A protein interaction network alignment method and system

Abstract

The application relates to a protein interaction network alignment method and system, which comprises the following steps: obtaining structure information of two protein interaction networks; extracting hierarchical structure features of all nodes in each network, and calculating structure scores between nodes based on the hierarchical structure features; constructing a multi-layer weighted graph based on the structure scores, so that each node performs random walk on the weighted graph to generate a context sequence; inputting the context sequence corresponding to each node in the two networks into a word vector model to obtain a low-dimensional vector corresponding to each node, and calculating the distance between the low-dimensional vectors corresponding to each node as the topological similarity between the nodes; combining the sequence similarity between the nodes and the topological similarity between the nodes to calculate the similarity score between the nodes, and constructing a similarity matrix between the nodes based on the similarity score; and generating a node alignment result of the two networks according to the similarity matrix between the nodes. The application effectively improves the protein interaction network alignment efficiency.

Description

Technical Field

[0001] This invention relates to the field of protein interaction network alignment technology, and in particular to a method and system for protein interaction network alignment. Background Technology

[0002] The development of high-throughput technologies and large-scale computing has led to the acquisition of vast amounts of protein-protein interaction (PPI) network data. Proteins are fundamental to various biological functions in cells, and analyzing these PPI networks can provide a better understanding of biological tissue characteristics and the discovery of important functional modules. Network alignment is a more efficient method for analyzing PPI networks, used to discover conserved functional modules and predict unknown functions of proteins. Network alignment matches different networks based on their similarity, thereby discovering conserved proteins and transferring information from well-studied species to less studied species. Currently, numerous global network alignment algorithms have been proposed.

[0003] SoRank is the first network alignment algorithm. Based on the idea of ​​PageRank, it aggregates the feature values ​​of a node's neighbors and combines them with sequence similarity to calculate the similarity score between nodes, then generates the alignment using spectral clustering. The problem with this algorithm is that it only uses neighbor information and does not explore the deep topological information in the network, resulting in low topological and biological functional quality of the final alignment results.

[0004] NAI GO first divides the network into subnets, and then selects the module with the highest similarity based on degree label similarity and GO terminology similarity as a seed for expansion, thus obtaining the final alignment result. The problems with this algorithm are: over-reliance on GO similarity, inability to align networks with unknown GO terms, and low alignment efficiency.

[0005] The MAGNA algorithm introduces heuristic algorithms into network alignment, using a genetic algorithm for the process. Each alignment result is treated as an individual in a population, and through iterative crossover and mutation of the initial population, individuals with higher objective function scores are generated, effectively avoiding the problem of local optima during alignment. However, this algorithm has drawbacks: the initial population is generally obtained randomly, and the alignment process involves multiple iterations to obtain the optimal result, resulting in a long runtime. Furthermore, the initial population and the number of iterations can negatively impact the alignment outcome.

[0006] Despite extensive research on network alignment, finding an efficient and meaningful algorithm remains a challenge. Existing algorithms often rely on network topology information to extract more conserved functional information, and how to better quantify this topology information is a significant challenge. Furthermore, further research is needed on alignment efficiency. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the problem of low protein network alignment efficiency in the prior art.

[0008] To address the aforementioned technical problems, this invention provides a method for aligning protein-protein interaction networks, comprising:

[0009] Step S1: Obtain the structural information of the interaction networks of any two proteins to be compared;

[0010] Step S2: Extract the structural features of all nodes in each protein-protein interaction network hierarchically to obtain hierarchical structural features, and calculate the structural scores between nodes based on the hierarchical structural features;

[0011] Step S3: Construct a multi-layered weighted graph, where each layer of the weighted graph is an undirected complete graph composed of all nodes in the protein-protein interaction network, and use the structural scores between the nodes as the edge weights of the weighted graph, so that each node performs a random walk on the weighted graph to generate a context sequence.

[0012] The context sequence corresponding to each node in any two protein-protein interaction networks to be compared is input into the word vector model for training, and the low-dimensional vector corresponding to each node is obtained. The distance between the low-dimensional vectors corresponding to each node is calculated as the topological similarity between nodes.

[0013] Step S4: Calculate the similarity score between nodes based on the sequence similarity and topological similarity between nodes, and construct a node similarity matrix based on the similarity score between nodes;

[0014] Step S5: Generate node alignment results for the two protein interaction networks based on the node similarity matrix.

[0015] In one embodiment of the present invention, the formula for the hierarchical structure feature in step S2 is:

[0016]

[0017] Among them, R k (u) represents the degree sequence vector of the node u's k-hop neighbor nodes, x i k Let d represent the number of k-hop neighbors of node u with degree i, and d represent the ideal maximum degree of the network.

[0018] In one embodiment of the present invention, in step S2, the structural score between nodes is calculated based on the hierarchical structure features, using the following formula:

[0019]

[0020] Where, N k (u) represents the number of k-hop neighbors of node u in the network, N k (v) represents the number of k-hop neighbors of node v in the network, R k (u) represents the degree sequence vector of the node u's k-hop neighbor nodes, R k (v) represents the degree sequence vector of the k-hop neighbor nodes at a distance of k hops from node v, and dis(R) k (u),R k (v) represents the similarity of the degree sequence vectors of node u and node v, and dis(R) k (u),R k (v))=||R k (u),R k (v)||2,||||2 represents the L2 norm.

[0021] In one embodiment of the present invention, the method of generating a context sequence by having each node perform a random walk on the weighted graph in step S3 includes: the node walking on the current layer of the weighted graph, and the node switching from the current layer to other layers in the weighted graph to generate a context sequence.

[0022] In one embodiment of the present invention, the probability formula for node u to move to node v in the k-th layer weighted graph when a node moves through the current layer of the weighted graph in step S3 is:

[0023]

[0024] Among them, f k (u,v) represents the structural score between node u and node v in the weighted graph of the k-th layer, where u' and v' represent any node in the node set V, and ∑ u',v'∈V f k (u',v') represents the sum of the weights of the edges between all nodes in the k-th layer weighted graph.

[0025] In one embodiment of the present invention, the probability formula for a node switching from the current layer to other layers in the weighted graph in step S3 is:

[0026]

[0027] Where, p k+1 (u) represents the probability that node u enters the corresponding node in the (k+1)th layer from the kth layer, p k-1 w(u) represents the probability that node u enters the corresponding node in the (k-1)th layer from the kth layer. k ,u k+1 ) represents the edge weight connecting different layers in a multi-layer weighted graph, uk Let u be a node at the k-th level. k+1 Let u represent the node at level k+1.

[0028] In one embodiment of the present invention, the edge weight formula connecting different layers in the multi-layer weighted graph is as follows:

[0029] w(u k ,u k+1 ) = log(e + ∑num|f k (u,v)>avg(f k )|)

[0030] Where e represents a constant value, f k (u,v) represents the structural score between node u and node v in the k-th layer weighted graph, avg(f k ) represents the average of the sum of the weights of all edges in the k-th layer weighted graph, and num|·| represents the number.

[0031] In one embodiment of the present invention, the low-dimensional vectors corresponding to each node in the two protein interaction networks in step S3 are located in the same vector space.

[0032] In one embodiment of the present invention, step S4 combines the sequence similarity between nodes and the topological similarity between nodes to calculate the similarity score between nodes, using the following formula:

[0033]

[0034] Where T(u,v) represents the topological similarity between nodes u and v and r(u) represents the low-dimensional vector of node u, r(v) represents the low-dimensional vector of node v, |·| represents the magnitude of the vector, Seq(u,v) represents the BLAST score of node u and node v, maxSeq represents the maximum BLAST score among all node pairs, and α represents the balance parameter and α∈(0,1).

[0035] To address the aforementioned technical problems, this invention provides a protein interaction network alignment system, comprising:

[0036] Acquisition module: Used to acquire structural information of the interaction networks of any two proteins to be compared;

[0037] Hierarchical extraction and calculation module: used to extract the structural features of all nodes in each protein-protein interaction network hierarchically to obtain hierarchical structural features, and calculate the structural scores between nodes based on the hierarchical structural features;

[0038] Topological similarity calculation module: used to construct a multi-layered weighted graph, each layer of which is an undirected complete graph composed of all nodes in the protein-protein interaction network, and uses the structural scores between the nodes as the edge weights of the weighted graph, so that each node performs a random walk on the weighted graph to generate a context sequence;

[0039] It is also used to input the context sequence corresponding to each node in any two protein-protein interaction networks to be compared into a word vector model for training, to obtain a low-dimensional vector corresponding to each node, and to calculate the distance between the low-dimensional vectors corresponding to each node as the topological similarity between nodes.

[0040] Similarity matrix construction module: used to calculate the similarity score between nodes based on the sequence similarity between nodes and the topological similarity between nodes, and to construct the similarity matrix of nodes based on the similarity score between nodes;

[0041] Alignment module: Used to generate node alignment results between two protein interaction networks based on the similarity matrix of the nodes.

[0042] The technical solution of the present invention has the following advantages compared with the prior art:

[0043] This invention extracts the structural features of nodes in a network in a hierarchical manner, which can discover the deep structural features of nodes, making two nodes that are not only structurally similar themselves but also have similar neighboring nodes more similar, thereby increasing the distinguishability between nodes.

[0044] This invention models the structural features of nodes using random walks and word vector models, so that nodes with similar topological structures have similar representation vectors in the embedding space (i.e., low-dimensional vectors of nodes), and uses the distance between vectors to quantify the topological score of nodes.

[0045] This invention combines sequence similarity information to preserve biological information in the network to a greater extent, and uses a greedy approach during comparison to expand the similarity relationships between nodes that are well preserved, thereby improving the scores of biological and topological indicators.

[0046] This invention can accurately match similar proteins between different species and find similar functional regions between different species; at the same time, this invention maintains the topological and biological balance of the alignment results well, without sacrificing topological quality for excessive pursuit of biological indicators. Attached Figure Description

[0047] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0048] Figure 1 This is a flowchart of the method of the present invention;

[0049] Figure 2 This is a schematic diagram of the overall framework of the protein interaction network alignment method according to an embodiment of the present invention;

[0050] Figures 3(a)-3(c) These are the scores of various metrics of the comparison results of the five methods in the embodiment on the synthetic network datasets CG and DMR;

[0051] Figures 4(a)-4(d) These are the scores of various metrics of the comparison results of the six methods in the embodiment on the real network dataset I soBase;

[0052] Figures 5(a)-5(d) These are the scores of various metrics for the comparison results of the six methods in the embodiment on the real network dataset BioGrid. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0054] Example 1

[0055] Reference Figure 1 and Figure 2 As shown, the protein interaction network alignment method of the present invention includes:

[0056] Step S1: Obtain the structural information of any two protein-protein interaction networks to be compared (i.e., Figure 2 (The input network part);

[0057] Step S2: Extract the structural features of all nodes in each protein-protein interaction network hierarchically to obtain hierarchical structural features, and calculate the structural scores between nodes based on these hierarchical structural features (i.e., ...). Figure 2 Extracting structural features from the data;

[0058] Step S3: Construct a weighted graph (the weighted graph is multi-layered) for each protein-protein interaction network. Each layer of the weighted graph is an undirected complete graph composed of all nodes in the protein-protein interaction network. The structural scores between the nodes are used as the edge weights of the weighted graph, and each node performs a random walk on the weighted graph to generate a context sequence.

[0059] The context sequence corresponding to each node in any two protein-protein interaction networks to be compared is input into the word vector model for training, resulting in a low-dimensional vector (i.e., ...) for each node. Figure 2 In the network embedding part, the distance between the low-dimensional vectors corresponding to each node is calculated as the topological similarity between nodes;

[0060] Step S4: Calculate the similarity score between nodes based on the sequence similarity and topological similarity between nodes, and construct a node similarity matrix based on the similarity score (i.e., ...). Figure 2 (Constructing the similarity matrix part);

[0061] Step S5: Generate the node alignment results of the two protein interaction networks based on the node similarity matrix (i.e., Figure 2 (Comparison results section).

[0062] The following is a detailed description of this embodiment:

[0063] Specifically, in step S1: For the input protein-protein interaction network (PPI network), it is modeled as an undirected graph G = (V, E), where V represents the set of nodes (i.e., proteins) in the network, E represents the set of edges (i.e., protein-protein interactions), and |V| and |E| represent the number of nodes and edges in the network, respectively. For PPI network alignment, the goal is to find a mapping relationship between two given networks G1 = (V1, E1) and G2 = (V2, E2), where |V1| ≤ |V2|, to achieve a one-to-one matching of nodes in G1 and G2, thereby achieving cross-species information transfer and inferring the functions of similar proteins. Figure 2 Network G1 has 6 nodes, and network G2 has 7 nodes.

[0064] Furthermore, in step S2, neighborhood information of nodes in the network is extracted hierarchically. For any node u∈V, different levels are divided based on u's k (k≥1) hop neighbors, using a d-dimensional vector R. k (u) represents the hierarchical structure feature, as shown in the following formula:

[0065]

[0066] Among them, R k (u) represents the degree sequence vector of the node u's k-hop neighbor nodes, x i k Let d represent the number of k-hop neighbors of node u with degree i, and d represent the ideal maximum degree of the network.

[0067] Since the PPI network is a sparse network with significant noise, directly using the maximum degree in network G as d might increase computational complexity. In this embodiment, to control the vector dimension and reduce the impact of noise, the following formula is used to revalue d and i:

[0068]

[0069] Where D represents the maximum degree in network G, u* Let deg(u) be the k-hop neighbor of u. * ) represents node u * The degree, and rounded down for d and i.

[0070] In step S2, having obtained neighbor information, for different nodes u,v∈V, the structural score formula between nodes in layer k is defined as follows:

[0071]

[0072] Where, N k (u) represents the number of k-hop neighbors of node u in the network, N k (v) represents the number of k-hop neighbors of node v in the network, R k (u) represents the degree sequence vector of the node u's k-hop neighbor nodes, R k (v) represents the degree sequence vector of the k-hop neighbor nodes at a distance of k hops from node v, and dis(R) k (u),R k (v) represents the similarity of the degree sequence vectors of node u and node v, and dis(R) k (u),R k (v))=||R k (u),R k (v)||2, || ||2 represents the L2 norm.

[0073] Furthermore, in step S3, this embodiment first constructs a k-layer weighted graph M. Each layer of the weighted graph M is a completely undirected weighted graph composed of all nodes in the network, and the weight of each edge is the structural score f. k (u,v). Then, for each node u∈V, a random walk is performed on the weighted graph M to generate the context sequence of node u.

[0074] In step S3, when each node walks through the weighted graph M, the weights between edges are always used as the basis for selecting the next node to walk through, so that nodes that are structurally similar to the current vertex are more likely to be selected. Furthermore, the method for each node to generate a context sequence through random walks on the weighted graph M includes: the node walking through the current layer of the weighted graph M, and the node switching from the current layer to other layers in the weighted graph M to generate a context sequence.

[0075] Specifically, in step S3, when a node walks through the current layer of the weighted graph M, the probability p of node u walking to node v in the k-th layer weighted graph is... k The formula for (u,v) is:

[0076]

[0077] Among them, f k(u,v) represents the structural score between node u and node v in the k-th layer weighted graph, where u' and v' represent any node in the node set V, and ∑ u',v'∈V f k (u',v') represents the sum of the weights of the edges between all nodes in the k-th layer weighted graph.

[0078] Meanwhile, because the weighted graph M is multi-layered, during a walk, it's necessary to consider whether to switch layers or remain in the current layer. Entering a higher layer means a higher probability of walking to a node similar to the current node. If switching layers is desired, the probabilities of entering the corresponding nodes in layers k+1 and k-1 (i.e., the probability of a node switching from its current layer to another layer in the weighted graph) are defined in the following formula:

[0079]

[0080] Where, p k+1 (u) represents the probability that node u enters the corresponding node in the (k+1)th layer from the kth layer, p k-1 w(u) represents the probability that node u enters the corresponding node in the (k-1)th layer from the kth layer. k ,u k+1 ) represents the edge weight connecting different layers in a multi-layer weighted graph, u k Let u be a node at the k-th level. k+1 Let u represent the node at level k+1.

[0081] Furthermore, the formula for the edge weights connecting different layers in a multi-layered weighted graph M is:

[0082] w(u k ,u k+1 ) = log(e + ∑num|f k (u,v)>avg(f k )|)

[0083] Where e represents a constant value, f k (u,v) represents the structural score between node u and node v in the k-th layer weighted graph, avg(f k ) represents the average of the sum of the edge weights in the k-th layer weighted graph, and num|·| represents the quantity. In this embodiment, a random walk starts from the lowest layer of the corresponding node in the weighted graph M, and by determining the step size of the walk, a context sequence of fixed length is obtained each time.

[0084] In step S3, the low-dimensional vectors corresponding to each node in the two protein interaction networks are located in the same vector space, which can be understood as having the same vector dimension, facilitating subsequent topological similarity calculation.

[0085] Further, in step S4, having obtained the context sequence of the nodes, the next step is to learn the features of the nodes from these sequences and represent them using vectors. In this embodiment, the obtained node context sequence is trained using a word vector model (Word2Vec model) to obtain low-dimensional vectors for the nodes. The goal of the Word2Vec model is to maximize the average log probability of the central word's context in the sequence, learning meaningful representations of words given a sentence context. For the obtained low-dimensional vectors of the nodes, a cosine function is used to calculate the distance between low-dimensional vectors, which is then used as the topological similarity between nodes, as shown in the following formula:

[0086]

[0087] Where T(u,v) represents the topological similarity between nodes u and v (i.e., the distance between low-dimensional vectors), r(u) represents the low-dimensional vector of node u, and r(v) represents the low-dimensional vector of node v.

[0088] Furthermore, in step S4, this embodiment, when constructing the similarity matrix for all nodes in the network, considers not only the topological similarity of the network but also sequence similarity during the alignment process to generate more biologically meaningful alignment results. BLAST is a commonly used sequence alignment tool, so this embodiment uses BLAST values ​​as the sequence similarity between different protein nodes. Based on the network's topological similarity and sequence similarity, the similarity score between nodes is calculated using the following formula:

[0089]

[0090] Where T(u,v) represents the topological similarity between nodes u and v, Seq(u,v) represents the BLAST scores of nodes u and v, maxSeq represents the maximum BLAST score among all node pairs, and α represents the balancing parameter ∈ (0,1); if α = 1, it means that only topological similarity is used for comparison; if α = 0, it means that only sequence similarity is used for comparison. The final similarity matrix is ​​generated based on the Score value between each pair of nodes. Figure 2 It is not difficult to see that the top 7 nodes of the similarity matrix correspond to the G2 network. Figure 2 The six nodes at the leftmost end of the similarity matrix correspond to network G1. That is, all nodes in network G2 and all nodes in network G1 have a one-to-one similarity score, which in turn forms a similarity matrix.

[0091] Further, in step S5, a global one-to-one comparison result is generated based on the obtained similarity matrix. Since the PPI network structure is sparse, in this embodiment, the two networks to be compared are divided into multiple disconnected sub-modules (a sub-module can be understood as a sub-network composed of several nodes in the network) during the comparison. The comparison is performed in different sub-modules, and then the nodes that are repeatedly compared and isolated nodes are removed. Specifically, it includes: (1) For the comparison within two modules (i.e., comparing the sub-module of network G1 with the sub-module of network G2), the node with the highest similarity score is first selected and added to the comparison result. Then, the comparison is expanded in a greedy way: that is, for the nodes of the two networks to be compared, if the two nodes are matched (the highest similarity score means the comparison is successful), the next comparison will prioritize their neighboring nodes. The neighboring nodes of the nodes that have been successfully matched will be added to the comparison queue. Then, the next set of nodes to be compared will be selected from the neighboring nodes according to the similarity score. This is because functional modules and protein complexes are densely connected and are often separated from other sub-network modules. Considering that when comparing submodules of network G1 with submodules of network G2, submodules of network G1 may be compared with several submodules of network G2, resulting in the same node being compared repeatedly, this embodiment only requires one-to-one comparison. Therefore, inter-submodule comparison is also required, specifically including: (2) For inter-submodule comparison, the ultimate goal of this embodiment is one-to-one comparison. Therefore, for intra-submodule comparisons with repeated comparisons, this embodiment calculates the number of conservative edges to retain comparisons with more conservative edges. Conservative edges refer to edges that exist in both networks, and the corresponding nodes are correctly compared, i.e., the edges that are compared. Finally, the remaining isolated nodes are processed. The processing method for isolated nodes is the prior art and will not be described here.

[0092] The experimental comparison results are as follows:

[0093] To verify the effectiveness and correctness of the method in this embodiment (i.e., the EmbAlign algorithm, which will be used in all subsequent comparative experimental figures), experiments were conducted on both synthetic and real networks. The datasets, alignment algorithms, and evaluation metrics used in the experiments are described below, followed by a detailed analysis of the experimental results.

[0094] Datasets: The synthetic networks are derived from the CG and DMR datasets of NAPAbench, while the real networks are derived from five eukaryotic organisms in IsoBase and four eukaryotic organisms in BioGrid.

[0095] Comparison Algorithm: This embodiment selects existing methods for comparison with the method in this embodiment. The selected existing methods are: MAGNA++, ModuleAlign, IBNAL, AligNet, and BatAlign.

[0096] Evaluation metrics: This embodiment selected four evaluation metrics from both topological and biological perspectives. The topological metrics include S... 3 ICS and LCCE, the biomarker selected is FC, and the detailed definitions are as follows:

[0097] S 3 The Symmetric Substructure Score is an evaluation metric used to penalize the increased proportion of conservative edges caused by comparing dense regions with sparse regions.

[0098] ICS (Induced Conservative Structure) is the score for the conserved inducible structure.

[0099] LCCE (largest Common Connected Edge) represents the number of edges in the largest common connected subgraph in the alignment results.

[0100] FC (Functional consistency) is a biological evaluation index that represents the functional coherence score.

[0101] Experimental Analysis: As shown in Figure 3, the experimental results on the synthetic network show that, on the CG dataset, the method of this embodiment outperforms the method on S... 3 Both achieved optimal results in the FC index, such as Figure 3(a) , 3(c) As shown in Figure 3(b), the method in this embodiment performs slightly worse than the IBNAL algorithm on ICS. On the DMR dataset, the method in this embodiment performs better on S... 3 It performs best on ICS and FC, and its scores are significantly higher than those of MAGNA++ and IBNAL algorithms. In experiments on two combined network datasets, the method of this embodiment showed good performance, which indicates that under ideal conditions without network noise, the method of this embodiment can capture the topological features of the network better than other algorithms, and achieves a high comparison of biometrics while maintaining a high topological score.

[0102] Experiments were conducted on five species datasets in the IsoBase database, and the results are shown in Figure 4 (the order of each comparison group on the horizontal axis of Figure 4 is obtained according to the order of 123456 in the upper right corner). In the experiment, five different species, MUS, CEL, SCE, DME and HSA, were selected and divided into eight different species pairs, and the results are shown in Figure 4. Figure 4(a) , 4(b) S was shown 3The results from ICS show that the method in this embodiment achieved the highest scores on both metrics in the eight sets of experiments, and the scores were significantly higher than those of IBNAL and BatAlign algorithms. Regarding LCCE, as shown in Figure 4(c), the method in this embodiment achieved the best results in 7 out of the 8 sets of experiments, only falling below ModuleAlign in CEL-HSA. In the FC experiments shown in Figure 4(d), the method in this embodiment did not achieve the best results in all species pairs, but it achieved the best results in the first 3 sets of experiments, falling below BatAlign in the 4th set, and slightly falling below AligNet and BatAlign in the other species pair experiments. However, it can be observed that AligNet and BatAlign have higher scores in S... 3 The performance of ICS and LCCE was poor, indicating that they only focused on biological indicators and ignored the importance of network topology. The method in this embodiment achieved a better balance between biological and topological indicators.

[0103] The method of this embodiment was simultaneously tested on four species networks—RN, SP, CE, and MM—from another database, BioGrid. The experimental results are shown in Figure 5 (the order of each comparison group on the horizontal axis of Figure 5 is obtained according to the order of 1, 2, 3, 4, 5, and 6 in the upper right corner). Figure 5(a) shows that this method still achieved the best scores on the S3 index. Similarly, the method of this embodiment also performed well on the ICS and LCCE scores, such as... Figure 5(b) , 5(c) As shown, ICS and LCCE scores were optimal in 3 out of 4 experimental groups. Only in the SP-CE species pair did the ICS score slightly lower than MAGNA++ and ModuleAlign, and the LCCE score was lower than MAGNA++ in RN-CE. The FC scores are shown in Figure 5(d). The method in this embodiment is nearly 2 times higher than MAGNA++ and nearly 1.5 times higher than ModuleAlign and IBNAL. Specifically, the method in this embodiment performs similarly to AligNet in RN-SP and RN-CE, but lower than BatAlign in SP-MM and CE-MM species. However, AligNet and BatAlign perform poorly in S3, ICS, and LCCE metrics.

[0104] Overall, across 16 experiments involving four species pairs, this embodiment achieved the best results in 10 experiments, with good performance in the other groups as well. The real-world network reflects real-world protein-protein interactions. Experiments on two real biological datasets demonstrate that this embodiment can accurately match similar proteins between different species and identify similar functional regions. Furthermore, this embodiment maintains good topological and biological balance in the alignment results, avoiding excessive pursuit of biological indicators at the expense of topological quality.

[0105] Example 2

[0106] This embodiment provides a protein interaction network alignment system, including:

[0107] Acquisition module: Used to acquire structural information of the interaction networks of any two proteins to be compared;

[0108] Hierarchical extraction and calculation module: used to extract the structural features of all nodes in each protein-protein interaction network hierarchically to obtain hierarchical structural features, and calculate the structural scores between nodes based on the hierarchical structural features;

[0109] Topological similarity calculation module: used to construct a multi-layered weighted graph, each layer of which is an undirected complete graph composed of all nodes in the protein-protein interaction network, and uses the structural scores between the nodes as the edge weights of the weighted graph, so that each node performs a random walk on the weighted graph to generate a context sequence;

[0110] It is also used to input the context sequence corresponding to each node in any two protein-protein interaction networks to be compared into a word vector model for training, to obtain a low-dimensional vector corresponding to each node, and to calculate the distance between the low-dimensional vectors corresponding to each node as the topological similarity between nodes.

[0111] Similarity matrix construction module: used to calculate the similarity score between nodes based on the sequence similarity between nodes and the topological similarity between nodes, and to construct the similarity matrix of nodes based on the similarity score between nodes;

[0112] Alignment module: Used to generate node alignment results between two protein interaction networks based on the similarity matrix of the nodes.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0118] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for aligning protein-protein interaction networks, characterized in that: include: Step S1: Obtain the structural information of the interaction networks of any two proteins to be compared; Step S2: Extract the structural features of all nodes in each protein-protein interaction network hierarchically to obtain hierarchical structural features, and calculate the structural scores between nodes based on the hierarchical structural features; The formula for the hierarchical structure feature in step S2 is: Among them, R k (u) represents the degree sequence vector of the k-hop neighbor nodes of node u. d represents the number of k-hop neighbors of node u with degree i, and d represents the ideal maximum value of the network degree. Step S3: Construct a multi-layered weighted graph, where each layer of the weighted graph is an undirected complete graph composed of all nodes in the protein-protein interaction network, and use the structural scores between the nodes as the edge weights of the weighted graph, so that each node performs a random walk on the weighted graph to generate a context sequence. The method for generating a context sequence by having each node perform a random walk on the weighted graph in step S3 includes: the node walking in the current layer of the weighted graph, and the node switching from the current layer to other layers in the weighted graph to generate a context sequence. In step S3, when a node moves through the current layer of the weighted graph, the probability formula for node u moving to node v in the k-th layer of the weighted graph is: Among them, f k (u,v) represents the structural score between node u and node v in the k-th layer weighted graph, where u' and v' represent any node in the node set V, and ∑ u',v'∈V f k (u',v') represents the sum of the weights of the edges between all nodes in the k-th layer weighted graph; In step S3, the probability formula for a node switching from the current layer to other layers in the weighted graph is: Where, p k+1 (u) represents the probability that node u enters the corresponding node in the (k+1)th layer from the kth layer, p k-1 w(u) represents the probability that node u enters the corresponding node in the (k-1)th layer from the kth layer. k ,u k+1 ) represents the edge weight connecting different layers in a multi-layer weighted graph, u k Let u be a node at the k-th level. k+1 Represents node u at level k+1; The context sequence corresponding to each node in any two protein-protein interaction networks to be compared is input into the word vector model for training, and the low-dimensional vector corresponding to each node is obtained. The distance between the low-dimensional vectors corresponding to each node is calculated as the topological similarity between nodes. Step S4: Calculate the similarity score between nodes based on the sequence similarity and topological similarity between nodes, and construct a node similarity matrix based on the similarity score between nodes; Step S5: Generate node alignment results for the two protein interaction networks based on the node similarity matrix.

2. The protein-protein interaction network alignment method according to claim 1, characterized in that: In step S2, the structural score between nodes is calculated based on the hierarchical structure features, using the following formula: Where, N k (u) represents the number of k-hop neighbors of node u in the network, N k (v) represents the number of k-hop neighbors of node v in the network, R k (u) represents the degree sequence vector of the node u's k-hop neighbor nodes, R k (v) represents the degree sequence vector of the k-hop neighbor nodes at a distance of k hops from node v, and dis(R) k (u),R k (v) represents the similarity of the degree sequence vectors of node u and node v, and dis(R) k (u),R k (v))=||R k (u),R k (v)||2, || ||2 represents the L2 norm.

3. The protein-protein interaction network alignment method according to claim 1, characterized in that: The formula for the edge weights connecting different layers in the multi-layer weighted graph is: w(u k ,u k+1 )=log(e+∑num|f k (u,v)>avg(f k )|) Where e represents a constant value, f k (u,v) represents the structural score between node u and node v in the k-th layer weighted graph, avg(f k ) represents the average of the sum of the weights of all edges in the k-th layer weighted graph, and num|·| represents the number.

4. The protein-protein interaction network alignment method according to claim 1, characterized in that: In step S3, the low-dimensional vectors corresponding to each node in the two protein interaction networks are located in the same vector space.

5. The protein-protein interaction network alignment method according to claim 1, characterized in that: In step S4, the sequence similarity between nodes and the topological similarity between nodes are combined to calculate the similarity score between nodes, using the following formula: Where T(u,v) represents the topological similarity between nodes u and v and r(u) represents the low-dimensional vector of node u, r(v) represents the low-dimensional vector of node v, |·| represents the magnitude of the vector, Seq(u,v) represents the BLAST score of node u and node v, maxSeq represents the maximum BLAST score among all node pairs, and α represents the balance parameter and α∈(0,1).

6. A protein interaction network alignment system for implementing the protein interaction network alignment method as described in any one of claims 1-5, characterized in that: include: Acquisition module: Used to acquire structural information of the interaction networks of any two proteins to be compared; Hierarchical extraction and calculation module: used to extract the structural features of all nodes in each protein-protein interaction network hierarchically to obtain hierarchical structural features, and calculate the structural scores between nodes based on the hierarchical structural features; Topological similarity calculation module: used to construct a multi-layered weighted graph, each layer of which is an undirected complete graph composed of all nodes in the protein-protein interaction network, and uses the structural scores between the nodes as the edge weights of the weighted graph, so that each node performs a random walk on the weighted graph to generate a context sequence; It is also used to input the context sequence corresponding to each node in any two protein-protein interaction networks to be compared into a word vector model for training, to obtain a low-dimensional vector corresponding to each node, and to calculate the distance between the low-dimensional vectors corresponding to each node as the topological similarity between nodes. Similarity matrix construction module: used to calculate the similarity score between nodes based on the sequence similarity between nodes and the topological similarity between nodes, and to construct the similarity matrix of nodes based on the similarity score between nodes; Alignment module: Used to generate node alignment results between two protein interaction networks based on the similarity matrix of the nodes.

Images