A method for predicting protein function based on transfer learning and three-channel combined GNN

Abstract

The present invention belongs to the field of deep learning and bioinformatics, and discloses a method for predicting protein function based on transfer learning and three-channel combined GNN. The method extracts the contact graph of amino acid residues with different connection densities by parsing the protein spatial structure file; constructs an amino acid sequence mixed feature extractor based on transfer learning to obtain the residue-level mixed feature map of the amino acid sequence; introduces a graph convolutional network and a GATv2 network with a multi-head graph attention mechanism to construct a three-channel combined graph neural network for protein function prediction; and obtains the probability value of each protein gene ontology function label through a multi-label classifier. The present invention improves the protein function prediction performance and facilitates researchers to efficiently and comprehensively obtain the complex association information of protein entries in different authoritative databases.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and bioinformatics, and in particular to a method for predicting protein function based on transfer learning and three-channel combined GNN. Background Art

[0002] Protein is essential for life activities and cell metabolism. Research on protein function will directly clarify various special mechanisms of protein in physiology or pathology, reveal the process of protein synthesis, degradation, processing and modification, and enable people to better understand gene expression and regulation. At present, protein function is mainly determined by enzyme activity assay, cell biology experiment and other technologies. Based on Gene Ontology (GO), protein function is divided into three categories: Molecular Function (MF); Biological Process (BP); Cellular Component (CC).

[0003] With the development of protein sequencing and structural analysis technologies such as X-ray single crystal diffraction and cryo-electron microscopy, as well as the application of protein structure prediction models, the number of known amino acid sequences and structures is very different from the corresponding functional annotations. In order to avoid the high equipment and time costs of directly determining protein functions through experiments, researchers are increasingly turning to computational methods for function prediction. The sequence and molecular structure characteristics of proteins determine their specific functions to a certain extent. By combining amino acid sequence or molecular structure information, designing efficient and accurate protein function prediction methods has become the focus of protein function research in recent years.

[0004] At first, researchers inferred the function of proteins by comparing sequence similarity and homology. That is, sequences with high homology are likely to have similar functions, and their known functional annotations can be transferred to similar proteins. Based on this theory, researchers have developed the Blast2GO tool to transfer and analyze functional annotations using sequence similarity. There are also methods to infer that proteins have similar functions by comparing structural similarities. Some literature has proposed calculating the similarity between protein structures through contact maps to predict protein functions.

[0005] Directly predicting protein functions based on sequence or structural similarity and homology is prone to errors. With the outstanding performance of deep learning technology in solving problems in multiple fields, people have begun to explore its application in protein function prediction. For example, the DeepGO model uses CNN to extract amino acid sequence features and combines them with the protein interaction network extracted from the STRING database to predict protein functions. The DeepGO model only performs feature modeling on sequences, ignoring the huge impact of the molecular structure of proteins on their specific functions. Based on this, some researchers proposed a model DeepFRI that combines molecular structure prediction functions. They sampled 10M sequences from the entire sequence collection of the Pfam database and trained a LSTM-based protein language model as a feature extractor for amino acid sequences. The residue-level feature matrix of the amino acid sequence and the residue adjacency matrix extracted from the protein spatial structure file were combined and input into the GCN graph convolutional neural network to predict protein function. However, LSTM has obvious defects in processing some long sequence text data, and there are limitations when using only GCN to process some complex irregular graph data. In the selection of experimental data, DeepFRI only uses high-precision protein structures in the PDB protein database and homology-based structural models constructed by SWISS-MODEL, but most proteins lack these data. At the same time, in actual operation, since each protein database is not directly related, key information such as protein sequence, molecular structure, functional annotation, etc. must be retrieved from different source databases through database cross-references, which greatly hinders researchers from obtaining complete and convenient protein-related information. In addition, some databases that provide downloads, such as the protein-ligand interaction database ChEMBL, are MySQL-type relational databases, which are lower in access convenience and retrieval efficiency than document-type NoSQL databases. Summary of the invention

[0006] In view of the above problems, the present invention proposes a method for predicting protein function based on transfer learning and three-channel combined GNN, which improves the prediction accuracy. At the same time, the present invention also provides a protein document-type association database, which is convenient for efficiently obtaining detailed association information of related proteins in protein databases with different focuses. Based on this, users can extract benchmark data sets for data-driven protein downstream prediction tasks from the protein document-type association database.

[0007] The technical solution adopted by the present invention is: a method for predicting protein function based on transfer learning and three-channel combined GNN, comprising the following steps:

[0008] Step S1, obtaining the amino acid sequence, protein spatial structure file, and protein Gene Ontology (GO) function label of the protein from the public protein database, cross-correlating and cleaning the obtained data, and then constructing a NoSQL type protein document-based association database, obtaining a benchmark data set for the protein function prediction task from the protein document-based association database, and dividing it into a training set, a validation set, and a test set;

[0009] Step S2, parsing the protein spatial structure file, extracting the spatial coordinates of the Ca atom in each amino acid residue in the protein molecule, and based on the spatial contact relationship of the Ca atoms in all residues, extracting the amino acid residue contact maps with different connection densities according to different contact thresholds, so as to map the complex spatial structure of the protein molecule;

[0010] Step S3, constructing an amino acid sequence hybrid feature extractor based on transfer learning, extracting multiple types of residue-level features from the amino acid sequence through the amino acid sequence hybrid feature extractor to form a primary embedding representation of the sequence; performing feature fusion and enrichment operations on the primary embedding representation of the sequence to obtain a residue-level hybrid feature map of the amino acid sequence;

[0011] Step S4, introducing a graph convolutional network and a GATv2 network with a multi-head graph attention mechanism to construct a three-channel combined graph neural network, using the residue-level mixed feature map of the amino acid sequence as the initial feature matrix input of the three-channel combined graph neural network, and the amino acid residue contact map will be used as the adjacency matrix input of different feature channels to capture and learn the complex graph relationship between amino acid residues in protein molecules, and combine the residue-level mixed feature map of the amino acid sequence and the multi-level structural features to predict protein function;

[0012] Step S5, construct a multi-label classifier, perform global pooling and aggregation operations on the output node features of the three-channel combined graph neural network, map the aggregated features to a 0-1 probability space, and obtain the probability value of each protein gene ontology function label;

[0013] Step S6, a complete protein function prediction model is formed by a sequence hybrid feature extractor, a three-channel combined graph neural network, and a multi-label classifier. The protein function prediction model is pre-trained on a training set, and the performance of the protein function prediction model at different training stages is evaluated using a validation set and a test set. The protein function prediction model that has passed the training is used to perform protein function prediction.

[0014] Further preferably, the spatial contact relationship of the Ca atoms in the residues in step S2 is: a protein is composed of N amino acid residues, the protein spatial structure file is traversed, the spatial structure coordinates of the Ca atoms in each amino acid residue are extracted, the spatial distances between all Ca atoms are calculated based on the coordinates, and a distance threshold is set to determine whether two residues are in contact, thereby obtaining an N*N amino acid residue contact matrix, and then only the edges r with contact relationships are taken, and the distance matrix is ​​converted into a 2*2r adjacency matrix according to the source node and target node of the edge r, which is used as the amino acid residue contact map.

[0015] Further preferably, the amino acid residue contact maps of different connection densities in step S2 are: respectively calculating the amino acid residue contact maps of the protein spatial structure file actually determined by PDB and the protein spatial structure file predicted by AlphaFold, and calculating the amino acid residue contact maps representing different contact densities according to different distance threshold sizes.

[0016] Further preferably, in step S3, the complete sequence of the protein or a subsequence is input into the amino acid sequence hybrid feature extractor, the sequence embedding representation extracted by the Bi-LSTM network is first, and then the ESM2 network composed of a 33-layer Transformer encoder continues to extract the sequence embedding representation, and the sequence embedding representations extracted by the Bi-LSTM network and the ESM2 network are jointly used as the primary sequence embedding representation of the amino acid sequence.

[0017] Further preferably, the feature fusion and enrichment operations on the primary embedding representation of the sequence in step S3 are: slicing the primary embedding representation of the sequence into: sequence one-hot encoding mapping, residue sequence deep contextual features and residue biochemical property encoding according to categories, performing learnable nonlinear mapping on the slices, and performing feature enrichment and dimensionality reduction on the primary embedding representation of the sequence by adjusting the dimensionality of each hidden layer. Finally, through ReLU nonlinear mapping and feature fusion operations, a residue-level mixed feature map of the amino acid sequence is obtained.

[0018] Further preferably, the three-channel combined graph neural network consists of a first feature channel composed of a two-layer GCN network in series; another identical two-layer GCN network in series constitutes a second feature channel, and the first feature channel is parallel to the second feature channel to form a 2*2 parallel dual-channel GCN network; a GATv2 network with a multi-head graph attention mechanism is introduced to form the third feature channel, and the GATv2 network is connected in series to the parallel dual-channel GCN network.

[0019] Further preferably, the feature aggregation and propagation method of the three-channel combined graph neural network is: using the residue-level mixed feature map of the amino acid sequence as the initial feature matrix input of the 2*2 parallel dual-channel GCN network, the sparse amino acid residue contact map extracted with a smaller contact threshold is used as the adjacency matrix input of one of the feature channels of the parallel network, and the dense amino acid residue contact map extracted with a larger contact threshold is used as the adjacency matrix input of another feature channel; the output features of the 2*2 parallel dual-channel GCN network are feature fused to obtain fused features, which are used as the feature matrix input of the GATv2 network, and the sparse amino acid residue contact map is used as the adjacency matrix input of the GATv2 network; the GATv2 network obtains the output node features of the three-channel combined graph neural network after further noise filtering and feature extraction of the fused features.

[0020] Further preferably, in step S1, a NoSQL type protein document-based relational database is constructed based on MongoDB.

[0021] Further preferably, the steps of constructing a NoSQL type protein document-based relational database are as follows:

[0022] Step S101, use the interface test script and the test tool PostMan to test the open data interface of the Uniprot database, obtain the Uniprot database identifier of the corresponding protein entry through the screening conditions: UPID, the screened protein entry must have the protein spatial structure file determined experimentally or the protein spatial structure file predicted by the AlphaFold model, select the protein gene ontology as the functional label of the protein, the total number of protein gene ontology labels possessed by each screened protein entry should be greater than or equal to 3, and the screened Uniprot database identifier is used as the source data ID of the database;

[0023] Step S102, accessing the Uniprot database data interface through the Uniprot database identifier obtained by screening, obtaining the cross-index ID of the corresponding protein entry in the PDB and AlphaFold databases from the data packet returned by the interface, and at the same time associating the primary key target ID of the target protein information table of the ChEMBL database through the Uniprot database identifier, and obtaining the remaining secondary table IDs in the ChEMBL database through this association;

[0024] Step S103, classify and request the data interface corresponding to each type of ID, parse the XML and JSON data packets returned by the interface through BeautifulSoup, Jsonpath framework and regular expression, and obtain a JSON dictionary consisting of protein association information in the three databases of Uniprot, PDB and AlphaFold, and a JSON dictionary consisting of target protein association data in the ChEMBL database;

[0025] Step S104: Based on the JSON dictionary, a protein information master table is constructed using the Pymongo framework with the Uniprot database identifier as the primary key, and a ChEMBL associated information table containing target information is constructed with the primary key target ID as the primary key;

[0026] Step S105, cross-correlate the primary key target ID in the ChEMBL association information table with the Uniprot database identifier in the protein information main table to obtain the final protein document-type association database. Users can retrieve key information of related protein entries in Uniprot, PDB, AlphaFold, and ChEMBL databases only through the Uniprot database identifier.

[0027] The beneficial effects of the present invention are as follows:

[0028] (1) A method for building a NoSQL protein document-based relational database based on MongoDB is proposed. Through the NoSQL protein document-based relational database, users can retrieve detailed related data of related protein entries based on a primary key target ID, without having to cross-index in multiple databases, which greatly improves the retrieval efficiency of protein-related information. Compared with the complete MySQL-type public database that is directly downloaded, the NoSQL database has a huge improvement in access convenience.

[0029] (2) Most of the current protein function prediction models, including DeepFRI, only use high-precision protein structures in the PDB protein database, and homology-based structural models constructed by SWISS-MODEL. However, most proteins lack corresponding data. To solve this problem, the experimental dataset obtained from the document-based NoSQL protein association database can simultaneously have high-precision PDB structural data determined by experiments and structural data predicted by the AlphaFold model, which can reduce the cost of data acquisition and improve the generalization performance of the model.

[0030] (3) The amino acid sequence features extracted directly through sequence one-hot encoding mapping or LSTM model lack residue deep context information. The present invention constructs an amino acid sequence hybrid feature extractor based on transfer learning. The transfer learning is based on two types of protein language models, Bi-LSTM and Transformer, combined with sequence one-hot encoding mapping, residue sequence deep context features and residue biochemical property encoding, to jointly constitute a primary sequence embedding representation of the amino acid sequence. During the model training process, a learnable nonlinear mapping is performed on the primary sequence embedding representation, which reduces the experimental overhead while further enriching the residue features. The feature expression ability of the residue-level hybrid feature mapping obtained in this way will be more powerful.

[0031] (4) The DeepFRI model architecture is composed of a series of GCN networks, which has certain limitations when processing complex irregular graphs such as protein spatial structures. The present invention introduces the GATv2 network with a multi-head attention mechanism, which allows the protein function prediction model to learn multiple sets of attention weights at the same time. Each attention head can learn different attention weights, thereby enhancing the protein function prediction model's ability to express complex graph structures.

[0032] (5) A single feature channel cannot efficiently learn multiple sets of structural characteristics of protein molecules. The present invention combines the GCN network and the GATv2 network with a multi-head graph attention mechanism to form a three-channel combined graph neural network architecture. Through a 2*2 parallel GCN network, two types of residue contact graphs extracted by different contact thresholds are simultaneously learned. Subsequently, the features output by the dual-channel parallel network are further filtered through the series-connected GATv2 network. This architecture can combine the deep-level sequence and structural characteristics of protein molecules to better capture the complex relationship between residues in the protein structure. At the same time, in the final multi-label classification stage of the model, the present invention enables the protein function prediction model to further obtain more node features by fusing two types of global pooling information. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 This is a schematic diagram of the principle structure of the present invention.

[0035] Figure 2 This is a diagram of the architecture of the amino acid sequence mixed feature extractor in the present invention.

[0036] Figure 3This is a diagram of the three-channel combination graph neural network architecture in the present invention.

[0037] Figure 4 This is a diagram of the multi-label classifier architecture in the present invention. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0039] like Figure 1 As shown, the method for predicting protein function based on transfer learning and three-channel combined GNN includes the following steps:

[0040] Step S1, obtain the amino acid sequence, protein spatial structure file, and protein Gene Ontology (GO) functional label of the protein from the public protein database, cross-correlate and clean the obtained data, and then build a NoSQL type protein document-type association database based on MongoDB, obtain the benchmark data set for the protein function prediction task from the protein document-type association database, and divide it into training set, validation set, and test set; the training set is used to train the protein function prediction model proposed in the present invention; the validation set is used to evaluate the specific performance of each epoch during the model training process; the test set is used to test the overall performance of the model after the training is completed. Among them, the public protein database includes four open source databases: Uniprot, PDB, AlphaFold, and ChEMBL; the protein spatial structure file includes the protein spatial structure file determined experimentally and predicted by the AlphaFold model;

[0041] Step S2, parsing the protein spatial structure file, extracting the spatial coordinates of the Ca atom in each amino acid residue in the protein molecule, and based on the spatial contact relationship of the Ca atoms in all residues, extracting the amino acid residue contact maps with different connection densities according to different contact thresholds, so as to map the complex spatial structure of the protein molecule;

[0042] Step S3, constructing an amino acid sequence hybrid feature extractor based on transfer learning, extracting multiple types of residue-level features from the amino acid sequence through the amino acid sequence hybrid feature extractor to form a primary embedding representation of the sequence; performing feature fusion and enrichment operations on the primary embedding representation of the sequence to obtain a residue-level hybrid feature map of the amino acid sequence;

[0043] Step S4, introducing a graph convolutional network and a GATv2 network with a multi-head graph attention mechanism to construct a three-channel combined graph neural network, using the residue-level mixed feature map of the amino acid sequence as the initial feature matrix input of the three-channel combined graph neural network, and the amino acid residue contact map will be used as the adjacency matrix input of different feature channels to capture and learn the complex graph relationship between amino acid residues in protein molecules, and combine the residue-level mixed feature map of the amino acid sequence and the multi-level structural features to predict protein function;

[0044] Step S5, construct a multi-label classifier, perform global pooling and aggregation operations on the output node features of the three-channel combined graph neural network, map the aggregated features to a 0-1 probability space, and obtain the probability value of each protein gene ontology function label;

[0045] Step S6, a complete protein function prediction model is formed by a sequence hybrid feature extractor, a three-channel combined graph neural network, and a multi-label classifier. The protein function prediction model is pre-trained on a training set, and the performance of the protein function prediction model at different training stages is evaluated using a validation set and a test set. The protein function prediction model that has passed the training is used to perform protein function prediction.

[0046] The step S1 of constructing a NoSQL protein document-based relational database based on MongoDB includes the following steps:

[0047] Step S101, use the interface test script and the test tool PostMan to test the open data interface of the Uniprot database, and obtain the Uniprot database identifier of the corresponding protein entry through the screening conditions: UPID, the screened protein entries must have the protein spatial structure file determined experimentally or the protein spatial structure file predicted by the AlphaFold model, the length of the amino acid sequence of each protein does not exceed 1022 (the ESM network processes a maximum of 1022 residues of the amino acid sequence of the protein at a time, if it exceeds, it needs to be sliced), select the protein gene ontology (GO) as the functional label of the protein, and the total number of protein gene ontology labels possessed by each screened protein entry should be greater than or equal to 3. In this embodiment, 26,335 Uniprot database identifiers that meet the above conditions are screened out as the source data ID of the database.

[0048] Step S102, accessing the Uniprot database data interface through the screened Uniprot database identifier, obtaining the cross-index ID of the corresponding protein entry in the PDB and AlphaFold databases from the data packet returned by the interface, and at the same time associating the primary key target (Target) ID of the target protein information table of the ChEMBL database through the Uniprot database identifier, and obtaining the remaining secondary table IDs in the ChEMBL database through this association;

[0049] Step S103, classify and request the data interface corresponding to each type of ID, parse the XML and JSON data packets returned by the interface through BeautifulSoup, Jsonpath framework and regular expression, and obtain a JSON dictionary consisting of protein association information in the three databases of Uniprot, PDB and AlphaFold, and a JSON dictionary consisting of target protein association data in the ChEMBL database;

[0050] Step S104: Based on the JSON dictionary, a protein information master table is constructed using the Pymongo framework with the Uniprot database identifier as the primary key, and a ChEMBL associated information table containing target information is constructed with the primary key target ID as the primary key;

[0051] Step S105, cross-correlate the primary key target ID in the ChEMBL association information table with the Uniprot database identifier in the protein information main table to obtain the final protein document-type association database. Users can retrieve key information of related protein entries in Uniprot, PDB, AlphaFold, and ChEMBL databases only through the Uniprot database identifier.

[0052] The benchmark data set for protein function prediction task obtained from the protein document-type association database specifically refers to: In this embodiment, 5,000 protein entries with high-precision experimental structures are first selected from a total of 26,335 protein data entries, and this data set is called Dataset1. In addition, 2,000 protein entries with AlphaFold model structures are selected and expanded to Dataset1 to obtain a mixed data set of 7,000 entries recorded as Dataset2. In order to reduce experimental overhead, this embodiment only takes GO terms with evidence coding (EXP, IDA, IPI, IMP, IGI, IEP, TAS, IC, HTP, HDA, HMP, HGI, HEP) as the protein gene ontology function label of the protein, and divides the extracted data set into a training set, a validation set, and a test set at a ratio of 8:1:1, wherein the protein entries in the test set have three protein gene ontology function labels of MF, BP, and CC at the same time. The distribution of protein gene ontology function labels of the data set is shown in Table 1.

[0053] Table 1: Distribution of protein gene ontology function labels of the dataset extracted in Example

[0054]

[0055] The spatial contact relationship of the Ca atoms in the residues in step S2 is: assuming that a protein is composed of N amino acid residues, traverse the protein spatial structure file, extract the spatial structure coordinates of the Ca atoms in each amino acid residue, calculate the spatial distance between all Ca atoms based on the coordinates, set a distance threshold to determine whether two residues are in contact, thereby obtaining an N*N amino acid residue contact matrix, and then only take the edge r with a contact relationship, and convert the distance matrix into a 2*2r adjacency matrix according to the source node and target node of the edge r, which is used as the amino acid residue contact map.

[0056] The amino acid residue contact maps of different connection densities in step S2 are: respectively calculating the amino acid residue contact maps of the protein spatial structure file actually determined by PDB and the protein spatial structure file predicted by AlphaFold, and calculating the amino acid residue contact maps representing different contact densities according to different distance thresholds. Specifically, for the protein entry with the Uniprot database identifier P05067, its corresponding PDB experimental structure index ID is: 8OTF, and the AlphaFold model predicted structure index ID is:

[0057] AF-P05067-F1, thereby obtaining the protein spatial structure file of P05067. By parsing this file, the spatial coordinates of the Ca atom in each amino acid residue in the protein molecule are obtained. The complete P05067 protein is composed of 770 amino acid residues. The spatial distance between Ca atoms is calculated to construct the amino acid residue contact matrix , Represents the dimension, only the edges with contact relationship in the contact matrix of the amino acid residue are taken, and the conversion is the adjacency matrix , as the amino acid residue contact map of protein P05067.

[0058] Specifically, the residue contact calculation formula can be defined as:

[0059] ;

[0060] in are the residues extracted from the protein spatial structure file and residues The three-dimensional coordinates of the Ca atom in is the contact distance threshold, Residue and residues When the calculated spatial distance is less than this threshold, the two amino acid residues are considered to have a contact relationship in space. In this embodiment, two types of amino acid residue contact maps under contact thresholds of 8Å and 10Å are calculated, representing sparse amino acid residue contact maps and dense amino acid residue contact maps, respectively.

[0061] like Figure 2 In step S3, the complete sequence of the protein or a subsequence is input into the amino acid sequence hybrid feature extractor. The Bi-LSTM network first extracts a 6155-dimensional sequence embedding representation. Then, the ESM2 network (Evolutionary Scale Modelling) composed of 33 layers of Transformer encoders continues to extract a 1280-dimensional sequence embedding representation, which are used together as the primary sequence embedding representation of the amino acid sequence.

[0062] The feature fusion and enrichment operations for the primary embedding representation of the sequence described in step S3 are: when the length of the amino acid sequence is too long, the overall computational complexity of the model will be significantly increased. In order to reduce the training cost and obtain more effective features, the primary embedding representation of the sequence is sliced ​​by category into: sequence one-hot encoding mapping, residue sequence deep context features, and residue biochemical property encoding. During the training process of the entire model, the above slices are subjected to learnable nonlinear mapping, and the primary embedding representation of the sequence is feature enriched and reduced by adjusting the dimensionality of each hidden layer. Finally, the residue-level mixed feature mapping of the amino acid sequence is obtained through ReLU nonlinear mapping and feature fusion operations.

[0063] Specifically, for the protein entry with UPID P05067, its complete sequence (770 amino acids in total) is input into the amino acid sequence hybrid feature extractor, and the Bi-LSTM network extracts the sequence one-hot encoding mapping (20 standard amino acids and one unknown amino acid) and deep context features of residue sequences , the ESM2 network will extract the biochemical properties of its residues , thus obtaining the primary embedding representation of the sequence , In order to avoid wasting resources due to the large number of amino acid residues, Slice by category and perform three learnable nonlinear mappings. Adjust the size of each hidden layer during linear changes to 21, 512, and 320; perform feature enrichment and dimensionality reduction on the primary embedding representation of the sequence. Finally, after ReLU nonlinear mapping, perform feature fusion on the three types of output features to obtain a residue-level mixed feature map of the amino acid sequence. ,During the above learnable nonlinear transformation process, all weight parameters and the corresponding bias values ​​will be trained together with all the parameters in the subsequent ,protein function prediction model.

[0064] like Figure 3 As shown in the figure, the three-channel combined graph neural network consists of a two-layer GCN network in series to form the first feature channel; another identical two-layer GCN network in series to form the second feature channel, the first feature channel is parallel to the second feature channel, forming a 2*2 parallel dual-channel GCN network; the GATv2 network with a multi-head graph attention mechanism is introduced to form the third feature channel, and the GATv2 network is connected in series to the parallel dual-channel GCN network. The activation function of the GCN network and the GATv2 network adopts the ReLU activation function.

[0065] Specifically, the feature aggregation and propagation method of the three-channel combined graph neural network is as follows: using the residue-level mixed feature mapping of the amino acid sequence As the initial feature matrix input of the 2*2 parallel dual-channel GCN network, the sparse amino acid residue contact map extracted with an 8Å contact threshold is used as the adjacency matrix input of one feature channel of the parallel network, and the dense amino acid residue contact map extracted with a 10Å contact threshold is used as the adjacency matrix input of the other feature channel. The output features of the 2*2 parallel dual-channel GCN network are fused to obtain the fused features. , which is used as the third feature channel, namely the feature matrix input of the GATv2 network. At the same time, the sparse amino acid residue contact map extracted using the 8Å threshold is used as the adjacency matrix input of the GATv2 network. After further noise filtering and feature extraction, the output node features of the three-channel combined graph neural network are obtained. .

[0066] In step S5, the process of performing global pooling and aggregation operations on the output node features of the three-channel combined graph neural network is as follows: Figure 4 As shown;

[0067] Output node features of three-channel combined graph neural network Average pooling and cumulative pooling are performed separately, and then the two types of pooling information are aggregated to obtain more comprehensive node feature information.

[0068] Specifically, the global pooling operation can be defined as:

[0069] ;

[0070] ;

[0071] ;

[0072] in, is the output node feature The feature vector of a node, Represents the accumulated pooling information, represents the average pooling information, Represents the node features after aggregation, N is the number of nodes; concat represents the aggregation operation.

[0073] The node features after aggregation Batch normalization, ReLU activation, and Dropout operations are performed in sequence. After a fully connected layer mapping, the sigmoid function is used to map the feature output to a 0-1 probability space to obtain the probability value of each protein gene ontology functional label.

[0074] The method of pre-training the protein function prediction model on the training set in step S6 is: selecting the training set in the benchmark data set as the training data of the protein function prediction model; using MF, BP, and CC labels as protein gene ontology function labels during supervised learning; selecting Adam as the training optimizer, and setting the learning rate to 0.0001; selecting the binary cross entropy function as the loss function during training; setting the training batch size to 64; and setting the maximum iteration round (epoch) to 50. During the training process, the validation set is used to evaluate the performance of the protein function prediction model after each iteration round (epoch), that is, after each round of training, the protein function prediction model will predict the protein gene ontology function labels of all protein entries in the validation set, and respectively calculate the scores of the prediction results of the protein function prediction model on the three evaluation indicators of micro precision recall rate curve area (m-AUPR), macro precision recall rate curve area (M-AUPR), and maximum F1 score (F-MAX). If there is an improvement in the score compared with the previous round, the protein function prediction model pre-trained in this round is saved. The training is stopped after reaching the maximum iteration round (epoch). Finally, the performance of the pre-trained protein function prediction model is evaluated on the test set.

[0075] The protein function prediction model proposed in the present invention was pre-trained on the training sets corresponding to Dataset1 and Dataset2, and tested on the test set of Dataset1. The scores of the protein function prediction model for predicting MF labels (482 in total) on three types of evaluation indicators were calculated, and the highest score among the five tests was taken. The results are shown in Table 2.

[0076] Table 2: Effect of expanding AlphaFold’s protein structure prediction data on model performance

[0077]

[0078] The DeepFRI model and the protein function prediction model of the present invention were evaluated and compared on the dataset Dataset2. The model was pre-trained for 50 epochs, and the scores of the models on the three evaluation indicators were calculated respectively. The highest score in the five tests was taken, and the prediction performance of the models on the three types of protein gene ontology function labels, MF (517 in total), BP (2124 in total), and CC (408 in total), were discussed respectively. The results are shown in Tables 3 and 4.

[0079] Table 3: Protein functions predicted by DeepFRI model

[0080]

[0081] Table 4: Protein function prediction model of the present invention predicts protein function

[0082]

[0083] It can be seen from the above experimental results that the present invention surpasses the DeepFRI model in most cases, and uses AlphaFold predicted protein structure data as the training input of the model, avoiding the problem of protein entries due to the lack of experimentally measured spatial structure, and improving the generalization performance of the model. At the same time, the protein document-based association database proposed by the present invention can greatly improve the retrieval efficiency of protein association information and facilitate the downstream protein prediction tasks.

[0084] Finally, it should be noted that the above is only a preferred example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting protein function based on transfer learning and three-channel combined GNN, characterized in that: The following steps are involved: Step S1, obtaining the amino acid sequence, protein spatial structure file, and protein gene ontology function label of the protein from the protein public database, cross-correlating and cleaning the obtained data, and then constructing a NoSQL type protein document-based association database, obtaining a benchmark data set for the protein function prediction task from the protein document-based association database, and dividing it into a training set, a validation set, and a test set; Step S2, parsing the protein spatial structure file, extracting the spatial coordinates of the Ca atom in each amino acid residue in the protein molecule, and based on the spatial contact relationship of the Ca atoms in all residues, extracting the amino acid residue contact maps with different connection densities according to different contact thresholds, so as to map the complex spatial structure of the protein molecule; Step S3, constructing an amino acid sequence hybrid feature extractor based on transfer learning, extracting multiple types of residue-level features from the amino acid sequence through the amino acid sequence hybrid feature extractor to form a primary embedding representation of the sequence; performing feature fusion and enrichment operations on the primary embedding representation of the sequence to obtain a residue-level hybrid feature map of the amino acid sequence; The complete sequence of the protein or a subsequence is input into the amino acid sequence hybrid feature extractor. The sequence embedding representation is first extracted by the Bi-LSTM network. Then, the ESM2 network composed of 33 layers of Transformer encoders continues to extract the sequence embedding representation. The sequence embedding representations extracted by the Bi-LSTM network and the ESM2 network are used together as the primary sequence embedding representation of the amino acid sequence. The feature fusion and enrichment operation on the primary embedding representation of the sequence is as follows: the primary embedding representation of the sequence is sliced ​​by category into: sequence one-hot encoding mapping, residue sequence deep context features and residue biochemical property encoding, the slices are subjected to learnable nonlinear mapping, the primary embedding representation of the sequence is subjected to feature enrichment and dimensionality reduction by adjusting the dimensionality of each hidden layer, and finally, a residue-level mixed feature map of the amino acid sequence is obtained by ReLU nonlinear mapping and feature fusion operation; Step S4, introducing a graph convolutional network and a GATv2 network with a multi-head graph attention mechanism to construct a three-channel combined graph neural network, using the residue-level mixed feature map of the amino acid sequence as the initial feature matrix input of the three-channel combined graph neural network, and the amino acid residue contact map will be used as the adjacency matrix input of different feature channels, capturing and learning the complex graph relationship between amino acid residues in protein molecules, and combining the residue-level mixed feature map of the amino acid sequence and the multi-level structural features to predict protein function; the three-channel combined graph neural network, the first feature channel is composed of a two-layer GCN network in series; another identical two-layer GCN network in series constitutes the second feature channel, the first feature channel is parallel to the second feature channel, forming a 2*2 parallel dual-channel GCN network; introducing a GATv2 network with a multi-head graph attention mechanism to form a third feature channel, and the GATv2 network is connected in series to the parallel dual-channel GCN network; Step S5, construct a multi-label classifier, perform global pooling and aggregation operations on the output node features of the three-channel combined graph neural network, map the aggregated features to a 0-1 probability space, and obtain the probability value of each protein gene ontology function label; Step S6, a complete protein function prediction model is formed by a sequence hybrid feature extractor, a three-channel combined graph neural network, and a multi-label classifier. The protein function prediction model is pre-trained on a training set, and the performance of the protein function prediction model at different training stages is evaluated using a validation set and a test set. The protein function prediction model that has passed the training is used to perform protein function prediction.

2. The method for predicting protein function based on transfer learning and three-channel combined GNN according to claim 1, characterized in that: The spatial contact relationship of the Ca atoms in the residues in step S2 is: a protein is composed of N amino acid residues, and the protein spatial structure file is traversed to extract the spatial structure coordinates of the Ca atoms in each amino acid residue. The spatial distance between all Ca atoms is calculated based on the coordinates, and a distance threshold is set to determine whether two residues are in contact, thereby obtaining an N*N amino acid residue contact matrix. Subsequently, only the edge r with a contact relationship is taken, and the distance matrix is ​​converted into a 2*2r adjacency matrix according to the source node and the target node of the edge r, which is used as the amino acid residue contact map.

3. The method for predicting protein function based on transfer learning and three-channel combined GNN according to claim 2, characterized in that: The amino acid residue contact maps with different connection densities described in step S2 are: the amino acid residue contact maps of the protein spatial structure file actually determined by PDB and the protein spatial structure file predicted by AlphaFold are calculated respectively, and the amino acid residue contact maps representing different contact densities are calculated according to different distance threshold sizes.

4. The method for predicting protein function based on transfer learning and three-channel combined GNN according to claim 1, characterized in that: The feature aggregation and propagation method of the three-channel combined graph neural network is as follows: using the residue-level mixed feature map of the amino acid sequence as the initial feature matrix input of the 2*2 parallel dual-channel GCN network, the sparse amino acid residue contact map extracted with a smaller contact threshold is used as the adjacency matrix input of one of the feature channels of the parallel network, and the dense amino acid residue contact map extracted with a larger contact threshold is used as the adjacency matrix input of another feature channel; the output features of the 2*2 parallel dual-channel GCN network are feature fused to obtain fused features, which are used as the feature matrix input of the GATv2 network, and the sparse amino acid residue contact map is used as the adjacency matrix input of the GATv2 network; the GATv2 network obtains the output node features of the three-channel combined graph neural network after further noise filtering and feature extraction of the fused features.

5. The method for predicting protein function based on transfer learning and three-channel combined GNN according to claim 1, characterized in that: In step S1, a NoSQL protein document-based relational database is constructed based on MongoDB.

6. The method for predicting protein function based on transfer learning and three-channel combined GNN according to claim 5, characterized in that: The steps to build a NoSQL protein document-based relational database are as follows: Step S101, use the interface test script and the test tool PostMan to test the open data interface of the Uniprot database, obtain the Uniprot database identifier of the corresponding protein entry through the screening conditions: UPID, the screened protein entry must have the protein spatial structure file determined experimentally or the protein spatial structure file predicted by the AlphaFold model, select the protein gene ontology as the functional label of the protein, the total number of protein gene ontology labels possessed by each screened protein entry should be greater than or equal to 3, and the screened Uniprot database identifier is used as the source data ID of the database; Step S102, accessing the Uniprot database data interface through the Uniprot database identifier obtained by screening, obtaining the cross-index ID of the corresponding protein entry in the PDB and AlphaFold databases from the data packet returned by the interface, and at the same time associating the primary key target ID of the target protein information table of the ChEMBL database through the Uniprot database identifier, and obtaining the remaining secondary table IDs in the ChEMBL database through this association; Step S103, classify and request the data interface corresponding to each type of ID, parse the XML and JSON data packets returned by the interface through BeautifulSoup, Jsonpath framework and regular expression, and obtain a JSON dictionary consisting of protein association information in the three databases of Uniprot, PDB, and AlphaFold, and a JSON dictionary consisting of target protein association data in the ChEMBL database; Step S104: Based on the JSON dictionary, a protein information master table is constructed using the Pymongo framework with the Uniprot database identifier as the primary key, and a ChEMBL associated information table containing target information is constructed with the primary key target ID as the primary key; Step S105, cross-correlate the primary key target ID in the ChEMBL association information table with the Uniprot database identifier in the protein information main table to obtain the final protein document-type association database. The user only needs to use the Uniprot database identifier to retrieve the key information of the relevant protein entry in the Uniprot, PDB, AlphaFold, and ChEMBL databases.

Images