An artificial intelligence-based method for predicting drug-target interactions

Abstract

The present invention belongs to the technical field of drug-target interaction prediction, and discloses a drug-target interaction prediction method based on artificial intelligence. The present method proposes to use a multi-scale multimodal neural network model to predict drug-target interaction, taking into account the local and overall characteristics of chemical molecular structures. The custom mode pooling layer and the custom discretization layer inside the twin feature pyramid effectively protect the intrinsic discreteness of the chemical molecular structure in the process of shrinking and extracting the molecular three-dimensional structure model. The present invention solves the problem that the existing drug-target interaction prediction method ignores the intrinsic discreteness of the chemical molecular structure and cannot capture the connection between the local and the overall through the above technical means, and realizes multimodal, multi-scale and high-precision drug-target interaction prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug-target interaction prediction, and specifically refers to a drug-target interaction prediction method based on artificial intelligence. Background Art

[0002] Drug-target interaction prediction methods are computational methods used to predict possible interactions between drugs and biological molecules. These methods can help researchers screen out the most promising drug candidates before experimental verification. For this reason, the industry has invented a series of drug-target interaction prediction methods, among which molecular docking technology is the most typical. Molecular docking technology attempts to restore the reaction between drugs and biological molecules in the form of microscopic physical simulation. However, its computing power and storage requirements increase exponentially with the simulation precision, and are limited to the structures of drugs and biological molecules themselves. It is unable to handle the disturbances and latent variables in the human body. For this reason, the industry has invented a series of artificial intelligence models to solve such problems, fitting the inherent laws of drug-target interactions in the form of multimodal fusion. However, this type of method simply and crudely applies various methods used to process continuous images to molecular three-dimensional structure models, ignoring the inherent discreteness of chemical molecular structures, resulting in information loss in the calculation process. Compared with linear neural network structures, it is also unable to capture the connection between the local and the whole, which ultimately leads to unsatisfactory performance of this type of method in drug-target interaction prediction tasks. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a drug-target interaction prediction method based on artificial intelligence. In view of the problem that the existing drug-target interaction prediction method ignores the inherent discreteness of the chemical molecular structure and cannot capture the connection between the local and the whole, the present invention adopts a multi-scale multimodal neural network model to predict drug-target interactions. The multi-scale multimodal neural network model adopts a twin feature pyramid to take into account the local and overall characteristics of the chemical molecular structure from the overall to the local. The custom mode pooling layer and the custom discretization layer inside the twin feature pyramid effectively protect the inherent discreteness of the chemical molecular structure in the process of shrinking and extracting the molecular three-dimensional structure model, and organically combines computational chemistry theory with artificial intelligence technology to achieve multimodal, multi-scale and high-precision drug-target interaction prediction.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides a drug-target interaction prediction method based on artificial intelligence, and the method comprises the following steps:

[0005] Step S1: collecting drug-target pair information, target protein data and drug data;

[0006] Step S2: construct positive samples of the entire data set;

[0007] Step S3: construct negative samples of the entire data set;

[0008] Step S4: Aggregate all positive samples of the data set and all negative samples of the data set into a data set;

[0009] Step S5: constructing a multi-scale multi-modal neural network model, wherein the multi-scale multi-modal neural network model includes a text feature extraction branch, a twin feature pyramid, a graphic feature fusion branch, and a re-representation branch;

[0010] Step S6: training a multi-scale multi-modal neural network model using the data set;

[0011] Step S7: Use a multi-scale multimodal neural network model for prediction.

[0012] Furthermore, the step S2 comprises the following steps:

[0013] Step S21: extracting a pair of drug and target protein from the drug-target pair information;

[0014] Step S22: constructing a drug text sequence, a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix corresponding to the drug and a target protein text sequence, a target protein multi-channel three-dimensional matrix and a binding site multi-channel three-dimensional matrix corresponding to the target protein, and summarizing them into a data set positive sample;

[0015] Step S23: adding a matching marker variable to the positive sample of the data set, where the value of the matching marker variable of the positive sample of the data set is 1;

[0016] Step S24: Repeat steps S21 to S23 for each pair of drug and target protein in the drug-target pair information until each pair of drug and target protein in the drug-target pair information has a corresponding data set positive sample, thereby generating positive samples for all data sets.

[0017] Furthermore, the step S22 specifically includes the following steps:

[0018] Step S221: extracting a target protein three-dimensional structure model from the target protein data, extracting a drug three-dimensional structure model from the drug data, extracting a binding site three-dimensional structure model from the target protein three-dimensional structure model, and extracting a pharmacophore three-dimensional structure model from the drug three-dimensional structure model;

[0019] Step S222: Matrixing the three-dimensional structural model of the target protein and the three-dimensional structural model of the binding site to construct a multi-channel three-dimensional matrix of the target protein and a multi-channel three-dimensional matrix of the binding site;

[0020] Step S223: extracting and encoding the data corresponding to the target protein except the target protein three-dimensional structure model from the target protein data to generate a target protein text sequence;

[0021] Step S224: Matrixing the drug three-dimensional structure model and the pharmacophore three-dimensional structure model to construct a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix;

[0022] Step S225: extracting data corresponding to the drug except the drug three-dimensional structure model from the drug data and encoding them to generate a drug text sequence;

[0023] Step S226: Aggregate the drug text sequence, drug multi-channel three-dimensional matrix, pharmacophore multi-channel three-dimensional matrix, target protein text sequence, target protein multi-channel three-dimensional matrix and binding site multi-channel three-dimensional matrix into a data set positive sample.

[0024] Furthermore, the process of constructing the target protein multi-channel three-dimensional matrix and the binding site multi-channel three-dimensional matrix in step S222 specifically includes the following steps:

[0025] Step S2221: constructing a target protein three-dimensional matrix with the target protein three-dimensional structure model in units of 1 cubic angstrom, wherein a zero element of the target protein three-dimensional matrix indicates that there is no atom at this position in the target protein three-dimensional structure model, and a value of a non-zero element is determined by the atomic number of the corresponding atom in the target protein three-dimensional structure model;

[0026] Step S2222: numbering different types of binding points in the three-dimensional structure model of the binding points to generate binding point numbers;

[0027] Step S2223: partitioning the three-dimensional matrix of the target protein based on the three-dimensional structural model of the binding sites, assigning the non-zero elements of the regions corresponding to the binding sites in the three-dimensional structural model of the target protein to the corresponding binding site numbers, and generating a binding site region matrix;

[0028] Step S2224: superimposing the target protein three-dimensional matrix and the binding site region matrix into a target protein multi-channel three-dimensional matrix;

[0029] Step S2225: splitting the target protein three-dimensional matrix based on the binding site region matrix to generate a binding site three-dimensional matrix group, wherein a single matrix in the binding site three-dimensional matrix group corresponds to the region where a single binding site of the target protein three-dimensional matrix is ​​located;

[0030] Step S2226: superimposing the combination point three-dimensional matrix groups in the order of the combination point numbers to generate a combination point multi-channel three-dimensional matrix.

[0031] Furthermore, the process of constructing the drug multi-channel three-dimensional matrix and the pharmacophore multi-channel three-dimensional matrix in step S224 specifically includes the following steps:

[0032] Step S2241: constructing a three-dimensional drug matrix with the drug three-dimensional structure model as a unit of 1 cubic angstrom, wherein a zero element of the drug three-dimensional matrix indicates that there is no atom at this position in the drug three-dimensional structure model, and a value of a non-zero element is determined by the atomic number of the corresponding atom in the drug three-dimensional structure model;

[0033] Step S2242: numbering different types of pharmacophores in the pharmacophore three-dimensional structure model to generate pharmacophore numbers;

[0034] Step S2243: partitioning the drug three-dimensional matrix based on the pharmacophore three-dimensional structure model, assigning the non-zero elements in the region corresponding to the pharmacophore in the drug three-dimensional structure model to the corresponding pharmacophore number, and generating a pharmacophore region matrix;

[0035] Step S2244: superimposing the drug three-dimensional matrix and the pharmacophore region matrix into a drug multi-channel three-dimensional matrix;

[0036] Step S2245: splitting the drug three-dimensional matrix based on the pharmacophore region matrix to generate a pharmacophore three-dimensional matrix group, wherein a single matrix in the pharmacophore three-dimensional matrix group corresponds to a region where a single pharmacophore of the drug three-dimensional matrix is ​​located;

[0037] Step S2246: superimposing the pharmacophore three-dimensional matrix groups in the order of pharmacophore numbers to generate a pharmacophore multi-channel three-dimensional matrix.

[0038] Furthermore, in step S3, the process of constructing negative samples of the entire data set specifically includes the following steps:

[0039] Step S31: randomly generating drug-target pairs that do not exist in the drug-target pair information, and extracting mismatched drugs and mismatched target proteins from the non-existent drug-target pairs;

[0040] Step S32: Aggregate the drug text sequence, drug multi-channel three-dimensional matrix and pharmacophore multi-channel three-dimensional matrix of the unmatched drug and the target protein text sequence, target protein multi-channel three-dimensional matrix and binding site multi-channel three-dimensional matrix of the unmatched target protein into a data set negative sample;

[0041] Step S33: adding a matching marker variable to the negative sample of the data set, and the value of the matching marker variable of the negative sample of the data set is 0;

[0042] Step S34: Repeat steps S31 to S33 until the number of negative samples in the data set is the same as the number of positive samples in the data set, thereby generating negative samples for all data sets.

[0043] Furthermore, in step S5, the process of constructing a multi-scale multi-modal neural network model specifically includes the following steps:

[0044] Step S51: constructing a text feature extraction branch, wherein the text feature extraction branch extracts latent variable features by analyzing drug text sequences and target protein text sequences;

[0045] Step S52: constructing a twin feature pyramid, wherein the twin feature pyramid extracts overall matching features, overall drug features, and overall target protein features by analyzing a drug multi-channel three-dimensional matrix and a target protein multi-channel three-dimensional matrix, and generates local matching features and image summary features by analyzing the overall drug features and the overall target protein features with the pharmacophore multi-channel three-dimensional matrix and the binding site multi-channel three-dimensional matrix;

[0046] Step S53: constructing a graphic feature fusion branch, wherein the graphic feature fusion branch integrates the overall matching feature and the local matching feature into the latent variable feature step by step to generate a fused matching feature;

[0047] Step S54: constructing a re-representation branch, wherein the re-representation branch fuses and re-represents the fused matching features and the image summary features, and outputs the drug-target interaction probability;

[0048] Step S55: construct a multi-scale multi-modal neural network model.

[0049] Furthermore, the process of constructing the text feature extraction branch in step S51 specifically includes the following steps:

[0050] Step S511: construct a first concatenation layer, an embedding layer, a first bidirectional LSTM layer, and a second bidirectional LSTM layer;

[0051] Step S512: Connect the first concatenation layer, the embedding layer, the first bidirectional LSTM layer and the second bidirectional LSTM layer in sequence to construct a text feature extraction branch.

[0052] Furthermore, the process of constructing the twin feature pyramid in step S52 specifically includes the following steps:

[0053] Step S521: constructing a drug molecular structure feature extraction branch and a target protein molecular structure feature extraction branch;

[0054] Step S522: constructing a binding site structure feature extraction branch, a pharmacophore structure feature extraction branch and a feature summary branch;

[0055] Step S523: constructing a first cosine similarity merging layer and a second cosine similarity merging layer.

[0056] Furthermore, the step S521 specifically includes the following steps:

[0057] Step S5211: constructing a first custom three-dimensional mode pooling layer, a first shared three-dimensional convolutional layer, a first custom discretization layer, a first Flatten layer and a third fully connected layer, wherein the first custom three-dimensional mode pooling layer takes the non-zero mode in its pooling window for pooling, and the first custom discretization layer first inverses the maximum-minimum normalization of the value in its input data and then rounds it, and then re-performs the maximum-minimum normalization on the rounded value and outputs it;

[0058] Step S5212: connecting the first custom three-dimensional mode pooling layer to the first shared three-dimensional convolutional layer, connecting the first shared three-dimensional convolutional layer to the first custom discretization layer and the first Flatten layer, connecting the first Flatten layer to the third fully connected layer, and constructing a drug molecular structure feature extraction branch;

[0059] Step S5213: constructing a second custom three-dimensional majority pooling layer, a second shared three-dimensional convolutional layer, a second custom discretization layer, a second Flatten layer and a fourth fully connected layer, wherein the first shared three-dimensional convolutional layer and the second shared three-dimensional convolutional layer share weights;

[0060] Step S5214: Connect the second custom three-dimensional majority pooling layer to the second shared three-dimensional convolutional layer, connect the second shared three-dimensional convolutional layer to the second custom discretization layer and the second Flatten layer, connect the second Flatten layer to the fourth fully connected layer, and construct a target protein molecular structure feature extraction branch.

[0061] Furthermore, the step S522 specifically includes the following steps:

[0062] Step S5221: construct a fifth splicing layer, a third shared three-dimensional convolutional layer, a third Flatten layer and a fifth fully connected layer, connect the fifth splicing layer, the third shared three-dimensional convolutional layer, the third Flatten layer and the fifth fully connected layer in sequence, connect the first custom discretization layer to the fifth splicing layer, and construct a pharmacophore structure feature extraction branch;

[0063] Step S5222: construct a sixth splicing layer, a fourth shared three-dimensional convolutional layer, a fourth Flatten layer and a sixth fully connected layer, connect the sixth splicing layer, the fourth shared three-dimensional convolutional layer, the fourth Flatten layer and the sixth fully connected layer in sequence, connect the second custom discretization layer to the sixth splicing layer, and construct a combined point structure feature extraction branch. The fourth shared three-dimensional convolutional layer and the third shared three-dimensional convolutional layer share weights;

[0064] Step S5223: Construct the seventh splicing layer, the fifth three-dimensional convolutional layer, the fifth Flatten layer and the third Highway layer, connect the seventh splicing layer, the fifth three-dimensional convolutional layer, the fifth Flatten layer and the third Highway layer in sequence, connect the third shared three-dimensional convolutional layer and the fourth shared three-dimensional convolutional layer to the seventh splicing layer, and construct a feature summary branch.

[0065] Furthermore, the step S523 specifically includes the following steps:

[0066] Step S5231: constructing a first cosine similarity merging layer and a second cosine similarity merging layer;

[0067] Step S5232: Connect the third fully connected layer and the fourth fully connected layer to the first cosine similarity merging layer, and connect the fifth fully connected layer and the sixth fully connected layer to the second cosine similarity merging layer.

[0068] Furthermore, the process of constructing the image-text feature fusion branch in step S53 specifically includes the following steps:

[0069] Step S531: construct the second splicing layer, the first Highway layer, the third splicing layer and the second Highway layer, and connect the second splicing layer, the first Highway layer, the third splicing layer and the second Highway layer in sequence;

[0070] Step S532: Connect the second bidirectional LSTM layer and the first cosine similarity merging layer to the second splicing layer, and connect the second cosine similarity merging layer to the third splicing layer to construct a graphic feature fusion branch.

[0071] Furthermore, the process of constructing the re-representation branch in step S54 specifically includes the following steps:

[0072] Step S541: construct a fourth concatenated layer, a first fully connected layer, and a second fully connected layer;

[0073] Step S542: connect the fourth splicing layer, the first fully connected layer, and the second fully connected layer in sequence, and connect the third Highway layer and the second Highway layer to the fourth splicing layer.

[0074] Furthermore, the step S55 specifically includes the following steps:

[0075] Step S551: specifying the first splicing layer, the first custom three-dimensional majority pooling layer, the second custom three-dimensional majority pooling layer, the fifth splicing layer and the sixth splicing layer as the input layer of the multi-scale multimodal neural network model, and specifying the second fully connected layer as the output layer of the multi-scale multimodal neural network model;

[0076] Step S552: Specify the loss function as binary_crossentropy and the optimizer as adam.

[0077] Furthermore, the step S6 specifically includes the following steps:

[0078] Step S61: designate the drug text sequence and target protein text sequence in the data set as the input of the first splicing layer, designate the drug multi-channel three-dimensional matrix in the data set as the input of the first custom three-dimensional majority pooling layer, designate the target protein multi-channel three-dimensional matrix in the data set as the input of the second custom three-dimensional majority pooling layer, designate the pharmacophore multi-channel three-dimensional matrix in the data set as the input of the fifth splicing layer, designate the binding site multi-channel three-dimensional matrix in the data set as the input of the sixth splicing layer, and designate the matching marker variable in the data set as the target output of the second fully connected layer;

[0079] Step S62: Divide the data set into a training set and a test set in a ratio of 8:2, and call the model.fit method to perform model training.

[0080] Furthermore, the step S7 specifically includes the following steps:

[0081] Step S71: collecting target protein data and target drug data;

[0082] Step S72: constructing a drug text sequence, a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix of the target drug and a target protein text sequence, a target protein multi-channel three-dimensional matrix and a binding site multi-channel three-dimensional matrix of the target protein according to the target protein data and the target drug data, and inputting them into a multi-scale multi-modal neural network model;

[0083] Step S73: The multi-scale multimodal neural network model outputs the interaction probability between the target drug and the target protein.

[0084] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0085] (1) The present invention provides a drug-target interaction prediction method based on artificial intelligence. In order to solve the problem that the drug-target interaction prediction method ignores the intrinsic discreteness of the chemical molecular structure and cannot capture the connection between the local and the whole, the present invention adopts a multi-scale multi-modal neural network model to predict drug-target interactions. It adopts a twin feature pyramid to take into account the local and overall characteristics of the chemical molecular structure from the whole to the local. The custom mode pooling layer and the custom discretization layer inside the twin feature pyramid effectively protect the intrinsic discreteness of the chemical molecular structure in the process of shrinking and extracting the molecular three-dimensional structure model, and organically combines computational chemistry theory with artificial intelligence technology to achieve multi-modal, multi-scale and high-precision drug-target interaction prediction.

[0086] (2) The design of the custom mode pooling layer and the custom discretization layer introduces the prior knowledge of the discreteness of the atomic number, that is, the intrinsic meaning of the atomic number is the difference in the type of elements. Common pooling methods include average pooling and maximum pooling, but average pooling will make the values ​​continuous, thereby causing irreversible damage to the information in the matrix, and maximum pooling will unreasonably emphasize the importance of heavy elements. Both of these phenomena are undesirable to us. However, the scale of the matrix is ​​too large, and it is impractical not to pool it from the perspective of computing power cost. Therefore, the present invention proposes a pooling method called mode pooling. , majority pooling, while maintaining the discreteness of the data, emphasizes the importance of the majority elements in the pooling window. The custom majority pooling layer can recover the information lost in the convolution process by rounding, improve the interpretability of the overall characteristics of the drug and the target protein, and make it possible to better combine the overall and local characteristics in the subsequent binding site structure feature extraction branch and pharmacophore structure feature extraction branch processing. The design of the custom majority pooling layer and the custom discretization layer effectively protects the chemical information in the matrix and achieves a more accurate and interpretable drug-target interaction prediction.

[0087] (3) The weight sharing mechanism introduced in the present invention reduces the model parameters while forcing the convolutional layer to learn more universal molecular structure features, thereby reducing computing power expenditure and improving the generalization performance of the multi-scale multi-modal neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 A schematic diagram of the multi-scale multi-modal neural network model proposed in the present invention;

[0089] Figure 2 Schematic diagram of the drug molecular structure feature extraction branch and the target protein molecular structure feature extraction branch of the multi-scale multi-modal neural network model;

[0090] Figure 3Schematic diagram of the binding site structure feature extraction branch and the pharmacophore structure feature extraction branch of the multi-scale multi-modal neural network model;

[0091] Figure 4 Schematic diagram of the feature summary branch of the multi-scale multimodal neural network model.

[0092] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0093] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0094] Embodiment 1, the present invention provides a drug-target interaction prediction method based on artificial intelligence, the method comprising the following steps:

[0095] Step S1: collecting drug-target pair information, target protein data and drug data. In this embodiment, the target protein data comes from Protein Data Bank, and the drug-target pair information and drug data come from DrugBank;

[0096] Step S2: construct positive samples of the entire data set;

[0097] Step S3: construct negative samples of the entire data set;

[0098] Step S4: Aggregate all positive samples of the data set and all negative samples of the data set into a data set;

[0099] Step S5: constructing a multi-scale multi-modal neural network model, wherein the multi-scale multi-modal neural network model includes a text feature extraction branch, a twin feature pyramid, a graphic feature fusion branch, and a re-representation branch;

[0100] Step S6: training a multi-scale multi-modal neural network model using the data set;

[0101] Step S7: Use a multi-scale multimodal neural network model for prediction.

[0102] Embodiment 2: This embodiment is based on the above embodiment, and step S2 includes the following steps:

[0103] Step S21: extracting a pair of drug and target protein from the drug-target pair information;

[0104] Step S22: constructing a drug text sequence, a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix corresponding to the drug and a target protein text sequence, a target protein multi-channel three-dimensional matrix and a binding site multi-channel three-dimensional matrix corresponding to the target protein, and summarizing them into a data set positive sample;

[0105] Step S23: adding a matching marker variable to the positive sample of the data set, where the value of the matching marker variable of the positive sample of the data set is 1;

[0106] Step S24: Repeat steps S21 to S23 for each pair of drug and target protein in the drug-target pair information until each pair of drug and target protein in the drug-target pair information has a corresponding data set positive sample, thereby generating positive samples for all data sets.

[0107] Embodiment 3: This embodiment is based on embodiment 2, and step S22 specifically includes the following steps:

[0108] Step S221: extracting the target protein three-dimensional structure model from the target protein data, extracting the drug three-dimensional structure model from the drug data, extracting the binding site three-dimensional structure model from the target protein three-dimensional structure model, and extracting the pharmacophore three-dimensional structure model from the drug three-dimensional structure model. In this embodiment, the extraction of the binding site three-dimensional structure model is implemented by using the SiteMap function of the Schrodinger software, and the extraction of the pharmacophore three-dimensional structure model is implemented by using the Phase function of the Schrodinger software;

[0109] Step S222: Matrixing the three-dimensional structural model of the target protein and the three-dimensional structural model of the binding site to construct a multi-channel three-dimensional matrix of the target protein and a multi-channel three-dimensional matrix of the binding site;

[0110] Step S223: extracting and encoding the data corresponding to the target protein except the target protein three-dimensional structure model from the target protein data to generate a target protein text sequence;

[0111] Step S224: Matrixing the drug three-dimensional structure model and the pharmacophore three-dimensional structure model to construct a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix. In this embodiment, the matrixing of the target protein three-dimensional structure model, the binding site three-dimensional structure model, the drug three-dimensional structure model and the pharmacophore three-dimensional structure model is implemented using the BioPython library;

[0112] Step S225: extracting the corresponding data of the drug except the drug three-dimensional structure model from the drug data and encoding them to generate a drug text sequence. In this embodiment, the drug text sequence and the drug text sequence are generated by using Keras word segmenter;

[0113] Step S226: Aggregate the drug text sequence, drug multi-channel three-dimensional matrix, pharmacophore multi-channel three-dimensional matrix, target protein text sequence, target protein multi-channel three-dimensional matrix and binding site multi-channel three-dimensional matrix into a data set positive sample.

[0114] Embodiment 4, based on Embodiment 3, the process of constructing a target protein multi-channel three-dimensional matrix and a binding site multi-channel three-dimensional matrix in step S222 specifically includes the following steps:

[0115] Step S2221: constructing a target protein three-dimensional matrix with the target protein three-dimensional structure model in units of 1 cubic angstrom, wherein a zero element of the target protein three-dimensional matrix indicates that there is no atom at this position in the target protein three-dimensional structure model, and a value of a non-zero element is determined by the atomic number of the corresponding atom in the target protein three-dimensional structure model;

[0116] Step S2222: numbering different types of binding points in the three-dimensional structure model of the binding points to generate binding point numbers;

[0117] Step S2223: partitioning the three-dimensional matrix of the target protein based on the three-dimensional structural model of the binding sites, assigning the non-zero elements of the regions corresponding to the binding sites in the three-dimensional structural model of the target protein to the corresponding binding site numbers, and generating a binding site region matrix;

[0118] Step S2224: superimposing the target protein three-dimensional matrix and the binding site region matrix into a target protein multi-channel three-dimensional matrix;

[0119] Step S2225: splitting the target protein three-dimensional matrix based on the binding site region matrix to generate a binding site three-dimensional matrix group, wherein a single matrix in the binding site three-dimensional matrix group corresponds to the region where a single binding site of the target protein three-dimensional matrix is ​​located;

[0120] Step S2226: superimposing the combination point three-dimensional matrix groups in the order of the combination point numbers to generate a combination point multi-channel three-dimensional matrix.

[0121] Embodiment 5, based on Embodiment 4, the process of constructing a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix in step S224 specifically includes the following steps:

[0122] Step S2241: constructing a three-dimensional drug matrix with the drug three-dimensional structure model as a unit of 1 cubic angstrom, wherein a zero element of the drug three-dimensional matrix indicates that there is no atom at this position in the drug three-dimensional structure model, and a value of a non-zero element is determined by the atomic number of the corresponding atom in the drug three-dimensional structure model;

[0123] Step S2242: numbering different types of pharmacophores in the pharmacophore three-dimensional structure model to generate pharmacophore numbers;

[0124] Step S2243: partitioning the drug three-dimensional matrix based on the pharmacophore three-dimensional structure model, assigning the non-zero elements in the region corresponding to the pharmacophore in the drug three-dimensional structure model to the corresponding pharmacophore number, and generating a pharmacophore region matrix;

[0125] Step S2244: superimposing the drug three-dimensional matrix and the pharmacophore region matrix into a drug multi-channel three-dimensional matrix;

[0126] Step S2245: splitting the drug three-dimensional matrix based on the pharmacophore region matrix to generate a pharmacophore three-dimensional matrix group, wherein a single matrix in the pharmacophore three-dimensional matrix group corresponds to a region where a single pharmacophore of the drug three-dimensional matrix is ​​located;

[0127] Step S2246: superimposing the pharmacophore three-dimensional matrix groups in the order of pharmacophore numbers to generate a pharmacophore multi-channel three-dimensional matrix.

[0128] Embodiment 6, this embodiment is based on embodiment 5, in step S3, the process of constructing negative samples of the entire data set specifically includes the following steps:

[0129] Step S31: randomly generating drug-target pairs that do not exist in the drug-target pair information, and extracting mismatched drugs and mismatched target proteins from the non-existent drug-target pairs. In this embodiment, randomly generating drug-target pairs that do not exist in the drug-target pair information is implemented using the Random library;

[0130] Step S32: Aggregate the drug text sequence, drug multi-channel three-dimensional matrix and pharmacophore multi-channel three-dimensional matrix of the unmatched drug and the target protein text sequence, target protein multi-channel three-dimensional matrix and binding site multi-channel three-dimensional matrix of the unmatched target protein into a data set negative sample;

[0131] Step S33: adding a matching marker variable to the negative sample of the data set, and the value of the matching marker variable of the negative sample of the data set is 0;

[0132] Step S34: Repeat steps S31 to S33 until the number of negative samples in the data set is the same as the number of positive samples in the data set, thereby generating negative samples for all data sets.

[0133] Embodiment 7, see Figure 1 This embodiment is based on the sixth embodiment. In step S5, the process of constructing a multi-scale multi-modal neural network model specifically includes the following steps:

[0134] Step S51: constructing a text feature extraction branch, wherein the text feature extraction branch extracts latent variable features by analyzing drug text sequences and target protein text sequences;

[0135] Step S52: constructing a twin feature pyramid, wherein the twin feature pyramid extracts overall matching features, overall drug features, and overall target protein features by analyzing a drug multi-channel three-dimensional matrix and a target protein multi-channel three-dimensional matrix, and generates local matching features and image summary features by analyzing the overall drug features and the overall target protein features with the pharmacophore multi-channel three-dimensional matrix and the binding site multi-channel three-dimensional matrix;

[0136] Step S53: constructing a graphic feature fusion branch, wherein the graphic feature fusion branch integrates the overall matching feature and the local matching feature into the latent variable feature step by step to generate a fused matching feature;

[0137] Step S54: constructing a re-representation branch, wherein the re-representation branch fuses and re-represents the fused matching features and the image summary features, and outputs the drug-target interaction probability;

[0138] Step S55: construct a multi-scale multi-modal neural network model.

[0139] Embodiment 8, based on Embodiment 7, the process of constructing the text feature extraction branch in step S51 specifically includes the following steps:

[0140] Step S511: construct a first concatenation layer, an embedding layer, a first bidirectional LSTM layer, and a second bidirectional LSTM layer;

[0141] Step S512: Connect the first concatenation layer, the embedding layer, the first bidirectional LSTM layer and the second bidirectional LSTM layer in sequence to construct a text feature extraction branch.

[0142] Embodiment 9: This embodiment is based on embodiment 8, and the process of constructing the twin feature pyramid in step S52 specifically includes the following steps:

[0143] Step S521: constructing a drug molecular structure feature extraction branch and a target protein molecular structure feature extraction branch;

[0144] Step S522: constructing a binding site structure feature extraction branch, a pharmacophore structure feature extraction branch and a feature summary branch;

[0145] Step S523: constructing a first cosine similarity merging layer and a second cosine similarity merging layer.

[0146] Embodiment 10, see Figure 2 This embodiment is based on the ninth embodiment, and the step S521 specifically includes the following steps:

[0147] Step S5211: constructing a first custom three-dimensional mode pooling layer, a first shared three-dimensional convolutional layer, a first custom discretization layer, a first Flatten layer and a third fully connected layer, wherein the first custom three-dimensional mode pooling layer takes the non-zero mode in its pooling window for pooling, and the first custom discretization layer first inverses the maximum-minimum normalization of the value in its input data and then rounds it, and then re-performs the maximum-minimum normalization on the rounded value and outputs it;

[0148] Step S5212: connecting the first custom three-dimensional mode pooling layer to the first shared three-dimensional convolutional layer, connecting the first shared three-dimensional convolutional layer to the first custom discretization layer and the first Flatten layer, connecting the first Flatten layer to the third fully connected layer, and constructing a drug molecular structure feature extraction branch;

[0149] Step S5213: construct a second custom three-dimensional majority pooling layer, a second shared three-dimensional convolutional layer, a second custom discretization layer, a second Flatten layer and a fourth fully connected layer, wherein the first shared three-dimensional convolutional layer and the second shared three-dimensional convolutional layer share weights. In this embodiment, the first custom discretization layer, the second custom discretization layer, the first custom three-dimensional majority pooling layer and the second custom three-dimensional majority pooling layer are implemented using Keras layers.core.Lambda;

[0150] Step S5214: Connect the second custom three-dimensional majority pooling layer to the second shared three-dimensional convolutional layer, connect the second shared three-dimensional convolutional layer to the second custom discretization layer and the second Flatten layer, connect the second Flatten layer to the fourth fully connected layer, and construct a target protein molecular structure feature extraction branch.

[0151] Example 11, see Figure 3 and Figure 4 This embodiment is based on the tenth embodiment, and the step S522 specifically includes the following steps:

[0152] Step S5221: construct a fifth splicing layer, a third shared three-dimensional convolutional layer, a third Flatten layer and a fifth fully connected layer, connect the fifth splicing layer, the third shared three-dimensional convolutional layer, the third Flatten layer and the fifth fully connected layer in sequence, connect the first custom discretization layer to the fifth splicing layer, and construct a pharmacophore structure feature extraction branch;

[0153] Step S5222: construct a sixth splicing layer, a fourth shared three-dimensional convolutional layer, a fourth Flatten layer and a sixth fully connected layer, connect the sixth splicing layer, the fourth shared three-dimensional convolutional layer, the fourth Flatten layer and the sixth fully connected layer in sequence, connect the second custom discretization layer to the sixth splicing layer, and construct a combined point structure feature extraction branch. The fourth shared three-dimensional convolutional layer and the third shared three-dimensional convolutional layer share weights;

[0154] Step S5223: Construct the seventh splicing layer, the fifth three-dimensional convolutional layer, the fifth Flatten layer and the third Highway layer, connect the seventh splicing layer, the fifth three-dimensional convolutional layer, the fifth Flatten layer and the third Highway layer in sequence, connect the third shared three-dimensional convolutional layer and the fourth shared three-dimensional convolutional layer to the seventh splicing layer, and construct a feature summary branch.

[0155] Embodiment 12: This embodiment is based on embodiment 11, and step S523 specifically includes the following steps:

[0156] Step S5231: constructing a first cosine similarity merging layer and a second cosine similarity merging layer;

[0157] Step S5232: Connect the third fully connected layer and the fourth fully connected layer to the first cosine similarity merging layer, and connect the fifth fully connected layer and the sixth fully connected layer to the second cosine similarity merging layer.

[0158] Embodiment 13, based on Embodiment 12, the process of constructing the image-text feature fusion branch in step S53 specifically includes the following steps:

[0159] Step S531: construct the second splicing layer, the first Highway layer, the third splicing layer and the second Highway layer, and connect the second splicing layer, the first Highway layer, the third splicing layer and the second Highway layer in sequence;

[0160] Step S532: Connect the second bidirectional LSTM layer and the first cosine similarity merging layer to the second splicing layer, and connect the second cosine similarity merging layer to the third splicing layer to construct a graphic feature fusion branch.

[0161] Embodiment 14: This embodiment is based on embodiment 13, and the process of constructing the re-representation branch in step S54 specifically includes the following steps:

[0162] Step S541: construct a fourth concatenated layer, a first fully connected layer, and a second fully connected layer;

[0163] Step S542: connect the fourth splicing layer, the first fully connected layer, and the second fully connected layer in sequence, and connect the third Highway layer and the second Highway layer to the fourth splicing layer.

[0164] Embodiment 15: This embodiment is based on Embodiment 14, and step S55 specifically includes the following steps:

[0165] Step S551: specifying the first splicing layer, the first custom three-dimensional majority pooling layer, the second custom three-dimensional majority pooling layer, the fifth splicing layer and the sixth splicing layer as the input layer of the multi-scale multimodal neural network model, and specifying the second fully connected layer as the output layer of the multi-scale multimodal neural network model;

[0166] Step S552: Specify the loss function as binary_crossentropy and the optimizer as adam.

[0167] Embodiment 16: This embodiment is based on Embodiment 15, and step S6 specifically includes the following steps:

[0168] Step S61: designate the drug text sequence and target protein text sequence in the data set as the input of the first splicing layer, designate the drug multi-channel three-dimensional matrix in the data set as the input of the first custom three-dimensional majority pooling layer, designate the target protein multi-channel three-dimensional matrix in the data set as the input of the second custom three-dimensional majority pooling layer, designate the pharmacophore multi-channel three-dimensional matrix in the data set as the input of the fifth splicing layer, designate the binding site multi-channel three-dimensional matrix in the data set as the input of the sixth splicing layer, and designate the matching marker variable in the data set as the target output of the second fully connected layer;

[0169] Step S62: Divide the data set into a training set and a test set in a ratio of 8:2, and call the model.fit method to perform model training.

[0170] Embodiment 17: This embodiment is based on Embodiment 16, and step S7 specifically includes the following steps:

[0171] Step S71: collecting target protein data and target drug data;

[0172] Step S72: constructing a drug text sequence, a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix of the target drug and a target protein text sequence, a target protein multi-channel three-dimensional matrix and a binding site multi-channel three-dimensional matrix of the target protein according to the target protein data and the target drug data, and inputting them into a multi-scale multi-modal neural network model;

[0173] Step S73: The multi-scale multimodal neural network model outputs the interaction probability between the target drug and the target protein.

[0174] Example 18. This example is based on Example 17. This example runs in a Windows operating system environment. The experiment relies on the Anaconda3 environment, and uses Keras as the model framework of the multi-scale multimodal neural network model. The Pandas library, Numpy library, and Scikit-learn library are used to assist in completing the data set construction work.

[0175] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0176] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0177] The present invention and its embodiments are described above, which is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical scheme, which should all fall within the scope of protection of the present invention.

Claims

1. A drug-target interaction prediction method based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1: collecting drug-target pair information, target protein data and drug data; Step S2: construct positive samples of the entire data set; Step S3: construct negative samples of the entire data set; Step S4: Aggregate all the positive samples of the data set and all the negative samples of the data set into a data set; Step S5: constructing a multi-scale multi-modal neural network model, wherein the multi-scale multi-modal neural network model includes a text feature extraction branch, a twin feature pyramid, a graphic feature fusion branch, and a re-representation branch; Step S6: training the multi-scale multi-modal neural network model using the data set; Step S7: using a multi-scale multi-modal neural network model for prediction; In step S5, the process of constructing a multi-scale multi-modal neural network model specifically includes the following steps: Step S51: construct a text feature extraction branch; Step S52: constructing a twin feature pyramid; Step S53: constructing a graphic-text feature fusion branch; Step S54: construct a re-representation branch; Step S55: constructing a multi-scale multi-modal neural network model; The step S52 specifically includes the following steps: Step S521: constructing a drug molecular structure feature extraction branch and a target protein molecular structure feature extraction branch; Step S522: constructing a binding site structure feature extraction branch, a pharmacophore structure feature extraction branch and a feature summary branch; Step S523: constructing a first cosine similarity merging layer and a second cosine similarity merging layer; The step S521 specifically includes the following steps: Step S5211: constructing a first custom three-dimensional mode pooling layer, a first shared three-dimensional convolutional layer, a first custom discretization layer, a first Flatten layer and a third fully connected layer, wherein the first custom three-dimensional mode pooling layer takes the non-zero mode in its pooling window for pooling, and the first custom discretization layer first inverses the maximum-minimum normalization of the value in its input data and then rounds it, and then re-performs the maximum-minimum normalization on the rounded value and outputs it; Step S5212: connecting the first custom three-dimensional mode pooling layer to the first shared three-dimensional convolutional layer, connecting the first shared three-dimensional convolutional layer to the first custom discretization layer and the first Flatten layer, connecting the first Flatten layer to the third fully connected layer, and constructing a drug molecular structure feature extraction branch; Step S5213: constructing a second custom three-dimensional majority pooling layer, a second shared three-dimensional convolutional layer, a second custom discretization layer, a second Flatten layer and a fourth fully connected layer, wherein the first shared three-dimensional convolutional layer and the second shared three-dimensional convolutional layer share weights; Step S5214: Connect the second custom three-dimensional majority pooling layer to the second shared three-dimensional convolutional layer, connect the second shared three-dimensional convolutional layer to the second custom discretization layer and the second Flatten layer, connect the second Flatten layer to the fourth fully connected layer, and construct a target protein molecular structure feature extraction branch.

2. The method for predicting drug-target interaction based on artificial intelligence according to claim 1, characterized in that: The step S2 comprises the following steps: Step S21: extracting a pair of drug and target protein from the drug-target pair information; Step S22: constructing a drug text sequence, a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix corresponding to the drug and a target protein text sequence, a target protein multi-channel three-dimensional matrix and a binding site multi-channel three-dimensional matrix corresponding to the target protein, and summarizing them into a data set positive sample; Step S23: adding a matching marker variable to the positive sample of the data set, where the value of the matching marker variable of the positive sample of the data set is 1; Step S24: Repeat steps S21 to S23 for each pair of drug and target protein in the drug-target pair information until each pair of drug and target protein in the drug-target pair information has a corresponding data set positive sample, thereby generating positive samples for all data sets.

3. The method for predicting drug-target interaction based on artificial intelligence according to claim 2, characterized in that: The step S22 specifically includes the following steps: Step S221: extracting a target protein three-dimensional structure model from the target protein data, extracting a drug three-dimensional structure model from the drug data, extracting a binding site three-dimensional structure model from the target protein three-dimensional structure model, and extracting a pharmacophore three-dimensional structure model from the drug three-dimensional structure model; Step S222: Matrixing the three-dimensional structural model of the target protein and the three-dimensional structural model of the binding site to construct a multi-channel three-dimensional matrix of the target protein and a multi-channel three-dimensional matrix of the binding site; Step S223: extracting and encoding the data corresponding to the target protein except the target protein three-dimensional structure model from the target protein data to generate a target protein text sequence; Step S224: Matrixing the drug three-dimensional structure model and the pharmacophore three-dimensional structure model to construct a drug multi-channel three-dimensional matrix and a pharmacophore multi-channel three-dimensional matrix; Step S225: extracting data corresponding to the drug except the drug three-dimensional structure model from the drug data and encoding them to generate a drug text sequence; Step S226: Aggregate the drug text sequence, drug multi-channel three-dimensional matrix, pharmacophore multi-channel three-dimensional matrix, target protein text sequence, target protein multi-channel three-dimensional matrix and binding site multi-channel three-dimensional matrix into a data set positive sample.

4. The method for predicting drug-target interaction based on artificial intelligence according to claim 3, characterized in that: The process of constructing the target protein multi-channel three-dimensional matrix and the binding site multi-channel three-dimensional matrix in step S222 specifically includes the following steps: Step S2221: constructing a target protein three-dimensional matrix with the target protein three-dimensional structure model in units of 1 cubic angstrom, wherein a zero element of the target protein three-dimensional matrix indicates that there is no atom at this position in the target protein three-dimensional structure model, and a value of a non-zero element is determined by the atomic number of the corresponding atom in the target protein three-dimensional structure model; Step S2222: numbering different types of binding points in the three-dimensional structure model of the binding points to generate binding point numbers; Step S2223: partitioning the three-dimensional matrix of the target protein based on the three-dimensional structural model of the binding sites, assigning the non-zero elements of the regions corresponding to the binding sites in the three-dimensional structural model of the target protein to the corresponding binding site numbers, and generating a binding site region matrix; Step S2224: superimposing the target protein three-dimensional matrix and the binding site region matrix into a target protein multi-channel three-dimensional matrix; Step S2225: splitting the target protein three-dimensional matrix based on the binding site region matrix to generate a binding site three-dimensional matrix group, wherein a single matrix in the binding site three-dimensional matrix group corresponds to the region where a single binding site of the target protein three-dimensional matrix is ​​located; Step S2226: superimposing the combination point three-dimensional matrix groups in the order of the combination point numbers to generate a combination point multi-channel three-dimensional matrix.

Images