcriptional expression profiles Two PPI (protein-protein interaction) networks were constructed, and two SGCNs (spectral-domain based graph convolutional networks) were used to extract the transcriptome expression features of drugs and target proteins from the two PPI network graph structures respectively; the PPI network is represented by a graph structure, in which each node represents a protein, and the attribute of the node is the expression value of the corresponding gene in the expression spectrum that guides the synthesis of the protein. The entire PPI network contains 978 nodes, corresponding to 978 landmark genes in the expression spectrum; the edges connecting the nodes represent the interactions between proteins.

[0064] The process of extracting features from the PPI network using SGCN is expressed as:

[0065] H l =f(Ug θ U T H l-1 ) (1)

[0066] Where f(·) is the Relu activation function, H l and H l-1 are the feature vectors of the lth layer and the l-1th layer respectively, and They represent the orthogonal matrix and its transposed matrix obtained by spectral decomposition of the Laplace matrix L. The Laplace matrix L is the degree matrix of the PPI network graph structure. Reduced adjacency matrix Calculated, g θ are learnable parameters. Finally, the drug and target protein transcriptome expression modality features output by the two SGCNs are and After splicing, we get the output of the transcriptome expression mode learning module

[0067] Furthermore, the structural modal learning module, such as Figure 2 (B) shows that the Attentive FP graph neural network with an attention mechanism is used to extract the drug molecule graph structure from the drug molecule graph structure. The structural features of the drug are extracted by using three parallel convolutional neural networks (CNNs) to construct The structural features of the target protein are extracted from the target protein; the process of Attentive FP learning for each atomic feature is as follows:

[0068]

[0069]

[0070] in, represents the features that node v aggregates from its neighbor node u during the i-1th layer iteration, N(v) represents the set of neighbor nodes of node v, elu(·) is the elu activation function, W is the trainable weight matrix, is the feature of node v in the i-1th layer iteration, is the attention weight of neighbor node u relative to node v in the process of aggregating features. The calculation process is shown in formula (3). represents the aligned feature vector of node v and node u. The calculation process is shown in formula (4). <·, ·> represents the concatenation operation, f(·) represents the Leaky Relu activation function, and GRU (gated recurrent unit) is used to update the feature of node v. and As the input of GRU, the feature of node v updated in the i-th layer iteration is obtained, as shown in formula (5). Attentive FP finally outputs the drug structure modal feature Then, three CNNs with different convolution kernel sizes are used to learn the target protein features, and finally the three features are concatenated as the final target protein structural features. The structural features of the drug and the target protein are spliced ​​to obtain the output of the structural modality learning module

[0071] Furthermore, the heterogeneous network modality learning module, such as Figure 2 As shown in (C), the random walk process in the heterogeneous network is first guided by the preset meta-path until the entire heterogeneous network graph structure is traversed; then the Skip-Gram model is used to learn the heterogeneous network modal features of drugs and target proteins from the graph structure. The graph structure is represented by G h =(V, E, T), where node v i ∈V represents biological entities such as drugs, target proteins, and diseases, and the edge e∈E represents the interaction between them. V and T E They represent the node type set (e.g., drug, target protein, disease) and edge type set (e.g., drug-protein interaction, drug-drug interaction, drug-disease interaction). Metapaths describe the composite relationships among different types of nodes in the graph structure. A given metapath S can be expressed as where V∈T V , E∈T E Given a heterogeneous network graph structure G h and a meta-path S, traveling to node v i+1 The probability is:

[0072]

[0073] in, Represents nodes i, v of type t i+1 for The neighbor nodes of Representation Node The neighbor node set of all nodes in the set is of type V t+1 , Represents node v i+1 and The connected edges, S(v i+1 ) represents the node v in the S-element path i+1 The heterogeneous network modal features of drugs and target proteins extracted by the Skip-Gram model are expressed as and After splicing, we get the output of the heterogeneous network modality learning module.

[0074] Furthermore, the multimodal feature fusion network based on the attention mechanism in step 3 includes a feature fusion module based on the attention mechanism, such as Figure 2 As shown in (D), multiple modal features output by the transcriptome expression modality learning module, the structural modality learning module, and the heterogeneous network modality learning module are fused into a feature vector; the multimodal feature fusion module uses a gated attention mechanism to calculate the attention weight of each modality feature to reduce the collinearity generated when fusing multimodal features. The calculation process of the gated attention mechanism is as follows:

[0075] g m =softmax(W sht →m· s ,u h ,u t >) (7)

[0076] z m =g m *u m (8)

[0077] Among them, s, h, and t represent structural mode, heterogeneous network mode, and transcriptome expression mode, respectively, and g m represents the attention weight of the modality m∈{s,h,t} relative to the structural modality, heterogeneous network modality, and transcriptome expression modality, W sht→m Represents a learnable weight matrix, where the attention weight of each modality is multiplied by its feature to obtain the feature z for multimodal fusion m ;

[0078] ​The Kronecker product is used to map three unimodal features (transcriptome expression modality features, drug and target protein structure features, and heterogeneous network modality features) into a common representation space to learn the data relationship between different modalities; in order to avoid excessive focus on the data relationship between modalities and ignore the data relationship within a single modality, when fusing multimodal features, an additional constant dimension with a value of 1 is attached to each unimodal feature representation, so as to retain the unique features of each unimodality while learning the data relationship between modalities:

[0079]

[0080] in, They represent the structural modal features, heterogeneous network modal features, and transcriptome expression modal features after attention mechanism operation, respectively. represents the Kronecker product, is the final fusion feature.

[0081] Furthermore, the DTI classification network of step 3 includes a classification prediction module, such as Figure 2 As shown in (E), the classification prediction module uses a fully connected layer for classification prediction. The fully connected layer contains 1 neuron to fusion feature z fusion To classify, firstly fusion Perform a linear transformation and then calculate the category probability. The calculation process is as follows:

[0082] z L =W L ·z fusion (10)

[0083] y=softmax(W·z L ) (11)

[0084] in, and represents the weight matrix, Represents the predicted probability, and finally makes classification prediction based on the probability.

[0085] Performance evaluation: AUROC (Area under the receiver operating characteristic curve) and AUPRC (Area under the precision-recall curve) are used as classification evaluation indicators. Then, the method of the present invention is compared with the prior art.

[0086] Provide relevant data

[0087] The above-mentioned multimodal feature fusion network based on the attention mechanism is used to identify DTI. Experimental results show that its classification indicators are better than the existing multimodal methods. Moreover, the multimodal feature module based on the attention mechanism proposed in the present invention can effectively fuse the multimodal features of drugs and target proteins, so that the technology of the present invention can not only pay attention to the data relationship between different modalities when identifying DTI, but also fully learn the data relationship within the same modality, which significantly increases the efficiency of identifying DTI. In addition, the present invention has important practical application value in identifying cancer drug targets and screening anti-cancer drugs.

[0088] The model was trained using NVIDIA Quadro GV100 32GB GPU, and the deep learning framework used was Pytorch (https: / / pytorch.org / ). The collected DTI data were divided into four datasets according to the type of transcriptome expression data: breast cancer (MCF7) dataset, non-small cell lung cancer (A549) dataset, melanoma (A375) dataset, and colon cancer (HT29) dataset. Table 1 details the distribution of the data. If there is a clear interaction between the drug and the target protein, the drug-target interaction (DTI) is defined as a positive sample with a label of 1. If there is no clear interaction, it is defined as a negative sample with a label of 0. In order to obtain a balanced dataset, samples equal to the number of positive samples were randomly selected from the negative samples and formed a complete dataset with the positive samples for model training and evaluation. The dataset was randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1. During the training process, the model selection was based on the effect of the validation set, that is, the model with the best effect on the validation set was saved, and the saved model was finally tested on the test set. In order to improve the generalization ability of the present invention in identifying DTI, additional independent test sets were constructed for the four data sets to simulate the actual situation of predicting DTI. The ratio of positive samples to negative samples in the independent test set was 1:2, and the drugs and target proteins in the test set did not appear in the training set and the validation set. The number of training iterations was 50, and the optimizer used the adaptive moment (Adam) random optimization method, the batch size was 128, and the learning rate was set to 0.001.

[0089] Table 1. Data distribution of different datasets

[0090]

[0091] Identify the results of DTI

[0092] The model AttMFFN of the present invention is compared with the existing multimodal DTI recognition methods MultiDTI, DeepCompoundNet (DCN), KGE_NFM, and DTI-multi-modal (DTI-MM) on four datasets (MCF7, A549, A375, and HT29). The comparison results are as follows: Figure 2 As shown. The algorithm of the present invention achieved the best AUROC and AUPRC values ​​in all data sets, significantly outperforming all comparison methods that only consider the two modalities of structure and heterogeneity network. The results show that the transcriptome expression modality can significantly improve the prediction performance of the model. In order to evaluate the generalization of the model AttMFFN of the present invention, AttMFFN was compared with the above four multimodal methods on the independent test sets of MCF7, A549, A375, and HT29. The comparison results are shown in Figure 3 As shown, the results show that the algorithm of the present invention is better than the existing multimodal method in terms of generalization performance.

[0093] Results of learning relationships between data within the same modality

[0094] The principal components analysis (PCA) method is used to visualize the unimodal features in AttMFFN using different feature fusion methods. The visualization results are shown in Figure 2. Figure 4 As shown in the figure. The dark and light points represent the unimodal features marked as negative samples and positive samples, respectively. The horizontal axis and the total axis represent the proportion of the original variable variance that can be explained by the first principal component and the second principal component, respectively. By comparing Figure A and Figure D, it can be seen that the feature fusion module based on the attention mechanism enables AttMFFN to learn the data relationship within the structural modality more fully than the splicing operation. The comparison results of Figures B and E also show that the feature fusion module based on the attention mechanism significantly improves the learning ability of AttMFFN on the data relationship within the heterogeneous network modality. Although AttMFFN (Figure F) using the splicing operation is superior to the feature fusion module based on the attention mechanism (Figure C) in terms of the proportion of the original variable variance that can be explained in the transcriptome expression modality feature, this is mainly due to the complexity of the transcriptome expression modality (containing 4421740 parameters), which causes AttMFFN to overfit the transcriptome expression modality feature when using the splicing operation. These results show that compared with the splicing operation, the feature fusion module based on the attention mechanism enables AttMFFN to learn the data relationship within the three modalities more effectively.

[0095] In order to verify the ability of the feature fusion module based on the attention mechanism in improving the prediction performance, the prediction performance of this module is compared with that of AttMFFN using the splicing operation. The results are as follows Figure 5The comparison results show that AttMFFN using the feature fusion module based on the attention mechanism performs better than AttMFFN using the concatenation operation on all datasets, which confirms the effectiveness of the feature fusion module based on the attention mechanism in improving the prediction performance.

[0096] Screening results for anticancer drugs

[0097] The algorithm AttMFFN of the present invention was used to screen drugs with the potential to treat breast cancer and non-small cell lung cancer in the DrugBank database. First, unlisted small molecule drugs with structural, heterogeneous network and transcriptome expression modality characteristics were collected from the DrugBank database. Then, therapeutic targets for breast cancer and non-small cell lung cancer were collected. Finally, the pre-trained AttMFFN was used to screen drugs for the treatment of breast cancer and non-small cell lung cancer. Autodock Vina was used to perform molecular docking experiments on drug-target protein interactions with higher prediction probabilities to confirm the reliability of the screening results. The molecular docking visualization results are shown in Figure 2. Figure 6 As shown in Figure 1. The screening results of AttMFFN showed that tamibarotene and flubendazole had high probability values ​​of interaction with breast cancer target serine-threonine kinase Akt1 (AKT1) and cyclin-dependent kinase 4 (CDK4), which were 0.921 and 0.93, respectively. Tamibarotene has a low binding free energy (-10.5 kcal / mol) with AKT1. The lower the binding free energy, the tighter the binding. As shown in Figure A, tamibarotene forms hydrogen bonds with tryptophan No. 80, asparagine No. 204, and aspartic acid No. 292 of AKT1, and forms a strong π-π stacking interaction with tryptophan No. 80. In addition, it also forms hydrophobic interactions with other key amino acids of AKT1. Figure B shows that the binding free energy of flubendazole with CDK4 is -9.6 kcal / mol. Flubendazole forms strong hydrogen bond interactions with tyrosine 17, lysine 142, glutamic acid 144, and aspartic acid 158 of CDK4. AttMFFN found that quinacrine has the potential to treat non-small cell lung cancer. It has a strong interaction with the epidermal growth factor receptor (EGFR), a therapeutic target of non-small cell lung cancer, and their binding free energy is -8 kcal / mol. As shown in Figure C, quinacrine forms strong hydrogen bonds with methionine 793 and threonine 854 of EGFR, and the distance between them is less than And it forms hydrophobic interactions with other key amino acids of EGFR. These results show that the algorithm AttMFFN of the present invention has important practical application value in discovering anti-tumor drugs.

[0098] Implementation Method 2

[0099] This embodiment provides a system for predicting the interaction between a drug and a target protein. The system uses the method for identifying the interaction between a drug and a target protein based on a semi-supervised neural network as described in Embodiment 1. The system includes:

[0100] Acquisition module: Acquisition of drug structural modal data, target protein structural data, and transcriptome expression data of drugs and target proteins in breast cancer, non-small cell lung cancer, melanoma, and colon cancer;

[0101] Preprocessing module: preprocessing the structural modal data and heterogeneous network modal data of drugs and target proteins;

[0102] Model building module: build a model, which includes a unimodal feature extraction network, a multimodal feature fusion network based on attention mechanism, and a DTI classification network;

[0103] Calculation module: Input the preprocessed data into the model constructed in step 2 to obtain the probability of interaction between the drug and the target protein; when identifying DTI, it can not only pay attention to the data relationship between different modalities, but also fully learn the data relationship within the same modality, which significantly increases the efficiency of DTI identification.

[0104] Implementation Method 3

[0105] An embodiment of the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory is used to store software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory and the processor are connected via a bus. Specifically, the processor implements any step in the first embodiment above by running the computer program stored in the memory.

[0106] It should be understood that in the embodiments of the present invention, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0107] The memory may include a read-only memory, a flash memory, and a random access memory, and provides instructions and data to the processor. A part or all of the memory may also include a nonvolatile random access memory.

[0108] It should be understood that if the above-mentioned integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned implementation method, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method implementations when executed by the processor. Among them, the above-mentioned computer program includes computer program code, and the above-mentioned computer program code can be in source code form, object code form, executable file or some intermediate form. The above-mentioned computer-readable medium may include: any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the above-mentioned computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0109] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest range consistent with the principles and novel features disclosed herein.

[0110] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the implementation method can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method implementation method, which will not be repeated here.

[0111] It should be noted that the methods and detailed examples provided in the above embodiments can be combined with the devices and equipment provided in the embodiments, and references can be made to each other, and no further details will be given.

[0112] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0113] In the embodiments provided by the present invention, it should be understood that the disclosed device / terminal equipment and method can be implemented in other ways. For example, the device / equipment implementation described above is only illustrative, for example, the division of the above modules or units is only a logical function division, and in actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0114] The above-mentioned 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 above-mentioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above-mentioned embodiments, or replace some of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for predicting the interaction between a drug and a target protein, characterized in that: The method comprises the following steps: Step 1: Obtain the structural modal data of the drug, the structural data of the target protein, and the transcriptome expression data of the drug and target protein in breast cancer, non-small cell lung cancer, melanoma, and colon cancer; Step 2: Preprocess the acquired structural modal data and heterogeneous network modal data of drugs and target proteins; Step 3: Build a model, which includes a unimodal feature extraction network, a multimodal feature fusion network based on an attention mechanism, and a DTI classification network; Step 4: Input the data preprocessed in step 2 into the model constructed in step 3; this allows the model to simultaneously focus on the data relationship between different modalities and fully learn the data relationship within the same modality when identifying DTI.

2. The method according to claim 1, characterized in that The step 1 preprocesses the structural modal data of the drug and the target protein as follows: The chemical structure of small molecule drugs is converted into graph structure data, where nodes represent atoms in molecules, edges represent chemical bonds between atoms, and each amino acid in the target protein is encoded as a feature vector with a dimension of 100. Therefore, the target protein can be represented by a two-dimensional feature matrix, where the rows of the matrix represent amino acids and the columns represent feature values. The step 1 of preprocessing the heterogeneous network modal data is specifically as follows: Heterogeneous networks are constructed using multiple interactions of drug-protein, drug-drug, and drug-disease, and are represented as graph structures, where nodes represent a biological entity in a drug, protein, or disease, and edges represent the interactions between them.

3. The method according to claim 1, characterized in that Specifically, step 2 includes taking the structural modal data and heterogeneous network modal data of the drug and target protein pre-processed in step 1 and the transcriptome expression modal data without pre-processing as input, and performing single-modal feature extraction through the structural, heterogeneous network, and transcriptome expression modal feature extraction networks respectively; The multimodal feature fusion network based on the attention mechanism controls the information expressed by each modal feature, reduces the collinearity between multimodal features, and retains each single modal feature when fusing multimodal features, thereby effectively fusing multimodal features; The fused features are input into the classification network of DTI to obtain the probability of interaction between the drug and the target protein.

4. The method according to claim 3, characterized in that The single-modal feature extraction network includes a transcriptome expression modality learning module, a structural modality learning module, and a heterogeneous network modality learning module. The transcriptome expression mode learning module specifically uses the gene transcription expression profile perturbed by small molecule drugs and gene expression profiles perturbed by gene expression changes Two PPI networks were constructed, and two SGCNs were used to extract the transcriptome expression features of drugs and target proteins from the two PPI network graph structures respectively; The PPI network is represented by a graph structure, in which each node represents a protein, and the attribute of the node is the expression value of the corresponding gene in the expression spectrum that guides the synthesis of the protein. The entire PPI network contains 978 nodes, corresponding to the 978 iconic genes in the expression spectrum; the edges connecting the nodes represent the interactions between proteins.

5. The method according to claim 4, characterized in that The structural modality learning module uses the graph neural network Attentive FP with an attention mechanism to learn from the drug molecule graph structure The structural features of the drug are extracted by using three parallel convolutional neural networks to construct Extract the structural features of the target protein; Attentive FP finally outputs the drug structural modal features Then, three CNNs with different convolution kernel sizes are used to learn the target protein features, and finally the three features are concatenated as the final target protein structural features. The structural features of the drug and the target protein are spliced ​​to obtain the output of the structural modality learning module 6. The method according to claim 5, characterized in that The heterogeneous network modality learning module first uses a preset meta-path to guide a random walk process in the heterogeneous network until the entire heterogeneous network graph structure is traversed; and then uses the Skip-Gram model to learn the heterogeneous network modality characteristics of drugs and target proteins from the graph structure.

7. The method according to claim 3, characterized in that The multimodal feature fusion network based on the attention mechanism in step 2 includes a feature fusion module based on the attention mechanism, which fuses multiple modal features output by the transcriptome expression modality learning module, the structural modality learning module, and the heterogeneous network modality learning module into a feature vector; The Kronecker product is used to map the three unimodal features into a common representation space to learn the data relationship between different modalities. When fusing multimodal features, an additional constant dimension with a value of 1 is appended to each unimodal feature representation in order to retain the unique characteristics of each unimodality while learning the data relationship between modalities.

8. The method according to claim 3, characterized in that The DTI classification network of step 2 includes a classification prediction module, which uses a fully connected layer for classification prediction. The fully connected layer contains 1 neuron to fusion feature z fusion To classify, firstly fusion Perform a linear transformation, then calculate the category probability, and finally make a classification prediction based on the probability.

9. A system for predicting the interaction between a drug and a target protein, characterized in that: The system uses a method for predicting the interaction between a drug and a target protein as described in any one of claims 1 to 8, and the system comprises: Acquisition module: Acquisition of drug structural modal data, target protein structural data, and transcriptome expression data of drugs and target proteins in breast cancer, non-small cell lung cancer, melanoma, and colon cancer; Preprocessing module: preprocessing the structural modal data and heterogeneous network modal data of drugs and target proteins; Model building module: build a model, which includes a unimodal feature extraction network, a multimodal feature fusion network based on attention mechanism, and a DTI classification network; Calculation module: Input the preprocessed data into the model constructed in step 2 to obtain the probability of interaction between the drug and the target protein; when identifying DTI, it can not only pay attention to the data relationship between different modalities, but also fully learn the data relationship within the same modality.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images