Protein interaction prediction method based on anti-fact sampling knowledge distillation

Abstract

The invention relates to the field of protein interaction prediction, in particular to a protein interaction prediction method based on anti-fact sampling knowledge distillation. The method comprises the following steps: constructing an amino acid level protein map based on space and sequence relative positions of amino acids in proteins, carrying out map convolution operation on the map, and constructing a self-prediction task optimization feature extraction module by utilizing type information of the amino acids in the proteins; relationship-level knowledge distillation is adopted, so that relation-level knowledge which is more suitable for protein interaction is better transferred between a teacher GNN model and a student MLP model; in the knowledge distillation process, an anti-fact sampling strategy is adopted, and more reliable positive / negative samples are selected for nodes; and link prediction is performed on the features by using a target node updated by a student MLP model, a real label loss is constructed by using binary cross entropy, and a total loss is constructed by combining two kinds of relationship-level knowledge distillation losses.

Description

Technical Field

[0001] The present invention relates to the field of protein interaction prediction, and in particular to a protein interaction prediction method based on counterfactual sampling knowledge distillation. Background Art

[0002] Protein interaction is a core concept in the field of life sciences, involving the process of two or more protein molecules forming a protein complex through non-covalent bonds. This interaction constitutes the main component of the cellular biochemical reaction network and plays a very important role in all life activities. Therefore, predicting PPIs plays a very important role in drug design and disease treatment. In the past few decades, a large number of experimental methods for high-throughput PPI prediction have been proposed, such as fluorescence resonance energy transfer (FRET) and bimolecular fluorescence complementation (BiFC). However, laboratory determination methods for determining PPIs are very expensive and time-consuming, and are difficult to apply to large-scale PPI predictions. Therefore, there is a need for an efficient and effective computational method that can be used to predict unknown PPIs.

[0003] In recent years, deep learning technology has made rapid progress and has been successfully applied to PPI prediction tasks. Its general pattern is to obtain the initial representation of proteins through sequence-based or structure-based methods. Then the protein is regarded as a node in the PPI graph, and the GNNs message passing encoding mechanism is used to update the protein node representation. The two updated target protein node representations are spliced ​​as the input of the predictor to predict whether there is an interaction between them. The success of GNNs in PPI prediction tasks is attributed to its mechanism of aggregating the contextual information of neighbors around the node, but this also leads to GNNs' heavy reliance on neighbor acquisition mode and aggregation mode. During the aggregation process, the number of neighbor nodes will explode, so GNNs will cause a lot of time consumption in the training and inference stages. This makes it difficult to generalize the PPIs prediction method based on GNNs to a large number of PPI datasets.

[0004] Compared to graph structures, tabular data does not have complex graph topology and is very efficient to process. Therefore, compared to GNNs, MLPs that process tabular data consume very little time in the training and inference stages. In order to alleviate the problem of large computational complexity and low time efficiency of GNNs, some work has begun to study how to combine GNNs and MLPs to simultaneously obtain the effectiveness of GNNs and the efficiency of MLPs. To achieve this goal, some recent works have proposed the use of knowledge distillation technology to transfer the knowledge learned by teacher GNNs to student MLPs. However, these works all focus on node or graph-level tasks, and few studies study the link relationship between nodes, making it difficult to apply to PPI prediction tasks that focus more on link relationships.

[0005] At the same time, in the process of knowledge distillation, existing methods usually use random sampling to select positive / negative samples of anchor nodes, but the samples selected by this method, especially negative samples, are not very accurate. The reason is: on the one hand, some of the interactions that originally existed will be divided into the validation set and the test set, and these interactions are invisible during the training process. Therefore, the positive samples of the anchor nodes may be mistakenly sampled as negative samples; on the other hand, for those interaction relationships that have not been discovered or recorded, they may also be mistakenly marked as negative samples during the modeling process. The above incorrect markings will mislead model learning, thereby limiting the performance of the model.

[0006] In general, there are two main challenges in transferring the knowledge learned by GNNs from the PPIs network to MLP through knowledge distillation to adapt to the PPI prediction task: (1) How to explicitly extract the knowledge of protein-protein interaction relationships rather than node-level or graph-level representation knowledge. (2) During the distillation process, how to select reasonable positive / negative samples for anchor nodes to improve model performance. Therefore, how to accurately transfer relation-level knowledge during knowledge distillation is crucial. Summary of the invention

[0007] In view of the shortcomings of the existing technology, this paper proposes a protein interaction prediction method based on counterfactual sampling knowledge distillation. This method comprehensively considers the transfer of relational knowledge and the selection of more reliable samples, and aims to optimize the efficiency and accuracy of protein interaction prediction through knowledge distillation.

[0008] The technical solution of the present invention is as follows:

[0009] A protein interaction prediction method based on counterfactual sampling knowledge distillation, including:

[0010] A protein graph is constructed based on the spatial and sequence relative positions of amino acids in the protein; the features of amino acid nodes in the protein graph are updated using a graph convolution model, and a self-prediction task is constructed in the protein graph by masking some amino acid type information through random sampling to guide the optimization of the graph convolution model; the protein features are obtained using the optimized graph convolution model;

[0011] Based on the known initial features of proteins and the interaction relationship between proteins, a PPI graph is constructed, and the teacher GNN model is trained on the PPI graph. The trained teacher GNN model is used to convolve the nodes in the PPI graph to obtain the updated feature matrix H.

[0012] Use the student MLP model to update the protein node attributes and obtain the updated feature matrix

[0013] Based on the counterfactual sampling strategy, the feature matrix H obtained by the teacher GNN model is used to sample each node in the training set to construct a positive / negative sample set;

[0014] Based on relation-level knowledge distillation, the target sample and its corresponding positive / negative sample set are used to separate the feature matrix H and feature matrix H. The features in the model are used to transfer the relational knowledge of protein interactions from the teacher GNN model to the student MLP model, and the student MLP model is optimized to obtain the representation of the two target amino acid nodes.

[0015] The representations of the two target amino acid nodes obtained by the student MLP model are concatenated as the input of the predictor, and the output is the possibility of the existence of a connection between the two nodes. The true label loss is constructed through binary cross entropy, and the total loss of the model is constructed by combining the two knowledge distillation losses.

[0016] Further, the protein map construction step comprises:

[0017] Based on the basic information of amino acids, nodes in the protein graph are constructed. Amino acids are used as nodes, and 20-dimensional one-hot encoding representing the type information of amino acids is used as the initial feature representation of the nodes.

[0018] Based on the relative position of the amino acid in the sequence seq , spatial relative position relationship d radius , and K nearest neighbor relationship construction respectively construct sequence edge, space edge and KNN edge; there are amino acid node pairs with sequence edge The following relationship needs to be satisfied:

[0019]

[0020] in, Indicates the relative position of the amino acid pair in the sequence, f seq (·) returns the position of the amino acid node in the sequence, Indicates a preset threshold;

[0021] Similarly, the amino acid node pairs connecting the spatial edges The following conditions must be met:

[0022]

[0023] in, represents the relative position of the amino acid pair in space, f coor (·) Return to amino acid C α The three-dimensional coordinates of the atoms, is a pre-set threshold; at the same time, KNN is used to construct KNN edges for each amino acid node;

[0024] Based on the obtained node representation and the edge relationship between nodes, construct the protein graph G p =<V r ,E,R>, where V r Represents the amino acid node in the protein graph, E represents the edge set, and R represents the type of edge in the edge set E, which includes three types: sequence edge, spatial edge and KNN edge.

[0025] Furthermore, based on the obtained protein graph, a relational graph convolution model is used to perform feature update operations on the nodes, and the calculation formula is as follows:

[0026]

[0027] in, represents the feature of node v in the k+1 layer, σ represents the activation function, W k represents the weight matrix of the kth layer, represents the feature of node v at the kth layer, r∈R represents different edge relationships, Indicates that u is a neighbor node of node v under edge relationship r, represents the weight matrix of the kth layer under relation r, Represents the features of node u at the kth layer.

[0028] Furthermore, a self-prediction task is constructed by randomly sampling and masking part of the amino acid type information in the protein graph, and the steps are as follows:

[0029] Randomly sample a certain proportion of amino acid nodes in the protein graph and mask their category information; use the updated features of these amino acids to predict their categories, and use the cross entropy function as the loss function of the model to guide the optimization of the relationship graph convolution model; the calculation formula is as follows:

[0030]

[0031] Among them, L mask represents the masking loss of the self-prediction task, n represents the number of masked amino acids, CE(·) represents the cross entropy function, and h i represents the updated representation, f r (h i ) returns the value of node v i The predicted results of amino acid types, res(v i ) represents the actual type of the node;

[0032] Use the relational graph convolution model to obtain protein features x i, the steps are as follows:

[0033] The trained graph convolution model is used to convolve the protein graph to obtain an updated representation of the node features. The calculation formula is as follows:

[0034]

[0035] in, represents the n-row and d-column feature matrix of all amino acid nodes in the graph, n represents the number of amino acid nodes in the graph, d represents the node feature dimension, and f(·) represents the trained relational graph convolution model. represents the initial feature representation of the amino acid node;

[0036] Use the READOUT function to read the node feature matrix X res The protein graph is compressed and the calculation process is as follows:

[0037] x i =READOUT(X res );

[0038] in, represents the feature of protein i with dimension d, READOUT(·) represents the READOUT function, X res is the amino acid feature matrix in the protein graph.

[0039] Further, the teacher GNN model is trained based on the PPI graph, and the operation method is as follows:

[0040] Use the teacher GNN model to perform convolution operations on the PPI graph to update the representation of each protein node;

[0041] Splice two target nodes v i ,v j The updated features are used as the input of the predictor to predict the existence of links between target nodes. The calculation formula is as follows:

[0042]

[0043] in, Represents the teacher GNN model for node v i ,v j The predicted value of the existence of the link between them, F(·) represents the predictor, concate(·) represents the concatenation operation, Respectively represent the updated features of the two nodes at layer l;

[0044] After obtaining the predicted value of the link, the binary cross entropy is used as the loss of the teacher GNN model to guide the optimization of the teacher GNN model. The calculation formula is as follows:

[0045]

[0046] in, Represents the teacher GNN model for node v i ,v j The loss value of the link existence prediction value, y i,j Indicates v i ,v j The actual situation of the link between them. If it exists, the value is 1, otherwise it is 0. Represents the teacher model for node v i ,v j The predicted value of link existence;

[0047] The trained teacher GNN model is used to convolve the nodes in the PPI graph to obtain the updated representation H, which is calculated as follows:

[0048] H=GNN T (X);

[0049] Among them, H represents the feature matrix of the nodes in the PPI graph after being updated using the teacher GNN model, GNN T represents the trained teacher GNN model, and X represents the original feature matrix of the node in the PPI graph.

[0050] Furthermore, the counterfactual sampling strategy steps include:

[0051] Calculate the processing matrix T and similarity matrix S:

[0052]

[0053] Among them, c(v i ) and c(v j ) represents node v i and v j The community index, T i,j Represents node v i and node v j Whether they are in the same community, S i,j Represents node v i and node v j similarity;

[0054] Construct the treatment matrix T for the counterfactual situation CF and the adjacency matrix A CF ;

[0055]

[0056] Among them, γ is a hyperparameter, which represents the minimum value of the similarity between two pairs of nodes; at the same time, S i,aand S j,b The sum is used as the anchor node v i Sample v j The likelihood of being a sample:

[0057] p i,j = S i,a + S j,b

[0058] The node pair v i , v j In the counterfactual case, the values in the processing matrix and the adjacency matrix are defined as:

[0059]

[0060] Construct the candidate matrix M and the probability matrix P for the positive / negative samples for the anchor node v i Construct the candidate positive sample set

[0061]

[0062] Select k positive samples through the above steps and put them into as the k candidate positive sample sets for the anchor node v i ; at the same time, put the sampling likelihood p i,j of these k samples through softmax into P i pos as the final sampling probability for each candidate positive sample:

[0063]

[0064] The sampling process for the candidate negative samples of the anchor node v is:

[0065]

[0066] Select the final positive / negative sample set C i , and obtain M pos , P pos , M neg , P neg After that, during the training process, these four matrices are needed to select the final positive / negative samples for the anchor node v i ; first, randomly sample n (n < k) values according to the probability values in P i pos , and then find the corresponding index nodes in and to form the sample set C i for the anchor node v i .

[0067] Furthermore, the relation-level knowledge distillation includes a sorting-based matching strategy and a distribution-based matching strategy, and the calculation steps include:

[0068] The calculation formula based on the sorting matching strategy is as follows:

[0069]

[0070] Among them, L rank It represents the ranking matching loss, and δ is a margin hyperparameter;

[0071] The calculation method based on the sorting matching strategy is:

[0072]

[0073] Among them, L distribution It represents the distribution matching loss, and τ is a temperature hyperparameter used to control the softness of softmax.

[0074] Further, the step of combining the two knowledge distillation losses to construct the total loss of the model includes:

[0075] During the training process, the student MLP model obtains the node pair v i , characteristics of vj and Concatenate them together as the input of the predictor, and the output of the predictor is the predicted value of the link:

[0076]

[0077] Among them, concat(·|·) represents the concatenation operation. That is, the model for node v i and v j The final prediction result of the possibility of direct link existence;

[0078] In the process of training the student MLP, the overall loss function L used is as follows:

[0079] L=α·L sup +β·L rank +γ·L distribution ;

[0080] Among them, α, β, and γ control L sup , L rank , L distribution importance.

[0081] The beneficial effects of the present invention are as follows: the present invention enhances the prediction process of protein interaction by modeling protein graphs and using relational knowledge distillation methods, and at the same time, can effectively improve the time efficiency of prediction. First, since the function of protein is mainly determined by the structure of protein, the present invention can extract more effective protein features by taking into account the relative positions of amino acids in space and sequence in the protein into the modeling process of protein graphs; and constructs self-prediction tasks by randomly masking some amino acid types, thereby enhancing the pre-training process of proteins. Secondly, the present invention transfers the knowledge of protein interaction from a complex teacher GNN model to a simple student MLP model through knowledge distillation, thereby improving the prediction efficiency while ensuring the accuracy; at the same time, the existing knowledge distillation methods are mainly aimed at node-level or graph-level knowledge, and there is little edge-level knowledge applicable to protein interaction. Therefore, the present invention uses a relational distillation strategy that is more suitable for protein interaction, which is based on a ranking matching strategy and a distribution matching strategy, respectively, and can effectively transfer relational knowledge in the process of knowledge distillation. Finally, the present invention proposes to use a counterfactual node sampling strategy to sample the positive / negative samples of the target node, thereby ensuring the reliability of the positive / negative samples and further improving the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 Schematic diagram of the protein interaction prediction method for counterfactual sampling knowledge distillation;

[0083] Figure 2 The process of constructing a protein map;

[0084] Figure 3 This is a diagram of the relation-level knowledge distillation framework;

[0085] Figure 4 Schematic diagram of the counterfactual sampling process. DETAILED DESCRIPTION

[0086] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0087] See also Figure 1 , which shows a protein interaction prediction method based on counterfactual sampling knowledge distillation provided by an embodiment of the present invention, comprising:

[0088] S100: Construct a protein graph based on the spatial and sequence relative positions of amino acids in the protein; use a graph convolution model to update the features of amino acid nodes in the protein graph, and construct a self-prediction task in the protein graph by randomly sampling and masking some amino acid type information to guide the optimization of the graph convolution model; use the optimized graph convolution model to obtain protein features;

[0089] S200: construct a protein-protein interaction (PPI) graph based on the known initial features of proteins and the interaction relationship between proteins, train the teacher GNN model on the PPI graph, and use the trained teacher GNN model to convolve the nodes in the PPI graph to obtain the updated feature matrix H;

[0090] S300: Use the student MLP model to update the protein node attributes and obtain the updated feature matrix

[0091] S400: Based on the counterfactual sampling strategy, the feature matrix H obtained by the teacher GNN model is used to sample each node in the training set and construct a positive / negative sample set.

[0092] S500: Based on relational knowledge distillation, the target sample and its corresponding positive / negative sample set are used to calculate the feature matrix H and feature matrix H. The features in the ,relationship-level knowledge of protein interactions is transferred from the teacher GNN model to the student MLP model, and the student MLP model is optimized to obtain the representation of the two target amino acid nodes.

[0093] S600: The representations of the two target amino acid nodes obtained by the student MLP model are concatenated as the input of the predictor, and the output is the possibility of the existence of a connection between the two nodes. The true label loss is constructed through binary cross entropy, and the total loss of the model is constructed by combining the two knowledge distillation losses.

[0094] like Figure 2 As shown in the figure, S100 performs a series of operations around the protein graph. First, nodes are constructed based on the basic information of amino acids in proteins, and three types of edges are constructed using their relative positions in sequence and space and K nearest neighbor relationships, thereby constructing a protein graph. Then, the node features are updated using a graph convolution model according to a specific formula. After that, a self-prediction task is constructed by randomly sampling and masking some amino acid type information in the protein graph, and the model is optimized using the cross entropy function as the loss function. Finally, protein features are obtained with the help of the trained graph convolution model and the READOUT function. This series of steps lays the foundation for subsequent protein analysis.

[0095] The specific steps are as follows:

[0096] The n-row and d-column feature matrix is ​​, where n represents the number of amino acid nodes in the graph, d represents the node feature dimension, and f(·) represents the trained relational graph convolution model. Represents the initial feature representation of amino acid nodes.

[0097] Use the READOUT function to read the node feature matrix X res The protein graph is compressed and the calculation process is as follows:

[0098] x i =READOUT(X res );

[0099] in, represents the feature of protein i with dimension d, READOUT(·) represents the READOUT function, X res is the amino acid feature matrix in the protein graph.

[0100] like Figure 3 As shown in the figure, S200 focuses on PPI graph related operations. First, a PPI graph is constructed based on known proteins and their interaction relationships, and the nodes, link sets, adjacency matrix, and protein node feature matrix in the graph are determined. Then, the teacher GNN model is used to perform convolution operations on the PPI graph, and the existence of links is predicted by splicing the updated features of the target nodes, and the binary cross entropy is used as the loss function to guide model optimization. Finally, the trained teacher GNN model is used to convolve the PPI graph nodes again to obtain the updated node representation, preparing for the subsequent further processing of the information in the PPI graph.

[0101] The specific contents are as follows:

[0102] S210: Construct a PPI map based on the known initial features of proteins and the interaction relationship between proteins. The steps are as follows:

[0103] Given N proteins P = {P1, P2, ..., P N}, and the interaction between them, we can construct a PPI graph G = (V, E), where each node v i ∈V represents a protein P i ∈P, E∈V×V represents the set of observed links (interactions). Use A∈{0,1} N×N represents the adjacency matrix of the PPI graph G, where A i,j =1 indicates node v i and node v j There is a link between them, otherwise there is no link. The protein node feature matrix in the graph It is obtained through step 1, where N represents the number of protein nodes in the PPI graph and F represents the feature dimension of the protein node.

[0104] S220: Train the teacher GNN model based on the PPI graph. The specific steps are as follows:

[0105] A teacher GNN model is used to perform convolution operations on the PPI graph to update the representation of each protein node.

[0106] Splice two target nodes v i ,v j The updated features are used as the input of the predictor to predict the existence of links between target nodes. The calculation formula is as follows:

[0107]

[0108] in, Represents the teacher GNN model for node v i ,v j The predicted value of the existence of the link between them, F(·) represents the predictor, concate(·) represents the concatenation operation, They represent the updated features of the two nodes at the lth layer respectively.

[0109] After obtaining the predicted value of the link, the binary cross entropy is used as the loss of the teacher GNN model to guide the optimization of the teacher GNN model. The calculation formula is as follows:

[0110]

[0111] in, Represents the teacher GNN model for node v i ,v j The loss value of the link existence prediction value, y i,j Indicates v i ,v j The actual situation of the link between them. If it exists, the value is 1, otherwise it is 0. Represents the teacher model for node v i ,v j The predicted value of link existence.

[0112] S230: Use the trained teacher GNN model to convolve the nodes in the PPI graph to obtain the updated representation H, which is calculated as follows:

[0113] H=GNN T (X);

[0114] Among them, H represents the feature matrix of the nodes in the PPI graph after being updated using the teacher GNN model, GNN Trepresents the trained teacher GNN model, and X represents the original feature matrix of the node in the PPI graph.

[0115] like Figure 3 As shown, S300 uses the student MLP model to update the features of the nodes in the PPI graph and obtains the updated representation The calculation formula is as follows:

[0116]

[0117] in, Represents the updated feature matrix of the node obtained using the student MLP model, MLP S represents the student model, and X represents the original feature matrix of the nodes in the PP I graph.

[0118] like Figure 3 As shown, S400 is about graph mining and counterfactual sample construction. Using a specific algorithm, the feature matrix H obtained by the teacher GNN model is used to divide the nodes in the protein-protein interaction graph into different communities, and the processing matrix and similarity matrix are calculated based on the community situation. Complex calculations are performed on the existence of links between node pairs in counterfactual situations, involving operations such as finding similar node pairs and determining sampling possibilities, and then constructing positive / negative sample candidate matrices and probability matrices. Through a series of sampling and sample supplementation steps, the positive / negative sample sets used for training are finally obtained, providing a diverse data basis for model training.

[0119] Specific content includes:

[0120] S410: Calculate the processing matrix T and the similarity matrix S. Use a specific graph mining / clustering algorithm c to divide the nodes V in the protein-protein graph into different communities. Based on the community situation of the nodes, calculate the processing matrix T and the similarity matrix S:

[0121]

[0122] Among them, c(v i ) and c(v j ) represents node v i and v j The community index, T i,j Represents node v i and node v j Whether they are in the same community, S i,j Represents node v i and node v j There are many ways to measure the similarity of . Here we use cosine similarity, h i With h j Respectively represent the nodes v obtained by the teacher GNN model iand v j The feature representation of .

[0123] S420: Construct the processing matrix T for the counterfactual situation CF and the adjacency matrix A CF For node pair v i and v j , we can only observe whether the link between them exists in the factual situation, but we cannot know whether they exist in the counterfactual situation, that is, T i,j The opposite case is the case of a link. Therefore, we use the node pair v i , v j Processing T i,j The pair of nodes v with opposite and most similar features a , v b The existence of the link between them is the counterfactual case for the node pair v i , v j If the link exists, the specific calculation process is as follows:

[0124]

[0125] Among them, γ is a hyperparameter, which represents the minimum value of the similarity between two pairs of nodes. i,a and S j,b The sum is used as the anchor node v i Sampling v j As a sample of the likelihood size:

[0126] p i,j =S i,a +S j,b

[0127] S430: Node pair v i , v j The values ​​in the treatment matrix and the adjacency matrix in the counterfactual case are defined as:

[0128]

[0129] S440: Construct the candidate matrix M and probability matrix P of positive / negative samples. First, consider the case of positive samples. For anchor node v i Constructing candidate positive sample sets

[0130]

[0131] S450: Select k positive samples through the above steps and put them into As the anchor node v i At the same time, the sampling probability p of these k samples is i,j After softmax, put it into Pi pos as the final sampling probability for each candidate positive sample.

[0132]

[0133] If the number of positive samples that satisfy the counterfactual sampling constraint is less than k, the anchor node v is used i to supplement the candidate positive sample set with other neighbor nodes of v. Their corresponding sampling probabilities are equal, but less than the sampling probability of the counterfactual samples.

[0134] S460: The sampling process of the candidate negative samples of the anchor node v is similar to that of the positive samples:

[0135]

[0136] As Figure 4 shown, select the final positive / negative sample set C i . Obtain M pos , P pos , M neg , P neg . After that, during the training process, these four matrices are needed to select the final positive / negative samples for the anchor node v i . First, according to the probability values in P i pos , , randomly sample n (n < k) values, and then find the nodes corresponding to the indices in and to form the sample set C i of the anchor node v i .

[0137] As Figure 3 shown, S500 transfers relation-level knowledge between the teacher GNN model and the student MLP model by using the relation information between the target sample and each sample in its corresponding positive / negative sample set through relation-level knowledge distillation, which includes two relation-level knowledge distillation strategies, namely the ranking matching strategy and the distribution matching strategy. Its calculation steps include:

[0138] S510: For the node v i , using the updated feature representation matrix H obtained in step three, the teacher GNN model calculates the logit values of the connection possibility between each sample in the positive / negative sample set and the node v i to form the set Y i . At the same time, the student MLP also calculates the logit values of the connection possibility between each sample in the positive / negative sample set and the node v according to the updated feature representation matrix i obtained in step three, and forms the set Use the teacher GNN model to get the set Y i The ranking information of the logit values ​​guides the student MLP model to obtain the set The ranking of the logit values ​​in the equation makes the ranking of the predicted values ​​of the student MLP model as close as possible to that of the teacher GNN model. The calculation formula is as follows:

[0139]

[0140] Among them, L rank It represents the ranking matching loss, δ is a margin hyperparameter, which is usually a very small value. Introducing r = 0 in the formula, that is, |y i,a -y i,b |<δ, which enables the student MLP to effectively distinguish the tiny logit value differences caused by the noise of the teacher GNN.

[0141] S520: The sorting matching strategy is a coarse-grained sorting method. In order to make full use of numerical information, a distribution-based matching strategy is also adopted. Specifically, its purpose is to make the predicted value set of the student MLP model and the predicted value Y of the teacher GNN model i The distribution of logit values ​​is as close as possible to:

[0142]

[0143] Among them, L distribution It represents the distribution matching loss, and τ is a temperature hyperparameter used to control the softness of softmax.

[0144] like Figure 3 As shown, S600 involves MLP model prediction and total loss calculation. The MLP model obtains node pair features and then splices the input predictor to predict the possibility of link existence, and calculates the true label loss through binary cross entropy loss. When training the student MLP model, its overall loss function is mainly based on the true label loss, and at the same time combines the relation-level distillation loss based on the sorting matching strategy and the distribution matching strategy. The importance of each loss is controlled by hyperparameters. By optimizing the total loss function, the student MLP model continuously improves its performance in protein interaction prediction.

[0145] Specific content includes:

[0146] S610: During the training process, MLP can obtain the node pair v i , characteristics of vj and Concatenate them together as the input of the predictor, and the output of the predictor is the predicted value of the link:

[0147]

[0148] Among them, concat(·|·) represents the concatenation operation. That is, the model for node v i and v j The final prediction result of the direct link existence possibility is then used to obtain the true label loss using binary cross entropy loss.

[0149] S620: The total loss is mainly based on the true label loss, combined with the two relation-level distillation losses in Section 4.1. Therefore, in the process of training the student MLP, the overall loss function L used is as follows:

[0150] L=α·L sup +β·L rank +γ·L distribution ;

[0151] Among them, α, β, and γ control L sup , L rank , L distribution importance.

[0152] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0153] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0154] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0155] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features of the present invention.

[0156] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A protein interaction prediction method based on counterfactual sampling knowledge distillation, characterized in that: include: Construct a protein graph based on the spatial and sequence relative positions of amino acids in the protein; The graph convolution model is used to update the amino acid node features in the protein graph, and a self-prediction task is constructed in the protein graph by randomly sampling and masking some amino acid type information to guide the optimization of the graph convolution model. The protein features are obtained using the optimized graph convolution model. Based on the known initial features of proteins and the interaction relationship between proteins, a PPI graph is constructed, and the teacher GNN model is trained on the PPI graph. The trained teacher GNN model is used to convolve the nodes in the PPI graph to obtain the updated feature matrix H. Use the student MLP model to update the protein node attributes and obtain the updated feature matrix Based on the counterfactual sampling strategy, the feature matrix H obtained by the teacher GNN model is used to sample each node in the training set to construct a positive / negative sample set; Based on relation-level knowledge distillation, the target sample and its corresponding positive / negative sample set are used to separate the feature matrix H and feature matrix H. The features in the model are used to transfer the relational knowledge of protein interactions from the teacher GNN model to the student MLP model, and the student MLP model is optimized to obtain the representation of the two target amino acid nodes. The representations of the two target amino acid nodes obtained by the student MLP model are concatenated as the input of the predictor, and the output is the possibility of the existence of a connection between the two nodes. The true label loss is constructed through binary cross entropy, and the total loss of the model is constructed by combining the two knowledge distillation losses.

2. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 1, characterized in that: The protein map construction step comprises: Based on the basic information of amino acids, nodes in the protein graph are constructed. Amino acids are used as nodes, and 20-dimensional one-hot encoding representing the type information of amino acids is used as the initial feature representation of the nodes. Based on the relative position of the amino acid in the sequence seq , spatial relative position relationship d radius , and K nearest neighbor relationship construction respectively construct sequence edge, space edge and KNN edge; there are amino acid node pairs with sequence edge The following relationship needs to be satisfied: in, Indicates the relative position of the amino acid pair in the sequence, f seq (·) returns the position of the amino acid node in the sequence, Indicates a preset threshold; Similarly, the amino acid node pairs connecting the spatial edges The following conditions must be met: in, represents the relative position of the amino acid pair in space, f coor (·) Return to amino acid C α The three-dimensional coordinates of the atoms, is a pre-set threshold; at the same time, KNN is used to construct KNN edges for each amino acid node; Based on the obtained node representation and the edge relationship between nodes, construct the protein graph G p =<V r ,E,R>, where V r Represents the amino acid node in the protein graph, E represents the edge set, and R represents the type of edge in the edge set E, which includes three types: sequence edge, spatial edge and KNN edge.

3. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 2, characterized in that: Based on the obtained protein graph, the relationship graph convolution model is used to perform feature update operation on the nodes, and the calculation formula is as follows: in, represents the feature of node v in the k+1 layer, σ represents the activation function, W k represents the weight matrix of the kth layer, represents the feature of node v at the kth layer, r∈R represents different edge relationships, Indicates that u is a neighbor node of node v under edge relationship r, represents the weight matrix of the kth layer under relation r, Represents the features of node u at the kth layer.

4. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 2, characterized in that: In the protein graph, a self-prediction task is constructed by randomly sampling and masking part of the amino acid type information, and the steps are as follows: Randomly sample a certain proportion of amino acid nodes in the protein graph and mask their category information; use the updated features of these amino acids to predict their categories, and use the cross entropy function as the loss function of the model to guide the optimization of the relationship graph convolution model; the calculation formula is as follows: Among them, L mask represents the masking loss of the self-prediction task, n represents the number of masked amino acids, CE(·) represents the cross entropy function, and h i represents the updated representation, f r (h i ) returns the value of node v i The predicted results of amino acid types, res(v i ) represents the actual type of the node; Use the relational graph convolution model to obtain protein features x i , the steps are as follows: The trained graph convolution model is used to convolve the protein graph to obtain an updated representation of the node features. The calculation formula is as follows: in, represents the n-row and d-column feature matrix of all amino acid nodes in the graph, n represents the number of amino acid nodes in the graph, d represents the node feature dimension, and f() represents the trained relational graph convolution model. represents the initial feature representation of the amino acid node; Use the READOUT function to read the node feature matrix X res The protein graph is compressed and the calculation process is as follows: x i =READOUT(X res ); in, represents the protein i feature with dimension d, READOUT() represents the READOUT function, X res is the amino acid feature matrix in the protein graph.

5. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 1, characterized in that: The operation method of training the teacher GNN model based on the PPI graph is as follows: Use the teacher GNN model to perform convolution operations on the PPI graph to update the representation of each protein node; Splice two target nodes v i ,v j The updated features are used as the input of the predictor to predict the existence of links between target nodes. The calculation formula is as follows: in, Represents the teacher GNN model for node v i ,v j The predicted value of the existence of the link between them, F(·) represents the predictor, concate(·) represents the concatenation operation, Respectively represent the updated features of the two nodes at layer l; After obtaining the predicted value of the link, the binary cross entropy is used as the loss of the teacher GNN model to guide the optimization of the teacher GNN model. The calculation formula is as follows: in, Represents the teacher GNN model for node v i ,v j The loss value of the link existence prediction value, y i,j Indicates v i ,v j The actual situation of the link between them. If it exists, the value is 1, otherwise it is 0. Represents the teacher model for node v i ,v j The predicted value of link existence; The trained teacher GNN model is used to convolve the nodes in the PPI graph to obtain the updated representation H, which is calculated as follows: H=GNN T (X); Among them, H represents the feature matrix of the nodes in the PPI graph after being updated using the teacher GNN model, GNN T represents the trained teacher GNN model, and X represents the original feature matrix of the node in the PPI graph.

6. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 1, characterized in that: The counterfactual sampling strategy steps include: Calculate the processing matrix T and similarity matrix S: Among them, c(v i ) and c(v j ) represents node v i and v j The community index of T i,j Represents node v i and node v j Whether they are in the same community, S i,j Represents node v i and node v j similarity; Construct the treatment matrix T for the counterfactual situation CF and the adjacency matrix A CF ; Among them, γ is a hyperparameter, which represents the minimum value of the similarity between two pairs of nodes; at the same time, S i,a and S j,b The sum is used as the anchor node v i Sampling v j As a sample of the likelihood size: p i,j =S i,a +S j,b Node pair v i , v j The values ​​in the treatment matrix and the adjacency matrix in the counterfactual case are defined as: Construct the candidate matrix M of positive / negative samples and the probability matrix P for the anchor node v i Constructing candidate positive sample sets Through the above steps, k positive samples are selected and put into As the anchor node v i k candidate positive sample sets; at the same time, the sampling probability p of these k samples i,j After softmax, put , as the final sampling probability of each candidate positive sample: The sampling process of candidate negative samples of anchor node v is: Select the final positive / negative sample set C i , to obtain M pos , P pos , M neg , P neg After that, during the training process, these four matrices need to be used to select samples for the anchor node v i Select the final positive / negative samples; first, randomly sample n (n < k) values according to the probability values in , and then find the nodes with the corresponding indices in and to form the sample set C i of the anchor node v i .

7. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 1, characterized in that: The relation-level knowledge distillation includes a sorting-based matching strategy and a distribution-based matching strategy, and its calculation steps include: The calculation formula based on the sorting matching strategy is as follows: Among them, L rank It represents the ranking matching loss, and δ is a margin hyperparameter; The calculation method based on the sorting matching strategy is: Among them, L distribution It represents the distribution matching loss, and τ is a temperature hyperparameter used to control the softness of softmax.

8. The protein interaction prediction method based on counterfactual sampling knowledge distillation according to claim 1, characterized in that: The step of combining the two knowledge distillation losses to construct the total loss of the model includes: During the training process, the student MLP model obtains the node pair v i , v j Features and Concatenate them together as the input of the predictor, and the output of the predictor is the predicted value of the link: Among them, concat(·|·) represents the concatenation operation. That is, the model for node v i and v j The final prediction result of the possibility of direct link existence; In the process of training the student MLP, the overall loss function L used is as follows: L=α·L sup +β·L rank +γ·L distribution ; Among them, α, β, and γ control L sup , L rank , L distribution importance.

Images