A method and device for mining a protein complex, an electronic device and a storage medium

Abstract

The present application relates to the technical field of crop breeding, and discloses a protein complex mining method and device, an electronic device and a storage medium, the method comprising: constructing a protein entity attribute table and an interaction relationship table based on protein information of a target crop, and then constructing an interacting protein knowledge graph; constructing a training set based on the protein entity attribute table and the interaction relationship table and inputting the training set into a protein interaction prediction model to obtain an interaction probability score of interacting protein pairs; then constructing a weighted protein interaction network according to the interaction probability score; and determining an actual protein complex of the target crop based on the weighted protein interaction network and the interacting protein knowledge graph. The present application can realize accurate prediction of protein interaction of the target crop from the perspectives of global topological regularity and local biological characteristics, and the mined protein complex has topological stability and biological rationality, thereby effectively improving the accuracy and interpretability of protein complex mining.

Description

Technical Field

[0001] This invention relates to the field of crop breeding technology, specifically to a method, apparatus, electronic device, and storage medium for discovering protein complexes. Background Technology

[0002] In crop breeding research, protein complexes, formed through protein interactions, often bind to the promoters of downstream genes to regulate gene transcription, playing crucial biological functions in living organisms. Therefore, the potential discovery of protein complexes can help reveal the structure of protein-protein interaction networks, uncover downstream regulatory genes, and better elucidate the molecular regulatory mechanisms of traits, which is key to supporting the breeding of high-quality, high-yield, and multi-resistant new varieties. However, existing methods for predicting protein interactions lack deep semantic connections from multi-dimensional scientific data, limiting themselves to the consideration of single influencing factors and ignoring characteristic interactions such as protein families and subcellular localization. This results in low accuracy in predicting protein interactions, making it difficult to discover crop protein complex structures. Furthermore, the results of protein complex discovery lack biological interpretability and cannot trace key influencing factors. Summary of the Invention

[0003] This invention provides a method, apparatus, electronic device, and storage medium for mining protein complexes, in order to solve the problems of protein complexes being difficult to mine and lacking interpretation.

[0004] In a first aspect, the present invention provides a method for mining protein complexes, the method comprising: acquiring protein information of a target crop, constructing a protein entity attribute table and an interaction relationship table based on the protein information, and constructing an interaction protein knowledge graph based on the protein entity attribute table and the interaction relationship table; constructing a training set based on the protein entity attribute table and the interaction relationship table, and inputting the training set into a pre-constructed protein interaction prediction model to obtain interaction probability scores of interaction protein pairs; constructing a weighted protein interaction network based on the interaction probability scores based on the protein complex discovery model, and determining the actual protein complexes of the target crop based on the weighted protein interaction network and the interaction protein knowledge graph.

[0005] The protein complex mining method provided by this invention constructs a protein entity attribute table and an interaction relationship table based on the protein information of the target crop, and then constructs an interaction protein knowledge graph. A training set is built based on the protein entity attribute table and the interaction relationship table and input into a protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs. A weighted protein interaction network is then constructed based on the interaction probability scores. The actual protein complexes of the target crop are determined based on the weighted protein interaction network and the interaction protein knowledge graph. This invention, by constructing an interaction protein knowledge graph, a fusion protein interaction prediction model, and a protein complex discovery model, overcomes the limitations of traditional methods that rely solely on a single topological structure or feature. It achieves accurate prediction of target crop protein interactions from both global topological patterns and local biological characteristics. The final output protein complexes possess both topological stability and biological rationality, effectively improving the accuracy and interpretability of protein complex mining, and providing reliable technical support for crop breeding and functional genomics research.

[0006] In one optional implementation, protein information of the target crop is obtained, and a protein entity attribute table and an interaction relationship table are constructed based on the protein information. This includes: obtaining a multi-dimensional attribute information file of the proteins contained in the target crop from the Uniprot database, and obtaining a protein interaction relationship dataset between different proteins in the target crop from the PlaPPISite database; performing data preprocessing on the multi-dimensional attribute information file and the protein interaction relationship dataset, and after the data preprocessing is completed, performing entity alignment of the same protein in the multi-dimensional attribute information file and the protein interaction relationship dataset based on the standardized identifier of the protein to obtain the protein information of each protein; and performing structured transformation based on the protein information of all proteins to obtain the protein entity attribute table and the interaction relationship table.

[0007] This invention acquires high-quality data from the Uniprot and PlaPPISite databases and performs entity alignment based on standardized identifiers. This enables the accurate fusion of multi-source heterogeneous protein data, ensuring the accuracy and consistency of the protein entity attribute table and interaction relationship table. This lays a high-quality data foundation for subsequent model training and knowledge graph construction, avoiding errors caused by data heterogeneity or inconsistent identifiers.

[0008] In one optional implementation, a training set is constructed based on a protein entity attribute table and an interaction relation table. The training set is then input into a pre-constructed protein interaction prediction model to obtain the interaction probability scores of interacting protein pairs. This includes: generating positive sample pairs based on the protein interaction relation dataset, randomly sampling to generate negative sample pairs, and determining the multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs based on the protein entity attribute table, thus obtaining a training set containing positive sample pairs, negative sample pairs, and multi-dimensional contextual feature information; calling the protein interaction prediction model and inputting the standardized identifiers of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set into the matrix factorization branch of the protein interaction prediction model to obtain global latent feature vectors; inputting the multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set into the deep factorization machine branch of the protein interaction prediction model to obtain local interaction feature vectors; concatenating the global latent feature vectors and local interaction feature vectors corresponding to each interacting protein pair and inputting them into the gated fusion attention layer of the protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs.

[0009] This invention constructs a dual-branch fusion model of matrix factorization and deep factorization machine to learn protein interaction patterns from two dimensions: global topological regularity and local biological feature interaction. Then, it achieves feature weighted fusion through a gated fusion attention layer. This can preserve the global structural information of the interaction network and explore the complex interaction patterns between multi-dimensional protein features, significantly improving the accuracy of protein interaction prediction and solving the problems of low prediction accuracy and poor interpretability of traditional models.

[0010] In one optional implementation, the standardized identifiers of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set are input into the matrix factorization branch of the protein interaction prediction model to obtain a global latent feature vector. This includes: converting the standardized identifiers of each protein in the interacting protein pair into an identifier vector; and performing a dot product operation on the identifier vectors between proteins in the interacting protein pair based on the dot product interaction layer of the matrix factorization branch to obtain the global latent feature vector.

[0011] This invention enables the matrix factorization branch to efficiently learn the global topological representation of proteins in the interaction network by converting standardized protein identifiers into identifier vectors and performing dot product operations. It captures the overall interaction trend between proteins without relying on specific biological features, complementing the local features of the deep factorization branch. This allows the model prediction to conform to both the network topology and the actual interaction logic within the organism.

[0012] In one optional implementation, the multi-dimensional contextual feature information of the interacting protein pairs corresponding to positive and negative sample pairs in the training set is input into the deep factorization machine branch of the protein interaction prediction model to obtain a local interaction feature vector. This includes: converting the multi-dimensional contextual feature information into a contextual feature vector; weighting the contextual feature vector based on the feature saliency gating layer of the deep factorization machine branch; calculating the weighted contextual feature vector using a dot product attention mechanism on the interaction-aware cross-attention layer of the deep factorization machine branch to obtain a cross-feature vector; concatenating the contextual feature vector and the cross-feature vector, and then inputting the concatenated vector into the factorization machine and deep neural network of the deep factorization machine branch to obtain a local interaction feature vector.

[0013] This invention introduces a feature saliency gating layer and an interaction-aware cross-attention layer into the deep factorization machine branch, which can perform biologically weighted screening and feature matching of multi-dimensional contextual features. This makes the model more focused on biological features that play a key role in protein-protein interactions, rather than simply combinations of mathematical features, which greatly improves the biological interpretability of local interaction feature vectors and further enhances the reliability of prediction results.

[0014] In one optional implementation, based on a protein complex discovery model, a weighted protein interaction network is constructed according to interaction probability scores, and the actual protein complex of the target crop is determined based on the weighted protein interaction network and the interaction protein knowledge graph. This includes: after the protein interaction prediction model converges, calling the protein complex discovery model, selecting interaction protein pairs with interaction probability scores higher than a first preset threshold as interaction edges, and constructing a weighted protein interaction network; using a community detection algorithm to perform topological partitioning on the weighted protein interaction network, identifying dense subgraphs of the weighted protein interaction network, and using these dense subgraphs as candidate protein complexes; obtaining protein information for each protein in the candidate protein complex based on the interaction protein knowledge graph, and calculating the protein complex score of the candidate protein complex based on the protein information; and selecting candidate protein complexes with protein complex scores higher than a second preset threshold as the actual protein complex of the target crop.

[0015] This invention employs a three-step process: threshold screening to construct a weighted PPI network, community discovery to identify candidate complexes, and multi-dimensional scoring verification. This process enables the efficient screening of high-confidence protein complexes from massive prediction results. It not only relies on the network topology but also combines interacting protein knowledge graphs for biological verification, effectively eliminating false positive candidates and ensuring that the final output protein complexes have high authenticity and application value.

[0016] In one optional implementation, calculating the protein complex score of a candidate protein complex based on protein information includes: calculating the sum of the weights of all interacting edges within the candidate protein complex, determining the total number of proteins within the candidate protein complex, and calculating the topological density score of the candidate protein complex based on the sum of weights and the total number of proteins; determining the subcellular localization of proteins within the candidate protein complex based on protein information, and calculating the spatial consistency score of the candidate protein complex based on the number of shared subcellular localizations between interacting protein pairs within the candidate protein complex and the total number of all subcellular localizations; calculating the terminological semantic similarity of each interacting protein pair within the candidate protein complex based on protein information, and averaging the terminological semantic similarity of all interacting protein pairs to obtain the functional similarity score of the candidate protein complex; and weighting and summing the topological density score, spatial consistency score, and functional similarity score based on preset weight coefficients to obtain the protein complex score.

[0017] This invention employs a weighted scoring system based on three dimensions: topological density, spatial consistency, and functional similarity. This system comprehensively quantifies the structural stability, spatial feasibility, and functional synergy of candidate protein complexes, avoiding the limitations of single-dimensional scoring. Furthermore, the introduction of semantic similarity in terminology more accurately reflects the functional synergy between proteins, making the scoring results more consistent with the actual complex formation patterns in vivo, thus enhancing the scientific rigor and accuracy of the screening process.

[0018] Secondly, the present invention provides a protein complex mining device, comprising: a protein information acquisition module, used to acquire protein information of a target crop, and construct a protein entity attribute table and an interaction relationship table based on the protein information, and construct an interaction protein knowledge graph based on the protein entity attribute table and the interaction relationship table; an interaction probability determination module, used to construct a training set based on the protein entity attribute table and the interaction relationship table, and input the training set into a pre-constructed protein interaction prediction model to obtain interaction probability scores of interaction protein pairs; and a protein complex mining module, used to construct a weighted protein interaction network based on the interaction probability scores according to the protein complex discovery model, and determine the actual protein complex of the target crop based on the weighted protein interaction network and the interaction protein knowledge graph.

[0019] Thirdly, the present invention provides an electronic device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the protein complex mining method of the first aspect or any corresponding embodiment described above.

[0020] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the protein complex mining method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific 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 from these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the first process of a protein complex mining method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a second process for a protein complex mining method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the protein knowledge graph construction process of the protein complex mining method according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the decomposition and DeepFM fusion algorithm architecture of the protein complex mining method according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the third process of the protein complex mining method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the protein complex mining process according to an embodiment of the present invention. Figure 7 This is a structural block diagram of a protein complex mining device according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0025] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0026] As an optional application scenario of this invention, the constructed protein complex mining system is Stahl 3.10.0, implemented using the Torch 2.5.0 + CUDA 12.4 framework in a Python 3.8.20 environment, and configured with the Neo4j graph database service. Simultaneously, a pre-built protein interaction prediction model and a protein complex discovery model are included. When executing the protein complex mining method, the system can invoke these models to accurately discover protein complexes that actually exist in the target crop. Therefore, this invention provides a method for mining protein complexes, which effectively improves the accuracy and interpretability of protein complex mining by constructing an interacting protein knowledge graph and a fusion of protein interaction prediction and discovery models.

[0027] According to an embodiment of the present invention, a method for mining protein complexes is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] This embodiment provides a method for mining protein complexes, which can be used on mobile terminals, such as computers. Figure 1 This is a flowchart of a protein complex mining method according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain protein information of the target crop, construct a protein entity attribute table and an interaction relationship table based on the protein information, and construct an interaction protein knowledge graph based on the protein entity attribute table and the interaction relationship table.

[0029] Specifically, in this embodiment of the invention, once the specific type of the target crop (such as Arabidopsis thaliana, rice, wheat, and other common crops) is identified, all protein-related data corresponding to the target crop are obtained from publicly available bioinformatics data sources. The obtained protein information is not single-dimensional data, but a complete set covering the core basic information, structural information, functional information, spatial location information, and potential or verified interaction relationships between proteins. The core basic information includes the protein name, UniProt accession number (normalized unique identifier of a protein), amino acid sequence, etc. The structural information includes Pfam domains, spatial conformation, etc. The functional information includes GO annotations (molecular function, biological process, cellular component, etc.), etc. The spatial location information includes the specific distribution of the protein in the cell (such as the nucleus, chloroplasts, cytoplasm, etc.).

[0030] Furthermore, based on the preprocessed protein information, this embodiment of the invention constructs a protein entity attribute table and an interaction table, respectively. The protein entity attribute table records a single protein as an independent entity, with each row corresponding to a unique protein and each column corresponding to a specific attribute of the protein. The interaction table uses protein pairs as the core recording unit, with each row corresponding to a pair of proteins and recording whether the pair of proteins interacts.

[0031] Furthermore, based on the constructed protein entity attribute table and interaction relationship table, an interaction protein knowledge graph is constructed. This knowledge graph adopts a graph structure, treating each protein as an independent node, the interaction relationships between proteins as edges connecting the nodes, and the attribute information in the protein entity attribute table as the features of the corresponding nodes, achieving an integrated representation of "protein entity - attribute features - interaction relationship". The types of triplet relationships existing in the knowledge graph are shown in the table below:

[0032] For example, Arabidopsis proteins At1g01010 and At1g01020 are used as two nodes, with "interaction" as the edge connecting them. The amino acid sequence, subcellular localization, GO annotation, and other attributes of the two proteins are used as features of the corresponding nodes. This constructs a structured knowledge system that can intuitively reflect the individual characteristics and interaction relationships of proteins, providing standardized and high-quality data support for subsequent training set construction, model prediction, and protein complex mining. This solves the problem of traditional protein data being fragmented and difficult to integrate and utilize.

[0033] Step S102: Construct a training set based on the protein entity attribute table and the interaction relationship table, and input the training set into the pre-constructed protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs.

[0034] Specifically, in this embodiment of the invention, based on the constructed protein entity attribute table and interaction relationship table, a training set for training the protein interaction prediction model is constructed. This process is not a simple data splicing, but rather a targeted sample screening and organization combined with the model training requirements: First, protein pairs that have been verified to interact are extracted from the interaction relationship table and designated as positive sample pairs, i.e., samples labeled "1", representing that the protein pair has a real interaction; at the same time, protein pairs that do not appear in the interaction relationship table and for which there is no evidence of interaction are randomly sampled from all proteins and designated as negative sample pairs, i.e., samples labeled "0", representing that the protein pair is likely not interacting. Simultaneously, according to the protein entity attribute table, all attribute information of the two proteins in each positive and negative sample pair is extracted and integrated into the multi-dimensional contextual feature information of the protein pair, ultimately forming a complete training set containing positive sample pairs, negative sample pairs, multi-dimensional contextual feature information of each sample pair, and sample labels.

[0035] After the training set is constructed, a portion of it is input into a pre-built protein interaction prediction model. This model is a computational model based on deep learning algorithms, capable of quantifying the probability of protein interactions. Its core function is to accurately predict the interaction probability of unknown protein pairs by learning the correlation between the features and labels of protein pairs in the training set. During model training, the model parameters are continuously adjusted to minimize the error between the prediction results and the true labels of the samples until the model converges, i.e., the prediction error stabilizes within a preset range and no longer decreases significantly.

[0036] After the model converges, all samples in the training set (or independent test samples) are input into the model for inference. The model will output a quantitative value between 0 and 1 for each pair of proteins, namely the interaction probability score. The magnitude of the score directly reflects the confidence level of the interaction between the pair of proteins. The closer the score is to 1, the higher the probability that the pair of proteins will interact. The closer the score is to 0, the lower the probability that the pair of proteins will interact.

[0037] Step S103: Based on the protein complex discovery model, a weighted protein interaction network is constructed according to the interaction probability scores, and the actual protein complex of the target crop is determined based on the weighted protein interaction network and the interaction protein knowledge graph.

[0038] Specifically, in this embodiment of the invention, after obtaining the interaction probability scores of all interacting protein pairs, the actual protein complexes of the target crop are mined based on a pre-constructed protein complex discovery model. This embodiment first constructs a weighted protein interaction network based on the interaction probability scores. Specifically, each protein is treated as a node in the network, and the interaction probability score of each protein pair is used as the weight of the edge connecting the two nodes. A first preset threshold (e.g., 0.5) is set, and only protein pairs with interaction probability scores higher than this threshold are retained as edges in the network. Unreliable interaction edges with probability scores lower than the threshold are removed. This constructs a structurally reliable weighted protein interaction network that reflects the strength of connections between proteins. The higher the weight of an interaction edge, the more reliable the interaction relationship between the protein pairs and the tighter the connection between nodes. For example, protein pairs with interaction probability scores ≥ 0.5 are used as network edges, and the edge with a probability score of 0.92 has a higher weight than the edge with a probability score of 0.85, indicating that the interaction relationship between the corresponding protein pairs is more reliable.

[0039] Furthermore, the protein complex discovery model combines the biological information such as protein attributes, functions, and locations contained in the already constructed interaction protein knowledge graph to analyze and screen the weighted protein interaction network. It identifies protein sets with dense structure, synergistic function, matching spatial location, and biological significance from the network structure, and finally determines the actual protein complexes that truly exist in the target crop and can jointly perform specific biological functions.

[0040] The protein complex mining method provided by this invention constructs a protein entity attribute table and an interaction relationship table based on the protein information of the target crop, and then constructs an interaction protein knowledge graph. A training set is built based on the protein entity attribute table and the interaction relationship table and input into a protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs. A weighted protein interaction network is then constructed based on the interaction probability scores. The actual protein complexes of the target crop are determined based on the weighted protein interaction network and the interaction protein knowledge graph. This invention, by constructing an interaction protein knowledge graph, a fusion protein interaction prediction model, and a protein complex discovery model, overcomes the limitations of traditional methods that rely solely on a single topological structure or feature. It achieves accurate prediction of target crop protein interactions from both global topological patterns and local biological characteristics. The final output protein complexes possess both topological stability and biological rationality, effectively improving the accuracy and interpretability of protein complex mining, and providing reliable technical support for crop breeding and functional genomics research.

[0041] This embodiment provides a method for mining protein complexes, which can be used in the aforementioned mobile terminals, such as computers. Figure 2 This is a flowchart of a protein complex mining method according to an embodiment of the present invention, such as... Figure 2As shown, the process includes the following steps: Step S201: Obtain protein information of the target crop, construct a protein entity attribute table and an interaction relationship table based on the protein information, and construct an interaction protein knowledge graph based on the protein entity attribute table and the interaction relationship table.

[0042] Specifically, step S201 includes: Step S2011: Obtain the multi-dimensional attribute information file of the proteins contained in the target crop from the Uniprot database, and obtain the protein interaction relationship dataset between different proteins in the target crop from the PlaPPISite database.

[0043] Specifically, in this embodiment of the invention, Arabidopsis thaliana is used as an example, such as Figure 3 As shown, in the Uniprot database, proteins that have been manually reviewed, annotated, and validated by experts and included in Reviewed (Swiss-Prot) were selected. Multi-dimensional attribute information files for Arabidopsis thaliana were downloaded and filtered to identify the target species (TARGET_SPECIES). These files are typically in FASTA, text, or XML formats. A total of 16,397 Arabidopsis thaliana data entries were obtained, including sequence data, GO annotations, Pfam domains, and subcellular localization data.

[0044] At the same time, such as Figure 3 As shown, the protein interaction relationship dataset for this target crop was obtained from the PlaPPISite database (a specialized database of plant protein interaction sites). This dataset is stored in tabular form, with core fields including "Protein A Identifier", "Protein B Identifier", "Interaction Verification Method", "Interaction Confidence", and "Literature Citations". The "Interaction Verification Method" covers experimental verification methods such as yeast two-hybrid, immunoprecipitation, and fluorescence resonance energy transfer to ensure that the obtained interaction relationships have experimental basis.

[0045] Step S2012: Perform data preprocessing on the multi-dimensional attribute information file and the protein interaction relationship dataset. After the data preprocessing is completed, perform entity alignment of the same protein in the multi-dimensional attribute information file and the protein interaction relationship dataset based on the standardized identifier of the protein to obtain the protein information of each protein.

[0046] Specifically, in embodiments of the present invention, such as Figure 3As shown, data preprocessing of the aforementioned multi-dimensional attribute information file and protein interaction relationship dataset is a core step to ensure the accuracy of subsequent data integration. Specifically, it includes the following sub-operations: ① Data cleaning: For the multi-dimensional attribute information file, invalid protein entries are removed; for the protein interaction relationship dataset, low-quality interaction entries are deleted; ② Format standardization: The multi-dimensional attribute information file is converted from XML format to a unified CSV format for easier structured processing. Simultaneously, protein identifiers in the protein interaction relationship dataset are uniformly converted to UniProt accession numbers; ③ Deduplication: For the multi-dimensional attribute information file, duplicate protein entries are removed based on UniProt accession numbers. For the protein interaction relationship dataset, duplicate interaction entries such as "protein A-protein B" and "protein B-protein A," as well as multiple duplicate records of the same protein pair, are removed.

[0047] Furthermore, after completing data preprocessing, such as Figure 3 As shown, entity alignment is performed based on standardized protein identifiers (i.e., UniProt accession numbers). The specific logic is as follows: using the UniProt accession number as the unique matching key, all attribute information corresponding to a certain accession number in the multi-dimensional attribute information file is associated with all interaction relationships involving that accession number in the protein interaction relationship dataset. This ensures that the attribute information and interaction relationships of the same protein correspond precisely. For example, the attribute information of rice protein with UniProt accession number Q6Z7X8 (amino acid sequence, subcellular location as chloroplast, GO annotation as photosynthesis) is bound to all interaction entries of that protein in the interaction relationship dataset (such as the interaction relationships with Q0J8U9 and Q9LRG7). Finally, complete protein information containing the attribute information and associated interaction relationships of each protein is obtained, solving the problem of inconsistent identification and scattered information of the same protein in multi-source data.

[0048] Step S2013: Perform structural transformation based on the protein information of all proteins to obtain a protein entity attribute table and an interaction table.

[0049] Specifically, in this embodiment of the invention, structured transformation is performed based on the complete protein information of all proteins to construct a protein entity attribute table and an interaction table. Each protein is recorded as a single row, with various protein attributes as column fields. These column fields specifically include UniProt accession number, protein name, gene name, amino acid sequence length, subcellular location, GO molecular function annotation, GO biological process annotation, Pfam domain, etc. Each row is filled with the specific attribute values ​​corresponding to the protein, resulting in the protein entity attribute table. This is merely an example and not a limitation.

[0050] Furthermore, each protein pair is recorded as a row, with core column fields including the UniProt accession number of protein A, the UniProt accession number of protein B, the interaction verification method, the interaction confidence level, and the literature ID. Each row is filled with details of the interaction relationship between a pair of proteins, resulting in an interaction relationship table. This is just an example and is not a limitation.

[0051] Step S202: Construct a training set based on the protein entity attribute table and the interaction relationship table, and input the training set into the pre-constructed protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs.

[0052] Specifically, step S202 includes: Step S2021: Generate positive sample pairs based on the protein interaction relationship dataset, generate negative sample pairs by random sampling, and determine the multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs based on the protein entity attribute table, thereby obtaining a training set containing positive sample pairs, negative sample pairs and multi-dimensional contextual feature information.

[0053] Specifically, in this embodiment of the invention, protein pairs with interaction verification methods of experimental verification (such as yeast two-hybrid, immunoprecipitation) and interaction confidence ≥ 0.8 are selected from the protein interaction relationship dataset, and the standardized identifier (UniProt accession number) of each protein pair is combined as a positive sample pair.

[0054] Meanwhile, to ensure the balance of the training set and the generalization ability of the model, a random sampling method is used to generate negative sample pairs. The specific process is as follows: extract all standardized protein identifiers of the target crop from the protein entity attribute table to construct a complete protein list; exclude all protein pairs already included in the positive sample pairs, including...<A,B> and<B,A> The negative sample pairs are in the form of reverse pairs to ensure that there is no overlap between the negative sample pairs and the positive sample pairs; according to the ratio of 1:1 to 1:2 of the number of positive sample pairs, protein pairs that do not interact are randomly selected from the complete protein list as negative sample pairs.

[0055] Furthermore, after generating positive and negative sample pairs, multi-dimensional contextual feature information corresponding to each sample pair is extracted based on the protein entity attribute table. This feature information is the core dimension characterizing the biological properties of proteins, specifically covering: ① Basic features: protein amino acid sequence length, molecular weight, and isoelectric point; ② Structural features: number of Pfam domains, domain type, and whether it contains a transmembrane region; ③ Functional features: molecular function category, biological process category, and cellular component category annotated by GO (all converted to one-hot encoded vectors); ④ Localization features: subcellular localization type (such as chloroplasts and nuclei, converted to numerical codes); ⑤ Modification features: number of phosphorylation sites and number of glycosylation sites.

[0056] Step S2022: Call the protein interaction prediction model and input the standardized identifiers of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set into the matrix factorization branch of the protein interaction prediction model to obtain the global latent feature vector.

[0057] Specifically, in this embodiment of the invention, a pre-built protein interaction prediction model is invoked. This model is a dual-branch fusion model combining a matrix factorization (MF) branch and a deep factorization machine (DeepFM) branch. For example... Figure 4 As shown, the standardized identifiers of all sample pairs in the training set are input into the matrix factorization branch to learn the global topological features of protein interactions.

[0058] In some optional implementations, step S2022 above includes: Step a1: Convert the normalized identifiers of each protein in the interacting protein pair into an identifier vector.

[0059] Step a2: The dot product interaction layer based on the matrix factorization branch performs a dot product operation on the identifier vectors between proteins in the interacting protein pair to obtain the global latent feature vector.

[0060] Specifically, in this embodiment of the invention, the core input of the matrix factorization branch is the UniProt accession number (normalized identifier) ​​of the protein. First, the discrete identifier is converted into a low-dimensional dense identifier vector through an embedding layer. The dimension is usually set to 64 or 128. For example, the identifier of rice protein Q6Z7X8 is converted into a 64-dimensional vector of [0.12, 0.35, -0.21, ..., 0.47], and Q0J8U9 is converted into a 64-dimensional vector of [0.08, 0.41, -0.18, ..., 0.52].

[0061] Furthermore, the dot product interaction layer of the matrix factorization branch performs a dot product operation on the identifier vectors of each pair of proteins. The calculation process involves multiplying the corresponding elements of the two vectors and then summing them. After normalization, the result forms a global latent feature vector with the same dimension as the identifier vector. This global latent feature vector encodes the global topological association rules of proteins in the entire interaction network. It does not depend on specific biological characteristics and only reflects the overall interaction trend between proteins.

[0062] Step S2023: Input the multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set into the deep factorization machine branch of the protein interaction prediction model to obtain the local interaction feature vector.

[0063] Specifically, in embodiments of the present invention, such as Figure 4As shown, the multi-dimensional contextual feature information of all sample pairs in the training set is input into the Deep Factorization Machine (DeepFM) branch to explore the local interaction patterns between protein features.

[0064] In some optional implementations, step S2023 above includes: Step b1: Transform the multi-dimensional contextual feature information into a contextual feature vector.

[0065] Step b2: The feature saliency weighting of the context feature vector is performed by the feature saliency gating layer based on the deep factorization machine branch.

[0066] Step b3: The interaction-aware cross-attention layer based on the deep factorization machine branch calculates the weighted context feature vector using a dot product attention mechanism to obtain the cross feature vector.

[0067] Step b4: Concatenate the context feature vector and the cross feature vector, and then input the concatenated vector into the factorization machine and deep neural network of the deep factorization machine branch to obtain the local interaction feature vector.

[0068] Specifically, in this embodiment of the invention, the DeepFM branch converts multi-dimensional contextual feature information into a contextual feature vector of a uniform dimension, for example, normalizing all features to the 0~1 range and merging them into a 256-dimensional vector, but is not limited to this. The DeepFM branch then performs the following operations in sequence: ① Feature saliency gating layer: Assigns weight coefficients to each dimension of the context feature vector, and the weight coefficients are determined by the contribution of the features to the protein interaction; ② Interaction-aware cross attention layer: Performs dot product attention mechanism calculation on the weighted context feature vector to capture the interaction relationship between different feature dimensions and outputs cross feature vectors; ③ Core layer (Factorization Machine (FM) + Deep Neural Network (DNN)): After concatenating the original context feature vector and the cross feature vector, the FM layer first learns low-order feature interactions, such as the direct relationship between two structural domains, and then the DNN layer learns high-order feature interactions, such as the complex relationship between structural domains, localization, and function.

[0069] The DeepFM branch ultimately outputs a local interaction feature vector, typically with 128 dimensions, but not limited to this. This local interaction feature vector encodes the local interaction patterns between proteins based on biological characteristics, effectively complementing the global topological features.

[0070] Step S2024: The global latent feature vector and local interaction feature vector corresponding to each interacting protein pair are concatenated and input into the gated fusion attention layer of the protein interaction prediction model to obtain the interaction probability score of the interacting protein pair.

[0071] Specifically, in embodiments of the present invention, such as Figure 4 After extracting the global latent feature vector and the local interaction feature vector, the two feature vectors of each protein pair are concatenated: the global latent feature vector (e.g., 64-dimensional) and the local interaction feature vector (e.g., 128-dimensional) are merged in dimensional order to form a fused feature vector with a dimension of 192. Furthermore, the concatenated fused feature vector is input into the gated fusion attention layer of the model. The core function of this layer is to adaptively adjust the contribution weights of global and local features: the importance coefficients of the two feature branches are learned through a gating mechanism, such as a global feature weight of 0.4 and a local feature weight of 0.6. The fused feature vector is weighted and fused, and then the fusion result is mapped to the 0~1 interval through the Sigmoid activation function. Finally, the interaction probability score of the protein pair is output. The closer the score is to 1, the higher the probability that the protein pair will interact, and the closer it is to 0, the lower the probability.

[0072] This invention employs a dual-branch feature extraction and gated fusion design. It utilizes matrix factorization branches to capture the global topological patterns of protein interactions and DeepFM branches to mine local biological feature interactions, thus addressing the shortcomings of traditional single-branch models that lack global structure or local biological significance. The gated fusion attention layer avoids feature redundancy caused by simple splicing, allowing the model to adaptively allocate global / local feature weights based on the characteristics of different protein pairs, significantly improving the accuracy and interpretability of interaction probability prediction.

[0073] Step S203: Based on the protein complex discovery model, a weighted protein interaction network is constructed according to the interaction probability scores. Then, based on the weighted protein interaction network and the interacting protein knowledge graph, the actual protein complex of the target crop is determined. For details, please refer to [link to details]. Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0074] The protein complex mining method provided by this invention constructs a protein entity attribute table and an interaction relationship table based on the protein information of the target crop, and then constructs an interaction protein knowledge graph. A training set is built based on the protein entity attribute table and the interaction relationship table and input into a protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs. A weighted protein interaction network is then constructed based on the interaction probability scores. The actual protein complexes of the target crop are determined based on the weighted protein interaction network and the interaction protein knowledge graph. This invention, by constructing an interaction protein knowledge graph, a fusion protein interaction prediction model, and a protein complex discovery model, overcomes the limitations of traditional methods that rely solely on a single topological structure or feature. It achieves accurate prediction of target crop protein interactions from both global topological patterns and local biological characteristics. The final output protein complexes possess both topological stability and biological rationality, effectively improving the accuracy and interpretability of protein complex mining, and providing reliable technical support for crop breeding and functional genomics research.

[0075] This embodiment provides a method for mining protein complexes, which can be used in the aforementioned mobile terminals, such as computers. Figure 5 This is a flowchart of a protein complex mining method according to an embodiment of the present invention, such as... Figure 5 As shown, the process includes the following steps: Step S501: Obtain protein information of the target crop, construct a protein entity attribute table and an interaction relationship table based on the protein information, and construct an interaction protein knowledge graph based on the protein entity attribute table and interaction relationship table. For details, please refer to [link to details]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0076] Step S502: Construct a training set based on the protein entity attribute table and interaction relationship table, and input the training set into a pre-constructed protein interaction prediction model to obtain the interaction probability scores of interacting protein pairs. For details, please refer to... Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0077] Step S503: Based on the protein complex discovery model, a weighted protein interaction network is constructed according to the interaction probability scores, and the actual protein complex of the target crop is determined based on the weighted protein interaction network and the interaction protein knowledge graph.

[0078] Specifically, step S503 includes: Step S5031: After the protein interaction prediction model converges, the protein complex discovery model is invoked, and interacting protein pairs with interaction probability scores higher than the first preset threshold are selected as interaction edges to construct a weighted protein interaction network.

[0079] Specifically, in this embodiment of the invention, a pre-defined criterion for determining the convergence of the protein interaction prediction model is set. For example, if the model's prediction error (loss value) on the validation set stabilizes within a preset range for 10 consecutive training epochs and no longer shows a significant downward trend, then the model has fully learned the characteristic patterns of protein interactions in the training set and can stably output reliable interaction probability scores, avoiding prediction bias caused by model non-convergence. Subsequently, interacting protein pairs with interaction probability scores higher than a first preset threshold are selected as interaction edges to construct a weighted protein interaction network.

[0080] The first preset threshold is determined based on experimental verification data calibration, preferably set to 0.5, and can be flexibly adjusted according to the target crop type and data volume. Taking a first preset threshold of 0.5 as an example, such as... Figure 6 As shown, in this embodiment of the invention, only protein pairs with an interaction probability score > 0.5 are retained as interaction edges in the weighted protein interaction network, while low-confidence interaction relationships with a probability score ≤ 0.5 are removed to ensure the reliability of the network structure.

[0081] The constructed weighted protein interaction network uses each protein of the target crop (identified by the UniProt accession number) as a network node, and the interaction probability score of the corresponding protein pair as the weight of the interaction edge connecting the two nodes. The larger the weight value of the interaction edge, the higher the confidence level of the interaction between the protein pair and the tighter the connection between the network nodes.

[0082] Step S5032: The community detection algorithm is used to perform topological partitioning of the weighted protein interaction network, identify the dense subgraphs of the weighted protein interaction network, and use the dense subgraphs as candidate protein complexes.

[0083] Specifically, in this embodiment of the invention, the community detection algorithm built into the protein complex discovery model is used to perform topological partitioning of the weighted protein interaction network. The core objective is to identify dense subgraphs within the network. For example, the Louvain algorithm (or the MCL algorithm, which can be selected based on network complexity) is preferably used as the community detection algorithm. This algorithm can automatically identify communities with tightly connected nodes in the network, i.e., dense subgraphs, and has high computational efficiency and is suitable for topological analysis of large-scale protein interaction networks.

[0084] For example, firstly, the degree (number of connected edges) and the sum of edge weights of each node in the weighted protein interaction network are calculated to measure the tightness of node connections. Then, through iterative optimization, the nodes are divided into different communities, maximizing the sum of edge weights within each community and minimizing the sum of edge weights between communities. Finally, each community (i.e., a dense subgraph with tightly connected nodes and dense interactions) is selected as a candidate protein complex. Since the interactions between proteins within a real biological protein complex are usually extremely tight, the criteria for a dense subgraph are: the average edge weight of all nodes in the subgraph is ≥0.7, and the connection density (actual number of edges / maximum possible number of edges) between nodes in the subgraph is ≥0.6, ensuring that the candidate protein complex possesses the core topological characteristics of a protein complex. This is merely an example and not a limitation.

[0085] Step S5033: Obtain protein information of each protein in the candidate protein complex based on the interacting protein knowledge graph, and calculate the protein complex score of the candidate protein complex based on the protein information.

[0086] Specifically, in this embodiment of the invention, after the candidate protein complex is identified, complete protein information for all proteins in each candidate protein complex is obtained based on the constructed interaction protein knowledge graph. The knowledge graph stores multi-dimensional attribute information such as protein standardized identifiers (UniProt accession numbers), subcellular localization, GO functional annotations (molecular function, biological process), Pfam domains, and protein interaction relationships. Using the protein's UniProt accession number as a search key, all attribute information for each protein in the candidate complex can be quickly retrieved, ensuring the accuracy and completeness of the information. Furthermore, based on the protein information and the structural information of the dense subgraph corresponding to the candidate protein complex, a protein complex score is calculated for that candidate protein complex.

[0087] In some optional implementations, step S5033 includes: Step c1: Calculate the sum of weights of all interacting edges within the candidate protein complex, determine the total number of proteins within the candidate protein complex, and calculate the topological density score of the candidate protein complex based on the sum of weights and the total number of proteins.

[0088] Step c2: Determine the subcellular localization of proteins within the candidate protein complex based on protein information, and calculate the spatial consistency score of the candidate protein complex based on the number of shared subcellular localizations between interacting protein pairs within the candidate protein complex and the total number of all subcellular localizations.

[0089] Step c3: Calculate the semantic similarity of terms for each interacting protein pair within the candidate protein complex based on the protein information, and average the semantic similarity of terms for all interacting protein pairs to obtain the functional similarity score of the candidate protein complex.

[0090] Step c4: The topological density score, spatial consistency score, and functional similarity score are weighted and summed based on preset weight coefficients to obtain the protein complex score.

[0091] Specifically, in this embodiment of the invention, the topological density score of the candidate protein complex is calculated based on the topological structure of the dense subgraph corresponding to the candidate complex in the weighted protein interaction network. The calculation formula is as follows:

[0092] in, Let be the weight of any interacting edge within the candidate protein complex. This represents the number of nodes in a candidate protein complex, i.e., the total number of proteins within the candidate protein complex. Topology density score. It can measure the tightness of the internal interactions of a candidate complex, with a value ranging from 0 to 1. The closer it is to 1, the more stable the topological structure.

[0093] In some optional implementations, embodiments of the present invention, based on the subcellular localization information of each protein in the candidate complex, use the Jaccard coefficient to calculate the spatial consistency score of the candidate protein complex. First, the intersection and union ratios of the subcellular localizations of each protein pair are calculated, and then the average value of all protein pairs is taken. The calculation formula is as follows:

[0094] in, The number of subcellular locations shared by the two proteins. Spatial consistency score is the total number of subcellular localizations of two proteins. It can measure the degree of overlap of the subcellular localization of all proteins in a protein complex in a prior knowledge graph. The value ranges from 0 to 1. The closer to 1, the more consistent the spatial localization of the proteins and the easier it is to form a complex.

[0095] In some optional implementations, embodiments of the present invention, based on the GO annotation information of each protein in the candidate complex, use Resnik semantic similarity calculation. First, the Resnik similarity of each pair of protein GO terms is calculated, and then the average value of all protein pairs is taken to obtain the functional similarity score of the candidate protein complex. Resnik similarity and functional similarity score The calculation formula is as follows:

[0096]

[0097] in, for and In the GO tree structure, all common categories are set. parent node The probability of appearing in the database. That is, nodes The information content. Based on the "biological processes" and "molecular functions" of gene ontology, the functional synergy between proteins in candidate complexes is calculated. Functional similarity score. The value ranges from 0 to 1, and the closer it is to 1, the more synergistic the protein function.

[0098] Furthermore, in this embodiment of the invention, the scores of the above three dimensions are weighted and summed based on preset weighting coefficients to obtain the protein complex score. The preset weighting coefficients include: , and Protein complex score The calculation formula is as follows:

[0099] in, =1, Protein Complex Score The higher the value, the stronger the authenticity and biological plausibility of the candidate protein complex.

[0100] Step S5034: Select candidate protein complexes with protein complex scores higher than the second preset threshold as the actual protein complexes of the target crop.

[0101] Specifically, in this embodiment of the invention, after calculating the scores of all candidate protein complexes of the target crop, the protein complex scores are selected. Candidate protein complexes that exceed a second preset threshold are considered as actual protein complexes of the target crop.

[0102] The second preset threshold is preferably set to 0.7, which can be adjusted according to the proteomic characteristics of the target crop. The purpose of setting this threshold is to eliminate false positive candidate complexes with low scores, loose structures, non-synergistic functions, or mismatched spatial locations, ensuring that the final output actual protein complexes have a high degree of authenticity and biological significance. That is, the screened actual protein complexes are functional units that actually exist in organisms, are formed by the coordinated assembly of multiple proteins, and can jointly perform specific biological functions.

[0103] For example, if the score of the above rice candidate complex is 0.904 > 0.7, it is identified as an actual protein complex (photosynthesis-related complex) of rice; if the score of a candidate complex is 0.65 < 0.7, it indicates that the internal protein functional synergy is insufficient or the topological structure is loose, and it is judged as a false positive candidate and is eliminated.

[0104] This invention summarizes all candidate protein complexes with scores higher than a second preset threshold to form an actual protein complex set for the target crop. This provides reliable research objects for subsequent crop functional genomics research and crop breeding (such as the discovery of stress resistance and high yield-related complexes). At the same time, it solves the technical pain points of high false positives and weak biological significance of protein complexes discovered by traditional methods.

[0105] The protein complex mining method provided by this invention constructs a protein entity attribute table and an interaction relationship table based on the protein information of the target crop, and then constructs an interaction protein knowledge graph. A training set is built based on the protein entity attribute table and the interaction relationship table and input into a protein interaction prediction model to obtain the interaction probability scores of the interacting protein pairs. A weighted protein interaction network is then constructed based on the interaction probability scores. The actual protein complexes of the target crop are determined based on the weighted protein interaction network and the interaction protein knowledge graph. This invention, by constructing an interaction protein knowledge graph, a fusion protein interaction prediction model, and a protein complex discovery model, overcomes the limitations of traditional methods that rely solely on a single topological structure or feature. It achieves accurate prediction of target crop protein interactions from both global topological patterns and local biological characteristics. The final output protein complexes possess both topological stability and biological rationality, effectively improving the accuracy and interpretability of protein complex mining, and providing reliable technical support for crop breeding and functional genomics research.

[0106] This embodiment also provides a protein complex mining device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0107] This embodiment provides a device for excavating protein complexes, such as... Figure 7 As shown, it includes: The protein information acquisition module 701 is used to acquire protein information of the target crop, construct a protein entity attribute table and an interaction relationship table based on the protein information, and construct an interaction protein knowledge graph based on the protein entity attribute table and the interaction relationship table.

[0108] The interaction probability determination module 702 is used to construct a training set based on the protein entity attribute table and the interaction relationship table, and input the training set into the pre-constructed protein interaction prediction model to obtain the interaction probability score of the interacting protein pairs.

[0109] The protein complex mining module 703 is used to construct a weighted protein interaction network based on the interaction probability score according to the protein complex discovery model, and to determine the actual protein complex of the target crop based on the weighted protein interaction network and the interaction protein knowledge graph.

[0110] The protein complex mining apparatus provided in this embodiment of the invention can execute the protein complex mining method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0111] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0112] The following is a detailed reference. Figure 8 This diagram illustrates a suitable structural schematic for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 801, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 802 or a program loaded from memory 808 into random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the electronic device. The processor 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0113] Typically, the following devices can be connected to I / O interface 805: input devices 806 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 807 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 808 including, for example, magnetic tapes, hard disks, etc.; and communication devices 809. Communication device 809 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 8 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0114] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 809, or installed from a memory 808, or installed from a ROM 802. When the computer program is executed by a processor 801, it performs the functions defined in the protein complex mining method of the embodiments of the present invention.

[0115] Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0116] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the protein complex mining method shown in the above embodiments is implemented.

[0117] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0118] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for mining protein complexes, characterized in that, The method includes: The protein information of the target crop is obtained, and a protein entity attribute table and an interaction relationship table are constructed based on the protein information. An interaction protein knowledge graph is then constructed based on the protein entity attribute table and the interaction relationship table. A training set is constructed based on the protein entity attribute table and the interaction relationship table, and the training set is input into a pre-constructed protein interaction prediction model to obtain the interaction probability scores of interacting protein pairs. Based on the protein complex discovery model, a weighted protein interaction network is constructed according to the interaction probability scores, and the actual protein complex of the target crop is determined based on the weighted protein interaction network and the interaction protein knowledge graph.

2. The method according to claim 1, characterized in that, The process of obtaining protein information of the target crop and constructing a protein entity attribute table and an interaction table based on the protein information includes: Obtain multi-dimensional attribute information files of proteins contained in the target crop from the Uniprot database, and obtain a dataset of protein interaction relationships between different proteins in the target crop from the PlaPPISite database; The multi-dimensional attribute information file and the protein interaction relationship dataset are preprocessed. After the data preprocessing is completed, the same protein in the multi-dimensional attribute information file and the protein interaction relationship dataset is aligned based on the protein's standardized identifier to obtain the protein information of each protein. Based on the protein information of all proteins, a structured transformation is performed to obtain the protein entity attribute table and the interaction relationship table.

3. The method according to claim 2, characterized in that, The step of constructing a training set based on the protein entity attribute table and the interaction relation table, and inputting the training set into a pre-constructed protein interaction prediction model to obtain the interaction probability scores of interacting protein pairs includes: Positive sample pairs are generated based on the protein interaction dataset, negative sample pairs are generated by random sampling, and multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs is determined based on the protein entity attribute table, so as to obtain a training set containing the positive sample pairs, the negative sample pairs and the multi-dimensional contextual feature information. The protein interaction prediction model is invoked, and the normalized identifiers of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set are input into the matrix factorization branch of the protein interaction prediction model to obtain the global latent feature vector. The multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set is input into the deep factorization machine branch of the protein interaction prediction model to obtain the local interaction feature vector. The global latent feature vector and the local interaction feature vector corresponding to each of the interacting protein pairs are concatenated and input into the gated fusion attention layer of the protein interaction prediction model to obtain the interaction probability score of the interacting protein pair.

4. The method according to claim 3, characterized in that, The step of inputting the standardized identifiers of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set into the matrix factorization branch of the protein interaction prediction model to obtain the global latent feature vector includes: The normalized identifiers of each protein in the interacting protein pair are converted into identifier vectors; The dot product interaction layer based on the matrix factorization branch performs a dot product operation on the identifier vectors between proteins in the interacting protein pair to obtain the global latent feature vector.

5. The method according to claim 3, characterized in that, The step involves inputting the multi-dimensional contextual feature information of the interacting protein pairs corresponding to the positive and negative sample pairs in the training set into the deep factorization machine branch of the protein interaction prediction model to obtain local interaction feature vectors, including: The multi-dimensional contextual feature information is transformed into a contextual feature vector; The context feature vector is weighted based on the feature saliency gating layer of the deep factorization machine branch; Based on the interaction-aware cross-attention layer of the deep factorization machine branch, the weighted context feature vector is calculated by a dot product attention mechanism to obtain the cross feature vector. The context feature vector and the cross feature vector are concatenated, and the concatenated vector is then input into the factorization machine and deep neural network of the deep factorization machine branch to obtain the local interaction feature vector.

6. The method according to claim 1, characterized in that, The protein complex discovery model constructs a weighted protein interaction network based on the interaction probability scores, and determines the actual protein complex of the target crop based on the weighted protein interaction network and the interaction protein knowledge graph, including: Once the protein interaction prediction model converges, the protein complex discovery model is invoked, and the interacting protein pairs with interaction probability scores higher than a first preset threshold are selected as interaction edges to construct a weighted protein interaction network. The weighted protein interaction network is topologically partitioned using a community detection algorithm to identify dense subgraphs of the network and to identify these dense subgraphs as candidate protein complexes. Based on the interacting protein knowledge graph, protein information of each protein in the candidate protein complex is obtained, and protein complex score of the candidate protein complex is calculated based on the protein information. Candidate protein complexes with scores higher than a second preset threshold are selected as the actual protein complexes of the target crop.

7. The method according to claim 6, characterized in that, The calculation of the protein complex score of the candidate protein complex based on the protein information includes: Calculate the sum of weights of all interacting edges within the candidate protein complex, determine the total number of proteins within the candidate protein complex, and calculate the topological density score of the candidate protein complex based on the sum of weights and the total number of proteins. The subcellular localization of proteins within the candidate protein complex is determined based on the protein information, and the spatial consistency score of the candidate protein complex is calculated based on the number of shared subcellular localizations between interacting protein pairs within the candidate protein complex and the total number of all subcellular localizations. The semantic similarity of terms for each interacting protein pair in the candidate protein complex is calculated based on the protein information, and the semantic similarity of terms for all interacting protein pairs is averaged to obtain the functional similarity score of the candidate protein complex. The protein complex score is obtained by weighting and summing the topological density score, spatial consistency score, and functional similarity score based on preset weighting coefficients.

8. A device for excavating protein complexes, characterized in that, The device includes: The protein information acquisition module is used to acquire protein information of the target crop, construct a protein entity attribute table and an interaction relationship table based on the protein information, and construct an interaction protein knowledge graph based on the protein entity attribute table and the interaction relationship table. The interaction probability determination module is used to construct a training set based on the protein entity attribute table and the interaction relationship table, and input the training set into a pre-constructed protein interaction prediction model to obtain the interaction probability score of the interacting protein pair. The protein complex mining module is used to construct a weighted protein interaction network based on the interaction probability scores according to the protein complex discovery model, and to determine the actual protein complex of the target crop based on the weighted protein interaction network and the interaction protein knowledge graph.

9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the protein complex mining method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the method for mining the protein complex according to any one of claims 1 to 7.

Images