Drug target matching recommendation method based on multi-view neural network

Abstract

The invention belongs to the field of computational biology, and particularly relates to a drug target matching recommendation method based on a multi-view neural network, and the method comprises the steps: obtaining to-be-predicted drug target information, inputting the to-be-predicted drug target information into a trained drug target recommendation model, selecting a drug target whose prediction probability is greater than an interaction threshold, and recommending the drug target to a user; according to the method, the feature vectors of the drug molecules are extracted by using the graph neural network for carrying out message passing based on points and edges, modeling and feature extraction are carried out on the drug molecules from two angles of atoms and chemical bonds, and the globality of drug sequence information is considered at the same time; the atomic characteristics, chemical bond characteristics and sequence characteristics of the drug molecules are interacted by using a cross attention mechanism, the complex relationship between the drug molecules and the target protein is captured more accurately, and the accuracy of drug-target interaction prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the field of computational biology, and in particular relates to a drug target matching recommendation method based on a multi-view graph neural network. Background Art

[0002] The drug development process is very difficult. Traditional methods based on biological experiments can more accurately screen out drug-target pairs that can react, such as high-throughput screening tests. However, this method is too cumbersome and requires a lot of time and money. In addition, the existence of millions of compounds and hundreds of potential targets makes this technology face great challenges. With the rapid development of bioinformatics, computational biology and drug development technology, bioinformatics databases have accumulated a large amount of biological experimental data, and the focus of drug research has gradually shifted from traditional laboratory experiments to computational methods. In the process of drug development, drug-target interaction prediction has become a key area.

[0003] The current mainstream method is to use the graph structure data of drug molecules and the sequence data of protein molecules to predict drug-target interactions. However, existing graph neural networks often only focus on the information connection between different atoms in drug molecules, ignoring that the chemical bonds connecting atoms are equally important in the chemical properties of drug molecules. The feature extraction of graph neural networks should not only consider the information transfer between atoms, but should also consider the information transfer between atoms and chemical bonds. At the same time, the sequence information of drugs is often discarded after being converted into graph structures. However, the graph structure data of drugs can only express the local information of drug molecules to a certain extent, while the molecular sequence can reflect the global information of drugs. Making full use of the graph structure data and sequence structure data of molecules can further improve the accuracy of model prediction. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a drug target matching recommendation method based on a multi-view graph neural network, comprising:

[0005] Obtain the information of the drug target to be predicted, input it into the trained drug target recommendation model, and select the drug targets with a prediction probability greater than the interaction threshold to recommend to the user;

[0006] The training process of the drug target recommendation model includes:

[0007] S1: Obtain drug target information, including drug SMILES sequence, molecular graph structure and target amino acid sequence;

[0008] S2: According to the atomic and chemical bond properties of the drug, the atomic initial feature vector and the chemical bond initial feature vector are constructed respectively, and the atomic feature vector h of the drug molecule graph is extracted using the graph neural network based on point feature message passing and the graph neural network based on edge feature message passing respectively. atom and the chemical bond characteristic vector h bond ;

[0009] S3: Obtain the sequence feature h of the drug based on the drug SMILES sequence seq , the cross-attention mechanism is used to interact the atomic feature vector and chemical bond feature vector of the drug with the drug sequence feature vector respectively, and the atomic and chemical bond feature vectors after interaction are concatenated as the final feature vector H of the drug molecule d ;

[0010] S4: Obtain the target sequence feature vector H according to the target amino acid sequence p , the feature vector of the drug molecule and the feature vector of the target sequence are interacted to obtain the matching probability of the drug target;

[0011] S5: The weight of the recommendation model is updated by calculating the loss based on the probability output by the model and the acquired drug target matching information to obtain a trained drug target recommendation model.

[0012] Beneficial effects of the present invention:

[0013] The present invention utilizes a graph neural network based on point and edge message passing to extract feature vectors of drug molecules, models and extracts features of drug molecules from the perspectives of atoms and chemical bonds, and takes into account the globality of drug sequence information. It utilizes a cross-attention mechanism to interact with the atomic features, chemical bond features, and sequence features of drug molecules, so as to more accurately capture the complex relationship between drug molecules and target proteins and improve the accuracy of drug-target interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a schematic diagram of the overall process of a drug target matching recommendation method based on a multi-view graph neural network of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] Embodiment 1: In this embodiment, the present invention provides a drug target matching recommendation method based on a multi-view graph neural network, such as Figure 1 As shown, including:

[0017] The SMILES sequence and molecular graph structure of the drug to be predicted, as well as the amino acid sequence information of the target to be predicted, are obtained and input into the trained drug target recommendation model. The drug targets with a prediction probability greater than the interaction threshold are selected and recommended to the user.

[0018] The training process of the drug target recommendation model includes:

[0019] S1: Obtain drug target information, including drug SMILES sequence, molecular graph structure and target amino acid sequence;

[0020] S2: According to the atomic and chemical bond properties of the drug, the atomic initial feature vector and the chemical bond initial feature vector are constructed respectively, and the atomic feature vector h of the drug molecule graph is extracted using the graph neural network based on point feature message passing and the graph neural network based on edge feature message passing respectively. atom and the chemical bond characteristic vector h bond ;

[0021] S3: Obtain the sequence feature h of the drug based on the drug SMILES sequence seq , the cross-attention mechanism is used to interact the atomic feature vector and chemical bond feature vector of the drug with the drug sequence feature vector respectively, and the atomic and chemical bond feature vectors after interaction are concatenated as the final feature vector H of the drug molecule d ;

[0022] S4: Obtain the target sequence feature vector H according to the target amino acid sequence p , the feature vector of the drug molecule and the feature vector of the target sequence are interacted to obtain the matching probability of the drug target;

[0023] S5: The weight of the recommendation model is updated by calculating the loss based on the probability output by the model and the acquired drug target matching information to obtain a trained drug target recommendation model.

[0024] According to the atomic and chemical bond properties of the drug, the atomic initial feature vector and the chemical bond initial feature vector are constructed respectively, including:

[0025] S21: One-hot encode the atomic number, number of chemical bonds, formal charge, chirality (unspecified, tetrahedral CW / CCW and other) and hybridization mode (sp, sp2, sp3, sp3d and sp3d2) of each atom, and the obtained 01 feature vector is used as the initial feature vector h of the atom. 0 (v), where v represents an atom and the characteristic dimension is d 1 .

[0026] S22: One-hot encode the chemical bond type (single, double, triple and aromatic), whether it is a conjugated bond, whether it is part of a ring and the stereochemistry (none, any, E / Z and cis / trans) of each chemical bond to obtain a 01 feature vector as the initial feature vector h of the chemical bond 0 (e uv ), where e uv Represents the chemical bond between atoms u and v, with a characteristic dimension of d 2 .

[0027] The graph neural network based on point feature message passing includes:

[0028] The graph neural network based on point feature message passing contains T time steps. In each time step t, the atom v is based on the message function Aggregate information from atoms adjacent to atom v to obtain vector m t+1 (v), update function Using m t+1 (v), the initial eigenvector h of atom v 0 (v) and the feature vector h at time step t t (v) Get the feature vector h of atom v at the next time step t+1 t+1 (v) m t+1 (v) and h t+1 The calculation formula of (v) is as follows.

[0029]

[0030]

[0031] Where N(v) represents the set of atoms adjacent to atom v, m t+1 (v) and h t+1 The feature dimension of (v) is d 1 .

[0032] Message function of graph neural network based on point message passing at time step t The calculation formula is:

[0033]

[0034] Among them, ⊙ represents element-by-element multiplication, k uv represents the attention weight between atom u and atom v. The message function first calculates the maximum value of each feature dimension in the set of adjacent atoms of atom v to obtain the most important adjacent atom information. Then weightedly add up the feature vectors of all adjacent atoms to obtain ∑ u∈N(v) k uv h t (u). Multiply the two feature vectors element by element to obtain the aggregate information vector m of atom v at the next time step. t+1 (v).

[0035] The weight k between atom u and atom v uv The calculation formula is:

[0036]

[0037] Among them, W k is the third trainable parameter matrix, dimension is 2d 1 ×d 1 || represents the concatenation of feature vectors. τ is the temperature coefficient hyperparameter used to control k uv The steepness of the distribution, exp represents the exponential function, ReLU represents the activation function, h(u) represents the feature representation of atom u, h(v) represents the feature representation of atom v, N(u) represents the set of atoms adjacent to atom u, and h(l) represents the feature representation of atom l.

[0038] Update function of graph neural network based on point message passing at time step t The calculation formula is:

[0039]

[0040] Update Function The aggregate information vector m of atom v t+1 (v) and the initial eigenvector h 0 (v) Update its feature vector h at the next time step t+1 (v). ELU is the Exponential Linear Units activation function; is a trainable parameter matrix with dimension 2d 1 ×d 1 , where a is the number of atoms in the drug molecule; h t+1 The feature dimension of (b) is d 1 .

[0041] At the last time step, all atoms in the drug molecule are spliced ​​to obtain the atomic feature vector h of the drug molecule. atom , with dimensions a×d 1 , a represents the number of atoms.

[0042] The graph neural network based on edge feature message passing consists of T time steps. In each time step t, the feature vector of the chemical bond between atom u and atom v is calculated according to the message function The vector m is obtained by aggregating the neighboring chemical bonds. t+1 (e uv ), the update function uses the vector m t+1 (e uv ) and chemical bond e uv The initial eigenvector h 0 (e uv ) to get e uv The feature vector h at the next time step t+1 t+1 (e uv ). t+1 (e uv ) and h t+1 (e uv ) is calculated as:

[0043]

[0044]

[0045] Where N(v)\u represents the set of atoms adjacent to atom v but not including atom u. The message function of the graph neural network based on edge message passing at time step t The calculation formula is:

[0046]

[0047] The update function calculation formula of the graph neural network based on edge message passing at time step t is:

[0048]

[0049] is a trainable parameter matrix with dimension 2d 2 ×d 2 . m t+1 (e uv ) and h t+1 (e uv ) has a feature vector dimension of d 2 In the last time step T of the graph neural network based on edge feature message passing, the chemical bond e uv The eigenvector of is converted into the eigenvector of atom u and atom v. The conversion formula is:

[0050]

[0051]

[0052] Among them, We is the fourth trainable parameter matrix with dimension d 2 ×d 1 .

[0053] In the last time step, all atoms in the drug molecule are spliced ​​to obtain the chemical bond feature vector h of the drug molecule. bond , with dimensions a×d 1 , a represents the number of atoms.

[0054] Obtain the sequence feature h of the drug based on the drug SMILES sequence seq , the cross-attention mechanism is used to interact the atomic feature vector and chemical bond feature vector of the drug with the drug sequence feature vector respectively, and the atomic and chemical bond feature vectors after interaction are concatenated as the final feature vector H of the drug molecule d ,include:

[0055] S31: Set a randomly initialized feature vector of dimension d for each character in the SMILES sequence 1 , a one-dimensional gated convolutional neural network is used for feature extraction to obtain the SMILES sequence feature h seq , with dimensions of c×d 1 .

[0056] S32: The atomic feature vector h of the drug molecule atom As the Query in the attention mechanism, the sequence feature h seq As Key and Value, the atomic feature vector H after interaction is calculated using the attention mechanism. atom ,include:

[0057]

[0058] Among them, H atom The feature dimension is d 1 , T1 represents matrix transpose, and Softmax represents the activation function.

[0059] S33: The chemical bond feature vector h of the drug molecule bond As the Query in the attention mechanism, the sequence feature h sep As Key and Value, the atomic feature vector H after interaction is calculated according to the attention mechanism. bond ,include:

[0060]

[0061] Among them, H bond The feature dimension is d 1 .

[0062] S34: H atom and H bond After splicing, the final feature vector H of the drug molecule is obtained d .

[0063] According to the amino acid sequence of the target, the target sequence feature vector H is obtained p , the feature vector of the drug molecule and the feature vector of the target sequence are interacted to obtain the matching probability of the drug target, including:

[0064] S41: Divide the target amino acid sequence into multiple overlapping ternary amino acid subsequences, and assign a randomly initialized feature vector with dimension d to each ternary amino acid subsequence. 1 . Feature extraction is performed through a one-dimensional gated convolutional neural network to obtain the target sequence feature vector H p , with dimensions of b×d 1 ;

[0065] The calculation formula for the i-th layer of a one-dimensional gated convolutional neural network is:

[0066]

[0067] Among them, X i+1 is based on the i-th layer input X i The obtained feature map vector, and X 0 A randomly initialized feature vector assigned to each triad of amino acid subsequences. and is the fifth and sixth trainable parameter matrix, with dimension f×m 1 ×m 2 , f represents the convolution kernel size, m 1 and m 2 The distribution represents the dimensions of input and output. The output of the last layer is H p .

[0068] S42: The characteristic vector H of the drug molecule d As the Query in the attention mechanism, the target sequence feature H p As Key and Value, the interactive feature vector is calculated according to the attention mechanism. The formula is:

[0069]

[0070] S43: The characteristic vector H of the drug molecule p As the Query in the attention mechanism, the target sequence feature H d As Key and Value, the interactive feature vector is calculated according to the attention mechanism. The formula is:

[0071]

[0072] S44: Transform the eigenvector by linear transformation and Map them to the same dimension d, concatenate the two and send them to the binary classification linear layer to get the probability of drug and target matching The formula is:

[0073]

[0074] Among them, W out represents the trainable parameter matrix of the binary linear layer, and σ represents the Sigmoid activation function.

[0075] The loss is calculated based on the probability of the model output and the acquired drug target matching information to update the weight of the recommendation model and obtain a trained drug target recommendation model. The loss function calculation formula is:

[0076]

[0077] in represents the total loss of the model, N represents the number of samples, and y i represents the label of the i-th sample, Represents the model's prediction result for the i-th sample,

[0078] Obtain the drug molecule information and target protein information to be predicted, input them into the trained model, obtain the interaction prediction probability of the drug target, and select the drug targets with interaction probability greater than the interaction threshold to recommend to the user; preferably, set the interaction threshold to 0.65.

[0079] 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.

Claims

1. A drug target matching recommendation method based on multi-view graph neural network, characterized in that: include: Obtain the information of the drug target to be predicted, input it into the trained drug target recommendation model, and select the drug targets with a prediction probability greater than the interaction threshold to recommend to the user; The training process of the drug target recommendation model includes: S1: Obtain drug target information, including drug SMILES sequence, molecular graph structure and target amino acid sequence; S2: According to the atomic and chemical bond properties of the drug, the atomic initial feature vector and the chemical bond initial feature vector are constructed respectively, and the atomic feature vector h of the drug molecule graph is extracted using the graph neural network based on point feature message passing and the graph neural network based on edge feature message passing respectively. atom and the chemical bond characteristic vector h bond ; S3: Obtain the sequence feature h of the drug based on the drug SMILES sequence seq , the cross-attention mechanism is used to interact the atomic feature vector and chemical bond feature vector of the drug with the drug sequence feature vector respectively, and the atomic and chemical bond feature vectors after interaction are concatenated as the final feature vector H of the drug molecule d ; S4: Obtain the target sequence feature vector H according to the target amino acid sequence p , the feature vector of the drug molecule and the feature vector of the target sequence are interacted to obtain the matching probability of the drug target; S5: The weight of the recommendation model is updated by calculating the loss based on the probability output by the model and the acquired drug target matching information to obtain a trained drug target recommendation model.

2. According to claim 1, a drug target matching recommendation method based on multi-view graph neural network is characterized in that: According to the atomic and chemical bond properties of the drug, the atomic initial feature vector and the chemical bond initial feature vector are constructed respectively, including: S21: One-hot encode the atomic number, number of chemical bonds, formal charge, chirality and hybridization mode of each atom, and the obtained 01 feature vector is used as the initial feature vector h of the atom. 0 (v), where v represents an atom and the characteristic dimension is d1. S22: One-hot encode the chemical bond type, whether it is a conjugated bond, whether it is part of a ring, and the stereochemistry of each chemical bond, and the resulting 01 feature vector is used as the initial feature vector h of the chemical bond. 0 (e uv ), where e uv Represents the chemical bond between atoms u and v, with a characteristic dimension of d2.

3. The drug target matching recommendation method based on multi-view graph neural network according to claim 1, characterized in that: Graph neural network based on point feature message passing, including: The graph neural network based on point feature message passing contains T time steps. In each time step t, the atom v is transmitted according to the message function of the graph neural network based on point feature message passing. Aggregate information from atoms adjacent to atom v to obtain the vector m at the next time step t+1 t+1 (v) Update function of graph neural network based on point feature message passing Using vector mt+ 1 (v), the initial eigenvector h of atom v 0 (v) and the feature vector h at time step t t (v) Get the feature vector h of atom v at the next time step t+1 t+1 (v); At the last time step T, all atomic features are integrated to obtain the atomic feature vector h atom ; m t+1 (v) and h t+1 The calculation formula of (v) is as follows: Where N(v) represents the set of atoms adjacent to atom v, m t+1 (v) and h t+1 The feature dimensions of (v) are d1, h t (u) represents the eigenvector of atom u adjacent to atom v at time step t.

4. The drug target matching recommendation method based on multi-view graph neural network according to claim 3 is characterized in that: The message function of the graph neural network based on point feature message passing include: Where N(v) represents the set of atoms adjacent to atom v, ⊙ represents element-by-element multiplication, and ku v represents the attention weight between atoms u and v, h t (u) represents the eigenvector of atom u adjacent to atom v at time step t; The update function of the graph neural network based on point feature message passing include: Among them, ELU represents the ExponentialLinear Units activation function, represents the first trainable parameter matrix with a dimension of 2d1×d1; h t+1 The feature dimensions of (v) are d1, h t (v) represents the eigenvector of atom v at time step t, m t+1 (v) represents the vector of atom v at the next time step t+1, h 0 (v) represents the initial eigenvector of atom v.

5. A drug target matching recommendation method based on a multi-view graph neural network according to claim 1, characterized in that: The graph neural network based on edge feature message passing includes: The graph neural network based on edge feature message passing contains T time steps. In each time step t, the feature vector of the chemical bond between atom u and atom v is calculated according to the message function of the graph neural network based on edge feature message passing. Aggregate from neighboring chemical bonds to get the vector m at the next time step t+1 t+1 (e uv ), the update function of the graph neural network based on edge feature message passing Using vector m t+1 (e uv ), chemical bond e uv The initial eigenvector h 0 (e uv ) and the feature vector h at time step t t (e uv ) to get e uv The feature vector h at the next time step t+1 t+1 (e uv ); At the last time step T, all atomic features are integrated to obtain the atomic feature vector h bond ; m t+1 (e uv ) and h t+1 (e uv ) is calculated as: Where N(v)\u represents the set of atoms adjacent to atom v but not including u, and h t (e vk ) represents the chemical bond e between atom u and atom k vk The feature vector at time step t.

6. The drug target matching recommendation method based on multi-view graph neural network according to claim 5, characterized in that: The message function of the graph neural network based on edge feature message passing include: The update function of the graph neural network based on edge feature message passing include: Among them, ELU represents the Exponential Linear Units activation function, N(v)\u represents the set of atoms adjacent to atom v but not including u, ⊙ represents element-by-element multiplication, Represents the second trainable parameter matrix with a dimension of 2d2×d2; m t+1 (eu v ) and h t+1 (e uv ) has a feature vector dimension of d2, h t (e vk ) represents the chemical bond e between atom u and atom k vk The feature vector at time step t, h t (e uv ) represents the chemical bond e between atom u and atom v uv The feature vector at time step t, h 0 (e uv ) represents the chemical bond e between atom u and atom v uv The initial eigenvector, m t+1 (e uv ) represents the chemical bond e between atom u and atom v uv The feature vector at the next time step t+1.

7. The drug target matching recommendation method based on multi-view graph neural network according to claim 1, characterized in that: The drug sequence feature hseq is obtained according to the drug SMILES sequence. The atomic feature vector and chemical bond feature vector of the drug are interacted with the drug sequence feature vector respectively by using the cross attention mechanism. The atomic and chemical bond feature vectors after interaction are concatenated as the final feature vector H of the drug molecule. d ,include: S31: Set a randomly initialized feature vector for each character in the SMILES sequence with a dimension of d1, and use a one-dimensional gated convolutional neural network for feature extraction to obtain the SMILES sequence feature h seq , the dimension is c×d1, where c represents the length of the SMILES sequence; S32: The atomic feature vector h of the drug molecule atom As the Query in the attention mechanism, the sequence feature h seq As the Key and Value in the attention mechanism, the atomic feature vector H after interaction is calculated using the attention mechanism. atom ; S33: The chemical bond feature vector h of the drug molecule bond As the Query in the attention mechanism, the sequence feature h seq As the Key and Value in the attention mechanism, the atomic feature vector H after interaction is calculated according to the attention mechanism. bond ; S34: H atom and H bond After splicing, the final feature vector H of the drug molecule is obtained d .

8. The drug target matching recommendation method based on multi-view graph neural network according to claim 1, characterized in that: The target sequence feature vector Hp is obtained according to the target amino acid sequence, and the feature vector of the drug molecule and the target sequence feature vector are interacted to obtain the matching probability of the drug target, including S41: Divide the target amino acid sequence into multiple overlapping ternary amino acid subsequences, assign a randomly initialized feature vector with dimension d1 to each ternary amino acid subsequence, and extract features through a one-dimensional gated convolutional neural network to obtain the target sequence feature vector H p , the dimension is b×d1, where b represents the length of the ternary amino acid subsequence; S42: The characteristic vector H of the drug molecule d As the Query in the attention mechanism, the target sequence feature H p As Key and Value, the interactive feature vector is calculated according to the attention mechanism. S43: The characteristic vector H of the drug molecule p As the Query in the attention mechanism, the target sequence feature H d As Key and Value, the interactive feature vector is calculated according to the attention mechanism. S44: Transform the eigenvector by linear transformation and Map them to the same dimension d, concatenate the two and send them into the binary classification linear layer to obtain the probability of drug and target matching.

Images