Deep learning cancer risk prediction method and system based on methylation sequence

Abstract

The invention relates to the technical field of cancer risk prediction, in particular to a deep learning cancer risk prediction method and system based on a methylation sequence. The method comprises the following steps: acquiring DNA methylation sequencing data and reference genome annotation data in a blood sample of a cancer patient; performing functional region classification processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter subregion methylation data, enhancement subregion methylation data and transcription factor binding site methylation data; performing dynamic sequence similarity calculation processing on the methylation data of the promoter region to obtain a first feature vector containing importance weight of the promoter region of the cancer-related gene; the methylation sites are classified according to the functions of the methylation sites in the transcriptional regulation pathway, so that the biological interpretation of the characteristics is improved, and the synergistic effect among different functional regions can be captured.

Description

Technical Field

[0001] The present invention relates to the technical field of cancer risk prediction, and particularly to a deep learning cancer risk prediction method and system based on methylation sequences. Background Art

[0002] As an important epigenetic marker, DNA methylation has been widely used in cancer risk prediction, early diagnosis, and discovery of therapeutic targets. DNA methylation refers to the addition of methyl groups to the carbon-5 position of certain cytosines in the DNA molecule, which plays a crucial role in gene expression regulation. Especially during cancer development, abnormal methylation patterns are often closely related to the occurrence and development of cancer. However, traditional cancer risk prediction methods based on methylation sequences have significant limitations in data processing and feature extraction, making it difficult to handle multi-dimensional and non-linear biological data. Especially in the context of multi-pathway co-regulation, existing methods are difficult to comprehensively consider the interactions of various functional regions and biological pathways in the gene regulatory network, resulting in insufficient accuracy and biological interpretability of prediction results. Traditional cancer risk prediction methods usually rely on single genes or clinical data, and these methods often suffer from problems such as insufficient sensitivity, large individual differences, and limited prognosis judgment, making it difficult to accurately reveal the complex mechanisms and multi-factor interactions of cancer. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a deep learning cancer risk prediction method and system based on methylation sequences to solve at least one of the above technical problems.

[0004] To achieve the above object, a deep learning cancer risk prediction method based on methylation sequences includes the following steps: Step S1: Obtain DNA methylation sequencing data and reference genome annotation data from the blood samples of cancer patients; perform functional region classification processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data; Step S2: Perform dynamic sequence similarity calculation processing on the promoter region methylation data to obtain a first feature vector containing the importance weights of the promoter regions of cancer-related genes; perform expression regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity status; perform binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability; Step S3: Perform biological pathway correlation analysis processing on the first feature vector, the second feature vector, and the third feature vector according to the preset gene regulatory network topological structure to obtain multi-level feature data with pathway correspondence relationships; Step S4: Perform attention mechanism processing with biological pathway constraints on the multi-level feature data to obtain pathway activity score data; perform multi-task learning processing based on cancer risk prediction and key regulatory pathway interpretation according to the pathway activity score data to obtain risk distribution mapping data and pathway-specific analysis data.

[0005] The present invention also provides a deep learning cancer risk prediction system based on methylation sequences for performing the above-mentioned deep learning cancer risk prediction method based on methylation sequences. The deep learning cancer risk prediction system based on methylation sequences includes: A data acquisition and functional region classification module for acquiring DNA methylation sequencing data and reference genome annotation data in a blood sample of a cancer patient; performing functional region classification processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data; A feature extraction and analysis module for performing dynamic sequence similarity calculation processing on the promoter region methylation data to obtain a first feature vector containing the importance weights of the promoter regions of cancer-related genes; performing performance regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity status; performing binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability; A biological pathway correlation analysis module for performing biological pathway correlation analysis processing on the first feature vector, the second feature vector, and the third feature vector according to a preset gene regulatory network topological structure to obtain multi-level feature data with pathway correspondence relationships; A multi-task learning and risk prediction module for performing attention mechanism processing with biological pathway constraints on the multi-level feature data to obtain pathway activity score data; performing multi-task learning processing based on cancer risk prediction and key regulatory pathway interpretation according to the pathway activity score data to obtain risk distribution mapping data and pathway-specific analysis data.

[0006] In the present invention, the data acquisition and functional region classification module obtains DNA methylation sequencing data and reference genome annotation data in a blood sample of a cancer patient, and classifies the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data. This provides an accurate data basis for subsequent feature extraction, enables specialized analysis for different functional regions, fully considers the unique roles of each functional region in gene regulation and cancer occurrence, and improves the scientificity and reliability of the entire method. Description of the Drawings

[0007] Other features, objects, and advantages of the present invention will become more apparent by reading the detailed description of non - restrictive embodiments with reference to the following drawings: Figure 1 It is a schematic flowchart of the steps of the deep - learning cancer risk prediction method based on methylation sequences of the present invention; Figure 2 For Figure 1 it is a detailed flowchart of step S1 in Figure 3 For Figure 1 it is a detailed flowchart of step S2 in Detailed implementation mode

[0008] To achieve the above object, please refer to Figures 1 to 3 The present invention provides a deep - learning cancer risk prediction method based on methylation sequences. The method includes the following steps: Step S1: Obtain DNA methylation sequencing data and reference genome annotation data in the blood samples of cancer patients; classify the DNA methylation sequencing data according to the reference genome annotation data for functional region processing to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data; When the embodiment of the present invention obtains DNA methylation sequencing data and reference genome annotation data in the blood samples of cancer patients, first, plasma is separated from the peripheral blood samples of cancer patients, and free DNA is extracted. Then, the free DNA samples are methylated using the bisulfite conversion method, and PCR amplification and sequencing are performed using high - throughput sequencing technology to obtain the original methylation sequencing data. Subsequently, through quality control and data filtering steps, low - quality sequencing data is removed to ensure data accuracy. To further analyze the DNA methylation status, it is necessary to rely on the reference genome sequence in the UCSC genome database to perform functional element annotation analysis on the sequencing data. This process divides the genome into different functional regions, including promoter regions, enhancer regions, and transcription factor binding sites. Through alignment and annotation analysis, reference genome annotation data containing genomic functional region information is obtained, and the methylation data is matched and classified with these functional regions. Finally, promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data are obtained. This series of operations can ensure the high quality and accuracy of the data and lay a foundation for subsequent analysis.

[0009] Step S2: Perform dynamic sequence similarity calculation processing on the promoter region methylation data to obtain a first feature vector containing the importance weights of the promoter regions of cancer-related genes; perform performance regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity status; perform binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability; In the embodiment of the present invention, when performing dynamic sequence similarity calculation on the promoter region methylation data, it is first necessary to extract the sequence fragments of the promoter region from the methylation data. Then, the dynamic time warping (DTW) algorithm is used to standardize these sequences for comparison with the sequences in the pre-established promoter region reference pattern library. The DTW algorithm can calculate the similarity between different sequences and find the best match through sequence alignment, thereby identifying the promoter regions of cancer-related genes. After the calculation, based on the similarity values, a feature vector containing the importance weights of the promoter regions is generated, and this feature vector represents the importance of the promoter regions in cancer occurrence. For the enhancer region methylation data, the performance regulation pattern is first identified. Specifically, the data in the enhancer region is annotated by epigenetic modification markers to identify the enhancer activity status. By processing the enhancer activity data through a convolutional neural network (CNN), the activity pattern of the enhancer region can be extracted, and then a feature vector containing the enhancer activity status is generated. For the transcription factor binding site methylation data, potential transcription factor binding sites are obtained through sequence window sliding analysis, and the conservation analysis is performed using the transcription factor binding site database to evaluate the conservation degree of these sites in different species. According to the site conservation score, combined with the deep neural network model, the transcription factor binding probability is calculated, and a feature vector containing the binding probability is generated.

[0010] Step S3: Perform biological pathway correlation analysis processing on the first feature vector, the second feature vector, and the third feature vector according to the preset gene regulation network topological structure to obtain multi-level feature data with pathway correspondence relationships; In the process of biological pathway relevance analysis and processing in the embodiments of the present invention, first, a gene regulatory network database is obtained and its topological structure is analyzed. The topological structure analysis can reveal the regulatory relationships between different genes and functional regions. For example, certain promoter regions may have strong regulatory relationships with specific enhancer regions, or certain transcription factors may play important roles in multiple pathways. Through this analysis, the interaction data between various regions in the gene regulatory network can be obtained. Based on these regulatory relationship data, feature relevance analysis is performed on the three groups of feature vectors (feature vectors of promoter regions, enhancer regions, and transcription factor binding sites) obtained in step S2. This analysis calculates the interaction strength between different features to obtain the relevance data of each feature with other features. These data provide a basis for the next step of pathway enrichment analysis and multi-level feature data generation. Finally, the analysis results will include the relevance of each functional region in the gene regulatory network and their relationships with cancer-related pathways.

[0011] Step S4: Perform attention mechanism processing with biological pathway constraints on the multi-level feature data to obtain pathway activity score data; perform multi-task learning processing based on the pathway activity score data for cancer risk prediction and key regulatory pathway interpretation to obtain risk distribution mapping data and pathway-specific analysis data.

[0012] In the process of attention mechanism processing of the multi-level feature data in the embodiments of the present invention, first, the prior knowledge of biological pathways is used to perform pathway node mapping on the generated multi-level feature data. Each node represents a functional region or a gene, and the mapping result is pathway-specific association data. Then, through weight calculation based on the attention mechanism, the relative importance of each pathway is evaluated. Specifically, the attention mechanism assigns different weights to each pathway feature and constrains these weights according to the pathway regulatory relationships to ensure that the calculation results conform to biological laws. Finally, pathway interaction weight data is obtained, and these data are used for subsequent risk prediction and pathway interpretation. When performing multi-task learning, based on the pathway activity score data, a dual-task optimization model is constructed, and the goal is to predict cancer risk and explain the role of key regulatory pathways simultaneously. By optimizing the loss function, multi-task optimization target data is obtained, and a deep neural network is used for network training. During the training process, a dynamic weight balancing strategy is applied to coordinate the learning process between cancer risk prediction and pathway analysis. Finally, the network model parameters are optimized through the backpropagation algorithm to obtain the final risk distribution mapping data and pathway-specific analysis data.

[0013] Preferably, step S1 includes the following steps: Step S11: Obtain a peripheral blood sample of a cancer patient, perform plasma separation processing on the peripheral blood sample to obtain a plasma-free DNA sample; When obtaining a peripheral blood sample from a cancer patient in an embodiment of the present invention, first, a peripheral blood sample of the patient is collected, usually obtained by venipuncture, and the collection volume is about 10 - 20 milliliters. To separate the plasma, the blood sample needs to be centrifuged. The conventional procedure is to centrifuge the blood sample at 4000 rpm for 10 minutes. After centrifugation, the upper layer of the plasma is separated and then transferred to a new centrifuge tube for re - centrifugation to remove residual blood cell components. The obtained plasma is stored in a low - temperature environment of - 80°C to ensure the integrity of plasma - free DNA. The plasma - free DNA sample mainly comes from the extracellular DNA of the patient and contains methylation information related to cancer, serving as the basis for subsequent analysis.

[0014] Step S12: Perform bisulfite conversion treatment on the plasma - free DNA sample and perform PCR amplification treatment to obtain methylation sequencing data; When performing bisulfite conversion treatment on the plasma - free DNA sample in an embodiment of the present invention, first, the plasma - free DNA is extracted. The phenol - chloroform extraction method or a kit is used to extract the free DNA. After extraction, the concentration and purity of the DNA need to be detected by an ultraviolet spectrophotometer. Then, bisulfite conversion is carried out. The common method is to mix the DNA with a bisulfite solution, and the reaction is carried out under specific conditions. The common conditions are to react at 37°C for 6 - 8 hours. The purpose of the conversion process is to keep the methylated cytosine (C) unchanged, while the unmethylated cytosine is converted to uracil. After the conversion is completed, the converted DNA is amplified by PCR amplification technology. The primers required for PCR amplification are designed according to specific sequences of known genes or regions, and the common primer length is 18 - 22 bases. The amplification conditions are usually: the initial denaturation temperature is 94°C, and the denaturation time is 30 seconds; the annealing temperature is 55 - 60°C, and the annealing time is 30 seconds; the extension temperature is 72°C, and the extension time is 1 minute; the final number of amplification cycles is 30 - 35 times. Through this step, DNA fragments rich in methylation information can be obtained for subsequent methylation sequencing.

[0015] Step S13: Perform high - throughput sequencing on the methylation sequencing data to obtain raw sequencing data; When the embodiment of the present invention performs high-throughput sequencing on the amplified DNA after PCR amplification, it is first necessary to construct a library for the amplified DNA. A common method is to add A to the ends of DNA fragments and ligate sequencing adapters. The design of the adapters needs to take into account the requirements of the sequencing platform. For example, the Illumina platform uses specific adapter sequences and ligates the adapters to the ends of each DNA fragment. After the library construction is completed, high-throughput sequencing platforms such as Illumina NovaSeq or HiSeq are used for sequencing. The selected read length is usually 150bp or 250bp, depending on the experimental requirements. During the sequencing process, each DNA fragment is individually labeled and sequenced in parallel to generate a large amount of short sequence read data. These short sequences will be processed through subsequent alignment and analysis steps to finally generate complete genomic methylation data.

[0016] Step S14: Perform quality control and data filtering on the original sequencing data to obtain DNA methylation sequencing data; When the embodiment of the present invention performs quality control and data filtering on the original sequencing data, it first uses tools such as FastQC to evaluate the quality of the original sequencing data, checking issues such as the quality distribution, GC content, and adapter contamination of the sequencing data. If there are low-quality sequences or adapter contamination in the data, it is necessary to use software such as Trim Galore or Cutadapt to remove the adapters and filter out the low-quality sequences. Common filtering criteria are to remove sequences with an average quality score lower than Q30 and sequences with a length shorter than 100bp. The filtered data will further remove possible redundant sequences and merge the data of different samples. The data after quality control and data filtering will be used for downstream methylation analysis to ensure the reliability and accuracy of the data.

[0017] Step S15: Obtain reference genome sequence data using the UCSC genome database and perform functional element annotation analysis on the reference genome sequence data to obtain reference genome annotation data containing genomic functional region information; In the embodiments of the present invention, when obtaining reference genome sequence data using the UCSC Genome Database, first, a species genome sequence matching the experimental sample is selected. For example, for a human sample, the human genome reference sequence (hg38) is selected. Relevant genomic data, including annotation information of functional regions such as gene structures, transcription factor binding sites, and enhancer regions, are downloaded through the UCSC Genome Browser. The functional element annotation analysis of the genomic data is performed through relevant tools provided by UCSC, such as Genome Browser or BEDTools, etc. Annotations are made based on various functional elements in the genome, and further analysis is carried out on specific gene regions (such as promoters, enhancers, etc.). The output of this process is an annotation file containing information on the functional regions of the reference genome, and the file format is usually in BED or GFF3 format, which contains the position information and annotation tags of each functional region in the genome.

[0018] Step S16: Classify and process the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data.

[0019] In the embodiments of the present invention, when classifying and processing the DNA methylation sequencing data according to the reference genome annotation data, first, the genomic annotation data obtained in step S15 is compared with the DNA methylation data obtained in step S14. Commonly used comparison tools include BWA or Bowtie2. These tools can align the short sequences in the methylation sequencing data to the corresponding positions of the reference genome. Based on the comparison results, tools such as BEDTools are used for functional region classification, and the methylation data is divided into different functional regions such as promoter regions, enhancer regions, and transcription factor binding sites. The division of the promoter region is based on the promoter sites in the genomic annotation file, which is usually defined as the 2000bp region upstream of the gene. The enhancer region is divided according to the enhancer markers in the genomic annotation. The transcription factor binding sites are located based on the annotation data of the transcription factor binding sites. Finally, the obtained results are promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data. These data will provide a basis for subsequent functional analysis and cancer correlation analysis.

[0020] The present invention obtains plasma-free DNA samples by acquiring peripheral blood samples from cancer patients and performing plasma separation, which provides a key biological material basis for subsequent methylation sequencing. This enables the capture of DNA methylation information related to cancer from the blood, opening up a way for non-invasive cancer risk prediction, greatly reducing the difficulty and risk of sample acquisition, and improving the feasibility of this method in clinical applications. The plasma-free DNA samples are subjected to bisulfite conversion and PCR amplification to obtain methylation sequencing data. Bisulfite conversion can convert unmethylated cytosine into uracil, while methylated cytosine remains unchanged, thus achieving the distinction of DNA methylation status. PCR amplification further amplifies the target fragment, providing a sufficient amount of template for subsequent high-throughput sequencing, ensuring the accuracy and reliability of the methylation sequencing data, and laying a solid data foundation for accurate cancer risk prediction. The methylation sequencing data is subjected to high-throughput sequencing to obtain raw sequencing data. High-throughput sequencing technology can sequence a large number of DNA fragments in a short time, greatly improving the sequencing efficiency, enabling the rapid acquisition of a large amount of methylation site information, providing the possibility for a comprehensive analysis of DNA methylation patterns, and helping to more deeply explore methylation characteristics related to cancer. The raw sequencing data is subjected to quality control and data filtering to obtain DNA methylation sequencing data. This step can remove low-quality sequencing reads, adapter sequence contamination, and data that do not meet quality standards, ensuring that the finally obtained DNA methylation sequencing data has high quality and high accuracy, thereby improving the credibility of subsequent analysis, avoiding false predictions caused by data quality problems, and guaranteeing the scientificity and effectiveness of the entire method. The reference genome sequence data is obtained using the UCSC Genome Database and subjected to functional element annotation analysis to obtain reference genome annotation data containing genomic functional region information. The UCSC Genome Database provides comprehensive and authoritative genomic sequence information. By performing functional element annotation on the reference genome sequence data, the functional attributes of different regions in the genome, such as promoters, enhancers, transcription factor binding sites, etc., can be clarified, providing a key reference basis for the subsequent precise matching of DNA methylation sequencing data with genomic functional regions, enabling more accurate localization of the functional positions of methylation sites in the genome, and providing strong support for in-depth analysis of the relationship between methylation and gene regulation. According to the reference genome annotation data, the DNA methylation sequencing data is classified into promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data.This step realizes the precise division of DNA methylation sequencing data according to genomic functional regions, enabling subsequent specialized analysis and processing of methylation data in different functional regions. It fully considers the unique roles of different functional regions in gene regulation and cancer development, providing a more refined data basis for in-depth exploration of the mechanism of DNA methylation in cancer risk prediction, and helping to improve the accuracy and biological interpretability of cancer risk prediction.

[0021] Preferably, step S16 includes the following steps: Step S161: Perform alignment and positioning processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain the genomic coordinate data of the methylation sites; In the embodiment of the present invention, when performing alignment and positioning processing on the DNA methylation sequencing data according to the reference genome annotation data, first use an alignment tool such as BWA or Bowtie2 to align the methylation sequencing data with the reference genome. During the alignment process, each sequencing fragment is located at a specific position on the reference genome to generate genomic coordinate data. Common alignment parameters include setting an appropriate error rate (such as 0.05), the maximum number of matches (such as 1), and allowing small-scale insertions and deletions. The SAM format file generated during the alignment process contains the position information of the sequencing fragments and their corresponding coordinates on the genome. To ensure the accuracy of the alignment results, it is necessary to use Quality Control (QC) tools to evaluate the data quality, exclude low-quality alignment results, and retain high-quality and accurate alignment data. After alignment, the obtained genomic coordinate data is used for subsequent methylation analysis, specifically including the chromosomal position, start and end positions of each methylation site.

[0022] Step S162: Perform methylation level quantification processing on the genomic coordinate data to obtain methylation status matrix data; When the embodiments of the present invention perform methylation level quantification processing on genomic coordinate data, first, according to the obtained genomic coordinate data by alignment, the methylation status information is combined with its corresponding genomic position. The methylation level is usually quantified by calculating the methylation ratio of cytosine at a specific genomic position, and this process is estimated through the methylation information in each sequencing read obtained by high-throughput sequencing. Specifically, for each alignment position, the ratio of methylated and non-methylated cytosines at this position is statistically calculated. For example, by comparing the number of methylated C and unmethylated C in each sequencing fragment, the methylation ratio at this position is obtained. A methylation status matrix is generated through these ratio data. Each row in this matrix represents a specific genomic position, and each column represents the methylation status of different samples. Each element of the matrix is the methylation level of this position in a specific sample, usually represented by a decimal between 0 and 1, where 1 represents complete methylation and 0 represents complete unmethylation.

[0023] Step S163: Perform functional region annotation processing on the methylation status matrix data according to the functional element information in the reference genome annotation data, so as to obtain functional region annotation data; When the embodiments of the present invention perform functional region annotation processing on the methylation status matrix data according to the functional element information in the reference genome annotation data, first, the functional element information of the reference genome is obtained, which usually includes regions such as promoter regions, enhancer regions, transcription factor binding sites, exons, and introns. The functional element information usually comes from the UCSC Genome Browser or the Ensembl database. Using these annotation information, each row (i.e., each methylation site) in the methylation status matrix is associated with the corresponding genomic functional region. For example, through tools such as BEDTools, the coordinates of the methylation sites are compared with the functional regions of the reference genome, and these methylation sites are classified into the corresponding functional regions according to the positions of the functional regions. Through this method, a methylation status matrix with functional region annotation can be obtained, and the data of each row will be annotated as the corresponding functional region category, such as promoter region, enhancer region, or transcription factor binding site, etc. The output of this step is an updated matrix, which contains the information of the functional regions to which the methylation sites belong, providing a basis for subsequent feature classification.

[0024] Step S164: Perform feature classification processing on the functional region annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data respectively.

[0025] When the embodiment of the present invention performs feature classification processing on the functional region annotation data, first, according to the functional region annotation data obtained in step S163, each methylation site is classified according to the functional region to which it belongs. For example, all methylation sites labeled as promoter regions are classified as promoter region methylation data, all methylation sites labeled as enhancer regions are classified as enhancer region methylation data, and methylation sites labeled as transcription factor binding sites are classified as transcription factor binding site methylation data. At this time, programming tools such as the Pandas library in Python can be used to operate and classify the data, splitting the matrix by columns or rows and classifying according to the labels of the functional regions. After classification, the data for each functional region will be used as an independent data set for subsequent analysis. The purpose of this process is to separate different types of methylation data so as to more specifically analyze the roles of various functional regions in cancer occurrence. For example, the methylation data in the promoter region may be related to gene expression regulation, while the methylation data in the enhancer region and transcription factor binding sites may be closely related to gene activity regulation.

[0026] The present invention aligns and locates DNA methylation sequencing data based on reference genome annotation data, accurately obtaining the genomic coordinate data of methylation sites, which clarifies the specific positions of methylation sites in the genome for subsequent analysis, enabling the accurate correspondence of methylation information with genomic functional regions, providing key spatial localization information for in-depth exploration of the role of methylation in gene regulation, and ensuring the accuracy and pertinence of subsequent analysis. Quantitative analysis of the genomic coordinate data for methylation levels yields methylation status matrix data, which realizes the quantitative description of methylation sites, converts methylation information into a data form suitable for mathematical calculation and statistical analysis, provides a quantitative data basis for subsequent exploration of the relationship between methylation patterns and cancer risk, and enables more precise assessment of changes in methylation levels and their impact on gene expression and cancer development. Using the functional element information in the reference genome annotation data to label the methylation status matrix data with functional regions gives functional region annotation data, which effectively associates the methylation status with the functional regions of the genome, endows methylation data with biological function connotations rather than just being dot-like information on the genome, clearly shows the methylation status of different functional regions, provides clear guidance for subsequent methylation feature analysis of different functional regions, and helps to deeply understand the mechanism of action of methylation in the gene regulatory network. Classifying the features of the functional region annotation data respectively gives promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data, which realizes the refined classification of methylation data according to genomic functional regions, enabling subsequent methylation analysis to be carried out separately for key gene regulatory elements such as promoters, enhancers, and transcription factor binding sites, fully considering the unique roles and importance of different functional regions in gene expression regulation and cancer development, providing a more accurate and detailed data basis for in-depth exploration of the mechanism of action of DNA methylation in cancer risk prediction, and helping to improve the accuracy and biological interpretability of cancer risk prediction.

[0027] Preferably, step S2 includes the following steps: Step S21: Perform sequence fragment truncation processing according to the promoter region methylation data to obtain promoter region sequence data of a fixed length; When the embodiment of the present invention performs sequence fragment truncation processing according to the methylation data of the promoter region, it first determines the specific range of the promoter region, and these ranges can be obtained according to the reference genome annotation data. Generally speaking, the promoter region refers to 1,000 to 2,000 base pairs near the transcription start site (TSS) of a gene. These regions can be accurately obtained using a genome browser tool (such as the UCSC Genome Browser). On this basis, through programming methods (such as the Biopython library in Python), the DNA sequence data of each promoter region is truncated to ensure that the truncated sequence has a fixed length. For example, the length of the promoter region can be set to 1,000 base pairs. For each promoter sequence, 500 base pairs upstream and 500 base pairs downstream of it are truncated to ensure that the length of each fragment is the same. The purpose of this step is to convert the methylation data of the promoter region into sequence fragments of a unified length, which is convenient for subsequent analysis and processing.

[0028] Step S22: Perform dynamic time warping algorithm processing on the promoter region sequence data to obtain standardized sequence data that can be compared with the pre-obtained reference pattern library of the promoter region; When the embodiment of the present invention performs dynamic time warping (DTW) algorithm processing on the promoter region sequence data, it first needs to obtain a reference pattern library of the promoter region, which contains standard promoter region sequence data from multiple healthy individuals or known cancer patients. The DTW algorithm can compare and align sequences of different lengths or different change speeds by calculating the similarity between different sequences. In specific implementation, first, the promoter region sequences of each sample are preprocessed to ensure the unified format and data cleanliness of the sequences. Next, the DTW algorithm is used to compare the sample sequences with each reference sequence in the reference pattern library. The DTW algorithm will find the best alignment path between the sample sequence and the reference sequence by minimizing the cost function, and through normalization processing, convert the result into standardized sequence data. This standardized sequence data can eliminate the deviations caused by different methylation levels or gene sequence lengths between samples, enabling the sequences between different samples to be compared under the same standard, which is convenient for subsequent analysis.

[0029] Step S23: Calculate the sequence similarity with the promoter region of cancer-related genes according to the standardized sequence data to obtain a first feature vector containing the importance weights of the promoter regions of cancer-related genes; When calculating the sequence similarity of the promoter regions of cancer-related genes based on standardized sequence data in the embodiments of the present invention, first, according to the known list of cancer-related genes, the promoter regions of these genes are screened out. Cancer-related genes are usually obtained through genomic data, literature analysis, or public databases (such as TCGA). Then, the similarity between the standardized promoter region sequences and the promoter region sequences of these cancer-related genes is calculated. Commonly used similarity measurement methods include sequence alignment algorithms, such as the Smith-Waterman or Needleman-Wunsch algorithms. By calculating the similarity between the standardized promoter sequence and the promoter sequence of cancer-related genes, a similarity score is obtained. The promoter regions of genes with higher similarity scores are considered to have a stronger correlation with cancer occurrence. Based on these similarity scores, a first feature vector is constructed, which contains the similarity scores of the promoter regions of each cancer-related gene, reflecting the potential contribution of each promoter region to cancer.

[0030] Step S24: Perform performance regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity state; When the embodiments of the present invention perform performance regulation pattern recognition processing on the enhancer region methylation data, first, the data of the reference enhancer region needs to be obtained, usually from the functional annotation information in public databases such as the ENCODE project. Using these data, the correlation between the methylation state of the enhancer region and the gene expression level can be analyzed. Through multivariate statistical analysis methods (such as principal component analysis PCA, clustering analysis, etc.), the regulation patterns shown by different enhancer regions during cancer occurrence are identified. These regulation patterns can be represented by the classification of the enhancer activity state, such as the active state, the inhibitory state, or the neutral state. The distribution of each activity state will be converted into a second feature vector, indicating the activity state of each enhancer region and its potential regulatory role in cancer. The core of this processing step is to discover how the regulation patterns of enhancer regions affect the expression of cancer-related genes by comparing the methylation patterns of enhancer regions in different cancer samples.

[0031] Step S25: Perform binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability.

[0032] When the embodiments of the present invention perform the binding site conservation analysis on the methylation data of transcription factor binding sites, it is first necessary to identify the transcription factor binding sites in the genome, which are usually provided by experimental data or databases (such as JASPAR). Then, by analyzing the conservation of these binding sites in different species, the importance of each transcription factor binding site in cancer is evaluated. Specifically, in implementation, first analyze the methylation status of the transcription factor binding sites and calculate the methylation level of these sites in cancer samples. Then, use conservation analysis tools (such as PhyloP, PhastCons, etc.) to evaluate the conservation of each transcription factor binding site in different species. Combining the methylation level and conservation, the binding probability of each binding site is obtained, which represents the activity degree of the transcription factor binding site. Finally, the obtained third eigenvector contains the binding probability of each transcription factor binding site, reflecting the potential role of the transcription factor binding site in the process of cancer occurrence.

[0033] The present invention intercepts sequence fragments based on promoter region methylation data to obtain promoter region sequence data of a fixed length, which enables subsequent analysis to be carried out on a unified sequence length scale, is conducive to eliminating the interference of sequence length differences on the analysis results, ensures the accurate extraction and comparison of promoter region methylation characteristics, and provides a standardized data basis for in-depth exploration of the relationship between promoter region methylation and cancer risk. The dynamic time warping algorithm is used to process the promoter region sequence data to obtain standardized sequence data, enabling it to be effectively compared with a pre-acquired promoter region reference pattern library. The dynamic time warping algorithm can handle the inconsistencies in time or space between sequences, thus more accurately matching and comparing the promoter region sequence data of different samples, helping to discover the subtle differences in promoter region methylation patterns, further improving the recognition accuracy of promoter region methylation characteristics, and providing strong support for accurately evaluating the role of promoter regions in cancer risk prediction. Calculate the sequence similarity with the promoter regions of cancer-related genes based on the standardized sequence data to obtain a first feature vector containing the importance weights of the promoter regions of cancer-related genes. By calculating the similarity with the reference pattern library of the promoter regions of cancer-related genes, the similarity degree between the promoter region methylation pattern and the known cancer-related pattern can be quantified, thereby determining the importance weights of each promoter region in cancer risk prediction. This enables subsequent analysis to focus on those promoter regions highly related to cancer risk, improving the accuracy and pertinence of cancer risk prediction, and providing a key basis for accurately identifying promoter regions related to cancer risk. Identify the performance regulation pattern of enhancer region methylation data to obtain a second feature vector containing the enhancer activity state. This step can reveal the enhancer activity state reflected by the enhancer region methylation data. As an important gene regulation element, the change in the activity state of the enhancer is closely related to gene expression regulation. By identifying the performance regulation pattern of the enhancer region methylation data, the activity of the enhancer in different samples can be accurately grasped, providing important information for evaluating the role of the enhancer in the occurrence and development of cancer, further enriching the feature dimension of cancer risk prediction, and helping to more comprehensively understand the molecular mechanism of cancer. Conduct binding site conservation analysis on transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability. The binding site conservation analysis can evaluate the conservation degree of transcription factor binding sites in different samples, and then infer the probability of transcription factor binding to DNA. Transcription factors play a key role in gene expression regulation, and the methylation state of their binding sites directly affects the binding ability of transcription factors. By conducting conservation analysis on transcription factor binding site methylation data, the transcription factor binding probability can be accurately predicted, providing an important basis for in-depth study of the role of transcription factors in cancer risk prediction, further improving the feature system of cancer risk prediction, and helping to improve the accuracy and reliability of prediction.

[0034] Preferably, step S24 includes the following steps: Step S241: Perform sequence segmentation processing on the enhancer region methylation data to obtain enhancer fragment data including upstream and downstream regulatory regions; When the embodiment of the present invention performs sequence segmentation processing on the enhancer region methylation data, it is first necessary to accurately identify and extract the sequence data of the enhancer region. Usually, these regions can be obtained through enhancer annotation information in genomic databases (such as UCSC Genome Browser or ENCODE). Enhancer regions are generally located upstream or downstream of the promoter of a gene, and their functions are closely related to the regulation of gene expression. In implementation, first, by analyzing the methylation data of these enhancer regions, the upstream and downstream regulatory regions of these regions are intercepted. Usually, the lengths of the upstream and downstream can be set to 500 to 1,000 base pairs. Through programming tools, such as the Biopython library in Python, the sequences of each enhancer region and its upstream and downstream regulatory regions can be extracted and segmented to ensure the integrity and accuracy of the data. The enhancer fragment data obtained at this time will provide basic data for subsequent epigenetic state annotation and activity feature extraction.

[0035] Step S242: Perform epigenetic state annotation processing on the enhancer fragment data according to a preset histone modification marker database to obtain enhancer activity feature data; When the embodiment of the present invention performs epigenetic state annotation processing on the enhancer fragment data according to a preset histone modification marker database, it is first necessary to obtain relevant histone modification marker information. Common histone modification markers, such as H3K27ac, H3K4me3, etc., these markers reflect gene activity and epigenetic states. The histone modification marker data of enhancer regions in different cell types can be obtained by referring to public databases such as the ENCODE project or the Roadmap Epigenomics database. In specific operations, bioinformatics tools (such as bedtools, deepTools, etc.) are used to annotate the enhancer fragment data, and mark the possible histone modifications on each enhancer region. These histone modification markers can help identify the activity states of enhancers. For example, active enhancers usually have H3K27ac modifications, while repressed enhancers may have H3K9me3 modifications. Through this step, the enhancer fragment data will be given corresponding epigenetic state information and transformed into enhancer activity feature data, providing a basis for subsequent regulatory pattern analysis.

[0036] Step S243: Perform convolutional neural network processing on the enhancer activity feature data to obtain local epigenetic regulation pattern data; When the enhancer activity feature data is processed by a convolutional neural network (CNN) in an embodiment of the present invention, a convolutional neural network model needs to be constructed first to extract local epigenetic regulation patterns from the enhancer activity feature data. Specifically, during the operation, the enhancer activity feature data is converted into a two-dimensional matrix. The rows of the matrix represent different features, while the columns represent the methylation levels or histone modification information of each enhancer region. The convolutional neural network (CNN) extracts local features through convolutional layers, and activation functions (such as ReLU) stimulate highly active epigenetic patterns. The pooling layer is used to reduce the computational amount and strengthen key features. In this step, the convolutional kernel size of the convolutional neural network can be set to 10 base pairs, the stride is 1, and the number of convolutional kernels is 32 to extract epigenetic regulation features of different scales. By training the model, local regulation patterns can be identified from the epigenetic data of the enhancer region, and local epigenetic regulation pattern data is generated, which will provide input for subsequent deep learning feature extraction.

[0037] Step S244: Perform deep learning feature extraction processing based on the local epigenetic regulation pattern data to obtain enhancer region activity score data; When the embodiment of the present invention performs deep learning feature extraction processing based on the local epigenetic regulation pattern data, a deep neural network (DNN) model is used to further extract complex features of the enhancer region. In this step, the local epigenetic regulation pattern data is first input into the deep neural network. Usually, the model includes multiple fully connected layers (such as 4 to 6 layers), and the number of neurons in each layer can be set to 256 to 512 to capture deep non-linear relationships in the data. By training the neural network model, the parameters of the model are adjusted using an optimization algorithm (such as the Adam optimizer) to gradually reduce the error. During this process, regularization techniques (such as Dropout) can be used to prevent overfitting, and batch normalization technology is adopted to accelerate the training. Finally, through the deep learning feature extraction model, key feature information of the enhancer region can be extracted from the local epigenetic regulation pattern data, and the activity score of the enhancer region is calculated based on these features to generate enhancer region activity score data.

[0038] Step S245: Perform feature vectorization processing on the enhancer region activity score data to obtain a second feature vector containing the enhancer activity state.

[0039] When the enhancer region activity score data is subjected to feature vectorization processing in the embodiments of the present invention, it is first necessary to convert the obtained enhancer activity score data into a vector form for subsequent analysis and calculation. Usually, the enhancer activity score data is a multi-dimensional data containing multiple variables, and each variable represents the activity score of a different enhancer region. At this stage, feature vectorization techniques (such as principal component analysis PCA) are used to reduce the dimensionality of these activity score data, retain the most representative features, and reduce redundant information. Specifically, principal component analysis transforms the multi-dimensional data into a new coordinate system through a linear transformation, where the first few principal components retain the maximum variance of the data and can effectively describe the main features of the data. Through this step, the final obtained feature vector will contain the activity state information of the enhancer region, representing the potential regulatory role of each enhancer region in cancer development. This second feature vector serves as the input data for subsequent cancer risk assessment and can provide accurate feature support for subsequent multi-task learning.

[0040] The present invention performs sequence segmentation processing on enhancer region methylation data to obtain enhancer fragment data containing upstream and downstream regulatory regions, which helps to comprehensively capture the regulatory information of enhancers. Since the upstream and downstream regions may contain regulatory elements crucial for enhancer activity, it provides a more complete data basis for subsequent analysis, making the assessment of enhancer activity more accurate and comprehensive. Using a preset histone modification marker database to annotate the epigenetic status of the enhancer fragment data to obtain enhancer activity characteristic data. This step can reveal the active state of enhancers at the epigenetic level by comparing the enhancer fragments with known histone modification markers. Histone modification is closely related to gene expression regulation, and accurate epigenetic status annotation provides key biological information for subsequent activity assessment, helping to deeply understand the mechanism of action of enhancers in gene regulation. Using a convolutional neural network to process the enhancer activity characteristic data to obtain local epigenetic regulation pattern data. The convolutional neural network can automatically extract local features and patterns in the data, which helps to mine local epigenetic regulation patterns related to regulation from complex enhancer activity characteristic data, improving the accuracy and efficiency of identifying the enhancer activity state and laying a solid foundation for subsequent feature extraction and activity scoring. Based on the local epigenetic regulation pattern data, deep learning feature extraction is performed to obtain enhancer region activity scoring data. Deep learning can learn complex feature representations from a large amount of data. This step further extracts and analyzes features from the local epigenetic regulation pattern data through deep learning algorithms, enabling more accurate assessment of the enhancer activity state, providing more representative and discriminative features for cancer risk prediction, and helping to improve the accuracy and reliability of prediction. Feature vectorization processing is performed on the enhancer region activity scoring data to obtain a second feature vector containing the enhancer activity state. Feature vectorization can convert complex activity scoring data into a vector form convenient for subsequent analysis and processing, enabling the enhancer activity state to be incorporated into the entire cancer risk prediction model in a concise and efficient manner, facilitating integration and analysis with other feature vectors, further enhancing the performance and prediction ability of the model, and providing strong data support for deeply exploring the role of enhancers in cancer risk prediction.

[0041] Preferably, step S25 includes the following steps: Step S251: Perform sequence window sliding processing on the transcription factor binding site methylation data to obtain candidate sequence data containing potential binding sites; When the embodiments of the present invention perform sequence window sliding processing on the methylation data of transcription factor binding sites, it is first necessary to extract potential transcription factor binding sites in the DNA sequence. In this step, the sliding window technique is used to process the methylation data of transcription factor binding sites. Specifically, the DNA sequence is divided into multiple consecutive windows, and each window contains a fixed number of base pairs (for example, 100 base pairs). The size of the window can be adjusted according to experimental requirements. The sliding window slides step by step with a fixed step size (such as 5 base pairs) to analyze the methylation status of each window region. In this way, potential transcription factor binding site regions can be obtained from the whole genome range. The methylation data of each window region can be characterized by calculating the methylation level (such as methylation ratio) of this region. Finally, the candidate sequence data obtained after sliding window processing contains the sequence information of all potential transcription factor binding sites, providing basic data for subsequent sequence conservation analysis and affinity prediction.

[0042] Step S252: Perform sequence conservation analysis processing on the candidate sequence data according to a preset transcription factor binding site database, so as to obtain site conservation score data; When the embodiments of the present invention perform sequence conservation analysis processing on the candidate sequence data according to a preset transcription factor binding site database, it is first necessary to obtain the conservation data of transcription factor binding sites from a public database. Common databases include TRANSFAC, JASPAR, and ENCODE, etc. These databases provide the conserved binding site data of different transcription factors in multiple species. In this step, by comparing the candidate sequence data with the transcription factor binding site information in these databases, the conservation of each potential binding site is evaluated. Specifically, sequence alignment tools (such as BLAST, MUSCLE, etc.) can be used to calculate the similarity and conservation degree between the candidate sequence and the known transcription factor binding sites in the database. Generally, binding sites with higher conservation also have similar sequence characteristics in different species. Through this alignment analysis, the conservation score data of each candidate binding site can be obtained, and these data can help screen out transcription factor binding sites with higher conservation, providing a reliable basis for subsequent affinity analysis and activity prediction.

[0043] Step S253: Perform sequence specificity detection processing on the site conservation score data to obtain transcription factor binding affinity data, where the sequence specificity detection processing includes sequence pattern recognition, position weight matrix calculation, and thermodynamic parameter integration; When the embodiment of the present invention performs sequence-specific detection processing on the site conservation score data, it is first necessary to perform sequence pattern recognition on the candidate binding sites to determine whether the site conforms to the characteristics of transcription factor binding. For this purpose, the sequence patterns of known transcription factor binding sites (such as position weight matrices, PWM) can be used for alignment. The position weight matrix describes the base preference of the transcription factor binding site in the DNA sequence. By aligning the candidate sequence with the PWM of the known transcription factor, potential binding sites can be identified. Then, thermodynamic parameters are combined to evaluate the affinity of the binding site. For example, the binding free energy (ΔG) is calculated to judge the binding strength between the transcription factor and DNA. Specifically, by calculating thermodynamic parameters (such as binding free energy, entropy change, etc.), sites with higher binding affinity can be further screened out. Through sequence pattern recognition, position weight matrix calculation, and thermodynamic parameter integration, transcription factor binding affinity data are obtained, and these data provide key features for subsequent deep learning processing.

[0044] Step S254: Perform deep neural network processing based on the transcription factor binding affinity data to obtain binding site activity prediction data; When the embodiment of the present invention performs deep neural network processing based on the transcription factor binding affinity data, it is first necessary to input the obtained transcription factor binding affinity data into the deep neural network model for multi-level feature learning and prediction. The input layer of the deep neural network accepts the feature vector containing the transcription factor binding affinity data. The hidden layers in the network process these features through non-linear activation functions (such as ReLU, Sigmoid, etc.) to gradually extract the potential laws in the data. The output layer of the network is used to predict the activity state of the binding site, and a classification or regression model can be used for prediction. Specifically in operation, the model can include multiple hidden layers, and each layer contains a certain number of neurons (such as 256 or 512) to extract more complex features. During training, optimization algorithms (such as Adam, SGD, etc.) are used to adjust the parameters of the model to minimize the prediction error. Through this deep learning model, after the binding affinity data are processed by the network, binding site activity prediction data can be generated to evaluate the activity level of each transcription factor binding site, further helping to identify key regulatory sites.

[0045] Step S255: Perform feature integration processing on the binding site activity prediction data to obtain a third feature vector containing the transcription factor binding probability.

[0046] When the embodiment of the present invention performs feature integration processing on the binding site activity prediction data, it is first necessary to integrate the binding site activity prediction data output by the deep neural network with other relevant features (such as site conservation scores, epigenetic states, etc.). The integration method can use weighted average, principal component analysis (PCA), or other multi-feature fusion algorithms to combine data from different sources into a unified feature vector. Each feature is assigned a certain weight according to its importance in the prediction, and these weights can be automatically learned through the training process. Through this feature integration processing, a more comprehensive feature vector can be obtained, representing the activity state of each binding site. Finally, the data after feature integration will be transformed into a third feature vector containing the transcription factor binding probability, and these feature vectors provide strong data support for subsequent cancer risk assessment and regulatory mechanism analysis.

[0047] The present invention performs sequence window sliding processing on the methylation data of transcription factor binding sites to obtain candidate sequence data containing potential binding sites. This helps to comprehensively scan and identify possible transcription factor binding sites, expands the search scope, increases the possibility of discovering potential binding sites, and provides a richer data basis for subsequent analysis. Using a preset transcription factor binding site database to perform sequence conservation analysis on the candidate sequence data, site conservation score data is obtained. With the help of the known transcription factor binding site information in the database, the conservation of the candidate sequences is evaluated, thereby screening out binding sites with higher conservation during evolution. These sites are more likely to have important biological functions, improving the accuracy and reliability of subsequent analysis. Performing sequence specificity detection on the site conservation score data to obtain transcription factor binding affinity data, where sequence specificity detection includes sequence pattern recognition, position weight matrix calculation, and thermodynamic parameter integration. This step can further evaluate the specificity and affinity of transcription factor binding to DNA. By identifying specific sequence patterns and calculating the position weight matrix, the exact position and strength of the transcription factor binding site can be determined, and thermodynamic parameter integration helps to understand the energy changes during the binding process, thus more accurately predicting the binding behavior of transcription factors and providing more accurate data support for subsequent activity prediction. Performing deep neural network processing based on the transcription factor binding affinity data to obtain binding site activity prediction data. The deep neural network can learn deep features and patterns from the complex affinity data, thereby accurately predicting the activity of the binding site, improving the accuracy and reliability of the prediction, and providing strong data support for subsequent feature integration and cancer risk prediction. Performing feature integration processing on the binding site activity prediction data to obtain a third feature vector containing the transcription factor binding probability. Through feature integration, prediction data from different sources and types are integrated together to form a comprehensive feature vector. This not only improves the utilization efficiency of the data but also enhances the expression ability and discrimination of the features, making the evaluation of the transcription factor binding probability more comprehensive and accurate, and providing more representative and predictive feature data for the final cancer risk prediction.

[0048] Preferably, step S253 includes the following steps: Performing sequence pattern scanning processing on the site conservation score data, and performing pattern matching analysis according to a preset transcription factor binding motif database to obtain sequence specificity pattern data; Performing position weight matrix construction processing according to the sequence specificity pattern data, and performing frequency statistical analysis on each nucleotide position to obtain position-specific weight data; Performing DNA conformation analysis processing on the position-specific weight data based on DNA double helix structure parameters and base pairing stability to obtain sequence physicochemical feature data; Based on the sequence physicochemical characteristic data, perform the calculation and processing of the binding free energy based on the thermodynamic model, so as to obtain the transcription factor-DNA interaction energy data; Perform the conversion processing of the transcription factor-DNA interaction energy data based on the equilibrium thermodynamics equation, so as to obtain the transcription factor binding affinity data.

[0049] When the embodiment of the present invention performs sequence pattern scanning processing on the site conservation score data, it is first necessary to perform pattern matching analysis using a preset transcription factor binding motif database. The transcription factor binding motif database usually includes the DNA binding sequence characteristics of known transcription factors. Common databases include TRANSFAC, JASPAR, etc. In this step, the candidate sequence (i.e., the transcription factor binding site) is aligned with the known transcription factor binding motifs in these databases. The specific operation is to scan the entire candidate sequence through a sequence pattern scanning algorithm, such as the method based on the position weight matrix (PWM), to identify the regions that match the motifs in the database. The PWM provides the binding preference information of the transcription factor at each nucleotide position. The scanning process uses the sliding window technique. Each time, a sequence with a window size of 10-20 base pairs is scanned, and the scanned sequence is aligned with the position weight matrix in the database to calculate the matching degree of each sequence. The regions with high matching degrees are regarded as potential transcription factor binding sites. Finally, the obtained sequence-specific pattern data characterizes the characteristic information of each potential binding site, including the similarity between the sequence and the known transcription factor binding motif. When performing position weight matrix construction processing based on the sequence-specific pattern data, it is first necessary to count the occurrence frequency of each nucleotide position in all matching sequences. The specific operation is to extract the base composition of each position in the transcription factor binding site sequence according to the sequence-specific pattern data. By statistically analyzing the occurrence frequency of these bases at each position, the PWM of each transcription factor binding site can be constructed. The PWM is a matrix, where each row represents a nucleotide (A, C, G, T), each column represents a position in the sequence, and the values in the matrix are the occurrence frequencies or probabilities of the respective bases at that position. In order to further improve the accuracy, the maximum likelihood estimation (MLE) method is often used to optimize the weights in the matrix. The PWM constructed in this way can reflect the binding preference of each nucleotide at a specific position and provide important characteristic data for further binding affinity prediction. When performing DNA conformation analysis processing on the position-specific weight data based on the DNA double helix structure parameters and base pair stability, it is first necessary to combine the physical and chemical properties of DNA to further evaluate the stability of the binding site. The DNA double helix structure has specific physical properties, such as the binding force between different base pairs in the genome and the folding form of the DNA molecule in space. In this step, by simulating the DNA double helix conformation and considering factors such as hydrogen bonds and stacking forces between base pairs, the stability of the target DNA sequence is analyzed. Molecular dynamics simulation methods (such as GROMACS or AMBER) can be used to simulate the three-dimensional structure of the DNA sequence, evaluate the stacking force and hydrogen bond strength of base pairs, and further calculate the conformational stability of each transcription factor binding site.In addition, the physicochemical characteristics of a sequence can be evaluated by analyzing the stability of base pairing (such as the difference in stability between A-T and G-C pairings), thereby obtaining physicochemical characteristic data of the sequence. These data help determine the stability of transcription factor binding sites and provide a basis for calculating binding energy. When performing binding free energy calculation processing based on a thermodynamic model according to the sequence physicochemical characteristic data, it is first necessary to apply a thermodynamic model to evaluate the binding free energy between a transcription factor and DNA. The binding free energy is an important indicator for measuring the binding strength between DNA and a transcription factor, usually represented by ΔG. According to the interaction model between DNA and a transcription factor, the binding free energy can be determined by the following factors: the stability of base pairing, the change in DNA conformation, the interaction between the transcription factor and DNA, etc. In this step, a thermodynamic calculation model (such as a thermodynamic binding model) is used to calculate the free energy change of each transcription factor-DNA binding site. Specific methods include calculating the free energy difference between DNA and the transcription factor before and after binding, and obtaining the final free energy value by statistically analyzing the cooperative effect of transcription factor and DNA binding. Usually,... Binding sites with lower values indicate stronger binding strength between the transcription factor and DNA. The calculation result of this step is transcription factor-DNA interaction energy data, which characterizes the binding ability of each binding site. When performing conversion processing on the transcription factor-DNA interaction energy data based on the equilibrium thermodynamics equation, first use the equilibrium thermodynamics equation to convert the transcription factor-DNA interaction energy into binding affinity data. Under thermodynamic equilibrium, there is the following relationship between the binding affinity (Ka) and the binding free energy ( ): , where R is the gas constant and T is the temperature. By substituting the binding free energy into this formula, the binding affinity Ka can be calculated. The binding affinity reflects the stability of the binding between the transcription factor and DNA. The larger the Ka value, the stronger the binding affinity. In this step, through the conversion processing of the equilibrium thermodynamics equation, the transcription factor-DNA interaction energy data is converted into binding affinity data, thereby providing a basis for the quantitative evaluation of the binding ability of transcription factors. The finally obtained binding affinity data can be used to identify key transcription factor binding sites and provide important information for subsequent regulatory analysis of cancer-related genes.

[0050] The present invention performs sequence pattern scanning on the site conservation score data and combines it with a preset transcription factor binding motif database for pattern matching, which can accurately identify specific patterns in the sequence. This provides crucial sequence information for subsequent analysis and helps determine potential binding sites of transcription factors. By constructing a position weight matrix and performing frequency statistical analysis, the specific weights of each nucleotide position can be quantified. This step further refines the description of the binding site, making it more operable and valuable for analysis. Based on the DNA double helix structure parameters and base pairing stability, DNA conformation analysis can convert the position-specific weight data into physicochemical characteristic data of the sequence, thus providing a deeper perspective for understanding the interaction between transcription factors and DNA. The calculation of the binding free energy is based on these physicochemical characteristic data, and the interaction energy between transcription factors and DNA is accurately evaluated through a thermodynamic model. This step provides important energy parameters for predicting the binding behavior of transcription factors. Finally, by performing transformation processing on the binding free energy data using the equilibrium thermodynamic equation, transcription factor binding affinity data is obtained. This not only provides a direct basis for evaluating the binding probability of transcription factors but also provides key data support for in-depth understanding of the mechanism of action of transcription factors in gene regulation, thus significantly improving the accuracy and reliability of the entire method in predicting the binding behavior of transcription factors.

[0051] Preferably, step S3 includes the following steps: Step S31: Obtain a gene regulatory network database; perform topological structure analysis on the gene regulatory network database to obtain regulatory relationship data between functional regions; When the embodiment of the present invention obtains a gene regulatory network database, it is first necessary to obtain information related to gene regulation from public databases, including but not limited to ENCODE, Gene Ontology (GO), KEGG (Kyoto Encyclopedia of Genes and Genomes), etc. The obtained databases contain information related to genes, transcription factors, miRNAs, epigenetic markers, etc. involved in gene regulation, including specific data on regulatory elements, regulatory relationships, and their interactions. Next, topological structure analysis is performed on these gene regulatory network data. The analysis process includes constructing a network graph, where nodes represent genes or functional regions, and edges represent regulatory relationships (such as the regulatory role of transcription factors) between genes or functional regions. Topological structure analysis uses methods in graph theory, such as calculating the degree distribution, clustering coefficient, path length, etc. of the network. Through these indicators, the regulatory strength and regulatory mode between different functional regions can be evaluated, and thus regulatory relationship data between functional regions can be obtained. These data will provide a basis for subsequent research on gene regulatory mechanisms.

[0052] Step S32: Perform topological structure analysis on the gene regulatory network database to obtain the regulatory relationship data between functional regions; When the embodiment of the present invention performs topological structure analysis on the gene regulatory network database, algorithms such as depth-first search (DFS) and breadth-first search (BFS) in graph theory are further used to analyze the relationships and hierarchical structures between nodes in the gene regulatory network. By calculating the topological properties of the network, such as the degree of nodes, the shortest path, and network connectivity, it can be revealed which genes or functional regions are key nodes in the network, and which genes play a role by regulating other genes or functional regions. The specific operation is to first model the nodes and edges in the network to form a directed graph (the regulatory direction from transcription factors to target genes). Then, algorithms (such as the shortest path algorithm or the minimum spanning tree algorithm) are applied to reveal the core modules and regulatory paths in the gene regulatory network. These analysis results will help identify key functional regions and their interaction patterns, and finally obtain the regulatory relationship data between functional regions. These data will provide a basis for subsequent feature analysis and pathway enrichment analysis.

[0053] Step S33: Perform feature correlation analysis on the first feature vector, the second feature vector, and the third feature vector according to the regulatory relationship data to obtain feature interaction intensity data; According to the regulatory relationship data obtained in steps S31 and S32, the embodiment of the present invention combines the foregoing first feature vector, second feature vector, and third feature vector to perform feature correlation analysis on them. The purpose of feature correlation analysis is to evaluate the interaction and influence relationships between various features, so as to identify important interaction patterns. Commonly used methods include Pearson correlation coefficient, Spearman rank correlation coefficient, mutual information analysis, etc. The specific operation is to combine the first, second, and third feature vectors with the regulatory relationship data to construct a multi-dimensional feature vector matrix, where each dimension represents the methylation data, activity data, and binding site data of different functional regions. Then, by calculating the correlation between different feature vectors, the possible interaction intensity between functional regions is identified. Through correlation analysis, an interaction intensity matrix can be obtained, which can quantitatively represent the interaction intensity of features between different functional regions and provide important information for subsequent pathway enrichment analysis.

[0054] Step S34: Perform gene pathway enrichment analysis on the feature interaction intensity data to obtain pathway association probability data; When performing gene pathway enrichment analysis on the feature interaction intensity data in the embodiments of the present invention, it is first necessary to use known gene pathway databases, such as KEGG, Reactome or WikiPathways, for enrichment analysis. The purpose of enrichment analysis is to find out which gene pathways or biological processes are highly correlated with the obtained feature interaction intensity data. The specific operation is to map the interaction intensity data to the gene sets in these gene pathway databases, and use statistical methods (such as hypergeometric distribution, Fisher’s Exact Test, GSEA) to evaluate whether the genes in each gene pathway are significantly enriched in the interaction intensity matrix. Through this analysis, pathways related to cancer, cell division, apoptosis, etc. can be identified, and the enrichment probability of each pathway can be given, and finally pathway association probability data can be obtained. These data can reflect the roles of different gene pathways in feature interaction, providing an important basis for subsequent network analysis and model construction.

[0055] Step S35: Construct a graph attention network model according to the pathway association probability data, and perform network node mapping processing on the feature vectors of the functional regions, so as to obtain feature network representation data; In the embodiments of the present invention, according to the pathway association probability data obtained in step S34, a graph attention network (GAT) model is constructed to perform network node mapping processing on the feature vectors. The graph attention network is a neural network model based on graph convolution, which can effectively process graph-structured data and dynamically adjust the weights of each node according to the relationships between the nodes in the graph. The specific operation is to use the feature vectors of the functional regions as the nodes in the graph and the pathway association probability data as the weights of the edges to construct a weighted graph model. Then, through the graph attention mechanism, the importance of each node in the graph is calculated through the self-attention mechanism, and the information transfer weights between the nodes are adjusted according to the adjacency relationship of the nodes. During this process, the feature vector of each node will be updated according to the information of other nodes in the graph, and finally the feature representation data after network node mapping is obtained. These feature network representation data reflect the relative importance of each functional region in the gene regulatory network, providing rich feature information for subsequent multi-level feature extraction.

[0056] Step S36: Perform hierarchical feature extraction processing on the feature network representation data, so as to obtain multi-level feature data with pathway correspondence relationships.

[0057] When the embodiment of the present invention performs hierarchical feature extraction processing on the feature network representation data, it is first necessary to apply multi-level feature extraction methods, such as deep graph convolutional network (GCN), graph autoencoder (GAE), etc. These methods can extract the high-order feature relationships between nodes from the graph and perform hierarchical processing on the data according to the hierarchy of the network structure. The specific operation is to gradually extract the high-level semantic information of the functional region features through multiple graph convolutional layers. In each layer, the feature vector of the node is updated by combining the features of its adjacent nodes, and finally a comprehensive feature representation containing all hierarchical information is formed. These hierarchical features can reflect the deeper interaction patterns between functional regions. Finally, the obtained multi-level feature data will have a corresponding relationship with the pathways and can be used for further biological analysis or the construction of disease-related prediction models.

[0058] The present invention obtains the gene regulatory network database and conducts topological structure analysis to clarify the regulatory relationships between functional regions, providing a basic framework for subsequent analysis and helping to understand the gene regulatory mechanism. Using the regulatory relationship data to perform feature correlation analysis on the feature vectors to obtain feature interaction intensity data, this step reveals the interaction between different feature vectors and provides an important basis for subsequent pathway analysis. Performing gene pathway enrichment analysis on the feature interaction intensity data to obtain pathway association probability data, which helps to identify the biological pathways related to cancer risk and provides key information for subsequent cancer risk prediction. According to the pathway association probability data, a graph attention network model is constructed and the functional region feature vectors are mapped to network nodes to obtain the feature network representation data. This step can better capture the complex relationships between features through the graph attention network model and improve the prediction ability of the model. Performing hierarchical feature extraction on the feature network representation data to obtain multi-level feature data with a corresponding relationship with the pathways, this step further refines the feature data, making it more representative and discriminative, and providing strong support for the final cancer risk prediction.

[0059] Preferably, step S4 includes the following steps: Step S41: Perform pathway node mapping processing on the multi-level feature data according to the pre-obtained prior knowledge of biological pathways, so as to obtain pathway-specific association data; In the embodiments of the present invention, it is necessary to perform pathway node mapping processing on multi-level feature data according to pre-acquired prior knowledge of biological pathways. The prior knowledge of biological pathways can be sourced from public databases such as KEGG, Reactome, BioCarta, etc., which provide rich gene regulation pathway information. In actual operation, first, by matching the multi-level feature data, each functional region or gene is mapped to the corresponding pathway. This process requires using the gene set information in the pathway database to map each gene or functional region in the multi-level feature data to its corresponding biological pathway node. The result of the mapping process is the performance of each gene or functional region in different biological pathways, and thus pathway-specific association data can be obtained. This data reflects the regulatory role of the functional region in a specific pathway and can provide a basis for subsequent weight calculation and pathway activity assessment.

[0060] Step S42: Perform weight calculation processing based on the attention mechanism on the pathway-specific association data, and constrain the attention scores based on the pathway regulatory relationship, so as to obtain pathway interaction weight data; The pathway-specific association data in the embodiments of the present invention undergoes weight calculation processing based on the attention mechanism. The specific operation is to first assign a preliminary weight to each pathway node, and based on the relationship between the nodes in the network, use the self-attention mechanism to calculate the attention score of each node. During this process, by analyzing the influence of each node on other nodes, the weight of the node is updated, and the weight of the pathway node with an important regulatory role is enhanced. On this basis, in order to combine the biological regulatory relationship, the attention scores are further constrained according to the regulatory structure of the biological pathway. The constraint process is to model the known gene regulatory relationships in the pathway, for example, by the directionality of the edges or the existing transcription factor regulatory information, to dynamically adjust the attention weights, so that the nodes consistent with the pathway regulatory relationship have higher weights. Finally, through this process, pathway interaction weight data is obtained, which represents the interaction strength between each node in the pathway.

[0061] Step S43: Perform multi-level feature fusion processing based on weighted summation according to the pathway interaction weight data, so as to obtain pathway activity score data; In the embodiment of the present invention, multi-level feature fusion processing based on weighted summation is performed on the path interaction weight data obtained according to step S42. The purpose of weighted summation is to synthesize feature data at multiple levels and highlight key functional regions and gene information. The specific operation is to first perform weighted summation on the feature vectors of each path node, where the weight of each node is determined by the attention score obtained in step S42. The process of weighted summation is to fuse feature vectors at different levels according to the interaction relationship and weight distribution between nodes. At this time, the feature vector of each node not only contains its own information but also combines the information of surrounding nodes, and a new fused feature vector is obtained through weighted summation. Through this fusion process, path activity score data can be obtained, which reflects the activity level of each path in the overall gene regulatory network and can provide a basis for subsequent risk prediction and path analysis.

[0062] Step S44: Perform dual-task objective function construction processing on the path activity score data, set a risk prediction loss function and a path interpretation loss function, so as to obtain multi-task optimization target data; In the embodiment of the present invention, dual-task objective function construction processing is performed on the path activity score data, and a risk prediction loss function and a path interpretation loss function are set. The risk prediction loss function is used to evaluate the accuracy of the model when performing risk prediction. Commonly used loss functions include the cross-entropy loss function, which is used to measure the difference between the prediction result and the true label in a classification task. The path interpretation loss function is used to measure the rationality of the model when interpreting the path regulation mechanism. Commonly used loss functions can be the mean square error (MSE) or a custom path-specific interpretation loss. The design of the dual-task objective function enables the model to learn two objectives of risk prediction and path interpretation simultaneously, and this process is achieved by weighted summation of the loss functions. In a specific application, weight parameters are used to adjust the relative importance of the two tasks to ensure that the model can balance and optimize the two tasks to obtain multi-task optimization target data.

[0063] Step S45: Construct a shared feature extraction network according to the multi-task optimization target data, and perform representation learning processing on the path activity features through a deep neural network, so as to obtain shared feature representation data; Based on the multi-task optimization target data obtained in step S44, an embodiment of the present invention constructs a shared feature extraction network. The shared feature extraction network is a deep neural network used to extract common feature representations from input data. This network usually consists of multiple convolutional layers and fully connected layers, aiming to extract general high-level features from gene regulation data. Specifically, first, a shared feature extraction module is used to perform feature learning on all input gene regulation data to extract shared features representing pathway activity and gene interaction. Then, these shared feature representations are passed to subsequent task-specific networks. The purpose of the shared feature extraction network is to enable two tasks to utilize the same feature learning module, reduce the consumption of computing resources, and at the same time extract features that can serve the multi-task optimization target.

[0064] Step S46: Perform task-specific network construction processing on the shared feature representation data, and respectively design a classification network for risk prediction and a regression network for pathway interpretation, so as to obtain dual-task network structure data; An embodiment of the present invention constructs task-specific networks for the risk prediction task and the pathway interpretation task according to the shared feature representation data obtained in step S45. The specific operation is to design a dedicated network structure for each task. The risk prediction task is modeled using a classification network, and the last layer of the classification network is usually a softmax layer, which outputs the probability of each risk category. The pathway interpretation task is modeled using a regression network, and the output of the regression network is the regression value of the pathway activity score or related biological interpretation. The task-specific network ensures that the two tasks can be optimized based on the same feature space by sharing some feature extraction modules, and then obtains dual-task network structure data. This network structure can complete multiple tasks simultaneously, further improving the comprehensive prediction ability and interpretation ability of the model.

[0065] Step S47: Perform parameter initialization processing according to the dual-task network structure data, and perform iterative optimization processing based on the backpropagation algorithm, and coordinate the dual-task learning process through a dynamic weight balancing strategy, so as to obtain model training data.

[0066] In the embodiments of the present invention, parameter initialization and iterative optimization processing are performed based on the data of the dual-task network structure. First, it is necessary to use a conventional initialization method, such as Xavier or He initialization, to initialize the parameters of the neural network. After initialization, the network parameters are optimized through the backpropagation algorithm. The backpropagation algorithm calculates the error of each neuron and updates the parameters according to the error through the gradient descent algorithm. In order to balance the optimization processes of the two tasks, a dynamic weight balancing strategy is introduced. This strategy dynamically adjusts the weights between tasks according to the magnitudes of the loss function values of each task, ensuring that the two tasks are balanced during the optimization process and preventing one task from being over-optimized and affecting the performance of the other task. Through multiple iterations of optimization, the trained model parameters are finally obtained, and the model training data is obtained.

[0067] Step S48: Perform risk probability calculation processing and pathway importance analysis processing respectively based on the model training data, so as to obtain risk distribution mapping data and pathway-specific analysis data.

[0068] In the embodiments of the present invention, risk probability calculation and pathway importance analysis processing are performed respectively based on the trained model parameters. First, the model is used to calculate the risk probability. For the input gene regulation data, the model will output the risk prediction probability of each sample. According to the prediction results, the samples can be divided into different risk levels to generate risk distribution mapping data. Then, pathway importance analysis processing is performed. By calculating the importance of the pathway in the network, the pathways that play a key role in risk prediction are identified. The analysis results can be obtained by interpreting the attention scores or feature importance scores in the model to obtain pathway-specific analysis data, which helps to further understand the contributions of each pathway to specific diseases or biological processes.

[0069] The present invention performs pathway node mapping processing on multi-level feature data by utilizing pre-acquired prior knowledge of biological pathways to obtain pathway-specific association data. This step combines multi-level feature data with biological pathway information, endowing the data with stronger biological significance and interpretability, and providing richer background information for subsequent analysis. Weight calculation processing based on the attention mechanism is performed on the pathway-specific association data, and the attention scores are constrained based on pathway regulatory relationships to obtain pathway interaction weight data. Through the attention mechanism, important pathway interaction relationships can be highlighted, and at the same time, the attention scores are constrained using pathway regulatory relationships, ensuring the rationality and accuracy of weight calculation, and further enhancing the model's ability to capture pathway interactions. Multi-level feature fusion processing based on weighted summation is performed according to the pathway interaction weight data to obtain pathway activity score data. This step fuses feature data at different levels, comprehensively considering the interaction effects of each pathway through weighted summation to obtain a more representative pathway activity score, providing a key quantitative indicator for subsequent risk prediction and pathway interpretation. Dual-task objective function construction processing is performed on the pathway activity score data, setting a risk prediction loss function and a pathway interpretation loss function to obtain multi-task optimization target data. By constructing a dual-task objective function, the two tasks of risk prediction and pathway interpretation are combined, enabling the model to simultaneously consider the requirements of these two tasks during the optimization process, improving the model's comprehensive performance and generalization ability. A shared feature extraction network is constructed according to the multi-task optimization target data, and representation learning processing of pathway activity features is performed through a deep neural network to obtain shared feature representation data. The shared feature extraction network can extract more representative and discriminative feature representations from pathway activity features, providing strong support for subsequent dual-task learning and further enhancing the model's prediction ability and interpretation ability. Task-specific network construction processing is performed on the shared feature representation data, respectively designing a classification network for risk prediction and a regression network for pathway interpretation to obtain dual-task network structure data. By designing specific network structures for the two tasks respectively, the requirements of each task can be better met, improving the accuracy and reliability of the model in both risk prediction and pathway interpretation. Parameter initialization processing is performed according to the dual-task network structure data, and iterative optimization processing based on the backpropagation algorithm is carried out. The dual-task learning process is coordinated through a dynamic weight balancing strategy to obtain model training data. This step enables the model to gradually learn the optimal parameters through parameter initialization and iterative optimization of the backpropagation algorithm, and at the same time, the dynamic weight balancing strategy is used to coordinate the dual-task learning process, ensuring the balanced development of the model on the two tasks and finally obtaining high-quality model training data.Perform risk probability calculation processing and pathway importance analysis processing respectively based on the model training data to obtain risk distribution mapping data and pathway-specific parsing data. This step uses the trained model to calculate the risk probability and analyze the pathway importance of the data. The obtained risk distribution mapping data can intuitively display the risk situation of the samples, while the pathway-specific parsing data can reveal the pathway information related to the risk, providing the final results and basis for cancer risk prediction and pathway interpretation.

[0070] The present invention also provides a deep learning cancer risk prediction system based on methylation sequences for performing the above-mentioned deep learning cancer risk prediction method based on methylation sequences. The deep learning cancer risk prediction system based on methylation sequences includes: A data acquisition and functional region classification module for acquiring DNA methylation sequencing data and reference genome annotation data in the blood samples of cancer patients; performing functional region classification processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data; A feature extraction and analysis module for performing dynamic sequence similarity calculation processing on the promoter region methylation data to obtain a first feature vector containing the importance weights of the promoter regions of cancer-related genes; performing performance regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity status; performing binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability; A biological pathway relevance analysis module for performing biological pathway relevance analysis processing on the first feature vector, the second feature vector, and the third feature vector according to a preset gene regulatory network topology structure to obtain multi-level feature data with pathway correspondence relationships; A multi-task learning and risk prediction module for performing an attention mechanism processing with biological pathway constraints on the multi-level feature data to obtain pathway activity score data; performing multi-task learning processing based on cancer risk prediction and key regulatory pathway interpretation according to the pathway activity score data to obtain risk distribution mapping data and pathway-specific parsing data.

Claims

1. A deep learning cancer risk prediction method based on methylation sequences, characterized in that: The following steps are involved: Step S1: obtaining DNA methylation sequencing data and reference genome annotation data from a blood sample of a cancer patient; performing functional region classification processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data, and transcription factor binding site methylation data; Step S2: Perform dynamic sequence similarity calculation processing on the promoter region methylation data to obtain a first feature vector containing the importance weight of the cancer-related gene promoter region; perform expression regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity state; perform binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability; Step S3: performing biological pathway correlation analysis on the first feature vector, the second feature vector and the third feature vector according to a preset gene regulatory network topology structure to obtain multi-level feature data with pathway correspondence; Step S4: Process the multi-level feature data with a biological pathway-constrained attention mechanism to obtain pathway activity score data; perform multi-task learning processing based on cancer risk prediction and key regulatory pathway interpretation on the pathway activity score data to obtain risk distribution mapping data and pathway-specific analysis data.

2. The deep learning cancer risk prediction method based on methylation sequences according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining a peripheral blood sample from a cancer patient, and performing plasma separation on the peripheral blood sample to obtain a plasma free DNA sample; Step S12: performing bisulfite conversion treatment on the plasma free DNA sample and performing PCR amplification treatment to obtain methylation sequencing data; Step S13: performing high-throughput sequencing processing on the methylation sequencing data to obtain original sequencing data; Step S14: performing quality control and data filtering on the original sequencing data to obtain DNA methylation sequencing data; Step S15: using the UCSC genome database to obtain reference genome sequence data, and performing functional element annotation analysis on the reference genome sequence data, thereby obtaining reference genome annotation data containing genome functional region information; Step S16: Perform functional region classification processing on the DNA methylation sequencing data according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data and transcription factor binding site methylation data.

3. The deep learning cancer risk prediction method based on methylation sequences according to claim 2, characterized in that: Step S16 includes the following steps: Step S161: performing alignment and positioning processing on the DNA methylation sequencing data according to the reference genome annotation data, thereby obtaining the genome coordinate data of the methylation site; Step S162: performing methylation level quantitative processing on the genome coordinate data to obtain methylation state matrix data; Step S163: performing functional region annotation processing on the methylation state matrix data according to the functional element information in the reference genome annotation data, thereby obtaining functional region annotation data; Step S164: performing feature classification processing on the functional region annotation data to obtain promoter region methylation data, enhancer region methylation data and transcription factor binding site methylation data respectively.

4. The deep learning cancer risk prediction method based on methylation sequences according to claim 3, characterized in that: Step S2 includes the following steps: Step S21: performing sequence segment interception processing according to the promoter region methylation data to obtain promoter region sequence data of fixed length; Step S22: Processing the promoter region sequence data using a dynamic time warping algorithm to obtain standardized sequence data that can be compared with a pre-acquired promoter region reference pattern library; Step S23: calculating the sequence similarity with the cancer-related gene promoter region according to the standardized sequence data, thereby obtaining a first feature vector containing the importance weight of the cancer-related gene promoter region; Step S24: performing expression regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity state; Step S25: performing binding site conservation analysis on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability.

5. The deep learning cancer risk prediction method based on methylation sequences according to claim 4, characterized in that: Step S24 includes the following steps: Step S241: performing sequence segmentation processing on the enhancer region methylation data, thereby obtaining enhancer segment data including upstream and downstream regulatory regions; Step S242: performing epigenetic annotation processing on the enhancer fragment data according to a preset histone modification mark database, thereby obtaining enhancer activity characteristic data; Step S243: performing convolutional neural network processing on the enhancer activity feature data to obtain local epigenetic regulation pattern data; Step S244: performing deep learning feature extraction processing according to the local epigenetic regulation pattern data, thereby obtaining enhancer region activity score data; Step S245: performing feature vector quantization processing on the enhancer region activity score data, thereby obtaining a second feature vector containing the enhancer activity state.

6. The deep learning cancer risk prediction method based on methylation sequences according to claim 5, characterized in that: Step S25 includes the following steps: Step S251: performing sequence window sliding processing on the transcription factor binding site methylation data, thereby obtaining candidate sequence data containing potential binding sites; Step S252: performing sequence conservation analysis on the candidate sequence data according to a preset transcription factor binding site database, thereby obtaining site conservation score data; Step S253: performing sequence-specific detection processing on the site conservation score data to obtain transcription factor binding affinity data, wherein the sequence-specific detection includes sequence pattern recognition, position weight matrix calculation, and thermodynamic parameter integration; Step S254: performing deep neural network processing on the transcription factor binding affinity data to obtain binding site activity prediction data; Step S255: performing feature integration processing on the binding site activity prediction data to obtain a third feature vector containing the transcription factor binding probability.

7. The deep learning cancer risk prediction method based on methylation sequences according to claim 6, characterized in that: Step S253 includes the following steps: The site conservation score data is subjected to sequence pattern scanning and pattern matching analysis based on a preset transcription factor binding motif database to obtain sequence-specific pattern data; A position weight matrix is ​​constructed based on the sequence-specific pattern data, and a frequency statistical analysis is performed on each nucleotide position to obtain position-specific weight data; The position-specific weight data are processed by DNA conformation analysis based on DNA double helix structure parameters and base pairing stability to obtain sequence physicochemical characteristic data; The binding free energy calculation based on the thermodynamic model is performed according to the sequence physicochemical characteristic data to obtain the transcription factor-DNA interaction energy data; The transcription factor-DNA interaction energy data is converted based on the equilibrium thermodynamic equation to obtain the transcription factor binding affinity data.

8. The deep learning cancer risk prediction method based on methylation sequences according to claim 7, characterized in that: Step S3 includes the following steps: Step S31: acquiring a gene regulatory network database; performing topological structure analysis on the gene regulatory network database to obtain regulatory relationship data between functional regions; Step S32: performing topological structure analysis on the gene regulatory network database to obtain regulatory relationship data between functional regions; Step S33: performing feature correlation analysis on the first feature vector, the second feature vector and the third feature vector according to the control relationship data, so as to obtain feature interaction intensity data; Step S34: performing gene pathway enrichment analysis on the feature interaction intensity data to obtain pathway association probability data; Step S35: constructing a graph attention network model according to the pathway association probability data, and performing network node mapping processing on the functional area feature vectors to obtain feature network representation data; Step S36: performing hierarchical feature extraction processing on the feature network representation data, thereby obtaining multi-level feature data with pathway correspondence.

9. The deep learning cancer risk prediction method based on methylation sequences according to claim 8, characterized in that: Step S4 includes the following steps: Step S41: performing pathway node mapping processing on the multi-level feature data according to the pre-acquired biological pathway prior knowledge, thereby obtaining pathway-specific association data; Step S42: performing weight calculation processing based on the attention mechanism on the pathway-specific association data, and constraining the attention score based on the pathway regulation relationship, so as to obtain pathway interaction weight data; Step S43: performing multi-level feature fusion processing based on weighted summation according to the pathway interaction weight data, thereby obtaining pathway activity score data; Step S44: constructing a dual-task objective function for the pathway activity score data, setting a risk prediction loss function and a pathway interpretation loss function, thereby obtaining multi-task optimization target data; Step S45: constructing a shared feature extraction network according to the multi-task optimization target data, and performing representation learning processing on the pathway activity features through a deep neural network, thereby obtaining shared feature representation data; Step S46: performing task-specific network construction processing on the shared feature representation data, respectively designing a classification network for risk prediction and a regression network for pathway interpretation, thereby obtaining dual-task network structure data; Step S47: Perform parameter initialization processing according to the dual-task network structure data, and perform iterative optimization processing based on the back propagation algorithm, coordinate the dual-task learning process through a dynamic weight balancing strategy, and thus obtain model training data. Step S48: performing risk probability calculation processing and pathway importance analysis processing respectively according to the model training data, thereby obtaining risk distribution mapping data and pathway specific analysis data.

10. A deep learning cancer risk prediction system based on methylation sequences, characterized in that: Used to execute the deep learning cancer risk prediction method based on methylation sequences as claimed in claim 1, the deep learning cancer risk prediction system based on methylation sequences comprises: The data acquisition and functional region classification module is used to obtain DNA methylation sequencing data and reference genome annotation data in blood samples of cancer patients; the DNA methylation sequencing data is subjected to functional region classification processing according to the reference genome annotation data to obtain promoter region methylation data, enhancer region methylation data and transcription factor binding site methylation data; The feature extraction and analysis module is used to perform dynamic sequence similarity calculation processing on the promoter region methylation data to obtain a first feature vector containing the importance weight of the cancer-related gene promoter region; perform expression regulation pattern recognition processing on the enhancer region methylation data to obtain a second feature vector containing the enhancer activity state; perform binding site conservation analysis processing on the transcription factor binding site methylation data to obtain a third feature vector containing the transcription factor binding probability; A biological pathway correlation analysis module, used to perform biological pathway correlation analysis on the first feature vector, the second feature vector and the third feature vector according to a preset gene regulatory network topology structure, to obtain multi-level feature data with pathway correspondence; The multi-task learning and risk prediction module is used to process the multi-level feature data with the attention mechanism of biological pathway constraints to obtain pathway activity score data; multi-task learning processing based on cancer risk prediction and key regulatory pathway interpretation is performed on the pathway activity score data to obtain risk distribution mapping data and pathway-specific analysis data.

Images