A multi-species protein function prediction method and system based on heterogeneous networks

Abstract

The present invention discloses a method and system for predicting the function of multi-species proteins based on heterogeneous networks, the method comprising: extracting features from protein sequences using an ESM-2 model to obtain sequence features; constructing a structural contact graph of proteins, training the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extracting structural features; splicing sequence features and structural features, and constructing a cross-species heterogeneous network based on a PPI network and a homology similarity network; in the training phase, using structural features and sequence features to propagate on a heterogeneous network, and using a graph attention mechanism to update node vectors during propagation; in the prediction phase, adding the GO label of a training set for network propagation, linearly combining the propagation results of protein representation and GO labels, and obtaining the final GO label prediction probability. The present invention improves the prediction effect of multi-species protein function prediction and realizes cross-species propagation of functional labels.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method, system, terminal and computer-readable storage medium for predicting multi-species protein functions based on a heterogeneous network. Background Art

[0002] Proteins are responsible for almost all life activities in organisms. Cell differentiation and renewal, transmission of biological information, energy supply and storage are all inseparable from the role of proteins. Even viruses with the simplest cell structure must be protected by a protein shell. Undoubtedly, proteins play a very critical role. Accurately locating the function, regulation and expression of proteins is the primary goal of proteomics. With the development of high-throughput sequencing technology, protein sequence data has been accumulated in large quantities. However, only less than 0.1% of proteins can be functionally annotated through biological experiments. Therefore, it is particularly important to use bioinformatics technology to annotate the functions of proteins.

[0003] In order to facilitate the prediction of protein functions, the Gene Ontology (GO) library standardizes protein functions into GO terms in three fields, namely: biological process (BP), molecular function (MF) and cellular component (CC). In this way, the protein function prediction problem is abstracted into a multi-label classification problem. According to the first central dogma, the sequence of a protein determines its structure, and the structure determines its function. Therefore, the earliest prediction technology focused on sequence data, using sequence similarity as the basis for transferring annotations, and thus measured the functions of some proteins. With the development of technology, a large amount of protein structural data has been measured. The use of structural data has broken through the bottleneck of sequence similarity and once again improved the accuracy of prediction. With the increase of biological data year by year, the function prediction model based on a single species has achieved some results. However, there is still a problem of data imbalance between different species. The protein function labels of some species are sparse, which is a huge challenge for the function prediction of the species. In order to solve this problem, protein function prediction technology based on multiple species came into being. Different from previous prediction methods, protein function prediction based on multi-species networks establishes a cross-species network by analyzing the PPI (protein-protein interaction) networks of multiple species and the similarities of proteins between species, and mines the functional connections between proteins of different species in the network through isomorphic networks or heterogeneous networks. It not only predicts the functions of proteins of this species, but also provides label transfer for other species.

[0004] Through InterProScan (non-redundant database), protein sequences are used to obtain information such as protein families and domains as the initial features of PPI network nodes. Multi-layer graph convolution is used to update the representation vector of each node to capture high-order information. The final node representation is input into the multi-label classifier, and the probability of each functional label of each protein is output. However, the functional labels in this model can only be propagated independently in the PPI network of each species and cannot provide information for network nodes of other species. Based on the idea of ​​homology similarity, the sequence and PPI network information of multiple species are integrated by calculating the IsoRank similarity score to create a network map of proteins of multiple species, which is essentially a homologous network. This method uses a multi-species network as input, expands the training set, and trains a neural network with GO annotation as the target label. Although this method has significantly improved the performance of multi-species function prediction, the number of protein nodes input is limited by the size of the GPU. Heterogeneous networks can also be used to establish homology relationships between different species, but it quantifies the step size of each network node and limits the randomness of the search. To solve this problem, a composite heterogeneous network is constructed by performing asynchrony on heterogeneous networks, using complementary annotations of different species, GO hierarchical structures and multi-level biological data related to proteins. The step size of each node is used to perform asynchronous random walks to predict the association between proteins and functional labels. In order to enable labels to be effectively propagated in multi-species networks, S2F introduced a new label diffusion algorithm. Based on the label diffusion algorithm proposed by S2F, SPROF-GO was proposed. This model uses the DIAMOND tool to search for homologous proteins and construct a multi-species network. The embedding obtained by the attention mechanism is input into the MLP to obtain a preliminary GO prediction, and the model parameters are updated with the preliminary prediction as the node attribute. In the testing phase, the initial prediction and the homologous network are fed into the label diffusion algorithm to output the final protein function prediction. However, SPROF-GO cannot solve the noise problem existing in homology similarity networks.

[0005] In summary, there are two main problems in the field of multi-species protein function prediction:

[0006] (1) Multi-species PPI networks are large and complex, and it is difficult to directly use structural data to propagate on the network. On the one hand, structural data consumes more memory and time resources than sequence data; on the other hand, the protein structures of many species are unknown, which limits multi-species datasets. Current methods all use sequence features to propagate on PPI networks, but sequence features can only provide less information, so the model prediction accuracy is limited.

[0007] (2) In order to avoid overly complex models, most methods only use one isomorphic network, such as the PPI network or the sequence homology similarity network. However, no matter which network is used, there are defects. The method using the PPI network cannot propagate labels across the PPI network, and the method using the sequence homology similarity network will introduce noise, because even proteins with similar sequences may exhibit different functions.

[0008] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0009] The main purpose of the present invention is to provide a method, system, terminal and computer-readable storage medium for predicting multi-species protein functions based on a heterogeneous network, aiming to solve the problem of low model prediction accuracy and introduction of noise when predicting multi-species protein functions in the prior art, resulting in poor prediction effect.

[0010] To achieve the above object, the present invention provides a method for predicting the function of multi-species proteins based on a heterogeneous network, and the method for predicting the function of multi-species proteins based on a heterogeneous network comprises the following steps:

[0011] Use the ESM-2 model to extract features from protein sequences and obtain sequence features;

[0012] Constructing a structural contact graph of the protein, training the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extracting structural features;

[0013] Splicing the sequence features and the structural features to obtain splicing node features, constructing a cross-species heterogeneous network based on a PPI network and a homology similarity network, and inputting the splicing node features into the heterogeneous network;

[0014] In the training phase, the structural features and the sequence features are used to propagate on the heterogeneous network, and a graph attention mechanism is used to update node vectors during propagation;

[0015] In the prediction stage, the GO tags of the training set are added for network propagation, and the propagation results of protein representation and GO tags are linearly combined to obtain the final GO tag prediction probability.

[0016] In addition, to achieve the above-mentioned purpose, the present invention also provides a multi-species protein function prediction system based on a heterogeneous network, wherein the multi-species protein function prediction system based on a heterogeneous network comprises:

[0017] A sequence feature extraction module is used to extract features from protein sequences using the ESM-2 model to obtain sequence features;

[0018] A structural feature extraction module is used to construct a structural contact graph of the protein, train the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extract structural features;

[0019] A feature splicing and heterogeneous network construction module, used to splice the sequence features and the structural features to obtain splicing node features, construct a cross-species heterogeneous network based on the PPI network and the homology similarity network, and input the splicing node features into the heterogeneous network;

[0020] A training module, used for propagating the structural features and the sequence features on the heterogeneous network during a training phase, and updating the node vectors using a graph attention mechanism during propagation;

[0021] The prediction module is used to add the GO tags of the training set for network propagation in the prediction stage, linearly combine the propagation results of protein representation and GO tags, and obtain the final GO tag prediction probability.

[0022] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a multi-species protein function prediction program based on a heterogeneous network stored in the memory and executable on the processor, wherein the multi-species protein function prediction program based on a heterogeneous network implements the steps of the multi-species protein function prediction method based on a heterogeneous network as described above when executed by the processor.

[0023] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a multi-species protein function prediction program based on a heterogeneous network, and when the multi-species protein function prediction program based on a heterogeneous network is executed by a processor, the steps of the multi-species protein function prediction method based on a heterogeneous network as described above are implemented.

[0024] In the present invention, the ESM-2 model is used to extract features of protein sequences to obtain sequence features; a structural contact graph of proteins is constructed, and the structural contact graph is trained based on a structural model of graph convolution and hierarchical graph pooling to extract structural features; the sequence features and the structural features are spliced ​​to obtain splicing node features, a cross-species heterogeneous network is constructed according to the PPI network and the homology similarity network, and the splicing node features are input into the heterogeneous network; in the training stage, the structural features and the sequence features are used to propagate on the heterogeneous network, and the node vector is updated using the graph attention mechanism during propagation; in the prediction stage, the GO label of the training set is added for network propagation, and the propagation results of the protein representation and the GO label are linearly combined to obtain the final GO label prediction probability. In the field of protein function prediction of multiple species, the present invention combines structural features with the PPI network of multiple species, uses graph convolution and hierarchical graph pooling to process the structural contact graph, uses a heterogeneous network propagation algorithm based on the graph attention mechanism, obtains low-dimensional representations of proteins from different levels, improves the prediction effect, and realizes cross-species propagation of functional labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a preferred embodiment of the method for predicting multi-species protein functions based on heterogeneous networks of the present invention;

[0026] Figure 2 It is a schematic diagram of the principle of the multi-species protein function prediction process in a preferred embodiment of the multi-species protein function prediction method based on heterogeneous network of the present invention;

[0027] Figure 3 It is a schematic diagram of a graph convolution pooling module framework in a preferred embodiment of the method for predicting multi-species protein functions based on heterogeneous networks of the present invention;

[0028] Figure 4 It is a schematic diagram of a pseudo code of a structural model in a preferred embodiment of the method for predicting multi-species protein functions based on heterogeneous networks of the present invention;

[0029] Figure 5 It is a schematic diagram of a pseudo code of a multi-species protein function prediction algorithm in a preferred embodiment of the multi-species protein function prediction method based on a heterogeneous network of the present invention;

[0030] Figure 6 It is a structural diagram of a preferred embodiment of the heterogeneous network-based multi-species protein function prediction system of the present invention;

[0031] Figure 7 It is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] The present invention proposes a multi-species protein function prediction method based on heterogeneous networks and protein structures, which can realize cross-species function propagation by using homology similarity and PPI networks, and introduces structural data for the first time to make the protein representation more complete. First, sequence features are extracted from protein sequences using ESM-2, and a structural model is trained for protein structure to extract structural features. Then, a heterogeneous network composed of homology similarity and PPI networks is used. Structural features and sequence features are propagated on the heterogeneous network during the training phase, and the node vector is updated using the graph attention mechanism during propagation. Finally, the GO label of the training set is added to the prediction phase for network propagation, and the propagation results of protein representation and GO label are linearly combined to obtain the final GO label prediction probability; the present invention utilizes the complementarity of the two networks to alleviate the noise problem to a certain extent, successfully improves the prediction effect, and makes the model have a certain robustness.

[0034] The multi-species protein function prediction method based on heterogeneous network described in the preferred embodiment of the present invention is as follows: Figure 1 and Figure 2 As shown, the multi-species protein function prediction method based on heterogeneous networks includes the following steps:

[0035] Step S10: Use the ESM-2 model to extract features from the protein sequence to obtain sequence features.

[0036] Specifically, the ESM-2 model is used to extract features from protein sequences and obtain the sequence feature H se The ESM-2 model is a large protein language model that uses a transformer-based architecture and attention mechanism to learn the interaction patterns between pairs of amino acids in the input sequence, which enables the ESM-2 model to capture the amino acid sequence evolution information about proteins and their properties. Each protein sequence is represented as a 1280-dimensional feature vector after the ESM-2 model (each sequence is represented as a 1280-dimensional vector after ESM-2); where H se A matrix representing multiple protein sequence feature vectors.

[0037] Step S20: construct a structural contact graph of the protein, train the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extract structural features.

[0038] Specifically, the original data is obtained, wherein the original data is the three-dimensional atomic coordinates of the amino acid sequence, based on the distance between the Cα atoms being less than Connect the edges between amino acids to construct a structural contact graph of the protein; the present invention plans to obtain the initial embedding representation of the amino acid nodes from the structural contact graph. Among the existing graph representation learning methods, the present invention selects the node2vec algorithm to obtain several embedding vectors. The node2vec algorithm improves the DeepWalk algorithm. Deepwalk uses a random walk strategy to select neighbor nodes to form a "sentence", and then uses word2vec to generate word embeddings for each node in the corpus. Node2vec changes the strategy of selecting neighbor nodes to a biased random walk. In the specific implementation, two hyperparameters p and q are used. The two hyperparameters are related to the probability of selecting the next hop node, and the selection probability of all domain nodes of the current node is calculated. Finally, the value with the largest probability will be selected as the next node until the walk length reaches a threshold or there is no neighborhood, and then continue to select a node that has not been the starting node to start the next walk.

[0039] If the current node is v, the probability of selecting x as the next node is:

[0040]

[0041] Among them, P(n i =|n i-1 =) indicates the probability that the next hop is x when the current node is v, n i represents the node selected by the i-th hop, n i-1 represents the node selected in the i-1th hop, π vx represents the probability of walking from vertex v to vertex x, π vx The value of is affected by weights and hyperparameters, Z represents the normalization constant, and E represents the edge set of the graph.

[0042] If the last walk was to reach node v via node t, π vx The calculation is:

[0043] π vx = pq ( ,x)·w vx ; (2)

[0044]

[0045] Among them, α pq (,x) represents the influencing factor that determines whether to walk inward or outward. p and q represent the hyperparameters introduced by the node2vec algorithm. By setting the hyperparameters to control the influencing factor, the bias of the random walk strategy can be controlled. w vx represents the weight of the (v, x) edge, dtx Represents the distance between node x and node t; node2vec will walk all the nodes in the structural contact graph and generate an initial embedding vector for each amino acid node.

[0046] Since the degree of the network involved in the present invention is generally large, in the downstream network propagation task, a neighborhood aggregation needs to read in the features of tens of thousands of proteins. Considering the time and resource constraints, the structure contact graph cannot be directly input into the network propagation layer for propagation. The method of the present invention is inspired by the Struct2GO model, and uses a structural model based on graph convolution and hierarchical graph pooling to train the structural contact graph, thereby extracting a low-dimensional feature representation of the structure. The structural model mainly includes a graph convolution pooling module and an MLP classifier. The graph convolution pooling module process is as follows Figure 3 As shown, the pseudo code of the structural model algorithm is as follows Figure 4 As shown (i.e. Algorithm 3-1: structural model pseudo code).

[0047] The graph convolution pooling module performs graph convolution on the structural contact graph, aggregates the neighborhood information, and then screens the nodes, maps the node features to one dimension to obtain the attention score of the corresponding node, screens out some nodes with scores higher than the preset score, and aggregates the features after graph convolution again according to the attention score to obtain the aggregated subgraph. The aggregated subgraph is pooled to obtain the output of the graph convolution pooling module of this layer. After aggregating the neighborhood information, the graph convolution pooling module does not perform graph pooling on the whole world, but selects the subgraph composed of the nodes that have the greatest impact on the structural features for pooling operation. This can retain the most obvious structural features and filter out noise. Therefore, by superimposing multiple layers of graph convolution pooling modules, effective structural features H can be learned. st , the present invention is set to two layers here, the output of the module is the splicing of the two layers of graph pooling 1024 dimensions, the output of the two layers of modules is added together to obtain a 1024-dimensional structural feature, and H st To represent the structural feature matrix of multiple proteins.

[0048] The structural model uses an MLP classifier to classify the structural contact graph, and the structural model is trained by calculating the cross entropy loss between the probability of predicting the GO label and the true value. Finally, the 1024-dimensional features before the MLP classifier are extracted as the structural features input for the downstream prediction task.

[0049] The heterogeneous network used in the present invention comes from the PPI network based on functional expression on the one hand, and from the homology similarity network based on sequence homology similarity on the other hand. The PPI network of multiple species can be downloaded directly from the STRING database, and the homology similarity network is obtained by comparing the similarities between sequences of multiple species using the Diamond tool, with the comparison threshold set to 10 -4 Thus, we get the cross-species heterogeneous network G = (V, E1 , E 2 ), V is all the nodes in the network, E 1 represents the edge in the PPI network, E 2 Represents the edge in the homology similarity network. The present invention hopes to use the underlying homology similarity network to provide essential attributes for proteins, and use the surface PPI network to supplement the functional expression level information for proteins. Therefore, the present invention proposes a multi-species protein function prediction model based on the above heterogeneous network propagation. The algorithm pseudo code is as follows: Figure 5 As shown (i.e., Algorithm 3-2: pseudo code of multi-species protein function prediction algorithm).

[0050] Step S30, splicing the sequence features and the structural features to obtain splicing node features, constructing a cross-species heterogeneous network based on the PPI network and the homology similarity network, and inputting the splicing node features into the heterogeneous network.

[0051] Specifically, the sequence feature H se and structural features H st Perform splicing to obtain the splicing node features:

[0052] H=H se | st ; (4)

[0053] Among them, H represents the splicing node feature, | represents the splicing operation;

[0054] In order to save time and memory, it is also necessary to map the splicing node features H into a low-dimensional vector space and use the MLP classifier to continue mapping to avoid problems such as over-smoothing caused by subsequent convolution processes:

[0055]

[0056] in, represents the input of the first layer of MLP, σ represents the nonlinear activation function, W e and b e represents the learnable parameters of the linear layer, represents the output of the l-th layer MLP, R d represents the real number domain, d is the hidden dimension of the output, d is set to 512 dimensions in the experiment, LN(·) represents the layer regularization function, represents the input of the l-th layer MLP, and represents the learnable parameters of the lth layer in the MLP.

[0057] Denote the PPI network as G p =V,E 1 ), the homology similarity network is represented by G s=(V, E 2 ).

[0058] Step S40: During the training phase, the structural features and the sequence features are used to propagate on the heterogeneous network, and a graph attention mechanism is used to update node vectors during propagation.

[0059] Specifically, the present invention adopts the graph attention mechanism to adjust the intensity of propagation between network nodes and learns an attention score for each edge of the node:

[0060]

[0061] Among them, α uv represents the attention score of protein node u to node v, a T and W t is a learnable weight, T represents a transposition operation, the | symbol represents a concatenation operation, and h u and h v Represent the node features of proteins u and v respectively, N u represents the neighbor set of u, i represents the neighbor set from N u The neighbor protein nodes selected in i Represents the node characteristics of the neighbor protein node i. In the experiment, the activation function σ is selected as the LeakyReLU function.

[0062] Using the same graph attention parameters for the PPI network G p and homology similarity network G s After updating the weights between nodes, they are represented as A p , A s ∈R n×n , where A p , A s denote the attention score matrices in the PPI network and homology similarity network, respectively, and n denotes the total number of protein nodes.

[0063] In order to improve the operation efficiency and save time and memory, only protein features are brought into the heterogeneous network propagation in the training stage, and the network propagation of functional labels is added in the prediction stage. The protein label is represented as a binary vector:

[0064]

[0065]

[0066] Formulas (8) and (9) indicate that the propagation is performed on the same network propagation layer, so the same weight H is used. () .

[0067] in, represents the propagation result on the PPI network of the first layer, which is a matrix composed of the node feature vectors after the protein is updated. () represents the input of the l-th layer network propagation, W () represents the weight, represents the propagation result on the l-th layer homology similarity network, H (+1) represents the output of the l-th layer network propagation, combining different information from the similarity network and the PPI network; in the prediction stage, the label vector Y is also propagated through a heterogeneous network based on the attention mechanism, so that the protein can obtain a direct transfer of functional labels. The label vector of the first layer input during initialization and They represent the node feature vector matrices after label propagation on the l-th layer PPI network and homology similarity network, represents the input of the l-th layer of label propagation, represents the output of the l-th layer of label propagation, and Noim(·) represents L2 regularization.

[0068] However, not all nodes will update their label vectors. In the prediction stage, the present invention only updates the label vector for the test set without labels. The label vectors of the remaining nodes are reset to Y after each layer of propagation. i , the final output after label propagation is expressed as

[0069] After sufficient network propagation, the nodes have learned network information from different expression levels. Next, the node features need to be mapped to the output vector space to represent the probability of the function label. A linear layer and a sigmoid function are used. Finally, the binary cross entropy loss between the output probability and the true value is calculated, and the binary cross entropy loss is used for back propagation to update the parameters:

[0070]

[0071]

[0072] in, represents the mapping of node features into the label vector space, n represents the total number of protein nodes, c represents the dimension of the label vector, and H out Represents the output of the last layer of network propagation, W oit and b out Represents the parameters that need to be learned in the current linear mapping layer, Loss represents the binary cross entropy loss, Y represents the true value label matrix of the protein, and ⊙ represents the element-by-element multiplication operation.

[0073] Step S50: In the prediction stage, the GO tags of the training set are added to the network for propagation, and the propagation results of the protein representation and the GO tags are linearly combined to obtain the final GO tag prediction probability.

[0074] Specifically, in the prediction phase, the last layer of the network propagates the updated label vector and Perform a linear combination:

[0075]

[0076] in, represents the final GO tag prediction probability, c' represents the number of tags in the current gene ontology, represents the probability that the protein has label i, It represents a parameter used to measure the importance of label vectors and protein features, with a step size of 0.1. Different branches may take different values ​​depending on the experimental results. value.

[0077] In the field of protein function prediction of multiple species, the present invention combines structural features with PPI networks of multiple species, adopts graph convolution and hierarchical graph pooling to process structural contact graphs, and uses a heterogeneous network propagation algorithm based on a graph attention mechanism to obtain low-dimensional representations of proteins from different levels, thereby improving the prediction performance and realizing cross-species propagation of functional labels.

[0078] Furthermore, there are many ways to extract structural features. For example, GVP-GNN can replace the structural contact graph and node2vec algorithm, and the graph neural network in the structural model can also be changed to GAT, HGCN, GraphSAG, etc.; the network propagation algorithm can use RGCN instead of GAT for propagation in heterogeneous networks, and can also replace heterogeneous networks with homogeneous networks to realize label propagation.

[0079] Furthermore, if Figure 6 As shown, based on the above-mentioned multi-species protein function prediction method based on heterogeneous network, the present invention also provides a multi-species protein function prediction system based on heterogeneous network, wherein the multi-species protein function prediction system based on heterogeneous network includes:

[0080] A sequence feature extraction module 51 is used to extract features from protein sequences using an ESM-2 model to obtain sequence features;

[0081] A structural feature extraction module 52 is used to construct a structural contact graph of the protein, train the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extract structural features;

[0082] A feature splicing and heterogeneous network construction module 53 is used to splice the sequence features and the structural features to obtain splicing node features, construct a cross-species heterogeneous network based on the PPI network and the homology similarity network, and input the splicing node features into the heterogeneous network;

[0083] A training module 54 is used to propagate the structural features and the sequence features on the heterogeneous network during a training phase, and update node vectors using a graph attention mechanism during propagation;

[0084] The prediction module 55 is used to add the GO tags of the training set to perform network propagation in the prediction stage, linearly combine the propagation results of the protein representation and the GO tags, and obtain the final GO tag prediction probability.

[0085] Furthermore, if Figure 7 As shown, based on the above-mentioned heterogeneous network-based multi-species protein function prediction method and system, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Figure 7 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0086] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a multi-species protein function prediction program 40 based on a heterogeneous network is stored on the memory 20, and the multi-species protein function prediction program 40 based on a heterogeneous network can be executed by the processor 10, thereby realizing the multi-species protein function prediction method based on a heterogeneous network in the present application.

[0087] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the heterogeneous network-based multi-species protein function prediction method.

[0088] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0089] In one embodiment, when the processor 10 executes the heterogeneous network-based multi-species protein function prediction program 40 in the memory 20, the steps of the heterogeneous network-based multi-species protein function prediction method described above are implemented.

[0090] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a multi-species protein function prediction program based on a heterogeneous network, and when the multi-species protein function prediction program based on a heterogeneous network is executed by a processor, the steps of the multi-species protein function prediction method based on a heterogeneous network as described above are implemented.

[0091] In summary, the present invention provides a method, system, terminal and storage medium for predicting multi-species protein functions based on heterogeneous networks. The method comprises: using the ESM-2 model to extract features of protein sequences to obtain sequence features; constructing a structural contact graph of the protein, training the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extracting structural features; splicing the sequence features and the structural features to obtain splicing node features, constructing a cross-species heterogeneous network based on a PPI network and a homology similarity network, and inputting the splicing node features into the heterogeneous network; in the training stage, using the structural features and the sequence features to propagate on the heterogeneous network, and using a graph attention mechanism to update the node vector during propagation; in the prediction stage, adding the GO label of the training set for network propagation, linearly combining the propagation results of the protein representation and the GO label, and obtaining the final GO label prediction probability. In the field of protein function prediction of multiple species, the present invention combines structural features with multi-species PPI networks, adopts graph convolution and hierarchical graph pooling to process structural contact graphs, and uses a heterogeneous network propagation algorithm based on a graph attention mechanism to obtain low-dimensional representations of proteins from different levels, thereby improving the prediction effect and realizing cross-species propagation of functional labels.

[0092] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or terminal including the element.

[0093] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0094] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A multi-species protein function prediction method based on heterogeneous networks, characterized in that: The multi-species protein function prediction method based on heterogeneous networks includes: Use the ESM-2 model to extract features from protein sequences and obtain sequence features; Constructing a structural contact graph of the protein, training the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extracting structural features; Splicing the sequence features and the structural features to obtain splicing node features, constructing a cross-species heterogeneous network based on a PPI network and a homology similarity network, and inputting the splicing node features into the heterogeneous network; In the training phase, the structural features and the sequence features are used to propagate on the heterogeneous network, and a graph attention mechanism is used to update node vectors during propagation; In the prediction stage, the GO tags of the training set are added for network propagation, and the propagation results of protein representation and GO tags are linearly combined to obtain the final GO tag prediction probability; The step of constructing a structural contact graph of a protein, training the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extracting structural features specifically includes: The original data is obtained, and the original data is the three-dimensional atomic coordinates of the amino acid sequence, based on the distance between the Cα atoms being less than Connect the edges between amino acids to build the structural contact graph of the protein; If the current node is v, the probability of selecting x as the next node is: Among them, P(n i =x|n i-1 =v) indicates the probability that the next hop is x when the current node is v, n i represents the node selected by the i-th hop, n i-1 represents the node selected in the i-1th hop, π vx represents the probability of walking from vertex v to vertex x, Z represents the normalization constant, and E represents the edge set of the graph; If the last walk was to reach node v via node t, π vx The calculation is: p vx =a pq (t,x)·w vx ;(2) Among them, α pq (t,x) represents the influencing factor that determines whether to walk inward or outward. p and q represent hyperparameters. By setting hyperparameters, the influencing factor can be controlled to control the bias of the random walk strategy. w vx represents the weight of the (v, x) edge, d tx Represents the distance between node x and node t; Walk all nodes in the structural contact graph and generate an initial embedding vector for each amino acid node; The structural model includes a graph convolution pooling module and an MLP classifier, and the structural contact graph is trained based on the graph convolution pooling module and the MLP classifier to extract structural features; The structure contact graph is trained based on the graph convolution pooling module and the MLP classifier to extract the structural features, specifically including: The graph convolution pooling module performs graph convolution on the structural contact graph, aggregates the neighborhood information and then filters the nodes, maps the node features to one dimension to obtain the attention scores of the corresponding nodes, filters out some nodes with scores higher than the preset scores, and aggregates the features after graph convolution again according to the attention scores to obtain the aggregated subgraph, and performs graph pooling on the aggregated subgraph to obtain the output of the graph convolution pooling module of this layer; The outputs of the multi-layer graph convolutional pooling modules are superimposed to obtain the effective structural feature H st , H st A matrix representing the structural features of multiple proteins; The MLP classifier classifies the structural contact graph, trains the structural model by calculating the cross entropy loss between the probability of predicting the GO label and the true value, and extracts the 1024-dimensional features before the MLP classifier as the structural features input for the downstream prediction task.

2. The method for predicting multi-species protein functions based on heterogeneous networks according to claim 1, characterized in that: The ESM-2 model is used to extract features from protein sequences to obtain sequence features, specifically including: The ESM-2 model is used to extract features from protein sequences and obtain the sequence feature H se The ESM-2 model is used to capture the amino acid sequence evolution information about proteins and properties. Each protein sequence is represented as a 1280-dimensional feature vector after the ESM-2 model. Among them, H se A matrix representing multiple protein sequence feature vectors.

3. The method for predicting multi-species protein functions based on heterogeneous networks according to claim 2, characterized in that: The step of splicing the sequence feature and the structural feature to obtain a splicing node feature specifically includes: The sequence feature H se and structural features H st Perform splicing to obtain the splicing node features: H=H se |H st ;(4) Among them, H represents the splicing node feature, | represents the splicing operation; Map the splicing node features H into a low-dimensional vector space and continue mapping using the MLP classifier: in, represents the input of the first layer of MLP, σ represents the nonlinear activation function, W e and b e represents the learnable parameters of the linear layer, represents the output of the l-th layer MLP, R d represents the real number domain, d is the hidden dimension of the output, LN(·) represents the layer regularization function, represents the input of the l-th layer MLP, and represents the learnable parameters of the lth layer in the MLP.

4. The method for predicting multi-species protein functions based on heterogeneous networks according to claim 3, characterized in that: The heterogeneous network includes a PPI network and a homology similarity network; the homology similarity network is used to provide essential properties for proteins, and the PPI network is used to supplement information at the functional expression level for proteins; The heterogeneous network is represented as G = (V, E1, E2), where V represents all nodes in the network, E1 represents an edge in the PPI network, and E2 represents an edge in the homology similarity network; The PPI network is represented as G p =(V, E1), the homology similarity network is represented by G s =(V, E2); In the training phase, the structural features and the sequence features are used to propagate on the heterogeneous network, and the node vector is updated using a graph attention mechanism during propagation, specifically including: The graph attention mechanism is used to adjust the intensity of propagation between network nodes, and an attention score is learned for each edge of the node: Among them, α uv represents the attention score of protein node u to node v, a T and W t is a learnable weight, T represents a transposition operation, the | symbol represents a concatenation operation, and h u and h v Represent the node features of proteins u and v respectively, N u represents the neighbor set of u, i represents the neighbor set from N u The neighbor protein nodes selected in i represents the node characteristics of neighbor protein node i; Using the same graph attention parameters for the PPI network G p and homology similarity network G s After updating the weights between nodes, they are represented as A p , A s ∈R n×n , where A p , A s denote the attention score matrices in the PPI network and homology similarity network, respectively, and n denotes the total number of protein nodes; In the training phase, only protein features are introduced into the heterogeneous network propagation, and in the prediction phase, the network propagation of functional labels is added. The protein labels are represented as binary vectors: in, represents the propagation result on the PPI network at layer l, H (l) represents the input of the l-th layer network propagation, W (l) represents the weight, represents the propagation result on the l-th layer homology similarity network, H (l+1) represents the output of the l-th layer network propagation, combining different information from the similarity network and the PPI network, and They represent the node feature vector matrices after label propagation on the l-th layer PPI network and homology similarity network, represents the input of the l-th layer of label propagation, represents the output of the l-th layer label propagation, and Norm(·) represents L2 regularization; In the prediction phase, only the label vector is updated for the unlabeled test set. The label vectors of the remaining nodes are reset to Y after each layer of propagation. i , the final output after label propagation is expressed as Map the node features into the output vector space to represent the probability of the function label. Use a linear layer and a sigmoid function to calculate the binary cross entropy loss between the output probability and the true value. Use the binary cross entropy loss for back propagation to update the parameters: in, represents the mapping of node features into the label vector space, n represents the total number of protein nodes, c represents the dimension of the label vector, and H out Represents the output of the last layer of network propagation, W out and b out Represents the parameters that need to be learned in the current linear mapping layer, Loss represents the binary cross entropy loss, Y represents the true value label matrix of the protein, and ⊙ represents the element-by-element multiplication operation.

5. The method for predicting multi-species protein functions based on heterogeneous networks according to claim 4, characterized in that: In the prediction stage, the GO tags of the training set are added for network propagation, and the propagation results of the protein representation and the GO tags are linearly combined to obtain the final GO tag prediction probability, which specifically includes: In the prediction phase, the last layer of the network propagates the updated label vector and Perform a linear combination: in, represents the final GO tag prediction probability, c' represents the number of tags in the current gene ontology, Represents a parameter used to measure the importance of label vectors and protein features.

6. A multi-species protein function prediction system based on a heterogeneous network, wherein the multi-species protein function prediction system based on a heterogeneous network is applied to the multi-species protein function prediction method based on a heterogeneous network according to any one of claims 1 to 5, characterized in that: The heterogeneous network-based multi-species protein function prediction system comprises: A sequence feature extraction module is used to extract features from protein sequences using the ESM-2 model to obtain sequence features; A structural feature extraction module is used to construct a structural contact graph of the protein, train the structural contact graph based on a structural model of graph convolution and hierarchical graph pooling, and extract structural features; A feature splicing and heterogeneous network construction module, used to splice the sequence features and the structural features to obtain splicing node features, construct a cross-species heterogeneous network based on the PPI network and the homology similarity network, and input the splicing node features into the heterogeneous network; A training module, used for propagating the structural features and the sequence features on the heterogeneous network during a training phase, and updating the node vectors using a graph attention mechanism during propagation; The prediction module is used to add the GO tags of the training set for network propagation in the prediction stage, linearly combine the propagation results of protein representation and GO tags, and obtain the final GO tag prediction probability.

7. A terminal, characterized in that: The terminal includes: a memory, a processor, and a heterogeneous network-based multi-species protein function prediction program stored in the memory and executable on the processor. When the heterogeneous network-based multi-species protein function prediction program is executed by the processor, the steps of the heterogeneous network-based multi-species protein function prediction method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a multi-species protein function prediction program based on a heterogeneous network, and when the multi-species protein function prediction program based on a heterogeneous network is executed by a processor, the steps of the multi-species protein function prediction method based on a heterogeneous network as described in any one of claims 1 to 5 are implemented.

Images