A biological risk factor identification method based on feature decoupling

Abstract

The present application belongs to the field of data mining, and particularly relates to a biological risk factor identification method based on feature decoupling. The present application collects biological safety literature data, extracts entities and semantic relationships, thereby constructing a biological academic network, performing representation and feature decoupling, and distinguishing biological risk factors from non-biological risk factors through abnormal score calculation. The biological safety related academic literature network is modeled, and the biological risk factor is detected based on the feature decoupling anomaly detection method with better interpretability, thereby improving the accuracy and reliability of the biological risk factor identification.

Description

Technical Field

[0001] This invention belongs to the field of data mining, specifically relating to a method for identifying biological risk factors based on feature decoupling. Background Technology

[0002] With the rapid development of biotechnology and the increasing prominence of biosafety issues, identifying and predicting biological risk factors has become an important research task. Biological risk factors refer to organisms, substances, or processes that pose a potential threat to human, animal, or plant health or the environment. Current research has proposed several methods for identifying biological risk factors, including biomarker-based analysis, statistical model building, and the application of machine learning algorithms.

[0003] Scientific literature, as an important form of reflecting cutting-edge research, carries researchers' explorations and discoveries in the field of biosafety. These documents not only record the latest achievements in biosafety but also include in-depth research and discussions on related issues. By exploring the important connections and hot topics in these biosafety-related documents, as well as the patterns and regularities within these connections, we can provide theoretical and empirical support for further prediction of biological risk factors. These connections and hot topics involve aspects such as the discovery of emerging pathogens, providing valuable clues and information sources for the identification and prediction of biological risk factors.

[0004] Therefore, based on research literature in the field of biosafety, we construct a complex academic network including pathogenic factors and related factors such as biotechnology to explore the potential relationships and interactions among biorisk-related research. Traditional complex network analysis methods often rely on the analysis of single or a small number of features, neglecting the interrelationships and complexity between features, and are prone to mispredicting statistical features unrelated to labels as key features. This approach leads to an incomplete and inaccurate identification and understanding of features and relationships in the network. We utilize feature decoupling technology to transform the original feature space into a new feature space to capture the truly relevant key features influencing biorisk factors and eliminate the influence of irrelevant statistical features. Through feature decoupling, the statistical correlation of features in the original feature space can be reduced, thereby better exploring the potential patterns and relationships among biorisk research and further mining biorisk factors.

[0005] The ultimate goal of this invention is to construct a complex network integrating multiple features based on massive academic literature data related to biosafety. By studying the intrinsic relationships and attribute characteristics of the nodes in the constructed network, a model suitable for effective risk identification research with multiple features is established, which can further accurately and reliably identify biological risk factors, thereby improving the ability and efficiency to cope with biosafety challenges. This invention has significant theoretical significance and practical application value. Summary of the Invention

[0006] The technical solution of this invention is a method for identifying biological risk factors based on feature decoupling. It primarily utilizes a feature-decoupling anomaly detection technique to uncover potential relationships and interactions among biological risk studies, building upon the complex academic network comprised of numerous factors related to biosafety, such as pathogens, biotechnology, and personnel / institutions. This reveals the complex relationships between potential biological risk studies and develops a more accurate and reliable method for identifying biological risk factors. This method will contribute to a deeper understanding of biological risk factors from both theoretical and empirical perspectives, providing a scientific basis for formulating biosafety policies and preventing biosafety threats, and promoting research and application in the field of biosafety.

[0007] This invention collects biosafety literature data, extracts entities and semantic relationships to construct a biological academic network, performs characterization and feature decoupling, and distinguishes between biological and non-biological risk factors through anomaly score calculation.

[0008] This method specifically includes the following steps:

[0009] (1) Constructing a complex academic network for biological risks:

[0010] The title of the paper is used as a unique identifier for each node in the network, and deduplication is performed.

[0011] Papers are treated as entities, the abstracts and keywords contained in the papers are treated as attribute information of the papers, the citation relationships between papers are defined as association relationships, and complex networks are constructed by obtaining entities and their association relationships based on the acquired paper database through entity extraction technology.

[0012] The formal definition of the constructed complex network is: ,in It is an adjacency matrix, each element Represents a node and nodes Are there any connecting edges between them? If so, then Otherwise, it is 0. It is an attribute matrix. Corresponding to node The attribute representation, yes The set of all contained nodes;

[0013] (2) Characterizing complex academic networks of biological risks: Using graph convolutional neural networks, all nodes in the network are mapped to a unified representation space, and a low-dimensional representation of the joint network structure and attribute features is learned.

[0014] Specifically, for the network Construct a mapping function g, i.e., a 2-layer GCN, to... Using the adjacency matrix A and attribute matrix X as inputs to the mapping function, we obtain the low-dimensional representation matrix of the nodes in the network:

[0015] (1)

[0016] in, It is a non-linear activation function. , This is the weight matrix. It is a node degree matrix. It is the identity matrix. It is a degree-normalized adjacency matrix.

[0017] (3) Decoupling of biological factors characteristics

[0018] Assume that the low-dimensional representation of each node consists of a semantic factor C, i.e., relevant key features, and an environmental factor S, i.e., irrelevant statistical features;

[0019] Define encoder For semantic factor encoder, i.e. , For environmental factor encoders, D stands for decoder, i.e. ;

[0020] By combining the semantic factor C with new environmental factors randomly sampled from a standard Gaussian distribution. Construct new representations that differ from the original features but whose key features remain unchanged. ;

[0021] Similarly, for the newly generated representations It is also composed of semantic factors and environmental factors;

[0022] Then, the encoder and decoder are trained using a bidirectional reconstruction loss and an adversarial loss;

[0023] Reconstruction loss includes the original representation Reconstruction, semantic factors Reconstruction and environmental factors Reconstruct the equations, corresponding to formulas (2), (3), and (4) respectively.

[0024] (2)

[0025] (3)

[0026] (4)

[0027] Furthermore, in order to generate representations through constraints With the original representation To ensure semantic factor encoders are as similar as possible A discriminator is introduced to capture the key features that determine the label. ,in Representation of Generative Representation The generated representation is based on the similarity probability score between the original representation and the original representation. These are negative samples. The discriminator is then trained using the following method:

[0028] (5)

[0029] Finally, the encoder, decoder and discriminator are jointly trained based on formulas (2)-(5) to capture semantic factors and environmental factors, thereby capturing the key features that are truly relevant to biological factors and eliminating the influence of irrelevant statistical features;

[0030] (4) Identify biological risk factors

[0031] Define an anomaly scorer Evaluate the anomaly score for each node. To more accurately guide the calculation of outlier scores, a reference score based on Gaussian prior is introduced. The reference score is defined as the average of the outlier scores of a randomly selected set of normal nodes, serving as a reference for quantifying the degree of deviation between the outlier scores and the scores of normal nodes.

[0032] First, draw from the Gaussian prior distribution. The abnormal scores, i.e. Each This represents the anomaly score of a random normal node;

[0033] Abnormal scores and reference scores The calculation process is shown in formulas (6) and (7):

[0034] (6)

[0035] (7)

[0036] in, It is a node Abnormal scores, and These are the learnable weight matrix and the weight vector, respectively. and It is the corresponding bias term;

[0037] Furthermore, nodes The deviation of outlier scores from reference scores is defined using standard scores: ,in Indicates sampling anomaly score The standard deviation; then, an adversarial loss is introduced instead of bias loss to train the anomaly scorer:

[0038] (8)

[0039] in, It is a node The true label;

[0040] If node It is an abnormal node. ,otherwise , It is the deviation turning radius;

[0041] Finally, after obtaining the anomaly scores of all nodes, they are sorted, and nodes exceeding a given anomaly threshold are identified as biological risk nodes.

[0042] Beneficial effects

[0043] Based on literature data in the field of biosafety, this invention maps a complex network composed of many factors related to biosafety (such as pathogens, biotechnology, etc.) to a low-dimensional dense vector space. Based on this vector space, an abnormal feature extraction model based on feature decoupling is established to achieve accurate identification of biorisk nodes.

[0044] The method proposed in this invention models the network of academic literature related to biosafety in a more concrete and vivid way, and detects biological risk factors based on a more interpretable feature decoupling anomaly detection method, thereby improving the accuracy and reliability of biological risk factor identification.

[0045] This method has good robustness and transferability, and can be applied not only to risk factor identification in biological academic networks, but also to other types of networks. Attached Figure Description

[0046] Figure 1 This is a framework diagram of a model for identifying biological risk factors based on feature decoupling;

[0047] Figure 2 This is a flowchart for identifying biological risk factors based on literature data in the biological field. Detailed Implementation

[0048] The present invention will be further described below with reference to the accompanying drawings.

[0049] The proposed biological risk factor identification method based on special decoupling is mainly applied to modeling the potential relationships and interactions among biological risk factors, and to uncover important correlations and hot issues in biosafety literature. This method can provide important references for the timely identification of biological risk factors. Figure 1 This is a framework diagram of a model for identifying biological risk factors based on feature decoupling. Figure 2 This is a schematic diagram of the main process of the present invention.

[0050] This approach consists of three main steps. First, based on biosafety-related scientific literature data, entity and semantic extraction techniques are used to extract entities containing elements such as biotechnology and pathogenic factors, along with their corresponding attribute values, to construct a complex academic network of biorisk. Then, graph neural network methods are used to map the constructed complex network to a unified representation space, and feature decoupling techniques are used to capture key features and statistically irrelevant features. Finally, anomaly detection methods are used to evaluate anomaly scores based on the captured features, thereby identifying biorisk factors.

[0051] Example 1: Taking the study of identifying infectious disease risks based on biosafety literature as an example, the implementation steps are as follows:

[0052] First, based on biosafety-related scientific literature data, entity extraction techniques using infectious disease-related keywords (such as "infection") are employed to extract relevant paper titles, paper attributes (publication year, publication location, author, institution, abstract, keywords, etc.), and citation relationships between papers. The titles of all papers are used as unique identifiers for each node in the network, and deduplication is performed to create a node set. The attributes of each node are encoded into vectors of a certain length using semantic representation techniques. The existence of citation relationships between papers is determined to construct edges between them, thus building a complex biosafety literature network. Finally, graph neural network algorithms (such as GCN) are used to map all nodes in the constructed complex network to a unified low-dimensional space, obtaining a low-dimensional representation of each node.

[0053] Semantic and environmental factor representations of nodes are learned based on attribute reconstruction and adversarial training. The semantic factor representations are used as input to an anomaly evaluator to calculate anomaly scores for all nodes, and a Gaussian prior reference score is introduced to guide the anomaly evaluator's learning. Reconstruction loss, adversarial loss, and anomaly loss are jointly optimized. An anomaly ranking list is formed by sorting the anomaly scores of all nodes, and a literature list is formed by finding the corresponding literature for each node. The literature list corresponding to the top 100 nodes in the anomaly ranking list is traversed. If the keywords and abstract of the corresponding literature explicitly indicate that the virus or other infectious agent studied in the paper is infectious and has not been studied in previous research, then the relevant research in that literature is defined as risk research.

[0054] Example 2

[0055] Taking research on identifying infectious disease risks based on biosafety literature as an example, the implementation steps are as follows:

[0056] First, based on biosafety-related scientific literature data, entity extraction techniques using infectious disease-related keywords such as monkeypox are employed to extract relevant paper titles, paper attributes (publication year, publication location, author, institution, abstract, keywords, etc.), and citation relationships between papers. The titles of all papers are used as unique identifiers for each node in the network, and after deduplication, a node set is created. The attributes of each node are encoded into vectors of a certain length using semantic representation techniques. The existence of citation relationships between papers is determined to construct edges between them, thus building a complex biological literature network. Graph neural network algorithms (such as GCN) are used to map all nodes in the constructed complex network to a unified low-dimensional space, obtaining a low-dimensional representation for each node. Semantic factor representations and environmental factor representations of nodes are learned based on attribute reconstruction and adversarial training. The semantic factor representations are used as input to an anomaly evaluator to calculate anomaly scores for all nodes, and a Gaussian prior reference score is introduced to guide the anomaly evaluator's learning. Reconstruction loss, adversarial loss, and anomaly loss are jointly optimized. The anomaly scores of all nodes are sorted to form an anomaly ranking list, and the corresponding literature for each node is used to form a literature list. Traverse the list of literature corresponding to the top 200 in the anomaly ranking list. If the keywords and abstract of the corresponding literature clearly indicate that the virus or other pathogen studied in the paper is infectious and has not been studied in previous studies, then the relevant research in the literature is defined as risk research.

[0057] Example 3: Taking research on identifying infectious disease risks based on biosafety literature as an example, the implementation steps are as follows:

[0058] First, based on biosafety-related scientific literature data, entity extraction techniques using infectious disease-related keywords such as anthrax are employed to extract relevant paper titles, paper attributes (publication year, publication location, author, institution, abstract, keywords, etc.), and citation relationships between papers. The titles of all papers are used as unique identifiers for each node in the network, and after deduplication, a node set is created. The attributes of each node are encoded into vectors of a certain length using semantic representation techniques. The existence of citation relationships between papers is determined to construct edges between them, thus building a complex biosafety literature network. Graph neural network algorithms (such as GCN) are used to map all nodes in the constructed complex network to a unified low-dimensional space, obtaining a low-dimensional representation for each node. Semantic factor representations and environmental factor representations of nodes are learned based on attribute reconstruction and adversarial training. The semantic factor representations are used as input to an anomaly evaluator to calculate anomaly scores for all nodes, and a Gaussian prior reference score is introduced to guide the anomaly evaluator's learning. Reconstruction loss, adversarial loss, and anomaly loss are jointly optimized. The anomaly scores of all nodes are sorted to form an anomaly ranking list, and the corresponding literature for each node is used to form a literature list. Traverse the list of literature corresponding to the top 200 in the anomaly ranking list. If the keywords and abstract of the corresponding literature clearly indicate that the virus or other pathogen studied in the paper is infectious and has not been studied in previous studies, then the relevant research in the literature is defined as risk research.

Claims

1. A method for identifying biological risk factors based on feature decoupling, characterized in that, By collecting biosafety literature data, extracting entities and semantic relationships, constructing a bio-academic network, performing characterization and feature decoupling, and distinguishing between biological and non-biological risk factors through anomaly score calculation; Includes the following steps: (1) Constructing a complex academic network for biological risks: The title of the paper is used as a unique identifier for each node in the network, and deduplication is performed. Papers are treated as entities, the abstracts and keywords contained in the papers are treated as attribute information of the papers, the citation relationships between papers are defined as association relationships, and complex networks are constructed by obtaining entities and their association relationships based on the acquired paper database through entity extraction technology. The formal definition of the constructed complex network is: ,in It is an adjacency matrix, each element Represents a node and nodes Are there any connecting edges between them? If so, then Otherwise, it is 0. It is an attribute matrix. Corresponding to node The attribute representation, yes The set of all contained nodes; (2) Characterizing complex academic networks of biological risks: Using graph convolutional neural networks, all nodes in the network are mapped to a unified representation space, and a low-dimensional representation of the joint network structure and attribute features is learned. Specifically, for the network Construct a mapping function g, i.e., a 2-layer GCN, to... Using the adjacency matrix A and attribute matrix X as inputs to the mapping function, we obtain the low-dimensional representation matrix of the nodes in the network: （1） in, It is a non-linear activation function. , This is the weight matrix. It is a node degree matrix. It is the identity matrix. It is a degree-normalized adjacency matrix; (3) Decoupling of biological factors characteristics Assume that the low-dimensional representation of each node consists of a semantic factor C, i.e., relevant key features, and an environmental factor S, i.e., irrelevant statistical features; Define encoder For semantic factor encoder, i.e. , For environmental factor encoders, D stands for decoder, i.e. ; By combining the semantic factor C with new environmental factors randomly sampled from a standard Gaussian distribution. Construct new representations that differ from the original features but whose key features remain unchanged. ; Similarly, for the newly generated representations It is also composed of semantic factors and environmental factors; Then, the encoder and decoder are trained using a bidirectional reconstruction loss and an adversarial loss; Reconstruction loss includes the original representation Reconstruction, semantic factors Reconstruction and environmental factors Reconstruct the equations, corresponding to formulas (2), (3), and (4) respectively. （2） （3） （4） Furthermore, in order to generate representations through constraints With the original representation To ensure semantic factor encoders are as similar as possible A discriminator is introduced to capture the key features that determine the label. ,in Representation of Generative Representation The generated representation is based on the similarity probability score between the original representation and the original representation. The samples are negative; then the discriminator is trained using the following method: （5） Finally, the encoder, decoder and discriminator are jointly trained based on formulas (2)-(5) to capture semantic factors and environmental factors, thereby capturing the key features that are truly relevant to biological factors and eliminating the influence of irrelevant statistical features; (4) Identify biological risk factors Define an anomaly scorer Evaluate the anomaly score for each node. To more accurately guide the calculation of outlier scores, a reference score based on Gaussian prior is introduced. The reference score is defined as the average of the abnormal scores of a randomly selected set of normal nodes, serving as a reference for quantifying the degree of deviation between the abnormal scores and the scores of normal nodes.

2. The biological risk factor identification method based on feature decoupling according to claim 1, characterized in that, The specific step (4) is as follows: First, draw from the Gaussian prior distribution. The abnormal scores, i.e. Each This represents the anomaly score of a random normal node; Abnormal scores and reference scores The calculation process is shown in formulas (6) and (7): （6） （7） in, It is a node Abnormal scores, and These are the learnable weight matrix and the weight vector, respectively. and It is the corresponding bias term; Furthermore, nodes The deviation of outlier scores from reference scores is defined using standard scores: ,in Indicates sampling anomaly score The standard deviation; then, an adversarial loss is introduced instead of bias loss to train the anomaly scorer: （8） in, It is a node The true label; If node It is an abnormal node. ,otherwise , It is the deviation turning radius; Finally, after obtaining the anomaly scores of all nodes, they are sorted, and nodes exceeding a given anomaly threshold are identified as biological risk nodes.

Images