Gene function diagnosis system based on twin networks and multi-order tensor information

Abstract

The present invention discloses a gene function diagnosis system based on twin networks and multi-order tensor information, including: a data management module, which is responsible for storing and managing genome data; a data import module, which pre-processes the genome data and generates a high-quality gene function feature matrix; a sample similarity module, which uses an improved similarity calculation method to generate corresponding sample similarity information for the high-quality gene function feature matrix; a functional diagnosis training module, which uses an automatic encoder to convert the high-quality gene function feature matrix into a deep representation, and combines the sample similarity module with the multi-order tensor similarity fusion method to train the improved twin network; a gene function diagnosis prediction module, which performs intelligent diagnosis of gene function based on the trained improved twin network, combines gene ontology annotations and provides an explanation of the diagnosis results. The present invention combines multi-order tensor information, twin networks and genomics data to improve diagnostic capabilities and the interpretability of diagnostic results.

Description

Technical Field

[0001] The present invention relates to the technical field of gene function diagnosis, and in particular to a gene function diagnosis system based on twin networks and multi-order tensor information. Background Art

[0002] Advances in next-generation gene sequencing technology have made it possible to collect biological information at higher dimensionalities. This rich data provides a rich source of features for supervised learning tasks. Consequently, researchers tend to use neural networks, which excel in supervised learning, to predict gene function. However, due to the high cost of sequencing and limited prior knowledge, the number of genetic samples available for training is often limited. This often presents challenges with high-dimensional, small-sample data, where the number of features far outnumbers the number of samples. In this scenario, the main challenges faced by traditional deep network algorithms when diagnosing high-dimensional, small-sample genetic data can be categorized into two categories: 1. Concentration effect: This refers to the tendency for distance metrics between data points to collapse to a constant in high-dimensional space as the number of dimensions increases; 2. Overfitting: Due to the sparsity of samples, deep networks are prone to overfitting, as the model may overfit the training data. Given these two challenges, designing a gene function prediction system tailored to the features of high-dimensional, small-sample data is an urgent task.

[0003] The current high-dimensional small-sample gene function diagnosis systems are mainly divided into three categories: 1) Systems based on high-order statistical similarity measurement: Drawing on the traditional point-to-point second-order similarity calculation idea, the gene function attribution is determined by calculating the distance information between multiple samples; 2) Systems based on traditional machine learning methods: Using traditional machine learning methods such as support vector machines, decision trees, ensemble learning, multi-kernel learning, tensor decomposition technology and perceptrons, they find functional difference representations in genetic data by simply extracting features; 3) Systems based on distance measurement deep methods: High-dimensional features are mapped to low-dimensional space through deep networks, and similarity measurement criteria are used for representation learning for downstream gene function difference differentiation tasks.

[0004] In general, existing high-dimensional, small-sample gene function diagnosis systems primarily focus on addressing the concentration effect brought about by high-dimensional features. They use different data processing criteria for subsequent processing, but fail to adequately consider computational complexity and sample sparsity. This means that current systems are unable to fully tap into the intrinsic spatial information between high-dimensional samples, nor can they effectively achieve the complementary fusion of high-order information, leading to low accuracy in gene function prediction. Furthermore, most current diagnostic systems fail to account for the over-adaptation problem brought about by sample sparsity, which makes the model prone to overfitting, lacks interpretability, and fails to provide evidence to support diagnostic results. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings and deficiencies of the existing technology, and proposes a gene function diagnosis system based on twin networks and multi-order tensor information. In response to the concentration effect problem brought by high dimensions, high-order correlations between binary, ternary and quaternary data groups are introduced, and high-order information is losslessly fused through tensor multiplication. A random batch training method is used to reduce resource consumption and guide the generation of gene functional representations. Considering the overfitting challenge caused by sample sparsity, the system integrates high-order similarity into the twin network framework, and re-evaluates the relationship between two data combinations, three data combinations and four data combinations, which is similar to enhancing sample information to reduce the overfitting effect. The system provides a certain degree of explainability for the gene function diagnosis results, thereby enhancing the credibility of the diagnosis results.

[0006] To achieve the above objectives, the present invention provides a technical solution: a gene function diagnosis system based on twin networks and multi-order tensor information, comprising:

[0007] The data management module is responsible for storing and managing genomic data. The genomic data is stored in a database and provides data storage and query functions for the system;

[0008] The data import module is used to load gene sequencing data, pre-process the data in the data management module through statistical analysis and semantic similarity analysis, and then generate a high-quality gene function feature matrix based on the biological background;

[0009] The sample similarity module uses an improved similarity calculation method to generate corresponding sample similarity information for high-quality gene function feature matrices. This improved similarity calculation method uses multiple sample point metrics to replace the traditional point-by-point metrics, learns the intrinsic spatial representation between sample points, and uses tensor product technology to achieve lossless fusion of sample representation and sample similarity information.

[0010] The functional diagnosis training module uses an autoencoder to convert the high-quality gene function feature matrix generated in the data import module into the corresponding deep gene representation. Based on the deep gene representation, the improved twin network is trained in combination with the sample similarity module. The improved twin network calculates the spatial distribution of the deep gene representation based on the multiple sample point measurement criteria designed by the sample similarity module, and uses tensor product technology to achieve lossless fusion of the deep gene representation and sample similarity information. Finally, the sample similarity module is combined with the gene function diagnosis task to form a unified learning framework, and finally a trained improved twin network, i.e., the optimal diagnostic model, is obtained.

[0011] The gene function diagnosis and prediction module performs intelligent diagnosis of gene function based on the optimal diagnostic model; among them, the gene function diagnosis and prediction module divides gene function into three key categories: molecular function, cellular component and biological process, combines with the external gene ontology annotation database and provides explanations of the diagnostic results.

[0012] Furthermore, in the data management module, the genomic data is stored in a database built by SQLite3. SQLite3 is a lightweight embedded relational database management system. Python provides a built-in SQLite3 module responsible for establishing an index structure to speed up the retrieval of genomic data and provide efficient query functions, including:

[0013] a. Index structure design: To achieve efficient genomic data retrieval, the choice of specific index structure depends on the characteristics of the data, query requirements, and performance requirements;

[0014] b. Gene name query: supports querying based on gene name. It can provide the common name or identifier of the gene, quickly locate the corresponding genome data based on the index structure, and return relevant data fragments or complete genome data;

[0015] c. Genome location query: Supports query of genome coordinates for the location information of genome data. By specifying the chromosome number, starting position, and ending position, the corresponding genome data will be quickly retrieved and returned using the index structure.

[0016] d. Multi-condition query: Supports multi-condition query and can perform combined query based on key information, including gene name, genome location and other annotation information. Based on the index structure and query optimization algorithm, it will efficiently filter out genomic data that meets the conditions;

[0017] e. Data fragment return: When performing genome data retrieval, data fragments of the specified area can be returned;

[0018] f. High-performance query optimization: To further improve the performance of genomic data retrieval, query optimization techniques can be used;

[0019] g. Query result sorting and filtering: When query results are returned, they can be sorted and filtered as needed.

[0020] Furthermore, the data import module includes a data loading module and a data preprocessing module, wherein:

[0021] The data loading module reads gene sequencing data from the local machine and supports reading data from a variety of common genomic data formats. By supporting different formats, it can easily import various types of genomic data. During the loading process, the data loading module parses and verifies the data in the data management module to ensure the integrity and accuracy of the data. For large-scale genomic data in the data management module, the data loading module can support data partitioning and sharding storage.

[0022] The data preprocessing module is used to construct a gene function set and generate a high-quality gene function feature matrix. The gene ontology is a classification system that provides a standardized description of gene and protein functions. It divides gene functions into three levels of components, including molecular functions, cellular components, and biological processes. Each component has a series of terms and annotations used to describe the participation of genes in specific functions or processes. Based on the hierarchical structure and relationships between terms in the gene ontology, the functional similarity between genes is measured by calculating the similarity scores between gene terms in gene annotations. In statistical analysis, the use of semantic similarity methods can be used to reduce the importance of statistical p-values, reduce false positive problems, and improve the reliability of results.

[0023] Furthermore, the sample similarity module includes a similarity tensor construction process and a high-order similarity tensor fusion process, wherein:

[0024] a. Regarding the similarity tensor construction process, there are:

[0025] a1. Second-order similarity tensor: Select Euclidean distance as the metric and judge the similarity of samples by the distance between sample pairs. Assuming that there is samples, select sample pair x i 、x j , their distance metric d in Euclidean space ij Expressed as:

[0026] d ij =||x i -x j ||2

[0027] Where, d ij Represents the sample pair x i 、x j The distance in Euclidean space, ||·||2 represents the matrix norm, so it contains the sample point x i 、x j The second-order similarity tensor of Corresponding elements for:

[0028]

[0029] Where exp is the natural exponential function and σ is a scaling constant;

[0030] a2. Third-order similarity tensor: Considering that high-order similarity metrics help to expand limited sample information, this is achieved by integrating information between multiple samples to achieve full interaction of sample information. Therefore, based on the second-order similarity tensor metric, a similarity metric criterion for triplet data sets is introduced, including sample point x i 、x j 、x k The third-order similarity tensor of Corresponding elements for:

[0031]

[0032] Where, d jk Represents the sample pair x j 、x k The distance in Euclidean space, for each index i, j, is the total number of samples, elements Indicates the sample x at the anchor point j Under the condition that the sample x i and sample x k similarities between;

[0033] a3. Further capture the relationship between the four-dimensional data group and introduce the similarity tensor of the four-dimensional data group, including the sample point x i 、x j 、x k 、x l The fourth-order similarity tensor of Corresponding elements for:

[0034]

[0035] Where, d kl Represents the sample pair x k 、x l The distance in Euclidean space, d ik Represents the sample pair x i 、x k The distance in Euclidean space, d jl Represents the sample pair x j 、x l The distance in Euclidean space, ε is a very small number, Represents the fourth-order similarity tensor, measuring the sample x i 、x j 、x k 、x lThe intrinsic correlation information between the four data groups;

[0036] b. For the high-order similarity tensor fusion process, there are:

[0037] The key lies in effectively utilizing the inherent high-order correlation interaction features between samples to guide the generation of representations, and using the tensor mode-k product to losslessly integrate the inherent high-order information of genomic data. The mode-k product is defined as follows:

[0038]

[0039] Where, It is the defined mode-k product symbol, k is a variable parameter that can set a specific value. is a tensor of order m, represents the field of real numbers, Representing a tensor Corresponding subscripts i1...i k-1 i k i k+1 ...i m The specific element values, n1, n2, ..., n m Represents the specific value of the m-th tensor dimension, i m Represents the sub-sequence number in the m-th tensor dimension, 0≤i m ≤n m , n k is the specific value of the k-th tensor dimension, i k Represents the sub-sequence number in the k-th tensor dimension, 0≤i k ≤n k , Indicates the subscript ji corresponding to the two-dimensional matrix V k The specific element value of

[0040] The purpose of fusing tensor similarity is to guide the generation of deep gene representation through the intrinsic spatial information between multiple samples, so the mode-k product is used to losslessly fuse the similarity tensor. Intrinsic spatial correlation with the deep gene representation F:

[0041]

[0042] Where, represents the m-order similarity tensor, is the mode-k product defined above, Set the variable parameter k of mode-k to a tensor The order m, It is the deep representation of the high-quality gene function feature matrix, and its corresponding column j is F :,j .

[0043] Furthermore, the functional diagnosis training module includes a feature representation module, a high-order similarity fusion module, and a twin network training module that fuses multi-order similarity tensors, wherein:

[0044] The feature representation module is used for feature representation of gene sample information. It uses an autoencoder to convert the high-quality gene function feature matrix into a corresponding deep vector form. Specifically, the encoder f(·) is used to encode the high-quality gene function feature matrix:

[0045] F=f(X)

[0046] Where, It is a high-quality gene function feature matrix generated in the data import module; the encoder consists of a 4-layer fully connected neural network. The first layer is the input layer of the original feature. The number of nodes is the second dimension m of the original feature. The number of nodes in the second, third, and fourth layers are 500, 500, and 2000 respectively. The node dimension corresponding to the fourth layer corresponds exactly to the second dimension of the deep representation F. The activation function between each layer is ReLU, and a normalization layer is added after the last layer of activation function;

[0047] The high-order similarity fusion module uses high-order similarity measurement criteria to perform sample relationship space progression on the deep representation of high-quality gene function feature matrix, including:

[0048] a、For the sample point x i 、x j The second-order similarity tensor of The fusion method is:

[0049]

[0050] Where, It is the mode-k product, where the variable parameter k is 1 or 2;

[0051] b. For the sample point x i 、x j 、x k The third-order similarity tensor of According to the third-order similarity tensor construction formula, the corresponding similarity fusion method is:

[0052]

[0053] Where, is the mode-k product, where the variable parameter k is 1, 2, or 3;

[0054] c. For the sample point x i 、xj 、x k 、x l The fourth-order similarity tensor of According to the fourth-order similarity tensor construction formula, the corresponding similarity fusion method is:

[0055]

[0056] Where, is the mode-k product, where the variable parameter k is 1, 2, 3, or 4;

[0057] The twin network training module integrating multi-order similarity tensors utilizes the sample similarity module to improve the twin network and combines the cross entropy loss function training to obtain the optimal diagnostic model, including:

[0058] a. Construct the target variable T of the sample pair required by the twin network based on the genome label ij :

[0059]

[0060] In the formula, the subscript ij indicates that the target variable measures the gene sample x i 、x j ;

[0061] b. According to the encoding process designed above, the high-quality gene function feature matrix X is mapped into the deep representation space:

[0062] F=f(X)

[0063] c. Combining the second-order, third-order, and fourth-order similarity tensor formulas designed by the sample similarity module with the deep representation F of the high-quality gene function feature matrix to construct the corresponding second-order, third-order, and fourth-order similarity tensors

[0064] d. Based on the designed tensor fusion method, the similarity tensors of different orders are losslessly fused with the deep representation F of the high-quality gene function feature matrix using the mode-k product:

[0065]

[0066] e. Improve the twin network based on the multi-order tensor similarity fusion method designed in step d, and design the loss function in combination with the gene function diagnosis task Perform representation learning:

[0067]

[0068] Where, is the cross entropy loss function, y iu Represents sample xi Whether it belongs to category u; is the sample x i Predictions belonging to category u; T ij It is the sample pair target variable constructed by the genome label. The subscript ij indicates that the target variable measures the gene sample x i 、x j , α, β, s are hyper parameters, is the number of samples, e is the natural logarithm, c represents the number of categories, F ij and It is the deep representation of high-quality gene function feature matrix F corresponding to subscript ij, The specific element value of

[0069] f. Train the improved twin network to obtain the optimal diagnostic model, that is, the trained improved twin network.

[0070] Furthermore, the gene function diagnosis and prediction module is used to perform intelligent diagnosis on the high-quality gene function feature matrix and provide an explanation of the diagnosis results, including the following steps:

[0071] 1) using the data import module to load the gene data set, and processing the gene data according to the data preprocessing module to obtain a high-quality gene function feature matrix;

[0072] 2) Inputting the high-quality gene function feature matrix obtained after processing in step 1) into the functional diagnosis training module for training, and training through the feature representation module, high-order similarity fusion module, and twin network training module that integrates multi-order similarity tensors to obtain the optimal diagnostic model;

[0073] 3) Extracting a deep representation of the high-quality gene functional feature matrix in the optimal diagnostic model in step 2) above, and then performing spatial analysis on the gene representation to provide interpretability for the functional partitioning of the genome.

[0074] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0075] 1. Gene feature data are interacted at the micro-data level, and high-order similarity tensor criteria measured by multiple sample points are used to enrich the spatial distribution of deep gene representation. The similarity tensor information of binary, triple, and quadruple gene samples is comprehensively considered to alleviate the concentration effect caused by high-dimensional small sample gene data.

[0076] 2. Select a sophisticated twin network metric to evaluate the relationship between data pairs, three-data combinations, and four-data combinations, similar to enhancing sample information to alleviate the overfitting effect caused by the sparsity of genetic samples.

[0077] 3. By losslessly fusing the multi-group information of the genome through tensor multiplication, we can achieve a progressive representation of the deep spatial relationships of the genome, combine the downstream gene function diagnosis tasks, and form a unified learning framework. Ultimately, we obtain a trained improved twin network, that is, the optimal diagnostic model, which enhances the accuracy of diagnosis.

[0078] 4. The random batch training method is used to reduce resource consumption. It is not only applicable to high-dimensional small-sample genetic phenotypic data, but can also be extended to ultra-high-dimensional large-sample phenotypic data, accelerating the training process of the twin network with high-order similarity improvement, which is conducive to revealing the phenotypic differences of different gene functions and enhancing the accuracy and credibility of gene function diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 Schematic diagram of the relationship between the various modules of the system of the present invention.

[0080] Figure 2 Flowchart of training and prediction of the system of the present invention.

[0081] Figure 3 Schematic diagram of the automatic encoder used in the system of the present invention.

[0082] Figure 4 Schematic diagram of the twin network of multi-order similarity tensor fusion of the system of the present invention. DETAILED DESCRIPTION

[0083] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0084] This embodiment discloses a gene function diagnosis system based on twin networks and multi-order tensor information. It is a gene function diagnosis system developed in Python and can be run on Windows devices. The relationship between the modules of the system is as follows: Figure 1 As shown, the system training and prediction process is as follows Figure 2 It includes:

[0085] The data management module is responsible for storing and managing genomic data. The genomic data is stored in a database and provides data storage and query functions for the system;

[0086] The data import module is used to load gene sequencing data, pre-process the data in the data management module through statistical analysis and semantic similarity analysis, and then generate a high-quality gene function feature matrix based on the biological background;

[0087] The sample similarity module uses an improved similarity calculation method to generate corresponding sample similarity information for high-quality gene function feature matrices. This improved similarity calculation method uses multiple sample point metrics to replace the traditional point-by-point metrics, learns the intrinsic spatial representation between sample points, and uses tensor product technology to achieve lossless fusion of sample representation and sample similarity information.

[0088] The functional diagnosis training module uses an autoencoder to convert the high-quality gene function feature matrix generated in the data import module into the corresponding deep gene representation. Based on the deep gene representation, the improved twin network is trained in combination with the sample similarity module. The improved twin network calculates the spatial distribution of the deep gene representation based on the multiple sample point measurement criteria designed by the sample similarity module, and uses tensor product technology to achieve lossless fusion of the deep gene representation and sample similarity information. Finally, the sample similarity module is combined with the gene function diagnosis task to form a unified learning framework, and finally a trained improved twin network, i.e., the optimal diagnostic model, is obtained.

[0089] The gene function diagnosis and prediction module performs intelligent diagnosis of gene function based on the optimal diagnostic model; among them, the gene function diagnosis and prediction module divides gene function into three key categories: molecular function, cellular component and biological process, combines with the external gene ontology annotation database and provides explanations of the diagnostic results.

[0090] Specifically, in the data management module, the genomic data is stored in a database built using SQLite3. When genomic data needs to be stored or called, the corresponding function code is triggered to run, retrieve and extract the corresponding genomic data from the database, and transfer the data to the target location. SQLite3 is a lightweight embedded relational database management system. Python provides a built-in SQLite3 module that is responsible for establishing an index structure to speed up the retrieval of genomic data and provide efficient query functions, including:

[0091] a. Index structure design: To achieve efficient genomic data retrieval, the choice of specific index structure depends on the characteristics of the data, query requirements, and performance requirements;

[0092] b. Gene name query: supports querying based on gene name. It can provide the common name or identifier of the gene, quickly locate the corresponding genome data based on the index structure, and return relevant data fragments or complete genome data;

[0093] c. Genome location query: Supports query of genome coordinates for the location information of genome data. By specifying the chromosome number, starting position, and ending position, the corresponding genome data will be quickly retrieved and returned using the index structure.

[0094] d. Multi-condition query: Supports multi-condition query and can perform combined query based on key information, including gene name, genome location and other annotation information. Based on the index structure and query optimization algorithm, it will efficiently filter out genomic data that meets the conditions;

[0095] e. Data fragment return: When performing genome data retrieval, data fragments of the specified area can be returned;

[0096] f. High-performance query optimization: To further improve the performance of genomic data retrieval, query optimization techniques can be used;

[0097] g. Query result sorting and filtering: When query results are returned, they can be sorted and filtered as needed.

[0098] Specifically, the data import module includes a data loading module and a data preprocessing module, wherein:

[0099] The data loading module reads gene sequencing data from the local machine and supports reading data from a variety of common genomic data formats. By supporting different formats, it can easily import various types of genomic data. During the loading process, the data loading module also parses and verifies the data in the data management module to ensure the integrity and accuracy of the data. For large-scale genomic data in the data management module, the data loading module can support data partitioning and sharding storage.

[0100] The data preprocessing module is used to construct a gene function set and generate a high-quality gene function feature matrix. Gene Ontology (GO) is a classification system for standardized description of gene and protein functions. It divides gene functions into three levels of components, including molecular function, cellular component and biological process. Each component has a series of terms and annotations used to describe the participation of genes in specific functions or processes. Based on the hierarchical structure and relationship between terms in the gene ontology, the functional similarity between genes is measured by calculating the similarity score between gene terms in gene annotations. In statistical analysis, the use of semantic similarity method can be used to reduce the importance of statistical p-values, reduce false positive problems, and improve the reliability of results.

[0101] Specifically, we first use online tools for gene function annotation and analysis to convert gene identifiers into the internal format of the gene set. Then, we analyze the gene tables of various organisms based on the gene ontology semantic annotation tool and use the semantic similarity metric (Sim) to reduce the redundancy of the gene ontology term set. It is defined as follows:

[0102]

[0103] Where g1 and g2 are two gene ontology terms, P() is the relative frequency of gene ontology terms in the UniProt gene sequence information database, MIA is the ancestral term shared by the annotations of the two gene ontologies g1 and g2, and P(MIA) is the probability of this informative ancestor;

[0104] The SimRel score is a functional similarity metric that is often used to compare two GO gene ontologies. Resnik )as follows:

[0105] Sim Resnik (g1,g2)=IC(MIA)=-log(P(MIA))

[0106] Where, IC(MIA) is used to represent the information content of this informative ancestral term, which is obtained by calculating the negative of the logarithm of the frequency of occurrence of the gene annotated by this term;

[0107] The probability of each gene ontology for functional enrichment analysis and subsequent semantic similarity analysis The values ​​are defined as follows:

[0108]

[0109] Where N is the number of genes in the genome belonging to the same gene ontology level; M is the number of genes in this gene ontology term; The formula for the number of combinations is: Taking z (z ≤ M) elements from M different elements and grouping them together is called a combination of z elements from M different elements; o is the number of genes belonging to the same gene ontology level in the input data; z is the number of genes belonging to gene ontology terms in the input data. To further reduce redundancy, a gene duplication matrix R for gene ontology nodes was constructed. The calculation formula for each element in the matrix is ​​as follows:

[0110]

[0111] Where R ij Indicates the specific element value corresponding to the subscript ij in the gene repetition rate matrix R;

[0112] in:

[0113] N(GO i ,GO j )=gene_num(GO i ∩GO j )

[0114] M(GO i ,GO j )=min{gene_num(GO i ),gene_num(GO j )}

[0115] In the formula, gene_num(GO i ) represents the i-th gene ontology GO i The number of genes, ∩ represents the union;

[0116] Gene ontology terms are merged using a gene duplication matrix. Specifically, if an element of a gene ontology term belongs to another gene ontology term, the gene ontology term containing a large number of genes is retained, while the gene ontology term containing a smaller number of genes is deleted. A duplication threshold is then set to merge the gene ontology term sets into a new term set to significantly reduce the redundancy of the gene ontology term sets. The gene expression matrix on each new term set is named a high-quality gene function feature matrix.

[0117] Specifically, the sample similarity module includes a similarity tensor construction process and a high-order similarity tensor fusion process, wherein:

[0118] a. Regarding the similarity tensor construction process, there are:

[0119] a1. Second-order similarity tensor: Select Euclidean distance as the metric and judge the similarity of samples by the distance between sample pairs. Assuming that there is samples, select sample pair x i 、x j , their distance metric d in Euclidean space ij Expressed as:

[0120] d ij =||x i -x j ||2

[0121] Where, d ij Represents the sample pair x i 、x j The distance in Euclidean space is the distance between the sample pairs x. i 、x j The second-order similarity in the Euclidean space, ||·||2 represents the second norm of the matrix, so it contains the sample point x i 、xj The second-order similarity tensor of Corresponding elements for:

[0122]

[0123] Where exp is the natural exponential function, that is, exp(x) = e x , σ is a scaling constant, usually 10 -6 ;

[0124] a2. Third-order similarity tensor: Considering that high-order similarity metrics help to expand limited sample information, this is achieved by integrating information between multiple samples to achieve full interaction of sample information. Therefore, based on the second-order similarity tensor metric, a similarity metric criterion for triplet data sets is introduced, including sample point x i 、x j 、x k The third-order similarity tensor of Corresponding elements for:

[0125]

[0126] Where, d jk Represents the sample pair x j 、x k The distance in Euclidean space, for each index i, j, is the total number of samples, elements Indicates the sample x at the anchor point j Under the condition that the sample x i and sample x k similarities between;

[0127] a3. Further capture the relationship between the four-dimensional data group and introduce the similarity tensor of the four-dimensional data group, including the sample point x i 、x j 、x k 、x l The fourth-order similarity tensor of Corresponding elements for:

[0128]

[0129] Where, d kl Represents the sample pair x k 、x l The distance in Euclidean space, d ik Represents the sample pair x i 、x k The distance in Euclidean space, djl Represents the sample pair x j 、x l The distance in the Euclidean space, ε is a very small number (usually set to 10 -3 ), Represents the fourth-order similarity tensor, measuring the sample x i 、x j 、x k 、x l The intrinsic correlation information between the four data groups;

[0130] b. For the high-order similarity tensor fusion process, there are:

[0131] The key lies in effectively utilizing the inherent high-order correlation interaction features between samples to guide the generation of representations, and using the tensor mode-k product to losslessly integrate the inherent high-order information of genomic data. The mode-k product is defined as follows:

[0132]

[0133] Where, It is the defined mode-k product symbol (k is a variable parameter and can be set to a specific value). is a tensor of order m, represents the field of real numbers, Representing a tensor Corresponding subscripts i1...i k- 1i k i k+1 ...i m The specific element values, n1, n2, ..., n m Represents the specific value of the m-th tensor dimension, i m Represents the sub-sequence number (0≤i m ≤n m ), n k is the specific value of the k-th tensor dimension, i k Represents the sub-sequence number (0≤i k ≤n k ), V jik Indicates the subscript ji corresponding to the two-dimensional matrix V k The specific element value of

[0134] The purpose of fusing tensor similarity is to guide the generation of deep gene representation through the intrinsic spatial information between multiple samples, so the mode-k product is used to losslessly fuse the similarity tensor. Intrinsic spatial correlation with the deep gene representation F:

[0135]

[0136] Where, represents the m-order similarity tensor, is the mode-k product defined above ( Set the variable parameter k of mode-k to a tensor The order m), It is the deep representation of the high-quality gene function feature matrix, and its corresponding column j is F :,j .

[0137] Specifically, the functional diagnosis training module includes a feature representation module, a high-order similarity fusion module, and a twin network training module that fuses multi-order similarity tensors, wherein:

[0138] The feature representation module is used for feature representation of gene sample information, and uses an autoencoder to convert the high-quality gene function feature matrix into a corresponding deep vector form, such as Figure 3 As shown in Figure 2, the encoder f(·) is used to encode the high-quality gene function feature matrix:

[0139] F=f(X)

[0140] Where, It is a high-quality gene function feature matrix generated in the data import module; the encoder consists of a 4-layer fully connected neural network, the first layer is the input layer of the original feature, and the number of nodes is the second dimension of the original feature The number of nodes in the second, third, and fourth layers are 500, 500, and 2000 respectively. The node dimension corresponding to the fourth layer corresponds exactly to the second dimension of the depth representation F. The activation function between each layer is ReLU, and a normalization layer is added after the last layer of activation function;

[0141] The high-order similarity fusion module uses high-order similarity measurement criteria to perform sample relationship space progression on the deep representation of high-quality gene function feature matrix, including:

[0142] a、For the sample point x i 、x j The second-order similarity tensor of The fusion method is:

[0143]

[0144] Where, It is the mode-k product, where the variable parameter k is 1 or 2;

[0145] b. For the sample point x i 、x j 、x k The third-order similarity tensor of According to the third-order similarity tensor construction formula, the corresponding similarity fusion method is:

[0146]

[0147] Where, is the mode-k product, where the variable parameter k is 1, 2, or 3;

[0148] c. For the sample point x i 、x j 、x k 、x l The fourth-order similarity tensor of According to the fourth-order similarity tensor construction formula, the corresponding similarity fusion method is:

[0149]

[0150] Where, is the mode-k product, where the variable parameter k is 1, 2, 3, or 4;

[0151] The twin network training module that integrates multi-order similarity tensors uses the sample similarity module to improve the twin network, such as Figure 4 As shown in Figure 2, the optimal diagnostic model is obtained by combining the cross entropy loss function training, including:

[0152] a. Construct the target variable T of the sample pair required by the twin network based on the genome label ij :

[0153]

[0154] In the formula, the subscript ij indicates that the target variable measures the gene sample x i 、x j ;

[0155] b. According to the encoding process designed above, the high-quality gene function feature matrix X is mapped into the deep representation space:

[0156] F=f(X)

[0157] c. Combining the second-order, third-order, and fourth-order similarity tensor formulas designed by the sample similarity module with the deep representation F of the high-quality gene function feature matrix to construct the corresponding second-order, third-order, and fourth-order similarity tensors

[0158] d. Based on the designed tensor fusion method, the similarity tensors of different orders are losslessly fused with the deep representation F of the high-quality gene function feature matrix using the mode-k product:

[0159]

[0160] e. Improve the twin network based on the multi-order tensor similarity fusion method designed in step d, and design the loss function in combination with the gene function diagnosis task Perform representation learning:

[0161]

[0162] Where, is the cross entropy loss function, y iu Represents sample x i Whether it belongs to category u; is the sample x i Predictions belonging to category u; T ij is the sample pair target variable constructed by the genome label (the subscript ij indicates that the target variable measures the gene sample x i 、x j ), α, β, s are hyper parameters, is the number of samples, e is the natural logarithm, c represents the number of categories, F ij and It is the deep representation of high-quality gene function feature matrix F corresponding to subscript ij, The specific element value of

[0163] f. Train the improved twin network to obtain the optimal diagnostic model, that is, the trained improved twin network.

[0164] Specifically, the gene function diagnosis and prediction module is used to perform intelligent diagnosis on a high-quality gene function feature matrix and provide an explanation of the diagnosis results, including the following steps:

[0165] 1) using the data import module to load the gene data set, and processing the gene data according to the data preprocessing module to obtain a high-quality gene function feature matrix;

[0166] 2) Inputting the high-quality gene function feature matrix obtained after processing in step 1) into the functional diagnosis training module for training, and training through the feature representation module, high-order similarity fusion module, and twin network training module that integrates multi-order similarity tensors to obtain the optimal diagnostic model;

[0167] 3) Extracting a deep representation of the high-quality gene functional feature matrix in the optimal diagnostic model in step 2) above, and then performing spatial analysis on the gene representation to provide interpretability for the functional partitioning of the genome.

[0168] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A gene function diagnosis system based on twin networks and multi-order tensor information, characterized by: include: The data management module is responsible for storing and managing genomic data. The genomic data is stored in a database and provides data storage and query functions for the system; The data import module is used to load gene sequencing data, pre-process the data in the data management module through statistical analysis and semantic similarity analysis, and then generate a high-quality gene function feature matrix based on the biological background; The sample similarity module uses an improved similarity calculation method to generate corresponding sample similarity information for high-quality gene function feature matrices. The improved similarity calculation method uses multiple sample point metrics to replace traditional metrics, learns the intrinsic spatial expression between sample points, and uses tensor product technology to achieve lossless fusion of sample representation and sample similarity information. The sample similarity module includes a similarity tensor construction process and a high-order similarity tensor fusion process. For the high-order similarity tensor fusion process, there are: The key lies in effectively utilizing the inherent high-order correlation interaction features between samples to guide the generation of representations, and using the tensor mode-k product to losslessly integrate the inherent high-order information of genomic data. The mode-k product is defined as follows: Where, It is the defined mode-k product symbol, k is a variable parameter that can set a specific value. is a tensor of order m, represents the field of real numbers, Representing a tensor Corresponding to subscript i1...i k-1 i k i k+1 ...i m The specific element values, n1, n2, ..., n m Represents the specific value of the m-th tensor dimension, i m Represents the sub-sequence number in the m-th tensor dimension, 0≤i m ≤n m , n k is the specific value of the k-th tensor dimension, i k Represents the sub-sequence number in the k-th tensor dimension, 0≤i k ≤n k , V jik Indicates the subscript ji corresponding to the two-dimensional matrix V k The specific element value of The purpose of fusing tensor similarity is to guide the generation of deep gene representation through the intrinsic spatial information between multiple samples, so the mode-k product is used to losslessly fuse the similarity tensor. Intrinsic spatial correlation with the deep gene representation F: Where, represents the m-order similarity tensor, is the mode-k product defined above, Set the variable parameter k of mode-k to a tensor The order m, It is the deep representation of the high-quality gene function feature matrix, and its corresponding column j is F :,j ; The functional diagnosis training module uses an autoencoder to convert the high-quality gene function feature matrix generated in the data import module into the corresponding deep gene representation. Based on the deep gene representation, the improved twin network is trained in combination with the sample similarity module. The improved twin network calculates the spatial distribution of the deep gene representation based on the multiple sample point measurement criteria designed by the sample similarity module, introduces high-order associations between binary, ternary and quaternary data groups, and uses tensor product technology to achieve lossless fusion of deep gene representation and sample similarity information. Finally, the sample similarity module is combined with the gene function diagnosis task to form a unified learning framework, and finally a trained improved twin network, i.e., the optimal diagnostic model, is obtained. The gene function diagnosis and prediction module performs intelligent diagnosis of gene function based on the optimal diagnostic model; among them, the gene function diagnosis and prediction module divides gene function into three key categories: molecular function, cellular component and biological process, combines with the external gene ontology annotation database and provides explanations of the diagnostic results.

2. The gene function diagnosis system based on twin networks and multi-order tensor information according to claim 1, characterized in that: In the data management module, the genomic data is stored in a database built using SQLite3. SQLite3 is a lightweight, embedded relational database management system. Python provides a built-in SQLite3 module that is responsible for establishing an index structure to speed up the retrieval of genomic data and provide efficient query functions, including: a. Index structure design: To achieve efficient genomic data retrieval, the specific index structure selected depends on the characteristics of the data, query requirements, and performance requirements; b. Gene name query: supports querying based on gene name. It can provide the common name or identifier of the gene, quickly locate the corresponding genome data based on the index structure, and return relevant data fragments or complete genome data; c. Genome location query: Supports query of genome coordinates for the location information of genome data. By specifying the chromosome number, starting position, and ending position, the corresponding genome data will be quickly retrieved and returned using the index structure. d. Multi-condition query: Supports multi-condition query and can perform combined query based on key information, including gene name, genome location and other annotation information. Based on the index structure and query optimization algorithm, it will efficiently filter out genomic data that meets the conditions; e. Data fragment return: When performing genome data retrieval, data fragments of the specified area can be returned; f. High-performance query optimization: To further improve the performance of genomic data retrieval, query optimization techniques can be used; g. Query result sorting and filtering: When query results are returned, they can be sorted and filtered as needed.

3. The gene function diagnosis system based on twin networks and multi-order tensor information according to claim 2, characterized in that: The data import module includes a data loading module and a data preprocessing module, wherein: The data loading module reads gene sequencing data from the local machine and supports reading data from a variety of common genomic data formats. By supporting different formats, it can easily import various types of genomic data. During the loading process, the data loading module parses and verifies the data in the data management module to ensure the integrity and accuracy of the data. For large-scale genomic data in the data management module, the data loading module can support data partitioning and sharding storage. The data preprocessing module is used to construct a gene function set and generate a high-quality gene function feature matrix. The gene ontology is a classification system that provides a standardized description of gene and protein functions. It divides gene functions into three levels of components, including molecular functions, cellular components, and biological processes. Each component has a series of terms and annotations used to describe the participation of genes in specific functions or processes. Based on the hierarchical structure and relationships between terms in the gene ontology, the functional similarity between genes is measured by calculating the similarity scores between gene terms in gene annotations. In statistical analysis, the use of semantic similarity methods can be used to reduce the importance of statistical p-values, reduce false positive problems, and improve the reliability of results.

4. The gene function diagnosis system based on twin networks and multi-order tensor information according to claim 3 is characterized in that: For the similarity tensor construction process, there are: a1. Second-order similarity tensor: Select Euclidean distance as the metric and judge the similarity of samples by the distance between sample pairs. Assuming that there is samples, select sample pair x i 、x j , their distance metric d in Euclidean space ij Expressed as: d ij =||x i -x j ||2 Where, d ij Represents the sample pair x i 、x j The distance in Euclidean space, ||·||2 represents the matrix norm, so it contains the sample point x i 、x j The second-order similarity tensor of Corresponding elements for: Where exp is the natural exponential function and σ is a scaling constant; a2. Third-order similarity tensor: Considering that high-order similarity metrics help to expand limited sample information, this is achieved by integrating information between multiple samples to achieve full interaction of sample information. Therefore, based on the second-order similarity tensor metric, a similarity metric criterion for triplet data sets is introduced, including sample point x i 、x j 、x k The third-order similarity tensor of Corresponding elements for: Where, d jk Represents the sample pair x j 、x k The distance in Euclidean space, for each index i, j, is the total number of samples, elements Indicates the sample x at the anchor point j Under the condition that the sample x i and sample x k similarities between; a3. Further capture the relationship between the four-dimensional data group and introduce the similarity tensor of the four-dimensional data group, including the sample point x i 、x j 、x k 、x l The fourth-order similarity tensor of Corresponding elements for: Where, d kl Represents the sample pair x k 、x l The distance in Euclidean space, d ik Represents the sample pair x i 、x k The distance in Euclidean space, d jl Represents the sample pair x j 、x l The distance in Euclidean space, ε is a very small number, Represents the fourth-order similarity tensor, measuring the sample x i 、x j 、x k 、x l The intrinsic correlation information between the four data groups.

5. The gene function diagnosis system based on twin networks and multi-order tensor information according to claim 4 is characterized in that: The functional diagnosis training module includes a feature representation module, a high-order similarity fusion module, and a twin network training module that fuses multi-order similarity tensors, wherein: The feature representation module is used for feature representation of gene sample information. It uses an autoencoder to convert the high-quality gene function feature matrix into a corresponding deep vector form. Specifically, the encoder f(·) is used to encode the high-quality gene function feature matrix: F=f(X) Where, It is a high-quality gene function feature matrix generated in the data import module; the encoder consists of a 4-layer fully connected neural network, the first layer is the input layer of the original feature, and the number of nodes is the second dimension of the original feature The number of nodes in the second, third, and fourth layers are 500, 500, and 2000 respectively. The node dimension corresponding to the fourth layer corresponds exactly to the second dimension of the depth representation F. The activation function between each layer is ReLU, and a normalization layer is added after the last layer of activation function; The high-order similarity fusion module uses a high-order similarity metric to perform a progressive sample relationship space on the deep representation of high-quality gene function feature matrices, including: a、For the sample point x i 、x j The second-order similarity tensor of The fusion method is: Where, It is the mode-k product, where the variable parameter k is 1 or 2; b. For the sample point x i 、x j 、x k The third-order similarity tensor of According to the third-order similarity tensor construction formula, the corresponding similarity fusion method is: Where, is the mode-k product, where the variable parameter k is 1, 2, or 3; c. For the sample point x i 、x j 、x k 、x l The fourth-order similarity tensor of According to the fourth-order similarity tensor construction formula, the corresponding similarity fusion method is: Where, is the mode-k product, where the variable parameter k is 1, 2, 3, or 4; The twin network training module integrating multi-order similarity tensors utilizes the sample similarity module to improve the twin network and combines the cross entropy loss function training to obtain the optimal diagnostic model, including: a. Construct the target variable T of the sample pair required by the twin network based on the genome label ij : In the formula, the subscript ij indicates that the target variable measures the gene sample x i 、x j ; b. According to the encoding process designed above, the high-quality gene function feature matrix X is mapped into the deep representation space: F=f(X) c. Combining the second-order, third-order, and fourth-order similarity tensor formulas designed by the sample similarity module with the deep representation F of the high-quality gene function feature matrix to construct the corresponding second-order, third-order, and fourth-order similarity tensors d. Based on the designed tensor fusion method, the similarity tensors of different orders are losslessly fused with the deep representation F of the high-quality gene function feature matrix using the mode-k product: e. Improve the twin network based on the multi-order tensor similarity fusion method designed in step d, and design the loss function in combination with the gene function diagnosis task Perform representation learning: Where, is the cross entropy loss function, y iu Represents sample x i Whether it belongs to category u; is the sample x i Predictions belonging to category u; T ij It is the sample pair target variable constructed by the genome label. The subscript ij indicates that the target variable measures the gene sample x i 、x j , α, β, s are hyper parameters, is the number of samples, e is the natural logarithm, c represents the number of categories, F ij and It is the deep representation of high-quality gene function feature matrix F corresponding to subscript ij, The specific element value of f. Train the improved twin network to obtain the optimal diagnostic model, that is, the trained improved twin network.

6. The gene function diagnosis system based on twin networks and multi-order tensor information according to claim 5, characterized in that: The gene function diagnosis and prediction module is used to perform intelligent diagnosis on a high-quality gene function feature matrix and provide an explanation of the diagnosis results, including the following steps: 1) using the data import module to load the gene data set, and processing the gene data according to the data preprocessing module to obtain a high-quality gene function feature matrix; 2) Inputting the high-quality gene function feature matrix obtained after processing in step 1) into the functional diagnosis training module for training, and training through the feature representation module, high-order similarity fusion module, and twin network training module that integrates multi-order similarity tensors to obtain the optimal diagnostic model; 3) Extracting a deep representation of the high-quality gene functional feature matrix in the optimal diagnostic model in step 2) above, and then performing spatial analysis on the gene representation to provide interpretability for the functional partitioning of the genome.

Images