Method for constructing drug-protein interaction prediction model based on multimodal fusion

Abstract

This invention belongs to the field of deep learning technology and discloses a method for constructing a drug-protein interaction prediction model based on multimodal fusion. Due to the lack of three-dimensional protein data, this invention uses the AlphaFold 2 prediction model to accurately complete the three-dimensional protein data. Regarding the modal fusion problem, this invention fuses drug molecules at a finer-grained atomic level, further enhancing the multimodal fusion effect. Furthermore, Cross-Attention is used to fuse the processed 1D and 3D protein data. Using Cross-Attention for multimodal fusion can precisely capture the interaction relationships between modalities. In summary, this invention improves the performance of the drug-protein interaction prediction (DTI) task by designing a multimodal fusion method.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and in particular relates to a method for constructing a drug-protein interaction prediction model based on multimodal fusion. Background Technology

[0002] With the development of internet technology, the technology in the field of drug discovery is constantly being updated and iterated. Drug-target interaction (DTI) refers to the binding between a drug and its target, which is crucial for discovering new drugs and reusing existing ones. Numerous methods based on deep learning technology have been proposed to solve drug-target interaction prediction, and research shows that deep learning-based methods can significantly improve the efficiency of drug discovery, which is of great significance. Deep learning computational methods typically require the 3D structure of the target protein to evaluate drug-target interactions. However, the 3D structure of many proteins is unknown, and obtaining this structure is expensive. Traditional techniques cannot adequately address the problem of missing protein 3D structures, which has led to some current deep learning computational methods being unable to fully learn protein structural information for drug-target interaction research. With the advent of large-scale models, the contributions of AlphaFold (AlphaFold Protein Structure Database) in protein structure prediction have solved the problem of missing protein 3D structures. Traditional deep learning techniques have several significant shortcomings in drug-target interaction (DTI) tasks, primarily in the following aspects: First, traditional deep learning methods typically utilize features from only a single modality. For example, the proposed DLM-DTI (Dual Language Model-Drug Target Interaction) algorithm focuses only on the sequence features of drugs and the sequence information of proteins. In this case, because it only focuses on one modality, it cannot comprehensively capture the multi-dimensional feature information of drugs or targets. For instance, using only molecular sequences cannot capture the three-dimensional spatial information of drugs, while using only protein sequences makes it difficult to obtain their structural information, resulting in information loss. Second, most existing multimodal fusion techniques focus on processing each modality separately and then performing coarse-grained fusion at the molecular level. For example, the proposed CAT-DTI (cross-attention and Transformer-Drug Target Interaction) model processes molecular graph features and protein sequence features separately before performing multimodal fusion. However, individual processing does not capture the inherent correlation between 2D and 3D data in the initial data.

[0003] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0004] Deep learning computational methods typically require the 3D structure of the target protein to assess drug-target interaction (DTI). However, the 3D structure of many proteins is unknown, and obtaining the 3D structure is expensive. Traditional techniques rarely use 3D structural data because they cannot adequately address the problem of missing 3D structures. Furthermore, multimodal techniques are less commonly used in the field of drug-target interaction (DTI), resulting in limited applications of multimodal methods, particularly in fine-grained modality fusion. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0006] This invention is implemented as follows: the method for constructing a drug-protein interaction prediction model based on multimodal fusion includes:

[0007] Step 1: Search for proteins in the dataset based on their codes. For proteins not included in the database, use the AlphaFold 2 (AlphaFold Protein Structure Database 2.0) model to perform high-accuracy prediction and completion.

[0008] Step 2, Data Preprocessing: The three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, for the SMILES structure of the drug, the RDKit tool is first used to convert it into a molecular graph and calculate the shortest path distance and edge type. Then, the RDKit tool is used to generate three-dimensional coordinates based on the SMILES structure of the molecule and calculate the Euclidean distance to generate a distance matrix.

[0009] Step 3: Fuse the shortest path distance, edge type, and Euclidean distance calculated from the three-dimensional coordinates of the molecule in the two-dimensional mode. Construct a mixing matrix through random sampling and inject the mixing matrix into the self-attention module to enhance the expression of the relationship between atomic pairs, similar to relative position encoding.

[0010] Step 4: The distance graph of the protein is processed using GCN (Graph Convolutional Network), while the protein sequence data is processed using CNN.

[0011] Step 5: The processed protein 1D data and protein 3D data are fused using Cross-Attention (Cross-Attention Mechanism), and the fused data and the molecular mixing matrix are spliced ​​together as the final input for drug and target prediction.

[0012] Step 6: The fused drug molecule features and protein features are spliced ​​together, and then the spliced ​​features are input into a multilayer perceptron for processing. Finally, the binary cross-entropy loss function is used for model training.

[0013] Furthermore, the specific process of step 1 is as follows:

[0014] Step 1.1: Obtain protein sequence data from the original dataset;

[0015] Step 1.2: Input the protein sequence data into AlphaFold2. AlphaFold2 generates a .pdb file of the protein 3D structure by predicting the torsion angle, frame (a local reference frame that defines the position and orientation of each amino acid residue in the protein).

[0016] Furthermore, the specific process of step 2 is as follows:

[0017] Step 2.1: First, for the protein, calculate the specific path for each pair of residues: Generate an initial distance matrix D of size N×N. Initially, for directly connected residue pairs (i,j), the distance is the Euclidean distance between them. If the atomic pairs are not directly connected, the initial distance is set to infinity. Use Dijkstra's algorithm (Dijkstra's Shortest Path Algorithm) to calculate the shortest path in D, where D... [i][j] This represents the shortest distance from residue i to residue j. For each pair of residues (i,j), the shortest distance matrix is ​​updated to represent the total Euclidean distance along the shortest path. The shortest distance matrix is ​​calculated for drug molecules. Update and Iteration Where i, j, and k are all different atoms, using This represents the shortest distance between atom i and atom j;

[0018] Step 2.2: Construct an edge matrix To represent the bond types between atoms in a drug molecule, a feature vector is defined: each edge has a type E. n It has a set of characteristics that represent the type of chemical bond (single bond, double bond, ring), and calculates the edge type matrix. Where N is the number of edges on the shortest path from atom i to j. It is the number of characteristic features (types of chemical bonds) of the nth edge on the path. It is related to each edge feature e n Associated learnable weight parameters, It is the characteristic matrix; T is the transpose of the matrix;

[0019] Step 2.3: Using the Euclidean distance formula Generate a matrix where each element [i,j] represents the distance from atom i to atom j, where the coordinates of each atom can be represented as (x1, y1, z1). This represents the spatial distance between atom pairs (i,j) in the 3D molecular structure; this distance is encoded by the Gaussian kernel function using the following formula.

[0020] Where d ij It is the Euclidean distance between atomic pairs (i,j), A(d; γ,β) = γ.d + β. The parameters γ and β are learnable. Where μ and σ are Gaussian kernel functions.

[0021] It is a set of basic kernel vectors. and Let W be the learnable weights and W be the weight matrix.

[0022] Furthermore, the specific process of step 3 is as follows:

[0023] Step 3.1: Obtain the different relation matrices mentioned above: To integrate them into a new single relation matrix, we first define a multinomial distribution S, where the probability vector p = (p1, p2, p3) represents the probability of the three relations appearing in the fusion matrix;

[0024] Step 3.2: For each position (i,j) in the matrix, randomly sample S from the distribution S to obtain Si. ij ∈{1,2,3}, and then determine the relation matrix value corresponding to that position, using the formula: Where the function ||S ij =K controls the value, if S ij =1, then choose If S ij =2, then choose If S ij =3, then choose The multimodal fusion matrix is ​​constructed through the above steps.

[0025] Step 3.3: Generated Modality Fusion Matrix Injected into the self-attention module to enhance the expression of relationships between atomic pairs, similar to relative position encoding, as shown in the following formula: Query, Key, and Value Space: Q = F P .W Q +b Q K = F P .W K +b K V = F P .W V +b V in It is a learnable weight matrix. Then, the attention weights are calculated using Query and Key to obtain a matrix M with n*d atomic features, which is used as the input for the final DTI prediction.

[0026] Furthermore, the specific process of step 4 is as follows:

[0027] Step 4.1: For the protein distance map, in order to extract the features of the protein distance map, GCN is used for processing. An input map is constructed based on its characteristics. Each GCN layer captures information about neighboring nodes by updating the features of each node. For the l-th layer of graph convolution, the feature update formula for each node is:

[0028] in It is the adjacency matrix of the distance graph, plus the identity matrix. that is The degree matrix, H (l) W is the node feature of layer l. (l) This is the weight matrix for learning, and σ is the activation function. Each convolution updates the features of the nodes, gradually integrating more neighbor information, capturing local structural information, and then performing max pooling to obtain an N*d feature matrix S. E , where N is the number of amino acids and d is the final feature dimension;

[0029] Step 4.2: For a one-dimensional protein sequence, denote the input feature matrix of the protein sequence as... A three-layer protein feature encoder is used, each layer including self-attention, a 1DCNN (Convolutional Neural Network), and a feedforward neural network. The self-attention layer is based on the input feature matrix F. p Through a linear layer, the input matrix F p Projecting onto the query, key, and value space: Q = F P .W Q +b Q K = F P .W K +bK V = F P .W V +b V ,in It is a learnable weight matrix. Then, attention weights are calculated using the Query and Key, and then applied to the Value. Here d k This is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally a linear transformation is applied to obtain the final output: X = Concat(head1,...,head).W M Q in which W M These are learnable weights, and the final output feature X M The dimension of the input matrix is ​​the same as that of the input matrix, both being l. p ×D p The output features include local information for each location and global contextual dependencies.

[0030] Furthermore, the specific process of step 5 is as follows:

[0031] Step 5.1, the protein 1D structure feature matrix is ​​as follows: The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is as follows: Calculate the 1D structural features and Euclidean distance map features respectively:

[0032] Query, Key, and Value of 1D Structural Features: Q 1D =X M ×W Q ,K 1D =X M ×W K V 1D =X M ×W V Query, Key, and Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K V Euc =S E ×W V ,in The weight matrix of the linear layer;

[0033] Step 5.2: Calculate the attention weight matrix using Query and Key to capture the interaction information between the 1D structure and the Euclidean distance map features.

[0034]

[0035] The outputs of each attention point are concatenated and subjected to a linear transformation to obtain the fused 1D structural features and Euclidean distance feature representations:

[0036]

[0037] The fused feature Z 1D and Z Euc Fusing with the original input features to further capture local and global information:

[0038]

[0039] Step 5.3: Perform global pooling on the fused 1D structural features and Euclidean distance map features respectively, and concatenate them into the final joint feature representation:

[0040]

[0041]

[0042] The joint features are represented as:

[0043] This joint feature vector f contains information about the interaction between the drug and the protein, and serves as the input for the final DTI (Drug-Target Interaction) prediction;

[0044] In step 6:

[0045] Step 6.1: Represent the drug's features using the matrix M generated in the previous steps. This matrix has dimensions N×d, where N is the number of atoms in the drug and d is the feature dimension of each atom. The joint feature vector f obtained in the previous steps has dimensions l. p ×D p Then feature concatenation is performed, which can be represented as: F concat =Concat(M,F) forms a new matrix F. concat ;

[0046] Step 6.2: Combine the concatenated feature matrix F concat The input is then fed into a multilayer perceptron (MLP) for further processing. The calculation process is as follows:

[0047] H1=σ(W1·F concat +b1)

[0048] H2=σ(W2·H1+b2) ......

[0050] H L =σ(WL ·H L-1 +b L )

[0051] Among them, W i b is the weight matrix of the i-th layer. i It is the bias, and σ is the activation function ReLU (RectifiedLinearUnit).

[0052] Step 6.3: Use the binary cross-entropy loss function (BCE loss) to calculate the model output and the true label. Assume the final output of the model is... Given the predicted probability and the true label y, the loss function can be expressed as:

[0053]

[0054] Where N is the number of samples, y i It is the true label of the i-th sample. That is the corresponding predicted probability.

[0055] Another objective of this invention is to provide a system for constructing a drug-protein interaction prediction model based on multimodal fusion, comprising:

[0056] The prediction completion module is used to find proteins in the dataset based on their codes. For proteins not included in the library, this module performs high-accuracy prediction completion using the AlphaFold 2 model.

[0057] The data preprocessing module is used for data preprocessing: the three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, for the SMILES structure of the drug, the RDKit tool is first used to convert it into a molecular graph and calculate the shortest path distance and edge type. Then, the RDKit is used to generate three-dimensional coordinates based on the SMILES structure of the molecule and calculate the Euclidean distance to generate a distance matrix.

[0058] The fusion module is used to fuse the shortest path distance, edge type, and Euclidean distance of the molecular two-dimensional mode. It constructs a mixing matrix through random sampling and injects the mixing matrix into the self-attention module to enhance the expression of the relationship between atomic pairs, similar to relative position encoding.

[0059] The processing module is used to process the distance map of proteins using GCN, and to process the protein sequence data using CNN.

[0060] The stitching module is used to fuse the processed protein 1D data and protein 3D data using Cross-Attention, and stitches the fused data with the molecular mixing matrix as the final input for drug and target prediction.

[0061] The training module is used to concatenate the fused drug molecule features with protein features, then input the concatenated features into a multilayer perceptron for processing, and finally use the binary cross-entropy loss function for model training.

[0062] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0063] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0064] Another objective of this invention is to provide an information data processing terminal for implementing the drug-protein interaction prediction model construction system based on multimodal fusion.

[0065] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:

[0066] First, due to the lack of 3D protein data, this invention uses the prediction model AlphaFold 2 to complete the 3D protein data with high accuracy. Addressing the multimodal fusion problem, drug molecules are fused at a finer atomic level, further enhancing the multimodal fusion effect. Furthermore, Cross-Attention is used to fuse the processed 1D and 3D protein data. Using Cross-Attention for multimodal fusion can precisely capture the interaction relationships between modalities. In summary, this invention improves the performance of the DTI task by designing a multimodal fusion method.

[0067] The technical solution of this invention solves the problems of insufficient data utilization, low prediction accuracy, and insufficient model generalization in drug-protein interaction prediction in industrial applications. Traditional methods usually rely only on single-modality drug or protein data (such as molecular sequences or three-dimensional structures) and ignore the integration of multimodal information, resulting in inaccurate interaction prediction results and limiting their application in actual drug development.

[0068] This invention utilizes a multimodal fusion method to combine the two-dimensional and three-dimensional structural features of drugs with the sequence and three-dimensional structural data of proteins, significantly improving the model's data utilization and expressive power. Specifically, this invention not only uses AlphaFold 2 to complete data when protein structural data is missing, but also incorporates shortest path distance, edge type, and three-dimensional distance matrix into drug molecule processing, enhancing the expressive power of relationships between drug molecules. The introduction of a self-attention mechanism enables the model to more accurately capture the atomic relationships and interaction patterns within drug molecules, significantly improving prediction accuracy.

[0069] Furthermore, this invention further integrates the 1D and 3D features of proteins through Cross-Attention and concatenates them with the drug's hybrid matrix before inputting them into the model. This deep fusion strategy enhances the synergistic effect of different modal features, enabling the model to more comprehensively and accurately characterize the complex interactions between drugs and proteins. Finally, under the processing of a multilayer perceptron, the model is optimized using a binary cross-entropy loss function, further improving the accuracy and robustness of interaction prediction.

[0070] Therefore, this invention represents a significant technological advancement in industrial applications, not only effectively improving the prediction accuracy of drug-protein interactions but also possessing strong generalization capabilities. This technical solution has broad application prospects in new drug development, drug screening, and precision medicine, providing strong support for accelerating drug discovery and personalized treatment.

[0071] Secondly, as supporting evidence of the inventiveness of this invention, it is also reflected in the following important aspects:

[0072] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0073] The technical solution of this invention provides a highly efficient and accurate drug-protein interaction prediction model, which is expected to generate significant economic benefits in drug development, disease treatment, and personalized medicine. By improving the accuracy and efficiency of drug screening, it can significantly reduce drug development costs and time, accelerating the new drug launch process. Furthermore, this model can provide researchers with deeper biological insights, promoting the discovery of new drug targets, thereby driving innovation and development throughout the pharmaceutical industry. Its commercial value lies in reducing drug development failure rates, enhancing drug market competitiveness, and providing more effective personalized treatment options for different patients.

[0074] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0075] This invention addresses the problems of incomplete data, low prediction accuracy, and low computational efficiency in traditional drug-protein interaction prediction methods by combining multimodal data and advanced deep learning technology. Most existing technologies, both domestically and internationally, rely on a single data source, while this invention integrates the three-dimensional structure of proteins and the SMILES structures of drugs, providing a more fine-grained multimodal fusion prediction method. This method not only improves prediction accuracy but also broadens the range of predictable proteins and drugs. Particularly for proteins not included in traditional databases, this invention uses the AlphaFold 2 model for high-accuracy prediction completion, filling a technological gap in this field both domestically and internationally.

[0076] (3) The technical solution of this invention successfully solves the long-standing problem of deep learning in drug-target interaction (DTI) prediction:

[0077] 1) Traditional methods suffer from data gaps due to the unknown 3D structure of the target protein and the high cost of obtaining such data, leading to inaccurate model results in the DTI prediction process as a result of the inability to fully utilize the spatial structural information of the protein. To address this issue, this invention employs the AlphaFold 2 model for high-accuracy prediction completion, enabling the acquisition of 3D structural information of the protein even in the absence of experimental data, thus providing a crucial structural foundation for DTI prediction.

[0078] 2) Due to the insufficient application of multimodal technology in the DTI field, existing technologies have shortcomings in fine-grained modality fusion. To address this issue, this invention not only focuses on the features of a single modality but also integrates the molecular map and sequence features of the drug, as well as the sequence and structural information of the protein, through fine-grained multimodal fusion technology. This fusion method can comprehensively capture the multidimensional feature information of drugs and targets, making up for the shortcomings of single-modal methods in information capture. This gives the model stronger cross-dataset transferability, enabling it to adapt to different types of drugs and targets, and improving the flexibility and practicality of the DTI model.

[0079] (4) The technical solution of the present invention effectively overcomes the technical bias that may exist in traditional prediction models.

[0080] 1) This invention not only considers the three-dimensional structural information of proteins but also incorporates the structural information of drug samples, providing a comprehensive and balanced prediction model. This fusion approach reduces reliance on a single data type, thereby mitigating bias caused by data type imbalance.

[0081] 2) By employing self-attention and cross-attention mechanisms, the model of this invention can automatically learn the complex interactions between proteins and drugs, reducing reliance on prior knowledge. This reduces bias caused by algorithm design flaws or over-reliance on specific features.

[0082] 3) This invention not only considers modal fusion at the molecular level, but also prioritizes modal fusion at the atomic level in the processing of raw multidimensional data, so as to obtain fine alignment and promote the understanding of structural molecules.

[0083] 4) The multimodal and adaptive learning approach of this invention enables the model to adapt to data from different sources and types, improving the model's adaptability and accuracy in diverse data environments and effectively overcoming technical biases caused by data imbalance or improper algorithm design. Attached Figure Description

[0084] Figure 1 This is a flowchart of the method for constructing a drug-protein interaction prediction model based on multimodal fusion provided in an embodiment of the present invention.

[0085] Figure 2 This is a block diagram of the drug-protein interaction prediction model construction system based on multimodal fusion provided in this embodiment of the invention.

[0086] Figure 3 This is a detailed flowchart of the method for constructing a drug-protein interaction prediction model based on multimodal fusion provided in this embodiment of the invention.

[0087] Figure 4 This is a schematic diagram of protein multimodal fusion provided in an embodiment of the present invention.

[0088] Figure 5-1 Comparison of model performance on the Human dataset; Figure 5-2 Comparison chart of model performance on the BioSNAP dataset;

[0089] Figure 5-3 A comparison chart of AUC metrics across different datasets and models;

[0090] Figure 5-4 A comparison chart of the F1-Score metric across different datasets and models. Detailed Implementation

[0091] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0092] like Figure 1 As shown in the figure, the method for constructing a drug-protein interaction prediction model based on multimodal fusion provided by this invention includes the following steps:

[0093] S101. For proteins in the dataset, search for them based on their codes. For proteins not included in the database, use the AlphaFold 2 model to perform high-accuracy prediction and completion.

[0094] S102. Data preprocessing: The three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, for the SMILES structure of the drug, the RDKit tool is first used to convert it into a molecular graph and calculate the shortest path distance and edge type. Then, the RDKit is used to generate three-dimensional coordinates based on the SMILES structure of the molecule, and then the Euclidean distance is calculated to generate a distance matrix.

[0095] S103. The shortest path distance, edge type, and 3D Euclidean distance of the molecular two-dimensional mode are fused together, a mixing matrix is ​​constructed by random sampling, and the mixing matrix is ​​injected into the self-attention module to enhance the expression of the relationship between atomic pairs, similar to relative position encoding.

[0096] S104. GCN is used to process the distance map of proteins, and CNN is used to process the protein sequence data.

[0097] S105. The processed protein 1D data and protein 3D data are fused using Cross-Attention, and the fused data and the molecular mixing matrix are spliced ​​together as the final input for drug and target prediction.

[0098] S106. The fused drug molecule features and protein features are spliced ​​together, and then the spliced ​​features are input into a multilayer perceptron for processing. Finally, the binary cross-entropy loss function is used for model training.

[0099] This method for constructing a drug-protein interaction prediction model based on multimodal fusion utilizes the multimodal features of proteins and drugs, improving the model's accuracy in predicting interactions by fusing features from different data types. First, in step 1, protein codes are retrieved from the database. For proteins not included in the database, 3D structure prediction is performed using the AlphaFold 2 model to fill in the missing protein data. This step ensures that all proteins have accurate 3D structures, providing complete data for subsequent model training.

[0100] In step 2, data preprocessing transforms the structural features of proteins and drugs. The 3D structure of the protein is used to generate a distance matrix using Euclidean distance to represent the spatial relationships between atoms. The SMILES structure of the drug molecule is converted into a molecular graph using the RDKit tool, and the shortest path distance and edge type of the molecule are further calculated. Simultaneously, the 3D coordinates of the drug are generated based on the SMILES structure, and Euclidean distance is calculated to generate the corresponding distance matrix. In this way, both proteins and drugs acquire multi-dimensional structural information, providing rich input data for the model.

[0101] In step 3, the two-dimensional and three-dimensional features of the drug molecule are fused. Specifically, the shortest path distance, edge type, and three-dimensional Euclidean distance of the drug are fused into a hybrid matrix, and multiple feature combinations are constructed using random sampling. This hybrid matrix is ​​input into the self-attention module, enhancing the expressive power of intramolecular atom pair relationships. This is equivalent to adding relative position encoding between atoms, which helps the model better capture the feature relationships of the drug. In step 4, the three-dimensional structure data of the protein is processed by a GCN (Graph Convolutional Network) to extract spatial relationship information; the protein sequence data is processed by a CNN (Convolutional Neural Network) to extract the sequence features of the protein from a sequence perspective.

[0102] In steps 5 and 6, protein features from different modalities are fused with drug features at multiple levels. First, the processed 1D protein sequence data and 3D structural data are fused using Cross-Attention, and then concatenated with the mixture matrix of the drug molecules to form the final input. The fused drug molecule features and protein features are further concatenated and input into a multilayer perceptron, and finally trained using a binary cross-entropy loss function. This multimodal feature fusion method effectively improves the model's ability to predict drug-protein interactions, providing a more accurate predictive model for drug target discovery.

[0103] The specific process of S101 provided in this embodiment of the invention is as follows:

[0104] Step 1.1: Obtain protein sequence data from the original dataset;

[0105] Step 1.2: Input the protein sequence data into AlphaFold2. AlphaFold2 will generate a .pdb file of the protein 3D structure after predicting the torsion angle, frame, etc.

[0106] The specific process of S102 provided in this embodiment of the invention is as follows:

[0107] Step 2.1: First, for the protein, calculate the specific path for each pair of residues: Generate an initial distance matrix D of size N×N. Initially, for directly connected residue pairs (i,j), the distance is the Euclidean distance between them. If the atomic pairs are not directly connected, the initial distance is set to infinity. Use Dijkstra's algorithm to calculate the shortest path in D, where D... [i][j] This represents the shortest distance from residue i to residue j. For each pair of residues (i,j), the shortest distance matrix is ​​updated to represent the total Euclidean distance along the shortest path. The shortest distance matrix is ​​calculated for drug molecules. Update and Iteration Where i, j, and k are all different atoms This represents the shortest distance between atom i and atom j;

[0108] Step 2.2: Construct an edge matrix To represent the bond types between atoms in a drug molecule, a feature vector is defined: each edge has a type E. n It has a set of characteristics that represent the type of chemical bond (single bond, double bond, ring), and calculates the edge type matrix. Where N is the number of edges on the shortest path from atom i to j. It is the number of characteristic features (types of chemical bonds) of the nth edge on the path. It is related to each edge feature e n Associated learnable weight parameters, It is the characteristic matrix;

[0109] Step 2.3: Using the Euclidean distance formula Generate a matrix where each element [i,j] represents the distance from atom i to atom j, where the coordinates of each atom can be represented as (x1, y1, z1). This represents the spatial distance between atom pairs (i,j) in the 3D molecular structure; this distance is encoded by the Gaussian kernel function using the following formula.

[0110] Where d ij It is the Euclidean distance between atomic pairs (i,j), A(d; γ,β) = γ.d + β. The parameters γ and β are learnable. Where μ and σ are Gaussian kernel functions, and d is the shortest distance calculated by Dijkstra.

[0111] It is a set of basic kernel vectors. and For learnable weights, R K×K and R K×1 Both are characteristic matrices.

[0112] The specific process of S103 provided in this embodiment of the invention is as follows:

[0113] Step 3.1: Obtain the different relation matrices mentioned above: To integrate them into a new single relation matrix, we first define a multinomial distribution S, where the probability vector p = (p1, p2, p3) represents the probability of the three relations appearing in the fusion matrix;

[0114] Step 3.2: For each position (i,j) in the matrix, randomly sample S from the distribution S to obtain Si. ij ∈{1,2,3}, and then determine the relation matrix value corresponding to that position, using the formula: Where the function ||S ij =K controls the value, if S ij =1, then choose If S ij =2, then choose If S ij =3, then choose The multimodal fusion matrix is ​​constructed through the above steps.

[0115] Step 3.3: Generated Modality Fusion Matrix It is injected into the self-attention module to enhance the expression of relationships between atomic pairs, similar to relative position encoding; the specific formula is as follows: Query, Key, and Value Space: Q = F c .W Q +b Q K = F c .W K +b K V = F c .W V +b V ,in Both are learnable weight matrices, F c Given the input feature matrix, b Q 、b K 、b v All are learnable bias terms. Then, the attention weights are calculated using Query and Key to obtain a matrix M with n*d atomic features, which serves as the input for the final DTI prediction.

[0116] The specific process of S104 provided in this embodiment of the invention is as follows:

[0117] Step 4.1: For the protein distance map, in order to extract the features of the protein distance map, GCN is used for processing. An input map is constructed based on its characteristics. Each GCN layer captures information about neighboring nodes by updating the features of each node. For the l-th layer of graph convolution, the feature update formula for each node is:

[0118] in It is the adjacency matrix plus the identity matrix of the distance graph. yes The degree matrix, H (l) W is the node feature of layer l. (l) This is the weight matrix for learning, and σ is the activation function. Each convolution updates the features of the nodes, gradually integrating more neighbor information, capturing local structural information, and then performing max pooling to obtain an N*d feature matrix S. E , where N is the number of amino acids and d is the final feature dimension;

[0119] Step 4.2: For a one-dimensional protein sequence, denote the input feature matrix of the protein sequence as... A three-layer protein feature encoder is used, each layer consisting of a self-attention layer, a 1D CNN, and a feedforward neural network. The self-attention layer is based on the input feature matrix F. p Through a linear layer, the input matrix F p Projecting onto the query, key, and value space: Q = F P .W Q +b Q K = FP .W K +b K V = F P .W V +b V ,in It is a learnable weight matrix, b Q 、b K 、b v Both are learnable biases. Then, attention weights are calculated using the Query and Key, and then applied to the Value. Here d k This is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally a linear transformation is applied to obtain the final output: X = Concat(head1,...,head).W M W M These are learnable weights, and the final output feature X M The dimension of the input matrix is ​​the same as that of the input matrix, both being l. p ×D p The output features include local information for each location and global contextual dependencies.

[0120] The specific process of S105 provided in this embodiment of the invention is as follows:

[0121] Step 5.1, the protein 1D structure feature matrix is ​​as follows: The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is as follows: Calculate the 1D structural features and Euclidean distance map features respectively:

[0122] Query, Key, and Value of 1D Structural Features: Q 1D =X M ×W Q ,K 1D =X M ×W K V 1D =X M ×W V Query, Key, and Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K V Euc =S E ×W V ,in The weight matrix of the linear layer;

[0123] Step 5.2: Calculate the attention weight matrix using Query and Key to capture the interaction information between the 1D structure and the Euclidean distance map features.

[0124]

[0125] The outputs of each attention point are concatenated and subjected to a linear transformation to obtain the fused 1D structural features and Euclidean distance feature representations:

[0126]

[0127] The fused feature Z 1D and Z Euc Fusing with the original input features to further capture local and global information:

[0128]

[0129] Step 5.3: Perform global pooling on the fused 1D structural features and Euclidean distance map features respectively, and concatenate them into the final joint feature representation:

[0130]

[0131] The joint features are represented as:

[0132] This joint feature vector f contains information about the interaction between the drug and the protein, and serves as the input for the final DTI prediction;

[0133] In S106:

[0134] Step 6.1: Represent the drug's features using the matrix M generated in the previous steps. This matrix has dimensions N×d, where N is the number of atoms in the drug and d is the feature dimension of each atom. The joint feature vector f obtained in the previous steps has dimensions l. p ×D p Then feature concatenation is performed, which can be represented as: F concat =Concat(M,F) forms a new matrix F. concat ;

[0135] Step 6.2: Combine the concatenated feature matrix F concat The data is then fed into a multilayer perceptron (MLP) for further processing. The calculation process is as follows:

[0136] H1=σ(W1·F concat +b1)

[0137] H2=σ(W2·H1+b2) ......

[0139] H L =σ(W L ·H L-1 +b L )

[0140] Among them, W i b is the weight matrix of the i-th layer. i σ is the bias, and σ is the activation function ReLU;

[0141] Step 6.3: Use the binary cross-entropy loss function (BCEloss) to calculate the model output and the true label. Assume the final output of the model is... Given the predicted probability and the true label y, the loss function can be expressed as:

[0142]

[0143] Where N is the number of samples, y i It is the true label of the i-th sample. That is the corresponding predicted probability.

[0144] like Figure 2 As shown in the figure, the drug-protein interaction prediction model construction system based on multimodal fusion provided by this invention includes:

[0145] The prediction completion module is used to find proteins in the dataset based on their codes. For proteins not included in the library, this module performs high-accuracy prediction completion using the AlphaFold 2 model.

[0146] The data preprocessing module is used for data preprocessing: the three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, for the SMILES structure of the drug, the RDKit tool is first used to convert it into a molecular graph and calculate the shortest path distance and edge type. Then, the RDKit is used to generate three-dimensional coordinates based on the SMILES structure of the molecule and calculate the Euclidean distance to generate a distance matrix.

[0147] The fusion module is used to fuse the shortest path distance, edge type, and 3D Euclidean distance of the molecular two-dimensional mode. It constructs a mixing matrix through random sampling and injects the mixing matrix into the self-attention module to enhance the expression of the relationship between atomic pairs, similar to relative position encoding.

[0148] The processing module is used to process the distance map of proteins using GCN, and to process the protein sequence data using CNN.

[0149] The stitching module is used to fuse the processed protein 1D data and protein 3D data using Cross-Attention, and stitches the fused data with the molecular mixing matrix as the final input for drug and target prediction.

[0150] The training module is used to concatenate the fused drug molecule features with protein features, then input the concatenated features into a multilayer perceptron for processing, and finally use the binary cross-entropy loss function for model training.

[0151] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0152] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0153] Another objective of this invention is to provide an information data processing terminal for implementing the drug-protein interaction prediction model construction system based on multimodal fusion.

[0154] Specific implementation of the present invention:

[0155] like Figure 3 As shown, a drug-protein interaction prediction method based on multimodal fusion includes the following steps:

[0156] Step 1: For proteins in the dataset, search for them based on their codes. For proteins not included in the database, use the AlphaFold2 model to perform high-accuracy prediction and completion. The specific process is as follows:

[0157] Step 1.1: Obtain the sequence data of all different proteins from the original dataset;

[0158] Step 1.2: Input the protein sequence data into AlphaFold2. The prediction model AlphaFold2 uses the provided protein sequence structure to predict the torsion angle, frame, etc., and generates a .pdb file of the protein 3D structure.

[0159] Step 2, Data Preprocessing: Read the .pdb file and extract the coordinate information of each atom. Typically, each line contains detailed information about an atom, including atom type, residue type, sequence number, and three-dimensional coordinates (x, y, z). Traverse the .pdb file to extract the coordinates of the desired atoms or residues. If only a specific alpha-carbon atom (referring to the carbon atom directly bonded to a functional group (C) for short, CA) is desired, the coordinates of "CA" atoms can be filtered out. For the extracted coordinate list, calculate the Euclidean distance matrix D between each pair of residues. [i][j] For the SMILES structure of the drug, the RDKit tool is first used to convert the SMILES structure into a molecular graph. Then, the shortest path distance and edge type of the molecular graph are calculated sequentially. Next, the RDKit tool is used to generate three-dimensional coordinates based on the SMILES structure of the molecule, and then Euclidean distance is calculated to generate a distance matrix. The specific process is as follows:

[0160] Step 2.1: First, for the protein, calculate the specific path for each pair of residues: Generate an initial distance matrix D of size N×N. Initially, for directly connected residue pairs (i,j), the distance is the Euclidean distance between them. If the atomic pairs are not directly connected, the initial distance is set to infinity. Use Dijkstra's algorithm to calculate the shortest path in D, where D... [i][j] This represents the shortest distance from residue i to residue j. For each pair of residues (i,j), the shortest distance matrix is ​​updated to represent the total Euclidean distance along the shortest path. The shortest distance matrix is ​​calculated for drug molecules. Update and Iteration Where i, j, and k are all different atoms This represents the shortest distance between atom i and atom j;

[0161] Step 2.2: Construct an edge matrix To represent the bond types between atoms in a drug molecule, a feature vector is defined: each edge has a type E. n It has a set of characteristics that represent the type of chemical bond (single bond, double bond, ring), and calculates the edge type matrix. Where N is the number of edges on the shortest path from atom i to j. It is the number of characteristic features (types of chemical bonds) of the nth edge on the path. It is related to each edge feature e n Associated learnable weight parameters, It is the characteristic matrix;

[0162] Step 2.3: Using the Euclidean distance formula Generate a matrix where each element [i,j] represents the distance from atom i to atom j, where the coordinates of each atom can be represented as (x1, y1, z1). This represents the spatial distance between atom pairs (i,j) in the 3D molecular structure; this distance is encoded using the Gaussian kernel function with the following formula.

[0163] Where d ij It is the Euclidean distance between atomic pairs (i,j), A(d; γ,β) = γ.d + β, where the parameters γ and β are learnable. Where μ and σ are Gaussian kernel functions.

[0164] It is a set of basic kernel vectors. and These are learnable weights;

[0165] Step 3, as follows Figure 4 As shown, the shortest path distance, edge type, and three-dimensional Euclidean distance of the molecule's two-dimensional modality are fused. A mixture matrix is ​​constructed through random sampling and injected into the self-attention module to enhance the expression of relationships between atom pairs, similar to relative position encoding. The specific process is as follows:

[0166] Step 3.1: Obtain the different relation matrices mentioned above: To merge them into a new single relation matrix, we first define a multinomial distribution S, where the probability vector p = (p1, p2, p3) represents the probability of the three relations appearing in the fusion matrix;

[0167] Step 3.2: For each position (i,j) in the matrix, randomly sample S from the distribution S to obtain Si. ij ∈{1,2,3}, and then determine the relation matrix value corresponding to that position, using the formula: Where the function ||S ij =K controls the value, if S ij =1, then choose If S ij =2, then choose If S ij =3, then choose The multimodal fusion matrix is ​​constructed through the above steps.

[0168] Step 3.3: Generated Modality Fusion Matrix This is injected into the self-attention module to enhance the representation of relationships between atomic pairs, similar to relative position encoding. The specific formula is as follows: The resulting matrix M, with atomic features of type n*d, is used as the input for the final DTI prediction.

[0169] Step 4: In this study, targeted processing strategies were adopted for different protein representation methods. Specifically, a Graph Convolutional Network (GCN) was used to process the protein distance graph, while a Convolutional Neural Network (CNN) was used to process the protein sequence data. The specific process is as follows:

[0170] Step 4.1: For the protein distance map, in order to extract its features, GCN is used for processing. An input graph is constructed based on its characteristics, and the adjacency matrix A and degree matrix D are constructed according to these characteristics. The adjacency matrix A represents the connection relationships between node residues, and the degree matrix D is a diagonal matrix representing the number of connections for each node. Each layer of GCN captures information about neighboring nodes by updating the features of each node. For the l-th layer of graph convolution, the feature update formula for each node is:

[0171] in It is obtained by adding the identity matrix to the adjacency matrix of the distance graph. yes The degree matrix, H (l) W is the node feature of layer l. (l) S is the weight matrix learned, and σ is the activation function. Each convolution updates the node's features, gradually integrating more neighbor information and capturing local structural information before performing max pooling to obtain an N*d feature matrix S. E , where N is the number of amino acids and d is the final feature dimension;

[0172] Step 4.2: For a one-dimensional protein sequence, denote the input feature matrix of the protein sequence as... A three-layer protein feature encoder is used, with each layer consisting of a self-attention layer, a 1D CNN, and a feedforward neural network. The self-attention layer is based on the input feature matrix F. p Through a linear layer, the input matrix F p Projecting onto the query, key, and value space: Q = F P .W Q +b Q K = F P .W K +b K V = F P .W V +b V ,in It is a learnable weight matrix. Then, attention weights (Attention(Q,K,V)) are calculated using the Query and Key, and then applied to the Value. Here d kThis is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally a linear transformation is applied to obtain the final output: X = Concat(head1,...,head).W M Q, where W M These are learnable weights, and the final output feature X M The dimension of the input matrix is ​​the same as that of the input matrix, both being l. p ×D p Through computation using the self-attention layer, the output feature matrix not only contains local information at each location but also captures global contextual dependencies. This enables the model to better understand the structural features and functions of protein sequences in subsequent tasks.

[0173] Step 5: The processed protein 1D data and protein 3D data are fused using Cross-Attention, and the fused data is then concatenated with the molecular mixing matrix as the final input for drug and target prediction.

[0174] The specific process is as follows:

[0175] Step 5.1, the protein 1D structure feature matrix is ​​as follows: The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is as follows: Calculate the 1D structural features and Euclidean distance map features respectively:

[0176] Query, Key, and Value of 1D Structural Features: Q 1D =X M ×W Q ,K 1D =X M ×W K V 1D =X M ×W V Query, Key, and Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K V Euc =S E ×W V ,in The weight matrix of the linear layer;

[0177] Step 5.2: Calculate the attention weight matrix using Query and Key to capture the interaction information between the 1D structure and the Euclidean distance map features.

[0178]

[0179] The outputs of each attention point are concatenated and subjected to a linear transformation to obtain the fused 1D structural features and Euclidean distance feature representations:

[0180]

[0181] The fused feature Z 1D and Z Euc Fusing with the original input features to further capture local and global information:

[0182]

[0183] Step 5.3: Perform global pooling on the fused 1D structural features and Euclidean distance map features respectively, and concatenate them into the final joint feature representation:

[0184]

[0185] The joint features are represented as:

[0186] This joint feature vector f contains information about the interaction between the drug and the protein, and serves as the input for the final DTI prediction.

[0187] Step 6: The fused drug molecule features and protein features are spliced ​​together, and then the spliced ​​features are input into a multilayer perceptron (MLP) for processing. Finally, the binary cross-entropy loss function (BCEloss) is used for model training.

[0188] The specific process is as follows:

[0189] Step 6.1: Represent the drug's features using the matrix M generated in the previous steps. This matrix has dimensions N×d, where N is the number of atoms in the drug and d is the feature dimension of each atom. The joint feature vector f obtained from the previous steps has dimensions l. p ×D p Then feature concatenation is performed, which can be represented as: F concat =Concat(M,F) forms a new matrix F. concat ;

[0190] Step 6.2: Combine the concatenated feature matrix F concat The data is then fed into a multilayer perceptron (MLP) for further processing. The calculation process is as follows:

[0191] H1=σ(W1·F concat +b1)

[0192] H2=σ(W2·H1+b2) ......

[0194] H L =σ(W L ·H L-1 +b L )

[0195] Among them, W i b is the weight matrix of the i-th layer. i σ is the bias, and σ is the activation function ReLU.

[0196] Step 6.3: Use the binary cross-entropy loss function (BCEloss) to calculate the model output and the true label. Assume the final output of the model is... Given the predicted probability and the true label y, the loss function can be expressed as:

[0197]

[0198] Where N is the number of samples, y i It is the true label of the i-th sample. This corresponds to the predicted probability. Through the above steps, the fused drug molecule features and protein features are spliced ​​together and input into the MLP. Finally, BCEloss is used for model training, thereby achieving effective prediction of drug-target interactions.

[0199] To verify the effectiveness of the proposed multimodal fusion-based drug-protein interaction prediction model construction method, two datasets, BioSNAP and Human, were selected for validation. Statistical information for the two datasets is shown in Table 1. In the experiment, the dropout probability was set to 0.6, the AdamW optimizer was used for training, the number of MLP layers was set to 3, the epoch was set to 200, and the bitsize was set to 32. AOC, AUC, Precision, F1, and Recall values ​​were used as metrics for the DTI task. Specific experimental results are shown in Tables 2 and 3.

[0200] Table 1 Statistical information for the two datasets

[0201]

[0202] Table 2. Experimental results (%) of the DTI task on the Human dataset.

[0203]

[0204] Table 3. Experimental results (%) of the BioSNAP dataset DTI task

[0205]

[0206] As can be seen from the experimental results in Tables 2 and 3, although the Recall and F1 scores of the proposed method in this embodiment are slightly lower than those of the DrugVQA method on the Human dataset, it achieves very good results on other metrics and the BioSNAP dataset. Moreover, the P (Precision) value of the proposed method in this embodiment exceeds that of most existing methods.

[0207] The technical solution of this invention is mainly applied in the field of drug discovery and development, particularly in the prediction of drug-target interaction (DTI). The following are the specific application areas and related products involved in this invention:

[0208] Drug Screening and Optimization: Utilizing the predictive model of this invention, drug candidate molecules with high affinity for specific disease-related targets can be rapidly screened, accelerating the new drug development process. For example, in the development of anticancer drugs, the predictive model can identify potential inhibitors of key proteins related to tumor growth.

[0209] Personalized medicine: In personalized medicine, this invention can help predict an individual's response to a specific drug, thereby providing patients with customized treatment plans. For example, by analyzing a patient's genomic data and protein expression patterns, the efficacy and side effects of a drug can be predicted, enabling precision medicine.

[0210] Drug Reuse: This invention can be used to discover new uses for existing drugs, explore the potential effects of known drugs on new targets through predictive models, and provide a scientific basis for drug reuse. For example, by analyzing the molecular structure of drugs and the three-dimensional structure of proteins, the potential for new treatments of rare diseases using old drugs can be discovered.

[0211] Drug safety assessment: In the early stages of drug development, the models of this invention can be used to predict potential side effects and toxicities of drugs, reducing risks in later clinical trials. For example, the cardiac safety of drugs can be assessed by predicting interactions between drugs and non-target proteins.

[0212] Biomarker discovery: This invention can be used to identify biomarkers associated with the occurrence and development of diseases, providing new molecular targets for disease diagnosis and treatment. For example, by analyzing protein-drug interaction networks, new biomarkers can be discovered for early diagnosis and treatment monitoring.

[0213] Drug design software: This invention can be integrated into drug design software, providing drug design tools based on multimodal data to help researchers design more effective drug molecules. For example, through the software interface, users can input the SMILES structure of a drug, and the software automatically predicts its interaction with proteins, assisting in drug design.

[0214] The technical solution of this invention demonstrates significant technical effectiveness in the field of drug-target interaction (DTI) prediction, and these effectiveness has been verified through a series of experimental and simulation tests. The following is a detailed description and analysis:

[0215] To verify the effectiveness of the proposed method for constructing a drug-protein interaction prediction model based on multimodal fusion, two widely recognized datasets were selected for validation: BioSNAP and Human. These two datasets contain biomedical information on drug-protein interactions, providing rich experimental scenarios for evaluating model performance.

[0216] The BioSNAP dataset is a publicly available biomedical network dataset containing a wealth of information on drug-protein interactions, which is crucial for understanding drug mechanisms of action and discovering new drug targets.

[0217] The Human dataset focuses on drug-target interactions related to human diseases, providing a data environment that is closer to clinical applications, making the model's predictions more valuable for practical use.

[0218] This invention uses AUC, Precision, F1, and Recall values ​​as key metrics for evaluating the performance of DTI tasks. These metrics comprehensively reflect the predictive accuracy and stability of the model. Two industry-recognized, high-performing models, DrugVQA and DrugBAN, were selected for comparison. The comparison results are as follows. Figure 5-1 As shown in Figure 5-2.

[0219] The model of this invention outperforms the DrugVQA and DrugBAN models in terms of metrics such as AUC, Precision, F1, and Recall. This result is not only theoretically significant but also demonstrates remarkable technical effectiveness in practical applications. By more comprehensively integrating multimodal information about drugs and proteins, the model of this invention provides more accurate interaction predictions, which has important practical value for drug discovery and development.

[0220] To more comprehensively evaluate and demonstrate the performance of the model proposed in this invention, this invention uses four charts, each corresponding to one of the four key evaluation metrics: AUC and F1 Score, to visually represent the performance comparison of the model of this invention with other state-of-the-art models on two different datasets (BioSNAP and Human).

[0221] AUC value is an important indicator for evaluating the ability of a model to distinguish between drug-protein interactions and non-interaction pairs. Figure 1-3As can be clearly seen, the model of this invention achieved the highest AUC values ​​on both datasets, demonstrating its superior performance in predicting drug-protein interactions. Particularly on the BioSNAP dataset, the model of this invention leads with an AUC value of 0.966. Figure 5-3 A comparison of the AUC metric across different datasets and models.

[0222] The F1 score, the harmonic mean of precision and recall, is an important metric for measuring the overall performance of a model. Figure 1-4 As can be seen, the model of this invention leads with an F1 score of 0.962 on the Human dataset and ranks among the top on the BioSNAP dataset with an F1 score of 0.869. This also shows that the model of this invention achieves a good balance between precision and recall, and has excellent overall performance. Figure 5-4 A comparison of the F1-Score metric across different datasets and models.

[0223] In summary, the model of this invention, through fine-grained multimodal fusion technology and structure-aware methods, successfully overcomes the limitations of traditional models in drug-protein interaction (DTI) prediction, achieving high-precision prediction of DTI. This achievement is not only theoretically innovative but also demonstrates significant technical effectiveness in practical applications, providing strong technical support for drug discovery and precision medicine.

[0224] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0225] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a drug-protein interaction prediction model based on multimodal fusion, characterized in that, Includes the following steps: Step 1: Search for proteins in the dataset based on their codes. For proteins not included in the database, use the AlphaFold 2 model to perform high-accuracy prediction and completion. Step 2, Data Preprocessing: The three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, for the SMILES structure of the drug, the RDKit tool is first used to convert it into a molecular graph and calculate the shortest path distance and edge type. Then, the RDKit is used to generate three-dimensional coordinates based on the SMILES structure of the molecule and calculate the Euclidean distance to generate a distance matrix. Step 3: The shortest path distance, edge type, and 3D Euclidean distance of the molecular two-dimensional mode are fused, a mixing matrix is ​​constructed by random sampling, and the mixing matrix is ​​injected into the self-attention module to enhance the expression of the relationship between atomic pairs, similar to relative position encoding; Step 4: Process the distance map of proteins using GCN, and process the protein sequence data using CNN; Step 5: The processed protein 1D data and protein 3D data are fused using Cross-Attention, and the fused data and the molecular mixing matrix are spliced ​​together as the final input for drug and target prediction. Step 6: The fused drug molecule features and protein features are spliced ​​together, and then the spliced ​​features are input into a multilayer perceptron for processing. Finally, the binary cross-entropy loss function is used for model training. The specific process of step 2 is as follows: Step 2.1: Calculate the shortest distance matrix Update and iterate ,use This represents the shortest distance between atom i and atom j; Step 2.2: Calculate the edge type matrix Where N is the number of edges on the shortest path from atom i to j. It is the number of features of the nth edge on the path, i.e., the type of key. It is related to each edge feature Associated learnable weight parameters; Step 2.3: Using the Euclidean distance formula Each element of the generated matrix The distance from atom i to atom j is expressed as... This represents the spatial distance between atom pairs (i,j) in the 3D molecular structure; this distance is encoded using the Gaussian kernel function with the following formula. , , in It is the Euclidean distance between the atomic pair (i,j). parameter It is learnable. in For Gaussian kernel function, It is a set of basic kernel vectors. These are learnable weights; The specific process of step 3 is as follows: Step 3.1: Obtain the different relation matrices mentioned above: To merge them into a new single relation matrix, we first define a multinomial distribution S, where the probability vector p = (p1, p2, p3) represents the probability of the three relations appearing in the fusion matrix; Step 3.2: For each position (i,j) of the matrix, randomly sample from distribution S. Then determine the relation matrix value corresponding to that position, using the formula: Where the function Control the value, if ,like ,like The multimodal fusion matrix is ​​constructed through the above steps. ; Step 3.3: Generated Modality Fusion Matrix It is injected into the self-attention module to enhance the expression of relationships between atomic pairs, similar to relative position encoding; The specific formula is as follows: The atomic characteristic is obtained as n The matrix M of d is used as the input for the final DTI prediction; The specific process of step 5 is as follows: Step 5.1, the protein 1D structure feature matrix is ​​as follows: The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is as follows: Calculate the 1D structural features and Euclidean distance map features respectively: Query, Key, and Value of 1D structural features: Query, Key, and Value of 3D structural features: in This is the weight matrix of the linear layer; Step 5.2: Calculate the attention weight matrix using Query and Key to capture the interaction information between the 1D structure and the Euclidean distance map features. The outputs of each attention point are concatenated and subjected to a linear transformation to obtain the fused 1D structural features and Euclidean distance feature representations: , ； Features of fusion Fusing with the original input features to further capture local and global information: , ; Step 5.3: Perform global pooling on the fused 1D structural features and Euclidean distance map features respectively, and concatenate them into the final joint feature representation: , ； The joint features are represented as: This joint feature vector It contains information about the interaction between the drug and the protein, and serves as input for the final DTI prediction.

2. The method for constructing a drug-protein interaction prediction model based on multimodal fusion as described in claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Obtain protein sequence data from the original dataset; Step 1.2: Input the protein sequence data into AlphaFold2. AlphaFold2 generates a .pdb file of the protein 3D structure after predicting the torsion angle and frame.

3. The method for constructing a drug-protein interaction prediction model based on multimodal fusion as described in claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: For the protein distance map, in order to extract the features of the protein distance map, GCN is used for processing. The input map is constructed according to its characteristics. Each layer of GCN captures the information of neighboring nodes by updating the features of each node. For the l-th layer graph convolution, the feature update formula for each node is: , in It is the adjacency matrix plus the identity matrix of the distance graph. yes The degree matrix, The node features of layer l It is the weight matrix for learning. It is an activation function; each convolution updates the node's features, gradually integrating more neighbor information, capturing local structural information, and then performing max pooling to obtain N. eigenmatrix of d , where N is the number of amino acids and d is the final feature dimension; Step 4.2: For a one-dimensional protein sequence, denote the input feature matrix of the protein sequence as... A three-layer protein feature encoder is used, with each layer including self-attention, a 1D CNN, and a feedforward neural network. The self-attention layer is based on the input feature matrix. Through a linear layer, the input matrix Projected onto the Query, Key, and Value spaces: in It is a learnable weight matrix. Then, attention weights are calculated using the Query and Key, and then applied to the Value. , here This is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally a linear transformation is performed to obtain the final output. Q in which The final output features The dimensions of the input matrix are the same. The output features include local information for each location and global contextual dependencies.

4. The method for constructing a drug-protein interaction prediction model based on multimodal fusion as described in claim 1, characterized in that, In step 6: Step 6.1: Represent the drug's features using the matrix M generated in the previous steps, with dimensions [missing information]. , where N is the number of atoms in the drug and d is the feature dimension of each atom; The joint feature vector obtained through the previous steps Its dimensions are Then, feature concatenation is performed, which can be represented as follows: , forming a new matrix ; Step 6.2: Combine the concatenated feature matrix The input is then fed into a multilayer perceptron (MLP) for further processing. The calculation process is as follows: ...... in, It is the weight matrix of the i-th layer. It's a bias. It is the activation function ReLU; Step 6.3: Use the binary cross-entropy loss function (BCEloss) to calculate the model output and the true label, assuming the final output of the model is the predicted probability. The real label is Then the loss function can be expressed as: Where N is the number of samples. It is the true label of the i-th sample. That is the corresponding predicted probability.

5. A drug-protein interaction prediction model construction system based on multimodal fusion, implementing the drug-protein interaction prediction model construction method based on multimodal fusion as described in any one of claims 1-4, characterized in that, The drug-protein interaction prediction model construction system based on multimodal fusion includes: The prediction completion module is used to find proteins in the dataset based on their codes. For proteins not included in the database, high-accuracy prediction completion is performed using the AlphaFold 2 model. The data preprocessing module is used for data preprocessing: the three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, for the SMILES structure of the drug, the RDKit tool is first used to convert it into a molecular graph and calculate the shortest path distance and edge type. Then, the RDKit is used to generate three-dimensional coordinates based on the SMILES structure of the molecule and calculate the Euclidean distance to generate a distance matrix. The fusion module is used to fuse the shortest path distance, edge type, and 3D Euclidean distance of the molecular two-dimensional mode. It constructs a mixing matrix through random sampling and injects the mixing matrix into the self-attention module to enhance the expression of the relationship between atomic pairs, similar to relative position encoding. The processing module is used to process the distance map of proteins using GCN and the protein sequence data using CNN. The stitching module is used to fuse the processed protein 1D data and protein 3D data using Cross-Attention, and stitches the fused data with the molecular mixing matrix as the final input for drug and target prediction. The training module is used to concatenate the fused drug molecule features with protein features, then input the concatenated features into a multilayer perceptron for processing, and finally use the binary cross-entropy loss function for model training.

6. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion as described in any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion as described in any one of claims 1-4.

8. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the drug-protein interaction prediction model construction system based on multimodal fusion as described in claim 5.

Images