Detection of somatic single nucleotide variants from cell-free nucleic acids for minimal residual disease monitoring

Abstract

The present disclosure provides a probability model for accurate and sensitive somatic single nucleotide variant (SNV) detection in a cell-free nucleic acid sample, the sample comprising a sequence data set. A joint genotype can be determined for each locus in the sequence data set, and germline mutations can be inherently removed. A set of filters can be applied to eliminate low-quality somatic variant calls. In addition, overall tumor cell-free deoxyribonucleic acid (cfDNA) scores and overlapping read partners can be considered, so that accurate SNV detection and variant allele frequency estimation can be performed from samples with low tumor cfDNA scores. By using a probability model and a machine learning model to distinguish true variants from sequencing errors, a sensitive early detection of minimal residual disease (MRD) is designed.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Patent Application No. 62 / 559,366, filed on September 15, 2017, which is expressly incorporated herein by reference in its entirety.

[0003] Government Interest Statement

[0004] This invention was made with government support from the National Institutes of Health under Grant HL108634. The government has certain rights in this invention. Background Art

[0005] Somatic mutations can accumulate in cells throughout the life of a subject. Although most such mutations may have little or no obvious effect, some may change genes and / or key cellular functions, and thus produce phenotypic changes. The product of somatic mutations can be cancer, which is due to cell clonal expansion and escape from both the built-in (in-built) program of normal somatic cell behavior and exogenous restrictions on cell proliferation. Somatic mutations that trigger cancer progression can be referred to as "driver mutations", and somatic mutations that do not cause phenotypic or biological consequences can be referred to as "passenger mutations".

[0006] Analysis of driver mutations that may be commonly found in tumors may be essential for analyzing cancer pathology, cancer diagnosis, precision oncology, and prognosis. Therefore, being able to obtain the genetic profile of a tumor may be important in both medical treatment and clinical research. According to some methods, the genetic profile of a tumor may be obtained from a surgical specimen or a biopsy specimen. However, since these methods may be expensive, invasive, and may endanger the health of the subject or patient, these methods cannot always be performed.

[0007] Notably, both somatic cells and tumors can release their nuclear and mitochondrial deoxyribonucleic acid (DNA) and / or ribonucleic acid (RNA) into the bloodstream of a subject during cell destruction (e.g., apoptosis and necrosis). As a result, genetic information of tumor cells can be found in the bloodstream of a subject and extracted from plasma.

[0008] Recently, cell-free DNA (cfDNA) and cell-free RNA (cfRNA) found in plasma have been considered as biomarkers with great potential in cancer diagnosis and prognosis. However, due to the nature of the mixture of cfDNA (e.g., containing both tumor and non-tumor), it may be difficult to detect cancer-related alleles from cfDNA sequencing data. For example, specialized analysis may be required to detect tumor-derived DNA (ctDNA) and its mutations because its quantity and quality in cfDNA can vary significantly. For example, the fraction of ctDNA in ctDNA can be 0.01% to 97%, while for early cancer, the fraction can usually be less than 10%. Due to the non-invasive nature of cfDNA, plasma samples can be easily and sequentially obtained for diagnosis and prognosis, thereby providing a large amount of information related to cancer over a period of time. Therefore, detecting cancer mutations in ctDNA can provide a promising solution for early cancer diagnosis, monitoring, and monitoring of minimal residual lesions.

[0009] Regarding the detection of single nucleotide variants (SNVs) or somatic mutations, many existing methods designed for impure solid tumor samples (such as VarScan2 and MuTect) may not be easily used for samples with tumor fractions as low as cfDNA. Regardless of the special characteristics of cfDNA, such methods may not be able to identify somatic SNVs even with sufficient evidence at low tumor fractions. In addition, methods developed for detecting somatic SNVs in cfDNA include Capp-Seq with iDES and SiNVICT. Capp-Seq with iDES may refer to an experimental error suppression technology based on a targeted sequencing group, which can sensitively detect cancer mutations in blood cfDNA samples of subjects (such as patients with intermediate to advanced cancer). This method can solve the problem of low tumor fractions by deep sequencing to capture tumor signals in cfDNA. However, due to the expensive cost of deep sequencing, it can only sequence a small part of the genomic region, thereby limiting its detection capabilities for different cancer types and its application to personalized treatment and outcome prediction. SiNVICT may refer to the computation of SNV callers that use a Poisson function to simulate the allele distribution in cfDNA. This probabilistic model may be too simple and may not take into account the impure nature of cfDNA, resulting in a large number of false positives. In addition, these methods may not take into account overlapping read mates in cfDNA sequencing data that can provide valuable information for detecting mutations in cfDNA. SUMMARY OF THE INVENTION

[0011] In view of the foregoing, the present disclosure provides a new probabilistic method to detect somatic single nucleotide variants (SNVs) from cell-free deoxyribonucleic acid (cfDNA) or cell-free ribonucleic acid (cfRNA), which not only solves the unique challenges in cfDNA / cfRNA sequencing data (e.g., low tumor fraction and low genome coverage and high overlap rate (for double-end sequencing data)), but also estimates the tumor fraction in cfDNA / cfRNA (hereinafter generally referred to as cfDNA or cfRNA) for cancer diagnosis, prognosis and minimal residual disease (MRD) detection after surgery. Specifically, the present disclosure provides a method for non-invasive cancer detection and diagnosis using cfDNA examination. In doing so, the methods and systems of the present disclosure can use a Bayesian-based probabilistic framework to estimate the likelihood of a certain genotype from a sequencing data set. Posterior estimation maximization can be used to determine the genotype of an object (e.g., a patient) and call somatic SNV candidates. Then, a set of criteria that adapts to the characteristics of cfDNA in terms of strand bias, variant quality, sequence background, and known single nucleotide polymorphisms (SNPs) can be used to filter somatic SNV candidates. In addition, both sequencing errors and low tumor variant allele frequencies in plasma samples can be considered. This method can help detect somatic SNVs and ensure good performance. This method may include one or more steps as described below.

[0012] According to one embodiment, the first step includes estimating the tumor score by combining the locus with the variant allele population observed in the sequencing data. The overall tumor score estimate can effectively suppress sequencing noise and avoid the serious effects of sampling fluctuations and sequencing errors (especially low or medium depth sequencing). This can not only improve accurate somatic SNV detection, but also can be a marker with clinical value in cancer diagnosis and MRD detection.

[0013] According to one embodiment, the second step comprises determining SNVs by maximizing the posterior of the combined tumor-normal genotype given the overall tumor fraction.This approach can be contrasted with other approaches that compare reads containing a reference allele and a variant allele.

[0014] According to one embodiment, the third step includes filtering somatic SNV candidates using a set of cfDNA feature criteria for strand bias, variant quality, sequence context, and known SNPs. Such post-call filtering ensures high quality of variant calls in cfDNA.

[0015] According to one embodiment, the fourth step comprises re-estimating somatic SNVs after filtering. As described with respect to the first step, this can be done using the overall tumor score.

[0016] The methods and systems of the present disclosure can be particularly advantageous because (1) the overall tumor fraction obtained in the first step can provide the expected level of tumor reads at individual loci even at low tumor fractions, so that somatic SNVs with low mutant frequencies can still be distinguished from random sequencing errors; (2) in the presence of mixed normal and tumor-derived cfDNA, the joint genotype model can better fit the cfDNA sequencing data because it fully utilizes both normal and tumor signals by genotyping normal and tumor cfDNA simultaneously; (3) computational sequencing error suppression by machine learning models can effectively distinguish true variants from sequencing errors based on individual read information; and (4) post-filtration of somatic SNV candidates can be specifically designed based on cfDNA characteristics.

[0017] In view of these advantageous features, the methods and systems of the present disclosure can be applied not only to deeply targeted cfDNA sequencing data, but also to medium-depth cfDNA sequencing data, such as whole exome sequencing data. Therefore, the methods and systems of the present disclosure can provide sensitive mutation calls and mutation spectra of cancer from cfDNA based on input data.

[0018] As an exemplary application of the present invention, the mutation spectrum obtained using the method and system of the present disclosure can be used for early detection of minimal residual disease (MRD) after surgery. The detection can be based on known somatic mutations, which are detected from the patient's matched tumor normal sample or from cfDNA in the patient's plasma before the patient's surgery. Truncated mutations can be selected for monitoring based on paired mutation call results. Compared with the background in follow-up (follow-up) plasma samples, MRD can be determined by the statistically significant cumulative occurrence of truncated mutations.

[0019] According to one aspect, the present disclosure provides a method for detecting somatic single nucleotide variants (SNVs) from cell-free nucleic acids, such as deoxyribonucleic acid (cfDNA) and ribonucleic acid (cfRNA), the method comprising: estimating an overall tumor cfDNA fraction for a cfDNA sample comprising a sequence dataset; determining a genotype likelihood for at least one locus k in the sequencing dataset; eliminating germline polymorphisms from the cfDNA sample comprising the sequence dataset; filtering SNV candidates through a set of filters; and analyzing the cfDNA sample using a probability model incorporating the estimated overall tumor cfDNA fraction to determine the fraction of tumor-derived DNA (ctDNA) in the cfDNA sample.

[0020] In some embodiments, for cfDNA samples, estimating the overall tumor cfDNA score includes: combining information from all potential SNV sites in the cfDNA sample to reduce the noise caused by sequencing errors occurring at one or more potential SNV sites. In some embodiments, determining the genotype likelihood includes: calculating the likelihood of observing the reading of the covering locus given different joint genotypes. In some embodiments, determining the genotype likelihood includes: estimating the genotype at the locus, which maximizes the posterior probability. In some embodiments, estimating the genotype at the locus includes: determining the joint genotype of each locus in the sequence data set. In some embodiments, filtering SNV candidates includes at least one of the following: filtering SNV candidates based on a chain bias filter; filtering SNV candidates based on a base quality filter; filtering sequencing reads based on a read partner filter; filtering SNV candidates based on a read partner filter; and filtering sequencing reads based on a machine learning model that classifies true variants and sequencing errors. In some embodiments, filtering sequencing reads based on a machine learning model that classifies true variants from sequencing errors includes: establishing ground-truth training data of sequencing reads that contain true variants or sequencing errors; creating a signature profile for each sequencing read that has information including: sequencing quality of each base in the read, read alignment information, sequence background (e.g., read sequence and insertion / deletion), and insert size for paired-end sequencing data; training a classifier based on the training data to classify sequencing reads with true variants and reads with sequencing errors by using the signature profile of each read; and classifying each cfDNA sequencing read as a read with a true variant or a read with a sequencing error using the trained classifier.

[0021] According to another aspect, the present disclosure provides a system for detecting somatic single nucleotide variants (SNVs) from cell-free nucleic acids, such as deoxyribonucleic acid (cfDNA) and ribonucleic acid (cfRNA), the system comprising: a computer memory; one or more computer processors communicatively coupled to the computer memory, the one or more computer processors configured to implement a method comprising: estimating an overall tumor cfDNA fraction for a cfDNA sample comprising a sequence data set; determining a genotype likelihood for at least one locus k in the sequencing data set; eliminating germline polymorphisms from the cfDNA sample comprising the sequence data set; filtering SNV candidates through a set of filters; and analyzing the cfDNA sample using a probability model incorporating the estimated overall tumor cfDNA fraction to determine the fraction of tumor-derived DNA (ctDNA) in the cfDNA sample.

[0022] In some embodiments, for cfDNA samples, estimating the overall tumor cfDNA score includes: combining information from all potential SNV sites in the cfDNA sample to reduce the noise caused by sequencing errors occurring at one or more potential SNV sites. In some embodiments, determining the genotype likelihood includes: calculating the likelihood of observing the reading of the covering locus given different joint genotypes. In some embodiments, determining the genotype likelihood includes: estimating the genotype at the locus, which maximizes the posterior probability. In some embodiments, estimating the genotype at the locus includes: determining the joint genotype of each locus in the sequence data set. In some embodiments, filtering SNV candidates includes at least one of the following: filtering SNV candidates based on a chain bias filter; filtering SNV candidates based on a base quality filter; filtering sequencing reads based on a read partner filter; filtering SNV candidates based on a read partner filter; and filtering sequencing reads based on a machine learning model that classifies true variants and sequencing errors. In some embodiments, filtering sequencing reads based on a machine learning model that classifies true variants from sequencing errors includes: establishing baseline true training data of sequencing reads containing true variants or sequencing errors; creating a signature profile for each sequencing read having information including: sequencing quality of each base in the read, read alignment information, sequence background (e.g., read sequence and insertion / deletion), and insert size for paired-end sequencing data; training a classifier based on the training data to classify sequencing reads with true variants and reads with sequencing errors by using the signature profile of each read; and classifying each cfDNA sequencing read as containing a read with a true variant or containing a read with a sequencing error using the trained classifier.

[0023] According to another aspect, the present disclosure provides a method for detecting minimal residual disease (MRD) from plasma samples, leukocytes, and resected tumor samples (if any) collected before and after surgery, the method comprising: identifying one or more truncating mutations and a mutation spectrum of each of the one or more truncating mutations from a pre-operative blood sample and / or a resected tumor sample; and detecting MRD using a follow-up plasma cfDNA sample after surgery.

[0024] In some embodiments, detecting MRD using a follow-up plasma cfDNA sample after surgery includes: extracting reads covering the position of the truncated mutation, and using the reads classified as having true variants to calculate the MRD prediction score. In some embodiments, using the reads classified as having true variants to calculate the MRD prediction score includes: sampling k sites in the genome that do not contain the identified mutations but match the characteristics of those k truncated mutations; filtering reads identified as containing errors; and generating an MRD prediction score.

[0025] According to another aspect, the present disclosure provides a system for detecting minimal residual disease (MRD) from plasma samples, white blood cells, and resected tumor samples (if any) collected before and after surgery, the system comprising: a computer memory; one or more computer processors communicatively coupled to the computer memory, the one or more computer processors configured to implement a method comprising: identifying one or more truncating mutations and a mutation spectrum for each of the one or more truncating mutations from a pre-operative blood sample and / or a resected tumor sample; and detecting MRD using a follow-up plasma cfDNA sample after surgery.

[0026] In some embodiments, detecting MRD using a follow-up plasma cfDNA sample after surgery includes: extracting reads covering the position of the truncated mutation, and using the reads classified as having true variants to calculate the MRD prediction score. In some embodiments, using the reads classified as having true variants to calculate the MRD prediction score includes: sampling k sites in the genome that do not contain the identified mutations but match the characteristics of those k truncated mutations; filtering reads identified as containing errors; and generating an MRD prediction score.

[0027] According to another aspect, the present disclosure provides a system for detecting minimal residual disease (MRD) from plasma samples, white blood cells, and resected tumor samples (if any) collected before and after surgery, the system comprising: a computer memory; one or more computer processors communicatively coupled to the computer memory, the one or more computer processors configured to implement a method comprising: identifying one or more truncating mutations and a mutation spectrum for each of the one or more truncating mutations from a pre-operative blood sample and / or a resected tumor sample; and detecting MRD using a follow-up plasma cfDNA sample after surgery.

[0028] In some embodiments, detecting MRD using a follow-up plasma cfDNA sample after surgery includes: extracting reads covering the position of the truncated mutation, and using the reads classified as having true variants to calculate the MRD prediction score. In some embodiments, using the reads classified as having true variants to calculate the MRD prediction score includes: sampling k sites in the genome that do not contain the identified mutations but match the characteristics of those k truncated mutations; filtering reads identified as containing errors; and generating an MRD prediction score.

[0029] According to another aspect, the present disclosure provides a non-transitory storage medium storing a set of instructions that, when executed, cause one or more computer processors to detect somatic single nucleotide variants (SNVs) from cell-free nucleic acids, such as deoxyribonucleic acid (cfDNA) and ribonucleic acid (cfRNA), the set of instructions comprising the following instructions: for a cfDNA / cfRNA sample comprising a sequencing data set, overlapping read mates; for a cfDNA / cfRNA sample comprising a sequence data set, an overall tumor cfDNA score; for a locus in the sequencing data set, a genotype likelihood; eliminating germline polymorphisms from a cfDNA / cfRNA sample comprising a sequence data set; using an incorporated The methods include analyzing cfDNA / cfRNA samples to determine the fraction of tumor-derived DNA (ctDNA / ctRNA) in cfDNA / cfRNA samples using a probabilistic model for the estimated overall tumor cfDNA fraction; eliminating low-quality somatic SNV candidates using a set of filters that incorporate characteristics of cfDNA / cfRNA samples; eliminating inconsistent overlapping read partners from cfDNA / cfRNA samples comprising sequencing datasets; eliminating reads with sequencing errors using a machine learning model that distinguishes sequencing errors from true variants; and determining early minimal residual disease (MRD) from cfDNA / cfRNA samples using truncated mutation profiles and a machine learning model.

[0030] In some embodiments, for cfDNA / cfRNA samples, estimating the overall tumor cfDNA score includes: combining information from all potential SNV sites in the cfDNA / cfRNA sample to reduce the noise caused by sequencing errors occurring at one or more potential SNV sites. In some embodiments, determining the genotype likelihood includes: calculating the likelihood of observing the reading of the coverage locus given different joint genotypes. In some embodiments, determining the genotype likelihood includes: estimating the genotype at the locus, which maximizes the posterior probability. In some embodiments, estimating the genotype at the locus includes: determining the joint genotype of each locus in the sequence data set. In some embodiments, eliminating reads with sequencing errors includes: training a machine learning model with samples sequenced multiple times; determining bases from true variants or sequencing errors in the sequencing data set. In some embodiments, determining early minimal residual disease (MRD) from cfDNA / cfRNA samples includes: determining truncated mutations from mutation spectra and variant allele frequencies called from cfDNA / cfRNA samples; combining multiple variant loci to enhance tumor signals in ultra-low tumor score samples; and determining MRD status by within-sample statistical tests.

[0031] According to another aspect, the present disclosure provides a method for detecting somatic single nucleotide variants (SNVs) from cell-free nucleic acids, such as deoxyribonucleic acid (cfDNA) and RNA (cfRNA), the method comprising: for a cfDNA / cfRNA sample comprising a sequencing data set, merging overlapping read partners; for a cfDNA / cfRNA sample comprising a sequence data set, estimating an overall tumor cfDNA fraction; for a locus in the sequencing data set, determining a genotype likelihood; eliminating germline polymorphisms from the cfDNA / cfRNA sample comprising the sequence data set; analyzing the cfDNA / cfRNA sample using a probability model incorporating the estimated overall tumor cfDNA fraction to determine the fraction of tumor-derived DNA (ctDNA / ctRNA) in the cfDNA / cfRNA sample; eliminating low-quality somatic SNV candidates using a set of filters incorporating characteristics of the cfDNA / cfRNA sample; eliminating inconsistent overlapping read partners from the cfDNA / cfRNA sample comprising the sequencing data set; eliminating reads with sequencing errors using a machine learning model that distinguishes sequencing errors from true variants; and determining early minimal residual disease (MRD) from the cfDNA / cfRNA sample using a truncated mutation spectrum and a machine learning model.

[0032] In some embodiments, for cfDNA / cfRNA samples, estimating the overall tumor cfDNA score includes: combining information from all potential SNV sites in the cfDNA / cfRNA sample to reduce the noise caused by sequencing errors occurring at one or more potential SNV sites. In some embodiments, determining the genotype likelihood includes: calculating the likelihood of observing the reading of the coverage locus given different joint genotypes. In some embodiments, determining the genotype likelihood includes: estimating the genotype at the locus, which maximizes the posterior probability. In some embodiments, estimating the genotype at the locus includes: determining the joint genotype of each locus in the sequence data set. In some embodiments, eliminating reads with sequencing errors includes: training a machine learning model with samples sequenced multiple times; determining bases from true variants or sequencing errors in the sequencing data set. In some embodiments, determining early minimal residual disease (MRD) from cfDNA / cfRNA samples includes: determining truncated mutations from mutation spectra and variant allele frequencies called from cfDNA / cfRNA samples; combining multiple variant loci to enhance tumor signals in ultra-low tumor score samples; and determining MRD status by intra-sample statistical tests.

[0033] According to another aspect, the present disclosure provides a method for detecting somatic single nucleotide variants (SNVs) from a plurality of cell-free nucleic acid (cfNA) molecules of a subject, comprising: (a) retrieving a plurality of sequence reads generated by a sequencer, wherein at least a subset of the plurality of sequence reads comprises sequences from the plurality of cfNA molecules or derivatives thereof; (b) applying a probabilistic model to the plurality of sequence reads at each of a plurality of genetic loci to estimate an overall tumor burden of the plurality of cfNA molecules, wherein the estimated overall tumor burden comprises a quantitative measure of tumor-derived cfNA molecules in the plurality of cfNA molecules, wherein the plurality of genetic loci comprise potential SNVs; V site; (c) for each of the multiple genetic loci, determining the likelihood of one or more genotypes of the subject based at least in part on the overall tumor burden, wherein the one or more genotypes are selected from a normal genotype, a tumor genotype, and a combined normal-tumor genotype; (d) for each of the multiple genetic loci, detecting one or more SNVs based at least in part on the likelihood of the one or more genotypes determined in (c) and the quantitative measurement of tumor-derived cfNA molecules in the multiple cfNA molecules; and (e) filtering out one or more germline polymorphisms from the one or more SNVs detected in (d) to obtain one or more somatic SNVs.

[0034] In some embodiments, the method further comprises (f) filtering out the one or more somatic SNVs obtained in (e) using one or more filters selected from the following, thereby obtaining a filtered set of somatic SNVs: a strand bias filter, a base quality filter, a read partner filter, a sequencing error filter, an insertion or deletion (indel) and a homopolymer-induced error filter, and a public database filter. In some embodiments, the method further comprises (g) applying the probability model to the multiple sequence reads at each of the multiple genetic loci to re-estimate the overall tumor burden, wherein the multiple genetic loci include the filtered set of somatic SNVs obtained in (f). In some embodiments, the multiple cell-free nucleic acid (cfNA) molecules include cell-free deoxyribonucleic acid (cfDNA) molecules. In some embodiments, the multiple cell-free nucleic acid (cfNA) molecules include cell-free ribonucleic acid (cfRNA) molecules. In some embodiments, estimating the overall tumor burden includes combining information from the multiple sequence reads spanning the potential SNV sites to reduce the noise in the multiple sequence reads caused by errors in the sequencing at one or more of the potential SNV sites. In some embodiments, re-estimating the overall tumor burden includes combining information from the multiple sequence reads spanning the filtered set of somatic SNVs to reduce the noise in the multiple sequence reads caused by errors in the sequencing at one or more of the filtered set of somatic SNVs. In some embodiments, combining the information includes calculating the total proportion of sequence reads containing SNVs in multiple predetermined SNV hotspots in the multiple sequence reads. In some embodiments, combining the information includes calculating the value that maximizes the likelihood of observing the multiple sequence reads at each of the multiple predetermined SNV hotspots given the overall tumor burden. In some embodiments, combining the information includes calculating the maximum value of the variant allele frequency value in multiple predetermined SNV hotspots. In some embodiments, the information includes base calls, base quality, mapping quality, or a combination thereof. In some embodiments, determining the likelihood of the genetic locus includes determining the likelihood of observing several sequence reads in the multiple sequence reads covering the genetic locus given one or more genotypes. In some embodiments, determining the likelihood of the genetic locus includes calculating the maximum a posteriori probability estimate. In some embodiments, determining the likelihood of the genetic locus comprises determining a joint normal tumor genotype of the genetic locus.In some embodiments, the filtering is performed using one or more filters selected from the group consisting of a strand bias filter, a base quality filter, a read partner filter, a sequence background filter, and a sequencing error filter.In some embodiments, the probabilistic model comprises a machine learning model configured to classify sequence reads having true variants from sequence reads having sequencing errors.

[0035] In some embodiments, the method further comprises subjecting the multiple cfNA molecules to amplification. In some embodiments, the amplification comprises polymerase chain reaction (PCR). In some embodiments, the method further comprises processing the detected somatic SNV for reference. In some embodiments, the reference comprises a second set of somatic SNVs detected from multiple cfNA molecules of one or more other objects. In some embodiments, the multiple cfNA molecules are obtained from a body sample of the object. In some embodiments, the body sample is selected from plasma, serum, bone marrow, cerebrospinal fluid, pleural fluid, saliva, feces and urine. In some embodiments, the method further comprises processing the detected somatic SNV to generate the likelihood that the object suffers from or is suspected of having a disease or condition. In some embodiments, the disease or condition is selected from the following cancers: pancreatic cancer, liver cancer, lung cancer, colorectal cancer, leukemia, bladder cancer, bone cancer, brain cancer, breast cancer, cervical cancer, endometrial cancer, esophageal cancer, gastric cancer, head and neck cancer, melanoma, ovarian cancer, testicular cancer, kidney cancer, sarcoma, bile duct cancer and prostate cancer. In some embodiments, the method further comprises sequencing the plurality of cfNA molecules or derivatives thereof to generate the plurality of sequence reads. In some embodiments, the somatic SNV comprises a cancer-driven mutation.

[0036] According to another aspect, the present disclosure provides a system for detecting somatic single nucleotide variants (SNVs) from a plurality of cell-free nucleic acid (cfNA) molecules of a subject, comprising: a database storing a plurality of sequence reads, wherein at least a subset of the plurality of sequence reads comprises sequences from the plurality of cfNA molecules or derivatives thereof; and one or more computer processors operatively coupled to the database, wherein the one or more computer processors are individually or collectively programmed to: (1) retrieve the plurality of sequence reads from the database; (2) apply a probabilistic model to the plurality of sequence reads at each of a plurality of genetic loci to estimate an overall tumor burden of the plurality of cfNA molecules, wherein the estimated overall tumor burden comprises a probability distribution of tumor-derived cfNA molecules in the plurality of cfNA molecules; (3) for each of the plurality of genetic loci, determining a likelihood of one or more genotypes of the subject based at least in part on the overall tumor burden, wherein the one or more genotypes are selected from a normal genotype, a tumor genotype, and a combined normal-tumor genotype; (4) for each of the plurality of genetic loci, detecting one or more SNVs based at least in part on the likelihood of the one or more genotypes determined in (3) and the quantitative measurement of tumor-derived cfNA molecules in the plurality of cfNA molecules; and (5) filtering out one or more germline polymorphisms from the one or more SNVs detected in (4) to obtain one or more somatic SNVs.

[0037] In some embodiments, the one or more computer processors are individually or collectively programmed to additionally (6) filter out the one or more somatic SNVs obtained in (5) using one or more filters selected from the following, thereby obtaining a filtered set of somatic SNVs: strand bias filter, base quality filter, read mate filter, sequencing error filter, insertion or deletion (indel) and homopolymer-induced error filter, and public database filter. In some embodiments, the one or more computer processors are individually or collectively programmed to additionally (7) apply the probability model to the multiple sequence reads at each of the multiple genetic loci to re-estimate the overall tumor burden, wherein the multiple genetic loci include the filtered set of somatic SNVs obtained in (6). In some embodiments, the multiple cell-free nucleic acid (cfNA) molecules include cell-free deoxyribonucleic acid (cfDNA) molecules. In some embodiments, the multiple cell-free nucleic acid (cfNA) molecules include cell-free ribonucleic acid (cfRNA) molecules. In some embodiments, estimating the overall tumor burden includes combining information from the multiple sequence reads spanning the potential SNV sites to reduce the noise in the multiple sequence reads caused by errors in the sequencing at one or more of the potential SNV sites. In some embodiments, re-estimating the overall tumor burden includes combining information from the multiple sequence reads spanning the filtered set of somatic SNVs to reduce the noise in the multiple sequence reads caused by errors in the sequencing at one or more of the filtered set of somatic SNVs. In some embodiments, combining the information includes calculating the total proportion of sequence reads containing SNVs in multiple predetermined SNV hotspots in the multiple sequence reads. In some embodiments, combining the information includes calculating the value that maximizes the likelihood of observing the multiple sequence reads at each of the multiple predetermined SNV hotspots given the overall tumor burden. In some embodiments, combining the information includes calculating the maximum value of the variant allele frequency value in multiple predetermined SNV hotspots. In some embodiments, the information includes base calls, base quality, mapping quality, or a combination thereof. In some embodiments, determining the likelihood of the genetic locus comprises determining the likelihood of observing several of the plurality of sequence reads covering the genetic locus given one or more genotypes. In some embodiments, determining the likelihood of the genetic locus comprises calculating a maximum a posteriori probability estimate. In some embodiments, determining the likelihood of the genetic locus comprises determining a joint normal tumor genotype of the genetic locus.In some embodiments, the filtering is performed using one or more filters selected from the group consisting of: a strand bias filter, a base quality filter, a read mate filter, a sequence background filter, and a sequencing error filter. In some embodiments, the probabilistic model comprises a machine learning model configured to classify sequence reads with true variants from sequence reads with sequencing errors.

[0038] In some embodiments, the one or more computer processors are individually or collectively programmed to process the detected somatic SNVs for a reference. In some embodiments, the reference comprises a second set of somatic SNVs detected from a plurality of cfNA molecules of one or more additional subjects. In some embodiments, the plurality of cfNA molecules are obtained from a body sample of the subject. In some embodiments, the body sample is selected from plasma, serum, bone marrow, cerebrospinal fluid, pleural fluid, saliva, feces and urine. In some embodiments, the one or more computer processors are individually or collectively programmed to process the detected somatic SNVs to generate a likelihood that the subject suffers from or is suspected of having a disease or condition. In some embodiments, the disease or condition is selected from the following cancers: pancreatic cancer, liver cancer, lung cancer, colorectal cancer, leukemia, bladder cancer, bone cancer, brain cancer, breast cancer, cervical cancer, endometrial cancer, esophageal cancer, gastric cancer, head and neck cancer, melanoma, ovarian cancer, testicular cancer, kidney cancer, sarcoma, bile duct cancer and prostate cancer. In some embodiments, the somatic SNVs comprise cancer-driving mutations.

[0039] According to another aspect, the present disclosure provides a non-transitory computer-readable medium comprising machine-executable code, which, when executed by one or more computer processors, implements a method for detecting somatic single nucleotide variants (SNVs) from a plurality of cell-free nucleic acid (cfNA) molecules of a subject, the method comprising: (a) retrieving a plurality of sequence reads generated by a sequencer, wherein at least a subset of the plurality of sequence reads comprises sequences from the plurality of cfNA molecules or derivatives thereof; (b) applying a probabilistic model to the plurality of sequence reads at each of a plurality of genetic loci to estimate an overall tumor burden of the plurality of cfNA molecules, wherein the estimated overall tumor burden comprises a quantitative measurement of tumor-derived cfNA molecules in the plurality of cfNA molecules; (c) for each of the plurality of genetic loci, determining a likelihood of one or more genotypes of the subject based at least in part on the overall tumor burden, wherein the one or more genotypes are selected from a normal genotype, a tumor genotype, and a combined normal-tumor genotype; (d) for each of the plurality of genetic loci, detecting one or more SNVs based at least in part on the likelihood of the one or more genotypes determined in (c) and the quantitative measurement of tumor-derived cfNA molecules in the plurality of cfNA molecules; and (e) filtering out one or more germline polymorphisms from the one or more SNVs detected in (d) to obtain one or more somatic SNVs.

[0040] In some embodiments, the method further comprises (f) filtering out the one or more somatic SNVs obtained in (e) using one or more filters selected from the following, thereby obtaining a filtered set of somatic SNVs: a strand bias filter, a base quality filter, a read partner filter, a sequencing error filter, an insertion or deletion (indel) and a homopolymer-induced error filter and a public database filter. In some embodiments, the method further comprises (g) applying the probability model to the multiple sequence reads at each of the multiple genetic loci to re-estimate the overall tumor burden, wherein the multiple genetic loci are included in the filtered set of somatic SNVs obtained in (f). In some embodiments, the multiple cell-free nucleic acid (cfNA) molecules comprise cell-free deoxyribonucleic acid (cfDNA) molecules. In some embodiments, the multiple cell-free nucleic acid (cfNA) molecules comprise cell-free ribonucleic acid (cfRNA) molecules. In some embodiments, estimating the overall tumor burden comprises combining information from the multiple sequence reads spanning the potential SNV sites to reduce the noise in the multiple sequence reads caused by errors in the sequencing at one or more of the potential SNV sites. In some embodiments, re-estimating the overall tumor burden includes combining information from the multiple sequence reads spanning the filtered set of somatic SNVs to reduce the noise in the multiple sequence reads caused by errors in the sequencing at one or more of the filtered set of somatic SNVs. In some embodiments, combining the information includes calculating the total proportion of sequence reads containing SNVs in multiple predetermined SNV hotspots in the multiple sequence reads. In some embodiments, combining the information includes calculating the value that maximizes the likelihood of observing the multiple sequence reads at each of the multiple predetermined SNV hotspots given the overall tumor burden. In some embodiments, combining the information includes calculating the maximum value of the variant allele frequency value in multiple predetermined SNV hotspots. In some embodiments, the information includes base calls, base quality, mapping quality, or a combination thereof. In some embodiments, determining the likelihood of the genetic locus includes determining the likelihood of observing several sequence reads in the multiple sequence reads covering the genetic locus given one or more genotypes. In some embodiments, determining the likelihood of the genetic locus includes calculating the maximum a posteriori probability estimate. In some embodiments, determining the likelihood of the genetic locus comprises determining a joint normal tumor genotype of the genetic locus.In some embodiments, the filtering is performed using one or more filters selected from the group consisting of a strand bias filter, a base quality filter, a read partner filter, a sequence background filter, and a sequencing error filter.In some embodiments, the probabilistic model comprises a machine learning model configured to classify sequence reads having true variants from sequence reads having sequencing errors.

[0041] In some embodiments, the method further comprises processing the detected somatic SNVs for a reference. In some embodiments, the reference comprises a second set of somatic SNVs detected from multiple cfNA molecules of one or more additional subjects. In some embodiments, the multiple cfNA molecules are obtained from a body sample of the subject. In some embodiments, the body sample is selected from plasma, serum, bone marrow, cerebrospinal fluid, pleural fluid, saliva, feces and urine. In some embodiments, the method further comprises processing the detected somatic SNVs to generate a likelihood that the subject suffers from or is suspected of having a disease or condition. In some embodiments, the disease or condition is selected from the following cancers: pancreatic cancer, liver cancer, lung cancer, colorectal cancer, leukemia, bladder cancer, bone cancer, brain cancer, breast cancer, cervical cancer, endometrial cancer, esophageal cancer, gastric cancer, head and neck cancer, melanoma, ovarian cancer, testicular cancer, kidney cancer, sarcoma, bile duct cancer and prostate cancer. In some embodiments, the somatic SNVs comprise cancer-driven mutations.

[0042] The features and technical advantages of the present invention have been outlined quite extensively above so that the following detailed description of the present invention may be better understood. The following will describe additional features and advantages of the present invention that form the subject matter of the claims of the present invention. It will be appreciated by those skilled in the art that the disclosed concepts and specific embodiments may be easily used as a basis for modifying or designing other structures to implement the same purpose of the present invention. It will also be appreciated by those skilled in the art that such equivalent constructions do not depart from the spirit and scope of the present invention as set forth in the appended claims. When considered in conjunction with the accompanying drawings, the novel features that are considered to be features of the present invention, together with additional objects and advantages, will be better understood from the following description, with respect to both the organization and method of operation of the present invention. However, it should be clearly understood that each of the accompanying drawings is provided for the purpose of illustration and description only and is not intended to be a limitation to the limitation of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] For a more complete understanding of the present disclosure, reference is now made to the following description in conjunction with the accompanying drawings, in which:

[0045] Figure 1 A panel showing a set of variant genes and their variant frequencies in different sample types according to one embodiment;

[0046] Figure 2 A flow chart of a method for detecting single nucleotide variants (SNVs) from plasma cell-free deoxyribonucleic acid (cfDNA) according to one embodiment is shown;

[0047] Figure 3 Specific concepts for estimating the overall tumor cfDNA fraction according to one embodiment are shown;

[0048] Figure 4 A specific concept of studying a locus according to one embodiment is shown;

[0049] Figure 5 A method for extracting features for a machine learning model that can distinguish true variants from sequence errors based on individual read information is shown according to one embodiment;

[0050] Figure 6 A method for generating training data for a machine learning model that can distinguish true variants from sequence errors based on individual read information is shown according to one embodiment;

[0051] Figure 7 shows a timeline for performing early detection of minimal residual disease (MRD) according to one embodiment;

[0052] Figure 8 Certain components of a system for performing early detection of minimal residual disease are shown according to one embodiment;

[0053] Fig. 9 Some aspects of using the disclosed method to detect minimal residual disease are shown according to one embodiment; and

[0054] Fig.10 Certain components of a system for performing the disclosed method are shown according to one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0056] The present disclosure provides a probability model for accurate and sensitive somatic single nucleotide variant (SNV) detection in cell-free deoxyribonucleic acid (cfDNA) samples and / or cell-free RNA (cfRNA) samples, the samples comprising sequence data sets. The cfDNA sample can be taken from a plasma sample. The model can receive sequencing data as input from cfDNA and a matched normal sample (e.g., from white blood cells in the blood). In addition, since a joint genotype can be determined for each locus in the sequence data set, germline mutations can be removed. In addition, since the overall tumor cfDNA score can be considered, accurate SNV detection can be performed from a sample with a low tumor score, which may not be available with other methods.

[0057] As described herein, cfDNA from a subject (e.g., a cancer patient) may include a mixture of DNA or RNA from cells (e.g., tumor cells) of a diseased organ and DNA or RNA from normal cells. Therefore, the genotype of cfDNA can be modeled as a normal tumor combined genotype. For a given locus, the prior variant allele frequency (prior VAF) of the tumor genotype represented as θ can be approximated as the fraction of mutation reads. (e.g., tumor cfDNA fraction). The tumor cfDNA fraction in early cancer patients can typically be less than 10%, so in order to accurately estimate somatic SNVs in tumor cells, the tumor cfDNA fraction can be estimated first. It can be assumed that the tumor cfDNA fraction is uniform across the genome in a single sample.

[0058] The methods and systems of the present disclosure can maximize the likelihood P(X|θ) of observing sequencing reads to estimate the tumor cfDNA fraction or use the average of the variant allele frequencies of predefined, frequently observed variants in cancer. Since the tumor cfDNA fraction can be assumed to be uniform, data from hot spots can be used to estimate the tumor cfDNA fraction. According to one embodiment, the following criteria can be used to select hot spots: (1) the major allele can be a reference base; (2) the locus in the matched normal sample can be homozygous and can be the same as the reference; (3) the sequencing error probability of the read supporting the minor allele can be less than the observed allele frequency; (4) at least one read with a variant allele can be observed; and (5) the read covering the locus must pass a strand bias filter. Thus, sequencing noise can be eliminated. The probability of observing a read or base can be calculated using the mapping quality of the read (which is the probability of misalignment) and the base quality of the sequenced base (which is the probability of sequencing error). To determine the genotype of a locus, the likelihood of observing a read covering the locus given different joint genotypes can be calculated.

[0059] According to one embodiment, the posterior probability of the joint genotype is calculated from the sequence data set. The joint genotype with the maximum posterior probability can be regarded as the true genotype at a given locus. Post-filtering can be applied to the detected somatic variants, which can exclude loci with chain bias, low coverage or tandem duplication. The final tumor score can be estimated as the maximum variant allele frequency of the somatic variants through filtering.

[0060] The input data may include raw sequencing reads of cfDNA from a subject (e.g., a patient). Moreover, embodiments may provide a somatic SNV report after analysis. When detecting SNVs from cfDNA, embodiments may accept a panel in which loci with the highest likelihood of containing tumor-derived point mutations are evaluated. For example, Figure 1A panel 100 is shown, which may be obtained by analyzing public cancer databases or collecting known mutations in the open literature.

[0061] The main challenge of detecting SNVs from cfDNA can be the low cfDNA fraction of tumor origin in the bloodstream, which makes it difficult to distinguish mutations from sequencing errors. Without paying special attention to the tumor cfDNA fraction, detecting mutations may be unreliable. The methods and systems of the present disclosure can solve the challenge by incorporating low tumor cfDNA fractions into a probabilistic framework, for example, explicitly considering low tumor-derived cfDNA fractions. To this end, the overall tumor cfDNA fraction θ can be estimated by combining information from all potential SNV sites, and especially sites known to cause mutations. Combining sites in this way can eliminate noise caused by sequencing errors at a single site. The estimated θ can then be used to accurately call SNVs at a single site. In some embodiments, SNVs at a single site can be called with an accuracy of at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or more.

[0062] Figure 2 A general flow chart of a method 200 for detecting SNVs from plasma cfDNA according to the disclosed embodiments is shown. According to the method 200, raw sequencing data 201 may include a plasma cfDNA sample and a matched normal sample (e.g., white blood cells from the same subject) and may be input to box 202. The raw sequencing data 201 may include input files, such as alignment files (.sam files or .bam files) and pileup files (.pileup files). Extract the SNVs along the small groups (e.g., Figure 1 Data in the area of ​​group 100) shown in .

[0063] The raw sequencing data can be obtained by sequencing a plasma cfDNA sample or a derivative thereof (e.g., an amplified product or an enriched product). The sequencing method may include, but is not limited to, high-throughput sequencing, pyrosequencing, sequencing-by-synthesis, single-molecule sequencing, nanopore sequencing, semiconductor-based sequencing, sequencing-by-ligation, sequencing-by-hVbridization, RNA-Seq (e.g., Illumina), digital gene expression (Helicos), next-generation sequencing (e.g., Illumina, Pacific Biosciences of California, Ion Torrent), single-molecule sequencing by synthesis (SMSS) (e.g., Helicos), massively-parallel sequencing, clonal single-molecule arrays (e.g., Solexa), shotgun sequencing, Maxim-Gilbert sequencing, and primer walking.

[0064] The raw sequencing data can be preprocessed to improve the mapping quality. Those reads with low-quality mapping can be recalibrated and locally realigned based on the reference genome. Bases with low base quality can also be eliminated by setting a hard threshold, which can ensure a low sequencing error rate.

[0065] The tumor cfDNA score estimation block 202 can estimate the overall tumor cfDNA score using pre-processed sequencing data, which contains information extracted from SAM or BAM files and / or PILEUP files. That is, at step 202, tumor score estimation is performed by combining the loci with the overall combination of variant alleles observed in the sequencing data. The overall tumor score estimation can effectively suppress sequencing noise and avoid the serious effects of sampling fluctuations and sequencing errors (especially low or medium depth sequencing). This can not only improve accurate somatic SNV detection, but also can be a marker with clinical value in cancer diagnosis and MRD detection.

[0066] Step 203 may include determining SNVs by maximizing the posterior of the combined tumor-normal genotype given the overall tumor fraction. This may be contrasted with other methods that compare reads containing a reference allele and a variant allele.

[0067] Step 204 may include filtering somatic SNV candidates using a set of cfDNA feature criteria in terms of strand bias, variant quality, sequence context, and known SNPs. Such post-call filtering ensures high quality of variant calls in cfDNA.

[0068] Step 205 may include re-estimating somatic SNVs after filtering. As described with respect to the first step, this can be done using the overall tumor score. At box 205, in examining the cfDNA sample and its matched normal sample, the variant allele frequency may be particularly considered. For example, since the true variant allele frequency may be closely related to the amount of tumor-derived cfDNA present in the bloodstream, the variant allele frequency cannot be simply considered to be 0.5 for heterozygous variants and 1.0 for homozygous variants. This method may represent an improvement over other SNV detection methods.

[0069] Traditional SNV methods may assume that all observed alleles are derived from diploid chromosomes, where the first chromosome copy has the genotype g1 and the second chromosome copy has g2. Therefore, the biallelic genotype at a locus can be represented by a couplet G=(g1, g2), where the prior allele frequencies of g1 and g2 are 50% and 50%. This abstraction may assume that the tumor sample being examined is pure or nearly pure (e.g., the tumor genotype dominates the sample). However, as discussed herein, the methods and systems of the present disclosure may account for the fact that cfDNA may actually contain a mixture of tumor-derived DNA and non-tumor-derived DNA. Therefore, such an approach may focus only on loci without germline polymorphisms and model the genotype of cfDNA as G=(g n , g t ), where gn = (gn1, gn2) is the genotype from normal cfDNA, and gt = (gt1, gt2) is the genotype from tumor-derived cfDNA. For a specific locus, the gt prior variant allele frequency (prior VAF) denoted as θ can be approximated as the fraction of mutation reads or tumor cfDNA fraction. Then (1-θ) is the prior frequency of gn. However, given a very low proportion of tumor-derived cfDNA, such an approximation of θ at a single locus may be very noisy. To address such challenges, it can be assumed that the entire sample shares the same prior allele frequency, so the overall tumor cfDNA fraction (e.g., overall θ) is used for operation.

[0070] Figure 2 The method shown in Figure 2It can be advantageous because (1) the overall tumor score obtained in the first step can generate the expected level of tumor reads at individual loci even at low tumor scores, so that somatic SNVs with low mutant frequencies can still be distinguished from random sequencing errors; (2) in the presence of mixed normal and tumor-derived cfDNA, the joint genotype model can better fit the cfDNA sequencing data by fully utilizing both normal and tumor signals by simultaneously genotyping normal and tumor cfDNA; (3) computational sequencing error suppression by machine learning models can effectively distinguish true variants from sequencing errors based on individual read information; and (4) post-filtering of somatic SNV candidates can be specifically designed based on cfDNA characteristics. In view of these advantageous features, the methods and systems of the present disclosure can be applied not only to deeply targeted cfDNA sequencing data, but also to medium-depth cfDNA sequencing data, such as whole exome sequencing data. Therefore, the methods and systems of the present disclosure can provide sensitive mutation calls and mutation spectra of cancer from cfDNA based on input data.

[0071] Reference Figure 3 300. θ can be approximated using the overall fraction of mutation reads of all possible hot spots, which may include loci that are highly likely to harbor tumor variants. That is, Figure 3 A concept is shown where, to estimate the overall tumor cfDNA fraction, data sequencing contains sequencing reads at all hotspots, which are believed to have lower sequencing noise effects, such as sequencing errors.

[0072] Probabilistic methods can be applied while taking into account possible sequencing errors. For example, θ can be estimated as the value that maximizes the likelihood P(X|θ), that is, given the overall tumor cfDNA fraction θ, the sequencing data X=(X|θ) observed at K hotspots (1) , …, X (K) ) probability.

[0073] The likelihood can be calculated by marginalizing the genotype G,

[0074]

[0075] Where RH represents the pool of reads covering at least one hotspot; X rrepresents the information contained in read r; Gh(r) represents the genotype of the hotspot covered by read r; and h(r) is the set of hotspots covered by read r. The probability P(Xr|Gh(r), θ) can then be decomposed by further specifying the starting point of the read (e.g., tumor cell or normal cell). In addition, when the overall tumor cfDNA fraction is θ, a random read from the sequencing data can have a probability of θ of being tumor-derived according to the following formula:

[0076]

[0077] in represents the joint tumor genotype of the hotspot on read r; and represents the joint normal genotype of hotspot h(r) on read r.

[0078] In some cases, different reads in RH may cover different hot spots (mutually exclusive) or have shared hot spots. Considering the maximum likelihood estimation framework, the joint combination of genotypes for all hot spots can be marginalized to eliminate the influence of genotype. Because this step can only include the most suspicious loci, a single read can cover a large number of hot spots, making such an approach reasonable for calculation.

[0079] Single reads covering more than one hotspot can be considered independent loci rather than accounting for linkage disequilibrium which can further complicate the issue. Likewise, the information in the reads can also be extended to base calling, base quality, and mapping quality. This can be expressed by the following equation:

[0080]

[0081] The read r contains K hot spots; X r (k) Indicates the information of the bases covering the hotspot k on read r; B r (k) , Q r (k) and M r denote the base call, base quality, and mapping quality of read r at hotspot k, respectively, and represents the tumor genotype at hotspot k. For normal cases, this is applicable. This can be expressed by the following equation:

[0082]

[0083] in represents the normal genotype at hotspot k.

[0084] Based on the foregoing, the probability of observing a base in the sequencing data given a genotype can be calculated. It can be assumed that for a given hotspot, a read with base A is observed, and the given genotype is AB. The base can then be generated by allele A with probability to determine, and with probability From allele B, a value of 1 / 3 represents all three (3) possible incorrect base calls. However, instead of simply using 1 / 3, an uneven probability can be used to account for different trends in sequencing errors. This can be represented by the following equation:

[0085]

[0086] where ∈ represents the probability that a base comes from a sequencing error. Similarly, P(A, Q (k) , M r |A) means if the base comes from sequencing allele A, a base with quality Q r (k) , M r The probability of base A, and P(A, Q (k) , M r |B) means if the base comes from sequencing allele B, a base with quality Q r (k) , M r The probability of an A base in the sequence is . Therefore, the above equation can account for two possible sources of base A. Base quality and mapping quality can indicate the probability of an incorrectly sequenced base or an incorrectly placed read in the Phred scale, so it can be converted back to probability. This can be represented by the following equation:

[0087]

[0088] By solving the above equation, an estimate of the overall tumor cfDNA fraction can be made with high accuracy (e.g., at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or more accuracy).

[0089] Assuming that reads and hot spots are independent of each other, the probability of observing such a sequencing data set can be calculated by quantifying the probability of sequencing errors and mapping errors. Alternatively, a simplified method can be used in which the overall tumor cfDNA fraction is estimated as the maximum value of the variant allele frequency in the locus. Even if the accuracy of the tumor cfDNA fraction estimate is not optimized, the accuracy of SNV calls may not be significantly reduced as long as the estimated tumor cfDNA fraction is scaled correctly.

[0090] The genotype calculation block 203 may use the information received from the tumor cfDNA fraction estimation block 202 to calculate the probabilities of the normal genotype, the tumor genotype, and the combined genotype. The framework for calculating genotypes (or calling SNVs) may include converting the genotype likelihood P(X) at locus k to (k) |G (k) ) is calculated as Where X (k) represents the observed number of sequencing reads covering a specific site on the genome of collection.

[0091] This concept is partly based on Figure 4 Refer to Figure 4 , when identifying somatic SNVs, only bases covering the locus of interest on the reads can be considered. As shown, for each sequencing read x (k) , which can be used to study tumor genotypes Wherein the sequenced allele A is shown by a solid line, and the sequenced allele B is shown by a dashed line.

[0092] According to one embodiment, maximizing the posterior is performed to predict the joint normal tumor genotype. Given the observed sequencing data, the joint normal tumor genotype with the highest posterior probability can be considered as the true genotype of the sample.

[0093] When analyzing cfDNA, the genotype likelihood may need to incorporate low tumor DNA frequencies by adding a new parameter θ. This can be expressed by the following equation:

[0094]

[0095] In general, the likelihood of each read r covering locus k can be decomposed into its own likelihood: g N (k) and g T (k) . This can be expressed by the following equation:

[0096]

[0097] Note that the prior probability P (that read r is from tumor-derived cfDNA) can actually be g at the current site k. T (k) The prior allele frequency θ is θ. This approach yields:

[0098]

[0099] In the above equation, or It can be calculated the same as in the previous discussion.

[0100] However, in contrast to estimating the overall tumor cfDNA fraction, when genotyping (e.g., calling SNVs), only the bases at the genotyped loci can be correlated on a single read. Then, using Bayes' theorem, the posterior genotype G at locus k that maximizes the posterior probability can be estimated. (k) . This can be expressed by the following equation:

[0101]

[0102]

[0103] The prior distribution of the joint normal tumor genotype G can be determined before calculation. For example, if the locus of interest is annotated in a public cancer mutation database such as a single nucleotide polymorphism database (dbSNP, https: / / www.ncbi.nlm.nih.gov / projects / SNP / ) or a catalog of somatic mutations in cancer (COSMIC, https: / / cancer.sanger.ac.uk / cosmic), the variant allele frequency in the general population can be directly obtained. Alternatively, sequencing data in public data sources can be explored, such as the Cancer Genome Atlas Database (TCGA, https: / / cancergenome.nih.gov / ), 1000 Genome Database (http: / / www.internationalgenome.org / ) and the International Cancer Genome Joint Database (ICGC, https: / / icgc.org / ). By considering a specific locus relative to a set of samples, the allele frequency of a group of a population portion or object (e.g., patient) can be determined.

[0104] In order to detect tumor-derived SNVs from cfDNA, germline polymorphisms can be eliminated. This can be done by considering both normal and tumor genotypes to account for germline polymorphisms.

[0105] Refer again Figure 2 At block 201, information from normal blood cells (e.g., leukocytes) from the same subject can be used to incorporate normal reads into the analysis framework. That is, this information can be used to further calculate the likelihood of normal genotypes, tumor genotypes, or combined genotypes. This can be a similar situation to the removal of germline SNVs in matched tumor normal tissue samples. The posterior probability can be changed to:

[0106]

[0107] Where X w (k) is the data from a matched normal sample (e.g., leukocytes), and X p(k) Represents cfDNA reads sequenced from plasma.

[0108] Filter block 204 may perform post-call filtering. A set of filters may be developed to remove unreliable mutation candidates for cfDNA data. Base quality and mapping quality may be important for removing low-quality reads containing sequencing errors. Probabilistic models may incorporate such quality scores. In addition to the types of errors that can be distinguished by quality scores, there may also be systematic errors that cannot be eliminated by using only quality scores. Embodiments may combine read information near mutation candidates, and low-quality mutation candidates may be filtered by all reads observed at the locus.

[0109] The strand bias filter can be applied by filtering mutation candidates if the mutation candidate meets one of the following criteria: (1) the percentage of reads from a single strand is greater than a threshold; (2) reads containing the variant allele are observed from a single strand; (3) among all non-reference reads, the proportion of reads containing the variant allele is above a given threshold; and (4) the proportion of reads containing the non-reference allele on a strand is above a given threshold.

[0110] Base quality filters can be applied by comparing base quality between a candidate and its neighboring loci. A mutation candidate can be filtered if it satisfies one of the following conditions: (1) the T statistic of base quality at the locus and its neighboring loci is above a given threshold; and (2) the T statistic of non-reference base quality at the locus and its neighboring loci is above a given threshold.

[0111] If read mates do not agree at their overlapping sequences, a read mate filter can be applied by excluding the read mate in subsequent filtering.

[0112] Sequence context can also affect the sequencing error rate. Therefore, a sequence context filter can be applied by identifying mutation candidates in close regions or within homopolymers.

[0113] Sequencing error filters can be applied based on machine learning models. Figure 2, at box 206, a machine learning model is trained to classify reads with true variants and reads with sequencing errors. Reads classified as containing sequencing errors can be eliminated in the following filtering. A classifier can be trained to classify the reads based on intrinsic differences in multiple aspects (e.g., base quality, sequencing background, PHRED score, etc.) of reads containing tumor-derived mutations and sequencing errors. In particular, different features of alleles in the reads can be extracted, such as: read alignment-related features (e.g., alignment quality and alignment information of each surrounding base encoded in a CIGAR string in a SAM file), genomic background around the allele (e.g., read sequence and insertion / deletion), sequencing quality of each base in the read, and insertion size of double-end sequencing data.

[0114] Figure 5 A method for extracting features of a machine learning model is shown. Figure 5 , the feature spectrum is input to the classifier to distinguish between signals and errors. Since cfDNA can be naturally fragmented, with a length distribution concentrated around about 166 base pairs (bp), the double-end read partner can have overlapping bases after mapping to the reference. Overlapping bases can be sequenced from both the forward and reverse strands of the same cfDNA fragment, so such bases can provide additional information or confirm that they are located at the variant locus. Therefore, the features extracted from the read partner can be combined together. In order to evaluate the effectiveness of this feature spectrum, conventional classification methods (such as random forests) can be used to evaluate the distinguishability of the extracted features. The results show that such a feature spectrum achieves good performance.

[0115] In addition to traditional classifiers, different classifiers, such more complex classifiers (such as deep learning classifiers) can be used to further improve the discrimination performance of true mutations and sequencing errors. The deep learning model may include any descending structure of a deep artificial neural network, such as a classifier based on a convolutional neural network (CNN). A CNN-based classifier can be used, which includes (1) a convolutional layer, (2) a subsampling layer that can extract informative and abstract features at different scales, and (3) an output prediction layer that can use the extracted features for the final prediction of true mutations and sequencing errors. These layers can be stacked to form a deep neural network, and their parameters can be learned by an optimization algorithm based on stochastic gradients. For example, such a CNN-based structure can be applied to imaging classification applications.

[0116] In order to train the machine learning model, high-quality benchmark real training data may be needed, which comprises the readings that have been verified to have sequence errors or true variants. Such benchmark real data can be obtained by using different sequencing data from the same sample with sufficient coverage, for example, whole exome sequencing data and high coverage PCR-free whole genome sequencing data are obtained from NA12878 and metastatic breast cancer patients (MBC315) in the 1000 genome project. The variant in the target area of ​​whole exome sequencing can be called from two samples respectively. Those reads included in the variants identified with high confidence in two samples can be regarded as the training data of the readings with true variants, and those reads included in the variants identified with low confidence in a sample can be regarded as the training data of the readings with sequencing errors.

[0117] Figure 6 Some aspects of collecting training data using NA12878 are shown. That is, Figure 6 A method for training data generation for a machine learning model is shown.

[0118] Known SNP filters can be applied by filtering mutation candidates observed in public databases such as dbSNP, Exome Aggregation Consortium (ExAC, http: / / exac.broadinstitute.org / ) and Genome Aggregation Database (gnomAD, http: / / gnomad.broadinstitute.org / ). Candidates observed in COSMIC can be retained.

[0119] After calling somatic SNVs between solid tumor tissue (resected during surgery or biopsy) and normal blood samples, the estimate of the overall tumor fraction can be further refined using the somatic SNVs identified from the above steps: blocks 202, 203, and 204 ( Figure 2 In particular, truncated mutations can be selected from the called somatic mutations for monitoring, wherein the variant allele frequencies of the called somatic mutations are calculated. Solid tumor tissue may not be required in this step. Many truncated mutations can be combined to ensure sensitive detection of small fractions of tumor-derived cfDNA.

[0120] Application of MRD detection: cfDNA-based monitoring methods

[0121] As shown at box 207, the method can have a wide range of applications, such as cancer detection, monitoring, prognosis, and more specifically blood tumor mutation burden (bTMB) calculation, MRD detection and resistance monitoring. For example, MRD detection is an application that can demonstrate the clinical utility of the methods and systems of the present disclosure.

[0122] Figure 7 A timeline for early detection of minimal residual disease (MRD) is shown. Figure 7 As shown in, the timeline for MRD detection using the methods and systems of the present disclosure may include two steps. At box 701, truncated mutations and mutation spectra from a preoperative plasma cfDNA sample ("baseline" sample) may be identified. Since plasma cfDNA contains DNA fragments originating from multiple tumor sites, the comprehensive tumor mutation spectrum (SNV and its variant allele frequency) calculated in the subject's preoperative plasma cfDNA sample can be used to identify "truncated mutations", which generally refer to those variants with high variant allele frequencies among all detected variants.

[0123] In particular, the cfDNA SNV detection method can be applied to a preoperative plasma cfDNA sample and its matched normal sample (e.g., leukocytes from the same subject) to identify all somatic SNVs and their variant frequencies. Truncating mutations can then be selected based on the mutation call results from cfDNA and matched normal samples (e.g., from leukocytes), and their variant allele frequencies can be calculated.

[0124] Alternatively, somatic SNVs called from solid tumor tissue (resected during surgery or biopsy) and its matched normal samples can be used to determine a set of truncating mutations. Many different mutation calling tools (e.g., VarScan2, MuTect, SomaticSniper, etc.) can be used to call somatic SNVs. However, truncating mutations cannot always be determined from solid tumor tissue taken from a single tumor site. Therefore, solid tumor tissue may not be required to make the methods and systems of the present disclosure effective. In addition, truncating mutations can be determined based only on mutation calling results from blood samples (e.g., plasma cfDNA and leukocyte DNA).

[0125] At box 702, follow-up plasma cfDNA samples are used to detect minimal residual disease (MRD). That is, after surgery, follow-up plasma samples can be used to monitor and detect MRD. Because the tumor has been treated or removed, the tumor fraction in the follow-up plasma can be lower than the tumor fraction in the baseline plasma. The scale of the variant allele frequency can be the same as the sequencing error rate, or even lower. Therefore, MRD detection may require sensitive and reliable detection of reads containing true mutations. For example, a machine learning method is used to distinguish true mutations from sequencing errors. (As shown by Figure 2In particular, this machine learning method can be applied to extract those reads that cover the positions of identified truncated mutations and are classified as carrying true variants. Then, only those reads classified as containing true variants can be used to calculate the MRD prediction score (or MRD indicator score). The score can be defined as the proportion of reads containing truncated mutations among the reads covering truncated mutations.

[0126] Figure 8 The foregoing is described in more detail. That is, Figure 8 Certain components of a system for performing early detection of minimal residual disease (MRD) are shown.

[0127] In order to evaluate the statistical significance of MRD prediction score in determining the presence of tumor, the background distribution generation method in the sample can be used to model the normal background without introducing additional experimental bias and without the need for additional sample groups. It can be assumed that k truncated mutations have been identified, and k sites can be randomly sampled in the genome, and the sites do not include the identified mutations but match the features (e.g., number of mutations, nucleotide distribution, and depth of coverage) of those k truncated mutations. The readings of the k sites covering the random sampling can be extracted, and then input into the sequencing error suppression workflow to filter out the sequencing errors. MRD prediction scores can be generated for the sites of k samples. Samples can be repeated N times (e.g., wherein N is at least about 5, about 10, about 25, about 50, about 100, or greater than about 100), and N MRD scores are generated. These N MRD prediction scores can include background distribution to evaluate the statistical significance of the MRD prediction scores of k truncated mutations. This step can be repeated over time to detect minimal residual disease (MRD).

[0128] Nucleic acid sequencing

[0129] A variety of nucleic acid sequencing methods can be used to sequence the samples of the present disclosure. Such samples can be processed before sequencing, for example, by purification, separation, enrichment, nucleic acid amplification (e.g., polymerase chain reaction (PCR)). Sequencing can be performed using, for example, the following: Sanger sequencing, high-throughput sequencing, pyrophosphate sequencing, synthesis sequencing, single-molecule sequencing, nanopore sequencing, semiconductor sequencing, connection sequencing, hybridization sequencing, RNA-Seq (Illumina), digital gene expression (Helicos), next-generation sequencing (e.g., Illumina, Pacific Biosciences of California, Ion Torrent), single-molecule synthesis sequencing (SMSS) (Helicos), large-scale parallel sequencing, cloned single-molecule arrays (Solexa), shotgun sequencing, Maxim-Gilbert sequencing, primer walking, sequencing performed using PacBio, SOLiD, Ion Torrent or Nanopore platforms, and any other sequencing methods known in the art. Multiple sequencing can be used to perform simultaneous sequencing reactions.

[0130] Sequencing can produce sequencing reads ("reads") that can be processed by a computer. In some embodiments, the reads can be processed against one or more references to identify single nucleotide variants (SNVs).

[0131] In some embodiments, sequencing can be performed on a cell-free polynucleotide that can contain a variety of different types of nucleic acids. The nucleic acids can be polynucleotides or oligonucleotides. Nucleic acids include, but are not limited to, deoxyribonucleic acid (DNA) or ribonucleic acid (RNA), single-stranded or double-stranded DNA, complementary DNA (cDNA), or RNA / cDNA pairs.

[0132] result

[0133] The performance of the methods and systems of the present disclosure in detecting somatic SNVs was compared with the results of four selected methods (SiNVICT, VarScan 2, MuTect, and MuTect 2) under the same testing conditions.

[0134] Simulation results of data generated by Wgsim

[0135] The sequencing reads in the first simulated data set were generated by wgsim. Reads containing the reference sequence and reads containing variant alleles were mixed in predefined proportions (0.1%, 0.5%, 1%, and 5%). Low variant allele frequencies can simulate the conditions encountered in cfDNA sequencing data.

[0136] In the first simulation, the results obtained using the methods and systems of the present disclosure were compared for VarScan2 and SiNVICT on simulated data obtained by applying wgsim to the human reference genome (hg19). In particular, a region (chr19: 10602252-10602938) with known pathogenic mutations and simulated ultra-deep sequencing of non-small cell lung cancer (NSCLC) was extracted. Thirteen somatic variants were added to this region.

[0137] The average read depth obtained in four different tumor-derived cfDNA fractions (5%, 1%, 0.5%, 0.1%) was about 1448. In order to better approximate the true read, sequencing errors and insertions were introduced into the read. As shown in Table 1, for samples with low tumor cfDNA fractions, the methods and systems of the present disclosure performed better than the other two methods. That is, the simulation results show that, in contrast to the other methods, the methods and systems of the present disclosure are able to meaningfully detect SNVs from plasma DNA with very low ctDNA fractions.

[0138] Table 1

[0139] Tumor cfDNA fraction 0.10％ 0.50％ 1％ 5％ Described embodiments 0％ 46％ 62％ 100％ VarScan2 (p<0.05) 0％ 15％ 62％ 85％ SiNVICT 0％ 0％ 46％ 85％

[0140] Table 1 illustrates the precision of the three methods on simulated sequencing samples with different tumor cfDNA fractions. Note that recall is not shown in Table 1 because all methods demonstrated perfect recall for simulated reads.

[0141] Simulation results for data generated from patients with metastatic cancer

[0142] The second simulated data set contained a spike-in data set of whole exome sequencing data from buffy coat samples and metastatic tumor samples from the same subject. The sample mixing ratio was set to ensure that the variant allele frequencies of somatic SNVs in the subject's genome were approximately 20%, 10%, 5%, and 1%.

[0143] In the second simulation, the methods and systems of the present disclosure were tested on real whole exome sequencing data from metastatic cancer patients. Real samples may contain higher complexity than simulated reads. Exon regions in 7 genes reported to have somatic mutations in cfDNA were extracted, 6 of which were identified in liver metastasis samples, and the remaining 1 was only detected in primary breast cancer.

[0144] In order to obtain samples with different tumor-derived cfDNA fractions, the buffy coat samples were mixed with liver metastasis samples with different spike ratios. Assuming that all somatic variants are heterozygous, the first spike sample is one of six different tumor-derived cfDNA fractions (about 20%, 10%, 5%, and 1%). The read depth at the mutation locus ranged from 46 to 195. Therefore, somatic variants that only appear in primary breast cancer samples are considered controls that should not be identified in spike samples.

[0145] As shown in Table 2, in a sample with a 10% tumor cfDNA fraction, the methods and systems of the present disclosure did not detect somatic variants with a variant allele frequency of 1%. In real sequencing samples, although the total tumor cfDNA fraction can be as high as 10%, the variant allele frequency may be lower for a given locus. When the tumor cfDNA fraction is low, the spiked samples do not contain any reads supporting the variant allele at most somatic variant loci.

[0146] SiNVICT performed moderately on simulated sequencing data but poorly on real sequencing data. It identified all somatic variants as germline variants.

[0147] Table 2

[0148]

[0149] Table 2 shows the precision and recall of the three methods on real sequencing data with different tumor cfDNA fractions. The read depth ranged from 46 to 195.

[0150] The computational method SiNVICT, designed to detect somatic SNVs from cfDNA, failed in most early cancer detections (e.g., when the tumor cfDNA fraction was relatively low). Another method, CAPP, which combines both experimental design and computational methods, failed in at least half of early cancer detections even with 10k coverage at the genotyped loci.

[0151] In contrast, the methods and systems of the present disclosure achieve comparable somatic SNV detection with 7 to 70 times less coverage. In addition, no special sequencing design or target panel may be required for analysis. Therefore, the methods and systems of the present disclosure can be applied to general cancer detection or screening, rather than specific detection of a single cancer type. This is because whole exome sequencing data can be accepted as input. If a specific cancer type is designated for examination using the methods and systems of the present disclosure, a genotyping panel can be designed for somatic SNV detection.

[0152] The sample was obtained from a 41-year-old woman diagnosed with ER+HER2+ breast cancer that had spread to the lymph nodes. Blood was collected about 30 minutes before a liver biopsy to produce a metastasis sample. According to the sequencing results, the average percentage of variant alleles obtained in cfDNA was 14%, indicating that about 28% of the cfDNA was tumor-derived. The blood sample was obtained at an advanced stage of the cancer, so the percentage of tumor DNA was relatively high.

[0153] Sequencing statistics for the analyzed samples are shown in Tables 3 and 4.

[0154] Table 3

[0155]

[0156] Table 4

[0157]

[0158] As described in the present disclosure, cfDNA provides unique opportunities for cancer diagnosis, prognosis and precision oncology. However, the computational tools currently used to analyze mutations (especially ctDNA in plasma) may not be optimally accurate. Given a sequencing data set, the methods and systems of the present disclosure can use a Bayesian-based probability framework to estimate the likelihood of a genotype. Maximizing the posterior estimate can be used to determine the genotype of an object (e.g., a patient) and call somatic SNVs. Both sequencing errors and low tumor variant allele frequencies in plasma samples can be considered, thereby facilitating the detection of somatic SNVs and ensuring good performance.

[0159] As can be seen from the simulation results (e.g., using spiked real sequencing data or simulated sequencing data), the methods and systems of the present disclosure can outperform other methods because they can account for the special biological properties of cfDNA. Since the methods and systems of the present disclosure can detect somatic SNVs from low tumor cfDNA fraction plasma samples, they can be applied to early cancer diagnosis as well as liquid biopsy and post-cancer treatment monitoring. In the case of certain driver mutations detected, clinicians can determine the cancer type and provide personalized treatment for patients.

[0160] Simulation results generated on the data by mixing two germline samples

[0161] To simulate the different dilutions of ctDNA in plasma, two high-quality germline samples obtained from public databases (e.g., normal (NA24385) and tumor (NA12878)) were mixed. Somatic mutations can include 679 high-quality mutations that were identified by VarScan in NA12878 (as tumor) but not in NA24385 (as normal).

[0162] According to the simulation, 20 samples can be generated for dilution (e.g., 0.001%, 0.002%, 0.005%, 0.008%, 0.01%, 0.02%, 0.05%, 0.08%, 0.1%, 0.2%, 0.5%, 0.8% and 1%). For example, if the p-value of the test is not greater than about 0.05, the sample can be identified as indicating resistance.

[0163] Fig. 9 The performance of minimal residual disease (MRD) detection using the methods and systems of the present disclosure is shown 900. Fig. 9 As shown, MRD can be perfectly detected in samples with tumor fraction (TF) as low as about 0.008%. Therefore, the methods and systems of the present disclosure can be applied to the early detection of MRD.

[0164] Fig.10 An exemplary system in which the method according to the present disclosure is applicable to detecting somatic single nucleotide variants (SNVs) from cell-free nucleic acids such as deoxyribonucleic acid (cfDNA) and ribonucleic acid (cfRNA) is shown. The electronic device 1010 may include a variety of configurations of the device. For example, the electronic device 1010 may include a computer, a laptop computer, a tablet computer, a server, a dedicated space processing component or device, a smart phone, a personal digital assistant (PDA), an Internet of Things (IoT or IOTA) device, a network device (e.g., a router, an access point, a femtocell, a picocell, etc.), etc.

[0165] The electronic device 1010 may include any number of components operable to facilitate the functions of the electronic device 1010 according to the method of the present disclosure, such as the processor 1011, system bus 1012, memory 1013, input interface 1014, output interface 1015, and encoder 1016 of the illustrated embodiment. The processor 1011 may include one or more processing units, such as a central processing unit (CPU) (e.g., a processor from the Intel CORE multi-processor unit family), a field programmable gate array (FPGA), and / or an application specific integrated circuit (ASIC), which is operable under the control of one or more instruction sets defining logic modules, and the logic modules are configured to provide the operations described herein. The system bus 1012 couples various system components such as memory 1013, input interface 1014, output interface 1015, and / or encoder 1016 to the processor 1011. Thus, the system bus 1012 of some embodiments may be any of a variety of types of bus structures, such as a memory bus or memory controller, a peripheral bus, and / or a local bus using any of a variety of bus structures. In addition or in lieu thereof, other interfaces and bus structures may be utilized, such as a parallel port, a game port, or a universal serial bus (USB). The memory 1013 may include various configurations of volatile and / or non-volatile computer-readable storage media, such as RAM, ROM, EPSOM, flash memory or other storage technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, cassettes, tapes, magnetic disk storage or other magnetic storage devices, or other tangible and / or non-temporary media that may be used to store the desired information. The input interface 1014 facilitates coupling one or more input components or devices to the processor 1011.

[0166] For example, a user may input commands and information into the electronic device 1010 through one or more input devices (e.g., a keypad, a microphone, a digital pointing device, a touch screen, etc.) coupled to the input interface 1014. Image capture devices such as cameras, scanners, 3-D imaging devices, etc. may be coupled to the input interface 1014 of the embodiment, for example to provide source video herein. The output interface 1015 facilitates coupling one or more output components or devices to the processor 1011. For example, the output of data, images, videos, sounds, etc. from the electronic device 1010 may be provided to the user through one or more output devices (e.g., a display monitor, a touch screen, a printer, a speaker, etc.) coupled to the output interface 1015. The output interface 1015 of the embodiment may provide an interface to other electronic components, devices, and / or systems (e.g., memory, video decoder, radio transmitter, network interface card, device such as a computer, laptop, tablet device, server, dedicated space processing component or device, smart phone, PDA, IOTA device, network equipment, set-top box, cable head-end system, smart TV, etc.).

[0167] Although the present invention and its advantages have been described in detail, it should be understood that, without departing from the spirit and scope of the present invention defined by the appended claims, various changes, substitutions and variations can be made herein. In addition, the scope of the present application is not intended to be limited to some specific embodiments of the process, machine, manufacture, material composition, means, method and step described in the specification. As those of ordinary skill in the art will easily understand from the disclosure of the present invention, according to the present invention, there can be utilized currently existing or to be developed to perform the function substantially the same as the corresponding embodiment described herein or to realize the process, machine, manufacture, material composition, means, method or step of the result substantially the same as it. Therefore, the appended claims are intended to include such process, machine, manufacture, material composition, means, method or step within its scope.

[0168] Moreover, the scope of the present application is not intended to be limited to the particular embodiments of the process, machine, manufacture, composition of matter, means, methods and steps described in the specification.

Claims

1. A computer-implemented method for detecting somatic single nucleotide variants (SNVs) from cell-free deoxyribonucleic acid (cfDNA), the computer-implemented method comprising: (a) Estimation of the overall tumor cfDNA fraction based on sequence datasets of cfDNA samples; (b) determining a genotype likelihood for at least one potential SNV site in the sequence data set based at least in part on the estimated total tumor cfDNA fraction; (c) based at least in part on the genotype likelihood, eliminating germline polymorphisms from a sequence dataset of the cfDNA sample to detect a plurality of somatic SNV candidates; (d) filtering the plurality of somatic SNV candidates from the sequence dataset through a set of filters to generate a plurality of filtered somatic SNV candidates, wherein the filtering is based on a machine learning model that classifies true variants from sequencing errors; as well as (e) analyzing the sequence dataset using a probability model that incorporates the estimated overall tumor cfDNA fraction and the plurality of filtered somatic SNV candidates to determine a re-estimated overall tumor cfDNA fraction in the cfDNA sample.

2. The computer-implemented method of claim 1, wherein estimating the overall tumor cfDNA fraction based on the sequence dataset of the cfDNA sample comprises: Information from all potential SNV sites in the sequence data set from the cfDNA sample is combined to reduce noise caused by sequencing errors occurring at one or more of at least one potential SNV site.

3. The computer-implemented method of claim 1 , wherein determining the genotype likelihood comprises: The likelihood that the observed sequencing reads cover the potential SNV site is calculated given the different combined genotypes.

4. The computer-implemented method of claim 1 , wherein determining the genotype likelihood comprises: Estimate the genotype at the potential SNV site that maximizes the posterior probability.

5. The computer-implemented method of claim 4, wherein estimating the genotype at the potential SNV site comprises: The joint genotype of each potential SNV site in the sequence dataset is determined.

6. The computer-implemented method of claim 1, wherein filtering the plurality of somatic SNV candidates from the sequence dataset further comprises at least one of: filtering the plurality of somatic SNV candidates from the sequence dataset based on a strand bias filter; filtering the plurality of somatic SNV candidates from the sequence dataset based on a base quality filter; filtering a plurality of somatic sequencing reads from the sequence dataset based on a read mate filter; filtering the plurality of somatic SNV candidates from the sequence dataset based on a read mate filter; filtering the plurality of somatic SNV candidates from the sequence dataset based on a sequence background filter; as well as The plurality of somatic SNV candidates are filtered from the sequence dataset based on a public database.

7. The computer-implemented method of claim 1, wherein filtering the plurality of somatic SNV candidates from the sequence dataset based on the machine learning model that classifies true variants versus sequencing errors comprises: Establishing baseline real training data of cfDNA sequencing reads containing real variants or sequencing errors; Creating a profile for each cfDNA sequencing read having information including sequencing quality of each base in the cfDNA sequencing read, read alignment information, sequence background, and insert size of paired-end sequencing data; training a classifier based on the benchmark true training data to classify cfDNA sequencing reads with true variants and cfDNA sequencing reads with sequencing errors by using a signature profile of each cfDNA sequencing read; as well as Each cfDNA sequencing read was classified as having a true variant or a sequencing error using a trained classifier.

8. The computer-implemented method of claim 7, wherein the cfDNA sequencing read is a single read.

9. The computer-implemented method of claim 1, wherein estimating the overall tumor cfDNA fraction in (a) comprises maximizing the likelihood of observing the sequence dataset at multiple potential SNV sites in the sequence dataset.

10. A system for detecting somatic single nucleotide variants (SNVs) from cell-free deoxyribonucleic acid (cfDNA), the system comprising: Computer memory; One or more computer processors communicatively coupled to the computer memory, the one or more computer processors configured to implement a method comprising: (a) Estimation of the overall tumor cfDNA fraction based on sequence datasets of cfDNA samples; (b) determining a genotype likelihood for at least one potential SNV site in the sequence data set based at least in part on the estimated total tumor cfDNA fraction; (c) based at least in part on the genotype likelihood, eliminating germline polymorphisms from a sequence dataset of the cfDNA sample to detect a plurality of somatic SNV candidates; (d) filtering the plurality of somatic SNV candidates from the sequence dataset through a set of filters to generate a plurality of filtered somatic SNV candidates, wherein the filtering is based on a machine learning model that classifies true variants from sequencing errors; as well as (e) analyzing the sequence dataset using a probability model that incorporates the estimated overall tumor cfDNA fraction and the plurality of filtered somatic SNV candidates to determine a re-estimated overall tumor cfDNA fraction in the cfDNA sample.

11. The system of claim 10, wherein estimating the overall tumor cfDNA fraction based on the sequence dataset of the cfDNA sample comprises: Information from all potential SNV sites in the sequence data set from the cfDNA sample is combined to reduce noise caused by sequencing errors occurring at one or more of at least one potential SNV site.

12. The system of claim 10, wherein determining the genotype likelihood comprises: The likelihood that the observed sequencing reads cover the potential SNV site is calculated given the different combined genotypes.

13. The system of claim 10, wherein determining the genotype likelihood comprises: Estimate the genotype at the potential SNV site that maximizes the posterior probability.

14. The system of claim 13, wherein estimating the genotype at the potential SNV site comprises: The joint genotype of each potential SNV site in the sequence dataset is determined.

15. The system of claim 10, wherein filtering the plurality of somatic SNV candidates from the sequence dataset further comprises at least one of: filtering the plurality of somatic SNV candidates from the sequence dataset based on a strand bias filter; filtering the plurality of somatic SNV candidates from the sequence dataset based on a base quality filter; filtering a plurality of somatic sequencing reads from the sequence dataset based on a read mate filter; filtering the plurality of somatic SNV candidates from the sequence dataset based on a read mate filter; filtering the plurality of somatic SNV candidates from the sequence dataset based on a sequence background filter; as well as The plurality of somatic SNV candidates are filtered from the sequence dataset based on a public database.

16. The system of claim 10, wherein filtering the plurality of somatic SNV candidates from the sequence dataset based on a machine learning model that classifies true variants versus sequencing errors comprises: Establishing baseline real training data of cfDNA sequencing reads containing real variants or sequencing errors; Creating a profile for each cfDNA sequencing read having information including sequencing quality of each base in the cfDNA sequencing read, read alignment information, sequence background, and insert size of paired-end sequencing data; training a classifier based on the benchmark true training data to classify cfDNA sequencing reads with true variants and cfDNA sequencing reads with sequencing errors by using a signature profile of each cfDNA sequencing read; as well as Each cfDNA sequencing read was classified as having a true variant or a sequencing error using a trained classifier.

17. The system of claim 16, wherein the cfDNA sequencing read is a single read.

18. The system of claim 10, wherein estimating the overall tumor cfDNA fraction in (a) comprises maximizing the likelihood of observing the sequence dataset at multiple potential SNV sites in the sequence dataset.

19. A system for detecting minimal residual disease (MRD) from plasma samples, leukocytes, and resected tumor samples collected before and after surgery, the system comprising: Computer memory; One or more computer processors communicatively coupled to the computer memory, the one or more computer processors configured to implement a method comprising: identifying one or more truncating mutations and a mutational profile of the one or more truncating mutations from at least one of a pre-operative blood sample and a resected tumor sample; and Detecting MRD using a follow-up plasma cfDNA sample after the surgery by performing the method of claim 1 to determine a re-estimated total tumor cfDNA fraction in the plasma cfDNA sample.

20. The system of claim 19, wherein detecting MRD using a follow-up plasma cfDNA sample after the surgery comprises: extracting sequencing reads covering the genomic locations of the one or more truncating mutations; classifying the sequencing reads as having a true variant or having a sequencing error; as well as Sequencing reads classified as having true variants were used to calculate the MRD prediction score.

21. The system of claim 20, further comprising evaluating a statistical significance of the MRD prediction score compared to a background MRD prediction score, wherein evaluating the statistical significance comprises: Sampling k sites in the genome that do not contain true variants but match the characteristics of the one or more truncating mutations; determining a second set of sequencing reads covering the genomic positions of the sampled k sites; classifying a second set of sequencing reads as having true variants or having sequencing errors; filtering out sequencing reads identified as having sequencing errors from the second sequencing read set; The second set of sequencing reads classified as having true variants was used to calculate the background MRD prediction score for the k sites sampled.

22. A non-transitory storage medium storing a set of instructions that, when executed, cause one or more computer processors to detect somatic single nucleotide variants (SNVs) from cell-free deoxyribonucleic acid (cfDNA) or cell-free ribonucleic acid (cfRNA), the set of instructions comprising the following instructions: For sequence datasets of cfDNA or cfRNA samples, overlapping sequencing read mates were merged; estimating a total tumor cfDNA fraction based at least in part on the sequence dataset of the cfDNA or cfRNA sample; For at least one potential SNV site in the sequence dataset, determining a genotype likelihood based at least in part on the estimated total tumor cfDNA fraction; Based at least in part on the genotype likelihoods, eliminating germline polymorphisms from a sequence dataset of the cfDNA or cfRNA sample to detect a plurality of somatic SNV candidates; analyzing the sequence dataset using a probability model that incorporates the estimated overall tumor cfDNA fraction and the plurality of somatic SNV candidates to determine a re-estimated overall tumor cfDNA fraction in the cfDNA or cfRNA sample; eliminating low-quality somatic SNV candidates from the sequence dataset using a set of filters that incorporate characteristics of the cfDNA or cfRNA sample; Eliminating inconsistent overlapping sequencing read mates from a sequence dataset of the cfDNA or cfRNA sample; Eliminate sequencing reads from sequence datasets with sequencing errors using a machine learning model that classifies sequencing errors from true variants; as well as Early minimal residual disease (MRD) is detected from the cfDNA or cfRNA sample using a truncated mutation spectrum and the machine learning model.

23. A computer-implemented method for detecting somatic single nucleotide variants (SNVs) from cell-free deoxyribonucleic acid (cfDNA) or cell-free ribonucleic acid (cfRNA), the computer-implemented method comprising: For sequence datasets of cfDNA or cfRNA samples, overlapping sequencing read mates were merged; estimating a total tumor cfDNA fraction based on the sequence dataset of the cfDNA or cfRNA sample; For at least one potential SNV site in the sequence dataset, determining a genotype likelihood based at least in part on the estimated total tumor cfDNA fraction; Based at least in part on the genotype likelihoods, eliminating germline polymorphisms from a sequence dataset of the cfDNA or cfRNA sample to detect a plurality of somatic SNV candidates; analyzing the sequence dataset using a probability model that incorporates the estimated overall tumor cfDNA fraction and the plurality of somatic SNV candidates to determine a re-estimated overall tumor cfDNA fraction in the cfDNA or cfRNA sample; eliminating low-quality somatic SNV candidates from the sequence dataset using a set of filters that incorporate characteristics of the cfDNA or cfRNA sample; Eliminating inconsistent overlapping sequencing read mates from a sequence dataset of the cfDNA or cfRNA sample; Eliminate sequencing reads from sequence datasets with sequencing errors using a machine learning model that classifies sequencing errors from true variants; as well as Early minimal residual disease (MRD) is detected from the cfDNA or cfRNA sample using a truncated mutation spectrum and the machine learning model.

24. A computer-implemented method for detecting somatic single nucleotide variants (SNVs) from a plurality of cell-free nucleic acid (cfNA) molecules from a subject, comprising: (a) retrieving a plurality of sequence reads generated by a sequencer, wherein at least a subset of the plurality of sequence reads comprises sequences from the plurality of cfNA molecules or derivatives thereof; (b) applying a probabilistic model to the plurality of sequence reads at each of a plurality of genetic loci to estimate an overall tumor burden of the plurality of cfNA molecules, wherein the estimated overall tumor burden comprises a quantitative measure of tumor-derived cfNA molecules in the plurality of cfNA molecules, wherein the plurality of genetic loci comprise potential SNV sites; (c) for each of the plurality of genetic loci, determining a likelihood of one or more genotypes for the subject based at least in part on the overall tumor burden, wherein the one or more genotypes are selected from a normal genotype, a tumor genotype, and a combined normal-tumor genotype; (d) for each of the plurality of genetic loci, detecting one or more SNVs based at least in part on the likelihood of the one or more genotypes determined in (c) and the quantitative measurement of tumor-derived cfNA molecules in the plurality of cfNA molecules; and (e) filtering out one or more germline polymorphisms from the one or more SNVs detected in (d) to obtain one or more somatic SNVs.

25. The computer-implemented method of claim 24, further comprising: (f) filtering out the one or more somatic SNVs obtained in (e) using one or more filters selected from the following to obtain a filtered set of somatic SNVs: a strand bias filter, a base quality filter, a read partner filter, a sequencing error filter, an insertion or deletion and homopolymer-induced error filter, and a public database filter.

26. The computer-implemented method of claim 25, further comprising: (g) applying the probabilistic model to the plurality of sequence reads at each of the plurality of genetic loci to re-estimate the overall tumor burden, wherein the plurality of genetic loci comprise the filtered set of somatic SNVs obtained in (f).

27. The computer-implemented method of claim 24, wherein the plurality of cell-free nucleic acid (cfNA) molecules comprises cell-free deoxyribonucleic acid (cfDNA) molecules or cell-free ribonucleic acid (cfRNA) molecules.

28. The computer-implemented method of claim 24, wherein estimating the overall tumor burden comprises combining information from the plurality of sequence reads across the potential SNV sites to reduce noise in the plurality of sequence reads caused by errors in the sequencing at one or more of the potential SNV sites.

29. The computer-implemented method of claim 26, wherein re-estimating the overall tumor burden comprises combining information from the plurality of sequence reads across the filtered set of somatic SNVs to reduce noise in the plurality of sequence reads caused by errors in the sequencing at one or more of the filtered set of somatic SNVs.

30. The computer-implemented method of claim 29, wherein combining the information comprises calculating a value that maximizes the likelihood of observing the plurality of sequence reads at each of a plurality of predetermined SNV hotspots, given the overall tumor burden.

31. The computer-implemented method of claim 24, wherein determining the likelihood of the one or more genotypes for each of the plurality of genetic loci comprises determining the likelihood of observing a number of the plurality of sequence reads covering each of the plurality of genetic loci given the one or more genotypes.

32. The computer-implemented method of claim 24, wherein determining the likelihood of the one or more genotypes for each of the plurality of genetic loci comprises calculating a maximum a posteriori probability estimate.

33. The computer-implemented method of claim 24, wherein the probabilistic model comprises a machine learning model configured to classify sequence reads having true variants from sequence reads having sequencing errors.

34. A system for detecting somatic single nucleotide variants (SNVs) from a plurality of cell-free nucleic acid (cfNA) molecules from a subject, comprising: A database storing a plurality of sequence reads, wherein at least a subset of the plurality of sequence reads comprises sequences from the plurality of cfNA molecules or derivatives thereof; and One or more computer processors operatively coupled to the database, wherein the one or more computer processors are individually or collectively programmed to: (1) retrieving the plurality of sequence reads from the database; (2) applying a probabilistic model to the plurality of sequence reads at each of a plurality of genetic loci to estimate an overall tumor burden of the plurality of cfNA molecules, wherein the estimated overall tumor burden comprises a quantitative measure of tumor-derived cfNA molecules in the plurality of cfNA molecules, wherein the plurality of genetic loci comprise potential SNV sites; (3) for each of the plurality of genetic loci, determining a likelihood of one or more genotypes of the subject based at least in part on the overall tumor burden, wherein the one or more genotypes are selected from a normal genotype, a tumor genotype, and a combined normal-tumor genotype; (4) for each of the plurality of genetic loci, detecting one or more SNVs based at least in part on the likelihood of the one or more genotypes determined in (3) and the quantitative measurement of tumor-derived cfNA molecules in the plurality of cfNA molecules; as well as (5) Filtering out one or more germline polymorphisms from the one or more SNVs detected in (4) to obtain one or more somatic SNVs.

35. The system of claim 34, wherein the one or more computer processors are individually or collectively programmed to additionally: (6) Filtering out the one or more somatic SNVs obtained in (5) using one or more filters selected from the following to obtain a filtered set of somatic SNVs: chain bias filter, base quality filter, read partner filter, sequencing error filter, insertion or deletion and homopolymer-induced error filter, and public database filter.

36. The system of claim 35, wherein the one or more computer processors are individually or collectively programmed to additionally: (7) applying the probabilistic model to the multiple sequence reads at each of the multiple genetic loci to re-estimate the overall tumor burden, wherein the multiple genetic loci include the filtered set of somatic SNVs obtained in (6).

37. The system of claim 34, wherein the plurality of cell-free nucleic acid (cfNA) molecules comprises cell-free deoxyribonucleic acid (cfDNA) molecules or cell-free ribonucleic acid (cfRNA) molecules.

38. The system of claim 34, wherein estimating the overall tumor burden comprises combining information from the multiple sequence reads across the potential SNV sites to reduce noise in the multiple sequence reads caused by errors in the sequencing at one or more of the potential SNV sites.

39. The system of claim 36, wherein re-estimating the overall tumor burden comprises combining information from the multiple sequence reads across the filtered set of somatic SNVs to reduce noise in the multiple sequence reads caused by errors in the sequencing at one or more of the filtered set of somatic SNVs.

40. The system of claim 38, wherein combining the information comprises calculating a value that maximizes the likelihood of observing the plurality of sequence reads at each of a plurality of predetermined SNV hotspots, given the overall tumor burden.

41. The system of claim 34, wherein determining the likelihood of the one or more genotypes for each of the multiple genetic loci comprises determining the likelihood of observing several of the multiple sequence reads covering each of the multiple genetic loci given the one or more genotypes.

42. The system of claim 34, wherein determining the likelihood of the one or more genotypes for each of the plurality of genetic loci comprises calculating a maximum a posteriori probability estimate.

43. The system of claim 34, wherein the probabilistic model comprises a machine learning model configured to classify sequence reads having true variants from sequence reads having sequencing errors.

Images