Knowledge graph-based SLC protein-disease association prediction method

Abstract

The invention discloses an SLC protein-disease association prediction method based on a knowledge graph. The method comprises the following steps: 1) constructing an SLC protein-disease knowledge graph based on SLC protein and disease related data; 2) analyzing and mining the knowledge graph by using a convolutional neural network model to obtain the potential association of the SLC protein and the disease in different biological pathways, and the method comprises the following steps: taking two nodes without connecting edges in the knowledge graph as to-be-processed objects; performing graph structure coding on the knowledge graph, and generating embedded vectors of corresponding nodes according to semantic features and structural features of the nodes; the semantic features are text vector representations of the nodes, and the structural features are adjacency relations and path information of the nodes; and inputting the embedded vectors of the two nodes corresponding to the to-be-processed object into the relation scoring function to obtain a potential relation score of the to-be-processed object, and determining potential association existing in different biological pathways of the two nodes corresponding to the to-be-processed object according to a scoring result.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and artificial intelligence, and specifically relates to a knowledge graph-based method for predicting SLC protein-disease associations. It aims to effectively mine and analyze the complex associations between SLC proteins and various diseases across nine biological pathways by constructing and utilizing a multi-attribute relational knowledge graph, combined with R-GCN (Relational Graph Convolutional Network). Background Technology

[0002] The SLC (Solute Carrier) protein family is an important class of transmembrane transport proteins responsible for transporting various small molecules across the cell membrane. Abnormalities in their function are closely associated with a variety of human diseases, such as metabolic disorders, neurological diseases, and cancer. Therefore, systematically identifying the association between SLC proteins and diseases can not only deepen our understanding of the molecular mechanisms of diseases but also provide important clues for the discovery of new drug targets. However, with the expansion of their applications, the complex relationship between SLC proteins and diseases is becoming increasingly prominent, and researchers urgently need an efficient tool to systematically explore and analyze these relationships.

[0003] In traditional research, the association between SLC proteins and diseases relies primarily on the accumulated knowledge of experts and experimental validation. These research methods typically include literature review, experimental validation, and data analysis. Researchers review a large amount of scientific literature to screen for SLC proteins that may be associated with diseases, and then validate these findings experimentally. The main advantage of this method is the high reliability of its results; the experimental data directly support the association between SLC proteins and diseases.

[0004] However, this method also has significant drawbacks. For example, manual literature retrieval and screening are inefficient. As the number of scientific papers increases, researchers face a massive amount of information, making it difficult to comprehensively and systematically review and analyze all relevant literature. This not only increases the workload for researchers but may also lead to the omission of some potential associations between SLC proteins and diseases. Secondly, experimental validation is time-consuming and costly. Each association requires experimental validation, typically involving complex processes such as cell culture and gene editing. These experiments are not only time-consuming but also require substantial amounts of experimental materials and equipment. Furthermore, due to the complexity of experimental conditions, results may differ between different laboratories, further increasing research uncertainty. Therefore, traditional research methods are ill-suited to diverse research needs. Researchers not only focus on the direct relationship between SLC proteins and diseases but also hope to understand their potential associations in different biological pathways. These complex associations are difficult to fully reveal using traditional methods.

[0005] In recent years, with the development of big data technology, knowledge graphs have gradually become an important tool for studying complex systems and uncovering connections. A knowledge graph is a structured representation of knowledge that visualizes different entities and their relationships through nodes and edges. Knowledge graphs can not only integrate heterogeneous data from multiple sources, but also reveal potential connections within complex systems through graph algorithms.

[0006] In the field of bioinformatics, the application of knowledge graphs is becoming increasingly widespread. Researchers have used knowledge graphs to integrate vast amounts of biological information data, such as genomic data, proteomic data, and metabolic network data. Integrating and analyzing this data through knowledge graphs provides researchers with entirely new perspectives, revealing potential connections that are difficult to discover using traditional methods. For example, researchers have used knowledge graphs to discover potential links between certain genes and diseases, driving the development of personalized medicine.

[0007] Knowledge graphs also hold great promise for predicting SLC protein-disease associations. By using knowledge graphs to link SLC proteins with various diseases, researchers can systematically integrate multi-source data, including literature and experimental data, to structurally represent SLC proteins, diseases, their attributes, and relationships. Knowledge graphs not only visually demonstrate the connections between SLC proteins and diseases but also enable the discovery of potential associations through graph algorithms.

[0008] While knowledge graphs have shown great potential in mining SLC protein-disease associations, their application still faces many challenges. For example, data acquisition and processing remain one of the main difficulties in constructing knowledge graphs. In SLC protein and disease research, data sources are diverse, including scientific literature, experimental data, and databases. These data are often heterogeneous, with varying formats, structures, and contents. Effectively collecting, cleaning, and standardizing this data to ensure accuracy and consistency is the primary task in knowledge graph construction. Furthermore, maintaining structural simplicity and information accuracy while ensuring the graph's scale is crucial for constructing high-quality knowledge graphs. As scientific research progresses and new data and knowledge emerge, knowledge graphs need continuous updates and maintenance to ensure their timeliness and practicality. Association mining in knowledge graphs relies on advanced graph algorithms and machine learning techniques. Traditional graph algorithms, such as path analysis and cluster analysis, can reveal some simple relationships in the graph, but their capabilities are limited for complex biological systems. Therefore, researchers began to explore the application of emerging technologies such as deep learning and graph neural networks to the mining of associations in knowledge graphs. Through these technologies, researchers can mine more complex potential associations from large-scale knowledge graphs, such as the implicit association between SLC protein and potential diseases.

[0009] However, these challenges also present new opportunities for researchers. By overcoming difficulties in data acquisition, graph construction, and algorithm application, the application prospects of knowledge graphs in SLC protein-disease association prediction research will be even broader. Researchers can use knowledge graphs to comprehensively and systematically analyze the complex relationship between SLC proteins and diseases, thereby providing theoretical support for disease diagnosis and treatment. Summary of the Invention

[0010] To address the aforementioned problems, this invention provides a method for predicting SLC protein-disease associations based on knowledge graphs. This method systematically integrates data from multiple sources to construct a knowledge graph of SLC proteins and diseases, and utilizes graph algorithms and text correlation techniques to deeply explore the potential associations between SLC proteins and diseases. The following sections will detail the method and steps of this invention from multiple aspects, including overall architecture, data acquisition and preprocessing, knowledge graph construction, association prediction algorithms, result analysis, and applications.

[0011] 1) Overall architecture of the present invention The core of this invention is a system architecture that integrates multi-source data integration, knowledge graph construction, algorithm mining, and prediction result analysis. The overall system architecture includes the following main modules.

[0012] 11) Data annotation module: Obtain disease-related data on SLC protein from different sources, including literature data, experimental data, etc.

[0013] 12) Data preprocessing module: Cleans and standardizes the raw data, removes noisy data, and ensures the accuracy and consistency of the data.

[0014] 13) Knowledge Graph Construction Module: The preprocessed data is structured to construct a knowledge graph of SLC protein-disease, and the nodes and edges in the graph are defined.

[0015] 14) Association Prediction Module: Through graph and convolutional neural network algorithms, the knowledge graph is analyzed and mined to reveal the potential associations between SLC proteins and diseases in different biological pathways.

[0016] 15) Results Analysis and Visualization Module: Analyzes and interprets the discovered relationships, uses visualization technology to display the analysis results, and provides support for research and application.

[0017] Appendix Figure 1 The overall architecture of the present invention is shown.

[0018] 2) Data Acquisition and Preprocessing 21) Data Acquisition Data acquisition is the first and most fundamental step in this invention. The sources of data mainly include the following categories: 211) Annotated scientific literature data: Literature data related to SLC protein and diseases is annotated by experts. Literature data usually includes the title of the paper, SLC, biological pathway, correlation, disease, etc.

[0019] 212) GeneCards and SLC Table databases: These provide a wealth of information about SLC proteins, genomics, protein properties, transport substrates, etc. This data can provide fundamental support for SLC node attributes and node relationships in knowledge graphs.

[0020] 213) MCE database data: Provides a wealth of information on SLC protein agonists and inhibitors. This data can support the node attributes related to SLC in the knowledge graph.

[0021] 214) Diseases Database: Provides a wealth of information on both cancer and non-cancer diseases. This data can support disease-related nodes and edges in a knowledge graph.

[0022] 22) Data Preprocessing Due to significant differences in data formats and structures from different sources, a unified preprocessing operation is required before applying this data to knowledge graph construction. This mainly includes the following steps: 221) Data cleaning: The cleaning operation mainly includes removing duplicate data, handling missing values, and correcting errors in the data. For example, redundant data such as SLC with overlapping markers and disease relationships in the labeled literature, or disease names not clearly stated in the literature, are cleaned.

[0023] 222) Data Standardization: Standardization ensures that data from different sources have a uniform format and structure for subsequent Neo4j map construction. For example, “SLC6A2”, “NAT1”, and “NET” in literature annotations all refer to the same material, but after data standardization, they are all represented as “SLC6A2”.

[0024] 223) Data Integration: This involves integrating data from different sources to form a logically unified whole. The data integration process includes operations such as data mapping, entity alignment, and relation integration. First, data mapping maps field names from different data sources to predefined patterns in the knowledge graph. For example, "cancers" and "Non-tumor disease" in the Diseases database are uniformly mapped to "disease." Second, entity alignment unifies the identification of synonymous entities. For instance, aliases such as "NET" and "NAT1" are aligned to "SLC6A2," and aliases such as "BC" and "Breast carcinoma" are aligned to "breast cancer (BC)." Finally, relation integration merges and removes duplicates of various semantic relationships ("promote," "inhibit," "correlated") between SLC pathways and diseases from different sources, ensuring that the relationship between each pair of entities is unique and semantically clear within the graph. The integrated data forms structured triples and is input into the Neo4j database in a standardized node and edge attribute format for subsequent knowledge graph construction and inference analysis.

[0025] 3) Construction of knowledge graphs Knowledge graph construction is one of the core steps of this invention. Its goal is to express the preprocessed data in a structured way so as to systematically demonstrate the complex relationship between SLC protein and disease.

[0026] 31) The basic structure of knowledge graphs The basic structure of a knowledge graph consists of nodes (entities) and edges (relationships). In the knowledge graph of this invention, nodes can represent entities such as SLC, tumor diseases, non-tumor diseases, agonists, and inhibitors, while edges represent the relationships between these entities, such as "belongs to", "promotes / inhibits", "aggravates", etc.

[0027] 311) Nodes: Nodes represent entities in a knowledge graph. In this invention, the main types of nodes include: SLC nodes: represent SLC proteins that are associated with the disease, such as SLC6A5 and SLC8A1.

[0028] Drug nodes: These represent agonist or inhibitory drugs associated with the SLC protein, such as BAY-876 and Punicic acid. Disease nodes: These represent common cancer and non-cancer diseases that are associated with the SLC protein.

[0029] SLC family node: Represents each SLC family.

[0030] Protein nodes: Represent aliases for each SLC protein.

[0031] Biological pathway nodes: represent the biological pathways between nine different SLCs and diseases.

[0032] 312) Edges: Edges represent relationships between entities. In this invention, the main types of relationships include the following: Promotion / inhibition relationship: This indicates that a certain drug can promote or inhibit the expression of a certain SLC protein.

[0033] Worsening / Delaying: This indicates that a certain SLC has the effect of worsening or delaying the disease on a specific biological pathway.

[0034] Belonging to: Indicates which SLC family each SLC belongs to.

[0035] Alias ​​Relationship: Indicates an alias for SLC.

[0036] 32) Knowledge Graph Construction Process The knowledge graph construction process mainly includes the following steps: entity selection, relation selection, graph construction and storage. 321) Entity selection: Select some data from the labeled data as entity nodes in the knowledge graph.

[0037] 322) Relationship selection: SLC proteins and disease nodes in the same literature are connected through a certain biological pathway using aggravation or delay relationships.

[0038] 323) Knowledge Graph Construction: After entities and relations have been extracted, they are combined to construct a knowledge graph. The constructed graph is a multi-attribute directed graph, with nodes connected by labeled edges. Each node and edge contains multiple attribute information; for example, an SLC node contains attributes such as Protein name, Aliases, Transport type, and Protein function.

[0039] 324) Graph Storage: The completed knowledge graph needs to be stored for subsequent querying and analysis. In this invention, the Neo4j graph database is used for storage.

[0040] 4) Association prediction algorithm After the knowledge graph is constructed, the potential associations between microorganisms and synthetic materials are analyzed in depth using graph information and convolutional neural network algorithms. The association prediction algorithm is one of the key steps in this invention. In this invention, the association prediction algorithm is used to predict the type of disease-related action (promotion or inhibition) of SLC proteins in specific biological pathways based on the constructed SLC-disease knowledge graph. This algorithm combines graph structure information with deep learning methods to achieve intelligent inference of unlabeled relationships.

[0041] 41) Objects to be processed: Objects to be processed are entity pairs in the knowledge graph whose relationships are not yet clearly labeled, mainly including the "SLC—Biological pathway—Disease" triple structure. The attributes of the SLC node include protein name, transport substrate, and functional type; the attributes of the disease node include disease name and classification; and the attributes of the biological pathway node include the specific pathway name. The goal is to determine the effect of SLC on the disease through a specific pathway in this triple, i.e., to predict whether the relationship type is "promoting" or "inhibiting".

[0042] 42) Feature Extraction and Graph Encoding: First, the knowledge graph is encoded using graph structure, transforming nodes and their attribute features into low-dimensional embedding representations. Node features include semantic features (such as BERT text vector representations of SLC and disease names) and structural features (such as node adjacency relationships and path information). At the algorithmic level, a Relational Graph Convolutional Network (R-GCN) is used to model the heterogeneous graph, allowing the propagation weights of each relation type to be learned independently, thereby capturing structural information under multiple relations. After multiple convolutions, the embedding vectors of SLC nodes and disease nodes can fuse semantic information from neighboring nodes and multi-hop paths.

[0043] 43) Relationship Scoring and Prediction: After completing the node embedding calculation, the potential relationships are scored and classified using a relationship scoring function. This invention employs the DistMult model based on tensor decomposition to calculate the relationship between SLC and the disease. For each candidate triplet, its association strength score is obtained through inner product operation. Based on the scoring results, the predicted probability is output using the Sigmoid function to determine whether the relationship belongs to the "promoting" or "inhibiting" type.

[0044] 44) Model Training and Inference: The model is trained using labeled SLC-disease relationships as supervised samples and optimized using a binary cross-entropy loss function. During training, both positive samples (known facilitating / inhibiting relationships) and negative samples (randomly sampled unconnected node pairs) are considered to ensure the model learns the ability to distinguish different relationship types. After training, the model is applied to inference with unknown triples. That is, the input is the SLC, pathway, and disease information to be predicted. After encoding, aggregation, and scoring, the output is the predicted relationship type and its confidence level.

[0045] The following is a detailed description of the four modules.

[0046] I. Data Preprocessing Module Data preprocessing is a crucial step in the entire knowledge graph construction and association prediction process, directly impacting the accuracy and efficiency of subsequent analysis. The main task of this module is to transform raw data into a format suitable for knowledge graph construction and efficient processing by convolutional network algorithms, including data labeling, cleaning, and standardization. The following is a detailed description of the data preprocessing module.

[0047] Data labeling includes data cleaning and standardization.

[0048] Data cleaning and standardization are important steps in data preprocessing, aiming to remove or correct noise, redundancy, and errors in the data. Common data cleaning operations include missing value handling and data standardization.

[0049] Missing value handling: Missing data is handled differently depending on the entity. If the data label is missing the SLC protein name or detailed disease entity, then the labeled data is deleted. If the missing relationship edge between SLC and disease is the relationship between SLC and disease under a certain biological pathway, then the corresponding relationship between SLC and disease in existing literature is searched. If there is no corresponding relationship in existing literature, then experts in the relevant field are consulted to see if a corresponding relationship can be found in existing experiments.

[0050] Data standardization: Data labeling errors, inconsistent capitalization, and other issues require standardization. For example, the labeled data "SLC6A2", "NAT1", and "NET" all represent SLC6A2, so they should be standardized to "SLC6A2".

[0051] II. Knowledge Graph Construction Module The knowledge graph construction module is responsible for transforming the processed data into a structured knowledge graph, which is used for predicting and mining the association between SLC proteins and diseases. The knowledge graph not only displays entities and relationships within the data but also supports subsequent algorithms in discovering implicit connections.

[0052] To construct a domain knowledge dataset, this invention designs an ontology. Within this ontology, we provide entity labels, including data such as the SLC family, SLC proteins, NCBI gene sequence numbers, protein functions, protein sizes, transport substrates, three-dimensional structures, tumors, non-tumor diseases, agonists, inhibitors, drug descriptions, aliases, gene sequences, and immune cells.

[0053] III. Model Training Module In this invention, to fully integrate the structural information of knowledge graphs and the semantic information of document texts, a FastText-RGCN fusion model is proposed. This model employs a dual-branch architecture for joint modeling. The graph structure branch captures the topological relationship features between SLC and the disease, while the text semantic branch extracts semantic vector representations of node-related documents or attributes. These two branches are then merged in a fusion layer to achieve relationship prediction based on multimodal features. The specific training process is as follows: 1. Overall Model Architecture The FastText-RGCN model consists of four core modules: Input and Sample Construction Module: Responsible for extracting training samples from knowledge graphs and labeled documents; Graph Structure Branch (R-GCN Encoder): Used to extract the graph structure features of nodes; The text semantic branch (FastText Encoder): used to generate the text semantic embeddings of nodes; Feature Fusion and Relationship Prediction Module (Fusion&Classifier): After fusing two types of features, it performs relationship prediction, outputs the prediction results, and optimizes the model parameters based on the loss function.

[0054] 2. Model Training Process (1) Graph structure branches: Given the node set and edge set of the knowledge graph, the R-GCN layer first performs convolutional propagation on different relation types: in, For the first Layer node features For a set of relations, The weight matrix is ​​for the relation type. After multi-layer propagation, the model can capture the high-order structural information between different SLC, pathway, and disease nodes. Then, dimensionality reduction and feature aggregation are performed through pooling layers, and fully connected layers are connected to obtain the high-level semantic feature representation of the nodes. (2) Text semantic branch For each node (such as SLC or disease), its corresponding text attributes are collected and input into the FastText word embedding layer. The FastText model converts each word into a 300-dimensional pre-trained vector to represent the morphological and semantic features of the word. Subsequently, all word vectors of the text of the node are averaged: A fixed-length node text semantic vector is obtained. This vector reflects the global semantic information of the node. (3) Feature fusion and relation prediction For each candidate triple, the model obtains: graph structure embedding from the R-GCN branch. and text embedding from the FastText branch The two are then spliced ​​and weighted in the fusion layer: in For learnable weights, This refers to the final node features that integrate structural and textual information. Feature fusion is performed on SLC and disease separately to obtain the fused features. and Features are then input into the DistMult relation scoring function: in A trainable diagonal matrix representing a specific biological pathway is used to map scores to probability values ​​in the range [0, 1] using a sigmoid function, with a threshold of 0.5. If... If a synergistic relationship is predicted, it indicates that SLC tends to promote disease progression through this pathway. Conversely, if , indicating that no predicted relationship was obtained. (4) Loss function and optimization: The binary cross-entropy loss function is used for training: in This represents the true labeling relationship (1 indicates promotion, 0 indicates inhibition). The model predicts probabilities. The model parameters are updated using the Adam optimizer to make the predicted relationships as close as possible to the true labels.

[0055] Appendix Figure 3 This describes the overall architecture of the model training module.

[0056] Compared with the prior art, the positive effects of the present invention are as follows: 1) Fine-grained relationship prediction for SLC protein mechanism of action: Existing drug synergy prediction methods mainly focus on the synergistic effect of drug combinations, while this invention focuses on the specific type of action (promotion / inhibition) of SLC transporter proteins in regulating diseases at the biological pathway level, realizing fine-grained reasoning from "whether it is related" to "how it works", which is more biologically interpretable and clinically applicable.

[0057] 2) Multi-source node embedding that integrates semantic and structural features: Compared with traditional knowledge graph embedding methods (such as the DistMult model), this invention introduces relational graph convolutional networks (RGCN) to integrate semantic features and graph structural features (such as multi-hop path information) in node embedding, which effectively improves the richness of node representation and context awareness.

[0058] 3) Relationship-aware graph convolution that supports heterogeneous graph relationship modeling: Compared with ordinary GNN methods, this invention uses RGCN to independently model different relationship types, so that different types of relationships such as "promotion" and "inhibition" have different weights in the information propagation process, which is more suitable for modeling complex regulatory relationships in biological pathways.

[0059] 4) Relationship scoring and interpretability output based on DistMult: Based on node embedding, this invention uses the DistMult model to score relationships and combines it with the Sigmoid function to output the probability of "promotion / inhibition". This not only provides high confidence in the prediction results, but also allows the scoring to be traced back to key path nodes, thus enhancing the interpretability of the model.

[0060] 5) Strategy design for balanced training of positive and negative samples: Compared with the method that only relies on positive samples for training, this invention introduces a negative sampling strategy during the training process. By randomly sampling unconnected node pairs to construct negative samples, the model can be effectively avoided from overfitting and the generalization ability to unknown relationships can be improved. Attached Figure Description

[0061] Figure 1 This is a diagram of the overall architecture of the present invention.

[0062] Figure 2 A flowchart for data preprocessing.

[0063] Figure 3 This describes the overall architecture of the model training module. Detailed Implementation

[0064] The details of this invention will now be described in detail.

[0065] I. Data Conversion 1. Text Data Conversion To incorporate the textual attributes of nodes into the R-GCN model, we employ the FastText model to convert text into vector representations. FastText, an efficient and powerful word embedding tool, is an extension and optimization of Word2Vec. It learns the semantic relationships between words, mapping vocabulary to a low-dimensional real-valued vector space, providing a foundation for subsequent model computation. FastText primarily implements word embeddings through two models: Continuous Bag of Words (CBOW) and Skip-Gram. CBOW predicts the word itself based on its contextual vocabulary, emphasizing the aggregation of semantics from the context. Skip-Gram, given a word, predicts the surrounding contextual vocabulary, focusing on allowing the semantics of a word to diffuse into its context.

[0066] When training the FastText model, the first step is to concatenate the structured attributes of each SLC node into a single attribute description string. This string represents the "document" of that node. This process is repeated for all SLC nodes to obtain a corpus consisting of multiple documents. Where N is the total number of SLC nodes. Next, for each document obtained... After standardization, the result is ', using the preprocessed entire corpus To train the FastText model.

[0067] The pre-trained FastText model is essentially a look-up table; given a word w, it outputs a d-dimensional word vector. The final vector of one of the words w It is the n-gram vector of all its characters. The sum of . Let the set of n-grams of the word w be . Its vector representation is: in This is the vector learned by the model for each n-gram g. This process is automatically performed during training using the Skip-gram or CBOW objective function.

[0068] This mechanism enables FastText to construct vectors for words that have never appeared in the training corpus (OOV, Out-of-Vocabulary), greatly enhancing the model's generalization ability, especially when dealing with biomedical terms with complex roots and affixes.

[0069] Finally, the trained FastText model and the preprocessed word list for each node v are used. Convert the text information into a FastText average vector. For a given text, convert it into a FastText average vector. each word element The described node Its initial embedding vector (This will also serve as the 0th layer representation of R-GCN) The calculation is as follows: in It is a word Obtained from the FastText model Dimensional vector.

[0070] After this step, each SLC node is represented as a dense, fixed-dimensional vector that captures the semantic information of its textual attributes.

[0071] 2. Graph structure information Graph Convolutional Networks (GCNs) are a class of specialized neural networks designed to process graph-structured data, effectively capturing the inherent relational information within the graph. The propagation rules of GCNs involve the graph's adjacency and degree matrices, ensuring appropriate scaling and aggregation of features to maintain the graph's structural integrity. They are particularly useful in fields such as social network analysis, recommender systems, and molecular biology, where the underlying data naturally forms graphs. Despite their advantages, GCNs also face challenges such as computational complexity and the oversmoothing problem, where node representations become difficult to distinguish as the number of layers increases.

[0072] RGCN is an extension of Graph Neural Networks (GNNs) for processing heterogeneous graphs, i.e., graphs with multiple different types of nodes and edges. In standard Graph Convolutional Networks (GCNs), edge type information is typically ignored, while RGCN preserves edge type information by introducing relation matrices and corresponding transformations, thus better capturing the rich semantic relationships in the graph. In this invention, RGCN is extended and customized for the SLC-disease interaction prediction task, focusing on solving the representation learning and relation reasoning problems of multi-relationship nodes (such as SLC proteins, biological pathways, and diseases) in heterogeneous knowledge graphs. Compared with the commonly used RGCN models in the prior art, this invention has made several optimizations and innovations in structural design, feature fusion, and training strategies. In standard GCNs, edge type information is ignored, while R-GCN preserves semantic differences by introducing independent transformation matrices for each relation type, thus better capturing the diverse relationships between entities such as SLC, diseases, and pathways. Let the input of the RGCN model be a multi-relationship graph, represented as: in, It is a set of nodes. It is a set of edges. It is a set of relations, and each layer of RGCN updates the representation of a node, a process that depends on the node's neighbors and their connections. For a node... The representation of exist In this layer, RGCN updates by aggregating information from neighboring nodes. The specific update formula is as follows: in Indicates the relationship Connect to node The set of neighboring nodes; It is a relationship In the The transformation matrix of the layer; It is a transformation matrix of the node's own information; It is a normalization constant used to adjust the contribution of different relation types to node representation updates; It is a non-linear activation function, such as ReLU.

[0073] In RGCN, different types of relations may contribute differently to node representation updates. By applying different weights to each relation type, RGCN can flexibly aggregate information from different relations; these weights are transformed through a matrix. ,accomplish: This formula demonstrates how to perform a weighted summation of information from different relation types and apply a non-linear transformation to the node representation using an activation function.

[0074] Because the number of parameters in the RGCN model increases with the number of relation types, it can lead to overfitting, especially with limited training data. To address this issue, RGCN introduces edge-type-based regularization, which is achieved by applying a transformation matrix W... r Decomposing the model into a product of low-rank matrices reduces the number of model parameters. in It is a shared fundamental transformation matrix. and It is a matrix specific to relation r.

[0075] II. Model Training (1) Graph structure branches After inputting the node set and edge set of the knowledge graph, the R-GCN layer first performs convolutional propagation on different relation types: Through multi-layer convolutional propagation, the model can capture the high-order structural dependencies between SLC, disease, and biological pathway nodes. Then, the pooling layer is used for feature aggregation and dimensionality reduction to obtain the structural feature vector of the node. After passing through a fully connected layer and the ReLU activation function for nonlinear mapping, the structural features of the node are projected into a new feature space, which is convenient for fusion with semantic features. (2) Text semantic branch For each node (e.g., SLC or disease), its corresponding text attributes are collected and input into the FastText word embedding layer. FastText maps each word to a 300-dimensional word vector to capture the morphological and semantic features of the word. All word vectors of the node text are then averaged. This yields a fixed-length text semantic vector. As a semantic representation of a node, this vector reflects the global meaning of the node in the semantic space of the corpus. (3) Feature fusion and relation prediction For each candidate triple, the model extracts the structural embeddings from the R-GCN branch respectively. and semantic embeddings from the FastText branch The two are then weighted and spliced ​​at the fusion layer: in For learnable weights, This refers to the final node features that integrate structural and textual information. Feature fusion is performed on SLC and disease separately to obtain the fused features. and Features are then input into the DistMult relation scoring function: in A trainable diagonal matrix representing a specific biological pathway is used to map scores to probability values ​​in the range [0, 1] using a sigmoid function, with a threshold of 0.5. If... If a synergistic relationship is predicted, it indicates that SLC tends to promote disease progression through this pathway. Conversely, if , indicating that no predicted relationship was obtained. (4) Loss function and optimization model training adopts the binary cross-entropy loss function: in This represents the true labeling relationship (1 indicates promotion, 0 indicates inhibition). These are the model's predicted values. The Adam optimizer is used for generational updates to minimize the prediction error in each training iteration.

[0076] Furthermore, to measure the model's performance in relation prediction tasks, this invention uses the Mean Reciprocal Rank (MRR) as the primary evaluation metric. Its calculation formula is as follows: in The total number of queries in the set. Indicates the first The ranking position of the first correct result in a query; the higher the MRR value, the stronger the model's ability to rank true relationships at a high position.

[0077] The FastText-RGCN model achieves high-precision prediction of the relationship between SLC and disease by jointly modeling graph structure and semantic information. It takes into account both the structural constraints of the graph and the semantic richness of the literature information, effectively overcoming the problem of incomplete information in single-modality methods, and providing important intelligent technical support for the study of SLC protein and tumor disease mechanisms.

[0078] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention should be determined by the claims.

Claims

1. A knowledge graph-based method for predicting SLC protein-disease associations, comprising the following steps: 1) Construct a knowledge graph of SLC proteins and diseases based on data related to SLC proteins and diseases. The knowledge graph includes nodes such as SLC nodes, drug nodes, disease nodes, SLC family nodes, protein nodes, and biological pathway nodes. SLC nodes represent SLC proteins associated with diseases; drug nodes represent agonist or inhibitor drugs associated with SLC proteins; disease nodes represent common oncology and non-oncology diseases associated with SLC proteins; SLC family nodes represent an SLC family; protein nodes represent aliases for each SLC protein; and biological pathway nodes represent biological pathways between SLC proteins and diseases. Edges between nodes are determined based on their relationships, including membership, facilitation, inhibition, aggravation, delay, and aliasing relationships. The attributes of SLC nodes include protein name, transport substrate, and functional type; the attributes of disease nodes include disease name and classification; and the attributes of biological pathway nodes include specific pathway names. 2) The knowledge graph is analyzed and mined using a convolutional neural network model to obtain the potential associations between SLC proteins and diseases in different biological pathways. The method is as follows: 21) Select two nodes in the knowledge graph that do not have a connecting edge as objects to be processed; 22) Perform graph structure encoding on the knowledge graph, and generate the embedding vector of the corresponding node based on the semantic features and structural features of the node; the semantic features of the node are the text vector representation of the node, and the structural features of the node are the adjacency relationship and path information of the node; 23) Input the embedding vectors of the two nodes corresponding to the object to be processed into the relationship scoring function to obtain the potential relationship score of the object to be processed, and determine the potential associations of different biological pathways between the two nodes corresponding to the object to be processed based on the scoring results.

2. The method according to claim 1, characterized in that, The convolutional neural network model includes a graph structure branch, a text semantic branch, and a feature fusion and relationship prediction module. The graph structure branch is used to capture the topological relationship features of nodes. The text semantic branch extracts the semantic vector representation of nodes. The feature fusion and relationship prediction module merges the topological relationship features and semantic vector representation of nodes to obtain the embedding vector of nodes. Then, a relationship scoring function is used to score the potential relationship of the object to be processed based on the embedding vector of two nodes in the object to be processed. Based on the scoring results, the potential association between the two nodes corresponding to the object to be processed in different biological pathways is determined.

3. The method according to claim 1 or 2, wherein the method for training the convolutional neural network model is as follows: using the triples in the knowledge graph as training samples to train the convolutional neural network model; and using the binary cross-entropy loss function to optimize the convolutional neural network model during training.

4. The method according to claim 1, characterized in that, Based on the semantic features of the nodes and structural features Weighted fusion is performed to generate the embedding vector of the corresponding node. .

5. The method according to claim 1, 2, or 3, characterized in that, The relational scoring function is the DistMult model based on tensor decomposition.

6. A server, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images