Drug target binding affinity prediction method and system based on multi-scale feature fusion

Abstract

The invention discloses a drug target binding affinity prediction method based on multi-scale feature fusion, and the method comprises the steps: introducing the multi-scale structural features of atoms, atomic groups and molecular levels in the expression of drug molecules, and combining a graph neural network and a sequence modeling module as a feature extractor; and deep interaction and fusion between different scale features of the protein and the drug are realized by using a multi-head bilinear cross attention mechanism, so that potential binding site information is effectively captured, and the accuracy and interpretation capability of affinity prediction are improved. The method not only overcomes the problems that a traditional machine learning method depends on manpower and is low in efficiency in feature construction, but also solves the technical bottlenecks that an existing deep learning method is only limited to local neighborhood information and cannot model global structural features, and the interaction relationship modeling capability is insufficient due to direct splicing of drugs and protein representation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bioinformatics and computational chemistry, and more particularly relates to a drug target binding affinity prediction method and system based on multi-scale feature fusion. BACKGROUND

[0002] With the continuous development of artificial intelligence, deep learning and computational biology, drug target binding affinity (DTA) prediction based on computational methods in the drug discovery process has become one of the important directions of drug research and development. Compared with the high cost and long period required by traditional experimental methods such as high-throughput screening (HTS), drug target binding affinity prediction models based on machine learning and deep learning have the advantages of high accuracy, low cost and fast feedback, and are widely used in virtual screening, drug repositioning and molecular optimization and other stages.

[0003] In existing drug target binding affinity prediction methods, deep learning methods have become the mainstream technical path, which can be mainly divided into two categories: (1) sequence-based deep learning methods, which usually use the sequences of drugs and targets as input, and the modeling methods include models based only on one-dimensional sequence information, and models further introducing two-dimensional structure information such as molecular structure graph and predicted protein contact map. This kind of method has the advantages of convenient data acquisition and low computational overhead; (2) structure-based deep learning methods, which rely on three-dimensional structure information of drugs and proteins, use three-dimensional coordinates, binding pockets, surface residues or spatial adjacency graphs for modeling, often combined with graph neural networks, three-dimensional convolutional networks or geometric deep learning techniques, to achieve fine description of spatial interaction at the atomic level, and perform well in improving prediction accuracy.

[0004] Although the above methods have achieved certain results in drug target binding affinity prediction, there are still some defects that cannot be ignored:

[0005] (1) Structure-based methods rely on high-quality structure data, but such data are not always available, and cannot avoid the high cost of structure determination and the uncertainty of structure prediction;

[0006] (2) Sequence-based methods have the highest adaptability, as they do not need to rely on three-dimensional structure information of proteins and ligands, but their large number of parameters make the model prone to learn noise and error patterns, leading to a decrease in prediction accuracy. SUMMARY

[0007] In view of the above defects or improvement needs of the prior art, the present application provides a drug target binding affinity prediction method and system based on multi-scale feature fusion, which aims to solve the technical problems that the existing structure-based method relies on high-quality structure data, but such data is not always available, and cannot avoid high structure determination cost and uncertainty of structure prediction, and the existing sequence-based method has the highest adaptability, which does not need to rely on three-dimensional structure information of proteins and ligands, but a large number of parameters make the model easily learn noise and error patterns, resulting in a decrease in prediction accuracy.

[0008] To achieve the above-mentioned purpose, according to one aspect of the present application, a drug target binding affinity prediction method based on multi-scale feature fusion is provided, comprising the following steps:

[0009] (1) Obtain a drug-protein data pair to be predicted, and preprocess the drug-protein data pair to be predicted to obtain a preprocessed drug-protein data pair, which includes a simplified molecular linear input canonical (SMILES) sequence of a drug, an amino acid sequence of a protein, and drug target binding affinity data;

[0010] (2) For the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained in step (1), use a pre-trained evolutionary scale model (ESM) to process the amino acid sequence to obtain an evolutionary language modeling representation and a contact probability matrix of the amino acid sequence, connect the residue pairs with a contact probability greater than or equal to a preset contact probability threshold in the contact probability matrix as edges in a protein structure graph, thereby obtaining a protein structure graph corresponding to the amino acid sequence, each node in the protein structure graph represents an amino acid, and the edges represent the contact between two amino acids; and map each amino acid in the amino acid sequence to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of length of the amino acid sequence x 33;

[0011] (3) For the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained in step (1), use the RDKit tool to parse the SMILES sequence to obtain a molecular structure graph corresponding to the SMILES sequence, obtain an atomic feature matrix according to the physicochemical properties of each atom in the molecular structure graph, generate an atomic group set according to the molecular structure graph corresponding to the obtained SMILES sequence, and generate an atomic group structure graph corresponding to the SMILES sequence according to the connection relationship between the atomic groups in the atomic group set;

[0012] (4) inputting the protein structure diagram and the residue feature matrix corresponding to each amino acid sequence obtained in step (2) and the molecular structure diagram, the atom group structure diagram and the atom feature matrix corresponding to the SMILES sequence of each drug obtained in step (3) into the pre-trained drug target binding affinity prediction model to obtain a drug target binding affinity prediction value.

[0013] Preferably, step (1) specifically comprises: firstly, screening the drug-protein data pairs to be predicted, and deleting data with data missing problems to obtain the SMILES sequence of the drug, the amino acid sequence of the protein, and the inhibition constant and the dissociation constant of the drug-protein interaction; and then obtaining the drug target binding affinity data according to the obtained inhibition constant and dissociation constant, respectively; wherein the calculation formula of the drug target binding affinity data is:

[0014]

[0015] wherein, wherein K i and K d respectively represent the inhibition constant and the dissociation constant, pK i represents the drug target binding affinity data calculated based on the inhibition constant, and pK d represents the drug target binding affinity data calculated based on the dissociation constant.

[0016] Finally, the obtained SMILES sequence of the drug, the amino acid sequence of the protein, and the drug target binding affinity data are combined into the preprocessed drug-protein data pairs.

[0017] Preferably, the drug target binding affinity prediction model comprises a protein encoder module, a drug encoder module, a feature fusion module, and a multi-layer perception module.

[0018] The protein encoder module comprises a protein feature extraction layer, a residue cluster clustering layer, and a protein pooling layer.

[0019] The protein feature extraction layer is realized based on Mamba and message passing neural network MPNN, and comprises a protein preprocessing layer, three protein local message passing layers, three structured state modeling layers, and three protein post-processing layers.

[0020] The residue cluster clustering layer comprises three residue cluster construction layers and three residue cluster pooling layers.

[0021] The protein pooling layer comprises three residue feature updating layers and a protein attention pooling layer.

[0022] The drug encoder module comprises a drug feature extraction layer, three multi-scale feature updating layers, and a drug pooling layer.

[0023] The drug feature extraction layer is implemented based on a bidirectional long short-term memory network BiLSTM and an MPNN, and includes a drug preprocessing layer, three drug local message passing layers, three sequence context modeling layers, and three drug post-processing layers.

[0024] The drug pooling layer includes three atomic feature update layers and a molecular attention pooling layer.

[0025] The feature fusion module includes three fusion layers.

[0026] Preferably, the input of the protein preprocessing layer is an evolutionary language modeling representation of a 32-amino acid sequence, a residue feature matrix, and a contact probability matrix; the protein preprocessing layer first processes the input contact probability matrix using a radial basis function to obtain an edge feature matrix with a size of E x 200, where E represents the number of edges in the protein structure graph, then processes the evolutionary language modeling representation through a multi-layer perception MLP, where the hidden layer uses a rectified linear unit ReLU as an activation function, and the output layer uses layer normalization, where the input of the first layer is an evolutionary language modeling representation with a dimension of V x 1280, and the output is an intermediate feature vector with a dimension of V x 400, the input of the second layer is an intermediate feature vector with a dimension of V x 400, and the final output is a preprocessed evolutionary language modeling representation with a dimension of V x 200, where V represents the length of the amino acid sequence, to obtain a preprocessed evolutionary language modeling representation; subsequently, the residue feature matrix is processed through a two-layer MLP, the hidden layer uses ReLU as an activation function, and the output layer uses layer normalization, where the input of the first layer is a residue feature matrix with a dimension of V x 33, the output is an intermediate feature vector with a dimension of V x 400, the input of the second layer is an intermediate feature vector with a dimension of V x 400, and the output is a preprocessed residue feature matrix with a dimension of V x 200, to obtain a preprocessed residue feature matrix; finally, the preprocessed evolutionary language modeling representation with a dimension of V x 200 and the preprocessed residue feature matrix with a dimension of V x 200 are added, and the final output is a comprehensive feature of protein residues with a dimension of V x 200;

[0027] The input of the first protein local message passing layer is the comprehensive feature of protein residues with a dimension of V x 200 output by the protein preprocessing layer and the protein structure graph, and the protein local message passing layer uses a primary neighborhood aggregation PNA algorithm to process the comprehensive feature of protein residues with a dimension of V x 200 and the edge index and edge feature corresponding to the protein structure graph, and finally outputs a feature of protein residues with a dimension of V x 200 after local message passing.

[0028] Preferably, the drug target binding affinity prediction model is obtained by training the following steps:

[0029] (4-1) Obtain a plurality of drug-protein data pairs to be predicted, preprocess all drug-protein pairs to be predicted to obtain a plurality of preprocessed drug-protein data pairs, which include the SMILES sequence of the drug, the amino acid sequence of the protein, and the drug target binding affinity data, and divide all preprocessed drug-protein data pairs into a training set and a test set according to a ratio of 4:1;

[0030] (4-2) For the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained in step (4-1), use the pre-trained ESM model to process the amino acid sequence to obtain the evolutionary language modeling representation (where each token is encoded as a 1280-dimensional vector) and the contact probability matrix of the amino acid sequence. The residue pairs in the contact probability matrix whose probabilities are greater than or equal to a preset contact probability threshold (in this application, the threshold is 0.5) are connected as edges in the protein structure graph, where the rows and columns of the contact probability matrix correspond to each residue in the amino acid sequence, and the values in the matrix represent the contact probability between the residue pairs. Thus, the protein structure graph corresponding to the amino acid sequence is obtained, where each node in the protein structure graph represents an amino acid, and the edges represent the contact between two amino acids. Each amino acid in the amino acid sequence is mapped to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of Vx33, where V is the length of the amino acid sequence.

[0031] (4-3) For the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained in step (4-1), use the RDKit tool to parse the SMILES sequence to obtain the molecular structure graph corresponding to the SMILES sequence. The atomic feature matrix is obtained according to the physicochemical properties of each atom in the molecular structure graph, and the atomic group set is generated according to the molecular structure graph corresponding to the obtained SMILES sequence, and the atomic group structure graph corresponding to the SMILES sequence is generated according to the connection relationship between the atomic groups in the atomic group set.

[0032] (4-4) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), input the protein structure graph, the evolutionary language modeling representation, and the residue feature matrix corresponding to the amino acid sequence obtained in step (4-2) into the first protein feature extraction layer in the protein encoder module in the drug target binding affinity prediction model, i.e., sequentially process the protein preprocessing layer, the first protein local message passing layer, the first structured state modeling layer, and the first protein post-processing layer to obtain the residue feature of the amino acid sequence corresponding to the first protein post-processing layer, which has a dimension of Vx200.

[0033] (4-5) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue feature corresponding to the amino acid sequence obtained in step (4-4) is input into the first residue cluster clustering layer in the protein encoder module of the drug target binding affinity prediction model, that is, the first residue cluster construction layer and the first residue cluster pooling layer are sequentially performed to obtain the residue cluster feature corresponding to the amino acid sequence, the dimensions of which are 32x5x200 and 32xV respectively m and the residue cluster assignment probability matrix

[0034] (4-6) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure graph, the atom group structure graph, and the atom feature matrix obtained in step (4-3) are input into the first drug feature extraction layer in the drug encoder module of the drug target binding affinity prediction model, that is, the drug preprocessing layer, the first drug local message passing layer, the first sequence context modeling layer, and the first drug post-processing layer are sequentially performed to obtain the atom feature corresponding to the SMILES sequence, the dimension of which is Nx200 after being processed by the first drug post-processing layer and the atom group feature corresponding to the SMILES sequence, the dimension of which is Cx200 after being processed by the first drug post-processing layer

[0035] (4-7) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the atom feature corresponding to the SMILES sequence obtained in step (4-6) is input into the first drug feature extraction layer in the drug encoder module of the drug target binding affinity prediction model, that is, the drug preprocessing layer, the first drug local message passing layer, the first sequence context modeling layer, and the first drug post-processing layer are sequentially performed to obtain the atom feature corresponding to the SMILES sequence, the dimension of which is Nx200 after being processed by the first drug post-processing layer and the atom group feature corresponding to the SMILES sequence, the dimension of which is Cx200 after being processed by the first drug post-processing layer and the drug molecular feature where 5 is the number of attention heads

[0036] ​​​​(4-8) For each drug-protein data pair in the training set obtained in step (4-1), the residue cluster features corresponding to the amino acid sequence obtained in step (4-5) have dimensions of 32x5x200 and 32xV m x5 and the residue cluster probability distribution matrix the atomic features corresponding to the SMILES sequence obtained in step (4-6) have dimensions of Nx200 after being processed by the first drug post-processing layer and the atom group features corresponding to the SMILES sequence obtained in step (4-7) have dimensions of Cx200 and 32x200 and the drug molecule features the first fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the drug molecule features corresponding to the SMILES sequence, which have dimensions of 32x1x200, 32xC m x200 and 32xN m x200 the atom group feature matrix and the atomic feature matrix

[0037] (4-9) For each amino acid sequence and the SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster features corresponding to the amino acid sequence obtained in step (4-5) have dimensions of 32x5x200 and the drug molecule features corresponding to the SMILES sequence obtained in step (4-8) have dimensions of 32x1x200, 32xC m x200 and 32xN m x200 the atom group feature matrix and the atomic feature matrix input into the first fusion layer of the feature fusion module in the drug target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and which have dimensions of 32x4xN m x5, 32x4xC m x5, 32x4x1x5, where 32 is the batch size, 4 is the number of attention heads, C m and N m are the maximum number of atom groups and atoms in each drug, respectively, and 5 is the number of residue clusters of the current multi-scale feature update layer;

[0038] (4-10) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the fusion attention score corresponding to the drug-protein data pair obtained in step (4-9) and the first fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the protein attention score corresponding to the amino acid sequence with dimensions of 32x5x4 and 160x200 respectively and residue cluster features and the group feature corresponding to the SMILES sequence with dimensions of Cx200, 32x200 and Nx200 respectively drug molecule features and atomic features

[0039] (4-11) For the SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the group feature corresponding to the SMILES sequence with dimensions of Cx200 obtained in step (4-10) and the atomic feature processed by the drug post-processing layer corresponding to the SMILES sequence with dimensions of Nx200 obtained in step (4-6) the first atomic feature update layer input into the drug target binding affinity prediction model to obtain the atomic feature processed by the first atomic feature update layer corresponding to the SMILES sequence with dimensions of Nx200

[0040] (4-12) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature corresponding to the amino acid sequence with dimensions of 160x200 obtained in step (4-10) and the residue feature processed by the first protein post-processing layer corresponding to the amino acid sequence with dimensions of Vx200 obtained in step (4-4) the first residue feature update layer input into the drug target binding affinity prediction model to obtain the residue feature processed by the first residue feature update layer corresponding to the amino acid sequence with dimensions of Vx200

[0041] (4-13) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure graph corresponding to the amino acid sequence obtained in step (4-2) and the residue feature of dimension V x 200 processed by the first residue feature update layer corresponding to the amino acid sequence obtained in step (4-12) The second protein feature extraction layer in the protein encoder module input into the drug target binding affinity prediction model, that is, sequentially performing the second protein local message passing layer, the second structured state modeling layer and the second protein post-processing layer to obtain the residue feature of dimension V x 200 processed by the second protein post-processing layer corresponding to the amino acid sequence

[0042] (4-14) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue feature of dimension V x 200 processed by the second protein post-processing layer corresponding to the amino acid sequence obtained in step (4-13) The second residue cluster clustering layer in the protein encoder module input into the drug target binding affinity prediction model, that is, sequentially performing the second residue cluster construction layer and the second residue cluster pooling layer to obtain the residue cluster feature of dimension 32 x 10 x 200 and 32 x V x 10 respectively m corresponding to the amino acid sequence And the residue cluster assignment probability matrix

[0043] (4-15) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure graph obtained in step (4-3), the atom group structure graph and the atom feature of dimension N x 200 processed by the first atom feature update layer corresponding to the SMILES sequence obtained in step (4-11) The second drug feature extraction layer in the drug encoder module input into the drug target binding affinity prediction model, that is, sequentially performing the second drug local message passing layer, the second sequence context modeling layer and the second drug post-processing layer to obtain the atom feature of dimension N x 200 processed by the second drug post-processing layer corresponding to the SMILES sequence And the atom group feature of dimension C x 200 processed by the second drug post-processing layer corresponding to the SMILES sequence

[0044] (4-16) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic features of dimension N×200 corresponding to the SMILES sequence obtained in step (4-15) after being processed by the second drug post-processing layer are The atomic cluster features of the second drug post-processing layer corresponding to the SMILES sequence with a dimension of C×200 The second multi-scale feature update layer is input into the drug-target binding affinity prediction model to obtain the atomic group attention scores corresponding to the SMILES sequence with dimensions of C×5, C×200 and 32×200 respectively. Cluster characteristics and drug molecular characteristics Where 5 is the number of attention heads;

[0045] (4-17) For each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-14) is converted to a matrix with dimensions of 32×10×200 and 32×V. m ×10 residue cluster features and the residue cluster probability assignment matrix The atomic features of the SMILES sequence obtained in step (4-15) with a dimension of N×200 after the second drug post-processing layer The atomic cluster features corresponding to the SMILES sequence obtained in step (4-16) with dimensions of C×200 and 32×200 respectively and drug molecular characteristics The second fusion layer of the feature fusion module in the drug-target binding affinity prediction model is input to obtain the corresponding SMILES sequence with dimensions of 32×1×200, 32×C m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix

[0046] (4-18) For each amino acid sequence in each drug-protein data pair and each drug SMILES sequence in the training set obtained in step (4-1), the residue cluster feature with a dimension of 32×5×200 corresponding to the amino acid sequence obtained in step (4-14) is converted into The dimensions of the SMILES sequence obtained in step (4-17) are 32×1×200 and 32×C respectively. m ×200 and 32×N mdrug molecule feature matrix of 32x200 atomic group feature matrix and atomic feature matrix the second fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the corresponding fusion attention score of the drug-protein data pair and with dimensions of 32x4xN m x10, 32x4xC m x10, 32x4x1x10

[0047] (4-19) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the fusion attention score corresponding to the drug-protein data pair obtained in step (4-18) and the second fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the corresponding protein attention score with dimensions of 32x10x4 and 320x200 respectively for the amino acid sequence and residue cluster feature atomic group feature with dimensions of Cx200, 32x200 and Nx200 respectively corresponding to the SMILES sequence drug molecule feature and atomic feature

[0048] (4-20) for the SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the atomic group feature with dimension Cx200 corresponding to the SMILES sequence obtained in step (4-19) and the atomic feature with dimension Nx200 corresponding to the SMILES sequence obtained in step (4-15) and processed by the second drug post-processing layer the second atomic feature update layer input into the drug target binding affinity prediction model to obtain the atomic feature with dimension Nx200 corresponding to the SMILES sequence and processed by the second atomic feature update layer

[0049] (4-21) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature with dimension 320x200 corresponding to the amino acid sequence obtained in step (4-19) the residue feature corresponding to the amino acid sequence obtained in step (4-13) and processed by the second protein post-processing layer, with a dimension of V x 200 the residue feature corresponding to the amino acid sequence obtained in step (4-13) and processed by the second residue feature updating layer, with a dimension of V x 200

[0050] (4-22) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure graph corresponding to the amino acid sequence obtained in step (4-2) and the residue feature corresponding to the amino acid sequence obtained in step (4-21) and processed by the second residue feature updating layer, with a dimension of V x 200 the residue feature corresponding to the amino acid sequence obtained in step (4-13) and processed by the second protein post-processing layer, with a dimension of V x 200

[0051] (4-23) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue feature corresponding to the amino acid sequence obtained in step (4-22) and processed by the third protein post-processing layer, with a dimension of V x 200 the residue cluster feature corresponding to the amino acid sequence, with a dimension of 32 x 20 x 200 and 32 x V x 20 m respectively, obtained in step (4-23) and processed by the third residue cluster clustering layer and the residue cluster assignment probability matrix

[0052] (4-24) for each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure graph obtained in step (4-3), the radical structure graph, and the atom feature corresponding to the SMILES sequence obtained in step (4-20) and processed by the second atom feature updating layer, with a dimension of N x 200 The third drug feature extraction layer in the drug encoder module in the drug-target binding affinity prediction model is input, that is, the third drug local message passing layer, the third sequence context modeling layer and the third drug post-processing layer are sequentially performed to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the third drug post-processing layer. And the atomic cluster characteristics of the SMILES sequence with a dimension of C×200 after the third drug post-processing layer

[0053] (4-25) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic features of dimension N×200 corresponding to the SMILES sequence obtained in step (4-24) after processing by the third drug post-processing layer are The atomic cluster features of the SMILES sequence with a dimension of C×200 after the third drug post-processing layer are shown. Input to the third multi-scale feature update layer in the drug-target binding affinity prediction model to obtain the atomic group attention scores corresponding to the SMILES sequence with dimensions of C×5, C×200 and 32×200 respectively Cluster characteristics and drug molecular characteristics Where 5 is the number of attention heads;

[0054] (4-26) For each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-23) has dimensions of 32×20×200 and 32×V m ×20 residue cluster features and the residue cluster probability assignment matrix The atomic features of the SMILES sequence obtained in step (4-24) with a dimension of N×200 after the third drug post-processing layer The atomic cluster features corresponding to the SMILES sequence obtained in step (4-25) with dimensions of C×200 and 32×200 respectively and drug molecular characteristics The third fusion layer of the feature fusion module in the drug-target binding affinity prediction model is input to obtain the corresponding SMILES sequence with dimensions of 32×1×200, 32×C m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix

[0055] (4-27) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the residue cluster feature of dimension 32x20x200 corresponding to the amino acid sequence obtained in step (4-23) and the drug molecule feature matrix of dimensions 32x1x200, 32xC m x200 and 32xN m x200 corresponding to the SMILES sequence obtained in step (4-26) atomic group feature matrix and atomic feature matrix input into the third fusion layer of the feature fusion module in the drug target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and of dimensions 32x4xN m x20, 32x4xC m x20, 32x4x1x20, respectively;

[0056] (4-28) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the fusion attention score of the drug-protein data pair corresponding to the SMILES sequence obtained in step (4-27) and input into the third fusion layer of the feature fusion module in the drug target binding affinity prediction model to obtain the protein attention score of dimensions 32x20x4 and 640x200 corresponding to the amino acid sequence and residue cluster feature and the atomic group feature of dimensions Cx200, 32x200 and Nx200 corresponding to the SMILES sequence drug molecule feature and atomic feature

[0057] (4-29) For the SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the atomic group feature of dimension Cx200 corresponding to the SMILES sequence obtained in step (4-28) and the atomic feature of dimension Nx200 corresponding to the SMILES sequence obtained in step (4-24) after being processed by the third drug post-processing layer a third atom feature update layer input into the drug target binding affinity prediction model to obtain atom features corresponding to the SMILES sequence after being processed by the third atom feature update layer, with a dimension of N x 200

[0058] (4-30) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature corresponding to the amino acid sequence obtained in step (4-28) has a dimension of 640 x 200 and the residue feature corresponding to the amino acid sequence obtained in step (4-22) has a dimension of V x 200 after being processed by the third protein post-processing layer a third residue feature update layer input into the drug target binding affinity prediction model to obtain residue features corresponding to the amino acid sequence after being processed by the third residue feature update layer, with a dimension of V x 200

[0059] (4-31) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster assignment probability matrix corresponding to the amino acid sequence obtained in step (4-5) has a dimension of 32 x V m x 5 the protein attention score corresponding to the amino acid sequence obtained in step (4-10) has a dimension of 32 x 5 x 4 the residue cluster assignment probability matrix corresponding to the amino acid sequence obtained in step (4-14) has a dimension of 32 x V m x 10 the protein attention score corresponding to the amino acid sequence obtained in step (4-19) has a dimension of 32 x 10 x 4 the residue cluster assignment probability matrix corresponding to the amino acid sequence obtained in step (4-23) has a dimension of 32 x V m x 20 the protein attention score corresponding to the amino acid sequence obtained in step (4-28) has a dimension of 32 x 20 x 4 and the residue features corresponding to the amino acid sequence obtained in step (4-30) have a dimension of V x 200 after being processed by the third residue feature update layer a protein attention pooling layer input into the drug target binding affinity prediction model to obtain a final protein feature vector corresponding to the amino acid sequence, with a dimension of 32 x 200

[0060] (4-32) for each drug in each drug-protein data pair in the training set obtained in step (4-1), the group attention score corresponding to the SMILES sequence of the drug obtained in step (4-7) is Cx5 the group attention score corresponding to the SMILES sequence obtained in step (4-16) is Cx5 the group attention score corresponding to the SMILES sequence obtained in step (4-25) is Cx5 and the atom feature processed by the third atom feature update layer corresponding to the SMILES sequence obtained in step (4-13) is Nx200 the final drug feature vector corresponding to the SMILES sequence obtained by inputting the drug attention pooling layer in the drug target binding affinity prediction model is 32x200

[0061] (4-33) for each drug-protein data pair in the training set obtained in step (4-1), the final protein feature vector corresponding to the amino acid sequence output in step (4-31) is 32x200 and the final drug feature vector corresponding to the SMILES sequence output in step (4-32) is 32x200 input into the multi-layer perception module in the drug target binding affinity prediction model, the final protein feature vector and the final drug feature vector are spliced to obtain a spliced vector with a dimension of 32x400, and the spliced vector is processed by a two-layer MLP to obtain a drug target binding affinity prediction value corresponding to the drug-protein data pair

[0062] (4-34) for each drug-protein data pair in the training set obtained in step (4-1), calculate the loss function according to the drug target binding affinity data corresponding to the drug-protein data pair obtained in step (4-1) and the drug target binding affinity prediction value corresponding to the drug-protein data pair obtained in step (4-33) y, and train the drug target binding affinity prediction model using the loss function until the drug target binding affinity prediction model converges, thereby obtaining a preliminarily trained drug target binding affinity prediction model;

[0063] (4-35) For each drug-protein data pair in the training set obtained in step (4-1), the drug target binding affinity prediction model preliminarily trained in step (4-34) is tested using the test set obtained in step (4-1) until the value of the loss function reaches the optimum, thereby obtaining the trained drug target binding affinity prediction model.

[0064] According to another aspect of the present application, a drug target binding affinity prediction system based on multi-scale feature fusion is provided, comprising:

[0065] A first module is configured to obtain a drug-protein data pair to be predicted, and pre-process the drug-protein data pair to be predicted to obtain a pre-processed drug-protein data pair, which includes a simplified molecular linear input specification (SMILES) sequence of a drug, an amino acid sequence of a protein, and drug target binding affinity data.

[0066] A second module is configured to, for each amino acid sequence in the pre-processed drug-protein data pair obtained by the first module, process the amino acid sequence using a pre-trained evolutionary scale model (ESM) to obtain an evolutionary language modeling representation and a contact probability matrix of the amino acid sequence, connect residue pairs with a contact probability greater than or equal to a preset contact probability threshold in the contact probability matrix as edges in a protein structure graph, thereby obtaining a protein structure graph corresponding to the amino acid sequence, wherein each node in the protein structure graph represents an amino acid, and the edges represent the contact between two amino acids; and map each amino acid in the amino acid sequence to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of length of the amino acid sequence x 33.

[0067] A third module is configured to, for each SMILES sequence of a drug in the pre-processed drug-protein data pair obtained by the first module, parse the SMILES sequence using an RDKit tool to obtain a molecular structure graph corresponding to the SMILES sequence, obtain an atomic feature matrix according to the physicochemical properties of each atom in the molecular structure graph, generate a set of atom groups according to the molecular structure graph corresponding to the obtained SMILES sequence, and generate an atom group structure graph corresponding to the SMILES sequence according to the connection relationship between the atom groups in the set of atom groups.

[0068] A fourth module is configured to input the protein structure graph and the residue feature matrix corresponding to each amino acid sequence obtained by the second module, and the molecular structure graph, the atom group structure graph, and the atomic feature matrix corresponding to each SMILES sequence of a drug obtained by the third module into a pre-trained drug target binding affinity prediction model to obtain a drug target binding affinity prediction value.

[0069] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0070] (1) The present method uses step (1) in the data preprocessing stage, which only relies on the SMILES sequence of the drug and the amino acid sequence information of the protein, thereby eliminating the dependence on protein structure data, avoiding the high cost of structure determination and the uncertainty of structure prediction, and significantly improving the universality and scalability of the method;

[0071] (2) The present application uses steps (2) to (4), which proposes an efficient sequence-based feature representation and interaction modeling mechanism, reduces the learning of noise and error patterns by the model, and improves the prediction accuracy; Specifically, the potential semantic features of drug molecules and protein sequences are automatically extracted through a multi-layer neural network structure, and the nonlinear interaction relationship is modeled, thereby effectively avoiding the dependence on artificial feature engineering in traditional methods, significantly improving the model development efficiency and generalization ability; In addition, the present application makes full use of the advantages of deep neural networks in representation learning, even in the case of limited and unevenly distributed biological data, the potential laws of drug-target interaction can still be accurately mined, thereby achieving higher prediction accuracy; In particular, combining the extraction ability of convolutional neural network (CNN) for local structure patterns and the modeling ability of attention mechanism for global association, the model can automatically identify the key functional domains in the drug molecular structure and protein amino acid sequence and their interaction feature regions, providing more accurate and interpretable prediction results for drug target binding affinity prediction;

[0072] (3) The present application constructs a protein structure graph and a residue feature matrix based on the amino acid sequence in step (4-2), and combines structured state space modeling and local message passing mechanism of graph neural network in steps (4-4), (4-13) and (4-22), to systematically integrate the local conformation information and global topological structure perception ability of the protein. This design effectively alleviates the technical problem of insufficient modeling of protein spatial structure in traditional methods;

[0073] (4) The present application constructs a molecular structure graph, an atom group structure graph and an atom feature matrix based on the SMILES sequence of the drug in step (4-3), and extracts drug sub-features by context sequence modeling and local message passing in steps (4-6), (4-15) and (4-24), taking into account the local structure information of the drug, providing a global receptive field and modeling the atom, atom group and molecular features of the drug in multiple scales, this design can effectively depict the key structure sites in the interaction between the drug and the protein, and improve the adaptability and generalization ability of the model in downstream tasks;

[0074] (5) The present application designs a multi-head bilinear cross-attention mechanism in steps (4-8) to (4-10), (4-17) to (4-19) and (4-26) to (4-28), respectively models the interaction between the atomic, atomic group and molecular characteristics of the drug and the protein residue cluster characteristics, realizes the ability to capture the potential interaction area from different scales, and significantly improves the recognition ability of the model to the drug-target binding relationship. The mechanism effectively solves the problem of lacking an efficient feature fusion scheme in the existing method, and improves the precision and interpretability of the drug-target binding affinity prediction. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 is the overall flowchart of the drug-target binding affinity prediction method of the present application based on multi-scale feature fusion;

[0076] Figure 2 is the network structure diagram of the drug-target binding affinity prediction model used in the method of the present application;

[0077] Figure 3 is the detailed working flow diagram of the drug-target binding affinity prediction model used in the method of the present application. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0079] The application provides a drug target binding affinity prediction method based on multi-scale feature fusion.

[0080] To make the technical solutions and advantages of the present application clearer, the following further describes the whole drug target binding affinity prediction method in detail in combination with the drawings.

[0081] As shown in Figure 1 The application provides a drug target binding affinity prediction method based on multi-scale feature fusion, which comprises the following steps:

[0082] (1) Obtain a drug-protein data pair to be predicted, and pre-process the drug-protein data pair to be predicted to obtain a pre-processed drug-protein data pair, which comprises a simplified molecular input line entry system (SMILES) sequence of a drug, an amino acid sequence of a protein, and drug target binding affinity data.

[0083] Specifically, first, the drug-protein data pair to be predicted is screened to delete data with missing data to obtain the SMILES sequence of the drug, the amino acid sequence of the protein, and the inhibition constant and dissociation constant of the drug-protein interaction; then, the drug target binding affinity data is obtained according to the obtained inhibition constant and dissociation constant, respectively; wherein the calculation formula of the drug target binding affinity data is:

[0084]

[0085] Wherein, K i and Kd respectively represent inhibition constant and dissociation constant, pK i represent drug target binding affinity data calculated based on inhibition constant, pK d represent drug target binding affinity data calculated based on dissociation constant.

[0086] Finally, the obtained SMILES sequence of the drug, the amino acid sequence of the protein, and the drug target binding affinity data are combined into the preprocessed drug-protein data pair.

[0087] The advantage of this step (1) is that it can map the excessively large or small inhibition constant and dissociation constant to an appropriate size of affinity value, which is convenient for model training.

[0088] (2) For the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained in step (1), a pre-trained evolutionary scale model (ESM) (in this application, it is esm2_t33_650M_UR50D) is used to process the amino acid sequence to obtain an evolutionary language modeling representation (where each token is encoded as a 1280-dimensional vector) and a contact probability matrix of the amino acid sequence. The residue pairs in the contact probability matrix with a probability greater than or equal to a preset contact probability threshold (in this application, the threshold is 0.5) are connected as edges in the protein structure graph, so as to obtain the protein structure graph corresponding to the amino acid sequence. Each node in the protein structure graph represents an amino acid, and the edge represents the contact between two amino acids; and each amino acid in the amino acid sequence is mapped to a fixed-dimensional feature vector (in this application, it is equal to the sum of 12-dimensional physicochemical properties and 21-dimensional amino acid types, i.e. 33). The feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of length of the amino acid sequence x 33.

[0089] (3) For the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained in step (1), the RDKit tool (a widely used open source toolkit in the field of chemical information) is used to parse the SMILES sequence to obtain the molecular structure graph corresponding to the SMILES sequence. According to the physicochemical properties of each atom in the molecular structure graph, an atomic feature matrix is obtained, and according to the obtained molecular structure graph corresponding to the SMILES sequence, an atomic group set is generated, and according to the connection relationship between the atomic groups in the atomic group set, an atomic group structure graph corresponding to the SMILES sequence is generated.

[0090] The step specifically comprises the following steps. First, the RDKit tool is used to parse each SMILES sequence to obtain a corresponding molecular structure graph, wherein each node in the molecular structure graph represents an atom, and an edge represents a covalent bond between atoms. A vector information containing 43-dimensional structural atomic features is extracted for each atom, mainly including the connectivity of the atom, the number of hydrogen atoms, the hybrid orbital type, the aromaticity, the chiral center marker, etc. The vector information corresponding to all atoms in the molecular structure graph forms an atomic feature matrix with a dimension of N x 43 (N is the number of nodes in the molecular structure graph); then, a tree decomposition (Junction Tree Decomposition) operation is performed based on the bond connection relationship between the atoms in the molecular structure graph to generate an atomic group set; finally, a tree-like connection structure is constructed according to the connection relationship between the atomic groups in the atomic group set, thereby forming an atomic group structure graph with structural connectivity (the purpose is to further mine the hierarchical information between high-order structural units such as ring structures or functional groups in the molecule), which contains the following information: the connection relationship between the atomic groups, the mapping relationship of the atoms to the atomic groups to which they belong, the total number of atomic groups, and the type code of each atomic group.

[0091] The step (3) has the advantages that by parsing the SMILES sequence of the drug into a molecular structure graph and an atomic group structure graph, a multi-scale substructure representation of the drug is initially established, the structural hierarchical features from the atomic level, the atomic group level to the molecular level are effectively captured, a structured input basis is provided for subsequent multi-scale fusion and attention mechanism, and thus the recognition ability of the model for the potential drug-target binding mode is enhanced.

[0092] (4) The protein structure graph and the residue feature matrix corresponding to each amino acid sequence obtained in step (2) and the molecular structure graph, the atomic group structure graph and the atomic feature matrix corresponding to each drug SMILES sequence obtained in step (3) are input into the pre-trained drug-target binding affinity prediction model to obtain a drug-target binding affinity prediction value.

[0093] As shown in Figure 2 and Figure 3 , the drug-target binding affinity prediction model of the application comprises a protein encoder module, a drug encoder module, a feature fusion module, and a multi-layer perception module.

[0094] The protein encoder module comprises a protein feature extraction layer, a residue cluster clustering layer, and a protein pooling layer.

[0095] The protein feature extraction layer is implemented based on Mamba and Message Passing Neural Networks (MPNN), and includes a protein preprocessing layer, three protein local message passing layers, three structured state modeling layers, and three protein postprocessing layers.

[0096] The residue cluster clustering layer includes three residue cluster construction layers and three residue cluster pooling layers.

[0097] The protein pooling layer includes three residue feature update layers and a protein attention pooling layer.

[0098] The drug encoder module includes a drug feature extraction layer, three multi-scale feature update layers, and a drug pooling layer.

[0099] The drug feature extraction layer is implemented based on Bidirectional Long Short-Term Memory (BiLSTM) and MPNN, and includes a drug preprocessing layer, three drug local message passing layers, three sequence context modeling layers, and three drug postprocessing layers.

[0100] The drug pooling layer includes three atom feature update layers and a molecule attention pooling layer.

[0101] The feature fusion module includes three fusion layers.

[0102] The input of the protein pre-processing layer is the evolutionary language modeling representation (with a dimension of 1280) of multiple amino acid sequences (the number of which is equal to the batch size in the present application, i.e., 32), a residue feature matrix (with a dimension of 33) and a contact probability matrix; the protein pre-processing layer first processes the input contact probability matrix using a radial basis function to obtain an edge feature matrix with a size of E x 200 (where E represents the number of edges in the protein structure graph), and then processes the evolutionary language modeling representation through a multilayer perceptron (MLP) (where the hidden layer uses a rectified linear unit (ReLU) as an activation function, and the output layer uses layer normalization, wherein the input of the first layer is the evolutionary language modeling representation with a dimension of V x 1280, the output is an intermediate feature vector with a dimension of V x 400, the input of the second layer is the intermediate feature vector with a dimension of V x 400, and the final output is the pre-processed evolutionary language modeling representation with a dimension of V x 200, where V represents the length of the amino acid sequence). Subsequently, the residue feature matrix is processed through a two-layer MLP (the hidden layer uses a ReLU as an activation function, and the output layer uses layer normalization, wherein the input of the first layer is the residue feature matrix with a dimension of V x 33, the output is an intermediate feature vector with a dimension of V x 400, the input of the second layer is the intermediate feature vector with a dimension of V x 400, and the output is the pre-processed residue feature matrix with a dimension of V x 200), to obtain the pre-processed residue feature matrix with a dimension of V x 200; finally, the pre-processed evolutionary language modeling representation with a dimension of V x 200 and the pre-processed residue feature matrix with a dimension of V x 200 are added, and the final output is a comprehensive feature of protein residues with a dimension of V x 200.

[0103] The input of the first protein local message passing layer is the comprehensive feature of protein residues with a dimension of V x 200 output by the protein pre-processing layer and the protein structure graph, and the protein local message passing layer processes the comprehensive feature of protein residues with a dimension of V x 200 and the edge index and edge feature corresponding to the protein structure graph using a principal neighborhood aggregation (PNA) algorithm, and finally outputs the feature of protein residues with a dimension of V x 200 after local message passing.

[0104] The input of the first structured state modeling layer is the integrated features of protein residues with a dimension of Vx200 obtained by the protein preprocessing layer and the protein structure graph. The structured state modeling layer sorts the nodes of the protein structure graph using a degree-based and batch-based node sorting strategy, processes the integrated features of protein residues through the Mamba model, and finally outputs the features of protein residues processed by global attention with a dimension of Vx200.

[0105] The input of the first protein post-processing layer is the features of protein residues processed by local message passing with a dimension of Vx200 output by the first protein local message passing layer and the features of protein residues processed by global attention with a dimension of Vx200 output by the first structured state modeling layer. The protein post-processing layer first merges the two input features to obtain a merged feature vector of protein residues with a dimension of Vx200. Then, the merged feature vector of protein residues is processed through a two-layer MLP (the hidden layer uses ReLU as the activation function, and the output layer uses graph normalization, where the input of the first layer is the merged feature vector of protein residues with a dimension of Vx200, and the output is an intermediate feature vector of protein residues with a dimension of Vx400; the input of the second layer is the intermediate feature vector of protein residues with a dimension of Vx400, and the output is the features of protein residues processed by the first protein post-processing layer with a dimension of Vx200).

[0106] The input of the first residue cluster construction layer is the features of protein residues processed by the first protein post-processing layer with a dimension of Vx200 and the protein structure graph. The residue cluster construction layer first processes the features through a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of Vx5 (where 5 is the preset cluster number of this layer). Then, the intermediate residue cluster assignment probability matrix is subjected to softmax operation and batch padding conversion according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32xV m x5 (where V m represents the maximum length in all input amino acid sequences).

[0107] The input of the first residue cluster pooling layer is the residue cluster assignment probability matrix with a dimension of 32xV ma residue cluster assignment probability matrix of size 5x5, and a feature of size Vx200 of the protein residues after being processed by a first protein post-processing layer, the residue cluster pooling layer first uses a dense mincut pooling algorithm to perform soft clustering on all residues in the feature to obtain a probability distribution of each residue belonging to different clusters as a residue cluster feature, and finally outputs a residue cluster feature of size 32x5x200.

[0108] The input of the drug preprocessing layer includes a molecular structure graph, an atomic feature matrix of size Nx43, and an atom group structure graph, the drug preprocessing layer first processes the atomic feature matrix by an MLP to obtain an atomic feature of size Nx200, and then processes the atom group structure graph by an MLP to obtain an atom group feature of size Cx200, where C represents the number of nodes in the atom group structure graph.

[0109] The input of the first drug local message passing layer is the atomic feature of size Nx200 obtained by the drug preprocessing layer and the molecular structure graph, the drug local message passing layer uses a PNA algorithm to process the atomic feature, and edge indexes and edge features corresponding to the molecular structure graph, and finally outputs a drug atom feature of size Nx200 after local message passing.

[0110] The input of the first sequence context modeling layer is the atomic feature of size Nx200 obtained by the drug preprocessing layer, the sequence context modeling layer uses a BiLSTM algorithm to process the atomic feature, and finally outputs an atomic feature of size Nx200 after sequence context modeling.

[0111] The input of the first drug post-processing layer is the drug atom feature of size Nx200 after local message passing output by the first drug local message passing layer and the atomic feature of size Nx200 after sequence context modeling output by the first sequence context modeling layer, the drug post-processing layer first splices the two input features to obtain a merged feature vector of size Nx400 of the atoms, then reduces the merged feature vector to Nx200 by a linear layer, and normalizes the reduced merged feature vector, and finally obtains an atomic feature of size Nx200.

[0112] The input of the first multi-scale feature update layer is the atomic feature with the dimension of N x 200 obtained by the first drug post-processing layer, the group feature with the dimension of C x 200 obtained by the drug preprocessing layer, the molecular structure graph and the group structure graph, the multi-scale feature update layer performs residual processing on the group feature to obtain the updated group feature with the dimension of C x 200 (firstly, the atomic feature is averaged and aggregated into the corresponding group based on the index, and then linearly transformed and activated by the ReLU function to obtain the group feature with the dimension of C x 200, and then the updated group feature is processed by the multi-head attention mechanism (the updated group feature is split into 5 attention heads, the feature of each attention head is processed by MLP to obtain the score, and then the dimension of the intermediate group attention score is C x 5, then dropout is applied to prevent overfitting, and the group attention score is converted to C x 5 by the softmax function. Then the group attention score and the group feature are multiplied element by element, and the dimension of the comprehensive feature of the drug molecule is 32 x 200 after weighted summation according to the group batch index), to obtain the drug molecule feature with the dimension of 32 x 200 and the group attention score with the dimension of C x 5.

[0113] The input of the first fusion layer is the residue cluster feature with the dimension of 32 x 5 x 200 output by the first residue cluster pooling layer, the group feature with the dimension of C x 200 and the drug molecule feature with the dimension of 32 x 200 output by the first multi-scale feature update layer, and the atomic feature with the dimension of N x 200 obtained by the first drug post-processing layer. The fusion layer performs structured preprocessing on the input drug molecule feature, group feature and atomic feature to obtain the drug molecule feature with the dimension of 32 x 1 x 200, the group feature with the dimension of 32 x C m x 200 and the atomic feature with the dimension of 32 x N m x 200, wherein C m and N m are the maximum number of groups and the number of atoms in all drugs respectively; then, the residue cluster feature is interacted with the drug molecule feature with the dimension of 32 x 1 x 200, the group feature with the dimension of 32 x C m x 200 and the atomic feature with the dimension of 32 x N m x 200 by the multi-head bilinear cross attention to obtain three fusion attention scores with the dimensions of 32 x 4 x 1 x 5, 32 x 4 x C m x 5 and 32 x 4 x N mX5 (wherein 32 is batch size, 4 is the number of attention heads, 5 is the number of the current multi-scale feature update layer residual cluster); then, the fusion attention score with the dimension of 32x4x1x5 is reshaped and the dimension is replaced (the fusion attention score with the dimension of 32x4x1x5 is reshaped to 32x4x5, and then the second dimension and the third dimension are replaced to obtain the protein attention score with the dimension of 32x5x4, which represents the score of the residual cluster to the drug feature under 4 attention heads), to obtain the protein attention score with the dimension of 32x5x4; then, the atom group feature Cx200, the drug molecule feature 32x200, and the atom feature Nx200 obtained by the first drug post-processing layer are updated by attention according to the three fusion attention scores (the third dimension of the three fusion attention scores is respectively subjected to softmax, which represents which substructure of the drug molecule is more concerned by the residual cluster, and is used to update the feature of the residual cluster; the fourth dimension of the three fusion attention scores is respectively subjected to softmax, which represents which residual cluster is more concerned by the drug molecule multi-scale feature, and is used to update the drug molecule multi-scale feature, then the output of each attention head is averaged and merged for the batch to obtain the residual cluster feature 160x200, the atom group feature Cx200, the drug molecule feature 32x200, and the atom feature Nx200), to obtain the protein attention score with the dimension of 32x5x4, the residual cluster feature 160x200, the atom group feature Cx200, the drug molecule feature 32x200, and the atom feature Nx200 output by the first fusion layer.

[0114] The input of the first residue feature update layer is the residual cluster feature with the dimension of 160x200 processed by the first fusion layer (wherein 160 is the batch size 32 multiplied by the number of residual clusters in the batch 5) and the residue feature Vx200 output by the first protein feature extraction layer, and the residue feature update layer performs mean aggregation processing on the residue feature (first, all the residual cluster features related to each residue are aggregated to the residue feature by averaging, the aggregated residue feature is connected with the original residue feature by residual connection, and finally dropout is applied and graph normalization is performed to obtain the residue feature with the dimension of Vx200 processed by the residue feature update layer), to obtain the residue feature matrix with the dimension of Vx200 processed by the residue feature update layer.

[0115] The input of the second protein local message passing layer is the residue feature matrix of dimension V×200 and the protein structure graph processed by the first residue feature update layer. The protein local message passing layer uses the PNA algorithm to process the edge index and edge features corresponding to the residue feature matrix and the protein structure graph, and finally outputs the features of the protein residues after local message passing with a dimension of V×200.

[0116] The input of the second structured state modeling layer is the residue feature matrix of dimension V×200 processed by the first residue feature update layer and the protein structure graph. The structured state modeling layer uses a degree- and batch-based node sorting strategy to sort all nodes in the protein structure graph, and processes the residue feature matrix through the Mamba model, and finally outputs the features of protein residues with dimension V×200 after global attention processing.

[0117] The input of the second protein post-processing layer is the V×200 features of protein residues after local message passing output by the second protein local message passing layer and the V×200 features of protein residues after global attention processing output by the second structured state modeling layer. This protein post-processing layer first merges the two features to obtain a merged feature vector of protein residues with a dimension of V×200. Subsequently, the merged feature vector of protein residues is processed by a two-layer MLP (the hidden layer uses ReLU as the activation function and the output layer uses graph normalization, where the first layer input is the merged feature vector of protein residues with a dimension of V×400 and the output is the intermediate feature vector of protein residues with a dimension of V×400. The second layer input is the intermediate feature vector of protein residues with a dimension of V×400 and the output is the features of protein residues with a dimension of V×200 after being processed by the second protein post-processing layer) to obtain the features of protein residues with a dimension of V×200 after being processed by the second protein post-processing layer.

[0118] The input of the second residue clustering layer is the features of protein residues processed by the second protein post-processing layer and the protein structure graph with a dimension of V×200. The residue clustering layer first processes the features with a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of V×10 (where 10 is the number of clusters preset in this layer). Then, the intermediate residue cluster assignment probability matrix is ​​subjected to a softmax operation and a batch filling conversion according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32×V m ×10.

[0119] The input of the third residue cluster pooling layer is the output of the second residue cluster construction layer with a dimension of 32×V ma residue cluster assignment probability matrix of size 10x10, and a feature of size Vx200 of the protein residues after being processed by a second protein post-processing layer, the residue cluster pooling layer using a dense mincut pooling algorithm to soft cluster all residues in the feature to obtain a probability distribution of each residue belonging to different clusters as residue cluster features, and finally outputting residue cluster features of size 32x10x200.

[0120] The input of the first atom feature update layer is the Cx200 atom group features processed by the first multi-scale feature update layer after being processed by the first fusion layer and the Nx200 atom features output by the first drug feature extraction layer, the atom feature update layer performing mean aggregation processing on the Nx200 atom features output by the first drug feature extraction layer (firstly, aggregating all atom group features related to each atom to the atom features, then performing residual connection with the original atom features after a linear layer, finally applying dropout and performing graph normalization to obtain updated atom features of size Nx200), to obtain an atom feature matrix of size Nx200 processed by the atom feature update layer.

[0121] The input of the second drug local message passing layer is the atom feature matrix processed by the first atom feature update layer and the molecular structure graph, the drug local message passing layer using a PNA algorithm to process the atom features and the edge index and edge features corresponding to the molecular structure graph, and finally outputting drug atom features of size Nx200 after being processed by the second local message passing.

[0122] The input of the second sequence context modeling layer is the atom feature matrix of size Nx200 processed by the first atom feature update layer, the sequence context modeling layer using a BiLSTM algorithm to process the atom feature matrix, and finally outputting atom features of size Nx200 after being processed by the sequence context modeling.

[0123] The input of the second drug post-processing layer is the atom features of size Nx200 of the drug atoms after being processed by the local message passing output by the second drug local message passing layer and the atom features of size Nx200 after being processed by the sequence context modeling output by the second sequence context modeling layer, the drug post-processing layer first merging the two input features to obtain a merged feature vector of size Nx400 atoms, then reducing the dimension of the merged feature vector to Nx200 through a linear layer, and normalizing the reduced merged feature vector, and finally obtaining atom features of size Nx200.

[0124] The inputs to the second multi-scale feature update layer are the N×200 atomic features obtained by the second drug post-processing layer, the C×200 atomic cluster features obtained by the drug pre-processing layer, the molecular structure graph, and the atomic cluster structure graph. This multi-scale feature update layer first performs residual processing on the atomic cluster features to obtain updated atomic cluster features of dimension C×200. Then, the updated atomic cluster features are processed through a multi-head attention mechanism to obtain drug molecular features of dimension 32×200 and atomic cluster attention scores of dimension C×5.

[0125] The input of the second fusion layer is the residue cluster features with a dimension of 32×10×200 output by the second residue cluster pooling layer, the atomic group features with a dimension of C×200 and the drug molecule features with a dimension of 32×200 obtained by the second multi-scale feature update layer, and the atomic features with a dimension of N×200 obtained by the second drug post-processing layer. The fusion layer performs structured preprocessing on the input drug molecule features, atomic group features and atomic features to obtain drug molecule features with a dimension of 32×10×200, drug molecule features with a dimension of 32×C×200 and drug molecule features with a dimension of 32×10×200. m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features; then, the residue cluster features are respectively combined with the drug molecular features of dimension 32×1×200 and the 32×C m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features to obtain three fusion attention scores with dimensions of 32×4×1×10, 32×4×C m ×10, 32×4×N m ×10 (where 10 is the number of residue clusters in the current multi-scale feature update layer); then, the fusion attention score with a dimension of 32×4×1×10 is reshaped and permuted to obtain a protein attention score with a dimension of 32×10×4. Then, according to the three fusion attention scores, the residue cluster features with a dimension of 32×10×200 output by the second residue cluster pooling layer and the atomic group features of C×200 output by the first multi-scale feature update layer, the drug molecule features of 32×200, and the atomic features of N×200 obtained by the second drug post-processing layer are updated to obtain the protein attention score with a dimension of 32×10×4 output by the first fusion layer, the residue cluster features of 320×200, the atomic group features of C×200, the drug molecule features of 32×200, and the atomic features of N×200.

[0126] The input of the second residue feature update layer is the residue cluster feature with the dimension of 320x200 processed by the second fusion layer (where 320 is the batch size 32 multiplied by the number of residue clusters within the batch 10) and the residue feature with the dimension of Vx200 output by the second protein feature extraction layer, and the residue feature update layer performs mean aggregation processing on the residue feature to obtain the residue feature matrix with the dimension of Vx200 processed by the residue feature update layer.

[0127] The input of the third protein local message passing layer is the residue feature matrix with the dimension of Vx200 processed by the second residue feature update layer and the protein structure graph, and the protein local message passing layer uses the PNA algorithm to process the residue feature matrix with the dimension of Vx200 and the edge index and edge feature corresponding to the protein structure graph, and finally outputs the feature of the protein residue after local message passing with the dimension of Vx200.

[0128] The input of the third structured state modeling layer is the residue feature matrix with the dimension of Vx200 processed by the second residue feature update layer and the protein structure graph, and the structured state modeling layer uses the node ordering strategy based on degree and batch to order the nodes of the protein structure graph and processes the residue feature matrix through the Mamba model, and finally outputs the feature of the protein residue after global attention processing with the dimension of Vx200.

[0129] The input of the third protein post-processing layer is the feature of the protein residue after local message passing with the dimension of Vx200 output by the third protein local message passing layer and the feature of the protein residue after global attention processing with the dimension of Vx200 output by the third structured state modeling layer. The protein post-processing layer first merges the two input features to obtain the merged feature vector of the protein residue with the dimension of Vx200, and then processes the merged feature vector of the protein residue through a two-layer MLP (the hidden layer uses ReLU as the activation function, and the output layer uses graph normalization, where the input of the first layer is the merged feature vector of the protein residue with the dimension of Vx200, the output is the intermediate feature vector of the protein residue with the dimension of Vx400, the input of the second layer is the intermediate feature vector of the protein residue with the dimension of Vx400, and the output is the feature of the protein residue after processing by the second protein post-processing layer with the dimension of Vx200).

[0130] The input of the third residue cluster construction layer is the feature of the protein residues after being processed by the third protein post-processing layer and the protein structure graph, which has a dimension of Vx200. The residue cluster construction layer first processes the feature by a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of Vx20 (where 20 is the preset number of clusters of the layer), and then performs a softmax operation and a batch padding conversion on the intermediate residue cluster assignment probability matrix according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32xVx20. m

[0131] The input of the third residue cluster construction layer is the feature of the protein residues after being processed by the third protein post-processing layer and the protein structure graph, which has a dimension of Vx200. The residue cluster construction layer first processes the feature by a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of Vx20 (where 20 is the preset number of clusters of the layer), and then performs a softmax operation and a batch padding conversion on the intermediate residue cluster assignment probability matrix according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32xVx20. m The input of the third residue cluster construction layer is the feature of the protein residues after being processed by the third protein post-processing layer and the protein structure graph, which has a dimension of Vx200. The residue cluster construction layer first processes the feature by a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of Vx20 (where 20 is the preset number of clusters of the layer), and then performs a softmax operation and a batch padding conversion on the intermediate residue cluster assignment probability matrix according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32xVx20.

[0132] The input of the second atom feature update layer is the atom group feature of Cx200 processed by the second multi-scale feature update layer through the second fusion layer and the atom feature of Nx200 output by the second drug feature extraction layer. The atom feature update layer performs mean aggregation processing on the atom feature of Nx200 output by the second drug feature extraction layer to obtain an atom feature matrix with a dimension of Nx200 processed by the atom feature update layer.

[0133] The input of the third drug local message passing layer is the atom feature matrix with a dimension of Nx200 processed by the second atom feature update layer and the molecular structure graph. The drug local message passing layer uses the PNA algorithm to process the atom feature with a dimension of Nx200 and the edge index and edge feature corresponding to the molecular structure graph, and finally outputs the feature of the drug atom after three local message passing with a dimension of Nx200.

[0134] The input of the third sequence context modeling layer is the atom feature matrix with a dimension of Nx200 processed by the third atom feature update layer. The sequence context modeling layer processes the atom feature matrix by the BiLSTM algorithm, and finally outputs the atom feature after sequence context modeling with a dimension of Nx200.

[0135] ​The input of the third drug post-processing layer is the drug atom features with a dimension of N x 200 after local message passing output by the third drug local message passing layer and the atom features with a dimension of N x 200 after sequence context modeling output by the third sequence context modeling layer. The drug post-processing layer first merges the two input features to obtain a merged feature vector of atoms with a dimension of N x 400, then reduces the dimension of the merged feature vector to N x 200 through a linear layer, and normalizes the reduced merged feature vector, and finally obtains atom features with a dimension of N x 200.

[0136] The input of the third multi-scale feature updating layer is the atom features with a dimension of N x 200 processed by the third drug post-processing layer, the atom group features with a dimension of C x 200 obtained by the drug preprocessing layer, the molecular structure graph, and the atom group structure graph. The multi-scale feature updating layer performs residual processing on the atom group features to obtain updated atom group features with a dimension of C x 200, and then processes the updated atom group features through a multi-head attention mechanism to obtain drug molecule features with a dimension of 32 x 200 and atom group attention scores with a dimension of C x 5.

[0137] The input of the third fusion layer is the residue cluster features with a dimension of 32 x 20 x 200 output by the third residue cluster pooling layer and the atom group features with a dimension of C x 200, the drug molecule features with a dimension of 32 x 200, and the atom features with a dimension of N x 200 output by the third multi-scale feature updating layer. The fusion layer performs structured preprocessing on the input drug molecule features, atom group features, and atom features to obtain drug molecule features with a dimension of 32 x 1 x 200, atom group features with a dimension of 32 x C m x 200, and atom features with a dimension of 32 x N m x 200; then, through multi-head bilinear cross-attention, the residue cluster features are interacted with the drug molecule features with a dimension of 32 x 1 x 200, the atom group features with a dimension of 32 x C m x 200, and the atom features with a dimension of 32 x N m x 200 to obtain three fusion attention scores with dimensions of 32 x 4 x 1 x 20, 32 x 4 x C m x 20, and 32 x 4 x N m×20 (20 is the number of current multi-scale feature update layer residual cluster) ; then, the fusion attention score with the dimension of 32x4x1x20 is reshaped and dimension transposed to obtain the protein attention score with the dimension of 32x20x4, and then the three fusion attention scores are used to update the residue cluster features with the dimension of 32x20x200 output by the third residue cluster pooling layer and the group features with the dimension of Cx200 output by the first multi-scale feature update layer, the drug molecule features with the dimension of 32x200, and the atom features with the dimension of Nx200 obtained by the third drug post-processing layer, to obtain the protein attention score with the dimension of 32x20x4, the residue cluster features with the dimension of 640x200, the group features with the dimension of Cx200, the drug molecule features with the dimension of 32x200, and the atom features with the dimension of Nx200 output by the first fusion layer.

[0138] The input of the third residue feature update layer is the residue cluster features with the dimension of 640x200 processed by the third fusion layer (where 640 is the batch size 32 multiplied by the number of residue clusters in the batch 20) and the residue features with the dimension of Vx200 output by the third protein feature extraction layer, and the residue feature update layer performs mean aggregation on the residue features with the dimension of Vx200 output by the third protein feature extraction layer to obtain the residue feature matrix with the dimension of Vx200 processed by the residue feature update layer.

[0139] The input of the protein attention pooling layer is the attention scores with the dimensions of 32x5x4, 32x10x4, and 32x20x4 output by the three fusion layers, the residue cluster features with the dimensions of 32xV m x5, 32xV m x10, and 32xV m×20 residue cluster assignment probability matrix, and the residue feature matrix of dimension V×200 after processing by the third residue feature update layer. The protein attention pooling layer processes the attention score and the residue cluster assignment probability matrix to obtain the residue attention score (weighted aggregation is performed by matrix multiplication to obtain the intermediate residue attention score V×12, and then a linear layer and softmax are passed to finally generate the residue attention score of dimension V×1). Then, the residue feature matrix is ​​weighted summed (the feature vector of each residue is multiplied by its corresponding residue attention score to obtain a protein representation of dimension 32×200) to obtain the protein representation. The protein representation is passed through a two-layer MLP (its hidden layer uses ReLU as the activation function and the output layer uses layer normalization, where the input of the first layer is the 32×200 protein representation and the output is the 32×400 intermediate feature vector. The input of the second layer is the 32×400 intermediate feature vector and the output is the 32×200 final protein feature vector) to obtain the final protein feature vector of dimension 32×200.

[0140] The input of the third atomic feature update layer is the C×200 atomic cluster features after the third multi-scale feature update layer is processed by the third feature fusion module and the N×200 atomic features output by the third drug feature extraction layer. The atomic feature update layer performs mean aggregation processing on the N×200 atomic features output by the third drug feature extraction layer to obtain an atomic feature matrix with a dimension of N×200 after processing by the atomic feature update layer.

[0141] The input of the molecular attention pooling layer is the atomic group attention scores of dimension C×5 obtained by the three multi-scale feature update layers and the atomic feature matrix of dimension N×200 after being processed by the atomic feature update layer output by the third atomic feature update layer. The molecular attention pooling layer processes the three atomic group attention scores (first, all the atomic group attention scores related to each atom are averaged and aggregated to the atomic attention score of dimension N×15, and then a linear layer and softmax are used to finally generate the atomic attention score of dimension N×1) to obtain the atomic attention score of dimension N×1, and then Finally, the atomic feature matrix is ​​weighted summed (the feature vector of each atom is multiplied by its corresponding atomic attention score to obtain a drug representation with a dimension of 32×200) to obtain the drug representation, and the drug representation is passed through a two-layer MLP (its hidden layer uses ReLU as the activation function, and the output layer uses layer normalization, where the input of the first layer is the drug representation of 32×200, and the output is the intermediate feature vector of 32×400, and the input of the second layer is the intermediate feature vector of 32×400, and the output is the final drug feature vector of 32×200) to obtain the final drug feature vector with a dimension of 32×200.

[0142] The input of the multi-layer perception module is the final drug feature vector with a dimension of 32x200 output by the molecular attention pooling layer and the final protein feature vector with a dimension of 32x200 output by the protein attention pooling layer, and the multi-layer perception module performs dimension reduction processing on the final drug feature vector and the final protein feature vector (firstly, the final drug feature vector and the final protein feature vector are spliced to obtain a spliced vector with a dimension of 32x400, and then a two-layer MLP is used to obtain a drug target binding affinity prediction value, wherein the hidden layer uses ReLU as the activation function, the input of the first layer is the spliced vector with a dimension of 32x400, and the output is an intermediate feature vector with a dimension of 32x200, and the input of the second layer is the intermediate feature vector with a dimension of 32x200, and the output is a drug target binding affinity prediction value with a dimension of 32x1) to obtain a drug target binding affinity prediction value with a dimension of 32x1.

[0143] Specifically, the drug target binding affinity prediction model of the application is obtained by the following steps:

[0144] (4-1) Obtain a plurality of to-be-predicted drug-protein data pairs, preprocess all to-be-predicted drug-protein pairs to obtain a plurality of preprocessed drug-protein data pairs, which include SMILES sequences of drugs, amino acid sequences of proteins, and drug target binding affinity data, and divide all preprocessed drug-protein data pairs into a training set and a test set according to a ratio of 4:1;

[0145] (4-2) For the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained in step (4-1), use a pre-trained ESM model to process the amino acid sequence to obtain an evolutionary language modeling representation (wherein each token is encoded as a 1280-dimensional vector) and a contact probability matrix of the amino acid sequence, connect the residue pairs (the rows and columns of the contact probability matrix correspond to each residue in the amino acid sequence, and the values in the matrix represent the contact probability between the residue pairs) with a probability greater than or equal to a preset contact probability threshold (in the present application, the threshold is 0.5) as edges in a protein structure graph, thereby obtaining a protein structure graph corresponding to the amino acid sequence, wherein each node in the protein structure graph represents an amino acid, and the edges represent the contact between two amino acids; and map each amino acid in the amino acid sequence to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of Vx33, where V is the length of the amino acid sequence.

[0146] (4-3) For each SMILES sequence of each drug in the pre-processed drug-protein data pair obtained in step (4-1), the SMILES sequence is parsed using the RDKit tool to obtain a molecular structure graph corresponding to the SMILES sequence, an atomic feature matrix is obtained according to the physicochemical properties of each atom in the molecular structure graph, an atomic group set is generated according to the molecular structure graph corresponding to the obtained SMILES sequence, and an atomic group structure graph corresponding to the SMILES sequence is generated according to the connection relationship between the atomic groups in the atomic group set;

[0147] (4-4) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure graph, the evolutionary language modeling representation, and the residue feature matrix corresponding to the amino acid sequence obtained in step (4-2) are input into the first protein feature extraction layer in the protein encoder module in the drug target binding affinity prediction model, that is, the protein preprocessing layer, the first protein local message passing layer, the first structured state modeling layer, and the first protein post-processing layer are processed in sequence to obtain the residue feature corresponding to the amino acid sequence after the first protein post-processing layer, which has a dimension of V x 200

[0148] The advantage of the above step (4-4) is that local message passing, structured state modeling, and pre-processing and post-processing modules are introduced to realize deep feature extraction of residues in the protein graph, capture long-range dependency and spatial structure characteristics in the protein graph, and improve the accuracy and hierarchy of protein representation.

[0149] (4-5) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue feature corresponding to the amino acid sequence obtained in step (4-4) after the first protein post-processing layer, which has a dimension of V x 200 is input into the first residue cluster clustering layer in the protein encoder module in the drug target binding affinity prediction model, that is, the first residue cluster construction layer and the first residue cluster pooling layer are processed in sequence to obtain the residue cluster feature corresponding to the amino acid sequence, which has dimensions of 32 x 5 x 200 and 32 x V m x 5, respectively. and the residue cluster assignment probability matrix

[0150] The advantage of the above step (4-5) is that the residues are clustered into multiple clusters, a higher-level local representation of the protein is constructed, and the feature dimension is reduced; it is beneficial for subsequent information alignment in the local area during cross-modal interaction and improves the efficiency of feature fusion.

[0151] (4-6) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure diagram, atomic group structure diagram, and atomic feature matrix obtained in step (4-3) are input into the first drug feature extraction layer in the drug encoder module in the drug target binding affinity prediction model, that is, the drug preprocessing layer, the first drug local message passing layer, the first sequence context modeling layer and the first drug post-processing layer are sequentially performed to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the first drug post-processing layer. And the atomic cluster features of the SMILES sequence with a dimension of C×200 after the first drug post-processing layer

[0152] (4-7) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic features of dimension N×200 corresponding to the SMILES sequence obtained in step (4-6) after being processed by the first drug post-processing layer are converted to The atomic cluster features of the SMILES sequence with a dimension of C×200 after the first drug post-processing layer Input to the first multi-scale feature update layer in the drug-target binding affinity prediction model to obtain the atomic group attention scores corresponding to the SMILES sequence with dimensions of C×5, C×200 and 32×200 respectively Cluster characteristics and drug molecular characteristics Where 5 is the number of attention heads;

[0153] The advantage of the above steps (4-7) is that the multi-head attention mechanism is used to aggregate atomic features layer by layer to the atomic cluster and molecular levels, completing information fusion between scales; retaining the key substructure information at different levels of the drug, which is conducive to interaction alignment with proteins.

[0154] (4-8) For each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-5) is converted to a matrix with dimensions of 32×5×200 and 32×V. m ×5 residue cluster features and the residue cluster probability assignment matrix The atomic features of the SMILES sequence obtained in steps (4-6) with a dimension of N×200 after the first drug post-processing layer The atomic cluster features corresponding to the SMILES sequence obtained in steps (4-7) with dimensions of C×200 and 32×200 respectively and drug molecule features The first fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the drug molecule feature matrix corresponding to the SMILES sequence with dimensions of 32x1x200, 32xC m x200 and 32xN m x200 group feature matrix and atom feature matrix

[0155] (4-9) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the residue cluster feature corresponding to the amino acid sequence obtained in step (4-5) with a dimension of 32x5x200 and the drug molecule feature matrix corresponding to the SMILES sequence obtained in step (4-8) with dimensions of 32x1x200, 32xC m x200 and 32xN m x200 group feature matrix and atom feature matrix The first fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and with dimensions of 32x4xN m x5, 32x4xC m x5, 32x4x1x5, where 32 is the batch size, 4 is the number of attention heads, C m and N m are the maximum number of groups and atoms in each drug, respectively, and 5 is the number of residue clusters of the current multi-scale feature update layer;

[0156] (4-10) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the fusion attention score corresponding to the drug-protein data pair obtained in step (4-9) and The first fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the protein attention score corresponding to the amino acid sequence with dimensions of 32x5x4 and 160x200 and residue cluster feature and the group feature corresponding to the SMILES sequence with dimensions of Cx200, 32x200 and Nx200 Drug molecule features and atomic features

[0157] The steps (4-8) to (4-10) have the advantages that, based on a bilinear cross-attention mechanism, matching of residue clusters and different scale drug features is achieved; meanwhile, cross-modal interaction relationships at multiple granularities are modeled, and the recognition ability of the model for complex binding sites is enhanced. The updated residue and drug features have higher pertinence, and the recognition accuracy of the binding site is improved.

[0158] (4-11) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), input the group feature corresponding to the SMILES sequence obtained in step (4-10) and the atomic feature obtained in step (4-6) to a first atomic feature updating layer of the drug target binding affinity prediction model, to obtain an atomic feature processed by the first atomic feature updating layer and corresponding to the SMILES sequence, the atomic feature processed by the first atomic feature updating layer and corresponding to the SMILES sequence has a dimension of N x 200. (4-11) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), input the group feature corresponding to the SMILES sequence obtained in step (4-10) and the atomic feature obtained in step (4-6) to a first atomic feature updating layer of the drug target binding affinity prediction model, to obtain an atomic feature processed by the first atomic feature updating layer and corresponding to the SMILES sequence, the atomic feature processed by the first atomic feature updating layer and corresponding to the SMILES sequence has a dimension of N x 200.

[0159] (4-12) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), input the residue cluster feature corresponding to the amino acid sequence obtained in step (4-10) and the residue feature corresponding to the amino acid sequence obtained in step (4-4) to a first residue feature updating layer of the drug target binding affinity prediction model, to obtain a residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence, the residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence has a dimension of V x 200. (4-12) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), input the residue cluster feature corresponding to the amino acid sequence obtained in step (4-10) and the residue feature corresponding to the amino acid sequence obtained in step (4-4) to a first residue feature updating layer of the drug target binding affinity prediction model, to obtain a residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence, the residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence has a dimension of V x 200.

[0160] (4-13) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), input the protein structure graph corresponding to the amino acid sequence obtained in step (4-2) and the residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence obtained in step (4-12) to the first residue feature updating layer of the drug target binding affinity prediction model, to obtain a residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence, the residue feature processed by the first residue feature updating layer and corresponding to the amino acid sequence has a dimension of V x 200. ​​the second protein post-processing layer, to obtain the residue feature corresponding to the amino acid sequence and processed by the second protein post-processing layer, with a dimension of V x 200

[0161] (4-14) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue feature corresponding to the amino acid sequence and processed by the second protein post-processing layer obtained in step (4-13) is obtained, with a dimension of V x 200 the second residue cluster construction layer and the second residue cluster pooling layer, to obtain the residue cluster feature corresponding to the amino acid sequence, with a dimension of 32 x 10 x 200 and 32 x V x 10, respectively m and the residue cluster assignment probability matrix

[0162] (4-15) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure graph, the atom group structure graph obtained in step (4-3), and the atom feature corresponding to the SMILES sequence and processed by the first atom feature update layer obtained in step (4-11) with a dimension of N x 200 the second drug feature extraction layer in the drug encoder module of the drug target binding affinity prediction model, that is, the second drug local message passing layer, the second sequence context modeling layer and the second drug post-processing layer are sequentially processed, to obtain the atom feature corresponding to the SMILES sequence and processed by the second drug post-processing layer, with a dimension of N x 200 and the atom group feature corresponding to the SMILES sequence and processed by the second drug post-processing layer, with a dimension of C x 200

[0163] (4-16) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the atom feature corresponding to the SMILES sequence and processed by the second drug post-processing layer obtained in step (4-15) with a dimension of N x 200 and the atom group feature corresponding to the SMILES sequence and processed by the second drug post-processing layer, with a dimension of C x 200 ​the second multi-scale feature update layer input into the drug target binding affinity prediction model to obtain the group attention score corresponding to the SMILES sequence, with dimensions of Cx5, Cx200 and 32x200 group features and drug molecule features wherein 5 is the number of attention heads;

[0164] (4-17) for each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature corresponding to the amino acid sequence obtained in step (4-14) has dimensions of 32x10x200 and 32xV m x10 and residue cluster probability distribution matrix the atomic feature corresponding to the SMILES sequence obtained in step (4-15) after being processed by the second drug post-processing layer, with a dimension of Nx200 and the group feature corresponding to the SMILES sequence obtained in step (4-16), with dimensions of Cx200 and 32x200 and drug molecule features the second fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the drug molecule feature matrix corresponding to the SMILES sequence, with dimensions of 32x1x200, 32xC m x200 and 32xN m x200 group feature matrix and atomic feature matrix

[0165] (4-18) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the residue cluster feature corresponding to the amino acid sequence obtained in step (4-14) has a dimension of 32x5x200 and the drug molecule feature matrix corresponding to the SMILES sequence obtained in step (4-17), with dimensions of 32x1x200, 32xC m x200 and 32xN m x200 group feature matrix and atomic feature matrix the second fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and with dimensions of 32x4xN m× 10, 32 × 4 × C m × 10, 32 × 4 × 1 × 10;

[0166] (4-19) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the fused attention score corresponding to the drug-protein data pair obtained in step (4-18) and is input into the second fusion layer of the feature fusion module in the drug target binding affinity prediction model to obtain the protein attention score corresponding to the amino acid sequence with dimensions of 32 × 10 × 4 and 320 × 200 and the residue cluster feature and the group feature corresponding to the SMILES sequence with dimensions of C × 200, 32 × 200 and N × 200 drug molecule feature and atomic feature

[0167] (4-20) For the SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the group feature corresponding to the SMILES sequence obtained in step (4-19) with dimensions of C × 200 and the atomic feature processed by the second drug post-processing layer corresponding to the SMILES sequence obtained in step (4-15) with dimensions of N × 200 is input into the second atomic feature update layer in the drug target binding affinity prediction model to obtain the atomic feature processed by the second atomic feature update layer corresponding to the SMILES sequence with dimensions of N × 200

[0168] (4-21) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature corresponding to the amino acid sequence obtained in step (4-19) with dimensions of 320 × 200 and the residue feature processed by the second protein post-processing layer corresponding to the amino acid sequence obtained in step (4-13) with dimensions of V × 200 is input into the second residue feature update layer in the drug target binding affinity prediction model to obtain the residue feature processed by the second residue feature update layer corresponding to the amino acid sequence with dimensions of V × 200

[0169] (4-22) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure graph corresponding to the amino acid sequence obtained in step (4-2) and the residue feature of dimension V x 200 processed by the second residue feature update layer corresponding to the amino acid sequence obtained in step (4-21) The third protein feature extraction layer in the protein encoder module input into the drug target binding affinity prediction model, i.e., sequentially performing the third protein local message passing layer, the third structured state modeling layer and the third protein post-processing layer to obtain the residue feature of dimension V x 200 processed by the third protein post-processing layer corresponding to the amino acid sequence

[0170] (4-23) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue feature of dimension V x 200 processed by the third protein post-processing layer corresponding to the amino acid sequence obtained in step (4-22) The third residue cluster clustering layer in the protein encoder module input into the drug target binding affinity prediction model, i.e., sequentially performing the third residue cluster construction layer and the third residue cluster pooling layer to obtain the residue cluster feature of dimension 32 x 20 x 200 and 32 x V x 20 respectively m corresponding to the amino acid sequence And the residue cluster assignment probability matrix

[0171] (4-24) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure graph obtained in step (4-3), the atom group structure graph and the atom feature of dimension N x 200 processed by the second atom feature update layer corresponding to the SMILES sequence obtained in step (4-20) The third drug feature extraction layer in the drug encoder module input into the drug target binding affinity prediction model, i.e., sequentially performing the third drug local message passing layer, the third sequence context modeling layer and the third drug post-processing layer to obtain the atom feature of dimension N x 200 processed by the third drug post-processing layer corresponding to the SMILES sequence And the atom group feature of dimension C x 200 processed by the third drug post-processing layer corresponding to the SMILES sequence

[0172] (4-25) For each SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the atom features of dimension N x 200 processed by the third drug post-processing layer corresponding to the SMILES sequence obtained in step (4-24) and the atom group features of dimension C x 200 processed by the third drug post-processing layer corresponding to the SMILES sequence input into the third multi-scale feature update layer of the drug target binding affinity prediction model to obtain the atom group attention score of dimension C x 5, C x 200 and 32 x 200 respectively corresponding to the SMILES sequence atom group features and drug molecule features wherein 5 is the number of attention heads;

[0173] (4-26) For each drug-protein data pair in the training set obtained in step (4-1), the residue cluster features of dimension 32 x 20 x 200 and 32 x V m x 20 respectively corresponding to the amino acid sequence obtained in step (4-23) and the residue cluster probability distribution matrix the atom features of dimension N x 200 processed by the third drug post-processing layer corresponding to the SMILES sequence obtained in step (4-24) and the atom group features of dimension C x 200 and 32 x 200 respectively corresponding to the SMILES sequence obtained in step (4-25) and drug molecule features input into the third fusion layer of the feature fusion module of the drug target binding affinity prediction model to obtain the drug molecule feature matrix of dimension 32 x 1 x 200, 32 x C m x 200 and 32 x N m x 200 respectively corresponding to the SMILES sequence atom group feature matrix and atom feature matrix

[0174] (4-27) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the residue cluster features of dimension 32 x 20 x 200 corresponding to the amino acid sequence obtained in step (4-23) and the drug molecule feature matrix of dimension 32 x 1 x 200, 32 x C m x 200 and 32 x N ma drug molecule feature matrix of 200 a group feature matrix and an atom feature matrix a third fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain a fusion attention score corresponding to the drug-protein data pair and with dimensions of 32x4xN respectively m x20, 32x4xC m x20, 32x4x1x20

[0175] (4-28) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1) and the SMILES sequence of each drug, the fusion attention score corresponding to the drug-protein data pair obtained in step (4-27) and a third fusion layer of the feature fusion module input into the drug target binding affinity prediction model to obtain a protein attention score corresponding to the amino acid sequence with dimensions of 32x20x4 and 640x200 respectively and residue cluster features and group features corresponding to the SMILES sequence with dimensions of Cx200, 32x200 and Nx200 respectively drug molecule features and atom features

[0176] (4-29) for the SMILES sequence of each drug in each drug-protein data pair in the training set obtained in step (4-1), the group features corresponding to the SMILES sequence obtained in step (4-28) with dimensions of Cx200 and the atom features corresponding to the SMILES sequence obtained in step (4-24) with dimensions of Nx200 processed by the third drug post-processing layer a third atom feature update layer input into the drug target binding affinity prediction model to obtain atom features corresponding to the SMILES sequence with dimensions of Nx200 processed by the third atom feature update layer

[0177] (4-30) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster features corresponding to the amino acid sequence obtained in step (4-28) with dimensions of 640x200 the residue feature corresponding to the amino acid sequence obtained in step (4-22) and processed by the third protein post-processing layer the residue feature corresponding to the amino acid sequence obtained in step (4-22) and processed by the third residue feature updating layer

[0178] (4-31) for each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster assignment probability matrix of dimension 32xV m x5 corresponding to the amino acid sequence obtained in step (4-5) the protein attention score of dimension 32x5x4 corresponding to the amino acid sequence obtained in step (4-10) the residue cluster assignment probability matrix of dimension 32xV m x10 corresponding to the amino acid sequence obtained in step (4-14) the protein attention score of dimension 32x10x4 corresponding to the amino acid sequence obtained in step (4-19) the residue cluster assignment probability matrix of dimension 32xV m x20 corresponding to the amino acid sequence obtained in step (4-23) the protein attention score of dimension 32x20x4 corresponding to the amino acid sequence obtained in step (4-28) and the residue feature of dimension Vx200 corresponding to the amino acid sequence obtained in step (4-30) and processed by the third residue feature updating layer the final protein feature vector of dimension 32x200 corresponding to the amino acid sequence obtained in step (4-30) and processed by the protein attention pooling layer

[0179] (4-32) for each SMILES sequence of each drug in the training set obtained in step (4-1), the atom group attention score of dimension Cx5 corresponding to the SMILES sequence obtained in step (4-7) the atom group attention score of dimension Cx5 corresponding to the SMILES sequence obtained in step (4-16) the atom group attention score of dimension Cx5 corresponding to the SMILES sequence obtained in step (4-25) and the atomic features of the SMILES sequence obtained in step (4-13) after being processed by the third atomic feature update layer with the dimension of N x 200 the drug attention pooling layer of the drug target binding affinity prediction model to obtain the final drug feature vector with the dimension of 32 x 200 corresponding to the SMILES sequence

[0180] the final protein feature vector with the dimension of 32 x 200 corresponding to the amino acid sequence output by step (4-31) for each drug-protein data pair in the training set obtained in step (4-1) and the final drug feature vector with the dimension of 32 x 200 corresponding to the SMILES sequence output by step (4-32) the multi-layer perception module of the drug target binding affinity prediction model to obtain the final protein feature vector and the final drug feature vector are concatenated to obtain a concatenated vector with the dimension of 32 x 400, and the concatenated vector is processed by a two-layer MLP to obtain the drug target binding affinity prediction value corresponding to the drug-protein data pair

[0181] (4-34) for each drug-protein data pair in the training set obtained in step (4-1), the drug target binding affinity data corresponding to the drug-protein data pair obtained in step (4-1) and the drug target binding affinity prediction value corresponding to the drug-protein data pair obtained in step (4-33) a loss function is calculated, and the drug target binding affinity prediction model is trained using the loss function until the drug target binding affinity prediction model converges, thereby obtaining a preliminarily trained drug target binding affinity prediction model.

[0182] Specifically, the loss function used in this step is:

[0183]

[0184] where y represents the drug target binding affinity data corresponding to the drug-protein data pair.

[0185] (4-35) for each drug-protein data pair in the training set obtained in step (4-1), the test set obtained in step (4-1) is used to test the drug target binding affinity prediction model preliminarily trained in step (4-34) until the value of the loss function reaches the optimal value, thereby obtaining a trained drug target binding affinity prediction model.

Claims

1. A drug-target binding affinity prediction method based on multi-scale feature fusion, characterized in that: The following steps are involved: (1) obtaining a drug-protein data pair to be predicted, and preprocessing the drug-protein data pair to be predicted to obtain a preprocessed drug-protein data pair, which includes a simplified molecular linear input specification SMILES sequence of the drug, an amino acid sequence of the protein, and drug-target binding affinity data; (2) For each amino acid sequence of the protein in the pre-processed drug-protein data pair obtained in step (1), the amino acid sequence is processed using a pre-trained evolutionary scale model (ESM) to obtain an evolutionary language modeling representation and a contact probability matrix of the amino acid sequence, and the residue pairs in the contact probability matrix whose contact probability is greater than or equal to a preset contact probability threshold are connected as edges in the protein structure graph, thereby obtaining a protein structure graph corresponding to the amino acid sequence, wherein each node in the protein structure graph represents an amino acid and an edge represents the contact between two amino acids; and each amino acid in the amino acid sequence is mapped to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of the length of the amino acid sequence × 33; (3) For each drug SMILES sequence in the preprocessed drug-protein data pair obtained in step (1), the SMILES sequence is parsed using the RDKit tool to obtain a molecular structure diagram corresponding to the SMILES sequence, an atomic feature matrix is ​​obtained based on the physical and chemical properties of each atom in the molecule in the molecular structure diagram, an atomic group set is generated based on the molecular structure diagram corresponding to the obtained SMILES sequence, and an atomic group structure diagram corresponding to the SMILES sequence is generated based on the connection relationship between the atomic groups in the atomic group set; (4) The protein structure diagram and residue feature matrix corresponding to each amino acid sequence obtained in step (2), and the molecular structure diagram, atomic cluster structure diagram and atomic feature matrix corresponding to the SMILES sequence of each drug obtained in step (3) are input into the pre-trained drug-target binding affinity prediction model to obtain the drug-target binding affinity prediction value.

2. The method for predicting drug-target binding affinity based on multi-scale feature fusion according to claim 1, characterized in that: Step (1) specifically includes: first, screening the drug-protein data pairs to be predicted, deleting the data with missing data problems, so as to obtain the SMILES sequence of the drug, the amino acid sequence of the protein, and the inhibition constant and dissociation constant of the drug-protein interaction; then, obtaining the drug-target binding affinity data based on the obtained inhibition constant and dissociation constant; wherein the calculation formula of the drug-target binding affinity data is: Among them, K i With K d represent the inhibition constant and dissociation constant, pK i represents the drug-target binding affinity data calculated based on the inhibition constant, pK d represents the drug-target binding affinity data calculated based on the dissociation constant; Finally, the obtained drug SMILES sequence, protein amino acid sequence, and drug-target binding affinity data are combined into preprocessed drug-protein data pairs.

3. The drug-target binding affinity prediction method based on multi-scale feature fusion according to claim 1 or 2, characterized in that: The drug-target binding affinity prediction model includes a protein encoder module, a drug encoder module, a feature fusion module, and a multi-layer perceptron module; The protein encoder module includes a protein feature extraction layer, a residue clustering layer, and a protein pooling layer; The protein feature extraction layer is implemented based on Mamba and message passing neural network MPNN, and includes a protein pre-processing layer, three protein local message passing layers, three structured state modeling layers, and three protein post-processing layers; The residue clustering layer includes three residue cluster construction layers and three residue cluster pooling layers; The protein pooling layer includes three residue feature update layers and a protein attention pooling layer; The drug encoder module includes a drug feature extraction layer, three multi-scale feature update layers, and a drug pooling layer; The drug feature extraction layer is based on a bidirectional long short-term memory network (BiLSTM) and MPNN, and includes a drug pre-processing layer, three drug local message passing layers, three sequence context modeling layers, and three drug post-processing layers. The drug pooling layer consists of three atomic feature update layers and one molecular attention pooling layer; The feature fusion module consists of three fusion layers.

4. The method for predicting drug-target binding affinity based on multi-scale feature fusion according to any one of claims 1 to 3, characterized in that: The input of the protein preprocessing layer is the evolutionary language modeling representation of the 32 amino acid sequence, the residue feature matrix and the contact probability matrix; the protein preprocessing layer first uses the radial basis function to process the input contact probability matrix to obtain an edge feature matrix of size E×200, where E represents the number of edges in the protein structure graph, and then processes the evolutionary language modeling representation through a multi-layer perceptron MLP, in which the hidden layer uses the rectified linear unit ReLU as the activation function, and the output layer uses layer normalization, in which the input of the first layer is the evolutionary language modeling representation of dimension V×1280, and the output is the intermediate feature vector of dimension V×400, the input of the second layer is the intermediate feature vector of dimension V×400, and the final output is the preprocessed evolutionary language modeling representation of dimension V×200. Language modeling representation, where V represents the length of the amino acid sequence, to obtain the preprocessed evolutionary language modeling representation; then, the residue feature matrix is ​​processed by a two-layer MLP, the hidden layer uses ReLU as the activation function, and the output layer uses layer normalization, where the input of the first layer is the residue feature matrix of dimension V×33, and the output is the intermediate feature vector of dimension V×400, and the input of the second layer is the intermediate feature vector of dimension V×400, and the output is the preprocessed residue feature matrix of dimension V×200, to obtain the preprocessed residue feature matrix; finally, the preprocessed evolutionary language modeling representation of dimension V×200 and the preprocessed residue feature matrix of dimension V×200 are added, and the comprehensive features of protein residues of dimension V×200 are finally output; The first protein local message passing layer takes as input the comprehensive features of protein residues with a dimension of V × 200 and the protein structure graph output by the protein preprocessing layer. This protein local message passing layer uses the principal neighborhood aggregation (PNA) algorithm to process the comprehensive features of protein residues with a dimension of V × 200 and the edge indices and edge features corresponding to the protein structure graph, and finally outputs the features of protein residues with a dimension of V × 200 after local message passing. The first structured state modeling layer takes as input the V×200-dimensional comprehensive features of protein residues obtained by the protein preprocessing layer and the protein structure graph. This structured state modeling layer uses a degree- and batch-based node sorting strategy to sort the nodes of the protein structure graph, processes the comprehensive features of protein residues using the Mamba model, and ultimately outputs V×200-dimensional features of protein residues after global attention processing. The input of the first protein post-processing layer is the features of the protein residues after local message passing with a dimension of V×200 output by the first protein local message passing layer and the features of the protein residues after global attention processing with a dimension of V×200 output by the first structured state modeling layer. The protein post-processing layer first merges the two input features to obtain a merged feature vector of protein residues with a dimension of V×200. Subsequently, the merged feature vector of protein residues is processed by a two-layer MLP, in which the hidden layer uses ReLU as the activation function and the output layer uses graph normalization. The input of the first layer is the merged feature vector of V×200 protein residues, and the output is the intermediate feature vector of protein residues with a dimension of V×400. The input of the second layer is the intermediate feature vector of protein residues with a dimension of V×400, and the output is the features of protein residues with a dimension of V×200 after being processed by the first protein post-processing layer, so as to obtain features of protein residues with a dimension of V×200 after being processed by the first protein post-processing layer; The input of the first residue clustering layer is the features of protein residues after being processed by the first protein post-processing layer and the protein structure graph with a dimension of V×200. The residue clustering layer first processes the features with a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of V×5. Then, the intermediate residue cluster assignment probability matrix is ​​subjected to a softmax operation and a batch filling conversion according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32×V m ×5, where V m Indicates the maximum length of all input amino acid sequences; The input of the first residue cluster pooling layer is the output of the first residue cluster construction layer with a dimension of 32×V m ×5 residue cluster assignment probability matrix, and a V×200 feature of protein residues after the first protein post-processing layer. The residue cluster pooling layer first uses the dense mincut pooling algorithm to soft cluster all residues in the feature to obtain the probability distribution of each residue belonging to different clusters as the residue cluster feature, and finally outputs the residue cluster feature with a dimension of 32×5×200.

5. The method for predicting drug-target binding affinity based on multi-scale feature fusion according to claim 4, characterized in that: The input of the drug preprocessing layer includes the molecular structure diagram, the atomic feature matrix with a dimension of N×43, and the atomic cluster structure diagram. The drug preprocessing layer first processes the atomic feature matrix through MLP to obtain atomic features with a dimension of N×200. Then, the atomic cluster structure diagram is processed by MLP to obtain atomic cluster features with a dimension of C×200, where C represents the number of nodes in the atomic cluster structure diagram. The first drug local message passing layer takes as input the atomic features and molecular structure graph of dimension N×200 obtained by the drug preprocessing layer. This drug local message passing layer processes the atomic features and the edge indices and edge features corresponding to the molecular structure graph using the PNA algorithm, and ultimately outputs the features of the drug atoms of dimension N×200 after local message passing. The first sequence context modeling layer takes as input the atomic features of dimension N × 200 obtained by the drug preprocessing layer. This sequence context modeling layer processes these atomic features using the BiLSTM algorithm and ultimately outputs atomic features of dimension N × 200 after sequence context modeling. The input of the first drug post-processing layer is the N×200 dimension of the drug atom features after local message passing output by the first drug local message passing layer and the N×200 dimension of the atomic features after sequence context modeling output by the first sequence context modeling layer. The drug post-processing layer first concatenates the two input features to obtain a merged feature vector of N×400 atoms, then reduces the dimension of the merged feature vector to N×200 through a linear layer, and normalizes the merged feature vector after dimensionality reduction to finally obtain the atomic features of N×200 dimensions. The input of the first multi-scale feature update layer is the atomic features of dimension N×200 obtained by the first drug post-processing layer, the atomic cluster features of dimension C×200 obtained by the drug pre-processing layer, the molecular structure diagram, and the atomic cluster structure diagram. The multi-scale feature update layer performs residual processing on the atomic cluster features to obtain updated atomic cluster features of dimension C×200. Then, the updated atomic cluster features are processed by the multi-head attention mechanism to obtain drug molecular features of dimension 32×200 and atomic cluster attention scores of dimension C×5. The input of the first fusion layer is the residue cluster features with a dimension of 32×5×200 output by the first residue cluster pooling layer, the atomic group features with a dimension of C×200 and the drug molecule features with a dimension of 32×200 output by the first multi-scale feature updating layer, and the atomic features with a dimension of N×200 obtained by the first drug post-processing layer. The fusion layer performs structured preprocessing on the input drug molecule features, atomic group features and atomic features to obtain drug molecule features with a dimension of 32×1×200, drug molecule features with a dimension of 32×C m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features, among which C m and N m are the maximum number of atomic groups and the number of atoms in all drugs respectively; then, the residue cluster features are respectively combined with the drug molecular features of dimension 32×1×200 and dimension 32×C by multi-head bilinear cross attention. m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features to obtain three fusion attention scores with dimensions of 32×4×1×5, 32×4×C m ×5, 32×4×N m ×5; then, the fusion attention score with a dimension of 32×4×1×5 is reshaped and permuted to obtain a protein attention score with a dimension of 32×5×4; then, according to the three fusion attention scores, the residue cluster features with a dimension of 32×5×200 output by the first residue cluster pooling layer, the atomic group features of C×200 output by the first multi-scale feature update layer, the drug molecular features of 32×200, and the atomic features of N×200 obtained by the first drug post-processing layer are updated to obtain the protein attention score with a dimension of 32×5×4 output by the first fusion layer, the residue cluster features of 160×200, the atomic group features of C×200, the drug molecular features of 32×200, and the atomic features of N×200; The input of the first residue feature update layer is the residue cluster features of dimension 160×200 processed by the first fusion layer and the residue features of dimension V×200 output by the first protein feature extraction layer. The residue feature update layer performs mean aggregation on the residue features to obtain a residue feature matrix of dimension V×200 processed by the residue feature update layer. The second protein local message passing layer takes as input the residue feature matrix (V × 200) processed by the first residue feature update layer and the protein structure graph. This protein local message passing layer processes the edge indices and edge features corresponding to the residue feature matrix and the protein structure graph using the PNA algorithm, and ultimately outputs the protein residue features (V × 200) after local message passing. The input of the second structured state modeling layer is the residue feature matrix of dimension V×200 processed by the first residue feature update layer and the protein structure graph. The structured state modeling layer uses a degree- and batch-based node sorting strategy to sort all nodes in the protein structure graph, and processes the residue feature matrix through the Mamba model. Finally, it outputs the protein residue features of dimension V×200 after global attention processing; The input of the second protein post-processing layer is the V×200 features of the protein residues after local message passing output by the second protein local message passing layer and the V×200 features of the protein residues after global attention processing output by the second structured state modeling layer. The protein post-processing layer first merges the two features to obtain a merged feature vector of protein residues with a dimension of V×200. Subsequently, the merged feature vector of protein residues is processed by a two-layer MLP to obtain a V×200 feature of protein residues after being processed by the second protein post-processing layer. The input of the second residue clustering layer is the features of protein residues processed by the second protein post-processing layer and the protein structure graph with a dimension of V×200. The residue clustering layer first processes the features with a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of V×10. Then, the intermediate residue cluster assignment probability matrix is ​​subjected to a softmax operation and a batch filling conversion according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32×V m ×10; The input of the third residue cluster pooling layer is the output of the second residue cluster construction layer with a dimension of 32×V m ×10 residue cluster assignment probability matrix and a V × 200 feature of protein residues after processing by the second protein post-processing layer. The residue cluster pooling layer uses the dense mincut pooling algorithm to soft cluster all residues in the feature to obtain the probability distribution of each residue belonging to different clusters as the residue cluster feature, and finally outputs the residue cluster feature with a dimension of 32 × 10 × 200.

6. The method for drug-target binding affinity prediction based on multi-scale feature fusion according to claim 5, characterized in that: The input of the first atomic feature update layer is the C×200 atomic cluster features processed by the first fusion layer of the first multi-scale feature update layer and the N×200 atomic features output by the first drug feature extraction layer. The atomic feature update layer performs mean aggregation on the N×200 atomic features output by the first drug feature extraction layer to obtain an atomic feature matrix with a dimension of N×200 after processing by the atomic feature update layer. The input of the second drug local message passing layer is the atomic feature matrix and molecular structure graph processed by the first atomic feature update layer. This drug local message passing layer uses the PNA algorithm to process the edge index and edge features corresponding to the atomic features and molecular structure graph, and finally outputs the features of the drug atoms after the second local message passing with a dimension of N×200. The second sequence context modeling layer takes as input the N×200 atomic feature matrix processed by the first atomic feature update layer. This layer processes the atomic feature matrix using the BiLSTM algorithm and ultimately outputs N×200 atomic features after sequence context modeling. The input of the second drug post-processing layer is the features of drug atoms with a dimension of N×200 after local message passing output by the second drug local message passing layer and the atomic features with a dimension of N×200 after sequence context modeling output by the second sequence context modeling layer. The drug post-processing layer first merges the two input features to obtain a merged feature vector with a dimension of N×400 atoms, then reduces the dimension of the merged feature vector to N×200 through a linear layer, and normalizes the merged feature vector after dimensionality reduction to finally obtain atomic features with a dimension of N×200. The input of the second multi-scale feature update layer is the atomic features of dimension N×200 obtained by the second drug post-processing layer, the atomic cluster features of dimension C×200 obtained by the drug pre-processing layer, the molecular structure diagram, and the atomic cluster structure diagram. The multi-scale feature update layer first performs residual processing on the atomic cluster features to obtain updated atomic cluster features of dimension C×200. Then, the updated atomic cluster features are processed through the multi-head attention mechanism to obtain drug molecular features of dimension 32×200 and atomic cluster attention scores of dimension C×5. The input of the second fusion layer is the residue cluster features with a dimension of 32×10×200 output by the second residue cluster pooling layer, the atomic group features with a dimension of C×200 and the drug molecule features with a dimension of 32×200 obtained by the second multi-scale feature update layer, and the atomic features with a dimension of N×200 obtained by the second drug post-processing layer. The fusion layer performs structured preprocessing on the input drug molecule features, atomic group features and atomic features to obtain drug molecule features with a dimension of 32×10×200, drug molecule features with a dimension of 32×C×200 and drug molecule features with a dimension of 32×10×200. m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features; then, the residue cluster features are respectively combined with the drug molecular features of dimension 32×1×200 and the 32×C m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features to obtain three fusion attention scores with dimensions of 32×4×1×10, 32×4×C m ×10, 32×4×N m ×10; then, the fusion attention score with a dimension of 32×4×1×10 is reshaped and permuted to obtain a protein attention score with a dimension of 32×10×4. Then, according to the three fusion attention scores, the residue cluster features with a dimension of 32×10×200 output by the second residue cluster pooling layer, the atomic group features of C×200 output by the first multi-scale feature update layer, the drug molecular features of 32×200, and the atomic features of N×200 obtained by the second drug post-processing layer are updated to obtain the protein attention score with a dimension of 32×10×4 output by the first fusion layer, the residue cluster features of 320×200, the atomic group features of C×200, the drug molecular features of 32×200, and the atomic features of N×200; The input of the second residue feature update layer is the residue cluster features with a dimension of 320×200 after processing by the second fusion layer and the residue features with a dimension of V×200 output by the second protein feature extraction layer. The residue feature update layer performs mean aggregation on the residue features to obtain a residue feature matrix with a dimension of V×200 after processing by the residue feature update layer.

7. The method for predicting drug-target binding affinity based on multi-scale feature fusion according to claim 6, characterized in that: The third protein local message passing layer takes as input the V×200 residue feature matrix processed by the second residue feature update layer and the protein structure graph. This protein local message passing layer uses the PNA algorithm to process the V×200 residue feature matrix and the edge indices and edge features corresponding to the protein structure graph, and ultimately outputs the V×200 features of the protein residues after local message passing. The input of the third structured state modeling layer is the residue feature matrix of dimension V × 200 processed by the second residue feature update layer and the protein structure graph. The structured state modeling layer uses a degree- and batch-based node sorting strategy to sort the nodes of the protein structure graph and processes the residue feature matrix through the Mamba model. The final output is the protein residue features of dimension V × 200 after global attention processing; The input of the third protein post-processing layer is the features of protein residues after local message passing (V×200) output by the third protein local message passing layer and the features of protein residues after global attention processing (V×200) output by the third structured state modeling layer. The protein post-processing layer first merges the two input features to obtain a merged feature vector of protein residues with a dimension of V×200. Subsequently, the merged feature vector of protein residues is processed by a two-layer MLP to obtain features of protein residues with a dimension of V×200 after being processed by the third protein post-processing layer. The input of the third residue clustering layer is the features of protein residues after processing by the third protein post-processing layer and the protein structure graph with a dimension of V×200. The residue clustering layer first processes the features with a graph convolution layer to obtain an intermediate residue cluster assignment probability matrix with a dimension of V×20. Then, the intermediate residue cluster assignment probability matrix is ​​subjected to a softmax operation and a batch filling conversion according to the protein structure graph to obtain a residue cluster assignment probability matrix with a dimension of 32×V m ×20; The input of the third residue cluster pooling layer is the output of the third residue cluster construction layer with a dimension of 32×V m ×20 residue cluster assignment probability matrix and a V × 200 feature of protein residues after processing by the third protein post-processing layer. The residue cluster pooling layer uses the dense mincut pooling algorithm to soft cluster all residues in the feature to obtain the probability distribution of each residue belonging to different clusters as the residue cluster feature, and finally outputs the residue cluster feature with a dimension of 32 × 20 × 200; The input of the second atomic feature update layer is the C×200 atomic cluster features processed by the second fusion layer of the second multi-scale feature update layer and the N×200 atomic features output by the second drug feature extraction layer. The atomic feature update layer performs mean aggregation processing on the N×200 atomic features output by the second drug feature extraction layer to obtain an atomic feature matrix with a dimension of N×200 after processing by the atomic feature update layer; The third drug local message passing layer takes as input the N×200 atomic feature matrix and the molecular structure graph processed by the second atomic feature update layer. This drug local message passing layer uses the PNA algorithm to process the N×200 atomic features and the edge indices and edge features corresponding to the molecular structure graph, ultimately outputting the N×200 features of the drug atoms after three local message passes. The third sequence context modeling layer takes as input the N×200 atomic feature matrix processed by the third atomic feature update layer. The sequence context modeling layer processes the atomic feature matrix using the BiLSTM algorithm and ultimately outputs N×200 atomic features after sequence context modeling. The input of the third drug post-processing layer is the features of drug atoms with a dimension of N×200 after local message passing output by the third drug local message passing layer and the atomic features with a dimension of N×200 after sequence context modeling output by the third sequence context modeling layer. The drug post-processing layer first merges the two input features to obtain a merged feature vector with a dimension of N×400 atoms, then reduces the dimension of the merged feature vector to N×200 through a linear layer, and normalizes the merged feature vector after dimensionality reduction to finally obtain atomic features with a dimension of N×200. The input of the third multi-scale feature update layer is the atomic features of dimension N×200 processed by the third drug post-processing layer, the atomic cluster features of dimension C×200 obtained by the drug pre-processing layer, the molecular structure diagram, and the atomic cluster structure diagram. The multi-scale feature update layer performs residual processing on the atomic cluster features to obtain updated atomic cluster features of dimension C×200. Then, based on the updated atomic cluster features, a multi-head attention mechanism is used to process them to obtain drug molecular features of dimension 32×200 and an atomic cluster attention score of dimension C×5. The input of the third fusion layer is the residue cluster features with a dimension of 32×20×200 output by the third residue cluster pooling layer, the atomic group features with a dimension of C×200 output by the third multi-scale feature updating layer, the drug molecular features with a dimension of 32×200, and the atomic features with a dimension of N×200 obtained by the third drug post-processing layer. The fusion layer performs structured preprocessing on the input drug molecular features, atomic group features and atomic features to obtain drug molecular features with a dimension of 32×1×200, drug molecular features with a dimension of 32×C m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features; then, the residue cluster features are respectively combined with the drug molecule features of dimension 32×1×200 and the 32×C m ×200 atomic cluster features and dimensions are 32×N m ×200 atomic features to obtain three fusion attention scores with dimensions of 32×4×1×20, 32×4×C m ×20, 32×4×N m ×20 (20 is the number of residue clusters in the current multi-scale feature update layer); then, the fusion attention score with a dimension of 32×4×1×20 is reshaped and permuted to obtain a protein attention score with a dimension of 32×20×4. Then, according to the three fusion attention scores, the residue cluster features with a dimension of 32×20×200 output by the third residue cluster pooling layer, the atomic group features of C×200 output by the first multi-scale feature update layer, the drug molecular features of 32×200, and the atomic features of N×200 obtained by the third drug post-processing layer are updated to obtain the protein attention score with a dimension of 32×20×4 output by the first fusion layer, the residue cluster features of 640×200, the atomic group features of C×200, the drug molecular features of 32×200, and the atomic features of N×200; The input of the third residue feature update layer is the residue cluster features of dimension 640×200 processed by the third fusion layer and the residue features of dimension V×200 output by the third protein feature extraction layer. The residue feature update layer performs mean aggregation processing on the residue features of dimension V×200 output by the third protein feature extraction layer to obtain a residue feature matrix of dimension V×200 after processing by the residue feature update layer.

8. The method for predicting drug-target binding affinity based on multi-scale feature fusion according to claim 7, characterized in that: The input of the protein attention pooling layer is the attention scores of the three fusion layers with dimensions of 32×5×4, 32×10×4, and 32×20×4, and the output of the three residue clustering layers with dimensions of 32×V m ×5, 32×V m ×10、32×V m ×20 residue cluster assignment probability matrix and the residue feature matrix of dimension V × 200 output by the third residue feature update layer after being processed by the residue feature update layer. The protein attention pooling layer processes the attention score and the residue cluster assignment probability matrix to obtain the residue attention score. Then, the residue feature matrix is ​​weighted summed to obtain the protein representation. The protein representation is passed through a two-layer MLP to obtain the final protein feature vector of dimension 32 × 200; The input of the third atomic feature update layer is the C×200 atomic cluster features processed by the third multi-scale feature update layer through the third feature fusion module and the N×200 atomic features output by the third drug feature extraction layer. The atomic feature update layer performs mean aggregation on the N×200 atomic features output by the third drug feature extraction layer to obtain an atomic feature matrix with a dimension of N×200 after processing by the atomic feature update layer. The input of the molecular attention pooling layer is the atomic group attention scores of dimension C×5 obtained by the three multi-scale feature update layers and the atomic feature matrix of dimension N×200 after being processed by the atomic feature update layer output by the third atomic feature update layer. The molecular attention pooling layer processes the three atomic group attention scores to obtain the atomic attention scores of dimension N×1. Then, the atomic feature matrix is ​​weighted summed to obtain the drug representation. The drug representation is processed by a two-layer MLP to obtain the final drug feature vector of dimension 32×200. The input of the multilayer perceptron module is the final drug feature vector with a dimension of 32×200 output by the molecular attention pooling layer and the final protein feature vector with a dimension of 32×200 output by the protein attention pooling layer. The multilayer perceptron module performs dimensionality reduction on the final drug feature vector and the final protein feature vector to obtain a drug-target binding affinity prediction value with a dimension of 32×1.

9. The method for predicting drug-target binding affinity based on multi-scale feature fusion according to claim 8, characterized in that: The drug-target binding affinity prediction model is trained through the following steps: (4-1) obtaining a plurality of drug-protein data pairs to be predicted, preprocessing all the drug-protein data pairs to be predicted to obtain a plurality of preprocessed drug-protein data pairs, which include drug SMILES sequences, protein amino acid sequences, and drug-target binding affinity data, and dividing all the preprocessed drug-protein data pairs into a training set and a test set in a ratio of 4:1; (4-2) For each protein amino acid sequence in the preprocessed drug-protein data pair obtained in step (4-1), the amino acid sequence is processed using a pre-trained ESM model to obtain an evolutionary language modeling representation of the amino acid sequence (wherein each token is encoded as a 1280-dimensional vector) and a contact probability matrix, and residue pairs in the contact probability matrix with a probability greater than or equal to a preset contact probability threshold (the threshold in the present invention is 0.5) (the rows and columns of the contact probability matrix correspond to each residue in the amino acid sequence, respectively, and the values ​​in the matrix represent the contact probability between the residue pairs) are connected as edges in the protein structure graph, thereby obtaining a protein structure graph corresponding to the amino acid sequence, wherein each node in the protein structure graph represents an amino acid, and an edge represents the contact between two amino acids; and each amino acid in the amino acid sequence is mapped to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of V×33, the length of the amino acid sequence; (4-3) For the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained in step (4-1), the SMILES sequence is parsed using the RDKit tool to obtain a molecular structure diagram corresponding to the SMILES sequence, an atomic feature matrix is ​​obtained based on the physical and chemical properties of each atom in the molecule in the molecular structure diagram, an atomic group set is generated based on the molecular structure diagram corresponding to the obtained SMILES sequence, and an atomic group structure diagram corresponding to the SMILES sequence is generated based on the connection relationship between the atomic groups in the atomic group set; (4-4) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure diagram, evolutionary language modeling representation and residue feature matrix corresponding to the amino acid sequence obtained in step (4-2) are input into the first protein feature extraction layer in the protein encoder module in the drug-target binding affinity prediction model, that is, the protein preprocessing layer, the first protein local message passing layer, the first structured state modeling layer and the first protein post-processing layer are sequentially processed to obtain the residue feature of the amino acid sequence with a dimension of V×200 after being processed by the first protein post-processing layer. (4-5) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue features corresponding to the amino acid sequence obtained in step (4-4) with a dimension of V×200 after being processed by the first protein post-processing layer are The first residue cluster clustering layer in the protein encoder module of the drug target binding affinity prediction model is input, that is, the first residue cluster construction layer and the first residue cluster pooling layer are sequentially performed to obtain the corresponding amino acid sequence with dimensions of 32×5×200 and 32×V m ×5 residue cluster features and residue cluster assignment probability matrix (4-6) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure diagram, atomic group structure diagram, and atomic feature matrix obtained in step (4-3) are input into the first drug feature extraction layer in the drug encoder module in the drug target binding affinity prediction model, that is, the drug preprocessing layer, the first drug local message passing layer, the first sequence context modeling layer and the first drug post-processing layer are sequentially performed to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the first drug post-processing layer. And the atomic cluster features of the SMILES sequence with a dimension of C×200 after the first drug post-processing layer (4-7) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic features of dimension N×200 corresponding to the SMILES sequence obtained in step (4-6) after being processed by the first drug post-processing layer are converted to The atomic cluster features of the SMILES sequence with a dimension of C×200 after the first drug post-processing layer Input to the first multi-scale feature update layer in the drug-target binding affinity prediction model to obtain the atomic group attention scores corresponding to the SMILES sequence with dimensions of C×5, C×200 and 32×200 respectively Cluster characteristics and drug molecular characteristics Where 5 is the number of attention heads; (4-8) For each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-5) is converted to a matrix with dimensions of 32×5×200 and 32×V. m ×5 residue cluster features and the residue cluster probability assignment matrix The atomic features of the SMILES sequence obtained in steps (4-6) with a dimension of N×200 after the first drug post-processing layer The atomic cluster features corresponding to the SMILES sequence obtained in steps (4-7) with dimensions of C×200 and 32×200 respectively and drug molecular characteristics The first fusion layer of the feature fusion module in the drug-target binding affinity prediction model is input to obtain the corresponding SMILES sequence with dimensions of 32×1×200 and 32×C m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix (4-9) For each amino acid sequence in each drug-protein data pair and each drug's SMILES sequence in the training set obtained in step (4-1), the residue cluster feature with a dimension of 32×5×200 corresponding to the amino acid sequence obtained in step (4-5) is converted into The dimensions of the SMILES sequence obtained in step (4-8) are 32×1×200 and 32×C respectively. m ×200 and 32×n m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix Input to the first fusion layer of the feature fusion module in the drug-target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and Its dimensions are 32×4×n m ×5, 32×4×C m ×5, 32×4×1×5, where 32 is the batch size, 4 is the number of attention heads, and C m and N m are the largest atomic group and the number of atoms in each drug, respectively, and 5 is the number of residue clusters in the current multi-scale feature update layer; (4-10) For each amino acid sequence and each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the fusion attention score corresponding to the drug-protein data pair obtained in step (4-9) is calculated. and Input to the first fusion layer of the feature fusion module in the drug-target binding affinity prediction model to obtain the protein attention scores corresponding to the amino acid sequence with dimensions of 32×5×4 and 160×200 respectively and residue cluster features The atomic cluster features corresponding to the SMILES sequence with dimensions of C×200, 32×200, and N×200 are Drug molecular characteristics and atomic features (4-11) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic cluster feature with a dimension of C×200 corresponding to the SMILES sequence obtained in step (4-10) is converted into The atomic features of the drug post-treatment layer corresponding to the SMILES sequence obtained in step (4-6) with a dimension of N×200 Input into the first atomic feature update layer of the drug-target binding affinity prediction model to obtain the atomic features of the SMILES sequence with a dimension of N×200 after being processed by the first atomic feature update layer (4-12) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature with a dimension of 160×200 corresponding to the amino acid sequence obtained in step (4-10) is converted into The residue features corresponding to the amino acid sequence obtained in step (4-4) with a dimension of V×200 after being processed by the first protein post-processing layer Input into the first residue feature update layer of the drug-target binding affinity prediction model to obtain the residue features corresponding to the amino acid sequence with a dimension of V×200 after being processed by the first residue feature update layer (4-13) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure corresponding to the amino acid sequence obtained in step (4-2) and the residue feature map corresponding to the amino acid sequence obtained in step (4-12) with a dimension of V×200 after being processed by the first residue feature update layer are added. The second protein feature extraction layer in the protein encoder module of the drug-target binding affinity prediction model is input, that is, the second protein local message passing layer, the second structured state modeling layer and the second protein post-processing layer are processed in sequence to obtain the residue features corresponding to the amino acid sequence with a dimension of V×200 after being processed by the second protein post-processing layer. (4-14) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue features corresponding to the amino acid sequence obtained in step (4-13) with a dimension of V×200 after being processed by the second protein post-processing layer are The second residue cluster clustering layer in the protein encoder module of the drug target binding affinity prediction model is input, that is, the second residue cluster construction layer and the second residue cluster pooling layer are sequentially performed to obtain the corresponding amino acid sequence with dimensions of 32×10×200 and 32×V m ×10 residue cluster features and residue cluster assignment probability matrix (4-15) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure diagram and atomic group structure diagram obtained in step (4-3) and the atomic features of the SMILES sequence obtained in step (4-11) with a dimension of N×200 after being processed by the first atomic feature update layer are converted into The second drug feature extraction layer in the drug encoder module in the drug-target binding affinity prediction model is input, that is, the second drug local message passing layer, the second sequence context modeling layer and the second drug post-processing layer are sequentially performed to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the second drug post-processing layer. And the atomic cluster features of the SMILES sequence with a dimension of C×200 after the second drug post-processing layer (4-16) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic features of dimension N×200 corresponding to the SMILES sequence obtained in step (4-15) after being processed by the second drug post-processing layer are The atomic cluster features of the second drug post-processing layer corresponding to the SMILES sequence with a dimension of C×200 The second multi-scale feature update layer is input into the drug-target binding affinity prediction model to obtain the atomic group attention scores corresponding to the SMILES sequence with dimensions of C×5, C×200 and 32×200 respectively. Atomic cluster characteristics and drug molecular characteristics Where 5 is the number of attention heads; (4-17) For each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-14) is converted to a matrix with dimensions of 32×10×200 and 32×V. m ×10 residue cluster features and the residue cluster probability assignment matrix The atomic features of the SMILES sequence obtained in step (4-15) with a dimension of N×200 after the second drug post-processing layer The atomic cluster features corresponding to the SMILES sequence obtained in step (4-16) with dimensions of C×200 and 32×200 respectively and drug molecular characteristics The second fusion layer of the feature fusion module in the drug-target binding affinity prediction model is input to obtain the corresponding SMILES sequence with dimensions of 32×1×200, 32×C m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix (4-18) For each amino acid sequence in each drug-protein data pair and each drug SMILES sequence in the training set obtained in step (4-1), the residue cluster feature with a dimension of 32×5×200 corresponding to the amino acid sequence obtained in step (4-14) is converted into The dimensions corresponding to the SMILES sequence obtained in step (4-17) are 32×1×200 and 32×C m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix Input to the second fusion layer of the feature fusion module in the drug-target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and Its dimensions are 32×4×N m ×10, 32×4×C m ×10, 32×4×1×10; (4-19) For each amino acid sequence in each drug-protein data pair and each drug SMILES sequence in the training set obtained in step (4-1), the fusion attention score corresponding to the drug-protein data pair obtained in step (4-18) is calculated. and Input to the second fusion layer of the feature fusion module in the drug-target binding affinity prediction model to obtain the protein attention scores corresponding to the amino acid sequence with dimensions of 32×10×4 and 320×200 respectively and residue cluster features The atomic cluster features corresponding to the SMILES sequence with dimensions of C×200, 32×200, and N×200 are Drug molecular characteristics and atomic features (4-20) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic cluster feature with a dimension of C×200 corresponding to the SMILES sequence obtained in step (4-19) is converted to The atomic features of the SMILES sequence obtained in step (4-15) with a dimension of N×200 after the second drug post-processing layer Input into the second atomic feature update layer in the drug-target binding affinity prediction model to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the second atomic feature update layer (4-21) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature with a dimension of 320×200 corresponding to the amino acid sequence obtained in step (4-19) is converted into The residue features corresponding to the amino acid sequence obtained in step (4-13) with a dimension of V×200 after being processed by the second protein post-processing layer Input into the second residue feature update layer in the drug-target binding affinity prediction model to obtain the residue features corresponding to the amino acid sequence with a dimension of V×200 after being processed by the second residue feature update layer (4-22) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein structure corresponding to the amino acid sequence obtained in step (4-2) and the residue feature map corresponding to the amino acid sequence obtained in step (4-21) with a dimension of V×200 after being processed by the second residue feature update layer are added. The third protein feature extraction layer in the protein encoder module of the drug-target binding affinity prediction model is input, that is, the third protein local message passing layer, the third structured state modeling layer and the third protein post-processing layer are processed in sequence to obtain the residue features corresponding to the amino acid sequence with a dimension of V×200 after being processed by the third protein post-processing layer. (4-23) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue features corresponding to the amino acid sequence obtained in step (4-22) with a dimension of V×200 after being processed by the third protein post-processing layer are The third residue cluster clustering layer in the protein encoder module of the drug target binding affinity prediction model is input, that is, the third residue cluster construction layer and the third residue cluster pooling layer are sequentially performed to obtain the corresponding amino acid sequence with dimensions of 32×20×200 and 32×V m ×20 residue cluster features and residue cluster assignment probability matrix (4-24) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the original molecular structure diagram and atomic group structure diagram obtained in step (4-3) and the atomic features of the SMILES sequence obtained in step (4-20) corresponding to the N×200 dimension after being processed by the second atomic feature update layer are updated. The third drug feature extraction layer in the drug encoder module in the drug-target binding affinity prediction model is input, that is, the third drug local message passing layer, the third sequence context modeling layer and the third drug post-processing layer are sequentially performed to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the third drug post-processing layer. And the atomic cluster characteristics of the SMILES sequence with a dimension of C×200 after the third drug post-processing layer (4-25) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic features of dimension N×200 corresponding to the SMILES sequence obtained in step (4-24) after processing by the third drug post-processing layer are The atomic cluster features of the SMILES sequence with a dimension of C×200 after the third drug post-processing layer are shown. Input to the third multi-scale feature update layer in the drug-target binding affinity prediction model to obtain the atomic group attention scores corresponding to the SMILES sequence with dimensions of C×5, C×200 and 32×200 respectively Atomic cluster characteristics and drug molecular characteristics Where 5 is the number of attention heads; (4-26) For each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-23) has dimensions of 32×20×200 and 32×V m ×20 residue cluster features and the residue cluster probability assignment matrix The atomic features of the SMILES sequence obtained in step (4-24) with a dimension of N×200 after the third drug post-processing layer The atomic cluster features corresponding to the SMILES sequence obtained in step (4-25) with dimensions of C×200 and 32×200 respectively and drug molecular characteristics The third fusion layer of the feature fusion module in the drug-target binding affinity prediction model is input to obtain the corresponding SMILES sequence with dimensions of 32×1×200, 32×C m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix (4-27) For each amino acid sequence in each drug-protein data pair and each drug SMILES sequence in the training set obtained in step (4-1), the residue cluster feature with a dimension of 32×20×200 corresponding to the amino acid sequence obtained in step (4-23) is converted into The dimensions of the SMILES sequence obtained in step (4-26) are 32×1×200 and 32×C respectively. m ×200 and 32×N m ×200 drug molecule feature matrix Cluster characteristic matrix and atomic feature matrix Input to the third fusion layer of the feature fusion module in the drug-target binding affinity prediction model to obtain the fusion attention score corresponding to the drug-protein data pair and Its dimensions are 32×4×N m ×20、32×4×C m ×20, 32×4×1×20; (4-28) For each amino acid sequence and each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the fusion attention score corresponding to the drug-protein data pair obtained in step (4-27) is and Input to the third fusion layer of the feature fusion module in the drug-target binding affinity prediction model to obtain the protein attention scores corresponding to the amino acid sequence with dimensions of 32×20×4 and 640×200 respectively and residue cluster features The atomic cluster features corresponding to the SMILES sequence with dimensions of C×200, 32×200, and N×200 are Drug molecular characteristics and atomic features (4-29) For each drug SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic cluster feature with a dimension of C×200 corresponding to the SMILES sequence obtained in step (4-28) is converted to The atomic features of the SMILES sequence obtained in step (4-24) with a dimension of N×200 after the third drug post-processing layer Input into the third atomic feature update layer in the drug-target binding affinity prediction model to obtain the atomic features corresponding to the SMILES sequence with a dimension of N×200 after being processed by the third atomic feature update layer (4-30) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the residue cluster feature with a dimension of 640×200 corresponding to the amino acid sequence obtained in step (4-28) is converted into The residue features corresponding to the amino acid sequence obtained in step (4-22) with a dimension of V×200 after being processed by the third protein post-processing layer Input into the third residue feature update layer in the drug-target binding affinity prediction model to obtain the residue features corresponding to the amino acid sequence with a dimension of V×200 after being processed by the third residue feature update layer (4-31) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the corresponding amino acid sequence obtained in step (4-5) with a dimension of 32×V m ×5 residue cluster assignment probability matrix The protein attention score corresponding to the amino acid sequence obtained in step (4-10) with a dimension of 32×5×4 The amino acid sequence obtained in step (4-14) corresponds to a dimension of 32×V m ×10 residue cluster assignment probability matrix The protein attention score corresponding to the amino acid sequence obtained in step (4-19) with a dimension of 32×10×4 The amino acid sequence obtained in step (4-23) corresponds to a dimension of 32×V m ×20 residue cluster assignment probability matrix The protein attention score corresponding to the amino acid sequence obtained in step (4-28) with a dimension of 32×20×4 And the residue features corresponding to the amino acid sequence obtained in step (4-30) with a dimension of V×200 after being processed by the third residue feature update layer Input to the protein attention pooling layer in the drug-target binding affinity prediction model to obtain the final protein feature vector of dimension 32×200 corresponding to the amino acid sequence (4-32) For each drug in each drug-protein data pair in the training set obtained in step (4-1), the atomic group attention score of dimension C×5 corresponding to the SMILES sequence obtained in step (4-7) is calculated. The atomic group attention score of dimension C×5 corresponding to the SMILES sequence obtained in steps (4-16) The atomic group attention score of dimension C×5 corresponding to the SMILES sequence obtained in step (4-25) And the atomic features corresponding to the SMILES sequence obtained in step (4-13) with a dimension of N×200 after being processed by the third atomic feature update layer Input to the drug attention pooling layer in the drug-target binding affinity prediction model to obtain the final drug feature vector corresponding to the SMILES sequence with a dimension of 32×200 (4-33) For each drug-protein data pair in the training set obtained in step (4-1), the final protein feature vector of dimension 32×200 corresponding to the amino acid sequence output in step (4-31) is converted to The final drug feature vector with a dimension of 32×200 corresponding to the SMILES sequence output in step (4-32) The final protein feature vector is input into the multi-layer perceptron module in the drug-target binding affinity prediction model. and the final drug feature vector The splicing is performed to obtain a splicing vector with a dimension of 32×400, and the splicing vector is processed by a two-layer MLP to obtain the drug-target binding affinity prediction value corresponding to the drug-protein data pair (4-34) For each drug-protein data pair in the training set obtained in step (4-1), the drug-target binding affinity data corresponding to the drug-protein data pair obtained in step (4-1) and the drug-target binding affinity prediction value corresponding to the drug-protein data pair obtained in step (4-33) are calculated. Calculating a loss function and using the loss function to train the drug-target binding affinity prediction model until the drug-target binding affinity prediction model converges, thereby obtaining a preliminarily trained drug-target binding affinity prediction model; (4-35) For each drug-protein data pair in the training set obtained in step (4-1), the drug-target binding affinity prediction model preliminarily trained in step (4-34) is tested using the test set obtained in step (4-1) until the value of the loss function reaches the optimal value, thereby obtaining a trained drug-target binding affinity prediction model.

10. A drug-target binding affinity prediction system based on multi-scale feature fusion, characterized in that: include: The first module is used to obtain drug-protein data pairs to be predicted and preprocess the drug-protein data pairs to be predicted to obtain preprocessed drug-protein data pairs, which include the simplified molecular linear input specification SMILES sequence of the drug, the amino acid sequence of the protein, and the drug-target binding affinity data; The second module is used to process the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained in the first module using the pre-trained evolutionary scaling model ESM to obtain an evolutionary language modeling representation and a contact probability matrix of the amino acid sequence, and connect the residue pairs in the contact probability matrix with a contact probability greater than or equal to a preset contact probability threshold as edges in the protein structure graph, thereby obtaining a protein structure graph corresponding to the amino acid sequence, where each node in the protein structure graph represents an amino acid and an edge represents the contact between two amino acids; and map each amino acid in the amino acid sequence to a feature vector with a dimension of 33, and the feature vectors corresponding to all amino acids constitute a residue feature matrix with a size of the length of the amino acid sequence × 33; The third module is used to parse the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained in the first module using the RDKit tool to obtain a molecular structure diagram corresponding to the SMILES sequence, obtain an atomic feature matrix based on the physical and chemical properties of each atom in the molecule in the molecular structure diagram, generate an atomic group set based on the molecular structure diagram corresponding to the obtained SMILES sequence, and generate an atomic group structure diagram corresponding to the SMILES sequence based on the connection relationship between atomic groups in the atomic group set; The fourth module is used to input the protein structure diagram and residue feature matrix corresponding to each amino acid sequence obtained in the second module, and the molecular structure diagram, atomic cluster structure diagram and atomic feature matrix corresponding to the SMILES sequence of each drug obtained in the third module into the pre-trained drug-target binding affinity prediction model to obtain the drug-target binding affinity prediction value.

Images