Protein complex identification method based on knowledge enhancement graph comparative learning

Abstract

The invention discloses a protein complex identification method based on knowledge enhancement graph comparative learning, and belongs to the technical field of bioinformatics. Edge trimming is performed by using a gene expression profile and subcellular localization data, so that a knowledge enhancement PPI network is constructed; and incorporating the molecular function and the biological process as network attributes as attributes for enhancing the network. Graph structure enhancement is applied to knowledge enhanced PPI attribute networks, and multiple perturbation views are generated with diverse random perturbation policies including edge addition, edge discarding, and feature masking to help mitigate data sparsity. A perturbation view generated by graph structure enhancement is coded through a randomized graph convolutional network layer, so that multi-scale features are extracted, key PPI interaction is reserved, and meanwhile prediction performance is improved. And finally, the original PPI network is weighted again, and the protein complex is identified by adopting a core-affiliated strategy.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of bioinformatics and artificial intelligence, and in particular relates to a protein complex recognition method based on knowledge-enhanced graph contrastive learning. Background Technology

[0002] Protein complexes are stable assemblies formed by the non-covalent interactions of two or more proteins. These complexes perform important biological functions in cells, such as signal transduction, metabolic regulation, and gene expression regulation. As the basic functional units of cellular life activities, the accurate identification of protein complexes is of great significance for understanding cellular molecular mechanisms, disease mechanisms, and drug target discovery.

[0003] Currently, protein-protein interaction networks (PPI networks) have become the main research framework for computationally identifying protein complexes. Algorithms based on network topology analysis (such as MCL and CFinder) can effectively identify protein modules with dense connections. With the development of multi-omics data, researchers have begun to integrate auxiliary information such as gene expression profiles and subcellular localization to improve identification accuracy. The introduction of deep learning methods, especially graph neural networks (GNNs), has provided new ideas for extracting high-order features from complex network data.

[0004] However, existing methods still have significant limitations: in terms of data integration, the fusion of multi-source biological data often leads to the destruction of network topology, making it difficult to effectively preserve global connectivity features; traditional graph neural network-based methods struggle to cope with the limited availability of data, which restricts their ability to learn complex protein-protein interaction patterns. While graph contrastive learning (GCL) offers a promising self-supervised alternative, applying random perturbation strategies to enhance graph structures and generate diverse contrastive views effectively enriches training signals and alleviates data sparsity. However, the perturbation strategies employed by traditional GCL methods, such as edge deletion and node feature masking, produce insufficient diversity in contrastive views and may destroy crucial biological interaction information. Furthermore, single encoder architectures struggle to capture the organizational features of protein complexes at different scales, limiting the model's ability to represent multi-level biological features. Summary of the Invention

[0005] In view of this, the present invention proposes a protein complex identification method based on knowledge-enhanced graph contrastive learning, which solves the problems of underutilization of global topological features caused by the current subnetting of biological resources, poor performance of traditional GNN-based methods under data sparsity, and the limitation of representation diversity by random perturbation strategies and uniform node encoders in GCL.

[0006] Therefore, the present invention provides the following technical solution: On one hand, the present invention provides a protein complex identification method based on knowledge-enhanced graph contrastive learning, comprising: The original PPI network is enhanced by integrating gene expression profiles and subcellular localization information. The two types of GO attribute nodes, MF and BP, from the gene ontology database are fused with the knowledge-enhanced PPI network to construct a knowledge-enhanced PPI attribute network. The knowledge-enhanced PPI attribute network is augmented with graph structure through edge addition, edge discarding, and feature masking, modifying the adjacency matrix and related feature matrix, and constructing two contrasting views; The adjacency matrix and feature matrix of the two comparison views are respectively input into the constructed graph convolutional neural network (GCN) encoder with random propagation layers to capture PPI topology and biological features, and the node embedding matrix of the corresponding view is obtained. The node embedding matrix is ​​compared using a loss function to obtain the final node embedding matrix. Based on the embedding of each protein node, the cosine similarity between protein nodes is calculated to obtain a weighted PPI network, from which protein complexes are identified.

[0007] Furthermore, the original PPI network is enhanced by integrating gene expression profiles and subcellular localization information, including: By integrating gene expression profiles and subcellular localization information, knowledge enhancement is achieved by removing edges between low-expressed or unexpressed proteins in the original PPI network, eliminating interactions between proteins located in different subcellular regions, and labeling these removed edges.

[0008] Furthermore, the edge addition includes: randomly adding only edges removed in knowledge enhancement, and the edge discarding and feature masking include: operating on all edges and features in the knowledge-enhanced PPI network.

[0009] Furthermore, the node embedding matrices are compared using a loss function to obtain the final node embedding matrix, which includes: The node embedding matrix is ​​compared and learned by using the loss function calculated by InfoNCE. The encoder parameters are continuously optimized until the set number of rounds is reached, and the final node embedding matrix is ​​obtained.

[0010] Furthermore, protein complexes, including:, were identified from the weighted PPI network. Identify high-density subgraphs as potential protein complexes; A clique mining algorithm is applied to the adjacency matrix of the PPI network to extract fully connected subgraphs containing at least three proteins; the edge weights are derived from protein embeddings to ensure that the cliques reflect topological density and functional synergy. Candidate groups are filtered by evaluating their internal connectivity to preserve a stable core structure; groups are ranked by density score, and the highest-ranked groups are selected as the initial core; overlapping or redundant groups are removed to refine the core set. Peripheral proteins are added to expand the identified core region; an additional score for each non-core protein relative to the core group is calculated, and proteins exceeding a threshold are assigned to the most compatible core; this process is iterated until no more proteins can be added, thereby ensuring the prediction of biologically relevant complexes.

[0011] In another aspect, the present invention also provides a computer-readable storage medium storing a computer instruction set, which, when executed by a processor, implements the aforementioned protein complex recognition method based on knowledge-enhanced graph contrastive learning.

[0012] The above technical solution has the following advantages compared with the existing technology: This invention enhances the original PPI network by integrating gene expression profiling and subcellular localization data, pruning temporally inactive or spatially mismatched edges to improve biological relevance. This knowledge-enhanced PPI network is further enriched by incorporating molecular function (MF) and biological process (BP) attributes extracted from gene ontology (GO). Diverse graph perturbation strategies, including random edge addition, are applied to generate diverse contrastive views on the enhanced PPI network. These views are then processed through random graph convolutional network (GCN) layers to achieve multi-view feature encoding, thereby capturing multi-scale structural and biological patterns. The learned embeddings are used to reweight the original PPI network, and a core-affiliation clustering algorithm is employed to better identify protein complexes. This addresses the underutilization of global topological features resulting from current methods that utilize biological resources for subnetting, and the poor performance of traditional GNN-based methods under data sparsity. Furthermore, the lack of a sufficient GCL in random perturbation strategies and a uniform node encoder limits representational diversity. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the process in an embodiment of the present invention; Figure 2This is a schematic diagram of the protein complex recognition method based on knowledge-enhanced graph contrastive learning in an embodiment of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0016] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] This invention proposes a knowledge-enhanced graph contrastive learning (PGI) method for protein complex identification, which enhances the original PPI network by integrating gene expression profiling and subcellular localization data (corresponding to cellular component annotations in GO). To improve biological relevance, temporally inactive or spatially mismatched edges are pruned. This knowledge-enhanced PPI network is enriched by introducing MF and BP attributes derived from GO. Diverse graph perturbation strategies are implemented on the enhanced PPI network, applying a random perturbation method including edge addition, edge discarding, and feature masking to generate diverse contrastive views. This strategy helps address data sparsity by enriching the graph structure and creating diverse learning signals. These views are then processed through randomized GCN layers, enabling multi-view feature encoding to capture multi-scale structural and biological patterns. The learned embeddings are used to reweight the original PPI network, and a core-affiliation clustering algorithm is applied to identify protein complexes.

[0018] like Figure 1 , Figure 2 As shown in this embodiment, a protein complex identification method based on knowledge-enhanced graph contrastive learning specifically includes the following steps: S1. Using gene expression profiles and subcellular localization data, remove temporally inactive or spatially mismatched edges from the original PPI network, and use two types of GO attribute nodes, MF and BP, from the gene ontology database as network attributes to construct a knowledge-enhanced PPI attribute network. The original PPI network was derived from protein-protein interactions extracted from the DIP, BioGrid, and Krogan14K datasets. Proteins active in the same spatiotemporal context are more likely to interact and form functional complexes. To reduce irrelevant interactions, edges between proteins with low or no expression under specific conditions were removed by integrating gene expression profiles and subcellular localization data, and interactions between proteins located in different subcellular regions were eliminated to enhance the PPI network. Furthermore, the MF and BP of GO were incorporated as network attributes.

[0019] For temporal information: Protein-protein interactions in the protein-protein interaction network are classified using temporal information provided by gene expression profiling. A dynamic threshold function is used to incorporate temporal features, as follows:

[0020] in and They represent proteins respectively. The time expression of mean and standard deviation, Typically set to 1, this threshold sensitivity is adjusted; proteins expressing below this threshold are classified as inactive, and their corresponding edges are removed from the original PIN. These edges are labeled rather than discarded, allowing for reintroduction under random perturbations.

[0021] For spatial information: The spatial context provided by subcellular localization information is crucial for the efficient construction of PPI networks. This information is used to confine interactions to proteins co-localized within the same subcellular compartment, removing edges between proteins in different compartments. To account for potential stochastic effects, these edges are labeled and may be reintroduced during random perturbations to maintain biologically plausible interactions.

[0022] By enhancing the PPI network with spatiotemporal information and introducing MF and BP as attributes, this improved PPI network enhances the performance of protein complex recognition.

[0023] S2: Enhance the graph structure of the knowledge-enhanced PPI attribute network through graph perturbation and construct two contrasting views; Graph perturbation methods include edge addition, edge discarding, and feature masking; these methods modify the adjacency matrix. and related feature matrices .

[0024] Edge Addition: The spatiotemporal knowledge-based PPI attribute network is enhanced by adding edges that were previously removed, based on temporal and spatial information. This is achieved by randomly selecting a subset of labeled edges and adding them back to the adjacency matrix. Specifically, a random matrix with a Bernoulli distribution is sampled. Its elements ,in The probability of adding an edge is defined. This indicates that the adjacency matrix follows a Bernoulli distribution. The updated adjacency matrix is:

[0025] in, It is A matrix of all 1s ( (Number of nodes), ensuring the generated Maintain a binary adjacency matrix, where each element represents the presence of an edge (0 or 1), to prevent duplicate edges from exceeding the binary state. It is a sampled Bernoulli distributed random matrix Its elements ,in The probability of adding an edge is defined.

[0026] Edge discard: Use edge discard with probability. The Bernoulli distribution randomly removes some edges.

[0027] The resulting adjacency matrix is:

[0028] in This represents the Hadamard product (element-by-element multiplication).

[0029] Feature mask: Randomly masks a subset of features in the feature matrix. Each feature dimension is independently masked with probability. It is used for masking.

[0030] The perturbated feature matrix is:

[0031] in It is a random mask vector that indicates which features to retain.

[0032] S3: Input the adjacency matrix and feature matrix of the two comparison views into the constructed GCN encoder with random propagation layers to capture PPI topology and biological features, obtain the node embedding matrix of the corresponding view, compare the node embedding matrix using the InfoNCE loss function, continuously optimize the encoder parameters, and obtain the final node embedding matrix. Specifically, this embodiment proposes a novel encoder framework in which the encoder for each view has a different number of propagation layers. Theoretically, this introduces a propagation depth for each stage. and Furthermore, the depth of each stage is sampled from a uniform distribution.

[0033]

[0034] in, The extent of the propagation layer is defined. The total encoder depth is defined as follows: and ,in To ensure uniqueness. The number of representative stages is set to 2 in this encoder framework. This is because the node embedding is calculated as... and ,in It is a weight matrix, which is used to decompose the graph filter through eigenvector decomposition. As shown below:

[0035] in It is an orthogonal matrix of eigenvectors. A diagonal matrix containing graph filters The eigenvalues, which reflect the frequency components of the graph signal, are given by the following equation based on this decomposition:

[0036] By using different encoder depths and This design enhances the diversity of training. Deeper encoders Emphasizing low-frequency components, while This suppresses high-frequency noise. Therefore, the enhanced view is encoded at a different depth in each iteration, while capturing task-relevant features, promoting randomness and generalization. Based on this principle, different... Operators and Build a view randomization encoder and and share transformation :

[0037]

[0038] Among them, symbols Indicates the composition of operations. Indicates the first Layer transformation operations, This indicates a propagation operation.

[0039] Therefore, the node embedding is generated as follows:

[0040] To optimize these embeddings, the InfoNCE loss is employed:

[0041] in, Represents cosine similarity. It's a temperature parameter. This represents protein pairs interacting in different views of the PPI network. (Negative sample pair set) It includes proteins that do not interact directly. It incorporates randomly associated negative samples from different functional modules within the same view, as well as inter-view negative samples sampled across views, to mitigate spurious correlations. By optimizing this loss, the model enhances its ability to capture biologically meaningful interactions, improves the distinction between functionally relevant and irrelevant proteins, and facilitates more reliable identification of protein complexes.

[0042] S4: Based on the embedding of each protein node, calculate the cosine similarity between protein nodes to obtain a weighted PPI network, and identify protein complexes from the weighted PPI network using a core-affiliated clustering algorithm. Specifically, a core-affiliation clustering algorithm is used to identify protein complexes based on protein embeddings learned from S3. This algorithm comprises three steps: mining high-density communities, filtering core structures, and expanding affiliated nodes. Wherein: S41. High-Density Community Mining: The first step is to identify high-density subgraphs as potential protein complexes. A community mining algorithm is applied to the adjacency matrix of the PPI network to extract fully connected subgraphs containing at least three proteins. Edge weights... From protein embedding and This ensures that the group reflects topological density and functional synergy.

[0043]

[0044] in Indicates protein and The interaction between them.

[0045] S42. Filtering Core Structure: Candidate clusters are filtered by evaluating their internal connectivity to preserve stable core structures. Specifically, density scores are calculated to measure the cohesion within each cluster.

[0046] in Indicates the first The groups are sorted by density score, and the highest-ranked group is selected as the initial core. Overlapping or redundant groups are removed to refine the core set.

[0047] S43, Extending Sub-nodes: Expanding the identified core by connecting peripheral proteins. Calculate each non-core protein. Compared to the core group Connection score:

[0048] Proteins exceeding a threshold are assigned to the most compatible core. This process is repeated iteratively until no more proteins can be added, thus ensuring the prediction of biologically relevant complexes.

[0049] Repeat the above steps until the candidate core set is empty. The small clusters in the seed core set are the cores of the protein complex.

[0050] In the above embodiments, the original PPI network is enhanced by integrating gene expression profiles and subcellular localization data, pruning temporally inactive or spatially mismatched edges to improve biological relevance. This knowledge-enhanced PPI network is further enriched by incorporating molecular function (MF) and biological process (BP) attributes extracted from gene ontology (GO). Diverse graph perturbation strategies are implemented on the enhanced PPI network, applying random perturbation methods including edge addition to generate diverse contrastive views. These views are then processed through random graph convolutional network (GCN) layers to achieve multi-view feature encoding, thereby capturing multi-scale structural and biological patterns. The learned embeddings are used to reweight the original PPI network, and a core-affiliation clustering algorithm is employed to better identify protein complexes. This addresses the underutilization of global topological features caused by current methods that utilize biological resources for subnetting, and the poor performance of traditional GNN-based methods on sparse labeled data. Furthermore, the random perturbation strategies and uniform node encoders with insufficient GCL limit representation diversity.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A protein complex recognition method based on knowledge-enhanced graph contrastive learning, characterized in that, include: The original PPI network is enhanced by integrating gene expression profiles and subcellular localization information. The two types of GO attribute nodes, MF and BP, from the gene ontology database are fused with the knowledge-enhanced PPI network to construct a knowledge-enhanced PPI attribute network. The knowledge-enhanced PPI attribute network is augmented with graph structure through edge addition, edge discarding, and feature masking, modifying the adjacency matrix and related feature matrix, and constructing two contrasting views; The adjacency matrix and feature matrix of the two comparison views are respectively input into the constructed graph convolutional neural network (GCN) encoder with random propagation layers to capture PPI topology and biological features, and obtain the node embedding matrix of the corresponding view. The node embedding matrix is ​​compared using a loss function to obtain the final node embedding matrix. Based on the embedding of each protein node, the cosine similarity between protein nodes is calculated to obtain a weighted PPI network, from which protein complexes are identified.

2. The method according to claim 1, characterized in that, Knowledge enhancement of the original PPI network is achieved by integrating gene expression profiles and subcellular localization information, including: By integrating gene expression profiles and subcellular localization information, knowledge enhancement is achieved by removing edges between low-expressed or unexpressed proteins in the original PPI network, eliminating interactions between proteins located in different subcellular regions, and labeling these removed edges.

3. The method according to claim 2, characterized in that, The edge addition includes: randomly adding only edges removed during knowledge enhancement; the edge discarding and the feature masking include: operating on all edges and features in the knowledge-enhanced PPI network.

4. The method according to claim 1, characterized in that, The node embedding matrices are compared using a loss function to obtain the final node embedding matrix, which includes: The node embedding matrix is ​​compared and learned by using the loss function calculated by InfoNCE. The encoder parameters are continuously optimized until the set number of rounds is reached, and the final node embedding matrix is ​​obtained.

5. The method according to claim 1, characterized in that, Protein complexes were identified from a weighted PPI network, including: Identify high-density subgraphs as potential protein complexes; A clique mining algorithm is applied to the adjacency matrix of the PPI network to extract fully connected subgraphs containing at least three proteins; the edge weights are derived from protein embeddings to ensure that the cliques reflect topological density and functional synergy. Candidate groups are filtered by evaluating their internal connectivity to preserve a stable core structure; groups are ranked by density score, and the highest-ranked groups are selected as the initial core; overlapping or redundant groups are removed to refine the core set. Peripheral proteins are added to expand the identified core region; an additional score for each non-core protein relative to the core group is calculated, and proteins exceeding a threshold are assigned to the most compatible core; this process is iterated until no more proteins can be added, thereby ensuring the prediction of biologically relevant complexes.

6. A computer-readable storage medium storing a computer instruction set, wherein the computer instruction set, when executed by a processor, implements a protein complex recognition method based on knowledge-enhanced graph contrastive learning as described in any one of claims 1 to 5.

Images