Protein function prediction method based on virtual atoms and distance hierarchical learning
By employing a method based on virtual atoms and distance hierarchy learning, the low efficiency and accuracy issues of protein function prediction in traditional methods are addressed. This approach enables effective modeling of protein three-dimensional structure and non-covalent interactions, thereby improving prediction accuracy.
Patent Information
- Application Number
- CN202511067559.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-07
AI Technical Summary
Traditional biochemical experimental methods are time-consuming, costly, and inefficient. Existing deep learning methods fail to effectively consider the three-dimensional structure of proteins and non-covalent interactions, affecting the accuracy of protein function prediction.
We employ a method based on virtual atoms and distance-level learning. By constructing an adjacency matrix and relation set, we combine virtual atoms and real atoms for feature updates and use graph neural networks to predict protein function.
It improves the accuracy and robustness of protein function prediction and effectively models the three-dimensional spatial structure and non-covalent interactions of proteins.
Smart Images
Figure CN120913635A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of protein function prediction, and particularly relates to a protein function prediction method based on virtual atoms and distance hierarchical learning. BACKGROUND
[0002] With the deepening of life science research, protein function prediction, as one of the important problems in life science, has attracted a large number of researchers' attention. Protein function prediction has a crucial influence in the fields of disease mechanism elucidation and drug target discovery, therefore, predicting protein function through computer methods has become a key research direction in bioinformatics. Traditional methods of determining protein function through biochemical experiments usually have the limitations of long time consumption, low efficiency, high cost, high professional requirement for operators, etc. At the same time, existing protein function prediction methods combined with deep learning usually use sequence information and structure information of proteins. Sequence information is the arrangement order of residues or atoms of proteins, while structure information is the atomic coordinates of proteins in three-dimensional space and the interaction information between atoms or residues.
[0003] [1] Q. M. Yuan, J. J. Xie, J. C. Xie, et al. Fast and accurate protein function prediction from sequence through pretrained language model and homology-based label diffusion[J]. Briefings in Bioinformatics, 2023, 24(3): bbad117. This paper proposes a protein function prediction model named SPROF-GO based on sequence features only. The model first extracts the initial embedding representation of the sequence using a pre-trained protein language model, and then generates the feature representation of the protein for function prediction through a multi-layer perception (MLP) and a self-attention pooling module. In addition, in order to ensure that the prediction results are consistent with the structure in the gene ontology (GO) database, a hierarchical learning strategy is also adopted. Finally, a homology-based label diffusion algorithm is used to optimize the prediction results.
[0004] [2] V. Gligorijevic, T. Kosciolek, J. K. Leman, et al. Structure-based protein function prediction using graph convolutional networks[J]. Nature Communications, 2021, 12(1):3168. This paper proposes a protein function prediction model named DeepFRI. The model is divided into two stages. In the first stage, a pre-trained long short-term memory (LSTM) model is used to extract residue-level features from protein sequences; in the second stage, the residue features obtained in the first stage are used as input together with the adjacency matrix, and a graph convolutional network (GCN) is used to propagate and integrate these features.
[0005] Defects of the prior art:
[0006] 1) Traditional biochemical experimental methods for determining protein function usually require professional personnel with relevant biochemical knowledge to operate and are high in cost, low in efficiency, time-consuming, and difficult to perform multiple experiments in a short period of time.
[0007] 2) The function of a protein is not only determined by the residue sequence, but also influenced by non-covalent bonds (such as hydrogen bonds, van der Waals forces, salt bridges, etc.) in its three-dimensional spatial layout. These interactions are crucial for the stability and function of the protein, and methods that rely solely on sequence feature input may overlook the three-dimensional structural characteristics and key non-covalent interactions of the protein.
[0008] 3) Methods based on multi-modal fusion (i.e., using both sequence information and structure information) usually extract information from different modalities through independent neural networks and simply fuse them, without fully considering the mutual relationship between protein sequence and structure and without effectively handling the characteristics of long-distance non-covalent interactions in proteins, especially the influence of non-covalent bonds with different spatial distances on the three-dimensional structure of the protein, limiting the predictive ability of the model.
[0009] Therefore, it is necessary to develop a protein function prediction method based on virtual atoms and distance hierarchical learning to solve the above problems. SUMMARY
[0010] The purpose of the present application is to design a protein function prediction method based on virtual atoms and distance hierarchical learning to solve the above problems.
[0011] The present application achieves the above-mentioned purpose through the following technical solutions:
[0012] A protein function prediction method based on virtual atoms and distance hierarchical learning, comprising the steps of:
[0013] S1, data preprocessing;
[0014] The original data is the SMILES formula of the protein;
[0015] First, the open source toolkit RDKit based on python is used to obtain the atomic sequence of the protein according to the SMILES formula, which is represented as Wherein, n represents the number of atoms, and the three-dimensional coordinates of each atom The Euclidean distance between each pair of atoms is calculated according to the three-dimensional coordinates of the atoms To obtain the distance matrix :
[0016] (1)
[0017] Then, 4Å is taken as the threshold to construct the adjacency matrix :
[0018] (2)
[0019] The atomic clusters are divided in units of residues Wherein, m represents the number of residues in the protein, and each atomic cluster Contains all atoms constituting the residue; then a virtual atom is added to each atomic cluster A virtual atom is added to the whole protein And added to the atomic sequence;
[0020] The atoms with subscript i from 1 to n are called real atoms, and the atoms with subscript i from n+1 to n+m+1 are called virtual atoms, wherein To Respectively represent the virtual atoms corresponding to the residues, And the virtual atom corresponding to the whole protein is represented as Therefore, the atomic sequence after addition Contains n+m+1 real atoms and virtual atoms;
[0021] S2, virtual atom initialization;
[0022] The non-weighted centroid coordinates of the atoms in each cluster Are calculated as the initial coordinates of the virtual atom :
[0023] (3)
[0024] wherein , , are respectively three-dimensional coordinates of represents the number of atoms in
[0025] For the virtual atom , the non-weighted centroid coordinates of all real atoms in the sequence are calculated as its initial coordinates:
[0026] (4)
[0027] All virtual atoms are added to the adjacency matrix, and for the virtual atom corresponding to the cluster, it is only connected to all atoms in the cluster:
[0028] (5)
[0029] For the virtual atom corresponding to the protein, it is connected to all other atoms; at this time the adjacency matrix ;
[0030] Then according to formula (1) the distances between all virtual atoms and all atoms in the sequence are calculated and added to the distance matrix R, at this time ;
[0031] The name of the virtual atom is identified by "$";
[0032] S3, training of a protein function prediction model based on virtual atoms and distance hierarchy learning; comprising:
[0033] S31, defining a relationship set;
[0034] S32, all atoms passing through an embedding layer to obtain initial hidden states;
[0035] S33, sampling a uniform distribution as a weight probability coefficient for relationship aggregation;
[0036] S34, according to the weight coefficient, using a message aggregation network based on the relationship set to perform K rounds of feature updates for real atoms and virtual atoms in turn;
[0037] S35, taking the virtual atom representing the protein as the center of the sphere to construct a plurality of concentric spheres with equal distance thresholds;
[0038] S36, performing weighted sum pooling and normalization for each atom in the concentric sphere to obtain a protein representation;
[0039] S37, inputting the protein representation into a two-layer feedforward network and a softmax layer for classification output;
[0040] S38, using multi-class cross-entropy as loss for reverse gradient update.
[0041] Step S31 specifically includes: defining a set of relationships P, each relationship p r contains a pair of atoms (c r1 , c r2 ), a distance range threshold s r , and a weight matrix W r to be learned, i.e. p r = {(c r1 , c r2 ), s r , W r}; each relationship p r describes a possible non-covalent bond.
[0042] Before step S31, step S30 is further included, which is to construct initial features of atoms, specifically including: using RDKit to obtain feature information of each real atom, including atom name, charge, degree, whether it is an aromatic atom, and whether it is on a ring, wherein the atom name, charge and degree features are encoded by one-hot, and the features of whether it is an aromatic atom and whether it is on a ring are encoded by 0-1; the atom name is represented by one-hot encoding of all atoms in the data set except hydrogen atom, and all atoms include “$”; the charge is represented by one-hot encoding of integers from -3 to 3, and values exceeding the domain take the nearest boundary value; the degree is represented by one-hot encoding of integers from 0 to 6; whether it is an aromatic atom is represented by 0 or 1; whether it is on a ring is represented by 0 or 1; for a virtual atom, except that the name is encoded by one-hot, the remaining features are padded with 0; the encoded features of the atoms are spliced to obtain initial features .
[0043] Steps S32-S34 specifically include:
[0044] First, use a two-layer fully connected layer as an embedding layer to embed the initial features of each atom into 256 dimensions to obtain an initial hidden state :
[0045] (6)
[0046] Unlike typical graph neural networks that use the same matrix for aggregation and updating node features during iterative processes, message aggregation networks based on relation sets obtain weight matrices for updating during a single iteration of message aggregation. Furthermore, since atomic pairs satisfying the relation set definition do not necessarily represent the relation, randomness needs to be introduced. Specifically, if the atomic pairs to be aggregated... and If a condition is satisfied by a certain relation in the relation set, then a random variable is sampled. Let t represent a uniform distribution between 0 and 1, and t be the probability coefficient of the weight matrix; if the atomic pairs to be aggregated do not satisfy any relational condition, then let the random variable t=1, that is:
[0047] (7)
[0048] Each round is divided into two phases of updates; in the first... In round-message aggregation, the first stage updates the characteristics of the real atoms:
[0049] (8)
[0050] Where the superscript k represents the k-th round of aggregation, This represents the hidden feature obtained by atom i after message aggregation in the k-th round. MLP stands for Multilayer Perceptron. These are learnable parameters. express The set of neighbors, if and Satisfy the definition of the r-th relation p in relation set P r Then W represents the corresponding weight matrix to be learned. r Otherwise W=W * W * express and The weight matrix to be learned when it does not satisfy any relation in the relation set P; in this stage... Only real atoms;
[0051] Then, in the second stage, the characteristics of the virtual atoms are updated according to formula (8), but in this stage... Virtual atoms only;
[0052] Then, K rounds of iterative updates are performed to obtain the final feature sequence of all atoms. K is a hyperparameter.
[0053] Steps S35-S37 specifically include:
[0054] Virtual atoms corresponding to proteins The coordinates of the center of the sphere, construct a spherical coordinate system, let alpha as the distance, construct A concentric sphere radius of an equal distance threshold At the same time, a weight is constructed for each distance threshold , weighted sum pooling is performed according to the weight to obtain the representation pro of the protein:
[0055] (9)
[0056] Wherein, represents the feature of the real atom whose distance with the virtual atom representing the protein is in the interval , norm represents the normalization of the features of all atoms in each distance threshold range, represents the distance between the center of the sphere and the atom farthest from the center of the sphere, that is , alpha is a hyperparameter, represents the upward rounding; The protein feature representation obtained by pooling is input into a two-layer feedforward network and a softmax layer to obtain the final output result
[0057] :
[0058] (10).
[0059] Step S38 specifically includes:
[0060] Since it is a multi-classification task, multi-classification cross entropy is used as the loss calculation, and the calculation formula is as follows:
[0061] (11)
[0062] Wherein, C represents the number of categories, represents the true category of the predicted sample, when the sample belongs to the i category, , otherwise , represents the probability that the predicted sample belongs to the i category.
[0063] The beneficial effects of the present application are:
[0064] The present application solves the problem that the traditional message aggregation network ignores the spatial structure in the protein modeling process, and constructs a relationship set, models the non-covalent bond interaction in the protein, and strengthens the perception of the network to the spatial structure. Compared with the current method of predicting protein function through a graph neural network, the accuracy of protein function prediction is effectively improved. BRIEF DESCRIPTION OF DRAWINGS
[0065] Figure 1 The flowchart of the present application is shown in the figure.
[0066] Figure 2 A flowchart of a training process of a protein function prediction model based on virtual atoms and distance hierarchy learning in the present application. DETAILED DESCRIPTION
[0067] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.
[0068] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without making creative efforts fall within the scope of protection of the present application.
[0069] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.
[0070] In the description of the present application, it should be understood that the terms "upper", "lower", "inner", "outer", "left", "right", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship commonly placed when the product of the present application is used, or the orientation or positional relationship commonly understood by those skilled in the art, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.
[0071] In addition, the terms "first", "second", etc. are only used for differentiation in description, and cannot be understood as indicating or implying relative importance.
[0072] In the description of the present application, it should also be noted that, unless otherwise explicitly specified and limited, the terms "arranged", "connected", etc. should be understood broadly, for example, "connected" can be fixedly connected, or detachably connected, or integrally connected; can be mechanically connected, or electrically connected; can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements. Those of ordinary skill in the art can understand the specific meaning of the above terms in the present application according to the specific circumstances.
[0073] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0074] like Figure 1 and 2 As shown, a protein function prediction method based on virtual atoms and distance-level learning is presented. This method introduces virtual atom representations and combines them with a spatial distance-level weighted learning mechanism to simultaneously model the protein's sequence information and three-dimensional spatial structure information, thereby improving the accuracy and robustness of protein prediction. The method includes the following steps:
[0075] S1, Data Preprocessing;
[0076] The original data is the SMILES (Simplified Molecular Input Line EntrySystem) formula for proteins; SMILES is a concise textual representation of chemical structures; it uses a series of characters to describe information about atoms and chemical bonds in a molecule, allowing complex chemical structures to be expressed using simple one-dimensional strings.
[0077] First, the open-source Python toolkit RDKit is used to obtain the atomic sequence of the protein according to the SMILES formula, represented as follows: Where n represents the number of atoms and the three-dimensional coordinates of each atom. ; Calculate each atom pair based on the three-dimensional coordinates of the atoms European distance To obtain the distance matrix :
[0078] (1)
[0079] Then, construct the adjacency matrix using 4Å as the threshold. :
[0080] (2)
[0081] Atomic clusters are divided into units based on residues. Where m represents the number of residues in the protein, and each atomic cluster It contains all the atoms that make up the residue; then for each cluster of atoms Each imagines an atom To create an imaginary atom for the protein as a whole. And add it to the atomic sequence;
[0082] Atoms with subscripts i from 1 to n Atoms with subscripts i from n+1 to n+m+1 are called real atoms, while those with subscripts i from n+1 to n+m+1 are called virtual atoms. to respectively represent the virtual atoms corresponding to the residue pairs, represent the virtual atom corresponding to the entire protein, then the added atom sequence contains n+m+1 real atoms and virtual atoms in total;
[0083] S2, virtual atom initialization;
[0084] Calculate the non-weighted centroid coordinates of the atoms in each cluster as the initial coordinates of the virtual atom :
[0085] (3)
[0086] wherein , , are the three-dimensional coordinates of , respectively, indicates the number of atoms in ;
[0087] For the virtual atom , calculate the non-weighted centroid coordinates of all real atoms in the sequence as its initial coordinates:
[0088] (4)
[0089] Add all virtual atoms to the adjacency matrix, for the virtual atom corresponding to the cluster, make it only connected to all atoms in the cluster:
[0090] (5)
[0091] For the virtual atom corresponding to the protein, make it connected to all other atoms; at this time, the adjacency matrix ;
[0092] Then calculate the distances between all virtual atoms and all atoms in the atom sequence according to formula (1) and add them to the distance matrix R, at this time ;
[0093] The name of the virtual atom is identified by "$";
[0094] S3, training of the protein function prediction model based on virtual atoms and distance hierarchical learning; including:
[0095] S31, defining a relationship set;
[0096] S32, all atoms pass through the embedding layer to obtain the initial hidden state;
[0097] S33, sampling a uniform distribution as a weight probability coefficient of relationship aggregation;
[0098] S34, according to the weight coefficient, performing K rounds of feature update of real atoms and virtual atoms in the message aggregation network based on the relationship set;
[0099] S35, constructing a plurality of concentric spheres with equal distance thresholds, with the virtual atom representing the protein as the sphere center;
[0100] S36, performing weighted sum pooling and normalization for atoms in each concentric sphere to obtain a protein representation;
[0101] S37, inputting the protein representation into a two-layer feedforward network and a one-layer softmax layer for classification output;
[0102] S38, using multi-class cross-entropy as loss for backward gradient update.
[0103] Step S30 is to construct the initial features of atoms, specifically including: using RDKit to obtain the feature information of each real atom, including atom name, charge, degree, whether it is an aromatic atom, and whether it is on a ring, wherein the atom name, charge and degree features are encoded by one-hot, and the features of whether it is an aromatic atom and whether it is on a ring are encoded by 0-1. The atom name is represented by one-hot encoding of all atoms in the data set except hydrogen atoms, and all atoms include “$”. The charge is represented by one-hot encoding of integers from -3 to 3, and values exceeding the domain take the nearest boundary value. The degree is represented by one-hot encoding of integers from 0 to 6. Whether it is an aromatic atom is represented by 0 or 1. Whether it is on a ring is represented by 0 or 1, as shown in Table 1.
[0104] Table 1. Atom feature table
[0105]
[0106] For the virtual atom, except that the name is encoded by one-hot, the remaining features are padded with 0. The encoded features of the atoms are spliced to obtain the initial features .
[0107] Step S31 specifically includes: in order to introduce prior knowledge, so that the atoms can aggregate the information of surrounding atoms according to the possible existence of non-covalent bonds, rather than using the same aggregation method for all neighbor atoms. Therefore, a relationship set P is defined, each relationship p r contains atoms (c r1 , c r2 ), distance range threshold s r , and a weight matrix W rThat is, p r = {(c r1 , c r2 ), s r W r}; Each relation p r This describes a possible non-covalent bond, such as defining relation one: ={(N, O), [1.5, 2.2], }, where N represents a nitrogen atom, O represents an oxygen atom, and [1.5, 2.2] represents , This represents the weight matrix corresponding to this relation. This can be understood as hydrogen bonds between nitrogen and oxygen atoms.
[0108] Steps S32-S34 specifically include:
[0109] First, two fully connected layers (FCNN) are used as embedding layers to embed each atom. initial features Embedded into 256 dimensions, the initial hidden state is obtained. :
[0110] (6)
[0111] Unlike typical graph neural networks that use the same matrix for aggregation and updating node features during iterative processes, message aggregation networks based on relation sets obtain weight matrices for updating during a single iteration of message aggregation. Furthermore, since atomic pairs satisfying the relation set definition do not necessarily represent the relation, randomness needs to be introduced. Specifically, if the atomic pairs to be aggregated... and If a condition is satisfied by a certain relation in the relation set, then a random variable is sampled. Let t represent a uniform distribution between 0 and 1, and t be the probability coefficient of the weight matrix; if the atomic pairs to be aggregated do not satisfy the conditions of any relation (i.e., no predefined relation exists), then let the random variable t=1, that is:
[0112] (7)
[0113] Each round is divided into two phases of updates; in the first... In round-message aggregation, the first stage updates the characteristics of the real atoms:
[0114] (8)
[0115] Where the superscript k represents the k-th round of aggregation, denotes the hidden feature of atom i after the k-th round of message aggregation, and MLP is a multi-layer perceptron, is a learnable parameter, denotes the neighbor set of atom i, if and satisfy the r-th relation p in the relation set P r , W denotes the corresponding weight matrix W r to be learned, otherwise W = W * , W * denotes and the weight matrix to be learned when any relation in the relation set P is not satisfied; in this stage, only real atoms;
[0116] Then the second stage updates the features of virtual atoms according to formula (8), but in this stage only virtual atoms;
[0117] Then K rounds of iteration are performed to obtain the final feature sequence of all atoms , K is a hyperparameter.
[0118] Steps S35-S37 specifically include:
[0119] A spherical coordinate system is constructed with the coordinates of the virtual atom corresponding to the protein as the center of the sphere, and let α Å be the interval, and construct a set of concentric spherical radii of equal distance thresholds , and construct a weight for each distance threshold, and perform weighted sum pooling according to the weight to obtain the representation pro of the protein:
[0120] (9)
[0121] wherein denotes the feature of a real atom whose distance to the virtual atom representing the protein is within the interval , norm denotes normalization of the features of all atoms within each distance threshold range, denotes the distance between the center of the sphere and the atom farthest from the center of the sphere, i.e. , and α is a hyperparameter, denotes the ceiling function;
[0122] The protein feature representation obtained by pooling is input into a two-layer feedforward network (FFN) and a one-layer softmax layer to obtain the final output result :
[0123] (10).
[0124] Step S38 specifically comprises:
[0125] Since it is a multi-classification task, multi-classification cross-entropy is used as the loss calculation, and the calculation formula is as follows:
[0126] (11)
[0127] Wherein, C represents the number of categories, represents the true category of the predicted sample, when the sample belongs to the i category, otherwise , represents the probability that the predicted sample belongs to the i category.
[0128] The application proposes a method for representing the overall information of amino acids and proteins by virtual atoms, solving the problem of insufficient feature representation of atomic clusters when only a single real atom is used as the only feature unit in the neural network in the traditional way
[0129] The application proposes a message aggregation network based on a relationship set, which can effectively model the non-covalent bond relationship that may exist between atoms by means of biochemical prior knowledge and the introduction of probability sampling, improve the perception of the model to the bond information in the molecule, and better capture the interaction between atoms in the protein molecule. A distance threshold weighted pooling method is used to obtain the final representation of the protein, which utilizes the length characteristics of the bond and the distribution characteristics of the atoms in the three-dimensional space, and strengthens the perception of the model to the spatial structure of the protein.
[0130] The above is only the preferred embodiment of the application, and it should be pointed out that for ordinary skilled persons in the art, some improvements and refinements can be made without departing from the technical principles of the application, and these improvements and refinements should also be regarded as the protection scope of the application.
Claims
1. A protein function prediction method based on virtual atom and distance hierarchy learning, characterized in that, Comprising steps of: S1, data preprocessing; The original data is the SMILES formula of the protein; First, the open-source Python toolkit RDKit is used to obtain the atomic sequence of the protein according to the SMILES formula, represented as follows: Where n represents the number of atoms and the three-dimensional coordinates of each atom. ; Calculate each atom pair based on the three-dimensional coordinates of the atoms European distance To obtain the distance matrix : （1）， Again, 4A is taken as the threshold to construct the adjacency matrix : （2）， Divide the atom cluster by residue Where m represents the number of residues in the protein, each atom cluster Contains all atoms that make up the residue; for each atom cluster Fictionalize one atom Fictionalize one atom for the protein as a whole And add to the atom sequence; atoms with index i from 1 to n atoms with index i from n+1 to n+m+1 are called virtual atoms, wherein to denote the virtual atoms corresponding to the residue pair, denote the virtual atoms corresponding to the whole protein, then the added atom sequence contains n+m+1 real atoms and virtual atoms in total; S2, virtual atom initialization; calculating the non-weighted centroid coordinates of the atoms in each cluster as initial coordinates of the virtual atoms : （3）， wherein , , are respectively three-dimensional coordinates, denotes the number of atoms in For virtual atoms The non-weighted centroid coordinates of all real atoms in the sequence are then calculated as their initial coordinates: （4）， All virtual atoms are added to the adjacency matrix, and for the virtual atoms corresponding to the cluster, they are only connected to all atoms in the cluster: （5）， For the virtual atoms corresponding to the protein, make them connected to all other atoms; at this point the adjacency matrix ; According to formula (1), the distances between all virtual atoms and all atoms in the atomic sequence are calculated and added to the distance matrix R, at which time ; The name of the virtual atom is identified by "$"; S3, training of the protein function prediction model based on virtual atoms and distance hierarchy learning; comprising: S31, defining a relationship set; S32, all atoms pass through the embedding layer to obtain the initial hidden state; S33, sampling a uniform distribution as the weight probability coefficient of relationship aggregation; S34, according to the weight coefficient, the real atoms and the virtual atoms are updated in K rounds of feature update by using the message aggregation network based on the relationship set in turn; S35, a plurality of concentric spheres with equal distance thresholds are constructed with the virtual atoms representing the protein as the sphere centers; S36, weighted sum pooling is performed for the atoms in each concentric sphere, and normalization is performed to obtain the protein representation; S37, the protein representation is input into a two-layer feedforward network and a softmax layer for classification output; S38, multi-classification cross-entropy is used as the loss for reverse gradient update. 2.The protein function prediction method based on virtual atom and distance hierarchy learning according to claim 1, wherein, Step S31 specifically includes: defining a set of relations P, each relation p r contains a pair of atoms (c r1 , c r2 ), a distance range threshold s r , and a weight matrix W r to be learned, that is, p r = {(c r1 , c r2 ), s r , W r}; each relation p r describes a possible non-covalent bond. 3.The protein function prediction method based on virtual atom and distance hierarchy learning according to claim 2, characterized in that, Before step S31, step S30 of constructing initial features of atoms is further included, specifically comprising: obtaining feature information of each real atom using RDKit, including atom name, charge, degree, whether it is an aromatic atom, and whether it is on a ring, wherein the atom name, charge and degree features are encoded by one-hot, and the features of whether it is an aromatic atom and whether it is on a ring are encoded by 0-1; the atom name is represented by one-hot encoding of all atoms except hydrogen in the data set, and all atoms include "$"; the charge is represented by one-hot encoding of integers from-3 to 3, and values exceeding the domain take the nearest boundary value; the degree is represented by one-hot encoding of integers from 0 to 6; whether it is an aromatic atom is represented by 0 or 1; whether it is on a ring is represented by 0 or 1; for a virtual atom, except that the name is encoded by one-hot, the remaining features are padded by 0; and the encoded features of the atoms are spliced to obtain initial features . 4.The protein function prediction method based on virtual atom and distance hierarchy learning according to claim 3, wherein, Steps S32-S34 specifically comprise: First, each atom is embedded into 256 dimensions using two fully connected layers as embedding layers with initial features to get initial hidden states : （6）， Unlike the iteration process of general graph neural networks, which uses the same matrix to aggregate and update the node features, in the iteration process of a round of message aggregation of the message aggregation network based on the relationship set, the weight matrix is obtained by matching the relationships in the relationship set for updating. At the same time, since the atom pair satisfying the definition of the relationship set does not necessarily exist the relationship, randomness needs to be introduced. Specifically, if the atom pair to be aggregated and satisfies the condition of a certain relationship in the relationship set, a random variable representing a uniform distribution between 0 and 1 is sampled, and t is taken as the probability coefficient of the weight matrix; if the atom pair to be aggregated does not satisfy the condition of any relationship, the random variable t = 1, namely: （7）， Each round is divided into two stages of update; in the kth round of message aggregation, the first stage first updates the features of the real atoms: （8）， wherein the superscript k represents the kth round of aggregation, represents the hidden feature obtained by the atom i after the kth round of message aggregation, and MLP is a multi-layer perceptron, is a learnable parameter, represents a neighbor set of and satisfies the rth relationship p in the relationship set P r , W represents the corresponding to-be-learned weight matrix W r , otherwise W = W * , W * represents and do not satisfy any relationship in the relationship set P; in this stage, only real atoms; Then the second stage updates the features of the dummy atoms according to equation (8), but in this stage only for dummy atoms; Then K rounds of iterative updates are performed to obtain the final feature sequence of all atoms , where K is a hyperparameter.
5. The method of claim 4, wherein the method is characterized by, Steps S35-S37 specifically comprise: coordinates of the virtual atoms corresponding to the proteins a spherical coordinate system with the coordinates of the virtual atoms corresponding to the proteins as the center of the sphere, and let α Å as the interval, construct a concentric sphere radius of an equal distance threshold , and construct a weight for each distance threshold , and obtain the representation of the protein pro by weighted sum pooling according to the weight. （9）， wherein, represents a feature of a real atom whose distance to a virtual atom representing a protein is in the interval norm represents a normalization over all features of atoms within each distance threshold range, represents the distance between the center of the sphere and the atom farthest from the center of the sphere, i.e. , a is a hyperparameter, represents a ceiling function; The pooled protein feature representation is input to two layers of feed forward networks and a layer of softmax layer to obtain the final output result : （10）。 6.The method of claim 5, wherein the method further comprises: Step S38 specifically comprises: Since it is a multi-classification task, multi-classification cross-entropy is used as the loss calculation, and the calculation formula is as follows: （11）， where C denotes the number of classes, denotes the true class of the predicted sample, when the sample belongs to class i, , otherwise , denotes the probability that the predicted sample belongs to class i.