DNA fragmentation gene detection data processing method based on artificial intelligence

Abstract

The present invention relates to the technical field of DNA fragment data processing, and in particular to a method for processing DNA fragmentation gene detection data based on artificial intelligence. The method divides DNA fragmentation detection data into several local data blocks, performs continuous homology analysis on each local data block, uses an attention mechanism to assign dynamic weights to the ring and branch structures in a multi-scale topological feature matrix, extracts base association patterns of base sequences in the DNA fragmentation detection data through one-dimensional convolution, and obtains sequence features. The weighted topological features and sequence features are tensor-joined in a latent space. A multi-layer perceptron performs coarse-grained classification on the fused high-order feature vectors, and outputs preliminary classification results. A lightweight neural network submodule is constructed based on the preliminary classification results to obtain the final classification results. The present invention effectively improves the accuracy and interpretability of the classification results.

Description

Technical Field

[0001] The present invention relates to the technical field of DNA fragment data processing, and in particular to a method for processing DNA fragmentation gene detection data based on artificial intelligence. Background Art

[0002] With the rapid development of high-throughput sequencing technology, the application of DNA fragmentation genetic testing data in personalized medicine and genomics research is becoming increasingly widespread. However, due to the characteristics of high dimensionality, low signal-to-noise ratio and complex spatial structure of the fragmented data generated during the sequencing process, traditional analysis methods are difficult to extract and utilize the key information efficiently and accurately. Existing technologies usually rely on single-dimensional feature extraction methods, such as sequence-based base association patterns or structure-based topological features, which fail to fully explore the collaborative information of sequence and spatial structure in DNA fragmentation data, resulting in insufficient classification accuracy and interpretability. In addition, traditional machine learning models often face problems such as high computational complexity and poor generalization ability when processing large-scale fragmented data, making it difficult to meet the requirements for efficiency and robustness in practical applications. In response to the above problems, there is an urgent need for an innovative method that can integrate multi-dimensional features and adaptively process complex data structures to improve the processing efficiency of DNA fragmentation genetic testing data. Summary of the Invention

[0003] The present invention overcomes the shortcomings of the prior art and provides a method for processing DNA fragmentation gene detection data based on artificial intelligence.

[0004] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is:

[0005] The first aspect of the present invention discloses a method for processing DNA fragmentation gene detection data based on artificial intelligence, comprising the following steps:

[0006] The DNA fragmentation detection data is divided into several local data blocks, and continuous homology analysis is performed on each local data block to generate a multi-scale topological feature matrix;

[0007] The attention mechanism is used to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features. The base association pattern of the base sequence in the DNA fragmentation detection data is extracted through one-dimensional convolution to obtain sequence features.

[0008] The weighted topological features and sequence features are tensor-concatenated in the latent space to generate a fused high-order feature vector;

[0009] The fused high-order feature vectors are coarsely classified through a multi-layer perceptron to output preliminary classification results; based on the preliminary classification results, a lightweight neural network sub-module is constructed to perform fine-grained classification to obtain the final classification results.

[0010] Preferably, the DNA fragmentation detection data is divided into several local data blocks, and continuous homology analysis is performed on each local data block to generate a multi-scale topological feature matrix, specifically:

[0011] Based on the preset sliding window size and step size, the DNA fragmentation detection data is divided into multiple local data blocks. The window size is determined according to the average length and mutation frequency of the fragments, and the step size is set to 50% of the window size.

[0012] A point cloud representation is constructed for each local data block, where each point represents a base or mutation site, and the distance matrix between points is calculated based on the physicochemical properties of the bases (such as charge and hydrophobicity);

[0013] Generate the Vietoris-Rips complex based on the distance matrix, and calculate the barcode of the Vietoris-Rips complex at different scales through continuous homology analysis, and extract the topological invariant features of the ring structure (H1) and branch connection point (H0);

[0014] If the continuous length of the ring structure in the barcode exceeds a preset length threshold, the corresponding local data block is determined to have topological features and is marked as a key area;

[0015] The topological features of each local data block are integrated into a multi-scale topological feature matrix, where each feature vector contains the persistence score of the ring structure, the number of branch points and their spatial distribution information.

[0016] Preferably, the attention mechanism is used to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features, specifically:

[0017] The multi-scale topological feature matrix is ​​used as input, where each row represents the topological feature vector of a local data block, including the persistence score of the ring structure, the number of branch points and their spatial distribution information;

[0018] Initialize the query, key, and value matrices in the attention mechanism. The query matrix is ​​generated by the global context features, the key matrix is ​​obtained by linearly transforming the topological feature vector, and the value matrix is ​​directly mapped to the weighted representation of the topological feature vector.

[0019] Calculate the dot product of the query matrix and the key matrix, and normalize it through the Softmax function to obtain the attention weight. If the persistence score of the ring structure exceeds the preset score threshold, increase its corresponding attention weight by the preset amplitude.

[0020] Multiply the attention weights by the value matrix to obtain the weighted topological feature vector. If there are multiple ring structures in the topological feature matrix, a multi-head attention mechanism is further introduced to calculate the weights of different ring structures separately and fuse their results.

[0021] The weighted topological feature vectors are integrated into weighted topological features.

[0022] Preferably, the base association pattern of the base sequence in the DNA fragmentation detection data is extracted by one-dimensional convolution to obtain the sequence features, specifically:

[0023] Convert DNA fragmentation detection data into a numerical sequence, where each base (A, T, C, G) is mapped to a preset numerical code and the sequence is padded to a fixed length;

[0024] Initialize the one-dimensional convolution kernel, input the digitized sequence into the one-dimensional convolution layer, and extract local base association patterns using a sliding convolution kernel. If the convolution kernel coverage area contains a known functional mutation site, increase the convolution kernel weight according to a preset ratio.

[0025] Perform nonlinear activation (such as ReLU) on the convolution output and reduce the feature dimension through the maximum pooling layer to retain the significant features of the base association pattern in the local area;

[0026] If there are repeated segments or low-complexity regions in the sequence, a residual connection is introduced to add the original sequence features to the convolution output;

[0027] The pooled feature vectors are concatenated into a complete sequence feature representation, where each feature dimension corresponds to a specific base association pattern.

[0028] Preferably, the weighted topological features and sequence features are tensor-concatenated in the latent space to generate a fused high-order feature vector, specifically:

[0029] The weighted topological feature matrix and the sequence feature matrix are input into the latent space mapping layer respectively; the weighted topological feature matrix contains the ring structure, branch connection points and their dynamic weight information, and the sequence feature matrix contains the base association pattern and its significant features;

[0030] The weighted topological features and sequence features are mapped to latent space representations of the same dimension through a fully connected layer. If the dimension of the topological feature matrix is ​​higher than that of the sequence feature matrix, the sequence feature matrix is ​​zero-filled or interpolated to ensure the consistency of the dimension.

[0031] Normalize the mapped latent space representation, and perform tensor splicing of the normalized topological features and sequence features in the latent space;

[0032] Perform nonlinear transformation (such as activation function or feature cross) on the concatenated high-order feature representation to generate a fused high-order feature vector.

[0033] Preferably, the fused high-order feature vectors are coarsely classified by a multi-layer perceptron, and a preliminary classification result is output, specifically:

[0034] The fused high-order feature vector is input into the input layer of the multilayer perceptron, where each feature dimension corresponds to the joint representation of the weighted topological feature and the sequence feature;

[0035] Perform nonlinear transformation on the feature vector through the hidden layer and use activation functions (such as ReLU) to extract the complex relationship between features;

[0036] The output of the hidden layer is passed to the Softmax output layer to calculate the probability distribution of each category. If the maximum probability value is lower than the preset probability threshold, the feature backtracking mechanism is triggered to readjust the fusion weights of the weighted topological features and sequence features and iteratively optimize the classification results.

[0037] The probability distribution of Softmax output is used as the preliminary classification result.

[0038] Preferably, a lightweight neural network submodule is constructed based on the preliminary classification results to perform fine-grained classification and obtain the final classification results, specifically:

[0039] The probability distribution of each category is obtained based on the preliminary classification results output by the multi-layer perceptron as the input of the lightweight neural network submodule;

[0040] Construct a lightweight neural network submodule to extract fine-grained features by performing nonlinear transformation on the probability distribution of each category in the hidden layer;

[0041] The hidden layer output is passed to the Softmax output layer to calculate the fine-grained probability distribution of each category. The maximum value in the fine-grained probability distribution is taken as the highest probability value, and the second largest value is taken as the second highest probability value. If the difference between the highest probability value and the second highest probability value is lower than the preset threshold, it is determined that the classification result is ambiguous. The fusion weight of the weighted topological features and sequence features is readjusted and the classification result is iteratively optimized.

[0042] Calculate the confidence of the classification result. The confidence is the difference between the highest probability value and the second highest probability value. If the confidence is lower than the preset confidence threshold, the corresponding classification result will be marked as "pending" and output to the manual review module. Otherwise, the category corresponding to the highest probability will be used as the final classification result.

[0043] The second aspect of the present invention discloses a DNA fragmentation gene detection data processing system based on artificial intelligence, wherein the DNA fragmentation gene detection data processing system includes a memory and a processor, wherein the memory stores a DNA fragmentation gene detection data processing method program. When the DNA fragmentation gene detection data processing method program is executed by the processor, any one of the steps of the DNA fragmentation gene detection data processing method is implemented.

[0044] The third aspect of the present invention discloses a computer-readable storage medium, which includes a DNA fragmentation gene detection data processing method program. When the DNA fragmentation gene detection data processing method program is executed by a processor, any one of the steps of the DNA fragmentation gene detection data processing method is implemented.

[0045] The present invention solves the technical defects existing in the background technology and has the following beneficial effects: dividing DNA fragmentation detection data into several local data blocks, performing continuous homology analysis on each local data block, and generating a multi-scale topological feature matrix; using the attention mechanism to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features; extracting the base association pattern of the base sequence in the DNA fragmentation detection data through one-dimensional convolution to obtain sequence features; performing tensor splicing on the weighted topological features and the sequence features in the latent space to generate a fused high-order feature vector; performing coarse-grained classification on the fused high-order feature vector through a multi-layer perceptron, and outputting a preliminary classification result; constructing a lightweight neural network submodule based on the preliminary classification result, performing fine-grained classification, and obtaining a final classification result. The present invention effectively improves the accuracy and interpretability of the classification results of DNA fragmentation detection data. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, without paying any creative work, they can also obtain drawings of other embodiments based on these drawings.

[0047] Figure 1 This is a flow chart of the first method of the DNA fragmentation gene detection data processing method;

[0048] Figure 2 This is a flow chart of the second method of the DNA fragmentation gene detection data processing method;

[0049] Figure 3 This is the system block diagram of the DNA fragmentation gene detection data processing system. DETAILED DESCRIPTION

[0050] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0052] like Figure 1 As shown, the first aspect of the present invention discloses a method for processing DNA fragmentation gene detection data based on artificial intelligence, comprising the following steps:

[0053] S102, dividing the DNA fragmentation detection data into several local data blocks, performing continuous homology analysis on each local data block, and generating a multi-scale topological feature matrix;

[0054] S104. Using an attention mechanism to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features; extracting base association patterns of base sequences in the DNA fragmentation detection data through one-dimensional convolution to obtain sequence features;

[0055] S106, performing tensor splicing on the weighted topological features and the sequence features in the latent space to generate a fused high-order feature vector;

[0056] S108. Perform coarse-grained classification on the fused high-order feature vectors through a multi-layer perceptron and output a preliminary classification result; construct a lightweight neural network submodule based on the preliminary classification result, perform fine-grained classification, and obtain the final classification result.

[0057] It should be noted that the present invention realizes the efficient extraction and fusion of topological structures and sequence association patterns in DNA fragmentation data by introducing technologies such as continuous homology analysis, attention mechanism and one-dimensional convolution, and combines the collaborative classification strategy of multi-layer perceptron and lightweight neural network sub-module to effectively improve the accuracy and interpretability of the classification results of DNA fragmentation detection data.

[0058] Preferably, the DNA fragmentation detection data is divided into several local data blocks, and continuous homology analysis is performed on each local data block to generate a multi-scale topological feature matrix, such as Figure 2 As shown, specifically:

[0059] S202, based on a preset sliding window size and step size, dividing the DNA fragmentation detection data into multiple local data blocks; wherein the window size is determined according to the average length and mutation frequency of the fragments, and the step size is set to 50% of the window size;

[0060] S204, constructing a point cloud representation for each local data block, where each point represents a base or mutation site, and calculating a distance matrix between points based on the physical and chemical properties of the bases;

[0061] It should be noted that each base or mutation site in the local data block is mapped to a point in three-dimensional space, whose coordinates are determined by the base's physicochemical properties (such as charge, hydrophobicity, and molecular weight) using preset transformation rules. Secondly, based on the coordinates of each point, the Euclidean distance between them is calculated. Due to the symmetry and non-negativity of the distance matrix, a complete distance matrix is ​​constructed, in which each element represents the weighted distance between two points.

[0062] S206. Generate the Vietoris-Rips complex based on the distance matrix, and calculate the barcode of the Vietoris-Rips complex at different scales through continuous homology analysis, and extract the topological invariant features of the ring structure (H1) and branch connection points (H0);

[0063] Among them, Vietoris-Rips complex is a topological structure construction method based on point cloud data, which is used to characterize high-order spatial relationships between data points.

[0064] It should be noted that the Vietoris-Rips complex is constructed based on the distance matrix using scale parameters that gradually increase in a preset ratio. Each scale parameter corresponds to a specific distance threshold. When the distance between points is less than the threshold, they are connected into edges, triangles, or higher-dimensional simplexes (i.e., the Vietoris-Rips complex). Persistent homology analysis tracks the topological structure changes of the complex at different scales, recording the generation and disappearance of ring structures (H1) and branch junctions (H0). Next, the generation and disappearance scale parameters of each topological feature are plotted as a barcode, where each bar represents the persistence interval of a topological feature. If the persistence length of a ring structure in the barcode exceeds a preset threshold, the ring structure is determined to be a significant topological feature. Finally, the persistence scores and spatial distribution information of all significant ring structures and branch junctions are extracted to generate a topologically invariant feature vector.

[0065] S208: If the continuous length of the ring structure in the barcode exceeds a preset length threshold, it is determined that the corresponding local data block has a topological feature and is marked as a key area;

[0066] S210 , integrating the topological features of each local data block into a multi-scale topological feature matrix, wherein each feature vector includes the persistence score of the ring structure, the number of branch points, and their spatial distribution information.

[0067] In summary, this method can efficiently extract multi-scale topological feature matrices by dividing DNA fragmentation detection data into local data blocks and performing continuous coherence analysis, thereby achieving accurate characterization of the complex spatial structure in DNA fragmentation data.

[0068] Preferably, the attention mechanism is used to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features, specifically:

[0069] The multi-scale topological feature matrix is ​​used as input, where each row represents the topological feature vector of a local data block, including the persistence score of the ring structure, the number of branch points and their spatial distribution information;

[0070] Initialize the query, key, and value matrices in the attention mechanism. The query matrix is ​​generated by the global context features, the key matrix is ​​obtained by linearly transforming the topological feature vector, and the value matrix is ​​directly mapped to the weighted representation of the topological feature vector.

[0071] Calculate the dot product of the query matrix and the key matrix, and normalize it through the Softmax function to obtain the attention weight. If the persistence score of the ring structure exceeds the preset score threshold, increase its corresponding attention weight by the preset amplitude.

[0072] Multiply the attention weights by the value matrix to obtain a weighted topological feature vector, which is then integrated into a weighted topological feature.

[0073] Among them, if there are multiple ring structures in the topological feature matrix, a multi-head attention mechanism is further introduced to calculate the weights of different ring structures separately and fuse their results.

[0074] In summary, this method improves the representation ability of key topological features by introducing an attention mechanism to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix. It can adaptively focus on ring structures with higher persistence scores and enhance their contribution to classification decisions. The weighted topological features finally generated provide high-dimensional and precise feature representation for subsequent classification and feature fusion, thereby improving the accuracy and interpretability of genetic testing data processing.

[0075] Preferably, the base association pattern of the base sequence in the DNA fragmentation detection data is extracted by one-dimensional convolution to obtain the sequence features, specifically:

[0076] Convert DNA fragmentation detection data into a numerical sequence, where each base is mapped to a preset numerical code and the sequence is padded to a fixed length;

[0077] Initialize the one-dimensional convolution kernel, input the digitized sequence into the one-dimensional convolution layer, and extract local base association patterns using a sliding convolution kernel. If the convolution kernel coverage area contains a known functional mutation site, increase the convolution kernel weight according to a preset ratio.

[0078] Perform nonlinear activation (such as ReLU) on the convolution output and reduce the feature dimension through the maximum pooling layer to retain the significant features of the base association pattern in the local area;

[0079] It should be noted that multiple one-dimensional convolution kernels are initialized based on the complexity of the base sequence and the distribution of functional mutation sites. For example, their size is set to 3 to 7 bases in length, and the number of convolution kernels is dynamically adjusted based on the sequence complexity. The numerical sequence is input into the one-dimensional convolution layer, and the local base association pattern is extracted using a sliding convolution kernel. If the convolution kernel coverage area contains a known functional mutation site, the convolution kernel weight is adjusted by a preset ratio (such as a 20% increase) to enhance the base pattern extraction effect in the key area. The convolution output is then activated nonlinearly (such as ReLU) to capture the nonlinear relationship in the base association pattern. The activated features are then reduced in dimension using a maximum pooling layer to retain the significant features of the base association pattern in the local area.

[0080] If there are repeated segments or low-complexity regions in the sequence, a residual connection is introduced to add the original sequence features to the convolution output;

[0081] It should be noted that before the convolution operation, the digitized sequence is preprocessed to mark the locations of repeated segments or low-complexity regions. After the convolution layer output, the original sequence features are added to the convolution output element by element through a residual connection.

[0082] The pooled feature vectors are concatenated into a complete sequence feature representation, where each feature dimension corresponds to a specific base association pattern.

[0083] A base sequence is a sequence formed by converting the bases (A, T, C, G) in a DNA fragment into a numerical representation according to their order in the genome. For example, a DNA fragment "ATCG" can be encoded as a numerical sequence [0, 1, 2, 3], where each number represents a specific base. This numerical sequence serves as the input for a subsequent one-dimensional convolution operation, which extracts local correlation patterns between bases and generates sequence features.

[0084] In summary, this method extracts local correlation patterns of base sequences in DNA fragmentation detection data through one-dimensional convolution, which can efficiently capture the nonlinear relationship between bases and improve the characterization ability of sequence features.

[0085] Preferably, the weighted topological features and sequence features are tensor-concatenated in the latent space to generate a fused high-order feature vector, specifically:

[0086] The weighted topological feature matrix and the sequence feature matrix are input into the latent space mapping layer respectively; the weighted topological feature matrix contains the ring structure, branch connection points and their dynamic weight information, and the sequence feature matrix contains the base association pattern and its significant features;

[0087] The weighted topological features and sequence features are mapped to latent space representations of the same dimension through a fully connected layer. If the dimension of the topological feature matrix is ​​higher than that of the sequence feature matrix, the sequence feature matrix is ​​zero-filled or interpolated to ensure the consistency of the dimension.

[0088] It should be noted that the weighted topological feature matrix and the sequence feature matrix are respectively input into the fully connected layer, where the output dimension of the fully connected layer is pre-set according to the dimension of the topological feature matrix; secondly, the weighted topological features are mapped to the latent space representation of the target dimension through the fully connected layer. If the dimension of the sequence feature matrix is ​​lower than the target dimension, the sequence feature matrix is ​​zero-padded, and zero values ​​are added at the end of the matrix to expand its dimension, or new feature values ​​are generated between existing features through interpolation methods (such as linear interpolation) to match the target dimension.

[0089] Normalize the mapped latent space representation, and perform tensor splicing of the normalized topological features and sequence features in the latent space;

[0090] It should be noted that the normalized topological feature matrix and sequence feature matrix are converted into tensor forms of the same dimension. The topological feature matrix contains information about ring structures, branch junctions, and their dynamic weights, while the sequence feature matrix contains base association patterns and their significance features. The two tensors are concatenated along the feature dimension and reshaped to conform to the format requirements of subsequent model input.

[0091] Perform nonlinear transformation (such as activation function or feature cross) on the concatenated high-order feature representation to generate a fused high-order feature vector.

[0092] It should be noted that the concatenated high-order feature representation is input into the nonlinear transformation layer, where the activation function (such as ReLU) is used to capture the nonlinear relationship between features. If the feature representation contains topological and sequence information of multiple dimensions, a feature crossover operation is introduced to enhance the synergy between features through element-by-element multiplication or weighted combination. The activated or crossed features are subjected to dimensionality reduction, and key feature information is retained through a fully connected layer or pooling operation. The reduced dimensionality features are normalized to eliminate dimensional differences between features and improve the convergence speed of the model; then, the normalized feature vectors are integrated to generate a fused high-order feature vector; finally, the generated high-order feature vector is used as the input of the subsequent classification or decision model to ensure the deep fusion and efficient utilization of topological and sequence information.

[0093] In summary, this method achieves a deep fusion of topological structure and sequence association patterns in DNA fragmentation data by tensor splicing weighted topological features and sequence features in latent space, thereby improving the comprehensiveness and robustness of feature representation.

[0094] Preferably, the fused high-order feature vectors are coarsely classified by a multi-layer perceptron, and a preliminary classification result is output, specifically:

[0095] The fused high-order feature vector is input into the input layer of the multilayer perceptron, where each feature dimension corresponds to the joint representation of the weighted topological feature and the sequence feature;

[0096] Perform nonlinear transformation on the feature vector through the hidden layer and use activation functions (such as ReLU) to extract the complex relationship between features;

[0097] The output of the hidden layer is passed to the Softmax output layer to calculate the probability distribution of each category. If the maximum probability value is lower than the preset probability threshold, the feature backtracking mechanism is triggered to readjust the fusion weights of the weighted topological features and sequence features and iteratively optimize the classification results.

[0098] The probability distribution of Softmax output is used as the preliminary classification result.

[0099] It's important to note that the Multilayer Perceptron (MLP) is a feedforward artificial neural network consisting of an input layer, hidden layers, and an output layer. It can model complex data by learning nonlinear mapping relationships. Its operating principle is as follows: the input layer receives a feature vector, the hidden layer transforms the input features using a nonlinear activation function (such as ReLU) to extract the complex relationships between features, and the output layer generates classification or regression results using a Softmax or Sigmoid function. The MLP optimizes network parameters using a backpropagation algorithm, adaptively capturing high-order features in the data. It is suitable for classification and regression tasks involving high-dimensional, nonlinear data, and exhibits strong fitting and generalization capabilities.

[0100] In summary, this method uses a multi-layer perceptron to perform coarse-grained classification on the fused high-order feature vectors, which can efficiently capture the complex relationship between topology and sequence information and improve the accuracy of the classification results. The final output preliminary classification results provide high-quality candidate categories for subsequent fine-grained classification, thereby improving the overall efficiency and accuracy of genetic testing data processing.

[0101] Preferably, a lightweight neural network submodule is constructed based on the preliminary classification results to perform fine-grained classification and obtain the final classification results, specifically:

[0102] The probability distribution of each category is obtained based on the preliminary classification results output by the multi-layer perceptron as the input of the lightweight neural network submodule;

[0103] Construct a lightweight neural network submodule to extract fine-grained features by performing nonlinear transformation on the probability distribution of each category in the hidden layer;

[0104] The hidden layer output is passed to the Softmax output layer to calculate the fine-grained probability distribution of each category. The maximum value in the fine-grained probability distribution is taken as the highest probability value, and the second largest value is taken as the second highest probability value. If the difference between the highest probability value and the second highest probability value is lower than the preset threshold, it is determined that the classification result is ambiguous. The fusion weight of the weighted topological features and sequence features is readjusted and the classification result is iteratively optimized.

[0105] It should be noted that the difference between the highest and second-highest probability values ​​in the fine-grained probability distribution generated by the Softmax output layer is calculated. If the difference between the two is lower than a preset threshold (such as 0.1), the classification result is determined to be ambiguous. Based on the ambiguous classification results, the algorithm goes back to the feature fusion stage and adjusts the fusion weights of the weighted topological features and sequence features to increase the contribution of key features according to the preset proportion. The adjusted feature vector is re-input into the lightweight neural network submodule to perform fine-grained classification and calculate a new probability distribution. The above steps are repeated until the difference between the highest and second-highest probability values ​​is no lower than the preset threshold or the maximum number of iterations is reached. Finally, the optimized classification result is used as the final output to ensure the rigor and accuracy of the classification process.

[0106] Calculate the confidence of the classification result. The confidence is the difference between the highest probability value and the second highest probability value. If the confidence is lower than the preset confidence threshold, the corresponding classification result will be marked as "pending" and output to the manual review module. Otherwise, the category corresponding to the highest probability will be used as the final classification result.

[0107] The final classification results include the following specific contents: First, the category corresponding to the highest probability is used as the main classification result, indicating the category to which the model believes the sample most likely belongs; second, the confidence of the classification result, that is, the difference between the highest probability value and the second highest probability value, is used to measure the reliability of the classification result; if the confidence is lower than the preset confidence threshold, the sample is marked as "pending" and output to the manual review module for further verification; in addition, it also includes the feature dimensions that contribute most to the classification results and their corresponding topological structures and base association patterns, which are used to assist users in understanding the basis of classification decisions.

[0108] It should be noted that this method further improves the accuracy and reliability of the classification results by constructing a lightweight neural network submodule to perform fine-grained classification on the preliminary classification results. Specifically, through nonlinear transformations in the hidden layer and probabilistic calculations in the Softmax output layer, it can adaptively extract fine-grained features and generate high-confidence classification results. If the classification result is ambiguous or the confidence level falls below a preset threshold, a feature fusion weight adjustment mechanism or a manual review process is triggered to ensure the rigor and accuracy of the classification process. The final output classification results provide high-precision discrimination results for genetic testing data, improving the credibility and practicality of data processing.

[0109] In this embodiment, the DNA fragmentation gene detection data processing method further includes the following steps:

[0110] The gene mutations or functional regions in the final classification results are extracted as query sequences. If the final classification results contain multiple mutation sites, they are sorted according to the preset priority. Secondly, the query sequences are compared with the gene function annotation database (such as GO, KEGG), and the corresponding entries of the query sequences in the database are identified through sequence alignment algorithms (such as BLAST). Then, biological function, pathway and disease association information are extracted based on the alignment results. If the query sequence has no direct matching entries in the database, its function is inferred through homologous genes or conserved regions. Then, the extracted functional information is associated with the classification results and annotated. If the same mutation site corresponds to multiple functions or pathways, the most relevant annotation is selected based on the functional significance score (such as P value). Finally, the annotation results are integrated into a functional annotation report, which includes the mutation site, functional description, pathway association, disease risk and references.

[0111] It should be noted that this method realizes the automated functional annotation and association analysis of gene mutation sites by dynamically linking the classification results with the gene function annotation database; it integrates multidimensional information of mutation sites, functional pathways and disease risks to generate structured reports, providing indirect guidance for clinical diagnosis and scientific research.

[0112] In this embodiment, the DNA fragmentation gene detection data processing method further includes the following steps:

[0113] After generating the fused high-order feature vector, epigenetic data (such as methylation and chromatin conformation capture data) are introduced;

[0114] The epigenetic data are converted into a three-dimensional space coordinate and interaction strength matrix, where the position of each DNA fragment is determined by its genomic coordinates and the interaction frequency of neighboring fragments;

[0115] It should be noted that the genomic coordinates of each DNA fragment are extracted based on epigenetic data (such as chromatin conformation capture data) and mapped to their initial positions in three-dimensional space. The spatial distance between fragments is calculated using the interaction frequency of adjacent fragments (such as the interaction strength in the Hi-C matrix). If the interaction frequency exceeds a preset threshold, the spatial distance between fragments is shortened, and vice versa. An interaction strength matrix is ​​constructed based on the three-dimensional coordinates, where the matrix elements represent the spatial interaction strength between fragments.

[0116] A graph neural network is constructed, where nodes represent DNA fragment regions and edges represent spatial interaction strengths (obtained from the interaction strength matrix). Features of neighboring nodes are aggregated through a graph convolutional layer to generate three-dimensional structural features.

[0117] It should be noted that DNA fragment regions are used as nodes in the graph neural network. The node feature vector includes its genomic coordinates, epigenetic markers (such as methylation levels), and the interaction strength of neighboring fragments. The graph edges are constructed based on the interaction strength matrix. If the interaction strength exceeds a preset threshold, an edge is established between the corresponding nodes, and the edge weight is determined by the interaction strength value. Next, the features of neighboring nodes are aggregated through the graph convolution layer, where each node's features are updated to a weighted sum of its own features and the features of neighboring nodes, with the weight determined by the edge interaction strength.

[0118] The fused high-order feature vector is cross-modally fused with the three-dimensional structural features to obtain cross-modal fusion features. The attention mechanism dynamically adjusts the contribution of the three-dimensional features in the classification decision according to their significance.

[0119] It should be noted that the fused high-order feature vectors and three-dimensional structural features are mapped to latent space representations of the same dimension. The query matrix (Query), key matrix (Key), and value matrix (Value) in the cross-modal attention mechanism are initialized. The dot product of the query matrix and the key matrix is ​​calculated and normalized using the Softmax function to obtain attention weights. If the significance score of certain dimensions in the three-dimensional structural features exceeds a preset threshold, the corresponding attention weights are increased by a preset proportion. The adjusted attention weights are then multiplied by the value matrix to obtain weighted three-dimensional structural features. Finally, the weighted three-dimensional structural features are concatenated with the high-order feature vectors to generate cross-modal fusion features.

[0120] The cross-modal fusion features are input into a lightweight neural network to generate the final classification results, and the segment regions that have the greatest impact on the classification results are located through inverse gradient saliency mapping (such as Grad-CAM);

[0121] It should be noted that the contribution of each feature dimension to the classification result is calculated through reverse gradient significance mapping. If the contribution is higher than the preset threshold, it will be marked as a key feature. According to the DNA fragment region corresponding to the key feature, the fragment region with the greatest impact on the classification result is located, and a visual significance map is generated as an interpretable output of the classification result.

[0122] The positioning results are superimposed with the epigenetic data to generate a visual interpretable map, annotate the functional annotations of key fragment regions and their contribution to the classification results, and output them through the user terminal.

[0123] In summary, this method realizes multimodal feature fusion and three-dimensional space modeling, thereby effectively improving the accuracy and interpretability of classification results.

[0124] like Figure 3As shown, the second aspect of the present invention discloses a DNA fragmentation gene detection data processing system 6 based on artificial intelligence, wherein the DNA fragmentation gene detection data processing system includes a memory 41 and a processor 52, wherein the memory 41 stores a DNA fragmentation gene detection data processing method program, and when the DNA fragmentation gene detection data processing method program is executed by the processor 52, any one of the steps of the DNA fragmentation gene detection data processing method is implemented.

[0125] The third aspect of the present invention discloses a computer-readable storage medium, which includes a DNA fragmentation gene detection data processing method program. When the DNA fragmentation gene detection data processing method program is executed by a processor, any one of the steps of the DNA fragmentation gene detection data processing method is implemented.

[0126] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0127] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0128] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0129] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0130] Alternatively, if the integrated units described above are implemented as software modules and sold or used as standalone products, they can also be stored on a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product, stored on a storage medium, includes instructions for enabling a computer device (such as a personal computer, server, or network device) to execute all or part of the methods of the various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as removable storage devices, ROM, RAM, magnetic disks, or optical disks.

[0131] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A DNA fragmentation gene detection data processing method based on artificial intelligence, characterized in that: The following steps are involved: The DNA fragmentation detection data is divided into several local data blocks, and continuous homology analysis is performed on each local data block to generate a multi-scale topological feature matrix; The attention mechanism is used to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features. The base association pattern of the base sequence in the DNA fragmentation detection data is extracted through one-dimensional convolution to obtain sequence features. The weighted topological features and sequence features are tensor-concatenated in the latent space to generate a fused high-order feature vector; The fused high-order feature vectors are coarsely classified using a multi-layer perceptron to output preliminary classification results. Based on the preliminary classification results, a lightweight neural network submodule is constructed to perform fine-grained classification to obtain the final classification results. The DNA fragmentation detection data is divided into several local data blocks, and continuous homology analysis is performed on each local data block to generate a multi-scale topological feature matrix, specifically: Based on the preset sliding window size and step size, the DNA fragmentation detection data is divided into multiple local data blocks; Construct a point cloud representation for each local data block, where each point represents a base or mutation site, and calculate the distance matrix between points based on the physical and chemical properties of the bases; Generate the Vietoris-Rips complex based on the distance matrix, and calculate the barcode of the Vietoris-Rips complex at different scales through continuous homology analysis to extract the topological invariant features of the ring structure and branch connection points; If the continuous length of the ring structure in the barcode exceeds a preset length threshold, the corresponding local data block is determined to have topological features and is marked as a key area; The topological features of each local data block are integrated into a multi-scale topological feature matrix, where each feature vector contains the persistence score of the ring structure, the number of branch points and their spatial distribution information; Among them, the base association pattern of the base sequence in the DNA fragmentation detection data is extracted by one-dimensional convolution to obtain the sequence features, specifically: Convert DNA fragmentation detection data into a numerical sequence, where each base is mapped to a preset numerical code and the sequence is padded to a fixed length; Initialize the one-dimensional convolution kernel, input the digitized sequence into the one-dimensional convolution layer, and extract the local base association pattern by sliding the convolution kernel; Perform nonlinear activation on the convolution output and reduce the feature dimension through the maximum pooling layer to retain the significant features of the base association pattern in the local area; If there are repeated segments or low-complexity regions in the sequence, a residual connection is introduced to add the original sequence features to the convolution output; The pooled feature vectors are concatenated into a complete sequence feature representation, where each feature dimension corresponds to a specific base association pattern.

2. The method for processing DNA fragmentation gene detection data based on artificial intelligence according to claim 1, characterized in that: The attention mechanism is used to assign dynamic weights to the ring and branch structures in the multi-scale topological feature matrix to obtain weighted topological features, specifically: The multi-scale topological feature matrix is ​​used as input, where each row represents the topological feature vector of a local data block, including the persistence score of the ring structure, the number of branch points and their spatial distribution information; Initialize the query, key, and value matrices in the attention mechanism, where the query matrix is ​​generated by the global context features, the key matrix is ​​obtained by linear transformation of the topological feature vector, and the value matrix is ​​directly mapped to the weighted representation of the topological feature vector; Calculate the dot product of the query matrix and the key matrix, and normalize it through the Softmax function to obtain the attention weight. If the persistence score of the ring structure exceeds the preset score threshold, increase its corresponding attention weight by the preset amplitude. Multiply the attention weights by the value matrix to obtain a weighted topological feature vector; and integrate the weighted topological feature vector into a weighted topological feature.

3. The method for processing DNA fragmentation gene detection data based on artificial intelligence according to claim 1, characterized in that: The weighted topological features and sequence features are tensor-concatenated in the latent space to generate a fused high-order feature vector, specifically: The weighted topological feature matrix and the sequence feature matrix are input into the latent space mapping layer respectively; the weighted topological feature matrix contains the ring structure, branch connection points and their dynamic weight information, and the sequence feature matrix contains the base association pattern and its significant features; The weighted topological features and sequence features are mapped to latent space representations of the same dimension through a fully connected layer. If the dimension of the topological feature matrix is ​​higher than that of the sequence feature matrix, the sequence feature matrix is ​​zero-filled or interpolated to ensure the consistency of the dimension. Normalize the mapped latent space representation, and perform tensor splicing of the normalized topological features and sequence features in the latent space; Perform nonlinear transformation on the concatenated high-order feature representation to generate a fused high-order feature vector.

4. The method for processing DNA fragmentation gene detection data based on artificial intelligence according to claim 1, characterized in that: The fused high-order feature vectors are coarsely classified through a multi-layer perceptron and the preliminary classification results are output, specifically: The fused high-order feature vector is input into the input layer of the multilayer perceptron, where each feature dimension corresponds to the joint representation of the weighted topological feature and the sequence feature; Perform nonlinear transformation on feature vectors through hidden layers and use activation functions to extract complex relationships between features; The output of the hidden layer is passed to the Softmax output layer to calculate the probability distribution of each category. If the maximum probability value is lower than the preset probability threshold, the feature backtracking mechanism is triggered to readjust the fusion weights of the weighted topological features and sequence features and iteratively optimize the classification results. The probability distribution of Softmax output is used as the preliminary classification result.

5. The method for processing DNA fragmentation gene detection data based on artificial intelligence according to claim 1, characterized in that: Based on the preliminary classification results, a lightweight neural network submodule is constructed to perform fine-grained classification and obtain the final classification results, specifically: The probability distribution of each category is obtained based on the preliminary classification results output by the multi-layer perceptron as the input of the lightweight neural network submodule; Construct a lightweight neural network submodule to extract fine-grained features by performing nonlinear transformation on the probability distribution of each category in the hidden layer; The hidden layer output is passed to the Softmax output layer to calculate the fine-grained probability distribution of each category. The maximum value in the fine-grained probability distribution is taken as the highest probability value, and the second largest value is taken as the second highest probability value. If the difference between the highest probability value and the second highest probability value is lower than the preset threshold, it is determined that the classification result is ambiguous. The fusion weight of the weighted topological features and sequence features is readjusted and the classification result is iteratively optimized. The confidence level of the classification result is calculated as the difference between the highest probability value and the second highest probability value. If the confidence level is lower than the preset confidence threshold, the corresponding classification result is marked as "pending" and output to the manual review module. Otherwise, the category corresponding to the highest probability is used as the final classification result.

6. The artificial intelligence-based DNA fragmentation gene detection data processing system is characterized by: The DNA fragmentation gene detection data processing system includes a memory and a processor, wherein the memory stores a DNA fragmentation gene detection data processing method program. When the DNA fragmentation gene detection data processing method program is executed by the processor, the DNA fragmentation gene detection data processing method steps as described in any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a DNA fragmentation gene detection data processing method program. When the DNA fragmentation gene detection data processing method program is executed by a processor, the steps of the DNA fragmentation gene detection data processing method according to any one of claims 1 to 5 are implemented.

Images