A method for designing a tumor diagnosis hybrid capture panel based on cfDNA whole genome sequencing fragmentomic signal mining

Abstract

The application provides a tumor diagnosis hybrid capture panel design method based on cfDNA whole genome sequencing fragmentomics signal mining. The method takes cfDNA WGS data as prior information for panel region selection, extracts features based on different functional annotation regions, and then realizes panel design according to the extracted features. The application integrates information mining and panel design into a single repeatable method, and does not require other prior information, effectively avoiding the interference of prior knowledge, and has good universality.

Description

Technical Field

[0001] This invention relates to the fields of computational genomics, liquid biopsy, and tumor diagnostics, and more specifically, to a method for designing a hybridization capture panel for tumor diagnostics based on cfDNA whole-genome sequencing fragment omics signal mining. Background Technology

[0002] Cell-free DNA (cfDNA) is DNA fragments released into the plasma after apoptosis or necrosis, predominantly consisting of single nucleosome fragments of approximately 167 bp due to nucleosome protection. Its fragment length, terminal position, and terminal base sequence encode information such as nucleosome localization, chromatin openness, transcription factor occupancy, and tissue origin of the released cell. Methods for characterizing and modeling these fragmentation patterns are collectively referred to as fragmentomics. Tumor-derived cfDNA in plasma carries abnormal fragmentation characteristics, reflecting dysregulated chromatin structure in malignant cells, thus allowing for cancer detection without the need to detect somatic mutations. In recent years, fragmentomics has evolved from analyzing fragment length ratios at the whole genome level to site-specific readouts such as nucleosome footprinting, transcription factor binding site coverage, terminal base sequence profiles, and fragment length distribution entropy, and has expanded to scenarios such as early pan-cancer detection, tissue origin tracing, and monitoring of minimal residual disease. Compared to ctDNA detection, which relies on somatic mutations, fragmentomics utilizes nucleosome and epigenetic signal and information dimensions to cover the entire genome. It has the advantages of being mutation-independent and having relatively controllable cost per sample, and has become an important complementary signal source for early cancer detection and subtyping.

[0003] Currently, cfDNA fragmentomics in cancer diagnosis mainly falls into two technical routes: one is the whole-genome approach, which uses 1-2× low-throughput whole-genome sequencing. This involves dividing the genome into fixed-size (5 Mb) bins and calculating the fragment size ratio (short / long) for each bin to extract signals. This route has the advantages of comprehensive genome coverage and high information content, but its drawback is that detecting sub-1% abundance ctDNA requires high-depth WGS sequencing of at least 30×, resulting in extremely high sequencing costs and poor economic viability for large-scale clinical deployment. The other route is the targeted panel sequencing approach based on prior biology. This approach relies on hybridization capture technology to achieve ultra-high-depth targeted sequencing of 500-1000×. The targeted regions are mainly pre-defined based on existing literature and biological experience, including cancer driver gene exons, transcription factor binding sites, and various regulatory regions. This approach achieves high sequencing depth, controllable cost per sample, and is more suitable for clinical testing scenarios. However, the target region screening of this approach lacks data-driven approaches, and the region selection relies entirely on researchers' experience and existing research. It does not mine the optimal interval based on the real signal characteristics of cfDNA fragmentomics, and the target region design is highly subjective and lacks scientific rigor and systematicity.

[0004] Current cfDNA fragmentomics targeted detection technologies suffer from several core shortcomings. First, traditional empirical panels often select exon regions, which are misaligned with real signal sources such as regulatory regions and transcription factor binding sites rich in fragmentomics signals, resulting in poor performance of core detection features. Second, panel design relies solely on known biological priors, leading to an inability to uncover unknown signal hotspots and a high risk of losing valuable diagnostic information. Third, there is a lack of a data-driven, systematic screening framework; panel design depends on manual trial and error, lacking standardized and reproducible protocols. Fourth, the methodologies of whole-genome sequencing (WGS) and targeted panel systems are disconnected, hindering the transfer and transformation of signaling results and severely limiting technological iteration and diagnostic performance improvement.

[0005] In summary, there is an urgent need in this field for a method that can systematically mine genomic regions enriched with fragmentomics diagnostic signals from existing cfDNA WGS data and transform these regions into clinically deployable hybridization capture panels, thereby enabling the panels to be used in scenarios such as combined cancer detection and tissue-derived typing. Summary of the Invention

[0006] The purpose of this invention is to provide a method for designing hybrid capture panels for tumor diagnosis based on fragment omics signal mining from cfDNA whole-genome sequencing.

[0007] In a first aspect, the present invention provides a method for designing panels based on cfDNA whole-genome sequencing, the method comprising the steps of: (S1) Provide a cfDNA whole genome sequencing dataset; the cfDNA whole genome sequencing dataset includes a training set and a validation set; the training set or validation set includes healthy control samples, samples of various cancer types and other disease samples; (S2) Preprocess the cfDNA whole genome sequencing dataset to obtain alignment fragment files; (S3) Based on different functional annotation regions, extract features from the sequencing data of the samples in the training set, including: (S3.1) Extracting features from the region panels includes: extracting fragments from the sequencing data of each sample based on the alignment fragment file, and screening the fragments; for a sample, mapping each screened fragment to one or more region panels that overlap with it; evaluating the 9-dimensional features and terminal base sequence diversity score of each region panel, and generating a region feature tensor for the sample based on the 9-dimensional features and terminal base sequence diversity score; and (S3.2) Extracting features from biological units includes: extracting fragments from the sequencing data of each sample based on the biological unit dataset, and filtering the fragments; for a sample, mapping each filtered fragment to one or more biological units, evaluating the entropy value of all fragments mapped to each biological unit, and generating an entropy matrix; (S4) Based on L1 regularization, sparse regions are filtered using steps selected from the following group: (S4.1) Based on the entropy matrix, a classification algorithm is used to perform binary classification modeling for healthy and cancerous individuals, obtaining the detection model and the corresponding preferred biological units; and / or (S4.2) Based on the entropy matrix, a classification algorithm is used to perform binary classification modeling of each cancer type with other cancer / disease types to obtain a unified classification model and the corresponding preferred biological unit; and / or (S4.3) Based on the different features obtained in step (S3), a classification algorithm is used to perform binary classification modeling for each cancer type and other cancer types to obtain a binary classification model corresponding to each feature; all the binary classification models are validated in the validation set to obtain a heterogeneity subtyping model for each cancer type and a preferred biological unit. (S5) Map the preferred biological units to their corresponding genomic sites from the preferred biological units, perform upstream and downstream sequence expansion on each genomic site, and merge and sort all genomic sites to obtain a panel.

[0008] In another preferred embodiment, the method further includes the step of: (S6) Perform significance verification on the panel using steps selected from the group below: (S6.1) Random zero-distribution control verification in areas with equal coverage length; (S6.2) Zero-distribution comparison verification with the same function annotation area; (S6.3) Genome-wide reference validation; and / or (S6.4) Perform validation on the validation set.

[0009] In another preferred embodiment, the cancers include: liver cancer, lung cancer, breast cancer, colorectal cancer, pancreatic cancer, and ovarian cancer.

[0010] In another preferred embodiment, the other diseases include: inflammation, autoimmune diseases.

[0011] In another preferred embodiment, the training set also includes cancer clinical staging samples.

[0012] In another preferred embodiment, the preprocessing refers to aligning the sequencing data in the cfDNA whole genome sequencing dataset to a reference genome to obtain an alignment fragment file.

[0013] In another preferred embodiment, the reference genome is the human reference genome hg38.

[0014] In another preferred embodiment, the software used for the comparison includes finaleDB.

[0015] In another preferred embodiment, the preprocessing further includes inter-sample sequencing depth normalization of the cfDNA whole-genome sequencing dataset.

[0016] In another preferred embodiment, the region panel is constructed on a reference genome using an equal-length sorted union and merge algorithm according to the target footprint size.

[0017] In another preferred embodiment, the target footprint size includes: 0.5 Mb, 0.8 Mb, 1.2 Mb, 1.8 Mb, and 2.4 Mb.

[0018] In another preferred embodiment, the region panel is derived from the set of exons encoding the protein, the set of promoters anchored by the TSS, the set of CpG islands, and the set of enhancers.

[0019] In another preferred embodiment, the region panel includes single-element panels and multi-element panels.

[0020] In another preferred embodiment, the single-unit panel includes: an exon panel, a promoter panel, a CpG island panel, an ATAC panel, and a transcription factor binding site (TFBS) panel.

[0021] In another preferred embodiment, the multi-element panel includes: a Tier-2 (T1) panel, a Tier-2 (T2) panel, a Reg panel, and an exonReg (ER) panel.

[0022] In another preferred embodiment, the T1 panel is the union of the promoter panel and the CpG island panel.

[0023] In another preferred embodiment, the T2 panel is the union of the promoter panel, the CpG island panel, and the ATAC panel.

[0024] In another preferred embodiment, the Reg panel is the union of a panel, a CpG island panel, an ATAC panel, and an enhancer.

[0025] In another preferred embodiment, the ER panel is a set consisting of exon panel and Reg panel data, each comprising half of the data.

[0026] In another preferred embodiment, the single primitive panel is constructed as follows: the regions in the source annotation are sorted from largest to smallest in length and continuously accumulated until the cumulative coverage reaches the target footprint. Then, the accumulated regions are sorted by coordinate and merged to obtain the single primitive panel.

[0027] In another preferred embodiment, the multi-primitive panel is constructed as follows: the budget is evenly distributed among the contributing sub-primitives, each sub-primitive contributing its share sorted by length, and then the union is sorted by coordinates and merged to obtain the multi-primitive panel.

[0028] In another preferred embodiment, in step (S3), the filtering refers to retaining fragments whose size and position meet certain conditions, thereby obtaining the filtered fragments.

[0029] In another preferred embodiment, during the screening, the fragment size L satisfies 20 ≤ L ≤ 500 bp.

[0030] In another preferred embodiment, the position in the screening satisfies being located on chromosomes 1-22, the X chromosome, or the Y chromosome.

[0031] In another preferred embodiment, the 9-dimensional features include: normalized depth, entropy of the fragment-size distribution, proportion of small fragments, and proportion of bins of 6 sizes.

[0032] In another preferred embodiment, the small segment refers to a segment length ≤ 150bp.

[0033] In another preferred embodiment, the six bin sizes include: 0-100bp, 101-150bp, 151-200bp, 201-250bp, 251-300bp, and >300bp.

[0034] In another preferred embodiment, the evaluation method for the terminal base sequence diversity score (MDS) is as follows: for each selected fragment, based on its alignment to the reference genome, the 5′ and 3′ n-mer terminal base sequences are extracted from the reference genome sequence, and the n-mer score of each region panel is calculated. 4 The entropy value on the possible terminal base sequences; for a region panel with <10 fragments, its MDS is set to NA; where n is selected from 4, 6 or 8.

[0035] In another preferred embodiment, the region feature tensor of a region panel of a sample is represented as: sample×region×10, where 10 represents the 9-dimensional features and the terminal base sequence diversity score.

[0036] In another preferred embodiment, the biological unit dataset is derived from a public database or literature report.

[0037] In another preferred embodiment, the biological unit dataset includes: a set of transcription factor binding sites and a set of ATAC peaks for multiple cancer types.

[0038] In another preferred embodiment, when the biological unit is a set of transcription factors, step (S3.2) includes: extracting fragments from the sequencing data of each sample based on a set of transcription factor binding sites with different density levels, and screening the fragments; at one density level, for a sample, mapping each screened fragment to one or more transcription factor binding sites, evaluating the entropy value of all fragments mapped to each transcription factor binding site, and generating an entropy matrix.

[0039] In another preferred embodiment, the density grade is selected from the group consisting of 250, 500, 1000, 2000 or 5000.

[0040] In another preferred embodiment, the entropy matrix is ​​represented as N×M, where N is the number of samples and M is the number of biological units.

[0041] In another preferred embodiment, the entropy value is selected from the group consisting of: Shannon entropy, Renyi entropy, or Tsallis entropy; preferably, Shannon entropy.

[0042] In another preferred embodiment, step (S3.2) further includes the steps of: estimating the mean and standard deviation of the healthy control samples for each biological unit in the entropy matrix; and mapping each sample value to an absolute standardized deviation relative to the mean of the healthy controls.

[0043] In another preferred embodiment, step (S3) further includes step (S3.3): GC content correction of the normalized coverage of the region panel.

[0044] In another preferred embodiment, the classification algorithm includes GLMnet, random forest, XGBoost, and neural networks.

[0045] In another preferred embodiment, the classification algorithm is GLMnet.

[0046] In another preferred embodiment, in (S4.1), when the entropy matrix is ​​an entropy matrix based on transcription factor binding sites, the specific steps include: (s4.1a) Preprocess the entropy matrices of different density levels to obtain the preprocessed entropy matrix; (s4.1b) On the preprocessed entropy matrix, a 5-fold cross-validation elastic network model is used to perform binary classification modeling of health and cancer to obtain a regularized path; (s4.1c) For each density level, the corresponding read operation point λ is determined according to multiple predetermined K values; K feature factors corresponding to λ are read from each read operation point λ in the regularization path to obtain multiple candidate detection models; (s4.1d) Cross-validate the multiple candidate detection models to determine the detection model and the corresponding preferred biological unit from the multiple candidate detection models.

[0047] In another preferred embodiment, the preprocessing of the entropy matrix in step (s4.1a) includes: deleting columns with NAfraction > 0.4, filling the remaining NA items with column mean, and deleting 0-variance columns.

[0048] In another preferred embodiment, in step (s4.1c), the plurality of predetermined K values ​​are selected from the following group: (r1) The minimum K value corresponding to when the detection model performance value (AUROC) is not lower than T × the optimal performance value (max_AUROC); where T satisfies 0 < T ≤ 1; (r2) Specifies the target K value; or (r3) Specifies the total genome coverage length of the target panel and the K value derived from the target footprint.

[0049] In another preferred example, T = 0.95.

[0050] In another preferred embodiment, in (S4.2), when the entropy matrix is ​​an entropy matrix based on transcription factor binding sites, it specifically includes: (s4.2a) Preprocess the entropy matrices of different density levels to obtain the preprocessed entropy matrix; (s4.2b) On the preprocessed entropy matrix, a resilient network model with 5-fold cross-validation is used to perform binary classification modeling for each cancer type and other cancer / disease types; (s4.2c) For each λ value on the regularized path in the elastic network model; under the condition that λ is the minimum value, record the top K biological units in each cancer type, thereby determining the unified subtyping model and the corresponding preferred biological units; K is an integer from 3 to 30.

[0051] In another preferred embodiment, in (S4.3), the features include: the entropy of the transcription factor and its absolute normalization bias, the coverage of each region, the Shannon entropy, the proportion of bins of the six sizes, and the diversity of terminal base sequences on the transcription factor, ATAC, exon, promoter, CpG island and enhancer regions.

[0052] In another preferred embodiment, step (S4.3) specifically includes the following steps: (S4.3a) For each feature, the elastic network model is used to perform binary classification modeling and calibration of each cancer type and other cancer types to obtain the calibration model of each cancer under each feature, and to construct several random panel models corresponding to each feature. (S4.3b) For a type of cancer, the calibration model for the cancer under each feature and several random panel models corresponding to the feature are validated on the validation set; the probability that the performance value of the calibration model for the cancer under each feature exceeds the performance value of the random panel model corresponding to the feature is evaluated, and the feature with the larger probability value is used as the design attribution feature for the cancer. (S4.3c) Under the design attribution features, for a cancer, the mean of multiple L1 regularized fits is taken as the heterogeneity subtyping model of the cancer; all biological units corresponding to the heterogeneity subtyping models of all cancers are merged as the preferred biological unit.

[0053] In another preferred embodiment, the design attribution features of the breast cancer, colorectal cancer, and pancreatic cancer are the terminal base sequence diversity of the transcription factors.

[0054] In another preferred example, the design attribution features for lung cancer and ovarian cancer are the diversity of exon terminal base sequences.

[0055] In another preferred embodiment, step (S5) specifically includes: mapping the preferred biological unit to its corresponding genomic site, performing upstream and downstream sequence expansion (W bp) on each genomic site, and merging and sorting all genomic sites to obtain a panel.

[0056] In another preferred example, W = 100 bp.

[0057] In another preferred embodiment, the panel includes: a Brittle panel and a Robust panel obtained by (S4.1), and a Multinomial panel obtained by (S4.2).

[0058] In another preferred embodiment, the Brittle panel includes: APOBEC3B, ARID4B, CBX1, CHD1, DAXX, DMC1, EBP, ELF3, GATA2, GFI1B, HBP1, HDGFL3, HES2, HNF1A, HNRNPK, HOXB13, HOXC5, JARID2, KDM6B, MBL2, MEN1, MITF, MYNN, MYOCD, NONO, NSD2, OVOL2, PATZ1, PCBP2, PCGF2, PRDM11, PRDM12, RARA, RBBP5, RERE, RNF2, TBP, YAP1, ZBTB17, ZBTB33, ZMIZ1, ZMYM3, ZNF366, ZNF394, ZNF426, ZNF518A, ZNF555, ZNF597, ZNF781, and p65.

[0059] In another preferred embodiment, the robust panel includes: BRD7, CEBPB, CENPA, CTCF, KMT2B, MYCN, SMARCB1, SMARCC1, ZNF426, and ZNF697.

[0060] In another preferred embodiment, the Multinomial panel includes: AHR, BAP1, BATF3, BCL11A, BCL6, BCLAF1, DMC1, E2F3, E2F8, EHF, ERF, ETS1, EWSR1, FOXA3, FOXO3, FOXP2, GLIS3, H2AFZ, HES2, HEYL, HHEX, HMBOX1, HOXA2, KLF12, KLF4, KLF6, KMT2D, MBL2, MEF2D, MNT, MUC22, MYCN, NFKBIZ, NFYA, NIPBL, NKX3-1, PADI2, PATZ1, PCBP1, P CF11, PHF6, PIAS1, PPARD, PROX1, PTEN, RAD51, RARG, RBM14, RBM39, RBPJ, RELB, SAP130, SFPQ, SKIL, SP4, SS18, SSRP1, STAT2, TAF15, TCF 7L1, TP53BP1, TP73, TSC22D4, ZBTB25, ZFP69B, ZHX1, ZIM3, ZNF134, ZNF280A, ZNF318, ZNF384, ZNF426, ZNF48, ZNF518A, ZNF623 and ZSCAN16.

[0061] In another preferred embodiment, step (S6.1) specifically includes the following steps: (s6.1a) Based on the chromosome length file of the reference genome, generate several random panels on chromosomes 1-22, X chromosome, or Y chromosome with the same coverage length as the panel obtained in step (S5); iteratively adjust the number of regions in each random panel so that its total coverage length converges to the target value of coverage length ±100 KB. (s6.1b) Extract the region feature tensor from each of the random panels and the panels obtained in step (S5) according to step (S3.1); (s6.1c) Model each of the random panels and the region feature tensors of the panels obtained in step (S5) using LASSO multinomial logistic regression to obtain the performance value (AUROC) of each classification model. (s6.1d) Collect the AUROC values ​​of all random panels for each cancer type and form a random distribution; (s6.1e) The random distribution is taken as the zero distribution, and the AUROC value of the panel obtained in step (S5) is used to evaluate the performance advantage index of the panel. The confidence of the performance value of the panel is determined by resampling, thereby verifying the significance of the panel.

[0062] In another preferred embodiment, step (S6.2) specifically includes the following steps: (s6.2a) Extract several subsets from the biological unit dataset that contain one or more biological units; (s6.2b) Extract the entropy matrix of each subset from the entropy matrix obtained in step (3.2); (s6.2c) The entropy matrix of each subset is modeled using LASSO binomial logistic regression to obtain the performance value (AUROC) of each classification model. (s6.2d) Collect the AUROC values ​​of all subsets for each cancer type and form a random distribution; (s6.2e) Using the random distribution as the zero distribution, evaluate the panel performance advantage index, and determine the confidence level of the panel performance value by resampling, thereby verifying the significance of the panel.

[0063] In another preferred embodiment, the performance advantage metrics include empirical P-value (POS) and raw margin.

[0064] In another preferred embodiment, step (s6.1e) or step (s6.2e) specifically includes: (x1) Compare the panel's performance value with the random distribution to obtain the POS; (x2) Evaluate the relative advantage based on the median performance value of a panel and the mean performance value of a random panel; (x3) Resample the mean of the random distribution using bootstrap to obtain a 95% confidence interval for the random mean distribution; determine the confidence level of the panel performance value based on the 95% confidence interval.

[0065] In another preferred embodiment, the POS is the proportion by which the AURCO of a panel is greater than the AURCO of a random panel.

[0066] In another preferred embodiment, the relative advantage raw margin is evaluated as follows: raw margin = designed panel median AUROC - random panel mean AUROC, Wherein, designed panel median AUROC is the median AURCO of the panel, and random panelmean AUROC is the mean AURCO of the random panel.

[0067] In another preferred embodiment, step (S6.3) specifically includes the following steps: (s6.3a) The autosomal genome is divided according to a fixed sequence length to obtain multiple sequence regions; (s6.3b) Extract features from the multiple sequence intervals according to step (S3.1) to obtain the whole genome (WGS)-bin feature matrix; (s6.3c) The panel is trained based on the WGS-bin feature matrix and the same nested cross-validation as the panel obtained in step (S5) is performed. The resulting performance value is used as the whole genome reference value. The performance value of the panel obtained in step (S5) is compared with the whole genome reference value to validate the panel obtained in step (S5).

[0068] In another preferred embodiment, before step (S6.3c), the step of correcting the GC content of the WGS-bin count is also included.

[0069] In another preferred embodiment, the nested cross-validation is a 5-seed × 10-fold nested cross-validation.

[0070] In another preferred embodiment, step (S6.4) specifically includes the following steps: (s6.4a) Use multinomial logistic regression to model the preferred biological units obtained in step (S4) to obtain a classification / detection model; (s6.4b) In the regional feature space or the entropy original feature space, the classification model obtained in the classification / detection model step (S4) is repeatedly used to predict the validation set, and the ability of the classification / detection model to distinguish between healthy controls and multiple cancer types, and between multiple cancer types and inflammation is evaluated, thereby performing significance validation on the panel.

[0071] In another preferred embodiment, step (S6.4) further includes the following step: (s6.4c) In the region-level feature space, the proportion of fragment coverage reorganization of each region in the cross-queue is statistically analyzed.

[0072] In another preferred embodiment, the method further includes the step of: (S7) Evaluate the detection limit of the panel in a series of dilutions of circulating tumor DNA.

[0073] In another preferred embodiment, step (S7) specifically includes the following steps: (S7.1) Generate a synthetic sample; (S7.2) Extract features from the synthesized sample to obtain a feature matrix; (S7.3) Train and evaluate the feature matrix to obtain the detection limit.

[0074] A second aspect of the present invention provides a system for clinical reasoning using a panel, the system comprising: (M1) Input module, the input module being configured to input data, the data including a cfDNA-based alignment fragment file; the alignment fragment file is obtained by hybridization capture and deep sequencing of a sample from the test object using a panel obtained according to the method of the first aspect of the present invention; (M2) Inference module, configured to perform the following operations: extract features from the comparison fragment file using step (S3) of the method in the first aspect of the present invention to obtain a region feature tensor and an entropy matrix; process the region feature tensor and the entropy matrix using the detection model and the classification model obtained in step (S4) of the method in the first aspect of the present invention to obtain a detection score and a classification score respectively; combine the detection score and the classification score according to a preset threshold to obtain an inference result; the classification model includes a unified classification model and a heterogeneous classification model; (M3) Output module, which is configured to output the reasoning result of the reasoning module.

[0075] In another preferred embodiment, the sample is a bodily fluid sample, preferably a plasma sample.

[0076] In another preferred embodiment, the clinical reasoning includes: i) determining whether the subject has cancer; ii) determining the type of cancer in the subject; and / or iii) determining the stage of cancer progression in the subject.

[0077] A third aspect of the present invention provides an electronic device including a processor and a memory, the memory having a plurality of executable instructions, the processor being configured to read the instructions and perform the steps of the method as described in the first aspect of the present invention: A fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when read and executed by a processor, implement the steps of the method described in the first aspect of the present invention: A fifth aspect of the present invention provides a computer program product comprising computer-executable instructions that, when executed by a processor, implement the steps of the method as described in the first aspect of the present invention: A sixth aspect of the invention provides the use of a panel and its detection reagents obtained by the method described in the first aspect of the invention for preparing diagnostic reagents or kits for: i) diagnosing / predicting cancer; ii) classifying cancer types; and / or iii) diagnosing / predicting cancer staging.

[0078] In another preferred embodiment, the panel includes: a Brittle panel, a Robust panel, and a Multinomial panel.

[0079] In another preferred embodiment, the Brittle panel includes: APOBEC3B, ARID4B, CBX1, CHD1, DAXX, DMC1, EBP, ELF3, GATA2, GFI1B, HBP1, HDGFL3, HES2, HNF1A, HNRNPK, HOXB13, HOXC5, JARID2, KDM6B, MBL2, MEN1, MITF, MYNN, MYOCD, NONO, NSD2, OVOL2, PATZ1, PCBP2, PCGF2, PRDM11, PRDM12, RARA, RBBP5, RERE, RNF2, TBP, YAP1, ZBTB17, ZBTB33, ZMIZ1, ZMYM3, ZNF366, ZNF394, ZNF426, ZNF518A, ZNF555, ZNF597, ZNF781, and p65.

[0080] In another preferred embodiment, the robust panel includes: BRD7, CEBPB, CENPA, CTCF, KMT2B, MYCN, SMARCB1, SMARCC1, ZNF426, and ZNF697.

[0081] In another preferred embodiment, the Multinomial panel includes: AHR, BAP1, BATF3, BCL11A, BCL6, BCLAF1, DMC1, E2F3, E2F8, EHF, ERF, ETS1, EWSR1, FOXA3, FOXO3, FOXP2, GLIS3, H2AFZ, HES2, HEYL, HHEX, HMBOX1, HOXA2, KLF12, KLF4, KLF6, KMT2D, MBL2, MEF2D, MNT, MUC22, MYCN, NFKBIZ, NFYA, NIPBL, NKX3-1, PADI2, PATZ1, PCBP1, P CF11, PHF6, PIAS1, PPARD, PROX1, PTEN, RAD51, RARG, RBM14, RBM39, RBPJ, RELB, SAP130, SFPQ, SKIL, SP4, SS18, SSRP1, STAT2, TAF15, TCF 7L1, TP53BP1, TP73, TSC22D4, ZBTB25, ZFP69B, ZHX1, ZIM3, ZNF134, ZNF280A, ZNF318, ZNF384, ZNF426, ZNF48, ZNF518A, ZNF623 and ZSCAN16.

[0082] In another preferred embodiment, the detection reagent is selected from the group consisting of antibodies, primers or primer pairs, probes, sequencing libraries, nucleic acid chips, protein chips, or combinations thereof.

[0083] In another preferred embodiment, the nucleic acid chip includes a substrate and specific oligonucleotide probes spotted on the substrate, the specific oligonucleotide probes including probes that specifically bind to polynucleotides (mRNA or cDNA) of the panel.

[0084] In another preferred embodiment, the protein chip includes a substrate and specific antibodies spotted on the substrate, the specific antibodies including specific antibodies against the panel.

[0085] In another preferred embodiment, the antibody is selected from the group consisting of monoclonal antibodies, polyclonal antibodies, or combinations thereof.

[0086] In another preferred embodiment, the detection is an in vitro sample detection.

[0087] In another preferred embodiment, the ex vivo sample is selected from the group consisting of: plasma samples, blood samples, serum samples, tissue samples, body fluid samples, cell samples, or combinations thereof.

[0088] In another preferred embodiment, the tissue sample is a cancer tissue sample.

[0089] In another preferred embodiment, the detection reagent is coupled with or carries a detectable marker.

[0090] In another preferred embodiment, the detectable marker is selected from the group consisting of chromophores, chemiluminescent groups, fluorophores, isotopes, enzymes, or combinations thereof.

[0091] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0092] Figure 1 The density-selectivity design space is shown. AUROC curves for detection are displayed on the grid combining site density and selectivity configuration. Rows represent site density for each factor, and columns represent... ℓ 1 The number of factors retained along the path forms 50 candidate cells. Both heatmaps use the same color scale. In each heatmap, robust cells are outlined in black, brittle cells in red, and "best" cells (those with AUROC above 0.80 both in-cohort and cross-cohort) are highlighted in green. Brittle cells are in-cohort optimal solutions but cannot be migrated. (A) Cristiano cohort facet, with higher values ​​in high-K cells, peaking at the brittle cell with 250 sites. (B) Snyder cross-cohort facet, peaking at the robust cell with 10 factors and 2,000 sites. (C) Comparison of Snyder's AUROC with the estimated capture footprint on the sweet-spot cell, colored by site density. The most compact sweet-spot cell is the robust panel.

[0093] Figure 2The following is an example of (A) intra- and inter-cohort comparison planes. Each point represents one of the five identical Snyder cancers, corresponding to one of its 15 feature matrices, and shows its Cristiano and Snyder “one-vs-rest” AUROC values. The area of ​​the point represents the footprint of the functional annotation region, the opacity represents the probability of superiority relative to a subset of 100 random identical functional annotation regions (solid indicates attributable to design, light indicates diffuse), and the black ring represents the deployed functional annotation region. The dashed line is y=x. Breast cancer, colorectal cancer, and pancreatic cancer are deployed on transcription factor regions, while lung cancer and ovarian cancer are deployed on exon regions. Within a joint region of 1.61 Mb, the design attribution probability for each cancer ranges from 0.92 to 0.99. (B) Snyder top-2 accuracy for the five identical cancers: random baseline, 12.65 Mb forced multivariate panel, and 1.61 Mb heterogeneous panel. (C) By selectively scanning, the Snyder top-2 accuracy of the multivariate panel is forced to have a numerical upper limit close to that of the labeled cell under each "site / factor" combination. The function annotation area pairs the region set with the feature type.

[0094] Figure 3The following shows (A) Cristiano (horizontal) and Snyder “whole vs. healthy” (vertical) detection planes. Blue crosses mark cells corresponding to entropy density and selectivity. Colored dots represent nine region panels (exons, Pro, CpG, ATAC, TFBS, Tier 1, Tier 2, Reg, ER) covering the 0.5–2.40 Mb range. Two solid markers represent the Robust entropy panel selected by Snyder and the Brittle entropy panel selected by Cristiano, respectively. Gray and light brown 95% ellipses represent panels with 100 random transcription factors and random region panels matched to the footprint, respectively. Dashed lines mark 0.80. Region panels overlap with light brown ellipses, while Robust panels are positioned above gray ellipses. (B) Snyder AUROC curves for the nine region panels (same color as Figure A) within the 0.5–2.40 Mb capture footprint range. The light brown band represents the 5-95% range of 100 random region panels that matched the footprint at each footprint, with its upper edge marked by a dashed line. These region panels did not exceed this matching range on any footprint, therefore their transfer performance did not show a consistent improvement over the random panels. (C) Superiority probabilities of the Cristiano and Snyder metrics on 100 random panels matching the captured footprints for 9 region panels and 2 entropy panels. The dashed line marks 0.95. Only the Robust panel exceeded this threshold on the Snyder axis.

[0095] Figure 4Cross-cohort validation, stage, and feature granularity decomposition are shown. Dashed lines indicate the level of randomness. (A) Snyder detection AUROC curves for 11 detection panels, Robust and Brittle entropy panels, and 9 2.40 Mb region panels. Evaluation metrics included overall, comparisons of cancer of the same and different types with healthy controls, and comparisons of cancer with controls with inflammation. (B) Specificity of entropy versus region readout at the same Robust and Brittle sites. Entropy scoring misclassified systemic lupus erythematosus (SLE) as cancer, with results far below the level of randomness. Region readouts restored this distinction, with the Brittle panel showing the most significant difference, distinguishing cancer from controls with inflammation more than 93 out of 100 footprint-matched random panels (random mean 0.70). In contrast, only SLE was distinguished by almost any matching panel. Inflammatory bowel disease (IBD) was close to the level of randomness. (C) Robust detection AUROCs by Cristiano cancer staging (sample sizes for stages I-IV: n = 44, 110, 35, and 37, respectively; control group: 246 healthy subjects). (D) Detection AUROCs for each region (2.40 Mb TFBS panel) and each factor granularity analysis in each cohort (coverage, fragment size entropy, proportion of short fragments, six-level size distribution, and terminal base sequence diversity). In the Cristiano cohort, these metrics are clustered by granularity. In the Snyder dataset, coverage calculated by region is transferable, while entropy calculated by region is not; when calculated by factor, size distribution metrics are transferable, while coverage collapses.

[0096] Figure 5A cohort study on the liver demonstrates how entropy detectors can be disrupted. (A) AUROC curves of the Robust panel are detected in the liver Jiang cohort based on signed entropy features and sign-invariant absolute bias features. Cancerous samples are compared with healthy, hepatitis, and cirrhosis samples. The bar chart shows the 95% DeLong confidence interval, with the dashed line representing the random level. Detection results show a reversal trend when using signed features, while using sign-invariant features restores detection capabilities above the random level. (B) Distribution of Robust binding site fragment lengths by group, expressed as the proportion of each length. An excess of long fragments is observed in the liver cohort, but not in the pan-cancer cohort. (C) Per-sample entropy bias relative to the cohort's healthy baseline, calculated as the Robust panel factor average, in standard deviation. The value for the healthy group is zero. In the pan-cancer cohort, cancer values ​​are lower than healthy values; while in the liver disease gradient, cancer values ​​rise above healthy values—this sign reversal causes the signed detector to produce a reversal.

[0097] Figure 6 The detection and genotyping panels primarily target cancer-related regulatory factors. (A) Functional categories of the 10 robust detection factors, each corresponding to a manually selected category. Eight of these belong to known chromatin regulation or oncogenic categories, while ZNF426 and ZNF697 are uncharacterized zinc finger proteins. (B) Gene set enrichment analysis of the "Seed Stability Detection Core," using a hypergeometric distribution model compared to a candidate factor set containing 808 factors. (C) Transcription factors overexpressed beyond binding site abundance in selected regions of each TFBS-based genotyping detector; known cancer-related factors are marked in red. (D) Overexpression of cancer marker gene sets in two exon-based genotyping detectors. (E) Functional composition of the 29-factor Seed Stability Detection Core, categorized by type. In (B), (C), and (D), the horizontal axis represents -log10 p, and the dashed line indicates p = 0.05. Detailed Implementation

[0098] Through extensive and in-depth research, the inventors have developed for the first time a method for designing panels for tumor diagnosis using signal mining based on cfDNA whole-genome sequencing data. This method utilizes a cfDNA WGS cohort containing various cancer samples, inflammation samples, and healthy controls. Based on two functional annotation regions—regional fragment omics features and biological units—features are extracted from the cfDNA WGS data. L1 regularization-based sparse modeling and screening are then performed based on these extracted features to obtain preferred panel regions and preferred biological units, thereby constructing the panel. Furthermore, the method includes steps such as significance verification of the output results of the aforementioned steps, evaluation of detection limits, and downstream deployment inference using the constructed panel. Based on these steps, the present invention is complete.

[0099] It should be understood that the specific methods and experimental conditions of the invention described below in varying degrees of detail are intended to provide a substantive understanding of the invention. Definitions of certain terms used in this specification are provided below. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0100] the term As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed-ended. In other words, the terms also include “consistently made of” or “made of”.

[0101] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0102] As used in this article, the term "significant" means that, in a hypothesis test, the observed effect (such as the difference between the experimental and control groups) is unlikely to be caused solely by random error. A hypothesis test includes: the null hypothesis (H0), which assumes that the observed effect does not exist (such as no difference between the experimental and control groups); the p-value, which is the probability of observing the current or more extreme effect when H0 is true; and the significance threshold (α). The significance threshold is typically used to determine whether a hypothesis test is significant. Generally, the significance threshold is 0.05. If the p-value ≤ α, then H0 is rejected, meaning the observed effect exists, and the result is called "significant."

[0103] As used herein, the term "functional annotation region" refers to different sources of features. These functional annotation regions can be regional panels or commonly used biological units such as transcription factor binding sites, exons, promoters, CpG islands, enhancers, and ATACs.

[0104] As used herein, the term "one-vs-rest" in this invention refers to training a single binary classifier for each of the C categories of cancer.

[0105] As used in this article, the terms “footprint,” “coverage length,” and “footprint” are used interchangeably and refer to the total genome length (Mb) of all captured regions after panel merging.

[0106] As used in this article, the terms “transfer” and “transfer” are used interchangeably and refer to the generalization ability of a model / metric set, that is, the ability of the patterns and rules learned by the model / metric set on the training set to be effectively applied to an unseen validation set and make accurate predictions.

[0107] As used herein, the terms "density level" and "density" are used interchangeably, referring to the selection of the first N binding sites for each transcription factor, and in this invention, only applicable to functional annotation regions based on transcription factors. The term "density level scan" refers to performing the same processing procedure on each density level, ultimately generating an entropy primitive for each density level.

[0108] As used herein, the term "DELFI protocol" is a cfDNA fragmentomics cancer detection method based on whole-genome shallow sequencing (WGS): the genome is divided into fixed 5 Mb bins, and the short / long fragment coverage ratio (fragmentation profile) is calculated within each bin. These fragment length ratio features across the entire genome are used to train a machine learning model, enabling the detection of various cancers without somatic mutations. The "whole-genome reference" of this invention is a replication of the 5 Mb bin protocol (631 bins) of DELFI as a baseline; these bins are referred to as WGS-bins.

[0109] The computer system is equipped with at least one processor and a memory. The processor invokes a sequence of computer-executable instructions stored in the memory to implement the evaluation process defined in the claims. Although the flowchart describes the operation steps in a specific logical order, in actual execution, the steps may be processed in parallel, their order adjusted, or partially omitted in some cases. As long as such adjustments do not deviate from the core features of the technical solution described in the claims and achieve the same technical effect, they all fall within the scope of protection of this invention. This flexibility in execution order is determined by the programmable nature of computer instructions.

[0110] panel, its design methods, and applications As used herein, the term "panel" refers to a pre-designed set of targeted sequencing regions consisting of multiple genomic regions, including exons, transcription factor binding sites, and open chromatin regions. In this invention, transcription factor binding sites are preferred. In cfDNA liquid assays, panels utilize hybridization capture or amplification techniques to specifically enrich these regions across the entire genome for high-depth sequencing, thereby improving detection sensitivity, reducing costs, and focusing on key disease-related signals.

[0111] Specifically, the present invention provides a method for designing panels based on cfDNA whole-genome sequencing, the method comprising the following steps: (S1) Provide a cfDNA whole genome sequencing dataset; the cfDNA whole genome sequencing dataset includes a training set and a validation set; the training set or validation set includes healthy control samples, samples of various cancer types and other disease samples; (S2) Preprocess the cfDNA whole genome sequencing dataset to obtain alignment fragment files; (S3) Based on different functional annotation regions, extract features from the sequencing data of the samples in the training set, including: (S3.1) Extracting features from the region panels includes: extracting fragments from the sequencing data of each sample based on the alignment fragment file, and screening the fragments; for a sample, mapping each screened fragment to one or more region panels that overlap with it; evaluating the 9-dimensional features and terminal base sequence diversity score of each region panel, and generating a region feature tensor for the sample based on the 9-dimensional features and terminal base sequence diversity score; and / or (S3.2) Extracting features from biological units includes: extracting fragments from the sequencing data of each sample based on the biological unit dataset, and filtering the fragments; for a sample, mapping each filtered fragment to one or more biological units, evaluating the entropy value of all fragments mapped to each biological unit, and generating an entropy matrix; (S4) Based on L1 regularization, sparse regions are filtered using steps selected from the following group: (S4.1) Based on the entropy matrix, a classification algorithm is used to perform binary classification modeling for healthy and cancerous individuals, obtaining the detection model and the corresponding preferred biological units; and / or (S4.2) Based on the entropy matrix, a classification algorithm is used to perform binary classification modeling of each cancer type with other cancer / disease types to obtain a unified classification model and the corresponding preferred biological unit; and / or (S4.3) Based on the different features obtained in step (S3), a classification algorithm is used to perform binary classification modeling for each cancer type and other cancer types to obtain a binary classification model corresponding to each feature; all the binary classification models are validated in the validation set to obtain a heterogeneity subtyping model for each cancer type and a preferred biological unit. (S5) Map the preferred biological units to their corresponding genomic sites from the preferred biological units, perform upstream and downstream sequence expansion on each genomic site, and merge and sort all genomic sites to obtain a panel.

[0112] The panel obtained using the above method can be further used for cancer prediction / diagnosis, cancer classification, or cancer staging. Specifically, a detection reagent is designed based on the panel. This reagent is used to hybridize and enrich plasma samples from the test subject, followed by deep targeted sequencing to obtain an alignment fragment file. Based on the alignment fragment file, step (S3) of the above method is performed to obtain a region feature tensor and an entropy matrix. The detection model and classification model obtained in step (S4) of the above method are then used to process the region feature tensor and the entropy matrix to obtain a detection score and a classification score, respectively. Finally, the detection score and the classification score are combined according to a preset threshold, and the combined result is used to predict / diagnose the test subject.

[0113] Using the method of this invention, three transcription factor-based panels were designed: a Brittle panel, a Robust panel, and a Multinomial panel. Among them, the Robust panel achieved the best transferability on the external validation set. Furthermore, the Robust panel can also achieve accurate staging prediction / diagnosis of cancer progression.

[0114] The system of the present invention This invention provides a system for clinical reasoning using a panel. The system can be embodied as an electronic device, including but not limited to: smartphones, tablets, personal computers, servers, terminals, or other intelligent terminals with data processing capabilities. The system can also be embodied as a computer-readable storage medium, such as a disk, optical disk, solid-state drive, read-only memory, or flash memory, on which a computer program is stored, which, when executed by one or more processors, enables the data processing flow. Furthermore, the system can also be embodied as a computer program product containing computer instructions that, when executed by a computer, cause the computer to perform all or part of the steps of the data processing flow.

[0115] The data processing flow can be loaded, deployed, or run on any of the aforementioned electronic devices. Through the collaboration of hardware resources and the software logic defined in the flow, the system can complete specific data acquisition, transmission, calculation, analysis, storage, or presentation tasks.

[0116] The main advantages of this invention include: (1) For the first time, cfDNA WGS data was used as prior information for panel region selection, thereby unifying the independent information mining and panel design processes into a single reproducible method.

[0117] (2) When screening panels based on biological priors, the present invention adopts a screening strategy based on L1 regularized regression, thereby obtaining more candidate panels.

[0118] (3) The candidate panels obtained by this invention are all obtained by sparsely solving the publicly published whole genome TF regulatory sites or ATAC peak diagrams and other functional annotation regions using pure data-driven L1 regularization. None of them use any cancer biology prior input, which can avoid the bias of prior knowledge, facilitate the discovery of new signals, and have good universality and transferability.

[0119] (4) The Robust model based on transcription factors obtained in this invention can not only classify a variety of cancers well, but also classify cancer stages well.

[0120] (5) When using the panel of the present invention to detect samples such as liver cancer whose fragmentation direction is opposite to that of most cancers, the present invention introduces a sign-invariant feature and scores the absolute deviation relative to the healthy baseline, so that the detection panel can maintain its ability to distinguish between the two opposite directions.

[0121] (6) The panel obtained by this invention can perform two independent clinical tasks, namely cancer / health detection and cancer / chronic validation, after a single wet experiment capture and sequencing, thereby reducing the time cost of clinical diagnosis.

[0122] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It is also understood that the purpose of describing the present invention in conjunction with the embodiments is to cover other options or modifications that may be derived based on the claims of the present invention. To provide a deep understanding of the invention, many specific details will be included in the following description. The invention may also be practiced without using these details. Furthermore, to avoid confusion or obscuring the focus of the invention, some specific details will be omitted in the description.

[0123] Processes and Methods 1. Process: (A) Input data acquisition and preprocessing: Data source: cfDNA WGS cohort.

[0124] File format: bgzip compressed file containing 5 columns of TSV (chromosome, start, end, identifier, fragmentlength) and a tabix index.

[0125] Phenotypic label: Each sample carries a cancer_type label, which is one of the following: bile_duct_cancer / breast_cancer / colorectal_cancer / gastric_cancer / lung_cancer / ovarian_cancer / pancreatic_cancer / healthy.

[0126] Specifically, the training set uses the Cristiano 2019 cohort (N=475, covering 7 cancer types + healthy controls), and hg38-aligned fragment files are obtained via finaleDB.

[0127] Cross-cohort validation set: Snyder 2016 cohort (N=60; 4 healthy controls, 4 inflammatory bowel disease (IBD) patients, 4 systemic lupus erythematosus (SLE) patients, and 48 cancer samples across 14 cancer types, 5 of which are the same as the Cristiano cohort), with hg38 fragments obtained using the same procedure. Samples from this cohort were not used for training.

[0128] Equal-depth normalization is not mandatory between samples.

[0129] (B) Construction of the region panel: Using an equal-length sorted union and merging algorithm, nine region panels matching the capture range were constructed with a footprint size of 2.4 Mb. These included single-primitive panels (i.e., exons, promoters, CpG islands, ATAC, and transcription factor binding sites (TFBS)) and multi-primitive panels (i.e., Tier-2 (T1), Tier-2 (T2), Reg, and exonReg (ER)). T1 is the union of promoters and CpG islands; T2 is the union of T1 and ATAC data; Reg is the union of T2 and enhancers; and ER is the set consisting of equal portions of exon and Reg data. For single-primitive panels, regions in the source annotations were sorted by length from largest to smallest and accumulated until the cumulative coverage reached 2.40 Mb. The accumulated regions were then sorted by coordinates and merged. For multi-primitive panels, the budget was evenly distributed among the contributing sub-primitives, with each sub-primitive contributing its share sorted by length. The unions were then sorted by coordinates and merged. The final coverage of all nine panels fell between 2.26 and 2.40 Mb. The target value of 2.40 Mb is comparable to the gene panel footprint used in previous studies that validated panel fragmentomics.

[0130] The same nine panel builds were repeated at target footprints of 0.5, 0.8, 1.2, and 1.8 Mb to track in-team accuracy and cross-team transfer dependency on footprint.

[0131] (C) Parallel feature extraction of dual-function annotation regions: (C.1) Region-level fragment omics features: for each candidate panel × each sample: (C.1.1) Extract the fragment using tabix -R panel.bed sample.bgz; (C.1.2) Retain segments with an insert size ∈ [20, 500] bp that are located at chr1-22 / chrX / chrY; (C.1.3) Use bedtools intersect -wa -wb to assign each fragment to its overlapping panel region; (C.1.4) Calculate 9-dimensional features for each region: (i) Normalized depth, normalized depth = number of segments × 10 9 / Region length / Total number of fragments; (ii) Shannon entropy of fragment-size distribution; (iii) Proportion of small fragments (fragments ≤ 150 bp); (iv-ix) Proportion of the 6 size bins (0-100, 101-150, 151-200, 201-250, 251-300, > 300 bp); (C.1.5) Terminal sequence diversity score (MDS): For each fragment, the 5′ and 3′ 4-mer terminal base sequences were extracted from the hg38 reference genome, and the Shannon entropy of each region on 256 possible motifs was calculated; regions with a fragment size <10 were labeled MDS = NA. (C.1.6) Output sample×region×10 tensors (9 region-level features + MDS) (C.2) Entropy primitive of biological units: (C.2.1) Functional annotation area: 808 transcription factor (TF) sets from the GTRD v19.10 Homo_sapiens_meta_clusters database; each TF takes the top 250, 500, 1,000, 2,000 and 5,000 binding sites supported by the experiment. (C.2.2) For each sample × each TF: extract fragments using tabix -R tf_sites.bed sample.bgz and filter as per step (C.1.2); (C.2.3) Use bedtools intersect to aggregate fragments to all binding sites of the TF; (C.2.4) Calculate the single Shannon entropy value on the merged fragments; (C.2.5) Output an N×808 entropy matrix for each density level (tfbs_entropy_matrix.tsv); The TCGA-ATAC pan-cancer peak set (a set of 23 cancer-type peaks) was processed in the same way, but without density level scanning, and an N×23 entropy matrix (atac_entropy_matrix.tsv) was output.

[0132] (C.2.6) Entropy Matrix Feature Transformation: For samples such as liver cancer, whose fragmentation direction is opposite to that of most cancers, the entropy matrix is ​​also corrected. Specifically, within each cohort, the healthy mean and standard deviation of the entropy of each TF are estimated, and each sample value is mapped to the absolute standardized deviation relative to the healthy mean (only the feature values ​​are changed, while the panel factor and downstream elastic network protocol are retained).

[0133] In this invention, the entropy matrix features are used in two places after transformation: ① detection under the reversed fragmentation direction in liver cohort verification (Example 5); ② as one of the 15 functional annotation regions for heterogeneous typing in subsequent steps (absolute bias of transcription factors).

[0134] (C.3) GC Correction: GC correction was performed on the normalized coverage and whole-genome sequencing (WGS) bin counts for each sample to correct for GC / library confounding robustness, and cross-cohort analyses were repeated. GC content was calculated for each region (i.e., exon / promoter / CpG / ATAC / TFBS ​​single- and multi-monotonic panels) or 5 Mb fixed bins using bedtools nuc. For each sample, locally weighted regression analysis was performed with GC content as the independent variable for each region's coverage, and the residuals after regression were re-normalized centered at the sample median according to existing whole-genome correction methods. Results showed that scale-invariant entropy, binning size, and terminal base sequence characteristics were unaffected.

[0135] (D) Construction and optimization of panels based on L1 regularization: The matching panel classifier used in this step is a multinomial elastic network model trained using glmnet and tuned based on Latin hypercubes. For each panel, seed, and fold, Latin hypercube sampling is performed using `dials::grid latinhypercube`, randomly sampling ten elastic network mixing parameter values ​​α∈[0, 1]. For each α, a regularized path is fitted using glmnet::cv.glmnet with 5-fold inner cross-validation and multinomial bias as the loss function. The (α, λ) combination that minimizes the inner cross-validation bias is finally selected. Class balance on the training partition is enforced using `caret::upSample`, and the data is limited to the training sample range to prevent data leakage. The outer loop uses `caret::createFolds` to run 5 seed × 10 layered outfolds, generating 50 training-test splits per panel. The AUROC for each fold is calculated using `pROC::roc`. Top-K accuracy refers to the proportion of out-of-fold scores that rank in the top K positions of the true class after aggregation. In the training cohort, the Top-K for cancer is limited to non-healthy samples.

[0136] (D.1) Detect panel: Under five density levels, joint multi-class regularization (pan-class L1multinomial) is applied, and density-selective boundary scan is used to obtain the N×808 TFBS entropy matrix output from step (C2).

[0137] (D.1.1) Preprocessing: Delete columns with NA fraction > 0.4; fill the remaining NA columns with the column mean; delete the 0-variance column.

[0138] (D.1.2) Use cv.glmnet (α = 1, binomial distribution and 5-fold internal cross-validation) to fit a ℓ1 regularized binary “cancer vs health” classification model to an N × 808 entropy matrix.

[0139] (D.1.3) For each λ value on the regularization path, record (K, AUROC), where K is the number of non-zero coefficients in the model and AUROC is the model performance (area under the ROC curve) estimated through inner-CV. (D.1.4) The operating point selection rules are selected from the following group: i) The minimum K value corresponding to the model performance not lower than T×optimal performance (max_AUROC); preferably, T=0.95; ii) User-specified target K value; iii) User-specified target footprint, and the K value is derived from the footprint.

[0140] (D.2) Typing panel: For each cancer category at each density level, the entropy matrix of multiple biological units output from step (C2) is obtained using density-selective scanning in parallel with the detection scan.

[0141] (D.2.1) Unified Classification Panel: Using cv.glmnet (α = 1, binomial distribution, AUC optimized, 5-fold internal cross-validation), with a seed value of 42, a binary classification model is independently trained for each cancer type (one-vs-rest). Under the condition of λ min, the top K rank models for each cancer type are retained. per_class The scan covered transcription factors, including K. per_class The range is from 3 to 30. Finally, the union of the specific TF lists for each cancer type is taken.

[0142] (D.2.2) Heterogeneity typing panel: 15 feature functional annotation regions generated in step (C), including the entropy and absolute bias of each transcription factor from step (C.2), the coverage, Shannon entropy, and 6 size bin distributions of each region from step (C.1), and the terminal base sequence diversity on transcription factors, ATAC, exons, promoters, CpG islands, and enhancer regions, involving 5 cancer types shared by the training and validation cohorts.

[0143] For each functional annotation region and cancer, cv.glmnet (α=1, binomial distribution) is trained on a subset of the training cohort containing only cancer to obtain the in-cohort cross-validation AUC and the "one-vs-rest" AUC for other common cancer types (4 types).

[0144] Design attribution refers to the probability that the detector surpasses 100 random subsets of the same size within the same Snyder cohort of transfers. Each cancer was assigned a functional annotation region for design attribution with a superiority probability close to 1, indicating that the annotation region provides higher transfer efficiency over a smaller coverage area. Ultimately, transcription factor-based region readouts were determined for breast, colorectal, and pancreatic cancers, while exon-based region readouts were used for lung and ovarian cancers.

[0145] Within the selected functional annotation area, the detector for each cancer is the mean of five ℓ1 regularized fits (once per seed), while the coverage of the detection panel is the union of all fitted regions in the selected area across the five cancer types.

[0146] During the testing phase, the detector for each cancer was calibrated using z-scores based on its Cristiano training score. The five calibrated scores were ranked, and the accuracy of the top K positions was defined as the proportion of samples whose true class was in the top K positions.

[0147] (E) Panel construction (mapping biological units to genome coordinates): (E.1) Extract the sites corresponding to the density level from the functional annotation area of ​​step (C.2.1) for the list of biological units output in step (D); (E.2) Each site is padded to both sides with W bp (for the Cristiano 2019 cohort, W=100 bp); (E.3) Sort and merge the sites; Sorting command line: `bedtools sort -k1,1 -k2,2n`; Merging command line: `bedtools merge`; (E.4) Output panel BED.

[0148] For the (D.1) detection panel: at each density level, counts of 10 retained factors were read from the regularization path, these counts being derived from the steps closest to K (K being 5, 8, 10, 14, 20, 25, 30, 40, 50, and 65 respectively), resulting in 50 candidate detection panels. Among the candidate detection panels, the Brittle panel was considered the unit with the largest AUROC value within the queue, with 250 sites per factor, K = 50, covering a region of 2.42 Mb; the Robust panel was considered the combination with the best cross-queue mobility, with 2,000 sites per factor, K = 10, covering ten factors: BRD7, CEBPB, CENPA, CTCF, KMT2B, MYCN, SMARCB1, SMARCC1, ZNF426, and ZNF697, for a total coverage of 3.63 Mb.

[0149] At a density level of 5,000 sites, a selection stability analysis was performed on five fitted seeds, each of which retained 37 to 45 factors, while the 29 factors selected from all five seeds defined the stable core.

[0150] For the (D.2) uniform typing panel: at 1,000 sites for each factor and K per_class When the value is 12, the result of maximizing the unit is a 12.65Mb panel (Multinomial panel) consisting of 76 transcription factors, which is scored by a forced multinomial classifier. Figure 2 C).

[0151] For (D.3) heterogeneous genotyping panel: the capture coverage is extended by 100 bp outside each region and merged to form a genotyping coverage of 1.61 Mb, which is combined with the 3.63 Mb detection panel to total 5.23 Mb.

[0152] (F) Significance verification: (1) Significance verification of random zero comparison Two random null hypothesis comparisons quantified whether the designed panel was superior to a size-matched randomized alternative. For regional panels within a 2.40 Mb matching coverage area, the aforementioned 100 randomized panels were scored using the same classification protocol as the 9 designed regional panels. The superiority probability is the empirical proportion in which the mean AUROC of the 100 randomized panels is significantly lower than the mean AUROC of the designed panels in 5-seed × 10-fold nested cross-validation. The original AUROC gap is the difference between the mean of the designed panels and the mean of the randomized null values. Bootstrap 95% confidence intervals were calculated based on 2,000 paired resamples from a randomized mean distribution.

[0153] For the transcription factor entropy panel, the corresponding zero values ​​were constructed from 100 uniformly randomly generated K×808 entropy column subsets, where K matches the panel size. For the Brittle panel, the zero values ​​were calculated when K=50; for the Robust panel, they were calculated when K=10. The i-th random subset was used to extract column indices using a random seed of 42+i, thus ensuring a completely deterministic distribution of the zero values. The same classifier protocol was used to design both the panel and the random panel.

[0154] Both null hypotheses were extended to the cross-cohort dimension. 100 randomly matched footprint region panels with 0.5 Mb footprints were extracted from both cohorts and scored according to the cross-cohort protocol described below. Training was performed on the Cristiano dataset and testing on the Snyder dataset to derive the cross-cohort superiority probability for each design region panel. For these 100 random subsets of K-of-808 entropy, scoring was performed both intra-cohort and cross-cohort, using nested cross-validation, thus generating paired intra-cohort and cross-cohort AUROC values ​​for each random panel. The robust panel was compared to its K = 10 cross-cohort null hypothesis, and the Brittle panel was compared to its K = 50 cross-cohort null hypothesis.

[0155] (F.1) Random null distribution control in areas with equal coverage length: (F.1.1) Using the human reference genome hg38 chromosome length file, generate N_random≥100 random panels on chr1-22 / chrX / chrY with the same target coverage length as the designed panel (e.g., 2.40 Mb); iteratively adjust the number of regions to converge the total coverage to the target ±100 KB; Command line: bedtools random -l 8000 -g hg38.chrom.sizes (F.1.2) Extract the region-level feature matrix for each random panel and designed panel according to the process in step (C.1); (F.1.3) A multinomial logistic regression model with an elastic network (α=0.5) is adopted. The generalization performance of the region-level feature matrix of each random panel and the designed panel is robustly evaluated by 10 repeated 10-fold nested cross-validation (inner 5-fold), and the bias is used as the criterion for selecting the regularization parameter. Command line: cv.glmnet (family="multinomial", alpha=0.5, 10 outer seeds ×10 outer folds × 5 inner folds, type.measure="deviance"); (F.1.4) Record the random null AUROC distribution for each cancer type; (F.1.5) Calculate the raw margin = designed panel median AUROC - random panel mean AUROC; POS = empirical P(AUROC_designed > AUROC_random_i). The designed panel median AUROC is the median AUROC of the designed panel across 10 seeds in the logistic regression model. If POS is close to 1, it indicates that the designed panel significantly outperforms the random expectation.

[0156] (F.1.6) Calculate the 95% CI (2,000 resamplings) of the random mean distribution based on bootstrap to determine whether the AUROC of the designed panel falls outside the confidence interval of the random performance.

[0157] (F.2) Random zero-distribution control within the same functional annotation area: (F.2.1) For each designed K value in the 808-TF universe, N_random=100 K-subsets are randomly selected; (F.2.2) Input the column subset corresponding to the matrix output in step (C.2); (F.2.3) A binary logistic regression model was trained using an elastic network (α=0.5) and 10-fold cross-validation, with AUC as the criterion, and the results were guaranteed to be reproducible by using a seed (42+i).

[0158] Command line: cv.glmnet(family="binomial", alpha=0.5, nfolds=10, type.measure="auc", seed = 42 + i); (F.2.4) Record designed vs random null AUROC, POS, bootstrap 95% CI.

[0159] (F.3) Cross-queue verification: (F.3.1) Make predictions for the Snyder queue using a multinomial logistic regression model (cv.glmnet(family="multinomial", alpha=0.5)) trained on the Cristiano queue; (F.3.2) Evaluate three analyses: (a) overlap cancer vs healthy; (b) non-overlap cancer vs healthy; (c) cancer vs IBD / SLE; (F.3.3) Repeat steps (F.3.1)-(F.3.2) in the following two feature spaces: region-level (features from step (C.1)) and entropy primitive (features from step (C.2)); (F.3.4) Calculate the cross-cohort survival rate for each region for regional features (the proportion of samples with sufficient coverage of the region fragments in the cross-cohort samples).

[0160] (2) Whole genome reference validation The 3,088 Mb autosomal genome was divided into 5 Mb fixed intervals, generating 631 intervals per sample. Feature extraction in step (C.1) was then performed to obtain the WGS-bin feature matrix. Based on the WGS-bin feature matrix, a matching panel classifier was trained and subjected to the same 5-seed × 10-fold nested cross-validation as the designed panel. The resulting in-pair AUROC values ​​are reported as genome-wide reference values ​​in the "Results" section.

[0161] (3) Evaluate the effectiveness of the Robust model in cancer staging tasks.

[0162] In the Cristiano cohort, the Robust model was evaluated based on cancer staging. The out-of-fold cancer probability derived from leak-free nested cross-validation for each sample was preserved, the probability of the five seed samples was averaged, and this was correlated with clinical staging. For each staging stratum, the detected AUROC value was the "cancer vs. healthy" discriminant between cancer samples within that stratum and a shared healthy control group; as internal validation, the merged staging strata reproduced the overall AUROC value of the detection panel within the cohort. Samples with stage 0 or undetermined staging were excluded from each staging stratum.

[0163] Characterization of the limit of detection (LOD) for (G) in silico ctDNA dilution: (G.1) Synthetic Sample Generation (Phase A): (G.1.1) Donor admission criteria: cancer donor admission threshold: n_fragments≥25,000,000; healthy donor admission threshold: n_fragments≥30,000,000. Cristiano queue: 215 out of 229 cancer donors meet the criteria, and 232 out of 246 healthy donors meet the criteria.

[0164] (G.1.2) 70 / 30 Training / Validation Set Splitting (executed at the donor level, before mixing): Layer by (cancer_type, disease_status), and within each layer, randomly shuffle the donor list (seed=42). The first 70% is used as the training set, and the last 30% is used as the validation set. If a single donor layer has only one donor, all donors are included in the training set.

[0165] (G.1.3) Cancer:Healthy mixing ratio (5 fixed ratios): 1:99, 5:95, 10:90, 25:75, 50:50. Corresponding cancer_fraction = 0.01, 0.05, 0.10, 0.25, 0.50.

[0166] (G.1.4) Target reads per synthetic sample: TOTAL_READS = 50,000,000 fragments.

[0167] (G.1.5) Mixed pairing generation mode (MIXING_MODE, choose one): i) Minimal: Match each cancer donor with one randomly selected healthy partner; match each healthy donor with one randomly selected healthy partner; ii) Sampled: For each cancer donor, K_HEALTHY_PER_CANCER = 8 random healthy partners are matched; for each healthy donor, K_HEALTHY_PER_HEALTHY = 4 random healthy partners are matched. iii) full: Take all cancer × healthy Cartesian products; take all healthy × healthy Cartesian products.

[0168] (G.1.6) Per-job hybrid execution: For each mixing task (cancer donor c, healthy donor h, ratio (c_pct, h_pct)): (a) Calculate target reads: target_C=TOTAL_READS×c_pct / 100; target_H=TOTAL_READS×h_pct / 100.

[0169] (b) Donor depth verification: if c.n_fragments < target_C or h.n_fragments < target_H, skip this task.

[0170] (c) Per-job seed derivation: •SID=SYN_C{job_idx:06d}_R{c_pct:02d} (cancer synth) or SYN_H{job_idx:06d}_R{c_pct:02d} (healthy synth); •JSEED = cksum(SID + master_seed) mod 2^31, where master_seed=42; •JSEED_C = JSEED; JSEED_H = (JSEED + 1) mod 2^31.

[0171] (d) Bernoulli subsampling: randomly extract a portion of fragments from the compressed fragment files of the cancer and healthy samples at a certain ratio for subsampling; •FRAC_C = target_C / c.n_fragments; FRAC_H = target_H / h.n_fragments; •For cancer donor BGZ: zcat c.bgz | awk -v p=FRAC_C -v s=JSEED_C 'BEGIN{srand(s)} rand() < p'; •For healthy donor BGZ: zcat h.bgz | awk -v p=FRAC_H -v s=JSEED_H 'BEGIN{srand(s)} rand() < p'.

[0172] (e) Merging, sorting, compression, indexing: • Concatenate (d) two outputs → sort -k1,1 -k2,2n -S 25% --parallel=2 → bgzip> {OUT}.tmp → mv {OUT}.tmp {OUT} → tabix -p bed {OUT}; • Output file path: {OUT_DIR} / {split} / {cancer_type} / {SID}.hg38.frag.tsv.bgz and .tbi.

[0173] (f) Phenotype TSV: Write to train_pheno.tsv and val_pheno.tsv respectively, with 6 columns: sample_id, phenotype (cancer | healthy), cancer_type, cancer_fraction, source_cancer, source_healthy.

[0174] (G.1.7) Healthy × healthy synthetic samples (executed when GENERATE_HEALTHY_MIXES=1): - For each pair (h_a, h_b) ∈ healthy donors (the number of partners is determined by (G1.5) MIXING_MODE) × 5 ratios, perform the per-job mixing process of (G1.6). The cancer_fraction label for this batch of samples is fixed at 0.00.

[0175] (G.1.8) Parallelism: Mixed tasks are executed in parallel using xargs -P N_THREADS (Cristiano queue uses N_THREADS=32).

[0176] (G.1.9) Idempotency: If both {OUT}.bgz and {OUT}.bgz.tbi already exist, the task is skipped.

[0177] (G.2) Synthetic Sample Feature Extraction (Phases B / C): Perform the following steps on the synthetic samples from the training and validation sets respectively: (G.2.1) Per-sample TFBS entropy (consistent with step (C.2)): (a) TFBS BED preprocessing: Expand each binding site on the 808-TF GTRD BED by 250 bp, sort -k1,1 -k2,2n, and then merge using bedtools to obtain expanded TFBS regions BED; (b) Extract fragments from each synthetic sample using Tabix-R based on the expanded BED file, filtering by insertion size ∈ [20, 500] bp + standard chromosome; (c) Calculate per-TF Shannon entropy on the original 808-TF BED; (d) Output {FEATURE_DIR} / tfbs_entropy / {SID}.tsv, with 3 columns: tf_name, n_fragments, entropy.

[0178] (G.2.2) Self-heal loop: (a) After the parallel pass is completed, count outputs the number of TSVs and the expected number of pheno TSVs; (b) Missing samples enter the sequential retry pass; (c) If the samples are still missing after the retry, the script hard-fails and exits (exit code is non-zero), without entering the downstream aggregate.

[0179] (G.2.3) Matrix aggregation: Merge all per-sample TSVs into a wide table {FEATURE_DIR} / tfbs_entropy_matrix.tsv with rows of sample_id and columns of {TF}_entropy.

[0180] (G.3) LOD Training and Assessment (Phase D) (G.3.1) Input: results / features / synth_train / tfbs_entropy_matrix.tsv; results / features / synth_val / tfbs_entropy_matrix.tsv; results / dilution / synthetic / {train,val}_pheno.tsv; TF lists: results / panels / tf / 14tf.txt, 29tf.txt, 37tf.txt.

[0181] (G.3.2) Training: Cut the column subset {TF}_entropy of the entropy matrix for each panel (14tf / 29tf / 37tf); fit the training set matrix with fit_lhs_glmnet (scripts / lib / lib_glmnet.R): ​​family="multinomial", N_LAMBDA=10 (α is sampled with Latin hypercube, α∈[0,1]), N_INNER=5 inner CV,caret::upSample training partition minority class upsampling; N_SEEDS=10 different master seeds repeated; for each seed, predict(s="lambda.min", type="response") on the validation set matrix and output an N_val×N_classprobability tensor.

[0182] (G.3.3) Per-bin AUROC Calculation: Divide the validation set samples into 6 bins according to cancer_fraction: 0_healthy_mix (cf == 0.00) / 1_50pct (cf ≥ 0.50) / 2_25pct (0.25 ≤ cf < 0.50) / 3_10pct (0.10 ≤ cf < 0.25) / 4_05pct (0.05 ≤ cf < 0.10) / 5_01pct (0 < cf < 0.05); Calculate one-vs-rest AUROC (pROC::roc, direction="<") for each class within each bin; Simultaneously calculate the binary cancer-vs-healthy AUROC within each bin: score = 1 - P[healthy], y_binary =(cancer_type != "healthy"); Skip pos < 2 or neg Combinations of (class, bin) with less than 2.

[0183] (G.3.4) Output: results / dilution / lod_entropy.tsv: raw long table, columns panel×seed×bin×task×n_pos×n_neg×auroc; results / dilution / lod_entropy_summary.tsv: median / mean / sd / IQRper(panel×bin×task), aggregated across seeds; results / dilution / lod_entropy_detection.png: cancer-vs-healthy AUROC vs cancer_fraction curves, each line represents a panel, and error bars are represented by IQR.

[0184] (H) Deployment of inference (dual primitive parallelism): (H.1) Perform hybrid-capture enrichment + deep targeted sequencing (typical depth ~1,000×) using the panel BED output from step (E). (H.2) Execute the two feature extraction processes (C.1) and (C.2) in parallel on the same BAM / fragment file; (H.3) Use the classifier obtained in step (D) to train the classifier and output the detection score and typing score respectively; (H.4) Combine two scores according to a preset threshold and output a multi-task report.

[0185] 2. Ablation experiments based on characteristic type and particle size: To distinguish fragmentomics metrics from aggregation granularity, TFBS functional annotation regions were scored at three granularity levels, following the aforementioned protocol. At the regional granularity level, metrics for each region within the 2.4 Mb TFBS panel were scored individually. These metrics included normalized coverage, fragment size Shannon entropy, six-region size distribution, proportion of short fragments, terminal base sequence diversity score, and the complete regional feature set. At the aggregation granularity level, fragments from each TFBS were aggregated to a single fragment size distribution for each sample and aggregated using a single Shannon entropy value. This aggregated value was scored using its univariate “Cancer vs. Healthy” AUROC and served as the lower limit of granularity for this ablation analysis. At the “by factor” granularity level, fragments overlapping with each transcription factor binding site were aggregated as a single value for each factor under each metric. This factor-based construction approach is consistent with the TFBS entropy metrics in previous panel fragmentomics. Regional granularity reads the 2.4 Mb panel data, while aggregation and factor-based granularity read all binding sites. Terminal base sequence diversity was calculated based on 4-mer 5′ and 3′ motifs read from a reference sequence. Region and factor level matrices were scored within the cohort using 5-seed leak-free nested cross-validation, subsequently trained on the Cristiano dataset, and applied to the Snyder dataset for cross-cohort transfer, with in-cohort z-score normalization performed in each case.

[0186] 3. Snyder cross-queue verification The detection and genotyping panels were trained on the full Cristiano cohort and applied directly to the Snyder cohort without retraining. To eliminate cross-cohort scale bias caused by independent cell-free DNA preparation processes, each entropy matrix was normalized by cohort z-scores before model fitting and prediction. Sample order within the cohort was preserved. Detection employed a binary ℓ1 regularized model trained on a subset of 10 transcription factor entropies labeled “cancer vs. healthy” from the Robust detection panel and fitted using the same Latin hypercube-tuned elastic net protocol as the in-cohort scorer. The multinomial genotyping baseline used an elastic net model based on a 76-transcription factor entropy detection panel, but limited to the cancer subset from the Cristiano cohort. The heterogeneity panel fitted a binary detector for each cancer on its designated functional annotation region and ordered after each detector calibration. Five fitting seeds characterized the model fit stability under Latin hypercube and internal cross-validation randomization conditions. The Snyder cohort itself served as an outer test partition, thus eliminating the need for nested cross-validation.

[0187] This paper reports the detection AUROC values ​​between identical cancer and healthy groups, dissimilar cancer and healthy groups, and cancer and each inflammation control group in two cohorts. Top-K accuracy is reported separately for overlapping cancer samples with true labels within the training set and dissimilar cancer samples with true labels outside the training set (therefore, the Top-K is structurally zero).

[0188] 4. Liver cohort and sign-invariant features The Robust model was directly applied to the Jiang liver cohort without retraining, and the same fragment entropy processing and Z-score normalization method used in the Snyder validation were employed. The Jiang cohort included 90 hepatocellular carcinoma (HCC) patients, 36 patients with cirrhosis, 67 hepatitis controls, and 32 healthy donors. AUROC values ​​for HCC versus healthy individuals, hepatitis patients, and patients with cirrhosis are reported, with HCC as a positive label and the controls as negative labels. The confidence intervals used were the DeLong 95% confidence intervals.

[0189] The sign-invariant features re-scored each factor by the absolute deviation of each factor from a healthy reference value. For each transcription factor, the healthy mean and standard deviation of its factor entropy are estimated within each cohort, and each sample value is mapped to the absolute standardized deviation relative to that healthy mean. The reference set is a cancer-free sample set, i.e., samples from truly healthy donors where available; for the Snyder cohort (which has a smaller number of truly healthy samples), the complete non-cancer sample set is used. This transformation preserves the panel factors and the downstream elastic-net protocol invariantly, only replacing the feature values.

[0190] The merged training analysis combined the Cristiano and Jiang cohorts into a single training set for a "cancer group and healthy group," with Z-scores standardized for each cohort before merging. Using 10-fold cross-validation, the out-of-fold AUROC values ​​for each feature were reported separately for both cohorts. Intra-cohort feature comparisons were performed only on the Cristiano cohort, fitting the same elastic network model using 10-fold cross-validation.

[0191] 5. Repeatability Each random sampling in the pipeline uses a seed value. Randomly matching footprint region panels use the panel index as the seed for the random region sampler. Random transcription factor subsets use 42 plus the panel index as the column sampling seed. The outer fold of nested cross-validation uses set.seed values ​​from 1 to 5. Latin hypercube sampling for each fit uses the product of the outer seed and the fold index. Expert model selection by category uses a fixed seed of 42 in the inner cv.glmnet fit. Snyder cross-cohort fitting uses sequential seeds from 1 to 5. Therefore, the entire process is deterministic and can be fully reproduced byte-by-byte on the same input cohort and annotation basis.

[0192] Example 1: Three Selected Panels for Transcription Factors Following the panel design process of this invention, three selected panels for transcription factors were ultimately determined: Brittle panel, Robust panel, and Multinomial panel. The transcription factors contained in the three panels are shown in Table 1.

[0193] Table 1. Complete list of transcription factor members included in the Brittle panel, Robust panel, and Multinomial panel Example 2: Site density and selectivity together define a Pareto surface with independent optima. By jointly scanning binding site density and panel selectivity on the entropy primitive, a discrete Pareto surface spanning panel size, site density, and detection AUROC was plotted. Figure 1 ).

[0194] The Cristiano AUROC peaked at 0.955 in a Brittlepanel covering 250 sites across 50 transcription factors (2.42 Mb). Cells located in the 250-site row achieved the highest Cristiano AUROC value. Cells with higher density and fewer retained factors had Cristiano AUROC values ​​between 0.85 and 0.91. Cristiano accuracy and Snyder transferability peaked in different cells. The optimal Snyder solution was a cell containing 2,000 sites across 10 transcription factors (Robust panel). In the five seed samples, the Cristiano AUROC was 0.906 ± 0.002. Figure 1 B shows the Snyder surface over the same set of 50 units.

[0195] A forced multivariate classifier for seven cancers in the Cristiano cohort achieved a Snyder top-2 accuracy of 0.679 in five of the same types and 0.607 across all training classes. Joint density and selective scanning did not improve this accuracy (Figure 2C). Grid analysis of 15 functional annotation regions and five Snyder cancers of the same type revealed that the cross-cohort discrimination ability for each cancer was demonstrated by the readout of transcription factors or exon regions: breast cancer, colorectal cancer, and pancreatic cancer were distinguished by transcription factors, while lung cancer and ovarian cancer were distinguished by exons. In all cases, the selected regions outperformed random subsets of the same functional annotation region, with superiority probabilities ranging from 0.92 to 0.99. These functional annotation regions for each cancer collectively constitute... Figure 2 The panel showing heterogeneity classification.

[0196] Example 3: Cross-queue transfers exhibit diffusion in region readouts, but specificity in entropy factors. In the Cristiano cohort, the designed regional genome panels were comparable to the footprint-matched random panels in terms of AUROC values. Within a compact 0.5 Mb footprint range, 100 footprint-matched random regulatory genome panels achieved a Cristiano AUROC mean of 0.82. The AUROC values ​​of the nine designed regional panels were all concentrated in a narrow interval around this random level. Only CpG islands and composite regulatory panels exceeded the Cristiano random zero value at the 95% confidence level. Figure 3 C).

[0197] In the Snyder cohort, these two basic types showed differences. For regional detection, the mean AUROC of 100 randomly matched footprint panels was 0.868, and no designed regional panel clearly exceeded this random zero value. Open chromatin and binding site panels showed the best transfer performance, with AUROCs of 0.92 and 0.90, respectively, but their performance was only better than up to 80% of the random panels. Transfer performance of exons and CpG islands was lower than the random mean and lower than their accuracy within their own cohorts (Figures 3A and 3C). The composite regulation panel was the only regional panel that both passed the Cristiano null hypothesis test and achieved transfer, with a Cristiano AUROC of 0.85 and a Snyder AUROC of 0.92. Its transfer performance remained within the random distribution range, with a superiority probability of 0.78. Within the footprint range of 0.5 to 2.40 Mb, regional transfers did not exceed the random bands matched with the footprints (…). Figure 3 B).

[0198] The entropy primitives of each transcription factor exhibit different characteristics. When the entropy panel, consisting of 10 out of 808 randomly selected transcription factors, was transferred to the Snyder cohort, its average AUROC value was only 0.69. The transferred AUROC value of the Robust panel was 0.962, exceeding that of 98 out of the 100 random panels by a margin of +0.27, and its 95% confidence interval did not include zero. Figure 3 A). Both designed entropy panels passed the Cristiano random null hypothesis test with a superiority probability of 1.00.

[0199] Screening based on Cristiano accuracy did not produce this migration effect. While the Brittle panel achieved a maximum Cristiano AUROC value of 0.955, its migration value was only 0.606, exceeding that of only 58 out of 100 random transcription factor panels. Figure 3 C). The Robust panel, specifically selected for migration performance, occupies the region with higher values ​​on both coordinate axes of the plane (3A). Among all the designed panels, Cristiano accuracy is almost uncorrelated with Snyder migration rate.

[0200] Example 4: Detection and cancer type subtyping have been moved to a separate cohort. The detection and typing panel was applied to the Snyder cohort without retraining. The cohort included 5 cancer types that were the same as those in the Cristiano training set and 9 types that did not appear in the training set.

[0201] The AUROC value of the panel was 0.963 for 5 identical cancers and 0.960 for 9 different cancers. Figure 4 The accuracy of the heterogeneous typing panel was higher than 0.906 on the Cristiano training set. The panel achieved a Top-2 accuracy of 0.750 across five identical cancer types, covering a 1.61 Mb region for these cancer types across 28 Snyder samples. The multivariate panel with a 12.65 Mb region achieved 0.679 accuracy, compared to 0.40 accuracy for the five-class randomized baseline.

[0202] This aggregate value is primarily influenced by lung cancer—the group with the largest sample size but the weakest test performance, achieving an AUROC of 0.62 and a Top-2 recall of 0.56 on 9 samples. Breast cancer and pancreatic cancer, as other types with sufficient sample sizes, achieved AUROCs of 0.79 on 6 and 8 samples respectively, with Top-2 recalls of 0.83 and 0.75. Colorectal cancer and ovarian cancer, with only three and two samples respectively, ranked in the top two in each sample (Top-2 recall of 1.00). The insufficient sample size only resulted in the inability to obtain stable AUROC values ​​for each cancer type. Using a uniform 100bp probe for padding, and merging the heterogeneous typing and detection panels, the capture footprint was 5.23 Mb.

[0203] The Snyder cohort also included controls for systemic lupus erythematosus (SLE) and inflammatory bowel disease (IBD). Neither of these diseases was present in the Cristiano training cohort, therefore these detection panels were never screened for inflammation during the selection process. When assessed by entropy per transcription factor, both detection panels showed significantly lower detection rates for SLE than random, with the Brittle panel having an AUROC of 0.06 and the Robust panel 0.10. The mean predicted cancer probability for SLE samples was 0.98. At the same binding site, the Brittle panel, targeting 50 transcription factors for the benign-inflammatory composite endpoint, achieved a region readout AUROC of 0.80, exceeding 93 of the 100 footprint-matched random panels (with a mean of 0.70). In contrast, the Robust panel, targeting 10 transcription factors, showed a detection rate decreasing from 0.96 to 0.73, and its region specificity was lower than its random null value. Figure 4 B). Regarding the comparison between "cancer and systemic lupus erythematosus (SLE)" alone, the AUROC value reached 0.93, but the matched random panel also reached 0.91. The comparison between "cancer and inflammatory bowel disease (IBD)" remained between 0.46 and 0.67 across all assessment indicators.

[0204] In the Cristiano cohort, the detection results across cancer stages were consistent. The cross-stage prediction results of the Robust panel under leak-free conditions were stratified by stage. The AUROC value of this panel for detecting stage I and II cancer in healthy controls was 0.91, for stage III and IV it was 0.90, and for detecting only stage I cancer it was 0.87. Figure 4 C). The combined AUROC value of the phases is consistent with the primary AUROC value of the detection panel in the Cristiano queue.

[0205] Decomposing the classifier by feature type and aggregation granularity reveals which signals are responsible for detection and which are responsible for transmission. Figure 4 (D) At the regional granularity of the TFBS panel, fragmentomics indices are distributed within a narrow range: short fragment ratio of 0.625, normalized coverage of 0.846, and fragment size entropy of 0.843. At the “per factor” granularity, all indices increase synchronously, from 0.924 for terminal base sequence diversity to 0.954 for coverage and size distribution. Within each granularity, the indices exhibit clustered distributions, and Cristiano accuracy increases with increasing clustering granularity. The sum of all binding sites yields a single entropy value of 0.73, the entropy value for each region reaches 0.84, and the entropy value for each factor reaches 0.944.

[0206] These metrics were then transferred to the Snyder cohort and categorized by type. At the "per factor" granularity, fragment length distribution shifted, with a fragment size distribution AUROC of 0.873, terminal base sequence diversity of 0.819, and entropy of 0.775. Factor-level coverage, however, dropped to 0.488, below the random level. The region granularity showed a completely opposite result: coverage shifted to 0.878, while region entropy was only 0.334. On the Robust panel, fragment length distribution for each factor was optimized to a Snyder AUROC of 0.962 through screening.

[0207] Example 5: In the case of fragment length pattern reversal, the detection result will be reversed, but it can be recovered by using a scoring method that does not change the sign. This metastasis model was tested not only in a pan-cancer context but also validated against a second external cohort limited to a single tissue. The robust panel was applied directly to the Jiang liver cohort (containing 90 hepatocellular carcinoma patients, 36 patients with cirrhosis, 67 hepatitis controls, and 32 healthy donors) without retraining.

[0208] The test results showed a reversal. At an AUROC of 0.24 (95% confidence interval 0.15 to 0.33), the score for hepatocellular carcinoma was lower than that of the healthy control group, while the score of the precancerous lesion control group was closer to cancer characteristics than that of hepatocellular carcinoma. Figure 5 The Brittle panel also exhibited the exact same inversion phenomenon at an AUROC of 0.236.

[0209] Compared to the healthy baseline of their respective cohorts, the entropy values ​​of each factor in cancer samples in the Cristiano cohort decreased, while the entropy values ​​of each factor in cancer samples in the Jiang cohort increased, indicating that the panel factors exhibited opposite trends. Figure 5 The liver cohort showed an excessive number of long fragments at the binding sites, a phenomenon not observed in the pan-cancer cohort. Three liver cancer samples from the Snyder cohort were assessed as having cancer characteristics, with a mean predicted cancer probability of 0.61.

[0210] The detection capability was restored using a sign-invariant feature. By scoring the absolute deviation of each factor from the healthy baseline, a discriminative ability above chance was restored in the liver cancer cohort. This feature achieved an AUROC of 0.76 in the comparison between cancerous and healthy controls and 0.63 in the comparison with cirrhotic controls, with confidence intervals excluding 0.5 for both. However, in Cristiano cross-validation, this feature caused the accuracy to drop from 0.913 to 0.779. After training with both cohorts combined, the cohort-specific AUROC of the signed feature was 0.39 in the liver cohort and 0.90 in the pan-cancer cohort, while the sign-invariant feature achieved 0.74 and 0.77, respectively.

[0211] Example 6: Feature set simplification method can achieve WGS performance level As a whole-genome reference, a feature set using WGS format, 5 Mb bins, and following the DELFI protocol achieved a Cristiano AUROC value of 0.962. A compact panel within this cohort achieved whole-genome accuracy. The optimal protocol within the cohort was the Brittle feature set, which scored 0.955 at 2.42 Mb. However, this panel failed to migrate across cohorts, with its score dropping to 0.606 in the Snyder cohort. Therefore, proximity to the whole-genome reference within the cohort was not the goal. The Robust panel scored slightly lower within the cohort at 0.906 (coverage of 3.63 Mb), with a coverage of only 1 / 850th of the WGS coverage. This panel achieved a transfer score of 0.962 in the independent Snyder cohort (Figure 4), comparable to the whole-genome reference sequence's score within the cohort. At standard panel sequencing depths, the number of sequencing bases required to obtain equivalent fragmentomics signals was reduced by approximately 25-fold compared to 30×WGS.

[0212] Example 7: Detection and Subtyping Panel primarily targets cancer-related regulatory factors. The selection method returned significantly different lists of transcription factors under different machine learning seeds and binding site densities, while achieving comparable accuracy. The Robust panel and Brittle panel shared only ZNF426. However, a stable core containing 29 factors was repeated in all five seeds (Table 1).

[0213] Both the Robust panel and the stable core belong to cancer-related chromatin regulatory factors. The 10 transcription factors in the Robust panel include 8 identified chromatin regulators or oncogens: the SWI / SNF remodeling subunits BRD7, SMARCB1, and SMARCC1; the chromatin organizer CTCF; the centromere histone variant CENPA; the H3K4 methyltransferase KMT2B; the MYC family oncogene MYCN; and the inflammatory coupler CEBPB; as well as two uncharacterized zinc finger factors, ZNF426 and ZNF697. Figure 6 A). Against a background of 808 candidate transcription factors, this detection core was enriched in cancer regulatory pathways ( Figure 6 B). Its stable core of 29 transcription factors restored known cancer regulators that were not provided as prior information, including the tumor suppressor PTEN, the oxidative stress regulator NFE2L2, and the DNA-modifying enzyme TET3. Figure 6 E, Table 2).

[0214] The 29 transcription factors are: ARID4B, BACH2, CEBPB, CENPA, CPSF3, CTCF, DMC1, DTL, E4F1, HEXIM1, HNF1A, HOXA9, KDM5C, NELFA, NFE2L2, PCBP1, PHF2, PHF5A, PHF8, PHOX2B, PPARD, PTEN, RBM39, TET3, TLE3, ZFP1, ZNF382, ZNF384, and ZNF654.

[0215] The genotyping classifier follows the same logic as tissue specificity. After binding site abundance correction for selected regions of each precancerous classifier, the overexpressed factors are all known cancer-related regulatory factors. These factors include: progesterone receptor PGR and proliferation factor FOXM1 in breast cancer; bromodomain protein BRD1, and intestinal factors KLF5 and NEUROD1 in colorectal cancer; and therapeutic targets BRD4, chromatin remodeling factor ATRX, serum response factor SRF, and CREB1 in pancreatic cancer. Figure 6 C). The ubiquitous transcriptional elongation complex, which dominated the raw count, was abundance-driven and not enriched. Genes selected by the two exon pathway classifiers were involved in cancer hallmark processes, cell migration, proliferation, and growth factor receptor signaling (C). Figure 6 D). The genotyping signal is the copy number variation in these regions, thus focusing on tissue-specific regulatory and coding sites, i.e., each region complements the entropy of the factor carrying the detection.

[0216] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for data mining and panel design based on cfDNA whole-genome sequencing, characterized in that, The method includes the following steps: (S1) Provide a cfDNA whole genome sequencing dataset; the cfDNA whole genome sequencing dataset includes a training set and a validation set; the training set or validation set includes healthy control samples, samples of various cancer types and other disease samples; (S2) Preprocess the cfDNA whole genome sequencing dataset to obtain alignment fragment files; (S3) Based on different functional annotation regions, extract features from the sequencing data of the samples in the training set, including: (S3.1) Extracting features from the region panels includes: extracting fragments from the sequencing data of each sample based on the alignment fragment file, and screening the fragments; for a sample, mapping each screened fragment to one or more region panels that overlap with it; evaluating the 9-dimensional features and terminal base sequence diversity score of each region panel, and generating a region feature tensor for the sample based on the 9-dimensional features and terminal base sequence diversity score; and (S3.2) Extracting features from biological units includes: extracting fragments from the sequencing data of each sample based on the biological unit dataset, and filtering the fragments; for a sample, mapping each filtered fragment to one or more biological units, evaluating the entropy value of all fragments mapped to each biological unit, and generating an entropy matrix; (S4) Based on L1 regularization, sparse regions are filtered using steps selected from the following group: (S4.1) Based on the entropy matrix, a classification algorithm is used to perform binary classification modeling for healthy and cancerous individuals, obtaining the detection model and the corresponding preferred biological units; and / or (S4.2) Based on the entropy matrix, a classification algorithm is used to perform binary classification modeling of each cancer type with other cancer / disease types to obtain a unified classification model and the corresponding preferred biological unit; and / or (S4.3) Based on the different features obtained in step (S3), a classification algorithm is used to perform binary classification modeling for each cancer type and other cancer types to obtain a binary classification model corresponding to each feature; all the binary classification models are validated in the validation set to obtain a heterogeneity subtyping model for each cancer type and a preferred biological unit. (S5) Map the preferred biological units to their corresponding genomic sites from the preferred biological units, perform upstream and downstream sequence expansion on each genomic site, and merge and sort all the above genomic sites to obtain a panel.

2. The method as described in claim 1, characterized in that, The method further includes the following steps: (S6) Perform significance verification on the panel using steps selected from the group below: (S6.1) Random zero-distribution control verification in areas with equal coverage length; (S6.2) Zero-distribution comparison verification with the same function annotation area; (S6.3) Genome-wide reference validation; and / or (S6.4) Perform validation on the validation set.

3. The method as described in claim 1, characterized in that, The nine-dimensional features include: normalized depth, entropy value of fragment-size distribution, proportion of small fragments, and proportion of bins of six sizes; The method for evaluating the terminal base sequence diversity score (MDS) is as follows: For each selected fragment, based on its alignment to the reference genome, the 5′ and 3′ n-mer terminal base sequences are extracted from the reference genome sequence, and the score of each region panel in the n-mer sequence is calculated. 4 The entropy value on the possible terminal base sequences; for a region panel with less than 10 fragments, its MDS is set to NA; where n is selected from 4, 6 or 8; The biological unit dataset includes: a set of transcription factor binding sites and ATAC peak sets for various cancer types.

4. The method as described in claim 3, characterized in that, In (S4.1), when the entropy matrix is ​​an entropy matrix based on transcription factor binding sites, the specific steps include: (s4.1a) Preprocess the entropy matrices of different density levels to obtain the preprocessed entropy matrix; (s4.1b) On the preprocessed entropy matrix, a 5-fold cross-validation elastic network model is used to perform binary classification modeling of health and cancer to obtain a regularized path; (s4.1c) For each density level, the corresponding read operation point λ is determined based on multiple predetermined K values; K feature factors corresponding to each read operation point λ in the regularization path are read to obtain multiple candidate detection models; (s4.1d) Cross-validate the multiple candidate detection models to determine the detection model and the corresponding preferred biological unit from the multiple candidate detection models; In (S4.2), when the entropy matrix is ​​an entropy matrix based on transcription factor binding sites, it specifically includes: (s4.2a) Preprocess the entropy matrices of different density levels to obtain the preprocessed entropy matrix; (s4.2b) On the preprocessed entropy matrix, a resilient network model with 5-fold cross-validation is used to perform binary classification modeling for each cancer type and other cancer / disease types; (s4.2c) For each λ value on the regularized path in the elastic network model; under the condition that λ is the minimum value, record the top K biological units in each cancer type, thereby determining the unified subtyping model and the corresponding preferred biological units; K is an integer from 3 to 30.

5. The method as described in claim 3, characterized in that, In (S4.3), the specific steps include: (S4.3a) For each feature, the elastic network model is used to perform binary classification modeling and calibration of each cancer type and other cancer types to obtain the calibration model of each cancer under each feature, and to construct several random panel models corresponding to each feature. (S4.3b) For a type of cancer, the calibration model for the cancer under each feature and several random panel models corresponding to the feature are validated on the validation set; the probability that the performance value of the calibration model for the cancer under each feature exceeds the performance value of the random panel model corresponding to the feature is evaluated, and the feature with the larger probability value is used as the design attribution feature for the cancer. (S4.3c) Under the design attribution features, for a cancer, the mean of multiple L1 regularized fits is taken as the heterogeneity subtyping model of the cancer; all biological units corresponding to the heterogeneity subtyping models of all cancers are merged as the preferred biological unit.

6. A system for clinical reasoning using a panel, characterized in that, The system includes: (M1) Input module, the input module being configured to input data, the data including a cfDNA-based alignment fragment file; the alignment fragment file is obtained by hybridization capture and deep sequencing of a sample from the test object using a panel obtained according to the method of claim 1; (M2) Inference module, configured to perform the following operations: extract features from the comparison fragment file using step (S3) of the method of claim 1 to obtain a region feature tensor and an entropy matrix; process the region feature tensor and the entropy matrix using the detection model and / or classification model obtained in step (S4) of the method of claim 1 to obtain a detection score and a classification score respectively; combine the detection score and the classification score according to a preset threshold to obtain an inference result; the classification model includes a unified classification model and a heterogeneous classification model; (M3) Output module, which is configured to output the reasoning result of the reasoning module.

7. An electronic device comprising a processor and a memory, characterized in that, The memory contains multiple executable instructions, and the processor is used to read the instructions and execute the steps in the method as described in claim 1.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when read and executed by a processor, implement the steps of the method as described in claim 1.

9. A computer program product comprising computer-executable instructions, characterized in that, When the computer-executable instructions are executed by the processor, they implement the steps of the method as described in claim 1.

10. The use of a panel obtained by the method of claim 1 and its detection reagent, characterized in that, This is used to prepare diagnostic reagents or kits for: i) diagnosing / predicting cancer; ii) classifying cancer types; and / or iii) diagnosing / predicting cancer staging.

Images