A Protein Sequence Design Method Based on Iterative Computational Graph Neural Networks

Abstract

This invention relates to the field of bioinformatics and discloses a protein sequence design method based on an iterative computational graph neural network. The method includes: acquiring a protein sequence dataset; extracting the main chain atomic coordinates of the protein sequence and adding noise to the end of the main chain atomic coordinates; designing virtual atoms based on the main chain atomic coordinates and obtaining their coordinates; establishing distance features between the main chain atomic coordinates and virtual atoms, and obtaining distance representations based on these features; constructing angle features between main chain atoms; obtaining the embedding expressions of node features based on the angle features, and obtaining the embedding expressions of edge features based on the distance representations; encoding standard amino acid types and unknown amino acids; constructing and training the iterative computational graph neural network; and inputting the embedding expressions of node features, edge features, and the encoded amino acid sequence into the trained iterative computational graph neural network to obtain the protein sequence. This invention can improve the accuracy of generated protein sequences.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a protein sequence design method based on iterative computational graph neural networks. Background Technology

[0002] Proteins are composed of one or more amino acid sequences. Amino acids are organic compounds, including amino and carboxyl functional groups, as well as amino acid-specific side chains (i.e., atomic groups). In bioengineering research, designing protein sequences capable of forming specific three-dimensional structures is a crucial topic; this process is known as structure-based protein design. With a deeper understanding of the relationship between protein structure and function, researchers have begun to attempt to artificially design and create proteins with specific functions to meet the urgent needs of medicine, agriculture, industry, and other fields. Currently, protein design largely utilizes physics-based methods, such as Rosetta's approach of treating sequence design as an energy optimization problem. However, even a single protein sequence of length 200 has a vast sequence search space, requiring enormous computational resources for such methods. Furthermore, the resulting sequences often lack precision and involve extensive trial and error, making them extremely costly.

[0003] In recent years, the rapid development of deep learning and artificial intelligence technologies has brought new revolutionary opportunities to the field of protein design. Traditional protein design methods often rely on experimental verification and expert experience, resulting in long design cycles, low accuracy, and difficulty in covering all possibilities of protein function. Summary of the Invention

[0004] The primary objective of this invention is to overcome the problems existing in the prior art and provide a protein sequence design method based on iterative computational graph neural networks, which can improve the accuracy of generated protein sequences.

[0005] To achieve the above objectives, this invention provides a protein sequence design method based on iterative computational graph neural networks, the method comprising the following steps:

[0006] Obtain a protein sequence dataset and divide the protein sequence dataset into a training set, a validation set, and a test set;

[0007] Extract the main chain atomic coordinates of the protein sequence and add noise at the end of the main chain atomic coordinates;

[0008] Virtual atoms are designed and their coordinates are obtained based on the main chain atom coordinates. The number of virtual atom coordinates is the same as the number of main chain atom coordinates, and the specific method for obtaining the virtual atom coordinates is as follows:

[0009]

[0010] in, , , All are user-defined constants with an absolute value less than 1. , , Main chain atoms;

[0011] Establish the distance features between the main chain atom coordinates and the virtual atoms, and input the distance features into the Gaussian radial function to obtain the distance characterization;

[0012] Construct the angular features between the atoms in the main chain;

[0013] The embedding representation of node features is obtained based on the angle features, and the embedding representation of edge features is obtained based on the distance representation;

[0014] The standard amino acid types that constitute the protein sequence and unknown amino acids are encoded to obtain the encoded amino acid sequence.

[0015] The iterative computation graph neural network constructed using the training set is trained iteratively. Each time it is trained, the generation accuracy of the trained iterative computation graph neural network is evaluated using the validation set. The iterative computation graph neural network with the highest accuracy is the trained iterative computation graph neural network. The trained iterative computation graph neural network is tested using the test set.

[0016] The embedded expressions of the node features, the embedded expressions of the edge features, and the encoded amino acid sequence are input into the trained iterative computational graph neural network. Sequential decoding is used, and graph updates are performed by iteratively calculating the attention between edge features and node features to obtain the protein sequence.

[0017] Furthermore, the step of obtaining the protein sequence dataset, and dividing the protein sequence dataset into a training set, a test set, and a validation set, specifically includes:

[0018] A first protein sequence dataset is obtained from a protein database according to a first screening condition, wherein the first screening condition is that the protein sequence length is less than 900.

[0019] The protein sequence dataset is obtained from the first protein sequence dataset according to the second screening condition, wherein the second screening condition is to delete protein sequences with a similarity higher than 40%.

[0020] The protein sequence dataset was divided into a training set, a validation set, and a test set in a 9:1:1 ratio.

[0021] Furthermore, the addition of noise to the main chain atomic coordinates is specifically done as follows:

[0022]

[0023] in, The set of main chain atom coordinates; This is a custom value used to adjust the noise intensity. It follows a normal distribution.

[0024] Furthermore, a distance feature is established between the main chain atom coordinates and the virtual atom. This distance feature is then input into a Gaussian radial function to obtain a distance representation, specifically including:

[0025] Construct the distance features between the main chain atom coordinates and the virtual atoms;

[0026] The distance feature is input into a Gaussian radial function, and by... Uniformly sample 16 different types of distance intervals Sixteen different distance representations were obtained for the same distance, and the specific calculation formulas are as follows:

[0027]

[0028] in, It is a radial basis function. The base of the natural index, The input is the distance feature. for Equal-interval sampling, with an interval of [missing information]. .

[0029] Furthermore, the angular features between the main chain atoms are specifically included in the construction of the first angular feature corresponding to the dihedral angle of the main chain atoms and the second angular feature corresponding to the torsion angle. The dihedral angle includes three types, denoted as follows: , and The torsion angle includes three types, namely: , and , Main chain The dihedral angle formed Main chain The dihedral angle formed Main chain The dihedral angle formed In the main chain The resulting angle of twist In the main chain The resulting angle of twist It represents the rotation angle on the backbone of two adjacent peptide bonds.

[0030] Furthermore, the distance representation is input into a linear layer to obtain the embedded representation of the edge features, and the angle features are input into a linear layer to obtain the embedded representation of the node features.

[0031] Furthermore, the iterative computational graph neural network comprises two parts: an encoder and a decoder, specifically:

[0032] The encoder is used to update the embedding representation of the edge features and the embedding representation of the node features. Taking the current residue as an example, the encoder calculates the attention of the node and the information of the K neighboring nodes, gives higher weight to the nodes with deeper connections, and then uses the attention to update the information of the current node. This process is repeated three times, that is, iteratively calculating the attention to update the embedding representation of the edge features and the embedding representation of the node features.

[0033] The decoder is used to decode information and iteratively generate residues. Taking the current node as an example, the decoder uses a multilayer perceptron to update the edge features, then calculates the attention of the node and the K neighboring nodes, and finally gathers the edge feature information of the residues in the K neighboring nodes. The information of the current node is updated through this information, and the residue probability information of the current node is calculated. Based on the probability information, the amino acid with the highest probability is selected as the residue of the node. After decoding the current node, the process is repeated for the next node.

[0034] Furthermore, it also includes configuring network parameters, setting the network drop rate to... Label smoothing rate The batch size is 2500 sequences in length, and the number of neighboring nodes K is set to a certain value. The loss function is defined as the cross-entropy between the generated sequence and the natural sequence.

[0035] Furthermore, the encoding of standard amino acid types and unknown amino acids is as follows: Alanine A: 0, Cysteine ​​C: 1, Aspartic acid D: 2, Glutamic acid E: 3, Phenylalanine F: 4, Glycine G: 5, Histidine H: 6, Isoleucine I: 7, Lysine K: 8, Leucine L: 9, Methionine M: 10, Asparagine N: 11, Proline P: 12, Glutamine Q: 13, Arginine R: 14, Serine S: 15, Threonine T: 16, Valine V: 17, Tryptophan W: 18, Tyrosine Y: 19, Unknown amino acid X: 20.

[0036] Compared with the prior art, the advantages of this invention are as follows:

[0037] 1. By calculating the distance feature in the edge features and the angle feature in the node features, the distance feature can intuitively reflect the spatial relationship between atoms in the protein sequence, while the angle feature can reveal the stability and directionality of chemical bonds. Calculating these two features can accurately capture high-order related information in the complex structure of proteins. By integrating these two features, we can gain a deeper understanding of the relationship between the internal structure and function of proteins, thereby designing more accurate and efficient proteins.

[0038] 2. Furthermore, this invention adds noise at the end of the main chain. This noise enables the network to cope with uncertainties and disturbances during the design process, while also simulating minute changes in protein structure in real-world environments. This improves the flexibility and robustness of the iterative graph neural network.

[0039] 3. In terms of algorithm design, this invention adopts an iterative calculation method. Through multiple iterations, the model can continuously adjust and optimize the design parameters to approximate the ideal protein structure. Attached Figure Description

[0040] Figure 1 This is a flowchart of a protein sequence design method based on an iterative computational graph neural network according to an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the encoder of the iterative computation graph neural network according to an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the decoder of the iterative computation graph neural network according to an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of sequential decoding according to an embodiment of the present invention. Detailed Implementation

[0044] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0045] In the description of this invention, it should be noted that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for 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. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0046] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0048] Example 1

[0049] like Figure 1-4 As shown, a preferred embodiment of the present invention provides a protein sequence design method based on an iterative computational graph neural network, comprising:

[0050] S1: Obtain the protein sequence dataset and divide the protein sequence dataset into a training set, a validation set, and a test set;

[0051] In one embodiment, the Protein Data Bank (PDB (https: / / www.rcsb.org / )) is currently the most commonly used protein structure database, storing over 70,000 labeled PDB structures. In this embodiment, representative proteins with sequence lengths less than 900 are selected from the PDB database. Using the highest sequence similarity of 40% as a boundary, the selected proteins are deredundant using CD-HIT software. The resulting protein set is the model's dataset, containing 23,622 proteins. The protein sequence dataset is divided into training, validation, and test sets in a 9:1:1 ratio, with a batch size of 2,500 sequences. The Adam optimizer and ReduceLROnPlateau learning rate scheduler are used, and the learning rate is adjusted when the perplexity of the validation set no longer decreases.

[0052] S2: Extract the main chain atomic coordinates of the protein sequence and add noise to the end of the main chain atomic coordinates;

[0053] In one feasible embodiment, taking protein 1a1x_A with a sequence length of 109 residues as an example, the main chain atoms of this protein sequence are composed of , (The carbon atom, located at the center of an amino acid, is attached to the amino group of the amino acid.) ),carboxyl( ) and a hydrogen atom ( ) and a sidechain ( base)), , Following the sequential composition, we used the Biopython package to extract the main chain atomic coordinates of the protein sequence from the dataset. During the extraction process, we excluded amino acids with incomplete three-dimensional coordinate information, i.e., any missing amino acids. , , or The amino acid at any atomic coordinate within an atom is used to ensure the completeness and accuracy of the extracted data. Taking a protein as an example, the backbone atoms are extracted sequentially for each residue in the protein. , , , The coordinates.

[0054] To improve the model's performance and enhance its generalization ability, this embodiment adds Gaussian noise to the ends of the extracted main chain atoms. This will blur local details to some extent and improve the model's focus on the overall shape of the protein. The formula is as follows:

[0055]

[0056] in, The set of main chain atom coordinates; This is a custom value, adjusted due to noise intensity. It follows a normal distribution.

[0057] S3: Design virtual atoms and obtain virtual atom coordinates based on the main chain atom coordinates;

[0058] In one feasible embodiment, based on the current residue , , The coordinates of the atoms are used to calculate the coordinates of the four virtual atoms: , , and The formula is as follows:

[0059]

[0060] in, , , Each atom represents the main atom coordinate of the current residue. , , All of them are randomized constants with an absolute value less than 1.

[0061] S4: Establish the distance features between the main chain atom coordinates and the virtual atom, and input the distance features into the Gaussian radial function to obtain the distance characterization;

[0062] In one feasible embodiment, distance features are constructed: This embodiment constructs pairwise distance features between main chain atoms and virtual atoms: - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - , - We input the distance feature into the Gaussian radial function RBF. By... Uniformly sample 16 different types of distance intervals Sixteen different representations of the same distance were obtained, as shown in the following formula:

[0063]

[0064] in, It is a radial basis function. The base of the natural index, The input is the distance feature. for Equal-interval sampling, with an interval of [missing information]. .

[0065] S5: Constructing the angular features between main chain atoms;

[0066] In one feasible embodiment, angular features between main chain atoms are constructed. These angular features specifically include a first angular feature corresponding to the dihedral angle of the main chain atoms and a second angular feature corresponding to the torsion angle. The dihedral angle includes three types, denoted as follows: , and The torsion angle includes three types, namely , and , Main chain The dihedral angle formed Main chain The dihedral angle formed Main chain The dihedral angle formed In the main chain The resulting angle of twist In the main chain The resulting angle of twist It represents the rotation angle on the backbone of two adjacent peptide bonds.

[0067] S6: Obtain the embedded representation of node features based on the angle features, and obtain the embedded representation of edge features based on the distance representation;

[0068] The angle features are input into the first linear layer to obtain the embedded representation of the node features, and the distance representation is input into the second linear layer to obtain the embedded representation of the edge features.

[0069] In one feasible embodiment, the linear layer can be an MLP.

[0070] S7: Encode the standard amino acid types that constitute the protein sequence and the unknown amino acids to obtain the encoded amino acid sequence;

[0071] In one feasible embodiment, 20 standard amino acid types and unknown amino acids are encoded, and then these encoded amino acid sequences are used as input data for the graph neural network decoder module, namely: alanine A:0, cysteine ​​C:1, aspartic acid D:2, glutamic acid E:3, phenylalanine F:4, glycine G:5, histidine H:6, isoleucine I:7, lysine K:8, leucine L:9, methionine M:10, asparagine N:11, proline P:12, glutamine Q:13, arginine R:14, serine S:15, threonine T:16, valine V:17, tryptophan W:18, tyrosine Y:19, and unknown amino acid X:20.

[0072] S8: The iterative computation graph neural network constructed by iteratively training the training set is used. Each time it is trained, the generation accuracy of the trained iterative computation graph neural network is evaluated using the validation set. The iterative computation graph neural network with the highest accuracy is the trained iterative computation graph neural network. The trained iterative computation graph neural network is tested using the test set.

[0073] In one feasible embodiment, the iterative computational graph neural network comprises two main parts: an encoder and a decoder, specifically:

[0074] The structure of the encoder is as follows Figure 2 As shown, the encoder is used to update the embedding representations of the edge features and the node features. Taking the current residue as an example, the encoder calculates the attention between the node and its K neighboring nodes, giving higher weights to nodes with deeper connections. Then, it uses the attention to update the information of the current node. This process is repeated three times, that is, iteratively calculating the attention to update the embedding representations of the edge features and the node features. It should be noted that the pseudocode of the encoder is as follows:

[0075]

[0076] The structure of the decoder is as follows Figure 3As shown, the decoder is used to decode information and iteratively generate residues. Taking the current node as an example, the decoder uses a multi-layer perceptron to update the edge features, then calculates the attention of the node and the information of the neighboring K nodes, and finally aggregates the edge feature information of the residues in the neighboring K nodes, and updates the information of the current node through this information to calculate the residue probability information of the current node. According to the probability information, the amino acid with the highest probability is selected as the residue of the node. After decoding the current node, go to the next node and repeat the above process; it should be noted that the pseudocode of the decoder is as follows:

[0077]

[0078] In the above pseudocode, V represents the node feature, E represents the edge feature, i represents the current node of the graph, j represents the jth neighbor node of the current node (j < k), MLP is a linear network layer, FFD is a feed-forward network layer, Attention is the attention calculation, Norm is the normalization operation, SUM is the summation operation, S represents the sequence information of the node of the residue, and the value of mask represents whether the current node is covered.

[0079] Configure network parameters: The dropout rate of the network is set to , and the label smoothing rate is , the batch size is 2500 sequence lengths, and the number of neighbor nodes K is taken as , and the loss function is defined as the cross-entropy between the generated sequence and the native sequence.

[0080] S9: Input the embedded expression of the node feature, the embedded expression of the edge feature, and the encoded amino acid sequence into the trained iterative computational graph neural network. Adopting the sequential decoding technique, the graph update adopts the way of iteratively calculating the attention between the edge feature and the node feature to obtain the protein sequence.

[0081] In a feasible embodiment, the sequential decoding method is the sequential decoding technique from end to end, that is, decoding sequentially from beginning to end. During the decoding process, decoding the current node requires the sequence context information of this node. The area in a sequence that does not need to be decoded (the segment that does not need to be designed) is called the fixed area. The process is as follows:

[0082] As Figure 4 shown, the first row in the figure represents the decoding order, the dotted-filled area in the second row represents the fixed area, the blue area represents the current node being decoded, the white area represents the area used when decoding the node, and the slanted area represents the area not used during decoding;

[0083] Sequential decoding means that the sequence information before the current node can be understood by the decoder, while the sequence information after the current node cannot be understood by the decoder.

[0084] The decoder can only understand the structural information after the current node.

[0085] Example 2

[0086] This embodiment compares the proposed iterative computational graph neural network model with the following methods. All comparative experiments were conducted on the Model dataset downloaded from the PDB database. final In the process, Model final The protein sequences in the dataset are all less than 900.

[0087] Model final The dataset is as follows:

[0088]

[0089] 1. Pearson

[0090] This embodiment selects Model final Pearson correlation coefficients were calculated for 100 target structures in the test set to compare the similarity between the designed sequences and the natural sequences:

[0091]

[0092] in: and These are the first and second sequences in the designed sequence and the natural sequence, respectively. The frequency of occurrence of each amino acid. and These are the average frequencies of amino acids in the designed and natural sequences, respectively. It is the number of amino acid types (for protein sequences, =20) The Pearson correlation coefficient between the designed sequence and the natural sequence in terms of amino acid composition is 0.961, indicating a high degree of similarity in their amino acid composition. This high similarity suggests that the designed sequence may possess certain optimized or improved characteristics while maintaining the functionality of the natural sequence.

[0093] The table below compares the design accuracy of other models on the test set of the Modelfinal dataset:

[0094]

[0095] As you can see, our method outperforms other mainstream methods in Model. final The design tasks in the test set have achieved the desired results.

[0096] In summary, this invention provides a protein sequence design method based on an iterative computational graph neural network. It calculates distance features in edge features and angle features in node features. Distance features directly reflect the spatial relationships between atoms in the protein sequence, while angle features reveal the stability and directionality of chemical bonds. Calculating these two features accurately captures high-order relevant information in the complex structure of proteins. By integrating these two features, a deeper understanding of the relationship between the internal structure and function of proteins can be achieved, leading to the design of more accurate and efficient proteins. Furthermore, by adding noise at the end of the main chain, the network can cope with uncertainties and interference during the design process, while also simulating small changes in protein structure in real-world environments. This improves the flexibility and robustness of the iterative graph neural network. In terms of algorithm design, iterative computation is employed. Through multiple iterations, the model can continuously adjust and optimize design parameters to approximate the ideal protein structure.

[0097] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A protein sequence design method based on iterative computational graph neural networks, characterized in that, The method includes the following steps: Obtain a protein sequence dataset and divide the protein sequence dataset into a training set, a validation set, and a test set; Extract the main chain atomic coordinates of the protein sequence and add noise at the end of the main chain atomic coordinates; Virtual atoms are designed and their coordinates are obtained based on the main chain atom coordinates. The number of virtual atom coordinates is the same as the number of main chain atom coordinates, and the specific method for obtaining the virtual atom coordinates is as follows: in, , , All are user-defined constants with an absolute value less than 1. , , Main chain atoms; Establish the distance features between the main chain atom coordinates and the virtual atoms, and input the distance features into the Gaussian radial function to obtain the distance characterization; Construct the angular features between the atoms in the main chain; The embedding representation of node features is obtained based on the angle features, and the embedding representation of edge features is obtained based on the distance representation; The standard amino acid types that constitute the protein sequence and unknown amino acids are encoded to obtain the encoded amino acid sequence. The iterative computation graph neural network constructed using the training set is trained iteratively. Each time it is trained, the generation accuracy of the trained iterative computation graph neural network is evaluated using the validation set. The iterative computation graph neural network with the highest accuracy is the trained iterative computation graph neural network. The trained iterative computation graph neural network is tested using the test set. The embedded expressions of the node features, the embedded expressions of the edge features, and the encoded amino acid sequence are input into the trained iterative computational graph neural network. Sequential decoding is used, and graph updates are performed by iteratively calculating the attention between edge features and node features to obtain the protein sequence.

2. The protein sequence design method based on iterative computational graph neural networks according to claim 1, characterized in that, The process of obtaining the protein sequence dataset, which involves dividing the dataset into a training set, a test set, and a validation set, specifically includes: A first protein sequence dataset is obtained from a protein database according to a first screening condition, wherein the first screening condition is that the protein sequence length is less than 900. The protein sequence dataset is obtained from the first protein sequence dataset according to the second screening condition, wherein the second screening condition is to delete protein sequences with a similarity higher than 40%. The protein sequence dataset was divided into a training set, a validation set, and a test set in a 9:1:1 ratio.

3. The protein sequence design method based on iterative computational graph neural networks according to claim 1, characterized in that, Noise is added to the main chain atomic coordinates, specifically in the following way: in, The set of main chain atom coordinates; This is a custom value used to adjust the noise intensity. It follows a normal distribution.

4. The protein sequence design method based on iterative computational graph neural networks according to claim 1, characterized in that, Establish the distance features between the main chain atom coordinates and the virtual atoms, and input the distance features into a Gaussian radial function to obtain the distance representation, specifically including: Construct the distance features between the main chain atom coordinates and the virtual atoms; The distance feature is input into a Gaussian radial function, and by... Uniformly sample 16 different types of distance intervals Sixteen different distance representations were obtained for the same distance, and the specific calculation formulas are as follows: in, It is a radial basis function. The base of the natural index, The input distance feature, for Equal-interval sampling, with an interval of [missing information]. .

5. The protein sequence design method based on iterative computational graph neural networks according to claim 1, characterized in that, The angular features between the main chain atoms are constructed, specifically including a first angular feature corresponding to the dihedral angle of the main chain atoms and a second angular feature corresponding to the torsion angle. The dihedral angles include three types, denoted as follows: , and The torsion angle includes three types, namely: , and , Main chain The dihedral angle formed Main chain The dihedral angle formed Main chain The dihedral angle formed In the main chain The resulting angle of twist In the main chain The resulting angle of twist It represents the rotation angle on the backbone of two adjacent peptide bonds.

6. The protein sequence design method based on iterative computational graph neural networks according to claim 1, characterized in that, The distance representation is input into a linear layer to obtain the embedded representation of the edge features, and the angle features are input into a linear layer to obtain the embedded representation of the node features.

7. The protein sequence design method based on iterative computational graph neural networks according to claim 1, characterized in that, Iterative computational graph neural networks consist of two parts: an encoder and a decoder, specifically: The encoder is used to update the embedding representation of the edge features and the embedding representation of the node features. Taking the current residue as an example, the encoder calculates the attention of the node and the information of the K neighboring nodes, gives higher weight to the nodes with deeper connections, and then uses the attention to update the information of the current node. This process is repeated three times, that is, iteratively calculating the attention to update the embedding representation of the edge features and the embedding representation of the node features. The decoder is used to decode information and iteratively generate residues. Taking the current node as an example, the decoder uses a multilayer perceptron to update the edge features, then calculates the attention of the node and the K neighboring nodes, and finally gathers the edge feature information of the residues in the K neighboring nodes. The information of the current node is updated through this information, and the residue probability information of the current node is calculated. Based on the probability information, the amino acid with the highest probability is selected as the residue of the node. After decoding the current node, the process is repeated for the next node.

8. The protein sequence design method based on iterative computational graph neural networks according to claim 7, characterized in that, Also includes: Configure network parameters; set the network drop rate to... Label smoothing rate The batch size is 2500 sequences in length, and the number of neighboring nodes K is set to a certain value. The loss function is defined as the cross-entropy between the generated sequence and the natural sequence.

9. A protein sequence design method based on an iterative computational graph neural network according to any one of claims 1 to 8, characterized in that, The standard amino acid types and unknown amino acids are coded as follows: Alanine A:0, Cysteine ​​C:1, Aspartic acid D:2, Glutamic acid E:3, Phenylalanine F:4, Glycine G:5, Histidine H:6, Isoleucine I:7, Lysine K:8, Leucine L:9, Methionine M:10, Asparagine N:11, Proline P:12, Glutamine Q:13, Arginine R:14, Serine S:15, Threonine T:16, Valine V:17, Tryptophan W:18, Tyrosine Y:19, Unknown amino acid X:20.

Images