A genomic-based prognosis prediction method and system for cervical cancer

Abstract

The present application relates to the technical field of biomedical engineering, in particular to a cervical cancer prognosis prediction method and system based on genomics. First, the transcription vector is obtained, and the element mean of each transcription vector is obtained; further, path search is performed in the preset network graph, and the graph topology interference matrix is obtained; further, the Gibbs kernel matrix is constructed, and the alternative scaling iteration is solved to obtain the feature transmission matrix; further, based on the mapping interference relationship of the feature transmission matrix and the graph topology interference matrix, combined with the transcription calibration coefficient and the element mean of each transcription vector, the immune interference difference value is obtained; finally, the spatial feature vector is weighted and mapped based on the feature regression weight, and the continuous offset index is obtained, which provides an objective index support for the prognosis prediction of cervical cancer independent of external dynamic queue, and effectively improves the accuracy and reliability of single sample prognosis prediction based on genomics.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering technology, specifically to a genomics-based method and system for predicting the prognosis of cervical cancer. Background Technology

[0002] High-throughput whole transcriptome sequencing data has wide applications in bioinformatics models and network feature analysis. Existing data processing pipelines typically collect the sequencing expression matrix of samples, extract numerical features using dimensionality reduction models, and then output genotyping results or evaluation indices.

[0003] In real-world microenvironment networks, biological pathways such as basic energy supply, cell proliferation, and immune defense exhibit strong connectivity and interference at the transcriptional expression level. However, existing data processing methods, when performing feature dimensionality reduction, generally treat all extracted genes as independent mathematical variables, completely ignoring the interference resistance caused by the topological distance between underlying network nodes, leading to severe distortion in feature weight allocation. Furthermore, the natural transcriptional abundance of different gene pathways varies significantly in scale, and unaligned multidimensional features are easily dominated by large numerical dimensions when performing spatial distance matching. Moreover, existing models typically rely on dynamic queue data across batches for online feature alignment, causing data distribution drift when the model is separated from the external queue and only receives independent single-sample input for inference, resulting in the inability to output stable classification labels. Additionally, when performing evaluation and prediction, existing methods fail to effectively integrate the complex network interference costs into the feature regression system, causing the output continuous prediction index to deviate from the true antagonistic state of the microenvironment, resulting in low accuracy and reliability of single-sample prognostic predictions. Summary of the Invention

[0004] To address the technical problem of inaccurate prediction results when using single-sample transcriptome data for prognostic prediction due to neglecting network topological interference of gene pathways and the high susceptibility to feature distribution drift, the present invention aims to provide a genomics-based method and system for cervical cancer prognostic prediction. The specific technical solution adopted is as follows:

[0005] A genomics-based method for predicting the prognosis of cervical cancer, the method comprising:

[0006] Retrieve the preset transcription calibration coefficients and feature regression weights; compare the whole transcriptome sequencing vectors to be analyzed with the preset static dictionary to obtain the transcription vectors for energy supply, proliferation and immunity, and obtain the mean value of elements in each transcription vector;

[0007] Gene nodes from the preset static dictionary are retrieved, and path search is performed in the preset network graph. A graph topological interference matrix is ​​obtained by combining all the transcription vectors. Based on the numerical distribution within all transcription vectors, an energy supply feature distribution vector and a consumption demand distribution vector are constructed. A Gibbs kernel matrix is ​​constructed based on the energy supply feature distribution vector, the consumption demand distribution vector, and the graph topological interference matrix, and iterative scaling is performed to obtain a feature transfer matrix. Based on the mapping interference relationship between the feature transfer matrix and the graph topological interference matrix, and combined with the transcription calibration coefficient and the element mean of each transcription vector, an immune interference difference is obtained.

[0008] A spatial feature vector is constructed based on the element mean of each transcription vector and the difference in immune interference; the spatial feature vector is weighted and mapped based on the feature regression weights to obtain a continuous offset index.

[0009] Furthermore, the method for obtaining the spectral topological interference matrix includes:

[0010] Determine the topological connectivity paths from each energy-supplying gene node to a proliferation gene node or an immune gene node within the preset static dictionary; obtain the minimum number of network nodes connected by the topological connectivity path as a resistance constant, and construct the graph topological interference matrix based on the resistance constant; the number of rows in the graph topological interference matrix is ​​the total number of terms in the energy-supplying transcription vectors; the number of columns is the total number of terms in the proliferation and immune transcription vectors; and the element values ​​in the matrix are the corresponding resistance constants.

[0011] If there is no topologically connected path between any two gene nodes, then the number of connected edges between the two gene nodes is set to a preset resistance constant.

[0012] Furthermore, the method for obtaining the feature transfer matrix includes:

[0013] The mean of the matrix elements of the topological interference matrix is ​​extracted, and a smoothing calculation coefficient is constructed by combining a preset regularization scaling factor and a preset division-by-zero positive parameter; the smoothing calculation coefficient is used to perform exponential mapping on each element of the topological interference matrix to obtain the Gibbs kernel matrix.

[0014] Initialize a column vector with the same length as the energy-generating transcription vector and all internal elements being 1, as the initial left scaling vector; initialize a column vector with the same length as the combined length of the proliferation transcription vector and the immune transcription vector and all internal elements being 1, as the initial right scaling vector.

[0015] Based on the Gibbs kernel matrix, the energy supply feature distribution vector, and the consumption demand distribution vector, an alternating scaling matrix iterative calculation is performed on the left scaling vector and the right scaling vector. After the iteration reaches the preset convergence condition, the left scaling vector, the right scaling vector, and the Gibbs kernel matrix output from the last iteration are multiplied together to construct the feature transfer matrix.

[0016] Furthermore, the method for obtaining the immune interference difference includes:

[0017] Extract a subset of elements corresponding to the immune transcription vector in the feature transfer matrix, multiply and sum the subset of elements with the matrix elements at the corresponding positions in the map topological interference matrix, and normalize the calculation to obtain the immune pathway interference weight coefficients.

[0018] The interference consumption is obtained by fusing the transcription calibration coefficient, the mean element of the power transcription vector, and the interference weight coefficient of the immune pathway; the difference between the mean element of the immune transcription vector and the interference consumption is truncated to a negative value and used as the immune interference difference.

[0019] Furthermore, the method for obtaining the spatial feature vector includes:

[0020] Following the order of the mean element of the energy-supplying transcription vector first, the mean element of the proliferation transcription vector in the middle, and the immune interference difference last, the three element values ​​are concatenated to form an initial multidimensional feature vector.

[0021] The initial multidimensional feature vector is subjected to data standardization processing to obtain the spatial feature vector.

[0022] Furthermore, the method for obtaining the continuous offset index includes:

[0023] Extract the preset regression coefficients corresponding to the energy supply dimension, proliferation dimension, and immune intervention dimension from the feature regression weights; use each preset regression coefficient to perform a weighted summation of the feature values ​​of the corresponding dimension in the spatial feature vector to obtain the overall offset scalar; apply the natural exponential function to the overall offset scalar for activation mapping, and use the result obtained after mapping as the continuous offset index.

[0024] Furthermore, the method for obtaining the energy supply characteristic distribution vector includes:

[0025] Logarithmic smoothing is performed on the numerical value of each element in the energy transcription vector, and the softmax function is applied to all elements in the smoothed vector to perform probability share mapping, thereby obtaining the energy feature distribution vector.

[0026] Furthermore, the method for obtaining the consumption demand distribution vector includes:

[0027] The proliferation transcription vector is placed first, followed by the immune transcription vector, and then a row-by-row concatenation operation is performed to generate a comprehensive consumption column vector. The values ​​in the comprehensive consumption column vector are logarithmically smoothed, and the softmax function is applied to all elements in the smoothed vector to perform probability share mapping, thereby obtaining the consumption demand distribution vector.

[0028] Furthermore, the preset static dictionary includes: a basic energy supply identifier sequence, a tumor proliferation identifier sequence, and an immune defense identifier sequence.

[0029] The present invention also proposes a genomics-based cervical cancer prognosis prediction system, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the steps of the genomics-based cervical cancer prognosis prediction method described above.

[0030] The present invention has the following beneficial effects:

[0031] This invention first obtains transcription vectors and the element-mean of each transcription vector as the original transcriptional numerical baseline for subsequent feature fusion operations. Next, path search is performed in a pre-defined network graph to obtain a graph topological interference matrix, quantifying the degree of network topological antagonism occurring at the transcriptional expression level in different biological pathways. Then, a Gibbs kernel matrix is ​​constructed and iteratively scaled to obtain a feature transfer matrix, analyzing the optimal network mapping relationship for the transfer from metabolic energy supply nodes to resource consumption channels under the constraint of network topological interference costs. Based on the mapping interference relationship between the feature transfer matrix and the graph topological interference matrix, combined with the transcriptional calibration coefficient and the element-mean of each transcription vector, an immune interference difference is obtained, providing a basis for constructing a highly stable single-sample prognostic prediction. Finally, a weighted mapping of spatial feature vectors is performed based on feature regression weights to obtain a continuous offset index, providing objective indicator support for cervical cancer prognostic prediction independent of external dynamic cohorts, effectively improving the accuracy and reliability of genomics-based single-sample prognostic prediction.

[0032] This invention constructs a topological interference matrix based on transcription vectors and iteratively solves it. It extracts the immune interference difference by combining transcription calibration coefficients, constructs a spatial feature vector with the element mean, and then weights and maps it to obtain a continuous offset index. This effectively integrates network interference costs and avoids feature drift, thereby improving the accuracy of single-sample prognosis prediction. Attached Figure Description

[0033] Figure 1 A flowchart illustrating a genomics-based cervical cancer prognosis prediction method provided in one embodiment of the present invention;

[0034] Figure 2This is a flowchart illustrating the acquisition of a feature transmission matrix according to an embodiment of the present invention. Detailed Implementation

[0035] The following description, in conjunction with the accompanying drawings, details the specific scheme of the genomics-based cervical cancer prognosis prediction method and system provided by this invention.

[0036] Please see Figure 1 The diagram illustrates a flowchart of a genomics-based cervical cancer prognosis prediction method according to an embodiment of the present invention, specifically including:

[0037] Step S1: Retrieve the preset transcription calibration coefficients and feature regression weights; compare the whole transcriptome sequencing vectors to be analyzed with the preset static dictionary to obtain the transcription vectors for energy supply, proliferation and immunity, and obtain the mean value of elements in each transcription vector.

[0038] Considering the natural differences in expression abundance among gene pathways and the large amount of redundant noise in massive whole transcriptome sequencing data, in order to facilitate rapid response to single sample input during the online inference stage and avoid repeated offline training calculations, the preset transcription calibration coefficients and feature regression weights are first retrieved, and a static dictionary is determined. Then, the whole transcriptome sequencing vectors to be analyzed are compared with the preset static dictionary to obtain transcription vectors for energy supply, proliferation, and immunity. A subset of vectors representing key biological pathways are accurately extracted, and the mean value of elements within each transcription vector is obtained as the original transcriptional numerical baseline for subsequent feature fusion operations.

[0039] As a preferred approach, the system calls the public gene annotation database set stored in the local storage module, including the pre-determined basic energy supply marker sequences, tumor proliferation marker sequences, and immune defense marker sequences, and uses these three sets of sequences as a preset static dictionary; then it retrieves the whole transcriptome sequencing vector to be analyzed.

[0040] Because genes exhibit strong pleiotropic effects, in real-world public gene annotation databases (such as KEGG and GO), the same core key gene (e.g., mTOR or AMPK) often participates in both basal energy metabolism and directly drives tumor proliferation. Since the three sets of dictionaries naturally overlap, when comparing with a pre-defined static dictionary to obtain transcription vectors for energy supply, proliferation, and immunity, a mutual exclusion check is performed on the whole transcriptome sequencing vectors to be analyzed, based on the priority order set for basal energy marker sequences, tumor proliferation marker sequences, and immune defense marker sequences. This ensures that any gene marker in the whole transcriptome sequencing vector belongs to only one corresponding transcription vector.

[0041] The system extracts energy-providing, proliferation-proliferating, and immune-providing transcriptional vectors from the whole transcriptome sequencing vectors to be analyzed. These are the energy-providing transcriptional vector, proliferation-proliferating transcriptional vector, and immune-providing transcriptional vector, respectively. Then, the mean value of each element within each vector is calculated, providing a basis for subsequent analysis. If a vector extraction result is empty, its mean value is directly assigned to 0.

[0042] The order of energy supply over proliferation over immunity is not a ranking of biological importance, but rather an engineering consideration based on algorithm stability. Energy supply has the highest priority to ensure the row-side distribution density of the Sinkhorn algorithm, providing ample optimal transport solution space for the interference matrix. Proliferation has the second highest priority as a noise buffer to isolate the interference of uncertain variables on the final immunity index. Immunization is the last priority to ensure that the final extracted immune interference difference has high specificity and purity, avoiding feature dilution and spurious interference bias caused by pleiotropic universal genes.

[0043] Before single-sample online inference, the system pre-acquires and solidifies static baseline parameters through an offline large-sample queue (such as the publicly available TCGA database). The specific steps for obtaining the preset transcription calibration coefficients and feature regression weights are as follows:

[0044] Whole transcriptome sequencing data of all samples in the large sample cohort were extracted. Combined with a pre-set static dictionary, the total mean of immune defense transcription and the total mean of basal energy transcription in the large sample cohort were calculated respectively. The total mean of immune defense transcription was divided by the total mean of basal energy transcription, and the resulting ratio was fixed as a transcription calibration coefficient, which was used to eliminate the dimensional differences in the natural expression abundance of different gene pathways.

[0045] Following the same dimensionality reduction logic as single-sample real-time processing, the mean baseline energy-providing transcriptional value, the mean tumor proliferation transcriptional value, and the difference in immune interference for each sample in the large sample cohort are pre-calculated to construct an offline three-dimensional feature set. Known phenotypic labels corresponding to the large sample cohort are extracted, and conventional regression training algorithms (such as Cox proportional hazards regression model or multiple linear regression model) are used to fit and train the offline three-dimensional feature set to the phenotypic labels. After training convergence, the numerical weights corresponding to each feature dimension are extracted and used as preset regression coefficients for the energy-providing dimension, proliferation dimension, and immune interference dimension, respectively, and solidified as feature regression weights.

[0046] The aforementioned pre-configuration process utilizes existing statistical and machine learning models to generate and solidify static constants offline before shipment. This mechanism eliminates the need for external queues for dynamic batch alignment during the online inference phase, completely preventing the risk of feature drift caused by single-sample input.

[0047] Step S2: Retrieve gene nodes from the preset static dictionary, perform path search in the preset network graph, and obtain the graph topological interference matrix by combining all transcription vectors; construct energy supply feature distribution vector and consumption demand distribution vector based on the numerical distribution of all transcription vectors; construct Gibbs kernel matrix based on energy supply feature distribution vector, consumption demand distribution vector and graph topological interference matrix, and perform alternating scaling iterative solution to obtain feature transfer matrix; obtain immune interference difference based on the mapping interference relationship between feature transfer matrix and graph topological interference matrix, combined with transcription calibration coefficient and the element mean of each transcription vector.

[0048] Since proteins translated from different genes often need to pass through intermediate nodes in biochemical reaction networks, the topological distance between these nodes objectively constitutes an obstacle to signal transduction or substance transformation. The system calls pre-installed standard biological interaction network graph data (such as known public databases like the PPI protein interaction network) as a preset network graph, and extracts the undirected connection edge data between all network nodes in the preset network graph.

[0049] Then, the gene nodes in the preset static dictionary are retrieved, and path search is performed in the preset network graph. The graph topology interference matrix is ​​obtained by combining all transcription vectors, and the degree of network topology antagonism that occurs at the transcriptional expression level of different biological pathways is quantified, breaking the feature extraction limitation of conventional models that treat genes as isolated mathematical variables.

[0050] Preferably, in one embodiment of the present invention, a conventional shortest path search algorithm (such as Dijkstra's algorithm or Breadth-First Search algorithm) is first started. Starting from each basic energy-supplying gene, the path between it and each tumor proliferation gene or immune defense gene is searched sequentially. The minimum number of network nodes connected by the search path is counted, and the topological connected path from each energy-supplying gene node to the proliferation gene node or immune gene node in the preset static dictionary is determined. The minimum number of network nodes connected by the topological connected path is obtained as the resistance constant. The larger the resistance constant, the greater the topological distance between the two genes.

[0051] Then, a topological interference matrix is ​​constructed based on the drag constant:

[0052] In this example, the number of rows in the topological interference matrix is ​​the total number of terms in the energy-supplying transcription vectors; the genes for proliferation and immunity are taken as the target energy-supplying genes, collectively referred to as the comprehensive energy-supplying genes, the number of columns is the total number of terms in the proliferation and immunity transcription vectors, and the element values ​​in the matrix are the resistance constants between the corresponding energy-supplying genes and the comprehensive energy-supplying genes.

[0053] If there is no topologically connected path between any two gene nodes, the number of connected edges between the two gene nodes is set to a preset resistance constant. In this example, the preset resistance constant is 100, which is used as the manually set limit transformation resistance constant.

[0054] The topological interference matrix quantifies the absolute topological interference distance of different biochemical pathways in cyberspace, analyzes the real physical obstacles to signal transduction between genes, and achieves a complete construction of the global interference cost while ensuring that the algorithm system does not cause deadlock due to isolated disconnections.

[0055] Since absolute transcription values ​​cannot directly reflect the energy supply ratio and consumption allocation preference within the same microenvironment system, and since the original transcription values ​​are often in the hundreds or thousands, directly inputting them into the exponential normalization function will cause the computer's underlying floating-point number to overflow, causing the weight of the largest value to approach 1 while the weight of the rest collapses to 0. Therefore, based on the numerical distribution within all transcription vectors, an energy supply characteristic distribution vector and a consumption demand distribution vector are constructed to characterize the relative weight quotas of the overall metabolic supply capacity and comprehensive consumption demand under the same probability dimension, and a pre-computational boundary is established to satisfy the supply and demand distribution conservation constraint of the optimal transport algorithm.

[0056] Existing dimensionality reduction schemes cannot handle complex network weight calculations with graph connectivity interference and supply and demand distribution conservation constraints, which can easily lead to feature drift and distortion. This system introduces an alternating scaling (Sinkhorn) iterative solution mechanism and performs cross-channel feature fusion calculation after the solution is completed, combined with a preset transcription calibration coefficient, to extract a network interference difference index with a single directionality, thereby cutting off the dependence of single-sample online inference on real-time batch alignment of large sample queues.

[0057] Therefore, a Gibbs kernel matrix is ​​constructed based on the energy supply feature distribution vector, the consumption demand distribution vector, and the graph topology interference matrix. Alternating scaling and iterative calculation are performed to obtain the feature transfer matrix, and the optimal network mapping relationship for the transfer from metabolic energy supply nodes to resource consumption channels under the constraint of network topology interference cost is analyzed.

[0058] Furthermore, considering that the network mapping transmission features calculated by simply solving are dimensionless proportional data, direct algebraic operations will cause dimension failure. Moreover, if the scale is not aligned with the absolute level of transcription of the sample itself, it will not be able to reflect the absolute resource shortage of biological pathways under real antagonistic conditions. Therefore, based on the mapping interference relationship between the feature transmission matrix and the graph topology interference matrix, combined with the transcription calibration coefficient and the mean of each transcription vector, the immune interference difference is obtained. This quantifies the degree of real transcriptional support deficiency caused by tumor network topology obstruction in the microenvironment resource competition of the immune defense system, providing a basis for constructing a highly stable single-sample prognostic prediction.

[0059] Preferably, in one embodiment of the present invention, in order to prevent data overflow and distribution collapse anomalies, logarithmic smoothing is introduced before normalization, and logarithmic smoothing is performed on the value of each element in the energy transcription vector.

[0060] In the field of gene sequencing, the numerical value of an element corresponds to the number of transcribed copies, representing the number of detected RNA fragments. Its physical meaning dictates that it is inherently a non-negative number greater than or equal to 0, and cannot be negative; the logarithmic function is... The element values ​​are substituted into the function as independent variables x for logarithmic smoothing. The constant 1 is used to ensure that the mapping result is positive, and the logarithm performs logarithmic smoothing compression on the original data.

[0061] Apply the softmax function to all elements in the smoothed vector to perform probability share mapping, obtain the energy supply feature distribution vector, and output an energy supply feature distribution vector with the same length as the energy supply transcription vector and a sum of normalization results of 1.

[0062] Furthermore, the two vectors corresponding to the consumption demand are merged. Specifically, the proliferation transcription vector is placed first, and the immune transcription vector is placed second. A row-by-row concatenation operation is performed to generate a comprehensive consumption column vector.

[0063] Logarithmically smooth the values ​​in the overall consumption column vector, and apply the softmax function to all elements in the smoothed vector to perform probability share mapping, thereby obtaining the consumption demand distribution vector.

[0064] The logarithmic smoothing and softmax function mapping here are the same as the process of obtaining the energy supply feature distribution vector, and will not be described again.

[0065] Subsequently, the feature transfer matrix is ​​constructed; please refer to [link / reference]. Figure 2 It illustrates a flowchart of a feature transfer matrix acquisition method provided by an embodiment of the present invention, specifically including:

[0066] Step S201: Extract the mean of the matrix elements of the topological interference matrix, and construct smoothing calculation coefficients by combining the preset regularization scaling factor and the preset division by zero positive parameter; use the smoothing calculation coefficients to perform exponential mapping on each element of the topological interference matrix to obtain the Gibbs kernel matrix.

[0067] Considering that directly using absolute physical interference resistance during matrix iteration in the optimal transmission algorithm can easily lead to floating-point underflow or division-by-zero crashes in exponential function calculations, it is necessary to smoothly transform absolute connectivity resistance into probabilistic transfer costs. Therefore, the mean of the matrix elements of the graph topology interference matrix is ​​extracted, and a smoothing calculation coefficient is constructed by combining a preset regularization scaling factor and a preset division-by-zero positive parameter. The smoothing calculation coefficient is then used to perform exponential mapping on each element of the graph topology interference matrix to obtain the Gibbs kernel matrix. This quantifies the smooth connectivity transfer probability between different gene nodes in the network space, ensuring the safety, stability, and convergence continuity of the underlying matrix iteration calculation.

[0068] As an example, with a preset regularization scaling factor of 0.1 and a preset division-by-zero positive parameter of 0.0001, the system calculates the mean of all matrix elements of the topological interference matrix, multiplies it by 0.1, and then adds 0.0001. The result is used as the smoothing coefficient. ;

[0069] Here, 10% (i.e., 0.1) of the matrix average cost is taken as... , is an adaptive empirical value commonly used in algorithm engineering implementation. Its purpose is to keep the input scale of the exponential function in a safe numerical range (usually between -10 and 0), ensuring that the calculation does not crash and has discriminative power. When set to 0.1, the exponential terms at the bottom layer of the kernel matrix can be amplified to a relative distance ratio of -10.

[0070] Then the elements of the topological interferometry matrix were... Substitute into function In this process, the element values ​​are used as independent variables x, and the functions are substituted to perform exponential mapping to obtain the Gibbs kernel matrix. For the natural constant exponential function, this operation will Transforming the original network into a positive number between 0 and 1, the connectivity resistance of the original network... The larger the value, the more likely the corresponding Gibbs kernel matrix elements will be generated. The closer the value is to zero, the better. i is the row number and j is the column number.

[0071] Step S202: Initialize a column vector with the same length as the energy-generating transcription vector and all internal elements being 1, as the initial left scaling vector; initialize a column vector with the same length as the combined length of the proliferation transcription vector and the immune transcription vector and all internal elements being 1, as the initial right scaling vector.

[0072] Since the iterative calculation of the alternating scaling matrix requires a pre-defined baseline scale variable to accommodate the probability scaling ratios of the supply and demand distributions in each cycle update, and the absolute neutrality of the iteration starting point must be ensured to prevent the introduction of artificial prior bias, a column vector with the same length as the energy transcription vector and all internal elements being 1 is initialized as the initial left scaling vector. , length is And initialize a column vector with a length equal to the total length of the proliferation transcription vector and the immune transcription vector, and with all internal elements being 1, as the initial right scaling vector. , length is An iterative starting reference container was established to satisfy the dimensional alignment of the supply and demand vector space, ensuring that the alternating scaling process starts its evolution from a completely uniform weight state.

[0073] Step S203: Based on the Gibbs kernel matrix, the energy supply feature distribution vector, and the consumption demand distribution vector, perform alternating scaling matrix iterative calculation on the left scaling vector and the right scaling vector; after the iteration reaches the preset convergence condition, perform a series multiplication operation using the left scaling vector, the right scaling vector, and the Gibbs kernel matrix output from the last iteration to construct the feature transfer matrix.

[0074] The iteration needs to find the mapping relationship that minimizes the overall network interference cost of the microenvironment under the dual edge constraints of supply-side distributed output quota conservation and consumption-side distributed receiving demand conservation. Therefore, alternating scaling matrix iterative calculation is performed, and after the iteration reaches the preset convergence condition, the characteristic transmission matrix is ​​constructed. The optimal flow allocation path for the power supply node to transfer to each consumption channel under the overall graph network interference constraint is analyzed, which represents the network game equilibrium state that minimizes the overall information and resource transmission loss within the microenvironment.

[0075] As an example, in the s-th loop, the system performs left scaling vector update and right scaling vector update calculations, specifically:

[0076] The system combines the Gibbs kernel matrix K with the right scaling vector Performing matrix multiplication yields a matrix of length . The system then performs element-wise division of the intermediate column vector: dividing each element of the energy feature distribution vector P by the value at its inner row position in this intermediate column vector; during this process, to prevent division by zero due to zero values ​​in the intermediate column vector, the system forcibly appends a preset minimum positive number 0.0001 to each denominator element of the intermediate column vector. After completing all element-wise divisions, the system generates the updated (s+1)th left scaling vector. ;

[0077] System extracts Gibbs kernel matrix transpose matrix ,Will With the newly obtained left scaling vector Performing matrix multiplication yields a matrix of length . The system then divides each element in the demand distribution vector Q by the value of its inner row position within this intermediate column vector. During this process, to prevent division by zero values ​​within the intermediate row vector, the system forcibly appends a preset minimum positive number 0.0001 to each denominator element of the intermediate row vector. After all elements are divided, the system generates the updated (s+1)th right scaling vector. ;

[0078] The preset convergence conditions include: First, the current iteration number reaches the preset maximum iteration number (100 times in this example); Second, the numerical difference between the left scaling vector (or right scaling vector) output by the current iteration and the corresponding scaling vector output by the previous iteration (e.g., the maximum value of the absolute difference of the vector elements) is less than the preset convergence tolerance threshold (e.g., 1e-5).

[0079] If any of the above conditions are met, the system determines that the preset convergence condition has been met, terminates the iteration, and confirms that the matrix has converged.

[0080] After terminating the iteration, the system extracts the left scaling vector from the output of the last loop. With right scaling vector Scale the vector to the left The elements are arranged along the diagonal to form a diagonal matrix. Similarly, the right-scaled vector The elements are constructed into a diagonal matrix. .

[0081] Diagonal matrix in sequence Gibbs kernel matrix With diagonal matrix Perform consecutive triple matrix multiplication operations, and use the result as the characteristic transfer matrix. This matrix shows the relative mapping weight distribution of the transfer from basic metabolic gene nodes to various demand channels under the overall topological interference constraint.

[0082] It should be noted that because the left and right scaling vectors are strongly coupled and updated alternately in the Sinkhorn algorithm, the system only needs to sample the difference in change of either side (e.g., only the left scaling vector) in each loop to determine the overall balance. As an alternative implementation, the system can also employ hard threshold truncation protection, that is, determining whether each element in the middle column vector and middle row vector is below a preset safe lower bound threshold (e.g., ...). If so, assign the element the value of the lower safety threshold (e.g., 0.0001); otherwise, retain the original value.

[0083] In a preferred embodiment of the present invention, considering that the feature transfer matrix only reflects the relative flow quota allocation output between gene nodes after Sinkhorn alternating scaling iteration, its element values ​​only represent the optimal probability quality transfer ratio from each energy-supplying gene node to each energy-consuming gene node under the constraint of supply-side and consumption-side distribution conservation, and do not directly reflect the actual topological distance cost overcome in the network signal transmission process; while the graph topological interference matrix records the network resistance constant corresponding to the number of shortest connected edges between genes in different pathways, its element values ​​are independent of any probability distribution, and purely reflect the physical topological distance in the biological network space.

[0084] Therefore, a subset of elements corresponding to the immune transcription vector in the feature transfer matrix is ​​extracted, and the subset of elements is multiplied and summed with the corresponding matrix elements in the graph topological interference matrix, and then normalized to obtain the immune pathway interference weight coefficient. This quantifies the final loss ratio of the microenvironment system tending to allocate transcription weight to the immune pathway under network topological constraints. Through element-wise multiplication and summation normalization operations, the mapping interference relationship between the feature transfer matrix and the graph topological interference matrix is ​​established.

[0085] Furthermore, considering that the extracted network interference weights are dimensionless proportional coefficients, directly using them for algebraic subtraction would lead to the failure of mathematical scales. Moreover, without combining them with the absolute basic expression levels extracted by sequencing, they would not be able to reflect the true supply and demand imbalance in the microenvironment. In addition, it is necessary to avoid model logic anomalies caused by negative gaps when the supply is sufficient. Therefore, the transcription calibration coefficient, the mean element of the energy-supplying transcription vector, and the immune pathway interference weight coefficient are integrated to obtain the interference consumption. The difference between the mean element of the immune transcription vector and the interference consumption is truncated to a negative value and used as the immune interference difference. This represents the real transcriptional support gap of the immune pathway caused by the topological interference of the tumor network after eliminating the dimensional differences in natural expression abundance. It quantifies the absolute scarcity of the immune defense system in the competition for microenvironment resources.

[0086] As an example, the formula for calculating the weighting coefficient of immune pathway interference includes:

[0087] ;

[0088] In the formula, This represents the weighting coefficient for immune pathway interference. This represents the total number of elements within the energy-providing transcription vector. This represents the total number of elements within the proliferating transcription vector. This represents the total number of elements in the consumption demand distribution vector. Based on the rules for vector concatenation, the features of the multiplication dimension occupy the first part of the feature transfer matrix. The system extracts the first column. List up to number The subset of all elements of the corresponding immune transcription vector in the column is used for summation calculation; This represents the matrix element located in the i-th row and j-th column of the feature transfer matrix. This represents the matrix element located in the i-th row and j-th column of the topological interference matrix. This represents the preset normalization extremum constant, which is the preset resistance constant used when constructing the topological interference matrix of the spectrum. In this example, it is set to 100.

[0089] Furthermore, the product of the transcription calibration coefficient, the mean of the elements of the power transcription vector, and the immune pathway interference weight coefficient is taken as the interference consumption; the maximum value between the difference between the mean of the elements of the immune transcription vector and the interference consumption and the constant 0 is taken as the immune interference difference.

[0090] Step S3: Construct a spatial feature vector based on the element mean of each transcription vector and the difference in immune interference; perform a weighted mapping on the spatial feature vector based on the feature regression weights to obtain the continuous offset index.

[0091] To transform pathway antagonism at the micro-gene network level into quantifiable numerical data features, this step concatenates the element mean of each transcription vector with the immune interference difference into a unified spatial feature vector. This preserves the independent semantic information of the three dimensions: energy supply baseline, proliferation burden, and immune interference loss. Subsequently, this spatial feature vector is weighted and mapped based on pre-trained feature regression weights. This compresses the multidimensional spatial features into a single continuous offset index while taking into account the differences in the contribution of each dimension. This provides objective indicator support for cervical cancer prognosis prediction, independent of external dynamic cohorts, and ultimately effectively improves the accuracy and reliability of single-sample prognosis prediction based on genomics.

[0092] Preferably, in one embodiment of the present invention, the three element values ​​are concatenated into an initial multidimensional feature vector in the order of the mean of the energy-supplying transcription vector first, the mean of the proliferating transcription vector in the middle, and the immune interference difference last.

[0093] The specific order in which the element features are concatenated is only to ensure the alignment of feature dimensions and the consistency of data structure in subsequent algebraic mapping operations. It does not have a substantial impact on the actual physical representation of the micro-environment features. In practical applications, it is only necessary to keep the feature concatenation order strictly consistent between the online inference and offline training stages.

[0094] Considering the huge difference in absolute numerical magnitude between the mean of each transcription and the difference of immune interference, when the feature vectors without scale unification are subsequently weighted, the smaller numerical dimension will be unidirectionally dominated by the larger numerical dimension, causing the multidimensional feature system to fail and leading to serious evaluation bias. Therefore, data standardization is performed on the initial multidimensional feature vectors to obtain spatial feature vectors.

[0095] As an example, the specific process of data standardization includes: retrieving the static mean and standard deviation constants of each dimension that have been pre-trained and solidified by an offline large sample queue; and using the static mean and standard deviation of the corresponding dimension, outputting the Z-score standardized result as the spatial feature vector by taking the element values ​​of each dimension in the initial multidimensional feature vector.

[0096] Finally, the preset regression coefficients corresponding to the energy supply dimension, proliferation dimension, and immune intervention dimension are extracted from the feature regression weights, respectively. , and ;

[0097] By using the preset regression coefficients, the feature values ​​of the corresponding dimensions in the spatial feature vector are weighted and summed to obtain the overall offset scalar.

[0098] The total offset scalar is activated by the natural exponential function (exp(x)), and the result obtained after mapping is used as the continuous offset index, providing objective quantitative data reference for clinicians to adjust targeted drug treatment plans or follow-up examination schedules.

[0099] An embodiment of the present invention also provides a genomics-based cervical cancer prognosis prediction system. The system includes a memory, a processor, and a computer program, wherein the memory is used to store the corresponding computer program, the processor is used to run the corresponding computer program, and the computer program, when running in the processor, can implement the genomics-based cervical cancer prognosis prediction method described in steps S1-S3.

[0100] In summary, to address the technical problem of inaccurate prediction results in prognostic analysis of single-sample transcriptome data due to neglecting network topological interference of gene pathways and the susceptibility to feature distribution drift, this invention provides a genomics-based method and system for cervical cancer prognostic prediction. This invention first obtains transcription vectors and the mean of each element of each transcription vector; then, it performs path search in a pre-set network graph to obtain the graph topological interference matrix; further, it constructs a Gibbs kernel matrix and performs alternating scaling iterative calculation to obtain the feature transfer matrix; further, based on the mapping interference relationship between the feature transfer matrix and the graph topological interference matrix, combined with the transcription calibration coefficient and the mean of each transcription vector, it obtains the immune interference difference; finally, it performs weighted mapping of spatial feature vectors based on feature regression weights to obtain a continuous offset index. This provides objective indicator support for cervical cancer prognostic prediction that is independent of external dynamic cohorts, effectively improving the accuracy and reliability of genomics-based single-sample prognostic prediction.

[0101] This invention constructs a topological interference matrix based on transcription vectors and iteratively solves it. It extracts the immune interference difference by combining transcription calibration coefficients, constructs a spatial feature vector with the element mean, and then weights and maps it to obtain a continuous offset index. This effectively integrates network interference costs and avoids feature drift, thereby improving the accuracy of single-sample prognosis prediction.

Claims

1. A genomics-based method for predicting the prognosis of cervical cancer, characterized in that, The method includes: Retrieve the preset transcription calibration coefficients and feature regression weights; compare the whole transcriptome sequencing vectors to be analyzed with the preset static dictionary to obtain the transcription vectors for energy supply, proliferation and immunity, and obtain the mean value of elements in each transcription vector; Gene nodes from the preset static dictionary are retrieved, and path search is performed in the preset network graph. A graph topological interference matrix is ​​obtained by combining all the transcription vectors. Based on the numerical distribution within all transcription vectors, an energy supply feature distribution vector and a consumption demand distribution vector are constructed. A Gibbs kernel matrix is ​​constructed based on the energy supply feature distribution vector, the consumption demand distribution vector, and the graph topological interference matrix, and iterative scaling is performed to obtain a feature transfer matrix. Based on the mapping interference relationship between the feature transfer matrix and the graph topological interference matrix, and combined with the transcription calibration coefficient and the element mean of each transcription vector, an immune interference difference is obtained. A spatial feature vector is constructed based on the element mean of each transcription vector and the difference in immune interference; the spatial feature vector is weighted and mapped based on the feature regression weights to obtain a continuous offset index.

2. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The method for obtaining the topological interference matrix of the graph includes: Determine the topological connectivity paths from each energy-supplying gene node to a proliferation gene node or an immune gene node within the preset static dictionary; obtain the minimum number of network nodes connected by the topological connectivity path as a resistance constant, and construct the graph topological interference matrix based on the resistance constant; the number of rows in the graph topological interference matrix is ​​the total number of terms in the energy-supplying transcription vectors; the number of columns is the total number of terms in the proliferation and immune transcription vectors; and the element values ​​in the matrix are the corresponding resistance constants. If there is no topologically connected path between any two gene nodes, then the number of connected edges between the two gene nodes is set to a preset resistance constant.

3. The genomics-based cervical cancer prognosis prediction method according to claim 1, characterized in that, The method for obtaining the feature transfer matrix includes: The mean of the matrix elements of the topological interference matrix is ​​extracted, and a smoothing calculation coefficient is constructed by combining a preset regularization scaling factor and a preset division-by-zero positive parameter; the smoothing calculation coefficient is used to perform exponential mapping on each element of the topological interference matrix to obtain the Gibbs kernel matrix. Initialize a column vector with the same length as the energy-generating transcription vector and all internal elements being 1, as the initial left scaling vector; initialize a column vector with the same length as the combined length of the proliferation transcription vector and the immune transcription vector and all internal elements being 1, as the initial right scaling vector. Based on the Gibbs kernel matrix, the energy supply feature distribution vector, and the consumption demand distribution vector, an alternating scaling matrix iterative calculation is performed on the left scaling vector and the right scaling vector. After the iteration reaches the preset convergence condition, the left scaling vector, the right scaling vector, and the Gibbs kernel matrix output from the last iteration are multiplied together to construct the feature transfer matrix.

4. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The method for obtaining the immune interference difference includes: Extract a subset of elements corresponding to the immune transcription vector in the feature transfer matrix, multiply and sum the subset of elements with the matrix elements at the corresponding positions in the map topological interference matrix, and normalize the calculation to obtain the immune pathway interference weight coefficients. The interference consumption is obtained by fusing the transcription calibration coefficient, the mean element of the power transcription vector, and the interference weight coefficient of the immune pathway; the difference between the mean element of the immune transcription vector and the interference consumption is truncated to a negative value and used as the immune interference difference.

5. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The method for obtaining the spatial feature vector includes: Following the order of the mean element of the energy-supplying transcription vector first, the mean element of the proliferation transcription vector in the middle, and the immune interference difference last, the three element values ​​are concatenated to form an initial multidimensional feature vector. The initial multidimensional feature vector is subjected to data standardization processing to obtain the spatial feature vector.

6. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The method for obtaining the continuous offset index includes: Extract the preset regression coefficients corresponding to the energy supply dimension, proliferation dimension, and immune intervention dimension from the feature regression weights; use each preset regression coefficient to perform a weighted summation of the feature values ​​of the corresponding dimension in the spatial feature vector to obtain the overall offset scalar; apply the natural exponential function to the overall offset scalar for activation mapping, and use the result obtained after mapping as the continuous offset index.

7. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The method for obtaining the energy supply characteristic distribution vector includes: Logarithmic smoothing is performed on the numerical value of each element in the energy transcription vector, and the softmax function is applied to all elements in the smoothed vector to perform probability share mapping, thereby obtaining the energy feature distribution vector.

8. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The method for obtaining the consumption demand distribution vector includes: The proliferation transcription vector is placed first, followed by the immune transcription vector, and then a row-by-row concatenation operation is performed to generate a comprehensive consumption column vector. The values ​​in the comprehensive consumption column vector are logarithmically smoothed, and the softmax function is applied to all elements in the smoothed vector to perform probability share mapping, thereby obtaining the consumption demand distribution vector.

9. The genomics-based cervical cancer prognostic prediction method according to claim 1, characterized in that, The preset static dictionary includes: basic energy supply identifier sequence, tumor proliferation identifier sequence, and immune defense identifier sequence.

10. A genomics-based cervical cancer prognosis prediction system, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the genomics-based cervical cancer prognosis prediction method as described in any one of claims 1 to 9.

Images