Traditional Chinese medicine multi-target interaction prediction method based on Transform architecture

Abstract

The invention discloses a traditional Chinese medicine multi-target interaction prediction method based on a Transform architecture. The method comprises the following steps: creating an interaction label of a medicine and a target; respectively carrying out feature extraction on the traditional Chinese medicine component data and the target protein data; fusing the traditional Chinese medicine component feature vector and the target feature vector through a cross attention mechanism; creating a data set by using the feature vector F and an interaction label of the drug and the target; the method comprises the following steps of: constructing and training a Transform encoder; inputting a to-be-detected medicine-target data set into the trained Transformer encoder, outputting a probability value of interaction between each target and the traditional Chinese medicine component, and when the probability value of interaction between the traditional Chinese medicine component and the target is greater than a preset threshold value, determining that interaction exists; the invention provides an efficient and accurate traditional Chinese medicine multi-target interaction prediction means, and is beneficial to acceleration of traditional Chinese medicine research and new medicine development processes.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence-assisted drug-target interaction research, and in particular, to a method for predicting multi-target interactions of traditional Chinese medicine based on a Transformer architecture. Background Art

[0002] Traditional Chinese medicine (TCM) is complex due to its multi-component and multi-target mechanism of action. Predicting the interactions between TCM components and targets is a critical task. TCM is composed of a variety of chemical components, which may have synergistic or antagonistic effects. The targets involve different biological pathways and systems, showing highly nonlinear characteristics of multiple targets and pathways. As a result, existing models are difficult to accurately capture complex relationships, making the prediction of multi-target interactions in TCM challenging. Traditional prediction methods usually rely on biological experiments and limited computational models. At the same time, due to the large amount of complex information between TCM components, drug-target predictions are uncertain. Therefore, an efficient and accurate method is urgently needed to solve these problems.

[0003] With the development of deep learning technology, especially the successful application of Transformer architecture, it has become possible to predict the interaction between drugs and targets through the combination of modules such as multi-head attention mechanism and feedforward neural network. However, the existing models have problems in how to effectively extract the characteristics of Chinese medicine components and targets and how to accurately predict the interaction of multiple targets when dealing with the prediction of multi-target interactions of traditional Chinese medicine. Summary of the invention

[0004] In order to effectively extract the characteristics of complex components and biological targets of traditional Chinese medicine and comprehensively capture the interactions between multiple components and multiple targets, the present invention provides a traditional Chinese medicine multi-target interaction prediction method based on Transformer architecture.

[0005] The technical solution adopted by the present invention is: the present invention provides a method for predicting multi-target interactions of traditional Chinese medicine based on the Transformer architecture, the method comprising the following steps:

[0006] Step S1, obtaining Chinese medicine component data, target protein data and drug-target data, and creating interactive labels between drugs and targets;

[0007] Step S2, performing feature extraction on the Chinese medicine component data and the target protein data respectively to obtain a Chinese medicine component feature vector and a target feature vector;

[0008] Step S3, fusing the Chinese medicine component feature vector and the target feature vector through a cross attention mechanism to obtain a feature vector F;

[0009] Step S4, creating a data set using the feature vector F and the interaction labels between the drug and the target, and dividing it into a training set, a test set, and a validation set;

[0010] Step S5, constructing a Transformer encoder, inputting the training set into the Transformer encoder, and the Transformer encoder performs deep representation learning on the input sequence and generates a probability value of the interaction between each target and the traditional Chinese medicine component;

[0011] Step S6, using binary cross entropy loss to train the Transformer encoder, and evaluating its performance on the validation set, and finally evaluating its generalization ability on the test set;

[0012] Step S7, input the drug-target data set to be tested into the trained Transformer encoder, and output the probability value of each target interacting with the Chinese medicine component, wherein when the probability value of the interaction between the Chinese medicine component and the target is greater than a predetermined threshold, it is determined that an interaction exists.

[0013] Further, step S1, obtaining Chinese medicine component data, target protein data and drug-target data, and creating interactive labels between drugs and targets, the specific steps are as follows:

[0014] The TCMID dataset was used to obtain the chemical information of Chinese medicine ingredients, including molecular formula, molecular weight, structural description and the corresponding simplified molecular linear input specification SMILES. The UniProt dataset was used to obtain the target protein sequence, structure and function information. The BindingDB dataset was used to obtain drug-target data. An interaction label of drug and target was created based on the Chinese medicine ingredient data, target protein data and drug-target data.

[0015] Furthermore, in step S2, feature extraction is performed on the Chinese medicine ingredient data, specifically: the simplified molecular linear input specification SMILES is used to extract graph embedding vectors as feature vectors of the Chinese medicine ingredients using the graph neural network GNN.

[0016] Furthermore, in step S2, feature extraction is performed on the target protein data, specifically: One-Hot Encoding is used to extract the amino acid sequence features of the target protein, and then the amino acid sequence features of the target protein are input into the sequence embedding model ProtBERT, and the embedding matrix obtained from the sequence embedding model ProtBERT is directly connected and merged with the amino acid sequence to obtain a target feature vector.

[0017] Furthermore, in step S3, the feature vector of the Chinese medicine component and the target feature vector are fused through a cross-attention mechanism to obtain a feature vector F. The specific steps are as follows:

[0018] The query matrix Q is obtained by linearly transforming the feature vector of traditional Chinese medicine ingredients drug , the target feature vector is transformed into the key matrix K by linear transformation target Sum value matrix V target :

[0019] Q drug =X drug W Q

[0020] K target =X target W K

[0021] V target =X target W V

[0022] Among them, W Q is the query weight matrix, W K is the key weight matrix, W V is the value weight matrix with dimension (d,d), X drug is the characteristic vector of Chinese medicine ingredients, X target is the target feature vector,

[0023] Calculate the attention score, and then apply the Softmax function to the attention score to get the attention weight:

[0024]

[0025] in, is the formula for calculating the attention score, is a scaling factor used to stabilize the gradient;

[0026] According to the attention weight, the value matrix V target Take a weighted average:

[0027] Weighted_Values ​​= Attention weights ·V target

[0028] Generate the feature vector F:

[0029] F = Weighted_Values ​​+ X drug .

[0030] Furthermore, the Transformer encoder is composed of a plurality of encoder layers stacked together, each layer of the encoder includes a positional encoding, a multi-head attention mechanism, a feedforward neural network, an activation function, a residual and layer normalization, and a binary classification output layer. The Transformer encoder performs deep representation learning on the input sequence and generates a probability value of the interaction between each target and the Chinese medicine component. The specific steps are as follows:

[0031] The feature vector F is concatenated in dimension after position encoding:

[0032] Input i =[F i PE i ]

[0033] Among them, the eigenvector F is composed of F i Composition, PE is the position encoding, i is the dimension index, Input i Constitute the input vector Input,

[0034] The input vector Input is transformed linearly to generate the query matrix Q', key matrix K' and value matrix V':

[0035] Q'=InputW Q’

[0036] K'=InputW K‘

[0037] V'=InputW V‘

[0038] Among them, W Q‘ is the weight matrix learned from the query matrix, W K’ The weight matrix learned by the key matrix, W V’ is the weight matrix obtained by learning the value matrix,

[0039] Calculate the attention weights:

[0040]

[0041] Among them, d k is the dimension of the key vector, and then the softmax function is applied to get the attention weight:

[0042] attention_weights=softmax(scores)

[0043] Use the attention weights to perform weighted summation on the value matrix V' to get the output of the jth attention head, denoted as head j, the multi-head attention mechanism calculates h independent attention heads in parallel, concatenates the outputs of all attention heads, and then transforms them linearly:

[0044] multi_head_output=concat(head1,head2,…,head h )W O

[0045] Among them, W O is the output weight matrix, j = 1, 2, 3…h,

[0046] Perform residual connection and layer normalization (Add&Norm) on the output of the multi-head attention mechanism:

[0047] Residual Connection:

[0048] Output res-MA =multi_head_output+Input

[0049] Layer Normalization:

[0050]

[0051] where μ is the feature mean, σ is the feature standard deviation, and ∈ is a small constant to avoid division by zero.

[0052] The output of the multi-head attention mechanism is residually connected and layer normalized before entering the feedforward neural network:

[0053] FFN(Norm MA )=ReLU(Norm MA ·W1+b1)W2+b2

[0054] Among them, W1 is the weight matrix of the first feedforward layer, W2 is the weight matrix of the second feedforward layer, b1 is the bias term of the first layer, and b2 is the bias term of the second layer.

[0055] Perform residual connection and layer normalization (Add&Norm) on the output of the feedforward neural network:

[0056] Output final =Norm(FFN(Norm(Output res-MA ))+Norm(Output res-MA ))

[0057] The fused feature vector is passed into the binary classification output layer to generate a probability value for prediction:

[0058] p=sigmoid(Output final ·W+b)

[0059] Among them, W and b are the weight and bias of the fully connected layer, sigmoid is the activation function, and p represents the probability value of the interaction between the traditional Chinese medicine ingredients and the target protein.

[0060] Furthermore, the loss function for training the Transformer encoder is:

[0061]

[0062] Where N is the number of samples, y i is the true label, is the predicted probability.

[0063] The present invention has the following beneficial effects: the powerful feature extraction and parallel processing capabilities of the Transformer architecture are utilized to improve the training efficiency of the model. High-quality feature vectors are extracted through pre-trained chemical language models and protein language models, thereby improving the accuracy of predictions. The interaction between multiple targets and Chinese medicine ingredients can be predicted simultaneously, and is suitable for the research needs of multi-target interactions of Chinese medicine. The model structure is flexible, and parameters and structures can be adjusted according to specific needs, and it is suitable for different Chinese medicine ingredients and target data. The present invention provides an efficient and accurate method for predicting multi-target interactions of Chinese medicine, which helps to accelerate the process of Chinese medicine research and new drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a method for predicting multi-target interactions of traditional Chinese medicine based on the Transformer architecture of the present invention.

[0065] Figure 2 It is a flow chart of the present invention.

[0066] Figure 3 This is a diagram of the Transformer encoder architecture of the present invention.

[0067] Figure 4 This is the architecture diagram of the multi-head attention mechanism of the present invention. DETAILED DESCRIPTION

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following examples are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0069] The present invention provides a method for predicting multi-target interactions of traditional Chinese medicine based on a Transformer architecture, the method comprising the following steps:

[0070] Step S1, obtaining Chinese medicine component data, target protein data, and drug-target data, and creating interactive tags between drugs and targets;

[0071] Specifically, this embodiment uses TCMID, UniProt and BindingDB data sets as the benchmark usage model; TCMID obtains chemical information of Chinese medicine ingredients, such as molecular formula, molecular weight and structural description, and the corresponding simplified molecular linear input specification (SMILES); UniProt obtains target protein sequence, structure and function information; BindingDB obtains drug-target data, and BindingDB is used to construct drug-target interaction tags;

[0072] To build the training data for the model, create a data table containing drug and target information:

[0073] Drug data includes drug ID (uniquely identifies each Chinese medicine ingredient), drug name, chemical molecule information (SMILES) (representation of chemical molecules for subsequent feature extraction);

[0074] Target data includes target ID (uniquely identifies each target protein), target name, protein sequence (amino acid sequence of protein, used for subsequent feature extraction), and functional information (describing the biological function of the target protein);

[0075] Based on the TCM component data, target protein data and drug-target data, a drug-target interaction matrix is ​​created, i.e., the drug-target interaction label, where the rows represent drugs, the columns represent targets, and the values ​​in the matrix represent the interaction relationship between drugs and targets, such as agonism, antagonism, inhibition, promotion, ligand, induction, non-competitive inhibition, competitive inhibition, complex formation, etc.

[0076] Definition and construction of drug-target interaction signatures:

[0077] Construct a two-dimensional matrix M with a dimension of N×T, where N represents the number of drugs, T represents the number of targets, and the element M[i][j] in the matrix is ​​used to indicate whether there is an interaction between drug i and target j, or the strength of the interaction;

[0078] Determine whether there is a known interaction between each drug and target based on the BindingDB database,

[0079] If there is an interaction between drug i and target j, the value of M[i][j] is marked as 1, otherwise it is marked as 0. This binary label can be used to describe the drug-target interaction relationship.

[0080] Step S2, performing feature extraction on the Chinese medicine component data and the target protein data respectively to obtain a Chinese medicine component feature vector and a target feature vector;

[0081] Specifically: feature extraction of traditional Chinese medicine ingredient data:

[0082] The SMILES of Chinese medicine ingredients are used to extract graph embedding vectors as the feature vectors of Chinese medicine ingredients using graph neural networks (GNNs), where atoms are nodes of the graph and chemical bonds are edges of the graph. Each node can also contain properties such as the type, valence, and electronegativity of the atom, and each edge can contain the type of bond (single bond, double bond, triple bond, aromatic bond, etc.);

[0083] Among them, the feature vector of the GNN convolution initialization node is encoded according to the attributes of the node, and the feature vector of the GNN convolution initialization edge is encoded according to the characteristics of the chemical bond; during the GNN convolution process, the feature vector of each node is updated by aggregating the information of neighboring nodes. The specific steps are as follows:

[0084] Each node receives messages from its neighboring nodes. Each node aggregates the messages transmitted by all neighboring nodes through a summation operation, combines the aggregated messages with the node's own features, and transforms them through a multi-layer perceptron MLP to obtain updated node features. This process usually performs multiple iterations according to the number of graph convolutional layers, so that information from more distant neighbors can be aggregated, and ultimately more global features of the compound can be captured. After multiple convolution operations, the feature representation of each node already includes the information of the surrounding neighboring nodes. The features of all nodes are globally pooled to obtain the embedding vector of the entire graph, which represents the chemical structure and properties of the entire Chinese medicine ingredient.

[0085] Feature extraction of target protein data:

[0086] One-Hot Encoding represents the amino acid sequence as a high-dimensional sparse matrix, which is used as the initial input for training in ProtBERT.

[0087] When ProtBERT processes an amino acid sequence, it generates an embedding vector for each amino acid to capture the contextual information of the amino acids in the sequence and the interaction between them. The embedding vector is max-pooled to generate a feature vector of fixed length, that is, to form an embedding matrix H, where H∈R L*D , L represents the length of the protein sequence (the number of amino acids), and D represents the dimension of the embedding matrix of each amino acid.

[0088] H=[h1,h2,…,h L ], where each h i ∈R d is a d-dimensional embedding vector representing the embedding of the i-th amino acid.

[0089] Initialize a zero vector, max pooling ∈R L*D Used to save the maximum value of all amino acid embedding vectors. For the embedding matrix H, the maximum value of all amino acid embeddings on each dimension is calculated dimension by dimension.

[0090] Specifically, for each dimension (in ), calculate the maximum value of all amino acid embeddings in this dimension: max pooling ∈R L*D

[0091] h i,j represents the value of the i-th amino acid embedding vector in the j-th dimension.

[0092] The result is a d-dimensional vector where the value of each dimension represents the maximum value of that dimension among all amino acid embeddings.

[0093] The result is a max pooling is a d-dimensional feature vector, representing the most significant feature of the entire protein sequence.

[0094] The embedding matrix obtained from ProtBERT is directly connected and merged with the amino acid sequence. The final feature vector contains the amino acid sequence information and three-dimensional structure information of the protein, that is, the target feature vector.

[0095] Step S3, fusing the Chinese medicine component feature vector and the target feature vector through a cross attention mechanism;

[0096] Specifically: Assume that the characteristic vector of Chinese medicine ingredients is X drug , whose dimension is (n drug ,d), where n drug is the number of characteristics of Chinese medicine ingredients, d is the characteristic dimension,

[0097] Assume that the target feature vector is X target , whose latitude is (n target ,d), where n target is the number of target features, d is the feature dimension,

[0098] The characteristic vector X of Chinese medicine ingredients drug and the target feature vector X target Perform linear transformations respectively to obtain the query matrix Q drug , key matrix K target Sum value matrix V target :

[0099] Q drug =X drug W Q

[0100] K target =X target W K

[0101] V target =X target W V

[0102] Among them, W Q is the query weight matrix, W K is the key weight matrix, W V is the value weight matrix, whose dimension is (d,d),

[0103] The query matrix is ​​dot-producted with each key matrix to calculate the attention score to reflect the correlation between the Chinese medicine ingredients and the target. Then the Softmax function is applied to the attention score to obtain the attention weight:

[0104]

[0105] in, is the step of calculating the attention score, is a scaling factor used to stabilize the gradient;

[0106] Use the attention weights to perform a weighted summation on the value matrix of the target protein:

[0107] Weighted_Values ​​= Attention weights ·V target

[0108] The fused target feature vector is merged with the TCM component feature vector, and the generated feature vector combines the feature information of TCM components and target proteins:

[0109] F = Weighted_Values ​​+ X drug

[0110] Make sure that the fused feature vector F has the shape , where N is the length of the fused sequence (the number of TCM ingredient features), d model is the fused feature dimension, which represents the feature representation of each drug-target pair.

[0111] Step S4, creating a dataset using the feature vector F and the drug-target interaction label, and dividing it into a training set (70%), a test set (15%), and a validation set (15%);

[0112] Step S5, construct a Transformer encoder, which is composed of multiple encoder layers stacked together, such as Figure 3 As shown, N=6, each layer of encoder includes position encoding, multi-head attention mechanism, feedforward neural network, activation function, residual and layer normalization, and binary classification output layer; the training set is input into the Transformer encoder. The Transformer encoder performs deep representation learning on the input sequence and generates the probability value of each target interacting with the Chinese medicine ingredient through the binary classification output layer;

[0113] Specifically: define the position encoding, generate a vector for each position pos, where the value of each dimension is determined by the sine and cosine functions, and define the formula to calculate the position encoding matrix:

[0114] Even dimensions (index 2i) are calculated using the sine function:

[0115]

[0116] Odd dimensions (index 2i) are calculated using the cosine function:

[0117]

[0118] Where pos is the position index (starting from 0), i is the dimension index in the positional encoding vector, and d is the feature dimension.

[0119] The position encoding PE and the feature vector F (given by F i The concatenated dimension will be the sum of the embedding dimension and the position encoding dimension.

[0120] Input i =[F i PE i ]

[0121] The input vector Input (Input is composed of Input i (Construction) Generate query matrix Q', key matrix K' and value matrix V' through linear transformation:

[0122] Q'=InputW Q’

[0123] K'=InputW K‘

[0124] V'=InputW V‘

[0125] Among them, W Q‘ is the weight matrix learned from the query matrix, W K’ The weight matrix learned by the key matrix, W V’It is the weight matrix learned from the value matrix, which is used to map the input vector Input to the representation space of query, key and value respectively.

[0126] Calculate the attention weight for each attention head:

[0127]

[0128] Among them, d k is the dimension of the key vector, and then the softmax function is applied:

[0129] attention_weights=softmax(scores)

[0130] Use the attention weights to perform a weighted summation of the value matrix V':

[0131] Output=attention weight ·V'

[0132] Perform weighted summation on the value matrix to obtain the output of the jth attention head, denoted as head j The multi-head attention mechanism calculates h independent attention heads in parallel and obtains h attention outputs, which are recorded as head1, head2, ..., head h ; Concatenate the outputs of all attention heads and then transform them linearly:

[0133] multi_head_output=concat(head1,head2,…,head h )W O

[0134] Among them, W O is the output weight matrix.

[0135] Perform residual connection and layer normalization (Add&Norm) on the output of the multi-head attention mechanism:

[0136] Residual Connection:

[0137] Output res-MA =multi_head_output+Input

[0138] Layer Normalization:

[0139]

[0140] Among them, μ is the characteristic mean, σ is the characteristic standard deviation, ∈ is a small constant,

[0141] The output of the multi-head attention mechanism is residually connected and layer normalized before entering the feedforward neural network, which consists of two fully connected layers and an activation function (such as ReLU):

[0142] FFN(Norm MA )=ReLU(Norm MA ·W1+b1)W2+b2

[0143] Among them, W1 is the weight matrix of the first feedforward layer, W2 is the weight matrix of the second feedforward layer, b1 is the bias term of the first layer, and b2 is the bias term of the second layer.

[0144] Perform residual connection and layer normalization (Add&Norm) on the output of the feedforward neural network:

[0145] Output final =Norm(FFN(Norm(Output res-MA ))+Norm(Output res-MA ))

[0146] The fused feature vector is passed into the binary classification output layer to generate a probability value for prediction:

[0147] p=sigmoid(Output final ·W+b)

[0148] Among them, W and b are the weight and bias of the fully connected layer, sigmoid is the activation function, and p represents the probability value of the interaction between the traditional Chinese medicine ingredients and the target protein.

[0149] The fused feature vector is passed to one or more fully connected layers (FC). The fully connected layer performs nonlinear mapping on the high-dimensional feature vector to generate a probability value for prediction. The output of the fully connected layer is:

[0150] p=sigmoid(Output final ·W+b)

[0151] Among them, W and b are the weight and bias of the fully connected layer, sigmoid is the activation function, and p represents the probability value of the interaction between the traditional Chinese medicine ingredients and the target protein.

[0152] Through the Sigmoid function, the score of each label is compressed into the range of [0,1], 1 means there is interaction, and 0 means there is no interaction. The Sigmoid function is expressed as:

[0153]

[0154] Furthermore, the full connection here outputs the probability value of the interaction between the traditional Chinese medicine ingredients and the target protein:

[0155]

[0156] The probability value P is used to determine whether there is an interaction between the drug and the target.

[0157] When p>predetermined threshold (set to 0.5 here), it is judged that there is interaction; otherwise, it is considered that there is no interaction.

[0158] In order to convert the probability value into a classification result (ie, “interaction exists” or “interaction does not exist”), a threshold t=0.5 needs to be set.

[0159] If the probability value p>t, it is considered that there is an interaction between the TCM ingredients and the target, and the model outputs 1 in the corresponding TCM ingredient-target interaction judgment.

[0160] If the probability value p≤t, it is considered that there is no interaction, and the model outputs 0 in the corresponding judgment of the interaction between the traditional Chinese medicine component and the target.

[0161] Step S6, use the binary cross entropy loss to train the Transformer encoder, evaluate the model performance on the validation set, and finally evaluate the generalization ability of the model on the test set;

[0162] Specifically:

[0163] Furthermore, the binary cross entropy loss is used to train the model, and the specific implementation steps are as follows:

[0164] The binary cross entropy loss is calculated for each label and averaged or summed over all labels and is defined as:

[0165]

[0166] Where N is the number of samples, y i is the true label of the i-th sample, which takes a value of 0 or 1. is the predicted probability.

[0167] Among drug-target pairs with annotated interactions:

[0168] For drug A and target B, if the experiment determines that there is an interaction, the true label y AB =1

[0169] For drug C and target D, if the experiment determines that there is no interaction, the true label y CD =0.

[0170] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

Claims

1. A method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture, characterized in that: The method comprises the following steps: Step S1, obtaining Chinese medicine component data, target protein data and drug-target data, and creating interactive labels between drugs and targets; Step S2, extracting features from the Chinese medicine component data and the target protein data respectively to obtain a Chinese medicine component feature vector and a target feature vector; Step S3, fusing the Chinese medicine component feature vector and the target feature vector through a cross-attention mechanism to obtain a feature vector F; Step S4, creating a data set using the feature vector F and the interaction labels between the drug and the target, and dividing it into a training set, a test set, and a validation set; Step S5, constructing a Transformer encoder, inputting the training set into the Transformer encoder, and the Transformer encoder performs deep representation learning on the input sequence and generates a probability value of the interaction between each target and the traditional Chinese medicine component; Step S6, using binary cross entropy loss to train the Transformer encoder, and evaluating its performance on the validation set, and finally evaluating its generalization ability on the test set; Step S7, input the drug-target data set to be tested into the trained Transformer encoder, and output the probability value of each target interacting with the Chinese medicine component, wherein when the probability value of the interaction between the Chinese medicine component and the target is greater than a predetermined threshold, it is determined that an interaction exists.

2. The method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture according to claim 1, characterized in that: Step S1, obtain Chinese medicine component data, target protein data and drug-target data, and create interactive labels between drugs and targets. The specific steps are as follows: The TCMID dataset was used to obtain the chemical information of Chinese medicine ingredients, including molecular formula, molecular weight, structural description and the corresponding simplified molecular linear input specification SMILES. The UniProt dataset was used to obtain the target protein sequence, structure and function information. The BindingDB dataset was used to obtain drug-target data. An interaction label of drug and target was created based on the Chinese medicine ingredient data, target protein data and drug-target data.

3. The method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture according to claim 2, characterized in that: In step S2, feature extraction is performed on the Chinese medicine ingredient data, specifically: the simplified molecular linear input specification SMILES is used to extract the graph embedding vector as the feature vector of the Chinese medicine ingredient using the graph neural network GNN.

4. The method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture according to claim 1, characterized in that: In step S2, feature extraction is performed on the target protein data, specifically: One-Hot Encoding is used to extract the amino acid sequence features of the target protein, and then the amino acid sequence features of the target protein are input into the sequence embedding model ProtBERT, and the embedding matrix obtained from the sequence embedding model ProtBERT is directly connected and merged with the amino acid sequence to obtain the target feature vector.

5. The method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture according to claim 1, characterized in that: Step S3, the Chinese medicine component feature vector and the target feature vector are fused through the cross attention mechanism to obtain the feature vector F. The specific steps are as follows: The query matrix Q is obtained by linearly transforming the feature vector of traditional Chinese medicine ingredients drug , the target feature vector is transformed into the key matrix K by linear transformation target Sum value matrix V target : Q drug =X drug W Q K target =X target W K V target =X target W V Among them, W Q is the query weight matrix, W K is the key weight matrix, W V is the value weight matrix with dimension (d, d), X drug is the characteristic vector of Chinese medicine ingredients, X targer is the target feature vector, Calculate the attention score, and then apply the Softmax function to the attention score to get the attention weight: in, is the formula for calculating the attention score, √f is the scaling factor used to stabilize the gradient; According to the attention weight, the value matrix V target Take a weighted average: Weighted_Values=Attention weights ·V target Generate the feature vector F: F=Weighted_Values+X drug 。 6. The method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture according to claim 1, characterized in that: The Transformer encoder is composed of multiple encoder layers stacked together. Each encoder layer includes position encoding, multi-head attention mechanism, feedforward neural network, activation function, residual and layer normalization, and binary classification output layer. The Transformer encoder performs deep representation learning on the input sequence and generates the probability value of each target interacting with the traditional Chinese medicine component. The specific steps are as follows: The feature vector F is concatenated in dimension after position encoding: Input i =[F i ;PE i ] Among them, the eigenvector F is composed of F i Composition, PE is the position encoding, i is the dimension index, Input i Constitute the input vector Input, The input vector Input is transformed linearly to generate the query matrix Q', key matrix K' and value matrix V': Q'=InputW Q’ K’=InputW K‘ V‘=InputW V‘ Among them, W Q‘ is the weight matrix learned from the query matrix, W K’ The weight matrix learned by the key matrix, W V’ is the weight matrix obtained by learning the value matrix, Calculate the attention weights: Among them, d k is the dimension of the key vector, and then the softmax function is applied to get the attention weight: attention_weights=softmax(scores) Use the attention weights to perform weighted summation on the value matrix V' to get the output of the jth attention head, denoted as head j , the multi-head attention mechanism calculates h independent attention heads in parallel, concatenates the outputs of all attention heads, and then transforms them linearly: multi_head_output=concat(head1,head2,...,head h )W O Among them, W O is the output weight matrix, j = 1, 2, 3…h, Perform residual connection and layer normalization on the output of the multi-head attention mechanism: Residual Connection: Output res-MA =multi_gead_output+Input Layer Normalization: Among them, μ is the characteristic mean, σ is the characteristic standard deviation, ∈ is a small constant, The output of the multi-head attention mechanism is residually connected and layer normalized before entering the feedforward neural network: <h2 style=";text-align:left;direction:ltr">FFN(Norm<h2 style=";text-align:left;direction:ltr"> MA <h2 style=";text-align:left;direction:ltr"> ReLU(Norm)<h2 style=";text-align:left;direction:ltr"> MA <h2 style=";text-align:left;direction:ltr"> W1+b1)W2+b2 Among them, W1 is the weight matrix of the first feedforward layer, W2 is the weight matrix of the second feedforward layer, b1 is the bias term of the first layer, and b2 is the bias term of the second layer. Perform residual connections and layer normalization on the output of the feedforward neural network: Output final =Norm(FFN(Norm(Output res-MA ))+Norm(Output res-MA )) The fused feature vector is passed into the binary classification output layer to generate a probability value for prediction: p=sigmoid(Output final ·W+b) Among them, W and b are the weight and bias of the fully connected layer, sigmoid is the activation function, and p represents the probability value of the interaction between the traditional Chinese medicine ingredients and the target protein.

7. The method for predicting multi-target interactions of traditional Chinese medicine based on Transformer architecture according to claim 1, characterized in that: The loss function for training the Transformer encoder is: Where N is the number of samples, y i is the true label, is the predicted probability.

Images