Tumor category detection system and method based on free DNA fragment omics characteristics

Abstract

The application belongs to the technical field of tumor category detection, and discloses a tumor category detection system and method based on free DNA fragment omics characteristics. The system comprises a sample processing module, a feature extraction module and a prediction module. The sample processing module is used for extracting free DNA from a sample to be detected, performing low-depth whole genome sequencing on the free DNA, obtaining sequencing data, and performing quality control and standardization processing on the sequencing data to obtain pre-processed sequencing data. The feature extraction module is used for extracting a plurality of free DNA fragment omics characteristics from the pre-processed sequencing data. The prediction module is used for inputting all the free DNA fragment omics characteristics into a pre-constructed integrated model, obtaining a detection result according to a prediction probability output by the integrated model, and the integrated model comprises a plurality of optimal base models, and each optimal base model corresponds to one free DNA fragment omics characteristic. The system can fully capture tumor-related characteristic information and improve the generalization ability and performance stability during detection.

Description

Technical Field

[0001] This invention belongs to the field of tumor category detection technology, specifically relating to a tumor category detection system and method based on cell-free DNA fragment omics characteristics. Background Technology

[0002] Bone and soft tissue tumors are diverse and have complex pathological characteristics. Accurate differentiation between benign and malignant tumors is a crucial prerequisite for clinical diagnosis and treatment decisions, directly impacting the choice of subsequent treatment plans and patient prognosis. Currently, common methods used in clinical practice to differentiate between benign and malignant bone and soft tissue tumors include imaging examinations and pathological biopsies. However, these methods all have significant limitations: imaging examinations have limited sensitivity and specificity, making it difficult to effectively differentiate early or atypical lesions; pathological biopsies are invasive procedures, potentially causing physical pain to patients, and are subject to problems such as sampling errors and long testing cycles, failing to meet the actual clinical need for non-invasive, rapid, and accurate differentiation between benign and malignant bone and soft tissue tumors.

[0003] Cell-free DNA (cfDNA) is mainly found in human peripheral blood. Its fragment omics characteristics (such as copy number variation, fragment size, and methylation status) are closely related to the occurrence and development of tumors, providing a new technical approach for non-invasive differentiation of benign or malignant bone or soft tissue tumors. Low-depth whole-genome sequencing (WGS) technology has significant advantages such as low detection cost, broad genome coverage, and small sample requirements, and has been gradually applied to early tumor screening and benign / malignant differentiation, providing technical support for non-invasive detection based on cfDNA.

[0004] While existing technologies include models for differentiating between benign and malignant tumors, these models have limited feature mining dimensions and lack systematic integration and complete extraction of multi-dimensional and multi-type cfDNA fragment omics features. They cannot fully capture tumor-associated feature information and are difficult to adapt to the complex pathological features of bone or soft tissue tumors. Furthermore, existing models have weak generalization ability and poor performance stability for different types of cfDNA fragment omics features, making it difficult to adapt to the actual application needs of different clinical scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a tumor category detection system and method based on cell-free DNA fragment omics features, which can fully capture tumor-associated feature information and improve the generalization ability and performance stability during detection.

[0006] The first aspect of this invention discloses a tumor category detection system based on cell-free DNA fragment omics characteristics, comprising:

[0007] The sample processing module is used to extract cell-free DNA from the sample to be tested, perform low-depth whole-genome sequencing on the cell-free DNA to obtain sequencing data, and perform quality control and standardization processing on the sequencing data to obtain preprocessed sequencing data.

[0008] The feature extraction module is used to extract various cell-free DNA fragment omics features from the preprocessed sequencing data. These features include copy number variation, fragment size, fragment omics-based methylation, griffins, and chromosome fragmentation.

[0009] The prediction module is used to input all the cell-free DNA fragment omics features into a pre-constructed ensemble model, and obtain the detection result based on the prediction probability output by the ensemble model. The ensemble model includes several optimal basic models, each of which corresponds to a cell-free DNA fragment omics feature. These models are obtained by training multiple machine learning algorithms through an automated machine learning platform and then filtering them. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal basic models.

[0010] In some implementations, a model building module is also included. This module is used to construct feature vectors for each cell-free DNA fragment omics feature and standardize them. All standardized feature vectors are input into an automated machine learning platform, which trains multiple machine learning algorithms. The optimal base model corresponding to each cell-free DNA fragment omics feature is selected using the area under the receiver operating characteristic curve as an evaluation metric to obtain the ensemble model. The predicted probability output by the ensemble model is a fusion result of the predicted probabilities output by all optimal base models using an arithmetic averaging strategy.

[0011] In some implementations, the various machine learning algorithms include at least random forest, extreme random tree, generalized linear model, extreme gradient boosting, gradient boosting machine and stacked ensemble model.

[0012] In some implementations, the model building module includes a screening unit, which is used to divide the training sample set into a training set and a validation set using a five-fold cross-validation strategy, calculate the area under the receiver operating characteristic (ROC) curve for each round of validation, and calculate the arithmetic mean of the ROC curves for all rounds of validation. For each cell-free DNA fragment omics feature, the machine learning algorithm corresponding to the largest arithmetic mean is selected as the optimal base model.

[0013] In some implementations, the predicted probability output by the integrated model is a predicted probability score of the sample to be detected being a malignant tumor. The prediction module is provided with a result determination unit, which is used to compare the predicted probability score with a preset determination threshold. When the predicted probability score is greater than the preset determination threshold, the detection result is set as malignant; otherwise, the detection result is set as benign.

[0014] In some embodiments, the feature extraction module includes a copy number variation extraction unit, a fragment size extraction unit, a fragmentomics-based methylation extraction unit, a griffin extraction unit, and a chromosome fragmentation extraction unit;

[0015] The copy number variation extraction unit is used to segment the genome into multiple fragments, analyze the sequencing depth of each fragment, compare it with the population baseline, calculate the copy number variation value of each fragment, and obtain a copy number variation map.

[0016] The fragment size extraction unit is used to select human autosomal regions, divide them into analysis intervals of preset length, and statistically analyze the number distribution of short and long free DNA fragments to obtain the proportion of short fragments and the proportion of long fragments in each analysis interval.

[0017] The fragmentomics-based methylation extraction unit is used to calculate the CGN to NCG ratio, the fragment ratios of CGC, NCG, CGN, and GCG within the CGCG sequence, and the CGN / NCG, CGC / NCG, and CGN / CGC motif ratios. The CGN / NCG motif ratio is calculated by dividing the number of cfDNA fragments carrying the 5' CGN motif by the number of fragments carrying the 5' NCG motif. The number of fragments carrying the 5' CGA, CGT, CGG, CGC, ACG, TCG, GCG, and CCG motifs are counted respectively.

[0018] The Griffin extraction unit is used to screen several transcription factors and analyze the average coverage, central coverage, and amplitude of the peak coverage around the binding site of each transcription factor within a preset area.

[0019] The chromosome fragmentation extraction unit is used to detect and analyze cell-free DNA sequencing data, extracting feature vectors at the sample level. These feature vectors quantify the burden of structural variations, breakpoint cluster size, number of copy number variant segments, and the chromosomes involved in each event that meets the high confidence criterion.

[0020] In some implementations, a validation module is also included, which is used to import multicenter prospective cohort sample data to validate the ensemble model and evaluate the ensemble model using AUC, sensitivity, and specificity metrics.

[0021] A second aspect of this invention discloses a method for detecting tumor categories based on the omics characteristics of cell-free DNA fragments, comprising:

[0022] Cell-free DNA is extracted from the sample to be tested, and low-depth whole-genome sequencing is performed on the cell-free DNA to obtain sequencing data. The sequencing data is then subjected to quality control and standardization to obtain preprocessed sequencing data.

[0023] Multiple cell-free DNA fragment omics features were extracted from the preprocessed sequencing data. These features included copy number variation, fragment size, fragment-based methylation, griffins, and chromosome fragmentation.

[0024] All the cell-free DNA fragment omics features are input into a pre-constructed ensemble model. The detection result is obtained based on the predicted probability output by the ensemble model. The ensemble model includes several optimal basic models, each of which corresponds to a cell-free DNA fragment omics feature. These models are obtained by training multiple machine learning algorithms through an automated machine learning platform and then filtering them. The predicted probability output by the ensemble model is the fusion result of the predicted probabilities output by all the optimal basic models.

[0025] In some implementations, constructing a classification model includes:

[0026] Each cell-free DNA fragment omics feature is constructed as a feature vector and standardized. All standardized feature vectors are input into an automated machine learning platform, and various machine learning algorithms are trained using the automated machine learning platform. The area under the receiver operating characteristic curve is used as the evaluation index to select the optimal base model corresponding to each cell-free DNA fragment omics feature, and the ensemble model is obtained. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal base models using an arithmetic averaging strategy.

[0027] In some implementations, multiple cell-free DNA fragment omics features are extracted from the sequencing data, including:

[0028] The genome is divided into multiple fragments, the sequencing depth of each fragment is analyzed, compared with the population baseline, the copy number variation value of each fragment is calculated, and a copy number variation map is obtained.

[0029] Human autosomal regions were selected and divided into analysis intervals of preset length. The number distribution of short and long free DNA fragments was statistically analyzed to obtain the proportion of short fragments and the proportion of long fragments in each analysis interval.

[0030] Calculate the CGN to NCG ratio, the fragment ratios of CGC, NCG, CGN, and GCG within the CGCG sequence, and the CGN / NCG, CGC / NCG, and CGN / CGC motif ratios. The CGN / NCG motif ratio is calculated by dividing the number of cfDNA fragments carrying the 5' CGN motif by the number of fragments carrying the 5' NCG motif. Count the number of fragments carrying the 5' CGA, CGT, CGG, CGC, ACG, TCG, GCG, and CCG motifs respectively.

[0031] Several transcription factors were screened, and the average coverage, central coverage, and amplitude of the peak coverage around the binding site of each transcription factor were analyzed.

[0032] The cell-free DNA sequencing data are analyzed to extract feature vectors at the sample level. These feature vectors quantify the burden of structural variations, the size of breakpoint clusters, the number of copy number variant segments, and the chromosomes involved in each event that meets the high confidence criterion.

[0033] The beneficial effects of this invention lie in its comprehensive capture of cfDNA differences between benign and malignant tumor patients by integrating multiple cfDNA fragment omics features. It combines various machine learning algorithms to construct an ensemble model, matching the optimal base model for each feature type and fusing all optimal base models. This overcomes the shortcomings of traditional single models that cannot adapt to multiple heterogeneous features, enabling the integration of effective information from multi-dimensional heterogeneous cfDNA features and significantly improving the overall generalization ability and robustness of the model. It also addresses the technical pain points of traditional models, such as large performance fluctuations, insufficient stability, and difficulty in simultaneously meeting the modeling needs of multiple feature types when adapting to different types of cfDNA fragment omics features, thus improving the generalization ability and performance stability during detection. Attached Figure Description

[0034] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0035] Unless otherwise specified or defined, the same reference numerals in different figures represent the same or similar technical features, and different reference numerals may be used to represent the same or similar technical features.

[0036] Figure 1 This is a schematic diagram of the structure of a tumor category detection system based on the omics characteristics of free DNA fragments disclosed in an embodiment of the present invention;

[0037] Figure 2 This is a comparison diagram of the copy number variation in the benign and malignant groups according to an embodiment of the present invention;

[0038] Figure 3 This is a comparison diagram of fragment size in the benign and malignant groups according to an embodiment of the present invention;

[0039] Figure 4 This is a comparison diagram of the effects of fragment omics-based methylation in benign and malignant groups according to an embodiment of the present invention;

[0040] Figure 5 This is a comparison diagram of chromosome fragmentation in the benign and malignant groups according to an embodiment of the present invention;

[0041] Figure 6 This is a volcano plot of differentially active transcription factors between the benign and malignant groups based on Griffin analysis, according to an embodiment of the present invention.

[0042] Figure 7 These are AUC curves of the base model and stacked model in the training and validation groups of this invention embodiment;

[0043] Figure 8 This is a schematic diagram illustrating the performance of the integrated model in predicting malignant tumor subgroups according to an embodiment of the present invention;

[0044] Figure 9 This is a schematic diagram illustrating the high specificity of the integrated model in this invention when distinguishing tumors that are easily misdiagnosed;

[0045] Figure 10 This is a flowchart of a tumor category detection method based on the omics characteristics of free DNA fragments, according to an embodiment of the present invention. Detailed Implementation

[0046] Unless otherwise specified or defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. When considering the technical solutions of this invention in a real-world context, all technical and scientific terms used herein may also have meanings corresponding to the objectives of achieving the technical solutions of this invention.

[0047] Unless otherwise specified or defined, the terms "described" or "the" as used herein refer to the technical features or technical content mentioned or described prior to the relevant section, which may be the same as or similar to the technical features or technical content mentioned herein. Furthermore, the terms "comprising" and "having," and any variations thereof, as used herein, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0048] The differentiation of benign or malignant bone or soft tissue tumors has limitations such as being invasive, having low sensitivity, insufficient specificity, and poor generalization ability. This invention addresses these limitations by integrating and analyzing multi-dimensional fragment omics features of cfDNA and combining them with machine learning techniques to construct an ensemble model, achieving non-invasive, efficient, and accurate differentiation of benign or malignant bone or soft tissue tumors. It should be noted that the tumor category detection system and method of this invention are not limited to bone or soft tissue tumors, but can also be used for various solid tumors with diverse tumor types and complex pathological characteristics.

[0049] like Figure 1 As shown, the tumor category detection system of this invention specifically includes: a sample processing module, a feature extraction module, and a prediction module.

[0050] The sample processing module is used to extract cell-free DNA from the sample to be tested, perform low-depth whole-genome sequencing on the cell-free DNA to obtain sequencing data, perform quality control and standardization on the sequencing data to obtain preprocessed sequencing data.

[0051] In this embodiment, the sample to be tested is a peripheral blood sample from a subject with bone or soft tissue tumors. The equipment includes EDTA blood collection tubes, a high-speed centrifuge, a cfDNA extraction kit (QIAamp Cyclic Nucleic Acid Kit), a quantitative kit (QubitdsDNA HS Detection Kit), a library construction kit (KAPA Hyper Prep Kit), a qPCR quantitative reagent (KAPA SYBR FASTqPCR Master Mix), and a NovaSeq sequencing platform, which are used to complete peripheral blood collection, plasma separation, cfDNA extraction, quantification, library construction, sequencing, and quality control and standardization of sequencing data. The specific procedure is as follows: 10 ml of peripheral blood was collected into an EDTA tube, and plasma was extracted by centrifugation at 16,000x for 10 minutes within 4 hours after blood collection; cfDNA was extracted from the thawed plasma sample using the QIAamp Cyclic Nucleic Acid Kit, and the cfDNA sample was stored at -80℃ and transported to the central laboratory via dry ice; cfDNA was quantified using the QubitdsDNA HS Detection Kit, and 5-10 nanograms of cfDNA were taken and DNA end repair, A-tail addition, and adapter ligation were performed using the KAPA Hyper Prep Kit to construct a WGS library; the WGS library was quantified using the KAPA SYBR FAST qPCRMaster Mix, and paired-end sequencing was performed on the NovaSeq platform; quality control of the sequencing data was performed: the sequence reading was optimized using Trimmomatic software to remove low-quality sequences; and PCR duplicates were eliminated using the Picard toolkit. The pruned sequences were aligned with the human reference genome GRCh37 / UCSC hg19 version using the Burrows-Wheeler Aligner for accurate alignment. The WGS data were downsampled to a uniform 5-fold coverage to eliminate potential biases caused by inconsistencies in depth, resulting in preprocessed sequencing data.

[0052] The feature extraction module is used to extract various cell-free DNA fragment omics features from preprocessed sequencing data. These features include copy number variation (CNV), fragment size (FS), fragment omics-based methylation (FBM), griffin, and chromosome fragmentation. Accordingly, the feature extraction module includes a copy number variation extraction unit, a fragment size extraction unit, a fragment omics-based methylation extraction unit, a griffin extraction unit, and a chromosome fragmentation extraction unit.

[0053] Specifically, in the copy number variation extraction unit, the ichorCNA method was used to divide each genome of chromosomes 1-22 into 1Mb fragments, resulting in a total of 2475 fragments. The sequencing depth of each fragment was analyzed using a hidden Markov model (HMM), and after comparison with the population baseline, the copy number variation value of each fragment was calculated to obtain a copy number variation map, which serves as a CNV feature.

[0054] In the fragment size extraction unit, human autosomal regions were selected and divided into analysis intervals of preset length. The number distribution of short and long cell-free DNA fragments was statistically analyzed to obtain the proportion of short and long fragments in each analysis interval. Specifically, the original BAM file was filtered, and each chromosome arm of autosomes 1 to 22 was divided into 5 megabase pair (Mb) intervals, resulting in 541 intervals. The number distribution of short cell-free DNA fragments (100-150 base pairs) and long cell-free DNA fragments (150-220 base pairs) within each interval was statistically analyzed; the proportion of short and long fragments in each interval was obtained. After local weighted scatter regression normalization, interval division, and Z-score standardization, a 1082-dimensional FS feature set was obtained.

[0055] In the fragmentomics-based methylation extraction unit, the correlation ratios of CGN and NCG were calculated, the number of 5' motif fragments of each type was counted, and after Z-fractionation, a standardized FBM feature set was obtained. Specifically, the CGN / NCG ratio was calculated to obtain 8 fragment ratios; the CGC, NCG, CGN, and GCG fragment ratios within the CGCG sequence were calculated to obtain 10 features; the ratios of CGN / NCG, CGC / NCG, and CGN / CGC motifs were calculated; the CGN / NCG motif ratio was calculated by dividing the number of cfDNA fragments carrying 5' CGN motifs (5'-CGA, CGT, CGG, and CGC motifs) by the number of fragments carrying 5' NCG motifs (-ACG, TCG, GCG, and CCG motifs); the number of fragments carrying CGA, CGT, CGG, CGC, ACG, TCG, GCG, and CCG motifs at the 5' end was counted, and after Z-fractionation, a standardized FBM feature set was obtained.

[0056] In the Griffin extraction unit, several transcription factors were screened, and the average coverage, central coverage, and amplitude of the peak coverage around the binding site of each transcription factor were analyzed. Specifically, 854 TFs (transcription factors) related to tumor development were selected from the GTRD database v21.12. The region from -5kbp upstream to 5kbp downstream of each TF binding site (TFBS) was extracted as the analysis window. The coverage of sequencing fragments within each analysis window was calculated and corrected, such as the average coverage, central coverage, and amplitude of the peak coverage around the center calculated by fast Fourier transform, forming the Griffin feature set.

[0057] In the chromosome fragmentation extraction unit, cell-free DNA sequencing data are analyzed to extract sample-level feature vectors. These feature vectors quantify the burden of structural variations, breakpoint cluster size, number of copy number variant segments, and the chromosomes involved in events meeting each high-confidence criterion. The specific process is as follows: First, structural variation (SV) analysis is performed using DELLY, CNV analysis using WisecondorX, and chromosome fragmentation event analysis using ShatterSeek. High-confidence chromosome fragmentation events are classified using a dual-criteria framework: Criterion 1 requires at least 6 intrachromosomal SVs, at least 7 oscillating CN segments, significant fragment rejoining events, and significant breakpoint clusters; Criterion 2 requires at least 3 intrachromosomal and 4 interchromosomal SVs, accompanied by a unique oscillating CN pattern. Sample-level feature vectors are extracted as Chromothripsis features.

[0058] Figures 2-6 This demonstrates the differences in cfDNA fragment characteristics between benign and malignant groups. From Figure 2 It can be seen that the copy number variation gain and deletion events were significantly more numerous in the malignant group than in the benign group; from Figure 3 It can be seen that the proportion of short fragments in malignant tumors is higher than that in benign lesions; from Figure 4 It can be seen that the methylation level of the malignant group based on fragmentomics is lower than that of the benign group; from Figure 5 It can be seen that the frequency of chromosome fragmentation-related CNV events was higher in the malignant group than in the benign group; Figure 6 A volcano plot of differentially active transcription factors between the benign and malignant groups based on Griffin analysis.

[0059] The prediction module is used to input all cell-free DNA fragment omics features into a pre-built ensemble model and obtain the detection result based on the prediction probability output by the ensemble model. The ensemble model includes several optimal base models, each of which corresponds to a cell-free DNA fragment omics feature. These models are obtained by training multiple machine learning algorithms through an automated machine learning platform and then filtering them. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal base models.

[0060] Specifically, the predicted probability output by the ensemble model is the predicted probability score of the sample to be detected as a malignant tumor. The five extracted fragment omics features are constructed into five independent feature vectors X_1, X_2, X_3, X_4, and X_5, respectively. The features of each dimension are standardized (e.g., Z-score normalization) to ensure that the numerical components of each feature are comparable. Then, these vectors are input into the ensemble model to obtain the predicted probability score of malignant tumors. Then, the predicted probability scores are compared through the result determination unit. and preset judgment threshold When the predicted probability score Greater than the preset judgment threshold If the test result is malignant, it is set as malignant; otherwise, it is set as benign.

[0061] This embodiment also includes a model building module, which is used to construct feature vectors for each free DNA fragment omics feature and standardize them. All standardized feature vectors are input into an automated machine learning platform, which trains various machine learning algorithms. The optimal base model corresponding to each free DNA fragment omics feature is selected using the area under the receiver operating characteristic curve as the evaluation index to obtain an ensemble model. The fusion result of the predicted probabilities output by all the optimal base models is fused using an arithmetic mean strategy as the predicted probability output by the ensemble model.

[0062] Specifically, 84 patients with benign bone or soft tissue tumors and 85 patients with malignant bone or soft tissue tumors were included as the modeling cohort. Peripheral blood (10 mL) was collected from all subjects into EDTA tubes. Plasma was separated by centrifugation at 16,000x for 10 minutes within 4 hours of blood collection. cfDNA was extracted from the thawed plasma samples using the QIAamp Cyclic Nucleic Acid Kit (Qiagen). The cfDNA samples were stored at -80°C and transported to the central laboratory of the clinical testing center using dry ice.

[0063] The samples were processed following the steps outlined in the sample processing and feature extraction modules, extracting five cfDNA fragment omics features and standardizing them. The preprocessed feature vectors were then input into the H2O AutoML automated machine learning platform. For each type of fragment omics feature, the system executed the following automated training process.

[0064] Candidate Algorithm Selection: The AutoML engine automatically invokes multiple machine learning algorithms in parallel, including but not limited to: Distributed Random Forest (DRF), Extreme Random Tree (XRT), Generalized Linear Model (GLM), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), and stacked ensemble models. Among these, Distributed Random Forest and Extreme Random Tree are used to handle nonlinear relationships and noise in features; Generalized Linear Model is used to provide baseline linear predictions with regularization (Elastic Net) constraints; Gradient Boosting Machine and Extreme Gradient Boosting are used to accurately capture deep interactions between features by iteratively optimizing the loss function.

[0065] Hyperparameter space search: Perform Cartesian grid search or random search for each algorithm to optimize key parameters such as decision tree depth, learning rate, and regularization coefficient.

[0066] Five-fold cross-validation and performance evaluation were performed using a selection unit: To prevent overfitting and ensure the robustness of the evaluation results, a five-fold cross-validation strategy was used for model training for each feature class. The training sample set D was randomly divided into five mutually exclusive subsets of similar size, F1…F5. Five iterations were performed. In the k-th iteration, the four subsets other than Fk were used as the training set, and Fk was used as the internal validation set. The area under the receiver operating characteristic (AUC) curve was used as the core evaluation metric. The arithmetic mean of the five validation results was calculated. For each cfDNA characteristic type, according to Ranking involves selecting the best-performing single model from all candidate algorithm instances, ultimately resulting in five optimal base models. , }, which correspond to five fragment omics feature categories respectively.

[0067] Multi-dimensional ensemble integration based on arithmetic averaging strategy: A probability fusion ensemble strategy is used to integrate the prediction results of the basic models in each dimension. The specific steps are as follows:

[0068] For the sample S to be detected, extract its corresponding five feature vectors { When input into the corresponding optimal base model, the output sample S of each model is the predicted probability score of malignancy. : .

[0069] An arithmetic averaging strategy is used to aggregate the predicted probability scores output by each optimal base model. This strategy effectively mitigates potential random fluctuations or prediction biases that may exist in a single feature dimension. The final ensemble model outputs the predicted probability scores. The calculation formula is: .

[0070] By leveraging H2O AutoML to automatically and concurrently call and train multiple machine learning algorithms, basic models corresponding to each of the five cfDNA fragmentomics features were constructed. A suitable algorithm architecture was matched for each feature class, overcoming the limitation of traditional single models being unable to adapt to multiple heterogeneous features. Then, through five-fold cross-validation and using the area under the receiver operating characteristic curve (AUC) as the evaluation metric, the optimal basic model for each feature was selected, avoiding the subjectivity and limitations of manual selection and ensuring optimal discriminative performance for each feature class. Finally, an arithmetic averaging strategy was used to integrate the optimal basic models, resulting in an ensemble model that integrates effective information from multi-dimensional heterogeneous cfDNA features, significantly improving the overall generalization ability and robustness of the model. This addresses the technical pain points of traditional models, such as large performance fluctuations, insufficient stability, and difficulty in meeting the modeling needs of multiple feature classes when adapting to different types of cfDNA fragmentomics features, resulting in a significant improvement in the discriminative accuracy and clinical applicability of the ensemble model.

[0071] A validation module is also included. This module imports multicenter prospective cohort sample data to validate the ensemble model and evaluates it using AUC, sensitivity, and specificity metrics. Specifically, a multicenter prospective cohort containing 58 benign, 57 malignant, and 51 intermediate-sized bone or soft tissue tumors was selected as the validation cohort. The validation cohort samples were preprocessed according to the aforementioned method, fragment omics features were extracted, input into the constructed ensemble model, and the detection results were output. See [link to relevant documentation]. Figure 7 As shown, the validation results demonstrate that the model can effectively distinguish between benign and malignant tumors in the validation cohort (AUC: 0.941; sensitivity: 84.2%; specificity: 98.3%), proving that the model has good discrimination effect and generalization ability.

[0072] Furthermore, this ensemble model demonstrated robust predictive performance in the malignant tumor subgroup, covering key clinical stratification variables such as AJCC stage, histological grade, histological type, and lung metastasis status; simultaneously, it exhibited high specificity in differentiating easily misdiagnosed tumors, effectively reducing the risk of clinical misdiagnosis. See also Figures 8-9 ,from Figure 8As shown in Figure A, in both the training and validation cohorts, the model scores for bone-derived tumors were significantly higher than those for soft tissue-derived tumors. Figure 8 As shown in Figure B, in the training queue, a higher AJCC score is significantly correlated with a higher model score. Figure 8 As shown in Figure C, in both the training and validation cohorts, the model scores for poorly differentiated tumors were significantly higher than those for well-differentiated tumors. Figure 8 The D-plot shows that in both the training and validation cohorts, patients without lung metastases had significantly lower model scores than those with lung metastases. Figure 9 Figure A shows significant differences in model scores among chondroblastoma (CB), enchondroma (ECH), and chondrosarcoma (CS). Figure 9 Figure B shows a significant difference in model scores between schwannomas (SN) and malignant peripheral nerve schwannomas (MPNST). Figure 9 As shown in Figure C, there is a significant difference in model scores between osteoblastoma (OB) and osteosarcoma (OS).

[0073] In summary, the tumor category detection system of this invention, based on peripheral blood cfDNA detection, eliminates the need for invasive pathological biopsy, effectively reducing patient suffering and the risk of sampling bias, making it suitable for early tumor screening and long-term monitoring. It integrates five cfDNA fragment omics features to comprehensively capture the differences in cfDNA between benign and malignant bone or soft tissue tumor patients, and combines multiple machine learning algorithms to construct an integrated model, improving the model's generalization ability and clinical applicability.

[0074] like Figure 10 As shown, based on the above-mentioned tumor category detection system based on cell-free DNA fragment omics features, this embodiment of the invention discloses a tumor category detection method based on cell-free DNA fragment omics features, including:

[0075] Step S100: Extract cell-free DNA from the sample to be tested, perform low-depth whole-genome sequencing on the cell-free DNA to obtain sequencing data, and perform quality control and standardization on the sequencing data to obtain preprocessed sequencing data;

[0076] Step S200: Extract multiple cell-free DNA fragment omics features from the preprocessed sequencing data. The multiple cell-free DNA fragment omics features include: copy number variation, fragment size, fragment omics-based methylation, griffin, and chromosome fragmentation.

[0077] Step S300: Input all the cell-free DNA fragment omics features into a pre-constructed ensemble model, and obtain the detection result based on the prediction probability output by the ensemble model. The ensemble model includes several optimal basic models, each of which corresponds to a cell-free DNA fragment omics feature. It is obtained by training multiple machine learning algorithms through an automated machine learning platform and then screening them. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal basic models.

[0078] In some implementations, constructing a classification model includes:

[0079] Each cell-free DNA fragment omics feature is constructed as a feature vector and standardized. All standardized feature vectors are input into an automated machine learning platform, and various machine learning algorithms are trained using the automated machine learning platform. The area under the receiver operating characteristic curve is used as the evaluation index to select the optimal base model corresponding to each cell-free DNA fragment omics feature, and the ensemble model is obtained. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal base models using an arithmetic averaging strategy.

[0080] In some implementations, multiple cell-free DNA fragment omics features are extracted from the sequencing data, including:

[0081] The genome is divided into multiple fragments, the sequencing depth of each fragment is analyzed, compared with the population baseline, the copy number variation value of each fragment is calculated, and a copy number variation map is obtained.

[0082] Human autosomal regions were selected and divided into analysis intervals of preset length. The number distribution of short and long free DNA fragments was statistically analyzed to obtain the proportion of short fragments and the proportion of long fragments in each analysis interval.

[0083] Calculate the CGN to NCG ratio, the fragment ratios of CGC, NCG, CGN, and GCG within the CGCG sequence, and the CGN / NCG, CGC / NCG, and CGN / CGC motif ratios. The CGN / NCG motif ratio is calculated by dividing the number of cfDNA fragments carrying the 5' CGN motif by the number of fragments carrying the 5' NCG motif. Count the number of fragments carrying the 5' CGA, CGT, CGG, CGC, ACG, TCG, GCG, and CCG motifs respectively.

[0084] Several transcription factors were screened, and the average coverage, central coverage, and amplitude of the peak coverage around the binding site of each transcription factor were analyzed.

[0085] The cell-free DNA sequencing data are analyzed to extract feature vectors at the sample level. These feature vectors quantify the burden of structural variations, the size of breakpoint clusters, the number of copy number variant segments, and the chromosomes involved in each event that meets the high confidence criterion.

[0086] For a detailed introduction to tumor category detection methods, please refer to the aforementioned content on tumor category detection systems, which will not be repeated here.

[0087] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

[0088] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A tumor classification detection system based on cell-free DNA fragment omics characteristics, characterized in that, include: The sample processing module is used to extract cell-free DNA from the sample to be tested, perform low-depth whole-genome sequencing on the cell-free DNA to obtain sequencing data, and perform quality control and standardization processing on the sequencing data to obtain preprocessed sequencing data. The feature extraction module is used to extract various cell-free DNA fragment omics features from the preprocessed sequencing data. These features include copy number variation, fragment size, fragment omics-based methylation, griffins, and chromosome fragmentation. The prediction module is used to input all the cell-free DNA fragment omics features into a pre-constructed ensemble model, and obtain the detection result based on the prediction probability output by the ensemble model. The ensemble model includes several optimal basic models, each of which corresponds to a cell-free DNA fragment omics feature. These models are obtained by training multiple machine learning algorithms through an automated machine learning platform and then filtering them. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal basic models.

2. The tumor category detection system as described in claim 1, characterized in that, It also includes a model building module, which is used to construct feature vectors for each cell-free DNA fragment omics feature and standardize them. All standardized feature vectors are input into an automated machine learning platform, and multiple machine learning algorithms are trained using the automated machine learning platform. The optimal base model corresponding to each cell-free DNA fragment omics feature is selected using the area under the receiver operating characteristic curve as the evaluation index to obtain the ensemble model. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal base models using an arithmetic averaging strategy.

3. The tumor category detection system as described in claim 2, characterized in that, The various machine learning algorithms mentioned include at least random forest, extreme random tree, generalized linear model, extreme gradient boosting, gradient boosting machine and stacked ensemble model.

4. The tumor category detection system as described in claim 2, characterized in that, The model building module includes a screening unit, which is used to divide the training sample set into a training set and a validation set using a five-fold cross-validation strategy, calculate the area under the receiver operating characteristic (ROC) curve for each round of validation, and calculate the arithmetic mean of the ROC curves for all rounds of validation. For each cell-free DNA fragment omics feature, the machine learning algorithm corresponding to the largest arithmetic mean is selected as the optimal base model.

5. The tumor category detection system as described in claim 1, characterized in that, The prediction probability output by the integrated model is a prediction probability score of the sample to be detected as a malignant tumor. The prediction module is equipped with a result determination unit, which is used to compare the prediction probability score with a preset determination threshold. When the prediction probability score is greater than the preset determination threshold, the detection result is set as malignant; otherwise, the detection result is set as benign.

6. The tumor category detection system as described in claim 1, characterized in that, The feature extraction module includes a copy number variation extraction unit, a fragment size extraction unit, a fragmentomics-based methylation extraction unit, a griffin extraction unit, and a chromosome fragmentation extraction unit; The copy number variation extraction unit is used to segment the genome into multiple fragments, analyze the sequencing depth of each fragment, compare it with the population baseline, calculate the copy number variation value of each fragment, and obtain a copy number variation map. The fragment size extraction unit is used to select human autosomal regions, divide them into analysis intervals of preset length, and statistically analyze the number distribution of short and long free DNA fragments to obtain the proportion of short fragments and the proportion of long fragments in each analysis interval. The fragmentomics-based methylation extraction unit is used to calculate the CGN to NCG ratio, the fragment ratios of CGC, NCG, CGN, and GCG within the CGCG sequence, and the CGN / NCG, CGC / NCG, and CGN / CGC motif ratios. The CGN / NCG motif ratio is calculated by dividing the number of cfDNA fragments carrying the 5' CGN motif by the number of fragments carrying the 5' NCG motif. The number of fragments carrying the 5' CGA, CGT, CGG, CGC, ACG, TCG, GCG, and CCG motifs are counted respectively. The Griffin extraction unit is used to screen several transcription factors and analyze the average coverage, central coverage, and amplitude of the peak coverage around the binding site of each transcription factor within a preset area. The chromosome fragmentation extraction unit is used to detect and analyze cell-free DNA sequencing data, extracting feature vectors at the sample level. These feature vectors quantify the burden of structural variations, breakpoint cluster size, number of copy number variant segments, and the chromosomes involved in each event that meets the high confidence criterion.

7. The tumor category detection system as described in claim 1, characterized in that, It also includes a validation module, which is used to import multicenter prospective cohort sample data to validate the ensemble model and evaluate the ensemble model using AUC, sensitivity and specificity metrics.

8. A tumor category detection method based on cell-free DNA fragment omics characteristics, characterized in that, include: Cell-free DNA is extracted from the sample to be tested, and low-depth whole-genome sequencing is performed on the cell-free DNA to obtain sequencing data. The sequencing data is then subjected to quality control and standardization to obtain preprocessed sequencing data. Multiple cell-free DNA fragment omics features were extracted from the preprocessed sequencing data. These features included copy number variation, fragment size, fragment-based methylation, griffins, and chromosome fragmentation. All the cell-free DNA fragment omics features are input into a pre-constructed ensemble model. The detection result is obtained based on the predicted probability output by the ensemble model. The ensemble model includes several optimal basic models, each of which corresponds to a cell-free DNA fragment omics feature. These models are obtained by training multiple machine learning algorithms through an automated machine learning platform and then filtering them. The predicted probability output by the ensemble model is the fusion result of the predicted probabilities output by all the optimal basic models.

9. The tumor category detection method as described in claim 8, characterized in that, Building a classification model includes: Each cell-free DNA fragment omics feature is constructed as a feature vector and standardized. All standardized feature vectors are input into an automated machine learning platform, and various machine learning algorithms are trained using the automated machine learning platform. The area under the receiver operating characteristic curve is used as the evaluation index to select the optimal base model corresponding to each cell-free DNA fragment omics feature, and the ensemble model is obtained. The prediction probability output by the ensemble model is the fusion result of the prediction probabilities output by all the optimal base models using an arithmetic averaging strategy.

10. The tumor category detection method as described in claim 8, characterized in that, Multiple cell-free DNA fragment omics features were extracted from the sequencing data, including: The genome is divided into multiple fragments, the sequencing depth of each fragment is analyzed, compared with the population baseline, the copy number variation value of each fragment is calculated, and a copy number variation map is obtained. Human autosomal regions were selected and divided into analysis intervals of preset length. The number distribution of short and long free DNA fragments was statistically analyzed to obtain the proportion of short fragments and the proportion of long fragments in each analysis interval. Calculate the CGN to NCG ratio, the fragment ratios of CGC, NCG, CGN, and GCG within the CGCG sequence, and the CGN / NCG, CGC / NCG, and CGN / CGC motif ratios. The CGN / NCG motif ratio is calculated by dividing the number of cfDNA fragments carrying the 5' CGN motif by the number of fragments carrying the 5' NCG motif. Count the number of fragments carrying the 5' CGA, CGT, CGG, CGC, ACG, TCG, GCG, and CCG motifs respectively. Several transcription factors were screened, and the average coverage, central coverage, and amplitude of the peak coverage around the binding site of each transcription factor were analyzed. The cell-free DNA sequencing data are analyzed to extract feature vectors at the sample level. These feature vectors quantify the burden of structural variations, the size of breakpoint clusters, the number of copy number variant segments, and the chromosomes involved in each event that meets the high confidence criterion.

Images