Multi-modal fusion-based drug-protein interaction prediction model construction method

Abstract

The invention belongs to the technical field of deep learning, and discloses a drug and protein interaction prediction model construction method based on multi-modal fusion. Due to the lack of protein three-dimensional data, the three-dimensional data of the protein is subjected to high-accuracy complementation by using a prediction model AlphaFold 2, and for the modal fusion problem, the drug molecules are fused at the atomic level with finer granularity, so that the multi-modal fusion effect is further enhanced, and the multi-modal fusion effect is improved. Protein 1D data and protein 3D data are processed by using Cross-Attention, the processed protein 1D data and protein 3D data are subsequently fused by using Cross-Attention, multi-modal fusion is performed by using Cross-Attention, so that the interaction relationship between modals can be carefully captured, and in a word, the DTI task effect is improved by designing a multi-modal fusion mode.

Description

Technical Field

[0001] The present 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 Art

[0002] With the development of Internet technology, the technology in the field of new drug discovery is also constantly updated and iterated. Drug-target interaction (DTI) refers to the combination between a drug and its target, which is the key to discovering new drugs and reusing existing drugs. A large number of methods based on deep learning technology to solve the prediction of drug-target interaction (Drug-Target Interaction) have been proposed, and studies have shown that methods based on deep learning can greatly improve the efficiency of new drug discovery, which is of great significance for the discovery of new drugs. Deep learning computational methods usually require the 3D structure of the target protein to evaluate the drug-target interaction (Drug-Target Interaction), but the 3D structure of many proteins is unknown, and the process of obtaining the structure is expensive. Traditional technology cannot solve the problem of missing protein 3D structure well, which also leads to the inability of some current deep learning computational methods to fully learn the structural information of proteins to study drug-target interactions. With the emergence of large models, the contribution of AlphaFold (AlphaFold Protein Structure Database: protein structure data) in protein structure prediction has solved the problem of missing protein 3D structure. Traditional deep learning technology has some obvious shortcomings in drug-target (DTI) tasks, which are mainly reflected in the following aspects: First, traditional deep learning methods usually only use the features of a single modality, such as the currently proposed DLM-DTI (Dual Language Model-drug target interaction) algorithm model, which only focuses on the sequence characteristics of drugs and the sequence information of proteins. In this case, since only one modality is focused on, it is impossible to fully capture the multi-dimensional feature information of drugs or targets. For example, when only the molecular sequence is used, the three-dimensional spatial information of the drug cannot be captured, and when only the protein sequence is used, it is difficult to obtain its structural information, resulting in information loss; second, most of the existing multimodal fusion technologies focus on processing each modality separately and then performing coarse-grained fusion at the molecular level. For example, the currently proposed CAT-DTI (cross-attention and Transformer-drug target interaction) model first processes the molecular graph features and protein sequence features separately and then performs multimodal fusion, but the separate processing does not capture the intrinsic correlation between 2D data and 3D data for the initial data.

[0003] Through the above analysis, the problems and defects of the prior art are as follows:

[0004] Deep learning computational methods usually require the 3D structure of the target protein to evaluate drug-target interactions (DTI), but the 3D structure of many proteins is unknown and the process of obtaining the 3D structure is expensive. Traditional technologies rarely use 3D structural data because they cannot solve the problem of missing 3D structures. In addition, since multimodal technology is rarely used in the field of drug-target interactions (DTI), there are relatively few applications of multimodality, especially in the field of fine-grained modal fusion. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0006] The present invention is implemented in this way. The method for constructing a drug-protein interaction prediction model based on multimodal fusion includes:

[0007] Step 1: Search the proteins in the dataset according to their codes. For proteins not included in the library, use the AlphaFold 2 (AlphaFold Protein Structure Database 2.0 protein structure data 2.0) model to make high-accuracy predictions and completions;

[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 then calculate the Euclidean distance to generate a distance matrix.

[0009] Step 3: The shortest path distance of the two-dimensional molecular mode, the edge type, and the Euclidean distance calculated by the three-dimensional coordinates of the molecule are fused, a mixing matrix is ​​constructed by random sampling, and the mixing matrix is ​​injected into the self-attention module to enhance the relationship expression between atomic pairs, similar to relative position encoding;

[0010] Step 4: GCN (Graph Convolutional Network) is used to process the distance graph of proteins, while CNN is used to process the protein sequence data;

[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 ​​as the final input for drug and target prediction;

[0012] Step 6: Concatenate the fused drug molecule features with the protein features, then input the concatenated features into a multi-layer perceptron for processing, and finally use the binary cross entropy loss function for model training.

[0013] Further, the specific process of step 1 is:

[0014] Step 1.1, obtain protein sequence data through the original data set;

[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 system that defines the position and direction of each amino acid residue in the protein).

[0016] Further, the specific process of step 2 is:

[0017] Step 2.1. First, for proteins, calculate the specific path of 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 the two. If the atom pairs are not directly connected, the initial distance is set to infinity. Use the Dijkstra (Dijkstra's Shortest Path Algorithm: Dijkstra's Shortest Path Algorithm) algorithm to calculate the shortest path of D, where D [i][j] Represents the shortest distance from residue i to residue j. For each pair of residues (i, j), update the distance of the shortest distance matrix so that it represents the total Euclidean distance on the shortest path. Calculate the shortest distance matrix for drug molecules Update iteration Where i, j, and k are all different atoms. represents the shortest distance between atom i and atom j;

[0018] Step 2.2: Build an edge matrix Used to represent the bond type between atoms in drug molecules, define the feature vector: each edge type E n With a set of features, representing the type of chemical bond (single bond, double bond, ring), calculate the edge type matrix where N is the number of edges on the shortest path from atom i to j, is the number of characteristics (type of chemical bonds) of the nth edge on the path, is related to each edge feature e n The associated learnable weight parameters, is the characteristic matrix; T is the transpose of the matrix;

[0019] Step 2.3: Use the Euclidean distance formula Generate a matrix where each element [i,j] is the distance between atom i and atom j, where the coordinates of each atom can be expressed as (x1, y1, z1), using Represents the spatial distance between the atomic pair (i, j) in the 3D structure of the molecule; this distance is encoded by the Gaussian kernel function, using the following formula

[0020] where d ij is the Euclidean distance of the atom pair (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] is a set of basis kernel vectors, and is the learnable weight and W is the weight matrix.

[0022] Further, the specific process of step 3 is:

[0023] Step 3.1: The different relationship matrices obtained above are: 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 fused matrix;

[0024] Step 3.2: For each position (i, j) of the matrix, randomly sample S from distribution S ij ∈{1,2,3}, and then determine the relationship matrix value corresponding to the position. The formula is The function Control value, if S ij =1, then select If S ij =2, then select If S ij =3, then choose After the above steps, a multimodal fusion matrix is ​​constructed

[0025] Step 3.3: Generated modal fusion matrix It is injected into the self-attention module to enhance the relationship between atom pairs, similar to relative position encoding. The specific formula is as follows: 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 the matrix M with atomic features of n*d as the input for the final DTI prediction.

[0026] Further, the specific process of step 4 is:

[0027] Step 4.1: For the protein distance graph, in order to extract the features of the protein distance graph, GCN is used for processing. The input graph 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 lth layer of graph convolution, the feature update formula for each node is:

[0028] in is the adjacency matrix of the distance graph, and after adding the identity matrix, that is The degree matrix, H (l) is the node feature of the lth layer, W (l) is the learned weight matrix, σ is the activation function; each convolution updates the node features, gradually integrates more neighbor information, captures local structural information, and performs the maximum pooling operation to obtain the 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 the one-dimensional sequence of proteins, the input feature matrix of the protein sequence is recorded as A three-layer protein feature encoder is used, each layer includes self-attention, 1DCNN (Convolutional Neural Network) and feedforward neural network. The self-attention layer is based on the input feature matrix F p Through the linear layer, the input matrix F p Projection into query, key and value space: Q = F P .W Q +b Q , K=F P .W K +b K , V=FP .W V +b V ,in is a learnable weight matrix, and then uses Query and Key to calculate the attention weight, and then applies it to Value Here is the k is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally the final output is obtained through linear transformation: X = Concat(head1,...,head).W M QW M is a learnable weight, and the final output feature X M The dimension is consistent with the input matrix, both are l p ×D p , the output features contain local information and global context dependencies at each location.

[0030] Further, the specific process of step 5 is:

[0031] Step 5.1: The protein 1D structural feature matrix is The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is Calculate 1D structural features and Euclidean distance graph features respectively:

[0032] Query, Key, 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, Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K ,V Euc =S E ×W V ,in is the weight matrix of the linear layer;

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

[0034]

[0035] The outputs of each attention head are concatenated and linearly transformed to obtain the fused 1D structural features and Euclidean distance feature representation:

[0036]

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

[0038]

[0039] Step 5.3: Globally pool 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 feature is expressed as:

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

[0044] In step 6:

[0045] Step 6.1: The matrix M generated by the previous step represents the characteristics of the drug, and its dimension is N×d, where N is the number of atoms in the drug and d is the characteristic dimension of each atom; the joint feature vector f obtained by the previous step has a dimension of l p ×D p , and then perform feature splicing. The specific operation can be expressed as: F concat =Concat(M,F), forming a new matrix F concat ;

[0046] Step 6.2: The concatenated feature matrix F concat Input to the multi-layer perceptron (MLPMultilayerPerceptron) 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 is the weight matrix of the i-th layer, b i is the bias, σ is the activation function ReLU (Rectified Linear Unit: linear correction unit);

[0052] Step 6.3: Use the binary cross-entropy loss function (BCEloss: Binary Cross-Entropy Loss) to calculate the model output and the true label. Assume that the final output of the model is (prediction probability), the true label is y, then the loss function can be expressed as:

[0053]

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

[0055] Another object of the present invention is to provide a drug-protein interaction prediction model construction system based on multimodal fusion, comprising:

[0056] The prediction and completion module is used to search for proteins in the dataset according to their encoding. For proteins not included in the library, this module uses the AlphaFold 2 model to perform high-accuracy prediction and completion;

[0057] Data preprocessing module, used for data preprocessing: the three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance, and then the SMILES structure of the drug is first converted into a molecular graph using the RDKit tool and the shortest path distance and edge type are calculated. Then, 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;

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

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

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

[0061] The training module is used to splice the fused drug molecule features with the protein features, and then input the spliced ​​features into the multi-layer 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, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes 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, enables the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0064] Another object of the present invention is to provide an information data processing terminal, which is used to implement the drug-protein interaction prediction model construction system based on multimodal fusion.

[0065] In combination with the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solutions to be protected by the present invention from the following aspects:

[0066] First, due to the lack of protein three-dimensional data, the present invention uses the prediction model AlphaFold 2 to complete the three-dimensional data of proteins with high accuracy, and fuses the drug molecules at a finer-grained atomic level to address the multimodal fusion problem, further enhancing the multimodal fusion effect. Subsequently, the processed protein 1D data and protein 3D data are fused using Cross-Attention. The use of Cross-Attention for multimodal fusion can accurately capture the interaction between modalities. In short, the present invention improves the effect of the DTI task by designing a multimodal fusion method.

[0067] The technical solution of the present invention solves the problems of insufficient data utilization, low prediction accuracy and insufficient model generalization in the existing technology in drug-protein interaction prediction in industrial applications. Traditional methods usually rely only on single-modality drug or protein data (such as molecular sequence or three-dimensional structure), while ignoring the integration of multimodal information, resulting in inaccurate interaction prediction results, which limits the application effect in actual drug development.

[0068] The present invention combines the two-dimensional and three-dimensional structural features of drugs with the sequence and three-dimensional structural data of proteins through a multimodal fusion method, greatly improving the data utilization and expression ability of the model. Specifically, the present invention not only uses AlphaFold 2 to complete the data when the protein lacks structural data, but also combines the shortest path distance, edge type and three-dimensional distance matrix in the drug molecule processing to enhance the expression ability of the relationship between drug molecules. The introduction of the self-attention mechanism enables the model to more accurately capture the atomic relationship and interaction pattern inside the drug molecule, significantly improving the prediction accuracy.

[0069] In addition, the present invention further integrates the 1D and 3D features of proteins through Cross-Attention, and splices them with the drug mixing matrix and inputs them into the model. This deep fusion strategy improves the synergy of different modal features, enabling the model to more comprehensively and accurately characterize the complex interaction relationship between drugs and proteins. Under the final processing of the multi-layer perceptron, the model is optimized with a binary cross entropy loss function, which further improves the accuracy and robustness of interaction prediction.

[0070] Therefore, the present invention has significant technical progress in industrial applications, which can not only effectively improve the prediction accuracy of drug-protein interactions, but also has strong generalization ability. This technical solution has broad application prospects in the fields of new drug development, drug screening and precision medicine, and provides strong support for accelerating drug discovery and personalized treatment.

[0071] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

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

[0073] The technical solution of the present invention provides an efficient and accurate prediction model for drug-protein interactions, which is expected to generate significant economic benefits in the fields of drug development, disease treatment, and personalized medicine. By improving the accuracy and efficiency of drug screening, the cost and time of drug development can be significantly reduced, and the process of new drug listing can be accelerated. In addition, the model can provide researchers with deeper biological insights, promote the discovery of new drug targets, and thus promote innovation and development of the entire pharmaceutical industry. The commercial value is reflected in the reduced failure rate of drug development, improved drug market competitiveness, and the provision of more effective personalized treatment plans for different patients.

[0074] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0075] The present invention solves 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 of the existing technologies at home and abroad rely on a single data source, while the present invention integrates the three-dimensional structure of the protein and the SMILES structure of the drug to provide a more fine-grained multimodal fusion prediction method. This method not only improves the accuracy of the prediction, but also broadens the scope of predictable proteins and drugs, especially for those proteins that are not included in traditional databases. The present invention uses the AlphaFold 2 model to perform high-accuracy prediction completion, filling the technical gap in this field at home and abroad.

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

[0077] 1) The traditional method is due to the lack of data caused by the unknown 3D structure of the target protein and the high cost of obtaining it, which leads to inaccurate model results due to the inability to fully utilize the spatial structure information of the protein for prediction during the DTI prediction process. To address this problem, the present invention uses the AlphaFold 2 model for high-accuracy prediction completion, so that even in the absence of experimental data, the 3D structure information of the protein can be obtained, thereby providing a key structural basis for DTI prediction.

[0078] 2) Due to the insufficient application of multimodal technology in the field of DTI, the existing technology is lacking in fine-grained modal fusion. To address this problem, the present invention not only focuses on the characteristics of a single modality, but also integrates the molecular graph, sequence characteristics of the drug, and the sequence and structural information of the protein through fine-grained multimodal fusion technology. This fusion method can comprehensively capture the multi-dimensional feature information of drugs and targets, making up for the shortcomings of single-modal methods in information capture, making the model have stronger cross-dataset migration capabilities, and can adapt to different types of drugs and targets, thereby 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) The present invention not only considers the three-dimensional structural information of proteins, but also combines the SMILES structural information of drugs to provide a comprehensive and balanced prediction model. This fusion method reduces the reliance on a single data type, thereby reducing the bias caused by imbalanced data types.

[0081] 2) Through the use of self-attention mechanism and Cross-Attention, the model of the present invention can automatically learn the complex interaction relationship between proteins and drugs, reducing the dependence on prior knowledge. This reduces the bias caused by algorithm design defects or over-reliance on specific features.

[0082] 3) The present invention not only considers the modal fusion method at the molecular level, but also gives priority to the modal fusion at the atomic level in the processing of raw multidimensional data to obtain fine alignment and promote structural molecular understanding.

[0083] 4) The multimodal and adaptive learning methods of the present invention enable the model to adapt to data from different sources and types, improve the adaptability and accuracy of the model in a diverse data environment, and effectively overcome the technical bias caused by data imbalance or improper algorithm design. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0085] Figure 2 It is a structural block diagram of a system for constructing a drug-protein interaction prediction model based on multimodal fusion provided in an embodiment of the present invention.

[0086] Figure 3 It is a detailed flow chart of the method for constructing a drug-protein interaction prediction model based on multimodal fusion provided in an embodiment of the present invention.

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

[0088] Figure 5-1 Comparison of model effects under the Human dataset; Figure 5-2 Comparison of model effects under the BioSNAP dataset;

[0089] Figure 5-3 Comparison chart of AUC indicators on different data sets and models;

[0090] Figure 5-4 Comparison chart of F1-Score indicator on different datasets and models. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0092] like Figure 1 As shown, a method for constructing a drug-protein interaction prediction model based on multimodal fusion provided by an embodiment of the present invention comprises the following steps:

[0093] S101. For proteins in the dataset, search based on their encodings. For proteins not included in the library, use the AlphaFold 2 model to make high-accuracy predictions and completions.

[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 according to 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 2D molecular mode are integrated, a mixing matrix is ​​constructed by random sampling, and the mixing matrix is ​​injected into the self-attention module to enhance the relationship expression between atomic pairs, which is similar to relative position encoding;

[0096] S104, the distance graph of the protein is processed using GCN, and the protein sequence data is processed using CNN;

[0097] S105, fusing the processed protein 1D data and protein 3D data using Cross-Attention, and concatenating the fused data with the molecular mixing matrix as the final input for drug and target prediction;

[0098] S106. The fused drug molecule features are concatenated with the protein features, and then the concatenated features are input into a multi-layer perceptron for processing. Finally, a 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, and improves the model's prediction accuracy for interactions by fusing features of different data types. First, in step 1, the protein codes in the dataset are searched from the database. For proteins that are not included, the three-dimensional structure is predicted using the AlphaFold 2 model to complete the missing protein data. This step ensures that all proteins have accurate three-dimensional structures, providing complete data for subsequent model training.

[0100] In step 2, data preprocessing transforms the structural features of proteins and drugs. The three-dimensional structure of the protein generates a distance matrix through Euclidean distance to represent the spatial relationship 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. At the same time, the three-dimensional coordinates of the drug are generated based on SMILES and the Euclidean distance is calculated to generate the corresponding distance matrix. In this way, both proteins and drugs have 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 molecules are fused. Specifically, the shortest path distance, edge type, and three-dimensional Euclidean distance of the drug are fused into a mixed matrix, and multiple feature combinations are constructed using random sampling. The mixed matrix is ​​input into the self-attention module, which enhances the expression ability of the relationship between atomic pairs within the molecule, which is equivalent to adding relative position encoding between atoms, helping the model to better capture the characteristic relationship of the drug. In step 4, the three-dimensional structure data of the protein is processed by GCN (graph convolutional network) to extract spatial relationship information; the protein sequence data is processed by CNN (convolutional neural network) to extract the sequence features of the protein from a sequence perspective.

[0102] In steps 5 and 6, protein features of different modalities are fused with drug features at multiple levels. First, the processed protein 1D sequence data is fused with the 3D structure data through Cross-Attention, and then spliced ​​with the mixing matrix of the drug molecule to form the final input. The fused drug molecule features are further spliced ​​with the protein features, input into the multi-layer perceptron, and finally trained with the binary cross entropy loss function. This multimodal feature fusion method effectively improves the model's ability to predict drug-protein interactions and provides a more accurate prediction model for drug target discovery.

[0103] The specific process of S101 provided in the embodiment of the present invention is:

[0104] Step 1.1, obtain protein sequence data through the original data set;

[0105] 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, etc.

[0106] The specific process of S102 provided in the embodiment of the present invention is:

[0107] Step 2.1. First, for proteins, calculate the specific path of 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 the two. If the atom pairs are not directly connected, the initial distance is set to infinity. Use the Dijkstra algorithm to calculate the shortest path of D, where D [i][j] Represents the shortest distance from residue i to residue j. For each pair of residues (i, j), update the distance of the shortest distance matrix so that it represents the total Euclidean distance on the shortest path. Calculate the shortest distance matrix for drug molecules Updates Among them, i, j, and k are all different atoms. represents the shortest distance between atom i and atom j;

[0108] Step 2.2: Build an edge matrix Used to represent the bond type between atoms in drug molecules, define the feature vector: each edge type E n With a set of features, representing the type of chemical bond (single bond, double bond, ring), calculate the edge type matrix where N is the number of edges on the shortest path from atom i to j, is the number of characteristics (type of chemical bonds) of the nth edge on the path, is related to each edge feature e n The associated learnable weight parameters, is the feature matrix;

[0109] Step 2.3: Use the Euclidean distance formula Generate a matrix where each element [i,j] is the distance between atom i and atom j, where the coordinates of each atom can be expressed as (x1, y1, z1), using Represents the spatial distance between the atomic pair (i, j) in the 3D structure of the molecule; this distance is encoded by the Gaussian kernel function, using the following formula

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

[0111] is a set of basis kernel vectors, and is the learnable weight, R K×K and R K×1 are all feature matrices.

[0112] The specific process of S103 provided in the embodiment of the present invention is:

[0113] Step 3.1: Substitute the different relationship matrices obtained above into: Integrate them into a new single relationship matrix. First, define the multinomial distribution S, where the probability vector p = (p1, p2, p3) represents the probability of the three relationships appearing in the fusion matrix;

[0114] Step 3.2: For each position (i, j) of the matrix, randomly sample S from distribution S ij ∈{1,2,3}, and then determine the relationship matrix value corresponding to the position. The formula is The function Control value, if S ij =1, then select If S ij =2, then select If S ij =3, then choose After the above steps, a multimodal fusion matrix is ​​constructed

[0115] Step 3.3: Generated modal fusion matrix It is injected into the self-attention module to enhance the relationship between atom 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 are all learnable weight matrices, F c is the input feature matrix, b Q 、b K 、b v They are all learnable bias items. Then, the Query and Key are used to calculate the attention weights, and the atomic feature matrix M is obtained as the input for the final DTI prediction.

[0116] The specific process of S104 provided in the embodiment of the present invention is:

[0117] Step 4.1: For the protein distance graph, in order to extract the features of the protein distance graph, GCN is used for processing. The input graph 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 lth layer of graph convolution, the feature update formula for each node is:

[0118] in is the adjacency matrix of the distance graph plus the identity matrix yes The degree matrix, H (l) is the node feature of the lth layer, W (l) is the learned weight matrix, σ is the activation function; each convolution updates the node features, gradually integrates more neighbor information, captures local structural information, and performs the maximum pooling operation to obtain the 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, the input feature matrix of the protein sequence is recorded as A three-layer protein feature encoder is used, each layer includes a self-attention layer, a 1DCNN and a feedforward neural network. The self-attention layer is based on the input feature matrix F p Through the linear layer, the input matrix F p Projection into 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 is the learnable weight matrix, b Q 、b K 、b v They are all learnable bias items. Then use Query and Key to calculate the attention weight and then apply it to Value. Here is the k is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally the final output is obtained through linear transformation: X = Concat(head1,...,head).W M , where W M is a learnable weight, and the final output feature X M The dimensions are consistent with the input matrix and are both l p ×D p , the output features contain local information and global context dependencies at each location.

[0120] The specific process of S105 provided in the embodiment of the present invention is:

[0121] Step 5.1: The protein 1D structural feature matrix is The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is Calculate 1D structural features and Euclidean distance graph features respectively:

[0122] Query, Key, 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, Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K ,V Euc =S E ×W V ,in is the weight matrix of the linear layer;

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

[0124]

[0125] The outputs of each attention head are concatenated and linearly transformed to obtain the fused 1D structural features and Euclidean distance feature representation:

[0126]

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

[0128]

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

[0130]

[0131] The joint feature is expressed as:

[0132] This joint feature vector f contains the interaction information between drugs and proteins and serves as the input for the final DTI prediction;

[0133] In said S106:

[0134] Step 6.1: The matrix M generated by the previous step represents the characteristics of the drug, and its dimension is N×d, where N is the number of atoms in the drug and d is the characteristic dimension of each atom; the joint feature vector f obtained by the previous step has a dimension of l p ×D p , and then perform feature splicing. The specific operation can be expressed as: F concat =Concat(M,F), forming a new matrix F concat ;

[0135] Step 6.2: The concatenated feature matrix F concat Input to the multi-layer 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 is the weight matrix of the i-th layer, b i is the bias, 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 that the final output of the model is (prediction probability), the true label is y, then the loss function can be expressed as:

[0142]

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

[0144] like Figure 2 As shown, a drug-protein interaction prediction model construction system based on multimodal fusion provided by an embodiment of the present invention includes:

[0145] The prediction and completion module is used to search for proteins in the dataset according to their coding. For proteins not included in the library, this module uses the AlphaFold 2 model to perform high-accuracy prediction and completion;

[0146] Data preprocessing module, 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 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 according to the SMILES structure of the molecule and then 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 2D mode, construct a mixing matrix through random sampling, and inject the mixing matrix into the self-attention module to enhance the relationship expression between atomic pairs, similar to relative position encoding;

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

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

[0150] The training module is used to splice the fused drug molecule features with the protein features, and then input the spliced ​​features into the multi-layer 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, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes 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, enables the processor to perform the steps of the method for constructing a drug-protein interaction prediction model based on multimodal fusion.

[0153] Another object of the present invention is to provide an information data processing terminal, which is used to implement the drug-protein interaction prediction model construction system based on multimodal fusion.

[0154] The present invention is specifically implemented:

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

[0156] Step 1: Search for proteins in the dataset according to their encoding. For proteins that are not included in the library, use the AlphaFold2 model to make high-accuracy predictions and completions. The specific process is as follows:

[0157] Step 1.1, obtain the sequence data of all different proteins through the original data set;

[0158] Step 1.2: Input the protein sequence data into AlphaFold2. The prediction model AlphaFold2 generates a .pdb file of the protein 3D structure through the provided protein sequence structure and prediction of torsion angles, frames, etc.

[0159] Step 2, data preprocessing: extract the coordinate information of each atom in the .pdb file by reading it. Usually each line contains detailed information of an atom, including atom type, residue type, serial number and three-dimensional coordinates (x, y, z). Traverse the contents of the .pdb file and extract the coordinates of the required atoms or residues. If you are only interested in specific alpha-carbon atoms (referring to carbon atoms (C) directly connected to functional groups), abbreviated as CA), you can filter out the coordinates of the "CA" atom, and calculate the Euclidean distance matrix D between each pair of residues for the extracted coordinate list. [i][j] For the SMILES structure of the drug, the RDKit tool is first used to convert the SMILES structure of the drug into a molecular graph, and then the shortest path distance and edge type of the molecular graph are calculated in turn. Secondly, the RDKit tool is used to generate three-dimensional coordinates according to the SMILES structure of the molecule, and then the Euclidean distance is calculated to generate a distance matrix. The specific process is as follows:

[0160] Step 2.1. First, for proteins, calculate the specific path of 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 the two. If the atom pairs are not directly connected, the initial distance is set to infinity. Use the Dijkstra algorithm to calculate the shortest path of D, where D [i][j] Represents the shortest distance from residue i to residue j. For each pair of residues (i, j), update the distance of the shortest distance matrix so that it represents the total Euclidean distance on the shortest path. Calculate the shortest distance matrix for drug molecules Update iteration Among them, i, j, and k are all different atoms. represents the shortest distance between atom i and atom j;

[0161] Step 2.2: Build an edge matrix Used to represent the bond type between atoms in drug molecules, define the feature vector: each edge type E n With a set of features, representing the type of chemical bond (single bond, double bond, ring), calculate the edge type matrix where N is the number of edges on the shortest path from atom i to j, is the number of characteristics (type of chemical bonds) of the nth edge on the path, is related to each edge feature e n The associated learnable weight parameters, is the feature matrix;

[0162] Step 2.3: Use the Euclidean distance formula Generate a matrix where each element [i,j] is the distance between atom i and atom j, where the coordinates of each atom can be expressed as (x1, y1, z1), using Indicates the spatial distance between the atomic pair (i, j) in the 3D structure of the molecule; this distance is encoded by the Gaussian kernel function using the following formula

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

[0164] is a set of basis kernel vectors, and is the learnable weight;

[0165] Step 3: Figure 4 As shown in the figure, the shortest path distance of the two-dimensional molecular mode, the edge type and the three-dimensional Euclidean distance of the molecule are fused, a mixing matrix is ​​constructed by random sampling, and the mixing matrix is ​​injected into the self-attention module to enhance the relationship expression between atomic pairs, which is similar to relative position encoding. The specific process is:

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

[0167] Step 3.2: For each position (i, j) of the matrix, randomly sample S from distribution S ij ∈{1,2,3}, and then determine the relationship matrix value corresponding to the position. The formula is The function Control value, if S ij =1, then select If S ij =2, then select If S ij =3, then choose After the above steps, a multimodal fusion matrix is ​​constructed

[0168] Step 3.3: Generated modal fusion matrix It is injected into the self-attention module to enhance the relationship between atom pairs, similar to relative position encoding. The specific formula is as follows: The obtained atomic feature matrix M is n*d as the input for the final DTI prediction.

[0169] Step 4: In the study, targeted processing strategies were adopted for different protein representations. Specifically, the distance graph of the protein was processed using a graph convolutional network (GCN), and the sequence data of the protein was processed using a convolutional neural network (CNN). The specific process is as follows:

[0170] Step 4.1: For the distance graph of the protein, in order to extract the features of the protein distance graph, GCN is used for processing. The input graph is constructed according to its characteristics. According to the characteristics of the protein distance graph, the adjacency matrix A and degree matrix D of the graph are constructed. The adjacency matrix A represents the connection relationship between node residues, and the degree matrix D is a diagonal matrix that represents the number of connections for each node. Each layer of GCN captures the information of neighboring nodes by updating the features of each node. For the lth layer of graph convolution, the feature update formula for each node is:

[0171] in It is the adjacency matrix of the distance graph plus the identity matrix, which is obtained yes The degree matrix, H (l) is the node feature of the lth layer, W (l) is the learned weight matrix, and σ is the activation function. Each convolution updates the node features, gradually integrates more neighbor information, captures local structural information, and then performs a maximum pooling operation 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 the one-dimensional sequence of proteins, the input feature matrix of the protein sequence is recorded as A three-layer protein feature encoder is used. Each layer of the encoder includes a self-attention layer, a 1DCNN, and a feedforward neural network. The self-attention layer is based on the input feature matrix F p Through the linear layer, the input matrix F p Projection into 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 use Query and Key to calculate the attention weight Attention(Q,K,V), and then apply it to Value Here is the kis the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally the final output is obtained through linear transformation: X = Concat(head1,...,head).W M Q, where W M is a learnable weight, and the final output feature X M The dimensions are consistent with the input matrix and are both l p ×D p , through the calculation of the self-attention layer, the output feature matrix not only contains the local information of each position, but also captures the global contextual dependencies. This enables the model to better understand the structural characteristics and functions of protein sequences in subsequent tasks. .

[0173] Step 5: Use Cross-Attention to fuse the processed protein 1D data and protein 3D data, and concatenate the fused data with the molecular mixing matrix as the final input for drug and target prediction.

[0174] The specific process is:

[0175] Step 5.1: The protein 1D structural feature matrix is The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is Calculate 1D structural features and Euclidean distance graph features respectively:

[0176] Query, Key, 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, Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K ,V Euc =S E ×W V ,in is the weight matrix of the linear layer;

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

[0178]

[0179] The outputs of each attention head are concatenated and linearly transformed to obtain the fused 1D structural features and Euclidean distance feature representation:

[0180]

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

[0182]

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

[0184]

[0185] The joint feature is expressed as:

[0186] This joint feature vector f contains the interaction information between drugs and proteins and serves as the input for the final DTI prediction.

[0187] Step 6: Concatenate the fused drug molecule features with the protein features, and then input the concatenated features into a multi-layer perceptron (MLP) for processing. Finally, the binary cross entropy loss function (BCEloss) is used for model training.

[0188] The specific process is:

[0189] Step 6.1: The matrix M generated by the previous step represents the characteristics of the drug, and its dimension is N×d, where N is the number of atoms in the drug and d is the characteristic dimension of each atom. The joint feature vector f obtained by the previous step has a dimension of l p ×D p , and then perform feature splicing. The specific operation can be expressed as: F concat =Concat(M,F), forming a new matrix F concat ;

[0190] Step 6.2: The concatenated feature matrix F concat Input to the multi-layer 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 is the weight matrix of the i-th layer, b i 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 that the final output of the model is (prediction probability), the true label is y, then the loss function can be expressed as:

[0197]

[0198] Where N is the number of samples, y i is the true label of the i-th sample, is the corresponding prediction probability. Through the above steps, the fused drug molecular features and protein features are spliced ​​and input into the MLP, and finally the BCEloss is used for model training to achieve effective prediction of drug-target interaction.

[0199] In order to verify the effectiveness of the method for constructing a drug-protein interaction prediction model based on multimodal fusion proposed in this embodiment, two data sets were selected for verification, namely BioSNAP and Human. The statistical information of the two data sets is shown in Table 1. In the experiment, the dropout probability was set to 0.6, the AdamW optimizer training parameters were used, the number of MLP layers was set to 3, the Epoch was set to 200, and the bitchsize was set to 32. AOC, AUC, Precision, F1 and Recall values ​​were used as measurement indicators for the DTI task. The specific experimental results are shown in Tables 2 and 3.

[0200] Table 1 Statistics of the two datasets

[0201]

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

[0203]

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

[0205]

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

[0207] The technical solution of the present invention is mainly used in the field of drug discovery and development, especially in the drug-target interaction

[0208] The following are the specific application fields and related products involved in the present invention:

[0209] Drug screening and optimization: The prediction model of the present invention can be used to quickly screen drug candidate molecules with high affinity for specific disease-related targets, accelerating the development process of new drugs. For example, in the development of anticancer drugs, the prediction model can be used to identify potential inhibitors of key proteins related to tumor growth.

[0210] Personalized medicine: In personalized medicine, the present invention can help predict an individual's response to a specific drug, thereby providing a customized treatment plan for the patient. For example, by analyzing the patient's genomic data and protein expression patterns, the efficacy and side effects of the drug can be predicted, achieving precision treatment.

[0211] Drug repurposing: 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 repurposing. For example, by analyzing the molecular structure of drugs and the three-dimensional structure of proteins, new therapeutic potential of old drugs for rare diseases can be discovered.

[0212] Drug safety assessment: In the early stages of drug development, the model of the present invention can be used to predict the potential side effects and toxicity of drugs and reduce the risks in later clinical trials. For example, by predicting the interaction between drugs and non-target proteins, the cardiac safety of drugs can be assessed.

[0213] Discovery of biomarkers: The present invention can be used to identify biomarkers related to the occurrence and development of diseases, and provide 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.

[0214] Drug design software: The present invention can be integrated into drug design software to provide 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 to assist in drug design.

[0215] The technical solution of the present invention has shown significant technical effects in the field of drug-target interaction (DTI) prediction, and these effects have been verified through a series of experiments and simulation tests. The following is a detailed description and analysis:

[0216] In order to verify the effectiveness of the method for constructing a drug-protein interaction prediction model based on multimodal fusion proposed in the present invention, two widely recognized data sets were selected for verification: BioSNAP and Human. These two data sets contain biomedical information on drug-protein interactions, respectively, and provide rich experimental scenarios for evaluating model performance.

[0217] The BioSNAP dataset is a public biomedical network dataset that contains a large amount of drug-protein interaction information, which is crucial for understanding the drug action mechanism and discovering new drug targets.

[0218] 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 prediction results more valuable for practical applications.

[0219] This paper uses AUC, Precision, F1 and Recall values ​​as key indicators to measure the performance of DTI tasks. These indicators can fully reflect the prediction accuracy and stability of the model. Two industry-recognized and excellent models are selected for comparison: DrugVQA and DrugBAN. The comparison results are as follows Figure 5-1 , as shown in 5-2.

[0220] The model of the present invention outperforms the DrugVQA and DrugBAN models in terms of AUC, Precision, F1, and Recall values. This result is not only of great theoretical significance, but also shows significant technical effects in practical applications. The model of the present invention provides more accurate interaction predictions by more comprehensively integrating multimodal information of drugs and proteins, which has important practical value for drug discovery and development.

[0221] In order to more comprehensively evaluate and demonstrate the performance of the model proposed in the present invention, four charts are used to intuitively show the performance comparison of the model of the present invention with other advanced models on two different data sets (BioSNAP and Human) corresponding to four key evaluation indicators: AUC and F1 Score.

[0222] The AUC value is an important indicator for evaluating the ability of the model to distinguish drug-protein interaction and non-interaction pairs. Figure 1-3 It can be clearly seen that the model of the present invention achieved the highest AUC value on both datasets, indicating that it has excellent performance in predicting drug-protein interactions. In particular, on the BioSNAP dataset, the model of the present invention leads with an AUC value of 0.966. Figure 5-3 Comparison of AUC indicators on different datasets and models.

[0223] F1 Score is the harmonic mean of precision and recall, and is an important indicator for measuring the overall performance of the model. Figure 1-4 It can be seen that the model of the present invention leads with an F1 Score of 0.962 on the Human dataset, and ranks first with an F1 Score of 0.869 on the BioSNAP dataset. It also shows that the model of the present invention has achieved a good balance between precision and recall, and has superior overall performance. Figure 5-4 Comparison of F1-Score indicators on different datasets and models.

[0224] In summary, the model of the present invention successfully overcomes the limitations of traditional models in DTI prediction through fine-grained multimodal fusion technology and structure-aware methods, and achieves high-precision prediction of drug-protein interactions. This achievement is not only innovative in theory, but also shows significant technical effects in practical applications, providing strong technical support for drug discovery and precision medicine.

[0225] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0226] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a drug-protein interaction prediction model based on multimodal fusion, characterized in that: The following steps are involved: Step 1: Search the proteins in the dataset according to their codes. For proteins not included in the library, use the AlphaFold 2 model to make high-accuracy predictions and completions; Step 2, data preprocessing: The three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance. Then, the SMILES structure of the drug is first converted into a molecular graph using the RDKit tool and the shortest path distance and edge type are calculated. Then, 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. Step 3: The shortest path distance, edge type, and 3D Euclidean distance of the 2D molecular modality are fused, a mixing matrix is ​​constructed by random sampling, and the mixing matrix is ​​injected into the self-attention module to enhance the relationship expression between atomic pairs, similar to relative position encoding. Step 4: Use GCN to process the protein distance graph and use CNN to process the protein sequence data; Step 5: Use Cross-Attention to fuse the processed protein 1D data and protein 3D data, and concatenate the fused data with the molecular mixing matrix as the final input for drug and target prediction; Step 6: Concatenate the fused drug molecule features with the protein features, then input the concatenated features into a multi-layer perceptron for processing, and finally use the binary cross entropy loss function for model training.

2. The method for constructing a drug-protein interaction prediction model based on multimodal fusion according to claim 1, characterized in that: The specific process of step 1 is as follows: Step 1.1, obtain protein sequence data through the original data set; 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, etc.

3. The method for constructing a drug-protein interaction prediction model based on multimodal fusion according to claim 1, characterized in that: The specific process of step 2 is: Step 2.1, calculate the shortest distance matrix D [i][j] , update iteration D [i][j] =min(D [i][j] ,D [i][k] +D [k][j] ),use 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, is the number of features (type of bonds) of the nth edge on the path, is related to each edge feature e n The associated learnable weight parameters; Step 2.3: Use the Euclidean distance formula Generate the distance between atom i and atom j for each element [i,j] of the matrix, using Indicates the spatial distance between the atomic pair (i, j) in the 3D structure of the molecule; this distance is encoded by the Gaussian kernel function using the following formula where d ij is the Euclidean distance of the atom pair (i, j), A(d;γ,β)=γ.d+β. The parameters γ and β are learnable. Where μ and σ are Gaussian kernel functions, is a set of basis kernel vectors, and are learnable weights.

4. The method for constructing a drug-protein interaction prediction model based on multimodal fusion according to claim 1, characterized in that: The specific process of step 3 is as follows: Step 3.1: Substitute the different relationship matrices obtained above into: In a new single relationship matrix of the same river channel, we first define the multinomial distribution S, where the probability vector p = (p1, p2, p3) represents the probability of the three relationships appearing in the fusion matrix; Step 3.2: For each position (i, j) of the matrix, randomly sample S from distribution S ij ∈{1,2,3} and then determine the relationship matrix value corresponding to the position. The formula is The function ||S ij =K control value, if S ij =1, then select If S ij =2, then select If S ij =3, then choose After the above steps, a multimodal fusion matrix is ​​constructed Step 3.3: Generated modal fusion matrix Injected into the self-attention module to enhance the relationship between atom pairs, similar to relative position encoding; The specific formula is as follows: The obtained atomic feature matrix M is n*d as the input for the final DTI prediction.

5. The method for constructing a drug-protein interaction prediction model based on multimodal fusion according to claim 1, characterized in that: The specific process of step 4 is as follows: Step 4.1: For the distance graph of the protein, in order to extract the features of the protein distance graph, GCN is used to process it and construct the input graph 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 of graph convolution, the feature update formula for each node is: in is the adjacency matrix of the distance graph plus the identity matrix yes The degree matrix, H (l) is the node feature of the lth layer, W (l) is the learned weight matrix, σ is the activation function; Each convolution updates the features of the node, gradually integrates more neighbor information, captures local structural information, and then performs a maximum pooling operation to obtain an N*d feature matrix S E , where N is the number of amino acids and d is the final feature dimension; Step 4.2: For a one-dimensional protein sequence, the input feature matrix of the protein sequence is recorded as A three-layer protein feature encoder is used, each layer includes self-attention, 1D CNN and feedforward neural network. The self-attention layer is based on the input feature matrix F p Through the linear layer, the input matrix F p Projection into 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 is a learnable weight matrix, and then uses Query and Key to calculate the attention weight, and then applies it to Value Here is the k is the dimension of the Key. Each attention head is calculated independently, then concatenated, and finally the final output is obtained through linear transformation: X = Concat(head1,...,head).W M QW M is a learnable weight, and the final output feature X M The dimensions are consistent with the input matrix and are both l p ×D p , the output features contain local information and global context dependencies at each location.

6. The method for constructing a drug-protein interaction prediction model based on multimodal fusion according to claim 1, characterized in that: The specific process of step 5 is as follows: Step 5.1: The protein 1D structural feature matrix is The feature matrix of protein 3D data, i.e., protein Euclidean distance features, is Calculate 1D structural features and Euclidean distance graph features respectively: Query, Key, 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, Value of 3D structural features: Q Euc =S E ×W Q ,K Euc =S E ×W K ,V Euc =S E ×W V in is the weight matrix of the linear layer; Step 5.2, use Query and Key to calculate the attention weight matrix to capture the interaction information between the 1D structure and the Euclidean distance map features: The outputs of each attention head are concatenated and linearly transformed to obtain the fused 1D structural features and Euclidean distance feature representation: The fused feature Z 1D and Z Euc Fusion with the original input features to further capture local and global information: Step 5.3: Globally pool the fused 1D structural features and Euclidean distance map features respectively, and concatenate them into the final joint feature representation: The joint feature is expressed as: This joint feature vector f contains the interaction information between drugs and proteins and serves as the input for the final DTI prediction; In step 6: Step 6.1: The matrix M generated by the previous step represents the characteristics of the drug, and its dimension is N×d, where N is the number of atoms in the drug and d is the characteristic dimension of each atom; the joint feature vector f obtained by the previous step has a dimension of l p ×D p , and then perform feature splicing. The specific operation can be expressed as: F concat =Concat(M,F), forming a new matrix F concat ; Step 6.2: The concatenated feature matrix F concat Input to the multi-layer perceptron (MLP) for further processing, the calculation process is as follows: H1=σ(W1·F concat +b1) H2=σ(W2·H1+b2) ...... H L =σ(W L ·H L-1 +b L ) Among them, W i is the weight matrix of the i-th layer, b i is the bias, σ 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. Assume that the final output of the model is (prediction probability), the true label is y, then the loss function can be expressed as: Where N is the number of samples, y i is the true label of the i-th sample, is the corresponding predicted probability.

7. A system for constructing a drug-protein interaction prediction model based on multimodal fusion, which implements the method for constructing a drug-protein interaction prediction model based on multimodal fusion as claimed in any one of claims 1 to 6, characterized in that: The drug-protein interaction prediction model construction system based on multimodal fusion includes: The prediction and completion module is used to search for proteins in the data set according to their codes and to perform high-accuracy prediction and completion for proteins that are not included in the library by using the AlphaFold 2 model; Data preprocessing module, used for data preprocessing: the three-dimensional structure of the protein is converted into a distance matrix by calculating the Euclidean distance, and then the SMILES structure of the drug is first converted into a molecular graph using the RDKit tool and the shortest path distance and edge type are calculated. Then, 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; The fusion module is used to fuse the shortest path distance, edge type and 3D Euclidean distance of the molecular 2D mode, construct a mixing matrix through random sampling, and inject the mixing matrix into the self-attention module to enhance the relationship expression between atomic pairs, similar to relative position encoding; A processing module is used to process the distance graph of proteins using GCN and to process the protein sequence data using CNN; The splicing module is used to fuse the processed protein 1D data and protein 3D data using Cross-Attention, and splice the fused data with the molecular mixing matrix as the final input for drug and target prediction; The training module is used to splice the fused drug molecule features with the protein features, and then input the spliced ​​features into the multi-layer perceptron for processing, and finally use the binary cross entropy loss function for model training.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes 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 to 6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes 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 to 6.

10. 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 7.

Images