A drug-target interaction prediction method based on Mamba

Abstract

The present invention relates to a drug-target interaction prediction method based on Mamba, and belongs to the technical field of natural language processing. It comprises the following steps: S1. Acquire the molecular sequences of a drug and a target protein, process the molecular sequences of the drug and the target protein, and obtain the substructure sequences of the drug and the target protein; S2. Construct a drug-target interaction prediction model based on Mamba, the model comprising an embedding layer, a Mamba layer, a mixed attention mechanism layer, a conditional parameter convolution layer, and a fully connected layer; S3. The substructure sequences of the drug and the target protein are passed through the drug-target interaction prediction model based on Mamba to obtain the drug-target interaction prediction probability. The present invention improves the accuracy and efficiency of drug-target interaction prediction by combining mixed attention and conditional parameter convolution.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to a drug-target interaction prediction method based on Mamba. Background Art

[0002] Understanding and predicting the interactions between drugs and their target proteins is a key research area in the process of drug discovery and development. Traditional methods for predicting drug-target interactions usually rely on large-scale experimental screening, which is not only costly but also inefficient. With the rapid development of bioinformatics and computational biology, using computational methods to predict the interactions between drugs and proteins has become an effective alternative strategy. Most of the existing computational methods are based on static molecular structure simulations or simple statistical learning models, which often ignore the complexity of the dynamic changes between drugs and target proteins. In addition, when dealing with large-scale multi-dimensional data, these models often cannot effectively capture the subtle interactions between drugs and proteins, resulting in insufficient prediction accuracy.

[0003] Therefore, the present invention proposes a drug-target interaction prediction method based on Mamba to solve the above problems. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention provides a drug-target interaction prediction method based on Mamba.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0006] A drug-target interaction prediction method based on Mamba, comprising the following steps:

[0007] S1. Obtaining the molecular sequences of the drug and the target protein, processing the molecular sequences of the drug and the target protein, and obtaining the substructure sequences of the drug and the target protein;

[0008] S2. Construct a drug-target interaction prediction model based on Mamba, wherein the model includes an embedding layer, an encoding layer, a conditional parameter convolution layer, and a fully connected layer, wherein the encoding layer includes a Mamba layer and a hybrid attention mechanism layer;

[0009] S3. The substructure sequences of the drug and target protein are subjected to the Mamba-based drug-target interaction prediction model to obtain the drug-target interaction prediction probability.

[0010] Furthermore, the step S1 specifically includes:

[0011] The molecular sequence of the drug is converted into a drug sequence in SMILES format, the molecular sequence of the target protein is converted into an amino acid sequence, and then the drug sequence and amino acid sequence in SMILES format are decomposed into substructure sequences of the drug and the target protein by using a BPE algorithm.

[0012] Furthermore, in step S3, the substructure sequences of the drug and the target protein are embedded and feature extracted through the embedding layer and the Mamba layer to obtain the drug feature vector and the target protein feature vector:

[0013] Embedding processing: Each word in the substructure sequence of the drug and target protein is word-embedded and position-embedded to obtain a preliminary vector representation of word embedding and position embedding, and then the preliminary vector representation is linearly combined to obtain the embedding vector of the drug and target protein substructure sequence, which is expressed as follows:

[0014] ,

[0015] ,

[0016] ,

[0017] ,

[0018] in, represents the word embedding matrix, Indicates the first substructure sequence of the drug and target protein The index of words, is the position embedding matrix, is the position of the word in the substructure sequence of the drug and target protein, and Represent the preliminary vector representations of word embedding and position embedding respectively, and Represent the number of words in the substructure sequence of the drug and the target protein, respectively. and The embedding vectors representing the drug and target protein substructure sequences, respectively;

[0019] Feature extraction: The embedding vectors of the drug and target protein substructure sequences are subjected to feature extraction by the selective state space model SSM to obtain the feature vectors of the drug and target protein. The formula is as follows:

[0020] ,

[0021] ,

[0022] in, and are the feature vectors representing drugs and target proteins, respectively. represents the feature extraction operation on the embedding vector of the drug substructure sequence, Represents the feature extraction operation on the target protein substructure sequence embedding vector.

[0023] Furthermore, in step S3, the drug feature vector and the target protein feature vector are feature enhanced through a hybrid attention mechanism layer to obtain a fused drug feature vector and target protein feature vector:

[0024] The hybrid attention mechanism layer includes a hybrid attention layer generation vector unit, a bilinear transformation unit, a self-attention score and output vector unit, and a feature vector fusion unit;

[0025] Mixed attention layer generates vector unit: the feature vectors of the drug and target protein are generated by the mixed attention layer vector unit to generate a query vector , key vector Sum value vector , the formula is as follows:

[0026] ,

[0027] ,

[0028] ,

[0029] in, represents the weight matrix corresponding to the query vector, represents the weight matrix corresponding to the key vector, represents the weight matrix corresponding to the value vector, represents the feature vector, is the drug feature vector Or target protein feature vector ;

[0030] Bilinear Transformation Unit: The key vector Perform bilinear transformation to obtain the key vector after bilinear transformation , the formula is as follows:

[0031] ,

[0032] in, Represents the weight matrix of the bilinear transformation;

[0033] Self-attention score and output vector unit: the query vector and the key vector after bilinear transformation Perform a dot product calculation to obtain the attention score S, and then normalize the attention score by the scaling factor to obtain the attention weight , the attention weights and the value vector Perform the product to get the output vector The output vector is passed through the output linear layer and layer normalization to obtain the final output vector , the formula is as follows:

[0034] ,

[0035] ,

[0036] ,

[0037] ,

[0038] in, Indicates the dimensions of each header part, represents the key matrix after bilinear transformation The transpose of is the weight matrix of the output linear transformation, The normalization operation of the representation layer is performed. After the above process, the drug feature vector is obtained. The corresponding output vector And the final output vector , and similarly obtain the target protein feature vector The corresponding output vector And the final output vector ;

[0039] Feature vector fusion unit: drug feature vector and the final drug output vector Splice to get the fused drug feature vector , and similarly obtain the fusion target protein feature vector , the calculation formula is as follows:

[0040] ,

[0041] ,

[0042] in, Represents a concatenation operation.

[0043] Furthermore, in step S3, the fused drug feature vector and target protein feature vector are subjected to feature interaction through a conditional parameter convolution layer to obtain an output representation of the drug-target pair:

[0044] The fused drug feature vector and target protein feature vector are interactively processed through an interactive function to obtain an interactive graph, which is expressed as follows:

[0045] ,

[0046] in, represents the interaction function, Represents an interaction graph, which is a graph with a dimension of A multidimensional tensor of is represented as the length of the fused drug feature vector, Represented as the length of the fused target protein feature vector, Representative function The output size of

[0047] The interaction diagram The output representation of the drug-target pair is obtained by processing with conditional parameter convolution, and the formula is as follows:

[0048] ,

[0049] in, represents the conditional parameter convolution operation, Represents the output representation of the drug-target pair.

[0050] Furthermore, in step S3, the output representation of the drug-target pair is passed through a fully connected layer to predict the drug-target interaction, and the final drug-target interaction prediction probability is obtained by a Sigmoid function, and the calculation formula is as follows:

[0051] ,

[0052] ,

[0053] in, represents the weight matrix, represents the bias parameter, is the flattened representation of the output representation of the drug-target pair, express Function operation, represents the output of the fully connected layer, represents the final predicted probability, .

[0054] The advantages of the present invention are:

[0055] The present invention proposes a new method, which combines hybrid attention and conditional parameter convolution, to improve the accuracy and efficiency in drug-target interaction prediction. First, the drug and target sequences are decomposed into corresponding substructures by substructure decomposition. Then, word embedding and position embedding are performed to obtain embedding vectors, and the input vectors are processed by the Mamba layer through the selective state space model (SSM) to extract and encode features layer by layer. The introduction of the hybrid attention mechanism allows the model to independently evaluate the interaction between drugs and proteins in multiple representation subspaces. Finally, the interaction features between drugs and targets are dynamically captured by conditional parameter convolution. Specifically, the present invention brings significant advantages in the following aspects: the introduction of the hybrid attention mechanism during drug and target feature extraction allows the model to independently evaluate the interaction between drugs and proteins in multiple representation subspaces, overcoming the limitation of incomplete capture of drug and target information; before the final prediction, the convolution kernel is dynamically adjusted according to the input features by conditional parameter convolution, so that the network can more flexibly adapt to the interaction characteristics of different drugs and proteins. The present invention not only improves the accuracy of drug-target interaction prediction, but also provides a new perspective and method in the scenario of developing new drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0057] Figure 1 It is a schematic diagram of the process of the drug-target interaction prediction method based on Mamba in the present invention;

[0058] Figure 2 It is a schematic diagram of the structure of the drug-target interaction prediction method based on Mamba in the present invention;

[0059] Figure 3 This is a visualization result image of the drug-target interaction prediction performed in the present invention. DETAILED DESCRIPTION

[0060] 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.

[0061] Example 1

[0062] In this embodiment, Figure 1As shown, the present invention provides a drug-target interaction prediction method based on Mamba, and the specific steps include:

[0063] Step 1: Obtain the molecular sequences of the drug and target protein, process the molecular sequences of the drug and target protein, and obtain the substructure sequences of the drug and target protein; use a selective subsequence vocabulary to encode the sequences of the drug and target protein to process the initial data. This vocabulary is the main reference for encoding, containing all subsequences and their corresponding frequencies, and then the BPE (Byte Pair Encoding) algorithm is used to effectively decompose each sequence into smaller and smaller subsequences until it is decomposed into substructure sequences of single atoms and amino acid symbols; the formula for decomposing the substructure sequence is:

[0064] ,

[0065] ,

[0066] in, represents the set of subsequences of drug molecules, represents the input drug SMILE molecule sequence, represents the subsequence set of the target protein, Represents the input target protein sequence.

[0067] Step 2, such as Figure 2 As shown, a drug-target interaction prediction model based on Mamba is constructed, wherein the model includes an embedding layer, an encoding layer, a conditional parameter convolution layer, and a fully connected layer, wherein the encoding layer includes a Mamba layer and a hybrid attention mechanism layer.

[0068] Step three, the substructure sequences of the drug and target protein are subjected to the Mamba-based drug-target interaction prediction model to obtain the drug-target interaction prediction probability.

[0069] Specifically, in step three, the substructure sequences of the drug and the target protein are embedded and feature extracted through the embedding layer and the Mamba layer to obtain the drug feature vector and the target protein feature vector.

[0070] Specifically, the embedding process is as follows: each word in the substructure sequence of the drug and the target protein is word-embedded and position-embedded to obtain a preliminary vector representation of the word embedding and position embedding, and then the preliminary vector representation is linearly combined to obtain the embedding vector of the drug and target protein substructure sequence, which is expressed as follows:

[0071] ,

[0072] ,

[0073] ,

[0074] ,

[0075] in, represents the word embedding matrix, Indicates the first substructure sequence of the drug and target protein The index of words, is the position embedding matrix, is the position of the word in the substructure sequence of the drug and target protein, and Represent the preliminary vector representations of word embedding and position embedding respectively, and Represent the number of words in the substructure sequence of the drug and the target protein, respectively. and The embedding vectors representing the drug and target protein substructure sequences, respectively;

[0076] Specifically, feature extraction: In order to reveal the relationship between the drug and the target protein subsequence, the embedding vector is passed into the Mamba encoder module to capture the context data of the embedding vector for feature extraction. The Mamba encoder adopts a selective state space model (SSM) to identify the internal structure and association of the drug and target protein subsequences using word embedding and position embedding. Word embedding gives semantic information to each word in the input drug and target protein subsequences, extracts contextual information in the sequence, and helps to understand the relationship between words. Position embedding provides position features for each word, helping the model to identify the relative position of the word in the sequence, so that the model can effectively extract the rich information contained in the drug and target protein subsequences. The embedding vectors of the drug and target protein substructure sequences are subjected to feature extraction by the selective state space model SSM to obtain the feature vectors of the drug and target protein, which are expressed as follows:

[0077] ,

[0078] ,

[0079] in, and are the feature vectors representing drugs and target proteins respectively, represents the feature extraction operation on the embedding vector of the drug substructure sequence, Represents the feature extraction operation on the target protein substructure sequence embedding vector.

[0080] Specifically, in step 3, the drug feature vector and the target protein feature vector are feature enhanced through a hybrid attention mechanism layer to obtain a fused drug feature vector and target protein feature vector:

[0081] Specifically, the hybrid attention mechanism layer includes a hybrid attention layer generation vector unit, a bilinear transformation unit, a self-attention score and output vector unit, and a feature vector fusion unit;

[0082] Specifically, in the hybrid attention layer vector generation unit, the feature vector is passed into the hybrid attention layer to generate a new query vector, key vector, and value vector. The multi-head attention mechanism of this layer can learn information from multiple subspaces. Compared with a single feature extraction method, the hybrid attention layer overcomes the limitation of incomplete capture of drug and target information. The feature vectors of the drug and target protein are generated by the hybrid attention layer vector generation unit to generate a query vector , key vector Sum value vector , the formula is as follows:

[0083] ,

[0084] ,

[0085] ,

[0086] in, represents the weight matrix corresponding to the query vector, represents the weight matrix corresponding to the key vector, represents the weight matrix corresponding to the value vector, represents the feature vector, is the drug feature vector Or target protein feature vector ;

[0087] Specifically, in the bilinear transformation unit, after generating a new query vector, key vector, and value vector, a bilinear transformation operation is performed to enhance the correlation between the key vector and the query vector. When applying a bilinear transformation to make its dimension match the depth, improve its interaction with the query vector; Perform bilinear transformation to obtain the key vector after bilinear transformation , the formula is as follows:

[0088] ,

[0089] in, represents the weight matrix of the bilinear transformation;

[0090] Specifically, in the self-attention score and output vector unit, the query vector and the key vector after bilinear transformation Perform a dot product calculation to obtain the attention score S, and then normalize the attention score by the scaling factor to obtain the attention weight , the attention weights and the value vector Perform the product to get the output vector The output vector is passed through the output linear layer and layer normalization to obtain the final output vector , the formula is as follows:

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] in, Indicates the dimensions of each header part, represents the key matrix after bilinear transformation The transpose of is the weight matrix of the output linear transformation, The normalization operation of the representation layer is performed. After the above process, the drug feature vector is obtained. The corresponding output vector And the final output vector , and similarly obtain the target protein feature vector The corresponding output vector And the final output vector ;

[0096] Specifically, in the feature vector fusion unit, the drug feature vector and the final drug output vector Splice to get the fused drug feature vector , and similarly obtain the fusion target protein feature vector , this fusion not only combines the advantages of global and local features, but also improves the model's ability to understand the dynamics of drug-target interactions. The calculation formula is as follows:

[0097] ,

[0098] ,

[0099] in, Represents a concatenation operation.

[0100] Specifically, in step 3, the fused drug feature vector and target protein feature vector are subjected to feature interaction through a conditional parameter convolution layer to obtain an output representation of the drug-target pair:

[0101] Specifically, in order to accurately analyze the complex interactions between drug subsequences and target protein subsequences, we use an interaction function that evaluates the interaction between two entities to process the interaction between drugs and target molecules. The fused drug feature vector and target protein feature vector are interactively processed through the interaction function to obtain an interaction graph, which is expressed as follows:

[0102] ,

[0103] in, represents the interaction function, Represents an interaction graph, which is a graph with a dimension of A multidimensional tensor of is represented as the length of the fused drug feature vector, Represented as the length of the fused target protein feature vector, Representative function The output size of

[0104] Specifically, the interaction diagram The interaction features between drugs and targets are further mapped and extracted through conditional parameter convolution. Conditional parameter convolution (CondConv) allows the model to dynamically adjust its convolution kernel according to different input features by integrating the attention mechanism, solving the problem that traditional convolution operations cannot capture complex interaction features. The output representation of the drug-target pair is obtained by processing with conditional parameter convolution, and the formula is as follows:

[0105] ,

[0106] in, represents the conditional parameter convolution operation, Represents the output representation of the drug-target pair.

[0107] Specifically, in step 3, the output representation of the drug-target pair is passed through a fully connected layer to predict the drug-target interaction, and the final drug-target interaction prediction probability is obtained through a Sigmoid function, and the calculation formula is as follows:

[0108] ,

[0109] ,

[0110] in, represents the weight matrix, represents the bias parameter, is the flattened representation of the output representation of the drug-target pair, express Function operation, represents the output of the fully connected layer, represents the final predicted probability, .

[0111] Example 2

[0112] In this embodiment, Table 1 is a comparison of the MSCN-DTI model on the BIOSNAP dataset and other existing methods in the drug-target interaction DTI prediction task. In order to prove the effectiveness and practicality of the present invention, the scheme of the present invention is subjected to an inference prediction experiment on the public dataset BIOSNAP. The experimental results are shown in Table 1. The results in Table 1 show that the MSCN-DTI in the present invention has a more obvious advantage in prediction ability compared with the existing drug-target interaction DTI prediction models LR, DNN, DeepDTI, Deep DTA, GNN-CPI, HyperAttentionDTI, BACPI and BERT-based Model. Specifically, in three reasoning scenarios, the other ROC curves, PR curves and specificity indicators all achieve better results. The results illustrate the excellence and feasibility of the model proposed in the present invention. In Table 1, the ROC curve represents the overall ability of the model to distinguish between positive and negative categories. The higher the value, the better the model effect; the PR curve represents the precision and recall rate of the model on the positive class. The higher the value, the better the model performs when processing unbalanced data; the specificity represents the ability of the model to correctly identify the negative class. The larger the value, the better.

[0113] Table 1 Experimental data comparison of MSCN-DTI model and other existing methods

[0114]

[0115] In this embodiment, Figure 3 As shown in Table 1, the MSCN-DTI model of the present invention is significantly better than other existing models in three key evaluation indicators (ROC curve, PR curve and specificity). The bar graphs corresponding to each model in the figure show the performance under different indicators. The column height of the MSCN-DTI model is significantly higher than that of other models, especially in terms of PR curve and ROC curve, which further proves its excellent performance in dealing with drug-target interaction (DTI) prediction tasks. These visualization results intuitively support the conclusions in Table 1, further proving the superiority of the model of the present invention in drug-target interaction prediction tasks.

[0116] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A drug-target interaction prediction method based on Mamba, characterized in that: The following steps are involved: S1. Obtaining the molecular sequences of the drug and the target protein, processing the molecular sequences of the drug and the target protein, and obtaining the substructure sequences of the drug and the target protein; S2. Construct a drug-target interaction prediction model based on Mamba, wherein the model includes an embedding layer, an encoding layer, a conditional parameter convolution layer, and a fully connected layer, wherein the encoding layer includes a Mamba layer and a hybrid attention mechanism layer; S3. The substructure sequences of the drug and the target protein are subjected to the Mamba-based drug-target interaction prediction model to obtain a drug-target interaction prediction probability; The details are as follows: S31. The substructure sequences of the drug and the target protein are embedded and feature extracted through the embedding layer and the Mamba layer of the Mamba-based drug-target interaction prediction model to obtain a drug feature vector and a target protein feature vector: Embedding processing: Each word in the substructure sequence of the drug and target protein is word-embedded and position-embedded to obtain a preliminary vector representation of word embedding and position embedding, and then the preliminary vector representation is linearly combined to obtain the embedding vector of the drug and target protein substructure sequence, which is expressed as follows: ， ， ， ， in, represents the word embedding matrix, Indicates the first substructure sequence of the drug and target protein The index of words, is the position embedding matrix, is the position of the word in the substructure sequence of the drug and target protein, and Represent the preliminary vector representations of word embedding and position embedding respectively, and Represent the number of words in the substructure sequence of the drug and the target protein, respectively. and The embedding vectors representing the drug and target protein substructure sequences, respectively; Feature extraction: The embedding vectors of the drug and target protein substructure sequences are subjected to feature extraction by the selective state space model SSM to obtain the feature vectors of the drug and target protein. The formula is as follows: ， ， in, and are the feature vectors representing drugs and target proteins respectively, represents the feature extraction operation on the embedding vector of the drug substructure sequence, represents the feature extraction operation on the embedding vector of the target protein substructure sequence; S32. The drug feature vector and the target protein feature vector are feature enhanced by the hybrid attention mechanism layer of the Mamba-based drug-target interaction prediction model to obtain a fused drug feature vector and target protein feature vector: The hybrid attention mechanism layer includes a hybrid attention layer generation vector unit, a bilinear transformation unit, a self-attention score and output vector unit, and a feature vector fusion unit; Mixed attention layer generates vector unit: the feature vectors of the drug and target protein are generated by the mixed attention layer vector unit to generate a query vector , key vector Sum value vector , the formula is as follows: ， ， ， in, represents the weight matrix corresponding to the query vector, represents the weight matrix corresponding to the key vector, represents the weight matrix corresponding to the value vector, represents the feature vector, is the drug feature vector Or target protein feature vector ; Bilinear Transformation Unit: The key vector Perform bilinear transformation to obtain the key vector after bilinear transformation , the formula is as follows: ， in, Represents the weight matrix of the bilinear transformation; Self-attention score and output vector unit: the query vector and the key vector after bilinear transformation Perform a dot product calculation to obtain the attention score S, and then normalize the attention score by the scaling factor to obtain the attention weight , the attention weights and the value vector Perform the product to get the output vector The output vector is passed through the output linear layer and layer normalization to obtain the final output vector , the formula is as follows: ， ， ， ， in, Indicates the dimensions of each header part, represents the key matrix after bilinear transformation The transpose of is the weight matrix of the output linear transformation, The normalization operation of the representation layer is performed. After the above process, the drug feature vector is obtained. The corresponding output vector And the final output vector , and similarly obtain the target protein feature vector The corresponding output vector And the final output vector ; Feature vector fusion unit: drug feature vector and the final drug output vector Splice to get the fused drug feature vector , and similarly obtain the fusion target protein feature vector , the calculation formula is as follows: ， ， in, Represents a splicing operation; S33. The fused drug feature vector and target protein feature vector are subjected to feature interaction through the conditional parameter convolution layer of the Mamba-based drug-target interaction prediction model to obtain an output representation of the drug-target pair: The fused drug feature vector and target protein feature vector are interactively processed through an interactive function to obtain an interactive graph, which is expressed as follows: ， in, represents the interaction function, Represents an interaction graph, which is a graph with a dimension of A multidimensional tensor of is represented as the length of the fused drug feature vector, Represented as the length of the fused target protein feature vector, Representative function The output size of The interaction diagram The output representation of the drug-target pair is obtained by processing with conditional parameter convolution, and the formula is as follows: ， in, represents the conditional parameter convolution operation, represents the output representation of the drug-target pair; S34. The output representation of the drug-target pair is passed through the fully connected layer of the Mamba-based drug-target interaction prediction model to predict the drug-target interaction, and the final drug-target interaction prediction probability is obtained through the Sigmoid function. The calculation formula is as follows: ， ， in, represents the weight matrix, represents the bias parameter, is the flattened representation of the output representation of the drug-target pair, express Function operation, represents the output of the fully connected layer, represents the final predicted probability, .

2. A drug-target interaction prediction method based on Mamba according to claim 1, characterized in that: The step S1 specifically includes: The molecular sequence of the drug is converted into a drug sequence in SMILES format, the molecular sequence of the target protein is converted into an amino acid sequence, and then the drug sequence and amino acid sequence in SMILES format are decomposed into substructure sequences of the drug and the target protein by using a BPE algorithm.

Images