A tumor screening system based on circulating free DNA multi-dimensional feature fusion and application thereof

Abstract

The application discloses a tumor screening system based on circulating free DNA multi-dimensional feature fusion and application thereof, and aims to solve the problem of lack of unified feature expression and fusion analysis framework for multi-omics data in existing tumor screening. The application integrates multi-dimensional features of methylation spectrum, mutation information and fragmentation mode of circulating free DNA, constructs a unified feature expression and fusion analysis model, and develops an efficient machine learning risk prediction method based on the model, so that early and accurate screening of tumors can be realized, and reliable auxiliary decision support for clinical tumor diagnosis is provided. Through deep mining and fusion of multi-dimensional features, the application effectively improves the sensitivity and specificity of tumor screening, is especially suitable for early tumor risk assessment of asymptomatic people, and has a wide clinical application prospect.

Description

Technical Field

[0001] This invention relates to the fields of biomedical detection, molecular diagnostics and artificial intelligence-assisted analysis, and in particular to a multi-omics feature fusion tumor screening system based on circulating cell-free DNA and its application, which is especially suitable for non-invasive early tumor screening and risk assessment. Background Technology

[0002] Cancer is a serious disease that severely threatens human health, and early screening is crucial for improving patient survival rates. Existing screening methods include imaging examinations, serum biomarker detection, and histopathological examinations, but they generally suffer from insufficient sensitivity, high invasiveness, or unsuitability for large-scale population screening.

[0003] In recent years, liquid biopsy technology has gradually become an important development direction for cancer screening. Among them, circulating cell-free DNA (cfDNA) is a DNA fragment released during cell apoptosis or necrosis, which can reflect tumor-related gene mutations, epigenetic changes, and chromatin structure information.

[0004] Current cfDNA detection technologies mainly focus on the following three pathways:

[0005] (1) Detection techniques based on a single mutation, such as high-throughput sequencing (NGS) or digital PCR (dPCR).

[0006] (2) Detection technologies based on DNA methylation, such as multi-site methylation detection panels;

[0007] (3) Risk prediction method based on machine learning model.

[0008] However, the aforementioned existing technologies still have the following problems:

[0009] (1) Single molecular features (such as mutation or methylation) are difficult to balance detection sensitivity and specificity;

[0010] (2) Data from different omics systems usually come from different samples or different detection systems, resulting in poor data consistency;

[0011] (3) The content of cfDNA in body fluids such as urine is low, the detection signal is weak, and the stability is insufficient;

[0012] (4) Multi-omics data lack a unified feature representation and fusion analysis framework.

[0013] Therefore, existing technologies have not yet established a systematic technical system for the unified extraction, structured expression, and fusion modeling of multidimensional molecular features based on the same cfDNA sample, making it difficult to achieve a balance between high sensitivity and high specificity. Summary of the Invention

[0014] To address the shortcomings of the aforementioned technologies, this invention provides a tumor screening system and its application based on the fusion of multidimensional features of circulating cell-free DNA. Within a unified framework, by introducing cross-modal feature alignment technology, multidimensional features such as mutation index, methylation index, and fragment feature index are mapped to the same feature space, constructing a structured fusion feature matrix, which effectively eliminates the heterogeneity of different omics data.

[0015] To achieve the above objectives, the present invention provides a tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA, comprising:

[0016] The sample processing module is used to receive body fluid samples from the subject and to extract and enrich cfDNA from the body fluid samples.

[0017] The molecular detection module is used to perform multi-omics detection on the cfDNA and obtain multidimensional molecular feature data of the cfDNA, including at least mutation features, methylation features and cfDNA fragment features.

[0018] The data processing module is used to preprocess the multidimensional molecular feature data obtained by detection to obtain mutation index, methylation index and fragment feature index, and map each multidimensional molecular feature data to the same feature space to construct a unified fusion feature matrix.

[0019] The multimodal fusion machine analysis module performs multimodal fusion analysis on the fusion feature matrix based on the multimodal fusion machine learning model, and generates the fusion analysis results.

[0020] The output module outputs the tumor risk score and risk level of the examined subject based on the results of the fusion analysis.

[0021] As an improvement of the present invention, the method for enriching the cfDNA in the body fluid sample is to achieve selective enrichment based on liquid phase partitioning or polymer phase separation.

[0022] As an improvement of the present invention, the mutation characteristics include mutation allele frequency, mutation load, number of mutation sites, hotspot mutation distribution characteristics, mutation combination patterns, and mutation spatial distribution characteristics.

[0023] The methylation features include the methylation level of a single CpG site, the average level of regional methylation, the features of differentially methylated regions, the combined methylation features of multiple sites, the methylation pattern, and the methylation heterogeneity index.

[0024] The cfDNA fragment characteristics include single nucleosome protected fragment characteristics, nucleosome periodic distribution characteristics, fragment length distribution curve characteristics, fragment ratios in different length intervals, and fragment length ratio characteristics.

[0025] As an improvement of the present invention, the method for preprocessing the various multidimensional molecular feature data in the data processing module includes:

[0026] For the mutation features, low-quality site data with sequencing depth below a preset threshold are removed, low-confidence mutation sites are removed, and missing mutation site data are filled with zero to obtain the mutation index.

[0027] For the methylation features, samples with bisulfite conversion efficiency lower than a preset conversion threshold were removed, and median filling was used for missing CpG site methylation data. Methylation levels were standardized using quantiles to eliminate batch effects and obtain the methylation index.

[0028] For the cfDNA fragment features, abnormal fragment data with lengths below a first preset threshold or above a second preset threshold are removed, as are low-quality fragment data with sequencing quality or alignment quality below a preset quality threshold. Based on the retained fragments, the fragment length distribution curve, the proportion of fragments in different length intervals, the fragment length ratio, the characteristics of single nucleosome protected fragments, and the periodic distribution characteristics of nucleosomes are statistically analyzed. Missing fragment feature data are filled with the mean or median. The fragment feature data are then standardized or normalized to obtain the fragment feature index.

[0029] As an improvement of the present invention, the process by which the data processing module constructs the fused feature matrix includes:

[0030] a. The pre-processed and screened mutation features, methylation features, cfDNA fragment features, and optional clinical information features are vectorized and concatenated to form a unified feature vector;

[0031] b. Perform Min-Max normalization on the concatenated feature vector to scale all feature values ​​to the [0,1] interval: =( - ) / ( - ), where X is the original feature vector, The minimum value of this eigenvector, The maximum value of this eigenvector, These are the normalized eigenvector values;

[0032] c. Construct a fusion feature matrix of dimension N×D. , where N is the number of samples and D is the total number of features after screening.

[0033] As an improvement of the present invention, the steps of the multimodal fusion machine analysis module in performing multimodal fusion analysis include:

[0034] The fused feature matrix The data are input into a machine learning model, which learns mutation features, methylation features, and fragment features through training.

[0035] Output the probability of cancer occurrence P, and calculate the risk score according to the formula. =Probability P × 100%;

[0036] Risk scoring is based on a preset risk threshold. It is divided into low-risk, medium-risk, and high-risk levels.

[0037] As an improvement of the present invention, the multimodal fusion machine learning model is any one of the following: random forest model, gradient boosting tree model, support vector machine model, or deep learning model.

[0038] As an improvement to the present invention, the fused feature matrix is ​​screened and optimized, and the screening method includes:

[0039] Variance-based feature filtering;

[0040] Feature screening based on statistical tests;

[0041] Feature optimization based on recursive feature elimination;

[0042] Redundancy removal based on feature correlation.

[0043] As an improvement of the present invention, the mutation detection method for obtaining the mutation characteristics includes multiplex PCR amplification, targeted capture sequencing, high-throughput sequencing, and digital PCR.

[0044] The methylation detection methods for obtaining the DNA methylation characteristics include bisulfite conversion sequencing, methylation-specific PCR, methylation-specific quantitative PCR, and targeted methylation high-throughput sequencing.

[0045] The cfDNA fragment extraction method for obtaining the cfDNA fragment characteristics includes pretreating the body fluid sample to remove cellular components and impurities, extracting and enriching cfDNA from the pretreated sample, constructing a library and sequencing the cfDNA to obtain cfDNA sequencing data.

[0046] This invention also provides an application of a tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA in early tumor screening, tumor risk assessment, or auxiliary diagnosis.

[0047] The beneficial effects of this invention are as follows: Compared with existing technologies, this invention provides a tumor screening system and its application based on the fusion of multidimensional features of circulating cell-free DNA. This invention integrates mutation features, methylation features, and fragment features of circulating cell-free DNA to construct a unified multidimensional feature expression system, breaking down information barriers between multi-omics data. Simultaneously, it establishes a standardized multimodal fusion analysis framework. By preprocessing, screening, and optimizing features of different dimensions, the fused feature matrix is ​​input into machine learning models such as random forests and gradient boosting trees for training and learning, achieving precise quantification and stratification of tumor risk. This integrated approach effectively solves the problem of the lack of a unified feature expression and fusion analysis framework for multi-omics data in existing technologies, significantly improving the accuracy and reliability of tumor screening and providing more scientific and efficient technical support for early clinical tumor detection. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the workflow of a tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA according to the present invention.

[0049] Figure 2 This is a flowchart of the multidimensional feature fusion process of cyclic free DNA according to the present invention;

[0050] Figure 3 Distribution map of cfDNA fragments from urine;

[0051] Figure 4 This is a performance evaluation curve for the model;

[0052] Figure 5 This is a graph showing the importance of model features. Detailed Implementation

[0053] To more clearly illustrate the present invention, the invention will be further described below with reference to the accompanying drawings.

[0054] In the following description, examples and details are given to provide a more in-depth understanding of the invention. It is obvious that the described embodiments are merely some, not all, of the embodiments of the invention. It should be understood that the specific embodiments are used only to explain the invention and are not intended to limit the invention.

[0055] It should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of a feature, whole, step, operation, element, or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, or combinations thereof.

[0056] Please see Figures 1-5The present invention discloses a tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA, comprising:

[0057] The sample processing module is used to receive body fluid samples from the subjects being tested and to extract and enrich cfDNA from the body fluid samples.

[0058] The molecular detection module is used to perform multi-omics detection on cfDNA and obtain multidimensional molecular feature data of cfDNA, including at least mutation features, methylation features and cfDNA fragment features.

[0059] The data processing module is used to preprocess the multidimensional molecular feature data obtained by detection, obtain mutation index, methylation index and fragment feature index, and map each multidimensional molecular feature data to the same feature space to construct a unified fusion feature matrix.

[0060] The multimodal fusion machine analysis module performs multimodal fusion analysis on the fusion feature matrix based on a multimodal fusion machine learning model, and generates the fusion analysis results.

[0061] The output module outputs the tumor risk score and risk level of the examined subjects based on the results of the fusion analysis.

[0062] This invention achieves simultaneous extraction and integration of mutation features, methylation features, and cfDNA fragment features using the same cfDNA sample and extraction system. This effectively solves the problems of scattered and inconsistent multi-omics data sources in existing technologies. For body fluid samples with low cfDNA content, such as urine, enrichment processing based on liquid phase partitioning or polymer phase separation significantly improves cfDNA recovery rate and detection signal intensity, enhancing detection stability. The data processing module ensures the reliability and comparability of feature data across all dimensions through targeted preprocessing procedures (such as filtering low-quality sites for mutation features, eliminating batch effects for methylation features, and standardizing fragment features). The multimodal fusion machine analysis module employs an early fusion strategy. The preprocessed multi-omics features are concatenated into a unified feature matrix, and machine learning models such as random forest and gradient boosting tree are used for fusion modeling to fully explore the complementary information between the features, thereby achieving accurate prediction of tumor risk. The output module classifies the examined subjects into three levels: low risk, medium risk, and high risk based on the risk score output by the model, providing clinicians with intuitive and quantitative decision-making basis. The tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA of this invention is suitable for early non-invasive screening of various common malignant tumors such as lung cancer, colorectal cancer, and liver cancer. It can be widely used in regular physical examinations of healthy people and risk monitoring of high-risk groups, which helps to achieve early detection, early diagnosis, and early treatment of tumors, and significantly improves the survival rate and quality of life of patients.

[0063] In this embodiment, the method for enriching cfDNA in body fluid samples is based on selective enrichment using either liquid phase partitioning or polymer phase separation. Liquid phase partitioning, through precise adjustment of the solution's pH, ionic strength, and organic solvent ratio, allows cfDNA to selectively accumulate at the interface between the aqueous and organic phases, effectively avoiding interference from impurities such as proteins and cell debris. Polymer phase separation, utilizing the affinity of specific polymers such as polyethylene glycol for cfDNA, promotes the precipitation of cfDNA from complex body fluid samples. This method is simple to operate and causes minimal damage to the molecular integrity of cfDNA. Using these enrichment methods can improve cfDNA recovery efficiency, enhance detection signal intensity, and improve detection stability.

[0064] The cfDNA enrichment treatment meets the following performance conditions:

[0065] The cfDNA concentration factor is ≥50 times, preferably ≥200 times, and even more preferably ≥500–1000 times;

[0066] cfDNA recovery rate ≥80%; short fragment cfDNA (100–200 bp) retention rate ≥90%.

[0067] The above enrichment process improves the detection capability of low-abundance molecular signals, providing a foundation for multi-omics joint detection.

[0068] Furthermore, the key to the cfDNA enrichment system lies in achieving selective enrichment of short cfDNA fragments based on differences in cfDNA fragment length, molecular weight, and conformation through liquid-phase partitioning or polymer-phase separation mechanisms. This maintains consistency in the following characteristics during the enrichment process: mutation characteristics (such as VAF), methylation characteristics (such as β value), and cfDNA fragment structural characteristics. Without this selective enrichment mechanism, the loss rate of short cfDNA fragments increases significantly, leading to a decrease in consistency between mutation and methylation characteristics, resulting in a performance degradation of more than 15–25% for the multi-omics fusion model. Therefore, the selective enrichment is not only used to increase cfDNA concentration but also to maintain structural consistency among multi-omics characteristics. Selective enrichment constitutes a key prerequisite for achieving multi-omics fusion modeling in this invention.

[0069] Furthermore, this selective enrichment characteristic differs from the non-selective extraction methods of traditional column purification or magnetic bead methods. The latter lacks the ability to distinguish different lengths of cfDNA during the enrichment process, which can easily lead to fragment distribution shifts and distortion of multi-omics features. The aforementioned selective enrichment mechanism works synergistically with the subsequent multi-omics feature extraction and fusion modeling process to form an inseparable overall technical solution.

[0070] cfDNA extraction and enrichment can be achieved using known nucleic acid separation techniques in the field, including but not limited to enrichment methods based on liquid phase partitioning, phase separation or selective precipitation, or magnetic bead method, column purification method, ultrafiltration method and their combination based on solid phase carrier.

[0071] In one embodiment, a non-solid-phase-dependent liquid-phase separation and enrichment system is used to concentrate cfDNA; the liquid-phase separation and enrichment system refers to a system in which cfDNA is selectively allocated or enriched with other impurities in a homogeneous liquid system through differences in intermolecular interactions (including but not limited to molecular weight, charge, hydrophilicity or polymer phase separation behavior).

[0072] Furthermore, liquid phase separation and enrichment systems include, but are not limited to, one or a combination of the following methods:

[0073] (1) Aqueous two-phase system based on polymer phase separation;

[0074] (2) Selective precipitation systems based on salt-polymer interactions;

[0075] (3) Liquid phase partitioning systems based on differences in molecular weight or charge;

[0076] In a preferred embodiment, the aqueous two-phase system may be constructed using polyethylene glycol (PEG) and polysaccharides (such as Dextran), but is not limited to this specific combination.

[0077] Preferably, the cfDNA isolation and enrichment system meets the following parameter conditions:

[0078] The PEG mass fraction is 5–15%;

[0079] The polysaccharide mass fraction is 5–10%;

[0080] pH range: 6.5–8.0;

[0081] Furthermore, the cfDNA enrichment system should meet the following functional indicators:

[0082] The cfDNA concentration factor is ≥50 times, preferably ≥200 times, and even more preferably ≥500 times;

[0083] cfDNA recovery rate ≥80%;

[0084] Fragment retention rate of 100–200bp ≥90%;

[0085] The key to this invention lies not in the specific type of separation system, but in achieving efficient concentration of low-abundance cfDNA through an enrichment system, thereby supporting the stable detection and unified modeling of subsequent multi-omics features.

[0086] Therefore, any method that can achieve an equivalent separation of the above-mentioned functional indicators falls within the protection scope of this invention.

[0087] The above enrichment method can achieve high-fold concentration of cfDNA, with a concentration factor of ≥50 times; preferably ≥200 times; further preferably ≥1000 times; and in some embodiments, it can reach more than 10000 times.

[0088] High enrichment can significantly improve the detection capability of low-abundance tumor-related molecular signals.

[0089] Furthermore, compared with traditional solid-phase extraction methods, the non-solid-phase separation and enrichment method of the present invention has the following technical advantages: reduced DNA loss and improved recovery rate; reduced background noise and improved signal-to-noise ratio; suitable for low-concentration cfDNA samples; and beneficial to the consistency of subsequent multi-omics joint detection.

[0090] Therefore, this invention provides a unified and high-quality molecular basis for subsequent mutation detection, methylation detection and fragment feature analysis by constructing an efficient cfDNA enrichment system.

[0091] In this embodiment, mutation characteristics include mutation allele frequency, mutation load, number of mutation sites, hotspot mutation distribution characteristics, mutation combination patterns, and mutation spatial distribution characteristics.

[0092] Methylation characteristics include the methylation level of a single CpG site, the average level of regional methylation, the characteristics of differentially methylated regions, the combined methylation characteristics of multiple sites, methylation patterns, and indices of methylation heterogeneity.

[0093] cfDNA fragment characteristics include single nucleosome protective fragment characteristics, nucleosome periodic distribution characteristics, fragment length distribution curve characteristics, fragment ratios in different length ranges, and fragment length ratio characteristics.

[0094] All of the above features are derived from the same cfDNA sample, achieving homology collection of multi-omics data and thus avoiding cross-sample bias and batch-to-batch effects. The extraction of these features requires a combination of specific experimental techniques and bioinformatics methods: for example, the mutational allele frequency (VAF) in the mutation feature is calculated using deep sequencing data of the target region, and the sequencing depth needs to be controlled ≥1000 to ensure the stability of the results; the methylation feature is detected using bisulfite conversion combined with high-throughput sequencing technology, and the average level of regional methylation is statistically analyzed for CpG sites in functional regions such as gene promoters and enhancers; the cfDNA fragment feature is based on the fragment length distribution data after library construction, and the peak position and periodicity characteristics of nucleosome protected fragments are analyzed using bioinformatics tools, and the proportion of fragments in different length intervals needs to be obtained by integral calculation of the fragment length distribution curve.

[0095] In this invention, tumor-related gene mutations are detected based on cfDNA, and the mutation sites preferably originate from functional pathways related to tumor development and progression, including but not limited to:

[0096] (1) Cell proliferation and signal transduction pathways, including the RAS / MAPK pathway and the PI3K / AKT pathway;

[0097] (2) Cell cycle regulation and tumor suppressor pathways;

[0098] (3) DNA damage repair pathway;

[0099] (4) Chromatin remodeling and transcriptional regulation pathways.

[0100] Mutation detection preferably covers the following types of gene regions:

[0101] Driver gene mutation sites;

[0102] Sites associated with loss of function of tumor suppressor genes;

[0103] High-frequency hotspot abrupt change regions;

[0104] Splice sites or regions related to structural variations.

[0105] In a preferred embodiment, the set of mutation sites may be selected from some or all of the tumor-associated mutation hotspots listed below, or mutation sites that are functionally equivalent to them.

[0106] Furthermore, the selection of mutation sites can be based on one or a combination of the following criteria: high frequency of occurrence in tumors, clear biological function or pathogenic significance, usability for tumor risk assessment or stratification management, and annotation as pathogenic or potentially pathogenic in public databases or literature.

[0107] This invention is not limited to a specific set of genes or sites; any equivalent mutation site selected based on the above functional criteria is within the scope of protection of this invention.

[0108] In this embodiment, the preprocessing methods for each multidimensional molecular feature data in the data processing module include:

[0109] For mutation features, low-quality site data with sequencing depth below a preset threshold (usually 500) are removed, low-confidence mutation sites are removed, and missing mutation site data are filled with zero.

[0110] For methylation features, samples with bisulfite conversion efficiency lower than a preset conversion threshold were removed. For missing CpG site methylation data, median filling was used, and quantile standardization was used to eliminate batch effects for methylation levels.

[0111] For cfDNA fragment features, abnormal fragment data with lengths below a first preset threshold or above a second preset threshold are removed, as are low-quality fragment data with sequencing quality or alignment quality below a preset quality threshold. Based on the retained fragments, the fragment length distribution curve, the proportion of fragments in different length intervals, the fragment length ratio, the characteristics of single nucleosome protected fragments, and the periodic distribution characteristics of nucleosomes are statistically analyzed. Missing fragment feature data are filled with the mean or median. The fragment feature data are then standardized or normalized to obtain the fragment feature index.

[0112] Low-quality site data with sequencing depth <500× were removed for mutation features. Missing mutation site data were filled with zero (i.e., it was assumed that there was no mutation at the site). The original value (0-100%) of the mutation allele frequency (VAF) was retained without additional standardization.

[0113] For methylation features, samples with bisulfite conversion efficiency <95% were removed. Median imputation was used for methylation data of missing CpG sites. Quantile normalization was used to eliminate batch effects for methylation levels (β value, range 0-1). Methylation features, mutation features, and cfDNA fragment features were derived from the same cfDNA sample and fused in a unified feature space. Methylation features can be jointly modeled with mutation features and cfDNA fragment features to improve the identification ability and model stability of low-abundance tumor signals.

[0114] For cfDNA fragment features, abnormal fragment data with lengths <50bp or >500bp were removed, as were low-quality fragment data with sequencing quality Q <30 or alignment quality MQ <30. Based on the retained high-quality alignment results, the fragment length distribution curve, the proportion of fragments in different length ranges (100-150bp, 150-200bp, etc.), the ratio of short to long fragment lengths, the proportion of fragments corresponding to the single nucleosome protection peak, and the periodic distribution characteristics of nucleosome spacing were statistically analyzed. Missing fragment feature data were filled with the median, and then all fragment features were standardized using the Z-score method to obtain standardized fragment feature indices.

[0115] cfDNA fragment characteristics reflect the origin, degradation pattern, and chromatin structure information of cfDNA, and can serve as an important supplement to tumor-related molecular characteristics.

[0116] Fragment features, mutation features, and methylation features originate from the same cfDNA sample and are fused together in a unified feature space, thereby improving the sensitivity and specificity of tumor detection.

[0117] In this embodiment, a feature selection and extraction method is also included, comprising:

[0118] The following features were extracted from the mutation detection results:

[0119] Binary features: whether a mutation exists at each detection site (0 / 1 encoding), totaling 218 features;

[0120] Quantitative features: VAF values ​​of detected mutation sites, totaling 218 features;

[0121] Summary features: total number of mutated genes, total number of mutation sites, average VAF value, and maximum VAF value, a total of 4 features.

[0122] The following features were extracted from the methylation detection results:

[0123] Quantitative characteristics: methylation β values ​​at each CpG site, totaling 120 characteristics;

[0124] Binary features: Whether each CpG site is hypermethylated (β≥0.3 is defined as hypermethylation), a total of 120 features;

[0125] Summary features: number of hypermethylation sites, average methylation level, and methylation variation coefficient, a total of 3 features.

[0126] The following methods were used to filter the initial features:

[0127] Variance filtering: Removes low-variance features with variance < 0.01;

[0128] Univariate feature selection: Chi-square test or mutual information method was used to select features that were significantly correlated with cancer labels (p<0.05);

[0129] Recursive Feature Elimination (RFE): Based on the random forest model, iteratively removes the least important features and retains the top 50 to 100 features.

[0130] Multicollinearity test: Calculate the Pearson correlation coefficient between features. For feature pairs with a correlation coefficient > 0.9, retain the feature with higher correlation to the label.

[0131] Table 1

[0132] age Continuous variables The risk of developing cancer is closely related to age, and the incidence of most cancers increases with age. gender Categorical variables Some cancers exhibit significant gender differences (such as breast cancer and prostate cancer), which can aid in model identification. smoking history Categorical variables Smoking is a significant risk factor for lung cancer and many other cancers, and can improve the accuracy of model predictions. drinking history Categorical variables Alcohol consumption is associated with the development of cancers such as liver cancer and esophageal cancer. Family history of cancer Categorical variables Cancer has a certain genetic predisposition, and those with a positive family history have an increased risk. BMI (Body Mass Index) Continuous variables Obesity is associated with an increased risk of various cancers, including colorectal cancer and breast cancer.

[0133] Among them, detailed medical history, medication history, imaging examination results, and biochemical test indicators are not included in the clinical information. Detailed medical history and medication history information are too complex and have large individual differences, making them difficult to standardize. Imaging examination results require additional testing, which does not conform to the non-invasive and simple design concept of this invention. Biochemical test indicators require blood samples, which is inconsistent with the urine detection positioning of this invention.

[0134] In this embodiment, the process of the data processing module constructing the fused feature matrix includes:

[0135] a. The pre-processed and screened mutation features, methylation features, cfDNA fragment features, and optional clinical information features are vectorized and concatenated to form a unified feature vector;

[0136] b. Perform Min-Max normalization on the concatenated feature vector to scale all feature values ​​to the [0,1] interval: =( - ) / ( - ), where X is the original feature vector, The minimum value of this eigenvector, The maximum value of this eigenvector, These are the normalized eigenvector values;

[0137] c. Construct a fusion feature matrix of dimension N×D. , where N is the number of samples and D is the total number of features after screening.

[0138] In constructing the fused feature matrix, different types of features are first aligned according to samples to ensure that features in the same row correspond to the same detection sample. Then, they are horizontally concatenated along the feature dimension. The concatenated feature vectors are then subjected to Min-Max normalization to scale all feature values ​​to the [0,1] interval, eliminating the interference of features of different magnitudes on model training and ensuring that the contributions of different types of features to the model are at a balanced level. After normalization, the complete fused feature matrix Xfused is constructed according to the number of samples N and the number of features D. Please refer to the appendix. Figure 5 , Figure 5 The study demonstrated the contribution weights of different features in the tumor risk assessment model, including mutation features (such as gene mutations in TP53 and PIK3CA), methylation features (such as gene methylation in PAX1 and SEPT9), and comprehensive indicators (such as mutation burden and methylation index). The results showed that multi-omics features have a synergistic effect in the model, jointly improving the accuracy of risk assessment. Mutation features and methylation features are derived from the same cfDNA sample and are jointly analyzed through a multi-feature fusion model, thereby improving the sensitivity and specificity of detection.

[0139] After constructing the fused feature matrix, the samples are divided into a training set and an independent validation set according to a preset ratio. The training set is used for model training and parameter tuning, while the independent validation set is used to evaluate the model's generalization performance. The division ratio can be adjusted according to the total sample size, and is usually set to training set:validation set = 7:3 or 8:2. This invention uses a multimodal fusion machine learning framework to construct a tumor risk prediction model. First, the mutation feature subset, methylation feature subset, and cfDNA fragment feature subset are independently encoded and learned to mine feature information at the single-omics level. Then, the feature vectors learned from each single-omics are integrated through a splicing and fusion layer. The input is then fed into a fully connected network to output the probability P that the sample is tumor-positive. The value of P ranges from 0 to 1, corresponding to the risk score. =Probability P × 100%, with a value ranging from 0 to 100. During model training, a cross-validation strategy is used to optimize the model hyperparameters, with binary cross-entropy as the loss function. The Adam optimizer is used for gradient descent updates, and L2 regularization and Dropout layers are added to suppress model overfitting. After model training, risk stratification thresholds are determined based on the risk score distribution of the validation set. Samples with scores below the low-risk threshold are classified as low-risk, those with scores between the low-risk and high-risk thresholds are classified as medium-risk, and those with scores above the high-risk threshold are classified as high-risk. High-risk samples are recommended for further clinical diagnostic testing, while low-risk samples are recommended for regular screening and follow-up.

[0140] This invention relates to a tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA. It can simultaneously integrate three different groups of tumor molecular features—mutation, methylation, and fragmentation—and combine them with optional clinical information. Compared to models constructed using single features, the multidimensional feature fusion method of this invention is expected to further improve the predictive performance of the model.

[0141] It effectively improves the detection sensitivity of early-stage tumors while maintaining high specificity, and can be used for early tumor screening and risk stratification management in healthy individuals. Please refer to the appendix. Figure 4 , attached Figure 4 The horizontal axis represents the false positive rate, and the vertical axis represents the true positive rate. The ROC curve shows the classification performance under different decision thresholds. The shaded area represents the confidence interval. The area under the curve (AUC) is used to comprehensively evaluate the model performance. The AUC of the method in this invention is approximately 0.73, indicating that it has good discriminative ability and stability.

[0142] In this embodiment, the steps of the multimodal fusion machine analysis module in performing multimodal fusion analysis include:

[0143] Fuse feature matrix The data is input into a machine learning model, which learns mutation features, methylation features, and fragment features through training.

[0144] Output the probability of cancer occurrence P, and calculate the risk score according to the formula. =Probability P × 100%;

[0145] Risk scoring is based on a preset risk threshold. The risk levels can be divided into low risk, medium risk, and high risk, or a risk assessment table can be created as shown in Table 2.

[0146] Table 2

[0147] 0-30 points Low risk Regular health checkups are recommended. 31-60 points Medium risk It is recommended to have a follow-up examination in 6 months, and to combine it with other examination methods. 61-100 points High risk Further clinical examinations (such as imaging and endoscopy) are recommended immediately.

[0148] The candidate methylation sites can be selected from some or all of the sites listed below, or methylation feature regions whose functional regions are equivalent to these sites: cg00000934, cg00001386, cg00001609, cg00001814, cg00002102, cg00002706… (listed here in order of the sites actually screened). This invention is not limited to the specific CpG sites mentioned above; any equivalent methylation sites obtained based on functional regions related to tumor development and progression are within the scope of protection of this invention.

[0149] For cfDNA fragment features, abnormal length fragments shorter than 50 bp and longer than 500 bp are first removed. Then, the proportion of fragments in different length intervals is statistically analyzed, and statistical measures such as skewness and kurtosis of the fragment length distribution are calculated. The interval proportions and statistical features are retained as the final features for subsequent modeling. Missing statistical features are filled using the mean of the corresponding features. After feature screening, the selected features include 15-25 mutation features, 25-35 methylation features, 15-25 fragment features, and 4-6 clinical features, totaling 60-90 features. This approach ensures predictive performance while avoiding model overfitting caused by excessive feature dimensionality.

[0150] In this embodiment, the multimodal fusion machine learning model can be any one of the following: random forest, gradient boosting tree, support vector machine, or deep learning model. Among them, XGBoost or LightGBM models based on gradient boosting trees perform well in handling high-dimensional features and nonlinear relationships, and can effectively capture the complex interactions between multidimensional features. The deep learning model can use convolutional neural networks (CNNs) or Transformers with attention mechanisms to further explore the potential correlations between features through automatic feature extraction, which is particularly suitable for large-scale sample datasets. Model selection should be combined with the actual application scenario and sample size: in small-scale sample scenarios, traditional machine learning models are preferred to reduce the risk of overfitting, while deep learning models can be used to improve prediction accuracy in large-scale sample scenarios. In addition, this invention supports model ensemble strategies, which use stacking or Bagging frameworks to weightedly fuse the prediction results of multiple single models to further enhance the stability and generalization ability of the model. For example, random forest, XGBoost, and support vector machines can be used as base models, and their outputs can be used as meta-features input into the logistic regression meta-model to obtain the final prediction probability.

[0151] In this embodiment, the feature matrix is ​​filtered and optimized. The feature matrix filtering method includes:

[0152] Variance-based feature filtering removes low-variance features with variance below a preset threshold. These features show very little difference across different samples and cannot provide effective distinguishing information for tumor classification prediction. Directly removing them can reduce feature dimensionality and decrease the computational load of the model.

[0153] Feature selection based on statistical tests involves calculating the correlation between each feature and the tumor-positive label individually, selecting features with statistically significant correlations, and quickly removing invalid features with weak associations to the label.

[0154] Feature optimization based on recursive feature elimination combines recursive feature elimination with feature importance scoring of a tree model to gradually remove the features with the lowest importance, and finally retain the set of core features that contribute the most to the prediction.

[0155] Redundancy removal based on feature correlation is performed, multicollinearity test is conducted, and the Pearson correlation coefficient between any two features is calculated. When the correlation coefficient is higher than a preset threshold, only the feature with higher correlation to the tumor label is retained, and redundant information is eliminated to avoid the interference of multicollinearity on the stability of model parameters and improve the interpretability of the model.

[0156] Through the above multiple rounds of screening and optimization, the feature dimension can be significantly compressed while retaining the core prediction information, thereby reducing the computational complexity of model training. At the same time, the interference of redundant features on model prediction is reduced, further improving the prediction stability of the model in practical applications. The final selected feature set can be directly used for subsequent model training and risk prediction, with the number of features controlled within a reasonable range to adapt to sample datasets of different sizes.

[0157] In this embodiment, the mutation detection methods for obtaining mutation features include multiplex PCR amplification, targeted capture sequencing, high-throughput sequencing, and digital PCR. Multiplex PCR amplification can be combined with targeted capture technology to enrich the target genomic region. High-throughput sequencing is used to obtain the gene sequence of the detection region. By comparing with the reference genome, different types of mutations such as single nucleotide variants, insertions / deletions, and copy number variations can be identified. The above detection methods have high throughput and good sensitivity, and can simultaneously detect mutations in hundreds of tumor-related genes, making them suitable for large-scale early screening applications. After obtaining the raw sequencing data, standard bioinformatics workflows are used for quality control and variant identification. Low-quality reads are removed, and after alignment with the reference genome, PCR repeats are removed. Then, the allele frequency (VAF) of each mutation site is calculated, and finally, the preset mutation features are extracted. By obtaining the mutation information of the target region through the above sequencing methods, susceptible sites that are clearly related to tumor occurrence and development are screened out. Corresponding features are extracted based on the presence and abundance of mutations at the sites. This invention does not limit the specific sequencing technology route, as long as the mutation information of the target site can be stably obtained.

[0158] Methods for obtaining DNA methylation features include bisulfite conversion sequencing, methylation-specific PCR, methylation-specific quantitative PCR, and targeted methylation high-throughput sequencing. Bisulfite conversion sequencing can convert unmethylated cytosine into uracil, while methylated cytosine remains unchanged. Subsequent sequencing can distinguish between methylated and unmethylated sites. Targeted methylation high-throughput sequencing can specifically enrich methylated fragments in the target region, reducing sequencing costs while ensuring detection accuracy, making it suitable for large-scale early screening applications. After obtaining the raw data, quality control and alignment to the reference genome are performed, and the methylation β value of each CpG site is calculated. The β value ranges from 0 to 1, with higher values ​​indicating higher methylation levels at that site. Binary features and summary features are then extracted based on preset thresholds to complete the construction of methylation features. The methylation level of the target CpG site is obtained through the above methods, and tumor-specific methylation sites are screened out. Corresponding features are extracted based on the methylation β value and high methylation determination of the site. This invention does not limit the specific methylation detection technology route, as long as the methylation level of the target site can be stably obtained.

[0159] The cfDNA fragment extraction method for obtaining cfDNA fragment features includes preprocessing body fluid samples to remove cellular components and impurities, extracting and enriching cfDNA from the preprocessed samples, constructing libraries and sequencing the cfDNA to obtain cfDNA sequencing data, determining the cfDNA fragment length based on the alignment position or sequencing read length information of paired reads in the sequencing data, statistically analyzing the proportion of fragments in different length intervals across the whole genome, and further calculating statistical measures such as the first moment, second moment, skewness, and kurtosis of the fragment length distribution. At the same time, fragment features at the genome level, such as fragment end preference and fragment proportion integral difference, are retained. After filtering out meaningless features with low variability, the final set of cfDNA fragment features used for modeling is obtained. For fragment features with missing features, the mean of the corresponding feature in all training samples is directly used to fill the missing features, ensuring that the feature matrix is ​​dimensionally complete and does not affect subsequent modeling and analysis.

[0160] This invention also provides an application of a tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA in cancer screening.

[0161] In another embodiment, a method for cleaning and characterizing cfDNA data is also included:

[0162] (1) Data cleaning process

[0163] Table 3

[0164] Sample quality screening Remove samples with cfDNA extraction amount <1ng Ensure the reliability of test data Outlier detection Use the 3σ principle to identify and mark outliers. Avoid the influence of extreme values ​​on the model Outlier handling Replace outliers with boundary values ​​(upper / lower 3σ). Preserve samples while reducing the impact of anomalies Duplicate sample removal Detect and remove duplicate samples Avoid data breaches Label verification Verify that the cancer / health label is consistent with the clinical diagnosis. Ensure label accuracy

[0165] (2) Feature encoding method

[0166] Table 4

[0167] Mutation Presence Binary encoding Mutation present = 1, no mutation present = 0 Mutant VAF Continuous values VAF = 2.5% → 0.025 methylation β value Continuous values β=0.45→0.45 gender One-hot encoding Male=[1,0], Female=[0,1] age Continuous numerical values ​​(standardized) 55 years old → Z-score smoking history ordered encoding Never = 0, Once = 1, Currently = 2 Family history Binary encoding Present = 1, Absent = 0

[0168] Methods for constructing training datasets

[0169] (1) Dataset source and scale

[0170] This invention constructs a training dataset containing multiple tested samples. In one embodiment, a training dataset consisting of cancer samples and healthy samples can be constructed. The size of the dataset can be determined according to the actual application scenario, for example, containing 100 to 300 samples. Cancer samples may include one or more types of cancer such as lung cancer, colorectal cancer, gastric cancer, and cervical cancer, while healthy samples may come from subjects without malignant tumors. The training dataset of this invention contains 120 urine samples:

[0171] Cancer group: 60 cases, including 15 cases of lung cancer, 15 cases of colorectal cancer, 10 cases of gastric cancer, 10 cases of cervical cancer, and 10 cases of other types of cancer;

[0172] Healthy control group: 60 cases, confirmed by clinical examination to have no malignant tumors.

[0173] (2) Inclusion and exclusion criteria

[0174] Inclusion criteria:

[0175] Cancer group: Patients with pathologically confirmed malignant tumors who have not yet undergone surgery, radiotherapy or chemotherapy;

[0176] Control group: Healthy individuals undergoing physical examinations who have not been diagnosed with malignant tumors within the past year.

[0177] Exclusion criteria:

[0178] Those whose urine samples are unqualified;

[0179] Patients with other serious systemic diseases;

[0180] Those with incomplete sample information.

[0181] (3) The dataset was split using stratified sampling, with the training and test sets divided in a 7:3 ratio:

[0182] Training set: 84 cases (42 cancer cases, 42 healthy cases)

[0183] Test set: 36 cases (18 with cancer, 18 healthy).

[0184] Stratified sampling ensures that the proportions of each cancer type, age distribution, and gender ratio remain consistent between the training set and the test set.

[0185] (4) Data augmentation (optional) For cancer types with small sample sizes, the SMOTE method can be used to generate synthetic samples to balance the number of samples in each category.

[0186] In one method of training an AI model:

[0187] (1) Model selection The present invention preferably adopts the random forest algorithm for the following reasons:

[0188] It has good processing capabilities for high-dimensional feature data;

[0189] Built-in feature importance assessment function;

[0190] It exhibits strong robustness against overfitting;

[0191] It has good interpretability and is easy to apply in clinical practice.

[0192] (2) Model parameter settings

[0193] Table 5

[0194] n_estimators (number of decision trees) 500 The optimal value determined by grid search max_depth (maximum depth) 10 Preventing overfitting min_samples_split (minimum number of split samples) 5 Preventing overfitting

[0195] (3) Training process

[0196] Step 1: Load the training set feature matrix X_train and label vector y_train

[0197] Step 2: Initialize the random forest classifier and set the parameters mentioned above.

[0198] Step 3: Call the fit(X_train, y_train) method to train the model.

[0199] Step 4: Save the trained model for subsequent predictions.

[0200] (4) In addition to random forest, the following models may also be used in this invention:

[0201] Logistic Regression: Suitable for scenarios where features and labels have a linear relationship;

[0202] Support Vector Machine: Suitable for small sample high-dimensional data;

[0203] Gradient Boosting: Excellent prediction performance, but long training time;

[0204] Deep Neural Networks (DNNs): Suitable for large-scale datasets.

[0205] Methods for model validation and optimization

[0206] (1) Cross-validation: 5-fold cross-validation was used to evaluate model performance.

[0207] The training set is divided into 5 equal-sized subsets;

[0208] Each training iteration uses 4 subsets and 1 subset for validation.

[0209] Repeat 5 times and calculate the average performance index.

[0210] (2) Hyperparameter optimization is performed using grid search combined with cross-validation:

[0211] Parameter search space:

[0212] -n_estimators:[100,300,500,800];

[0213] -max_depth:[5,10,15,None];

[0214] -min_samples_split:[2,5,10];

[0215] -min_samples_leaf:[1,2,4];

[0216] run;

[0217] Choose the parameter combination with the highest AUC from cross-validation as the final model parameters.

[0218] (3) Performance evaluation indicators

[0219] Table 6

[0220] Sensitivity TP / (TP+FN) The proportion of cancer patients correctly identified Specificity TN / (TN+FP) The proportion of people who are correctly identified as healthy Accuracy (TP+TN) / (TP+TN+FP+FN) Overall classification accuracy AUC Area under the ROC curve Model comprehensive discriminative ability Positive predictive value (PPV) TP / (TP+FP) The proportion of true positives among predicted positives Negative predictive value (NPV) TN / (TN+FN) The proportion of true negatives among those predicted as negative

[0221] Where: TP = true positive, TN = true negative, FP = false positive, FN = false negative

[0222] (4) Overfitting detection and prevention

[0223] Compare the performance differences between the training set and the test set; a difference greater than 10% indicates overfitting.

[0224] Overfitting can be prevented by employing methods such as regularization, early stopping, and reducing the number of features;

[0225] Plot the learning curve and observe the performance change trend as the number of training samples increases.

[0226] Basis for cancer risk prediction output

[0227] (1) Calculation method for predicting the probability of cancer P output by the random forest model:

[0228] P = (1 / N) × Σ (probability that Ti is predicted to be cancer);

[0229] Where N is the total number of decision trees (500), and Ti is the i-th decision tree. Each decision tree assigns samples to leaf nodes based on the input features, and the proportion of cancer samples in that leaf node is the predicted probability of that tree.

[0230] (2) Risk score calculation formula

[0231] Risk Score =Probability P × 100%;

[0232] Where P is the cancer probability (0-1) output by the model, Round is the rounding function, and the final score range is 0-100.

[0233] (3) Classification threshold setting

[0234] Threshold selection method: Based on the ROC curve, the point that maximizes the Youden Index is selected as the classification threshold.

[0235] Yoden Index = Sensitivity + Specificity - 1

[0236] In one embodiment, the model established by the present invention can achieve good classification performance, and its area under the ROC curve (AUC) preferably reaches 0.73 or above, more preferably 0.8 or above.

[0237] In this embodiment of the invention, the optimal classification threshold is calculated to be P=0.45 (i.e., risk score of 45) using training data.

[0238] (4) Basis for setting risk level thresholds

[0239] Risk level thresholds are determined based on training data using ROC analysis, Youden index, or other methods.

[0240] Table 7

[0241] Low risk 0-30 points The actual cancer prevalence rate within this range is <5%, corresponding to a higher NPV. Medium risk 31-70 points This range includes the classification threshold (45 points), and further observation is needed. High risk 71-100 points The actual cancer prevalence rate within this range is >80%, corresponding to a higher PPV.

[0242] (5) Output Report Content Model The final output includes:

[0243] Cancer risk score (0-100 points)

[0244] Risk level (low / medium / high)

[0245] Key influencing factors (Top 5 characteristics and their contribution)

[0246] Recommended measures

[0247] The bioinformatics analysis workflow is shown in Table 8 below. Sequencing data is processed through the following bioinformatics analysis workflow:

[0248] Table 8

[0249] Data quality control FastQC v0.11.9 Raw data quality assessment Connector removal Trimmomatic / fastp v0.39 / v0.23.2 Remove connector sequences and low-quality bases Sequence alignment BWA-MEM2 v2.2.1 Reads were aligned to the human reference genome (hg19 / GRCh37). Repeat sequence markers PicardMarkDuplicates v2.27.4 Labeled PCR repeat sequences Mutation identification VarDict / Mutect2 v1.8.3 / GATK4.3 Low-frequency mutation detection Mutation annotation ANNOVAR / VEP Latest version Mutation Functional Annotation VAF calculation Custom script - Calculate the frequency of mutant alleles

[0250] Key analysis parameters:

[0251] Alignment quality threshold: MAPQ ≥ 30;

[0252] Base quality threshold: Base Quality ≥ 20;

[0253] Mutation detection threshold: VAF ≥ 0.5%, supporting ≥ 3 reads;

[0254] Reference genome: GRCh37 / hg19;

[0255] The following is a specific embodiment of the present invention:

[0256] (a) Sample Acquisition

[0257] Obtain bodily fluid samples from the subject of the examination; urine is preferred.

[0258] In a preferred embodiment, a urine sample of 20–50 mL is collected, preferably a morning urine sample, to improve the cfDNA concentration and detection stability. In other embodiments, blood, cervical exfoliated cells or other bodily fluid samples may also be used. After sample collection, it is preferable to complete the processing within a short time to reduce cfDNA degradation.

[0259] In a preferred embodiment, the sample is pretreated within 2–6 hours after collection, or stored at 4°C for a short period of no more than 24 hours; furthermore, nucleic acid stabilizers or anti-degradation agents can be added to the sample to improve the integrity and stability of cfDNA.

[0260] The sample acquisition method of this invention is suitable for the enrichment and detection of low-concentration cfDNA, and provides a unified molecular basis for the joint analysis of subsequent mutation features, methylation features and cfDNA fragment features.

[0261] Please see the appendix Figure 3 The cfDNA fragments mainly formed a main peak at approximately 150 bp, corresponding to DNA fragments protected by mononucleosomes. Dinucleosome and trinucleosome-related peaks were observed at approximately 300 bp and 450 bp, respectively, exhibiting typical periodic nucleosome distribution characteristics. This feature originates from the protective effect of nucleosomes on DNA during apoptosis, indicating that the extracted cfDNA has good fragment integrity and stable biological origin, making it suitable for subsequent mutation detection and methylation analysis.

[0262] (II) Extraction and enrichment of cfDNA

[0263] 1. Sample preprocessing

[0264] Pretreatment of body fluid samples to remove cells and impurities preferably includes:

[0265] Centrifuge at low speed to remove cellular components;

[0266] High-speed centrifugation is used to remove residual cell debris and impurities;

[0267] The resulting supernatant was used for subsequent cfDNA extraction.

[0268] In a preferred embodiment, the low-speed centrifugation conditions are 1000–4000g for 5–15min; the high-speed centrifugation conditions are 10000–16000g for 5–15min.

[0269] 2. cfDNA extraction and enrichment methods

[0270] cfDNA extraction and enrichment can be achieved using known nucleic acid separation techniques in the field, including but not limited to enrichment methods based on liquid phase partitioning, phase separation or selective precipitation, or magnetic bead method, column purification method, ultrafiltration method and their combination based on solid phase carrier.

[0271] In a preferred embodiment, a non-solid-phase-dependent liquid-phase separation and enrichment system is used to concentrate cfDNA.

[0272] Liquid phase separation and enrichment systems refer to systems in which cfDNA is selectively allocated or enriched with other impurities in a homogeneous liquid system through differences in intermolecular interactions (including but not limited to molecular weight, charge, hydrophilicity / hydrophobicity, or polymer phase separation behavior).

[0273] Furthermore, liquid phase separation and enrichment systems include, but are not limited to, one or a combination of the following methods:

[0274] (1) Aqueous two-phase system based on polymer phase separation;

[0275] (2) Selective precipitation systems based on salt-polymer interactions;

[0276] (3) Liquid phase partitioning systems based on differences in molecular weight or charge;

[0277] In a preferred embodiment, the aqueous two-phase system may be constructed using polyethylene glycol (PEG) and polysaccharides (such as Dextran), but is not limited to this specific combination.

[0278] Preferably, the above system needs to meet the following parameter conditions:

[0279] The PEG mass fraction is 5–15%;

[0280] The polysaccharide mass fraction is 5–10%;

[0281] pH range: 6.5–8.0;

[0282] Furthermore, the cfDNA enrichment system should meet the following functional indicators:

[0283] The cfDNA concentration factor is ≥50 times, preferably ≥200 times, and even more preferably ≥500 times;

[0284] cfDNA recovery rate ≥80%;

[0285] Fragment retention rate of 100–200bp ≥90%;

[0286] The key to this invention lies not in the specific type of separation system, but in achieving efficient concentration of low-abundance cfDNA through enrichment systems, thereby supporting the stable detection and unified modeling of subsequent multi-omics features. Therefore, any equivalent separation method that can achieve the above-mentioned functional indicators falls within the scope of protection of this invention.

[0287] The above enrichment method can achieve high-fold concentration of cfDNA, with concentration factors of ≥50 times, ≥200 times, ≥1000 times, and in some embodiments, it can reach more than 10,000 times.

[0288] High enrichment can significantly improve the detection capability of low-abundance tumor-related molecular signals.

[0289] Furthermore, compared with traditional solid-phase extraction methods, the non-solid-phase separation and enrichment method of the present invention reduces DNA loss, improves recovery rate, reduces background noise, and improves signal-to-noise ratio; it is suitable for low-concentration cfDNA samples; and it is beneficial to the consistency of subsequent multi-omics joint detection.

[0290] Therefore, this invention provides a unified and high-quality molecular basis for subsequent mutation detection, methylation detection and fragment feature analysis by constructing an efficient cfDNA enrichment system.

[0291] (III) Mutation Detection

[0292] 1. Testing Content

[0293] This invention detects tumor-related gene mutations based on cfDNA, with the mutation sites preferably originating from functional pathways related to tumor development and progression, including but not limited to:

[0294] (1) Cell proliferation and signal transduction pathways, including the RAS / MAPK pathway and the PI3K / AKT pathway;

[0295] (2) Cell cycle regulation and tumor suppressor pathways;

[0296] (3) DNA damage repair pathway;

[0297] (4) Chromatin remodeling and transcriptional regulation pathways.

[0298] Mutation detection preferably covers the following types of gene regions:

[0299] Driver gene mutation sites;

[0300] Sites associated with loss of function of tumor suppressor genes;

[0301] High-frequency hotspot abrupt change regions;

[0302] Splice sites or regions related to structural variations.

[0303] In a preferred embodiment, the set of mutation sites may be selected from some or all of the sites listed below, or mutation sites that are functionally equivalent to them.

[0304] Furthermore, the selection of mutation sites can be based on one or a combination of the following criteria:

[0305] It has a high frequency of occurrence in tumors;

[0306] It has a clearly defined biological function or pathogenic significance;

[0307] It can be used for tumor risk assessment or stratified management;

[0308] It is labeled as pathogenic or potentially pathogenic in public databases or literature.

[0309] This invention is not limited to a specific set of genes or sites; any equivalent mutation site selected based on the above functional criteria is within the scope of protection of this invention.

[0310] 2. Mutation Detection Methods

[0311] Mutation detection can be performed using conventional and well-established methods in the field, including but not limited to:

[0312] Multiplex PCR amplification;

[0313] Targeted capture sequencing;

[0314] High-throughput sequencing (NGS);

[0315] Digital PCR (dPCR).

[0316] In a preferred embodiment, the detection method includes the following steps:

[0317] (1) Constructing a library of cfDNA, including end repair, adapter ligation and amplification;

[0318] (2) Targeted capture or multiplex amplification of the target mutation region;

[0319] (3) Perform high-throughput sequencing or quantitative detection;

[0320] (4) Obtain the raw sequencing data and perform subsequent analysis.

[0321] The preferred sequencing depth is ≥2000×, and the preferred depth is ≥5000×.

[0322] 3. Sequencing data processing

[0323] Processing sequencing data, including but not limited to:

[0324] Sequence alignment;

[0325] Repeated sequence removal (based on UMI or algorithmic deduplication);

[0326] Mutation identification and screening;

[0327] Mutant allele frequency (VAF) calculation.

[0328] In a preferred embodiment, the mutation detection threshold is VAF ≥ 0.1%–0.5%.

[0329] 4. Mutation Feature Extraction

[0330] A mutation feature system is constructed based on the detection results, including but not limited to:

[0331] Variant Allele Frequency (VAF);

[0332] Mutation burden;

[0333] Number of mutation sites;

[0334] Hotspot mutation distribution characteristics;

[0335] Mutational combination patterns;

[0336] Spatial distribution characteristics of mutations.

[0337] Furthermore, a comprehensive mutation index can be constructed to characterize the overall degree of mutational abnormality; mutation features, methylation features, and cfDNA fragment features originate from the same cfDNA sample and are fused and analyzed in a unified feature space.

[0338] (iv) DNA methylation detection

[0339] 1. Testing Content

[0340] This invention detects tumor-related DNA methylation features based on cfDNA. The methylation sites are preferably derived from epigenetic regulatory regions related to tumor development and progression, including but not limited to:

[0341] (1) Gene promoter region;

[0342] (2) CpG islands and their adjacent areas (CpG shores and CpG shelves).

[0343] (3) Enhancers and transcriptional regulatory regions;

[0344] (4) Chromatin open regions and nucleosome localization related regions.

[0345] Methylation detection preferably covers the following types of functional regions:

[0346] Abnormally hypermethylated regions in tumor suppressor gene promoters;

[0347] Tumor-associated differentially methylated regions (DMRs);

[0348] Tissue- or tumor-specific methylation marker regions;

[0349] A set of key CpG sites related to gene expression regulation.

[0350] In a preferred embodiment, the methylation detection site may be selected from some or all of the sites listed in Appendix B, or methylation sites that are functionally equivalent to them.

[0351] Furthermore, the selection of methylation sites can be based on one or a combination of the following criteria:

[0352] Significant differences in methylation levels exist between tumor tissues and normal tissues;

[0353] Related to gene expression regulation or tumorigenesis mechanisms;

[0354] It has been shown to have diagnostic, stratification, or predictive value in publicly available databases or literature;

[0355] It exhibits a stable detection signal and low background noise in multi-sample data.

[0356] This invention is not limited to specific methylation sites or gene sets; any equivalent methylation site screened based on the above functional criteria is within the scope of protection of this invention.

[0357] 2. Detection Method

[0358] DNA methylation detection can employ methylation detection techniques known in the art, including but not limited to:

[0359] Bisulfite conversion sequencing.

[0360] Methylation-specific PCR (MSP);

[0361] Methylation-specific quantitative PCR (qPCR);

[0362] High-throughput sequencing targeting methylation.

[0363] In a preferred embodiment, the detection method includes the following steps:

[0364] (1) Treat cfDNA with bisulfite to convert unmethylated cytosine into uracil;

[0365] (2) Primer amplification or targeted capture library construction targeting the target region;

[0366] (3) Perform high-throughput sequencing or real-time fluorescence quantitative detection;

[0367] (4) Calculate the methylation level of each CpG site or region.

[0368] The methylation level is preferably expressed using the β value, which is calculated as follows:

[0369] β value = number of methylated reads / total number of reads.

[0370] Preferred reaction conditions include:

[0371] Conversion temperature: 60–70℃;

[0372] Conversion time: 2–4 hours;

[0373] Conversion efficiency: ≥95%, preferably ≥98%.

[0374] 3. Methylation Feature Extraction

[0375] Methylation features are extracted based on the detection results, including but not limited to:

[0376] Methylation level (β value) at a single CpG site;

[0377] Average regional methylation level;

[0378] Differentially methylated region (DMR) characteristics;

[0379] Multisite combined methylation characteristics;

[0380] Methylation pattern;

[0381] Methylation heterogeneity index.

[0382] Furthermore, based on the above methylation characteristics, a comprehensive methylation index can be constructed to characterize the overall degree of methylation abnormality.

[0383] Methylation features, mutation features, and cfDNA fragment features all originated from the same cfDNA sample and were fused together in a unified feature space.

[0384] In a preferred embodiment, methylation features can be jointly modeled with mutation features and cfDNA fragment features to improve the ability to identify low-abundance tumor signals and the stability of the model.

[0385] (v) Extraction of cfDNA fragment features

[0386] Information on the length distribution of cfDNA fragments was obtained based on sequencing data, and relevant fragment features were extracted.

[0387] In a preferred embodiment, the cfDNA fragment length distribution curve is obtained by statistically analyzing the sequencing fragment lengths.

[0388] Fragment features include, but are not limited to:

[0389] A fragment of approximately 150 bp is used to characterize the features of the mononuclear body protected fragment;

[0390] The periodic distribution characteristics of nucleosomes include peaks at approximately 300 bp and 450 bp.

[0391] Characteristics of fragment length distribution curves;

[0392] Proportion of segments in different length intervals;

[0393] Fragment length ratio characteristics.

[0394] Furthermore, the fragment features preferably include the following computable indices:

[0395] (1) Short Fragment Ratio.

[0396] (2) Long Fragment Ratio.

[0397] (3) Short / Long Ratio.

[0398] (4) Nucleosome periodicity intensity index;

[0399] (5) Segment length distribution skewness or kurtosis.

[0400] Furthermore, a fragmentation index can be constructed to characterize the overall fragmentation features of cfDNA.

[0401] cfDNA fragment characteristics reflect the origin, degradation pattern, and chromatin structure information of cfDNA, and can serve as an important supplement to tumor-related molecular characteristics.

[0402] Fragment features, mutation features, and methylation features originate from the same cfDNA sample and are fused together in a unified feature space, thereby improving the sensitivity and specificity of tumor detection.

[0403] (vi) Data processing and feature construction

[0404] 1. Data Preprocessing

[0405] Standardize raw data from different sources to improve data consistency and comparability.

[0406] (1) Mutation data preprocessing, including:

[0407] Remove data with a sequencing depth below the threshold (preferably <500×);

[0408] Remove low-confidence mutation sites;

[0409] Missing value filling.

[0410] (2) Methylation data preprocessing, including:

[0411] The methylation level of CpG sites was standardized.

[0412] Missing values ​​were filled using the median;

[0413] Use quantile normalization or other standardization methods.

[0414] (3) Clinical data preprocessing (optional), including:

[0415] Z-score standardization of continuous variables;

[0416] Categorical variables are one-hot encoded.

[0417] 2. Multidimensional Feature Construction

[0418] A multidimensional molecular feature system was constructed based on the preprocessed data, including:

[0419] Mutation characteristics;

[0420] Methylation characteristics;

[0421] cfDNA fragment characteristics;

[0422] Optional clinical features.

[0423] Furthermore, the above features are uniformly mapped to the same feature space to construct a fused feature matrix:

[0424] F = [mutation signature + methylation signature + fragment signature + clinical signature]

[0425] In a preferred embodiment, the feature system includes the following comprehensive indicators:

[0426] Mutation Index;

[0427] Methylation Index;

[0428] Fragmentation Index.

[0429] 3. Feature Filtering and Optimization

[0430] The feature matrix is ​​filtered and optimized to improve model stability and prediction performance.

[0431] Filtering methods include, but are not limited to:

[0432] Variance-based feature filtering;

[0433] Feature screening based on statistical tests;

[0434] Feature optimization based on recursive feature elimination (RFE);

[0435] Redundancy removal based on feature correlation.

[0436] Furthermore, the features can be dynamically optimized by combining the feature importance assessment during the model training process.

[0437] By employing the aforementioned data processing and feature construction methods, we can achieve unified expression and fusion of data at different molecular levels, enabling collaborative analysis of multi-omics information within the same feature space, thereby improving the accuracy and stability of tumor risk prediction.

[0438] (vii) Multimodal fusion and model training

[0439] 1. Feature fusion

[0440] Based on the aforementioned multidimensional feature system, mutation features, methylation features, cfDNA fragment features, and optional clinical features are integrated to construct a fusion feature matrix.

[0441] In a preferred implementation, an early fusion strategy is employed to map features from different sources to the same feature space, forming a unified feature vector:

[0442] X = [mutation signature + methylation signature + fragment signature + clinical signature]

[0443] All features originated from the same cfDNA sample and were fused after standardization.

[0444] 2. Model Building

[0445] A predictive model is built based on the feature matrix to assess tumor risk.

[0446] The model can employ commonly used machine learning or deep learning methods in this field, including but not limited to: random forest models, gradient boosting tree models, support vector machine models, neural network models, or deep learning models.

[0447] In a preferred implementation, the model is fitted using training data and its performance is evaluated using cross-validation or independent validation sets; furthermore, features can be dynamically optimized based on feature importance or model output to improve model stability and generalization ability.

[0448] 3. Risk Score Calculation

[0449] A tumor risk score is calculated based on the model output.

[0450] In a preferred implementation, the model output is a probability value P(0–1), and a risk score P is obtained through a linear mapping. =Probability P × 100%, the risk score is used to characterize the relative risk level of the examinee to develop a tumor.

[0451] 4. Risk stratification

[0452] Based on risk scoring The subjects of the examination are stratified, including but not limited to:

[0453] Risk Score A score of 0–30 indicates low risk.

[0454] Risk Score A score of 31–70 indicates a medium risk level.

[0455] Risk Score A score of 71-100 indicates a high risk.

[0456] In other implementations, the stratification threshold can be adjusted according to different population characteristics or application scenarios.

[0457] This invention, through the above-mentioned technical system, achieves effective detection of low-abundance cfDNA signals and synergistic analysis of multi-omics information, thereby improving the sensitivity, specificity, and stability of tumor screening, and has good scalability and industrial application prospects.

[0458] This invention can also be used for early cancer screening in asymptomatic individuals, and is particularly suitable for health checkups and monitoring of high-risk populations. By enriching cfDNA in urine or other bodily fluid samples and simultaneously acquiring mutation, methylation, and fragment characteristics, and performing fusion analysis in a unified feature space, potential risks can be identified before tumors show obvious clinical symptoms or imaging changes. It is applicable to, but is not limited to, cervical cancer, lung cancer, colorectal cancer, breast cancer, gastric cancer, and liver cancer.

[0459] This invention can be used for risk stratification and dynamic management of high-risk groups for cancer, including: people with a family history of cancer; people with persistent HPV infection; people who smoke or drink alcohol for a long time; and patients with chronic inflammation or precancerous lesions.

[0460] The following are sites listed in this invention as a preferred embodiment; in practical applications, some of these sites or their functionally equivalent sites may be selected:

[0461] A. RAS / MAPK pathway (34 points)

[0462] KRAS: G12A, G12C, G12D, G12F, G12R, G12S, G12V, G13C, G13D, G13R, Q61H, Q61K, Q61L, Q61R

[0463] NRAS: G12A, G12C, G12D, G12S, G12V, G13D, Q61H, Q61K, Q61L, Q61R

[0464] HRAS: G12C, G12D, G12S, G12V, Q61L, Q61R

[0465] BRAF: V600E, V600K, K601E, G469A

[0466] B. RTK / PI3K / AKT pathway (42 points)

[0467] EGFR: Exon 18 G719A, Exon 18 G719C, Exon 18 G719S, Exon 19 deletion area, Exon 20 S768I, Exon 20 insertion area, Exon 20 T790M, Exon 20 C797S, Exon 21L858R, Exon 21 L861Q, Exon 18 E709K, Exon 21 exon-end hotspot area

[0468] ERBB2 / HER2: S310F, S310Y, L755S, V777L, Exon 20 insertion area, D769Y

[0469] ERBB3: V104L, E928G, D297Y

[0470] MET: Exon 14 skipping related splice area, Y1230C, Y1230H, D1228N

[0471] FGFR2: S252W, P253R, N550K, K660E

[0472] FGFR3: S249C, R248C, Y373C, K650E

[0473] PIK3CA: R88Q, N345K, C420R, E542K, E545A, E545D, E545G, E545K, H1047R

[0474] AKT1: E17K

[0475] AKT2: E17K

[0476] AKT3: E17K

[0477] MTOR: S2215F, R2505P

[0478] C. Tumor suppressor genes / cell cycle / DNA damage repair (66 sites)

[0479] TP53: R175H, G245S, R248Q, R248W, R249S, R273C, R273H, R273L, R273S, R282W, Y220C, C176F, H179R, M237I, C242S, G266R, V272M, E285K, P278A, splice hotspot area

[0480] PTEN: R130G, R130L, R130Q, R130*, R173C, R173H, Y68H, K267fs hotspot area, Exon 5 hotspot area, Exon 7 / 8 truncation hotspot area

[0481] ARID1A: Q372fs hotspot, R1276* hotspot, G2087fs hotspot, Exon 1 truncation area, central frameshift hotspot area, C-terminal truncation hotspot area

[0482] FBXW7: R465C, R465H, R479Q, R505C, R505G, R505L

[0483] RB1: splice hotspot area, Exon 17 truncation hotspot, Exon 20 truncation hotspot, pocket domain hotspot area

[0484] CDKN2A: R58, D84N, A148T, splice hotspot area

[0485] SMAD4: R361C, R361H, D351H, G386D

[0486] APC: R1450, Q1367*, 1309fs hotspot, mutation cluster region (MCR)

[0487] CTNNB1: S33C, S33F, S37C, S45F, T41A

[0488] PPP2R1A: P179R, R183W, S256F

[0489] POLE: P286R, V411L, S297F, A456P

[0490] D. DNA repair / homological recombination / mismatch repair (28 sites)

[0491] BRCA1: 185delAG hotspot, 5382insC hotspot, Exon 11 frameshift hotspot area, RING domain hotspot area

[0492] BRCA2: 6174delT hotspot, Exon 11 truncation hotspot region, DNA-binding domain hotspot region, C-terminal truncation hotspot region

[0493] ATMs: R337C, V2424G, truncation hotspot area

[0494] ATR: Kinase domain hotspot, truncation hotspot area

[0495] CHEK2: 1100delC, I157T

[0496] PALB2: truncation hotspot 1, truncation hotspot 2

[0497] MLH1: coding hotspot, splice hotspot

[0498] MSH2: coding hotspot, splice hotspot

[0499] MSH6: coding hotspot, splice hotspot

[0500] PMS2: coding hotspot, splice hotspot

[0501] E. Transcriptional regulation / chromatin / telomere and other gynecological-related genes (30 sites)

[0502] TERT promoter: C228T, C250T

[0503] IDH1: R132C, R132H

[0504] IDH2: R140Q, R172K

[0505] EP300: H1451 hotspot, truncation hotspot, HAT domain hotspot

[0506] CREBBP: HAT domain hotspot, truncation hotspot

[0507] KMT2D: SET domain hotspot, truncation hotspot

[0508] KMT2C: truncation hotspot 1, truncation hotspot 2

[0509] NOTCH1: HD domain hotspot, PEST domain hotspot

[0510] FAT1: truncation hotspot 1, truncation hotspot 2

[0511] FAT4: truncation hotspot 1, truncation hotspot 2

[0512] NFE2L2: E79Q, G31R

[0513] KEAP1: G333C, R470C

[0514] STK11: F354L, truncation hotspot

[0515] GNAS: R201C, R201H

[0516] RNF43: G659fs, R117fs

[0517] MAP2K1 / MEK1: K57N, Q56P

[0518] MAP2K2 / MEK2: Q60P

[0519] JAK1: V658F

[0520] JAK2: V617F

[0521] SRC: E527K

[0522] ESR1: Y537S, Y537N, D538G

[0523] PPM1D: Exon 6 truncating hotspot

[0524] SPOP: F133V

[0525] A total of 218 hotspot sites / hotspot regions were detected.

[0526] The following is a set of candidate methylation sites.

[0527] Methylation site hg19

[0528] Cervical cancer: chr1:3568566;

[0529] Multi-cancer screening:

[0530] chr1:156406599、chr12:114841155、chr16:86600277、chr2:68546553、chr3:27765348、chr5:1875863、chr7:27196189、chr7:139168349、chr1:4714033、chr1:158150985、chr11:125036498、chr13:37005997、chr17:35165640、chr2:73518649、chr3:50378122、chr7:27204728、chr7:152622582、chr1:13910245、chr11:133938851、chr14:37136251、chr17:46656093、chr2:105459166、chr3:62356223、chr5:42994093、chr7:27205159、chr7:153445808、chr1:14925876、chr1:223302841、chr12:25056207、chr14:60978096、chr17:48042533、chr2:111876726、chr3:139653774、chr5:78365599、chr7:27283528、chr7:156797121、chr1:33219657、chr1:248020908、chr12:25101965、chr14:61110649、chr18:55104979、chr2:121412136、chr4:8863338、chr7:29603279、chr7:156810972、chr1:50882021、chr10:17271503、chr12:50297721、chr14:102247968、chr2:154335110、chr4:13526852、chr5:115152494、chr7:157478340、chr1:63785841、chr14:102248105、chr2:175199881、chr4:24801842、chr5:178421646、chr7:38671009、chr8:17271027、chr1:77333903、chr10:105036745、chr12:63544103、chr15:48937573、chr20:21086198、chr4:141677849、chr6:28411301、chr8:41167045、chr12:95942193, chr16:1030641, chr19:30715705, chr20:21685481, chr4:147560279, chr6:32119691, chr7:7 0596939, chr8:57359311, chr1:119526933, chr10:131770969, chr12:108169051, chr16:25704534, chr20:4514 2329, chr4:168154865, chr6:137814788, chr8:70981935, chr1:145562905, chr10:133110829, chr12:11391751 8. chr16:51190145, chr22:33197647, chr4:176987040, chr7:1272935, chr7:100318589, chr8:70982082, chr1:1 51811370, chr11:31837607, chr12:113917635, chr16:66613195, chr19:46997242, chr4:176987240, chr7:1294 25828、chr8:97505843、chr1:119532849、chr1:170630712、chr1:63786015、chr1:63795374、chr10:118892493、 chr14:57278227, chr17:37321626, chr2:19556813, chr2:223161815, chr2:45231868, chr2:71116118, chr20:2 1492828, chr20:25062309, chr3:157812320, chr14:38725015, chr6:10417655, chr8:55370844, chr8:99986879,

[0531] The advantages of this invention are:

[0532] 1. It integrates multidimensional molecular markers such as mutation features, methylation features and fragment features of circulating cell-free DNA (cfDNA), breaking through the limitations of traditional single feature detection. It can capture molecular signals in the process of tumor occurrence and development more comprehensively and accurately, and significantly improve the sensitivity and specificity of early tumor screening.

[0533] 2. By employing standardized preprocessing, feature selection, and optimization strategies, combined with multimodal fusion algorithms, cross-dimensional features are organically integrated, effectively addressing the technical bottleneck of collaborative analysis of multi-omics data and enhancing the model's ability to capture tumor molecular patterns and its generalization performance. Simultaneously, the risk prediction system built based on the machine learning model, through quantitative risk scoring and stratification standards, enables accurate assessment of the tumor risk of examinees, providing clear and interpretable decision-making basis for clinicians, assisting in the early detection and intervention of tumors, and possessing significant clinical translational value.

[0534] The above-disclosed embodiments are merely a few specific examples of the present invention, but the present invention is not limited thereto. Any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA, characterized in that, include: The sample processing module is used to receive body fluid samples from the subject and to extract and enrich cfDNA from the body fluid samples. The molecular detection module is used to perform multi-omics detection on the cfDNA and obtain multidimensional molecular feature data of the cfDNA, including at least mutation features, methylation features and cfDNA fragment features. The data processing module is used to preprocess the multidimensional molecular feature data obtained by detection to obtain mutation index, methylation index and fragment feature index, and map each multidimensional molecular feature data to the same feature space to construct a unified fusion feature matrix. The multimodal fusion machine analysis module performs multimodal fusion analysis on the fusion feature matrix based on the multimodal fusion machine learning model, and generates the fusion analysis results. The output module outputs the tumor risk score and risk level of the examined subject based on the results of the fusion analysis.

2. The tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA according to claim 1, characterized in that, The molecular detection module enriches the cfDNA in the body fluid sample using methods such as selective precipitation based on liquid phase partitioning or polymer phase separation, or magnetic bead method, column purification method, ultrafiltration method, and combinations thereof based on solid-phase carrier.

3. The tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA according to claim 1, characterized in that, The mutation characteristics include mutation allele frequency, mutation load, number of mutation sites, hotspot mutation distribution characteristics, mutation combination patterns, and mutation spatial distribution characteristics. The methylation features include the methylation level of a single CpG site, the average level of regional methylation, the features of differentially methylated regions, the combined methylation features of multiple sites, the methylation pattern, and the methylation heterogeneity index. The cfDNA fragment characteristics include single nucleosome protected fragment characteristics, nucleosome periodic distribution characteristics, fragment length distribution curve characteristics, fragment ratios in different length intervals, and fragment length ratio characteristics.

4. The tumor screening system based on multidimensional feature fusion of circulating cell-free DNA according to claim 3, characterized in that, The data processing module includes the following methods for preprocessing the multidimensional molecular feature data: For the mutation features, low-quality site data with sequencing depth below a preset threshold are removed, low-confidence mutation sites are removed, and missing mutation site data are filled with zero to obtain the mutation index. For the methylation features, samples with bisulfite conversion efficiency lower than a preset conversion threshold were removed, and median filling was used for missing CpG site methylation data. Methylation levels were standardized using quantiles to eliminate batch effects and obtain the methylation index. For the cfDNA fragment features, abnormal fragment data with lengths below a first preset threshold or above a second preset threshold are removed, as are low-quality fragment data with sequencing quality or alignment quality below a preset quality threshold. Based on the retained fragments, the fragment length distribution curve, the proportion of fragments in different length intervals, the fragment length ratio, the characteristics of single nucleosome protected fragments, and the periodic distribution characteristics of nucleosomes are statistically analyzed. Missing fragment feature data are filled with the mean or median. The fragment feature data are then standardized or normalized to obtain the fragment feature index.

5. The tumor screening system based on multidimensional feature fusion of circulating cell-free DNA according to claim 4, characterized in that, The process by which the data processing module constructs the fused feature matrix includes: a. The pre-processed and screened mutation features, methylation features, cfDNA fragment features, and optional clinical information features are vectorized and concatenated to form a unified feature vector; b. Perform Min-Max normalization on the concatenated feature vector to scale all feature values ​​to the [0,1] interval: =( - ) / ( - ), where X is the original feature vector, The minimum value of this eigenvector, The maximum value of this eigenvector, These are the normalized eigenvector values; c. Construct a fusion feature matrix of dimension N×D. , where N is the number of samples and D is the total number of features after screening.

6. The tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA according to claim 5, characterized in that, The steps for multimodal fusion analysis performed by the multimodal fusion machine analysis module include: The fused feature matrix The data are input into a machine learning model, which learns mutation features, methylation features, and fragment features through training. Output the probability of cancer occurrence P, and calculate the risk score according to the formula. =Probability P × 100%; The output module scores the risk based on a preset risk threshold. It is divided into low-risk, medium-risk, and high-risk levels.

7. The tumor screening system based on multidimensional feature fusion of circulating cell-free DNA according to claim 6, characterized in that, The multimodal fusion machine learning model can be any one of the following: random forest model, gradient boosting tree model, support vector machine model, or deep learning model.

8. The tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA according to claim 5, characterized in that, The fused feature matrix is ​​filtered and optimized, and the filtering method includes: Variance-based feature filtering; Feature screening based on statistical tests; Feature optimization based on recursive feature elimination; Redundancy removal based on feature correlation.

9. The tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA according to claim 5, characterized in that, Mutation detection methods for obtaining the mutation characteristics include multiplex PCR amplification, targeted capture sequencing, high-throughput sequencing, and digital PCR; The methylation detection methods for obtaining the DNA methylation characteristics include bisulfite conversion sequencing, methylation-specific PCR, methylation-specific quantitative PCR, and targeted methylation high-throughput sequencing. The cfDNA fragment extraction method for obtaining the cfDNA fragment characteristics includes pretreating the body fluid sample to remove cellular components and impurities, extracting and enriching cfDNA from the pretreated sample, constructing a library and sequencing the cfDNA to obtain cfDNA sequencing data.

10. The application of the tumor screening system based on the fusion of multidimensional features of circulating cell-free DNA as described in any one of claims 1-9 in early tumor screening, tumor risk assessment or auxiliary diagnosis.

Images