An end-to-end drug target affinity prediction method based on feature interaction

Abstract

This invention discloses an end-to-end drug target affinity prediction method based on feature interaction, relating to the field of drug target affinity prediction. The technical solution of this invention considers the contextual information of proteins and ligands, performing feature interaction between protein and ligand features, focusing on the interaction characteristics between proteins and ligands. By focusing on relevant features, the importance and subtle differences of these interactions can be better captured. This allows target protein characterization to provide more biological information for affinity prediction, achieving improved accuracy while eliminating dependence on protein structure. Ultimately, this invention's end-to-end drug target affinity prediction method based on feature interaction can effectively predict drug target affinity.

Description

Technical Field

[0001] This invention relates to the field of drug target affinity prediction, and in particular to an end-to-end drug target affinity prediction method based on feature interaction. Background Technology

[0002] The therapeutic effects of active compounds are determined by observing drug-target affinity prediction (DTA), in which the pharmacological activity of the drug modulates the biological processes of the target. Therefore, identifying novel molecules with relevant binding activities is crucial for biologically significant targets in the early stages of drug discovery, as drug binding ability plays a vital role in the execution of its intrinsic activity. However, performing low-throughput or high-throughput bioassays to screen potential clues is time-consuming, labor-intensive, and impractical for the vast compound and protein space, thus impacting the effectiveness of these methods. Traditional machine learning methods, including random forests (RF), support vector machines (SVM), and k-nearest neighbors (KNN), are widely used in quantitative structure-activity relationship (QSAR) and protein chemometrics (PCM) to model DTA.

[0003] Compared to traditional machine learning methods, deep learning-based methods can automatically learn feature representations from raw inputs without any domain knowledge, thus achieving better performance in predicting drug-to-target (DTA). To date, various deep learning-based methods have been developed to predict drug-to-target binding affinity. These methods can be categorized into two types based on whether they possess three-dimensional structural information: structure-based methods and sequence-based methods. Structure-based methods treat the protein-ligand complex as a 3D mesh; however, in large-scale binding affinity prediction tasks, the 3D mesh representation of the protein-ligand complex leads to enormous computational costs. Furthermore, the performance of these structure-based methods heavily relies on the high-quality 3D structure of the complex, which significantly limits their application in practical tasks. Sequence-based methods, on the other hand, use only the one-dimensional sequences of the protein and ligand as model input, extracting useful information from simplified representations of the protein and ligand to predict binding affinity. Existing research has shown that the structural complementarity between the protein-binding capsule and the ligand significantly affects the binding strength between the protein and ligand.

[0004] However, existing sequence-based methods typically learn feature representations of the protein binding pocket and ligand through two separate modules, simply concatenating the extracted feature vectors while ignoring the individual contributions of each binding component and the shared contributions of the involved interactive substructures. Furthermore, using simple protein sequence input networks, which only contain linear amino acid arrangement information, fails to provide structural details. Due to model complexity, excessive complexity can lead to overfitting, while insufficient complexity can result in underfitting. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an end-to-end drug target affinity prediction method based on feature interaction. This method involves feature interaction between protein and ligand features, focusing on the interaction characteristics between them. By concentrating attention on relevant features, it better captures the importance and subtle differences of these interactions. This allows target protein characterization to provide more biological information for affinity prediction, achieving improved accuracy while eliminating dependence on protein structure.

[0006] The technical solution adopted by this invention to solve its technical problem is: an end-to-end drug target affinity prediction method based on feature interaction, comprising the following steps:

[0007] S1, Preprocessing of protein and drug molecule data: For drug compounds, molecular embedding is used to map discrete sequence data; for protein sequences, distance graphs are used to represent amino acid residues; and the preprocessed encoding vectors are input into the backbone network.

[0008] S2, Protein feature extraction based on dynamic CNN and drug feature extraction based on BiLSTM;

[0009] S3, after splicing protein features and drug features, inputs them into the drug target affinity prediction network and outputs the drug target affinity prediction score.

[0010] S4, Determine the optimal model architecture and parameter set: Use the K-Fold method with chemical genomics representation to divide the dataset into representative folds and determine hyperparameters; The entire network is trained end-to-end, continuously updating the network weights until the model converges.

[0011] Preferably, step S1 is specifically as follows: preprocessing the protein and drug sequence data and normalizing the discrete sequence data.

[0012] Preferably, step S1 is as follows:

[0013] Drug compounds are represented in SMILES format;

[0014] Proteins can be simply described as a list of amino acid residues P = (γ1,...,γ) n A linear sequence consisting of γ, where γ i Let l be the one-hot encoded representative vector of the i-th amino acid type with a length of 20, where l is the sequence length;

[0015] Calculate the pairwise distance graph in two dimensions using the following method:

[0016]

[0017] Where d(γ) i ,γ j ) is the C of residues i and j α The distance between atoms, d0, is set to set up Let r be an l-dimensional distance vector. i All residuals in P. Proteins can be represented by a distance matrix:

[0018]

[0019] Preprocessing the input discrete sequence data of proteins and drug compounds can preserve the information of the discrete data and provide distance information between residues in the protein structure. This distance information includes secondary structure, domain boundaries and spatial relationships between residues, quantifying the structural similarity or difference between different proteins.

[0020] Preferably, step S2 is as follows:

[0021] In the end-to-end drug target affinity prediction method based on feature interaction, the backbone network uses a combination of Dynamic Convolutional Neural Network (DynCNN) and Bidirectional Long Short-Term Memory (BiLSTM) based on an attention mechanism for multivariate discrete sequence feature extraction. The preprocessed multivariate discrete sequences are input separately into the backbone network, employing a dual-branch structure: a protein feature extraction network based on Dynamic CNN and a drug feature extraction network based on BiLSTM. The Dynamic CNN structure is used to extract protein features, preserving contextual information and capturing relationships between protein residues. The BiLSTM structure is used to extract drug features, capturing molecular response site features. This transforms the raw data into a low-dimensional latent variable representation, models the multivariate data distribution, and obtains representative protein and drug features.

[0022] Preferably, step S2 is as follows:

[0023] Step S2.1, Protein feature extraction based on dynamic CNN;

[0024] Step S2.2, Drug feature extraction based on BiLSTM.

[0025] Preferably, in step S2.1, an adaptive CNN is used to encode the protein distance map into a fixed-size vector representation; the CNN module consists of stacked residual blocks and sequential self-attention blocks; for the residual blocks, an improved residual network is used to process the protein input; given a protein distance map P∈R l×l Using a with N f1 A 3×3 convolutional layer is applied to the filter; then the output is processed using stacked residual blocks. The input to each intermediate residual unit can be x q Let q (1≤q≤Q) be the unit index of the block. Each residual block can be defined as:

[0026] y q =H(x) q )+F(x q +W q )

[0027] x q+1 =f(y q )

[0028] Where x q and x q+1 Let H(x) be the input and output of the q-th block unit, F be the residual function, and H(x) be the input and output of the q-th block unit. q )=x q Set as unit mapping; It is the set of weights associated with the q-th residual unit, where The number of filters is denoted by f; the f function is the activation function, using exponential linear units; each residual block consists of a stack of a 5×5 convolutional layer and a 3×3 convolutional layer.

[0029] Design a dynamic neural network capable of handling variable-length inputs while predicting the importance of each amino acid; remove pooling layers between residual blocks and use zero-padding on both sides of the input to ensure the output of the residual block has the same size as the input; the output of the last residual block is still l×l×N. f The dimensions are then used; average pooling is then used to compress the residual blocks to enrich the output for downstream processing;

[0030] A protein feature map is obtained by average pooling. P q This can be viewed as a sequence representation of a protein, where l is the number of amino acids (sites) in the protein, and N... f This represents the spatial features of each site; since most sites are not directly related to drug binding, identifying a small subset of binding sites is crucial for accurate DTA prediction; a multi-head sequential attention mechanism is employed to fully utilize these features for classification; such as... Figure 1 As shown, the attention mechanism is based on Pq Given a value as input, output a value with weight a. p vector:

[0031]

[0032]

[0033] in w p2 For a size of d p The parameter vector, where d p For adjustable hyperparameters; this vector representation typically focuses on a set of contiguous protein sequence sites; since the protein binding bag is composed of multiple spatially adjacent contiguous sites, further... p2 Expand to r p ×d p A matrix of size W is called W. p2 To capture the overall structural information of the combination. (This is followed by an unrelated instruction: "a") p Converted into bullish attention weights for:

[0034]

[0035] The above equation can be viewed as an unbiased 2-layer multilayer perceptron with hidden cell number d. a The parameter is {W p1 W p2}; By labeling matrix A p With feature map P q Calculate r by multiplication p Weighted sum:

[0036] P a =A p P q

[0037] Where P a It is a size of r p ×N f The feature map contains the potential relationships between the contributions of protein sites to the interaction.

[0038] Preferably, in step S2.1, each drug molecule SMILES string is encoded into a two-dimensional embedding matrix M∈R. n ×d The token vectors in the molecular matrix M are independent of each other; BiLSTM is used to process molecules.

[0039]

[0040]

[0041] and Connect to obtain a hidden state h. i To replace token embedding t i Therefore h t This becomes a more information-rich vector, acquiring a certain dependency between adjacent tokens in the molecule; all h at each time step i... i Let it be H.

[0042]

[0043] H=(h0,h i ,...,h n )

[0044] If the hidden cell number of each unidirectional LSTM is set to u, then the shape of H is n×2u; a multi-head self-attention mechanism is introduced to determine the relationship between tokens and interactions; the attention mechanism takes the entire LSTM hidden state H as input and outputs a weight A. m vector:

[0045] A m =softmax(MLP(H T ))

[0046] The number of hidden units in the MLP is d. m The parameter is {W m1 W m2}; By using annotation matrix A m The weighted sum is calculated by multiplying the LSTM hidden state H, and the resulting matrix is ​​the self-focused molecule embedding:

[0047] M a =A m H

[0048] Where M a It is a self-focused drug molecule characterization map that contains potential relationships between tokens in interactions; M a The size is r m ×2u, where r m It is an adjustable hyperparameter that represents the number of attention vectors.

[0049] Preferably, step S3 is as follows: For P a and M a The input is given to the cross-attention layer, and the corresponding attention weights are obtained. Then, the obtained weight vectors are normalized so that their sum is 1. This yields two information-rich one-dimensional vectors. and Then input it into the MLP layer; and Connect them, that is We obtain an output vector o∈R 2 :

[0050]

[0051] in, Let b be the weight matrix. o ∈R 2 This is the bias vector. Finally, a sigmoid function is added to the output layer o = [y0, y1], and the modeling is performed as follows:

[0052]

[0053] In the formula, t∈{0,1} represents a binary label (i.e., whether it is interactive or not), p t Let t represent the probability of occurrence. The probability is t=1; given a dataset D={(m i ,p i ,y i The training objective is to minimize the loss function L, whose cross-entropy loss is:

[0054]

[0055] The beneficial effects of this invention are:

[0056] This patent collects a large amount of information on protein-protein interactions and sequence similarities, and performs feature fusion between proteins and drugs. This allows target protein characterization to provide more biological information for affinity prediction, improving the accuracy of affinity prediction while eliminating dependence on protein structure. Specifically, this method predicts DTA directly from protein distance maps and chemical language. The protein distance map and drug SMILES are encoded by a Dynamic Convolutional Neural Network (DynCNN) and a Bidirectional Long Short-Term Memory (BiLSTM) based on an attention mechanism, respectively. The outputs are connected to a cross-attention layer, which uses the cross-attention mechanism to capture the interaction features between proteins and ligands. Then, dilated convolutions are used to learn multi-scale long-range features of proteins and ligands respectively. Finally, the learned feature vectors of proteins and ligands are concatenated and used as input to a fully connected neural network (FNN) to predict the binding affinity between proteins and ligands. The weight parameters are continuously updated through a hyperparameter optimization strategy until the network converges.

[0057] This algorithm is a drug target affinity prediction algorithm that combines dynamic convolutional neural networks, bidirectional long short-term memory networks, and cross-attention mechanisms. It considers the contextual information of proteins and ligands, performs feature interactions between protein and ligand features, and focuses on the interaction features between proteins and ligands. By concentrating attention on relevant features, it better captures the importance and subtle differences of these interactions. This allows target protein characterization to provide more biological information for affinity prediction, achieving improved accuracy while eliminating dependence on protein structure. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only three of the drawings in this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0059] Figure 1 The proposed model framework;

[0060] Figure 2 Dynamic CNN. It consists of two key components: a) residual blocks and b) attention blocks;

[0061] Figure 3 K-Fold method for chemical genomics representation. Detailed Implementation

[0062] To enhance understanding of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. These embodiments are only used to explain the invention and do not limit the scope of protection of the invention.

[0063] Example

[0064] like Figure 1 As shown, an end-to-end drug target affinity prediction method based on feature interaction includes the following steps:

[0065] S1, Preprocessing of protein and drug molecule data: For drug compounds, molecular embedding is used to map discrete sequence data; for protein sequences, distance graphs are used to represent amino acid residues; and the preprocessed encoding vectors are input into the backbone network.

[0066] Step S1 is as follows: Protein and drug sequence data are discrete data structures, with data values ​​that are discrete and finite. Considering subsequent data analysis and modeling, the discrete sequence data is first preprocessed. Considering the characteristics and distribution of discrete sequence data, protein sequence data often contains different lengths and residue distributions. To avoid the influence of different dimensions on data analysis and modeling, the discrete sequence data is normalized.

[0067] Step S1 is as follows:

[0068] Drug compounds are represented in SMILES format;

[0069] This is a text string containing topological information based on chemical bond rules. For example, a benzene ring can be encoded as "c1ccccc1". Assuming a linear symbol for a drug molecule containing n symbols, the molecule can be represented by a molecular embedding sequence as M = (t1,...,t...). n ), where t i It is the d-dimensional token embedding vector of the i-th token. Therefore, M∈R n×d It is a representation of a two-dimensional matrix in which all the markers are embedded and connected together.

[0070] Proteins can be simply described as a list of amino acid residues P = (γ1,...,γ) n A linear sequence consisting of γ, where γ i Let l be the one-hot encoded representative vector of the i-th amino acid type with a length of 20, where l is the sequence length;

[0071] Calculate the pairwise distance graph in two dimensions using the following method:

[0072]

[0073] Where d(γ) i ,γ j ) is the C of residues i and j α The distance between atoms, d0, is set to set up Let r be an l-dimensional distance vector. i All residuals in P. Proteins can be represented by a distance matrix:

[0074]

[0075] Preprocessing the input discrete sequence data of proteins and drug compounds can preserve the information of the discrete data and provide distance information between residues in the protein structure. This distance information includes secondary structure, domain boundaries and spatial relationships between residues, quantifying the structural similarity or difference between different proteins.

[0076] S2, Protein feature extraction based on dynamic CNN and drug feature extraction based on BiLSTM;

[0077] Step S2 is as follows:

[0078] In the end-to-end drug target affinity prediction method based on feature interaction, the backbone network uses a combination of Dynamic Convolutional Neural Network (DynCNN) and Bidirectional Long Short-Term Memory (BiLSTM) based on an attention mechanism for multivariate discrete sequence feature extraction. The preprocessed multivariate discrete sequences are input separately into the backbone network, employing a dual-branch structure: a protein feature extraction network based on Dynamic CNN and a drug feature extraction network based on BiLSTM. The Dynamic CNN structure is used to extract protein features, preserving contextual information and capturing relationships between protein residues. The BiLSTM structure is used to extract drug features, capturing molecular response site features. This transforms the raw data into a low-dimensional latent variable representation, models the multivariate data distribution, and obtains representative protein and drug features. The proposed model architecture is as follows: Figure 1 As shown.

[0079] Step S2.1: An adaptive CNN is used to encode the protein distance map into a fixed-size vector representation; the CNN module consists of stacked residual blocks and sequential self-attention blocks; for the residual blocks, an improved residual network is used to process the protein input; given a protein distance map P∈R l×l Using a with N f1 A 3×3 convolutional layer is applied to the filter; then the output is processed using stacked residual blocks. The input to each intermediate residual unit can be x q Let q (1≤q≤Q) be the unit index of the block. Each residual block can be defined as:

[0080] y q =H(x) q )+F(x q +W q )

[0081] x q+1 =f(y q )

[0082] Where x q and x q+1 Let H(x) be the input and output of the q-th block unit, F be the residual function, and H(x) be the input and output of the q-th block unit. q )=x q Set as unit mapping; It is the set of weights associated with the q-th residual unit, where This is the number of filters; the f function is the activation function, using exponential linear units (ELU); each residual block consists of a stack of a 5×5 convolutional layer and a 3×3 convolutional layer.

[0083] Design a dynamic neural network capable of handling variable-length inputs while predicting the importance of each amino acid; remove pooling layers between residual blocks and use zero-padding on both sides of the input to ensure the output of the residual block has the same size as the input; the output of the last residual block is still l×l×N. f The dimensions are then used; average pooling is then used to compress the residual blocks to enrich the output for downstream processing;

[0084] A protein feature map is obtained by average pooling. P q This can be viewed as a sequence representation of a protein, where l is the number of amino acids (sites) in the protein, and N... f This represents the spatial features of each site; since most sites are not directly related to drug binding, identifying a small subset of binding sites is crucial for accurate DTA prediction; a multi-head sequential attention mechanism is employed to fully utilize these features for classification; such as... Figure 1 As shown, the attention mechanism is based on P q Given a value as input, output a value with weight a. p vector:

[0085]

[0086]

[0087] in w p2 For a size of d p The parameter vector, where d p For adjustable hyperparameters; this vector representation typically focuses on a set of contiguous protein sequence sites; since the protein binding bag is composed of multiple spatially adjacent contiguous sites, further... p2 Expand to r p ×d p A matrix of size W is called W. p2 To capture the overall structural information of the combination. (This is followed by an unrelated instruction: "a") p Converted into bullish attention weights for:

[0088]

[0089] The above equation can be viewed as an unbiased 2-layer multilayer perceptron with hidden cell number d. a The parameter is {W p1 W p2}; By labeling matrix A p With feature map P q Calculate r by multiplication p Weighted sum:

[0090] P a =A p P q

[0091] Where P a It is a size of r p ×N f The feature map contains the potential relationships between the contributions of protein sites to the interaction.

[0092] In step S2.1, each drug molecule SMILES string is encoded into a two-dimensional embedding matrix M∈R. n×d The token vectors in the molecular matrix M are independent of each other; BiLSTM is used to process molecules.

[0093]

[0094]

[0095] and Connect to obtain a hidden state h. i To replace token embedding t i Therefore h t This becomes a more information-rich vector, acquiring a certain dependency between adjacent tokens in the molecule; all h at each time step i... i Let it be H.

[0096]

[0097] H=(h0,h i ,...,h n )

[0098] If the hidden cell number of each unidirectional LSTM is set to u, then the shape of H is n×2u; a multi-head self-attention mechanism is introduced to determine the relationship between tokens and interactions; the attention mechanism takes the entire LSTM hidden state H as input and outputs a weight A. m vector:

[0099] A m =softmax(MLP(H T ))

[0100] The number of hidden units in the MLP is d. m The parameter is {W m1 W m2}; By using annotation matrix A m The weighted sum is calculated by multiplying the LSTM hidden state H, and the resulting matrix is ​​the self-focused molecule embedding:

[0101] M a =A m H

[0102] Where M a It is a self-focused drug molecule characterization map that contains potential relationships between tokens in interactions; M a The size is r m ×2u, where r m It is an adjustable hyperparameter that represents the number of attention vectors.

[0103] S3, after splicing protein features and drug features, inputs them into the drug target affinity prediction network and outputs the drug target affinity prediction score.

[0104] For P a and M a The input is given to the cross-attention layer, and the corresponding attention weights are obtained. Then, the obtained weight vectors are normalized so that their sum is 1. This yields two information-rich one-dimensional vectors. and Then input it into the MLP layer; and Connect them, that is We obtain an output vector o∈R 2 :

[0105]

[0106] in, Let b be the weight matrix. o ∈R 2 This is the bias vector. Finally, a sigmoid function is added to the output layer o = [y0, y1], and the modeling is performed as follows:

[0107]

[0108] In the formula, t∈{0,1} represents a binary label (i.e., whether it is interactive or not), p t Let t represent the probability of occurrence. The probability is t=1; given a dataset D={(m i ,p i ,y i The training objective is to minimize the loss function L, whose cross-entropy loss is:

[0109]

[0110] S4, Determining the Optimal Model Architecture and Parameter Set: The most common approach to determining the optimal model architecture and parameter set is grid search with cross-validation, where the dataset is split into different folds under varying conditions depending on the method used. However, in the context of this problem, traditional cross-validation methods are often unsatisfactory or unrepresentative, especially considering the Davis dataset's performance on pK. d The value distribution is highly unbalanced, and 1D structural data is used to characterize compounds. The K-Fold method, based on chemical genomics representation, is used to segment the dataset into representative folds and determine hyperparameters; this method considers pK during segmentation. d The distribution of values, the similarity of protein sequences, and the similarity of SMILES strings. It initially included pK d DTI values ​​greater than 5 are distributed across different K-Folds based on the lowest similarity scores. This index corresponds to a weighted average between the median of all protein sequence similarity scores and the median of all SMILES string similarity scores, calculated between each entry in the sample and its corresponding set. Furthermore, this method guarantees that each set is of equal size, thus considering only sets that have not previously been assigned samples in each step. In pK... d After a pair with a value greater than 5, pair pK d This process is repeated for DTA with a value of 5. The entire network is trained end-to-end, continuously updating the network weights until the model converges. Figure 3 This demonstrates the K-Fold method for chemical genomics representation.

[0111] The beneficial effects of this invention are:

[0112] This patent collects a large amount of information on protein-protein interactions and sequence similarities, and performs feature fusion between proteins and drugs. This allows target protein characterization to provide more biological information for affinity prediction, improving the accuracy of affinity prediction while eliminating dependence on protein structure. Specifically, this method predicts DTA directly from protein distance maps and chemical language. The protein distance map and drug SMILES are encoded by a Dynamic Convolutional Neural Network (DynCNN) and a Bidirectional Long Short-Term Memory (BiLSTM) based on an attention mechanism, respectively. The outputs are connected to a cross-attention layer, which uses the cross-attention mechanism to capture the interaction features between proteins and ligands. Then, dilated convolutions are used to learn multi-scale long-range features of proteins and ligands respectively. Finally, the learned feature vectors of proteins and ligands are concatenated and used as input to a fully connected neural network (FNN) to predict the binding affinity between proteins and ligands. The weight parameters are continuously updated through a hyperparameter optimization strategy until the network converges.

[0113] This algorithm is a drug target affinity prediction algorithm that combines dynamic convolutional neural networks, bidirectional long short-term memory networks, and cross-attention mechanisms. It considers the contextual information of proteins and ligands, performs feature interactions between protein and ligand features, and focuses on the interaction features between proteins and ligands. By concentrating attention on relevant features, it better captures the importance and subtle differences of these interactions. This allows target protein characterization to provide more biological information for affinity prediction, achieving improved accuracy while eliminating dependence on protein structure.

[0114] The specific implementation details of each part of this invention are as follows:

[0115] 1. First, the discrete sequence data is preprocessed. Considering the characteristics and distribution of discrete sequence data, protein sequence data often contain different lengths and residue distributions. In order to avoid the influence of different dimensions on data analysis and modeling, the discrete sequence data is normalized.

[0116] 2. A dual-branch feature extraction structure is used, consisting of a protein feature extraction network based on dynamic CNN and a drug feature extraction network based on BiLSTM. The raw data is transformed into a low-dimensional latent variable representation, and the multivariate data distribution is modeled to obtain representative protein and drug features.

[0117] 3. After obtaining multiple sets of features, they are spliced ​​together and input into the MLP network to obtain the results of drug target affinity prediction.

[0118] 4. PK was considered during the verification process. d The distribution of values, similarity of protein sequences, and similarity of SMILES strings were analyzed. The K-Fold method, using chemigenomics representation, was used to divide the dataset into representative folds and determine hyperparameters. Simultaneously, the entire network was trained end-to-end, continuously updating weights and parameters until the network converged.

[0119] The foregoing description illustrates the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

[0120] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0121] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An end-to-end drug target affinity prediction method based on feature interaction, characterized in that, Includes the following steps: S1, Preprocessing of protein and drug molecule data: For drug compounds, molecular embedding is used to map discrete sequence data; for protein sequences, distance graphs are used to represent amino acid residues; and the preprocessed encoding vectors are input into the backbone network. S2, Protein feature extraction based on dynamic CNN and drug feature extraction based on BiLSTM; S3, after splicing protein features and drug features, inputs them into the drug target affinity prediction network and outputs the drug target affinity prediction score. S4, Determine the optimal model architecture and parameter set: Use the K-Fold method with chemical genomics representation to divide the dataset into representative folds and determine the hyperparameters; The entire network is trained end-to-end, continuously updating the network weights until the model converges.

2. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: Step S1 is as follows: preprocess the protein and drug sequence data and normalize the discrete sequence data.

3. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: Step S1 is as follows: Drug compounds are represented in SMILES format; Proteins can be simply described as a list of amino acid residues P = (γ1,...,γ) n A linear sequence consisting of γ, where γ i Let l be the one-hot encoded representative vector of the i-th amino acid type with a length of 20, where l is the sequence length; Calculate the pairwise distance graph in two dimensions using the following method: Where d(γ) i ,γ j ) is the C of residues i and j α The distance between atoms, d0, is set to set up Let r be an l-dimensional distance vector. i All residuals in P. Proteins can be represented by a distance matrix: Preprocessing the input discrete sequence data of proteins and drug compounds can preserve the information of the discrete data and provide distance information between residues in the protein structure. This distance information includes secondary structure, domain boundaries and spatial relationships between residues, quantifying the structural similarity or difference between different proteins.

4. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: Step S2 is as follows: In the end-to-end drug target affinity prediction method based on feature interaction, the backbone network uses a combination of Dynamic Convolutional Neural Network (DynCNN) and Bidirectional Long Short-Term Memory (BiLSTM) based on attention mechanism for multivariate discrete sequence feature extraction. The preprocessed multivariate discrete sequences are input separately into the backbone network, which employs a dual-branch structure: a protein feature extraction network based on Dynamic CNN and a drug feature extraction network based on BiLSTM. Using a dynamic CNN structure to extract protein features, preserving the contextual information of the features, and capturing the relationships between protein residues; Using BiLSTM structures to extract drug features and capture molecular response site features, the raw data is transformed into a low-dimensional latent variable representation, and the multivariate data distribution is modeled to obtain representative protein and drug features.

5. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: Step S2 is as follows: Step S2.1, Protein feature extraction based on dynamic CNN; Step S2.2, Drug feature extraction based on BiLSTM.

6. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: Step S2.1: An adaptive CNN is used to encode the protein distance map into a fixed-size vector representation; the CNN module consists of stacked residual blocks and sequential self-attention blocks; for the residual blocks, an improved residual network is used to process the protein input; given a protein distance map P∈R l×l Using a with N f1 A 3×3 convolutional layer is applied to the filter; then the output is processed using stacked residual blocks. The input to each intermediate residual unit can be x q Let q (1≤q≤Q) be the unit index of the block. Each residual block can be defined as: y q = H(x q ) + F(x q ) + W q ) x q+1 = f(y q ) Where x q and x q+1 Let H(x) be the input and output of the q-th block unit, F be the residual function, and H(x) be the input and output of the q-th block unit. q )=x q Set as unit mapping; It is the set of weights associated with the q-th residual unit, where It represents the number of filters; The f function is the activation function, using exponential linear units; each residual block consists of a stack of a 5×5 convolutional layer and a 3×3 convolutional layer. Design a dynamic neural network capable of handling variable-length inputs while predicting the importance of each amino acid; remove pooling layers between residual blocks and use zero-padding on both sides of the input to ensure the output of the residual block has the same size as the input; the output of the last residual block is still l×l×N. f The dimensions are then used; average pooling is then used to compress the residual blocks to enrich the output for downstream processing; A protein feature map is obtained by average pooling. P q This can be viewed as a sequence representation of a protein, where l is the number of amino acids (sites) in the protein, and N... f The spatial features of each site are represented; since most sites are not directly related to drug binding, identifying a small subset of binding sites is crucial for accurate DTA prediction; a multi-head sequential attention mechanism is employed to fully utilize these features for classification. As shown in Figure 1(b), attention mechanism is based on P q Given a value as input, output a value with weight a. p vector: in w p2 For a size of d p The parameter vector, where d p For adjustable hyperparameters; this vector representation typically focuses on a set of contiguous protein sequence sites; since the protein binding bag is composed of multiple spatially adjacent contiguous sites, further... p2 Expand to r p ×d p A matrix of size W is called W. p2 To capture the overall structural information of the combination. (This is followed by an unrelated instruction: "a") p Converted into bullish attention weights for: The above equation can be viewed as an unbiased 2-layer multilayer perceptron with hidden cell number d. a The parameter is {W p1 W p2 }; By labeling matrix A p With feature map P q Calculate r by multiplication p Weighted sum: P a =A p P q Where P a It is a size of r p ×N f The feature map contains the potential relationships between the contributions of protein sites to the interaction.

7. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: In step S2.1, each drug molecule SMILES string is encoded into a two-dimensional embedding matrix M∈R. n×d The token vectors in the molecular matrix M are independent of each other; BiLSTM is used to process molecules. and Connect to obtain a hidden state h. i To replace token embedding t i Therefore h t This becomes a more information-rich vector, acquiring a certain dependency between adjacent tokens in the molecule; all h at each time step i... i Let it be H. H=(h0,h i ,...,h n ) If the hidden cell number of each unidirectional LSTM is set to u, then the shape of H is n×2u; the relationship between tokens and interactions is determined by introducing a multi-head self-attention mechanism; the attention mechanism takes the entire LSTM hidden state H as input and outputs a weight A. m vector: A m =softmax(MLP(H T )) The number of hidden units in the MLP is d. m The parameter is {W m1 W m2 }; By using annotation matrix A m The weighted sum is calculated by multiplying the LSTM hidden state H, and the resulting matrix is ​​the self-focused molecule embedding: M a =A m H Where M a It is a self-focused drug molecule characterization map that contains potential relationships between tokens in interactions; M a The size is r m ×2u, where r m It is an adjustable hyperparameter that represents the number of attention vectors.

8. The end-to-end drug target affinity prediction method based on feature interaction according to claim 1, characterized in that: Step S3 is as follows: For P a and M a Input the cross-attention layer to obtain the corresponding attention weights, and then normalize the obtained weight vectors so that their sum is 1; Able to obtain two information-rich one-dimensional vectors and Then input it into the MLP layer; and Connect them, that is We obtain an output vector o∈R 2 : in, Let b be the weight matrix. o ∈R 2 This is the bias vector. Finally, a sigmoid function is added to the output layer o = [y0, y1], and the modeling is performed as follows: In the formula, t∈{0,1} represents a binary label (i.e., whether it is interactive or not), p t Let t represent the probability of occurrence. The probability is t=1; given dataset D={(m i ,p i ,y i The training objective is to minimize the loss function L, whose cross-entropy loss is:

Images