Protein-nucleic acid binding site prediction method based on graph neural network representation

Abstract

A method for predicting protein-nucleic acid binding sites based on graph neural network representation, by constructing a protein-nucleic acid interaction dataset, extracting the position and feature information of each residue in the protein and its structural context after sample fusion processing, and constructing a graph representation of the residue structural context based on this, and predicting the graph representation of the protein to be predicted through a hierarchical graph neural network, obtaining the probability of each residue binding to DNA / RNA, and realizing protein-nucleic acid binding site prediction. The present invention learns the key structure and feature pattern of the binding site from the graph representation through a graph representation of residues based on structural context and a hierarchical graph neural network model.

Description

Technical Field

[0001] The present invention relates to a technology in the field of bioengineering, specifically a method for predicting protein and nucleic acid binding sites based on graph neural network representation of protein local structure context. Background Art

[0002] Protein-nucleic acid interactions play an important role in many life activities, such as DNA replication, transcription, translation, gene expression, signal transduction and recognition, etc. Learning protein-nucleic acid interactions is of great significance for analyzing genes, protein functions and drug design. Since the experimental method for analyzing protein-nucleic acid interactions is expensive and time-consuming, it cannot meet the current needs of massive protein analysis, and computational methods are becoming more and more important. Currently, computational methods can be divided into protein sequence-based methods and protein structure-based methods. Sequence-based methods learn the local pattern of binding sites from protein sequences, but since binding sites often have spatial structural conservation, sequence-based methods may not be sufficient to obtain sufficient characteristics of binding sites, thus affecting the prediction accuracy. Structure-based methods attempt to learn the local tertiary structure pattern of binding sites from the structural information of proteins. Since the tertiary structure directly determines the function of proteins, structure-based protein binding site prediction algorithms often have higher accuracy. The main challenge of structure-based methods is how to encode structural information and learn the key structural and physicochemical properties of binding sites from it. Currently, there are some methods that use manually designed protein structure descriptors. Since the manually designed descriptors are predefined, they are unable to extract key information for downstream tasks (Li, S., Kazuo, Y., Mar, AK and Standley, DM (2014) Quantifying sequence and structural features of protein–RNA interactions. Nucleic Acids Research, 42, 10086-10098.). There are also some methods that map protein position information to a three-dimensional grid in space and use a three-dimensional convolutional neural network to learn the pattern of binding sites in the structure. The problem with this method is that the distribution of atoms in proteins in three-dimensional space is sparse, which is not conducive to being mapped into Euclidean space, and it is difficult to ensure the rotation and translation invariance of proteins (Jimenez, J., Doerr, S., Martinezrosell, G., Rose, A. S. and De Fabritiis, G. (2017) DeepSite: protein-binding site predictor using 3D-convolutional neural networks. Bioinformatics, 33, 3036-3042.). How to better represent the structural information and physicochemical properties of proteins and accurately predict nucleic acid binding sites remains a challenge. Summary of the invention

[0003] In order to solve the problem of low recognition accuracy of existing protein-nucleic acid binding sites, the present invention proposes a protein-nucleic acid binding site prediction method based on graph neural network representation, which learns the key structure and characteristic patterns of binding sites from the graph representation through a graph representation of residues based on structural context and a hierarchical graph neural network model.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a method for predicting protein and nucleic acid binding sites based on graph neural network representation. The method constructs a protein and nucleic acid interaction data set, extracts the position and feature information of each residue in the protein and its structural context after sample fusion processing, and constructs a graph representation of the residue structural context based on the information. The graph representation of the protein to be predicted is predicted by a hierarchical graph neural network to obtain the probability of each residue binding to DNA / RNA, thereby realizing the prediction of protein and nucleic acid binding sites.

[0006] The protein dataset, i.e., the protein dataset that interacts with DNA / RNA, is specifically constructed in the following manner: a set of protein-nucleic acid complexes and labels of protein-nucleic acid binding sites are extracted from BioLip, and a set of proteins that bind to DNA and a set of proteins that bind to RNA are extracted according to the types of bases in the complexes.

[0007] Preferably, based on the sequence release time, the sequences released before January 6, 2016 are used as the training set, and the rest are used as the test set.

[0008] The sample fusion processing refers to: enhancing the training set data to address the imbalance problem of positive and negative samples in the data set, fusing the labels of binding sites of protein clusters with high sequence and structural similarity, and increasing the proportion of positive samples in the training set, specifically including:

[0009] Step ① Cluster the proteins in the training set: use bl2seq and TM-align to calculate the sequence similarity and structural similarity of two proteins; cluster the proteins with sequence similarity greater than 0.8 and TM-score greater than 0.5 into one cluster; transfer the annotations of the true binding sites of proteins in the same cluster to the protein with the most residues in the cluster; remove proteins with sequence similarity greater than 30% in the training set;

[0010] Step ② uses CD-HIT to remove proteins in the test set whose sequence similarity with the training set is greater than 30%, and ensures that the sequence similarity of the test set itself is less than 30%.

[0011] The position and characteristic information of the residues include: atomic characteristics, secondary structure information, bond and torsion angle information, solubility, and evolutionary conservation information, which are specifically extracted by the following methods:

[0012] Step 1: For each residue, its position is defined as the coordinates of the center of mass of the residue;

[0013] Step 2: For each residue, extract the features of the atoms belonging to the residue, and then average the features of all atoms of the residue to obtain the atomic feature matrix of the residue, with a size of L×7, where L represents the number of residues in the protein;

[0014] The characteristics of the atom include: atomic mass, B-factor, whether it is a side chain atom, charge, the number of bound hydrogen atoms, and whether it is in a ring.

[0015] Step 3: For proteins with L residues, use DSSP to calculate a secondary structure feature matrix of size L × 14 based on the protein structure;

[0016] The secondary structure characteristic matrix includes: residue water-soluble surface area, bond and torsion angle, and one-hot encoded eight-state secondary structure.

[0017] Step 4: For proteins with L residues, use PSI-BLAST to search the NCBI NR database and calculate the evolutionary conservation score matrix PSSM with a size of L×20;

[0018] The evolutionary conservation scoring matrix PSSM represents the probability of an amino acid mutating into 20 amino acids. Preferably, the element x in the PSSM is normalized using a sigmoid function:

[0019] Step 5. For proteins with L residues, use HHblits to search the uniclust30 database and calculate another evolutionary conservation matrix HMM of size L×30;

[0020] The first 20 columns of the evolutionary conservative matrix HMM represent the probability of amino acid mutation into 20 kinds of amino acids, the first 21-27 columns represent the transfer frequency, and the last three columns represent the local diversity. Preferably, each element x in the matrix is ​​normalized using:

[0021] Step 6: Concatenate the atomic feature matrix, the secondary structure feature matrix, the evolutionary conservation scoring matrix PSSM and the evolutionary conservation matrix HMM to obtain an L×71 feature matrix for a protein with L residues.

[0022] The structural context is extracted by using a structure-based sliding ball for each residue of the protein according to the position information of the residue in the protein. Specifically, according to the position coordinates of the residue in the tertiary structure, a sliding ball along the polypeptide chain is used to extract the structural context of each residue. Specifically, for a target residue, the residue is taken as the center of the ball, r g Draw a sphere for the radius, and all the residues within the sphere and their positional relationships together constitute the structural context of the residue.

[0023] The graph representation is described as defining the residues in the structural context as nodes in the graph, constructing an adjacency matrix based on the relative position information of the nodes, and defining the characteristics of the nodes and edges of the graph based on the relative position relationship and the characteristics of the residues. The specific construction method is as follows:

[0024] Step i) The residues within the structural context are taken as nodes of the graph, and the distance from the node to the center of the sphere is calculated as the 1D position code of the residue. The residue position code is concatenated with the position information of the extracted residue to obtain a node feature of size 72. The node feature set is expressed as The node feature of node i is represented by v i ;

[0025] Step ii) Calculate the distance between two nodes to get the node distance matrix D, using the threshold r v Binarize the distance matrix to get the node adjacency matrix A. Specifically, for node i and node j, When A ij =1, it means there is an edge between node i and node j, namely edge ij;

[0026] Step iii) The edge feature set is represented as E = {e ij |A ij =1}, where the feature e of edge ij ij Includes: 1) the distance between node i and node j, 2) the vector from node i to the center of the sphere The vector from node j to the center of the sphere The angle θ ij The cosine value of

[0027] The hierarchical graph neural network learns the local geometry and local patterns of the features of the binding site from the graph representation of the residue structure context in a supervised manner based on the training set samples, and specifically includes: a graph network coding unit, a K-layer stacked graph neuron block based on a gated recurrent unit (GRU), and a binary classifier based on a multi-layer perceptron, wherein: the graph network coding unit sequentially performs edge coding, node coding, and graph coding; the K-layer stacked graph neuron block based on the gated recurrent unit sequentially performs edge updating, node updating, and graph updating; the binary classifier based on the multi-layer perceptron performs a binary classifier based on the updated graph feature {u k |k=1,…,K} predicts the binding probability of the residues to DNA / RNA.

[0028] The edge coding mentioned above refers to: splicing edge features and the node characteristics of its two end nodes and And through a nonlinear transformation layer Calculate the encoded edge features Node encoding refers to: splicing node features The sum of all encoded edge features starting from this node and passing through a nonlinear transformation layer Calculate the encoded node features Graph encoding means: calculating the sum of all the encoded node features on the graph and passing it through a nonlinear transformation layer Calculate the encoded graph feature u enc ; The implementation of the nonlinear transformation layer is MLP(X)=W2max(0,W1X+b1)+b2, where X is the input, W1, W2, b1, b2 are the network parameters to be learned.

[0029] The edge update is: first concatenate the input edge features The node characteristics of its two end nodes and And the graph feature u k-1 , and calculate the edge features of the intermediate state through a nonlinear transformation layer The features of these two states are fused through the gated recurrent unit to obtain the updated edge features. Node updating means: first concatenate the input node features The sum of all updated edge features starting from this node and the graph feature u k-1 , and calculate the node features of the intermediate state through a nonlinear transformation layer The features of these two states are fused through the gate recurrent unit to obtain the updated node features. Graph updating means: first concatenate the sum of the updated features of all nodes on the graph and the graph feature u k-1, and calculate the intermediate state graph feature u through a nonlinear transformation layer k′ , the graph features of these two states are fused through the gate recurrent unit to obtain the updated graph feature u k .

[0030] The classifier is defined as, in,

[0031] The hierarchical graph neural network uses 80% of the DNA / RNA binding protein dataset as a training set to supervise the training of the neural network, and the remaining 20% ​​is used as a validation set to adjust the hyperparameters of the model and define the binarization threshold T.

[0032] The hyperparameters include: g 、r v , K, the dimensions of W and b in nonlinear transformation.

[0033] The adjustment is specifically to select the hyperparameter combination that makes the Matthews coefficient of the validation set the highest as the hyperparameter of the model by grid search of the hyperparameters, and to select the binarization threshold that makes the Matthews coefficient of the validation set the highest as the binarization threshold T of the model.

[0034] The protein and nucleic acid binding site prediction is performed by comparing the probability of each residue binding to DNA / RNA predicted by the hierarchical graph neural network with a threshold T to obtain binding residues and non-binding residues.

[0035] The present invention relates to a system for implementing the above method, comprising: a feature extraction module, a structural context construction module, a graph representation module, and a hierarchical graph neural network module, wherein the feature extraction module extracts the position and feature information of the protein from the sequence and structure of the protein, the structural context construction module constructs the local context of the residue through the position information of the protein, the protein information extracted by the feature extraction module and the structural context of the residue constructed by the structural context module are transmitted to the graph representation module to construct a graph representation of the residue structural context, and the hierarchical graph neural network module uses the graph representation of the residue structural context as the input of the module to predict the probability that the residue belongs to a protein-nucleic acid binding residue.

[0036] Technical Effects

[0037] The present invention solves the problem of low prediction accuracy of protein and DNA / RNA binding sites in the prior art as a whole; compared with the prior art, the present invention represents the local neighborhood information of residues in three-dimensional space as a graph in non-Euclidean space, and the characteristics of nodes and edges in the graph encode the local geometric information and multiple biochemical characteristics of proteins. This representation does not require manual design of protein structure descriptors. Compared with the structural representation in Euclidean space, the graph representation is more suitable for sparse and irregular protein residue distribution, and has rotation and translation invariance. Through the designed hierarchical graph neural network, important structural and biochemical features in binding sites can be automatically learned end-to-end, thereby improving the prediction accuracy of protein and DNA / RNA binding sites. Compared with the existing inventions, the present invention significantly improves the prediction accuracy of protein and nucleic acid binding sites, and is easy to migrate to the prediction of binding sites between proteins and other ligands, and has good generalization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the present invention;

[0039] Figure 2 To construct a graph representation of residues (A) Feature extraction module (B) Structural context building module (C) Graph representation module;

[0040] Figure 3 It is a flowchart of the hierarchical graph neural network;

[0041] Figure 4 Flowchart of graph neuron blocks in a hierarchical neural network. DETAILED DESCRIPTION

[0042] This example uses protein and DNA\RNA binding site datasets as the training set and test set of this example. For the DNA binding dataset, the training set contains 573 proteins. After data enhancement, 14479 binding residues and 145404 non-binding residues are obtained. The test set contains 129 proteins, including 2249 binding residues and 35275 non-binding residues; for the RNA binding dataset, the training set contains 495 proteins. After data enhancement, 14609 binding residues and 122290 non-binding residues are obtained. The test set contains 117 proteins, including 2031 binding residues and 35314 non-binding residues. The process of data enhancement is to calculate the sequence and structural similarity of proteins and divide proteins with sequence similarity greater than 0.8 and TM-score greater than 0.5 into a cluster, and migrate the binding site labels of proteins in each cluster to the protein with the most residues in the cluster.

[0043] like Figure 1As shown, this embodiment relates to a method for predicting protein and nucleic acid binding sites based on graph neural network representation, comprising the following steps:

[0044] ① Based on the sequence and structure of proteins in the above database, the three-dimensional position information of each residue in the protein and the characteristics of the atoms it includes are extracted. The secondary structure information, bond and torsion angle information, and solubility of the protein are extracted using the DSSP algorithm. The evolutionary conservation information of the protein is calculated using the PSI-BLAST and HHblits algorithms. The features except the position information are normalized to between [0,1]. For a protein with L residues, an L×3 position information matrix and an L×71 feature matrix are obtained.

[0045] ② Design a radius of r g The three-dimensional sphere with the target residue as the center of the sphere, the residues in the sphere and their positional relationships together constitute the structural context of the residue.

[0046] ③ Construct a graph representation of the local neighborhood of the residues, take the residues within the structural context as the nodes of the graph, calculate the distance from the node to the center of the sphere as the 1D position code of the residue, concatenate the residue position code with the protein feature matrix to obtain a node feature of size 72, and the node feature of node i is represented as v i , calculate the distance between two nodes to get the node distance matrix D, use the threshold r v Binarize the distance matrix to get the node adjacency matrix A and calculate the edge feature e ij , including two position-related features: 1) the distance between node i and node j, 2) the vector from node i to the center of the sphere The vector from node j to the center of the sphere The cosine of the angle formed.

[0047] ④ 80% of the proteins in the training set were used as the training set and the remaining 20% ​​as the validation set, and the similarity between the training set and the validation set was ensured to be less than 30% under the CD-HIT algorithm. The parameters of the model were learned through supervised training based on the training set samples. The data in the training set were input into the constructed hierarchical graph neural network in mini-batches. The difference between the forward propagation results and the true labels on the training set was obtained based on the cross entropy loss function. The Adam optimizer was used to iteratively optimize until the Matthews coefficient on the validation set did not increase in 10 consecutive iterations. The grid search method was used to optimize the network hyperparameters. For each set of hyperparameters, the Matthews coefficient of the validation set obtained by the trained network was calculated, and the hyperparameter combination that maximized the value was selected as the final hyperparameter of the model. The specific parameters were mini-batch = 64, r g =20, r v=10, K=4, and the dimensions of the graph coding unit and the graph neuron block are 128. So far, the models included in the present invention have been trained.

[0048] ⑤ After the model training is completed, the same features are extracted from the proteins in the test set and the same graph representation is constructed. The trained graph neural network is input to predict the probability P of the residue as a binding site. The binary threshold T optimized on the validation set is used to determine whether the residue is a binding site. If P>T, it is judged as a binding site, otherwise it is not a binding site.

[0049] The evaluation indicators used in this embodiment include: recall rate Accuracy F1 value Matthews coefficient The area under the ROC curve and the coordinate axis is AUC, where TP, FP, TN, and FN are true positive, false positive, true negative, and false negative results, respectively.

[0050] In the experimental stage, this example is compared with other representative protein and nucleic acid binding site identification methods:

[0051] 1) Hu, J., Li, Y., Zhang, M., Yang,

[0052] 2) Yu, D.-J., Hu, J., Yang, J., Shen, H.-B., Tang, J. and Yang, J.-Y. (2013) Designing template-free predictor for targeting protein-ligand binding sites with classifier ensemble and spatial clustering. IEEE / ACM transactions on computational biology and bioinformatics, 10, 994-1008.

[0053] 3)Li,S.,Kazuo,Y.,Mar,A.K.and Standley,D.M.(2014)Quantifying sequenceand structural features ofprotein–RNAinteractions.NucleicAcidsResearch,42,10086-10098.

[0054] 4)Lam,J.H.,Li,Y.,Zhu,L.,Umarov,R.,Jiang,H.,Héliou,A.,Sheong,F.K.,Liu,T.,Long,Y.and Li,Y.(2019)A deep learning framework to predict bindingpreference of RNA constituents on protein surface.Nature communications,10,1-13.

[0055] 5)Su,H.,Liu,M.,Sun,S.,Peng,Z.and Yang,J.(2019)Improving theprediction of protein–nucleic acids binding residues via multiple sequenceprofiles and the consensus of complementarymethods.Bioinformatics,35,930-936.

[0056] 6)Zhu,Y.-H.,Hu,J.,Song,X.-N.and Yu,D.-J.(2019)DNAPred:accurateidentification of DNA-binding sites from protein sequence by ensembledhyperplane-distance-based support vector machines.Journal ofchemicalinformation and modeling,59,3057-3071.

[0057] 7)Walia, RR, Xue, LC, Wilkins, K., El-Manzalawy, Y., Dobbs, D. and Honavar, V. (2014) RNABindRPlus: a predictor that combines machine learning and sequencehomology-based methods to improve the reliability of predicted RNA-bindingresidues in proteins. PLoS One, 9, e97725.

[0058] 8) Liu, R. and Hu, J. (2013) DNABind: A hybrid algorithm for structure-based prediction of DNA-binding residues by combining machine learning-and template-based approaches. Proteins: Structure, Function, and Bioinformatics, 81, 1885-1899.).

[0059] The results are shown in the following table. This method is significantly better than other methods in all indicators, which shows that this method can identify more real binding samples, and the false positive rate is also lower than other methods. In summary, the present invention has a significant advantage in identifying protein and nucleic acid binding sites.

[0060]

[0061] Compared with the existing technology, this method avoids manually designed structural descriptors by encoding the neighborhood information of the residue structure into a graph representation in non-Euclidean space, which is more suitable for the representation of sparse protein structure. Through end-to-end neural network training, local patterns of structures and features closely related to the task of predicting binding sites can be learned from the graph representation. In the prediction of protein-DNA binding sites, the F1, MCC, and AUC were improved by 8.2%, 8.8%, and 6.9%, respectively. In the prediction of protein-RNA binding sites, the F1, MCC, and AUC were improved by 9.7%, 10.6%, and 6.6%, respectively. Specifically, the prediction accuracy of this method is higher than other sequence-based or structure-based methods in the field, and it has higher accuracy than the DNABind, aaRNA method or NucleicNet method using convolutional neural networks that use manually designed descriptors to characterize protein structures. On the one hand, this shows that this method retains more complete protein topological information and is more suitable for characterizing the local structure of proteins. On the other hand, it shows that end-to-end hierarchical graph neural networks can learn local patterns of structural and biochemical characteristics that are more relevant to downstream binding site prediction tasks from the graph representation of local protein structures.

[0062] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for predicting protein and nucleic acid binding sites based on graph neural network characterization, characterized in that: By constructing a protein data set, the position information and structural context of each residue in the protein are extracted after sample fusion processing, and a graph representation of the residue structural context is constructed based on this. The graph representation of the protein to be predicted is predicted through a hierarchical graph neural network to obtain the probability of each residue binding to DNA / RNA, thereby realizing the prediction of protein-nucleic acid binding sites; The protein dataset, i.e., the protein dataset interacting with DNA / RNA, is constructed in the following manner: extracting a set of protein-nucleic acid complexes and labels of protein-nucleic acid binding sites from BioLip, and extracting a set of proteins binding to DNA and a set of proteins binding to RNA according to the types of bases in the complexes; The hierarchical graph neural network learns the local geometry and local patterns of the features of the binding site from the graph representation of the residue structure context in a supervised manner based on the training set samples, and specifically includes: a graph network coding unit, a K-layer stacked graph neuron block based on a gated recurrent unit, and a binary classifier based on a multilayer perceptron, wherein: the graph network coding unit sequentially performs edge coding, node coding, and graph coding; the K-layer stacked graph neuron block based on the gated recurrent unit sequentially performs edge updating, node updating, and graph updating; the binary classifier based on the multilayer perceptron performs the updated graph feature {u k |k=1,…,K} predicts the binding probability of the residue to DNA / RNA; The protein and nucleic acid binding site prediction uses a hierarchical graph neural network to predict the probability of each residue binding to DNA / RNA, and compares it with a threshold T to obtain binding residues and non-binding residues.

2. The prediction method according to claim 1, characterized in that: The sample fusion processing refers to: enhancing the training set data to address the imbalance problem of positive and negative samples in the data set, fusing the labels of binding sites of protein clusters with high sequence and structural similarity, and increasing the proportion of positive samples in the training set, specifically including: Step ① Cluster the proteins in the training set: use bl2seq and TM-align to calculate the sequence similarity and structural similarity of two proteins; cluster the proteins with sequence similarity greater than 0.8 and TM-score greater than 0.5 into one cluster; transfer the annotations of the true binding sites of proteins in the same cluster to the protein with the most residues in the cluster; remove proteins with sequence similarity greater than 30% in the training set; Step ② uses CD-HIT to remove proteins in the test set whose sequence similarity with the training set is greater than 30%, and ensures that the sequence similarity of the test set itself is less than 30%.

3. The prediction method according to claim 1, characterized in that: The position and characteristic information of the residues include: atomic characteristics, secondary structure information, bond and torsion angle information, solubility, and evolutionary conservation information, which are specifically extracted by the following methods: Step 1: For each residue, its position is defined as the coordinates of the center of mass of the residue; Step 2: For each residue, extract the features of the atoms belonging to the residue, and then average the features of all atoms of the residue to obtain the atomic feature matrix of the residue, with a size of L×7, where L represents the number of residues in the protein; Step 3: For proteins with L residues, use DSSP to calculate a secondary structure feature matrix of size L × 14 based on the protein structure; Step 4: For proteins with L residues, use PSI-BLAST to search the NCBI NR database and calculate the evolutionary conservation score matrix PSSM with a size of L×20; Step 5. For proteins with L residues, use HHblits to search the uniclust30 database and calculate another evolutionary conservation matrix HMM of size L×30; Step 6: Concatenate the atomic feature matrix, the secondary structure feature matrix, the evolutionary conservation scoring matrix PSSM and the evolutionary conservation matrix HMM to obtain an L×71 feature matrix for a protein with L residues.

4. The prediction method according to claim 1, characterized in that: The structural context is extracted by using a structure-based sliding ball for each residue of the protein according to the position information of the residue in the protein. Specifically, according to the position coordinates of the residue in the tertiary structure, a sliding ball along the polypeptide chain is used to extract the structural context of each residue. Specifically, for a target residue, the residue is taken as the center of the ball, r g Draw a sphere for the radius, and all the residues within the sphere and their positional relationships together constitute the structural context of the residue.

5. The prediction method according to claim 1, characterized in that: The graph representation is described as defining the residues in the structural context as nodes in the graph, constructing an adjacency matrix based on the relative position information of the nodes, and defining the characteristics of the nodes and edges of the graph based on the relative position relationship and the characteristics of the residues. The specific construction method is as follows: Step i) The residues within the structural context are taken as nodes of the graph, and the distance from the node to the center of the sphere is calculated as the 1D position code of the residue. The residue position code is concatenated with the position information of the extracted residue to obtain a node feature of size 72. The node feature set is expressed as The node feature of node i is represented by v i ; Step ii) Calculate the distance between two nodes to get the node distance matrix D, using the threshold r v Binarize the distance matrix to get the node adjacency matrix A. Specifically, for node i and node j, When A ij =1, it means there is an edge between node i and node j, namely edge ij; Step iii) The edge feature set is represented as E = {e ij |A ij =1}, where the feature e of edge ij ij Includes: 1) the distance between node i and node j, 2) the vector from node i to the center of the sphere The vector from node j to the center of the sphere The angle θ ij The cosine value of 6. The prediction method according to claim 1, characterized in that: The edge coding mentioned above refers to: splicing edge features and the node characteristics of the nodes at both ends and And through a nonlinear transformation layer Calculate the encoded edge features Node encoding refers to: splicing node features The sum of all encoded edge features starting from this node and passing through a nonlinear transformation layer Calculate the encoded node features Graph encoding means: calculating the sum of all the encoded node features on the graph and passing it through a nonlinear transformation layer Calculate the encoded graph feature u enc ; The implementation of the nonlinear transformation layer is MLP(X)=W(X)2max(0,W1X+b1)+b2, where X is the input, W1, W2, b1, b2 are the network parameters to be learned.

7. The prediction method according to claim 1, characterized in that: The edge update is: first concatenate the input edge features The node characteristics of its two end nodes and And the graph feature u k-1 , and calculate the edge features of the intermediate state through a nonlinear transformation layer The features of these two states are fused through the gated recurrent unit to obtain the updated edge features. Node updating means: first concatenate the input node features The sum of all updated edge features starting from this node and the graph feature u k-1 , and calculate the node features of the intermediate state through a nonlinear transformation layer The features of these two states are fused through the gate recurrent unit to obtain the updated node features. The graph update is to first concatenate the sum of the updated features of all nodes on the graph and the graph feature u k-1 , and calculate the intermediate state graph feature u through a nonlinear transformation layer k′ , the graph features of these two states are fused through the gate recurrent unit to obtain the updated graph feature u k .

8. The prediction method according to claim 1, characterized in that: The classifier in, 9. The prediction method according to claim 1, characterized in that: By performing grid search on hyperparameters, the hyperparameter combination that makes the Matthews coefficient of the validation set the highest is selected as the hyperparameter of the model, and the binarization threshold T that makes the Matthews coefficient of the validation set the highest is selected as the binarization threshold.

10. A system for implementing the method according to any one of claims 1 to 9, characterized in that: include: A feature extraction module, a structural context building module, a graph representation module, and a hierarchical graph neural network module, wherein the feature extraction module extracts the position and feature information of the protein from the sequence and structure of the protein, the structural context building module builds the local context of the residues through the position information of the protein, the protein information extracted by the feature extraction module and the structural context of the residues built by the structural context module are transmitted to the graph representation module to build a graph representation of the residue structural context, and the hierarchical graph neural network module uses the graph representation of the residue structural context as the input of the module to predict the probability that the residue is a protein-nucleic acid binding residue.

Images