A Method and System for Constructing and Updating Degradation Protein Knowledge Graphs Based on Multi-Source Fusion

Abstract

This invention discloses a method and system for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion, belonging to the field of electronic digital data processing technology. First, the invention performs structured extraction of text, images, and tables from multi-source documents, establishing a semantic mapping between multimodal content and corresponding textual descriptions. Through contrastive learning, multimodal features are mapped to a unified semantic space, and adaptive weighted fusion is achieved based on an attention mechanism. Based on the fusion features, the catalytic relationship between degradation proteins and pollutants is extracted and confidence-weighted quadruples are constructed. Effective knowledge units are screened through multimodal evidence cross-validation. Meta-learning is combined to enhance the recognition ability of long-tail relationships, and external databases are integrated to complete entity attributes. Finally, a knowledge graph of degradation proteins is constructed, and a dynamic update mechanism including consistency verification and version retention is established. This invention improves the completeness and accuracy of knowledge extraction, ensuring the timeliness and reliability of the knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to a method and system for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion. Background Technology

[0002] With the rapid development of environmental bioremediation technologies, research on the catalytic effects of degradation proteins and organic pollutants has deepened, leading to an explosive growth in related academic literature, experimental reports, and professional database resources. Knowledge graphs, as a core technology of structured knowledge management, can transform degradation proteins, pollutants, and catalytic relationships into a structured network of entity-relationships, providing data support for degradation enzyme screening and pollution remediation scheme development. This has become a research hotspot in the interdisciplinary field of biomedicine and environmental information.

[0003] Currently, existing knowledge graph construction technologies in the field of degradation proteins still have many shortcomings: First, existing knowledge extraction schemes mostly process pure text documents, making low use of non-textual modal information such as mechanism diagrams, experimental data charts, and parameter tables, and failing to establish semantic relationships between different modalities. A large amount of key catalytic evidence from experimental data is missed, resulting in incomplete knowledge extraction results. Second, existing schemes mostly use single-modal features to independently complete entity recognition and relationship classification, failing to achieve deep fusion and cross-validation of cross-modal semantics. Noise and bias in single-modal information are difficult to correct, and the accuracy and confidence of catalytic relationship extraction lack multi-dimensional verification mechanisms. Third, the sample size of many low-frequency, long-tail catalytic relationships in the field of degradation proteins is limited, and traditional supervised learning models perform poorly in recognizing them. Furthermore, existing knowledge graphs mostly adopt a static construction mode, lacking standardized incremental updates, conflict arbitration, and version management mechanisms, making it impossible to adapt to the rapid iterative updates of domain literature, and the timeliness and reliability of the knowledge graph cannot be guaranteed.

[0004] In summary, existing technologies cannot efficiently integrate multi-source and multi-modal literature information on degradation proteins, and it is difficult to balance the completeness and accuracy of knowledge extraction with the dynamic updating capability of knowledge graphs, thus failing to meet the high-quality knowledge management needs of practical application scenarios. Summary of the Invention

[0005] This invention discloses a method and system for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion, in order to solve the problems of scattered and heterogeneous literature information on degradation proteins, incomplete extraction of single-modal knowledge, difficulty in identifying long-tailed catalytic relationships, and lack of dynamic verification mechanism in the knowledge graph in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion, comprising the following steps:

[0007] S1. Multi-source literature multimodal structured parsing: Acquire domain literature data from multiple sources, segment the text modality, image modality and table modality regions in the literature through document layout analysis algorithm, perform structured information extraction respectively, and establish semantic anchors based on the chapter level and reference markers of the literature to generate semantic mapping associations between each modality content and the corresponding text descriptions.

[0008] S2, Cross-modal unified semantic space alignment: The structured content of each modality is encoded by a dedicated encoder to obtain the initial feature vector of the corresponding dimension. The cross-modal alignment model trained by contrastive learning maps the initial feature vectors of all modalities to the same high-dimensional semantic space and outputs an alignment feature vector with unified dimension.

[0009] S3. Multimodal adaptive weighted fusion: Based on the attention mechanism, the semantic contribution weight of each modal alignment feature to the target knowledge extraction task is calculated, and the multimodal features are weighted and fused according to the weights to generate a fused feature vector containing complementary semantic information of multiple modalities.

[0010] S4. Fusion-driven knowledge unit extraction: Input the fusion feature vector into the entity recognition and relationship classification model to extract the subject entity, object entity and the interaction relationship between them in the domain. Based on the credibility of the recognition and classification results, construct a relation quadruple with initial confidence.

[0011] S5. Multimodal evidence cross-validation: Count the number of modal types that can independently corroborate the same relationship, adjust the initial confidence level based on the number of modal evidences, obtain the final confidence level, and select quadruplets that meet the preset confidence threshold as effective knowledge units.

[0012] S6. Knowledge Graph Construction and Dynamic Update: Integrate effective knowledge units with external structured databases to construct domain knowledge graphs and establish an incremental dynamic update mechanism that includes entity alignment, consistency verification, and version retention.

[0013] Preferably, in step S1, the specific process of associating structured parsing with semantic mapping is as follows:

[0014] The document page is segmented into three categories at the pixel level using a layout analysis model: main text area, independent chart area, and structured table area. For the main text area, optical character recognition, paragraph splitting, and chapter numbering are performed to extract text segments with hierarchical attributes. For the independent chart area, chart border detection, legend text recognition, and numerical annotation extraction are performed, and the corresponding chart numbers are recorded. For the structured table area, cell coordinate restoration, header hierarchy mapping, and cell content extraction are performed, and the corresponding table numbers are recorded.

[0015] Using chapter numbers, figure numbers, and table numbers as globally unique semantic anchors, text, image, and table data under the same anchor are bound and stored together to generate a multimodal structured dataset with anchor identifiers, achieving a one-to-one semantic correspondence between multimodal content from the same source.

[0016] Preferably, in step S2, the cross-modal semantic alignment model is trained based on a contrastive learning framework, and its alignment loss function satisfies the following relationship:

[0017] ,

[0018] In the formula: This represents the cross-modal semantic alignment loss value. Let be the initial feature vector of the text modality for the i-th training sample; is a linear projection matrix of the text modality, used to map the initial features of the text to a unified semantic space; Let be the initial feature vector of the image modality that matches the text in the i-th training sample; is a linear projection matrix of the image modalities, used to map the initial features of the image to a unified semantic space; Let be the initial feature vector of the negative sample image mode corresponding to the i-th training sample; is the cosine distance calculation function between feature vectors; m is a preset boundary threshold parameter, with a value range of 0.1 to 0.9; This is the hinge operation function, which takes the maximum value between the value inside the parentheses and 0; N is the total number of samples in the training batch.

[0019] This relation constrains the precise semantic alignment of modal features within a unified semantic space by reducing the semantic distance between positive sample pairs with the same anchor point and increasing the semantic distance between negative sample pairs with different anchor points. Table modalities follow the same alignment logic, and their corresponding linear projection matrices are denoted as follows. .

[0020] Preferably, in step S3, the adaptive weighted fusion is achieved through two steps: attention weight allocation and feature weighted summation, satisfying the following relationship:

[0021] ,

[0022] ,

[0023] In the formula: Let be the attention weights for the m-th modality alignment feature; t represents the semantic similarity score between the alignment feature vector of the m-th modality and the preset task query vector; v and g represent the text modality, image modality, and table modality, respectively. This is the final output cross-modal fusion feature vector; Let be the alignment feature vector of the m-th mode; This is the feature transformation matrix corresponding to the m-th mode, used to unify the dimensionality of features across modes;

[0024] This relation adaptively allocates the information contribution ratio based on the semantic relevance of each modality to the knowledge extraction task, so that the fused features focus on the core semantic information and improve the completeness and accuracy of subsequent knowledge extraction.

[0025] Preferably, in step S4, the specific process of knowledge unit extraction and quadruple construction is as follows:

[0026] The fused feature vectors are input into the sequence labeling model, and the BIO labeling system is used to identify the boundaries between the main entities and the object entities in the document, and the identification confidence of each entity is output. The fused feature fragments corresponding to the entity pairs are extracted and input into the relation classification model, and the classification confidence of the entity pairs for all corresponding candidate relations is output.

[0027] The relationship with the highest classification confidence is selected as the target relationship. Based on the average entity recognition confidence and the relationship classification confidence, the initial confidence is calculated by weighting according to the preset weights. Finally, a relationship quadruple is constructed with the structure of degradation protein entity, pollutant entity, catalytic relationship and initial confidence.

[0028] Preferably, in step S5, the specific correction formula for multimodal evidence cross-validation satisfies the following relationship:

[0029] ,

[0030] In the formula: The final confidence level after multimodal cross-validation; The initial confidence level of the relation quadruple; is the confidence gain coefficient corresponding to independent evidence of a single modality, with a value ranging from 0.05 to 0.3; n is the number of modalities that can independently corroborate the relationship, and n is a positive integer greater than or equal to 1; To minimize the value, the final confidence level is constrained to not exceed 1; the criterion for independent evidence is that the corresponding entity and the relationship between actions can be directly derived from the textual descriptions of other modalities without relying on the textual descriptions of other modalities.

[0031] This relation enhances the credibility of knowledge units through multimodal cross-validation, and selects quadruplets with a final confidence level greater than or equal to a preset threshold as valid knowledge units.

[0032] Preferably, step S6 includes steps for enhancing low-frequency long-tail relationships and completing attribute knowledge before constructing the knowledge graph, specifically:

[0033] For low-frequency long-tail relationships with a sample size below a preset threshold, a few-shot classification model under the meta-learning framework is used for recognition enhancement. Each type of low-frequency relationship is constructed as an independent few-shot task. Each task contains a support set consisting of a small number of labeled samples and a query set consisting of samples to be classified. The model quickly adapts to the task features through the support set and completes the relationship classification on the query set, thereby achieving accurate recognition of long-tail relationships with a small sample size.

[0034] Simultaneously, an external structured professional database is invoked, and the extracted entities are matched with the database entities through an entity linking algorithm to complete the entity attribute information. The long-tail relationships obtained from enhanced recognition and the completed attribute information are then integrated into the knowledge graph.

[0035] Preferably, in step S6, the incremental dynamic update mechanism includes four stages in sequence: incremental knowledge generation, entity alignment and fusion, conflict knowledge arbitration, and version difference retention. The specific process is as follows:

[0036] 1) Incremental knowledge generation: Acquire new domain literature according to the preset time period, reuse the entire process of multimodal parsing, alignment and fusion, extraction and verification from steps S1 to S5, and generate incremental knowledge units with final confidence.

[0037] 2) Entity alignment and fusion: The weighted calculation method of string edit distance and semantic vector similarity is used to obtain the comprehensive similarity between the incremental entity and the existing knowledge graph entities. Entities with a comprehensive similarity higher than a preset threshold are identified as the same entity, and the attribute information of the two are merged to complete the entity fusion.

[0038] 3) Conflict knowledge arbitration: For the same entity pair corresponding to the relationship, if the incremental relationship is inconsistent with the relationship type of the existing graph, it is determined to be a knowledge conflict; compare the final confidence values ​​of the two, retain the relationship entry with higher confidence, and mark the low confidence entry as a historical candidate and keep the conflict log.

[0039] 4) Version Differential Retention: After each update, a unique incremental version number is generated. The differential storage method is used to record only the three types of changes in this update: addition, modification and deletion. At the same time, a full version index is established to support the comparison of differences between any two historical versions and the full data backtracking.

[0040] This invention provides a domain knowledge graph construction and update system based on multi-source cross-modal fusion, comprising:

[0041] The multimodal structured parsing module is used to acquire domain literature data from multiple sources, segment multimodal regions through layout analysis algorithms and perform structured information extraction, and establish semantic mapping associations for each modality based on semantic anchors;

[0042] The cross-modal semantic alignment module is used to generate initial feature vectors for each modality through a dedicated encoder, and to map multimodal features to a unified semantic space based on a contrastive learning model, outputting aligned feature vectors with unified dimensions.

[0043] The adaptive feature fusion module is used to calculate the semantic contribution weights of each modality feature based on the attention mechanism, and generate a fusion feature vector containing multimodal complementary semantics through weighted fusion.

[0044] The knowledge unit extraction module is used to perform entity sequence labeling and relation classification based on fused feature vectors, and to construct relation quadruplets with initial confidence.

[0045] The multimodal evidence verification module is used to count the number of independent modal evidence, perform gain correction on the initial confidence level, and filter out effective knowledge units.

[0046] The knowledge graph construction and update module is used to integrate effective knowledge units and externally supplemented knowledge to construct a domain knowledge graph, and to perform dynamic update operations such as incremental knowledge generation, entity alignment and fusion, conflict knowledge arbitration, and version difference retention.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. This invention improves the completeness and accuracy of knowledge extraction. By using multimodal structured parsing and cross-modal semantic adaptive fusion, it fully mines the catalytic evidence in text, charts, and tables in the literature, makes up for the information loss in single-modal extraction, and effectively improves the comprehensiveness and accuracy of identifying catalytic relationships of degradation proteins.

[0049] 2. This invention enhances the reliability and coverage of the knowledge system. It quantifies knowledge confidence by relying on a multimodal evidence cross-validation mechanism and enhances the recognition ability of low-frequency long-tail catalytic relationships by combining a meta-learning framework. This reduces extraction errors caused by single-modal noise and expands the relationship coverage of the knowledge graph.

[0050] 3. This invention establishes an incremental dynamic update mechanism that includes entity alignment, conflict arbitration, and version difference retention, enabling continuous iterative updates of the knowledge graph. It also supports historical version backtracking and difference comparison, adapting to the rapid update needs of domain literature. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0052] In the attached diagram:

[0053] Figure 1This is a flowchart of a method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion, according to the present invention.

[0054] Figure 2 This is a simplified diagram of the module structure of a knowledge graph construction and updating system for degradation proteins based on multi-source fusion, as proposed in this invention. Detailed Implementation

[0055] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0056] Example: The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion disclosed in this invention is applied to the automated extraction and graph maintenance of degradation protein knowledge in the context of bioremediation of petroleum organic pollutants.

[0057] like Figure 1 As shown, the method in this embodiment is executed on a general-purpose server or cloud computing node, and its core is achieved through multimodal semantic fusion, confidence cross-validation, and incremental dynamic update mechanisms. The complete execution flow of this embodiment includes six core steps: multimodal structured parsing of multi-source documents, cross-modal unified semantic space alignment, multimodal adaptive weighted fusion, fusion-driven catalytic knowledge unit extraction, multimodal evidence cross-validation, knowledge graph construction, and incremental dynamic update. The specific implementation process is as follows:

[0058] S1. Multi-source document multimodal structured analysis

[0059] The input for this step is academic literature related to degradation proteins from multiple sources, including PDF documents obtained from databases such as PubMed, Web of Science, and CNKI. The literature content covers multimodal information such as degradation protein sequences, catalytic experimental data, pollutant degradation efficiency tables, and mechanism diagrams.

[0060] Specifically, a pre-trained document layout analysis model is first used to perform pixel-level region segmentation on each page of the document. The layout analysis model is a document region detection model obtained by fine-tuning LayoutLMv3. The segmentation categories include three types: main text area, independent chart area, and structured table area. The region segmentation confidence threshold is set to 0.7. Regions below the threshold are manually reviewed or discarded.

[0061] Furthermore, structured information extraction is performed on the three types of regions respectively:

[0062] 1) Text processing: The optical character recognition engine is used for character recognition; then paragraph splitting and reference resolution are performed, and the text is hierarchically marked according to the chapter number of the document, such as 1, 1.1, 2.1, etc., and the text fragments containing core semantics such as degradation proteins, organic pollutants, and catalytic effects are extracted and their chapter number and context position are recorded.

[0063] 2) Independent chart area processing: The outer frame of the chart is identified by the border detection algorithm and the chart number is extracted; the legend text, axis labels and numerical labels in the chart are identified by text recognition, and the corresponding experimental conditions, degradation efficiency, protein expression level and other labeling information are extracted and recorded.

[0064] 3) Structured table area processing: The cell coordinate detection algorithm is used to restore the row and column structure of the cells, identify the header level and merged cells, map the header fields to the corresponding cell contents, extract structured data such as protein names, pollutant types, degradation rates, and reaction conditions from the table, and record the corresponding table number.

[0065] Furthermore, chapter numbers, figure / table numbers, and table numbers are used as globally unique semantic anchors to bind and store text, images, and table data under the same anchor. By establishing semantic associations between table data labeled "Table 2" and text paragraphs labeled "as shown in Table 2" in the main text, a multimodal structured dataset with anchor labels is finally generated and stored in JSON format, achieving a one-to-one semantic correspondence between multimodal content from the same source.

[0066] S2, Cross-modal unified semantic space alignment

[0067] This step uses a dedicated encoder and a contrastive learning alignment model to map features from different dimensions and modalities to the same high-dimensional semantic space, thus eliminating the semantic gap between modalities.

[0068] Specifically, three types of dedicated encoders are used to encode the features of each modal data after structured processing:

[0069] The text encoder uses a BERT-based pre-trained model, with the input being the segmented text fragments and the output being a 768-dimensional initial feature vector of the text modality.

[0070] The image encoder uses a ResNet-50 pre-trained model. The input is a normalized image of the chart region, and the output is a 2048-dimensional initial feature vector of the image modality.

[0071] The table encoder uses the TableTransformer pre-trained model, takes the serialized content of a structured table as input, and outputs an initial feature vector of the table modality with a dimension of 1024.

[0072] Furthermore, a cross-modal semantic alignment model is trained based on a contrastive learning framework. During training, text-image and text-table feature pairs under the same semantic anchor are used as positive sample pairs, while feature pairs under different semantic anchors are used as negative sample pairs. The number of samples N in each training batch is set to 64. The alignment loss function of the model satisfies the following relationship:

[0073] ,

[0074] In the formula: This represents the cross-modal semantic alignment loss value. Let be the initial feature vector of the text modality for the i-th training sample; is the linear projection matrix of the text modality, used to map the initial 768-dimensional text features to a 512-dimensional unified semantic space; Let be the initial feature vector of the image modality that matches the text in the i-th training sample; is a linear projection matrix for the image modality, used to map the initial features of the 2048-dimensional image to a 512-dimensional unified semantic space; is the initial feature vector of the negative sample image mode corresponding to the i-th training sample; m is a preset boundary threshold parameter, with a value range of 0.1 to 0.9, and in this embodiment, the preferred value is 0.5; This is the hinge operation function, which takes the maximum value between the value inside the parentheses and 0; N is the total number of samples in the training batch. The cosine distance function between feature vectors is given by the following formula: ,in and These represent two high-dimensional feature vectors with the same dimension, representing the semantic distance to be calculated, and both reside in a unified semantic space. In the cross-modal alignment loss function of this embodiment, the two correspond to the projected cross-modal feature vectors, specifically: 1) When calculating the distance between positive sample pairs, correspond (The text feature vector after projection). correspond (Feature vector of the matched image after projection); 2) When calculating the distance between negative sample pairs, correspond , correspond (Feature vector of the mismatched image after projection).

[0075] This relation constrains the precise semantic alignment of modal features within a unified semantic space by reducing the semantic distance between positive sample pairs with the same anchor point and increasing the semantic distance between negative sample pairs with different anchor points. Table modalities follow the same alignment logic, and their corresponding linear projection matrices are denoted as follows. This is used to map 1024-dimensional table features to a 512-dimensional unified semantic space. After training, the output dimension is unified to 512-dimensional aligned feature vectors for text, images, and tables.

[0076] S3, Multimodal Adaptive Weighted Fusion

[0077] This step is based on an attention mechanism, which adaptively assigns weights according to the semantic relevance of each modality to the catalytic knowledge extraction task, generating fusion features that take into account the complementary information of multiple modalities.

[0078] Specifically, a 512-dimensional learnable query vector is pre-defined to correspond to the protein degradation catalysis knowledge extraction task. This vector is optimized synchronously during model training. Alignment feature vectors and query vectors are calculated for text, image, and table modalities respectively. The semantic similarity score is calculated using cosine similarity, i.e. .

[0079] Furthermore, the similarity scores are normalized using the softmax function to obtain the attention weights for each modality. Then, the aligned features of each modality are unified in dimension using a feature transformation matrix, and finally, weighted summation is performed according to the attention weights to obtain the final fused features. The calculation process satisfies the following relationship:

[0080] ,

[0081] ,

[0082] In the formula, Let be the attention weights for the m-th modality alignment feature; t represents the semantic similarity score between the aligned feature vector of the m-th modality and the preset task query vector; v and g represent the text modality, image modality, and table modality, respectively. The final output is a 512-dimensional cross-modal fusion feature vector; Let be the alignment feature vector of the m-th mode; This is the feature transformation matrix corresponding to the m-th mode, used to unify the dimension of the features of each mode to 512 dimensions.

[0083] This relational expression adaptively allocates the information contribution ratio based on the semantic relevance of each modality to the protein degradation catalytic knowledge extraction task: when a modality contains more explicit evidence of catalytic relationships, its attention weight automatically increases. When the literature table directly provides data on the degradation rate of pollutants by proteins, the weight of the table modality is significantly increased, allowing the fused features to focus on the core catalytic semantics and improving the completeness and accuracy of subsequent knowledge extraction.

[0084] S4. Fusion-driven catalytic knowledge unit extraction

[0085] This step performs entity recognition and relationship classification based on cross-modal fusion features, and constructs catalytic relationship quadruples with confidence.

[0086] Specifically, entity recognition is performed first: the fused feature vector is input into the sequence labeling model, which uses a BERT+CRF structure and the BIO labeling system, where B represents the start position of an entity, I represents the middle position, and O represents a non-entity position. The model outputs the entity label probability corresponding to each character, and takes the label with the highest probability as the labeling result, identifying degradation protein entities in the literature, such as alkane hydroxylase AlkB and catechol dioxygenase, and pollutant entities, such as petroleum hydrocarbons, polycyclic aromatic hydrocarbons, and benzene series compounds, and outputs the recognition confidence score for each entity. The value range is 0 to 1.

[0087] Calculate the average recognition confidence score of all identified core entities. ,Right now:

[0088] Where k is the number of core entities, in this embodiment It contains at least one degraded protein entity and one contaminant entity.

[0089] Further, relation classification is performed: The fused feature fragments corresponding to entity pairs are extracted and input into the relation classification model. The relation classification model employs a two-layer fully connected network structure with a Softmax output layer, and the hidden layer dimension is 256. Candidate relation types include three categories: catalytic degradation, inhibition, and no interaction. The model outputs the classification confidence score for each candidate relation corresponding to the entity pair. The catalytic relationship with the highest classification confidence score is selected as the target relation, and its classification confidence score is denoted as [value missing]. .

[0090] Furthermore, by combining the credibility of entity recognition and relation classification, an initial confidence level is calculated, and a catalytic relation quadruple is constructed. The initial confidence level satisfies the following relation:

[0091] ,

[0092] In the formula: The initial confidence level of the catalytic relationship quadruple; The weighting coefficient for the entity recognition result is preferably set to 0.4 in this embodiment, with a value ranging from 0.2 to 0.8. This represents the average confidence level for identifying degraded protein entities and contaminant entities. The weighting coefficient for the relation classification result is preferably set to 0.6 in this embodiment, with a range of 0.2 to 0.8; this satisfies the constraint. ; Classification confidence level of the target catalytic relationship.

[0093] The final structure consists of a quadruple of degradation protein entity, pollutant entity, catalytic relationship, and initial confidence level of catalytic relationship, thus completing the initial extraction of knowledge units.

[0094] S5, Multimodal Evidence Cross-verification

[0095] This step uses multimodal independent evidence to gain and correct the initial confidence level, thereby improving the reliability of the knowledge unit.

[0096] Specifically, the first step is to determine independent modal evidence: for the same catalytic relationship quadruple, determine whether the content of the text, image, and table modalities can be directly deduced from the relationship between the corresponding entity and the catalytic action without relying on other modal descriptions.

[0097] 1) If the text explicitly describes "a certain protein can catalyze the degradation of a certain pollutant", then the text modality is considered an independent piece of evidence;

[0098] 2) If the chart is a bar chart of degradation efficiency and the labeling is clear, and it can directly reflect the catalytic effect, then the image mode is determined to be an independent piece of evidence;

[0099] 3) If the table directly lists the corresponding degradation rate data of proteins and pollutants, then the table modality is considered as independent evidence.

[0100] The number of modal types n that can independently corroborate the catalytic relationship is counted. n is a positive integer greater than or equal to 1, and in this embodiment, the value is 1, 2 or 3.

[0101] Furthermore, a linear gain correction is applied to the initial confidence level based on the number of modal evidences, and the final confidence level is constrained to not exceed 1. The correction formula satisfies the following relationship:

[0102] ,

[0103] In the formula: The final confidence level after multimodal cross-validation; The confidence gain coefficient corresponding to independent evidence of a single modality is preferably 0.15 in this embodiment, and the value range is 0.05~0.3; n is the number of modal types that can independently corroborate the catalytic relationship; This is used for minimum value calculation to constrain the final confidence level to not exceed 1.

[0104] This relationship indicates that the more independent modal evidence supports a catalytic relationship, the higher its final confidence level. In this embodiment, a final confidence threshold of 0.7 is preset, and quadruples with a final confidence level greater than or equal to 0.7 are selected as valid knowledge units to reduce extraction errors caused by noise from single modal information.

[0105] S6. Knowledge Graph Construction and Incremental Dynamic Updates

[0106] This step first improves the knowledge system by enhancing long-tail relationships and supplementing it with external knowledge, then constructs an initial knowledge graph and establishes an incremental dynamic update mechanism.

[0107] (1) Enhancement of low-frequency long-tail catalytic relationship and completion of attribute knowledge

[0108] Before constructing the knowledge graph, low-frequency long-tail relationships are first identified and enhanced. The number of samples for all catalytic relationship types is counted, and relationships with a sample size below a preset threshold are identified as low-frequency long-tail catalytic relationships; in this embodiment, 10 such relationships are preferred.

[0109] A model-independent meta-learning (MAML) few-shot classification model under the meta-learning framework is used for recognition enhancement. Each class of low-frequency catalytic relations is constructed as an independent few-shot task, with each task consisting of a support set of 5 labeled samples and a query set of 15 samples to be classified. The model's meta-training loss satisfies the following relationship:

[0110] ,

[0111] In the formula: is the total adaptation loss value of meta-learning; S is the total number of meta-training tasks, which is preferably 20 in this embodiment. Let be the classification loss for the s-th task on the query set; These are the initial parameters for the relation classification model; The learning rate for the inner loop is preferably set to 0.01 in this embodiment, with a range of 0.001 to 0.1. Let be the gradient of the loss of the s-th task on the support set with respect to the model parameters.

[0112] This relational expression enables the model to quickly adapt to low-frequency relationships through multi-task meta-learning: in the inner loop, the model quickly adapts to task features on the support set, and in the outer loop, the model's initial parameters are optimized based on the query set loss. Finally, the model can complete the classification of new low-frequency catalytic relationships with a small number of samples, thereby improving the accuracy of long-tail relationship recognition.

[0113] Simultaneously, external biomedical structured databases (such as UniProt and KEGG) are invoked. An entity linking algorithm is used to match extracted degradation protein entities and contaminant entities with standard entities in the database, with a semantic similarity threshold of 0.85 for entity matching. Upon successful matching, entity attribute information is supplemented, including the amino acid sequence, molecular weight, and species origin of the protein, and the molecular formula and physicochemical properties of the contaminant. The long-tailed catalytic relationships obtained from enhanced recognition are integrated with the supplemented attribute information into a knowledge system to construct an initial degradation protein knowledge graph.

[0114] (2) Incremental dynamic update mechanism

[0115] This embodiment establishes an incremental dynamic update mechanism comprising four stages, the specific process of which is as follows:

[0116] 1) Incremental knowledge generation: According to the preset time period (preferably once a month in this embodiment), newly added literature related to degradation proteins is obtained, and the multimodal analysis, alignment fusion, extraction and verification process of steps S1 to S5 is reused to generate incremental catalytic knowledge units with final confidence.

[0117] 2) Entity Alignment and Fusion: A weighted calculation method combining string edit distance and semantic vector similarity is used to calculate the comprehensive similarity between the incremental entity and existing knowledge graph entities. The calculation formula is as follows:

[0118] ,

[0119] in The string similarity weight is preferably set to 0.3; This represents the string similarity calculated based on the Levenshtein distance. The cosine similarity is calculated based on the entity semantic vectors. Entities with a comprehensive similarity higher than a preset threshold are identified as the same entity; in this embodiment, a threshold of 0.8 is preferred. The attribute information of the two entities is merged, and when attributes conflict, the attribute value with higher source confidence is retained.

[0120] 3) Conflict Knowledge Arbitration: For the same entity pair corresponding to a catalytic relationship, if the incremental relationship is inconsistent with the relationship type in the existing graph, it is determined to be a knowledge conflict. The final confidence values ​​of the two are compared, and the relationship entry with the higher confidence value is retained; at the same time, the low confidence value entry is marked as a historical candidate, and information such as conflict time, conflicting parties, and confidence value is recorded to generate a conflict log.

[0121] 4) Differential Version Retention: After each update, a unique, incrementing version number is generated, in the format "V + Year + Serial Number", such as V2026001. Differential storage only records the three types of changes in this update: additions, modifications, and deletions. Compared to full storage, this reduces storage space usage by approximately 60%. Simultaneously, a full version index is established, storing a full snapshot pointer for each version, supporting comparison of differences between any two historical versions and full data backtracking.

[0122] like Figure 2 In addition, this embodiment of the invention also provides a system for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion, used to execute the above method steps. The system is a modular software system deployed on the server side, including the following modules:

[0123] The multimodal structured parsing module is used to acquire literature data related to degradation proteins from multiple sources. It segments text, images, and tables into multimodal regions through layout analysis algorithms, performs structured information extraction such as OCR recognition and cell structure restoration, and establishes semantic mapping associations for each modality based on semantic anchors such as chapter numbers and figure numbers.

[0124] The cross-modal semantic alignment module has built-in dedicated encoders for text, images, and tables to generate initial feature vectors for each modality; it also has a built-in cross-modal alignment model trained based on contrastive learning to map multimodal features to a unified semantic space and output aligned feature vectors with uniform dimensions.

[0125] The adaptive feature fusion module is used to calculate the semantic contribution weight of each modality feature for the catalytic knowledge extraction task based on the attention mechanism, and generate a fused feature vector containing multimodal complementary semantics through weighted fusion.

[0126] The catalytic knowledge extraction module has a built-in sequence labeling submodule and relation classification submodule, which are used to perform entity sequence labeling and relation classification based on fused feature vectors, calculate the initial confidence and construct catalytic relation quadruplets with confidence;

[0127] The multimodal evidence verification module is used to count the number of independent modal evidence, perform gain correction on the initial confidence level, and screen out effective catalytic knowledge units;

[0128] The knowledge graph construction and dynamic update module is used to integrate effective knowledge units and external databases to complete the knowledge graph of degradation proteins, and to perform dynamic update operations such as incremental knowledge generation, entity alignment and fusion, conflict knowledge arbitration and version difference retention.

[0129] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion, characterized in that, Includes the following steps: S1. Obtain domain literature data from multiple sources, segment the text modality, image modality and table modality regions in the literature, perform structured information extraction respectively, and establish semantic anchors based on the chapter level and reference markers of the literature to generate semantic mapping associations between each modality content and the corresponding text descriptions. S2. Each modal content after structuring is encoded with a dedicated encoder to obtain the initial feature vector of the corresponding dimension. The cross-modal alignment model trained based on contrastive learning maps the initial feature vectors of all modalities to the same high-dimensional semantic space and outputs an alignment feature vector with unified dimension. S3. Calculate the semantic contribution weight of each modality alignment feature for the target knowledge extraction task based on the attention mechanism, and perform weighted fusion of multimodal features according to the weight to generate a fusion feature vector containing complementary semantic information of multimodality. S4. Input the fused feature vector into the entity recognition and relationship classification model, extract the subject entity, object entity and the interaction relationship between them in the domain, and construct the relationship quadruple with initial confidence by comprehensively considering the credibility of the recognition and classification results. S5. Count the number of modal types that can independently corroborate the same relationship, adjust the initial confidence level based on the number of modal evidences, obtain the final confidence level, and select quadruplets that meet the preset confidence threshold as effective knowledge units. S6. Integrate effective knowledge units with external structured databases to construct a domain knowledge graph, and establish an incremental dynamic update mechanism that includes entity alignment, consistency verification, and version retention.

2. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 1, characterized in that: In step S1, the specific process of associating structured parsing with semantic mapping is as follows: The document page is segmented into three categories at the pixel level using a layout analysis model: main text area, independent chart area, and structured table area. For the main text area, optical character recognition, paragraph splitting, and chapter numbering are performed to extract text segments with hierarchical attributes. For the independent chart area, chart border detection, legend text recognition, and numerical annotation extraction are performed, and the corresponding chart numbers are recorded. For the structured table area, cell coordinate restoration, header hierarchy mapping, and cell content extraction are performed, and the corresponding table numbers are recorded. Using chapter numbers, figure numbers, and table numbers as globally unique semantic anchors, text, image, and table data under the same anchor are bound and stored to generate a multimodal structured dataset with anchor identifiers.

3. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 1, characterized in that: In step S2, the cross-modal semantic alignment model is trained based on a contrastive learning framework, and its alignment loss function satisfies the following relationship: ， In the formula: This represents the cross-modal semantic alignment loss value. Let be the initial feature vector of the text modality for the i-th training sample; is a linear projection matrix of the text modality, used to map the initial features of the text to a unified semantic space; Let be the initial feature vector of the image modality that matches the text in the i-th training sample; is a linear projection matrix of the image modalities, used to map the initial features of the image to a unified semantic space; Let be the initial feature vector of the negative sample image mode corresponding to the i-th training sample; is the cosine distance calculation function between feature vectors; m is the preset boundary threshold parameter; This is the hinge operation function, which takes the maximum value between the value inside the parentheses and 0; N is the total number of samples in the training batch. This relation constrains the precise semantic alignment of modal features within a unified semantic space by reducing the semantic distance between positive sample pairs with the same anchor point and increasing the semantic distance between negative sample pairs with different anchor points. Table modalities follow the same alignment logic, and their corresponding linear projection matrices are denoted as follows. .

4. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 1, characterized in that: In step S3, adaptive weighted fusion is achieved through two steps: attention weight allocation and feature weighted summation, satisfying the following relationship: ， ， In the formula: Let be the attention weights for the m-th modality alignment feature; t represents the semantic similarity score between the alignment feature vector of the m-th modality and the preset task query vector; v and g represent the text modality, image modality, and table modality, respectively. This is the final output cross-modal fusion feature vector; Let be the alignment feature vector of the m-th mode; This is the feature transformation matrix corresponding to the m-th mode, used to unify the dimensionality of features across modes; This relation adaptively allocates the information contribution ratio based on the semantic relevance of each modality to the knowledge extraction task, so that the fused features focus on the core semantic information.

5. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 1, characterized in that: In step S4, the specific process of knowledge unit extraction and quadruple construction is as follows: The fused feature vectors are input into the sequence labeling model, and the BIO labeling system is used to identify the boundaries between the main entities and the object entities in the document, and the identification confidence of each entity is output. The fused feature fragments corresponding to the entity pairs are extracted and input into the relation classification model, and the classification confidence of the entity pairs for all corresponding candidate relations is output. The relationship with the highest classification confidence is selected as the target relationship. Based on the average entity recognition confidence and the relationship classification confidence, the initial confidence is calculated by weighting according to the preset weights. Finally, a relationship quadruple is constructed with the structure of degradation protein entity, pollutant entity, catalytic relationship and initial confidence.

6. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 5, characterized in that: In step S5, the specific correction formula for multimodal evidence cross-validation satisfies the following relationship: ， In the formula: The final confidence level after multimodal cross-validation; The initial confidence level of the relation quadruple; is the confidence gain coefficient corresponding to independent evidence of a single modality; n is the number of modalities that can independently corroborate the relationship; This is an operation to find the minimum value; This relation enhances the credibility of knowledge units through multimodal cross-validation, and selects quadruplets with a final confidence level greater than or equal to a preset threshold as valid knowledge units.

7. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 1, characterized in that: Step S6 includes steps for enhancing low-frequency long-tail relationships and completing attribute knowledge before constructing the knowledge graph. Specifically: For low-frequency long-tail relationships with a sample size below a preset threshold, a few-shot classification model under the meta-learning framework is used for recognition enhancement. Each type of low-frequency relationship is constructed as an independent few-shot task. Each task contains a support set consisting of a small number of labeled samples and a query set consisting of samples to be classified. The model quickly adapts to the task features through the support set and completes the relationship classification on the query set, thereby achieving accurate recognition of long-tail relationships with a small sample size. Simultaneously, an external structured professional database is invoked, and the extracted entities are matched with the database entities through an entity linking algorithm to complete the entity attribute information. The long-tail relationships obtained from enhanced recognition and the completed attribute information are then integrated into the knowledge graph.

8. The method for constructing and updating a knowledge graph of degradation proteins based on multi-source fusion according to claim 1, characterized in that: In step S6, the incremental dynamic update mechanism includes four stages in sequence: incremental knowledge generation, entity alignment and fusion, conflict knowledge arbitration, and version differential retention. The specific process is as follows: 1) Acquire new domain literature according to the preset time period, reuse the entire process of multimodal parsing, alignment and fusion, extraction and verification in steps S1 to S5, and generate incremental knowledge units with final confidence. 2) The weighted calculation method of string edit distance and semantic vector similarity is adopted to obtain the comprehensive similarity between the incremental entity and the existing knowledge graph entity. Entities with a comprehensive similarity higher than the preset threshold are identified as the same entity, and the attribute information of the two are merged to complete the entity fusion. 3) For the same entity pair corresponding to a relationship, if the incremental relationship is inconsistent with the relationship type of the existing graph, it is determined to be a knowledge conflict; compare the final confidence values ​​of the two, retain the relationship entry with higher confidence, and mark the low confidence entry as a historical candidate and keep the conflict log. 4) After each update, a unique, incrementing version number is generated. Differential storage is used to record only the three types of changes in this update: additions, modifications, and deletions. At the same time, a full version index is established to support the comparison of differences between any two historical versions and the full data backtracking.

9. A domain knowledge graph construction and update system based on multi-source cross-modal fusion, applied to the method of any one of claims 1-8, characterized in that, include: The multimodal structured parsing module is used to acquire domain literature data from multiple sources, segment multimodal regions through layout analysis algorithms and perform structured information extraction, and establish semantic mapping associations for each modality based on semantic anchors; The cross-modal semantic alignment module is used to generate initial feature vectors for each modality through a dedicated encoder, and to map multimodal features to a unified semantic space based on a contrastive learning model, outputting aligned feature vectors with unified dimensions. The adaptive feature fusion module is used to calculate the semantic contribution weights of each modality feature based on the attention mechanism, and generate a fusion feature vector containing multimodal complementary semantics through weighted fusion. The knowledge unit extraction module is used to perform entity sequence labeling and relation classification based on fused feature vectors, and to construct relation quadruplets with initial confidence. The multimodal evidence verification module is used to count the number of independent modal evidence, perform gain correction on the initial confidence level, and filter out effective knowledge units. The knowledge graph construction and update module is used to integrate effective knowledge units and externally supplemented knowledge to construct a domain knowledge graph, and to perform dynamic update operations such as incremental knowledge generation, entity alignment and fusion, conflict knowledge arbitration, and version difference retention.

Images