Protein structure model quality assessment method based on graph neural network

Through a graph neural network-based method, combined with deep learning and field knowledge of protein structure, the shortcomings of the existing technology in the evaluation of high-precision protein structure model are solved, and more efficient and accurate protein structure model evaluation is achieved, especially in high-precision Excellent performance on the dataset.

CN114997366BActive Publication Date: 2024-11-05SHANGHAI JIAOTONG UNIV
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202210557804.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-19
Publication Date
2024-11-05
Estimated Expiration
2042-05-19

AI Technical Summary

Technical Problem

The existing protein structure model evaluation methods perform poorly on high-precision protein structure model sets, making it difficult to reliably predict the authenticity of protein-like bait structures, and it takes a long time and the accuracy is affected by the model pool distribution.

Method used

Using a graph neural network-based method, combining deep learning and domain knowledge of protein structure, neural network is constructed through triple attention mechanism and graph pooling technology, and using pre-trained model parameters to calculate the global and local scores of protein bait structure and natural structure. Improve the discernment and accuracy of the evaluation algorithm.

Benefits of technology

The evaluation accuracy of protein structure model is significantly improved on the high-precision dataset, shortens the evaluation time, achieves accuracy similar to existing first-class algorithms, and obtains optimal results on authoritative datasets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114997366B_ABST
    Figure CN114997366B_ABST
Patent Text Reader

Abstract

A protein structure model quality assessment method based on graph neural network, extracting the global coordinates of all 167 heavy atoms in the predicted protein-like bait in three-dimensional space as input, preprocessing it through dual-space relax of Rosetta software, analyzing and calculating the original node features and edge features and inputting them into the graph neural network model constructed based on attention mechanism and graph pooling technology, using the pre-trained model network parameters for calculation, and obtaining the global score and local score reflecting the gap between the protein-like bait structure and the real natural protein structure at the protein level and amino acid residue level. The present invention can more discriminately focus on those protein-like bait structures whose accuracy is close to that of the natural structure, thereby improving the accuracy of protein structure model assessment on high-precision data sets.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a technology in the field of bioengineering, specifically a method for evaluating the quality of protein structure models based on graph neural networks, which evaluates the authenticity of the model structure when the corresponding natural structure is unknown. Background Art

[0002] Protein structure model evaluation aims to quantify the accuracy of protein-like decoy structures generated by protein structure prediction models relative to their native structures without knowing the native structure. For structural bioinformatics involving proteins, protein structure model evaluation tools are extremely important infrastructure. Good evaluation methods can provide important guidance for tasks such as protein structure prediction, structure-based protein function prediction, and structure-based protein molecular docking. Generally speaking, protein structure model evaluation is divided into two types of methods: multi-model and single-model. The former requires the entire protein structure model pool as input, and evaluates the protein structure model by comparing the target protein structure model with other models in the pool and calculating the average structural similarity as a quality indicator; the latter only uses the sequence and structural information of a single decoy structure itself as input to predict its quality. Generally, multi-model methods can achieve false high accuracy when most models in the model pool are of good quality, but the time consumption often grows exponentially; the performance of single-model methods can be more consistent and more independent of the distribution of the quality of the model to be evaluated, which is closer to the expectation of the protein structure model evaluation algorithm, that is, learning the real structure that the sequence may form under natural conditions. In 2020, models led by Google's latest AI algorithm AlphaFold 2 proved in the officially organized CASP 14 structure prediction competition that deep learning of amino acid co-evolution data is sufficient to predict protein-like decoy structures with considerable accuracy for all protein sequences of unknown structures. Because past evaluation techniques tend to identify protein-like decoy structures with basic authenticity, the generation of a large number of high-precision protein-like decoy structures has caused a huge decline in the accuracy of almost all evaluation methods. In other words, as an important guiding part of the structure prediction algorithm, the existing model evaluation algorithm can no longer keep up with the rapid development of the structure prediction algorithm.

[0003] Existing multi-model quality assessment methods are more time-consuming and their accuracy is greatly affected by the distribution of the protein structure model pool itself; single-model quality assessment methods are closer to solving problems in this field, but the algorithm learning effect is not good; and almost all existing quality assessment methods find it difficult to make reliable predictions in high-precision protein models. Summary of the invention

[0004] In response to the defect that existing protein structure model evaluation methods perform poorly on high-precision protein structure model sets, the present invention proposes a protein structure model quality evaluation method based on graph neural network, which combines deep learning technology with domain knowledge of protein structure, and applies the neural network built based on the triple attention mechanism to the protein structure model evaluation algorithm. It takes the lead in focusing on the evaluation of high-precision protein structure models and improves and optimizes it. It can more discriminately focus on those protein-like bait structures whose accuracy is close to that of natural structures, thereby improving the accuracy of protein structure model evaluation on high-precision data sets.

[0005] The present invention is achieved through the following technical solutions:

[0006] The present invention relates to a protein structure model quality assessment method based on graph neural network. The global coordinates of all 167 heavy atoms in a protein decoy to be predicted in three-dimensional space are extracted as input, and after being preprocessed by dual-space relax of Rosetta software, original node features and edge features are obtained by analytical calculation and input into a graph neural network model constructed based on attention mechanism and graph pooling technology. The pre-trained model network parameters are used for calculation, and the global score and local score reflecting the gap between the structure of the protein decoy and the real natural protein structure at the protein level and the amino acid residue level are obtained respectively.

[0007] The 167 heavy atoms mentioned above are all atoms with relatively large atomic mass contained in the 20 standard amino acid residues that can constitute proteins in nature. Compared with other atoms, they can determine the skeleton posture and relative position of the protein in space.

[0008] The graph neural network model constructed based on the attention mechanism and graph pooling technology includes: a graph neural computing network with the same four-layer network layer architecture and built in series, and a graph pooling network connected to it consisting of two pooling layers in parallel, wherein: each network layer in the graph neural computing network is composed of a triple attention mechanism and a channel attention mechanism in series, and the two parallel pooling layers in the graph pooling network perform local pooling and global pooling on the node embeddings of the upstream graph respectively.

[0009] The pre-trained model network parameters are obtained by repeatedly training the above model architecture on the training set for ten rounds and then evaluating them on the validation set and the test set, respectively, and specifically include:

[0010] Step ① Take the sample protein bait structure from the training set, perform dual-space relax, and obtain the original node features and edge features from the structure analysis data to construct the training data queue Q train ={(X1,Y1),(X2,Y2),…,(Xn ,Y n )}, where: X is the protein map constructed from the original features of the sample, and Y is the true label value corresponding to the sample.

[0011] Step 2: Queue the training data Q train The data pairs in the batch are input into the neural network model in turn to get the output δ θ (X), and then calculate the loss function value Loss = ∑ batch (Y-δ θ (X) 2 +∑ label>0.5 (Y-δ θ (X) 2 , where: θ It is a graph neural network model.

[0012] Step ③ Update δ using the stochastic gradient descent algorithm based on the loss function θ The kth parameter θ in the network k , specifically Where: λ is the preset learning rate 0.001, is the corresponding gradient value obtained by the loss function.

[0013] Step ④ After traversing all samples in the training set, input the samples in the validation set into δ θ Test in the network, calculate the correlation coefficient and absolute value difference between the predicted values ​​and label values ​​of all samples in the validation set, compare them with the results of the previously trained model, and select the better version to save.

[0014] Step ⑤ After completing the specified number of training iterations, take out the historical optimal model parameters saved based on the validation set, input the samples in the test set, calculate the prediction score of the model on the test set and obtain the correlation coefficient. Then, based on the correlation coefficient, you can evaluate whether the current model design is reasonable and decide whether to terminate the training of the model or adjust the relevant hyperparameters.

[0015] The training set and validation set are obtained by using different prediction methods and prediction software to perturb different amino acid sequences of protein structures in the human protein database, which are known experimentally, and then dividing them into a large number of protein-like bait structures in a ratio of 8:2.

[0016] The test set refers to the CASP standard data set, which comes from the official data set provided by the CASP competition. CASP, also known as the Critical Assessment of Protein Structure Prediction, is a global experiment for the entire community. In each CASP competition, a group of protein sequences with unknown structures but whose true structures are about to be solved by the world's top laboratories will be sorted out, and they will be provided to all participating teams for structure prediction, and then all predicted structures will be collected and provided to the protein structure model evaluation team for calculation. Therefore, its official data set is the gold standard in the research field of this invention.

[0017] The protein graph is a data representation of a protein-like decoy structure. Each protein graph consists of defined nodes and edges, where nodes represent various heavy atoms in proteins and edges represent the relationship between each node atom and its 50 nearest neighbor atoms in three-dimensional space. Each node and edge has original features.

[0018] The original node feature refers to the one-hot vector encoding the atomic species at the node.

[0019] The edge features specifically include:

[0020] Category 1) Edge distance: For every two atoms that will be connected in the protein graph, their relative distance in the sample is calculated. In this work, Gaussian basis expansion is used to generate a vector representation of the distance. Its mean varies uniformly from [0, 15] with a step size of 0.4.

[0021] Category 2) Edge coordinates: To account for the apparent anisotropy of relative atomic positions within proteins, a set of directional features is used to supplement the above distances. For each heavy atom B, it can form a local reference frame with its preceding and succeeding atoms A and C in the amino acid sequence. The three unit vectors defining this reference frame are They are calculated as follows. So when considering any two atoms B in space j ,B j When the edge coordinate features are i ,B j The three-dimensional spatial coordinates are replaced into the local reference frame of the other party, then the edge (i, j) in the protein graph obtains 2×3 features, which are atoms B i (B j )'s spatial coordinates are in the atom B j (B i )'s local coordinate system.

[0022] Category 3) Chemical bond: If the connection between the two atoms is a chemical bond in an actual protein structure, it is set to 1, otherwise it is set to 0, that is, a Boolean feature.

[0023] The true label value refers to: the accurate score of a protein-like decoy structure relative to its corresponding true structure. Similarly, the true label values ​​are divided into two categories: global and local. When the natural true structure corresponding to the sequence is known, the global and local label values ​​are implemented by two algorithms, global distance test (GDT-TS) and local distance difference test (lDDT), respectively, where: the GDT-TS score is the maximum set within the defined distance cutoff range of the position of the α-carbon atom of the amino acid residue in the model structure in the natural structure after iterative superposition of the two structures. lDDT is a non-superposition score used to evaluate the local distance differences of all atoms in the model, including verifying the rationality of stereochemistry. It is calculated on all atomic pairs of the natural protein structure and measures the degree of reproduction of the environment in the natural structure in the protein-like decoy structure. These two methods are the gold standards for measuring the global and local fields of structural similarity, respectively.

[0024] The correlation coefficient mentioned is: Pearson correlation coefficient Among them: cov(X,Y) corresponds to the covariance of the two vectors, and σ corresponds to the standard deviation of the vector.

[0025] The present invention relates to a system for implementing the above method, comprising: a protein graph generation module, a feature extraction network based on an attention mechanism, a graph pooling network, a loss function module, a parameter updating module and a verification module, wherein: the protein graph generation module extracts original node features and edge features from sample protein bait structure data respectively, and constructs the extracted features into an original protein graph; the feature extraction network based on the attention mechanism obtains high-dimensional features of the protein graph through forward learning; the graph pooling network performs simple calculation and average processing on the high-dimensional features to obtain global and local prediction scores respectively; the obtained scores are sent to the loss function module together with the true scores corresponding to the samples to calculate the losses; the network parameters are updated in the parameter updating module according to the losses by using back propagation and gradient descent algorithms; the training iterations are repeated until the verification module determines that the model has achieved the expected effect and stops.

[0026] Technical Effects

[0027] The present invention can achieve similar accuracy to existing first-class evaluation algorithms on authoritative public datasets in a shorter running time than them, and obtain optimal results on high-precision subsets. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the present invention;

[0029] Figure 2Schematic diagram of the training process of the graph neural network model built based on the attention mechanism and graph pooling technology;

[0030] Figure 3 Schematic diagram of the architecture for implementing a graph neural network based on an attention mechanism. DETAILED DESCRIPTION

[0031] like Figure 1 As shown, this embodiment relates to an evaluation method of a protein structure model based on deep learning, comprising the following steps:

[0032] Step 1) First, the Cartesian coordinate information of the heavy atoms of each residue in the protein-like bait structure in three-dimensional space is extracted, and then the structure is preprocessed by dual-relax of Rosetta software to relax the structure under dual space constraints.

[0033] Step 2) Calculate the Euclidean distance between each pair of heavy atoms according to the coordinate information of each heavy atom, construct an adjacency matrix according to the distance, and calculate the original node features and edge features of each heavy atom, specifically including:

[0034] Step 2.1) Given a protein containing L residues, assume that it contains N heavy atoms in total, where: the Cartesian coordinates of the i-th heavy atom in three-dimensional space are c i =(x i ,y i ,z i ), the coordinates of the jth heavy atom are c j =(x j ,y j ,z j ), then the Euclidean distance between these two residues is d ij =||c i -c j ||, in practical applications, Gaussian basis expansion is used to generate a vector representation of the distance. Its mean changes uniformly from [0,15] with a step size of 0.4. The distance matrix of this protein is The atomic coordinates of the protein are

[0035] Step 2.2) Based on the distance matrix obtained above By traversing and finding the M nearest neighboring atoms of each heavy atom in three-dimensional space, its adjacency matrix can be obtained. Where: (i,j)∈a ij To determine whether two heavy atoms are connected by an edge in the protein graph, k is a hyperparameter used for normalization, and k = 50 is taken in practical applications.

[0036] Step 2.3) According to the properties of each heavy atom in the protein, the original node features of each heavy atom are obtained, and the set of original node features is recorded as v i is a 167-dimensional one-hot vector that independently identifies the atomic species.

[0037] Step 2.4) According to the adjacency matrix Distance Matrix The coordinate set C can calculate other edge features besides the Euclidean distance feature, such as contact features and projection features, where: Hyperparameter d0=8, Projection Matrix The coordinates of the i-th heavy atom are known to be c i =(x i ,y i ,z i ), then the coordinates of the preceding and succeeding heavy atoms in the amino acid sequence are c i-1 =(x i+1 ,y i+1 ,z i+1 ) and c i+1 =(x i+1 ,y i+1 ,z i+1 ), then the local reference system can be calculated from the coordinates of these three heavy atoms according to the following formula: Similarly, the local coordinate system of the jth heavy atom can be obtained and Then the projection vector p ij is the connection vector between two heavy atoms (c j -c i ) is the concatenation of the projections in the two local coordinate systems. This allows us to define Where: μ ij =[d ij ,b ij ,p ij ], a total of 47 dimensions of features.

[0038] Step 2.5) In summary, the protein graph can be defined as In the figure Used to show whether the i-th and j-th heavy atoms in the graph are connected. is the original node feature of the i-th heavy atom, Used to characterize the edge between the i-th and j-th heavy atoms.

[0039] Step 3) Build a training data queue in: is the protein graph constructed from the original features of the sample, Y is the true label value corresponding to the sample, and Y n ∈(0,1).

[0040] Step 4) Construct a graph neural network learning framework, which includes a graph convolution network based on an attention mechanism and a graph pooling calculation network. Input the training data into the graph convolution network to calculate the high-dimensional features of the protein graph, and then perform graph pooling calculations on the high-dimensional features to obtain the local and global scores of the sample. The obtained scores are sent to the loss function module together with the true scores corresponding to the sample to calculate the loss; the network parameters are updated in the parameter update module using the back propagation and gradient descent algorithms based on the loss; the training iterations are repeated until the verification module determines that the model has achieved the expected effect and stops.

[0041] Step 4.1) Take out a batch of protein images First, it is sent to the graph convolutional network based on the attention mechanism to obtain high-dimensional features. The network consists of 4 layers of the same architecture. For each layer of the network, the total calculation is as follows: Where: (k) is the high-dimensional feature map e (k) The attention weights obtained by the squeeze-and-excitation module are: e (k) is τ (k) After the full connector, τ (k) It is a high-dimensional feature map obtained by the triple attention mechanism module, specifically: in: are the original feature maps z (k) Four feature maps obtained through four multi-head full connectors with the same framework, the subscript h means multi-head: Where: the superscript k is the number of the current corresponding network layer, which can be 1, 2, 3, or 4; is the softmax function, σ(·) is the sigmoid function, δ(·) is the relu nonlinear activation function; ⊙ is the element-by-element multiplication, is the concatenation between vectors; and are the convolution weight matrix and convolution bias matrix of the kth layer, respectively, where: the subscript h means multi-head, and in this work the subscript h is the pre-set hyperparameter of 3.

[0042] Step 4.2) The new node embedding obtained by forward calculation is v′ i , using graph pooling technology to obtain global and local scores, specifically: the final global score The local score of the kth amino acid in: is the set of heavy atoms that constitute the kth amino acid, b l , b m are the weights and biases of the global and local fully connected layers, respectively.

[0043] Step 4.3) Update δ using the stochastic gradient descent algorithm based on the loss function θ The kth parameter θ in the network k , specifically Among them: λ is the preset learning rate, is the corresponding gradient value obtained by the loss function. Specifically, assuming that the sample consists of L amino acids, the final score obtained by the network is The corresponding true label value is also Then the loss function of this sample is Where: ε(·) is a step function.

[0044] Step 4.4) After completing the specified number of iterations, query the model parameters that perform best on the validation set under the storage path and make predictions on the test set by comparing the Pearson coefficients of the two vectors (Score, Y) on the test set. Determine whether hyperparameters need to be adjusted, where cov(·) is the covariance and Δ is the standard deviation.

[0045] Step 5) For the protein-like decoy structure waiting to be predicted, firstly perform the processing of steps 1 and 2 to obtain its protein graph, which is then input into the graph neural network to obtain the final result, which is the prediction score for the protein-like decoy structure.

[0046] This example uses DeepAccNet (from Hiranuma, N., Park, H., Baek, M. et al. Improved protein structure refinement guided by deep learning based accuracy estimation. Nat Commun 12, 1340 (2021). https: / / doi.org / 10.1038 / s41467-021-21511-x) and GNNRefine (from Jing, X., Xu, J. Fast and effective protein model refinement using deep graph neural networks. Nat Comput Sci 1,462–469(2021).https: / / doi.org / 10.1038 / s43588-021-00098-9) were used as training and validation sets. The two datasets contained 1,000,000 protein-like decoy structures generated from 7,900 protein sequences and 500,000 protein-like decoy structures generated from 29,400 protein sequences, respectively. The CASP13 and CASP14 competition datasets were used as test sets, containing 3,000 and 5,000 protein-like decoy structures, respectively. The Pearson correlation and AUC of the global scores were compared on the test sets as the criteria for evaluating the quality of the models.

[0047] In the biological definition, when the score of a protein-like decoy structure relative to the real protein structure reaches 0.5, it can be considered that the protein-like decoy structure has a certain degree of authenticity. In order to prove the superiority of the invention in high-precision structure evaluation, samples with real label values ​​greater than 0.5 in the CASP13 and CASP14 protein-like decoy structure data sets were screened to obtain two new test data sets CASP13_H and CASP14_H, and comparisons were also made on these two data sets.

[0048] The final results on the CASP13 and CASP14 ranking tasks are shown in Table 1-2. Compared with the top four single-model evaluation methods in the competition, the Pearson correlation value and AUC value of this method remain in the first-class echelon within an acceptable range, where: The AUC value is defined as the area under the ROC curve, which reflects the ability of the method to distinguish between protein-like value bait structures with poor authenticity and good authenticity (based on whether GDT-TS is greater than 0.5).

[0049] Table 1: Comparison on CASP13

[0050] method AUC PEARSON ProQ3D 95.83% 0.8510 MESHI-enrich-server 95.08% 0.8318 ProQ3 95.05% 0.8316 MASS1 94.28% 0.8096 This method 92.32% 0.8247

[0051] Table 2: Comparison on CASP14

[0052] method AUC PEARSON ProQ4 80.79% 0.6278 Bhattacharya-QDeep 78.88% 0.6249 QMEANDisCo 78.84% 0.6120 Bhattacharya-QDeepU 78.38% 0.6084 This method 82.81% 0.6383

[0053] The final results of the CASP13_H and CASP14_H ranking tasks are shown in Tables 3-4. Compared with the 14 best-ranked evaluation methods in the competition (including multi-model methods that often perform better in data indicators), it shows that they can achieve a significant lead in Pearson correlation values ​​and can better rank and select protein-like baits whose authenticity is closer to the natural structure.

[0054] Table 3: Comparison on CASP13_H

[0055] method PEARSON MUfoldQA_T 0.7636 MULTICOM_CLUSTER 0.8022 Davis-EMA consensus 0.7733 ModFOLDclust2 0.7777 FALCON-QA 0.7516 RaptorX-DeepQA 0.6791 Bhattacharya - ClustQ 0.7124 ProQ3D 0.6317 FaeNN 0.8115 Pcons 0.5867 Pcomb 0.5560 MESHI-enrich-server 0.7772 ProQ3 0.7188 MASS1 0.4661 This method 0.8646

[0056] Table 4: Comparison on CASP14_H

[0057]

[0058]

[0059] Similarly, for 1000 protein-like decoy structures randomly selected from the validation set, the prediction time comparison between this method and other first-class methods using CPU is shown in Table 5.

[0060] Table 5: Comparison of prediction time on 1000 randomly selected decoys

[0061] method Time ProQ4 27min 22s Bhattacharya-QDeep 31min 3s QMEANDisCo 2h 3min 0s Bhattacharya-QDeepU 33min 10s This method 10min 46s

[0062] Compared with the existing technology, the prediction time used by this method is less than that of the existing technology, and the evaluation accuracy is on par with the existing first-class technology; in the high-precision samples of the authoritative data set, the evaluation accuracy of this method is better than all the existing first-class technologies.

[0063] 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 protein structure model quality assessment method based on graph neural network, It is characterized in that The global coordinates of all 167 heavy atoms in the predicted protein-like decoy in three-dimensional space are extracted as input, and then preprocessed by dual-space relax of Rosetta software. The original node features and edge features are obtained by analytical calculation and input into the graph neural network model built based on attention mechanism and graph pooling technology. The pre-trained model network parameters are used for calculation to obtain the global score and local score reflecting the gap between the structure of this type of protein decoy and the real natural protein structure at the protein level and amino acid residue level. The graph neural network model constructed based on the attention mechanism and graph pooling technology includes: a graph neural computing network with the same four-layer network layer architecture and built in series, and a graph pooling network connected to it consisting of two pooling layers in parallel, wherein: each network layer in the graph neural computing network is composed of a triple attention mechanism and a channel attention mechanism in series, and the two parallel pooling layers in the graph pooling network perform local pooling and global pooling on the node embeddings of the upstream graph respectively.

2. The protein structure model quality assessment method based on graph neural network according to claim 1, Its characteristics are: The pre-trained model network parameters are obtained by repeatedly training the above model architecture on the training set for ten rounds and then evaluating them on the validation set and the test set, respectively, and specifically include: Step ① Take the sample protein bait structure from the training set, perform dual-space relax, obtain the original node features and edge features from the structure analysis data, and build the training data queue , where: X is the protein map constructed from the original features of the sample, and Y is the true label value corresponding to the sample; Step 2: Queue the training data The data pairs in the batch are input into the neural network model in turn to get the output , and then calculate the loss function value ,in: It is a graph neural network model; Step ③ Update using stochastic gradient descent algorithm based on the loss function The kth parameter in the network , specifically ,in: The default learning rate is 0.

001. is the corresponding gradient value obtained by the loss function; Step ④ After traversing all samples in the training set, input the samples in the validation set into Test in the network, calculate the correlation coefficient and absolute value difference between the predicted value and label value of all samples in the validation set, compare with the results of the previously trained model, and select the better version to save; Step ⑤ After completing the specified number of training iterations, take out the historical optimal model parameters saved based on the validation set, input the samples in the test set, calculate the prediction score of the model on the test set and obtain the correlation coefficient. Then, based on the correlation coefficient, you can evaluate whether the current model design is reasonable and decide whether to terminate the training of the model or adjust the relevant hyperparameters.

3. The protein structure model quality assessment method based on graph neural network according to claim 2, Its characteristics are: The training set and validation set are obtained by using different prediction methods and prediction software to perturb different amino acid sequences of protein structures in the human protein database, which are known through experiments, and then dividing them into a large number of protein-like decoy structures in a ratio of 8:2; The test set refers to the CASP standard data set, which is an official data set provided by the CASP competition. CASP, also known as the Critical Assessment of Protein Structure Prediction, is a global experiment for the entire community. In each CASP competition, a group of protein sequences with unknown structures but whose true structures are about to be solved by the world's top laboratories will be compiled, and they will be provided to all participating teams for structure prediction. All predicted structures will then be collected and provided to the protein structure model evaluation team for calculation. Therefore, its official data set is the gold standard in the research field of the present invention.

4. The protein structure model quality assessment method based on graph neural network according to claim 2, Its characteristics are: The protein graph refers to the data representation of the protein-like decoy structure. Each protein graph is composed of nodes and edges, wherein the nodes represent various heavy atoms in the protein, and the edges represent the relationship between each node atom and its 50 nearest neighbor atoms in three-dimensional space. Each node and edge has original features. The original node feature refers to the use of a one-hot vector to encode the atomic species at the node.

5. The protein structure model quality assessment method based on graph neural network according to claim 4, Its characteristics are: The edge features specifically include: Category 1) Edge distance: For every two atoms that will be connected in the protein graph, their relative distance in the sample is calculated. In this work, Gaussian basis expansion is used to generate a vector representation of the distance; Category 2) Edge coordinates: Each heavy atom B forms a local reference frame with its preceding atom A and succeeding atom C in the amino acid sequence; the three unit vectors of the reference frame are ; When considering any two atoms in space When the edge coordinate features are The three-dimensional coordinates of the protein graph are replaced by the local reference frame of the other party. get features, respectively, atoms The spatial coordinates of the atom The projection on the local coordinate system of Category 3) Chemical bonds: When atoms If the connection is a chemical bond in an actual protein structure, it is set to 1, otherwise it is set to 0, which is a Boolean feature.

6. The protein structure model quality assessment method based on graph neural network according to claim 2, Its characteristics are: The true label value refers to: the accurate score of a protein-like bait structure relative to its corresponding true structure; the true label value is divided into two categories: global and local: when the natural true structure corresponding to the sequence of the sample corresponding to the true label value is known, the global and local label values ​​are respectively implemented by the global distance test (GDT-TS) and local distance difference test (lDDT) algorithms.

7. The protein structure model quality assessment method based on graph neural network according to claim 2, Its characteristics are: The correlation coefficient mentioned is: Pearson correlation coefficient ,in: The covariance of the two vectors is The standard deviation of this vector.

8. A system for implementing the protein structure model quality assessment method based on graph neural network as described in any one of claims 1 to 7, It is characterized in that include: A protein graph generation module, a feature extraction network based on an attention mechanism, a graph pooling network, a loss function module, a parameter updating module and a verification module, wherein: the protein graph generation module extracts original node features and edge features from sample protein bait structure data respectively, and constructs the extracted features into an original protein graph; the feature extraction network based on the attention mechanism obtains the high-dimensional features of the protein graph through forward learning; the graph pooling network performs simple calculations and average processing on the high-dimensional features to obtain global and local prediction scores respectively; the obtained scores are sent to the loss function module together with the true scores corresponding to the sample to calculate the loss; the network parameters are updated in the parameter updating module according to the loss using back propagation and gradient descent algorithms; the training iterations are repeated until the verification module determines that the model has achieved the expected effect and stops.

Citation Information

Patent Citations

  • Histocyte pathology image diagnosis system based on graph neural network

    CN113674252A

  • Protein-ligand binding site prediction method based on deep learning

    CN113707213A

  • Protein design automation for protein libraries

    EP1482434A2

  • Accelerating convolutional neural network computation throughput

    US20190102640A1

  • Machine learning based methods of analysing drug-like molecules

    WO2020016579A2