Colorectal cancer early-stage accurate diagnosis model based on plasma cfDNA multi-dimensional fragment omics and construction method and application of colorectal cancer early-stage accurate diagnosis model

Abstract

The invention discloses a colorectal cancer early-stage accurate diagnosis model based on plasma cfDNA multi-dimensional fragment omics and a construction method and application thereof. The method comprises the following steps: a) collecting and processing a sample; b) performing whole genome sequencing (WGS) and bioinformatics processing; c) fragment omics feature extraction; and d) constructing and optimizing a machine learning model. According to the method, cfDNA is extracted from a plasma sample of a colorectal cancer patient, the fragment size distribution and terminal sequence characteristics of the cfDNA are analyzed by utilizing a high-throughput sequencing technology, and then a specific cfDNA fragment mode related to CRC is identified through bioinformatics analysis. Based on the identified cfDNA fragment features, an economically feasible and highly sensitive diagnosis model is constructed, and early diagnosis of CRC is realized.

Description

Technical Field

[0001] The present invention belongs to the field of medical detection technology and relates to an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics, and a construction method and application thereof. Background Art

[0002] Colorectal cancer (CRC) is one of the most common cancers worldwide, with an estimated 1.9 million newly diagnosed cases of CRC in 2020, resulting in approximately 940,000 deaths, making it the second leading cause of cancer-related deaths. From a demographic perspective, China has the largest number of new CRC cases, with an estimated 560,000 cases, followed by the United States with 160,000 cases. In addition, the incidence of CRC in people under 50 years of age is on the rise, with an annual increase of approximately 2.2% in young people from 2010 to 2017. The early symptoms of CRC are usually not obvious, and its slow progression makes population-based screening and prevention strategies particularly important.

[0003] Despite significant advances in CRC screening technology, only about 40% of CRC patients can currently achieve early diagnosis, and the five-year survival rate of these patients is as high as 90%. However, about 20% of CRC patients develop metastatic disease, and the five-year survival rate drops sharply to 12.5%. Colonoscopy is the gold standard for CRC screening, but this method has some limitations, including high cost, high dependence on medical resources, and public concerns caused by the discomfort caused by the examination process. Other less invasive screening methods, such as fecal occult blood test (FOBT), have been recommended for CRC screening by multiple guidelines, but their sensitivity is limited and the test results vary greatly between different laboratories. Although the fecal immunochemical test (FIT) is more advanced than the traditional FOBT, it still has shortcomings in sensitivity and consistency. Therefore, it is of great significance to develop an accurate and non-invasive method for CRC diagnosis.

[0004] In recent years, liquid biopsy based on plasma circulating free DNA (cfDNA) has shown great potential in the early detection of solid tumors including CRC. cfDNA is mainly derived from cell apoptosis, necrosis or secretion processes, and contains rich genetic and epigenetic information, such as mutations, copy number variations, chromosome rearrangements and methylation abnormalities. The characteristic differences between cancer cells and normal cells in cfDNA provide new opportunities for cancer detection. However, due to the low blood concentration of cfDNA in the early stages of cancer, its detection becomes complicated. Although deep sequencing can accurately locate genetic aberrations, its high cost limits its widespread application.

[0005] Although methylation-based detection methods (such as whole-genome bisulfite sequencing) show some potential for early cancer detection, these methods are technically complex and have lengthy procedures, which limits their clinical application. Therefore, there is an urgent need for an alternative detection method that simplifies the technical process while maintaining high sensitivity and specificity.

[0006] As an emerging field, fragmentomics focuses on cfDNA fragmentation patterns and shows great potential, especially when deep whole genome sequencing (WGS) data is not available. The fragmentation pattern of cfDNA is not evenly distributed in the genome and is affected by chromatin structure, providing valuable insights into nucleosome positioning and gene expression. The early fragmented DNA evaluation (DELFI) technology achieved the evaluation of cfDNA fragments in a 5Mb window. Features such as coverage and size have successfully revealed the differences between cancer patients and healthy individuals.

[0007] Fragmentomics studies have also found that the terminal coordinates and terminal motifs of cfDNA fragments have specific differences between cancer and healthy states. Recent studies have shown that the transcription factor binding patterns in cfDNA can also distinguish cancer patients from healthy individuals with a high accuracy. Therefore, fragmentomics analysis based on cfDNA provides a new direction for early non-invasive detection of cancer.

[0008] In order to overcome the limitations of the existing technology, the present invention proposes a new detection model, which utilizes the fragment omics characteristics of cfDNA and combines five optimized fragment features to develop an integrated stacked machine learning model, thereby achieving a highly sensitive and economically feasible non-invasive detection method for colorectal cancer. Summary of the invention

[0009] Technical problem to be solved: The present invention provides a method for constructing and applying an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics. The method extracts cfDNA from plasma samples of colorectal cancer patients, analyzes its fragment size distribution and terminal sequence characteristics using high-throughput sequencing technology, and then identifies specific cfDNA fragment patterns related to CRC through bioinformatics analysis. Based on the identified cfDNA fragment characteristics, an economically feasible and highly sensitive diagnostic model is constructed to achieve early diagnosis of CRC.

[0010] Technical solution: A method for constructing an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics, including the following clear steps:

[0011] a) Sample collection and processing:

[0012] Peripheral venous blood samples from patients suspected of colorectal cancer and healthy subjects were collected using EDTA anticoagulant tubes and centrifuged at 2000 rpm for 10 min at 4°C to extract the upper plasma and remove cell debris;

[0013] cfDNA was extracted from plasma using the QIAamp circulating nucleic acid kit;

[0014] The concentration of extracted cfDNA was measured using the Qubit dsDNA assay kit;

[0015] b) Whole genome sequencing (WGS) and bioinformatics processing:

[0016] The extracted cfDNA was sequenced using the Illumina NovaSeq platform for low-coverage whole-genome sequencing to generate sequencing data with a sequencing depth of at least 0.1x;

[0017] The sequencing data were processed by bioinformatics, including: trimming reads using Trimmomatic software, setting the parameters to ILLUMINACLIP:TruSeq3-PE.fa:2:30:10SLIDINGWINDOW:4:15MINLEN:36; aligning reads to the human reference genome (GRCh37 / UCSC hg19) using the BWA tool, setting the default parameters; and removing PCR duplicates using the Picard tool, setting the parameter to REMOVE_DUPLICATES=true;

[0018] c) Fragmentomics feature extraction:

[0019] The following cfDNA fragment omics features were extracted from the WGS data after bioinformatics processing:

[0020] Copy number variations (CNVs): Sequencing depth was assessed using a hidden Markov model and compared with the baseline of a healthy control group in the same interval. The log2 ratio between the sample and the baseline in each interval was calculated to obtain the copy number variation;

[0021] Fragment size coverage (FSC): For each fragment size group: short fragments <100 bp and long fragments ≥150 bp, the normalized coverage score (z-score) within each 5 Mb interval was calculated;

[0022] Fragment size distribution (FSD): The coverage of cfDNA fragments from 100 bp to 220 bp was examined in intervals of 5 bp on 39 chromosome arms. The original fragment size distribution characteristics were extracted and converted into z-scores after comparison with the overall mean.

[0023] Nucleosome footprint (NF): Analysis of transcription factor binding sites using nucleosome coverage: 1) Center coverage: measure the coverage value at 30 bp away from the specified binding site; 2) Average coverage: calculate the average coverage within the ±1000 bp region of the binding site; 3) The amplitude of the binding site is determined by fast Fourier transform analysis;

[0024] Fragment-based methylation (FM): The methylation status of the sample was determined by analyzing the cleavage pattern around the CpG site. In the Alu region, the fragment ratios of the 5'-end motifs of CGN, NCG, CGC, CGN / NCG, CGC / NCG, CGN / CGC, NCG(C1), and CGC(C2) were analyzed as feature sets, where N represents any nucleotide: A, C, T, or G;

[0025] d) Machine learning model building and optimization:

[0026] Based on the omics features of the extracted cfDNA fragments, an integrated stacking machine learning strategy was used to build a classification model for early diagnosis of colorectal cancer by combining five basic algorithms: generalized linear model, gradient boosting machine, random forest, deep learning algorithm or XGBoost.

[0027] Through the H2O AutoML framework, for each feature type, 200 base models were generated using random grid search within a wide range of hyperparameters for each algorithm, among which the top five base models with the area under the ROC curve (AUC) cancer score were compiled into a matrix, and then a second-level generalized linear model (GLM) algorithm was applied to create the final model.

[0028] The model building method also includes cross-validation, hyperparameter adjustment, and feature selection of the constructed model to further improve the diagnostic accuracy and generalization ability of the model.

[0029] The model building method also includes a step of performing a performance evaluation on the constructed model, which includes calculating the sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (ROC) (AUC) indicators to verify the diagnostic performance of the model.

[0030] The above construction method obtained an accurate early diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics.

[0031] A colorectal cancer early diagnosis system comprises: (1) an acquisition unit, used to acquire marker information of a sample to be tested; (2) a detection unit, used to use the marker information of the sample to be tested as an input feature of a machine learning model, and then use the prediction model trained by the method to obtain the result of the sample to be tested; the result is the cancer risk score of the sample to be tested.

[0032] An electronic device comprises: a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to apply the above system.

[0033] An executable storage medium stores computer program instructions. When the computer program instructions are executed on a processor, the processor executes the above system.

[0034] Beneficial effects: (1) Model construction and evaluation analysis

[0035] The five feature types mentioned above were used to generate models implemented by five basic algorithms, namely generalized linear model (GLM), gradient boosting machine (GBM), random forest (RF), deep learning (DL) and XGBoost. 200 base models were generated for each feature-algorithm pair. Each algorithm used 360 healthy and CRC samples from the training cohort to train its classifier and calculate the corresponding cancer score for each sample. The higher the score, the higher the probability of cancer. Using an ensemble approach, the cancer scores from the first five base models of each feature type were merged into a matrix, and then the second layer of GLM algorithm was used to create the base model. Fold cross-validation was performed on each of the five base models to prevent model overfitting. The process of model training was performed entirely in the training dataset, and the independent validation dataset remained locked until the model was completed. In addition, an independent validation cohort was used to evaluate the performance of the multidimensional ensemble machine learning model. Two critical values, FRAGTECT-1 / 2, were determined to generate sub-models with ultra-high specificity (99.5%) and high specificity (98%).

[0036] To assess the reproducibility of the model, the reproducibility of the model was evaluated by comparing the results of batch 1 and batch 2 for each participant using three scores obtained from each participant in the same batch. To further demonstrate the reproducibility of the prediction, the model was tested in a test involving four independent healthy participants to predict their cancer scores from cfDNA collected at four different time points during the day: before eating and exercising, and after eating and exercising. To understand the performance of our model in different CRC patient subgroups, the diagnostic sensitivity of the model for predicting each subgroup was calculated. The subgroups included age, sex, tumor stage, tumor location, tumor grade, tumor size, lymph node metastasis, microsatellite (MS) status, carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA 19-9), carbohydrate antigen 125 (CA 125), and alpha-fetoprotein (AFP). MS status was classified as microsatellite stable (MSS) or microsatellite unstable (MSI). Finally, the model was tested in a prospective cohort to confirm its performance and real-world applicability. The ranges of biomarkers in the references are as follows: CEA ≤ 5 ng / mL, CA 125 ≤ 24 units / mL, CA19 -9 ≤ 30 units / mL, and AFP < 7 ng / mL.

[0037] (2) Performance evaluation of the ensemble superposition model

[0038] The constructed model showed the highest predictive ability with an AUC of 0.997 (95% CI: 0.994-1.000) in the training cohort and was validated in an independent cohort with an AUC of 0.986 (95% CI: 0.973-0.999) ( Figure 2 : A). The base model used any single feature that produced an AUC ranging from 0.903 to 0.964, demonstrating the superior predictive power of the overall model ( Figure 2 :B). In the training data set that generated FRAGTECT-1, a cancer score cutoff of 0.60 was applied to achieve a specificity of 98.0%. An additional cutoff was set to a cancer score of 0.68 to capture a specificity of 99.5% in the training data set that generated FRAGTECT-2 (Table 3). These cutoffs maintained specificities of 0.9747 (95% CI: 0.9281, 0.9948) (Table 2) and 0.9832 (95% CI: 0.9406, 1.00) (Table 3) in the validation cohort, respectively. Accordingly, the resulting sensitivity in the validation data set was 0.9488 (95% CI: 0.9051, 0.9764), as predicted by FRAGTECT-1, reduced to 0.9231 (95% CI: 0.8590, 0.9642), as predicted by FRAGTECT-2. The distribution of cancer scores between cancer and healthy individuals was compared using a box plot, which was used as a visualization of the predictive power of the two models ( Figure 2 :C). The analysis showed that cancer patients showed significantly higher predicted cancer scores compared with healthy patients. Therefore, the multifaceted ensemble stacking model incorporating five cfDNA features showed better predictive ability.

[0039] (3) Economical and operable

[0040] FRAGTECT-1 / 2 correctly predicted disease status in triplicate samples of both cancer and healthy samples, regardless of health status at the time of sample collection ( Figure 5 ). When evaluating the durability of the model at lower WGS coverage, the AUC remained high (≥0.9185) even after reducing the WGS coverage to 4×, 3×, 2×, 1×, and 0.5× (Table 4). In summary, these tests highlight the stability and consistency of the ensemble stacking model in various scenarios. Based on low-coverage whole genome sequencing technology, the present invention reduces the cost of detection while simplifying the operating process, enabling the method to be promoted on a large scale in actual medical applications.

[0041] (4) Performance of the prediction model in cancer sample subgroups

[0042] The performance of the model in the validation cohort was further evaluated by analyzing various CRC subgroups. Figure 3: As shown in A and Table 5, FRAGTECT-1 reliably detected cancer patients at an early stage, with sensitivities of 0.8824 (95% CI: 0.6356-0.9854) and 0.9167 (95% CI: 0.8002-0.9768) in stages 0-I and II, respectively. FRAGTECT-2 showed similar performance (Table S5). An enhancement in predictive sensitivity was observed in the validation dataset as cancer stage progressed. FRAGTECT-1 showed superior sensitivity in stage III (0.9535, 95% CI: 0.8419-0.9943) compared to FRAGTECT-2 (0.9302, 95% CI: 0.8094-0.9854). Both models were able to predict stage 4 patients with 100% sensitivity (FRAGTECT-1 / 2: 95% CI: 0.6306-1). Notably, in patients with colon cancer, both models showed improved sensitivity in detecting right-sided tumors compared to left-sided tumors. The models had high sensitivity in CRC patients with lymph node metastasis, with FRAGTECT 1 / 2 rates of 0.9545 (95% CI: 0.8453-0.9944) and 0.9318 (0.8134-0.9857), respectively. Sixteen cases had tumors larger than 3 cm in size, while FRAGTECT-2 showed higher sensitivity when the tumor size was less than 3 cm. Biomarkers including CEA, CA125, CA19-9, and AFP are commonly used as indicators for cancer diagnosis and were tested for postoperative monitoring in CRC and 334 healthy participants ( Figure 3 :EF, Table 6). In this study, these biomarkers were evaluated by ROC analysis ( Figure 7 ) to evaluate its diagnostic ability to detect cancer in the validation cohort. The results showed that the AUC of CEA was 0.805 (95% CI: 0.749-0.861), while the AUC values ​​of other biomarkers ranged from 0.635-0.501. In addition, the FRAGTECT model was validated in different clinical scenarios, showing its applicability to different subgroups of colorectal cancer patients and its ability to remain stable in actual environments, showing its potential for wide application in a variety of populations, and its detection effect on subgroups such as right-sided colon cancer and patients with lymph node metastasis was particularly prominent.

[0043] (5) Model performance in prospective cohorts

[0044] The performance of the model was further evaluated using a cohort of prospectively collected samples. The predictive ability of the developed model was confirmed in this prospective cohort, achieving an AUC of 0.969 (95% CI: 0.947, 0.990) ( Figure 4:A). For the FRAGTECT-1 model, the specificity in the prospective cohort was 0.9558 (95% CI: 0.8998356-0.9855) (Table 1) and the sensitivity was 0.9147 (95% CI: 0.8525-0.9567). In contrast, the FRAGTECT-2 model had a target specificity of 99.5% in the training dataset, a specificity of 0.9646 (95% CI: 0.9118-0.9903) and a sensitivity of 0.8372 (95% CI: 0.7620359-0.8963) in the prospective dataset (Table 7). In addition, consistent with the observations from the validation cohort, there was a significant improvement in the predictive sensitivity in the validation dataset with the progression of cancer stage ( Figure 4 :C).

[0045] (6) Significant clinical benefits

[0046] To evaluate the potential clinical benefits that this model could provide in the real world, simulations were performed using previously published data. Based on the incidence and 5-year survival rates of CRC in China, FRAGTECT-1 ( Figure 7 ) and -2( Figure 8 ) will diagnose 98% of Chinese CRC patients at stage I, while 47% of them will be diagnosed at a late stage (stage III / IV) under usual care, increasing the 5-year survival rate from 66% to 84%. The potential impact of the model on the US population was also evaluated, showing that 26% and 21% of stage III and IV CRC patients, respectively, were diagnosed at an early stage of the disease (stage I). Fig. 9 ). These results show that the model can promote the early diagnosis of CRC patients. This can significantly improve the early diagnosis rate of CRC patients, thereby improving the survival expectancy of patients, and provide a solution with substantial clinical value for the early screening and diagnosis of colorectal cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the non-invasive detection method for colorectal cancer based on cfDNA fragmentomics. In this study, a total of 422 plasma samples from patients with colorectal cancer and 416 plasma samples from healthy volunteers were collected. All 838 samples were tested by WGS and then bioinformatics analysis was performed. Five cfDNA features, FSC, FSD, CNV, NF and FM, were extracted from the sequencing sequences. Each feature was integrated and learned using five algorithms, GLM, GBM, RF, DL and XGBoost, to finally obtain a basic model. By training the GLM algorithm, these basic models were integrated into the final ensemble superposition model in turn.

[0048] Figure 2Performance evaluation of the ensemble superposition model. A) ROC curves of the ensemble superposition model for training and independent validation datasets; B) ROC curves of the base model constructed based on individual features (FSC, FSD, CNV, NF, FM) and its ensemble superposition model. C) The ensemble superposition model predicts the distribution of cancer scores for colorectal cancer and healthy groups in the training and validation cohorts. Two cutoff points (98% and 99.5%) after training specificity were selected to generate two sub-models, FRAGTECT-1 / 2, as shown by the dotted lines.

[0049] Figure 3 To show the sensitivity of FRAGTECT-1 / 2 in subgroups of colorectal cancer patients in the validation cohort, the sensitivity expressed as a percentage and its 95% CI (represented by a bar graph) were calculated according to A) cancer stage, B) tumor location, C) tumor size, D) lymph node metastasis status, E) CEA level, and F) CA19-9 level.

[0050] Figure 4 To evaluate the performance of the ensemble superposition model in a prospective cohort. A) ROC curves of the ensemble superposition model for training and independent validation datasets; B) The ensemble superposition model predicts the distribution of cancer scores in colorectal cancer and healthy groups in the training and validation cohorts. Two cutoff points (98% and 99.5%) after training specificity were selected to generate two sub-models, FRAGTECT-1 / 2, as shown by the dotted lines; C) The dot plot shows the sensitivity of FRAGTECT-1 in patients with colorectal cancer at different stages in the prospective cohort.

[0051] Figure 5 Figure 1 shows the stability of the ensemble superposition model. A) Box plots evaluating the repeatability and reproducibility of the prediction model. Based on two batches of data (each batch containing three colorectal cancer patients and three healthy subjects), the performance of the model was evaluated by intra-batch analysis and inter-batch analysis; B) Box plots evaluating the stability of the model in predicting healthy individuals in four daily scenarios. FRAGTECT-1 / 2 is represented by dotted lines, which predicts cancer status in a binary manner.

[0052] Figure 6 ROC curves for serum biomarker testing in the validation cohort. ROC curve analysis was used to evaluate the ability of serum biomarkers (including carcinoembryonic antigen CEA, CA19-9, CA125, and AFP) to discriminate between colorectal cancer patients and healthy subjects.

[0053] Figure 7 This is a real-world analysis of FRAGTECT-1 in the Chinese population. The Sankey diagram shows the proportion of patients with each cancer stage at diagnosis, with or without FRAGTECT-1 intervention, and potential outcomes in the real world.

[0054] Figure 8 To analyze disease stage shifts in a real-world setting using the FRAGTECT-2 test in the Chinese population. The Sankey diagram shows the proportion of patients with each cancer stage at diagnosis, with or without the FRAGTECT-2 test, and the possible outcomes of these patients in the real world.

[0055] Fig. 9 This is a stage-by-stage analysis of the real-world FRAGTECT-1 intervention in the US population. The Sankey diagram shows the proportion of patients with each cancer stage at diagnosis with and without the FRAGTECT-1 cutoff intervention, as well as potential real-world outcomes. DETAILED DESCRIPTION

[0056] In order to better understand the present invention, the present invention is further described below in conjunction with specific embodiments and drawings, but it should be noted that the implementation of the present invention is not limited thereto.

[0057] The reagents and raw materials used in the present invention are all commercially available or can be prepared according to literature methods. The experimental methods without specific conditions specified in the examples of the present invention are all based on conventional conditions or the conditions recommended by the manufacturer.

[0058] Example 1

[0059] A method for constructing an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics. The specific technical solution is as follows: Figure 1 :

[0060] 1. Sample collection and processing

[0061] Peripheral venous blood samples were collected from the subjects using EDTA anticoagulant tubes and centrifuged at 2000 rpm for 10 minutes at 4°C to extract the upper plasma and remove cell debris. Plasma samples were processed using the QIAamp Circulating Nucleic Acid Kit (Qiagen) according to the manufacturer's instructions for free circulating deoxyribonucleic acid (cfDNA) extraction. To ensure that there was enough cfDNA for subsequent analysis, the Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific) was used and the cfDNA concentration in plasma was assessed according to the manufacturer's guidelines. After the extracted nucleic acid was quantified by cfDNA, the total amount of nucleic acid was not less than 5 ng, that is, the sample quality control was qualified. If the concentration of the extracted nucleic acid is less than 0.4 ng / uL, it should be dried and concentrated below 45°C to a concentration of 0.4 ng / μL or more.

[0062] 2 Whole genome sequencing (WGS) and bioinformatics analysis

[0063] 5-10 ng cfDNA per sample was used for multiplex polymerase chain reaction (PCR) whole genome sequencing (WGS) library construction using the library construction reagents included in the KAPA Hyper Prep Kit (KAPA Biosystems). After the library was quantified using the KAPASYBR FAST qPCR Master Mix (KAPA Biosystems) kit, 150 base pair (bp) paired-end sequencing was performed on the NovaSeq6000 platform. The sample library was quantified and normalized to the sequencing pool. The pooled sample library was fluorescently quantified, loaded on the sequencing flow cell, and sequenced on the NovaSeq 6000 platform. Clinical negative controls were introduced to ensure that there was no DNA contamination during library preparation.

[0064] To ensure data quality, quality control of sequencing results was performed. Trimmomatic was used for read trimming, and then PCR duplicates were removed using the Picard tool (Broad Institute, MA, USA). The trimmed reads were aligned to the human reference genome (GRCh37 / UCSC hg19) using the Burrows-Wheeler Aligner. To standardize the data and mitigate the impact of different sequencing depths, a downsampling procedure was used. Samples should have no less than 5 times the average coverage depth to pass sequencing quality control. Coverage depths exceeding 5X were reduced to a uniform 5X to ensure consistency between samples.

[0065] 3. Fragmentomics feature extraction

[0066] A variety of cfDNA fragmentomics features were extracted from the processed WGS data, including copy number variations (CNVs), fragment size coverage (FSC), fragment size distribution (FSD), nucleosome footprint (NF), and fragment-based methylation (FM). The fragmentomics feature profiles collected from the training dataset were used to build a multidimensional machine learning model for cancer detection and localization. Among them, CNVs were calculated using a hidden Markov model to evaluate sequencing depth, FSC was calculated based on the coverage of short and long fragments, FSD was calculated by analyzing fragment length patterns within 5bp intervals, NF was calculated through a nucleosome coverage analysis framework, and FM was based on analyzing the cleavage pattern around CpG sites.

[0067] 4. Machine Learning Model Building and Optimization

[0068] A custom analysis framework was designed using a multi-layer stacking ensemble approach, which processed WGS data through internal scripts to extract a set of multidimensional genomic and fragmentomics features, including CNVs, FSC, FSD, NF, and FM. Each feature was analyzed independently to generate basic data. Then, an integrated stacking machine learning strategy was used to combine multiple basic algorithms (such as generalized linear models, gradient boosting machines, random forests, XGBoost, convolutional neural networks, and fully connected neural networks) to build a classification model. Each algorithm generates multiple basic models, and the outputs of each basic model are integrated into a cancer risk score through an integrated approach.

[0069] 4.1 Copy Number Variation (CNVs) Model

[0070] Copy number variations (CNVs) refer to changes in the number of copies of DNA fragments in the human genome, that is, the increase (amplification) or decrease (deletion) of DNA fragments of at least 1kb in size. These variations constitute an important part of the genetic structural variation landscape. To analyze CNVs, the genome of each sample was first divided into 1-Mb (megabase) intervals. For each of these intervals, the sequencing depth after interval-level GC correction was evaluated by a hidden Markov model. The model compares the sequencing depth of the sample with the baseline from a healthy sample in the same interval. Subsequently, the log2 ratio between the sample and the baseline in each interval was calculated to reveal the pattern of copy number variations (CNVs).

[0071] To model copy number variation (CNV) features for a binary base model, a convolutional neural network (CNN) was used that excels at extracting spatial hierarchies and patterns from data. The CNN architecture consists of convolutional blocks that contain convolutional layers, ReLU activation functions, batch normalization, and pooling layers. These blocks process 1-Mb intervals of data sequentially to capture and abstract key features indicative of CNVs. After feature extraction by the convolutional blocks, a fully connected layer classifies the detected CNV patterns. This combination provides a sophisticated approach to identifying and classifying CNV features, making an important contribution to understanding the full picture of genetic structural variation.

[0072] 4.2 Fragment size coverage (FSC) model

[0073] Fragment size coverage (FSC) was generated by comparing the coverage of short fragments (65-150 bp) and long fragments (150-220 bp). The method for extracting fragment size features was adapted based on the method described by D.Mathios et al. First, a non-parametric method was used to correct for GC content and library size at the fragment level. Subsequently, the genome was divided into 100 kb intervals (bins). Next, the coverage of the two fragment size groups within each 100 kb interval was calculated and corrected for GC content. The coverage of every 50 consecutive 100 kb intervals was then merged to calculate the coverage within the corresponding 5 Mb window, and finally 541 windows were obtained (windows containing less than 50 intervals were discarded). For each fragment size group (short and long fragments), a standardized coverage score (z-score) within each 5 Mb interval was calculated.

[0074] To analyze the FSC data, a convolutional neural network (CNN) architecture was employed. By processing z-score sequences from 5Mb windows, the CNN was able to skillfully identify features associated with fragmentation patterns, thereby providing refined insights into genomic structural variation driven by fragment size differences.

[0075] 4.3 Fragment size distribution (FSD) model

[0076] Fragment size distribution (FSD) features were extracted at the chromosome arm level, examining the coverage of cfDNA fragments from 100 bp to 220 bp, with each 5 bp being an interval (e.g., 110-114 bp, 115-119 bp, …, 215-220 bp; 24 intervals in total). A total of 39 chromosome arms were examined, resulting in 936 (39 arms × 24 intervals) FSD features. Since the short arms of the five satellite chromosomes have remained largely unsequenced to date, they were not included in the analysis. The raw coverage of the FSD was also converted to a z-score by comparing it to the overall mean to ensure that the model focuses on the underlying patterns in the data and is not affected by differences between samples. This approach enables enhanced detection of high-resolution chromosome-level patterns, which may further reveal differences between cancer and non-cancer groups.

[0077] A convolutional neural network (CNN)-like architecture was used to systematically analyze FSD features. By processing z-score transformed sequences corresponding to stepwise varying cfDNA fragment lengths, the CNN effectively identified patterns indicative of chromosomal abnormalities. This approach highlights the potential of using a unified CNN architecture to mine deep insights from genomic data variation, particularly in distinguishing pathological states based on chromosomal-level fragment size distributions.

[0078] 4.4 Nucleosome footprint (NF) model

[0079] Nucleosomes protect DNA from degradation in the bloodstream and are the structural units that package and arrange DNA to form chromatin. These nucleosomes are essential for the integrity of cfDNA, as unprotected cfDNA is susceptible to damage, which results in reduced sequencing coverage. Nucleosomes preferentially bind to accessible genomic regions, such as open chromatin and active transcription factor binding sites, which influences the pattern of sequencing coverage. In these regions, nucleosome-protected sites have low coverage, while adjacent unprotected sites have high coverage. To analyze this pattern, we used the analytical framework proposed by Doebley et al. to assess nucleosome occupancy and the corresponding protection of cfDNA. We selected 854 transcription factor binding sites. For each transcription factor binding site, we calculated three values ​​from the GC-corrected coverage curves: 1) central coverage, obtained by measuring the coverage value at 30 bp away from the specified binding site; 2) average coverage, obtained by calculating the average coverage within the ±1000 bp region of the binding site; and 3) the amplitude of the binding site, determined by applying fast Fourier transform analysis.

[0080] A fully connected neural network (FCNN) was used in the modeling and analysis of nucleosome profiling data. This architectural choice is suitable for processing structured data generated by the analysis framework, enabling the network to learn and identify complex patterns indicative of cancer. The FCNN consists of multiple dense layers, each connected to its previous layer, and non-linear activation functions are incorporated between layers to model the complex relationship between nucleosome occupancy patterns and cancer signals. The network structure effectively processes input features - center coverage, average coverage, and amplitude - to classify samples and distinguish the presence of cancer from low-throughput whole genome sequencing (WGS) data, demonstrating the powerful ability to apply advanced neural network models to genomic data analysis.

[0081] 4.5 Fragment-based methylation (FM) model

[0082] The fragment-based methylation (FM) method determines the methylation status of a sample by analyzing the cleavage pattern around cytosine-phosphate-guanine (CpG) sites. It is hypothesized that an increased probability of cfDNA cleavage at adjacent cytosines is associated with methylated CpG sites, while unmethylated sites may not show this trend. Alu regions are the focus of the study because methylation levels are more significant in these regions compared to the entire human genome and standard CpG islands. Within these Alu regions, the fragment ratios of the 5'-end motifs of CGN and NCG (where N represents any nucleotide: A, C, T, or G) were analyzed, and eight different fragment ratios were identified. In addition, the ratio of CGN / NCG motifs was calculated. In CGCG sequencing, the cleavage rate was found to be significantly increased at both C positions. In the broader genomic context, all two tandem CpG nucleotides spanning the CGCG sequence were examined. The fragment ratios of the 5'-end motifs of CGC, NCG (C1), CGN, and CGC (C2) were determined as features. In total, ten fragment ratios were recorded for CGCG sequencing sites. In addition, the motif ratios of CGC / NCG and CGN / CGC were calculated and recorded as feature sets. A fully connected neural network (FCNN) was deployed to process FM features.

[0083] 4.6 Integrated Generalized Linear Models (GLM)

[0084] The outputs generated by these base models were then imported into a separate ensemble model consisting of a generalized linear model (GLM). A binary classification ensemble model responsible for detecting cancer signals, a logistic regression model was used to simulate the probability of an instance being classified as "cancer" versus "non-cancer". In model construction, the H2O AutoML R package was used to make full use of the five feature types mentioned above, and five major algorithms (generalized linear model GLM, gradient boosting machine GBM, random forest RF, deep learning DL, and XGBoost) were used to build the base model. Through the H2O AutoML framework, 200 base models were generated for each feature type using a random grid search within a wide range of hyperparameters for each algorithm. Each model uses healthy samples and colorectal cancer samples from the training cohort to train its classifier and calculate a corresponding cancer score for each sample. The score ranges from 0 to 1, and a higher score indicates a higher probability of cancer. An ensemble approach was then applied, where the cancer scores of the top five base models ranked by the area under the ROC curve (AUC) among the 200 base models for each feature type were compiled into a matrix, and then a second-level generalized linear model (GLM) algorithm was applied to create the final model. To prevent model overfitting, a five-fold cross-validation was performed on each of the five base models. The entire model training process was performed only on the training dataset, and the independent validation dataset remained locked until the model was finalized.

[0085] 4.7 Decision Call Threshold

[0086] During the training of the classifier, internal validation is used to set thresholds corresponding to predetermined specificity levels. For the binary classification ensemble model for detecting cancer, two thresholds were defined based on the validation set. The standard specificity threshold (SST) was set to a model specificity of 99.0%. Specificity refers to the ability of the test to correctly identify non-cancer cases (true negatives) among all non-cancer cases evaluated. A specificity of 99.0% means that the model can correctly identify 99 out of 100 non-cancer cases as non-cancer. The high specificity threshold (HST) is set at a higher specificity level of 99.9%. This stricter standard ensures that the model is more conservative when labeling cases as "cancer", thereby further reducing the false positive rate.

[0087] By utilizing these thresholding strategies, the classifier can be fine-tuned to meet specific performance metrics, especially in reducing false positives, which is crucial in medical diagnosis.

[0088] 5 Model verification and testing

[0089] The training of the model all comes from the training cohort, and the independent validation cohort verifies the efficiency and reliability of the constructed model. The prospective cohort verifies the performance of the model, evaluates the sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of the model, and ensures the reliability of the model through technical repeatability and inter-run reproducibility tests. In addition, an independent validation cohort was used to evaluate the performance of the multidimensional integrated machine learning model, and two thresholds, FRAGTECT-1 and FRAGTECT-2, were determined to generate sub-models with ultra-high specificity (99.5%) and high specificity (98%).

[0090] To validate the performance of the model in different subgroups of colorectal cancer patients, the diagnostic sensitivity of the model predicting each subgroup was calculated. These subgroups included age, sex, tumor stage, tumor site, tumor grade, tumor size, lymph node metastasis, microsatellite (MS) status, carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA19-9), CA125, and alpha-fetoprotein (AFP). Microsatellite status was divided into microsatellite stable (MSS) and microsatellite unstable (MSI). The reference ranges of biomarkers were as follows: CEA ≤ 5 ng / mL, CA125 ≤ 24 U / mL, CA19-9 ≤ 30 U / mL, and AFP < 7 ng / mL.

[0091] Finally, the model was tested in a prospective cohort to validate its performance and ensure consistent applicability across a wider range of clinical scenarios.

[0092] 6 Statistical analysis

[0093] All statistical analyses were performed in R (version 4.2.3). Receiver operating characteristic (ROC) curves and area under the curve (AUC) were analyzed using the pROC package (version 1.18.0). After the true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) were determined in the model prediction, the sensitivity [TP / (TP+FN)], specificity [TN / (TN+FP)], positive predictive value (PPV) [TP / (TP+FP)], negative predictive value (NPV) [TN / (TN+FN)], accuracy [(TP+TN) / (TP+FP+TN+FN)] and the corresponding 95% confidence intervals (CI) were calculated using the epiR package (version 2.0.60).

[0094] Example 2

[0095] Model application based on Example 1

[0096] 1. Distribution characteristics of the enrolled population

[0097] A total of 853 participants were recruited in this study, including patients with colorectal cancer (CRC) and healthy or non-CRC volunteers. Four participants were excluded due to substandard quality control of the second-generation sequencing (NGS), and another three voluntarily withdrew, resulting in a total of 838 participants. The training cohort consisted of 360 participants, including 176 CRC patients and 184 healthy non-CRC patients, whose data were used to train the multidimensional machine learning model. The independent validation cohort included 236 individuals, including 117 CRC patients and 119 healthy controls, for independent evaluation of model performance. The prospective cohort consisted of 129 CRC patients and 113 non-CRC patient controls to validate the constructed multidimensional machine learning model.

[0098] The samples of this cohort were collected at the same medical institution between January 2022 and February 2023. The first subcohort of the prospective validation included 136 individuals who were confirmed to be FIT-positive by hospital physical examination, 23 of whom were diagnosed with colorectal cancer after colonoscopy, and the second subcohort included 106 patients diagnosed with colorectal cancer. Due to the low incidence of colorectal cancer in FIT-positive individuals, this subcohort was established to enhance the persuasiveness of the prospective validation cohort.

[0099] Table 1 summarizes the demographic and clinical characteristics of all participants from the training and validation cohorts. The mean age of patients with colorectal cancer was 61.18 years in the training cohort and 60.38 years in the validation cohort. In comparison, the mean age of healthy participants was 60.07 years in the training cohort and 59.69 years in the validation cohort. The distribution of cancer stage was consistent, with most enrolled patients having early-stage disease (training cohort: 56.25% and validation cohort: 55.56%). There was no significant difference in the location of the primary tumor between the two cohorts. A higher proportion of patients with right-sided colon cancer was found in the validation cohort, at 26.5%, compared with 19.89% in the training cohort.

[0100] Results of biomarker testing were collected from colorectal cancer patients and healthy participants. In the colorectal cancer group, the proportion of carcinoembryonic antigen (CEA) negativity was 62.50% in the training cohort and 53.85% in the validation cohort. Similarly, compared with the healthy control group, the proportion of carbohydrate antigen 19-9 (CA19-9) and alpha-fetoprotein (AFP) negativity was higher in the training cohort and validation cohort patients with colorectal cancer (CA19-9 negative: 71.59% in the training cohort and 59.83% in the validation cohort; AFP negative: 89.2% in the training cohort and 92.31% in the validation cohort). However, the proportion of carbohydrate antigen 125 (CA125) positivity was higher in colorectal cancer patients in the training cohort and validation cohort, 83.52% and 92.3%, respectively. Among the healthy individuals in the training cohort and validation cohort, more than 90% of the participants tested negative for the remaining four biomarkers, except CA125, which was not included in the test panel for female participants.

[0101] Overall, the demographic characteristics of the training and validation cohorts were similar, ensuring the reliability of the performance estimates.

[0102] 2. Feature extraction and ensemble stacking model evaluation

[0103] The constructed model showed extremely high predictive ability in the training cohort, with an area under the curve (AUC) of 0.997 (95% confidence interval: 0.994–1.000), and was validated in an independent cohort, reaching 0.986 (95% confidence interval: 0.973–0.999; Figure 2 :A) AUC values. The base model using any single feature produced AUC values ​​ranging from 0.903 to 0.964, demonstrating the superior predictive power of the ensemble model ( Figure 2: B). In the training dataset, a cancer score of 0.60 was applied as a threshold to achieve a specificity of 98.0%, generating FRAGTECT-1. In addition, a cancer score of 0.68 was set as another threshold to capture a specificity of 99.5% in the training dataset, generating FRAGTECT-2 (Table 3). These thresholds maintained specificities of 0.9747 (95% confidence interval: 0.9281–0.9948; Table 2) and 0.9832 (95% confidence interval: 0.9406–1.00; Table 3), respectively, in the validation cohort. Accordingly, in the validation dataset, the sensitivity predicted by FRAGTECT-1 was 0.9316 (95% confidence interval: 0.8697–0.9700), while the sensitivity dropped to 0.9231 (95% confidence interval: 0.8590–0.9642) when using FRAGTECT-2. The positive predictive value (PPV) reflects the trend of sensitivity. In the validation dataset, the PPV value of FRAGTECT-1 was 0.9732 (95% confidence interval: 0.9237–0.9944), while the PPV value of FRAGTECT-2 was 0.9818 (95% confidence interval: 0.9359–0.9978). The distribution of cancer scores between cancer patients and healthy individuals was compared using box plots to visually visualize the predictive power of the two models ( Figure 2 :C). The analysis showed that the predicted cancer scores were significantly higher in cancer patients compared with healthy controls. Therefore, the multivariate ensemble stacking model combining the five cfDNA features showed excellent predictive ability.

[0104] 3. Evaluate the stability and robustness of the model

[0105] In an independent validation cohort, we first evaluated the repeatability and reproducibility of the model using intra- and inter-assay replicate samples from three cancer patients and three healthy controls. Figure 6 : As shown in A, FRAGTECT-1 / 2 predicted the disease status of three replicates of cancer and healthy samples with 100% accuracy, regardless of the condition. In addition, high consistency was observed within the three replicates of each tested condition (intra-batch testing) and between the first and second batches of samples from the same participant (inter-batch testing). To further examine the predictive stability of the model, we recruited four additional healthy participants and calculated their cancer scores using samples collected from them at four different time points throughout the day. Regardless of the health status at the time of sample collection, FRAGTECT-1 / 2 correctly predicted the disease status of all four healthy volunteers with 100% accuracy ( Figure 6:B). When evaluating the robustness of the model at low whole genome sequencing (WGS) coverage, the AUC value remained at a high level (≥0.9185) even when the WGS coverage was reduced to 4x, 3x, 2x, 1x, and 0.5x (Table 4). In summary, these tests demonstrate the robustness and consistent performance of the integrated stacking model in various situations.

[0106] 4. Performance of the prediction model in tumor subgroups

[0107] We further evaluated the performance of the model in the validation cohort by analyzing different subgroups of colorectal cancer. Figure 3 As shown in Figures A and Table 5, FRAGTECT-1 reliably detected cancer patients with a sensitivity of 0.8824 (95% CI: 0.6356-0.9854) and 0.9167 (95% CI: 0.8002-0.9768) in stage 0-I and stage II cancer patients, respectively. FRAGTECT-2 performed similarly (Table 6). The prediction sensitivity in the validation dataset increased with the progression of cancer stage. FRAGTECT-1 showed higher sensitivity in stage III (0.9535, 95% CI: 0.8419-0.9943) compared to FRAGTECT-2 (0.9302, 95% CI: 0.8094-0.9854). Both models were able to predict stage IV patients with 100% sensitivity (FRAGTECT-1 / 2: 95% CI 0.6306-1). Notably, in patients with colon cancer, both models showed increased sensitivity for detecting right-sided tumors compared with left-sided tumors ( Figure 3 : B). FRAGTECT-1 showed higher sensitivity when the tumor size was larger than 3 cm, while FRAGTECT-2 showed higher sensitivity when the tumor size was smaller than 3 cm ( Figure 3 : C). Our model showed high sensitivity in colorectal cancer patients with lymph node metastasis, with sensitivities of 0.9545 (95% confidence interval: 0.8453-0.9944) and 0.9318 (95% confidence interval: 0.8134-0.9857) for FRAGTECT-1 and FRAGTECT-2, respectively ( Figure 3 :D).

[0108] Biomarkers including carcinoembryonic antigen (CEA), carbohydrate antigen 125 (CA125), carbohydrate antigen 19-9 (CA19-9), and alpha-fetoprotein (AFP), which are commonly used as indicators for cancer diagnosis and postoperative monitoring, were measured in colorectal cancer patients and healthy participants. Figure 3 :E and F; Table 5-6). In this study, receiver operating characteristic (ROC) curve analysis ( Figure 7 ) evaluated the diagnostic ability of these biomarkers in detecting cancer in a validation cohort. The results showed that the area under the curve (AUC) of CEA was 0.805 (95% confidence interval: 0.749-0.861), while the AUC values ​​of other biomarkers ranged from 0.635 to 0.501. In addition, the FRAGTECT model showed good detection sensitivity in colorectal cancer patients with negative specific biomarkers. For example, in CEA-negative individuals, the detection sensitivities of FRAGTECT-1 and FRAGTECT-2 were 92.06% (95% confidence interval: 82.44%-97.37%) and 90.48% (95% confidence interval: 80.41%-96.42%), respectively ( Figure 3 : E). Similarly, for individuals with negative CA19-9, the sensitivities of FRAGTECT-1 and FRAGTECT-2 were 94.74% (95% confidence interval: 87.07%-98.55%) and 93.42% (95% confidence interval: 85.31%-97.83%), respectively ( Figure 3 :F). In CA125-negative individuals, the sensitivity of FRAGTECT-1 was 93.68% (95% confidence interval: 86.76%-97.65%), while the sensitivity of FRAGTECT-2 was 92.63% (95% confidence interval: 85.41%-96.99%). Similar results were observed in AFP-negative individuals, with the detection sensitivities of FRAGTECT-1 and FRAGTECT-2 being 94.27% (95% confidence interval: 0.894-0.9735) and 92.59% (95% confidence interval: 0.8593-0.9675), respectively.

[0109] 5. Performance of the prediction model in a prospective cohort

[0110] We further evaluated the performance of the model using a prospective cohort, which validated the predictive ability of the developed model, with an area under the curve (AUC) of 0.969 (95% confidence interval: 0.947, 0.990; Figure 4: A). For the FRAGTECT-1 model, this cohort showed an overall specificity of 0.9558 (95% CI: 0.8998-0.9855; Table 1), an overall sensitivity of 0.9147 (95% CI: 0.8525-0.9567), and a positive predictive value (PPV) of 0.9593 (95% CI: 0.9077-0.9867). The FRAGTECT-2 model, targeting 99.5% specificity in the training dataset, achieved a specificity of 0.9646 (95% CI: 0.9118-0.9903), a sensitivity of 0.8372 (95% CI: 0.7620-0.8963), and a positive predictive value of 0.9643 (95% CI: 0.9111-0.9902) in the entire prospective dataset (Table 7). Figure 4 The box plots in B show the predictive advantages of the two models. The FRAGTECT model performed consistently in both subcohorts, with slightly higher sensitivity in cohort 1 (Table 8). In addition, in line with the trend observed in the validation cohort, the predictive sensitivity in the prospective dataset improved significantly with the progression of cancer stage ( Figure 4 :C). In addition, we used a published trial (21) to evaluate the clinical potential of our model in the real world. Compared with the 47% detection rate of advanced stage (stage III / IV) with standard treatment, FRAGTECT-1 and FRAGTECT-2 could detect 98% of Chinese stage I colorectal cancer cases and improve the 5-year survival rate from 66% to 84% ( Figure 7 and 8 In the United States, these models have improved early diagnosis rates, with 26% and 21% of colorectal cancer patients originally diagnosed at stage III and IV being diagnosed at stage I, respectively ( Fig. 9 ), which indicates the effectiveness of this model in early detection of colorectal cancer.

[0111] In summary, our integrated stacking model, utilizing the multiplexed properties of circulating free DNA (cfDNA), performed well in detecting colorectal cancer risk with extremely high sensitivity and stability. This promising model is expected to enable early diagnosis and bring significant benefits to a wide range of colorectal cancer patients.

[0112] Table 1. Demographics and baseline characteristics of the study cohort

[0113]

[0114]

[0115] Table 2. Performance of FRAGTECT-1 in the training, validation, and prospective cohorts

[0116]

[0117] Table 3. Model performance at different ROC curve cutoff values

[0118]

[0119] Table 4. Evaluating the performance of FRAGTECT-2 on an independent validation dataset

[0120]

[0121] Table 5. Detection limits using downsampled data in an independent validation dataset

[0122]

[0123] Table 6. Performance of FRAGTECT-1 in colorectal cancer patient subgroups in the independent validation dataset

[0124]

[0125]

[0126] Table 7. Performance of FRAGTECT-2 in colorectal cancer patient subgroups in the independent validation dataset

[0127]

[0128]

[0129] Table 8. Performance of FRAGETCT in a prospective dataset

[0130]

[0131] Table 9. Comparison of performance of FRAGTECT and LUNAR-2 at 90.0% specificity in different prospective cohorts

[0132]

Claims

1. A method for constructing an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics, characterized in that The following steps are involved: a) Sample collection and processing: Peripheral venous blood samples from patients suspected of colorectal cancer and healthy subjects were collected using EDTA anticoagulant tubes, and the upper plasma was extracted and cell debris was removed; cfDNA was extracted from plasma using the QIAamp circulating nucleic acid kit; The concentration of extracted cfDNA was measured using the Qubit dsDNA assay kit; b) Whole genome sequencing (WGS) and bioinformatics processing: The extracted cfDNA was sequenced using the Illumina NovaSeq platform for low-coverage whole-genome sequencing to generate sequencing data with a sequencing depth of at least 0.1x; Bioinformatics processing of sequencing data was performed, including: trimming reads using Trimmomatic software, setting the parameters to ILLUMINACLIP:TruSeq3-PE.fa:2:30:10 SLIDINGWINDOW:4:15 MINLEN:36; aligning reads to the human reference genome (GRCh37 / UCSC hg19) using the BWA tool, setting the default parameters; and removing PCR duplicates using the Picard tool, setting the parameter to REMOVE_DUPLICATES=true; c) Fragmentomics feature extraction: The following cfDNA fragment omics features were extracted from the WGS data after bioinformatics processing: Copy number variations (CNVs): Sequencing depth was assessed using a hidden Markov model and compared with the baseline of a healthy control group in the same interval. The log2 ratio between the sample and the baseline in each interval was calculated to obtain the copy number variation; Fragment size coverage (FSC): For each fragment size group: short fragments <100 bp and long fragments ≥150 bp, the normalized coverage score (z-score) within each 5 Mb interval was calculated; Fragment size distribution (FSD): The coverage of cfDNA fragments from 100 bp to 220 bp was examined in intervals of 5 bp on 39 chromosome arms. The original fragment size distribution characteristics were extracted and converted into z-scores after comparison with the overall mean. Nucleosome footprint (NF): Analysis of transcription factor binding sites using nucleosome coverage: 1) Center coverage: measure the coverage value at 30 bp away from the specified binding site; 2) Average coverage: calculate the average coverage within the ±1000 bp region of the binding site; 3) The amplitude of the binding site is determined by fast Fourier transform analysis; Fragment-based methylation (FM): The methylation status of the sample was determined by analyzing the cleavage pattern around the CpG site. In the Alu region, the fragment ratios of the 5' end motifs of CGN, NCG, CGC, CGN / NCG, CGC / NCG, CGN / CGC, NCG(C1), and CGC(C2) were analyzed as feature sets, where N represents any nucleotide: A, C, T, or G; d) Machine learning model building and optimization: Based on the omics features of the extracted cfDNA fragments, an integrated stacking machine learning strategy was used to build a classification model for early diagnosis of colorectal cancer by combining five basic algorithms: generalized linear model, gradient boosting machine, random forest, deep learning algorithm or XGBoost. Through the H2O AutoML framework, for each feature type, 200 base models were generated using random grid search within a wide range of hyperparameters for each algorithm, among which the top five base models with the area under the ROC curve (AUC) cancer score were compiled into a matrix, and then a second-level generalized linear model (GLM) algorithm was applied to create the final model.

2. The method for constructing an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics according to claim 1, characterized in that: It also includes cross-validation, hyperparameter adjustment, and feature selection of the final model to further improve the diagnostic accuracy and generalization ability of the model.

3. The method for constructing an early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics according to claim 1, characterized in that: The method also includes a step of performing a performance evaluation on the constructed model, which includes calculating the sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (ROC) (AUC) indicators to verify the diagnostic performance of the model.

4. An early accurate diagnosis model for colorectal cancer based on plasma cfDNA multidimensional fragment omics obtained by the construction method described in any one of claims 1-3.

5. A colorectal cancer early diagnosis system, characterized in that: include: (1) an acquisition unit, used to acquire marker information of the sample to be tested; (2) A detection unit, used to use the marker information of the sample to be tested as an input feature of a machine learning model, and then use the prediction model trained by the method of any one of claims 1 to 3 to obtain the result of the sample to be tested; the result is a cancer risk score of the sample to be tested.

6. An electronic device, characterized in that: include: Memory and processor; The memory is used to store computer programs; The processor is configured to execute a computer program to apply the system according to claim 5.

7. An executable storage medium, characterized in that: The storage medium stores computer program instructions, which, when executed on a processor, cause the processor to execute the system as claimed in claim 5 .

Images