Entity semantic enhancement-based protein interaction extraction method
By employing entity semantic enhancement methods, combined with entity attention mechanisms and graph attention networks, the problems of data scarcity and information fusion in protein-protein interaction extraction from biomedical literature are addressed, achieving higher extraction accuracy and performance.
Patent Information
- Application Number
- CN202511725442.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies face challenges in protein-protein interaction extraction tasks from biomedical literature, including data scarcity, the separation of semantic and structural information caused by data augmentation methods, and the difficulty of effectively integrating entity semantic and structural information. These issues affect the model's accurate understanding of protein entity relationships.
We adopt an entity semantic enhancement approach, which obtains context-enhanced relational representations through entity attention mechanisms, integrates multi-dimensional semantic information, and uses BERT encoding combined with graph attention networks to generate fused representations of protein pairs, effectively integrating text semantics and graph structure information.
This method significantly improves the accuracy and performance of protein interaction extraction, surpassing existing methods. It demonstrates excellent micro-average F1 scores on multiple PPI datasets, validating the effectiveness and advancement of the method.
Smart Images
Figure CN121565262A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of biomedical literature mining technology, and in particular relates to a protein interaction extraction method based on entity semantic enhancement. Background Technology
[0002] In recent years, with the continuous deepening of biomedical research, the number of biomedical literatures has grown exponentially. These literatures contain a wealth of biomedical knowledge and are of great significance for advancing research and applications in the medical field. However, for researchers, reading such a large volume of literature presents enormous challenges. Therefore, automatically extracting structured knowledge from biomedical literature has become a crucial task. Among these tasks, the automatic extraction of protein-protein interaction (PPI) relationships is one of the key tasks in the field of biomedical literature mining and holds a very important position.
[0003] Automated extraction of protein-protein interactions is an important research direction in biomedical text mining. Its core challenge lies in how to accurately identify and structurally characterize the complex relationships between proteins from massive amounts of literature. With the exponential growth of biomedical literature, traditional manual annotation methods are no longer sufficient to meet research needs, prompting researchers to develop automated extraction techniques based on natural language processing (NLP).
[0004] Early PPI extraction systems primarily relied on rule templates and shallow machine learning methods, but these were limited by the complexity of feature engineering and domain adaptability issues. In recent years, deep learning techniques, especially the introduction of pre-trained language models (such as BERT), have significantly improved the performance of PPI extraction by capturing the unique semantic and syntactic features of biomedical text.
[0005] However, current PPI extraction techniques still face many challenges. First, the high cost of building high-quality PPI corpora and the limited size of existing labeled datasets make data scarcity a key factor restricting model performance. Second, traditional data augmentation methods often result in a disconnect between semantic and structural information, affecting the model's accurate understanding of protein entity relationships. Furthermore, effectively integrating semantic and structural information of entities to improve the accuracy of relationship representation remains a critical issue that urgently needs to be addressed. Summary of the Invention
[0006] The purpose of this invention is to provide a protein interaction extraction method based on entity semantic enhancement, which aims to solve the problems mentioned in the background art.
[0007] The present invention is implemented as follows: a protein interaction extraction method based on entity semantic enhancement includes the following steps: Step 1: Map the input sequence to an embedding representation containing contextual semantic information, and obtain a context-enhanced relational representation through an entity attention mechanism; Step 2: Integrate protein entities with multi-dimensional semantic information to construct an enhanced entity representation; Step 3: Use BERT encoding and graph attention network to obtain graph structure information of protein entities from the text, fuse the obtained graph structure information with text semantic information to generate a fused representation of protein pairs for interaction relationship determination.
[0008] A further technical solution, the specific steps of step 1 are as follows: Step 1.1: Input sequence. Insert entity markers into the input sequence to explicitly identify the locations of protein entities of interest. For sequences containing protein pairs... The sentence, i.e., the input sequence Convert it to ; (1); (2); in, For the input sequence The first in One word, For the input sequence The total number of words in the Chinese dictionary; and These are the classification and delimiter labels in the BERT model, respectively. , , and These are entity markers used to identify the location of two protein entities; Step 1.2: Encode the sequence using the BioBERT model to obtain the context hidden state representation corresponding to each token. ; (3); in, , express The first in The hidden state of each token, for The total number of tokens in the game. for 3D real space; for entity labeling and Corresponding hidden state and The attention weights of these weights for all locations are calculated. and ; (4); (5); in, express The first in The hidden state of each token.
[0009] Step 1.3: Add the attention probabilities of the two entities together, and select the one with the highest attention weight based on the merged attention score. The tokens are assigned to several positions, and max-pooling is performed on these tokens to construct a context representation. Finally, the pooled result is integrated into the relation representation to obtain the sum of the attention probabilities of the two entities. and the context representation after pooling Relationship representation after integration ; (6); (7); (8); in, This indicates a max pooling operation. This indicates selecting the option with the highest attention weight. One position.
[0010] A further technical solution, the specific steps of step 2 are as follows: Step 2.1: Utilize a pre-trained large language model to analyze protein entities in biomedical literature. The system identifies and generates a multi-dimensional semantic information set. (9); in, For the first dimension of semantic information A protein sub-semantic fragment; Protein entities A semantic set; This represents the semantic output obtained through a large language model; Step 2.2: Analyze each protein entity using BioBERT Encode to obtain vector BioBERT encoding is used for each multi-dimensional semantic information. Obtain the embedding vector For each protein entity and Cosine similarity is used to measure their similarity. ; (10); in, Describes the norm of vector L2. It is the dot product of the two; Step 2.3: Use a weighted average method to fuse the representations of the original entities and the newly added semantic information to obtain the enhanced entity representation vector. : (11); in, It is a balancing parameter used to control the weights of the original entity representation and the newly added semantic information representation.
[0011] A further technical solution, the specific steps of step 3 are as follows: Step 3.1: Model the protein-protein interaction network using a graph attention network (GAT); (12); in, For nodes In the Layered graph structure embedding representation; It is a node The set of neighboring nodes; It is a node For neighboring nodes Attention coefficient; For nodes In the Layered graph structure embedding representation; It is the first The layer weight matrix is calculated as follows: (13); in, The LeakyReLU activation function is used. This indicates a splicing operation. For nodes Graph embedding representation, This is the weight matrix. For each protein node, there is an attention parameter vector. go through After layered graph attention network, graph structure embedding is obtained ; (14); Step 3.2: Align the textual semantic representation and graph embedding representation of protein entities in the feature space; (15); (16); in, and It is an aligned representation of text and image; and These are the linear transformation matrix and bias term of the text modality, respectively; and These are the linear transformation matrix and bias term of the graph mode, respectively; Protein entities Text embedding vectors obtained after BioBERT encoding; Step 3.3: Define the dynamic gating mechanism, as shown in formula (18): (17); (18); in, and These are the weight matrix and bias term of the gating mechanism, respectively; This indicates the concatenation of two aligned representations; These are the calculated gating weights, used to control the contribution level of each mode; It is a protein entity The fusion representation; Step 3.4: Protein pair ( Introducing interaction terms to model the relationship between the two, we obtain protein pairs ( The final representation of ) ; (19); in, It is the element-order product of the fusion of two proteins. Protein entities The fusion representation.
[0012] This invention provides a protein interaction extraction method based on entity semantic enhancement. This method utilizes an attention mechanism to effectively identify and capture contextually relevant words through entity tagging and attention distribution, constructing a more accurate relation representation. Secondly, based on a large language model, it generates multiple semantic representations of protein entities and enhances the semantic richness and contextual adaptability of these representations through an adaptive fusion mechanism. Finally, through multimodal language interaction, it comprehensively models the complex relationships between proteins from both textual semantics and graph structure dimensions. Extensive experimental validation was conducted on five standard PPI datasets. Experimental results show that this method significantly outperforms existing methods in the PPI relation extraction task, achieving optimal performance and effectively improving the overall performance of protein interaction extraction from biomedical literature. Attached Figure Description
[0013] Figure 1 This is an overall architecture diagram of a protein interaction extraction method based on entity semantic enhancement provided in an embodiment of the present invention. Detailed Implementation
[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0015] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.
[0016] like Figure 1 As shown, an embodiment of the present invention provides a protein interaction extraction method based on entity semantic enhancement, comprising the following steps: Step 1: Map the input sequence to an embedding representation containing rich contextual semantic information, as follows: Step 1.1: Input sequence. Insert entity markers into the input sequence to explicitly identify the locations of protein entities of interest. For sequences containing protein pairs... The sentence, i.e., the input sequence Convert it to ; (1); (2); in, For the input sequence The first in One word, For the input sequence The total number of words in the Chinese dictionary; and These are the classification and delimiter labels in the BERT model, respectively. , , and These are entity markers used to identify the location of two protein entities; Step 1.2: Encode the sequence using the BioBERT model to obtain the context hidden state representation corresponding to each token. ; (3); in, , express The first in The hidden state of each token, for The total number of tokens in the game. for 3D real space; for entity labeling and Corresponding hidden state and The attention weights of these weights for all locations are calculated. and ; (4); (5); in, express The first in The hidden state of each token.
[0017] Step 1.3: Add the attention probabilities of the two entities together, and select the one with the highest attention weight based on the merged attention score. The tokens are assigned to several positions, and max-pooling is performed on these tokens to construct a context representation. Finally, the pooled result is integrated into the relation representation to obtain the sum of the attention probabilities of the two entities. and the context representation after pooling Relationship representation after integration ; (6); (7); (8); in, This indicates a max pooling operation. This indicates selecting the option with the highest attention weight. One location; Step 2: Integrate protein entities with multi-dimensional semantic information to construct a more comprehensive entity representation, as follows: Step 2.1: Utilize a pre-trained large language model (DeepSeek) to analyze protein entities in biomedical literature. The system identifies and generates a multi-dimensional semantic information set. (9); in, For the first dimension of semantic information A protein sub-semantic fragment; Protein entities A semantic set; This represents the semantic output obtained through a large language model; Step 2.2: Analyze each protein entity using BioBERT Encode to obtain vector BioBERT encoding is used for each multi-dimensional semantic information. Obtain the embedding vector For each protein entity and Cosine similarity is used to measure their similarity. ; (10); in, Describes the norm of vector L2. It is the dot product of the two; Step 2.3: Use a weighted average method to fuse the representations of the original entities and the newly added semantic information to obtain the enhanced entity representation vector. : (11); in, It is a balancing parameter used to control the weights of the original entity representation and the newly added semantic information representation.
[0018] Step 3: Use BERT encoding and a graph attention network to obtain the graph structure information of protein entities from the text, and then fuse the obtained graph structure information with the text semantic information, as follows: Step 3.1: Model the protein-protein interaction network using a graph attention network (GAT); (12); in, For nodes In the Layered graph structure embedding representation; It is a node The set of neighboring nodes; It is a node For neighboring nodes Attention coefficient; For nodes In the Layered graph structure embedding representation; It is the first The layer weight matrix is calculated as follows: (13); in, The LeakyReLU activation function is used. This indicates a splicing operation. For nodes Graph embedding representation, This is the weight matrix. For each protein node, there is an attention parameter vector. go through After layered graph attention network, graph structure embedding is obtained ; (14); Step 3.2: Align the textual semantic representation and graph embedding representation of protein entities in the feature space; (15); (16); in, and It is an aligned representation of text and image; and These are the linear transformation matrix and bias term of the text modality, respectively; and These are the linear transformation matrix and bias term of the graph mode, respectively; Protein entities Text embedding vectors obtained after BioBERT encoding; Step 3.3: Define the dynamic gating mechanism, as shown in formula (18): (17); (18); in, and These are the weight matrix and bias term of the gating mechanism, respectively; This indicates the concatenation of two aligned representations; These are the calculated gating weights, used to control the contribution level of each mode; It is a protein entity The fusion representation; Step 3.4: Protein pair ( Introducing interaction terms to model the relationship between the two, we obtain protein pairs ( The final representation of ) ; (19); in, It is the element-order product of the fusion of two proteins. Protein entities The fusion representation.
[0019] To validate the effectiveness of this method, experiments were conducted on five PPI benchmark corpora: AIMed, BioInfer, HPRD50, IPEA, and LLL. AIMed contains 225 abstracts related to human protein-protein interactions and is one of the most widely used PPI datasets; BioInfer is large, containing over 1100 sentences with a fine-grained annotation system covering complex nested structures; HPRD50, although containing only 50 abstracts, provides high-quality annotations of human protein-protein interactions; IEPA contains 486 sentences and focuses on drug-related protein interactions; LLL is a small but specialized corpus with 77 training samples, focusing on interaction information in bacterial gene regulatory networks. This method uses a unified version of the PPI benchmark dataset provided in the state-of-the-art (SOTA) model for processing and experimentation. In the dataset, PPI relationships are labeled as positive and negative, and the corpus statistics are shown in Table 1.
[0020] Table 1. Five PPI benchmark corpora for Positive and Negative keywords
[0021] The evaluation criterion used in the experiment was the micro-F1 score, which is the performance metric commonly used by the state-of-the-art (SOTA) models in the field.
[0022] The performance comparison results of this method with existing methods are shown in Table 2.
[0023] Table 2. Experimental Results
[0024] The experimental results presented in the table cover five recognized PPI benchmark corpora. The experimental data clearly show that our method achieves the best performance in the overall classification task and significantly outperforms existing state-of-the-art methods in terms of average F1 score. This result strongly demonstrates the effectiveness and advancement of our method.
[0025] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A protein interaction extraction method based on entity semantic enhancement, characterized in that, Includes the following steps: Step 1: Map the input sequence to an embedding representation containing contextual semantic information, and obtain a context-enhanced relational representation through an entity attention mechanism; Step 2: Integrate protein entities with multi-dimensional semantic information to construct an enhanced entity representation; Step 3: Use BERT encoding and graph attention network to obtain graph structure information of protein entities from the text, fuse the obtained graph structure information with text semantic information to generate a fused representation of protein pairs for interaction relationship determination.
2. The protein interaction extraction method based on entity semantic enhancement according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step 1.1: Input sequence. Insert entity markers into the input sequence to explicitly identify the locations of protein entities of interest; for sequences containing protein pairs... The sentence, i.e., the input sequence Convert it to ; （1）； （2）； in, For the input sequence The first in One word, For the input sequence The total number of words in the Chinese dictionary; and These are the classification and delimiter labels in the BERT model, respectively. , , and These are entity markers used to identify the location of two protein entities; Step 1.2: Encode the sequence using the BioBERT model to obtain the context hidden state representation corresponding to each token. ; （3）； in, , express The first in The hidden state of each token, for The total number of tokens in the game. for 3D real space; for entity labeling and Corresponding hidden state and The attention weights of these weights for all locations are calculated. and ; （4）； （5）； in, express The first in The hidden state of each token; Step 1.3: Add the attention probabilities of the two entities together, and select the one with the highest attention weight based on the merged attention score. The tokens are assigned to several positions, and max-pooling is performed on these tokens to construct a context representation. Finally, the pooled result is integrated into the relation representation to obtain the sum of the attention probabilities of the two entities. and the context representation after pooling Relationship representation after integration ; （6）； （7）； （8）； in, This indicates a max pooling operation. This indicates selecting the option with the highest attention weight. One position.
3. The protein interaction extraction method based on entity semantic enhancement according to claim 2, characterized in that, The specific steps of step 2 are as follows: Step 2.1: Utilize a pre-trained large language model to analyze protein entities in biomedical literature. The system identifies and generates a multi-dimensional semantic information set. （9）； in, For the first dimension of semantic information A protein sub-semantic fragment; Protein entities A semantic set; This represents the semantic output obtained through a large language model; Step 2.2: Analyze each protein entity using BioBERT Encode to obtain vector BioBERT encoding is used for each multi-dimensional semantic information. Obtain the embedding vector For each protein entity and Cosine similarity is used to measure their similarity. ; （10）； in, Describes the norm of vector L2. It is the dot product of the two; Step 2.3: Use a weighted average method to fuse the representations of the original entities and the newly added semantic information to obtain the enhanced entity representation vector. : （11）； in, It is a balancing parameter used to control the weights of the original entity representation and the newly added semantic information representation.
4. The protein interaction extraction method based on entity semantic enhancement according to claim 3, characterized in that, The specific steps of step 3 are as follows: Step 3.1: Model the protein-protein interaction network using a graph attention network (GAT); （12）； in, For nodes In the Layered graph structure embedding representation; It is a node The set of neighboring nodes; It is a node For neighboring nodes Attention coefficient; For nodes In the Layered graph structure embedding representation; It is the first The layer weight matrix is calculated as follows: （13）； in, The LeakyReLU activation function is used. This indicates a splicing operation. For nodes Graph embedding representation, This is the weight matrix. For each protein node, there is an attention parameter vector. go through After layered graph attention network, graph structure embedding is obtained ; （14）； Step 3.2: Align the textual semantic representation and graph embedding representation of protein entities in the feature space; （15）； （16）； in, and It is an aligned representation of text and image; and These are the linear transformation matrix and bias term of the text modality, respectively; and These are the linear transformation matrix and bias term of the graph mode, respectively; Protein entities Text embedding vectors obtained after BioBERT encoding; Step 3.3: Define the dynamic gating mechanism, as shown in formula (18): （17）； （18）； in, and These are the weight matrix and bias term of the gating mechanism, respectively; This indicates the concatenation of two aligned representations; These are the calculated gating weights, used to control the contribution level of each mode; It is a protein entity The fusion representation; Step 3.4: Protein pair ( Introducing interaction terms to model the relationship between the two, we obtain protein pairs ( The final representation of ) ; （19）； in, It is the element-order product of the fusion of two proteins. Protein entities The fusion representation.