Intelligent drug target recommendation method based on comparative learning

Abstract

The invention belongs to computational biology, and particularly relates to a drug target intelligent recommendation method based on comparative learning. The method comprises the following steps: acquiring drug and target information, including an SMILES sequence of a drug and an amino acid sequence of a target protein; using a drug encoder and a target protein encoder to obtain characterization information of the drug and the target protein; a self-adaptive attention mechanism is designed, interaction between a drug target characterization vector and a learnable query vector is achieved through different mask methods, and a drug characterization space and a target protein characterization space are aligned through comparative learning; the matching probability between the drug targets is obtained according to the query vector obtained after interaction; the drug targets with the probability larger than a threshold value are selected and recommended to the user; the prediction result is high in accuracy, the time for mining drug target reaction pairs by biomedical researchers is shortened, and the research burden is relieved.

Description

Technical Field

[0001] The present invention belongs to the field of computational biology, and specifically designs a drug target intelligent recommendation method based on contrastive learning. Background Art

[0002] Drug-target interaction prediction aims to reveal the potential interactions between drugs and target proteins, which is a key step in drug development. This prediction technology focuses on the interaction between specific proteins and drug compounds, providing an important basis for drug repositioning, multi-drug treatment strategy design, drug resistance analysis, and side effect evaluation. Although traditional in vitro experiments can provide accurate biomedical data, their high cost, lengthy cycle, and inefficiency in processing large-scale data sets limit their scope of application. Therefore, with the rapid advancement of bioinformatics, computational biology, and drug development technology, a large amount of biological experimental data has gradually been accumulated in bioinformatics databases, and the focus of drug research is gradually shifting from traditional laboratory experiments to computational methods.

[0003] At present, drug-target interaction prediction methods are mainly divided into two categories: similarity-based methods and feature-based methods. Feature-based methods include two key modules: feature extraction module and downstream prediction module. The feature extraction module maps the target protein and drug molecule to a discriminative feature space. However, existing methods often ignore the problem that the original sequences of drug molecules and target proteins are different, resulting in inconsistent feature space distribution after feature extraction. At the same time, simply merging the feature vectors of the two to predict interactions may affect the accuracy of the model. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a drug target intelligent recommendation method based on contrastive learning, 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 the SMILES sequence of the drug and the amino acid sequence of the target protein;

[0008] S2: Convert the drug's SMILES sequence into a molecular graph structure, and set an initial feature vector for it according to the chemical properties of each atom and chemical bond in the molecular graph; split the amino acid sequence of the target protein into overlapping ternary amino acid subsequences, and set a random vector as the initial feature vector for each subsequence;

[0009] S3: Extract the drug representation vector h using the drug encoder based on graph neural network d , using the sequence-based target protein encoder to extract the target protein representation vector h p ;

[0010] S4: Design a randomly initialized query vector h q , multi-layer adaptive attention module, to achieve query vector h q With drug characterization vector h d and target protein characterization vector h p Interact with each other to obtain the vector H that interacts with the query vector and the drug representation d , a vector H that interacts the query vector and the target protein representation p and the query vector H that interacts with the drug representation and target protein representation q ;

[0011] S5: Take the matched drug-target pairs as positive samples and use the interaction feature H p and H d Compute contrastive learning loss to align the representation spaces of drugs and target proteins, and add perturbations to the feature vectors of each atom in the drug molecule graph to improve the generalization ability of drug representation;

[0012] S6: Query vector H q The matching probability of the drug target is output through the linear layer, and the matching loss is calculated according to the known drug target matching labels. It is added to the contrastive learning loss as the final loss to update the model parameters and obtain a trained drug target recommendation model.

[0013] Beneficial effects of the present invention:

[0014] The present invention uses contrastive learning to calculate the contrastive loss between two separately trained drug target encoders, so that the feature vectors of drug target pairs that can react are closer, thereby aligning the feature space of drug molecules and target proteins, ignoring the noise and differences specific to individual data, retaining the interactive information that is only beneficial to interaction prediction, and further improving the accuracy and generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The figure is an overall flow chart of a drug target intelligent recommendation method based on contrastive learning of the present invention. DETAILED DESCRIPTION

[0016] 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.

[0017] Embodiment 1: In this embodiment, the present invention provides a drug target intelligent recommendation method based on contrastive learning, such as Figure 1 As shown, including:

[0018] Obtain the SMILES sequence of the drug to be predicted and the amino acid sequence of the target protein, input them 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;

[0019] The training process of the drug target recommendation model includes:

[0020] S1: Obtain drug target information, including the SMILES sequence of the drug and the amino acid sequence of the target protein;

[0021] S2: Convert the drug's SMILES sequence into a molecular graph structure, and set an initial feature vector for it according to the chemical properties of each atom and chemical bond in the molecular graph; split the amino acid sequence of the target protein into overlapping ternary amino acid subsequences, and set a random vector as the initial feature vector for each subsequence;

[0022] S3: Extract the drug representation vector h using the drug encoder based on graph neural network d , using the sequence-based target protein encoder to extract the target protein representation vector h p ;

[0023] S4: Design a learnable query vector h q , controls the mask matrix of the attention mechanism, and realizes the query vector h q With drug characterization vector h d and target protein characterization vector h p Interact with each other to obtain the vector H that interacts with the query vector and the drug representation d , which interacts the query vector and the target protein representation vector H p and the query vector H that interacts with the drug representation and target protein representation q .

[0024] S5: Take the matched drug-target pairs as positive samples and use the interaction feature H p and H dCompute contrastive learning loss to align the representation spaces of drugs and target proteins, and add perturbations to the feature vectors of each atom in the drug molecule graph to improve the generalization ability of drug representation;

[0025] S6: Query vector H q The matching probability of the drug target is output through the linear layer, and the matching loss is calculated according to the known drug target matching labels. The matching loss is added to the contrast loss as the final loss to update the model parameters and obtain a trained drug target recommendation model.

[0026] The drug representation vector h is extracted using a drug encoder based on a graph neural network. d , using the sequence-based target protein encoder to extract the target protein representation vector h p ,include:

[0027] Drug representation vector h d The calculation formula is:

[0028] h d =concat({GnnEncoder(v):v∈D})

[0029] Among them, the calculation formula of the drug encoder GnnEncoder based on graph neural network is:

[0030]

[0031] Where D represents the set of atoms in the drug molecule, h 0 (u) represents the initial eigenvector of atom u according to its chemical properties, N(v) represents the set of atoms adjacent to atom v, and W d represents the first trainable parameter matrix, concat() represents feature vector concatenation, and drug characterization vector h d The dimension is n×d, where n represents the number of atoms in the drug molecule and d represents the characteristic dimension of each atom;

[0032] Target protein characterization vector h p The calculation formula is:

[0033] h p =GnnEncoder(X)

[0034] The calculation formula of GnnEncoder is:

[0035]

[0036] in, are the second and third trainable parameter matrices, b represents the trainable bias parameter, σ represents the Sigmoid function, X represents the feature vector of the triplet subsequence in the initial amino acid sequence, and the target protein vector h p The dimension is l×d, where l represents the length of the triplet subsequence in the amino acid sequence of the target protein. Design a learnable query vector h q , controls the mask matrix of the attention mechanism, and realizes the query vector h q With drug characterization vector h d and target protein characterization vector h p The distribution interacts to obtain the vector H that interacts the query vector and the drug representation d , which interacts the query vector and the target protein representation vector H p and the query vector H that interacts with the drug representation and target protein representation q ,include:

[0037] S31: Set the randomly initialized query vector h q , with dimensions of n×d.

[0038] S32: drug characterization vector h d With the query vector h q The input mask strategy after splicing is In the multi-layer adaptive attention module, the interactive feature vector H is obtained d .

[0039] The input of the first layer adaptive attention module is: The dimension is (n+m)×d, and || represents the concatenation operation.

[0040] In this process, the calculation formula of each layer of the adaptive attention module is as follows:

[0041]

[0042] Among them, W0, W q , W k , W v are the fourth, fifth, sixth, and seventh trainable parameter matrices, with dimensions of d×d, Represents the feature vector of the drug molecule in the tth layer. At the last layer T, Where T=6.

[0043] S33: Target protein representation vector h p With the query vector h q The input mask strategy after splicing is In the multi-layer adaptive attention module, the interactive feature vector H is obtained p .

[0044] The input of the first layer adaptive attention module is: The dimension is (m+l)×d, and || represents the concatenation operation.

[0045] In this process, the calculation formula of each layer of the adaptive attention module is as follows:

[0046]

[0047] Represents the feature vector of the target protein in the tth layer. At the last layer T,

[0048]

[0049] S34: The query vector is respectively combined with the drug representation vector h d and target protein characterization vector h p After splicing, the input mask strategy is M cross In the multi-layer adaptive attention module, the feature vector and Take out the part belonging to the query vector to obtain the query vector H that interacts with the drug target representation information q .

[0050] In the feature vector During the calculation process of , the input of the first layer adaptive attention module is The dimension is (n+m)×d, and || represents the concatenation operation.

[0051] The calculation formula for each layer of the adaptive attention module is:

[0052]

[0053] in Represents the drug molecule feature vector that interacts with the query vector at the tth layer. The first layer is At the last layer T,

[0054] In the feature vector During the calculation process of , the input of the first layer adaptive attention module is The dimension is (m+l)×d, and || represents the concatenation operation.

[0055] The calculation formula for each layer of the adaptive attention module is:

[0056]

[0057] in Represents the target protein feature vector that interacts with the query vector at the tth layer. At the last layer T,

[0058] Mask and M cross The representation is as follows:

[0059]

[0060] The i, j distribution represents the element in the i-th row and j-th column of the mask matrix. When the input data is the drug characterization vector h d and the query vector h q After concatenation, M self This allows h to be used during attention calculation. d and h q Only calculate the correlation with itself, while M cross It can make h d and h q The correlation can be calculated to obtain the interaction information between them.

[0061] Get feature vector and The representation information of the query vector in is:

[0062]

[0063] in, express The eigenvector from the nth position to the n+mth position of the vector, express The feature vector from the first position to the mth position of the vector, α is a trainable parameter used to control and The importance of.

[0064] Using the interaction feature H p and H d Compute contrastive learning loss to align the representation spaces of drugs and target proteins, including:

[0065] S61: For each pair of input drug molecule d and target protein p, treat it as a positive sample pair;

[0066] S62: For each drug molecule d, randomly sample k component sets from all target protein data For each target protein p, k component sets are randomly sampled from all drug molecule data The calculation formula of contrast loss is:

[0067]

[0068] in Indicates the number of samples contained in a batch during model training, H d (d) represents the characterization vector of drug d, H p (p) represents the characterization vector of drug p, τ is the temperature coefficient, which controls the difference between positive and negative samples, and β is a trainable parameter ranging from 0 to 1, which is used to balance the importance of drug contrast loss and target protein contrast loss.

[0069] Add perturbations to the feature vectors of each atom in the drug molecule graph to improve the generalization ability of drug representation, including:

[0070] S71: Drug characterization vector h using all drug molecule data d Get the mean The dimension is n×d;

[0071] S72: Perturb each drug molecule d to obtain a characterization vector h corresponding to the drug molecule d d (d) Similar, but with different fine-grained feature vectors The formula is:

[0072]

[0073] Where ‖*‖2 means finding the square of a vector;

[0074] S73: In a training batch, the drug molecule d representation vector and the perturbed representation vector are used as positive samples, and other drug molecules are used as negative samples to calculate the contrast loss. The loss function formula is:

[0075]

[0076] in Represents the set of drug molecules within a batch during training.

[0077] Query vector H q The matching probability of the drug target is output through the linear layer, and the calculation formula is:

[0078]

[0079] Where σ represents the sigmoid activation function, W out are the trainable parameters in the linear layer, represents the probability of drug-target match.

[0080] The matching loss is calculated based on the known drug target matching labels, and the formula is:

[0081]

[0082] Where N represents the number of samples, y i represents the label of the i-th sample, It indicates that the model outputs the matching probability of the drug target for the i-th sample.

[0083] The final loss calculation formula is:

[0084]

[0085] 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.

[0086] 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 intelligent recommendation method based on contrastive learning, 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 the SMILES sequence of the drug and the amino acid sequence of the target protein; S2: Convert the drug's SMILES sequence into a molecular graph structure, and set an initial feature vector for it according to the chemical properties of each atom and chemical bond in the molecular graph; split the amino acid sequence of the target protein into overlapping ternary amino acid subsequences, and set a random vector as the initial feature vector for each subsequence; S3: Extract the drug representation vector h using the drug encoder based on graph neural network d , using the sequence-based target protein encoder to extract the target protein representation vector h p ; S4: Design a randomly initialized query vector h q , multi-layer adaptive attention module, to achieve query vector h q With drug characterization vector h d and target protein characterization vector h p Interact with each other to obtain the vector H that interacts with the query vector and the drug representation d , a vector H that interacts the query vector and the target protein representation p and the query vector H that interacts with the drug representation and target protein representation q ; S5: Take the matched drug-target pairs as positive samples and use the interaction feature H p and H d Compute contrastive learning loss to align the representation spaces of drugs and target proteins, and add perturbations to the feature vectors of each atom in the drug molecule graph to improve the generalization ability of drug representation; S6: Query vector H q The matching probability of the drug target is output through the linear layer, and the matching loss is calculated according to the known drug target matching labels. It is added to the contrastive learning loss as the final loss to update the model parameters and obtain a trained drug target recommendation model.

2. The drug target intelligent recommendation method based on contrastive learning according to claim 1, characterized in that: The drug representation vector h is extracted using a drug encoder based on a graph neural network. d , using the sequence-based target protein encoder to extract the target protein representation vector h p ,include: Drug representation vector h d The calculation formula is: h d =concat({CnnEncoder(v):v∈D}) Among them, the calculation formula of the drug encoder GnnEncoder based on graph neural network is: Where D represents the set of atoms in the drug molecule, h 0 (u) represents the initial eigenvector of atom u according to its chemical properties, N(v) represents the set of atoms adjacent to atom v, and W d represents the first trainable parameter matrix, concat() represents feature vector concatenation, and drug characterization vector h d The dimension is n×d, where n represents the number of atoms in the drug molecule and d represents the characteristic dimension of each atom; Target protein characterization vector h p The calculation formula is: h p =GnnEncoder(X) The calculation formula of GnnEncoder is: in, are the second and third trainable parameter matrices, b represents the trainable bias parameter, σ represents the Sigmoid function, X represents the feature vector of the triplet subsequence in the initial amino acid sequence, and the target protein vector h p The dimension is l×d, where l represents the length of the triplet subsequence in the amino acid sequence of the target protein.

3. The drug target intelligent recommendation method based on contrastive learning according to claim 1, characterized in that: Design a randomly initialized query vector h q , multi-layer adaptive attention module, to achieve query vector h q With drug characterization vector h d and target protein characterization vector h p Interact with each other to obtain the vector H that interacts with the query vector and the drug representation d , a vector H that interacts the query vector and the target protein representation p and the query vector H that interacts with the drug representation and target protein representation q ,include: S41: Set the randomly initialized query vector h q , the dimension is m×d, m represents the query vector h q The characteristic length of h q The characteristic dimension of S42: drug characterization vector h d With the query vector h q The input mask strategy after splicing is In the multi-layer adaptive attention module, the interactive feature vector H is obtained d ; The input of the first layer adaptive attention module is: The dimension is (n+m)×d, where || represents the splicing operation, and n represents the number of atoms in the drug molecule; S43: Target protein representation vector h p With the query vector h q The input mask strategy after splicing is In the multi-layer adaptive attention module, the interactive feature vector H is obtained p ; The input of the first layer adaptive attention module is: The dimension is (m+l)×d, and || represents the concatenation operation; S44: query vector h q and drug characterization vector h d and target protein characterization vector h p After splicing, the input mask strategy is M cross In the multi-layer adaptive attention module, the feature vector and Take out the part belonging to the query vector to obtain the query vector H that interacts with the drug target representation information q .

4. The drug target intelligent recommendation method based on contrastive learning according to claim 3, characterized in that: The mask strategy Mask strategy Mask strategy M cross ,include: Among them, i, j distribution represents the element in the i-th row and j-th column of the mask matrix; modify the mask in the attention calculation to achieve different data interaction modes: when the input data is the drug representation vector h d and query vector h q After concatenation, N self This allows g to be used during attention calculation. d and g q Only calculate the correlation with itself, while M cross It can make h d and h q The correlation can be calculated to obtain the interaction information between them.

5. The drug target intelligent recommendation method based on contrastive learning according to claim 3, characterized in that: From the feature vector and The part belonging to the query vector is taken out from the query vector to obtain the query vector H that interacts with the drug target representation information. q ,include: in, express The eigenvector from the nth position to the n+mth position of the vector, express The feature vector from the first position to the mth position of the vector, α represents the control and A trainable parameter of importance.

6. The drug target intelligent recommendation method based on contrastive learning according to claim 1, characterized in that: The matched drug-target pairs are taken as positive samples and the interaction feature H is used to identify p and H d Compute contrastive learning loss to align the representation spaces of drugs and target proteins, including: S5-11: For each pair of input drug molecule d and target protein p, treat it as a positive sample pair; S5-12: For each drug molecule d, randomly sample k component sets from all target protein data For each target protein p, k component sets are randomly sampled from all drug molecule data S5-13: At the Gathering In the above example, we use the interaction feature H p and H d Compute contrastive learning loss to align the representation spaces of drugs and target proteins.

7. The drug target intelligent recommendation method based on contrastive learning according to claim 6, characterized in that: In the collection In the above example, we use the interaction feature H p and H d Calculate the contrastive learning loss. Contrastive loss includes: in, represents the contrast loss of the positive sample pair, Indicates the number of samples contained in a batch during model training, H d (d) represents the characterization vector of drug d, H p (p) represents the characterization vector of the target protein p, τ represents the temperature coefficient that controls the difference between positive and negative samples, β represents a trainable parameter used to balance the importance of drug contrast loss and target protein contrast loss, and its value range is between 0 and 1, exp represents the exponential function, and H p (s) represents the characterization vector of target protein s, H d (t) represents the characterization vector of drug t, and log represents the logarithmic function.

8. The drug target intelligent recommendation method based on contrastive learning according to claim 1, characterized in that: Add perturbations to the feature vectors of each atom in the drug molecule graph to improve the generalization ability of drug representation, including: S5-21: Drug characterization vector h using all drug molecule data d Get the mean The dimension is n×d; S5-22: Perturb each drug molecule d to obtain a characterization vector h corresponding to drug molecule d d (d) Similar but fine-grained feature vectors S5-23: In a training batch, the drug molecule d representation vector and the perturbed feature vector As a positive sample, the contrast loss is calculated with other drug molecules as negative samples to enhance the generalization ability of drug molecule feature representation.

9. The drug target intelligent recommendation method based on contrastive learning according to claim 8, characterized in that: Perturb each drug molecule d to obtain the characterization vector h of drug molecule d d (d) Similar but fine-grained feature vectors include: Among them, h d (d) represents the characterization vector of drug d, ‖*‖2 represents the square of the vector, ε represents a trainable parameter used to control the degree of perturbation, ranging from 0 to 1, Represents the mean of the drug characterization vectors for all drug molecule data.

10. The drug target intelligent recommendation method based on contrastive learning according to claim 8, characterized in that: In a training batch, the drug molecule d representation vector and the perturbed feature vector As a positive sample, the contrast loss is calculated with other drug molecules as negative samples, including: in, represents the contrast loss of negative samples, represents the set of drug molecules in a batch during the training process, β represents a trainable parameter used to compare the importance of drug molecules, and its value range is between 0 and 1, exp represents the exponential function, log represents the logarithmic function, and h d (d) represents the characterization vector of drug d, τ represents the temperature coefficient that controls the difference between positive and negative samples, Indicates drug The perturbed eigenvector.

Images