Early detection technology for common digestive system cancers based on multi-dimensional features of cfDNA targeted methylation sequencing

Abstract

The present application provides a marker combination for early detection of digestive system cancer, tissue localization, diagnosis, prognosis detection and identification of benign and malignant, the marker combination is selected from 1656 chromosome regions in table 1. A multi-cancer early screening and localization technology GutSeer for five high mortality digestive system cancers is also provided, which proves that using a relatively small second-generation sequencing panel can realize accurate cancer detection and organ tissue localization by using multiple dimensional characteristics including methylation, copy number variation and terminal motif.

Description

Technical Field

[0001] This invention relates to the field of cancer diagnosis, specifically to an early detection technology for common digestive system cancers based on the multidimensional characteristics of cfDNA targeted methylation sequencing. Background Technology

[0002] Common malignant tumors of the digestive system, such as esophageal cancer, stomach cancer, liver cancer, pancreatic cancer, and colorectal cancer, cause hundreds of thousands of deaths each year. Because these cancers are often discovered at an advanced stage, their treatment outcomes are poor. Early detection of these cancers can significantly improve survival rates. Currently, the main methods for screening and diagnosing early digestive system cancers are medical imaging techniques, such as endoscopy, which is particularly effective for cancers of the digestive tract (esophagus, stomach, and colorectal), and ultrasound, which is useful for liver and pancreatic cancers. However, due to low patient compliance, endoscopy and ultrasound are difficult to widely adopt as screening methods. Furthermore, these techniques are highly dependent on the skills of endoscopists and pathologists. Therefore, there is an urgent need for a practical and efficient blood-based screening and diagnostic method for digestive system cancers.

[0003] Cell-free DNA (cfDNA) is a fragment of genomic DNA released into the bloodstream from apoptotic and necrotic cells. In cancer patients, cfDNA also includes circulating tumor DNA (ctDNA) from cancer cells. ctDNA carries specific information about cancer-related tissue damage and is therefore considered a promising biomarker for early cancer detection and diagnosis. ctDNA can be used to detect cancer-related genomic mutations, copy number variations (CNVs), gene fusions, etc., and the results have been applied to establish blood-based non-invasive cancer detection technologies.

[0004] Recent studies have found that cfDNA methylation is more accurate in detecting cancer than the aforementioned gene mutations. Therefore, cfDNA methylation has been used in the development of newer methods and is increasingly being developed as a biomarker or detection target for cancer screening. For example, the PanSeer study screened 595 methylation regions associated with five common cancers (gastric cancer, esophageal cancer, colorectal cancer, lung cancer, and liver cancer) as detection targets and established corresponding methylation-targeted sequencing technology for non-invasive cancer screening. EpiPanGIDx, through the methylation levels of 67,832 differentially methylated regions, can screen for six digestive system cancers (colorectal cancer, hepatocellular carcinoma, esophageal squamous cell carcinoma, gastric cancer, esophageal adenocarcinoma, and pancreatic ductal adenocarcinoma). CancerRadar uses whole-genome methylation sequencing and analysis technology to detect four types of cancer (colon cancer, liver cancer, lung cancer, and gastric cancer) by detecting the methylation levels of over 110,000 targets.

[0005] Besides ctDNA methylation, cfDNA fragmentomics detects and locates cancer by analyzing the sequence characteristics of cfDNA fragments, such as the non-random distribution of cfDNA length, motifs at ends or breakpoints, nucleosome footprints, and fragment coverage. For example, DELFI (DNA evaluation of fragments for early intervention) technology analyzes the overall condition of cfDNA fragments to detect seven types of cancer (breast cancer, colorectal cancer, lung cancer, ovarian cancer, pancreatic cancer, gastric cancer, and bile duct cancer), achieving an overall AUC of 0.94 and a tissue origin accuracy of 61% for cancer signals. DECIPHER (detecting early cancer by inspecting ctDNA features) further expands the application of fragmentomics in non-invasive screening for multiple cancer types. This technology integrates five fragmentomics features (fragment size coverage, fragment size distribution, terminal motifs, breakpoint motifs, and copy number variations) to detect three cancer types (primary liver cancer, colorectal adenocarcinoma, and lung adenocarcinoma), achieving an overall AUC of 0.983 and a tissue origin prediction accuracy of 93.1%. However, most fragmentomics-based methods rely on whole genome sequencing (WGS), resulting in relatively high experimental and data analysis costs. Although some studies have shown that low-throughput or even ultra-low-throughput WGS can meet the needs of multi-cancer detection and tissue and organ localization, these studies are limited by relatively small study scales or relatively limited cancer types. Therefore, these fragmentomics-based multi-cancer screening technologies lack validation in large-scale populations, particularly in assessing their ability to localize cancer tissues or organs.

[0006] Currently, there are no particularly efficient methylation markers for digestive system tumors, nor are there direct comparisons or combined applications between fragmentomics-based cancer screening technologies and similar technologies based on ctDNA methylation. Summary of the Invention

[0007] To address the aforementioned technical issues, this invention employs targeted methylation sequencing to detect samples. By integrating multi-dimensional features from sequencing data, a GutSeer model is established for cancer detection and tissue tracing.

[0008] Specifically, the first aspect of the present invention provides a combination of biomarkers for early detection, tissue localization, diagnosis, prognosis detection and benign / malignant identification of digestive system cancers, wherein the combination of biomarkers is selected from 1656 chromosomal regions in Table 1.

[0009] In some implementations, the combination of markers is selected from some or all of the 1656 chromosomal regions shown in Table 1.

[0010] In some embodiments, the combination of markers is selected from at least 5, at least 10, at least 50, at least 100, at least 200, at least 300, at least 400, at least 500, at least 600, at least 700, at least 800, at least 900, at least 1000, at least 1100, at least 1200, at least 1300, at least 1400, at least 1500, and at least 1600 regions from the 1656 chromosomal regions shown in Table 1.

[0011] In some embodiments, the digestive system cancer is selected from esophageal cancer, gastric cancer, liver cancer, pancreatic cancer, and colorectal cancer.

[0012] A second aspect of the present invention provides a kit for early detection, tissue localization, diagnosis, prognosis detection, and benign / malignant identification of digestive system cancers, the kit comprising reagents for detection of the biomarker combination described in the first aspect of the present invention.

[0013] In some implementations, the detection includes targeted methylation sequencing based on sample cfDNA.

[0014] In some implementations, methylation status is extracted based on the data from the targeted methylation sequencing.

[0015] In some implementations, fragment omics features are extracted based on the data from the targeted methylation sequencing.

[0016] In some implementations, methylation status and fragmentomics features are extracted from the data of the targeted methylation sequencing.

[0017] In some embodiments, the methylation state is determined by measuring its mean methylation level (AMF) and methylation haplotype fraction (MHF).

[0018] In some implementations, the fragment omics features include terminal motifs and copy number variations.

[0019] A third aspect of the present invention provides a GutSeer model for early detection, tissue localization, diagnosis, prognostic detection and benign / malignant identification of digestive system cancers, wherein the model integrates the methylation status of a combination of sample ctDNA biomarkers and optional fragment omics features.

[0020] In some embodiments, the combination of markers includes regions as described in the first aspect of the invention.

[0021] In some embodiments, the digestive system cancer is selected from esophageal cancer, gastric cancer, liver cancer, pancreatic cancer, and colorectal cancer.

[0022] In some embodiments, the methylation status is determined by measures including the average methylation level (AMF) and the methylation haplotype fraction (MHF).

[0023] In some implementations, the AMF value is calculated by dividing the number of reads at methylation sites within the target region of the sample by the total number of reads detected within that region, using the following formula:

[0024]

[0025] Where i is the index of the target region; Ni and Mi are the total number of reads detected in the region and the number of reads at methylation sites;

[0026] In some embodiments, the MHF is quantified by calculating the fractions of fully methylated and fully unmethylated haplotypes within the target region to evaluate the methylation level of that region, using the following formula:

[0027]

[0028] Where i is the ordinal number of a specific consecutive CpG combination within the marked region; h is the ordinal number of the haplotype with fully methylated / unmethylated CpG; Ni,h is the number of reads of haplotype h in consecutive CpG combination i; and Ni is the total number of reads covering consecutive CpG combination i.

[0029] In some implementations, the fragment omics features include terminal motifs and copy number variations (CNVs).

[0030] In some implementations, the CNV is quantified by the coverage depth of the corresponding genomic location for each target region in each sample, using the following formula:

[0031]

[0032] Where i is the index of the marked region; Ni is the number of reads mapped to region i; Li is the length of region i; and N is the number of reads located on the target region.

[0033] In some implementations, the terminal motif is calculated based on reads from all target regions using the following formula:

[0034]

[0035] Where m is the sequence number of the 6 base motifs; Fm is the proportion of motif m; Nm is the number of motif m; and 4096 is the total number of all possible motifs (46).

[0036] In some implementations, the method further includes calculating a methylation encoding score (MES) for each target site, which utilizes both methylation and fragmentomics information.

[0037] In some implementations, the calculation of the MES score includes the following steps: 1) Methylation haplotype vector for each target region: For each target region, the methylation state and fragment coverage state of CpG in that region, as reflected by each sequencing read in that target region, are converted into two binary vectors respectively; then these two vectors are concatenated into a single binary vector with a length of twice that of the target region; 2) Modeling the methylation haplotype vector for each target region: For each target region, 10 models are trained based on the methylation haplotype vector. These models can calculate the probability of cancer / health and tissue origin represented by the methylation haplotype; wherein, the 10 models include 5 cancer / health models and 5 tissue origin models; 3) Normalization of probability values ​​and target region score vector: For each sample, for each target region... 4) Region Model: For each target region, 10 neural networks are built based on the region score vector. These region models can score the probability of cancer / health and the probability of tissue origination for the region score vector. The probability of cancer and the probability of tissue origination are calculated by binning the probability values ​​of the target region into 14 groups (from 0 to 1). The probability values ​​of each group are normalized according to the total number of all values. Combined with the methylation haplotype count in each group and the total number of methylation haplotypes in the region, a vector of length 29 is obtained.

[0038] The fourth aspect of the present invention provides a method for constructing a GutSeer model for early detection, tissue localization, diagnosis, prognosis detection and benign / malignant identification of digestive system cancers. The method includes training two ensemble models, which are used to distinguish between cancer plasma and normal plasma and to identify cancer type (TOO identification).

[0039] In some implementations, both integrated models consist of two layers. In the first layer, four sub-models are trained using AMF / MHF, MES, terminal motifs, and CNV features, respectively. In the second layer, a logistic regression classifier is used to train and score the predicted scores output by the first-layer model, ultimately outputting the cancer-health prediction result and the probability of TOO tissue tracing.

[0040] In some implementations, the AMF / MHF sub-model training process includes the following steps: 1) For each target, using the MHF matrix and AMF matrix of the target, and combining it with the corresponding labels of its samples, an SVM model is trained, and the target is scored according to the model to predict the cancer-health binary value or the probability value of each tissue origin; 2) Based on the predicted score of the SVM model, a stacked model is trained using multiple modeling methods and the second round of prediction values ​​is output as the result of the AMF / MHF sub-model.

[0041] In some implementations, the corresponding label includes cancer or health, and cancer type.

[0042] In some implementations, the various modeling methods include logistic regression, random forest, SVM, and gradient boosting.

[0043] In some implementations, the second round of predictions is a binary value predicting cancer-health, or a probability value predicting the origin of each tissue.

[0044] In some implementations, the MES, terminal motif, or CNV sub-model is used to train a deep neural network sub-model or a logistic regression sub-model using the MES value, CNV value, or terminal motif frequency of the target region, and outputs a predicted score, a predicted cancer-health binary value, or a predicted probability of origin for each tissue.

[0045] The fifth aspect of the present invention provides a GutSeer model for early detection, tissue localization, diagnosis, prognostic detection, and benign / malignant identification of digestive system cancers obtained by the method described in the fourth aspect of the present invention.

[0046] The sixth aspect of this invention provides the application of the GutSeer model described in the third and fifth aspects of this invention in combination with conventional cancer biomarkers in the preparation of reagent kits or devices for early detection, tissue localization, diagnosis, prognostic detection, and benign / malignant identification of digestive system cancers.

[0047] In some embodiments, the conventional cancer markers include carcinoembryonic antigen (CEA), cancer antigen 199 (CA199), cancer antigen 724 (CA724), alpha-fetoprotein (AFP), and abnormal prothrombin (DCP).

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1) This invention developed GutSeer, a multi-cancer early screening and localization technology for five digestive system cancers with high mortality rates. It demonstrates that accurate cancer detection and organ and tissue localization can be achieved by using a relatively small second-generation sequencing panel and utilizing multiple dimensions of features, including methylation, copy number changes and terminal motifs.

[0050] 2) The GutSeer panel is specifically designed for detecting five types of digestive system cancers (gastric cancer, esophageal cancer, liver cancer, colorectal cancer, and pancreatic cancer), containing a total of 1656 methylation targets. These targets are genomic regions where the methylation status differs between adjacent normal tissues and cancerous tissues, particularly digestive system cancer tissues. Furthermore, this invention has validated the panel's quality and effectiveness through technical evaluation. The methylation target panel in this invention not only shows clear clusters between cancerous tissues and healthy plasma, but also exhibits significant differences between cancerous tissues of different cancers, demonstrating excellent tissue specificity and strong capabilities in digestive tract cancer detection and organ attribution.

[0051] 3) GutSeer outperforms single-type ctDNA features such as methylation or fragmentomics in multi-cancer detection and localization testing, and also has advantages over WGS technology. Furthermore, GutSeer is highly accurate while also being efficient and low-cost, making it a promising non-invasive method for multi-cancer screening and localization. Attached Figure Description

[0052] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0053] Figure 1 This shows the overall design of the GutSeer study.

[0054] Figure 2 This shows the calculation process for the "methylation coding score".

[0055] Figure 3 Displaying GutSeer's technical assessment - 1. U-MAP plot of cfDNA data for cancer, adjacent normal tissue, and healthy tissue. Color intensity represents tissue type. Shape represents pathological state.

[0056] Figure 4 The technical evaluation of GutSeer-2 shows that TOO and cancer targets exhibit different methylation patterns.

[0057] Figure 5 This shows the GutSeer sample composition and study design.

[0058] Figure 6 Displays an overview of the GutSeer process.

[0059] Figure 7Showing the prediction results for cancer and healthy samples and tissue origination prediction. (AB) Average ROC curves (A) and (B) based on cross-validation in the training set. This includes a comparison of the ROC curves of the integrated GutSeer model and models built using methylation features or fragmentomics alone. (C) Sensitivity for cancer sample prediction at 96.4% specificity in cross-validation on the training set; and sensitivity for cancer sample prediction at 96.7% specificity in the test set. Results include sensitivity for all cancers, sensitivity for each cancer type, and sensitivity for each clinical stage of cancer. (DE) Confusion matrix of tissue origination prediction results on cross-validation (D) in the training set and on the test set (E). (F) Accuracy for tissue origination of all cancers and specific cancers.

[0060] Figure 8 This shows the cross-validation results of GutSeer in cancer-health detection. (A) ROC curve of GutSeer based on methylation and fragmentomics integration in 10-fold cross-validation; (BC) ROC curve of model using methylation features alone (B) or model using fragmentomics features alone (C) in 10-fold cross-validation.

[0061] Figure 9 This demonstrates GutSeer's sensitivity at 99% specificity.

[0062] Figure 10 The results show tissue source identification predictions based on a 5-category classification (excluding esophageal and gastric cancer). (A) Based on the confusion matrix of the training set, the tissue source identification accuracy is 76.0%. (B) Based on the confusion matrix of the test set, the tissue source identification accuracy is 74.3%.

[0063] Figure 11 This demonstrates the clinical application of GutSeer in identifying benign diseases and its comparison with conventional cancer markers. (AB) Distribution of GutSeer scores in benign samples. (A) All benign samples. (B) Benign samples from each tissue type. (C) Comparison of GutSeer and conventional cancer markers.

[0064] Figure 12 This demonstrates the application of GutSeer in identifying benign diseases. The distribution of GutSeer scores in benign diseases across each tissue type is shown. Samples are further subdivided according to their stage and subclass. The thresholds for GutSeer scores are the scores corresponding to 95% and 99% specificity in the training set. (A), (B), (C), and (D) represent colorectal, esophageal, liver, and gastric tissues, respectively.

[0065] Figure 13The sensitivity of models built using methylation features or fragment omics features alone in predicting cancer and healthy samples in the training set is shown. (A) Sensitivity of the model using methylation features alone in the training set (10-fold cross-validation, 96.4% specificity) and the test set (96.4% specificity). Sensitivity for all cancers, for each cancer type, and for each clinical stage are compared. (B) Same as (A), the model is built based on fragment omics features, with specificities of 96.4% and 95.9%, respectively.

[0066] Figure 14 The diagram shows tissue origin prediction based on either methylation or fragmentomics alone. (AB) shows the confusion matrices obtained by using models with methylation features alone for tissue origin prediction on the training set (A) and test set (B), respectively. (CD) is the same as (AB), but uses fragmentomics features alone to build a model for tissue origin prediction. (EF) shows the accuracy of tissue origin prediction based on models built using methylation (E) or fragmentomics (F) features alone, respectively.

[0067] Figure 15 This shows the impact of age and gender on GutSeer performance. (AB) GutSeer score distribution for samples of different ages (A) and genders (B). Except for a statistically significant difference between healthy male and female samples, there were no significant differences between the same groups in other comparisons. (CD) ROC curves (C) and sensitivity and specificity (D) for cancer sample prediction results based on all samples, male samples, female samples, and propensity score-matched samples on the test set.

[0068] Figure 16 Covariate analysis was performed. The covariates analyzed included (A) smoking status, (B) the number of unique DNA molecules, (C) the medical center providing the sample, and (D) the total DNA extraction amount (ng). The results showed no significant differences in GutSeer scores between the healthy sample subgroups and the cancer sample subgroups, which were further grouped according to the above covariates.

[0069] Figure 17 The impact of sequencing depth on GutSeer performance is shown. (A) The sensitivity (light curve) of cancer and health detection and the accuracy (dark curve) of tissue tracing are evaluated by downsampling sequencing data to 10M, 5M, 2M, 1M, 0.5M, and 0.1M. (B) The correlation of AMF for each target between high sequencing depth (downsampling data to 20M) and low sequencing depth (downsampling raw data to 10M, 5M, 2M, 1M, 0.5M, 0.1M, and 0.01M, respectively).

[0070] Figure 18 Showing model prediction results based on 800 randomly selected sites. (AB) Average ROC curves (A) and (B) based on cross-validation in the training set. This includes a comparison of the ROC curves of the integrated GutSeer model and models built using methylation features or fragmentomics alone. (C) Sensitivity for cancer sample prediction at 96.45% specificity in cross-validation on the training set; and sensitivity for cancer sample prediction at 98.22% specificity in the test set. Results include sensitivity for all cancers, sensitivity for each cancer type, and sensitivity for each clinical stage of cancer. (DE) Confusion matrix of tissue origination prediction results on cross-validation (D) in the training set and on the test set (E). (F) Accuracy for tissue origination of all cancers and specific cancers.

[0071] Figure 19 Showing model prediction results based on 500 randomly selected sites. (AB) Average ROC curves (A) and (B) based on cross-validation in the training set. This includes a comparison of the ROC curves of the integrated GutSeer model and models built using methylation features or fragmentomics alone. (C) Sensitivity for cancer sample prediction at 96.45% specificity in cross-validation on the training set; and sensitivity for cancer sample prediction at 93.13% specificity in the test set. Results include sensitivity for all cancers, sensitivity for each cancer type, and sensitivity for each clinical stage of cancer. (DE) Confusion matrix of tissue origination prediction results on cross-validation (D) in the training set and on the test set (E). (F) Accuracy for tissue origination of all cancers and specific cancers. Detailed Implementation

[0072] The applicant will now provide a detailed description of the method for early detection and tissue tracing of common gastrointestinal cancers using the GutSeer targeted methylation sequencing technology of the present invention, with reference to specific embodiments, to facilitate a clear understanding of the present invention by those skilled in the art. However, it should be understood that the following embodiments should not be construed in any way as limiting the scope of protection claimed in this application.

[0073] Example 1: Target Screening for Target Sequencing Panels

[0074] Two types of genomic regions were selected as GutSeer targets: (1) genomic regions where methylation differs between cancerous (or adjacent) and normal tissues; and (2) genomic regions where methylation signals differ between tissues of seven cancers (five digestive system cancers, lung cancer, and breast cancer). The former serves as a target for cancer-health prediction, while the latter serves as a target for organ and tissue tracing of cancer.

[0075] For each candidate target of each cancer type, its methylation level in the corresponding cancer tissue sample DNA was compared with normal samples, other cancer types, and blood cell data to determine whether the candidate could distinguish between cancer and healthy samples, or perform tissue tracing in the presence of background noise from blood cells. Finally, 1656 regions were selected as targets for GutSeer (Table 1).

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] Example 2: Sample Processing and GutSeer Library Construction

[0127] The experimental procedures for plasma separation and cfDNA extraction followed previous research (Chen et al. Nat Commun 11, 3475, 2020). GutSeer methylation-targeted sequencing libraries were constructed using the MethylTitan method. First, following the manufacturer's instructions, cfDNA was converted to bisulfite using the Methylcode bisulfite conversion kit (ThermoFisher, MECOV50). The conversion product was dephosphorylated and ligated to a universal adapter with UMI. The ligation product was used as a template for second-strand synthesis. After purification, the synthesized product was used as a template for a first-round semi-targeted amplification using PCR technology and a GutSeer-specific primer panel. The purified PCR product was then used as a substrate for a second-round PCR to add sample-specific barcodes and full-length sequencing adapters. The purified second-round PCR product became the sequencing library. The library was quantified using the KAPA Library Quantification Kit (KK4844) and sequenced on an Illumina NextSeq 500 / 550 sequencer in paired-end 150 bases mode, with at least 4M reads required per sample.

[0128] The pear package (version 0.9.6) was used to merge paired-end reads derived from potentially identical fragments to select high-quality raw cfDNA fragments. The trim_galore package (version 0.4.0) was used to remove adapter sequences from the fragment ends, and then UMIs were extracted from each read. The processed reads were aligned to the CT and GA-converted human genome reference sequence (version hg19), and the methylated haplotypes were deduplicated using UMIs.

[0129] Example 3 Feature Extraction of Targeted Methylation Sequencing Data

[0130] 3.1 Methylation state

[0131] The methylation status of the target region for each sample was determined by measuring the mean methylation level (AMF) and the methylation haplotype fraction (MHF).

[0132] The AMF value of a target site is calculated by dividing the number of reads at methylation sites within the target region of the sample by the total number of reads detected within that region. See the formula below for details:

[0133]

[0134] Where i is the index of the target region; Ni and Mi are the total number of reads detected in the region and the number of reads at methylation sites.

[0135] MHF quantifies the methylation level of a target region by calculating the fractions of fully methylated and fully unmethylated haplotypes (as shown in the formula below). See the formula below for details:

[0136]

[0137] Where i is the ordinal number of a specific consecutive CpG combination within the marked region; h is the ordinal number of the haplotype with fully methylated / unmethylated CpG; Ni,h is the number of reads of haplotype h in consecutive CpG combination i; and Ni is the total number of reads covering consecutive CpG combination i.

[0138] 3.2 Copy Number Variation (CNV)

[0139] In addition to methylation metrics, fragmentomics features such as terminal motifs and copy number variations are also used to measure target sites. CNV is quantified by the coverage depth of the corresponding genomic location for each target region in each sample. Here, to eliminate the bias of the sample sequencing depth and target region length, the target size and the number of reads located within that target region are used to standardize the CNV value. See the following formula for details:

[0140]

[0141] Where i is the index of the marked region; Ni is the number of reads mapped to region i; Li is the length of region i; and N is the number of reads located on the target region.

[0142] 3.3 Terminal motif

[0143] Terminal motifs were calculated based on reads from all target regions. It has been reported that cancer-induced abnormal cfDNA fragmentation, different from that in healthy individuals, affects the enrichment of cfDNA terminal motifs. Therefore, terminal motifs were calculated by taking the proportion of motifs for each specific 6-base combination at the ends of all cfDNA fragments in each sample, as the terminal motif feature for each sample (Jiang et al., 2020). See the formula below for details:

[0144]

[0145] Where m is the sequence number of the 6 base motifs; Fm is the proportion of motif m; Nm is the number of motif m; and 4096 is the total number of all possible motifs (46).

[0146] 3.4 Methylation Encoding Score (MES)

[0147] In addition to the commonly used methylation and fragmentomics features mentioned above, the methylation encoding score (MES) for each target site was also calculated. Figure 2 The score utilizes both methylation and fragmentomics information. The specific steps are as follows: Step 1: Methylation haplotype vector for each target region. For each target region, the methylation state and fragment coverage state of CpG in that region, as reflected by each sequencing read, are converted into two binary vectors respectively; these two vectors are then concatenated into a single binary vector, twice the length. Step 2: Modeling the methylation haplotype vector for each target region. For each target region, 10 models (5 cancer / health models and 5 tissue origination models) are trained based on the methylation haplotype vector. These models can calculate the probability of cancer / health and tissue origination represented by the methylation haplotype. Step 3: Normalization of probability values ​​and target region score vector. For each target region of each sample, the probability values ​​are binned into 14 groups (from 0 to 1), and the values ​​in each group are normalized according to the total number of values. Combining the methylation haplotype count in each group with the total number of methylation haplotypes in that region, a vector of length 29 is obtained. Step 4: Region Modeling. For each target region, 10 neural networks (5 cancer / health models and 5 tissue origination models) are built based on the region score vector. These region models score the probability of cancer / health and the probability of tissue origination for that target region. Step 5: Methylation Encoding Model. Based on the scores from the region models, the neural networks are trained using 10-fold cross-validation. The neural networks calculate a methylation encoding score (MES) for each sample to assess the probability of cancer and the probability of tissue origination.

[0148] Example 4: Technical Evaluation of GutSeer

[0149] To test the ability of these targets to pinpoint the tissue origin of digestive system cancers, 175 tissue samples (85 cancerous tissues and 90 adjacent normal tissues) and 38 healthy plasma samples (Table 2) from five types of digestive system cancers were used to compare the methylation levels of these targets, and U-MAP clustering was performed using the mean methylation level (AMF) of the target region. The "umap" package in R was used to cluster the AMF matrices of the 175 tissue samples and 38 plasma samples. When generating the AMF matrix, if the coverage depth of a target in a sample was less than 50, the AMF value at that location was assigned NA. The KNN algorithm (N=5) was used to impute missing values ​​in the matrix, followed by dimensionality reduction using the U-MAP algorithm, and finally projected to 2D. The results showed that there were significant clusters not only between cancerous tissues and healthy plasma, but also significant differences between cancerous tissues of different cancers. Figure 3 Furthermore, heatmaps plotted based on the AMF values ​​of the targets also showed that these targets were tissue-specific and enriched with their unique biological functions. Figure 4 All these results demonstrate that the GutSeer target has strong capabilities in the detection of gastrointestinal cancers and organ attribution.

[0150] Table 2: Number of tissue samples used for panel validity validation

[0151] Organizational Source all cancer Next to cancer colon and rectum 34 16 18 esophagus 38 18 20 Stomach 38 21 17 liver 32 17 15 pancreas 33 13 20 all 175 85 90

[0152] Example 5 Subject Recruitment

[0153] Subjects were prospectively enrolled at multiple centers. A total of 2473 samples were collected by balancing age and sex (787 healthy donors, 629 benign patients, 209 esophageal cancer cases, 239 gastric cancer cases, 180 colorectal cancer cases, 342 liver cancer cases, and 87 pancreatic cancer cases); more than half of the cancer samples came from patients with earlier-stage (clinical stage I-II) cancer (stage I 35.6%, stage II 23.3%, stage III 21.7%, and stage IV 12.5%) (Table 3). Cancer and healthy samples were randomly assigned to the training and testing cohorts in a 1:1 ratio, while 629 benign case samples were used as an independent validation set. Figure 5 ).

[0154] Table 3: Case Information

[0155] A. Overall information about the case

[0156]

[0157]

[0158] B. Number of cases of each type of cancer by clinical stage

[0159]

[0160] Example 6: Construction of the GutSeer Prediction Model

[0161] First, the training set samples were randomly divided into 10 groups for subsequent 10-fold cross-validation (note that this grouping remains consistent throughout model training). When predicting the test set samples, the 10 cross-validated classifiers were used for prediction, and the scores from the 10 classifiers were averaged to obtain the final prediction result. Two ensemble models were trained to distinguish between cancer plasma and normal plasma and to identify cancer types (TOO identification), respectively. Each ensemble model consists of two layers. In the first layer, four sub-models were trained using features such as AMF / MHF, MES, terminal motifs, and CNV.

[0162] The AMF / MHF sub-model training process is as follows: First, for each target, use the MHF matrix and AMF matrix of that target, and combine them with the corresponding labels of its samples (cancer or healthy, cancer type) to train an SVM model, and score the target based on the model (predicting the binary value of cancer-health, or predicting the probability value of each tissue origin); Second, based on the predicted score of the SVM model, train a stacked model using multiple modeling methods (logistic regression, random forest, SVM and gradient boosting) and output the second round of prediction values ​​(predicting the binary value of cancer-health, or predicting the probability value of each tissue origin) as the result of the AMF / MHF sub-model.

[0163] The MES, terminal motif, or CNV sub-models respectively use the MES value, CNV value, or terminal motif frequency of the target region to train a deep neural network sub-model or a logistic regression sub-model, and output a prediction score (a binary value predicting cancer-health, or a probability predicting the origin of each tissue). The second-layer model uses a logistic regression classifier to train and score the prediction scores output by the first-layer model, and finally outputs the cancer-health prediction result and the probability of TOO tissue origin (…). Figure 6 ).

[0164] Example 7: GutSeer model's multi-cancer detection and organ origination function

[0165] The trained GutSeer model was first tested on the training set samples using 10× cross-validation and achieved a high area under the receiver operating characteristic curve (AUC). The GutSeer model had an average AUC of 0.967 when distinguishing between cancer and normal samples (specificity 96.4%, overall sensitivity 89.0%, see...). Figure 7 A and Figure 8 A) indicates that the GutSeer model has high robustness and accuracy in predicting gastrointestinal cancers. In the test set, the GutSeer model's AUC value for cancer predictions was 0.964 (specificity 96.7%, overall sensitivity 86.2%, see [reference]). Figure 7 B) further proves its performance. Figure 7 C demonstrates GutSeer's sensitivity across various cancer types and clinical stages, showing its good detection capability for early-stage cancers (the same conclusion can be achieved by increasing specificity to 99%, see [link]). Figure 8 ).

[0166] Regarding tissue attribution, the overall accuracy of tissue attribution for samples predicted as cancer in the training and test groups was 84.7% and 82.0%, respectively. Figure 7 For D-7E, see the cancer type-specific accuracy details. Figure 7 F; the accuracy rates for tissue tracing of esophageal cancer and gastric cancer separately were 76.0% and 74.3%, respectively, see [reference needed]. Figure 10 As can be seen, GutSeer can accurately locate the organs in which common digestive system cancers occur. In summary, these data demonstrate GutSeer's powerful capabilities as a non-invasive multi-cancer screening method in detecting common digestive tract cancers and tracing their origins.

[0167] Example 8: Application of GutSeer in the Detection of Benign Diseases

[0168] Further independent validation was performed using 629 benign samples (including 174 cases of esophageal disease, 361 cases of gastric disease, 43 cases of colorectal disease, and 41 cases of liver disease). Using thresholds based on GutSeer scores corresponding to 95% and 99% specificity in the training set, the specificity of the benign samples was 87.1% and 93.9%, respectively. Figure 11 A; Table 4). Furthermore, the specificity was high in all benign disease tissues (esophageal diseases 88.1% and 94.6%, gastric diseases 89.7% and 95.8%, colorectal diseases 81.4% and 90.7%), except for liver cases (65.9% and 78%), which may be related to its small sample size. Figure 11B; Table 4). This indicates that GutSeer can be used not only for cancer screening, but also for the identification of benign and malignant tumors.

[0169] Furthermore, because benign gastrointestinal diseases typically encompass multiple stages (e.g., mild or moderate inflammation, severe inflammation with polyps or hyperplasia, low-grade neoplasia), and several other subclasses (e.g., esophageal smooth muscle tumors and lipomas), these benign samples are categorized for further comparison. Clearly, the specificity of benign diseases at each stage and subclass is high in the esophagus, stomach, and colorectal region. Figure 12 (Table 4). In particular, the specificity exceeded 90% for low-grade lesions in the stomach and esophagus, and exceeded 85% and 95% for other tumor subclasses in the stomach and esophagus (below the GutSeer 95% and 99% specificity thresholds). Figure 12 (Table 4). Specificity was low in some benign liver diseases (e.g., hemangioma, cirrhosis, hepatic space-occupying lesions), possibly due to the small sample size. However, in other relatively mild liver diseases, the specificity reached 78.9% and 94.7%, indicating that GutSeer remains superior in the identification of various benign liver diseases. Figure 12 (Table 4). These observations further demonstrate that GutSeer has stable specificity at all stages of most benign gastrointestinal diseases, and therefore, GutSeer has the potential to monitor benign tumors.

[0170] Table 4: Specificity of GutSeer identification for benign samples

[0171]

[0172] Example 9: Comparison of GutSeer with tumor markers and its combined use

[0173] To investigate whether GutSeer is superior to commonly used clinical cancer markers such as carcinoembryonic antigen (CEA), cancer antigen 199 (CA199), cancer antigen 724 (CA724), alpha-fetoprotein (AFP), and abnormal prothrombin (DCP), 242 subjects (187 cancer samples, 54 benign samples, and 1 healthy sample) participated in the detection of at least one of these common cancer markers. GutSeer showed higher sensitivity than any of the common cancer markers, especially CEA, CA199, and CA724. Even compared to DCP, the best-performing of the five common cancer markers (with a sensitivity of 83.0% at 96.9% specificity in benign samples), GutSeer still showed higher sensitivity (91.8%). Figure 11 C; Table 5).

[0174] Table 5: Sensitivity of serum cancer markers used in clinical practice

[0175]

[0176] To further investigate whether combining GutSeer with conventional cancer biomarkers could enhance detection performance, a simple combination method was tested, where a positive result was considered positive regardless of whether GutSeer or a conventional tumor biomarker was used. The results showed that combining GutSeer with conventional cancer biomarkers slightly improved sensitivity and slightly decreased specificity (Table 6). Therefore, GutSeer has the potential to replace conventional cancer biomarkers.

[0177] Table 6: Comparison of GutSeer with conventional cancer markers

[0178]

[0179] Example 10 Advantages of integrating methylation and fragmentomics

[0180] GutSeer's superior performance is attributed to its integration of ctDNA methylation and fragmentomics features. This is confirmed by comparing GutSeer with models that use the same model framework but only employ methylation (AMF and MHF) or fragmentomics (CNV and terminal motif) features.

[0181] In distinguishing between cancer and healthy samples, the average AUC values ​​of models built solely based on methylation or fragment omics features in the training set were 0.947 and 0.933, respectively, significantly lower than the average AUC of 0.967 after integration (Delong test p-values ​​were 1.12 × 10⁻⁶). -7 and 1.02×10 -11 , Figure 7 A; The results of the 10-fold cross-validation are shown in... Figure 8 B-8C); In the test set, the AUC values ​​for distinguishing cancer-healthy samples using methylation or fragmentomics models alone were 0.940 and 0.945, respectively, which were also significantly lower than the integrated AUC of 0.964. Figure 7 B, the p-values ​​for the Delong test were 4.82 × 10⁻⁶. -6 and 4.07×10 -6 The sensitivity of the GutSeer model in distinguishing between cancer and healthy samples was compared for all cancer samples and individual cancer types. In the training set, the sensitivity of the GutSeer model (89.0%) was significantly higher than that of using methylation alone (81.5%, p = 7.59 × 10⁻⁶). -6 The sensitivity of models based on chi-square test or fragment omics (78.4%, p = 3.19 × 10⁻⁶) was assessed. -9 Chi-square test; Figure 13(Tables 7 and 8). On the test set, the sensitivity of the GutSeer model (86.2%) was also significantly higher than that of methylation alone (77.7%, p = 2.59 × 10⁻⁶). -6 The sensitivity of models based on chi-square test or fragment omics (77.1%, p = 6.69 × 10⁻⁶) was assessed. -7 Chi-square test; Figure 13 (Tables 7 and 9). Furthermore, this improvement is observed in all cancers and at all clinical stages. Figure 13 Tables 7 and 9 show that integrating methylation and fragmentomics can improve the accuracy of distinguishing between cancer and healthy samples.

[0182] Table 7: Specific thresholds for cancer-health detection

[0183]

[0184] Table 8: Comparison of sensitivity of GutSeer in the training set with models using methylation or fragment omics features alone in predicting cancer and healthy samples.

[0185]

[0186]

[0187] Table 9: Comparison of sensitivity of GutSeer and models using methylation or fragmentomics features alone in predicting cancer and healthy samples in the test set.

[0188]

[0189]

[0190] Beyond distinguishing between cancer and healthy samples, the GutSeer model, integrating methylation and fragmentomics features, significantly improves organ origination performance. Cross-validation using samples from the training set shows that the GutSeer model achieves an organ origination accuracy of 84.7%. Figure 7 D), significantly higher than methylation alone (75.9%, p = 3.99 × 10⁻⁶). -7 The accuracy of the chi-square test or fragment omics features was 77.8%, p = 2.42 × 10⁻⁶. -6 Chi-square test; Figure 14 A and 14C; Table 10 shows the comparison of each cancer type). Also on the test set, the GutSeer model achieved an accuracy of 82.0% (…). Figure 7 E), significantly higher than methylation alone (75.6%, p = 6.19 × 10). -5 The accuracy of the chi-square test or fragment omics features was 75.4%, p = 2.98 × 10⁻⁶.-5 Chi-square test; Figure 14 B and 14D; Table 10 shows a comparison of each type of cancer.

[0191] Table 10: Comparison of the accuracy of GutSeer in tissue tracing with models using methylation or fragment omics features alone. The p-value of the chi-square test was used to evaluate the improvement that GutSeer, after integrating features, brought to models using methylation or fragment omics features alone.

[0192]

[0193]

[0194] Example 11: The impact of covariates and sequencing depth on GutSeer prediction results

[0195] The impact of demographic factors (such as age and sex) of volunteers participating in GutSeer on GutSeer prediction results was specifically analyzed. The GutSeer model was also analyzed in different age groups (<55 years old, 55-75 years old, and >75 years old). Figure 15 A) or comparisons of prediction results for cancer and healthy samples within different gender groups revealed that, apart from a slight but statistically significant difference in GutSeer scores between healthy male and healthy female samples, Figure 15 B), there were no significant differences among the other groups. Regarding gender, there were no differences in AUC, sensitivity, and specificity of GutSeer for predicting cancer-healthy samples across male, female, or propensity score-matched (PSM) groups (balancing age and gender between cancer and normal samples) groups (differences in AUC were assessed using the DeLong test, and differences in sensitivity and specificity were assessed using the chi-square test). Figure 15 Covariate analysis further collected other covariates that might affect GutSeer scores, such as smoking status, sample collection site, number of unique DNA molecules, and cfDNA extraction amount, for both cancer and healthy samples in the test set. The cancer and healthy samples in the test set were further divided into different subgroups based on these factors. Covariate analysis of these subgroups showed no significant differences in GutSeer scores among these subgroups. Figure 16 These results show that the GutSeer model is largely unaffected by the aforementioned covariates in predicting cancer-healthy samples, demonstrating high stability.

[0196] The possibility that sequencing depth affects the predictive performance of GutSeer was also investigated. Here, GutSeer sequencing data were randomly downsampled to reduce sequencing depth, and the same analysis was performed. The results showed that when the sequencing data volume was reduced to 0.5M (number of reads; 0.5M is equivalent to less than 5% of the total sequencing data), the sensitivity of cancer-health prediction and the accuracy of tissue origin analysis decreased by only 5.49% and 5.46%, respectively, and the correlation of AMF values ​​for the same methylation target between samples showed only a slight change. Figure 17 (See Table 11). This indicates that GutSeer's accuracy is less dependent on sequencing depth, suggesting that it can maintain high predictive accuracy while reducing sequencing volume and costs. This will contribute to the popularization and promotion of GutSeer.

[0197] Table 11: Accuracy and Sensitivity of GutSeer Predictions After Downsampling

[0198] 0.1M 0.5M 1M 2M 5M 10M Accuracy of organizational traceability 66.7％ 76.5％ 77.9％ 79.6％ 80.2％ 80.4％ Sensitivity 71.0％ 80.7％ 82.4％ 83.5％ 84.3％ 85.0％ Specificity 94.7％ 95.9％ 96.9％ 97.5％ 95.7％ 96.7％

[0199] Example 12: Reducing the impact of target number on GutSeer prediction results

[0200] To verify GutSeer's robustness to targets, the impact of the number of targets on GutSeer's prediction results was examined. 800 targets were randomly selected from 1656 targets, and the number was further reduced to 500 (i.e., randomly selected 500 targets from 1656 targets). Based on the data from these 800 or 500 targets, the same framework was used to build the model. Similarly, 10-fold cross-validation was performed on the training set, and the model was tested on the test set to predict tissue localization in cancer samples.

[0201] Based on model predictions at 800 sites ( Figure 18The results show that the area under the ROC curve for the training set cross-validation reached 0.953, with an overall sensitivity of 84.5% at 96.45% specificity. The individual sensitivities for each cancer type were: 87.8% (colorectal cancer), 80.6% (esophageal cancer), 71.9% (gastric cancer), 94.2% (liver cancer), and 83.7% (pancreatic cancer). The area under the ROC curve for the test set reached 0.958, with an overall sensitivity of 78.4% at 98.22% specificity. The individual sensitivities for each cancer type were: 82.2% (colorectal cancer), 71.7% (esophageal cancer), 60.2% (gastric cancer), 91.8% (liver cancer), and 84.1% (pancreatic cancer). Regarding the accuracy of tissue localization, the overall accuracy of cross-validation on the training set was 84.34%, with prediction accuracies for each cancer type as follows: 82.19% (colorectal cancer), 82.95% (esophageal and gastric cancer), 94.08% (liver cancer), and 60.87% (pancreatic cancer). The overall accuracy on the test set was 79.71%, with prediction accuracies for each cancer type as follows: 77.61% (colorectal cancer), 83.92% (esophageal and gastric cancer), 84.91% (liver cancer), and 51.11% (pancreatic cancer).

[0202] Based on model predictions at 500 sites ( Figure 19 The results show that the area under the ROC curve for the training set cross-validation reached 0.950, with an overall sensitivity of 82.61% at 96.45% specificity. The individual sensitivities for each cancer type were: 88.9% (colorectal cancer), 76.7% (esophageal cancer), 69.4% (gastric cancer), 91.3% (liver cancer), and 86.0% (pancreatic cancer). The area under the ROC curve for the test set reached 0.948, with an overall sensitivity of 82.58% at 93.13% specificity. The individual sensitivities for each cancer type were: 92.2% (colorectal cancer), 78.3% (esophageal cancer), 62.7% (gastric cancer), 93.5% (liver cancer), and 84.1% (pancreatic cancer). Regarding the accuracy of tissue localization, the overall accuracy of cross-validation on the training set was 81.92%, with prediction accuracies for each cancer type as follows: 96.15% (colorectal cancer), 72.40% (esophageal and gastric cancer), 91.28% (liver cancer), and 80.0% (pancreatic cancer). The overall accuracy on the test set was 77.52%, with prediction accuracies for each cancer type as follows: 78.13% (colorectal cancer), 71.28% (esophageal and gastric cancer), 86.50% (liver cancer), and 57.14% (pancreatic cancer).

[0203] The results above indicate that the model built based on a subset of the 1656 targets has slightly lower predictive performance than the model built based on all targets, but it still exhibits better cancer prediction and tissue localization performance. Even when the number of targets is less than half (800) or even less than one-third (500) of the total number of targets, the GutSeer model still predicts cancer and cancer tissue localization quite accurately. This demonstrates that the GutSeer model has high robustness to targets.

[0204] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0205] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This description is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A biomarker combination based on sample cfDNA targeted methylation sequencing for early detection, tissue localization, diagnosis, and benign / malignant identification of digestive system cancers, characterized in that, The biomarker combination consists of 1656 chromosomal regions as shown in Table 1 of the specification; wherein, when used for early detection, tissue localization, and diagnosis of digestive system cancers, the digestive system cancers are selected from esophageal cancer, gastric cancer, liver cancer, pancreatic cancer, and colorectal cancer; when used for the identification of benign or malignant digestive system cancers, the digestive system cancers are selected from esophageal cancer, gastric cancer, liver cancer, and colorectal cancer.

2. A reagent kit for early detection, tissue localization, diagnosis, and benign / malignant identification of digestive system cancers, characterized in that, The kit includes reagents for detecting the biomarker combination of claim 1; wherein, when used for early detection, tissue localization, and diagnosis of digestive system cancers, the digestive system cancers are selected from esophageal cancer, gastric cancer, liver cancer, pancreatic cancer, and colorectal cancer; when used for benign or malignant identification of digestive system cancers, the digestive system cancers are selected from esophageal cancer, gastric cancer, liver cancer, and colorectal cancer; the detection includes targeted methylation sequencing based on sample cfDNA.

3. The kit according to claim 2, wherein methylation status and fragment omics features are extracted based on the data from the targeted methylation sequencing.

4. The kit according to claim 3, wherein the methylation status is determined by measuring its mean methylation level (AMF) and methylation haplotype fraction (MHF).

5. The kit according to claim 3, wherein the fragmentomics features include terminal motifs and copy number variations.

6. The application of the reagent for combined detection of biomarkers as described in claim 1, combined with reagents for the detection of conventional cancer biomarkers, in the preparation of kits or devices for early detection, tissue localization, diagnosis, and benign / malignant identification of digestive system cancers; wherein, When used for early detection, tissue localization, and diagnosis of digestive system cancers, the digestive system cancers are selected from esophageal cancer, gastric cancer, liver cancer, pancreatic cancer, and colorectal cancer; when used for the identification of benign or malignant digestive system cancers, the digestive system cancers are selected from esophageal cancer, gastric cancer, liver cancer, and colorectal cancer; the detection includes targeted methylation sequencing based on sample cfDNA.

7. The application according to claim 6, wherein the conventional cancer markers include carcinoembryonic antigen (CEA), cancer antigen 199 (CA199), cancer antigen 724 (CA724), alpha-fetoprotein (AFP), and abnormal prothrombin (DCP).

Images