Ultrasensitive detection of circulating tumor DNA via genome-wide integration

Abstract

The present disclosure relates to systems, software and methods for diagnosing neoplastic disease in a patient.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Application No. 62 / 636,135, filed on February 27, 2018, the entire contents of which are incorporated herein by reference. Technical Field

[0003] Embodiments of the present disclosure generally relate to the field of medical diagnosis. Specifically, embodiments of the present disclosure relate to compositions, methods and systems for tumor detection and diagnosis. Background Art

[0004] The enormous burden that cancer (such as solid tumors of the lung, breast, prostate, liver, and brain) places on human health is well documented in the medical literature. Most subjects are diagnosed with advanced neoplastic disease, which is associated with dismal outcomes. Recently, computed tomography (CT) has been found to improve early detection and is being used by a U.S. task force to screen high-risk populations. However, this method is limited by a high false-positive rate, leading to expensive and potentially dangerous follow-up evaluations.

[0005] A method for cancer diagnosis is to analyze the genetic clues or markers of tumor samples. Cancer genomes have acquired somatic mutations that drive their proliferation ability (Lawrence et al., Nature, 505 (7484): 495-501, 2014). Mutations in cancer genomes also provide key information about the evolutionary history and mutation processes active in each cancer (Martincorena et al., Cell, 171 (5): 1029-1041.e21, 2017; Alexandrov et al., Nature, 500 (7463): 415-421, 2013). Cancer mutation calls in patient biopsies have become a key step in assessing patient outcomes and therapeutic nominations. Identifying cancer driver mutations in liquid biopsy samples, such as cell-free circulating DNA (cfDNA), has been suggested as a translational platform for early cancer screening.

[0006] Statistical methods for analyzing genomic markers such as somatic mutations in DNA, such as single nucleotide variants (SNVs), require multiple independent observations (supporting reads) of the somatic variant at any genomic position to distinguish true mutations from sequencing errors. One technique for distinguishing true mutations from sequencing errors is to increase the depth of sequencing, which is useful as long as the tumor sample contains a high proportion of tumor cells. When the tumor cell content in the sample decreases, for example due to the presence of normal cells such as immune cells in the sample, multiple reads no longer support each somatic variant, thereby excluding the use of these mutation callers. For example, MUTECT is the current state-of-the-art low allele frequency somatic mutation caller. The core of MUTECT is to place SNVs in two Bayesian classifiers, one assuming that the SNV is generated by random noise and the other assuming that the site contains a true variant. Then, it filters SNVs based on the log-likelihood ratio from the two models. This is fundamentally different from the cfDNA setting. In the benchmark setting when the mutant allele frequency dropped to 0.05 and the tumor sample sequencing depth was reduced to 10X, the sensitivity of MUTECT dropped to less than 0.1 (Cibulskis et al., Nature Biotechnology, 31(3), 213, 203). Although MUTECT is currently the most advanced somatic mutation caller in the low-frequency setting, it still cannot identify somatic mutations in the tumor fraction, as observed in cfDNA.

[0007] A fundamental limitation of MUTECT and other mutation callers is that clinical sensitivity is below acceptable levels when input material is limited (e.g., in the setting of early-stage cancer disease). Such low cfDNA amounts translate only into hundreds to thousands of genome equivalents. Therefore, a limited number of physical fragments (e.g., 1000 genome equivalents in 6 ng cfDNA) to cover each site present in the sample may render ultra-deep sequencing (e.g., 100,000X) ineffective. Even with ultra-deep sequencing and advanced molecular error suppression, limited input material results in a detection limit of tumor fraction (TF) frequencies below 0.1-1%.

[0008] This limitation is exemplified by Abbosh et al. (Nature, 545(7655):446-451, 2017), who applied advanced sequencing methods, including technically challenging patient-specific targeted deep sequencing of lung adenocarcinoma, to identify approximately 18 mutations at a median sequencing depth of 42,000X. However, the paucity of cfDNA likely resulted in only 19% of early-stage subjects having cancer detected in cfDNA, even including more advanced stage III tumors in the study group. Furthermore, all of these patients identified as positive had lesions detectable by CT scan. These data suggest that even ultra-deep sequencing is currently inferior to imaging techniques in terms of inclusiveness and / or accuracy in the early stages of the disease.

[0009] Improved methods and systems are needed for identifying low-abundance disease markers, such as somatic mutations in cfDNA (including different subject-specific signatures), that are indicative of neoplastic disease. Additionally, systems and methods are needed to utilize such high-quality markers that can be used in the early diagnosis of tumors, thereby providing clinicians with better options for disease management and / or therapeutic intervention, and also substantially improving disease outcomes (e.g., improving survival and / or quality of life). Summary of the invention

[0010] Provided herein are programs, systems and methods for screening a subject for cancer and using the information obtained from the screening for early detection and disease stratification. In some embodiments, the programs, systems and methods of the present disclosure allow a user, such as a clinician, to diagnose cancer early.

[0011] In some aspects, the present disclosure provides a kind of classifier, it is trained to distinguish systematic errors and somatic mutations caused by cancer (such as tobacco-induced lung cancer).Using cancer mutations and sequencing errors are all systematic and controlled by unique signatures that can be learned and used for effective signal-to-noise ratio distinction, the classifier integrates such knowledge to improve the accuracy of cancer diagnosis and / or detection.For example, in the context of genomes, cancer signatures may include base substitutions that cause cancer-related mutagenesis.Such genomic signatures are particularly unique in cancers caused by exposure to tobacco and ultraviolet rays, including cancers related to dysregulated DNA checkpoints and / or repair enzyme activities (such as BRCA (BRCA1 or BRCA2), p53, APOBEC1, etc.).

[0012] The present disclosure also relates to multiple indicators that can indicate that the variant detected by sequencing is not a true somatic mutation but a sequencing or mapping technology artifact. In this case, previous studies have shown that sequencing errors are not random and may be related to the relevant technical factors of the DNA sequence context and sequencing technology. The fidelity of sequencing is also limited by the length of each sequencing read, and the error rate increases with the increase of the read length. Errors may occur when the read is mapped to the reference genome. Because the genome has variable regions, motifs and repeatable elements, this mapping process is computationally intensive and complex. Short nucleotide reads may be mapped to more than one position or may not be mapped at all. The system and method disclosed herein can be used to correct these limitations of existing methods for sequencing / mapping genomic data. The disclosed metrics are able to call true mutations from errors by analyzing multiple factors, including: (i) low base quality; (ii) low mapping quality; (iii) estimated fragment size (RP) of the read; (iv) estimated allele fraction (VAF) of the read; (v) sequence context; (vi) abundance; (vii) sequencing depth; and / or (viii) sequencing errors.

[0013] The present systems and methods are particularly suitable for detecting low-abundance markers that are predictive of cancer. The inventors of the present disclosure have recognized that sequencing width is not limited by the abundance of the input material and can replace methods that rely on deep sequencing. Since width sequencing is less dependent on the abundance of the input material, it can be used to improve the accuracy and sensitivity of detection. From a statistical point of view, the inventors first demonstrated that sequencing width (e.g., 10X sequencing of 10,000 mutations) is equivalent to depth (100,000X sequencing of a single mutation) and can be performed on as little as 1ng of cfDNA. Therefore, the analytical method of the present disclosure integrates whole genome mutation information for sensitive analysis of samples containing cfDNA, thereby easily and non-invasively detecting and / or accurately diagnosing tumors (e.g., tobacco-induced cancers).

[0014] In this case, simulated testing of plasma somatic mutation calls using synthetic mixtures of tumor and normal whole genome sequencing data from lung patients with variable fractions of tumor reads ranging from 1% to 0.001% (1 / 10,000) was used to demonstrate the strength and accuracy of the present method relative to the prior art. The performance of the technology was further benchmarked by the following steps: first, standard mutation calls were used to characterize patient-specific somatic cancer SNVs on pure tumor and normal samples from patients; and then several methods were used to detect cancer mutations in plasma samples, including the convolutional network of the present invention. The sensitivity and accuracy of each method using pure tumor mutation calls as a reference demonstrated the high signal and / or low noise of the analytical method of the present disclosure. Finally, a validation study conducted on actual cfDNA samples from early-stage lung cancer patients showed significantly superior sensitivity and precision compared to the current state-of-the-art methods.

[0015] The present disclosure relates to the following non-limiting embodiments:

[0016] According to various embodiments, a method for genetic screening of cancer in a subject is provided. The method includes receiving a subject-specific whole genome reading summary associated with multiple genetic markers from a biological sample of the subject. The biological sample may include a tumor sample. The reading summary may each include a reading of a single base pair length. The method may also include filtering artificial sites from the reading summary. Filtering may include removing the repeated sites generated on the reference healthy sample queue from the reading summary. Alternatively or in combination, filtering may include identifying germline mutations in the biological sample and / or identifying shared mutations between peripheral blood mononuclear cells of tumor samples and normal cell samples as germline mutations, and removing the germline mutations from the reading summary. The method may also include filtering noise from the whole genome reading summary using at least one error suppression scheme to generate a filtered reading set of the whole genome reading summary. The at least one error suppression scheme may include calculating the probability that any single nucleotide variation in the summary is an artificial mutation, and removing the mutation. The probability may be calculated as a function of a feature selected from a group including the following: mapping quality (MQ), variant base quality (MBQ), position in reading (PIR), average read base quality (MRBQ) and a combination thereof. Alternatively or in combination, the at least one error suppression scheme may include: using inconsistency testing between independent repeats of the same DNA fragment generated by polymerase chain reaction or sequencing processing, and / or repeat consistency to remove artificial mutations, wherein artificial mutations are identified and removed when there is a lack of consistency between the majority of a given repeat family. The method may include compiling a subject-specific signature using a filtered set of reads based on a comparison of a specific mutation signature associated with a predetermined mutagenesis process. The method may also include statistically quantifying the confidence estimate that the subject's biological sample includes a cancer-associated mutation signature through a subject-specific signature based on a comparison of the cancer-associated mutation signature exposure value with a background mutation signature cohort. The method may include: if the confidence estimate that the subject's biological sample contains a cancer-associated mutation signature exceeds a given threshold, the subject is screened for cancer.

[0017] According to various embodiments, a method for genetic screening of cancer in a subject is provided. The method includes receiving a subject-specific whole genome reading summary associated with multiple genetic markers from a biological sample of the subject. The biological sample may include a tumor sample. The reading summary may each include a copy number variation (CNV). The method may include dividing the reading summary into multiple windows. The method may include calculating a feature set for each window. These features may include a median depth coverage of each window and a representative fragment size of each window. The method may include filtering artificial sites from the reading summary. Filtering may include removing duplicate sites generated on a reference healthy sample cohort from the reading summary. The method may include normalizing the reading summary to generate a filtered reading set of the whole genome reading summary. The method may include calculating the estimated tumor score using the filtered reading set by calculating the linear relationship between the feature sets of each window and using a regression model to convert the calculated relationship into an estimated tumor score. Alternatively or in combination, the method may include calculating an estimated tumor score based on one or more integrated mathematical models as a function of the feature set of each window calculated in the subject-specific whole genome reading summary. The method may include screening the subject for cancer if the estimated tumor score exceeds an empirical threshold.

[0018] According to various embodiments, a system for genetic screening of cancer to a subject is provided. The system includes an analysis unit, the analysis unit includes a pre-filter engine, the pre-filter engine is configured and arranged to receive a subject-specific full genome reading summary associated with multiple genetic markers from a biological sample of the subject, the biological sample includes a tumor sample, wherein the reading summary each includes a reading of a single base pair length. The pre-filter engine can be configured and arranged to filter out artificial sites from the reading summary, wherein filtering includes removing the repeated sites generated on the reference healthy sample queue from the reading summary. The pre-filter engine can be configured and arranged to also or in combination identify germline mutations in the biological sample and / or identify shared mutations between peripheral blood mononuclear cells of tumor samples and normal cell samples as germline mutations, and remove the germline mutations from the reading summary. The analysis unit may include a correction engine, the correction engine is configured and arranged to use at least one error suppression scheme to filter the noise of the reading summary to produce a filtered reading set of a full genome reading summary. The at least one error suppression scheme may include calculating the probability that any single nucleotide variation in the summary is an artificial mutation, and removing the mutation, wherein the probability is calculated as a function of a feature selected from the group including the following: mapping quality (MQ), variant base quality (MBQ), position in read (PIR), median read base quality (MRBQ) and combinations thereof. The at least one error suppression scheme may also or in combination include using an inconsistency test between independent repeats of the same DNA fragment generated by polymerase chain reaction or sequencing processing, and / or repeat consistency to remove artificial mutations, wherein artificial mutations are identified and removed when there is a lack of consistency between the majority of a given repeat family. The system may include a computing unit configured and arranged to compile a subject-specific signature using a filtered set of readings based on a comparison of a specific mutation signature associated with a predetermined mutagenesis process. The computing unit may be configured and arranged to statistically quantify the confidence estimate of the subject's biological sample, including a cancer-related mutation signature, by a subject-specific signature based on a comparison of a cancer-related mutation signature exposure value with a background mutation signature cohort. The computing unit may be configured and arranged to screen the subject for cancer if a confidence estimate that the biological sample of the subject contains a cancer-associated mutation signature exceeds a given threshold.

[0019] According to various embodiments, a system for performing genetic screening of cancer on a subject is provided. The system includes an analysis unit, the analysis unit includes a binning engine, the binning engine is configured and arranged to receive a subject-specific full genome reading summary related to a plurality of genetic markers from a biological sample of a subject, the biological sample includes a tumor sample, wherein the reading summary each includes a copy number variation (CNV). The binning engine can be configured and arranged to divide the reading summary into a plurality of windows. The binning engine can be configured and arranged to calculate a feature set of each window, and these features include the median depth coverage of each window and the representative fragment size of each window. The system may include a pre-filter engine, the pre-filter engine is configured and arranged to filter out artificial sites from the reading summary, wherein filtering includes removing the repeated sites generated on the reference healthy sample queue from the reading summary. The system may include a normalization engine, the normalization engine is configured and arranged to normalize the reading summary to produce a filtered reading set of a full genome reading summary. The system may include a computing unit, the computing unit is configured and arranged to use a filtered reading set to calculate an estimated tumor score (eTF). The computing unit may calculate the eTF by calculating the linear relationship between the feature set of each window and converting the calculated relationship into an estimated tumor score using a regression model. Alternatively or in combination, the computing unit may calculate the eTF as a function of the feature set of each window calculated in the subject-specific whole genome read summary based on one or more integrated mathematical models. The computing unit may be configured and arranged to screen the subject for cancer if the estimated tumor score exceeds an empirical threshold.

[0020] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising: (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated read fragment size; and / or (4) estimated read allele fraction (VAF); (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by utilizing machine learning (ML) methods to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising multiple true reads in the summary; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on the match between the subject-specific signature and the cancer signature; and (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains the cancer-associated mutation signature.

[0021] In some embodiments of the present methods, the biological sample of the subject comprises plasma, cerebrospinal fluid, pleural fluid, ocular fluid, feces, urine, or a combination thereof.

[0022] In some embodiments of the present method, cancer signatures include COSMIC tobacco signature, UV signature, breast cancer (BRCA) signature, microsatellite instability (MSI) signature, apolipoprotein BmRNA editing enzyme, poly (ADP-ribose) polymerase (PARP) high activity signature, catalytic polypeptide-like (APOBEC) signature. In particular, in some embodiments, cancer signatures include patterns associated with tissue-specific epigenetic patterns, such as tissue-specific chromatin accessibility patterns.

[0023] In some embodiments of the present method, by using machine learning (ML) methods to distinguish cancer-related mutation features (true positives) from PCR or sequencing error-related features (false positives), the sequencing noise associated with each read in the summary is filtered. In some embodiments, the machine learning method includes a deep convolutional neural network (CNN), a recurrent neural network (RNN), a random forest (RF), a support vector machine (SVM), a discriminant analysis, a nearest neighbor analysis method (KNN), an integrated classifier, or a combination thereof. In some embodiments, ML has been trained to distinguish between sequencing reads altered by cancer and reads altered by sequencing or PCR errors. In some embodiments, ML has been trained on a large cancer data set of whole genome sequencing (WGS), which contains billions of reads across tumor mutations and normal sequencing errors. In some embodiments, ML is able to (a) identify sequencing or PCR artifacts with high accuracy, and (b) integrate sequence context and read-specific features.

[0024] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising: (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removing artificial reads from the summary by statistically classifying each read in the summary as signal or noise, where noise is detected (P N ) is a function of (1) the base quality (BQ) of the read; (2) the mapping quality (MQ) of the read; (3) the estimated fragment size of the read; and / or (4) the estimated variant allele fraction (VAF) of the read; (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by utilizing a machine learning (ML) method, the method comprising performing an optimal receiver operating characteristic (ROC) curve based on a joint base quality (BQ), mapping quality (MQ) score and fragment size, which includes a probabilistic classification of genetic markers in the summary; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising multiple real reads in the summary; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and a cancer signature; and (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains a cancer-associated mutation signature.

[0025] In some embodiments of the present method, the tumor is a tumor including the following: heterogeneous or homogeneous brain cancer, lung cancer, skin cancer, nasal cancer, laryngeal cancer, liver cancer, bone cancer, lymphoma, pancreatic cancer, skin cancer, intestinal cancer, rectal cancer, thyroid cancer, bladder cancer, kidney cancer, oral cancer, gastric cancer, solid tumors, non-small cell lung cancer (NSCLC), tobacco-induced cancer (TIC), ultraviolet-induced cancer, apolipoprotein B mRNA editing enzyme catalytic protein (APOBEC) activity-mediated cancer, including breast cancer protein (BRCA) mutations, including poly (ADP-ribose) polymerase (PARP) activity and including microsatellite instability (MSI) tumors. In some embodiments of the present method, the screening method allows diagnosis of early cancer disease in undiagnosed and / or asymptomatic patients. In particular, the subject is an early cancer patient in phase I to III.

[0026] In some embodiments, the present disclosure relates to a method for genetically screening a subject for cancer and recommending a therapy, comprising: (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N ) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated fragment size of the read; and / or (4) estimated allele fraction (VAF) of the read; (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by using machine learning (ML) methods to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising a plurality of true reads in the summary; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and the cancer signature; (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains a cancer-associated mutation signature; (g) nominating signature-based therapies based on the patient-specific signature employed in the diagnosis. In some embodiments, the therapy nominations include PARP inhibitors for the BRCA signature and immunotherapy for the MSI signature. In some embodiments, the PARP inhibitor is niraparib, olaparib, veliparib, rucaparib, and / or talazoparib. In some embodiments, the immunotherapy for the MSI signature comprises an anti-PD-1 antibody (e.g., nivolumab or pembrolizumab) or an anti-CTLA4 antibody (e.g., nivolumab or pembrolizumab). In some embodiments, the tumor comprises a heterogeneous or homogeneous brain cancer, lung cancer, skin cancer, nose cancer, laryngeal cancer, liver cancer, bone cancer, lymphoma, pancreatic cancer, skin cancer, intestinal cancer, rectal cancer, thyroid cancer, bladder cancer, kidney cancer, oral cancer, stomach cancer, solid tumors, lung adenocarcinoma, ductal adenocarcinoma (breast tumor), non-small cell lung cancer lung adenocarcinoma (NSCLC LUAD), skin melanoma, urothelial carcinoma (bladder tumor), colorectal cancer (Lynch) or osteosarcoma.

[0027] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N ) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated fragment size of the read; and / or (4) estimated variant allele fraction (VAF) of the read; (c) adaptively and / or systematically filtering sequencing noise associated with each read in the compendium by utilizing a machine learning (ML) approach to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising multiple true reads in the compendium; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and the cancer signature; (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains the cancer-associated mutation signature, wherein step (f) comprises solving a linear optimization problem -min‖Ax-b‖, x≥0, where A is a mutation signature sequence context matrix, x is the contribution of each cosmic mutation signature (variable), and b is a patient-specific sequence context compendium. In some embodiments, the optimization problem is solved by a non-negative least squares method (NNLS), a cross entropy global optimization method, a golden section search method, or a combination thereof. In some embodiments, the method further comprises verifying the confidence using a comparison of the cancer mutation signature with a plurality of random background signatures, for example, using a comparison of the cancer mutation signature with a plurality of random background signatures. In some embodiments, the comparing step comprises an evaluation of a zscore, wherein a zscore above a threshold indicates that the subject-specific signature is specific to the cancer signature and is not associated with a random signature.

[0028] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N ) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated fragment size of the read; and / or (4) estimated variant allele fraction (VAF); (c) adaptively and / or systematically filtering sequencing noise associated with each read in the compendium by utilizing a machine learning (ML) approach to distinguish between cancer-associated mutation signatures and PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising a plurality of true reads in the compendium; (e) statistically quantifying a confidence estimate that the subject's biological sample comprises circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and the cancer signature; (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample comprises the cancer-associated mutation signature, wherein step (b) comprises (1) removing reads with low mapping quality (e.g., <29, ROC optimization); establishing duplicate families (representing multiple PCR / sequencing copies of the same DNA fragment), and selecting the corresponding reads based on a consensus test (consensus test); test) to generate corrected reads; (3) removing reads with low base quality (e.g., <21, ROC optimization); and / or (4) removing reads with high fragment size (e.g., >160, ROC optimization), step (f) includes calculating sequence context similarity between the patient sequence-context summary and the specific cosmic sequence-context summary.

[0029] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated read fragment size; and / or (4) estimated read allele fraction (VAF); (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by utilizing a machine learning (ML) approach to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising multiple true reads in the summary; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and the cancer signature; and (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains a cancer-associated mutation signature, wherein step (f) comprises evaluating the similarity between the subject-specific signature and the cancer signature based on cosine similarity, correlation, mutual information, or a combination thereof. In some embodiments, the method further comprises verifying the confidence using a comparison of the cancer mutation signature to a plurality of random background signatures, e.g., using a comparison of the cancer mutation signature to a plurality of random background signatures. In some embodiments, the comparing step comprises an evaluation of a zscore, wherein a zscore above a threshold indicates that the subject-specific signature is specific to the cancer signature and is not correlated with the random background signature.

[0030] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated read fragment size; and / or (4) estimated read allele fraction (VAF); (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by utilizing a machine learning (ML) approach to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising multiple true reads in the summary; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and the cancer signature; and (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains a cancer-associated mutation signature, wherein step (f) comprises comparing the cancer-specific signature confidence (zscore) to an empirical threshold calculated by a background noise model. In some embodiments, the empirical noise model is defined by measuring cancer-specific signature confidences (zscores) in normal healthy samples and converting them to base noise zscore estimates, where the zscore estimate noise threshold is between 1 and 5.

[0031] In some embodiments of the aforementioned cancer screening / diagnosis methods, the subject-specific signature is matched to a cancer-specific mutation signature that includes markers that are differentially expressed in tumors but not in normal samples. In some embodiments, the tumor sample includes a lung tumor, a breast tumor, a melanoma, a bladder tumor, a colorectal tumor, or a bone tumor.

[0032] In some embodiments of the foregoing cancer screening / diagnosis methods, the method allows for early detection in at least 50% of subjects.

[0033] In some embodiments of the aforementioned cancer screening / diagnosis methods, the method includes further implementing computed tomography (CT) screening, wherein the CT screening step is performed before, simultaneously with, or after genetic screening. In some embodiments, the cancer is a solid tumor, and the CT screening includes, for example, detecting suspicious nodules in patients with benign lesions. In some embodiments, benign lesions are identified by advanced CT screening, histopathology, and / or biopsy.

[0034] In some embodiments of the aforementioned cancer screening / diagnosis methods, the method includes distinguishing between malignant nodules and benign nodules, which increases the positive predictive value (PPV) of CT screening by, for example, at least 30%, at least 40%, at least 50%, at least 60%, at least 80% or at least 90%.

[0035] In some embodiments of the foregoing cancer screening / diagnosis methods, the method comprises early detection (ED) of malignancies.

[0036] In some embodiments, the present disclosure relates to a method for genetic screening of a subject for cancer, comprising (a) receiving a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) determining a probability of detection based on noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N ) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; (3) estimated fragment size of the read; and / or (4) estimated allele fraction (VAF) of the read; (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by using a machine learning (ML) method to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the filtering step (b), compiling a subject-specific signature comprising multiple true reads in the summary; (e) statistically quantifying a confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA) based on a match between the subject-specific signature and the cancer signature; (f) screening the subject for cancer based on the confidence estimate that the subject's biological sample contains a cancer-associated mutation signature, wherein step (a) comprises aggregating whole genome mutation data by whole genome sequencing, and step (c) comprises detecting the mutation signature using a mathematical optimization step. In some embodiments, the mathematical optimization step comprises employing non-negative least squares.

[0037] In some embodiments, the present disclosure relates to a method for detecting a pre-malignant tumor signature in a subject, comprising (a) generating a subject-specific genome-wide read profile associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker profile is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) generating a probability of detection noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; and / or (3) estimated read fragment size (4) read allele fraction (VAF); (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by utilizing machine learning (ML) methods to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the optional filtering step (b), compiling a subject-specific signature comprising multiple true reads in the summary; (e) based on the match between the subject-specific signature and the cancer signature, statistically quantifying the confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA); (f) based on the confidence estimate that the subject's biological sample contains the cancer-associated mutation signature, detecting a pre-malignant tumor signature. In some embodiments, the pre-malignant tumor comprises a heterogeneous or homogeneous brain cancer, lung cancer, skin cancer, nose cancer, laryngeal cancer, liver cancer, bone cancer, lymphoma, pancreatic cancer, skin cancer, intestinal cancer, rectal cancer, thyroid cancer, bladder cancer, kidney cancer, oral cancer, stomach cancer, solid tumors, lung adenocarcinoma, ductal adenocarcinoma (breast tumor), non-small cell lung cancer lung adenocarcinoma (NSCLC LUAD), skin melanoma, urothelial carcinoma (bladder tumor), colorectal cancer (Lynch) or osteosarcoma, in particular Lynch syndrome or BRCA genetic defects.

[0038] In some embodiments of the foregoing methods, machine learning (ML) includes a deep convolutional neural network (CNN) that adaptively and / or systematically filters sequencing noise. In some embodiments, the CNN includes using a deep learning algorithm in a pan-tumor cohort to identify signatures that distinguish true tumor mutations from artifacts; assigning confidence estimates to each individual mutation detected in a tumor patient sample; integrating the confidence estimates across the genome; and applying non-negative least squares to a specific cosmic mutation signature in the sample.

[0039] In some embodiments, the present disclosure relates to a computer-readable medium comprising computer-executable instructions, which, when executed by a processor, cause the processor to perform a method or a set of steps for early detection of a tumor or detection of pre-malignant tumor lesions, the method or steps comprising: (a) generating a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of the subject, wherein the genetic marker summary is selected from the group consisting of: single nucleotide variations (SNVs), short insertions and deletions (indels), copy number variations, structural variants (SVs), and combinations thereof; (b) generating a probability of detection noise (P N ), removes artificial reads from the compendium by statistically classifying each read in the compendium as signal or noise, where the probability of detecting noise (P N) is a function of (1) base quality (BQ) of the read; (2) mapping quality (MQ) of the read; and / or (3) estimated read fragment size (4) read allele fraction (VAF); (c) adaptively and / or systematically filtering sequencing noise associated with each read in the summary by utilizing machine learning (ML) methods to distinguish cancer-associated mutation signatures from PCR or sequencing error-related signatures; (d) based on the noise removal step (c) and the optional filtering step (b), compiling a subject-specific signature comprising multiple true reads in the summary; (e) based on the match between the subject-specific signature and the cancer signature, statistically quantifying the confidence estimate that the subject's biological sample includes circulating tumor DNA (ctDNA); (f) based on the confidence estimate that the subject's biological sample contains the cancer-associated mutation signature, detecting a pre-malignant tumor signature. In some embodiments, the ML comprises a hierarchical convolutional neural network (CNN) with a single fully connected layer at one end, wherein the CNN remains spatially invariant when convolving over a three-nucleotide window; and maintains a quality map by collapsing the read fragments into multiple segments, each segment representing a region of approximately eight nucleotides.

[0040] In some embodiments of the aforementioned computer-readable medium or method, the CNN includes 8 layers, including a single fully connected layer at one end and two consecutive convolutional layers, and its output is downsampled by a maximum pooling (maxpooling) with a receptive field of 2 and a stride of 2, wherein the 8-layer CNN maintains a quality map by collapsing the read fragments into about 25 independent segments, and convolves on the columns at the positions of the genomic reads using a perceptive field of size 3; and wherein the output of the last convolutional layer is directly applied to the sigmoid fully connected layer, thereby finally classifying the markers. In some embodiments, the CNN includes a read representation that jointly captures the genomic context of the alignment, the complete read sequence, and the integral of the quality score per base. In some embodiments of the aforementioned computer-readable medium or method, compared with MUTECT, the CNN enriches tumor-specific markers containing somatic mutations in genomic reads by about 1.12 times to about 30 times.

[0041] In some embodiments, the present disclosure relates to a computer-readable medium comprising computer-executable instructions that, when executed by a processor, cause the processor to perform a method or a set of steps for diagnosing cancer in a subject in need of diagnosis, the medium comprising a convolutional neural network (CNN) developed by the following method: (A) receiving a genetic marker summary received from a sample of a subject, wherein the genetic markers include somatic single nucleotide variations (sSNVs); somatic copy number variations (sCNVs); insertions / deletions (indels); or structural variations (SVs) in genomic reads; (B) processing the genetic marker summary for each subject in a pan-tumor cohort to identify signatures that can distinguish between true cancer markers and artificial errors; (C) assigning a confidence estimate to each signature in the summary based on processing step (B); (D) integrating the confidence estimates for each signature of step (C) in the genomic reads to establish a tumor signature; (E) mathematically optimizing the tumor signature. In some embodiments, the assignment of confidence estimates comprises: (1) calculating a confidence measure of the contribution of a cosmic mutation signature using linear mixture optimization; or (2) calculating the similarity of a patient sequence context profile to a particular cosmic signature. In some embodiments, the linear mixture optimization comprises solving an algebraic function min‖Ax-b‖, x≥0, where A is a mutation signature sequence context matrix, x is the contribution of each cosmic mutation signature, and b is a patient-specific sequence context profile. In some embodiments, A in the algebraic function ‖Ax-b‖, x≥0 comprises at least 5, at least 10, at least 15, at least 20, at least 25, or at least 30 COSMIC signatures and 100 random mutation signatures. In some embodiments, the linear mixture optimization includes calculating the distribution of random signature contributions including extracting E_random (average contribution score) and std_random (std contribution score); checking the confidence of the contribution detection of each COSMIC signature by zscore, including calculating the metric (cosmic_sig_contribution-E_random) / std_random, wherein the metric represents the significance of the specific signature compared to the random set. In some embodiments, the mathematical optimization step includes employing non-negative least squares (NNLS).

[0042] In some embodiments, the present disclosure relates to a system for diagnosing a tumor in a subject in need thereof, comprising: a data acquisition unit configured to receive a plurality of reads comprising genetic markers amplified and sequenced from a biological sample comprising a plasma sample and a normal cell sample of the subject; a marker identification unit configured to identify a plurality of subject-specific markers in a subject-specific genetic marker summary, the marker identification unit being communicatively connected to the data acquisition unit, comprising: a noise removal unit that removes artificial noise based on base quality of the reads, mapping quality of the reads, fragment size of the reads, and / or variable allele frequency (VAF) of the reads; and a classification engine configured to statistically classify each noise-removed read in the summary based on a confidence interval score representing a statistical level of statistical association between the read and the tumor, wherein the classification engine utilizes machine learning (ML) to adaptively and systematically filter noise introduced in an amplification step or a sequencing step, and further matches the noise-removed ML-filtered reads in the summary with one or more known cancer signatures; and a diagnosis unit configured to diagnose the tumor based on the match.

[0043] In some embodiments of the systems of the present disclosure, the classification engine is further configured to match the noise-removed ML-filtered reads in the summary with one or more known cancer signatures by computing a confidence metric using a linear mixture optimization problem.

[0044] In some embodiments of the systems of the present disclosure, the linear mixed optimization comprises calculating a zscore confidence estimate of the association between tumor incidence and a tumor mediator selected from tobacco exposure, UV exposure, DNA repair dysregulation, DNA editing errors, microsatellite instability, or a combination thereof.

[0045] In some embodiments of the systems of the present disclosure, the artificial noise removal engine is configured to perform an optimal receiver operating characteristic (ROC) curve based on a base quality (BQ) score of the read, a mapping quality (MQ) score of the read, a fragment size of the read, or a variable allele frequency (VAF) of the read, the optimal receiver operating characteristic curve includes a probability classification of the reads in the summary. In some embodiments of the systems of the present disclosure, the artificial noise removal engine is also configured to filter noise based on (iii) position (RP) in the read; (iv) sequence context (SC) of the read; (v) abundance of the read; (vi) sequencing depth and / or (vii) sequencing errors.

[0046] In some embodiments of the system of the present disclosure, the confidence calculation includes a zscore confidence estimate by parsing an algebraic function including min‖Ax-b‖, x≥0, where A is a mutation signature sequence context matrix, x is the contribution of each cosmic mutation signature; b is a patient-specific sequence context summary. In some embodiments, the zscore confidence estimate includes parsing an algebraic function including min‖Ax-b‖, x≥0, where A includes 30 cosmic signatures and 100 random mutation signatures; calculating the contribution distribution of the cosmic signature (CSC) random signature (E_random) including an average contribution score (ACS) and a standard contribution score (std_random); checking the confidence of each cosmic signature contribution by calculating the zscore metric using the function (CSC-E_random) / std_random, where the zscore represents the significance of the contribution of a specific signature compared to a random set. In some embodiments, the zscore confidence estimate includes calculating the similarity of the patient sequence-context summary to the specific cosmic signature. In some embodiments, the zscore confidence estimate includes normalizing the patient sequence-context summary to obtain a density function; calculating the cosine similarity between the patient sequence-context density function and the cosmic signature density function; and normalizing the cosine similarity by dividing the cosine similarity between the patient sequence-context density function and the non-information uniform density function. In some embodiments, the zscore confidence estimate includes checking whether the zscore exceeds a detection threshold, wherein the threshold includes the basic noise in the healthy sample estimated empirically. In some embodiments, the cancer signature includes a tobacco signature, and the positive confidence interval includes a zscore greater than 2, 3, 4, preferably greater than 5 standard deviations.

[0047] In some embodiments of the methods and systems of the present disclosure, genetic markers include SNVs, CNVs, indels, and / or SVs in DNA, and the receiving unit receives whole genome sequencing (WGS) genetic data, such as genetic data from biological samples including plasma samples and normal cell samples, the plasma samples include cell-free DNA (cfDNA); the normal cell samples include peripheral mononuclear blood cells (PMBC), and wherein the genetic data includes multiple markers, the multiple markers include somatic single nucleotide variations (sSNVs) or somatic copy number variations (sCNVs) or a combination thereof. In some embodiments, the amount of cfDNA in the sample is between about 0.1ng / ml and about 20.0ng / ml. In some embodiments, the sample has a low tumor score (TF), as measured by the ratio of the amount of tumor DNA molecules to normal DNA molecules, for example, between about 0.0001% (1 to one million molecules) and about 20%. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The details of one or more embodiments of the present disclosure are set forth in the accompanying drawings / tables and the following description. Other features, objects, and advantages of the present disclosure will be apparent from the accompanying drawings / tables and detailed description, and from the claims.

[0049] Figure 1A A representative flow chart of the diagnostic method of the present disclosure is shown. In a first step 110, a subject-specific whole genome read summary associated with a plurality of genetic markers (e.g., somatic SNVs) is received from a sample of the subject, for example, generated by whole genome sequencing. In step 120, a probability of detecting noise (P N ), removing artifactual reads by statistically classifying each read as signal (S) or noise (N), where the probability of detecting noise (P N ) is a function of (1) base quality (BQ); (2) mapping quality (MQ); (3) estimated fragment size (4) estimated allele fraction (VAF). Other secondary parameters may also be used, such as (v) positional size of the read (vi) sequence context (SC); (vii) abundance; (viii) sequencing depth and / or (ix) sequencing error. The noise-reduced reads may be input into a convolutional neural network that has been trained using a computer dataset and / or a dataset from a pan-cancer cohort. The neural network adaptively and systematically filters sequencing noise in step 130. Next, in step 140, a subject-specific signature including multiple real reads in the summary is compiled based on the noise removal step 120 and the filtering step 130. Next, in step 150, a confidence estimate is made that the subject's biological sample includes circulating tumor DNA (ctDNA) by matching the subject-specific signature and the cancer signature. In step 160, the subject is screened for cancer based on the confidence estimate. Figure 1B A representative workflow for performing cancer screening in a subject is shown, according to various embodiments. Figure 1C A representative workflow for performing cancer screening in a subject is shown, according to various embodiments. Figure 1D A representative workflow for cancer screening in a subject based on the measurement of single nucleotide polymorphisms (SNVs) or indels is shown. Figure 1E A representative workflow for cancer screening in a subject based on measurement of copy number variation (CNV) or structural variation (SV) is shown. Figure 1F Shown is a representative scheme for generating a computer database of synthetic plasma for seven cancer patients - two melanomas, three lung adenocarcinomas, and two breast cancers (Scheme A).

[0050] Figure 2A representative flow chart is provided that outlines the use of the systems and methods of the present disclosure to aid in the early detection of cancer, which reduces, if not eliminates, the need for surgical and / or therapeutic interventions. Early cancer detection can bring many economic and health benefits, including avoiding surgical risks (e.g., pneumonia, bleeding, infection, clotting (hematoma), and reaction to anesthesia), chemotherapy or immunotherapy side effects (e.g., fatigue, hair loss, easy bruising and bleeding, infection, anemia, nausea and vomiting, appetite changes, constipation, diarrhea, mouth, tongue and throat problems, nerve and muscle problems, such as numbness, tingling and pain; skin and nail changes, such as dry skin and changes in skin color; urine and bladder changes and kidney problems; weight fluctuations).

[0051] Figures 3A-3C The parameter detection probability as a function of each parameter is shown. Figure 3A , the figure shows that the detection probability decreases rapidly using samples containing a low tumor fraction (TF). Figure 3B Shown are the predicted average number of sites detected and the probability of at least one detection as a function of the number of unique DNA fragments (genome equivalents or coverage), mutation load (N), and tumor fraction (TF). Figure 3C We show that even at a TF of 1:100,000, integrating more than 20,000 sSNVs (~10 mutations / megabase pair found in 17% of human cancers) provides high detection probability (up to 0.98) with a modest sequencing effort (20× coverage), which is easily achievable using standard whole-genome sequencing (WGS).

[0052] Figures 4A-4E Optimization of SNV markers is shown. Figure 4A A linear relationship is shown between the number of artificial SNV detections (errors) and the total number of unique reads examined. This represents an error rate of 1 error per 1000 reads, indicating that the errors are mainly due to the sequencing error rate (1 / 1000). Each point is a control sample (TF=0), which was generated from PBMC data from 6 different patients, with 3 different cancer types (lung cancer, melanoma, and breast cancer), multiple coverage ranges (from 2X to 25X), and multiple independent replicates. All of these are clearly on the same regression line, so this is invariant to the cancer type. Figure 4B Receiver operating characteristic (ROC) curves for base quality filtering (BQ) are shown. Figure 4CA line graph of the number of reads examined (x-axis) and the number of errors detected (y-axis) in the filtered multi-cancer error model is shown, demonstrating a linear relationship between the number of artificial SNV detections (errors) and the total number of unique reads examined. SNV detections (errors) were performed after applying the optimized BQ and MQ filters. Figure 4D The effect of applying a combined BQ and MQ optimized filter was shown to allow for approximately a seven-fold change in suppression of sequencing errors. The distribution of error rates across multiple replicates was assessed using control samples. For both lung cancer and melanoma cancer types, the pre-filter noise rate was ~2x10 -3 , and for both cancer types, the post-filter noise rate dropped to ~2x10 -4 . Figure 4E Heatmaps of error rates as a function of plasma coverage (x-axis) and tumor burden (y-axis) are shown (red indicates more errors, blue indicates fewer errors). Estimated error rates (e.g., number of SNVs detected divided by total number of unique reads examined) at various coverages and tumor mutation burdens (the tumor mutation burden was modified by subsampling the original patient-specific tumor mutation list) are shown. Each entry in the matrix is ​​the average of multiple independent replicates. For each mutation burden greater than 2000, this shows that the error rate is essentially unchanged with coverage and mutation burden (approximately 2-3*10 -4 ). This suggests that the above results are robust to any tumor with more than 1 mutation per megabase pair (>1 / Mbp).

[0053] Figure 5 A diagram of deep learning-based de-novo mutation detection and noise suppression is provided.

[0054] Figure 6 A typical pad comprising a matrix for genomic reads is shown (e.g., 16×200 base pairs for 150 base pair reads). The top panel shows a representation of the reads and their alignments as seen by the engine. The bottom panel shows the genomic context appended to the end of the read. For non-context features, padding is zero.

[0055] Figure 7The schematic diagram of the exemplary method of the present disclosure as applied in a clinical setting is shown.As shown in the figure, a biopsy sample containing cell-free DNA (cfDNA) obtained from a subject (e.g., a cancer patient or a subject suspected of having a tumor) is processed (e.g., sequenced), such as a plasma sample, to obtain the patient's genetic data (e.g., VCF file), which is classified using PILEUP (or a similar program). A VAF filter is applied to exclude germline markers (e.g., SNV, CNV, indel, or SV). Mapping quality (MQ), position filter (PIR), and / or base quality (BQ) filter are further applied to filter artificial noise. In the next step, deep learning is applied to filtered genetic data. The deep learning method involves training a machine with genetic data, which includes a summary of markers from mixed tumor biopsy samples and peripheral blood mononuclear cells (PMBC; control), which have been passed through the above-mentioned filter (e.g., by the manual reading directory of PILEUP, VAF filter to exclude germline variation, BQ filter to eliminate markers with low base quality, MQ to eliminate poorly mapped markers). The machine can also be trained with a data set. The product of the above-described systems and methods is the identification of multiple markers in cfDNA that are clinically relevant in the context of cancer diagnosis, which aids in the early diagnosis and prognosis of cancer.

[0056] FIG8 shows the properties of the dataset and the results of the engine signature analysis. Fig. 8A COSMIC signatures from Alexandrov et al. (ibid., 2013) associated with tobacco (top) and melanoma (bottom) are shown. Figure 8B Trinucleotide frequencies from sample-specific tumor and PBMC reads are shown. Specific trinucleotides associated with tobacco (purple) and UV radiation (green). Figure 8C Correlation of relative differences in trinucleotide frequencies with the average activation of the engines is shown.

[0057] FIG9 shows a line graph of various performance-related characteristics of the ENGINE of the present disclosure compared to mutation callers known in the art. Fig. 9A The sensitivity using synthetic plasma from patient CA0044 is shown. It can be seen that the ENGINE of the present invention (KITTYHAWK) outperforms mutation callers known in the art such as MUTECT, SNOOPER and / or STRELKA in terms of sensitivity. Fig. 9B A comparative line graph of the accuracy obtained using ENGINE on patient CA0044 synthetic plasma (as measured by positive predictive value or PPV) is shown. MUTECT was excluded because only two tests were performed. It can be seen that ENGINE outperforms the prior art mutation callers in terms of accuracy. Fig. 9CThe enrichment obtained using ENGINE on synthetic plasma from patient CA0044 is shown. MUTECT was excluded because only two tests were performed. It can be seen that ENGINE outperforms mutation callers known in the art in terms of enrichment.

[0058] Fig.10 Shown are the SNV detection rates in silico or in ctDNA samples obtained from control subjects (BB600; BB601) or cancer patients (BB1122 or BB1125) using the methods and systems of the present disclosure.

[0059] Fig.11 is a table showing clinical characteristics of subjects diagnosed with adenocarcinoma or benign nodules.

[0060] Figures 12A-12C Tumor-specific signatures that are differentially expressed in various tumors are shown. Fig. 12A It was shown that the application of tumor-specific signatures (UV, tobacco) in lung cancer and melanoma samples provided high specificity. Fig. 12B Differential expression of gene signatures in normal samples (PBMCs) and tumor samples from lung cancer patients (left panel) and / or melanoma patients (right panel) are shown. Fig. 12C Shown is the expression of various COSMIC signatures (and their associated zscores) in patients with breast cancer, melanoma, or lung adenocarcinoma.

[0061] Figures 13A-13C showed that cancer signatures can be detected in synthetic plasma at tumor fractions (TF) as low as approximately 1 / 1000. Fig.13A and Fig. 13B Graph representing data from two seeds, Seed 3 and Seed 4, showing that the tobacco signature can be detected in synthetic plasma down to a tumor fraction (TF) of approximately 1 / 1000. Fig. 13C Representative data from a single seed showing that the lung signature can be detected in synthetic plasma at tumor fractions (TF) as low as approximately 1 / 1000.

[0062] Figures 14A-14B The zscores for individual patient samples are shown. Fig.14A Mutation signature detection of tobacco-related signatures for lung cancer patients (blue) and patients with benign nodules (red, detected by CT) is shown in the form of zscore relative to background random signatures. This shows the ability to distinguish benign and malignant nodules based on non-invasive blood tests. Tobacco signatures (signature 4 / 8) were detected in early cancer plasma of patients exposed to tobacco, but tobacco signatures were not detected in patients with benign nodules or no smoking history. ND represents undetected samples. PY represents the pack-years of smoking for each patient. ED represents early detection. Fig. 14BThe zscore distribution of mutation signature detection in a cohort of samples obtained from subjects at various stages of lung cancer (e.g., stage IA, stage IB, stage IIA, stage IIb, and stage IIIa) compared to benign controls is shown. In most cancer samples, a baseline sensitivity of at least 67% was obtained, which increased to approximately 100% for all high-stage cases (e.g., stage IIIa and higher).

[0063] Fig.15 A schematic diagram of a computer system of the present disclosure is provided.

[0064] Figures 16A-16C Schematic diagrams of various systems of the present disclosure are provided. Shown are the various units included in a representative system.

[0065] Figure 17A-17E The use of orthogonal features (eg, fragment size) in the diagnostic methods of the present disclosure is shown, as well as the concomitant effects of applying such orthogonal features in SNV-based methods. Fig.17A The fragment size distribution displayed in healthy normal cfDNA samples is shown. Fig. 17B Shown are fragment size shifts in breast tumor cfDNA (red and purple) compared to normal cfDNA samples. Fig. 17C We show that in a mouse xenograft (PDX) model, circulating DNA from tumor origins is significantly shorter than that from normal origins. Fig.17D Shown are line graphs of fragment DNA size (x-axis; number of bases) plotted against the frequency with which fragments of that length were observed in tumor and normal samples. Fig.17E Patient-specific mutation detection using orthogonal features such as the correspondence of DNA fragments with tumor origin based on their fragment size distribution (x-axis) and GMM-joint log odds ratios (y-axis) are shown.

[0066] Figures 18A-18F The use of orthogonal features such as fragment size in the diagnostic methods of the present disclosure is shown, as well as the concomitant effects of applying such orthogonal features in CNV-based methods. Fig.18A Line graphs showing genomic region (bp) relative to cumulative plasma depth coverage skew (bottom graph), plasma relative to normal depth coverage skew (middle graph) and coverage (top graph). Fig.18B Shown is the relationship between the log2 of the depth coverage (log2>0.5=amplification, log2<-0.5=deletion) and the local fragment size centroid (COM) in that segment. Fig. 18CA dot plot of depth coverage Log2 versus fragment size centroid (COM) is shown. Using the estimated Log2 and COM values ​​for all windows in the genome, the median sample centroid (COM), slope, and R^2 of the Log2 / COM linear model are calculated at various time points (e.g., baseline day 0, day 21, and day 42). Fig.18D The correlation between the Log2 / FS estimate and the fraction of tumor DNA is shown. Fig.18E The relationship between deep coverage-based CNV detection and fragment size centroid (COM)-based CNV detection in patient samples is shown. Fig.18F A lack of relationship between CNV detection based on deep coverage and CNV detection based on fragment size centroid (COM) in normal (healthy) plasma samples was shown.

[0067] Detailed description

[0068] The present disclosure will now be described in more detail with reference to the accompanying drawings, in which preferred embodiments of the present disclosure are shown. However, the present disclosure may be embodied in different forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0069] Unless otherwise defined, the scientific and technical terms used in conjunction with the present teachings described herein shall have the meanings commonly understood by those of ordinary skill in the art. The terms used in the description of the disclosure herein are only for the purpose of describing a specific embodiment and are not intended to limit the disclosure. In addition, unless the context otherwise requires, singular terms shall include plural, and plural terms shall include singular. Generally, the nomenclature used in conjunction with the techniques of molecular biology, protein and oligonucleotide or polynucleotide chemistry and hybridization described herein and the above-mentioned techniques are well known and commonly used in the art. Standard techniques are used, for example, techniques for nucleic acid purification and preparation, chemical analysis, recombinant nucleic acid and oligonucleotide synthesis. The instructions of the manufacturer of the enzymatic reaction and purification technique are carried out, or carried out as is usually done in the art or carried out as described herein. The techniques and procedures described herein are generally performed according to conventional methods known in the art, and as described in various general and more specific references cited and discussed throughout this specification. See, for example, Sambrook et al., Molecular Cloning: A Laboratory Manual (3rd edition, Cold Spring Harbor Laboratory Press, Cold Spring Harbor, NY 2000). The nomenclature used in connection with, and the laboratory procedures and techniques described herein, are those well known and commonly used in the art.

[0070] Various embodiments of the present disclosure are described in further detail in the following paragraphs.

[0071] definition

[0072] As used in the specification and appended claims of the present disclosure, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. Also as used herein, "and / or" refers to and encompasses any one and all possible combinations of one or more of the associated listed items, as well as the lack of a combination when interpreted in an alternative manner ("or").

[0073] The word "about" means a range of plus or minus 10% of the value, for example, "about 5" means 4.5 to 5.5, "about 100" means 90 to 100, and so on, unless the context of the present disclosure indicates otherwise or is inconsistent with this interpretation. For example, in a numerical list such as "about 49, about 50, about 55", "about 50" refers to a range extending to less than half of the interval between the previous value and the subsequent value, for example, greater than 49.5 to less than 52.5. In addition, the phrase "less than about" a numerical value or "greater than about" a numerical value should be understood based on the definition of the term "about" provided herein.

[0074] Where a range of values ​​is provided in the present disclosure, it is intended that each intervening value between the upper and lower limits of the range and any other stated value or intervening value within the range is included in the present disclosure. For example, if the stated range is 1 μM to 8 μM, it is intended that 2 μM, 3 μM, 4 μM, 5 μM, 6 μM and 7 μM are also explicitly disclosed.

[0075] As used herein, the term "plurality" may be 2, 3, 4, 5, 6, 7, 8, 9, 10 or more.

[0076] As used herein, the term "screening" has a broad meaning. It includes processes intended to diagnose or determine the susceptibility, predisposition, risk, or risk assessment of an asymptomatic subject to develop a condition later in life. Screening also includes the prognosis of a subject, i.e., when a subject has been diagnosed with a condition, predetermining the progression of the condition and evaluating the efficacy of therapeutic options for treating the condition.

[0077] As used herein, the term "detection" refers to the process of determining a value or a set of values ​​associated with a sample by measuring one or more parameters in the sample, and may also include comparing the test sample to a reference sample. According to the present disclosure, the detection of a tumor includes identifying, determining, measuring and / or quantifying one or more markers.

[0078] As used herein, the term "diagnosis" refers to a method that can determine whether a subject is likely to have a given disease or condition, including but not limited to diseases or conditions characterized by genetic variation. Technicians usually diagnose based on one or more diagnostic indicators (e.g., markers), the presence, absence, quantity, or change in quantity of which indicates the presence, severity, or absence of a disease or condition. Other diagnostic indicators may include patient history; physical symptoms, such as unexplained weight loss, fever, fatigue, pain, or skin abnormalities; phenotype; genotype; or environmental or genetic factors. Technicians will understand that the term "diagnosis" refers to the increased likelihood of a particular course of disease or outcome; that is, a course of disease or outcome is more likely to occur in a patient exhibiting a given feature (e.g., the presence or level of a diagnostic indicator) compared to an individual who does not exhibit the feature. The diagnostic method disclosed herein can be used alone or in combination with other diagnostic methods to determine whether a course of disease or outcome is more likely to occur in a patient exhibiting a given feature.

[0079] As used herein, the term "early detection" of a disease (e.g., cancer) refers to the possibility of discovering disease manifestations before metastasis, for example in the context of cancer. Preferably, early detection refers to identifying a disease before morphological changes in tissues or cells are observed. In addition, the term "early detection" of cell transformation refers to the high probability that a cell will undergo transformation at an early stage before it is morphologically designated as transformed.

[0080] As used herein, the term "cell transformation" refers to a change in cell characteristics from one form to another, such as from normal to abnormal, from non-tumor to tumor, from undifferentiated to differentiated, from homogeneous to heterogeneous. In addition, transformation can be identified by the morphology, phenotype, biochemical properties (e.g., growth properties, apoptosis properties, detachment, invasion properties, etc.) of the cell.

[0081] As used herein, the term "tumor" includes any cell or tissue that may have undergone transformation at the genetic, cellular, or physiological level compared to a normal or wild-type cell. The term generally refers to a neoplastic growth, which may be benign (e.g., a tumor that does not form metastases and destroys adjacent normal tissue) or malignant / cancerous (e.g., a tumor that invades surrounding tissue and is often capable of producing metastases, which may recur after attempted removal and is likely to result in death of the host unless appropriately treated). See Steadman's Medical Dictionary, 28 th Ed Williams & Wilkins, Baltimore, MD (2005).

[0082] The term "cancer" (used interchangeably with "tumor") refers to human cancers and carcinomas, sarcomas, adenocarcinomas, lymphomas, leukemias, solid cancers and lymphomas, etc. Examples of different types of cancer include, but are not limited to, lung cancer, pancreatic cancer, breast cancer, stomach cancer, bladder cancer, oral cancer, ovarian cancer, thyroid cancer, prostate cancer, uterine cancer, testicular cancer, neuroblastoma, squamous cell carcinoma of the head, neck, cervix and vagina, multiple myeloma, soft tissue and osteogenic sarcomas, colorectal cancer, liver cancer, kidney cancer (e.g., RCC), pleural cancer, cervical cancer, anal cancer, bile duct cancer, gastrointestinal carcinoids, esophageal cancer, gallbladder cancer, small intestine cancer, central nervous system cancer, skin cancer, choriocarcinoma, osteogenic sarcoma, fibrosarcoma, glioma, melanoma, etc. In some embodiments, "liquid" cancers, such as blood cancers, such as lymphomas and / or leukemias, are excluded.

[0083] Exemplary cancers include, but are not limited to, adrenocortical carcinoma, AIDS-related cancer, AIDS-related lymphoma, anal cancer, anorectal cancer, anal canal cancer, appendix cancer, childhood cerebellar astrocytoma, childhood brain astrocytoma, basal cell carcinoma, skin cancer (non-melanoma), biliary tract cancer, extrahepatic bile duct cancer, intrahepatic bile duct cancer, bladder cancer, urinary bladder cancer, cancer), bone and joint cancer, osteosarcoma and malignant fibrous histiocytoma, brain cancer, brain tumor, brain stem glioma, cerebellar astrocytoma, brain astrocytoma / malignant glioma, ependymoma, medulloblastoma, supratentorial primitive neuroectodermal tumor, visual pathway and hypothalamic glioma, breast cancer, bronchial adenoma / carcinoid, carcinoid, gastrointestinal tract, nervous system cancer, nervous system lymphoma, central nervous system cancer, central nervous system lymphoma, cervical cancer, childhood cancer, chronic lymphocytic leukemia, chronic myelocytic leukemia, chronic myeloproliferative disease, colon cancer, colorectal cancer, cutaneous T-cell lymphoma, lymphoid neoplasms, mycosis fungoides, Seziary syndrome, endometrial cancer, esophageal cancer, extracranial germ cell tumor tumor), extragonadal germ cell tumor, extrahepatic bile duct cancer, eye cancer, intraocular melanoma, retinoblastoma, gallbladder cancer, gastric cancer, gastrointestinal carcinoid, gastrointestinal stromal tumor (GIST), germ cell tumor, ovarian germ cell tumor, gestational trophoblastic tumor, glioma, head and neck cancer, hepatocellular (liver) cancer, Hodgkin lymphoma, hypopharyngeal cancer, intraocular melanoma, eye cancer, islet cell tumor (endocrine pancreas), Kaposi's sarcoma, kidney cancer, renal cancer cancer), laryngeal cancer, acute lymphocytic leukemia, acute myeloid leukemia, chronic lymphocytic leukemia, chronic myeloid leukemia, hairy cell leukemia, lip and oral cancer, liver cancer, lung cancer, non-small cell lung cancer, small cell lung cancer, AIDS-related lymphoma, non-Hodgkin lymphoma, primary central nervous system lymphoma, Waldenstram macroglobulinemia, medulloblastoma, melanoma, intraocular (eye) melanoma, Merkel cell carcinoma, malignant mesothelioma, mesothelioma, metastatic squamous neck cancer, oral cancer, tongue cancer, multiple endocrine neoplasia syndrome, mycosis fungoides, myelodysplastic syndrome, myelodysplastic / myeloproliferative disorders, chronic myeloid leukemia, acute myeloid leukemia, multiple myeloma, chronic myeloproliferative disorders, nasopharyngeal cancer, neuroblastoma, oral cancer, oral cancercavitycancer), oropharyngeal cancer, ovarian cancer, ovarian epithelial cancer, ovarian low-grade malignant potential tumor, pancreatic cancer, islet cell pancreatic cancer, sinus and nasal cavity cancer, parathyroid cancer, penile cancer, pharyngeal cancer, pheochromocytoma, pinealoma and supratentorial primitive neuroectodermal tumor, pituitary tumor, plasma cell neoplasms / multiple myeloma, pleuropulmonary blastoma, prostate cancer, rectal cancer, renal pelvis and ureter cancer, transitional cell carcinoma, retinoblastoma, rhabdomyosarcoma, salivary gland cancer, Ewing's sarcoma family, Kaposi's Sarcoma, uterine cancer, uterine sarcoma, skin cancer (non-melanoma), skin cancer (melanoma), Merkel cell skin cancer, small intestine cancer, soft tissue sarcoma, squamous cell carcinoma, gastric cancer, supratentorial primitive neuroectodermal tumor, testicular cancer, laryngeal cancer, thymoma, thymoma and thymic carcinoma, thyroid cancer, transitional cell carcinoma of the renal pelvis and ureter and other urinary organs, gestational trophoblastic tumor, urethral cancer, endometrial uterine cancer, uterine sarcoma, uterine body cancer, vaginal cancer, vulvar cancer and Wilm's Tumor.

[0084] As used herein, "high somatic mutation rate" refers to tumors with about 1, about 2, about 3, about 5, about 7, about 10, about 12, about 15, about 20, about 25, about 30, about 40, about 50, about 60, about 75, about 80, about 100, about 125, about 150 or more mutations per megabase pair of the genome (mutations / MBP). See Collisson et al., Nature, 511(7511):543-50, 2014.

[0085] As used herein, the term "non-small cell lung cancer" or NSCLC as used herein refers to all lung cancers that are not small cell lung cancers, and includes several subtypes, including but not limited to large cell carcinoma, squamous cell carcinoma, and adenocarcinoma. All stages and metastases are included. Squamous cell carcinomas account for 25% of lung cancers and usually begin near the central bronchus. Hollow cavities and associated necrosis are usually found in the center of the tumor. Well-differentiated squamous cell carcinomas usually grow slower than other types of cancer. Adenocarcinomas account for 40% of non-small cell lung cancers. It usually originates in the surrounding lung tissue. Most cases of adenocarcinoma are related to smoking; however, in people who have never smoked, adenocarcinoma is the most common form of lung cancer. See, Rosell et al., Lung Cancer, 46(2), 135-48, 2004; Coate et al., Lancet Oncol, 10, 1001-10, 2009.

[0086] As used herein, the term "cell" is used interchangeably with the term "biological cell". Non-limiting examples of biological cells include eukaryotic cells, plant cells, animal cells (e.g., mammalian cells, reptile cells, avian cells, fish cells, etc.), prokaryotic cells, bacterial cells, fungal cells, protozoan cells, etc., cells isolated from tissues (e.g., muscle, cartilage, fat, skin, liver, lung, neural tissue, etc.), immune cells (e.g., T cells, B cells, natural killer cells, macrophages, etc.), embryos (e.g., fertilized eggs), oocytes, egg cells, sperm cells, hybridomas, cultured cells, cells from cell lines, cancer cells, infected cells, transfected and / or transformed cells, reporter cells, etc. Mammalian cells can be, for example, from humans, mice, rats, horses, goats, sheep, cattle, primates, etc.

[0087] As used herein, the term "subject" refers to mammals, including humans, veterinary or farm animals, livestock or pets, and animals commonly used in clinical research. In particular, the subject is a human subject, such as a human patient diagnosed with a tumor or suspected of having a tumor.

[0088] As used herein, the term "subject-specific data set" refers to a variety of information unique to each individual, such as genomic information, phenotypic information, biochemical information, metabolic information, microbiome sequence information, electronic medical record data, electronic health record data, drug prescriptions, biometric data, nutritional information, exercise information, family medical history information (e.g., can be obtained through a family health history survey), in-application written chat records, subject's personal healthcare provider records and notes, subject's insurance provider, patient advocate (advocate) network information, social network information, etc. In some embodiments, one or more subject-specific data sets are routinely updated and / or supplemented. In some embodiments, one or more data sets are added to multiple subject-specific data sets.

[0089] The term “subject-specific genomic information” refers to the genetic makeup of an individual, including mutations (SNPs, Del / Dup, VUS, etc.) and mutation frequencies, familial genomic sequence information, structural genomic information (including mutations (sequence, deletion, insertion)), single nucleotide polymorphisms, personal immunogenomic information (i.e., the use of whole-genome approaches to study immune system regulation and response to pathogens), functional genomic information (functional genomics focuses on dynamic aspects such as gene transcription, translation, and protein-protein interactions), computational genomic information (the use of computational and statistical analyses to decipher, discover, or predict biological mechanisms from genomic sequences and related data), epigenomics (reversible modifications of DNA or histones that affect gene expression without changing DNA sequence (e.g., DNA methylation and histone modifications), etiological information (including personal genome-microbe interactions involved in disease states), regenomic information, behavioral genomic information, metagenomics (i.e., personal genetic material recovered directly from environmental samples).

[0090] The term "subject-specific phenotypic information" refers to sex, race, height, weight, hair color, eye color, heart rate, taste preferences, blood pressure, self-described medical symptoms, medically diagnosed symptoms, test results and / or diagnoses provided by medical professionals, proteomic profile, etc. The term "subject-specific biochemical information" refers to the results of clinical tests (e.g., sodium, magnesium, potassium, iron, blood urea nitrogen (BUN), uric acid, etc.), drug / agent levels in tissues / blood, etc.

[0091] The terms "subject-specific electronic medical record data" (EMR), "electronic health record" (EHR), and "personal health record" (PHR) refer to medical and clinical data from individual healthcare providers, clinics, hospitals, nursing facilities, subject health history, subject susceptibility conditions, subject medical history, diagnoses, medications / prescriptions, treatment plans, immunization dates, allergies, radiological images, laboratory and test results, advance directives, biopsies, data from home and mobile monitoring devices (e.g., FITBIT, iWatch, Withings scales, wireless blood pressure cuffs, etc.).

[0092] As used herein, the term "sample" refers to a composition obtained from or derived from a subject of interest, comprising, for example, cells and / or other molecular entities characterized and / or identified based on physical, biochemical, chemical and / or physiological properties. The source of a tissue sample can be blood or any blood component; body fluid; solid tissue from fresh, frozen and / or preserved organ or tissue samples or biopsies or aspirations; and cells or plasma at any time during the pregnancy or development of the subject. Samples include, but are not limited to, primary or cultured cells or cell lines, cell supernatants, cell lysates, platelets, serum, plasma, vitreous humor, eye fluid, lymph, synovial fluid, follicular fluid, semen, amniotic fluid, milk, whole blood, urine, cerebrospinal fluid (CSF), saliva, sputum, tears, sweat, mucus, tumor lysates and tissue culture media, as well as tissue extracts (e.g., homogenized tissues), tumor tissues and cell extracts. Samples also include biological samples that have been processed in any manner after being obtained, such as by reagent treatment, solubilization or enrichment of certain components (such as proteins or nucleic acids), or embedded in a semi-solid or solid matrix for sectioning purposes, such as thin sections of tissues or cells in histological samples. Preferably, the sample is obtained from blood or blood components, including, for example, whole blood, plasma, serum, lymph, etc.

[0093] As used herein, the term "marker" refers to a characteristic that can be objectively measured as an indicator of a normal biological process, a pathogenic process, or a pharmacological response to a therapeutic intervention (e.g., treatment with an anticancer drug). Representative types of markers include, for example, molecular changes in the structure (e.g., sequence) or quantity of a marker, including, for example, gene mutations, gene duplications, or multiple differences, such as somatic changes in cfDNA, copy number variations, tandem duplications, or combinations thereof.

[0094] As used herein, the term "genetic marker" refers to a DNA sequence with a specific position on a chromosome, which can be measured in a laboratory. The term "genetic marker" can also be used to refer to, for example, cDNA and / or mRNA encoded by a genomic sequence, and the genomic sequence itself. Genetic markers may include two or more alleles or variants. Genetic markers can be direct (for example, located in a target gene or locus (for example, a candidate gene)), indirect (for example, due to proximity but not in a target gene or locus, and closely connected to a target gene or locus). Moreover, genetic markers may also be unrelated to genes or loci (such as SNV, CNV or tandem repeats) present in the non-coding segments of the genome. Genetic markers include nucleic acid sequences that encode or do not encode gene products (such as proteins). In particular, genetic markers include single nucleotide polymorphisms / variations (SNP / SNV) or copy number variations (CNV) or combinations thereof. Preferably, compared with a reference sample, genetic markers include somatic variations in DNA, such as sSNV or sCNV or combinations thereof.

[0095] As used herein, the term "cell-free DNA" or "cfDNA" refers to cell-free deoxyribonucleic acid (DNA) strands extracted or isolated, for example, from plasma / serum of circulating blood, from lymph, cerebrospinal fluid (CSF), urine or other body fluids. The term "cfDNA" is distinct from "circulating tumor DNA" or "ctDNA". Cell-free DNA (cfDNA) is a broader term that describes DNA that circulates freely in the bloodstream but is not necessarily from a tumor.

[0096] As used herein, the term "single nucleotide polymorphism" or "single nucleotide variation" ("SNP" or "SNV") with respect to a mutation refers to a difference of at least one nucleotide in a sequence compared to another sequence. The term "copy number variation" or "CNV" refers to a comparative numerical change in the presence or absence / gain or loss of a gene fragment having the same nucleotide sequence.

[0097] As used herein and generally in the art, the term "indel" refers to a position on a genome where one or more bases are present in one allele but not in the other allele. From an evolutionary point of view, insertions or deletions are different, but in analytical procedures such as those described herein, they are not usually distinguished because an insertion in one allele is equivalent to a deletion in another allele. Therefore, the term indel refers to the position of an insertion / deletion between two alleles.

[0098] "Structural variants" involve changes in certain parts of chromosomes, rather than changes in the number of chromosomes or a group of chromosomes in the genome. There are four common types of mutations that lead to structural variants: deletions and insertions, such as duplications (respectively involving changes in the amount of DNA in the chromosome, loss and gain of genetic material), inversions (involving changes in the arrangement of chromosome fragments) and translocations (involving changes in the position of chromosome fragments, which can cause gene fusions). In the present invention, the term "structural variant" includes loss of genetic material, gain of genetic material, translocation, gene fusion and combinations thereof.

[0099] As used herein, the term "germline DNA" or "gDNA" refers to DNA isolated or extracted from a patient's peripheral mononuclear blood cells, including lymphocytes obtained from circulating blood.

[0100] As used herein, the term "variation" refers to a change or deviation. With respect to nucleic acids, variation refers to differences or changes between DNA nucleotide sequences, including copy number differences (CNVs). This actual difference in nucleotides between DNA sequences can be a change in SNP and / or DNA sequence, for example, a change observed when the sequence is compared with a reference (e.g., germline DNA (gDNA) or a reference human genome HG38 sequence), such as fusions, deletions, additions, duplications, etc. Preferably, variation refers to the difference between a cfDNA sequence and a control DNA sequence that is not from a tumor cell, for example, when cfDNA is compared with a reference HG38 sequence, when cfDNA is compared with gDNA. The differences identified in gDNA and cfDNA are considered "constitutional" and can be ignored.

[0101] "Locus" (plural "loci") corresponds to an identified position in a genome, and can span a continuous sequence of a single base or multiple bases. Loci are typically identified by using an identifier value or a range of identifier values ​​relative to a reference genome and / or its chromosome; for example, the range of identifier values ​​"5100001" to "5800000" can refer to a specific position on chromosome 1 in a reference human genome. "Heterozygous locus" (also referred to as "het") is a locus in a genome in which two copies of a chromosome do not have the same sequence. These different sequences at the locus are referred to as "alleles". If a reference genome position has two alleles that differ by a single base, het can be a single nucleotide polymorphism (SNP). "het" can also be a reference genome position in which there is an insertion or deletion (collectively referred to as "indel") of one or more nucleotides or one or more tandem repeats. "Homozygous locus" is a locus in a reference or baseline genome in which two copies of a chromosome have the same allele. The "haplotype" of a chromosome refers to whether the chromosome occurs once or twice in the genome; for the genome of a cancer or other tumor cell, the chromosome haplotype may be a non-integer and / or a value greater than 2. A "region" in a genome may include one or more loci.

[0102] "Fragment" refers to a nucleic acid molecule (e.g., DNA) contained in or derived from (e.g., by amplification) a biological sample extracted from a target organism (e.g., a human). A fragment may include an entire arm of a chromosome, an entire chromosome, or a portion thereof.

[0103] "Fragment size" refers to the length of the fragment and can be expressed in any acceptable unit, such as base pairs or Daltons, etc. Representative fragments may be shorter than 200 bps; 200-500 bps; 500-1 Kb (where 1 Kb = 1000 bps); 1 Kb-10 Kb, 10 Kb-50 Kb, 50 Kb-100 Kb and more than 100 Kb in length (e.g., 1 megabase pair). Sequencing is used to determine information identifying one or more sequences (reads) of nucleotides in a fragment. Partial as well as full sequence information of a fragment can be generated. Sequence information can be determined with varying degrees of statistical reliability or confidence.

[0104] As used herein, the term "variant allele frequency" (VAF) or "variant allele fraction" refers to the fraction of one allele to the total amount of alleles in a DNA sample after genotyping. By convention, for a biallelic polymorphic variant (PV), VAF refers to the B allele frequency (BAF), which is the fraction of the B allele in the PV typing data, which can be obtained from a DNA sample by a high-throughput genotyping method (e.g., SNP array or NGS). In some embodiments, VAF is the B allele frequency. Alternatively, the A allele frequency (AAF) can also be used. The B allele frequency includes the A allele frequency information, and vice versa.

[0105] Typically, VAF values ​​are expressed using values ​​from 0 to 1, because they refer to frequencies or fractions. In principle, VAF values ​​can be expressed using multiples of these values, for example, using values ​​from 0 to 100. For example, a VAF value of 0.5, which indicates that half of the total number of alleles have a polymorphic variant allele, can be expressed as, for example, 50. In this case, a VAF value of 1 (i.e., all alleles have a particular genotype) would be expressed as 100. Typically, VAF values ​​are expressed using multiples of these values, for example, using values ​​from 0 to 100. max represents the maximum VAF value (i.e., all alleles have a specific genotype), while VAF min represents the minimum VAF value (i.e., none of the alleles have a particular genotype). Throughout this application, values ​​from 0 to 1 are used to represent VAF (particularly BAF) values, so VAF min is 0 and VAF max is 1. However, embodiments of the present invention are not limited to VAF values ​​expressed using this particular range. Detailed guidance on VAF, including "flipped" VAF, is provided in US2016 / 0210402.

[0106] As used herein, "reading" refers to a set of one or more data values ​​representing one or more nucleotide bases. Reading can be generated by a sequencer and / or associated logic that has sequenced all or part of a nucleic acid fragment. "Mating (mate pair)" (also referred to as "mated read" or "paired-end read") refers to at least two readings (also referred to as "arm reads") determined from opposite ends of the same fragment. Two arm reads can be collectively referred to as pairings, wherein there is a gap between the two arm reads with respect to the fragment sequenced for the pairing. The two arm reads can be referred to as "left" arm reads and "right" arm reads, respectively; however, it should be understood that any "left" (or "right") designation is not limited to being strictly on the left (or on the right), because the position of the arm reads from the fragment can be reported with respect to various reference points, such as the direction of the observer, the directionality of the DNA chain (e.g., 5' end to 3' end, and vice versa), or the genomic coordinate system selected for the reference genome. Reads can be stored with various information, such as a unique read identifier, an identifier of a fragment, or a pair identifier for the read that is part of a pair.

[0107] As used herein, "artifact" refers to observations that are not naturally present in scientific investigations or experiments, but occur due to preparatory or research procedures. Artifacts in sequencing include, for example, artifact peaks (shaded bands) and template-related artifacts (false terminations). Artifact peaks refer to peaks visible in separation that do not correspond to fragments of the correct size terminated by the corresponding dideoxynucleotide triphosphates (ddNTPs), which are used to produce DNA chains of varying lengths for DNA sequencing in the Sanger dideoxy method. Artifact peaks can be subdivided into primer-induced artifact peaks and template-induced artifact peaks. When the primers used have an affinity for binding to other regions of the template that are not intended to be combined, the formation of DNA fragments unrelated to the target sequence will result, which will produce artifacts related to the primers. On the contrary, since DNA polymerase falls off from the template before adding ddNTPs, termination artifacts are produced. It is believed that the secondary structure of the template DNA is the cause of this false termination. DNA polymerase also has limited periodicity in terms of binding to the template, which is called processivity, and short processivity frequencies are believed to increase the number of artifacts. For example, Taq DNA polymerase has a processivity of approximately 40 base pairs and is thought to contain no artifactual peaks associated with primers. False stops can occur during Sanger chain termination, which occurs when the DNA polymerase is prevented from extending its growing chain when it encounters a ddNTP, and DNA chain extension is stopped without the inclusion of the ddNTP.

[0108] The term "allele" refers to one of two or more different nucleotide sequences occurring at a particular genetic locus.

[0109] "Allele fraction" refers to the percentage of one or more alleles of nucleic acid fragment sequencing included in a biological sample for a given locus in a genome. Except for some exceptions (e.g., Y chromosomes of human males), diploid organisms (e.g., humans) usually have two copies of each chromosome. Therefore, the locus in a genome can be homozygous (e.g., having the same allele on two chromosome copies) or heterozygous (e.g., having different alleles on two chromosome copies). Therefore, "equal allele fraction" value refers to a data value of 1.0 (e.g., 100% allele fraction for alleles at homozygous locus) or 0.5 (e.g., 50% allele fraction for alleles at heterozygous locus).

[0110] "Variable allele fraction" or "VAF" refers to a data value greater than zero but different from 0.5 and 1.0. Variable allele fraction values ​​can be used to address situations where the alleles of a given locus can be represented in nucleic acid fragments of a biological sample at fractions other than 0%, 50%, and 100%. Such situations may include, but are not limited to, heterogeneity, contamination, and aneuploidy. For example, a tumor sample (e.g., a cancer sample) may be heterogeneous due to normal / stromal tissue contamination within the sample or due to multiple different tumor populations within the same tumor sample. In another example, a tumor sample may be aneuploid, such that a chromosome (or region thereof) has a copy number different from two, resulting in an allele fraction that deviates from 50% of a het to 33% or 66% when there are three copies. Examples of variable allele fraction values ​​include, but are not limited to, values ​​in the following ranges and / or combinations of ranges: 0.005 to 0.10; 0.10 to 0.20; 0.20 to 0.30; 0.30 to 0.40; 0.40 to 0.49; 0.51 to 0.60; 0.60 to 0.70; 0.70 to 0.80; 0.80 to 0.90; 0.90 to 0.99; more generally, any value in the range between 0.005 to 0.49 and 0.51 to 0.99.

[0111] As used herein, the term "control" refers to a reference for a test sample, such as control DNA isolated from peripheral mononuclear blood cells and lymphocytes, where these cells are not cancer cells, etc. As used herein, a "reference sample" refers to a sample of tissue or cells for comparison that may or may not have cancer. Therefore, a "reference" sample thus provides a basis for comparison with another sample, such as a plasma sample containing cfDNA. In contrast, a "test sample" refers to a sample compared to a reference sample or a control sample. A reference sample does not have to be cancer-free, such as when a reference sample and a test sample are obtained from the same patient at different times.

[0112] In some embodiments, a reference sample or control may include a reference assembly. The term "reference assembly" refers to a digital nucleic acid sequence database, such as a human genome (HG38) database (assembled in December 2013) comprising a HG38 assembly sequence. The gateway can be accessed through the Homo sapiens University of California Santa Cruz (UCSC) genome browser gateway at the World Wide Web URL GENOME.UCSC.EDU. In addition, a reference assembly may also refer to the human genome assembly (version 38; assembled in June 2017) of the Genome Reference Association, which can be accessed on the Internet through the website of the US National Center for Biotechnology Information (NCBI).

[0113] As used herein, the term "sequencing" or "sequencing" as a verb refers to the process of determining the nucleotide sequence of DNA or the order of nucleotides, such as the nucleotide sequence AGTCC, etc. "Sequence" as a noun refers to the actual nucleotide sequence obtained from sequencing; for example, a DNA with the sequence AGTCC. Where "sequence" is provided and / or received in digital form, such as on a disk or remotely provided by a server, "sequencing" can refer to a collection of DNA propagated, manipulated and / or analyzed using the methods and / or systems of the present disclosure.

[0114] As used herein, "substantially" means sufficient for the intended purpose. Thus, the term "substantially" allows for minor, insignificant variations from an absolute or perfect state, dimension, measurement, result, etc., which would be expected by one of ordinary skill in the art but would not significantly affect overall performance. When used with respect to a value or parameter or characteristic that can be expressed as a numerical value, "substantially" means within ten percent.

[0115] As used herein, the term "substantially purified" refers to a cfDNA molecule that is isolated, separated or separated or extracted from its natural environment and is at least 60%, preferably 75%, more preferably 90%, and most preferably 99% free from other components with which it is naturally associated.

[0116] The term "whole genome sequencing" refers to a laboratory process that determines the DNA sequence of each DNA strand in a sample. The resulting sequence can be referred to as "raw sequencing data" or "reading". As used herein, when the sequence has similarity to a region of a reference chromosome DNA sequence, the reading is a "mappable" reading. The term "mappable" can refer to an area that shows similarity to a reference sequence and is therefore "mapped", for example, a cfDNA segment that shows similarity to a reference sequence in a database, such as a cfDNA with a high similarity percentage to the human chromosome region 8q248q24.3 in the human genome (HG38) database is a "mappable reading".

[0117] In addition to "WGS", targeted sequencing can be used to obtain a genomic summary. Compared with WGS, the term "targeted sequencing" used herein refers to a laboratory process for determining the DNA sequence of a selected DNA locus or gene in a sample, such as sequencing a selected group of cancer-related genes or markers (such as targets). In this context, the term "target sequence" refers to a selected target polynucleotide, such as a sequence present in a cfDNA molecule, and its presence, quantity and / or nucleotide sequence or changes therein need to be determined. The target sequence is queried for somatic mutations. The target polynucleotide can be a gene region associated with a disease such as cancer. In some embodiments, the region is an exon.

[0118] As used herein, the term "low abundance" for cfDNA refers to the amount of cfDNA in the sample being less than about 20 ng / mL, such as about 15 ng / mL, about 10 ng / mL or less, such as about 9 ng / mL, 8 ng / mL, 7 ng / mL, 6 ng / mL, 5 ng / mL, 4 ng / mL, 3 ng / mL, 2 ng / mL, 1 ng / mL, 0.7 ng / mL, 0.5 ng / mL, 0.3 ng / mL or less, such as 0.1 ng / mL or even 0.05 ng / mL. In some embodiments, the term "low abundance" can be understood in the context of the uniqueness of the mark, such as length or base composition. For example, although the subject's sample may contain a large amount of cfDNA (e.g., >20 ng / mL), the actual number of unique genetic markers (e.g., sSNVs) contained in the cfDNA may be very low. Typically, this parameter is expressed as genome equivalents (GE) or coverage, as described below. In some embodiments, the term "low abundance" can be understood in the context of tumor specificity of the mark. For example, although a subject's sample may contain a large amount of cfDNA (e.g., >20 ng / mL), the vast majority of genetic markers (e.g., sSNVs) contained in the cfDNA may be redundant and / or correlated with a reference (e.g., PBMC gDNA). Typically, this parameter is expressed as a tumor fraction (TF), as described below.

[0119] As used herein, the term "tumor-specific" or "tumor-associated" with respect to cfDNA refers to differences in the DNA sequence of cfDNA in a subject whose cancer forms a tumor (e.g., a lung cancer patient) when compared to a reference DNA, such as when cfDNA is compared to control DNA (gDNA) from non-tumor cells, as described herein. Alternatively, "tumor-specific" may relate to cfDNA before treatment when compared to cfDNA collected during or after treatment.

[0120] As used herein, the term "genome equivalent" or "GE" refers to the number of unique DNA fragments. In some embodiments, the sample comprises 5 to about 10000GE, preferably 100 to about 5000GE, particularly about 200 to about 2000GE, such as about 25, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200, 1400, 1600, 2000 or 5000GE. As understood in the art, a typical sample comprising about 6ng of cfDNA comprises about 1000 or less GE. Preferably, GE is greater than 1 (e.g., greater than 2, 5, 10, 15, 20, 25, 50, 100, 200, 500 or 1000). It is expected that 10-20ml of blood contains about 10,000GE. Thus, in some embodiments, a suitable sample may comprise about 20 ml, 15 ml, 10 ml, 5 ml, 4 ml, 3 ml, 2 ml, 1 ml, 0.5 ml, 0.1 ml, 0.01 ml, or 0.001 ml of plasma.

[0121] The term "coverage" or "depth of read" refers to sequencing effort. For example, a coverage of 20X indicates moderate sequencing effort, while a coverage of 35X or more indicates high sequencing effort, and a coverage of 5X indicates low sequencing effort. In embodiments of the present disclosure, coverage is typically between about 5X and about 100X, particularly between 15X and about 40X, such as 20X, 30X, 35X, 40X, 50X, 70X or greater.

[0122] As used herein, the term "mutation load" or "N" refers to the level (e.g., quantity) of changes (e.g., one or more genetic changes, particularly one or more individual cell changes) of each preselected unit (e.g., per megabase pair) in a predetermined genomic window. Mutation load can be measured, for example, based on the entire genome or exome or based on a subset of a genome or exome. In certain embodiments, the mutation load measured based on a subset of a genome or exome can be extrapolated to determine the entire genome or exome mutation load. In certain embodiments, mutation load is measured in a sample such as a tumor sample (e.g., a lung tumor sample or a sample obtained or derived from a lung tumor) from a subject such as a subject described herein. Preferably, mutation load is a measure of the number of mutations per megabase pair (1,000,000bp or MBP) of cfDNA. As known in the art, mutation load can vary according to the type of tumor, genetic pedigree, and other subject-specific features (e.g., age, sex, tobacco consumption, etc.). In the context of tumor diagnosis, the mutation load may be between about 1000 and about 10000 mutations per MBP, such as about 1000, 2000, 4000, 6000, 8000, 10000, 12000, 15000, 20000, 25000, 30000, 40000, 50000, 60000, 60000, 70000, 80000, 90000, 10000, 10000 or more, such as about 200000 per MBP. Typically, in non-smokers, the mutation load is about 8,000 per MBP, while in subjects with melanoma, the mutation load is more than 40,000 per MBP.

[0123] As used herein, the term "genomic window" refers to a DNA region within the boundaries of a selected nucleotide sequence. The windows can be separate from each other or overlap with each other.

[0124] As used herein, the term "tumor fraction" or "TF" refers to the level (e.g., quantity) of tumor DNA molecules relative to normal DNA molecules. In some embodiments, "tumor fraction" refers to the ratio of tumor DNA (ctDNA) without circulating cells to the total amount of cell-free DNA (cfDNA). It is believed that the tumor fraction indicates the size of the tumor. Typically, the tumor fraction (TF) is between about 0.001% and about 1%, for example, about 0.001%, 0.05%, 0.1%, 0.2%, 03%, 0.4%, 0.5%, 0.6%, 0.7%, 0.8%, 0.9%, 1% or more, for example 2%.

[0125] The term "abundance" can refer to binary (e.g., absent / present), qualitative (e.g., absent / low / moderate / high), or quantitative information (e.g., a value proportional to the amount, frequency, or concentration) indicating the presence of a particular molecular species. In this case, mutations present at higher relative concentrations are associated with a greater number of malignant cells, e.g., cells that are transformed earlier in the carcinogenic process than other malignant cells in the body (Welch et al., Cell, 150:264-278, 2012). Because of the higher relative abundance of these mutations, they are expected to exhibit higher diagnostic sensitivity for detecting cancer DNA than those mutations with lower relative abundance.

[0126] As used herein, "sequencing error rate" refers to the proportion of incorrectly sequenced nucleotides. For example, in the context of whole genome sequencing, a sequencing error rate of approximately 1 per 1000 bases has been reported in the literature (range: error rate per base call is approximately 0.1–1%; Wu et al., Bioinformatics, 33(15): 2322-2329, 2017).

[0127] As used herein, the term "sequencing depth" refers to the number of times a sequencing region is covered by sequence reads. For example, an average sequencing depth of 10x means that each nucleotide in the sequencing region is covered by an average of 10 sequence reads. As sequencing depth increases, the chance of detecting cancer-associated mutations is expected to increase. However, in practice, the chance of detection does not increase linearly with sequencing depth, as evidenced by the fact that even at a median depth of 42,000X, the fundamental limitation of cfDNA abundance only results in positive detection of approximately 19% of early lung adenocarcinomas (Abbosh et al., Nature, 545(7655): 446-451, 2017).

[0128] As used herein, the term "base quality" score of a given base in a sequencing read is the probability that the sequencer incorrectly calls a base. A quality score is assigned to each base in the read by a Phred-like algorithm (Ewing et al., Genome Res. 8 (3): 175-185, 1998; Ewing et al., Genome Res. 8 (3): 186-194, 1998, describing a representative method), which is similar to the method originally developed for Sanger sequencing experiments. In some embodiments, base quality (BQ) includes variable base quality (VBQ) or average reading base quality (MRBQ), both of which are variants of base quality metrics.

[0129] As used herein, the term "PCR error" refers to errors introduced during sequencing via the polymerase chain reaction (PCR) amplification step. A typical PCR error rate is 10 5There is approximately 1 error in each base pair (Barnes et al., PNAS USA, 91:2216, 1994).

[0130] As used herein, the term "mapping quality" score indicates the confidence that a particular sequence read is accurately placed relative to a reference sequence. Li et al., Genome Research, 18: 1851-1858, 2008 provides a method for determining a mapping quality score. After mapping a sequence read to a reference sequence, a mapping quality score can be provided by a mapping algorithm.

[0131] The term "read position" or "position in read (PIR)" refers to a position on a read (e.g., a marker) in a nucleotide sequence. As understood in genomics, many sequencing protocols are prone to various types of amplification-induced biases and errors, which can be reduced by implementing filters such as "read direction" and "read position" filters. The read direction filter removes variants that are almost exclusively present in forward or reverse reads. For many sequencing protocols, such variants are most likely the result of errors caused by amplification. The read position filter is implemented in a manner similar to the "read direction filter" to eliminate systematic errors, but is also applicable to hybridization-based data. It removes variants that are differently positioned in the reads that carry it compared to what would be expected from the general position of the reads covering the variant site. This is done by classifying each sequenced nucleotide (or gap) according to the mapping direction of the read and the position where the nucleotide is found in the read; each read is divided into several parts (e.g., 5 parts) along its length, and the part number of the nucleotide is recorded. For each sequenced nucleotide, a total of ten categories are given, and a given site will be distributed between these ten categories of reads covering the site. If a variant is present in the site, the variant nucleotides can be expected to follow the same distribution. Read position filters perform tests to measure the significance of a read position, for example, measuring whether the read position distribution of the variant carrying reads is different from the read position distribution of the total set of reads covering the site.

[0132] As used herein, the term "bin" refers to groups of DNA sequences grouped together, for example, in a "genomic bin." In certain cases, a bin may include groups of DNA sequences binned based on a "genomic bin window," which includes grouping DNA sequences using a genomic window.

[0133] By way of example only, and as a summary of the detailed description below, various embodiments herein relate to algorithms and software for running the diagnostic engine (engine) of the present disclosure. The engine uses a read representation that captures the genomic context of the alignment, the complete read sequence, and the integral of the quality score of each base. In contrast, the representation used in sequence analysis software known in the art treats a batch of reads as a single feature, thereby losing valuable information about the sequence alignment itself and the quality of each base associated with the read (Poplin et al., bioRxiv, pp. 092890, 2016; Torracinta & Campagne, bioRxiv, pp. 097469, 2016).

[0134] method

[0135] The systems and methods disclosed herein can be used for the diagnosis, prognosis and monitoring of various human diseases. For example, the methods and systems described herein can be used to detect a variety of cancers. Like most cells, cancer cells are characterized by a turnover rate, in which old cells die and are replaced by new cells. Typically, dead cells in contact with the vasculature of a given subject can release DNA or DNA fragments into the bloodstream. This is also true for cancer cells in various stages of the disease. Depending on the stage of the disease, cancer cells can also be characterized by various genetic abnormalities such as copy number variations and mutations. Using the methods and systems described herein, this phenomenon can be used to detect the presence or absence of cancer individuals.

[0136] According to the present disclosure, blood from a subject at risk of cancer can be extracted and prepared as described herein to produce a cell-free polynucleotide population. In one example, a colony may include cell-free DNA. The system and method disclosed herein can be used to detect markers (e.g., SNV, CNV, indel and / or SV) present in certain cancers. Although there are no symptoms or other disease characteristics, the method can still help detect the presence of cancer cells in vivo. The method disclosed herein can be applied to diagnose or prognose any type of cancer or tumor. Therefore, the cancer types that can be detected include but are not limited to blood cancer, brain cancer, lung cancer, skin cancer, nose cancer, throat cancer, liver cancer, bone cancer, lymphoma, pancreatic cancer, skin cancer, intestinal cancer, rectal cancer, thyroid cancer, bladder cancer, kidney cancer, oral cancer, gastric cancer, solid tumors. Both heterogeneous tumors and homogeneous tumors can be diagnosed or prognosed according to the present disclosure.

[0137] The system and method can be used to detect any number of genetic abnormalities that may be caused or caused by cancer. This may include but is not limited to mutation, mutation, indel, copy number variation, transversion, translocation, inversion, deletion, aneuploidy, partial aneuploidy, polyploidy, chromosome instability, chromosome structural change, gene fusion, chromosome fusion, gene truncation, gene amplification, gene duplication, chromosome damage, DNA damage, abnormal changes in nucleic acid chemical modification, abnormal changes in epigenetic patterns, nucleic acid methylation infection and abnormal changes in cancer. In addition, the system and method described herein can also be used to help characterize certain cancers. The genetic data produced by the system and method disclosed herein can allow doctors to help better characterize specific forms of cancer. Generally, the composition and staging of cancer are heterogeneous. Genetic profile data can allow characterization of specific subtypes of cancer, which may be important in the diagnosis or treatment of this specific subtype. This information can also provide clues about the prognosis of specific types of cancer for subjects or doctors. The system and method provided herein can be used to monitor known cancers or other diseases in specific subjects. This can allow subjects or doctors to adapt to treatment selection according to the progress of the disease. In this example, the systems and methods described herein can be used to construct a genetic profile of a particular subject during the course of a disease. In some cases, cancer may progress, become more aggressive and genetically unstable. In other examples, cancer may remain benign, inactive or dormant. The systems and methods of the present disclosure may be useful in determining disease progression.

[0138] In addition, the systems and methods described herein can be used to determine the efficacy of a particular treatment option. In one example, if the treatment is successful, the successful treatment option can actually increase the number of copy number variations or mutations detected in the subject's blood because more cancers may die and shed DNA. In other examples, this may not happen. In another example, perhaps certain treatment options may be correlated with the genetic profile of the cancer over time. This correlation can be used to select a therapy. In addition, if the cancer is observed to be in remission after treatment, the systems and methods described herein can be used to monitor residual disease or recurrence of the disease.

[0139] The methods and systems described herein may not be limited to detecting mutations and copy number variations associated only with cancer.Preferably, the methods and systems disclosed herein can be used for early diagnosis or early detection of cancer.

[0140] In addition, the method of the present disclosure can be used to characterize the heterogeneity of abnormal conditions in a subject, the method comprising generating a genetic profile of extracellular polynucleotides in a subject, wherein the genetic profile comprises a plurality of data from copy number variation and mutation analysis. In some cases, a disease may be heterogeneous, including but not limited to cancer. Disease cells may not be identical. In the example of cancer, it is known that some tumors include different types of tumor cells, some cells at different stages of cancer. In other examples, heterogeneity may include multiple lesions of the disease. Again, in the example of cancer, there may be multiple tumor foci, perhaps one or more of which are the result of metastasis that spreads from the primary site.

[0141] The methods disclosed herein can be used to generate or summarize (profile) fingerprints or data sets, which are the sum of genetic information obtained from different cells in heterogeneous diseases. The data set may include one or a combination of copy number variation and mutation analysis. In addition, the systems and methods disclosed herein can be used to diagnose, prognose, monitor or observe cancer or other diseases of fetal origin. In other words, these methods can be used for pregnant subjects to diagnose, prognose, monitor or observe cancer or other diseases of unborn subjects whose DNA and other polynucleotides can circulate together with maternal molecules.

[0142] The foregoing diagnostic methods can be used in conjunction with other common diagnostic procedures, such as review of health history, physical examination, laboratory tests (blood, urine, etc.), biopsy imaging tests (such as X-ray, PET / CT, MRI, ultrasound, etc.), nuclear medicine scans (e.g., bone scans), endoscopy, family history, etc.

[0143] Preferably, the diagnostic methods of the present disclosure improve the predictive prognostic value (PPV) of common diagnostic procedures (eg, CT scans) by at least 20%, at least 30%, at least 40% or more, for example, at least 50%.

[0144] In Figure 1 of the accompanying drawings, Figure 2 and Figure 7 A representative, non-limiting schematic diagram of a diagnostic method is provided in .

[0145] Workflow

[0146] Figure 1A1 is a flow chart illustrating a method 100 for diagnosing a tumor disease, such as an early-stage tumor disease, according to various embodiments of the present disclosure. The method 100 is illustrative only, and embodiments may use variations of the method 100. The method 100 may include the following steps: receiving a summary of markers; filtering noise associated with the markers according to a plurality of features; applying a convolutional neural network that has been trained with a computer data set and / or a patient data set to adaptively and systematically filter the noise; removing artificial noise markers from the summary to generate subject-specific markers that are statistically matched with the data set to generate confidence intervals; and diagnosing a disease based on the confidence intervals.

[0147] exist Figure 1A In step 110 of method 100, a genetic marker summary is received from a subject. In some embodiments, the summary of the genetic marker is received in a variant call format (VCF) file. As understood in the art, VCF files are used to store gene sequence variations in bioinformatics. The VCF format is developed with the emergence of large-scale genotyping and DNA sequencing projects (e.g., 1000 genome projects). Alternatively, a summary can be provided in a general feature format (GFF) containing all genetic data. Typically, GFF provides redundant functionality because they are shared between genomes. On the contrary, using VCF, only variation needs to be stored with a reference genome. In some embodiments, (e.g., using whole genome sequencing (WGS)) the sample of the subject is sequenced, and sequence files are processed, for example, using tools such as, for example, genome VCF (gVCF).

[0148] exist Figure 1A In step 120 of method 100, the probability of detecting noise (P N) to filter out artificial reads by statistically classifying each read as signal or noise, with a probability that is a function of (1) base quality (BQ); (2) mapping quality (MQ); (3) estimated fragment size and / or (4) estimated variant allele fraction (VAF). Other parameters may also be used, such as (5) position in the read (RP); (6) sequence context (SC); (7) abundance; (8) sequencing depth and / or (9) sequencing errors. The noise removal step 120 may include performing an optimal receiver operating characteristic (ROC) curve based on the joint base quality (BQ) and mapping quality (MQ) scores, which includes a probabilistic classification of genetic markers in the profile. Typically, the joint BQMQ score is provided in the form of a matrix (x, y), where x is the BQ score and y is the MQ score. In an exemplary embodiment, a joint BQMQ score between 10 and 50 (for each parameter) is typically used, for example, BQMQ scores are (10, 40), (15, 30), (20, 20), (20, 30), (30, 40). In some embodiments, classification of the marker includes measuring the area under the ROC curve (AUC), which typically represents the probability that a randomly selected candidate marker among potential markers will show a value higher than a randomly extracted control marker. For completely non-informative markers, the ROC curve will approach an ascending diagonal line (called the "chance diagonal" or "probability line"), and the AUC will tend to 0.5, which is the expected probability of classification due to chance alone. In contrast, in the case of perfect classification, the ROC curve will reach the point of highest theoretical accuracy (sensitivity and specificity are both 100%), and the AUC will tend to 1, the highest probability value. In Figure 3B Representative ROCs are provided in . The error models before and after filtering are shown in Figure 3A and Figure 3C middle.

[0149] Optionally, the genetic markers are weighted. In some embodiments where the markers are SNVs or CNVs, a weighting step is performed to distinguish between true markers (e.g., mutations that may be associated with a disease) and common variations (e.g., random somatic SNPs that are not associated with a disease). In some embodiments, the weighting step is based on a probability score (P D ) weights the labels. Preferably, the weighting step 120 includes weighting the labels based on the Bernoulli equation P D =1-[(1-TF)] GE To measure the probability of detection (P D ), where P Dis the probability of detection, TF is the tumor fraction, and GE is the number of genomic equivalents present in the patient's DNA. Implementing a weighting step is advantageous because it helps overcome sequencing depth by increasing the number of detection sites (SNVs) by increasing the width, which is generated by repeating the Bernoulli trial for each SNV (binomial distribution of Bernoulli trial probabilities). In some embodiments, the weighting step can additionally include binning the markers based on, for example, an increased weight or range of weights. For example, genetic markers can be binned based on PD, where markers with high P D The markers have low P D For example, the markers can be binned based on a P value of at least about 0.60, such as at least about 0.65, 0.70, 0.75, 0.80, 0.90, 0.95 or greater, such as at least about 0.98. D The threshold bins the genetic markers. Thus, if the P D If the sample is smaller than the threshold, it can be classified as a false positive and not included in the analysis.

[0150] exist Figure 1A In step 130 of method 100, the sequencing noise of each read in the summary is filtered by using a machine learning (ML) method to distinguish between cancer-related mutational signatures and signatures associated with PCR or sequencing errors. In some embodiments, the diagnostic method of the present disclosure can use a neural network to systematically remove or reduce noise. Although it may be advantageous to implement the neural network after removing artificial markers according to step 120 above, the neural network can be applied at any step of the method. In this regard, in Figure 1A In the purely exemplary method 100 of , a deep convolutional neural network (CNN) is optionally applied at step 130 to adaptively and / or systematically filter the sequencing noise present in the filtered dataset. Preferably, the CNN includes: using a deep learning algorithm in a pan-tumor cohort to identify signatures that distinguish true tumor mutations from artifacts; assigning confidence estimates to each individual mutation detected in a tumor patient sample; integrating the confidence estimates across the entire genome; and performing rigorous analysis of specific cosmic mutation signatures in the sample, such as using non-negative least squares (NNLS) for each marker.

[0151] In some embodiments, a CNN is trained with an in silico dataset. For example, a computer dataset may include synthetic plasma samples obtained from actual cancer patients, such as a cohort of breast cancer or lung cancer patients. The accuracy, sensitivity, and / or precision of a CNN may be evaluated according to the methods described below. For example, sensitivity may be determined as a ratio [TP / (TP+FN)], where TP is a true positive and FN is a false negative; precision may be determined as a ratio [TP / (TP+FP)], where TP is a true positive and FP is a false positive; specificity may be determined as a ratio [TN / (TN+FN)], where TN is a true negative and FN is a false negative. Under a representative validation method, the accuracy of a CNN may be evaluated based on an average F1 score. For example, the F1 score may be calculated as 2x[(accuracy x recall (Recall)) / (accuracy + recall (Recall))]]. In some embodiments, on a tumor control, a CNN may achieve an F1 score of at least about 0.5, about 0.6, about 0.7, about 0.8, or even about 0.9 or higher, such as 0.95.

[0152] In some embodiments, a CNN can be trained with an in silico patient-specific dataset comprising tumor and normal WGS reads mixed in different proportions at different tumor fractions (0.00001, 0.00005, 0.0001, 0.0005, 0.001, 0.005, 0.01) and coverages (5, 10, 15, 20, 35). Replication and / or randomization seed can further be used to increase variability in the training dataset.

[0153] The architecture of CNN will be discussed in more detail below.

[0154] exist Figure 1A In step 140 of method 100, a subject-specific signature comprising a plurality of real reads in the compendium is compiled by removing artificial noise (see step 120) and / or sequencing noise (see step 130). Although not being bound by any particular theory, in some embodiments, the deleting step filters out "noise" markers with low base quality and / or mapping quality from the marker compendium initially identified as strongly associated with the disease. In some embodiments, the deleting step may include obtaining a probability of detection that satisfies a threshold (P) based on step 120. N ), classify the marker as signal or noise based on the ROC curve; if the marker is classified as noise, delete the marker from the summary. Optionally, include, for example, the probability of detection (P D ) and noise probability (P N ) ratio can be used to remove tags that do not meet a preset threshold score.

[0155] exist Figure 1A In step 150 of method 100, a match is made between the subject-specific signature and the cancer signature, and a confidence estimate of the subject's biological sample including circulating tumor DNA (ctDNA) is quantified. This can be achieved using, for example, a probability density function (PDF) estimate and / or a zscore estimate, both of which are explained in detail below.

[0156] In the estimation of the confidence interval, a weighting step may be optionally employed. For example, all markers that have been classified as true positives based on the noise removal step 120 and the noise filtering step 130 may be weighted identically. For example, based on the scores assigned to markers in the pan-tumor network, an improved weighting system may be used. Diagnosis may also include the use of a threshold score, such as a score obtained by performing the same noise removal step 120 and noise filtering step 130 in a test marker (e.g., a marker known to be associated with a tumor). For example, such a test marker may include unique SNVs and / or CNVs in a cancer patient sample that are not present in a control (non-tumor) subject.

[0157] like Figure 1B As further provided by the illustrated example workflow 100, according to various embodiments, a method for genetic screening of a subject for cancer is provided. As provided in step 110, the method may include receiving a subject-specific whole genome read summary associated with a plurality of genetic markers from a biological sample of the subject. The biological sample may include a tumor sample. The read summaries may each include a read of a single base pair length.

[0158] As in Figure 1B As provided in step 120 of method 100, the method may include filtering artificial sites from the read summary. Filtering may include removing duplicate sites generated on a reference healthy sample cohort from the read summary. Alternatively or in combination, filtering may include identifying germline mutations in the biological sample and / or identifying shared mutations between peripheral blood mononuclear cells of a tumor sample and a normal cell sample as germline mutations, and removing the germline mutations from the read summary.

[0159] As in Figure 1BThe method 100 of step 130 provided in, the method may include filtering noise from the whole genome reading summary using at least one error suppression scheme to generate a filtered reading set of the whole genome reading summary. The at least one error suppression scheme may include calculating the probability that any single nucleotide variation in the summary is an artificial mutation, and removing the mutation. The probability can be calculated as a function of the feature selected from the group including the following: mapping quality (MQ), variant base quality (MBQ), position in reading (PIR), average reading base quality (MRBQ) and combinations thereof. Alternatively or in combination, the at least one error suppression scheme may include: using inconsistent tests between independent repetitions of the same DNA fragment produced by polymerase chain reaction or sequencing, and / or repeat consistency to remove artificial mutations, wherein, when lacking consistency between the majority of a given repeat family, identifying and removing artificial mutations.

[0160] As in Figure 1B As provided in step 140 of method 100, the method may include compiling a subject-specific signature using the filtered set of reads based on comparison to a specific mutation signature associated with a predetermined mutagenesis process.

[0161] As in Figure 1B As provided in step 150 of method 100, the method may include statistically quantifying a confidence estimate that the subject's biological sample includes a cancer-associated mutation signature based on a comparison of the cancer-associated mutation signature exposure value to a background mutation signature cohort through a subject-specific signature.

[0162] As in Figure 1B As provided in step 160 of method 100 , the method may include screening the subject for cancer if a confidence estimate that the biological sample of the subject contains a cancer-associated mutation signature exceeds a given threshold.

[0163] like Figure 1C The example workflow 100 shown in further provides a method for genetic screening of cancer in a subject. As provided in step 110, the method may include receiving a subject-specific whole genome read summary associated with multiple genetic markers from a biological sample of the subject. The biological sample may include a tumor sample. The read summary may each include a copy number variation (CNV).

[0164] As in Figure 1C As provided in step 120 of method 100, the method may include dividing the reading summary into a plurality of windows.

[0165] As in Figure 1CAs provided in step 130 of method 100, the method may include calculating a set of features for each window. These features may include the median depth coverage of each window and the representative fragment size of each window.

[0166] As in Figure 1C As provided in step 140 of method 100, the method may include filtering artificial sites from the read summary. Filtering may include removing duplicate sites generated on a reference healthy sample cohort from the read summary.

[0167] As in Figure 1C As provided in step 150 of method 100, the method may include normalizing the read summary to produce a filtered set of reads for the whole genome read summary.

[0168] As in Figure 1C As provided in step 160 of method 100 of , the method may include calculating an estimated tumor score using the filtered set of reads by calculating a linear relationship between the feature set for each window and converting the calculated relationship to an estimated tumor score using a regression model. Alternatively or in combination, the method may include calculating an estimated tumor score based on one or more integrated mathematical models as a function of the feature set for each window calculated in the subject-specific whole genome read summary.

[0169] As in Figure 1C As provided in step 170 of method 100, the method may include screening the subject for cancer if the estimated tumor score exceeds an empirical threshold.

[0170] Exemplary workflow for implementing a marker type-based screening approach

[0171] Figure 1D and Figure 1E A schematic workflow for practicing the method of the present disclosure is shown. Figure 1D An overview of the workflow commonly used when the marker of interest contains SNVs / indels; Figure 1E A workflow that is typically used when the marker of interest includes CNV / CV is outlined. It should be noted that although separate workflows are provided for illustrative purposes, they do not necessarily have to be performed separately to implement the methods of the present disclosure. For example, certain features / elements of the workflows can be used in combination to generate an output (e.g., an estimated tumor score based on a combination of SNV / indel and CNV / SV) that is associated with a target outcome (e.g., whether a subject has or is likely to develop cancer).

[0172] SNV-based cancer screening

[0173] The present disclosure provides systems, methods and algorithms for cancer screening based on detecting SNV / indel markers in biological samples of subjects. Figure 1B As shown, cancer diagnosis based on SNV / indel markers generally utilizes the following steps: receiving genetic data; detecting mutations (e.g., single mismatches); removing / filtering artificial sites; suppressing errors using algorithms including machine learning; correcting reads; detecting cancer based on one or more mathematical models; and, optionally, orthogonally integrating analysis of secondary features in the genomic data (e.g., fragment size shift) to improve the sensitivity, specificity, and / or reliability of detection.

[0174] exist Figure 1D In the first step, genetic data is received from a biological sample (usually a plasma sample). Next, sensitive mutation calls are performed on the plasma sample using PILEUP (or other single supported read callers). Germline SNPs are detected by using the GATK germline caller on the plasma sample or using mutation calls on matched peripheral blood mononuclear cells (PBMCs). Oral swabs can also be used instead of PBMCs. Serially or in parallel, duplicate artificial sites are generated in a healthy plasma sample cohort (normal panel (PON) blacklist or mask) and deleted from the detected mutations to remove common sequencing or alignment artifacts.

[0175] Next, a highly sensitive method capable of detecting single mutant fragments was adopted. This step includes one or more error suppression steps. In the first error suppression step, a filtering scheme is used to analyze on the basis of a single reading, and the probability that the reading represents an artificial mutation is quantified. In some embodiments, a multidimensional classification framework using a support vector machine (SVM) classification with a linear kernel can be implemented in this step. Compared with the low variant allele fraction (VAF) sequencing artifact in normal PBMC samples, the classification framework is trained on germline SNPs. Here, the classification decision boundary is defined on a multidimensional space including variant base quality (VBQ), mapping quality (MQ), position in reading (PIR) and / or average reading base quality (MRBQ). In order to evaluate the classification scheme, the validation metric of the SVM classification scheme is compared with the random forest after 10-fold cross validation under the same protocol. The SVM classification shows a higher classification performance, moderately better than the random forest model. Among all patients, the average sensitivity of SVM was 90.7% and the specificity was 83.9% (N=10 samples, F1=87.7%, PPV=84.9%).

[0176] In the second error suppression step, artificial mutations generated by PCR or sequencing are corrected by comparing independent repeats of the same original DNA fragment. In cfDNA samples, paired-end 150bp sequencing is usually used, which results in overlapping paired reads (R1 and R2 sequences overlap) given the short length of typical cfDNA fragments (~165bp). Therefore, any inconsistency between R1 and R2 pairs is considered a potential sequencing artifact, which is corrected back to the corresponding reference genome. In addition, recognizing the potential for independent duplications to occur in any DNA molecule that may be copied multiple times in sequencing and PCR, repeat families are identified by 5' and 3' similarities and alignment positions. Each repeat family is then used to check the consistency of specific mutations in independent repeats, thereby correcting artificial mutations that do not show consistency in most repeat families.

[0177] Using one or more identification steps, the resulting set of confident de-novo plasma mutations is used to identify a tumor signature. A first approach involves identifying a tumor signature in a result set using a mutation signature inference method (e.g., a non-negative least squares (NNLS) method). The method outputs a confidence score (e.g., a zscore) that can be used to determine whether a subject has cancer. In this regard, a threshold confidence score (e.g., a zscore of about 2) can be used to make a reliable determination that a subject has cancer. A second approach can be employed that utilizes a deep learning approach to detect mutation signatures. The method outputs a tumor proportion score (e.g., eTF) that can be used to determine whether a subject has cancer. These methods will be described in more detail below:

[0178] Cancer-specific mutagenic signatures

[0179] Cancer mutagenesis is dominated by sequence context-specific signatures associated with different mutagenic processes (e.g., smoking, UV light, etc.). These mutation signatures are unique to cancer tissues and do not appear in normal PBMC samples. Here, gene signatures are differentially expressed in lung cancer patients (exposed to tobacco) and melanoma patients (exposed to UV light) compared to normal samples (PBMC). Recognizing this feature, a new analytical method for sensitive detection was developed. The method is based on a non-negative least squares (NNLS) model that utilizes specific mutation signatures in a single plasma sample. The confidence of the signature detection was further verified by comparing the cancer-specific mutation signature exposure values ​​with the exposure values ​​inferred for 100 random background signatures and setting a confidence threshold of z-score>2std.

[0180] Deep learning mutation signature detection

[0181] To further suppress artificial sequencing errors and improve ctDNA sensitivity, a machine learning approach was developed to distinguish between sequencing reads altered by cancer and reads altered by sequencing errors, allowing for an adaptive and specific filter for systematic sequencing noise. A deep convolutional neural network (CNN) based on artificial intelligence techniques was applied. CNN allows a large number of features to be learned and integrated for classification problems in a supervised manner. This approach is based on a rethinking of the challenge of mutation calling, as it aims to distinguish reads containing true variants from reads containing sequencing artifacts. This allows the CNN to be trained on millions of reads and errors of true mutations using a large amount of tumor and normal WGS data, achieving very high sensitivity and specificity for a variety of patients and tumor types.

[0182] Implementation of the above features in deep CNN training resulted in the independent capture of sequence context features known to occur in lung cancer and melanoma. First, to apply CNNs into an early detection (ED) framework, a CNN algorithm was trained on a pan-lung cancer cohort (5 deep tumor and PBMC WGS patients) using supervised learning to identify signatures that distinguish true tumor mutations from artifacts. The resulting model was used to infer each individual mutation detected in our ED plasma samples from early-stage lung adenocarcinoma patients and assign them confidence estimates that can be integrated into an estimate of the proportion of tumor reads in a given sample. The model was able to identify specific tobacco and / or UV signatures that, when used on patient samples, enabled the detection of early-stage patients with each cancer with high accuracy.

[0183] In addition, the method was applied to plasma samples from 21 patients with early lung disease and 12 patients with benign nodules detected by CT to evaluate the ability of the method to improve the low positive predictive value (PPV) of current lung cancer CT screening in people at risk of tobacco exposure. The results showed that there were 14 positive detections in early lung cancer samples and 3 positive detection results in benign nodules, showing an improved PPV of 80% compared to the 40%-50% PPV in the current CT-based screening program. These data indicate a significant improvement over existing methods for early detection of lung cancer and melanoma patients.

[0184] Integration of Orthogonal Characteristics

[0185] Optionally, the above basic workflow can orthogonally integrate the minor features contained in the genetic data into the final analysis model. For example, in order to improve the robustness, accuracy and / or sensitivity / specificity of the detection method, the features based on the readings (such as the offset of the DNA fragment size) can be orthogonally integrated into the mathematical model. The significance of the orthogonal feature integral (in cancer detection) can be calculated using a probabilistic mixture model (e.g., a Gaussian mixture model). See the corresponding data in the Examples section and Figures 17 and 18.

[0186] CNV-based cancer screening

[0187] Alternatively or additionally, the present disclosure provides systems, methods and algorithms for cancer screening based on detecting CNV / SV markers in biological samples of subjects. Figure 1E As shown, cancer detection based on CNV / SV markers generally utilizes the following steps: receiving genetic data; extracting window-based feature vectors in the genetic data; filtering artificial CNV windows; normalizing the filtered genetic data using one or more normalization steps; detecting tumors after feature vector partitioning; and, optionally, orthogonally integrating analysis of secondary features in the genomic data (e.g., analysis of fragment size offsets) to improve the sensitivity, specificity and / or reliability of detection.

[0188] exist Figure 1E In the first step, receive the genetic data from biological sample (usually plasma sample). Next, extract the feature vector based on window from genetic data. For example, extract the depth coverage feature (represented by Log2) and / or the fragment size feature (represented by COM). Use the estimated Log2 and COM values ​​of all windows in genome to calculate the median sample centroid (median COM on neutral region), the slope and R^2 of Log2 / COM linear model. In addition, split reading can also be extracted. When a part of NGS reading is mapped to a position of genome and another part of the same reading is mapped to different positions of genome, split reading is usually produced.

[0189] Next, windows with low mappability and / or coverage are filtered. Continuously or in parallel, duplicate artificial sites are generated in a healthy plasma sample cohort (a normal panel (PON) blacklist or mask) and removed from the window to filter the artificial window. The filtered high confidence reference CNV / SV segments are normalized. Typically, the normalization step includes guanine-cytosine (GC) normalization and / or zscore normalization.

[0190] Next, one or more mathematical models are used to partition the feature vector. In some embodiments, a hidden Markov model (HMM) is used. In some embodiments, a self-organizing neural network (SONN) based on a mathematical model (e.g., adaptive resonance theory (ART) or self-organizing map (SOM)) is used. By using the data of one or more analysis partitions in these mathematical models, copy number variation (CNV) detection and cancer diagnosis are performed.

[0191] Here, it is also possible to orthogonally integrate secondary features of the genetic data in the final analysis model. For example, to improve the robustness, accuracy and / or sensitivity / specificity of the detection method. Summarizing the Log2 / COM correlation (R^2), the Log2 / COM slope and the sample median fragment size centroid (COM), a classification model can be defined to classify tumor samples from healthy samples and calculate estimated TF, for example, using a generalized linear model (GLM).

[0192] The significance of orthogonal feature integrals (in cancer detection) can be calculated using a probabilistic mixture model (eg, a Gaussian mixture model).

[0193] It should be understood that, with some modifications, the workflows disclosed herein may also be broadly used to detect residual disease during or after chemotherapy, immunotherapy, targeted therapy, or a combination thereof, and / or in monitoring the efficacy of such treatments.

[0194] Use of the aforementioned method for early tumor diagnosis

[0195] The method disclosed herein is particularly useful in early diagnosis of tumors. Preferably, the diagnostic method of the present invention is performed non-invasively. The diagnostic method can be performed before surgery or treatment of the tumor.

[0196] The methods of the present disclosure can be performed even at low tumor fractions (TF). Typically, in samples with lower TFs, the prior art is unable to accurately and reliably diagnose tumor diseases due to low detection probabilities. In contrast, the methods of the present disclosure allow for detection of markers and accurate diagnosis of tumor diseases at low tumor fractions, such as 1 / 1000 or 1 / 10,000 or even 1 / 20,000. The sensitivity of the methods and systems of the present disclosure is particularly demonstrated by the fact that even at very low tumor fractions (e.g., 1 / 10,000 or lower), the methods of the present disclosure can detect approximately 10 to 15 sSNVs contained in a single supporting read. This detection enables the distinction between normal samples and tumor samples with a high level of fidelity and accuracy that the prior art cannot provide. It should be understood that diagnosis is not limited to sSNV detection. For example, diagnosis can be made based on detection of about 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 or more, e.g., 150, 200 or 250 copy altered segments (genome-wide) that are frequently found in human cancers.

[0197] The present disclosure is particularly related to a method for early diagnosis of a tumor characterized by a high somatic mutation rate. Preferred types of tumors that can be diagnosed or detected according to the present disclosure include, for example, non-small cell lung cancer (NSCLC), tobacco-induced cancer (TIC), ultraviolet-induced cancer, apolipoprotein B mRNA editing enzyme catalytic protein (APOBEC) activity-mediated cancer, cancers comprising breast cancer protein (BRCA) mutations and / or cancers comprising poly (ADP-ribose) polymerase (PARP) high activity, tumors comprising microsatellite instability (MSI). The method may be suitable for diagnosing liquid tumors, solid tumors, or mixtures thereof, such as heterogeneous tumors, including, for example, lymphomas that have metastasized to extralymphatic organs such as liver, lungs, or brain.

[0198] According to the present disclosure, the following tumors can be diagnosed particularly early: lung adenocarcinoma, ductal adenocarcinoma (breast tumor), skin melanoma, urothelial carcinoma (bladder tumor) or osteosarcoma. In particular, the tumors include non-small cell lung cancer adenocarcinoma (NSCLCLUAD).

[0199] The present disclosure is particularly related to early diagnosis or detection of non-small cell lung cancer, which is preferably tobacco-induced lung cancer characterized by a high somatic mutation rate. Tobacco use (e.g., smoking or chewing) is a recognized risk factor or causative agent for epithelial cancers of the oral cavity, pharynx, larynx, esophagus, lung, stomach, cervix, and colon / rectum. See Sasco et al., Lung Cancer 45, Suppl 2, S3-9, 2004.

[0200] The present disclosure also relates to the diagnosis or detection of UV-induced cancers, such as skin cancer. Exposure to ultraviolet (UV) radiation is associated with approximately 65% ​​of melanoma cases and 90% of non-melanoma skin cancers (NMSC), including basal cell carcinoma (BCC) and squamous cell carcinoma (SCC). See, Kim et al., Genes & Disease, 1(2): 188-198, 2014. Preferably, the UV-induced cancer is selected from melanoma and SCC, both of which are characterized by a high somatic mutation rate. See Alexandrov et al., Curr Opin Genet Dev. 24, 52-60, 2014.

[0201] The present disclosure also relates to early diagnosis of cancers with high somatic mutation rates due to perturbations of enzymes associated with gene editing / DNA checkpoints. In some embodiments, the present disclosure relates to diagnosis of cancers mediated by gene editing enzymes such as apolipoprotein B mRNA editing enzyme catalytic protein (APOBEC). APOBEC-mediated mutation patterns are common in bladder cancer, cervical cancer, breast cancer, head and neck cancer, and lung cancer. See Roberts et al., Nat Genet., 45(9): 970-6, 2013.

[0202] In some embodiments, the present disclosure relates to early diagnosis of cancers mediated by breast cancer protein (BRCA) mutations, for example, cancers mediated by BRCA1 mutations or BRCA2 mutations or a combination thereof. Reports estimate that more than 50% of women with BRCA1 mutations will develop breast cancer by the age of 70, and more than 1 / 3 of them will develop ovarian cancer at that age. In addition to breast and ovarian cancer, BRCA2 mutations are also involved in the risk of male breast and pancreatic cancers and melanoma. Both BRCA1 / 2 mutations are associated with male prostate cancer risk. See Ngeow et al., npj Genomic Medicine 1, 15006, 2016.

[0203] In some embodiments, the present disclosure relates to early diagnosis of cancers induced by microsatellite instability (MSI). MSI-induced cancers are usually caused by mutations in DNA mismatch repair genes (e.g., MLH1, MSH2, or MSH6) and are characterized by errors in repeated sequences. MSI can occur in tumors of many organs, but it is primarily a hallmark of colorectal cancer. Kurzawski et al., Annals of Oncology, 15 (Supp. 4), 283–284, 2004. MSI has also been found in endometrial cancer, ovarian cancer, gastric cancer, sebaceous carcinoma, glioblastoma, lymphoma / leukemia, and Lynch syndrome tumors (hereditary non-polyposis CRC (HNPCC)). Vilar et al., Nat Rev Clin Oncol., 7 (3): 153-62, 2010.

[0204] In some embodiments, the present disclosure relates to early diagnosis of cancers induced by PPAR activity, such as cancers mediated by the compensatory homologous recombination mechanism of PARP. For example, tumors with defects in certain homologous recombination mechanisms may rely on PARP-mediated DNA repair for survival and be sensitive to PARP inhibition. Therefore, PARP inhibition is a potential synthetic lethality therapeutic strategy for treating cancers with specific DNA repair defects, such as those arising in BRCA1 or BRCA2 mutation carriers (Morales et al., Crit Rev Eukaryot Gene Expr., 24(1): 15-18, 2014; Fong et al., N Engl J Med., 361(2): 123-34, 2009).

[0205] Diagnostic methods disclosed herein relate to first receiving a subject sample comprising multiple genetic markers. In some embodiments, the subject sample comprising DNA / RNA is sequenced, and the genetic markers therein are received for analysis. In other embodiments, genetic markers can be received from a data set, such as genome sequencing information compiled and / or stored in a computer or remotely (e.g., in a server). Genetic markers can be received by sequencing various samples. Preferably, the sample includes a biological sample, such as a cell, a tissue, an organ, including a biological fluid, such as blood, blood plasma, lymph, etc. Or, the sample includes a primary or metastatic tumor.

[0206] Can use a variety of methods to obtain samples. Tissue biopsy is usually used to obtain representative tumor tissue. Tumor cells can also be obtained indirectly in the form of tissue or fluid known or thought to contain target tumor cells. For example, biological samples of lung cancer lesions can be obtained by excision, bronchoscopy, fine needle puncture, bronchial brushing or from sputum, pleural fluid or blood. Tumors that have metastasized can be sampled from nearby tissues or lymph nodes (primary metastasis) or from a distant part of the body (distant metastasis).

[0207] Preferably, the sample includes a plasma sample including circulating DNA and peripheral blood mononuclear cells (PMBC). In this case, conventional techniques can be used to obtain samples from subjects, such as blood drawing (phlebotomy), biopsy (including liquid biopsy), surgical resection, tracheal swabs, expectoration, etc. The sample thus obtained can be optionally processed, for example, to purify and / or separate markers useful in diagnosis. The presence of cfDNA in the sample can be checked using conventional methods, for example, using PCR with universal primers, followed by electrophoresis. The cfDNA in the subject's sample can be purified using conventional techniques, such as the DNA separation kit described in the "Examples" section of the present disclosure.

[0208] In some embodiments, the sample comprises a biological fluid selected from blood, cerebrospinal fluid, pleural fluid, ocular fluid, urine, or a combination thereof.

[0209] In some embodiments, samples containing somatic mutations in cfDNA are obtained using liquid biopsy technology (LBT), which is a transformative, non-invasive technology that allows the detection of tumor DNA in patient plasma cfDNA samples and allows the characterization of somatic malignant genomes.

[0210] In a specific embodiment, biological sample is a plasma sample comprising cell-free DNA (cfDNA). Typically, the amount of cfDNA in the sample is between about 0.1ng / ml and about 20.0ng / ml; preferably, between about 1ng / ml and about 10ng / ml. Normal cell samples comprising peripheral mononuclear blood cells (PMBC) can be used as controls. In two samples, genetic markers including single nucleotide variations (SNV) (preferably somatic SNV), copy number variations (CNV) (preferably somatic CNV), short insertions and deletions (indel), structural variants (SV) or combinations thereof can be analyzed.

[0211] In some embodiments, genetic markers include a combination of SNV and CNV. Such combinations are typically used in samples comprising low SNV mutation load but high CNV load. In an exemplary embodiment, a sample comprising a SNV mutation load of less than 8000 mutations per megabase pair (MBP) can be analyzed by additionally detecting CNV. Typically, in this case, it is desirable that the CNV load per megabase pair (MBP) of DNA is at least 50, such as at least 60, at least 70, at least 80, at least 90, at least 100 or more, such as 200CNV / MBP, because it may have diagnostic significance.

[0212] As known in the field of genetics, the significance of variations such as SNV or CNV is profoundly affected by the differences between germline and somatic cells. Mutations in somatic cells are not transferred to offspring. Mutations occurring in somatic cells (e.g., lungs) may damage cells, cancerate cells, or even kill cells. However, if there is mutated DNA in the germline of gametes, the mutated DNA can only be passed to the next generation. Therefore, comparing germline sequences (e.g., using them as controls) allows identification of changes in somatic cells or cancer cells specific to the subject, which do not exist in non-cancerous cells of the same subject. Although the comparison between germline sequences and gamete sequences indicates the presence of variations, the comparison between cancer cells and non-cancerous cells is also useful. For example, peripheral leukocytes or lymphocytes of the subject can be used as controls, which represent non-cancerous somatic cell sequences. In this way, mutations found in both cancer cells and non-cancerous cells can be ignored.

[0213] Preferably, the genetic markers of the present disclosure, such as sSNVs, sCNVs, indels or SVs in cfDNA, can be detected by comparing the cfDNA sequence to a reference sequence, such as a germline DNA sequence.

[0214] In some embodiments, the methods of the present disclosure may include detecting variations between genetic markers and reference (e.g., control) sequences. In some embodiments, the variations may be uniform, semi-uniform, or dynamic between samples. Temporally dynamic variations include, for example, differences between cfDNA collected during or after treatment compared to pre-treatment samples.

[0215] Variations in cfDNA can also be detected by generating a genome-wide summary of genetic markers and subtracting the genetic markers present in a control (e.g., germline) sample. In this article, the term "whole genome" refers to and includes the genetic material of the germline and somatic cells of an organism. The marker summary may include, for example, multiple sSNVs, sCNVs, indels, SVs in DNA, including other variations, such as fusions, etc.

[0216] Typically, the sample is characterized by a low tumor fraction (TF). In some embodiments, the TF is about 0.0001% to about 1%, preferably about 0.001% to about 0.1%, in particular less than 0.1%, such as 0.005%, 0.02%, 0.03%, 0.04%, 0.05%, 0.06%, 0.07%, 0.08%, 0.09%.

[0217] Additionally, the sample comprising cfDNA is characterized by a genomic equivalent (e.g., the number of unique DNA fragments, as determined by randomly sampling the entire pool of cfDNA fragments in the subject's sample) between about 100 and about 20,000; preferably between about 1000 and about 10000.

[0218] In some embodiments, the cfDNA sample is characterized by a mutation load (N) of about 3,000 to about 100,000; preferably about 5000 to about 40000.

[0219] Representative methods for generating a whole genome summary may include sequencing. Typically, sequencing is performed using a purified nucleic acid sample. In particular, whole genome sequencing is used to obtain a whole genome summary used in the diagnostic method and / or system disclosed herein. For example, WGS can be performed using conventional techniques and may include amplification (e.g., PCR amplification). Amplification-free sequencing may also be used using methods and reagents known in the art. See, Karlsson et al., Genomics, 105(3): 150-8, 2015. Purely as an exemplary method, in some embodiments, genetic markers in cfDNA can be detected by the following steps: whole genome sequencing (WGS) of a tumor of a subject, whole genome sequencing (WGS) of normal cells of a subject, mixing tumor and normal WGS reads in different proportions to generate a data set of subject-specific samples with different tumor scores and coverages, and downsampling the data set to generate a complementary data set of downsampled normal reads without mixing in reads from the tumor. The complementary data set can be filtered by removing noise-related markers, as described below.

[0220] Whole-genome summaries of genetic markers can also be generated by targeted sequencing (TS) or a combination of WGS and TS.

[0221] The following publications related to whole genome sequencing and / or targeted sequencing are incorporated herein by reference in their entirety: U.S. Patent Nos. 7,115,400; 7,718,403; 7,741,463; 8,932,812; 7,572,584 and 9,218,450.

[0222] Once the DNA sample is received, the diagnostic method can be implemented. The genetic markers contained in the sample are preferably analyzed for mutations, such as somatic mutations. The most common type of somatic mutation in DNA is single nucleotide variation (SNV), which occurs at a frequency of 1-100 / Mbp (megabase pair). These variations are usually identified in shotgun sequencing data by carefully comparing DNA sequencing reads that map to specific loci in cancer samples and germline normal DNA samples (controls). This complex process has been developed using increasingly sophisticated techniques / tools that refine the statistical comparison between the number of supporting reads with the variant in cancer samples and germline samples. See, Cibulskis et al., Nature Biotechnology, 31(3):213–219, 2013; Saunders et al., Bioinformatics, 28(14):1811-1817, 2012; Wilm et al., Nucleic acids research, 40(22):11189–11201, 2012.

[0223] Analysis of variants can be performed using a variety of techniques, including but not limited to array-based methods (e.g., DNA microarrays, etc.), real-time / digital / quantitative PCR instrument methods, and whole or targeted nucleic acid sequencing systems (e.g., whole genome sequencing (WGS) services provided by Illumina, Helicos Biosciences, Pacific Biosciences, Complete Genomics, Sequenom, ION Torrent Systems, and Halcyon Molecular).

[0224] Preferably, whole genome sequencing (WGS) is used to analyze somatic mutations and / or copy number variations of genetic markers. Whole genome sequencing methods can parse gene readings with single base resolution. In the case of DNA (deoxyribonucleic acid), these methods can parse readings at the level of the basic components of DNA (e.g., A (adenine), T (thymine), C (cytosine), and G (guanine). In the case of RNA (ribonucleic acid), these methods can parse readings at the level of the basic components of DNA (e.g., A, U (uracil), G, and C).

[0225] The products of the aforementioned sequencer / sequencing method include "sequencing data", "sequencing information" or "sequencing reads", which include information about the order of one or more of the aforementioned bases in a polynucleotide molecule (e.g., a whole genome, a whole transcriptome, an exome, an oligonucleotide, a polynucleotide, a fragment, etc.). By comparing the reading sequence of DNA in a sample (e.g., cfDNA contained in a patient's plasma sample) with a control (e.g., a whole genome sequence of PMBC), the target genetic marker (e.g., somatic SNV or somatic CNV) can be identified. It should be understood that the currently disclosed identification methods are applicable to all types of sequencing technologies, platforms or technologies, including but not limited to capillary electrophoresis, microarrays, ligation-based systems, polymerase-based systems, hybridization-based systems, direct or indirect nucleotide identification systems, pyrophosphate sequencing, ion-based or pH-based detection systems, electronic signature-based systems, etc.

[0226] The next step in the early diagnosis method of the present disclosure involves identifying low-abundance, tumor-specific markers.

[0227] The present disclosure relates to determining the error probability in a read based on a plurality of factors selected from the following: (1) base quality (BQ) of the read, (2) mapping quality (MQ) of the read; and / or (3) estimated read fragment size (4) estimated read allele fraction (VAF), which individually or collectively affect signal quality. Other auxiliary parameters may also be employed, such as (5) position in the read (PIR); (6) sequence context (SC); (7) abundance; (8) sequencing depth and / or (9) sequencing errors.

[0228] Typically, base quality (BQ) relates to the confidence of sequencing quality at each base, while mapping quality (MQ) score relates to the confidence estimate about the accuracy of mapping the marker to the genome. In the case of sSNV markers, base quality (BQ) score is a measure of the quality of identification of the nucleobases produced by automatic DNA sequencing. Conventional methods can be used to determine, for example, Phred quality score, which is assigned to each nucleotide base call in the automatic sequencer trace (trace). Phred quality score (Q) is defined as a property related to the logarithm of base call error probability (P). For example, if Phred is a base assigned quality score of 30, then the chance of miscalling the base is one in a thousand. Typically, the BQ of sequencing reads is between 10 and 50, such as a BQ score of 10, 15, 20, 25, 30, 35 or 40.

[0229] Also in the case of sSNV markers, the mapping quality (MQ) score is a measure of the confidence that the read actually comes from the position aligned by the mapping algorithm. It can be determined using conventional methods, such as mapping quality scores (see, Li et al., Genome Research 18: 1851-8, 2008). Typically, the MQ of a read is between 10 and 50, such as an MQ score of about 10, 15, 20, 25, 30, 35 or 40.

[0230] In some embodiments, the noise removal step includes performing an optimal receiver operating characteristic (ROC) curve based on a joint base quality (BQ) and mapping quality (MQ) score, wherein the curve includes a probability classification of the genetic markers in the summary. Typically, the joint BQMQ score is provided in the form of a matrix (x, y), where x is the BQ score and y is the MQ score. In an exemplary embodiment, a joint BQMQ score between 10 and 50 (for each parameter) is typically used, for example, a BQMQ score of (10, 40), (15, 30), (20, 20), (20, 30), (30, 40).

[0231] The noise removal step may include performing additional filters. For example, additional information sources contained in the read pairs from the DNA fragments can be used to determine the source chain (Watson or Crick) and estimate the size of the DNA fragments. It has been observed that ctDNA has a different fragment size distribution compared to conventional circulating healthy DNA (Underhill et al., PLoS Genetics, 12 (7): e1006162, 2016). More specifically, the fragment lengths from cell-free DNA between tumor patients and healthy controls found that mutant alleles more commonly appear at fragment lengths shorter than the fragment lengths of wild-type alleles. Similarly, selecting shorter cell-free DNA fragment lengths substantially increases the frequency of mutant alleles in human lung cancer (Jiang et al., PNAS USA, 112.11, E1317-E1325, 2015; Mouliere et al., bioRxiv, 134437, 2017; Underhill, supra). Therefore, a specific subset of fragment lengths from cell-free DNA detection can be used to improve ctDNA detection. In some embodiments, the fragment size of the reads is preferably less than 160 bp, such as 160 bp, 140 bp, 120 bp, 100 bp, 75 bp, 50 bp or less, such as 20 bp.

[0232] In addition, artificial noise can be eliminated based on variable allele frequency (VAF). In some embodiments, low allele group fraction mutation sites are removed from the sample, for example, about 1% or less VAF. In some embodiments, only markers (e.g., SNVs) with threshold VAF are retained for downstream analysis. For example, mutation sites with VAF of at least 1%, at least 2%, at least 3%, at least 4%, at least 5% (as determined by amplicon sequencing on a PGM instrument) can be retained. As known in the art, the VAF value of a specific allele (e.g., BRAFV600R) is not static, and may change over time (due to the development and / or progression of cancer), and also due to therapy (e.g., immunotherapy, chemotherapy or targeted therapy). However, a threshold VAF less than 1% (e.g., VAF of 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8) can be reliably used to estimate that a specific allele is not related to a tumor.

[0233] In a specific embodiment, artificial noise is removed by performing one or more, preferably all of the following steps: (a) removing reads with low mapping quality (e.g., <29, ROC optimization); (b) establishing duplicate families (e.g., representing multiple PCR / sequencing copies of the same DNA fragment) and generating corrected reads based on a consistency test; (c) removing reads with low base quality (e.g., <21, ROC optimization); and / or (d) removing reads with high fragment size (e.g., >160, ROC optimization).

[0234] In addition to using the aforementioned BQ / MQ, VAF and fragment size filters, other factors such as read position (RP or PIR) can also be used to filter artificial noise, because RP affects signal quality. In the case of sSNV markers, RP can be mapped, for example, by mapping the position of the starting base of sequencing reads. Other factors affecting marker quality include, for example, specific sequence contexts associated with higher sequencing error probabilities (Chen et al., Science, 355 (6326): 752-756, 2017). In this regard, true mutations can usually be mapped to their own specific sequence contexts, while errors cannot. For example, mutations associated with tobacco tend to occur in CC contexts, while mutations associated with APOBEC enzyme activity are more inclined to the TpC contexts for inserting somatic mutations (see Greenman et al., Nature, 446 (7132): 153-158, 2007). Therefore, sequence context can be used to help identify changes that are more likely to be caused by sequencing artifacts and changes that are more likely to be caused by general mutation processes.

[0235] In some embodiments, the Bernoulli equation P D =1-〖(1-TF)〗 GE The labels are further weighted by measuring the probability of detection, where P D is the probability of detection, TF is the tumor fraction, and GE is the number of genome equivalents present in the patient’s DNA. Then based on P D The genetic markers are weighted, where the markers with high P D For example, the binning can be based on a P value of at least about 0.60, such as at least about 0.65, 0.70, 0.75, 0.80, 0.90, 0.95 or greater, such as at least about 0.98. D The threshold bins the genetic markers. Thus, if the P D If the sample is smaller than the threshold, it can be classified as a false positive and not included in the analysis.

[0236] Once the artificial noise reads are removed from the read summary, the remaining markers are input into a deep learning inference model that is trained to separate between tumor-associated signatures and PCR / sequencing error signatures. This step involves classifying cancer mutation supporting reads and artificial mutation (error) reads in a read-based manner. In one embodiment, the sequence context distribution of cancer mutation supporting reads is calculated and the contribution to the known mutation signature is classified using machine learning (ML).

[0237] Reads that have been filtered for artificial noise and / or classified as supported by cancer mutations are matched to cancer signatures. In some embodiments, a dataset containing such cancer signatures can be used (e.g., Catalog of Somatic Mutations in Cancer; COSMIC). As of February 2018, 30 different cancer signatures have been collected in this database, the details of which are summarized as follows:

[0238] Signature 1 (found in all cancer types) is the result of an endogenous mutational process initiated by spontaneous deamination of 5-methylcytosine;

[0239] Signature 2 (found in 22 cancers) has been attributed to the activity of the AID / APOBEC family. Based on the similarity in the sequence context of cytosine mutations caused by APOBEC enzymes in experimental systems, it seems likely that APOBEC1, APOBEC3A, and / or APOBEC3B play a role in human cancers compared to other members of the family;

[0240] Signature 3 (breast, ovarian, and pancreatic cancer) is associated with failure of DNA double-strand break repair by homologous recombination;

[0241] Signature 4 (head and neck cancer, liver cancer, lung adenocarcinoma, lung squamous cell carcinoma, small cell lung cancer, and esophageal cancer) is associated with smoking and has features similar to mutational patterns found in experimental systems exposed to tobacco carcinogens such as benzo[a]pyrene. Signature 4 is likely due to tobacco mutagenesis.

[0242] Signature 5 (unknown etiology) has been found in all cancers and in most cancer samples;

[0243] Signature 6 (found in 17 cancer types, most frequently in colorectal and uterine cancers) is associated with defective DNA mismatch repair and is found in microsatellite unstable tumors;

[0244] Signature 7 (skin and lip cancers; head and neck or oral squamous cell carcinomas) is associated with UV exposure;

[0245] Signature 8 (found in breast cancer and medulloblastoma) has an unknown cause;

[0246] Signature 9 (found in CLL and malignant B-cell lymphomas) is attributed to polymerase η, which is associated with AID activity during somatic hypermutation;

[0247] Signature 10 (found in six cancer types, particularly colorectal and uterine cancer) is due to altered activity of the error-prone polymerase POLE. Recurrent POLE somatic mutations, Pro286Arg and Val411Leu, were mostly associated with signature 10 mutations.

[0248] Signature 11 (found in melanoma and glioblastoma) exhibits a mutation pattern similar to that of alkylating agents;

[0249] Signature 12 (found in liver cancer) has an unknown cause;

[0250] Signature 13 (found in 22 cancer types and appears to be most common in cervical and bladder cancers) is attributed to the activity of the AID / APOBEC family of cytosine deaminases, which convert cytosine to uracil;

[0251] Signature 14 (etiology unknown) has been found in four uterine cancers and a single adult low-grade glioma sample;

[0252] Signature 15 (found in several gastric cancers and a single small-cell lung cancer) is associated with defective DNA mismatch repair;

[0253] Signature 16 (found in liver cancer) has an unknown cause;

[0254] Signature 17 (found in esophageal cancer, breast cancer, liver cancer, lung adenocarcinoma, B-cell lymphoma, gastric cancer, and melanoma) has an unknown cause;

[0255] Signature 18 (found in neuroblastoma and also observed in breast and gastric cancer) has unknown etiology;

[0256] Signature 19 (found in pilocytic astrocytomas) has an unknown etiology;

[0257] Signature 20 (found in gastric and breast cancer) is associated with defective DNA mismatch repair;

[0258] Signature 21 (found in gastric cancer) has an unknown cause;

[0259] Signature 22 (found in urothelial (renal pelvis) cancer and liver cancer) is associated with exposure to aristolochic acid;

[0260] Signature 23 (found in liver cancer) has an unknown cause;

[0261] Signature 24 (found in a subset of liver cancers) was associated with exposure to aflatoxins;

[0262] Signature 25 (found in Hodgkin lymphoma) has an unknown cause;

[0263] Signature 26 (found in breast, cervical, gastric, and uterine cancers) is associated with DNA mismatch repair;

[0264] Signature 27 (found in a subset of clear cell renal carcinomas) has unknown etiology;

[0265] Signature 28 (found in gastric cancer) has an unknown cause;

[0266] Signature 29 (found in gingival-buccal oral squamous cell carcinoma) is associated with people who chew tobacco;

[0267] Signature 30 (found in a small subset of breast cancers) has an unknown etiology.

[0268] In some embodiments, the matching step involves a linear hybrid optimization (e.g., zscore confidence estimates of the contribution of tobacco exposure or BRCA mutations or APBEC1 activity) for calculating a confidence measure of the contribution of the COSMIC mutation signature. Purely as a representative, non-limiting example, the linear optimization problem can be solved using the algebraic function min‖Ax-b‖, x≥0, where A is the mutation signature sequence context matrix, x is the contribution (variable) of each cosmic mutation signature, and b is the patient-specific sequence context summary.

[0269] In some embodiments, in the linear optimization method used above, A may include any number of cosmic signatures, including random mutation signatures. For example, A may include about 20, 30, 40, 50 or more, such as 70 cosmic signatures and about 50, 60, 80, 100 or more, such as 150 random mutation signatures. The contribution distribution of random signatures is calculated using an extraction method, such as E_random, which calculates the average contribution score; and std_random, which calculates the standard contribution score. The confidence associated with the contribution of each cosmic signature can be statistically calculated, for example, using a zscore. For example, the Z score can be calculated as (cosmic_sig_contribution–E_random) / std_random. Therefore, like the permutation score (permutationscore), the Z score represents the significance of the signature contribution compared to the random set.

[0270] In some embodiments, a statistical method such as a probability density function (PDF) is used to calculate the similarity of the patient sequence-context summary to a specific cosmic signature. Purely as a representative example, to calculate the PDF, the patient sequence-context summary is normalized to generate a density function. The cosine similarity between the patient sequence-context density function and the cosmic signature density function is calculated. The cosine similarity is then normalized by dividing by the cosine similarity between the patient sequence-context density function and the non-informative uniform density function.

[0271] exist Figure 1A In step 160 of method 100, the confidence estimate calculated in step 150 is used to screen the subject for cancer, e.g., for early detection of a tumor. As is known in the art, a confidence interval consists of a range of values ​​(interval) that serves as a good estimate of an unknown population parameter (e.g., the likelihood that an asymptomatic patient has cancer). The desired confidence level is set by the researcher (rather than determined by the data). Most commonly, a 95% confidence level is used; however, other confidence levels may be used, such as any value between 80% and 99%, e.g., 80%, 90%, 98%, or even 99%.

[0272] In some embodiments, the confidence interval can be single (e.g., based on a single reading) or composite (e.g., based on multiple readings). Confidence bands or confidence regions can also be used. Confidence regions generalize the confidence interval concept to handle multiple quantities and can be used to reveal the degree of possible sampling error and / or the unreliability of quantities used in statistical analysis. Confidence bands can be used to represent uncertainty in the estimation of a curve or function based on limited or noisy data; prediction bands can be used to represent uncertainty (affected by noise) about the value of a new data point on a curve.

[0273] Optionally, the calculated confidence for the contribution of the COSMIC mutation signature can be checked against a detection threshold. In some embodiments, the threshold is defined by an empirically measured baseline noise detection estimate from healthy samples, for example, a zscore of at least 2 standard deviations (STDs) above the threshold, particularly at least 3STDs above the threshold, preferably at least 4STDs above the threshold, and especially at least 5STDs above the threshold, indicates that the confidence metric is significant.

[0274] Purely by way of example, in an exemplary embodiment, the method of the present disclosure involves diagnosing a tumor in a subject by the following steps: first receiving a plurality of genetic markers sequenced from a biological sample of the subject (e.g., a sample including a plasma sample and a normal cell sample) to generate a subject-specific whole-genome gene read summary comprising markers (e.g., sSNVs, CNVs, indels and / or SVs); filtering artificial noise from the read summary using one or more parameters selected from BQ, MQ, position in read (PIR), fragment size and / or VAF; inputting the noise-removed reads into a neural network that can distinguish true markers from noise generated by PCR and / or sequencing errors; generating a filtered, noise-removed subject-specific signature that matches a cancer signature (e.g., a COSMIC signature), wherein the matching comprises calculating zscores for all markers or a subset thereof, or estimating a probability density function between the subject signature and a reference cancer signature; outputting a confidence interval indicating that the subject's signature includes a tumor signature, thereby diagnosing a tumor in the subject. Figure 1A A representative method is shown in the flowchart of . Details of this method are provided in the example below.

[0275] In some embodiments, a cancer signature may include patterns associated with tissue-specific epigenetic patterns, such as tissue-specific chromatin accessibility patterns (eg, methylation status).

[0276] In some embodiments, the diagnostic method can further employ karyotyping. For example, a data set comprising tumor-specific, low-abundance markers can be further subjected to karyotyping, for example by removing markers close to the centrosome. This step can be performed using the above-mentioned mapping techniques. In addition, the data set comprising low-abundance markers can be orthogonally integrated with aneuploidy markers (e.g., markers indicating gene amplification or gene deletion).

[0277] Systems and devices for implementing diagnostic / screening methods

[0278] The computer system 400 can be used as a stand-alone device or on a distributed network (e.g., a cloud computing network) that shares computer processing resources to implement the methods described herein (e.g., method 100). In this way, a non-transitory computer-readable medium having a first program stored therein can be provided, the first program being used to cause a computer to perform the disclosed method for removing artificial noise (e.g., associated with the following markers: low BQ / MQ markers, markers larger than a threshold fragment size of about 160 bp; and markers with a VAF smaller than a threshold of about 4%). A non-transitory computer-readable medium can be provided, in which a second program is stored to adaptively and systematically filter noise (e.g., noise associated with PCR / sequencing errors). A non-transitory computer-readable medium can be provided, in which a second program is stored, for matching a noise-filtered, CNN-processed subject-specific signature with a cancer signature, for example, by determining a zscore or by analyzing a probability density function, and outputting a confidence interval (CI) for the match, wherein a CI ≥ a threshold value (e.g., 80%, 90%, 95%, or 99%) indicates that the subject has a tumor. In some embodiments, each of the first, second, and third programs may be provided or used separately (eg, in a standalone form); in some embodiments, each of the first, second, and third programs may be provided or used together (eg, as a program package).

[0279] It should also be understood that the foregoing embodiments may be provided in whole or in part as a system of components that are integrated to perform the described methods. Figure 1A The workflow can be provided as a system as components or stations to identify high-quality, low-abundance, tumor-specific markers present in cfDNA of cancer patients and further enable early diagnosis in a sensitive, precise and accurate manner.

[0280] As mentioned above, one of the notable features of the systems and methods of the present disclosure is the use of an engine that can adaptively and systematically filter noise. A representative engine is described in detail below. The engine can be based on, for example, Figure 1A The flowchart of is implemented in the diagnostic method of the present disclosure (as discussed in detail below). (Note: the location of the engine in the flowchart is only exemplary to fit the representative method). The engine may include a convolutional neural network (CNN) capable of capturing the invariance of the marker (e.g., somatic mutations including sSNVs). The CNN and its corresponding architecture will be discussed in detail below with reference to the "Convolutional Neural Network (CNN)" section.

[0281] The ability of this engine to eliminate low-quality markers can be evaluated in synthetic plasma samples and real plasma DNA samples. By randomly sampling from the patient's healthy DNA and the patient's tumor DNA, synthetic plasma samples can be generated from test samples (e.g., lung samples). For real plasma DNA analysis, plasma samples obtained from lung cancer patients who smoke can be used. As a control, the patient's PMBC can be used. Alternatively, plasma samples obtained from non-cancer or healthy subjects can be used as controls.

[0282] Figure 5 An exemplary overview of how machine learning (ML) can be used to suppress noise (e.g., errors during amplification (PCR), errors during sequencing, errors in mapping, and other false positive markers, such as mutations seen in control samples) and detect mutations in subject samples is provided in. As shown, genetic data are received from a subject in an appropriate format (e.g., VCF format for variant calls), which can be true positives or false positives. These data are input into a machine learning tool, such as an n-dimensional convolutional neural network (CNN). The CNN may have K filters per position, for a total of 32 D learnable filters, where D is the number of dimensions in the CNN. For example, a size of 2 and a stride of 2 are used to max pool genetic data. Any method that captures sequencing reads with discrete feature representations can be used. For example, a spatially oriented representation with a maximum of 1, 2, 3, 4...n feature lengths can be used.

[0283] An exemplary feature is provided in FIG8 . As shown, the first five rows represent the reference context (e.g., a sequence in the human genome), the next five rows represent the read sequence (base pairs in the read), rows 11-15 represent the alignment string (CIGAR), and the last row represents the quality score for each position in the read. Each column of the feature represents an indicator vector that indicates whether a specific base is present. The read, genomic context, and CIGAR rows are mutually exclusive, such as in one-hot encoding. Details on the construction and implementation of this feature are provided in the representative embodiments below.

[0284] The Engine can be used as a standalone tool or in combination with other mutation callers known in the art, such as PILEUP (Li et al., Bioinformatics, 25(16): 2078-2079, 2009), STRELKA (Saunders et al., Bioinformatics, 28(14): 1811-1718, 2012), LOFREQ (Wilm et al., Nucleic acids research, 40(22): 11189-11201, 2012), etc. Figure 7 An exemplary overview of the location and input / output of the Engine is shown in . Note: Although the Engine is located at the far end of the pipeline in this figure, in reality, the Engine can be placed at any level or stage in the process. To train the ENGINE, genetic data containing marker profiles from mixed tumor biopsy samples and peripheral blood mononuclear cells (PMBCs; control) can be optionally processed with the above filters (e.g., manual read classification via PILEUP; exclusion of germline variants using VAF; removal of markers with low base quality using appropriate BQ filters; and removal of poorly mapped markers using appropriate MQ filters). The machine can also be trained with a dataset.

[0285] When the engine was executed using independent samples from lung cancer patients, it was found to distinguish true somatic mutations from noise with a high degree of sensitivity and accuracy. The results are shown in Figures 8 and 9. Experiments with synthetic plasma revealed that the engine is particularly accurate and sensitive at low tumor fractions (TF) and outperforms state-of-the-art callers such as MUTECT (Cibulskis et al., Nature biotechnology, 31(3):213–219, 2013) and / or PILEUP. In particular, the engine demonstrated strong performance in both computer analysis and clinical settings. Compared to programs such as MUTECT, the engine performs particularly well in the context of balanced tumor fractions. For example, on the sensitivity metric, it outperforms MUTECT, SNOOPER (Spinella et al., BMC Genomics, 17(1):912, 2016), and STRELKA. See Fig. 9A . On the accuracy metric, it outperforms PILEUP at all tumor fractions, and at low TF (TF=0.0001), it outperforms PILEUP by about 25-fold. It maintains a large portion of its performance even in simulated plasma. The engine also achieves about 30-fold enrichment (over PILEUP) at TF of 0.0001, demonstrating its ability to capture relevant somatic mutations even when they are 10-fold less than the sequencing noise itself. See Fig. 9C . In contrast, MUTECT provided a modest improvement of approximately 2-fold across all tumor fractions (compared to PILEUP). In addition, the engine allows the user to minimize false negatives, and for applications where specificity is a priority, the engine can be set to minimize false positives. The engine variant identification system can minimize both false positives and false negatives, detecting variants with unparalleled precision and accuracy (see Table 4 for summary).

[0286] In particular, the engine can be optionally applied with noise elimination filters (e.g., mutation frequency (MF) filters and / or base quality map quality (BQMQ) filters) to significantly improve the accuracy of mutation callers known in the art. Representative channels using the mutation caller PILEUP and downstream noise elimination filters and engines are described in the Examples section below. In the case of real plasma samples, the above channels including PILEUP, noise elimination filters (based on mutation frequency (MF) and quality (BQMQ)) and the engine can significantly suppress false positives while significantly enriching samples for tumor DNA analysis. In summary, these results show that the engine can be used to greatly improve the performance of mutation callers with minimal (if any) loss of sensitivity.

[0287] The performance of the engine demonstrates that integrating features from reads and their alignments has laid the foundation for a new collection of somatic mutation callers that use the full mutational profile of a sample, not just those covered at high depth. Capturing mutations at such a sensitive level using simple measurement tools allows for new and improved diagnostic platforms that can be used in the treatment and / or management of cancer patients.

[0288] The present disclosure relates to at least three potential applications of the engine, namely, improved somatic SNV mutation detection, particularly in cancer diagnosis, prognosis and care and other clinical settings; improved structural variant detection for genetic disease diagnosis and disease risk assessment; and / or improved germline genomic SNV detection in biomedical research, disease diagnosis and / or treatment. Fig.10 .

[0289] Based on the state of the art, the engine is the first somatic mutation caller designed to work in a low allele frequency context, such as liquid biopsies for early cancer detection. To achieve the goal of early detection, a new read-along representation was implemented using a custom architecture to optimally capture expected features associated with reads and their alignments. Therefore, the present disclosure provides a new family of somatic mutation callers to assist in detection in liquid biopsies, paving the way for critical, non-invasive cancer diagnostic methods, particularly in the context of early tumor detection as well as residual tumor detection.

[0290] Computer System

[0291] In some embodiments, the diagnostic methods of the present disclosure are implemented on a computer system. As a representative example only, Fig.15 A schematic diagram of such a computer system is provided in . Fig.154 is a block diagram illustrating a computer system 400 on which embodiments or portions of embodiments of the present disclosure may be implemented. In various embodiments of the present disclosure, the computer system 400 may include a bus 402 or other communication mechanism for communicating information, and a processor 404 coupled to the bus 402 for processing information. In various embodiments, the computer system 400 may also include a memory, which may be a random access memory (RAM) 406 or other dynamic storage device, coupled to the bus 402 to determine instructions to be executed by the processor 404. The memory may also be used to store temporary variables or other intermediate information during the execution of instructions executed by the processor 404. In various embodiments, the computer system 400 may also include a read-only memory (ROM) 408 or other static storage device coupled to the bus 402 for storing static information and instructions for the processor 404. A storage device 410, such as a magnetic disk or optical disk, may be provided and coupled to the bus 402 to store information and instructions. In various embodiments, the computer system 400 can be coupled to a display 412, such as a cathode ray tube (CRT) or a liquid crystal display (LCD), via the bus 402 to display information to a computer user. An input device 414 including alphanumeric and other keys can be coupled to the bus 402 to communicate information and command selections to the processor 404. Another type of user input device is a cursor control 416, such as a mouse, trackball, or cursor direction keys, for communicating direction information and command selections to the processor 404 and controlling cursor movement on the display 412. This input device 414 typically has two degrees of freedom on two axes, a first axis (e.g., x) and a second axis (e.g., y), allowing the device to specify a position in a plane. However, it should be understood that input devices 414 that allow 3-dimensional (x, y, and z) cursor movement are also contemplated herein.

[0292] Consistent with certain embodiments of the present disclosure, results may be provided by computer system 400 in response to processor 404 executing one or more sequences of one or more instructions contained in memory 406. These instructions may be read into memory 406 from another computer-readable medium or computer-readable storage medium (e.g., storage device 410). Executing the sequences of instructions contained in memory 406 may cause processor 404 to perform the processes described herein. Alternatively, hard-wired circuitry may be used in place of or in combination with software instructions to implement the present teachings. Therefore, implementation of the present teachings is not limited to any specific combination of hardware circuitry and software.

[0293] As used herein, the term "computer-readable medium" (e.g., data retention, data storage, etc.) or "computer-readable storage medium" refers to any medium that participates in providing instructions to processor 404 for execution. Such media can take many forms, including but not limited to non-volatile media, volatile media, and transmission media. Examples of non-volatile media can include but are not limited to optical disks, solid-state disks, magnetic disks, such as storage device 410. Examples of volatile media can include but are not limited to dynamic memory, such as memory 406. Examples of transmission media can include but are not limited to coaxial cables, copper wire, and optical fiber, including the wires that comprise bus 402.

[0294] Common forms of computer readable media include, for example, floppy disks, diskettes, hard disks, magnetic tape or any other magnetic medium, CD-ROMs, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, RAM, PROM and EPROM, FLASH-EPROM, any other memory chip or cassette, or any other tangible medium from which a computer can read.

[0295] In addition to computer-readable media, data may be provided as signals on a transmission medium included in a communication device or system to provide one or more instruction sequences to the processor 404 of the computer system 400 for execution. For example, a communication device may include a transceiver with signals indicating instructions and data. The instructions and data are configured to cause one or more processors to implement the functions outlined in the present disclosure. Representative examples of data communication transmission connections may include, for example, telephone modem connections, wide area networks (WANs), local area networks (LANs), infrared data connections, NFC connections, etc.

[0296] It should be understood that the methods described herein, including the flow charts, diagrams, and accompanying disclosure, may be implemented using computer system 400 as a standalone device or on a distributed network (eg, a cloud computing network) that shares computer processing resources.

[0297] system

[0298] The present disclosure further relates to a system for performing the method of the present disclosure. Figures 16A-16C A representative system is provided in the schematic diagram of . Fig.16A An exemplary system for implementing the diagnostic method of the present disclosure is shown. As depicted herein, a system 500 is provided that may include a data acquisition unit 510, a marker identification unit 520, a diagnostic unit 550, and a display 412 for outputting data and for receiving user input via an associated input device (not shown). The marker identification unit 520 may include a noise removal unit 530 and a classification engine 540. It should be noted that Fig.16AOne configuration of the system is shown. The orientation and configuration of these components can be varied as desired. In addition, additional components can be added to the system (e.g., a convolutional neural network). These various components, their various operations, their various orientations, and various relationships between them will be discussed in detail below.

[0299] Fig.16A The data acquisition 510 unit can be configured and arranged to receive the genetic summary from the subject, for example, from a plurality of genetic markers of biological sample sequencing including the plasma sample and normal cell sample of the subject, to produce a subject-specific whole genome genetic marker summary. In some embodiments, the genetic marker summary is received in a physical disk (for example, CD, DVD) or through the Internet (for example, as provided by a server or cloud) in a variation call format (VCF) file. In some embodiments, for example, the sample of the subject is sequenced using whole genome sequencing (WGS), and the sequence file is directly transferred to the data acquisition unit 510. In some embodiments, the data acquisition unit 510 can reformat, organize, classify or otherwise reconfigure the received data to be further analyzed in the system 500. In some embodiments, the unit 510 can receive data (for example, via a display 412, data associated therewith or user input, a memory associated therewith, or another memory component associated with the computer system 400).

[0300] The data acquired by the data acquisition unit can be transmitted to the marker identification unit 520. The marker identification unit 520 may include one or more engines for analyzing the markers in the subject-specific genetic marker summary. As one of those components of the unit 520, the noise removal unit 530 may include one or more programs for filtering artificial noise by weighting the markers based on BQ, MQ, fragment size and / or VAF, including one or more secondary features detailed above, such as position (RP) in the read; sequence context (SC); abundance; sequencing depth and / or sequencing errors. Preferably, the noise removal unit includes a program for calculating the best receiver operating characteristic (ROC) curve, which includes a probability classification of the genetic markers in the summary based on a score (e.g., a joint base quality (BQ) and mapping quality (MQ) score), which is integrated with the fragment size score and / or VAF score. The noise removal unit may include a program for measuring the area under the ROC curve (AUC), which typically represents the probability that a candidate marker randomly selected from the potential markers shows a higher value than a randomly extracted control marker. The classifier may include a procedure for evaluating whether a particular binned marker is a "chance" marker or a "true" marker based on a ROC curve.

[0301] In some embodiments, the noise removal unit may be based on the probability score (P D ) weights the marks. Preferably, the program is based on the Bernoulli equation P D =1-〖(1-TF)〗 GE To measure the probability of detection (P D ), where P D is the probability of detection, TF is the tumor fraction, and GE is the number of genome equivalents present in the patient’s DNA. D Each genetic marker is weighted so that the one with the highest P D For example, the markers can be binned based on a P value of at least about 0.60, such as at least about 0.65, 0.70, 0.75, 0.80, 0.90, 0.95 or greater, such as at least about 0.98. D The threshold bins the genetic markers. Thus, if the P D Below the threshold, it can be classified as a false positive and not included in the analysis.

[0302] The marker identification unit 520 may include a classification engine 540 that can examine, for example, the likelihood that a marker is associated with noise. The classifier may include a classification scheme that includes an algorithm or neural network that can adaptively identify erroneous markers (e.g., errors due to PCR or sequencing). In a specific embodiment, the classification unit 540 includes a deep convolutional neural network (CNN) to adaptively and / or systematically filter sequencing noise that can affect accurate detection of tumor-specific low-abundance markers. The CNN can be provided as a separate engine within the marker identification unit 520, or can be provided as a separate unit (e.g., between the marker identification unit 520 and the diagnostic unit 550). Features of the CNN ( Fig.16A Detailed description is given below.

[0303] Finally, the subject-specific signature including the markers filtered by noise and processed by the CNN can be provided as a file to the diagnostic unit 550, which is configured and arranged to diagnose a disease (e.g., a tumor disease) based on a statistical score indicating a match between the subject-specific signature and the cancer signature. The diagnostic unit may include a repository containing cancer signatures, such as a Catalogue of Somatic Mutations in Cancer (COSMIC) database or a Latin American Consortium for the Investigation of Lung Cancer (CLICaP) database. The diagnostic unit 550 may include one or more software or algorithms for comparing known cancer mutation signatures (e.g., any one of COSMIC's signatures 1 to 30) with subject-specific mutation signatures. Representative examples of such comparison software include, for example, measuring confidence estimates at the level of a single marker and at the level of a pool including 2, 5, 10, 20, 50, 100, 200, 500, 1000 or more, such as 5000 unique markers. Representative methods include estimating the Z-score confidence level using linear optimization (as described above), or determining the similarity of normalized probability density functions (PDFs) using a cosine similarity function (as described above).

[0304] The output of the diagnostic engine can be output to, for example, a display 412 for viewing by a user. In some embodiments, the output can include a raw confidence interval (CI) score or an ordinal score (e.g., a score on a scale of 1 to 10, where 10 is a high probability that the subject has a neoplastic disease and 1 is a low probability that the subject has a neoplastic disease).

[0305] Because it involves orientation, Fig.16AThe marker identification unit 520 of the system 500 may be communicatively connected to the data acquisition unit 510. In addition, each component (e.g., engine, module, etc.) depicted as part of the marker identification unit 520 (and described herein) may be implemented as hardware, firmware, software, or any combination thereof. In various embodiments, the marker identification unit 520 may be implemented as an integrated instrument system component with the data acquisition unit 510. That is, the unit 520 and the unit 510 may be housed in the same housing assembly and may communicate via conventional device / component connection means (e.g., serial bus, optical cable, electrical cable, etc.). In various embodiments, the marker identification unit 520 may be implemented as an independent computing device (as shown in FIG. 16 ) that is communicatively connected to the data acquisition unit 510 via an optical, serial port, network, or modem connection, for example, via a LAN or WAN connection that allows imaging data acquired by the data acquisition unit 510 to be transferred to the marker identification unit 520 for analysis. In various embodiments, the functionality of the tag identification unit 520 may be implemented on a distributed network of shared computer processing resources (e.g., a cloud computing network) that is communicatively connected to the data acquisition unit 510 via a WAN (or equivalent) connection. For example, the functionality of the tag identification unit 520 may be divided to be distributed across a network such as AMAZON WEB SERVICES. TM The system is implemented in one or more computing nodes on a cloud processing service.

[0306] Fig. 16B A second exemplary system for implementing the diagnostic method of the present disclosure is shown. Fig. 16B As shown, an exemplary system 100 is provided, which is configured and arranged for performing genetic screening for cancer on a subject in need thereof. Fig. 16B , the system 100 may include an analysis unit 110 and a computation unit 140. The analysis unit 110 may include a pre-filter engine 120 and a correction engine 130. These system components and associated engines will be discussed in more detail below.

[0307] Reference again Fig. 16B , the pre-filter engine 120 of the analysis unit 110 can be configured and arranged to receive a subject-specific whole genome read summary associated with a plurality of genetic markers from a biological sample of the subject. As discussed with respect to the workflow herein, and according to various embodiments, the biological sample can include a tumor sample; the read summaries can each include reads of a single base pair length.

[0308] The pre-filter engine 120 may also be configured and arranged to filter artificial sites from the read summary. As discussed with respect to the workflow herein, and according to various embodiments, filtering may include removing duplicate sites generated in a reference healthy sample cohort from the read summary, and / or identifying germline mutations in the biological sample / or identifying shared mutations between peripheral blood mononuclear cells of a tumor sample and a normal cell sample as germline mutations, and removing the germline mutations from the read summary.

[0309] The correction engine 130 of the analysis unit 110 may be configured and arranged to receive the output from the engine 120. The correction engine 130 may be configured and arranged to filter noise from the whole genome read summary using at least one error suppression scheme to produce a filtered set of reads of the whole genome read summary.

[0310] As discussed with respect to the workflows herein, and according to various embodiments, the at least one error suppression scheme can include calculating the probability that any single nucleotide variation in the compendium is an artificial mutation, and removing that mutation.

[0311] As discussed with respect to the workflows herein, and according to various embodiments, the probability can be calculated as a function of a feature selected from the group consisting of mapping quality (MQ), variant base quality (MBQ), position in the read (PIR), mean read base quality (MRBQ), and combinations thereof.

[0312] As discussed with respect to the workflows herein, and according to various embodiments, the at least one error suppression scheme can include removing artificial mutations using inconsistency testing and / or repeat consistency between independent repeats of the same DNA fragment generated by polymerase chain reaction or sequencing processing, wherein artificial mutations are identified and removed when consistency is lacking between a majority of a given repeat family.

[0313] The computational unit 140 of the system 100 may be configured and arranged to receive the output from the correction engine 130 and compile a subject-specific signature using the filtered set of reads based on comparison to a specific mutation signature associated with a predetermined mutagenesis procedure.

[0314] The computing unit 140 may also be configured and arranged to statistically quantify the confidence estimate that the biological sample of the subject includes the cancer-associated mutation signature based on a comparison of the cancer-associated mutation signature exposure value with the background mutation signature cohort through the subject-specific signature. The computing unit 150 may be further configured and arranged to screen the subject for cancer if the confidence estimate that the biological sample of the subject includes the cancer-associated mutation signature exceeds a given threshold.

[0315] The system 100 may also include a display 150, such as Fig. 16BThe display may be configured and arranged to receive output from the computing unit 140. The output may include data related to cancer screening in the subject / user. Alternatively, the system 100 may exclude a display, and the data output from the computing unit 140 may be sent to any form of storage or display device or location external to the system 100. Also as discussed herein, the components of the system 100 may be integrated into a single unit, or may be broken down into separate components. Fig. 16B More separate physical units are shown. Furthermore, system 100 may be part of a distributed network of systems, each performing substantially similar tasks and transmitting data from each system to a hub.

[0316] Fig. 16C A third exemplary system for implementing the diagnostic method of the present disclosure is shown. Fig. 16C As shown, an example system 100 is provided, which is configured and arranged to perform cancer genetic screening on a subject in need. The system 100 may include an analysis unit 110 and a calculation unit 150. The analysis unit 110 may include a binning engine 120, a pre-filter engine 130, and a normalization engine 140. These system components and related engines will be discussed in more detail below.

[0317] Reference again Fig. 16C , the binning engine 120 can be configured and arranged to receive a subject-specific whole genome read summary associated with a plurality of genetic markers from a biological sample of a subject. As discussed with respect to the workflow herein, and according to various embodiments, the first biological sample can include a tumor sample; the first read summary can include copy number variations (CNVs).

[0318] The binning engine 120 can be configured and arranged to divide the read summary into a plurality of windows and calculate a set of features for each window. These features can include the median depth coverage of each window and the representative fragment size of each window.

[0319] The pre-filter engine 130 may be configured and arranged to filter artificial sites from the read summary. The filtering may include removing duplicate sites generated on a reference healthy sample cohort from the read summary.

[0320] The normalization engine 140 of the analysis unit 110 can be configured and arranged to receive the output from the engine 130. The normalization engine 140 can be configured and arranged to normalize the read summary to produce a filtered set of reads for the whole genome read summary. Normalization methods are discussed in detail herein and can be used in any desired combination to normalize the reads in question.

[0321] The computing unit 150 of the system 100 can be configured and arranged to receive the output from the normalization engine 140 and calculate an estimated tumor score (eTF) using the filtered set of reads by calculating the linear relationship between the feature set of each window and converting the calculated relationship into an estimated tumor score using a regression model. The computing unit 150 can also or additionally be configured and arranged to calculate the estimated tumor score based on one or more integrated mathematical models as a function of the feature set of each window calculated in the subject-specific whole genome read summary. The computing unit 150 can be further configured and arranged to screen the subject for cancer if the estimated tumor score exceeds an empirical threshold. Regression models, integrated mathematical models, and empirical thresholds are discussed in detail herein.

[0322] The system 100 may also include a display 160, such as Fig. 16C The display may be configured and arranged to receive output from the computing unit 150. The output may include data related to detecting residual disease in the subject / user. Alternatively, the system 100 may exclude a display, and data output from the computing unit 150 may be sent to any form of storage or display device or location external to the system 100. Also as discussed herein, the components of the system 100 may be integrated into a single unit, or may be broken down into more Fig. 16C More separate physical units are shown. Furthermore, system 100 may be part of a distributed network of systems, each performing substantially similar tasks and transmitting data from each system to a hub.

[0323] Convolutional Neural Networks (CNN)

[0324] The present disclosure also relates to systems and programs that utilize a convolutional neural network (CNN) (e.g., an engine) to adaptively and / or systematically filter sequencing noise.

[0325] The present disclosure further relates to a computer-readable storage medium containing a program for detecting a tumor marker comprising somatic mutations in genomic reads, the program comprising a hierarchical convolutional neural network (CNN).

[0326] As is known in the art, convolutional neural networks (CNNs) typically accomplish a high-level form of processing and classification / detection by first finding low-level features (such as, for example, repetitive sequences in a read), and then proceeding to more abstract (e.g., unique to the type of read being classified) concepts through a series of convolutional layers. A CNN can do this by passing the data through a series of convolutional, nonlinear, pooling (or downsampling, as described below), and fully connected layers, and obtaining an output. Again, the output can be a single class or a probability of a class that best describes the data or detects an object in the data.

[0327] Regarding the layers in a CNN, the first layer is typically a convolutional layer (conv). The first layer will process a representative array of readings using a series of parameters. Rather than processing the data as a whole, a CNN will use filters (or neurons or kernels) to analyze a collection of subsets of the data. The subset will include the focal point in the array as well as the surrounding points. For example, the filter is able to examine a series of 5x 5 regions (or areas) in a 32x 32 representation. These areas can be called receptive fields. Since the filter typically has the same depth as the input, a representation of size 32x 32x 3 will have a filter of the same depth (e.g., 5x5x 3). The actual steps of performing a convolution using the exemplary dimensions above will involve sliding the filter along the input data, multiplying the filter values ​​with the original representative values ​​of the data to calculate an element wise multiplication, and summing these values ​​to derive a single number for the inspection area of ​​the representation.

[0328] After this convolution step is complete, using 5x 5x 3 filters, you will get an activation map (or filter map) of size 28x 28x 1. With each additional layer used, the spatial dimensions are better preserved so that using two filters will result in an activation map of 28x 28x 2. Each filter typically has a unique feature that it represents, which together represent the feature identifier required for the final data output. When these filters are used in combination, the CNN is allowed to process the data input in order to detect those features that are present on each representation. So if the filter is used as a curve detector, the convolution of the filter along the data input will produce an array of numbers in the activation map that corresponds to a high likelihood of a curve (high summed elements multiplied in sequence), a low likelihood of a curve (low summed elements multiplied in sequence), or a zero value (where the input volume at certain points does not provide anything to activate the curve detector filter). In this way, the greater the number of filters (also called channels) in the Conv, the more depth (or data) is provided on the activation map and therefore provides more information about the input that will lead to a more accurate output.

[0329] Balanced against the accuracy of a CNN is the processing time and power required to produce the results. In other words, the more filters (or channels) used, the more time and processing power required to perform the Conv. Therefore, the choice and number of filters (or channels) should be specifically chosen to meet the needs of the CNN method to produce the most accurate output possible while taking into account the available time and power.

[0330] To further enable the CNN to detect more complex features, additional Convs can be added to analyze the output from previous Convs (e.g., activation maps). For example, if the first Conv looks for basic features (such as curves or edges), the second Conv looks for more complex features (such as shapes), which can be a combination of individual features detected in earlier Conv layers. By providing a series of Convs, the CNN is able to detect higher and higher level features, ultimately reaching a probability of detecting a specific desired object. In addition, because Convs are stacked on top of each other, analyzing previous activation map outputs, each Conv in the stack naturally analyzes a larger and larger receptive field through the proportional reduction that occurs at each Conv level, allowing the CNN to respond to gradually growing areas of the representation space when detecting the object of interest.

[0331] The CNN architecture typically consists of a set of processing blocks, including at least one processing block for convolving the input volume (data) and at least one processing block for deconvolution (or transposed convolution). In addition, the processing block may include at least one pooling block and an unpooling block. The pooling block can be used to scale down the data resolution to produce an output that can be used for Conv. This can provide computational efficiency (effective time and capacity), which in turn can improve the actual performance of the CNN. These pooling or subsampling, blocks keep the filters small and the computational requirements reasonable, and these blocks can make the output coarse (causing the loss of spatial information in the receptive field), thereby reducing the size of the input by a specific factor.

[0332] The unpooling block can be used to reconstruct these coarse outputs to produce an output volume with the same dimensions as the input volume. The unpooling block can be viewed as the reverse operation of the convolution block to return the activation output to the original input volume dimensions. However, the unpooling process often simply upscales the coarse output into a sparse activation map. To avoid this result, the deconvolution block densifies this sparse activation map to generate an enlarged and densified activation map, ultimately resulting in an output volume that is closer in size and density to the input volume after any further necessary processing. As the reverse operation of the convolution block, instead of reducing multiple array points in the receptive field to a single number, the deconvolution block associates a single activation output point with multiple outputs to upscale and densify the resulting activation output.

[0333] It should be noted that while pooling blocks can be used to downscale data and unpooling blocks can be used to upscale these downscaled activation maps, convolution and deconvolution blocks can be constructed to perform convolution / deconvolution and downscaling / upscaling without separate pooling blocks and unpooling blocks.

[0334] Depending on the intended objects to be detected in the data input, the pooling block and unpooling process may have drawbacks. Since pooling usually scales down the data by looking at sub-windows of the data without window overlap, there is a significant loss of spatial information when scaling down.

[0335] The processing block may include other layers wrapped with convolutional or deconvolutional layers. These layers may include, for example, rectified linear unit layers (ReLU) or exponential linear unit layers (ELU), which are activation functions that examine the output of the Conv in its processing module. The ReLU or ELU layer acts as a gating function, advancing only those values ​​corresponding to positive detections of the feature of interest specific to the Conv.

[0336] Given the basic architecture, the CNN is then prepared for a training process to hone its accuracy in data classification / detection (of the target object). This involves a process called backpropagation (backprop), which uses a training dataset or sample data used to train the CNN to allow it to update its parameters to achieve an optimal (or threshold) accuracy. Backpropagation involves a series of repeated steps (training iterations) that will train the CNN slowly or quickly, depending on the parameters of the backpropagation. The backpropagation step typically includes a forward pass, a loss function, a backward pass, and parameter (weight) updates according to a given learning rate. The forward pass involves passing the training data through the CNN. The loss function is a measure of the error in the output. The backward pass determines the factors that affect the loss function. Weight updates involve updating the parameters of the filters to move the CNN to an optimal state. The learning rate determines how much the weights are updated each iteration in order to achieve the optimal state. If the learning rate is too low, training may take a long time and involve too much processing power. If the learning rate is too fast, each weight update may be too large to accurately achieve a given optimal value or threshold.

[0337] The back-propagation process can complicate training, thus leading to the need for lower learning rates and more specific, carefully determined initial parameters at the beginning of training. One complication is that changes in Conv parameters are amplified with the depth of the network as weight updates occur at the end of each iteration. For example, if a CNN has multiple Convs, which allow for higher levels of feature analysis as described above, then parameter updates to the first Conv are multiplied at each subsequent Conv. The net result is that minimal changes to the parameters can have large effects depending on the depth of a given CNN. This phenomenon is called internal covariate shift.

[0338] Typically, the CNN disclosed herein can adaptively and / or systematically filter sequencing noise. In some embodiments, based on the inventor's understanding that the trinucleotide background contains different features involved in mutagenesis, the CNN architecture is designed. Therefore, the CNN uses a receptive field of size 3 to convolve on all features (columns) at a certain position. After two consecutive convolutional layers, downsampling is performed by maximum pooling with a receptive field of 2 and a stride of 2, thereby forcing the model in the engine to retain the most important features only in a smaller spatial region. When convolving on a trinucleotide window, the resulting architecture maintains spatial invariance and captures "quality mapping" by collapsing the read fragments into 25 segments (each segment represents a region of about 8 nucleotides). The final classification is performed by applying the output of the last convolutional layer directly to an S-type fully connected layer. CNN uses a simple logistic regression layer instead of a multilayer perceptron or global average pooling to retain features related to the position in the genome read.

[0339] In order to train the engine, first sample multiple lung cancer patients and their matched systematic error profiles. The goal of the training exercise is to use a training scheme that allows true somatic mutations to be detected with high sensitivity and also rejects candidate mutations caused by systematic errors. To this end, four separate samples are selected from different smoking lung cancer patients for training, each sample including a complete tumor sample and a healthy tissue sample from the same patient (see, e.g., Table 3). For example, the consistency of three callers known in the art (STRELKA, LOFREQ, and MUTECT) can be used to make the final call of somatic mutations. Then, the readings supporting these mutations are used as tumor readings for the training engine.

[0340] In order to ensure that the model engine learns to distinguish sequencing artifacts, the readings from healthy samples containing mutations that occur just once are collected. Since these variants are not supported by more than one reading, it can be highly certain that they are regarded as the product of system errors. Then filter low-quality variants. For example, the base quality score with less than 20 mutations or the readings with less than 40 mapping qualities (e.g., BQ20, MQ40) can be filtered. These thresholds are purely exemplary and can be identified by reading inspection. Optionally, if necessary, samples of lower quality can be included in the training engine. A subset of the training set can be used as a validation data set, which can be used to monitor the progress of the training and verify the performance of the model on independent readings.

[0341] According to various embodiments herein, a computer-readable medium is provided, the computer-readable medium comprising computer-executable instructions, which when executed by a processor causes the processor to perform a method or a set of steps for identifying low-abundance tumor-specific markers in a genetic marker summary received from a subject sample, wherein the genetic markers include SNVs (preferably sSNVs), CNVs (preferably sCNVs), indels and / or SVs (preferably translocations, gene fusions, or combinations thereof) in genomic reads. Preferably, the medium comprises a hierarchical convolutional neural network (CNN) having a single fully connected layer at one end, wherein the CNN maintains spatial invariance when convolved on a trinucleotide window; and maintains a quality map by collapsing the fragments of the reads into multiple segments (each segment representing a region of approximately 8 nucleotides), wherein the CNN weights each genetic marker in the summary. For example, the CNN of the present disclosure may include 8 layers, including a single fully connected layer at one end and two consecutive convolutional layers, whose outputs are downsampled by maximum pooling with a receptive field of 2 and a stride of 2; wherein the 8-layer CNN maintains a quality map by collapsing the read fragments into approximately 25 separate segments and convolving on the columns at a certain position of the genomic reads using a perceptive field of size 3; and wherein the output of the last convolutional layer is directly applied to a sigmoid fully connected layer, thereby performing a final classification of the markers.

[0342] The CNN may include a read representation that jointly captures the genomic context of the alignment, the complete read sequence, and an integral of the quality score for each base. Due in part to this arrangement and architecture, the CNN of the present disclosure enriches tumor-specific signatures containing somatic mutations in genomic reads by about 1.12 to about 30 times compared to the state-of-the-art mutation caller MUTECT.

[0343] The present disclosure also relates to a computer-readable medium including computer-executable instructions, which when executed by a processor cause the processor to perform a method or a set of steps for diagnosing cancer in a subject in need of diagnosis, the medium comprising a convolutional neural network (CNN). In some embodiments, the CNN is developed by the following steps: using a training data set containing tumor-related signatures and PCR / sequencing error signatures to train the CNN to distinguish between cancer mutation-supported reads and artificial mutation (error) reads; and optionally using actual samples from cancer patients or synthetic plasma obtained from the data set for verification, receiving a genetic marker summary obtained from a sample of the subject, wherein the genetic marker includes SNV (preferably sSNV), CNV (preferably sCNV), indel and / or SV (preferably translocation, gene fusion or a combination thereof) in the genomic reads.

[0344] In some embodiments, the mathematical optimization step used in CNN development includes using non-negative least squares (NNLS). Other representative methods include cross entropy global optimization method, golden section search method or a combination thereof.

[0345] Preferably, the CNN of the present disclosure comprises a single fully connected layer at one end, wherein the procedure maintains spatial invariance when convolving over a three-nucleotide window; and maintains a quality map by collapsing the read fragments into multiple segments, each segment representing a region of approximately eight nucleotides.

[0346] In some embodiments, the disclosed system includes an 8-layer CNN that maintains a quality map by collapsing the read fragments into about 25 separate segments, and which further convolves all features (columns) at a certain position of the genomic read using a perceptive field of size 3. In the case of analyzing genetic markers (e.g., sSNVs) in cfDNA, the CNN may include two consecutive convolutional layers, whose outputs are downsampled by max pooling with a receptive field of 2 and a stride of 2, and the output of the last convolutional layer is directly applied to a sigmoid fully connected layer, from which the marker is finally classified.

[0347] The CNN configured in the aforementioned manner takes into account spatial invariance in true somatic mutations and errors due to mapping, while maintaining the base quality of the entire read and providing a read representation that jointly captures the genomic context of the alignment, the complete read sequence, and an integral of the quality score for each base.

[0348] The embodiments disclosed herein have several advantages over known CNNs. These advantages include, for example, providing a CNN that can greatly improve accuracy and sensitivity. In particular, compared to programs known in the art (e.g., MUTECT), the systems and networks disclosed herein enrich tumor-specific markers containing somatic mutations in genomic reads (which measure the ratio of output accuracy to input accuracy) by about 1.12 times to about 12 times, such as about 2 times, about 3 times, about 3 times, about 4 times, about 5 times, about 6 times, about 7 times, about 8 times, about 9 times, about 10 times or more.

[0349] In some embodiments, the CNN includes employing a deep learning algorithm in a pan-cancer cohort to identify signatures that distinguish true tumor mutations from artifacts. The algorithm performs this function by assigning a confidence estimate to each individual mutation detected in a tumor patient sample; integrating the confidence estimates across the entire genome; and employing an algorithm to analyze mutation signatures in the sample. For example, in the case of diagnosing lung cancer, the algorithm can analyze the lung tumor signature in the sample. Similarly, in the case of diagnosing UV-induced melanoma, the algorithm can analyze the UV signature in the sample. Similarly, in the case of diagnosing breast cancer, the algorithm can analyze the breast tumor (BRCA) signature in the patient sample.

[0350] In some embodiments, the CNN of the present disclosure includes an algorithm capable of performing NNLS analysis on a sample using an art-recognized / accessed mutation signature (e.g., a mutation signature registered in the Catalog of Somatic Mutations in Cancer (COSMIC) database). The present disclosure further relates to a CNN of the present disclosure integrated with a specific genomic map, such as the TCGA pan-cancer dataset.

[0351] According to various embodiments, the CNN of the present disclosure can be trained using a deep learning algorithm developed in a pan-lung cancer cohort. In this case, the cohort may include WGS data for patients with deep tumors and PBMCs (controls). By utilizing supervised learning, CNNs can be trained to identify signatures that distinguish true tumor mutations from artificial errors. The model thus obtained can be used to infer each individual mutation detected in a plasma sample of a cancer patient (e.g., an early patient with lung adenocarcinoma) and assign confidence estimates to it. Next, a signal for tumor detection can be obtained by the following steps: integrating these confidence estimates across the genome, and then using a novel analytical method for sensitive detection using non-negative least squares (NNLS) of a specific cosmic mutation signature in a single plasma sample. The confidence of the detection signal can be further verified using a comparison of cosmic mutation exposure values ​​with exposure values ​​inferred for 100 random background signatures.

[0352] In some embodiments, the machine learning (ML) methods used in the systems and / or methods of the present disclosure include deep convolutional neural networks (CNNs), recurrent neural networks (RNNs), random forests (RFs), support vector machines (SVMs), discriminant analysis, nearest neighbor analysis (KNNs), ensemble classifiers, or combinations thereof.

[0353] The systems and / or methods of the present disclosure allow for early detection in at least 50%, at least 60%, at least 70%, at least 80% or a greater percentage, such as 90% or even 95% of subjects.

[0354] Other applications

[0355] Patient reports compiled according to the foregoing methods can be sent and accessed electronically via the Internet. For example, analysis of the sequence data can occur at a location other than the subject's location. A report is generated, optionally annotated, and sent to the subject's location, such as a computer authorized via the Internet. The annotated information can be used by a healthcare provider to select other drug treatment options and / or provide information about drug treatment options to an insurance company. The method may include providing, for example, NCCN Oncology TM NCCN Clinical Practice Guidelines in Oncology TM ) or the American Society of Clinical Oncology (ASCO) clinical practice guideline. Drug therapy options that are layered in the report can be annotated in the report by listing additional drug therapy options. Additional drug therapies can be medications that are approved by the FDA for off-label use. A provision in the Omnibus Budget Reconciliation Act of 1993 (OBRA) requires Medicare to cover anticancer drugs that are outside of the allowed uses included in the standard medical pharmacopoeia. Drugs used for annotated listings can be found in CMS-approved pharmacopeias, including the National Comprehensive Cancer Network (NCCN) Drug and Biologic pharmacopoeias. TM (Biologics Compendium TM )、Thomson Micromedex Drug Elsevier Gold Standard's Clinical Pharmacology compendium and the U.S. Hospital Formulary Service - Drug Information Compendium (American Hospital Formulary Service—DrugInformation ).

[0356] In some embodiments, drug treatment options can be annotated by listing experimental drugs that can be used to treat cancers with one or more molecular markers of a particular state. Experimental drugs can be drugs for which in vitro data, in vivo data, animal model data, preclinical trial data, or clinical trial data are available. Data may be published in the peer-reviewed medical literature in journals listed in the CMS Medicare Benefit Policy Manual, including, for example, American Journal of Medicine, Annals of Internal Medicine, Annals of Oncology, Annals of Surgical Oncology, Biology of Blood and Marrow Transplantation, Blood, Bone Marrow Transplantation, British Journal of Cancer, British Journal of Hematology, British Medical Journal, Cancer, Clinical Cancer Research, Drugs, European Journal of Cancer, Gynecologic Oncology, International Journal of Radiation, Oncology, Biology, and Physics, The Journal of the American Medical Association, Journal of Clinical Oncology, Journal of the National Cancer Institute, Journal of the National Comprehensive Cancer Network (NCCN), Journal of Urology, Lancet, Lancet Oncology, Leukemia, The New England Journal of Medicine, or Radiation Oncology.

[0357] Drug therapy options can be annotated by providing a link that connects the listed drug to the scientific information about the drug on an electronic report. For example, information (clinicaltrials.gov) linked to a drug clinical trial can be provided. If the report is provided by a computer or a computer website, the link can be a footnote, a hyperlink to the website, a pop-up box, or a fly-over box (fly-over box) with information, etc. The report and the annotation information can be provided on a printed form, and the annotation can be a footnote such as a reference. The information for annotating one or more drug therapy options in the report can be provided by a commercial entity that stores scientific information. The experimental drugs listed in the annotation information can be used by a health care provider to treat a subject (such as a cancer patient), and the health care provider can access the annotated drug therapy options, retrieve scientific information (such as, print medical journal articles), and submit it (such as, printed journal articles) to an insurance company together with a request for reimbursement for providing drug therapy. The physician can use any one of various diagnosis-related groups (DRG) codes for reimbursement.

[0358] The drug treatment options in the report may also be annotated with information about other molecular components in the pathway that the drug affects (e.g., information about a drug that targets a kinase that is downstream of a cell surface receptor that is a drug target). The drug treatment options may be annotated with information about drugs that target one or more other molecular pathway components. The identification and / or annotation of pathway-related information may be outsourced or subcontracted to another company.

[0359] The annotation information can be, for example, a drug name (e.g., a drug outside of an FDA-approved permitted use; a drug found in a CMS-approved pharmacopeia; and / or a drug described in a scientific (medical) journal article), scientific information related to one or more drug treatment options, one or more links to scientific information related to one or more drugs, clinical trial information related to one or more drugs, one or more links to scientific information citations related to the drug, etc. The annotation information can be inserted anywhere in the report. The annotation information can be inserted in multiple locations in the report. The annotation information can be inserted near the portion of the report regarding the stratified drug treatment options. The annotation information can be inserted into a page of the report that is separate from the stratified drug treatment options. Reports that do not contain stratified drug treatment options can be annotated with the available information.

[0360] The system may also include a report on the effect of the drug on a sample (e.g., tumor cell) isolated from a subject (e.g., a cancer patient). In vitro cultures of tumors from cancer patients using techniques known to those skilled in the art may be established. The system may also include high throughput screening of drugs or experimental drugs outside the permitted use range approved by the FDA using the in vitro culture and / or xenograft models. The system may also include monitoring tumor antigens for recurrence detection.

[0361] In a preferred embodiment, the annotation information may include treatment recommendations, including annotations on the effects of PARP inhibitors for the BRCA signature, immunotherapy for the MSI signature.

[0362] The foregoing embodiments of the present disclosure are further described based on the following non-limiting examples. Example

[0363] The structures, materials, compositions and methods described herein are intended to be representative examples of the present disclosure, and it should be understood that the scope of the present disclosure is not limited by the scope of the embodiments. Those skilled in the art will recognize that the present disclosure can be implemented by modifying the disclosed structures, materials, compositions and methods, and such modifications are considered to be within the scope of the present disclosure.

[0364] background

[0365] Sequencing width can replace sequencing depth to overcome the limitation of cfDNA abundance in sensitive cancer detection.

[0366] The above data indicate that the detection of single sSNVs in patient plasma samples results from two sequential statistical sampling processes. The first process provides the probability of sampling a mutant fragment with a limited number of genomic equivalents present in a typical blood sample. The second process, given its abundance, sequencing depth, and sequencing errors (signal-to-noise ratio), assesses the probability of detecting a mutant fragment in the sample. Although the latter process has been the focus of intensive research and technological development in the scientific community (e.g., ultra-deep error-free sequencing protocols), the stochastic process of the former has rarely been addressed. However, in the detection of ctDNA with low disease burden, both processes play an important role as shown above. If there is no physical fragment representing the targeted sSNV, then even ideal ultra-deep targeted sequencing will not be able to detect cancer signals, which is considered to be one of the main factors leading to the limited sensitivity of these methods (~40%, Rosenfeld et al.). In practice, this problem is further complicated by the fact that a single observation (mutation read) is rarely sufficient for reliable detection.

[0367] To determine the probability of sampling a mutant fragment in a given cfDNA sample, cfDNA sampling was modeled as a Bernoulli experiment using a mixture of cfDNA fragments originating from two populations, whose proportions were defined by the tumor fraction (TF), namely cfDNA fragments originating from normal cells and cfDNA fragments originating from malignant origin. Therefore, the genomic equivalents present in the plasma sample constitute a random sampling of the entire pool of cfDNA fragments in the patient's circulation. Therefore, the probability of sampling at least one mutant fragment in a plasma sample supporting a specific permutation can be defined as: P = 1-(1-TF) GE , where P is the probability; TF is the tumor fraction; and GE corresponds to the number of genomic equivalents present in the patient's cfDNA. The model predicts that the probability of detection of TF associated with early cancer scenarios (TF < 1%) will show a rapid decline for low TFs, and even at a frequency of 0.1% (1 / 1000), the probability of detection is expected to be less than 0.65 ( Figure 3A ). Notably, these limitations were observed even under ideal conditions where exhaustive sequencing could be efficiently utilized with 1000 genome equivalents (~6 ng cfDNA) and detection based on a single supporting DNA fragment with an ideal signal-to-noise ratio. These results suggest that plasma sampling probability imposes a strict upper limit on mutation detection for low TF scenarios such as MRD and early cancer stage detection.

[0368] On the contrary, this model also shows that the limitation on sequencing depth can be effectively overcome by increasing the number of detected sites (SNVs) by increasing the width, which is due to repeating the Bernoulli test for each SNV (binomial distribution of the probability of the Bernoulli test). The model can be represented by the Bin(N, P) binomial distribution, where N represents the number of sites (mutations) tested and P = 1-(1-TF) GE is the probability of detection of a single locus. Importantly, the mathematical model predicts the average number of loci detected and the probability of at least one detection as a function of the number of unique DNA fragments (genome equivalents or coverage), mutation load (N, which can also be used as panel size), and TF ( Figure 3B Using this model, we found that integrating more than 20,000 sSNVs (~10 mutations / mb found in 17% of human cancers) with a modest sequencing effort (20X coverage) provides high detection probability (up to 0.98) even at a TF of 1:100,000, and can therefore be easily achieved using standard whole genome sequencing (WGS) ( Figure 3C ).

[0369] Related applications

[0370] Non-invasive prenatal testing (NIPT) for chromosomal aberrations

[0371] The present disclosure further relates to non-invasive prenatal testing (NIPT) for chromosomal aberrations using the aforementioned systems, methods and algorithms. Preferably, Figure 1C and 1E NIPT is performed using the CNV / SV-based workflow outlined in . Here, de novo amplifications and deletions can be called and used to diagnose a subject's sample (e.g., amniotic fluid or blood from a pregnant woman carrying a fetus suspected of having a chromosomal aberration). This method uses the unique log2 / fragment size relationship (the same phenomenon occurs in fetal and normal DNA) to increase sensitivity and specificity, e.g. Fig.18E and Fig.18F .therefore, Figure 1C and Figure 1E The workflow allows researchers or clinicians to combine two sources of information that are only relevant in CNVs generated by fetal DNA, and not related to noise corresponding to sequencing, alignment, and GC artifacts. Therefore, even if prior information about the CNV segment is not readily available, the methods and systems of the present disclosure allow clinicians to obtain higher sensitivity and specificity of NIPT using de novo CNV detection.

[0372] Example 1: Design of a somatic mutation classifier

[0373] When designing a model for somatic mutation classification, it is important to identify sources of error that may lead to false positive somatic mutations. Real mutations may have high base quality regardless of their position in the read. Similarly, the read base, reference base, and alignment string (CIGAR) at the position of the real mutation may be independent of the read alignment. More specifically, it is expected that real somatic mutations are spatially invariant. It is well known that systematic errors in sequencing experiments depend on the position in the read, so although the mutation itself may be invariant in space, its position in the read is usually not invariant. Errors caused by mismapping are likely to contain repetitive sequences or very specific sequence motifs (such as TTAGGG in telomeres). Therefore, it is expected that the model can accurately represent the spatial invariance in real somatic mutations and the errors caused by mapping, while maintaining the base quality of the entire read. Therefore, it is concluded that any shallow convolutional network that relies on a fully connected layer of the target read for classification will not be able to capture the invariance of mutations.

[0374] Recognizing these constraints and / or requirements, the inventors designed a somatic mutation classifier engine. The engine uses a convolutional neural network, using an eight-layer convolutional neural network to correct for spatial dependencies, with a single fully connected layer inspired by the VGG architecture at the end (Simonyan & Zisserman, arXiv: 1409.1556, revised on April 10, 2015; Alexandrov et al., Nature, 500 (7463): 415-421, 2013). All features (columns) at a position are convolved using a perceptive field of size 3. After two consecutive convolutional layers, downsampling is performed by maximum pooling with a receptive field of 2 and a stride of 2, thereby forcing the model to retain only the most important features in a small spatial region. It is expected that this architecture has two main benefits: 1) spatial invariance is maintained when convolution is performed on a three-nucleotide window; 2) "quality mapping" is captured by collapsing the read fragments into 25 segments (each segment represents a region of approximately 8 nucleotides). The output of the last convolutional layer is directly applied to a sigmoid fully connected layer for final classification. In order to retain features related to the position in the readout, a simple logistic regression layer is used instead of a multilayer perceptron or global average pooling.

[0375] The disclosed model and training scheme are referred to as the Engine. The Engine is the first use of a read representation that jointly captures the genomic context of the alignment, the complete read sequence, and an integral of each base quality score. The performance of the Engine demonstrates that integrating the features of the reads and their alignments has laid the foundation for a new set of somatic mutation callers that use the complete mutational profile of a sample, not just those covered at high depth.

[0376] To evaluate the performance of the model, the predictive performance of the model was studied on an independently presented lung cancer dataset. The dataset was paired with healthy WGS data of the same patients. The model was evaluated using the metrics F1 score, accuracy, sensitivity, and specificity, which are defined as:

[0377] Sensitivity = TP / (TP+FN) (Equation 1)

[0378] Accuracy = TP / (TP+FP)……………(Equation 2)

[0379] Specificity = TN / (TN+FN) (Equation 3)

[0380] F1 score = 2 x (precision x recall) / (precision + recall) ... (Equation 4)

[0381] Table 1: Validation test metrics

[0382]

[0383]

[0384] The model was found to achieve an average F1-score of .961 on the validation set. The model achieved an F1 score of .71 on the tumor controls. Although the model remained sensitive on the tumor controls, it showed some decrease in specificity compared to the validation dataset. However, in the case of independent lung samples, an F1 of .92 was observed, indicating high specificity (Table 1). The low precision and specificity in the cancer controls suggest that the engine has learned specific mutation patterns associated with smoking lung cancer, while also learning common erroneous patterns.

[0385] To further examine the learning capabilities of the engine, additional samples from melanoma patients (CA0040; Table 1) were used in the analysis. Melanoma samples often exhibit distinct mutational profiles due to UV exposure compared to those associated with tobacco exposure ( Fig. 8A ). The engine model achieved an F1 score of 0.71 on melanoma samples. Therefore, while the model was still sensitive, the lower accuracy and specificity for melanoma samples suggests that the engine learned specific mutational patterns associated with lung cancer in tobacco exposure, while learning more general sequencing artifact patterns that apply to both tumor types.

[0386] To further examine the above questions, we examined the differences in trinucleotide context frequencies between true cancer mutation variant reads and sequencing artifacts containing reads from the following datasets: (i) lung cancer patient samples included in training (CA0046, validation dataset), (ii) lung cancer patients not included in training (CA0044), and (iii) melanoma patients (CA0040). The results are shown in Figure 8B .

[0387] Note that, as expected, tobacco-associated lung adenocarcinoma samples showed a high enrichment of C>A transitions consistent with the tobacco-associated mutational signature ( Figure 8B ). It was therefore hypothesized that the engine could learn specific sequence contexts that are prevalent in tumor mutation data (i.e., tumor-specific mutation signatures). To test this hypothesis, the difference in frequency between true cancer variants and sequencing artifacts in each trinucleotide context was measured and correlated with the average model prediction for these same reads. It was inferred that a high correlation between trinucleotide sequence frequencies and model outputs would be expected if the model was learning (lung) cancer-specific sequence contexts. By inference, a high correlation between model predictions and trinucleotide enrichment was observed in both CA0046 (included in training, Pearsons r=1) and CA0044 (not included in training, Pearsons r=0.95). The results are shown in Figure 8C .

[0388] To directly examine whether the high correlation is the result of accurate classification independent of sequence context (alternatives), a similar analysis was performed on a melanoma sample (CA0040). The results showed that there was a positive correlation between the trinucleotide context and the model prediction (Pearsons r = 0.64), indicating that accurate classification came from features other than the mutation signature, which was significantly lower than the lung cancer data exposed to tobacco. This finding is consistent with model learning of specific lung cancer mutation signatures. This finding paves the way for training a separate model specifically for detecting melanoma-associated somatic mutations. Using the above-mentioned NSCLC procedure, additional data sets from three melanoma patients were examined. The higher F1 scores in the melanoma validation dataset and independent melanoma samples indicate that these observations are similar in terms of performance, but the F1 score is lower when the model is applied to the NSCLC data (control).

[0389] Engine sensitivity and precision at low tumor fractions in synthetic plasma

[0390] To evaluate the performance of the system and / or method in the context of low tumor scores, the accuracy and sensitivity of the engine were compared with the state-of-the-art callers MUTECT, SNOOPER, and STRELKA. The results are shown in Fig. 9A , showing the superior sensitivity achieved with the engine, especially at low tumor fractions. In contrast, MUTECT was unable to detect more than two mutations in the synthetic samples at any tumor fraction, and made the same call across tumor fractions whenever it successfully made a mutation prediction. Thus, the engine improved sensitivity over MUTECT by more than 200-fold, while improving accuracy over the simple filter at a tumor fraction of 0.01. Based on these surprisingly better results, the systems and methods of the present disclosure were applied to the context of real plasma samples.

[0391] A comparison was also made between the engine and a simple call to the method PILEUP. The results are shown in Fig. 9B and Fig. 9C , the comparative evaluation was performed on the filters implemented using the engine. A comparative evaluation was further performed using another metric called enrichment, which provides information on the increase in the ratio of tumor to normal mutations when the filter is implied. The enrichment factor can be calculated using Equation 5 below.

[0392] Enrichment = precision out / Accuracy in …(Equation 5)

[0393] Although PILEUP has sufficient sensitivity to detect somatic mutations in simulated plasma, it includes all mutations. This is poorly reflected in terms of enrichment and precision metrics. In the next stage of the channel, filters by mutation frequency were used. Although MF and BQ+MQ filters actually depleted samples of tumor reads, an increase in enrichment was observed at TF=0.01. This is a good demonstration that this filter can be used not only to evaluate channels, but also to remove most of the noise before presenting it to CNN. When using the CNN filter, another (third) reduction in the magnitude of the noise was observed. Most importantly, the resulting reduction in sensitivity was only about 25%. In the complete channel, a 30-fold enrichment was observed at both a tumor fraction of 0.01 and a tumor fraction of 0.0001 (relative to PILEUP; green line). The data are shown in Fig. 9C middle.

[0394] Using the engine to analyze somatic mutations in actual cfDNA samples

[0395] In order to determine whether the methods and systems of the present disclosure are robust in an actual clinical setting, two different types of samples were evaluated in practice. The first was a cfDNA sample from a healthy individual (identifiers: BB600; BB601); the second was from an early-stage lung cancer patient, obtained before surgery (identifiers: BB1122; BB1125). In actual clinics, the clinicians performing the test do not have access to mutation information about the patient. However, because BB1125 underwent surgery, the clinicians were able to use a standard mutation calling pipeline to measure true somatic mutations. These calls can be used and combined with reads from cfDNA to obtain a qualified second estimate of the sensitivity, precision, and enrichment of the engine.

[0396] It was found that after applying the filtering channel, 27 of the 413 mutations present in the sample were successfully captured. Most significantly, the false positive rate in the control was suppressed from 266 errors to 3 errors (see Table 2). The results show that although the channel as a whole actually depleted about 50% of the tumor signal, in contrast, the engine enriched the samples it presented by about 1.7 times.

[0397] Table 2: Performance metrics for cell-free DNA samples, lung cancer (BB1125)

[0398]

[0399] The results suggest that differences in the preprocessing steps may have led to a poor setting of the BQMQ filter. It can be inferred that a base quality score of 20 may be too permissive for this sample.

[0400] Recognizing that it may be advantageous to use a training scheme that allows detection of true somatic mutations with high sensitivity while rejecting candidate mutations caused by systematic errors, multiple lung cancer patients were sampled and matched for their systematic error profiles. Four representative samples from multiple smoking lung cancer patients were selected for training to implement the scheme (Table 3).

[0401] Table 3: Training, validation and test datasets

[0402]

[0403]

[0404] Another lung cancer patient who smokes is used for testing. The samples are processed and provided by the Cancer Alliance of the New York Genome Center. These samples have complete tumor samples and healthy tissue samples from the same patient. The consistency of the three callers STRELKA, LOFREQ and MUTECT is used to make the final call of somatic mutations. The reads supporting these mutations are then used as tumor reads for training.

[0405] Since it is expected that the model learns to distinguish from sequencing artifacts, readings from healthy samples containing exactly one mutation are used. Since these variants are not supported by more than one reading, it is likely due to a system error. These low-quality variants are then filtered, and readings (BQ20, MQ40) with a base quality score of less than 20 or a mapping quality of less than 40 at the mutation are filtered. These threshold BQMQ values ​​are determined by inspection; however, a window is generated to allow samples of lower quality to be included in the training. A small subset of the training set is additionally set as a validation data set. This data set is not only used to monitor the progress of training, but also to verify the performance of the model on independent readings (but not independent mutations). The performance of the model is then evaluated on the test lung data set.

[0406] Synthetic plasma

[0407] To test the model's ability to detect somatic mutations at low frequencies, four simulated plasma samples from a test lung sample (CA0044, Table 3) were generated by randomly sampling from the patient's healthy DNA and the patient's tumor DNA. Sampling was performed at a coverage of 35 and a tumor mixture of 0%, 0.01%, 0.001%, and 0.0001%. For robustness, mixing was performed with three random seeds. A threshold ratio of approximately 0.1 was chosen as the somatic mutation ratio in cfDNA. Therefore, when preparing synthetic plasma reads for inference, only samples with less than 1 / 10 of the mixture were selected. th The mutations are supported by the coverage reads.

[0408] To evaluate the performance of the methods and / or systems of the present disclosure in the context of low tumor scores, parameters such as accuracy, sensitivity, and enrichment were compared between the engine and the prior art low-frequency caller MUTECT. Further comparisons were made by including a simple calling method called PILEUP, which allowed any mismatches observed. After PILEUP, the same filters used for the engine were repeatedly applied and the performance of each step was measured. The filters performed in this method were MF (mutation frequency) (which filters PILEUP reads with a higher frequency of occurrence than expected in plasma (10% of the number of mutation occurrences)), BQMQ (which filters reads with a base quality below 20 or a mapping quality below 40 at the mutation), and the final filtering method using the engine.

[0409] Evaluation of cfDNA samples

[0410] After evaluating the engine on synthetic samples, its performance on real plasma DNA samples was tested. Control samples (BB600; BB601) and samples of lung cancer patients who smoke (BB1125 or BB1122) were used in the analysis. Because tumor biopsies were also performed on these patients, true positives were measured by assuming that all mutations called by MUTECT from the biopsy were also present in cfDNA. Using these calls, the same analysis as in synthetic plasma was performed (same as above).

[0411] Assessing sensitivity, precision, and enrichment

[0412] In the case of controls, all measurements were performed for the mutation in subject BB1125.

[0413] Feature Construction

[0414] To fully capture the sequencing reads, alignments, and genomic context, a spatially oriented representation of the reads was created ( Figure 5 ). For insertions in the reference, an "N" is placed at the insertion position in the reference context to maintain spatial alignment. For deletions in the reference, an "N" is placed at the deletion position in the read sequence. The softmasked region is partitioned so that the reads are adjacent to the mapped portion of the read and the reference context is broken with consecutive "N"s until the end of the softmasked region. This is done for two reasons, to ensure that the signal for the softmasked region is strong and secondly, to maintain the idea of ​​reads being independent of their alignment.

[0415] Pad segments (e.g., + / - 25 bp) from the genomic context to both sides of the read ( Figure 6). For a 150 bp read, this would result in a 16x200 bp matrix, with additional context bases added in the case that the read was not 150 bp. The maximum base quality score was set to 40 (p=99:99%), and the scores were scaled to the interval [0,1]. Bases not covered by the read (genomic context) received a base quality score of zero. Deletions in a read received a quality score that was the average of the two flanking positions in the read.

[0416] Hyperparameters and implementation details

[0417] The model was trained using mini-batch stochastic gradient descent with an initial learning rate = :1 and momentum = :9. The learning rate was reduced by a factor of 10 when the validation loss reached a plateau as described by He et al. (in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016). A mini-batch size of 256 was used as it seemed to provide the best trade-off between validation loss and training speed. A base of 64 filters was used, doubling it after each downsampling layer to maintain a consistent number of parameters at each stage. This was chosen empirically after observing that a 32-base filter model did not perform adequately. Batch normalization was applied after each convolutional layer, followed by rectified linear units. A dropout with drop probability = .5 was applied before each pooling layer.

[0418] The engine showed strong performance in a balanced tumor fraction context. It maintained a large portion of its performance even in simulated plasma. The engine also achieved a 2-fold enrichment at a tumor fraction of 0.0001, suggesting that it was able to capture relevant somatic mutations even when they were 10-fold less than the sequencing noise itself. In contrast, MUTECT, a tool not designed to function in a cfDNA context, made fewer than two predictions at all tumor fractions. See Figures 9A-9C .

[0419] Table 4 gives a detailed summary of the engine results.

[0420] Table 4: Summary of engine performance metrics (F1 score, sensitivity, specificity, and precision) on melanoma and lung cancer datasets.

[0421]

[0422] Other Implementations

[0423] Based on the foregoing, the system and method can be developed into a complete early detection engine. Although the engine captures the position in the read by using a fully connected S-shaped layer, some architectures may be more suitable for capturing the relative position in the read. In addition, additional information sources contained in the read pairs from DNA fragments, which are not included in the initial detection, can be used to determine the source chain (Watson or Crick) and estimate the size of the DNA fragments. It has been observed that ctDNA has a different fragment size distribution compared to regular circulating healthy DNA (Underhill et al., PLoS Genetics, 12(7): e1006162, 2016).

[0424] The foregoing systems and methods can be integrated with recurrent neural networks (RNNs). RNNs have been shown to be a powerful tool for using length as a feature in bioinformatics at distances even up to 1 kb, far exceeding the size of ctDNA fragments (Hill et al., bioRxiv, pp. 200758, 2017). Integrating recurrent neural networks instead of logistic regression layers can further improve the performance of the methods and systems of the present disclosure.

[0425] Embodiment 2: Methods and systems for detecting and validating tumor-specific low-abundance tumor markers and their use in cancer diagnosis

[0426] The systems and methods disclosed herein can be used for early diagnosis of cancer. As known in the art, compared with metastatic cancer (characterized by high disease burden and significantly elevated ctDNA), the abundance of ctDNA limits the use of targeted sequencing technology in the case of early cancer or residual disease detection. Considering the known limited amount of cfDNA in the case of low tumor burden, first, the potential for optimizing cfDNA extraction was studied. First, in order to reduce differences and inter-individual differences from sample collection, a uniform cfDNA material generated by large-volume plasma collection (about 300cc) performed by plasmapheresis of healthy subjects and cancer patients undergoing hematopoietic stem cell collection was used to compare commercially available extraction kits and methods. Large amounts of plasma can test a variety of methods and protocol parameters on the same cfDNA input, making it possible to accurately measure subtle differences in yield and quality.

[0427] Kits and / or extraction methods from Capital Biosciences (Gaithersburg, MD, USA; Catalog # CFDNA-0050), Qiagen (Germantown, MD, USA), Zymo (Irvine, CA, USA; Catalog # D4076), Omega BIO-TEK (Norcross, GA, USA; Catalog # M3298), and NEOGENESTAR (Somerset, NJ, USA, Catalog # NGS-cfDNA-WPR) were used in this comparative study. These kits and reagents were used uniformly according to the manufacturer's instructions, and large-volume plasma samples of 1 ml were extracted. Multiple plasma aliquots were processed in parallel to assess inter- and intra-method variability. The yield and purity of each recovered cfDNA sample were determined using fluorescent quantification (total mass), UV absorbance (detection of salt and protein contaminants), and on-chip electrophoresis (size distribution and gDNA contamination).

[0428] The results showed that the MAG-BIND cfDNA extraction kit from Omega BIO-TEK outperformed all other methods tested. Each step of the manufacturer's protocol was further systematically optimized to reduce contaminant carryover and improve cfDNA recovery.

[0429] The optimized extraction protocol was then applied to samples from early lung cancer. The group included 11 preoperative early lung cancer plasma samples and 4 plasma samples from benign patients (controls). Exemplary patient characteristics are Fig.11 Despite optimization of the extraction method, the yield of cfDNA in low disease burden samples remained low and showed high variability between patients, ranging from 0.13 ng / mL to 1.6 ng / mL. These data confirm the small and variable number of DNA molecules available for cfDNA sequencing.

[0430] Sequencing width can replace sequencing depth to overcome the limitation of cfDNA abundance in sensitive cancer detection.

[0431] The above data indicate that the detection of single sSNVs in patient plasma samples comes from two consecutive statistical sampling processes. The first process provides the probability of sampling mutant fragments with a limited number of genomic equivalents present in a typical blood sample. The second process, given its abundance, sequencing depth, and sequencing errors (signal-to-noise ratio), assesses the probability of detecting mutant fragments in the sample. Although the latter process has been the focus of intensive research and technological development in the scientific community (e.g., ultra-deep error-free sequencing schemes), the stochastic process of the former has rarely been addressed. However, in the detection of ctDNA with low disease burden, both processes play an important role as shown above. If there is no physical fragment representing the targeted sSNV, then even ideal ultra-deep targeted sequencing cannot detect cancer signals, which is considered to be one of the main factors leading to the limited sensitivity of these methods (~40%, Rosenfeld et al.). In practice, the fact that a single observation (mutation read) is rarely sufficient for reliable detection further complicates this problem.

[0432] To determine the probability of sampling a mutant fragment in a given cfDNA sample, cfDNA sampling was modeled as a Bernoulli experiment using a mixture of cfDNA fragments originating from two populations, whose proportions were defined by the tumor fraction (TF), namely cfDNA fragments originating from normal cells and cfDNA fragments originating from malignant origin. Therefore, the genomic equivalents present in the plasma sample constitute a random sampling of the entire pool of cfDNA fragments in the patient's circulation. Therefore, the probability of sampling at least one mutant fragment in a plasma sample supporting a specific permutation can be defined as: P = 1-(1-TF) GE , where P is the probability; TF is the tumor fraction; and GE corresponds to the number of genomic equivalents present in the patient's cfDNA. The model predicts that the probability of detection of TF associated with early cancer scenarios (TF < 1%) will show a rapid decline for low TFs, and even at a frequency of 0.1% (1 / 1000), the probability of detection is expected to be less than 0.65 ( Figure 3A ). Notably, these limitations were observed even under ideal conditions where exhaustive sequencing could be efficiently utilized with 1000 genome equivalents (~6 ng cfDNA) and detection based on a single supporting DNA fragment with an ideal signal-to-noise ratio. These results suggest that plasma sampling probability imposes a strict upper limit on mutation detection for low TF scenarios such as MRD and early cancer stage detection.

[0433] On the contrary, this model also shows that the limitation on sequencing depth can be effectively overcome by increasing the number of detected sites (SNVs) by increasing the width, which is due to repeating the Bernoulli test for each SNV (binomial distribution of the probability of the Bernoulli test). The model can be represented by the Bin(N, P) binomial distribution, where N represents the number of sites (mutations) tested and P = 1-(1-TF) GE is the probability of detection of a single locus. Importantly, the mathematical model predicts the average number of loci detected and the probability of at least one detection as a function of the number of unique DNA fragments (genome equivalents or coverage), mutation load (N, which can also be used as panel size), and TF ( Figure 3B Using this model, we found that integrating more than 20,000 sSNVs (~10 mutations / mb found in 17% of human cancers) with a modest sequencing effort (20X coverage) provides high detection probability (up to 0.98) even at a TF of 1:100,000, and can therefore be easily achieved using standard whole genome sequencing (WGS) ( Figure 3C ).

[0434] In silico validation of integrated genome-wide sSNV detection

[0435] The model indicated that increasing the number of loci resulted in a significant increase in the probability of detection. To validate this prediction, cfDNA testing was simulated using in silico mixtures of tumor and normal WGS data from 11 cancer patients with various cancer types, including high-grade tumors from lung adenocarcinoma, ductal adenocarcinoma (breast), skin melanoma, urothelial carcinoma (bladder), and osteosarcoma (full clinical details for protocol A; Figure 1F ).

[0436] All samples were deeply sequenced with ~80X tumor WGS and ~40X PBMC WGS. To generate in silico mixtures, tumor and normal WGS reads were mixed in different ratios and patient-specific virtual plasma samples with different TFs (0.00001, 0.00005, 0.0001, 0.0005, 0.001, 0.005, 0.01) and coverages (5, 10, 15, 20, 35) were obtained. Each condition had 5 independent replicates, which were obtained by different random seeds used in the downsampling process. To simulate detection in the context of residual disease, somatic mutation calls were made on the original tumor and germline WGS data, and patient-specific sSNV summaries were obtained. The number of tumor-associated mutation sites in the in silico plasma simulation mixture was then measured by detecting at least one supporting read of the patient-specific sSNV summary. It has been found that integrating many sites also leads to the accumulation of noise caused by sequencing errors, which may limit the detection of the above signals. To estimate the degree of noise in WGS-based cfDNA detection, a complementary database of downsampled normal reads (TF=0, 20 replicates at 20X and 35X coverage) was generated without mixing reads from tumor WGS. These data enabled signal-to-noise measurements and demonstrated that integrated whole-genome SNV detection can confidently detect TF>1:2000 in high mutation burden tumors at 20X coverage for a variety of tumor types.

[0437] These data also show how noise from sequencing errors shapes the relationship between the number of sites detected and TF, as the relative contribution of noise increases with decreasing TF. Comparison with predictions from a mathematical model that incorporates estimated sequencing noise shows high consistency across different TF and coverage values ​​for all patients and cancer types. The analysis also demonstrates how increases in mutation load (N) and coverage increase the detection signal, with the number of detections at 1% TF varying from 40K mutation load (melanoma) to 8K mutation load (non-tobacco lung).

[0438] Characterization of the variables that form the basis of the estimated noise and the development of optimized filters can significantly improve the signal-to-noise ratio and detection sensitivity. The noise distribution is modeled by other independent variables such as mutation load, coverage, and cancer type. The results show an error probability independent of cancer type, which reflects the previously published sequencing error rate (~1 / 1000 bases). Moreover, the detected signal shows a patient-specific relationship with negligible germline-related noise.

[0439] The data showed that sequencing errors were found to be associated with parameters such as base quality (BQ), mapping quality (MQ), fragment length, and variable allele frequency (VAF). Therefore, in order to reduce the sequencing error rate, a filter that combined base quality (BQ) and mapping quality (MQ) optimization was developed through optimal receiver operating characteristic (ROC) analysis, which reduced the measured error rate by 3FC (to approximately 3x10 -4 ). Applying this filter with reduced 35X coverage allowed detection of markers even when the TF of the tumor dropped to 1 / 20,000 at a WG depth of 35X. These data support the use of patient-matched integrated genome-wide sSNV maps to achieve cancer detection probability even at very low TF, independent of cfDNA abundance (e.g., 1 ng of input allows 100X WGS). Furthermore, the high agreement between the experimental results and the mathematical model suggests that the number of loci detected (patient-specific sSNVs) can be converted to an estimate of plasma TF, allowing quantitative TF monitoring in the early detection setting.

[0440] Other parameters besides quality metrics can be further used to filter the remaining noise, including leveraging information about specific motifs, features, etc. Representative methods include, for example, implementing fragment size filters (e.g., only considering fragments of about 200 bp or shorter) and variable allele frequency (VAF) filters (e.g., only considering alleles with a VAF greater than a threshold value (e.g., 2%, 5%, 10%)). The various mutational signatures for tobacco exposure and UV exposure are shown in Fig. 12A The top and bottom panels of Figure 2 show the differentially expressed COSMIC signatures in lung tumors, breast tumors, and melanoma samples. Fig. 12B and 12C middle.

[0441] application

[0442] This sensitive de novo mutation detection was then applied to sequenced preoperative plasma from five early-stage patients, generating a genome-wide cfDNA mutation detection. The genome-wide mutation data were aggregated to calculate a mutational profile for each patient, and then a new analysis approach was performed using new machine learning algorithms and tools, such as convolutional neural networks (CNNs), for sensitive mutation signature detection.

[0443] The CNN is based on the application of a two-pronged strategy - first, a deep learning algorithm was trained on a pan-lung cancer cohort (5 patients with deep tumors and PBMC WGS) using supervised learning to identify signatures that distinguish true tumor mutations from artifacts. The obtained model was then used to infer and assign confidence estimates to each individual mutation detected in early detection plasma samples of patients with early lung adenocarcinoma. Secondly, the detection signal was obtained by integrating these confidence estimates across the entire genome, and then a new analytical method for sensitive detection of specific cosmic mutation signatures in individual plasma samples was performed using non-negative least squares (NNLS). The confidence of signature detection was further verified using a comparison of cosmic mutation exposure values ​​with those inferred for 100 random background signatures (zscore>2STD).

[0444] The results are shown in Figure 13, which shows that the CNN of the present disclosure is particularly useful in early tumor detection (ED). By this method, in each case, even if the TF is 1 / 1000 or lower, a tobacco-specific signature was detected in lung cancer patients; a UV-specific signature was detected in melanoma patients; and a BRCA-specific signature was detected in breast cancer patients. In order to evaluate the ability of this method to improve the low positive predictive value (PPV) of current lung cancer CT screening in subjects at risk of exposure to tobacco, the method was applied to plasma samples from 5 early lung patients and 4 benign nodules, all of which were detected as positive in CT screening. The data show that early lung cancer samples were positively detected, while benign nodules were low (false) positive detections, thus showing improved PPV.

[0445] Next, the patient-specific signature score (zscore) was mapped to patient characteristics, such as smoker or non-smoker, smoking history, such as the number of pack-years of smoking per patient (for smokers), including histopathological features such as nodule detection positive or negative (ND). The results are shown in Fig.14A , which showed that the tobacco signature was detected in the plasma of early cancers in patients with tobacco exposure, but not in patients with benign nodules or no smoking history. By this method, it was possible to detect the tobacco signature in 3 of 4 patients with early lung disease who had previous tobacco exposure, while the signature was not detected (ND) in plasma samples of non-smoking lung patients and 3 individuals who had undergone resection of benign lung nodules. In all stages except one, the specificity of the tobacco signature in the detection of lung cancer patients was at least 67%, and in patients with advanced stages (e.g., stage IIIa and later), the specificity was close to 100%.

[0446] The use of diagnostic methods and CT screening can improve PPV

[0447] To improve the low positive predictive value (PPV) of the CT screening approach, the aforementioned screening approach was applied to diagnose / predict at-risk tobacco-exposed subjects with or without CT screening. First, markers and signatures (including SNVs, CNVs, indels, and / or SVs) were detected de novo by whole genome sequencing (WGS), and the noise / error of the markers was analyzed using the above-mentioned approach. In this way, a total of 30 preoperative samples collected from early-stage NSCLC patients (stages I and II) were analyzed. In addition, WGS was performed on 30 age-matched patients with tobacco exposure who were found to have benign lesions through the institutional CT-based screening program. The detection signal from the cfDNA data was integrated with the CT-based reads in a blinded manner to determine whether the positive predictive value of CT screening could be improved using cfDNA information. It was estimated that the cohort had the power to detect a 20% increase in the effect size of the PPV, from approximately 40% with the present approach to approximately 60% using integrated cfDNA and CT screening. Based on the results of the study, a large-scale prospective institutional clinical trial can also be conducted.

[0448] discuss

[0449] The data show that the methods and / or systems disclosed herein are superior to existing methods, especially in the case of low-abundance markers, which are used for early detection (ED) of tumors. In early cancer detection, matching tumor DNA is not available, which requires challenging de novo cancer variant detection. The whole genome integration method disclosed herein uses sSNV sequence context information to detect mutation signatures associated with specific mutagenic processes (such as exposure to tobacco, UV light, excessive APOBEC activity, BRCA mutations, PARP activity, or MSI). In all tested samples, these signatures appeared specifically in tumor somatic mutations and were completely absent in PBMC somatic mutations.

[0450] Sensitive and specific de novo mutation detection in cfDNA of low TF samples is fundamentally challenging for existing mutation detection algorithms. All methods known in the art focus on comparing tumor and normal DNA at specific genomic sites. The ability to detect mutation sites in the genome derives from the observation of multiple supporting reads covering the site, which are then subjected to a statistical framework to distinguish these multiple observations from sources of sequencing noise (sequencing errors, mapping errors, etc.). However, in the case of early detection, the amount of mutated ctDNA is much lower than the sequencing depth (or the number of fragments available for sequencing at a specific site), so at best, only a single supporting read is observed per site. For example, when applying MUTECT on virtual plasma data, the number of true tumor-associated somatic mutations decreases rapidly with decreasing TF, even when considering all detections included in the calling statistics file (before variant filtering), but more such mutation sites are called when detections by a single supporting read are considered.

[0451] To allow error-free de novo single ctDNA detection at low TFs, a new framework is needed that enables the distinction between alternative reads originating from cancer mutations and reads generated by sequencing artifacts. Although mutation signatures have typically exploited trinucleotide context, recent data suggest that sequence context may extend beyond this range and may be difficult to capture through supervised feature selection.

[0452] The present disclosure provides a new method and channel for filtering sequencing errors. For example, tumors produced by specific mutation processes produce unique mutation patterns th...

Claims

1. A system for detecting cancer in a subject in need thereof, comprising: An analysis unit comprising a binning engine configured and arranged to receive a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of a subject, the biological sample comprising a plasma sample, wherein the subject has a tumor fraction in the sample of about 0.0001% to about 2%, wherein the read summaries each include copy number variations (CNVs); partition the read summary into a plurality of windows; and calculate a set of features for each window, the features comprising a median depth of coverage for each window and a representative fragment size for each window; a pre-filter engine configured and arranged to filter out artificial sites from the read summaries, wherein the filtering comprises removing duplicate sites generated on a reference healthy sample cohort from the read summaries, wherein the filtering is performed by machine learning, and the machine learning comprises a deep convolutional neural network (CNN); and a normalization engine configured and arranged to normalize the read summary to produce a filtered set of reads for the whole genome read summary; and a computing unit configured and arranged to compute an estimated tumor score using the filtered set of reads by: computing a linear relationship between the feature sets for each window and converting the computed relationship into an estimated tumor score using a regression model; and screening the subject for cancer if the estimated tumor score exceeds an empirical threshold; The tumor fraction refers to the ratio of tumor DNA (ctDNA) without circulating cells to the total amount of cell-free DNA (cfDNA).

2. A system for detecting cancer in a subject in need thereof, comprising: An analysis unit comprising a binning engine configured and arranged to receive a subject-specific genome-wide read summary associated with a plurality of genetic markers from a biological sample of a subject, the biological sample comprising a plasma sample, wherein the subject has a tumor fraction in the sample of about 0.0001% to about 2%, wherein the read summaries each include copy number variations (CNVs); partition the read summary into a plurality of windows; and calculate a set of features for each window, the features comprising a median depth of coverage for each window and a representative fragment size for each window; a pre-filter engine configured and arranged to filter out artificial sites from the read summaries, wherein the filtering comprises removing duplicate sites generated on a reference healthy sample cohort from the read summaries, wherein the filtering is performed by machine learning, and the machine learning comprises a deep convolutional neural network (CNN); and a normalization engine configured and arranged to normalize the read summary to produce a filtered set of reads for the whole genome read summary; and a computing unit configured and arranged to compute an estimated tumor score using the filtered set of reads as a function of a set of features for each window computed in the subject-specific genome-wide read summary based on one or more integrated mathematical models; and screening the subject for cancer if the estimated tumor score exceeds an empirical threshold; The tumor fraction refers to the ratio of tumor DNA (ctDNA) without circulating cells to the total amount of cell-free DNA (cfDNA).

3. The system of claim 1 or 2, wherein the marker comprises a copy number variation (CNV).

4. The system of claim 1 or 2, wherein in the binning engine, each window is at least 150 bp.

5. The system of claim 1 or 2, wherein the pre-filter engine extracts depth coverage and fragment size relationships from whole genome feature vectors.

6. The system of claim 1 or 2, wherein the pre-filter engine filters duplicate sites generated on a reference healthy plasma sample cohort by generating a normal panel (PON) blacklist or mask; and / or filters windows of low mappability or coverage.

7. The system of claim 1 or 2, wherein the normalizing step comprises normalizing the depth coverage values ​​by performing two LOESS regression curve fits on the bin-by-bin GC scores and the mappability scores to correct for GC content and mappability bias.

8. The system of claim 1 or 2, wherein the normalizing step comprises batch effect correction using robust zscore normalization, the robust zscore normalization being applied to each sample separately.

9. The system of claim 8, wherein the zscore normalization comprises calculating a median and a median absolute deviation (MAD) based on the neutral region for each sample, and normalizing all CNV bins by subtracting the median and dividing the differential by the MAD.

10. The system of claim 1 or 2, wherein the normalization engine calculates depth coverage skewness and / or fragment size centroid (COM) skewness in plasma samples compared to a normal panel (PON) healthy plasma samples.

11. The system of claim 1 or 2, wherein the computing unit uses copy number variation (CNV) calling and calculating the tumor score of the filtered read set by using a hidden Markov model or a self-organizing neural network.

12. The system of claim 11, wherein the self-organizing neural network comprises a neural network based on adaptive resonance theory (ART) or a self-organizing map (SOM).

13. The system of claim 1 or 2, further comprising performing orthogonal integration on secondary features including fragment size offsets.

14. The system of claim 13, wherein intra-patient fragment size deviations in tumor-specific markers and random marker lists are analyzed using statistical methods.

15. The system of claim 14, wherein the statistical method comprises a significance test or a Gaussian mixture model (GMM).

Images