A Method for Predicting Biomolecular Interactions Based on a High-Order Filtered Graph Convolutional Attention Network

Abstract

The present invention belongs to the field of intelligent cell biometric recognition, and particularly relates to a method for predicting biomolecular interactions based on a high-order filtered graph convolutional attention network. The method includes five stages: extraction of basic structural features of a biomolecular interaction network, extraction of high-order structural features of a biomolecular interaction network, mixing of basic structural features and high-order structural features, a feature filtering network, and a link predictor. In this method, the basic structural features and high-order structural features of a biomolecular interaction network are extracted separately first, and then the two structural features are mixed through the CONCAT mechanism, thus improving the extraction of structural features of a biomolecular interaction network. In the feature filtering network, the present method proposes a feature filtering mechanism of LN+SE+LSTM (normalization layer + attention layer + long short-term memory network), which successfully reduces the noise of high-order adjacent nodes.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent cell biometrics, and particularly relates to a method for predicting biomolecular interactions based on a high-order filtered graph convolutional attention network. Background Art

[0002] The prediction of biomolecular interactions is an interdisciplinary field of bioinformatics, biophysics, computer science, etc., aiming to predict and interpret the interactions between biomolecules through computational inventions and models. It is the basis for understanding and mastering the intermolecular interaction relationships in life sciences and has important theoretical and application values.

[0003] Specifically, the prediction of biomolecular interactions has important application values in the following aspects:

[0004] Drug R & D: The prediction of biomolecular interactions can be used to screen compound libraries, help find appropriate potential drug molecules, and can optimize the design of drug molecule structures.

[0005] Biological Research: In physiological and pathological processes, many details of biomolecular interactions are unknown to the present invention. The prediction of biomolecular interactions can help further understand important signal transduction mechanisms, cell recognition, antibody responses and other biological processes in life sciences.

[0006] Disease Diagnosis and Treatment: Biomolecular interactions are closely related to the occurrence and development of many diseases. By predicting the interactions between biomolecules, the mechanism of disease occurrence can be more accurately grasped and new drugs can be discovered.

[0007] Biological Resource Development: The prediction of biomolecular interactions can help discover new functional molecules, explore natural drugs, etc., and open up new biological resources.

[0008] In summary, the prediction of biomolecular interactions has wide applications in life sciences, computer science, drug R & D, etc., and is an important invention for studying intermolecular interactions.

[0009] Currently, the inventions for predicting biomolecular interactions include:

[0010] Molecular Docking: By computationally analyzing the interaction patterns between molecules and the energy of molecular binding, the most likely molecular binding pattern is found. Therefore, molecular docking is one of the most commonly used inventions for predicting intermolecular interactions.

[0011] Molecular Dynamics Simulation: This invention can explore the dynamic changes between molecules and their binding modes, and can simulate the interactions between biomolecules on a relatively long time scale.

[0012] Machine learning-based inventions: Such inventions make predictions according to different feature extraction methods and classifiers, and can train a model with prediction functions based on known binding information.

[0013] Conformation- and sequence-based inventions: Based on known molecular structure and sequence information, calculate, compare, and identify key features to predict possible intermolecular interactions.

[0014] Biological network-based inventions: Such inventions utilize biological network structures, represent biomolecules with nodes and edges, and predict intermolecular interactions by analyzing the features and topological structures of the network.

[0015] In addition, the research on predicting biomolecular interactions still faces a series of challenges and limitations:

[0016] Difficulty in ensuring accuracy: Due to the complexity of biomolecular interactions, innovative computational inventions or new data may not necessarily bring sufficient improvement, resulting in the accuracy of prediction results being much lower than experimental results.

[0017] Low data quality: The accuracy of predicting biomolecular interactions depends on the input data. However, some data may be noisy or have other limitations and are unlikely to be accurate. At the same time, biomolecular data may also have very cumbersome formats and may be incompatible between different databases and computer software.

[0018] Lack of a large amount of high-quality experimental data: Although many models for predicting biomolecular interactions have been improved, they still rely on experimental data in the real world to verify their prediction capabilities. However, in practical applications, these experimental data are very limited, resulting in possible biases in the models.

[0019] Diversity issues: The heterogeneity of biomolecules is very significant. In the same biological process, they may interact in multiple ways. Therefore, it is difficult for prediction models to handle such diverse systems.

[0020] In summary, in-depth research on predicting biomolecular interactions requires attention to the improvement of properties and data quality, as well as the appropriate handling of sensitive and diverse systems. Summary of the Invention

[0021] A method for predicting biomolecular interactions based on graph neural networks, from GCN (Graph Convolutional Network) to SkipGNN (Jump Graph Convolutional Network), demonstrates the importance of aggregating high-order adjacent nodes to update the information of biomolecular nodes. HOGCN (High-Order Graph Convolutional Network) introduces MixHop (a high-order graph convolutional network with sparse neighborhood mixing) to aggregate high-order adjacent nodes. However, as the order of aggregating adjacent nodes in the graph convolutional network deepens, noise is inevitable, and it is not advisable to directly aggregate the information of high-order adjacent nodes completely. First, the present invention proposes a basic structure feature extraction module and a high-order feature extraction module, which are respectively used to extract the basic structure features and high-order structure features of the biomolecular interaction network, and then fuse the two features, enabling it to better take into account the basic structure information and high-order structure information of the biomolecular interaction network. Second, the present invention proposes a new LN+SE+LSTM feature filtering mechanism (LayerNorm+Squeeze and Excitation networks+Long short-term memory), which can effectively filter the noise generated during the process of extracting high-order structure features. Finally, by inputting high-order structure features at different levels into the filtering mechanism, the present invention enables it to perform hierarchical processing on the features at different levels obtained by the high-order graph convolutional network. Compared with traditional methods such as MixHop (a high-order graph convolutional network with sparse neighborhood mixing) that directly and blindly splice the structural features at each level, the high-order structural features extracted by this method are more effective in updating the features of target nodes.

[0022] The technical solution of the present invention is as follows:

[0023] A method for predicting biomolecular interactions based on a high-order filtered graph convolutional attention network, the method includes five stages: extraction of basic structure features of the biomolecular interaction network, extraction of high-order structure features of the biomolecular interaction network, mixing of basic structure features and high-order structure features, a feature filtering network, and a link predictor. The five stages are as follows:

[0024] The first stage: the stage of extracting basic structure features of the biomolecular interaction network. This stage includes four steps, namely the construction of initial features of the biomolecular interaction network, and using one layer of GCN (Graph Convolutional Network) to extract the basic structure features of the biomolecular interaction network. The specific steps are as follows:

[0025] The present invention first uses one layer of traditional GCN (Graph Convolutional Network) to extract basic features. The core idea of GCN (Graph Convolutional Network) is to learn the basic representation information of biomolecules in the network structure by repeatedly aggregating the information of adjacent nodes directly connected to the target node, so as to extract the basic features of the network. The detailed extraction process is as follows:

[0026] First, based on the initialized feature matrix and the normalized adjacency matrix One layer of GCN (Graph Convolutional Network) can aggregate the node features directly connected to the target node; then, a linear transformation is performed through the weight matrix W and the result is output through a non-linear activation function. The calculation formula of the above one-layer graph convolution can be expressed as:

[0027]

[0028] where, H (0) is the feature initialized for the biomolecular interaction network. Generally, the nodes of the biomolecular interaction network are initialized by node embedding, such as DeepWalk, Node2vec. In this case, Node2vec is used to generate H (0) . is the adjacency matrix after normalization by the degree matrix, I is the identity matrix, D is the degree matrix and D |V| = ∑ ii A j A ij . W is the trainable weight matrix. σ(·) is the activation function, and ReLU(·) = max(0, ·) is used in the present invention. H (base) is the basic structural feature extracted by one layer of GCN (Graph Convolutional Network).

[0029] The specific steps of this stage are as follows:

[0030] Step 1: Use Node2Vec to initialize the node features of the original biomolecular interaction network H (0) .

[0031] Step 2: Use GCN (Graph Convolutional Network) to extract the basic structural feature H (base) .

[0032] Second stage: Use the high-order feature extraction module to extract the high-order structural features of the biomolecular interaction network. In the present invention, two layers of high-order graph convolution (MixHop (High-order Graph Convolutional Network with Sparse Neighborhood Mixing)) are established to extract high-order features. Specifically, the initialized feature H (0) obtained from Node2vec needs to pass through adjacency matrices of different orders in sequence. Then, the trainable weight matrix of the first layer is multiplied by in sequence. Finally, the structural features of different orders obtained are concatenated to complete the extraction of the features of high-order adjacent nodes. The above high-order graph convolution process can be expressed by the following formula:

[0033]

[0034] Among them, P represents the set of orders for aggregating higher-order adjacent nodes. In this case, P = {0, 1, 2, 3}, represents the adjacency matrix multiplied j times, and ||* represents concatenation in the vertical dimension of the matrix. Compared with traditional GCN (Graph Convolutional Network) and SkipGNN (Skip Graph Convolutional Network), traditional GCN (Graph Convolutional Network) only considers the first-order adjacency matrix information, that is, setting P = {1}; similarly, SkipGNN (Skip Graph Convolutional Network) considers the adjacency information of, that is, setting P = {1, 2}; the dimensions set for the weight matrices of different orders are the same. contains the features of adjacent nodes of four orders, which are the high-order structural features extracted by the first layer of MixHop (High-Order Graph Convolutional Network for Sparse Neighborhood Mixing).

[0035] After extracting high-order structural features through one layer of MixHop (High-Order Graph Convolutional Network for Sparse Neighborhood Mixing) with an order of 4, the second layer of MixHop (High-Order Graph Convolutional Network for Sparse Neighborhood Mixing) will perform hierarchical processing on the high-order structural features. After obtaining it will then pass through the adjacency matrices of different orders respectively to extract deeper high-order information. are the trainable weight matrices of different orders for the second layer of MixHop (High-Order Graph Convolutional Network for Sparse Neighborhood Mixing).

[0036] The specific steps of this stage are as follows:

[0037] Step 1: Use one layer of MixHop (High-Order Graph Convolutional Network for Sparse Neighborhood Mixing) with an order of 4 to extract the high-order features of the first layer

[0038] Step 2: Input the high-order features extracted in the first layer into the second layer of MixHop (High-Order Graph Convolutional Network for Sparse Neighborhood Mixing) respectively, but the high-order features of different orders are not concatenated and will be processed in subsequent stages.

[0039] The third stage: The hybrid feature module of the present invention mixes the basic structural features with the high-order structural features. The present invention fuses the basic structural features and the high-order structural features, and the basic structural features and the high-order structural features play different roles in updating the target node features. Specifically, the basic structural features provide the directly connected structural information on the first-order neighbors, and the high-order structural features contain the global structural information of the high-order neighbors. The above two kinds of structural information complement each other. Therefore, mixing these two kinds of information is very effective in improving the prediction performance of the present invention. The features extracted by the first-layer high-order graph convolution first pass through the adjacency matrices of different orders and then pass through the weight matrix to output the second-layer hierarchical high-order structural features, which are concatenated with the basic structural feature H (base) through the CONCAT mechanism. The calculation formula of the hybrid feature is as follows:

[0040]

[0041] The fourth stage: The present invention uses the mechanism of LN+SE+LSTM (LayerNorm+Squeeze and Excitation networks+Long short-term memory) to filter the noise in the high-order structural features. In the present invention, a new feature filtering network including LN, SE, and LSTM is proposed to alleviate the problem of noise generation by the model aggregating high-order adjacent nodes. First, the LN module is used to normalize the mixed features, so that the SE and LSTM modules can better process the mixed features. Then, the SE module uses the adjacent channels to filter the noise in the high-order adjacent nodes. Finally, the LSTM module filters the noise in the high-order neighbors through the serialized part formed by the first-order adjacent nodes and the high-order adjacent nodes. The detailed process is as follows:

[0042] The present invention represents the input mixed features as a matrix L is the feature dimension output by the graph convolution layer. In module b (LN layer), the mean value of M j is taken. The process of the LN layer re-centering and re-scaling M j to is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] Among them, μ and are the mean and standard deviation of the input M j , and ⊙ is the dot product operation. The bias b and the gain g are trainable parameters that ensure that the original data information is not destroyed during the normalization process.

[0048] The normalized features are input into the SE module to reassign the weights of the features between channels. First, the compression function compresses the extracted features, and the compressed features are input into the global average pooling layer to obtain global features. Second, two fully connected layers are used to expand the global features. Finally, through the Sigmoid activation function, the features between channels can be learned. Therefore, the channel attention mechanism enables the target node to focus on the adjacent nodes containing more information.

[0049] Specifically, the present invention expands the normalized features into where C = N is the number of adjacent channels and Y = 1 is the expanded dimension. The learning process of the channel features is as follows:

[0050] First, the parameter zc representing the feature relationship within the channel is calculated by compressing the global features:

[0051]

[0052] Then, the variable s representing the dependence relationship between channels and the reweighted channel features are calculated as follows:

[0053]

[0054]

[0055] where F sq is the compression function and and the excitation function F ex can effectively learn the connection between channels. The expansion function F scale can integrate channel information. The weight matrices W1 and W2 store the parameters of the interaction relationship between channels. The final output of module c is composed of the channel features combined as

[0056] After the adjacent channel features are reset, the LSTM is used to filter the high-order features, which is similar to GraphSAGE using the LSTM aggregator to aggregate features. The mixed base and feature and high-order features contain a sequence composed of first-order neighbors and high-order neighbors. The introduction of LSTM can effectively extract sequence features and filter noise.

[0057] The main structure of LSTM is the internal information processing setting, including three gate structures, namely the input gate, the forget gate, and the output gate. The forget gate is responsible for attenuating the main input. Then, the input gate controls the "supply size" to supplement energy to the main input to generate a new main input. This process of attenuation and supplementation completes the update of the main input. Finally, a new output is generated under the control of the output gate. The detailed process is as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] The three gates of LSTM all use the Sigmoid function as the activation function; represents the main output. Here When LSTM generates candidate memories, the hyperbolic tangent function tanh is used as the activation function. The part where the main input is forgotten is The input gate parameter is I i and the source parameter of the supplement is The output gate parameter is O i . The part where the gate line controls the supplement is

[0065] In essence, filtering is to extract better features to reduce the interference of noise. The establishment of the LSTM layer can effectively filter out the noise of higher-order neighbors. The features of different levels output by the filtering network are F0, F1, F2, and F3.

[0066] The specific steps of this stage are as follows:

[0067] Step 1: Use LN to normalize the mixed feature M j to

[0068] Step 2: Use the SE module to reset the adjacent channel features of to

[0069] Step 3: Use the LSTM module to process the sequential part in the features.

[0070] Phase 5: The present invention inputs the filtered node features into the link predictor to output the predicted numerical value of the edge. In this section, the node features will be transformed into edge features and finally a predicted numerical value of the edge will be output. The link predictor consists of a bilinear new transformation layer and two fully connected layers. The detailed process is as follows:

[0071] First, the features F0, F1, F2, and F3 of the filtered adjacent nodes with four different orders are concatenated to obtain the final feature Z. Secondly, according to the index numbers of the nodes in the dataset, the feature information of two corresponding nodes is extracted and input into the bilinear layer to obtain the feature information of the edge between the two nodes. Finally, the final predicted numerical value is output through two fully connected layers. For nodes v i and v j , the features z i and z j extracted through the network are input into the bilinear transformation layer, activated by the ELU function, and the output represents the interaction information e ij of the edge. The calculation process is as follows:

[0072]

[0073] Then, the lower-dimensional predicted information e ij output by the bilinear layer is reduced in dimension through two fully connected layers, and a predicted numerical value p ij is output. The calculation process is as follows:

[0074] p ij = sigmoid(FC2(ELU(FC1(e ij ))))

[0075] p ij The numerical value of represents the probability that there is an edge between the two target nodes. According to the preset threshold, generally set to 0.5, if it is greater than 0.5, there is an interaction, otherwise there is no interaction.

[0076] Advantages of the present invention:

[0077] (1) The present invention proposes a basic structure feature extraction module and a high-order feature extraction module, which are respectively used to extract the basic structure features and high-order structure features of the biomolecular interaction network, and then fuse the two features, which enables it to better take into account the basic structure information and high-order structure information of the biomolecular interaction network.

[0078] (2) The present invention proposes a new feature filtering mechanism (LayerNorm + Squeeze and Excitation networks + Long short-term memory), which can effectively filter the noise generated during the process of extracting high-order structure features.

[0079] (3) By inputting high-order structural features of different levels into the filtering mechanism, the present invention enables it to perform hierarchical processing on features of different levels obtained by the high-order graph convolutional network. Compared with traditional methods, the high-order structural features extracted by the present invention are more effective in updating the features of target nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 is the framework diagram of the present invention (Part A: basic feature extraction module; Part B: high-order feature extraction module; Part C: hybrid feature extraction module; Part D: filtering network; Part E: link predictor);

[0081] Figure 2 is the diagram of basic structural feature extraction of the present invention based on the traditional GCN (Graph Convolutional Network);

[0082] Figure 3 is the diagram of high-order feature extraction of the present invention based on high-order graph convolution;

[0083] Figure 4 is the diagram of hierarchical processing of high-order structural features of the present invention;

[0084] Figure 5 is the fusion of basic features and high-order features of the present invention;

[0085] Figure 6 is the diagram of the feature filtering mechanism of the present invention;

[0086] Figure 7 is the structural diagram of the link predictor of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0087] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0088] As Figures 1 to 7 shown, the present invention implements a method for predicting biomolecular interactions based on a high-order filtered graph convolutional attention network, and its architecture is as Figure 1As shown in the figure. The invention first uses a layer of GCN (Graph Convolutional Network) to extract direct basic structure features; secondly, uses two layers of MixHop (High-order Graph Convolutional Network with Sparse Neighborhood Mixing) to extract high-order structure features; thirdly, after the information propagation of two layers of MixHop (High-order Graph Convolutional Network with Sparse Neighborhood Mixing), the features of directly connected adjacent nodes are mixed with the features of high-order adjacent nodes. The reason for mixing the basic adjacent features with the high-order features is that the basic features contain most of the information required for updating the target node. However, MixHop (High-order Graph Convolutional Network with Sparse Neighborhood Mixing) sets the same weight for neighbors at each level, and because there is noise in the high-order neighbors, the invention uses the basic features as the filtering criterion; then, in order to reduce the noise pollution brought by the high-order adjacent features, the invention proposes a high-order feature filtering mechanism, that is, the LN+SE+LSTM (LayerNorm+Squeeze and Excitation networks+Long short-term memory) mechanism. That is, the regularization layer+attention layer+long short-term memory network. After splicing the information features of directly connected adjacent nodes with the information features of high-order adjacent nodes at different distances, this mechanism introduces the SE (Squeeze and Excitation networks) module to reset the channel features, and then uses the forgetting gate mechanism of LSTM to filter out some features that are not conducive to updating the target node, successfully alleviating the noise problem brought by aggregating high-order adjacent nodes. Finally, the features after filtering the noise are spliced, and the information of the edges between nodes is formed through the bilinear transformation module, and then the predicted values of the edges between nodes are output through the fully connected layer.

[0089] Example 1

[0090] This part explores the prediction performance of the method L3 for comparing network similarity of the present invention, the network embedding method DeepWalk, Node2vec, and the graph convolutional network methods GCN (Graph Convolutional Network), VGAE (Graph Autoencoder), SkipGNN (Skip Graph Convolutional Network), HOGCN (High-order Graph Convolutional Network) on four different biomolecular interaction datasets. The results of the present invention and the comparative inventions on the four datasets are shown in Table 1. As shown in the table, the present invention has achieved the best results on all four datasets among all the inventions.

[0091] On the DTI dataset, compared with the network embedding method DeepWalk, the present invention achieved improvements of 20.7% and 22.2% in AURPC and AUROC respectively, greatly making up for the deficiency of network embedding methods in extracting structural information. On the PPI dataset, compared with L3 which focuses on local structural information, the present invention achieved improvements of 3.7% and 7.3% in AURPC and AUROC respectively, indicating that considering global information shows great advantages in node feature extraction. On the GDI dataset, since the traditional GCN (Graph Convolutional Network) only aggregates adjacent nodes at a distance of 1, the present invention achieved improvements of 3.3% and 3.1% in AURPC and AUROC respectively in aggregating nodes at a farther distance. On the DDI dataset, compared with HOGCN (Higher-Order Graph Convolutional Network), the present invention achieved improvements of 7.3% and 6.0% in AURPC and AUROC respectively. This shows that compared with HOGCN (Higher-Order Graph Convolutional Network) which completely aggregates adjacent nodes at higher-order distances, the present invention can better capture the representation information of target nodes after filtering out noise nodes.

[0092] The above results also show that: compared with traditional network similar inventions, the present invention greatly improves the ability to extract structural features of biomolecular interaction networks through the application of graph convolution and higher-order graph convolution, and gets rid of the characteristics of being insensitive to structural similarity such as the PPI dataset. Compared with network embedding inventions, the present invention avoids the instability of the weights of target nodes to adjacent node features brought by random walk inventions. Compared with previous graph convolution and higher-order graph convolution inventions, the present invention makes the relationship between the features of directly connected adjacent nodes and the features of higher-order adjacent nodes more specific, and to a certain extent reduces the noise generated by higher-order adjacent nodes to the feature update of target nodes; for the features of directly connected adjacent nodes, the present invention increases the weight of directly connected adjacent nodes by splicing them with higher-order features of different levels multiple times.

[0093] Based on AURPC and AUROC, the present invention added a comparison of the F1_score index. Although the performance of the present invention in processing the GDI dataset with more nodes is not much improved compared with HOGCN (Higher-Order Graph Convolutional Network), in the experiment of predicting the existence of interactions with a set threshold of 0.5, the present invention performs better in F1 score, making up for the deficiency of HOGCN (Higher-Order Graph Convolutional Network) in F1 score. The improvement of the F1 score value reflects the improvement of the present invention in the quality of interaction prediction.

[0094] Table 1 Performance comparison of different methods

[0095]

[0096] Example 2

[0097] This section will explore the robustness of the proposed invention when faced with sparse biomolecular interaction networks. Currently, most of the interaction information in biomolecular interaction networks is sparse. Especially with the development of scientific research and the increase in the number of biological nodes, the interaction network is becoming even sparser. Therefore, it is necessary to study the robustness of the interaction prediction model for networks with a high degree of sparsity. To this end, the present invention explores the performance of the present invention in terms of the prediction metric AUPRC by gradually and uniformly reducing the number of connections of the nodes in the network, that is, making the connections of the graph gradually sparser. In the experiment, the present invention increased the ratio of training edges in the training set from 10% to 70% in sequence, with an increase of 20% at each level, and kept the test set unchanged. Specifically, the present invention compared the performance of the present invention with three benchmark inventions, namely HOGCN (Higher-Order Graph Convolutional Network), SkipGNN (Skip Graph Convolutional Network), and GCN (Graph Convolutional Network). The experimental results are shown in Table 2.

[0098] As shown in Table 2, the present invention shows better robustness than several other inventions on four different interaction datasets. Even under the setting of only containing 10% interaction information, the present invention can still maintain a high AUPRC metric.

[0099] Table 2 Comparison of the robustness of different methods

[0100]

[0101] Example 3

[0102] The ultimate goal of the research on biomolecular interaction networks is to be able to predict possible interactions through known interaction information, so as to provide experimental references for biological scientists. Therefore, in addition to observing the prediction accuracy of known interactions, the analysis of the prediction ability of unknown interactions is also very meaningful. This section conducts such an analysis. Taking the DDI dataset as an example, the present invention selects the top 10 pairs of drug molecule interactions with the highest interaction prediction scores given by the present invention except for the existing interactions in the dataset BioSNAP-DDI. For these results, the present invention verifies the results by comparing them in the DrugBank database.

[0103] As can be seen from Table 3, corresponding evidence for all 10 newly predicted interactions by the present invention can be found in DrugBank, which further demonstrates the effectiveness of the present invention.

[0104] Table 3 Verification results in DrugBank of the top 10 newly predicted DDI interactions by the present invention

[0105]

Claims

1. A method for predicting biomolecular interactions based on a high-order filtered graph convolutional attention network, characterized in that the steps are as follows: The first step: Initialize the node features in the biomolecular interaction network using the node embedding method Node2vec; The second step: Use one layer of GCN to extract the basic structural features of the biomolecular interaction network; The third step: Use two layers of MixHop to extract the high-order structural features of biomolecular interactions. The first layer is a 4th-order MixHop, and the second layer is a hierarchical 4th-order MixHop; The fourth step: Mix the obtained basic structural features and high-order structural features through the CONCAT splicing mechanism; The fifth step: Input the mixed features into a new feature filtering network LN+SE+LSTM, that is, the regularization layer + attention layer + long short-term memory network; After the mixed features are input into the filtering module, they first pass through the LN regularization layer to normalize the mixed features, then are input into the SE attention layer module to reset the feature weights of adjacent nodes, and finally are input into the LSTM long short-term memory network for first-order adjacent node and high-order adjacent node serialization filtering, and finally output features F0, F1, F2 and F3 of different orders; The sixth step: Concatenate the filtered hierarchical features and then input them into the link predictor module, which consists of a bilinear transformation layer and a fully connected layer. The final output of the link predictor module is the predicted value.

2. The biomolecular interaction prediction method based on the high-order filtered graph convolutional attention network according to claim 1, wherein: The specific operations of the second step are as follows: the initialized 128-dimensional features of the node embedding are input into the GCN encoding layer, where N represents the number of molecular nodes; the first-order adjacent node information is aggregated, and the basic structural features H of 32 dimensions are output (base) .

3. The biomolecular interaction prediction method based on a high-order filtered graph convolutional attention network according to claim 1 or 2, characterized in that: The specific operations in the third step are as follows: both layers of MixHop are 4 - level MixHop; the initialized 128 - dimensional features of the node embedding are input into the first - layer 4 - level MixHop. The feature output dimension of each level is 32 dimensions. The features of the first layer are directly concatenated to form an output dimension of 128 dimensions; the 32 - dimensional features of different levels of the second layer are not concatenated and are processed hierarchically in subsequent steps.

4. The biomolecular interaction prediction method based on a high-order filtered graph convolutional attention network according to claim 1 or 2, characterized in that: The specific operations of the fourth step are as follows: The 32-dimensional basic structure feature H output by the GCN (base) is concatenated with the 32-dimensional feature output by the second-level hierarchical MixHop of the high-order structure feature extraction module in sequence to form a mixture M of the high-order structure feature and the basic structure feature at 4 levels j (j = {0, 1, 2, 3}), and the mixed feature of each level has a dimension of 64.

5. The biomolecular interaction prediction method based on the high-order filtered graph convolutional attention network according to claim 3, characterized in that: The specific operations of the fourth step are as follows: The 32-dimensional basic structure feature H output by the GCN (base) is sequentially concatenated with the 32-dimensional feature output by the second-level hierarchical MixHop of the high-order structure feature extraction module to form a mixture M of the high-order structure feature and the basic structure feature at 4 levels j (j = {0, 1, 2, 3}), and the mixed feature of each level has a dimension of 64.

6. The biomolecular interaction prediction method based on a high-order filtered graph convolutional attention network according to claim 1 or 2 or 5, characterized in that: In the link predictor in the sixth step, there is 1 layer of bilinear transformation layer and two layers of fully connected layers; the link predictor first concatenates the filtered features F0, F1, F2 and F3 to form 128-dimensional node features, and then inputs them into the bilinear transformation layer to form edge features; through two layers of fully connected layers, a predicted value is output. The first layer of fully connected layer reduces the edge features from 128 dimensions to 32 dimensions, and the second layer of fully connected layer changes the 32-dimensional edge features into a 1-dimensional value output.

7. The biomolecular interaction prediction method based on the high-order filtered graph convolutional attention network according to claim 3, wherein: In the link predictor in the sixth step, there is 1 layer of bilinear transformation layer and two layers of fully connected layers; the link predictor first concatenates the filtered features F0, F1, F2 and F3 to form 128-dimensional node features, and then inputs them into the bilinear transformation layer to form edge features; through two layers of fully connected layers, a predicted value is output. The first layer of fully connected layer reduces the edge features from 128 dimensions to 32 dimensions, and the second layer of fully connected layer changes the 32-dimensional edge features into a 1-dimensional value output.

8. The biomolecular interaction prediction method based on the high-order filtered graph convolutional attention network according to claim 4, characterized in that: In the link predictor in the sixth step, there is 1 layer of bilinear transformation layer and two layers of fully connected layers; the link predictor first concatenates the filtered features F0, F1, F2 and F3 to form 128-dimensional node features, and then inputs them into the bilinear transformation layer to form edge features; through two layers of fully connected layers, a predicted value is output. The first layer of fully connected layer reduces the edge features from 128 dimensions to 32 dimensions, and the second layer of fully connected layer changes the 32-dimensional edge features into a 1-dimensional value output.

Images