Drug and target interaction prediction method based on multi-scale convolution feature fusion

Abstract

The invention discloses a drug and target interaction prediction method based on multi-scale convolution feature fusion, which comprises the following steps: acquiring drug molecule data, target protein data and drug and target interaction data, constructing a drug molecule map according to the drug molecule data, and coding a target protein sequence according to the target protein data; inputting the drug molecular map into a model, and obtaining drug features through a multi-scale map convolutional network and a dynamic gating attention mechanism; inputting a target protein sequence into the model, and obtaining target features through hierarchical cavity convolution and a bidirectional gating cycle unit; through multi-head cross attention, the drug features are aligned with the target features, local and global cross-modal fusion is carried out, and drug target fusion features are obtained; based on the drug target fusion features, outputting a drug and target interaction prediction probability; and training the model according to the drug and target interaction data and the prediction probability, and applying the trained model to drug and target interaction prediction.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics and deep learning technology, and more specifically, to a drug-target interaction prediction method based on a multi-scale convolutional neural network and a dynamic attention mechanism. Background Art

[0002] Among experimental methods, molecular docking technology is a typical one. This technology places molecules from a known three-dimensional structure database one by one at the active site of the target protein. By continuously optimizing the position, conformation, dihedral angles of rotatable bonds within the molecule, and the side chains and backbone of the receptor's amino acid residues, it explores the optimal conformation for the interaction between the receptor's small molecule compound and the target macromolecule, predicts its binding mode and affinity, and uses a scoring function to screen out ligands with the best affinity for the receptor and close to the natural conformation. Its principle is based on the shape complementarity and property complementarity of the receptor and ligand, while considering the electrostatic interaction, hydrogen bond interaction, van der Waals interaction, and hydrophobic interaction when the ligand binds to the receptor.

[0003] Among the early computational methods, ligand-based similarity search methods were widely used. From a two-dimensional perspective, the basic principle is that "molecules with similar structures may have similar biological activities because they target the same protein"; from a three-dimensional perspective, molecules with similar volumes may have the potential to effectively bind to the same or similar sized spaces in the active pocket of the protein (ligand-induced fit effects need to be considered). When using shape screening to predict the targets of small molecules, a database of small molecule ligands with protein target annotations is required. Subsequently, the overall shape similarity of the query molecule and each ligand in the database is measured separately, and the protein targets corresponding to those matching molecules with high similarity scores can be regarded as potential targets of the query molecule.

[0004] While traditional experimental methods can intuitively reveal drug-target interactions, they suffer from numerous limitations. For one thing, they are expensive. A comprehensive molecular docking experiment requires not only expensive laboratory equipment and specialized personnel, but also potentially the purchase of a large number of compound samples. Furthermore, experimental cycles are lengthy, often taking months or even years from experimental design to reliable results. This significantly prolongs the R&D cycle and increases costs.

[0005] Early computational methods also had shortcomings. For ligand-based similarity search methods, when the number of target-binding ligands is limited, the prediction effect is poor, making it difficult to accurately predict drug-target interactions. In addition, these methods do not mine complex biological data deeply enough, relying only on the structural information of the drug and target itself, ignoring the interactions between drug-target pairs and other rich semantic information in the biological system. Faced with the needs of large-scale drug research and development, the efficiency and accuracy of traditional methods are difficult to meet the requirements, and new technologies are urgently needed to improve the efficiency and accuracy of drug-target interaction prediction.

[0006] In summary, current DTI (Drug-Target Interaction) prediction methods have three major limitations: First, there is a bottleneck in drug feature representation. Traditional methods simplify drug molecules into fixed-dimensional fingerprint vectors (e.g., ECFP). This approach ignores the topological relationships between atoms in the molecular graph structure, resulting in loss of local chemical bond information. Although existing technologies use single-layer graph convolutional networks (GCNs) to extract molecular features, the use of fixed-size convolutional kernels fails to capture molecular substructures of varying scales, such as functional groups and ring systems. Second, target feature extraction is insufficient. Protein sequences are typically encoded using word embedding methods (e.g., Word2Vec) or unidirectional LSTMs. These methods struggle to model long-range residue dependencies and multi-level semantic features (e.g., α-helices and β-sheets). Finally, cross-modal fusion mechanisms are flawed. Existing technologies simply concatenate drug and target feature vectors, failing to fully consider the nonlinear interactions between the two, resulting in inadequate information fusion. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects and shortcomings of the existing technology and provide a drug-target interaction prediction method based on multi-scale convolutional feature fusion. By fusing the multi-scale local features of drug molecular graphs with the global semantic features of target protein sequences, combined with cross-modal feature alignment technology, the prediction accuracy and generalization ability are significantly improved, which is suitable for drug repositioning, virtual screening, side effect prediction and target discovery in new drug research and development.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A drug-target interaction prediction method based on multi-scale convolutional feature fusion includes the following steps:

[0010] S1: Obtain drug molecule data, target protein data, and drug-target interaction data, perform standardization preprocessing on the data, construct a drug molecule graph based on the drug molecule data, and encode the target protein sequence based on the target protein data;

[0011] S2: The drug molecule graph is input into the model, and the multi-scale features of the drug molecules are extracted through the multi-scale graph convolutional network. The multi-scale features of the drug molecules are fused through the dynamic gated attention mechanism to obtain the drug features;

[0012] S3: The target protein sequence is input into the model, and the sequence features of the target protein at different granularities are captured through hierarchical dilated convolution. The target protein sequence dependencies are captured through bidirectional gated recurrent units to obtain the target features.

[0013] S4: Through multi-head cross attention, drug features and target features are aligned, local-global cross-modal fusion is performed to obtain drug-target fusion features;

[0014] S5: Based on the drug-target fusion features, the predicted probability of drug-target interaction is output through residual connection, layer normalization and fully connected layer;

[0015] S6: Based on the drug-target interaction data and predicted probability, a joint loss function is used to train and optimize the model, and the trained model is used to predict the interaction between the drug to be tested and the target.

[0016] Furthermore, a drug molecule graph is constructed based on the drug molecule data, specifically:

[0017] Drug molecule data is in the form of SMILES strings. RDKit is used to parse the SMILES strings and generate atomic node features and chemical bond edge features of the drug molecules.

[0018] Atomic node features include atom type, atomic charge, valence state, hybridization state, and ring affiliation; chemical bond edge features include bond type, bond length, whether it is a ring bond, whether it is conjugated, and stereochemical properties;

[0019] According to the atomic node characteristics and chemical bond edge characteristics of the drug molecule, the graph structure of the drug molecule is constructed to obtain:

[0020] G drug =(V, E)

[0021] Where G drug Represents the drug molecule graph; V represents the atomic node feature set; E represents the chemical bond edge feature set.

[0022] Furthermore, the target protein sequence is encoded according to the target protein data, specifically:

[0023] The target protein data is an amino acid sequence in FASTA format. The amino acid sequence in FASTA format is mapped into a 128-dimensional vector through the embedding layer after one-hot encoding to generate the target protein sequence vector matrix, and the result is:

[0024]

[0025] Where, X target represents the target protein sequence vector matrix; L represents the target protein sequence length; R represents the real number set.

[0026] Furthermore, the multi-scale features of drug molecules are extracted through a multi-scale graph convolutional network, specifically:

[0027] The multi-scale graph convolutional network adopts a parallel multi-branch structure and uses three types of convolution kernels, 3×3, 5×5, and 7×7, to extract local features of atomic neighborhoods.

[0028] The update formula for the hidden state of node i is:

[0029]

[0030] Where, represents the hidden state of node i at the l+1th layer at scale k; ReLU represents the nonlinear activation function; N(i) represents the set of neighboring nodes of node i; represents the weight matrix of the lth layer at scale k; represents the hidden state of node j in layer l; e ij Represents the edge feature between node i and node j; Represents the feature concatenation operation; k = 1, 2, 3, corresponding to 3×3 convolution kernel, 5×5 convolution kernel, and 7×7 convolution kernel, respectively.

[0031] Furthermore, the multi-scale features of drug molecules are integrated through the dynamic gated attention mechanism, specifically:

[0032] The multi-scale features of drug molecules are adaptively aggregated through dynamic gated attention. The calculation formula is:

[0033]

[0034] Where g k represents the attention weight for scale k; σ represents the Sigmoid function; U and b represent learnable parameters; GlobalMaxPool represents the global maximum pooling operation; ) represents the output feature of the last layer, i.e., the Lth layer, of the graph convolutional network at scale k for node i; h drug Indicates drug characteristics.

[0035] Furthermore, hierarchical dilated convolution is used to capture sequence features of target proteins at different granularities, specifically:

[0036] A stack of dilated convolutional layers with expansion rates of 1, 2, and 4 is used to capture sequence features of target proteins at different granularities. The feature mapping formula is:

[0037]

[0038] Where C d (x) represents the output result after applying dilated convolution to the target protein sequence x; d represents the dilation rate; s represents the sliding step size; W i Represents the convolution kernel weight of node i.

[0039] Furthermore, the target protein sequence dependencies are captured through a bidirectional gated recurrent unit, specifically:

[0040] Capture the contextual dependencies of the target protein sequence and output the global semantic vector, i.e., the target feature h target ;

[0041] The hidden state update formula is:

[0042] h t =BiGRU(h t-1 ,x t )

[0043] Where h t represents the output hidden state at time t; BiGRU represents the bidirectional gated recurrent unit; h t-1 represents the output hidden state at time t-1; x t Represents the input vector of the target protein sequence at time t.

[0044] Furthermore, through multi-head cross attention, the drug features and target features are aligned, and local-global cross-modal fusion is performed to obtain the drug-target fusion features, specifically:

[0045] The drug characteristics h drug As a query condition, the target feature h target As the key value, the drug characteristics and target characteristics are aligned. The formula is:

[0046]

[0047] Where Attention() represents the attention function; Q represents the query matrix; K represents the key matrix; V represents the value matrix; softmax represents the activation function; d k Represents the dimension scaling factor.

[0048] Furthermore, the residual connection and layer normalization are formulated as follows:

[0049] h fusion =LayerNorm(h drug +Dropout(MHCA(h drug ,h target )))

[0050] Where h fusion Represents the final drug target fusion feature; LayerNorm represents layer normalization; Dropout represents random inactivation regularization; MHCA represents multi-head cross attention mechanism;

[0051] The predicted probability of drug-target interaction is as follows:

[0052] p=σ(W p ·h fusion +b p )

[0053] Where p represents the interaction probability; W p represents the weight matrix of the fully connected layer; b p represents the bias term of the fully connected layer;

[0054] The joint loss function is:

[0055] L=BCE(y,p)+λ||θ||2

[0056] Where y represents the true drug-drug interaction label; BCE() represents the binary cross entropy loss; λ represents the regularization coefficient; ||||2 represents the L2 norm; and θ represents the model parameter weight.

[0057] A drug-target interaction prediction system using multi-scale convolutional feature fusion, applying any of the above-mentioned drug-target interaction prediction methods using multi-scale convolutional feature fusion, comprising:

[0058] A data collection and processing unit is used to obtain drug molecule data, target protein data, and drug-target interaction data, perform standardization preprocessing on the data, construct a drug molecule graph based on the drug molecule data, and encode the target protein sequence based on the target protein data;

[0059] The drug feature extraction unit is used to extract multi-scale features of drug molecules based on the drug molecular graph of the input model through the multi-scale graph convolutional network module, and fuse the multi-scale features of drug molecules through the dynamic gated attention fusion module to obtain drug features;

[0060] The target feature extraction unit is used to capture the target protein sequence features of different granularities through the hierarchical dilated convolution module according to the target protein sequence of the input model, and to capture the target protein sequence dependency through the bidirectional gated recurrent unit module to obtain the target features;

[0061] The local-to-global fusion unit is used to align drug features with target features through a multi-head cross-attention module, perform local-global cross-modal fusion, and obtain drug-target fusion features;

[0062] The interaction prediction unit is used to output the predicted probability of drug-target interaction based on the drug-target fusion features through residual connection, layer normalization and fully connected layer;

[0063] The model training and optimization unit is used to train and optimize the model based on the drug-target interaction data and prediction probability using a joint loss function, and use the trained model to predict the interaction between the drug to be tested and the target.

[0064] Compared with the existing technology, the present invention significantly improves the prediction accuracy and generalization ability by fusing the multi-scale local features of drug molecule graphs with the global semantic features of target protein sequences, and combining cross-modal feature alignment technology. It is suitable for drug repositioning, virtual screening, side effect prediction and target discovery in new drug research and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Flowchart of the drug-target interaction prediction method based on multi-scale convolutional feature fusion.

[0066] Figure 2 Schematic diagram of the structure of the hierarchical void convolution HDCNN module and the bidirectional gated recurrent unit BiGRU module.

[0067] Figure 3 Schematic diagram of the multi-head cross attention MHCA mechanism.

[0068] Figure 4 Schematic diagram of the cross-modal fusion and interaction prediction architecture. DETAILED DESCRIPTION

[0069] The drug-target interaction prediction method based on multi-scale convolutional feature fusion of the present invention is further described below with reference to the accompanying drawings and specific examples.

[0070] See also Figure 1 The present invention discloses a drug-target interaction prediction method based on multi-scale convolutional feature fusion, comprising the following steps:

[0071] S1: Obtain drug molecule data, target protein data, and drug-target interaction data, perform standardization preprocessing on the data, construct a drug molecule graph based on the drug molecule data, and encode the target protein sequence based on the target protein data;

[0072] S2: The drug molecule graph is input into the model, and the multi-scale features of the drug molecules are extracted through the multi-scale graph convolutional network. The multi-scale features of the drug molecules are fused through the dynamic gated attention mechanism to obtain the drug features;

[0073] S3: The target protein sequence is input into the model, and the sequence features of the target protein at different granularities are captured through hierarchical dilated convolution. The target protein sequence dependencies are captured through bidirectional gated recurrent units to obtain the target features.

[0074] S4: Through multi-head cross attention, drug features and target features are aligned, local-global cross-modal fusion is performed to obtain drug-target fusion features;

[0075] S5: Based on the drug-target fusion features, the predicted probability of drug-target interaction is output through residual connection, layer normalization and fully connected layer;

[0076] S6: Based on the drug-target interaction data and predicted probability, a joint loss function is used to train and optimize the model, and the trained model is used to predict the interaction between the drug to be tested and the target.

[0077] By obtaining compound structures in SMILES format from databases such as PubChem and ChEMBL, covering small molecule drugs, natural products and candidate compounds, drug molecular data are obtained.

[0078] By obtaining protein FASTA sequence and 3D structure data from UniProt and PDB, including enzymes, receptors, ion channels and other categories, target protein data is obtained.

[0079] By extracting verified DTI affinity data (Ki, IC50, Kd values) from databases such as BindingDB, Davis, and KIBA, drug-target interaction data were obtained.

[0080] Data standardization involves: deduplication and debase processing of drug molecules and unification into the standardized SMILES format; removal of fragmented data of target protein sequences, retaining the full-length sequence and annotating the functional domains; logarithmic transformation of affinity values and normalization to the [0,1] range to construct a standardized labeling system.

[0081] Integrate chemical property databases (such as DrugBank), protein function annotations (such as GO terms), and pharmacological data (such as side effect targets) to form a multi-dimensional knowledge graph and construct a knowledge base.

[0082] Construct a drug molecule graph based on drug molecule data, specifically:

[0083] Drug molecule data is in the form of SMILES strings. RDKit is used to parse the SMILES strings and generate atomic node features and chemical bond edge features of the drug molecules.

[0084] Atomic node features include 12-dimensional features such as atom type, atomic charge, valence state, hybridization state, and ring affiliation. Chemical bond edge features include bond type, bond length, whether it is a ring bond, whether it is conjugated, and stereochemical properties.

[0085] According to the atomic node characteristics and chemical bond edge characteristics of the drug molecule, the graph structure of the drug molecule is constructed to obtain:

[0086] G drug =(V, E)

[0087] Where G drug Represents the drug molecule graph; V represents the atomic node feature set; E represents the chemical bond edge feature set.

[0088] In step S1, we also add virtual edges to connect The non-bonded atoms within the range can capture the spatial proximity effect and enhance the structural characteristics of the drug molecule graph.

[0089] Encode the target protein sequence according to the target protein data, specifically:

[0090] The target protein data is a FASTA format amino acid sequence, which is mapped into a 128-dimensional vector through the embedding layer after One-hot encoding.

[0091] The formula is:

[0092] x aa =Embedding(One-hot(aa i ))

[0093] Where x aa Represents the obtained amino acid embedding vector; Embedding represents the embedding encoding function, which maps the One-hot vector to a dense, trainable continuous vector; One-hot represents the One-hot encoding function, which maps the amino acid aa i Convert to a one-hot vector; aa i represents the i-th amino acid residue in the target protein sequence;

[0094] Generate the target protein sequence vector matrix and get:

[0095]

[0096] Where, X target represents the target protein sequence vector matrix; L represents the target protein sequence length; R represents the real number set.

[0097] Secondary structure tags, such as α-helix, β-sheet, and random coil, are extracted from PDB as additional feature channels for structural feature fusion.

[0098] The multi-scale features of drug molecules are extracted through the Multi-Scale GCN Multi-Scale Graph Convolutional Network (MSGCN), specifically:

[0099] The multi-scale graph convolutional network adopts a parallel multi-branch structure and uses three types of convolution kernels, 3×3, 5×5, and 7×7, to extract local features of atomic neighborhoods.

[0100] The update formula for the hidden state of node i is:

[0101]

[0102] Where, represents the hidden state of node i at the l+1th layer at scale k; ReLU represents the nonlinear activation function; N(i) represents the set of neighboring nodes of node i; represents the weight matrix of the lth layer at scale k; represents the hidden state of node j in layer l; e ij Represents the edge feature between node i and node j; Represents the feature concatenation operation; k = 1, 2, 3, corresponding to 3×3 convolution kernel, 5×5 convolution kernel, and 7×7 convolution kernel, respectively.

[0103] The multi-scale features of drug molecules are integrated through the Dynamic Gated Attention (DGA) mechanism, specifically:

[0104] The multi-scale features of drug molecules are adaptively aggregated through dynamic gated attention. The calculation formula is:

[0105]

[0106] Where g k represents the attention weight for scale k; σ represents the Sigmoid function; U and b represent learnable parameters; GlobalMaxPool represents the global maximum pooling operation; ) represents the output feature of the last layer, i.e., the Lth layer, of the graph convolutional network at scale k for node i, where L is the last layer; h drug Indicates drug characteristics.

[0107] See also Figure 2 , using Hierarchical Dilated CNN (HDCNN) to capture sequence features of target proteins at different granularities, specifically:

[0108] The hierarchical dilated convolution module uses a stack of dilated convolution layers with dilation rates of 1, 2, and 4 to capture sequence features of target proteins at different granularities. The feature mapping formula is:

[0109]

[0110] Where C d (x) represents what; d represents the expansion rate; s represents the sliding step size; W i Represents the convolution kernel weight of node i.

[0111] The target protein sequence dependencies are captured by the Bidirectional Gated Recurrent Unit (BiGRU), specifically:

[0112] The bidirectional gated recurrent unit module captures the contextual dependencies of the target protein sequence and outputs a global semantic vector, i.e., the target feature h target ;

[0113] The hidden state update formula is:

[0114] h t =BiGRU(h t-1 ,x t )

[0115] Where h t represents the output hidden state at time t, which combines the outputs of the forward and reverse GRU layers and represents the contextual semantic information of the current amino acid; BiGRU represents a bidirectional gated recurrent unit; h t-1 represents the output hidden state at time t-1; x t Represents the input vector of the target protein sequence at time t.

[0116] See also Figure 3 and Figure 4 Through Multi-Head Cross Attention (MHCA), drug features and target features are aligned and local-global cross-modal fusion is performed to obtain drug-target fusion features, specifically:

[0117] Multi-head cross attention module, the drug feature h drug As a query condition, the target feature h target As the key value, the drug characteristics and target characteristics are aligned. The formula is:

[0118]

[0119] Where Attention() represents the attention function; Q represents the query matrix; K represents the key matrix; V represents the value matrix; softmax represents the activation function; d k Represents the dimension scaling factor.

[0120] In this embodiment, the number of heads of multi-head cross attention is 8, and the dimensions of query matrix Q, key matrix K, and value matrix V are 64;

[0121] Residual connection and layer normalization, the formula is:

[0122] h fusion =LayerNorm(h drug +Dropout(MHCA(h drug ,h target )))

[0123] Where h fusion Represents the final drug target fusion feature; LayerNorm represents layer normalization; Dropout represents random inactivation regularization; MHCA represents multi-head cross attention mechanism;

[0124] The predicted probability of drug-target interaction is as follows:

[0125] p=σ(W p ·h fusion +b p )

[0126] Where p represents the interaction probability; W p represents the weight matrix of the fully connected layer; b p represents the bias term of the fully connected layer;

[0127] The joint loss function is:

[0128] L=BCE(y,p)+λ||θ||2

[0129] Where y represents the true drug-drug interaction label; BCE() represents the binary cross entropy loss; λ represents the regularization coefficient; ||||2 represents the L2 norm; and θ represents the model parameter weight.

[0130] Training and optimizing the model, including:

[0131] Data partitioning: The drug-target interaction dataset is divided into training, validation, and test sets, typically in a ratio of 7:1:2. The training set is used for model parameter learning, the validation set is used for adjusting model hyperparameters, and the test set is used for evaluating model performance.

[0132] Optimizer selection: AdamW optimizer is used to train the model. AdamW optimizer combines the adaptive learning rate adjustment capability and weight decay mechanism of Adam optimizer to effectively prevent model overfitting. The learning rate is set to 3×10 -4, the weight decay coefficient is 1×10 -5 .

[0133] Regularization strategy: During model training, we use the Dropout technique for regularization. A Dropout rate of 0.3 is set between hidden layers of the fully connected neural network to randomly drop some neuron connections, reducing the co-adaptability between neurons and improving the generalization ability of the model.

[0134] The present invention uses multiple metrics to evaluate model performance, including accuracy, recall, mean squared error, and consistency index. Validated through drug repositioning case studies, the model output predicted drug-target interaction probability (p=0.86) using both marketed drugs (such as metformin) and novel targets (such as mTOR pathway proteins). Experimental validation demonstrated an IC50 of 8.2 μM, with the predicted error between the actual interaction data and the prediction being less than 15%.

[0135] This invention significantly improves prediction accuracy and generalization ability by fusing the multi-scale local features of drug molecule graphs with the global semantic features of target protein sequences, combined with cross-modal feature alignment technology. It is suitable for drug repositioning, virtual screening, side effect prediction and target discovery in new drug research and development.

[0136] The present invention also discloses a drug-target interaction prediction system based on multi-scale convolutional feature fusion, which uses any of the above-mentioned drug-target interaction prediction methods based on multi-scale convolutional feature fusion, including:

[0137] A data collection and processing unit is used to obtain drug molecule data, target protein data, and drug-target interaction data, perform standardization preprocessing on the data, construct a drug molecule graph based on the drug molecule data, and encode the target protein sequence based on the target protein data;

[0138] The drug feature extraction unit is used to extract multi-scale features of drug molecules based on the drug molecular graph of the input model through the multi-scale graph convolutional network module, and fuse the multi-scale features of drug molecules through the dynamic gated attention fusion module to obtain drug features;

[0139] The target feature extraction unit is used to capture the target protein sequence features of different granularities through the hierarchical dilated convolution module according to the target protein sequence of the input model, and to capture the target protein sequence dependency through the bidirectional gated recurrent unit module to obtain the target features;

[0140] The local-to-global fusion unit is used to align drug features with target features through a multi-head cross-attention module, perform local-global cross-modal fusion, and obtain drug-target fusion features;

[0141] The interaction prediction unit is used to output the predicted probability of drug-target interaction based on the drug-target fusion features through residual connection, layer normalization and fully connected layer;

[0142] The model training and optimization unit is used to train and optimize the model based on the drug-target interaction data and prediction probability using a joint loss function, and use the trained model to predict the interaction between the drug to be tested and the target.

[0143] The drug-target interaction prediction system based on multi-scale convolution feature fusion of the present invention can execute the drug-target interaction prediction method based on multi-scale convolution feature fusion of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0144] The above description is a detailed description of the preferred embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications made under the technical spirit disclosed by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A drug-target interaction prediction method based on multi-scale convolutional feature fusion, characterized in that: The following steps are involved: S1: Obtain drug molecule data, target protein data, and drug-target interaction data, perform standardization preprocessing on the data, construct a drug molecule graph based on the drug molecule data, and encode the target protein sequence based on the target protein data; S2: The drug molecule graph is input into the model, and the multi-scale features of the drug molecules are extracted through the multi-scale graph convolutional network. The multi-scale features of the drug molecules are fused through the dynamic gated attention mechanism to obtain the drug features; S3: The target protein sequence is input into the model, and the sequence features of the target protein at different granularities are captured through hierarchical dilated convolution. The target protein sequence dependencies are captured through bidirectional gated recurrent units to obtain the target features. S4: Through multi-head cross attention, drug features are aligned with target features, local-global cross-modal fusion is performed to obtain drug-target fusion features; S5: Based on the drug-target fusion features, the predicted probability of drug-target interaction is output through residual connection, layer normalization and fully connected layer; S6: Based on the drug-target interaction data and predicted probability, a joint loss function is used to train and optimize the model, and the trained model is used to predict the interaction between the drug to be tested and the target.

2. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 1, characterized in that: Construct a drug molecule graph based on drug molecule data, specifically: Drug molecule data is in the form of SMILES strings. RDKit is used to parse the SMILES strings and generate atomic node features and chemical bond edge features of the drug molecules. Atomic node features include atom type, atomic charge, valence state, hybridization state, and ring affiliation; chemical bond edge features include bond type, bond length, whether it is a ring bond, whether it is conjugated, and stereochemical properties; According to the atomic node characteristics and chemical bond edge characteristics of the drug molecule, the graph structure of the drug molecule is constructed to obtain: G drug =(V,E) Where G drug Represents the drug molecule graph; V represents the atomic node feature set; E represents the chemical bond edge feature set.

3. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 1, characterized in that: Encode the target protein sequence according to the target protein data, specifically: The target protein data is an amino acid sequence in FASTA format. The amino acid sequence in FASTA format is mapped into a 128-dimensional vector through the embedding layer after one-hot encoding to generate the target protein sequence vector matrix, and the result is: Where, X target represents the target protein sequence vector matrix; L represents the target protein sequence length; R represents the real number set.

4. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 1, characterized in that: The multi-scale features of drug molecules are extracted through a multi-scale graph convolutional network, specifically: The multi-scale graph convolutional network adopts a parallel multi-branch structure and uses three types of convolution kernels, 3×3, 5×5, and 7×7, to extract local features of atomic neighborhoods. The update formula for the hidden state of node i is: Where, represents the hidden state of node i at the l+1th layer at scale k; ReLU represents the nonlinear activation function; N(i) represents the set of neighboring nodes of node i; represents the weight matrix of the lth layer at scale k; represents the hidden state of node j in layer l; e ij Represents the edge feature between node i and node j; Represents the feature concatenation operation; k = 1, 2, 3, corresponding to 3×3 convolution kernel, 5×5 convolution kernel, and 7×7 convolution kernel, respectively.

5. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 4, characterized in that: The multi-scale features of drug molecules are integrated through the dynamic gated attention mechanism, specifically: The multi-scale features of drug molecules are adaptively aggregated through dynamic gated attention. The calculation formula is: Where g k represents the attention weight for scale k; σ represents the Sigmoid function; U and b represent learnable parameters; GlobalMaxPool represents the global maximum pooling operation; ) represents the output feature of the last layer, i.e., the Lth layer, of the graph convolutional network at scale k for node i; h drug Indicates drug characteristics.

6. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 5, characterized in that: Hierarchical dilated convolution is used to capture sequence features of target proteins at different granularities, specifically: A stack of dilated convolutional layers with dilation rates of 1, 2, and 4 is used to capture sequence features of target proteins at different granularities. The feature mapping formula is: Where C d (x) represents the output result after applying dilated convolution to the target protein sequence x; d represents the dilation rate; s represents the sliding step size; W i Represents the convolution kernel weight of node i.

7. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 6, characterized in that: The target protein sequence dependencies are captured through a bidirectional gated recurrent unit, specifically: Capture the contextual dependencies of the target protein sequence and output the global semantic vector, i.e., the target feature h target ; The hidden state update formula is: h t =BiGRU(h t-1 ,x t ) Where h t represents the output hidden state at time t; BiGRU represents the bidirectional gated recurrent unit; h t-1 represents the output hidden state at time t-1; x t Represents the input vector of the target protein sequence at time t.

8. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 7, characterized in that: Through multi-head cross attention, the drug features are aligned with the target features, and local-global cross-modal fusion is performed to obtain the drug-target fusion features, specifically: The drug characteristics h drug As a query condition, the target feature h target As the key value, the drug characteristics and target characteristics are aligned. The formula is: Where Attention() represents the attention function; Q represents the query matrix; K represents the key matrix; V represents the value matrix; softmax represents the activation function; d k Represents the dimension scaling factor.

9. The drug-target interaction prediction method based on multi-scale convolutional feature fusion according to claim 8, characterized in that: Residual connection and layer normalization, the formula is: h fusion =LayerNorm(h drug +Dropout(MHCA(h drug ,h target ))) Where h fusion Represents the final drug target fusion feature; LayerNorm represents layer normalization; Dropout represents random inactivation regularization; MHCA represents multi-head cross attention mechanism; The predicted probability of drug-target interaction is as follows: p=σ(W p ·h fusion +b p ) Where p represents the interaction probability; W p represents the weight matrix of the fully connected layer; b p represents the bias term of the fully connected layer; The joint loss function is: L=BCE(y,p)+λ||θ||2 Where y represents the true drug-drug interaction label; BCE() represents the binary cross entropy loss; λ represents the regularization coefficient; ||||2 represents the L2 norm; and θ represents the model parameter weight.

10. A drug-target interaction prediction system based on multi-scale convolutional feature fusion, characterized in that: A method for predicting drug-target interactions using the multi-scale convolutional feature fusion method according to any one of claims 1 to 9, comprising: A data collection and processing unit is used to obtain drug molecule data, target protein data, and drug-target interaction data, perform standardization preprocessing on the data, construct a drug molecule graph based on the drug molecule data, and encode the target protein sequence based on the target protein data; The drug feature extraction unit is used to extract multi-scale features of drug molecules based on the drug molecular graph of the input model through the multi-scale graph convolutional network module, and fuse the multi-scale features of drug molecules through the dynamic gated attention fusion module to obtain drug features; The target feature extraction unit is used to capture the target protein sequence features of different granularities through the hierarchical dilated convolution module according to the target protein sequence of the input model, and to capture the target protein sequence dependency through the bidirectional gated recurrent unit module to obtain the target features; The local-to-global fusion unit is used to align drug features with target features through a multi-head cross-attention module, perform local-global cross-modal fusion, and obtain drug-target fusion features; The interaction prediction unit is used to output the predicted probability of drug-target interaction based on the drug-target fusion features through residual connection, layer normalization and fully connected layer; The model training and optimization unit is used to train and optimize the model based on the drug-target interaction data and prediction probability using a joint loss function, and use the trained model to predict the interaction between the drug to be tested and the target.

Images