A lightweight prediction ppi method based on node2vec

Abstract

The application discloses a kind of light prediction PPI methods based on Node2vec, comprising the following steps: protein initial representation module: using Node2vec from protein interaction network G=(V, E) Unsupervisedly learn protein initial representation;Multi-scale relationship structure representation module: set two different scales GNN, to capture shallow and deep structure information in protein interaction network, including SGCN variant for shallow structure information and Mamba module for deep structure information;The protein feature representation obtained by learning SGCN and Mamba is weighted and summed to obtain the final protein representation;In prediction module, element multiplication is carried out to deep protein representation and shallow protein representation, the semantic information of feature is enhanced, and then prediction is carried out.Introduce unsupervised learning method Node2vec to learn the initial representation of protein unsupervisedly, then adopt graph neural network+supervised attention to carry out fine learning, and co-ordinate local feature and global feature in protein initial representation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of PPI prediction, and particularly relates to a lightweight PPI prediction method based on Node2vec. BACKGROUND

[0002] Protein-protein interactions (PPIs) are essential in various cellular functions and biological processes, as they form the basis of various biological activities, such as signal transduction, cell cycle progression, and metabolic pathways. Therefore, understanding, identifying, and predicting PPIs are essential for medical, pharmaceutical, and genetic research. In recent decades, various experimental techniques for high-throughput PPI prediction have emerged, including yeast two-hybrid screening and mass spectrometry protein complex identification methods. However, experimental PPI detection in wet laboratories is often expensive and time-consuming, which poses a challenge to meet the growing demand for large-scale PPI prediction. Therefore, there is an urgent need to develop computational methods that can effectively predict unknown PPIs. With the advancement of artificial intelligence technology, computational methods for PPI prediction have shifted the paradigm from traditional machine learning methods to deep learning techniques.

[0003] Supervised learning methods can achieve excellent results when there is a sufficient amount of labeled data. However, in protein-protein interaction (PPI) prediction, obtaining and labeling large datasets is costly, and a large amount of unlabeled data remains underutilized. In recent years, semi-supervised and self-supervised methods have emerged, utilizing large amounts of unlabeled data to reveal hidden patterns and structures. Some researchers have improved the clustering performance of single-cell RNA sequencing data by incorporating prior biological knowledge. Others have used labeled and unlabeled data to predict protein-protein interactions. With the development of next-generation sequencing technology, thousands of protein sequences have been identified and added to various databases, such as the Protein Data Bank (PDB). As the protein sequence corpus continues to grow, many studies have attempted to apply pre-training methods from natural language processing (NLP) to the field of biological sequences. Baidu has proposed large-scale unsupervised protein pre-training models and tools, such as MSA Transformer, Protein, Text, and Helix, which utilize these vast sequence corpora to infer the structure and function of proteins from unlabeled data.

[0004] These unsupervised methods have an advantage when labeled data is scarce, as they do not rely on explicit annotations for training. However, while they can capture broad patterns from large datasets, they can also impose a significant computational burden. SUMMARY

[0005] The purpose of the present application is to provide a lightweight PPI prediction method based on Node2vec in view of the above-mentioned problems existing at present, by introducing an unsupervised learning method Node2vec to learn the initial representation of proteins unsupervisely, and then using a supervised graph neural network + attention to learn finely, coordinating the local features and global features in the initial representation of proteins, and capturing the local and global structures in the PPI network through the Node2vec walking strategy. Thus, the node is provided with rich feature representation, and the homogeneity and isomorphism of the PPI network are better balanced. After obtaining the initial representation of proteins through Node2vec, a supervised learning method GIN is used to further learn the local modal features of proteins, and a multi-layer graph neural network and a Mamba mechanism are used to capture relationship features of different scales, and finally a classifier is used for prediction to obtain their local modal features and global modal features.

[0006] The technical scheme of the present application is as follows:

[0007] A lightweight PPI prediction method based on Node2vec, comprising the following steps:

[0008] Protein initial representation module: using Node2vec to learn the initial representation of proteins unsupervisely from the protein interaction network G=(V,E);

[0009] Multi-scale relationship structure representation module: setting two GNNs of different scales to capture the shallow and deep structure information in the protein interaction network, including an SGCN variant for shallow structure information and a Mamba module for deep structure information;

[0010] The protein feature representations learned by SGCN and Mamba are weighted and summed to obtain the final protein representation;

[0011] In the prediction module, the deep protein representation and the shallow protein representation are element-wise multiplied to enhance the semantic information of the features while retaining detailed information, and then prediction is performed.

[0012] Further, the multi-scale relationship structure representation module specifically comprises the following steps:

[0013] After obtaining the initial representation of proteins, a PPI network with the initial representation of proteins as node information is constructed, and GNN is used to learn the multi-scale relationship modal features between proteins;

[0014] Different types of GNNs are set to capture the shallow relationship modal features and the deep relationship modal features of the PPI network at the same time, and then the shallow and deep relationship modal features are fused to represent the final protein feature representation.

[0015] Further, the multi-scale relationship structure representation module adopts a graph isomorphism network to learn the shallow feature representation of the protein, and the graph isomorphism network includes the following steps:

[0016] Each GIN layer adopts a summation aggregation function to integrate the features of the neighbor nodes, and uses a multilayer perceptron to update the aggregated features;

[0017] For each protein node p, the update rule of the feature representation is:

[0018]

[0019] Wherein, N k(p) represents the k-hop neighbor set of node p, σ (l) and ∈ (l) are the level-specific activation function and the learnable self-loop parameter respectively, through the recursive aggregation and update mechanism, GIN can learn and refine the advanced compact representation of the protein from the Node2vec pre-trained features.

[0020] Further, the multi-scale relationship structure representation module adopts Graph-Mamba to integrate the GAT and Mamba mechanisms to extract the deep relationship modal features in the PPI network, and specifically includes the following steps:

[0021] GAT adopts an attention mechanism to learn the importance of each adjacent node to the current node, and when calculating the attention coefficient α i,j , the edge features are connected with the features of the adjacent nodes and the current node:

[0022]

[0023] Wherein, T represents transposition, || represents connection, α is a learnable parameter for learning the attention coefficient in a single-layer feedforward neural network, W is a linear transformation shared in all node and edge features, X is a node feature, E is an edge feature, is a certain adjacent node of the current node i, k represents all adjacent nodes j of the current node i, once α i,j of all adjacent nodes is obtained, the feature of the current node is updated through the weighted sum of the features of its adjacent nodes;

[0024] The Graph-Mamba framework is set by learning long-distance features through Mamba, and the Graph-Mamba framework includes at least five stacked GMB layers.

[0025] Further, the structure of the Mamba includes:

[0026] MAMBA is mainly based on structured state space models, which are a type of sequential model, linear ordinary differential equations through a latent state h(t) e R N The features x(t) e R N learned by GAT that contain node importance are mapped to output sequences y(t) e R N ;

[0027] The core idea of SSM is to describe the dynamic behavior and input-output relationship of the system using state space representation:

[0028] h'(t) = Ah(t) + Bx(t),

[0029] y(t) = Ch(t),

[0030] where A e R N xN and B, C e R N represent state matrix, input matrix and output matrix, respectively;

[0031] The above calculation formula is discretized as:

[0032]

[0033] y t = Ch t ,

[0034] where, and Δ is the discretized state parameter, and Δ is the discretization step;

[0035] Consider N = 1, A = -1, B = 1, and Δ t = softplus(Linear(x t )), for selective SSM, the discretized recursive output is defined as:

[0036] g t = σ(Linear(x t )),

[0037] h t = (1-g t )h t-1 +g t x t

[0038] g t ranges between 0 and 1, so that the model can completely filter out irrelevant context when needed, and the current input x t can control the current input and the previous context when updating the hidden state through g t ​

[0039] h t-1 the balance between the two, the data dependence on the discretization step size and The projection matrices B and C are parameterized as linear projections of the input x to further control the influence of x t on the extent of the update of the hidden state h t and h t on the output y

[0040] Proof: Substitute the expression of Δt into the zero-order hold discretization formula: and

[0041]

[0042]

[0043] By parameterizing B, C and Δ on the input x, where B enables the model to control the influence of the input x t on the hidden state h t , C enables the model to control the influence of h t on the output y t , and Δ controls the extent of attention or neglect of x t , a larger Δ means more attention to x t than the previous hidden state;

[0044] Each node in the graph is a token, and Mamba's selection mechanism enables the model to minimize the influence of unimportant nodes in each step of hidden state calculation; Laplacian position encoding as Mamba's position information;

[0045] The formula of the position encoding is as follows:

[0046] L = D - A,

[0047] where A is the adjacency matrix of the PPI network and D is the degree matrix of the PPI network nodes. By calculating the eigenvalues and eigenvectors of the graph Laplacian matrix L, the position encoding P is generated, and the position encoding P is added to the input features of Mamba:

[0048]

[0049] where, is the position encoded feature of node i, h i is the initial feature of node i, p i is the position encoding of node i;

[0050] The updated node embeddings are combined through a set of MLP layers to generate output node embeddings, using the output of the previous layer as the input of the next layer, iterated in 5 GMB layers to obtain the final output node embeddings, the updating process of node embeddings in the MLP is as follows:

[0051] h (i+1) = ReLU(BatchNorm(W (2) h (1) + b (i) + b (1) ), (2) .

[0052] h(i) is the input node embedding of the i-th layer, W (1) and W (1) are the weight matrices of the linear layer, b (1) and b (2) are the bias terms of the linear layer, and the output node embedding is generated after two linear transformations, two ReLU activations and one batch normalization in one update.

[0053] Further, the protein initial representation module specifically comprises the following steps:

[0054] For a given protein interaction network G=(V,E), a random walk sequence is generated for each node, in Node2vec, the random walk is controlled by adjusting two parameters p and q, which represent the degree of breadth-first search and depth-first search in the random walk process, respectively;

[0055] Starting from node v i , the probability of moving to node v j in the next step is represented as:

[0056]

[0057] Where E represents the edge set in the graph.

[0058] After generating the node sequence, the node sequence is used to train the Word2Vec model to learn the low-dimensional representation of the node.

[0059] Further, the calculation formula of the Word2Vec model is:

[0060] For the node sequence (v1,v2,…,v T ), the conditional probability of the neighboring nodes v i of each node v i+j is maximized as:

[0061]

[0062] where p(v i+j |v i ; θ) represents the probability of predicting node v i given node v i+j , θ represents model parameters, and k represents the size of the context window.

[0063] Further, the prediction module will simultaneously receive shallow relationship modal feature H s and deep relationship modal feature H d output from the multi-scale relationship modal feature learning module, then assign weights for addition and fusion to obtain the final protein feature representation H, and the calculation formula is:

[0064] H = w1H s + w2H d ,

[0065] Protein interaction representation Z is obtained by Hadamard product on the relevant protein feature representation, and the calculation formula is:

[0066] z k = h i *h j , if A ij = 1,

[0067] Z = [z1...z n ],

[0068] The protein interaction representation Z is input into the MLP as a classifier to predict the protein interaction type.

[0069] Compared with the existing technology, the beneficial effects of the present application are:

[0070] 1. A lightweight PPI prediction method based on Node2vec, the protein initial representation module: an unsupervised graph representation learning method on the protein interaction network is designed by using Node2Vec, and the node representation obtained from the node2vec learning will be used as the node information in the PPIs network;

[0071] 2. A lightweight PPI prediction method based on Node2vec, the multi-scale relationship structure representation module: two GNNs of different scales are set to capture the shallow and deep structure information in the protein interaction network, which includes an SGCN variant for shallow structure information and a Mamba module for deep structure information, the protein feature representations obtained by learning SGCN and Mamba are weighted and summed to obtain the final protein representation, and finally in the prediction module, the deep protein representation and the shallow protein representation are element-wise multiplied to enhance the semantic information of the features while retaining detailed information, and then prediction is performed. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the architecture of this application.

[0073] Figure 2 A detailed grouping comparison plot of unknown protein-protein interactions (PPIs) between 82 and GNN-PPI under cross-dataset conditions.

[0074] Figure 3 This is a comparison chart of the F1 scores of 82 and GNN-PPI under three partitions, using SHS27k as the training set and string as the test set, under the condition of cross-dataset.

[0075] Figure 4 This is a comparison chart of the F1 scores of 82 and GNN-PPI under three partitions, using SHS148k as the training set and string as the test set, under the condition of cross-dataset. Detailed Implementation

[0076] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0077] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0078] Given N proteins P = {P1, P2, ..., P...} N} and a set of PPIs (protein-protein interactions) X={(P i ,P j )|P i ,P j Construct a PPI graph G for each ∈P, i≠j}. T =(V T E T ), where each node p i ∈V T Represents a protein P i ∈P, each edge ei,j ∈E T iff (P i ,P j )∈X. Consider a semi-supervised PPI prediction task, where only part of PPIs X L ∈X and their corresponding interaction types Y L are known, the labeled set can be represented as D L =(X L ,Y L ), and the unlabeled set as D U =(X U ,Y U ), where X U =X / X L . The goal of PPI prediction is to learn a mapping F∶X→Y in the transductive setting of the training data (X,Y L ) so that it can be used to infer the interaction types Y U of the unlabeled PPIs X U .

[0079] Referring to Figures 1-3 , a lightweight PPI prediction method based on Node2vec, as shown in Figure 1 , includes the following steps:

[0080] Protein initial representation module: using Node2vec to learn the initial representation of proteins unsupervised from the protein interaction network G=(V,E);

[0081] The protein initial representation module specifically includes the following steps:

[0082] For a given protein interaction network G=(V,E), a random walk sequence is generated for each node. In Node2vec, the random walk is controlled by adjusting two parameters p and q, which represent the degree of breadth-first search and depth-first search in the random walk process, respectively.

[0083] Starting from node v i , the probability of moving to node v j in the next step is represented as:

[0084]

[0085] Where E represents the edge set in the graph.

[0086] After generating the node sequence, the Word2Vec model is trained using the node sequence to learn the low-dimensional representation of the nodes.

[0087] The objective of the Word2Vec model is to maximize the conditional probability of neighboring nodes given a sequence of nodes.

[0088] The computational formula of the Word2Vec model is:

[0089] For a sequence of nodes (v1, v2, …, v T ), the conditional probability of the neighboring nodes v i of each node v itj is maximized:

[0090]

[0091] where p(v i+j |v i ; θ) represents the probability of predicting the node v i given the node v i+j , θ represents the model parameters, and k represents the size of the context window.

[0092] The Node2vec method can learn the low-dimensional representation of nodes in the graph in the protein-protein interaction network, thereby effectively capturing the structural information and semantic information between nodes. Finally, each learned low-dimensional representation of the node is taken as the initial representation of the protein.

[0093] The multi-scale relationship structure representation module: two GNNs of different scales are set to capture the shallow and deep structural information in the protein-protein interaction network, including an SGCN variant for shallow structural information and a Mamba module for deep structural information;

[0094] The multi-scale relationship structure representation module specifically includes the following steps:

[0095] After obtaining the initial representation of the protein, a PPI network with the initial representation of the protein as the node information is constructed, and GNN is used to learn the multi-scale relationship modal features between proteins;

[0096] Different types of GNNs are set to simultaneously capture the relationship modal features of the PPI network at the shallow and deep levels, and then the final protein feature representation is represented by fusing the shallow and deep relationship modal features. The protein representation actually contains protein relationship modal information at three scales, which are the relationship modal information learned by the protein initial representation module unsupervised learning, the shallow relationship modal information and the deep relationship modal information.

[0097] The multi-scale relationship structure representation module uses a graph isomorphism network to learn the shallow feature representation of the protein, and the graph isomorphism network includes the following steps:

[0098] Each GIN layer adopts a sum aggregation function to integrate the features of neighboring nodes, and uses a multi-layer perceptron (MLP) to update the aggregated features; such a design allows not only to consider the direct neighbors of each protein in each iteration, but also to gradually expand to a wider neighborhood range through a multi-layer structure.

[0099] For each protein node p, the update rule of its feature representation is:

[0100]

[0101] where N k(p) represents the k-hop neighbor set of node p, σ (l) and ∈ (l) are the layer-specific activation function and learnable self-loop parameter, respectively. Through the recursive aggregation and update mechanism, GIN can learn and refine the advanced compact representation of proteins from the Node2vec pre-trained features.

[0102] Through such a configuration, the GIN model can effectively refine the advanced features useful for PPI prediction on the basis of maintaining the original topology and functional attributes of proteins. In addition, the protein representation learned by GIN not only contains individual attributes, but also integrates the interaction information between proteins, thereby providing a powerful foundation for subsequent PPI prediction and biological function analysis.

[0103] The multi-scale relationship structure representation module uses Graph-Mamba to integrate GAT and Mamba mechanisms to extract deep relationship modal features in PPI networks, including the following steps:

[0104] GAT uses an attention mechanism to learn the importance of each neighboring node to the current node, and when calculating the attention coefficient α i,j , the edge features are connected with the features of neighboring nodes and the current node:

[0105]

[0106] where T represents transposition, || represents connection, α is a learnable parameter for learning attention coefficients in a single-layer feedforward neural network, W is a linear transformation shared in all node and edge features, X is a node feature, E is an edge feature, is a certain neighboring node of the current node i, and k represents all neighboring nodes j of the current node i. Once α i,j of all neighboring nodes is obtained, the feature of the current node is updated through the weighted sum of its neighboring node features;

[0107] The Graph-Mamba framework is configured to learn long-range features using Mamba. The Graph-Mamba framework comprises at least five stacked GMB layers. It provides better adaptability in modeling long sequences, enabling the model to adaptively select relevant information from the context.

[0108] Mamba introduces a selection mechanism that allows the model to selectively retain information, enabling it to learn deeper layers of information. In PPI graph learning, with protein nodes as the input sequence, the selection mechanism allows the hidden state to be updated based on relevant protein nodes in the previous sequence. This is controlled by the current input protein node and subsequently affects the output embedding of the current node. The ability to select and reset allows Mamba to extract protein-related dependencies given a long-range context, while minimizing the influence of unimportant protein nodes at each recursive step.

[0109] Mamba's structure includes:

[0110] MAMBA is primarily based on the Structured State-Space Model (SSM), a type of sequential model where a linear ordinary differential equation (ODE) passes through a latent state h(t) ∈ R. N The features x(t)∈R that contain node importance learned by GAT N Mapping to the output sequence y(t)∈R N ;

[0111] The core idea of ​​SSM is to use state-space representation to describe the dynamic behavior and input-output relationships of a system:

[0112] h′(t)=Ah(t)+Bx(t),

[0113] y(t) = Ch(t),

[0114] Where A∈R N ×N and B,C∈R N These represent the state matrix, input matrix, and output matrix, respectively.

[0115] Discretize the above calculation formula as follows:

[0116]

[0117] y t =Ch t ,

[0118] in, and Δ is the discretized state parameter, and Δ is the discretization step size. SSM has rich theoretical properties, but suffers from high computational cost and numerical instability. The Structured State-Space Sequence Model (S4) solves these limitations by imposing structural constraints on the state matrix A based on the HIPPO matrix, thereby significantly improving performance and efficiency.

[0119] S4 demonstrates better adaptability in modeling long sequences, but performs poorly when content-aware reasoning is required, due to its time invariance. Specifically, matrices A, B, and C are identical for all input labels in the sequence. Mamba achieves this by simply treating the SSM parameters B, C, and Δ as functions of the input x.

[0120] Using the reparameterized discretized step size Δ as an example, we can illustrate the intuition of the Mamba selection mechanism. Δt plays a generalized role similar to the gating mechanism in an RNN, facilitating the selection of input dependencies. For further reference to the SSM selection mechanism, the following theorem is given:

[0121] Consider N = 1, A = -1, B = 1, and Δ t =softplus(Linear(x) t For selective SSM, the discretized recursive output is defined as:

[0122] g t =σ(Linear(x) t )),

[0123] h t =(1-g t )h t-1 +g t x t

[0124] g t The range is between 0 and 1, allowing the model to completely filter out irrelevant context when needed, given the current input x. t Able to pass through g t Control the current input and previous context when updating the hidden state. t-1 The balance between these factors, and obtaining data dependencies within the discretization step size. and Projection matrices B and C are parameterized as linear projections of the input x to further control x. t Update the degree of hidden state and h t For output y t The impact.

[0125] Proof: Substitute the expression for Δt into the zero-order preserved discretization formula: and

[0126]

[0127] By parameterizing B, C and Δ on the input x, where B enables the model to control the input x t the influence of the hidden state h t , C enables the model to control the influence of h t on the output y t , and Δ controls the degree of attention or neglect of x t , a larger Δ means paying more attention to x t than the previous hidden state.

[0128] Each node in the graph is a token, and Mamba's selection mechanism enables the model to minimize the influence of unimportant nodes in each step of hidden state calculation; Laplacian position encoding as Mamba's position information;

[0129] The formula of position encoding is as follows:

[0130] L = D - A,

[0131] where A is the adjacency matrix of the PPI network, and D is the degree matrix of the PPI network nodes. By calculating the eigenvalues and eigenvectors of the graph Laplacian matrix L, the position encoding P is generated, and the position encoding P is added to the input features of Mamba:

[0132]

[0133] where, is the position encoded feature of node i, h i is the initial feature of node i, p i is the position encoding of node i.

[0134] The updated node embedding is combined through the MLP layer to generate the output node embedding, using the output of the previous layer as the input of the next layer, and iterating in 5 GMB layers to obtain the final output node embedding. The updating process of node embedding in MLP is as follows:

[0135] h (i+1) = ReLU (BatchNorm (W (2) ReLU (W (1) h (i) + b (1) + b (2) )),

[0136] h(i) is the input node embedding of the i-th layer, W (1) and W (1) are the weight matrices of the linear layer, and b (1) and b(2) is the bias term of the linear layer, and the output node embedding is generated after two linear transformations, two ReLU activations, and one batch normalization in one update.

[0137] Graph-Mamba combines the selection mechanism in Mamba module and achieves two layers of graph sparsification through cyclic scanning in sequence modeling. The first layer involves the selection mechanism in the Mamba module, which effectively filters the relevant information in the long-distance context. The second layer is achieved through the proposed node priority method, which enables important nodes in the graph to access more context. Therefore, these sequence modeling features provide a promising method for graph sparsification that combines data dependency and heuristic selection.

[0138] The final protein representation is obtained by weighting and summing the protein feature representations learned by SGCN and Mamba;

[0139] In the prediction module, the deep protein representation and the shallow protein representation are element-wise multiplied to enhance the semantic information of the features while preserving detailed information, and then prediction is performed.

[0140] The prediction module will simultaneously receive shallow relationship modal feature H s and deep relationship modal feature H d output from the multi-scale relationship modal feature learning module, then assign weights for summation fusion to obtain the final protein feature representation H, the calculation formula is:

[0141] H=w1H s +w2H d ,

[0142] The protein interaction representation Z is obtained by Hadamard product on the relevant protein feature representation, the calculation formula is:

[0143] z k =h i *h j ,if A ij =1,

[0144] Z=[z1...z n ],

[0145] The protein interaction representation Z is input into the MLP as a classifier to predict the protein interaction type.

[0146] Experiment:

[0147] For the lightweight PPI prediction method proposed in this application, it is referred to as "82" in the experiment.

[0148] Datasets: Extensive experiments were conducted on three public PPI datasets, STRING, SHS27k, and SHS148k, respectively. The STRING dataset contains human PPI data from the STRING database, with a total of 1,150,830 entries covering 14,952 proteins and 572,568 interactions. Each protein-protein interaction is annotated as at least one of the following seven types: Activation, Binding, Catalysis, Expression, Inhibition, Post-translational modification (Ptmod), and Reaction. "One interaction + one annotation" is referred to as one PPI data entry. From the human subset of the STRING database, proteins with more than 50 amino acids and less than 40% sequence identity were randomly selected to generate two more challenging datasets, SHS27k and SHS148k. The SHS27k dataset contains 16,912 PPI entries covering 1,663 proteins and 7,401 interactions. The SHS148k dataset contains 99,782 PPI entries covering 5,082 proteins and 43,397 interactions. In addition, three partitioning algorithms, including Random, Breath-First Search (BFS), and Depth-First Search (DFS), were used to divide each dataset into training and test sets in an 8:2 ratio. To better evaluate the generalization ability, the test data was further divided into three subsets according to whether the two proteins had been seen in the training data: (1) BS: both seen; (2) ES: seen one protein; (3) NS: both not seen. In practice, the BFS and DFS partitions are more challenging because their test data only contains the ES and NS subsets, and the test set obtained by BFS is very difficult for the model to predict interactions because it contains a larger proportion of unknown proteins. It is assumed that the data set partitioning using BFS is more consistent with practical applications, and to some extent, it may reflect the performance of the model in real-world scenarios.

[0149] Evaluation metrics and hyperparameters:

[0150] All experiments were performed using Python v3.8.3 and PyTorch v1.8.1. The experimental results are the average of three repeated experiments and were calculated on an NVIDIA RTX 3090 GPU with 24 GB of memory. In the model, the dimension of protein node features obtained by Node2vec is 256, and the hidden dimension of the classifier is set to 512. The number of layers of the GMB encoder is set to 5. The model is trained on the SHS27k dataset for 600 epochs with a batch size of 2056, and on the SHS148k dataset for 600 epochs with a batch size of 2056, using the Adam optimizer, a learning rate (lr) of 0.01, and a weight decay of 5e-4. The patience value of the learning rate decay is set to 30, and the decay rate is set to 0.5. The evaluation indicator for multi-label PPI prediction is Micro-F1. Micro-F1 is widely used in multi-label classification, and since the different PPI types in the data set used are extremely imbalanced, micro-average F1 (micro-F1) may be more suitable than macro-average F1 (macro-F1) for evaluating the performance of multi-label PPI type prediction. Micro-F1 is obtained by calculating the harmonic mean of the total precision and recall of all types.

[0151]

[0152] Baseline: The current effective PPI prediction methods include sequence-based DL-PPI, DNN-PPI, GNN-PPI, and structure-based DSSGN-PPI, HIGH-PPI. Further, some scholars use unlabeled data for pre-training, including AFTGAN, to more accurately predict PPI. The performance of the present application is superior to that of the un-pretrained method, and is superior to or at least comparable to the pre-trained method. It is worth mentioning that the pre-trained model consumes a large amount of computing resources and time. The present application model can achieve the pre-training effect using only the most basic PPI network.

[0153] Three important observations can be drawn from Table 1: (1) For models trained only on downstream labeled data, the two structure-based models, HIGH-PPI and DSSGNN-PPI, outperform other sequence-based baselines, but these two methods still lag behind the methods. (2) AFTGAN is a sequence-based model, and its performance is lower than that of structure-based models. However, because AFTGAN adds ESM-1b embeddings containing a lot of biological information as protein sequence features, AFTGAN's performance is better than other sequence-based models, but it is still lower than models that only use PPI networks, which further illustrates the efficiency of the model in predicting novel PPI relationships. (3) For the structure-based model HIGH-PPI, since protein interactions are closely related to protein structure, its performance is significantly improved compared to methods that only use sequence information. However, the introduction of structural information inevitably brings a large computational burden. The method in this application can outperform the method that uses structural information by using only sequence information.

[0154] Table 1

[0155]

[0156]

[0157] Generalization Analysis: Each PPI consists of two proteins. Unknown PPIs can be further categorized into more detailed classes based on the protein's visibility to the method. Proteins appearing during training are known proteins to the method; therefore, in testing, both proteins in an interaction are known protein representations (BS), any but not both proteins in an interaction are known protein representations (ES), and no protein in an interaction is a known protein representation (NS). ES and NS are interactions involving unknown proteins. Figure 2 As shown, GNN-PPI and 82 exhibit significant performance degradation in PPI prediction on the "ES" and "NS" datasets because these sets contain novel proteins that the model had never encountered during training. Nevertheless, our method shows a significant improvement over GNNPPI. Particularly in the "NS" category, our method achieves a high improvement (average 7%), where competing methods experience a rapid performance decline. This is because the "NS" PPI does not contain any known proteins, requiring the model to have high generalization ability. However, due to the corpus information of the PPI learned by the model during training and the learning of multi-scale shallow and deep interactions, our method has a clear advantage in handling such unknown protein interaction prediction tasks.

[0158] To better evaluate the generalization ability of the method, the method of training using one dataset and then testing the performance of the model on another larger dataset, i.e. the heterogenous train-test set, is evaluated. For example, a model trained on the SHS27k dataset will encounter proteins containing fewer amino acids on the STRING dataset. More importantly, unlike the BFS and DFS divisions on the SHS27k dataset, more than 80% of the PPIs in the test data are located in the NS subset when migrating to STRING. In Figure 3 and Figure 4 The performance of 82 and GNNPPI in the two domain migration settings is reported in 82 outperforms GNNPPI in all settings, especially when migrating from SHS148k to STRING.

[0159] Ablation experiments:

[0160] The proposed method mainly consists of three parts: the unsupervised generation of PPI initial representation part (referred to as Node), the supervised local feature extraction part (referred to as GIN), and the supervised global feature extraction part (referred to as GM). The initial protein representation learned by the proposed protein initial feature representation module is also from the PPI network, which is a scale of protein relationship modal information, so a MLP classifier is used to directly use the protein initial representation to predict protein-protein interaction. The Node in the table represents this case. To prove the effectiveness of each part, different combinations of these components are tested. As shown in Table 2, the best performance is obtained by the complete 82; when the GM module for extracting global modal information is removed, the performance is lower than that of the complete 82, which shows that the global feature contributes to improving the prediction of protein-protein interaction, and when the local feature module GIN is removed, the performance of the model is significantly decreased, because for the PPI network, the nearest neighbor contains most of the effective information, and when the local feature is missing, the model performance is greatly discounted. When only using the protein relationship modal information obtained at the unsupervised learning scale of the protein initial representation module, and using a simple MLP as the classifier, the model performance is sharply decreased,

[0161] Table 2

[0162]

[0163] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are within the protection scope of the present application.

Claims

1. A method for lightweight prediction of PPI based on Node2vec, characterized in that, The method comprises the following steps: Protein initial representation module: learning protein initial representations unsupervised from protein interaction networks using Node2vec from protein interaction networks using Node2vec The multi-scale relationship structure representation module: two GNNs of different scales are set to capture shallow and deep structural information in the protein interaction network, including an SGCN variant for shallow structural information and a Mamba module for deep structural information; The protein feature representations learned by SGCN and Mamba are weighted and summed to obtain the final protein representation; In the prediction module, the deep protein representation and the shallow protein representation are element-wise multiplied to enhance the semantic information of the features while retaining detailed information, and then prediction is performed.

2. The method of claim 1, wherein, The multi-scale relationship structure representation module specifically comprises the following steps: After obtaining the initial protein representation, a PPI network is constructed with the initial protein representation as node information, and GNN is used to learn the multi-scale relationship modal features between proteins; Different types of GNNs are set to simultaneously capture the shallow relationship modal features and the deep relationship modal features of the PPI network, and then the final protein feature representation is represented by fusing the shallow and deep relationship modal features.

3. The method of claim 1 or 2, wherein, The multi-scale relationship structure representation module uses a graph isomorphism network to learn the shallow feature representation of the protein, and the graph isomorphism network comprises the following steps: Each GIN layer uses a summation aggregation function to integrate the features of neighbor nodes, and uses a multi-layer perceptron to update the aggregated features; For each protein node The update rule for the feature representation is ， wherein, representing nodes of a set of hop neighbors, and are the activation function and the learnable self-loop parameter of the layer, respectively. Through a recursive aggregation and update mechanism, GIN is able to learn and distill high-level compact representations of proteins from the Node2vec pre-trained features.

4. The method of claim 1 or 2, wherein, The multi-scale relationship structure representation module uses Graph-Mamba to integrate GAT and Mamba mechanisms to extract deep relationship modal features in the PPI network, and specifically comprises the following steps: The GAT adopts an attention mechanism to learn the importance of each neighboring node to the current node, and in calculating the attention coefficient the edge feature is connected with the features of the neighboring nodes and the current node: ， in, Indicates transpose. Indicates a connection. These are learnable parameters for learning attention coefficients in a single-layer feedforward neural network. It is a linear transformation shared across all node and edge features. It is a node feature. It is an edge feature. The current node A neighboring node, Represents all current nodes neighboring nodes Once all neighboring nodes are obtained The features of the current node are updated by the weighted sum of the features of its neighboring nodes; The Graph-Mamba framework is set, and Mamba is used to learn long-distance features, and the Graph-Mamba framework comprises at least five stacked GMB layers.

5. The method of claim 4, wherein, The structure of the Mamba comprises: MAMBA is based on a structured state space model, which is a kind of sequential model, linear ordinary differential equations through a latent state Features learned by GAT that contain node importance are mapped to the output sequence ; The core idea of the SSM is to use a state space representation to describe the dynamic behavior and input-output relationship of the system: ， ， wherein, and denote state matrix, input matrix and output matrix, respectively; The above calculation formula is discretized as: ， ， wherein and , is a discretization step size; Consider , , , and For selective SSM, the discretized recurrent output is defined as: ， ranging between 0 and 1, enabling the model to completely filter out irrelevant context when needed, the current input can be controlled by controlling the balance between the current input and the previous context when updating the hidden state acquiring data dependencies in discrete steps and ; projection matrices B and C are parameterized as linear projections of the input to further control the degree of updating the hidden state and the influence on the output ; Proof: Substitute the expression for into the zeroth-order hold discretization formula: and : ； By parameterizing B, C and B enables the model to control the influence of the hidden state on the input C enables the model to control the influence of the hidden state on the output , controlling the degree of focus or neglect on , a larger means more focus on the hidden state than the previous one ; Each node in the graph is a token, and the selection mechanism of Mamba enables the model to minimize the influence of unimportant nodes in each step of hidden state calculation; the Laplacian position encoding serves as the position information of Mamba; The formula of the position encoding is as follows: ， wherein, is an adjacency matrix of the PPI network, is a degree matrix of the PPI network nodes, the position encoding is generated by computing eigenvalues and eigenvectors of the graph Laplacian matrix L , the position encoding is added to the input features of Mamba: ， wherein, is a position encoded feature of a node , is an initial feature of a node , is a position encoding of a node . The updated node embedding is combined through an MLP layer to generate an output node embedding, and the output of the previous layer is used as the input of the next layer, and the iteration is performed in the five GMB layers to obtain the final output node embedding, and the updating process of the node embedding in the MLP is as follows: ， is the first input node embedding of the layer, and is the weight matrix of the linear layer, and is the bias term of the linear layer, generating the output node embedding after two linear transformations, two ReLU activations, and one batch normalization in one update.

6. The method of claim 1, wherein, The protein initial representation module specifically comprises the following steps: For a given protein interaction network , a random walk is generated for each node, in Node2vec, the random walk is controlled by adjusting two parameters and , which represent the degree of breadth-first search and depth-first search in the random walk process, respectively; From node Start random walk, next move to node The probability of moving to node ， wherein represents a set of edges in the graph; After generating the node sequence, the node sequence is used to train a Word2Vec model to learn the low-dimensional representation of the node.

7. The method of claim 6, wherein, The calculation formula of the Word2Vec model is as follows: For a sequence of nodes , the conditional probability of the neighboring nodes of each node is maximized as: ， wherein, denotes the probability of predicting a node at a given node , denotes the model parameters, denotes the size of the context window.

8. The method of claim 1, wherein, The prediction module will simultaneously receive shallow relationship modal features output from the multi-scale relationship modal feature learning module , and deep relationship modal features , then assign weights for summation fusion to obtain the final protein feature representation , the calculation formula is: ， protein interaction representation The protein interaction representation is obtained by Hadamard product of the relevant protein feature representations, and the calculation formula is: ， ， The protein interaction representation Z is input into an MLP serving as a classifier to predict the protein interaction type.

Images