Protein affinity prediction method, device, electronic equipment and computer readable storage medium

Abstract

The present application relates to computer-aided drug design technology, and provides a protein affinity prediction method and device, an electronic equipment and a computer readable storage medium, the method comprises: obtaining the three-dimensional structure of the protein, the three-dimensional structure comprises a plurality of protein chains, each protein chain comprises a plurality of amino acids; determining the feature vector of each amino acid according to the connection relationship between the amino acids, the connection relationship comprises the intrachain connection between the amino acids belonging to the same protein chain and the interchain connection between the amino acids belonging to different protein chains; inputting the feature vectors of all amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, so as to calculate the affinity of the protein according to the binding free energy. The intrachain connection and the interchain connection between the amino acids make the determination of the characteristics of the amino acids more comprehensive, and the accurate prediction of the protein-protein binding affinity is realized.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug design technology, and more specifically, to a protein affinity prediction method, apparatus, electronic device, and computer-readable storage medium. Background Technology

[0002] Protein-protein interactions refer to the interactions between two or more proteins through physical contact. These interactions play a crucial role in many biological processes, such as signal transduction, immune responses, and metabolic control.

[0003] Affinity refers to the strength or stability of protein-protein interactions. It reflects the tightness of the binding between two proteins. With the improvement of data and computing power, artificial intelligence methods have shown great potential in predicting protein-protein binding affinity.

[0004] Improving the accuracy of protein-protein binding affinity prediction is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The present invention aims to provide a method, apparatus, electronic device, and computer-readable storage medium for predicting protein affinity. By utilizing intra-chain and inter-chain linkages between amino acids, the characteristics of amino acids can be determined more comprehensively. This allows for the accurate prediction of protein binding free energy using more comprehensive amino acid characteristics, ultimately achieving accurate prediction of protein-protein binding affinity.

[0006] The embodiments of the present invention can be implemented as follows:

[0007] In a first aspect, the present invention provides a method for predicting protein affinity, the method comprising:

[0008] Obtain the three-dimensional structure of a protein, the three-dimensional structure comprising multiple protein chains, each protein chain comprising multiple amino acids;

[0009] The feature vector of each amino acid is determined based on the linkage relationship between the amino acids, which includes intra-chain linkage between amino acids belonging to the same protein chain and inter-chain linkage between amino acids belonging to different protein chains.

[0010] The feature vectors of all the amino acids are input into a pre-trained prediction model to obtain the binding free energy of the protein, and the affinity of the protein is calculated based on the binding free energy.

[0011] In an optional implementation, the step of determining the feature vector of each amino acid based on the linkage relationships between the amino acids includes:

[0012] A graph structure is generated by using the amino acids as nodes and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between the amino acids.

[0013] The feature vector of each amino acid is updated according to the graph structure.

[0014] In an optional implementation, the step of generating a graph structure by using the amino acids as nodes and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between the amino acids includes:

[0015] For any two amino acids in the three-dimensional structure, the spatial distance between the two amino acids in the three-dimensional structure is obtained.

[0016] Based on the spatial distance, the preset interval segmentation, and the atom type of the two amino acids, a weight vector for the connection relationship between the two amino acids is determined;

[0017] The graph structure is generated by using the amino acids as nodes, the connections between the amino acids as edges, and the weight vectors of the connections as edge feature vectors.

[0018] In an optional implementation, before the step of generating a graph structure by using the amino acids as nodes and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between the amino acids, the following steps are included:

[0019] For any two amino acids in the three-dimensional structure, calculate the spatial distance between the two amino acids in the three-dimensional structure;

[0020] If the two amino acids belong to the same protein chain, the intra-chain connection of the two amino acids is determined according to the spatial distance and the first preset value.

[0021] If the two amino acids belong to different protein chains, the interchain connection of the two amino acids is determined according to the spatial distance and the second preset value, wherein the first preset value is less than the second preset value.

[0022] In an optional implementation, the step of determining the intra-chain linkage of the two amino acids based on the spatial distance and a first preset value, if the two amino acids belong to the same protein chain, includes:

[0023] If the spatial distance is less than or equal to the first preset value, it is determined that there is an intrachain connection between the two amino acids; otherwise, it is determined that there is no intrachain connection between the two amino acids.

[0024] If the two amino acids belong to different protein chains, the step of determining the interchain connection of the two amino acids based on the spatial distance and the second preset value includes:

[0025] If the spatial distance is less than or equal to the second preset value, it is determined that there is an interchain connection between the two amino acids; otherwise, it is determined that there is no interchain connection between the two amino acids.

[0026] In an optional implementation, each node in the graph structure corresponds to an amino acid with an initial vector, and the step of updating the feature vector of each amino acid according to the graph structure includes:

[0027] For each node in the graph structure, the initial vector of the amino acid corresponding to each node is updated according to the nodes associated with each node in the graph structure, so as to obtain the feature vector of the amino acid corresponding to each node.

[0028] In an optional implementation, the step of updating the initial vector of the amino acid corresponding to each node based on the nodes associated with each node in the graph structure to obtain the feature vector of the amino acid corresponding to each node includes:

[0029] For any target node, the target node is taken as the current node, and the initial vector of the amino acid corresponding to the target node is taken as the current feature vector;

[0030] The nodes in the graph structure that are directly connected to the current node are considered as neighboring nodes.

[0031] Select a preset number of nodes from the neighboring nodes as the selected nodes;

[0032] The current feature vector is updated based on the initial vector of each selected node and the connection relationship between the current node and each selected node, to obtain the updated current feature vector.

[0033] The updated feature vector is used as the current feature vector, all selected nodes are used as the current node, and the step of using the nodes directly connected to the current node in the graph structure as neighbor nodes is returned until the preset condition is met.

[0034] The current feature vector obtained last time is used as the feature vector of the amino acid corresponding to the target node.

[0035] In an optional implementation, after the step of inputting the feature vectors of all the amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, the method further includes:

[0036] The predicted affinity of the protein is calculated based on the binding free energy, the preset constant, and the temperature.

[0037] In a second aspect, the present invention provides a protein affinity prediction device, the device comprising:

[0038] An acquisition module is used to acquire the three-dimensional structure of a protein, wherein the three-dimensional structure includes multiple protein chains, and each protein chain includes multiple amino acids;

[0039] A determination module is used to determine the feature vector of each amino acid based on the linkage relationship between the amino acids, wherein the linkage relationship includes intra-chain linkage between amino acids belonging to the same protein chain and inter-chain linkage between amino acids belonging to different protein chains.

[0040] The prediction module is used to input the feature vectors of all the amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, and to calculate the affinity of the protein based on the binding free energy.

[0041] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory being used to store a program, and the processor being used to implement the protein affinity prediction method as described in any of the foregoing embodiments when executing the program.

[0042] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the protein affinity prediction method as described in any of the foregoing embodiments.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This embodiment determines the feature vector of each amino acid based on the linkage relationships between amino acids, and then inputs all the feature vectors of amino acids into a pre-trained prediction model to obtain the binding free energy of the protein. The protein affinity is then calculated based on the binding free energy. Since the linkage relationships include intra-chain linkages between amino acids belonging to the same protein chain and inter-chain linkages between amino acids belonging to different protein chains, the characteristics of the determined amino acids are more comprehensive. This reflects both the characteristics of intra-chain amino acid interactions and the characteristics of inter-chain amino acid interactions, thus enabling accurate prediction of the protein binding free energy using more comprehensive amino acid characteristics, ultimately achieving accurate prediction of protein-protein binding affinity. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the protein affinity prediction method provided in this embodiment.

[0047] Figure 2 This is an example diagram illustrating the feature vector update process provided in this embodiment.

[0048] Figure 3 This is a flowchart illustrating the overall algorithm framework provided in this embodiment.

[0049] Figure 4 Example diagram comparing the training and testing results of different models provided in this embodiment.

[0050] Figure 5 This is a block diagram illustrating the protein affinity prediction device provided in this embodiment.

[0051] Figure 6 This is a block diagram of the electronic device provided in this embodiment.

[0052] Icons: 10-Electronic device; 11-Processor; 12-Memory; 13-Bus; 100-Protein affinity prediction device; 110-Acquisition module; 120-Determination module; 130-Prediction module. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0054] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0055] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0056] In the description of this invention, it should be noted that if terms such as "upper," "lower," "inner," or "outer" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0057] Furthermore, the terms "first" and "second" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0058] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0059] Machine learning, as a typical method in artificial intelligence, typically includes three approaches when using machine learning models to predict protein-protein binding affinity, depending on the type of features used: sequence-driven approaches, structure-driven approaches, and hybrid approaches.

[0060] Sequence-driven protein-protein binding affinity methods use amino acid sequences as input to directly predict binding affinity. The characterization methods employed include SMR (Substitution Matrix Representation), PSSM (Position-Specific Scoring Matrix), and other embedding methods developed for natural language processing.

[0061] Structure-driven protein-protein binding affinity methods use the 3D structure of proteins as model input to predict binding affinity. Since the 3D structure of a protein contains a wealth of information about the protein-protein interface, essential information about the 3D structure of each protein can be captured through careful design of feature vectors.

[0062] Hybrid protein-protein binding affinity methods are designed from multiple perspectives, including protein amino acid sequence and protein three-dimensional structure. For example, sequence, structure, and other functional cues are integrated with Naive Bayes networks, combining all information into a Naive Bayes classifier for PPI (Protein-Protein Interactions) prediction.

[0063] Due to the limited availability and significant noise in existing protein-protein complex datasets, the results of PPI prediction using the methods described above are not ideal.

[0064] In view of this, this embodiment provides a protein affinity prediction method, apparatus, electronic device, and computer-readable storage medium, which can improve the accuracy of PPI prediction, and will be described in detail below.

[0065] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the protein affinity prediction method provided in this embodiment. The method includes the following steps:

[0066] Step S101: Obtain the three-dimensional structure of the protein, which includes multiple protein chains, each of which contains multiple amino acids.

[0067] In this embodiment, the three-dimensional structure of a protein refers to the specific arrangement of protein molecules in space, which determines its biological function. Amino acids are the basic building blocks of proteins; the amino acid sequence determines the protein structure, and the protein structure determines the protein function. Protein chains are biological macromolecules formed by the linear linkage of amino acids through peptide bonds, and are the basis of protein structure and function.

[0068] Step S102: Determine the feature vector of each amino acid based on the linkage relationship between amino acids. The linkage relationship includes intra-chain linkage between amino acids belonging to the same protein chain and inter-chain linkage between amino acids belonging to different protein chains.

[0069] In this embodiment, the feature vector of an amino acid is used to characterize the contextual information of the amino acid in the amino acid sequence, the spatial information in the three-dimensional structure, and the information on the interaction between the amino acid and other amino acids.

[0070] In this embodiment, the linkages between amino acids characterize the interactions between them. These interactions determine the structure, stability, and function of a protein. In this embodiment, the presence of a linkage between two amino acids means that their interaction cannot be ignored when predicting protein affinity; otherwise, it means there is no interaction between the two amino acids or the interaction is negligible. From the perspective of protein chains as biological macromolecules, intra-chain linkages can be considered as connections within the molecule, while inter-chain linkages can be considered as connections between molecules.

[0071] Step S103: Input the feature vectors of all amino acids into the pre-trained prediction model to obtain the binding free energy of the protein, and calculate the affinity of the protein based on the binding free energy.

[0072] In this embodiment, the training process of the prediction model and the process of obtaining the binding free energy of a protein using the prediction model are similar. The input data both include feature vectors of amino acids, and the methods for obtaining these feature vectors are the same. The difference lies in the following: When training the prediction model, the input data includes the feature vectors of the amino acids in the protein sample and a label, where the label is the binding free energy of the protein sample. When using the prediction model to obtain the prediction result for the protein sample, the prediction model needs to be tuned based on the prediction result, the label, and a preset loss function. Once the preset training termination condition is met, the trained prediction model can be obtained. However, when actually using the prediction model, the input data includes the feature vectors of the amino acids of the protein to be predicted, without inputting labels. The obtained prediction result is used as the binding free energy of the protein to be predicted, and no parameter tuning is required.

[0073] It should be noted that since predicting the binding free energy of a protein is a typical regression task, MSE (Mean Squared Error) can be used as the loss function for loss assessment, and then parameter tuning can be performed based on the assessed loss value. Besides MSE, other loss functions such as MAE (Mean Absolute Error) and Huber Loss can also be used.

[0074] It should also be noted that, due to the limited availability of existing protein-protein complex datasets for training and the high noise levels in these datasets, the accuracy of directly predicting protein affinity is greatly affected. This embodiment utilizes the close correlation between protein affinity and protein binding free energy to transform the prediction of protein affinity into the prediction of protein binding free energy. Then, the protein affinity is obtained based on the protein binding free energy, achieving indirect prediction of protein affinity. This enables accurate prediction of affinity even in scenarios with scarce data and high data noise.

[0075] The method provided in this embodiment utilizes intra-chain linkages between amino acids belonging to the same protein chain and inter-chain linkages between amino acids belonging to different protein chains. This allows the characteristics of the amino acids to be determined to reflect both the characteristics of intra-chain amino acid interactions and the characteristics of inter-chain amino acid interactions. By utilizing more comprehensive amino acid characteristics, the binding free energy of proteins can be accurately predicted, ultimately achieving accurate prediction of protein-protein binding affinity.

[0076] In this embodiment, in order to effectively remove noise from the input data, simplify the amount of input data that needs to be processed during model training, and ensure the high quality and reliability of the input data, this embodiment preprocesses the original protein data after obtaining it, and uses the three-dimensional structure of the preprocessed protein as the processing data of the method provided in this embodiment.

[0077] Preprocessing mainly includes: (1) removing water molecules, small molecules, and metal ions from the structure of the original protein data; (2) identifying similar sequences within the structure of the original protein data, removing duplicates from these similar sequences, and retaining the protein chain with the shortest centroid distance among the non-identical protein chains in the similar sequences; (3) reconstructing the data obtained in (2) to make the docking results between amino acids in the protein chains closer to the real physiological state. Reconstruction can be performed using the AutoDockFR tool. AutoDockFR is mainly used to consider the flexibility of the receptor when predicting the interaction between the ligand and the receptor (usually a protein), allowing the flexibility of the receptor to be simulated during the docking process, thereby providing more realistic docking results. During the reconstruction process, AutoDockFR adds polar hydrogen atoms and calculates partial charges, which is crucial for improving the accuracy of molecular representation.

[0078] In optional embodiments, the linkage relationships between amino acids and their spatial positions in the three-dimensional structure, as well as the spatial distance between two amino acids, are inseparable. To determine the linkage relationships between amino acids and accurately distinguish between intra-chain and inter-chain linkages, this embodiment provides an implementation method:

[0079] First, for any two amino acids in the three-dimensional structure, calculate the spatial distance between the two amino acids in the three-dimensional structure;

[0080] Secondly, if the two amino acids belong to the same protein chain, the intra-chain linkage of the two amino acids is determined based on the spatial distance and the first preset value.

[0081] The specific method for determining intra-chain linkages can be as follows: if the spatial distance is less than or equal to a first preset value, then it is determined that there is an intra-chain linkage between the two amino acids; otherwise, it is determined that there is no intra-chain linkage between the two amino acids.

[0082] Third, if two amino acids belong to different protein chains, the interchain connection between the two amino acids is determined based on the spatial distance and the second preset value, where the first preset value is less than the second preset value.

[0083] The specific method for determining inter-chain connections can be as follows: if the spatial distance is less than or equal to a second preset value, then it is determined that there is an inter-chain connection between the two amino acids; otherwise, it is determined that there is no inter-chain connection between the two amino acids.

[0084] In this embodiment, the steps of determining intra-chain connections and determining inter-chain connections can be executed simultaneously or sequentially. This embodiment does not limit the order of execution of the two steps.

[0085] Both the first and second preset values ​​can be set according to the actual scenario. For example, the first preset value can be set to... The second preset value is set to An angstrom is a unit used to represent extremely small lengths, and is widely used, especially in atomic-scale, molecular structure, and optical wavelength measurements.

[0086] After obtaining the connection relationships between amino acids, this embodiment introduces a graph representation to simplify the process of determining the feature vectors of amino acids. The graph represents the connection relationships between amino acids, and then the feature vectors of amino acids are determined based on the graph. One implementation method is as follows:

[0087] First, using amino acids as nodes, the edges and edge feature vectors between corresponding nodes are determined based on the connection relationships between amino acids, thus generating a graph structure.

[0088] In this embodiment, amino acids are treated as nodes in the graph. If two amino acids are connected, then an edge exists between the corresponding two nodes; otherwise, no edge exists. As one implementation, a one-dimensional structure can be extracted from the three-dimensional structure, i.e., a protein sequence comprising multiple amino acids. A pre-defined protein language model is used to generate a 1280-dimensional vector embedding for each amino acid in the protein sequence. This embedding vector combines local and global information of the protein sequence, protein structure, function, and evolutionary features, including secondary structure, three-dimensional spatial relationships, solvent accessibility, catalytic activity, conserved regions, and evolutionary-related features. This can capture the contextual information of amino acids in the sequence and their potential functions and interactions in the three-dimensional structure. The embedding vector of an amino acid is used as the initial vector of the amino acid. The protein language model can be an Evolutionary Scale Modeling-2 (ESM-2) model, an AlphaFold2 model, etc.

[0089] Secondly, the feature vector of each amino acid is updated according to the graph structure.

[0090] In this embodiment, the graph structure includes nodes, edges, and edge feature vectors. This embodiment provides an implementation method for generating a graph structure:

[0091] First, based on the connection relationship between any two amino acids in the three-dimensional structure, the spatial distance between the two amino acids in the three-dimensional structure is obtained.

[0092] Secondly, based on the spatial distance, the preset interval segmentation, and the atom type of the two amino acids, the weight vector of the connection relationship between the two amino acids is determined;

[0093] In this embodiment, the weight vector of the link relationship includes a spatial distance weight vector and an atom type weight vector.

[0094] To further differentiate the effects of different distances on molecular interactions and improve prediction accuracy, inter-chain and intra-chain connections can correspond to different preset intervals. For example, the preset interval for inter-chain connections could be [0, 15], and the preset interval for intra-chain connections could be [0, 3.5]. These preset intervals are then segmented according to a preset number of segments. The number of segments can be set based on the required accuracy; more segments are used for higher accuracy requirements, and fewer segments for lower accuracy requirements. Each interval segment corresponds to a segment identifier. The spatial distance weight vector is determined based on the segment identifier of the interval into which the spatial distance falls. This can be done by directly using the segment identifier, or by prefixing, suffixing, or encryption of the segment identifier, and then using the expanded or converted result as the spatial distance weight vector.

[0095] Taking inter-chain connections as an example, the preset number of segments can be 10. The range [0,15] is divided into 10 equal segments, and each segment corresponds to an interval segment. For example, segments 1 to 10 can be represented by 0 to 9, or by the letters aj. If the spatial distance is 0.5, then the corresponding spatial distance weight vector can be represented as 0.

[0096] To further differentiate the effects of different atom types on molecular interactions and improve prediction accuracy, this embodiment sets up 28 intentional combinations based on seven atom types (A, C, OA, N, SA, HD, NA), including: 'A_A', 'A_C', 'A_OA', 'A_N', 'A_NA', 'A_SA', 'A_HD', 'C_C', 'C_OA', 'C_N', 'C_NA', 'C_SA', 'C_HD', 'OA_OA', 'OA_N', 'OA_NA', 'OA_SA', 'OA_HD', 'N_N', 'N_NA', 'N_SA', 'N_HD', 'NA_NA', 'NA_SA', 'NA_HD', 'SA_SA', 'SA_HD', 'HD_HD'. Where A represents an aromatic carbon atom, C represents an aliphatic carbon atom, OA represents oxygen as a hydrogen bond acceptor, N represents a nitrogen atom, SA represents a sulfur atom, HD represents hydrogen as a hydrogen bond donor, and NA represents nitrogen as a hydrogen bond acceptor. Each combination corresponds to a combination identifier, which can be used as an atom type weight vector.

[0097] By combining or concatenating the spatial distance weight vector and the atom type weight vector, the weight vector of the link relationship is obtained.

[0098] Finally, a graph structure is generated using amino acids as nodes, the connections between amino acids as edges, and the weight vectors of the connections as edge feature vectors.

[0099] In this embodiment, to ensure that the feature vector of each amino acid reflects both its local features within the protein chain and its global features between protein chains, an intra-chain graph structure representing the local features of each amino acid can be generated, while an inter-chain graph structure representing the global features can be generated based on all inter-chain connections. This embodiment provides an implementation method:

[0100] (1) Generate the inter-chain graph structure of the protein by using the amino acids associated with the inter-chain connections as nodes and all inter-chain connections as edges;

[0101] (2) Generate the intrachain graph structure of each protein chain by using amino acids belonging to the same protein chain as nodes and intrachain connections between amino acids belonging to the same protein chain as edges.

[0102] (3) Inter-chain graph structures and intra-chain graph structures of all protein chains are used as graph structures.

[0103] In this embodiment, after obtaining the graph structure and the initial vector of each amino acid corresponding to each node in the graph, the initial vector of each amino acid can be updated according to the graph structure to finally obtain the feature vector of each amino acid. One implementation method is as follows:

[0104] For each node in the graph structure, the initial vector of the amino acid corresponding to each node is updated based on the nodes associated with each node in the graph structure, thus obtaining the feature vector of the amino acid corresponding to each node.

[0105] In this embodiment, the update of each node in the graph structure can be performed sequentially or simultaneously. For ease of description, this embodiment focuses on any target node, and its implementation can be as follows:

[0106] (1) For any target node, take the target node as the current node and take the initial vector of the amino acid corresponding to the target node as the current feature vector;

[0107] (2) Designate the nodes in the graph structure that are directly connected to the current node as neighbor nodes;

[0108] (3) Select a preset number of nodes from the neighboring nodes as the selected nodes;

[0109] In this embodiment, the preset number can be adjusted according to the actual training situation during the prediction model training. For example, the preset number can be set to 3. It should be noted that if the total number of neighboring nodes is less than the preset number, all neighboring nodes will be selected as nodes.

[0110] (4) Update the current feature vector based on the initial vector of each selected node and the edge feature vector between the current node and each selected node to obtain the updated current feature vector.

[0111] In this embodiment, if the graph structure includes an intra-chain graph structure and an inter-chain graph structure, the same method will be used to determine the corresponding feature vector for each amino acid corresponding to each node in the intra-chain graph structure and for each amino acid corresponding to each node in the inter-chain graph structure. Specifically, the initial vector of the selected node, the feature vector of the current node and the edge feature vector between each selected node, and the current feature vector can be concatenated or fused using pooling to obtain the updated current feature vector.

[0112] (5) Reuse the updated feature vector as the current feature vector, reuse all selected nodes as the current nodes, and return to the step of taking the nodes directly connected to the current node in the graph structure as neighbor nodes until the preset conditions are met.

[0113] In this embodiment, the preset conditions can be set according to actual needs. For example, the preset condition can be set to a preset number of updates; if the current feature vector has been updated a preset number of times, then no further updates will be made. As one implementation method, a network layer can be set up for each update to aggregate the vectors of the nodes involved in that update. Aggregation can use an aggregation function, which can be implemented through mean aggregation, max pooling, or long short-term memory networks. Aggregation not only ensures that the current feature vector includes the initial features of neighboring nodes but also incorporates the relationships between nodes, ensuring that the final feature vector has good generalization ability.

[0114] (6) Use the current feature vector obtained last time as the feature vector of the amino acid corresponding to the target node.

[0115] The process of updating the current feature vector described above, by progressively selecting the selected node from the neighboring nodes, not only controls the computational cost but also captures information from multiple levels of neighboring nodes, achieving a balance between computational efficiency and comprehensive feature information.

[0116] To illustrate the process of updating the current feature vector more intuitively, please refer to [link / reference]. Figure 2 , Figure 2 This is an example diagram illustrating the feature vector update process provided in this embodiment. Figure 2 In this algorithm, node 1 is the target node, and the current feature vector is the initial vector of node 1. First, node 1 is set as the current node. Node 1 has 4 neighboring nodes. Nodes 2 and 3 are selected from the 4 neighboring nodes. The current feature vector is updated based on the initial vectors of nodes 2 and 3. Nodes 2 and 3 are then set as the current nodes again. Each of nodes 2 and 3 has two neighboring nodes. The neighbors of node 2 (nodes 4 and 5) and the neighbors of node 3 (nodes 6 and 7) are selected as their respective nodes. The initial vector of node 2 is updated based on the initial vectors of nodes 4 and 5. The initial vector of node 3 is updated based on the initial vectors of nodes 6 and 7. Finally, the current feature vector of node 1 is updated based on the updated initial vectors of nodes 2 and 3.

[0117] It should be noted that the above feature vector update process can be implemented based on the GraphSAGE algorithm. For any graph structure, represented as G = (V, E), the initial vector of node v is represented as x. v , v∈V. The update process using the GraphSAGE algorithm is as follows: First, the input includes G, the initial vector x of the nodes. v Weight matrix W kLet k ∈ {1,…,K}, the nonlinear activation function σ, the aggregator function, and the neighbor function N be the initial node features. The initial vector is set as node v. In each iteration, for each node, the GraphSAGE algorithm first aggregates the features of its neighboring nodes from the previous round, then concatenates the aggregated result with the features of the current node, and processes it through an activation function. After K iterations, the node's feature vector is finally represented as... The node embedding representation is obtained through L2 normalization and formed.

[0118] To provide a comprehensive overview of the protein binding free energy prediction process described in this embodiment, please refer to [link to documentation]. Figure 3 , Figure 3 This is a flowchart illustrating the overall algorithm framework provided in this embodiment. Figure 3 In the diagram, purple and green graphs represent intra-chain graph structures corresponding to different protein chains, while mixed purple and green graphs represent inter-chain graph structures. For any intra-chain or inter-chain graph structure, feature aggregation is performed through multiple updates to determine the feature vector of each amino acid. The feature vectors of all amino acids in the same intra-chain graph structure are then aggregated to obtain a aggregated feature vector. Similarly, the feature vectors of all amino acids in the inter-chain graph structure are aggregated to obtain a aggregated feature vector. Aggregation methods can include average pooling or max pooling. Each aggregated feature vector is then used as an input to the Multi-Layer Perceptron (MLP) in the prediction model for prediction, ultimately yielding the protein's binding free energy.

[0119] In an optional implementation, since there is a preset relationship between the protein binding free energy and the protein affinity, in the step of obtaining the protein binding free energy, the protein binding free energy can also be converted into the protein affinity according to the preset relationship. Specifically, this can be achieved by:

[0120] The predicted value of protein affinity is calculated based on the binding free energy, a preset constant, and temperature.

[0121] In this embodiment, the preset relationship is related to a preset constant and temperature, and can be specifically expressed as: ΔG=RTlnK d Where R is a preset constant, usually the ideal gas constant, with a value of 8.314 J / (mol·K), T is the temperature, with a value of 298 K, and ΔG is the protein binding free energy, K. d The dissociation constant is used to characterize the affinity of a protein; the smaller the value, the stronger the affinity. In addition, the affinity of a protein can also be expressed using the inhibition constant K. i K indicatesi K is a thermodynamic parameter that directly reflects the affinity between the inhibitor and the target protein; the smaller the value, the stronger the affinity. For the binding of the inhibitor to free protein, K... i It can be regarded as a special kind of K d At this point, the two can be considered equivalent. Protein affinity can also be expressed using IC50, which is an indicator of inhibitory activity. However, its value may be indirectly affected by the affinity between the protein and its ligand (such as the inhibitor). IC50 and K... i The relationship is: IC50 = K i ×(1+[S] / K m ), where K m is the Michaelis constant, and [S] is the substrate concentration in the experiment.

[0122] To more clearly reflect the effectiveness of the method provided in this embodiment, this embodiment uses MAE (Mean Absolute Error) and R (Pearson Correlation Coefficient) as evaluation indicators.

[0123] 1. Mean Absolute Error (MAE): The average of the absolute values ​​of the differences between the predicted and actual values.

[0124] 2. Pearson correlation (R): The correlation between predicted and actual values.

[0125] This embodiment tested on two benchmark datasets. Dataset 1 is a subset of PDB. This database contains structural information on protein-protein complexes, which may consist of two or more protein chains, totaling 79 data points. Dataset 2 consists of only two protein chains and includes a homodimeric complex, totaling 90 data points. Finally, a fusion was implemented, merging the two datasets and conducting the test.

[0126] Table 1 compares the results of various methods.

[0127]

[0128] The results in the table show that the ProAffinity-GNN method achieves an MAE of 1.65 and an R-value of 0.72 on Dataset 1, demonstrating excellent performance and reliability in both prediction accuracy and relevance. However, the method in this embodiment further reduces the MAE to 1.42 on Dataset 1 while maintaining an R-value of 0.76, outperforming ProAffinity-GNN and surpassing PRODIGY, showcasing the best prediction accuracy. On Dataset 2, ProAffinity-GNN continues to perform strongly, while the method in this embodiment performs similarly to the best, indicating that the model in this embodiment remains effective when handling complexes composed of two chains. Furthermore, the method in this embodiment further improves performance on the combined set of Datasets 1 and 2, achieving an MAE of 1.54 and an R-value of 0.66, demonstrating better predictive ability on this combined set.

[0129] This embodiment also compares it with methods from other models; please refer to [link / reference]. Figure 4 , Figure 4 This is an example figure comparing the training and testing results of different models provided in this embodiment. Figure 4 Among them, ProAffinity-GNN is currently the best model; GAT_GNN is a method that combines Graph Attention Network (GAT) and Graph Neural Network (GNN), where the graph attention mechanism learns the importance of nodes and inputs them into the graph neural network for learning; HeteroGraphConv is a network model for heterogeneous graphs, where the original authors added covalent bonds between nodes, an interaction network between atoms, and a K-nearest neighbor network between atoms to represent the properties of edges in the graph; RCNN is a sequence-based model that uses bidirectional gated recurrent units (GRUs) to learn the relationships between sequences.

[0130] Figure 4 The training and testing performance of different models on the PDBbind database are shown respectively. Figure 4 The MSE comparison chart in the figure compares the MSE during the training and validation phases. The yellow line (corresponding to the Validation MSE in the figure) represents the MSE during the application phase, and the blue line (corresponding to the Training MSE in the figure) represents the MSE during the training phase. The smaller the MSE value, the higher the prediction accuracy. It can be seen that the method provided in this embodiment exhibits the best performance. Figure 4The R-comparison plot in the figure compares the Pearson correlation coefficient R during the training and validation phases. The pink line (corresponding to Validation R in the figure) represents the R during the application phase, and the green line (corresponding to R in the figure) represents the R during the training phase. The higher the R value, the more relevant the prediction results are. It can be seen that the method provided in this embodiment is similar to the ProAffinity-GNN method and also has good performance.

[0131] To perform the corresponding steps in the above embodiments and various possible implementations, an implementation of a protein affinity prediction device 100 is given below. Please refer to... Figure 5 , Figure 5 This is a block diagram of the protein affinity prediction device provided in this embodiment. It should be noted that the protein affinity prediction device 100 provided by the present invention has the same basic principle and technical effect as the corresponding embodiment described above. For the sake of brevity, it is not mentioned in this embodiment.

[0132] The protein affinity prediction device 100 includes an acquisition module 110, a determination module 120, and a prediction module 130.

[0133] The acquisition module 110 is used to acquire the three-dimensional structure of a protein, which includes multiple protein chains, each protein chain including multiple amino acids.

[0134] The determination module 120 is used to determine the feature vector of each amino acid based on the linkage relationship between amino acids. The linkage relationship includes intra-chain linkage between amino acids belonging to the same protein chain and inter-chain linkage between amino acids belonging to different protein chains.

[0135] The prediction module 130 is used to input the feature vectors of all amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, so as to calculate the affinity of the protein based on the binding free energy.

[0136] In an optional implementation, the determining module 120 is specifically used for:

[0137] A graph structure is generated by using amino acids as nodes and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between amino acids.

[0138] The feature vector of each amino acid is updated based on the graph structure.

[0139] In an optional implementation, the determining module 120 is specifically used to generate a graph structure by determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between amino acids, using amino acids as nodes:

[0140] For any two amino acids in the three-dimensional structure, obtain the spatial distance between the two amino acids in the three-dimensional structure.

[0141] The weight vector of the connection relationship between the two amino acids is determined based on the spatial distance, the preset interval segmentation, and the atom type of the two amino acids.

[0142] A graph structure is generated using amino acids as nodes, the connections between amino acids as edges, and the weight vectors of the connections as edge feature vectors.

[0143] In an optional implementation, the determining module 120 is further configured to:

[0144] For any two amino acids in the three-dimensional structure, calculate the spatial distance between the two amino acids in the three-dimensional structure;

[0145] If two amino acids belong to the same protein chain, the intra-chain linkage of the two amino acids is determined based on the spatial distance and the first preset value.

[0146] If two amino acids belong to different protein chains, the interchain connection between the two amino acids is determined based on the spatial distance and a second preset value, where the first preset value is less than the second preset value.

[0147] In an optional implementation, the determining module 120, when determining the intra-chain linkage of two amino acids based on the distance between them and a first preset value if the two amino acids belong to the same protein chain, is specifically used for:

[0148] If the spatial distance is less than or equal to the first preset value, it is determined that there is an intrachain connection between the two amino acids; otherwise, it is determined that there is no intrachain connection between the two amino acids.

[0149] The determination module 120 is specifically used to determine the inter-chain connection of two amino acids based on spatial distance and a second preset value if the two amino acids belong to different protein chains.

[0150] If the spatial distance is less than or equal to the second preset value, it is determined that there is an interchain connection between the two amino acids; otherwise, it is determined that there is no interchain connection between the two amino acids.

[0151] In an optional implementation, each node in the graph structure corresponds to an amino acid with an initial vector. The determination module 120, when updating the feature vector of each amino acid according to the graph structure, specifically performs the following:

[0152] For each node in the graph structure, the initial vector of the amino acid corresponding to each node is updated based on the nodes associated with each node in the graph structure, thus obtaining the feature vector of the amino acid corresponding to each node.

[0153] In an optional implementation, the determining module 120, when updating the initial vector of the amino acid corresponding to each node based on the nodes associated with each node in the graph structure to obtain the feature vector of the amino acid corresponding to each node, specifically performs the following:

[0154] For any target node, the target node is taken as the current node, and the initial vector of the amino acid corresponding to the target node is taken as the current feature vector;

[0155] Nodes in the graph structure that are directly connected to the current node are considered neighboring nodes.

[0156] Select a preset number of nodes from the neighboring nodes as the selected nodes;

[0157] Based on the initial vector of each selected node and the connection relationship between the current node and each selected node, the current feature vector is updated to obtain the updated current feature vector.

[0158] The updated feature vector is used as the current feature vector again, all selected nodes are used as the current nodes again, and the process of treating nodes directly connected to the current node in the graph structure as neighbor nodes is returned until the preset conditions are met.

[0159] The current feature vector obtained last time is used as the feature vector of the amino acid corresponding to the target node.

[0160] In an optional implementation, the prediction module 130 is further configured to:

[0161] The predicted value of protein affinity is calculated based on the binding free energy, a preset constant, and temperature.

[0162] This invention also provides a block diagram of an electronic device 10, which implements the protein affinity prediction method described in the foregoing embodiments. Please refer to... Figure 6 , Figure 6 This is a block diagram of the electronic device 10 provided in this embodiment. The electronic device 10 includes a processor 11, a memory 12 and a bus 13. The processor 11 and the memory 12 are connected through the bus 13.

[0163] Processor 11 can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the protein affinity prediction method described in the above embodiments can be completed by integrated logic circuits in the hardware of processor 11 or by instructions in software form. Processor 11 can be a general-purpose processor, including CPU (Central Processing Unit), NP (Network Processor), etc.; it can also be DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Logic Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0164] The memory 12 is used to store the program for implementing the protein affinity prediction method. The program may be a software function module stored in the memory 12 in the form of software or firmware or embedded in the OS (Operating System) of the electronic device 10.

[0165] After receiving the execution instruction, the processor 11 executes the program to implement the protein affinity prediction method in this embodiment.

[0166] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the protein affinity prediction method as described in the foregoing embodiments.

[0167] In summary, embodiments of the present invention provide a protein affinity prediction method, apparatus, electronic device, and computer-readable storage medium. The method includes: obtaining the three-dimensional structure of a protein, the three-dimensional structure including multiple protein chains, each protein chain including multiple amino acids; determining the feature vector of each amino acid based on the linkage relationships between amino acids, the linkage relationships including intra-chain linkages between amino acids belonging to the same protein chain and inter-chain linkages between amino acids belonging to different protein chains; inputting the feature vectors of all amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, and calculating the protein affinity based on the binding free energy. Compared with the prior art, this embodiment has at least the following advantages: (1) It determines the characteristics of amino acids more comprehensively through intra-chain and inter-chain linkages between amino acids, and then uses the more comprehensive characteristics of amino acids to achieve accurate prediction of protein binding free energy, reduce false alarm rate, and improve the reliability and economy of prediction; (2) It utilizes the close correlation between protein affinity and protein binding free energy to transform the prediction of protein affinity into the prediction of protein binding free energy, and then obtains the protein affinity based on the protein binding free energy, realizing indirect prediction of protein affinity, and can also achieve accurate prediction of affinity in scenarios with scarce data and high data noise; (3) By progressively selecting selected nodes from neighboring nodes, it controls the computational cost and can capture information from multi-level neighboring nodes, achieving a balance between computational efficiency and comprehensive feature information; (4) By predicting and dissociation constant K d Strongly correlated binding free energy accelerates research progress in protein-protein interactions (PPIs), supporting the rapid screening and validation of potential drug targets during drug discovery, thereby accelerating the development of new drugs.

[0168] The above descriptions are merely various embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting protein affinity, characterized in that, The method includes: Obtain the three-dimensional structure of a protein, the three-dimensional structure comprising multiple protein chains, each protein chain comprising multiple amino acids; The feature vector of each amino acid is determined based on the linkage relationship between the amino acids, which includes intra-chain linkage between amino acids belonging to the same protein chain and inter-chain linkage between amino acids belonging to different protein chains. The feature vectors of all the amino acids are input into a pre-trained prediction model to obtain the binding free energy of the protein, and the affinity of the protein is calculated based on the binding free energy. The step of determining the feature vector of each amino acid based on the connection relationship between the amino acids includes: Using the amino acids as nodes, and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between the amino acids, a graph structure is generated. The graph structure includes the amino acids as nodes, the connection relationships between the amino acids as edges, and the weight vectors of the connection relationships as edge feature vectors. The weight vectors of the connection relationships are obtained by combining or splicing spatial distance weight vectors and atom type weight vectors. The spatial distance weight vector is determined based on the segmentation identifier of the interval into which the spatial distance between two amino acids in the three-dimensional structure falls. The atom type weight vector is a combination identifier of a preset combination of preset atom types of two amino acids. The interval is one of multiple intervals obtained by segmenting a preset interval according to a preset number of segments. Update the feature vector of each amino acid according to the graph structure; and After the step of inputting the feature vectors of all the amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, the method further includes: The predicted affinity of the protein is calculated based on the binding free energy, the preset constant, and the temperature.

2. The method according to claim 1, characterized in that, The step of generating a graph structure by using the amino acids as nodes and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between the amino acids includes: For any two amino acids in the three-dimensional structure, the spatial distance between the two amino acids in the three-dimensional structure is obtained. Based on the spatial distance, the preset interval segmentation, and the atom type of the two amino acids, a weight vector for the connection relationship between the two amino acids is determined; The graph structure is generated by using the amino acids as nodes, the connections between the amino acids as edges, and the weight vectors of the connections as edge feature vectors.

3. The method according to claim 1, characterized in that, Before the step of generating a graph structure by using the amino acids as nodes and determining the edges and edge feature vectors between corresponding nodes based on the connection relationships between the amino acids, the following steps are included: For any two amino acids in the three-dimensional structure, calculate the spatial distance between the two amino acids in the three-dimensional structure; If the two amino acids belong to the same protein chain, the intra-chain connection of the two amino acids is determined according to the spatial distance and the first preset value. If the two amino acids belong to different protein chains, the interchain connection of the two amino acids is determined according to the spatial distance and the second preset value, wherein the first preset value is less than the second preset value.

4. The method according to claim 3, characterized in that, If the two amino acids belong to the same protein chain, the step of determining the intra-chain linkage of the two amino acids based on the spatial distance and a first preset value includes: If the spatial distance is less than or equal to the first preset value, it is determined that there is an intrachain connection between the two amino acids; otherwise, it is determined that there is no intrachain connection between the two amino acids. If the two amino acids belong to different protein chains, the step of determining the interchain connection of the two amino acids based on the spatial distance and the second preset value includes: If the spatial distance is less than or equal to the second preset value, it is determined that there is an interchain connection between the two amino acids; otherwise, it is determined that there is no interchain connection between the two amino acids.

5. The method according to claim 1, characterized in that, Each node in the graph structure corresponds to an amino acid with an initial vector. The step of updating the feature vector of each amino acid according to the graph structure includes: For each node in the graph structure, the initial vector of the amino acid corresponding to each node is updated according to the nodes associated with each node in the graph structure, so as to obtain the feature vector of the amino acid corresponding to each node.

6. The method according to claim 5, characterized in that, The step of updating the initial vector of the amino acid corresponding to each node in the graph structure based on the nodes associated with each node in the graph structure to obtain the feature vector of the amino acid corresponding to each node includes: For any target node, the target node is taken as the current node, and the initial vector of the amino acid corresponding to the target node is taken as the current feature vector; The nodes in the graph structure that are directly connected to the current node are considered as neighboring nodes. Select a preset number of nodes from the neighboring nodes as the selected nodes; The current feature vector is updated based on the initial vector of each selected node and the edge feature vector between the current node and each selected node to obtain the updated current feature vector. The updated feature vector is used as the current feature vector, all selected nodes are used as the current node, and the step of using the nodes directly connected to the current node in the graph structure as neighbor nodes is returned until the preset condition is met. The current feature vector obtained last time is used as the feature vector of the amino acid corresponding to the target node.

7. A protein affinity prediction device, characterized in that, The device includes: An acquisition module is used to acquire the three-dimensional structure of a protein, wherein the three-dimensional structure includes multiple protein chains, and each protein chain includes multiple amino acids; A determination module is used to determine the feature vector of each amino acid based on the linkage relationship between the amino acids, wherein the linkage relationship includes intra-chain linkage between amino acids belonging to the same protein chain and inter-chain linkage between amino acids belonging to different protein chains. The prediction module is used to input the feature vectors of all the amino acids into a pre-trained prediction model to obtain the binding free energy of the protein, and to calculate the affinity of the protein based on the binding free energy. The determining module is specifically used to: use the amino acid as a node, determine the edges and edge feature vectors between corresponding nodes according to the connection relationship between the amino acids, and generate a graph structure. The graph structure includes the amino acid as a node, the connection relationship between the amino acids as an edge, and the weight vector of the connection relationship as an edge feature vector. The weight vector of the connection relationship is obtained by combining or splicing the spatial distance weight vector and the atom type weight vector. The spatial distance weight vector is determined according to the segment identifier of the interval into which the spatial distance between the two amino acids in the three-dimensional structure falls. The atom type weight vector is a combination identifier of a preset combination of preset atom types of the two amino acids. The interval is one of multiple intervals obtained by segmenting a preset interval according to a preset number of segments. The prediction module is also used to calculate the predicted value of the affinity of the protein based on the binding free energy, the preset constant, and the temperature.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory being used to store a program, and the processor being used to implement the protein affinity prediction method as described in any one of claims 1-6 when executing the program.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the protein affinity prediction method as described in any one of claims 1-6.

Images