Geometric characterization-based drug target interaction prediction method

Abstract

The invention provides a drug target interaction method based on geometric characterization. The method comprises the following steps: acquiring drug structure features from a drug structure chart by using a geometric enhancement graph neural network GeoGNN; obtaining target point structure features from the target point structure diagram by using a geometric encoder; combining the drug comprehensive similarity, the target comprehensive similarity and the drug target incidence matrix to form a heterogeneous network, and performing feature extraction on the heterogeneous network by using a graph neural network GCN to obtain drug network features and target network features; performing multi-modal feature fusion on the drug structural features and the drug network features, and the target structural features and the target network features; extracting interaction information between the medicine and the target spot from the fused medicine multi-modal features and the fused target spot multi-modal features; based on the interaction information, better features of the drug and the target spot are obtained, and the probability of interaction between the drug and the target spot is predicted through a classifier. According to the method, the drug target interaction prediction accuracy is improved.

Description

Technical Field

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

[0002] Drug-Target Interaction (DTI) prediction is a technology that uses computational methods and models to predict the interaction between drugs and biological targets. DTI prediction can help researchers gain a deeper understanding of the drug's mechanism of action, discover new drug targets, and screen for potential drug candidates.

[0003] In recent years, with the development of deep learning technology, multimodal features based on sequence, structure and network features have been widely used in DTI prediction. At present, many DTI prediction models have adopted the fusion of features from multiple modalities. These features include but are not limited to the chemical structure, sequence information and related biological network features of drugs or targets. Through deep learning models, researchers can extract useful information from complex data, thereby improving the accuracy and reliability of DTI prediction.

[0004] Although existing DTI prediction models can use the structural information of drugs or targets for feature extraction, these models usually only represent the topological structure of molecules through atomic properties and the connection relationship between atoms, without fully capturing the complex three-dimensional geometric information and the spatial arrangement of atoms in the drug. In addition, existing multimodal feature fusion methods often simply splice the features of multiple modalities or fuse them by introducing learnable weight parameters, without considering how to effectively integrate these multimodal features to further improve the prediction performance. Therefore, it is particularly important to develop a new DTI prediction method that can effectively fuse multimodal features and fully consider 3D geometric information and atomic spatial arrangement. Summary of the invention

[0005] In view of this, the present invention provides a drug-target interaction prediction method based on geometric representation to improve the accuracy of drug-target interaction prediction.

[0006] To this end, the present invention provides the following technical solutions:

[0007] The present invention provides a drug-target interaction method based on geometric characterization, the method comprising:

[0008] Obtaining a drug structure diagram and a drug fingerprint from a drug's SMILES sequence; and obtaining drug structural features from the drug structure diagram using a geometrically enhanced graph neural network GeoGNN;

[0009] Obtain a target structure map from the Fasta sequence of the target; obtain target structural features from the target structure map using a geometric encoder;

[0010] The drug fingerprint similarity of the drugs and the target sequence similarity of the targets were calculated separately;

[0011] The drug-target association matrix was used to obtain the drug-target Gaussian similarity and the target-target Gaussian similarity;

[0012] Performing similarity fusion on the drug fingerprint similarity and the drug Gaussian similarity to obtain a comprehensive drug similarity;

[0013] Performing similarity fusion on the target sequence similarity and the target Gaussian similarity to obtain a comprehensive target similarity;

[0014] The comprehensive drug similarity, the comprehensive target similarity and the drug-target association matrix are combined to form a heterogeneous network, and the graph neural network GCN is used to extract features of the heterogeneous network to obtain drug network features and target network features;

[0015] Performing drug multimodal feature fusion on the drug structure feature and the drug network feature;

[0016] Performing target multimodal feature fusion on the target structure features and the target network features;

[0017] Extracting the interaction information between the drug and the target from the fused drug multimodal features and the fused target multimodal features;

[0018] Based on the extracted interaction information between drugs and targets, updated features of drugs and targets are obtained, and the interaction probability between drugs and targets is predicted by a classifier.

[0019] Further, calculating the drug fingerprint similarity of the drugs includes: calculating the drug fingerprint similarity using the Tanimoto coefficient.

[0020] Furthermore, calculating the target sequence similarity of the target includes: using the Smith-Waterman algorithm to calculate the target sequence similarity.

[0021] Furthermore, a target structure diagram is obtained from the Fasta sequence of the target, including: using a protein language model to process the Fasta sequence of the target to obtain the target structure diagram.

[0022] Furthermore, a drug structure diagram and a drug fingerprint are obtained from the SMILES sequence of the drug, including: using the Rdkit tool to process the SMILES sequence of the drug to obtain the drug structure diagram and the drug fingerprint.

[0023] Furthermore, the drug fingerprint similarity and the drug Gaussian similarity are similarly fused to obtain a comprehensive drug similarity, including:

[0024] Normalize the drug fingerprint similarity and drug Gaussian similarity;

[0025] The K-nearest neighbor algorithm is used to evaluate the local affinity between drugs in the drug fingerprint similarity matrix and the drug Gaussian similarity matrix, and the top 10% of local affinities are screened out, and the others are set to zero; this process is iterated continuously until a specific stopping condition is met;

[0026] After the iteration, the drug fingerprint similarity matrix and the drug Gaussian similarity matrix are added and their average values ​​are calculated;

[0027] Add the mean value matrix to its transposed matrix and calculate the mean value again to obtain the comprehensive drug similarity matrix;

[0028] The target sequence similarity and the target Gaussian similarity are similarly fused to obtain a comprehensive target similarity, including:

[0029] Normalize the target sequence similarity and target Gaussian similarity;

[0030] The K-nearest neighbor algorithm is used to evaluate the local affinity between targets in the target sequence similarity matrix and the target Gaussian similarity matrix, and the top 10% of local affinities are screened out, and the others are set to zero; this process is iterated continuously until a specific stopping condition is met;

[0031] After the iteration, the target sequence similarity matrix and the target Gaussian similarity matrix are added and their average is calculated;

[0032] The average matrix is ​​added to its transposed matrix and the average is calculated again to obtain the target comprehensive similarity matrix.

[0033] Furthermore, the drug structure graph includes an atom-atom graph and a bond-bond graph, wherein the atom-atom graph regards atoms in the drug as nodes and covalent bonds as edges, and the bond-bond graph regards covalent bonds as nodes and angles between covalent bonds as edges; using a geometrically enhanced graph neural network GeoGNN to obtain drug structural features from the drug structure graph, including: using GeoGNN to extract drug structural features from the atom-atom graph and the bond-bond graph, thereby incorporating the atom-bond-angle geometric relationship of the drug.

[0034] Furthermore, the target structure graph uses amino acid residues as nodes and the contact graph of residues as edges, and uses a geometric encoder to obtain target structural features from the target structure graph, including:

[0035] Use GCN to extract residue features from the target structure graph; GCN only considers sparse interactions and ignores the global interactions between all residue pairs. It updates the representation of each residue through the self-attention mechanism, calculates the attention weight using the query matrix and the bond matrix, and performs weighted summation on the value matrix through the weights to obtain the updated residue representation;

[0036] To incorporate the geometric information of the target, a local framework was constructed for each residue using carbon, nitrogen, and α carbon;

[0037] v i1 =x N(i) -x Cα(i) ;

[0038] v i2 =x C(i) -x Cα(i) ;

[0039] v i1 and v i2 is a vector determined by the coordinates of the nitrogen N, carbon C, and α-carbon Cα atoms of residue i;

[0040] u i1 ,u i2 =GramSchmidt(v i1 ,v i2 );

[0041] The GramSchmidt procedure is used to i1 and v i2 Orthogonalize to ensure they are orthogonal to each other in space;

[0042]

[0043] R i is the local frame of residue i, represented by the orthogonal vector u i1 and u i2 and their cross product, these vectors are normalized to form unit vectors;

[0044]

[0045] p i is the spatial representation of residue i, which is calculated by weighted summing the information of all other residues j;

[0046] When the entire structure rotates, the local frame R i and the relative position x Cα(i) -x Cβ(j) will also rotate accordingly, thus keeping the calculation result unchanged.

[0047] The updated residue representation and spatial representation are fused to obtain the final residue attention representation.

[0048] Furthermore, the drug structure feature and the drug network feature are fused into drug multimodal features, including:

[0049] The drug network features and drug structure features are concatenated, and three linear transformations are applied to obtain the query matrix, key matrix, and value matrix. The query matrix and key matrix are used to calculate the attention weights, including the correlation between network features and network features, the correlation between network features and structural features, the correlation between structural features and network features, and the correlation between structural features and structural features. The attention weights are used to weight the elements in the value matrix to obtain the fusion features of the drug.

[0050] The target structure feature and the target network feature are subjected to target multimodal feature fusion, including:

[0051] The target network features and target structural features are concatenated, and three linear transformations are applied to obtain the query matrix, key matrix and value matrix. The query matrix and key matrix are used to calculate the attention weights, including the correlation between network features and network features, the correlation between network features and structural features, the correlation between structural features and network features, and the correlation between structural features and structural features. The attention weights are used to weight the elements in the value matrix to obtain the fusion features of the target.

[0052] Furthermore, the interactive information between the drug and the target is extracted from the fused drug multimodal features and the fused target multimodal features, including:

[0053] For drugs, the query matrix is ​​obtained using drug features, the key matrix and value matrix are obtained using target features, and the attention weights are obtained through the query matrix and key matrix. Finally, the attention weights are used to weight the elements in the value matrix, thereby updating the drug features with the help of target features.

[0054] For the target, the query matrix is ​​obtained using the target features, the key matrix and the value matrix are obtained using the drug features, the attention weights are obtained through the query matrix and the key matrix, and finally, the attention weights are used to weight the elements in the value matrix, thereby updating the target features with the help of the drug features.

[0055] Advantages and positive effects of the present invention: Compared with common deep learning methods, the present invention can not only extract complex 3D geometric information, but also use an efficient multimodal feature fusion method to fuse structural features and network features, and also extract the interactive information between drugs and targets, and finally use a multi-layer perceptron (MLP) to predict drug-target interactions. The method of the present invention helps to improve the accuracy of drug-target interaction prediction, and provides valuable insights for actual disease diagnosis and treatment and subsequent drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0057] Figure 1 is a flow chart of a method for predicting drug-target interaction based on geometric representation in an embodiment of the present invention;

[0058] Figure 2 A flowchart of heterogeneous network construction in an embodiment of the present invention;

[0059] Figure 3 This is a diagram of the GeoGNN model architecture in an embodiment of the present invention;

[0060] Figure 4 It is a multimodal feature fusion module diagram in an embodiment of the present invention;

[0061] Figure 5 Schematic diagram of an interactive information extraction module in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0063] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0064] like Figure 1 As shown, a drug-target interaction prediction method based on geometric representation in an embodiment of the present invention mainly includes three parts: data preparation, model building and model testing.

[0065] (I) Data preparation:

[0066] For drugs, the Rdkit tool (RDKit is an open source chemical informatics and machine learning software suite written in C++ and Python and designed to process and analyze chemical data) is used to process its SMILES sequence (Simplified Molecular Input Line Entry System, a specification that explicitly describes the molecular structure using ASCII strings) to obtain drug structure diagrams and drug fingerprints MACCS Keys (Molecular ACCessSystem keys, a widely used two-dimensional structural fingerprint, mainly used in chemical informatics research fields such as molecular similarity comparison, molecular classification, molecular clustering and molecular screening), and the Tanimoto coefficient (also known as the generalized Jaccard similarity coefficient, is an indicator used to compare the similarity of two sets (usually molecules in chemical informatics)) is used to calculate the drug fingerprint similarity.

[0067] For the target, the protein language model (ESM, a method that uses deep learning technology to predict protein structure and function) is used to process its Fasta sequence (Fasta sequence is a text format widely used to represent nucleic acid (DNA or RNA) and protein sequences) to obtain the target structure diagram, and the Smith-Waterman algorithm (Smith-Waterman algorithm is a dynamic programming algorithm for sequence alignment) is used to calculate the target sequence similarity. The drug Gaussian similarity and target Gaussian similarity are obtained through the drug-target association matrix.

[0068] (II) Model construction:

[0069] Step 1: Similarity fusion.

[0070] Combine the two similarities of drugs and targets. Taking drugs as an example, normalize the fingerprint similarity and Gaussian similarity; then, use the K nearest neighbor (KNN) algorithm to evaluate the local affinity between drugs in these two similarity matrices, and screen out the top 10% of local affinities, setting the others to zero. This process will be iterated until a specific stopping condition is met. After the iteration, add the two similarity matrices and calculate their average. Finally, in order to obtain the comprehensive similarity matrix of the drug, add this average matrix to its transposed matrix and calculate the average again. For the target, the same method is used to calculate its comprehensive similarity matrix.

[0071] Step 2: Network feature extraction.

[0072] A drug-target heterogeneous network is constructed based on the comprehensive drug similarity, comprehensive target similarity and drug-target association matrix, such as Figure 2 As shown in Figure 2, GCN is used to learn drug network features and target network features from heterogeneous networks.

[0073] Step 3: Structural feature extraction.

[0074] For drugs, drug structure graphs include atom-atom graphs and bond-bond graphs. The atom-atom graph considers atoms in the drug as nodes and covalent bonds as edges, while the bond-bond graph considers covalent bonds as nodes and angles between covalent bonds as edges. GeoGNN is used to extract drug structural features from these two graphs, thereby incorporating the atom-bond-angle geometric relationship of the drug. The GeoGNN workflow is as follows: Figure 3 shown.

[0075] For the target, the target structure graph uses amino acid residues as nodes and the contact graph of residues as edges, and GCN is used to extract residue features from the target structure graph. Limited by computational complexity, GCN only considers sparse interactions and ignores the global interactions between all residue pairs. The representation of each residue is updated through the self-attention mechanism, the attention weight is calculated using the query matrix and the bond matrix, and the value matrix is ​​weighted summed by the weight to obtain the updated residue representation.

[0076] To incorporate the geometric information of the target, a local framework was constructed for each residue using carbon, nitrogen, and alpha carbon.

[0077] v i1 =x N(i) -x Cα(i) ;

[0078] vi2 =x C(i) -x Cα(i) ;

[0079] v i1 and v i2 is a vector defined by the coordinates of the nitrogen (N), carbon (C), and alpha carbon (Cα) atoms of residue i.

[0080] u i1 ,u i2 =GramSchmidt(v i1 ,v i2 );

[0081] The Gram-Schmidt procedure is an algorithm in linear algebra that converts a set of linearly independent vectors into an orthogonal or orthonormal basis. i1 and v i2 Orthogonalize to ensure they are orthogonal to each other in space.

[0082]

[0083] R i is the local frame of residue i, represented by the orthogonal vector u i1 and u i2 and their cross product, these vectors are normalized to form unit vectors.

[0084]

[0085] p i is the spatial representation of residue i, which is calculated by weighted summing the information of all other residues j.

[0086] When the entire structure rotates, the local frame R i and the relative position x Cα(i) -x Cβ(j) This ensures that no matter how the structure is rotated, the model can correctly handle the relative spatial relationship between amino acid residues and is not affected by the rotation.

[0087] The updated residue representation and spatial representation are fused to obtain the final residue attention representation, which not only considers the characteristics of the residues themselves, but also their relative positions in space, enhancing the model's sensitivity to geometric structure.

[0088] Step 4: Multimodal feature fusion.

[0089] The drug network features and drug structure features are concatenated, and three linear transformations are applied to obtain the query matrix, key matrix, and value matrix. The query matrix and key matrix are used to calculate the attention weights, including the correlation between network features and network features, the correlation between network features and structural features, the correlation between structural features and network features, and the correlation between structural features and structural features. Finally, the attention weights are used to weight the elements in the value matrix to obtain the fusion features of the drug. The fusion features of the target are obtained in the same way. This fusion method takes into account both intra-modal and inter-modal correlations. The multimodal feature fusion process is as follows: Figure 4 shown.

[0090] Step 5: Interaction information extraction.

[0091] In order to effectively extract interaction information, an interaction information extraction module is designed to estimate the contribution of targets to drugs or drugs to targets. Specifically, for drugs, the query matrix is ​​obtained using drug features, the key matrix and value matrix are obtained using target features, and the attention weights are obtained through the query matrix and the key matrix. Finally, the attention weights are used to weight the elements in the value matrix, thereby updating the drug features with the help of target features. For targets, the query matrix is ​​obtained using target features, the key matrix and value matrix are obtained using drug features, and the attention weights are obtained through the query matrix and the key matrix. Finally, the attention weights are used to weight the elements in the value matrix, thereby updating the target features with the help of drug features. The interaction information extraction process is as follows: Figure 5 shown.

[0092] (III) Model testing:

[0093] Step 1: Use the Rdkit tool to obtain the drug structure diagram and drug fingerprint from the SMILES sequence of the drug, use the protein language model (ESM) to obtain the target structure diagram from the Fasta sequence of the target, use the Tanimoto coefficient to calculate the drug fingerprint similarity, apply the Smith-Waterman algorithm to calculate the target sequence similarity, obtain the Gaussian similarity between the drug and the target through the drug-target association matrix, apply similarity fusion to the two similarities of the drug and the target respectively to obtain the comprehensive similarity of the drug and the target, use GeoGNN to obtain the drug structure features from the drug structure diagram, use the geometric encoder to obtain the target structure features from the target structure diagram, combine the comprehensive similarity of the drug and the target and the drug-target association matrix to form a heterogeneous network, then use GCN to obtain the network features of the drug and the target, fuse the structural features and network features of the drug and the target, then extract the interaction information between the drug and the target, and finally predict the interaction probability between the drug and the target through the multi-layer perceptron (MLP).

[0094] Step 2: Test various parameters and important modules that affect model performance. Observe the impact of structural feature extraction, network feature extraction, multimodal feature fusion module, and interactive information extraction module on model performance. Develop model variants that remove structural feature extraction, network feature extraction, multimodal feature extraction module, and interactive information extraction module, and test the impact of different modules on model performance.

[0095] Compared with common deep learning methods, the prediction method in the above embodiment can not only extract complex 3D geometric information, but also use an efficient multimodal feature fusion method to fuse structural features and network features, and also extract the interactive information between drugs and targets, and finally use a multilayer perceptron (MLP) to predict drug-target interactions. The method of the present invention helps to improve the accuracy of drug-target interaction prediction, and provides valuable insights for actual disease diagnosis and treatment and subsequent drug development.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for drug-target interaction based on geometric characterization, characterized in that: The method comprises: Obtaining a drug structure diagram and a drug fingerprint from a drug's SMILES sequence; and obtaining drug structural features from the drug structure diagram using a geometrically enhanced graph neural network GeoGNN; Obtain a target structure map from the Fasta sequence of the target; obtain target structural features from the target structure map using a geometric encoder; The drug fingerprint similarity of the drugs and the target sequence similarity of the targets were calculated separately; The drug-target association matrix was used to obtain the drug-target Gaussian similarity and the target-target Gaussian similarity; Performing similarity fusion on the drug fingerprint similarity and the drug Gaussian similarity to obtain a comprehensive drug similarity; Performing similarity fusion on the target sequence similarity and the target Gaussian similarity to obtain a comprehensive target similarity; The comprehensive drug similarity, the comprehensive target similarity and the drug-target association matrix are combined to form a heterogeneous network, and the graph neural network GCN is used to extract features of the heterogeneous network to obtain drug network features and target network features; Performing drug multimodal feature fusion on the drug structure feature and the drug network feature; Performing target multimodal feature fusion on the target structure features and the target network features; Extracting the interaction information between the drug and the target from the fused drug multimodal features and the fused target multimodal features; Based on the extracted interaction information between drugs and targets, updated features of drugs and targets are obtained, and the interaction probability between drugs and targets is predicted by a classifier.

2. A drug-target interaction method based on geometric characterization according to claim 1, characterized in that: Calculate the drug fingerprint similarity of drugs, including: Calculate drug fingerprint similarity using Tanimoto coefficient.

3. A drug-target interaction method based on geometric characterization according to claim 1, characterized in that: Calculating target sequence similarity of the target, including: calculating target sequence similarity using the Smith-Waterman algorithm.

4. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: Obtaining a target structure diagram from a Fasta sequence of the target, including: using a protein language model to process the Fasta sequence of the target to obtain a target structure diagram.

5. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: Obtaining a drug structure diagram and a drug fingerprint from a drug SMILES sequence, including: using an Rdkit tool to process the drug SMILES sequence to obtain a drug structure diagram and a drug fingerprint.

6. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: The drug fingerprint similarity and the drug Gaussian similarity are similarly fused to obtain a comprehensive drug similarity, including: Normalize the drug fingerprint similarity and drug Gaussian similarity; The K-nearest neighbor algorithm is used to evaluate the local affinity between drugs in the drug fingerprint similarity matrix and the drug Gaussian similarity matrix, and the top 10% of local affinities are screened out, and the others are set to zero; this process is iterated continuously until a specific stopping condition is met; After the iteration, the drug fingerprint similarity matrix and the drug Gaussian similarity matrix are added and their average values ​​are calculated; Add the mean value matrix to its transposed matrix and calculate the mean value again to obtain the comprehensive drug similarity matrix; The target sequence similarity and the target Gaussian similarity are similarly fused to obtain a comprehensive target similarity, including: Normalize the target sequence similarity and target Gaussian similarity; The K-nearest neighbor algorithm is used to evaluate the local affinity between targets in the target sequence similarity matrix and the target Gaussian similarity matrix, and the top 10% of local affinities are screened out, and the others are set to zero; this process is iterated continuously until a specific stopping condition is met; After the iteration, the target sequence similarity matrix and the target Gaussian similarity matrix are added and their average is calculated; The average matrix is ​​added to its transposed matrix and the average is calculated again to obtain the target comprehensive similarity matrix.

7. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: The drug structure graph includes an atom-atom graph and a bond-bond graph, wherein the atom-atom graph regards atoms in the drug as nodes and covalent bonds as edges, and the bond-bond graph regards covalent bonds as nodes and angles between covalent bonds as edges; using a geometrically enhanced graph neural network GeoGNN to obtain drug structure features from the drug structure graph, including: using GeoGNN to extract drug structure features from the atom-atom graph and the bond-bond graph, thereby incorporating the atom-bond-angle geometric relationship of the drug.

8. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: The target structure graph uses amino acid residues as nodes and the contact graph of residues as edges. The geometric encoder is used to obtain the target structural features from the target structure graph, including: Use GCN to extract residue features from the target structure graph; GCN only considers sparse interactions and ignores the global interactions between all residue pairs. It updates the representation of each residue through the self-attention mechanism, calculates the attention weight using the query matrix and the bond matrix, and performs weighted summation on the value matrix through the weights to obtain the updated residue representation; To incorporate the geometric information of the target, a local framework was constructed for each residue using carbon, nitrogen, and α carbon; v i1 =x N(i) -x Cα(i) ; v i2 =x C(i) -x Cα(i) ; v i1 and v i2 is a vector determined by the coordinates of the nitrogen N, carbon C, and α-carbon Cα atoms of residue i; u i1 ,u i2 =GramSchmidt(v i1 ,v i2 ); The GramSchmidt procedure is used to i1 and v i2 Orthogonalize to ensure they are orthogonal to each other in space; R i is the local frame of residue i, represented by the orthogonal vector u i1 and u i2 and their cross product, these vectors are normalized to form unit vectors; p i is the spatial representation of residue i, which is calculated by weighted summing the information of all other residues j; When the entire structure rotates, the local frame R i and the relative position x Cα(i) -x Cβ(j) will also rotate accordingly, thus keeping the calculation result unchanged. The updated residue representation and spatial representation are fused to obtain the final residue attention representation.

9. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: The drug structure feature and the drug network feature are subjected to drug multimodal feature fusion, including: The drug network features and drug structure features are concatenated, and three linear transformations are applied to obtain the query matrix, key matrix, and value matrix. The query matrix and key matrix are used to calculate the attention weights, including the correlation between network features and network features, the correlation between network features and structural features, the correlation between structural features and network features, and the correlation between structural features and structural features. The attention weights are used to weight the elements in the value matrix to obtain the fusion features of the drug. The target structure feature and the target network feature are subjected to target multimodal feature fusion, including: The target network features and target structural features are concatenated, and three linear transformations are applied to obtain the query matrix, key matrix and value matrix. The query matrix and key matrix are used to calculate the attention weights, including the correlation between network features and network features, the correlation between network features and structural features, the correlation between structural features and network features, and the correlation between structural features and structural features. The attention weights are used to weight the elements in the value matrix to obtain the fusion features of the target.

10. The method for drug-target interaction based on geometric characterization according to claim 1, characterized in that: Extract the interaction information between the drug and the target from the fused drug multimodal features and the fused target multimodal features, including: For drugs, the query matrix is ​​obtained using drug features, the key matrix and value matrix are obtained using target features, and the attention weights are obtained through the query matrix and key matrix. Finally, the attention weights are used to weight the elements in the value matrix, thereby updating the drug features with the help of target features. For the target, the query matrix is ​​obtained using the target features, the key matrix and the value matrix are obtained using the drug features, the attention weights are obtained through the query matrix and the key matrix, and finally, the attention weights are used to weight the elements in the value matrix, thereby updating the target features with the help of the drug features.

Images