EA048297B1 - MUTATIONAL ANALYSIS OF DNA IN PLASMA FOR CANCER DETECTION - Google Patents

Description

Links to related applications

This application is a non-provisional application and claims priority to U.S. Provisional Patent Application No. 61/662,878, entitled Mutational analysis of plasma DNA for cancer detection, filed June 21, 2012; U.S. Provisional Patent Application No. 61/682,725, entitled Pakamutational analysis of plasma DNA for cancer detection, filed August 13, 2012; U.S. Provisional Patent Application No. 61/695,795, entitled Mutational analysis of plasma DNA for cancer detection, filed August 31, 2012; and U.S. Provisional Patent Application No. 61/711,172, entitled Mutational analysis of plasma DNA for cancer detection, filed October 8, 2012, which are hereby incorporated by reference in their entireties.

State of the art

Tumor-derived DNA has been shown to be present in cell-free plasma/serum from cancer patients (Chen XQ et al. Nat Med 1996; 2: 1033–1035). Most recent approaches rely on direct analysis of mutations known to be associated with cancer (Diehl F et al. Proc Natl Acad Sci 2005; 102: 16368–16373; Forshew T et al. Sci Transl Med 2012; 4: 136ra68). Another approach examines cancer-associated copy number variations detected by random DNA sequencing in plasma (US Patent Publication No. 2013/0040824, Lo et al.).

It is known that over time, more than one cancer cell will gain a growth advantage and give rise to several clones of daughter cells. Ultimately, the growing tumor and/or its metastases will contain a conglomerate of groups of clonal cancer cells. This phenomenon is commonly referred to as tumor heterogeneity (Gerlinger M et al. N Engl J Med 2012; 366: 883–892; Yap TA et al. Sci Transl Med 2012; 4: 127ps10).

It is known that cancers are highly heterogeneous, i.e. the mutational profile of cancers of the same tissue type can vary significantly.

Consequently, direct analysis of specific mutations typically detects only a subset of cases in a particular cancer type for which association with those specific mutations is known. Furthermore, tumor-derived DNA typically accounts for only a small proportion of the DNA in human plasma; the absolute concentration of DNA in plasma is low. Consequently, when directly detecting a single cancer-associated mutation or a small subset of such mutations in plasma or serum, the assay sensitivity may be low even among patients with cancers in which the target mutations are known to be present. Furthermore, it has been shown that there is substantial intratumoral heterogeneity of mutations, even within a single tumor. Mutations may be detectable only in a subset of tumor cells. The differences in mutational profiles between the primary tumor and metastatic lesions are even greater. One example of heterogeneity between the original tumor and primary metastases involves the KRAS, BRAF, and PIK3CA genes in patients with colorectal cancer (Baldus et al. Clin Cancer Research 2010. 16:790-9.).

In a case where a patient has a primary tumor (carrying a KRAS mutation but not a PIK3CA mutation) and an occult metastatic lesion (carrying a PIK3CA mutation but not a KRAS mutation), if we focus on detecting the KRAS mutation in the primary tumor, the occult metastatic lesion will not be detected. However, if we include both mutations in the assay, we can detect both the primary tumor and the occult metastatic lesion. Therefore, an assay that includes both mutations will have greater sensitivity in detecting residual tumor tissue. This simple example becomes much more complex if we are screening for cancer and there is little or no information about the type of mutations that may occur.

Therefore, there is a need for new ways to perform widespread cancer screening, detection or assessment.

Brief description of the invention

In embodiments, the frequency of somatic mutations in a biological sample (e.g., plasma or serum) from a subject being screened or monitored for cancer may be examined compared to those in the constitutive DNA of the same subject. Random sequencing may be used to determine these frequencies. A parameter may be derived from these frequencies and used to establish a classification of the cancer level. False positives may be eliminated by requiring that any variant locus must have at least a certain number of read sequences (markers) at any altered locus, which allows for a more accurate parameter to be obtained. The relative frequencies of different altered loci may be analyzed to determine the level of tumor heterogeneity in a patient.

In one embodiment, the parameter may be compared to the same parameter obtained for a group of subjects without cancer or with a low risk of developing cancer. A significant difference between the parameter obtained for the test subject and the same parameter obtained for the group of subjects without cancer or with a low risk of developing cancer may indicate an increased risk that the test subject has cancer or a pre-malignant condition, or that he will develop cancer in the future. Thus, in one embodiment,

- 1 048297 DNA analysis of plasma can be performed without prior genomic information about the tumor. This implementation option is therefore particularly suitable for cancer screening.

In another embodiment, the described embodiments may also be used to monitor a patient with cancer after treatment and to detect the presence of residual tumor or tumor recurrence. For example, a patient with residual tumor or whose tumor has recurred will have an increased frequency of somatic mutations compared to a patient who does not have residual tumor or who does not have tumor recurrence. Monitoring may include obtaining samples from the patient with cancer at multiple time points after treatment to determine temporal changes in tumor-associated genetic aberrations in bodily fluids or other samples with extracellular nucleic acids, such as plasma or serum.

According to one embodiment, the method allows detecting cancer or a pre-malignant change in a subject. A constitutive genome of the subject is obtained. One or more sequence markers are obtained for each of a plurality of DNA fragments in a biological sample of the subject, if the biological sample contains extracellular DNA. The positions of the sequence markers in the genome are determined. The sequence markers are compared with the constitutive genome to determine a first number of first loci. In each of the first loci, the number of sequence markers with a variant sequence relative to the constitutive genome is above a threshold value, where the threshold value is greater than one. Based on the count of sequence markers with a variant sequence in the first loci, a parameter is determined. The parameter is compared with the threshold value to determine the classification of the cancer level in the subject.

According to another embodiment, the method analyzes heterogeneity of one or more tumors in a subject. A constitutive genome of the subject is obtained. One or more sequence markers are obtained for each of a plurality of DNA fragments in a biological sample of the subject, if the biological sample contains extracellular DNA. The positions of the sequence markers in the genome are determined. The sequence markers are compared with the constitutive genome to determine the first number of first loci. In each of the first loci, the number of sequence markers with a variant sequence relative to the constitutive genome is above a threshold value, where the threshold value is greater than one. A measure of heterogeneity of one or more tumors is calculated based on the corresponding first numbers from the set of first positions in the genome.

According to another embodiment of the method, the fractional concentration of tumor DNA in a biological sample containing extracellular DNA is determined. One or more sequence markers are obtained for each of a plurality of DNA fragments in the biological sample. The positions of the sequence markers in the genome are determined. For each of a plurality of genome regions, the corresponding number of DNA fragments belonging to this genomic region is determined using the sequence markers located in the genome within this genomic region. The corresponding number is normalized to obtain the corresponding density. The corresponding density is compared with the reference density in order to detect the loss of 1 copy or the gain of 1 copy in this genomic region. The first density is calculated by the corresponding densities with the detected loss of 1 copy or by the corresponding densities with the detected gain of 1 copy. The fractional concentration is calculated by comparing the first density with another density to obtain the difference, wherein the difference is normalized to the reference density.

Other embodiments are directed to systems and computer-readable media associated with the methods described in this application.

A better understanding of the nature and advantages of the present invention can be obtained from the following detailed description and accompanying drawings.

Brief description of the figures

Fig. 1 is a flow chart of a method 100 for detecting cancer or a pre-cancerous lesion in a subject according to embodiments of the present invention.

Fig. 2 shows a flow chart of a method for directly comparing a sample genome (SG) with a constitutive genome (CG) according to embodiments of the present invention.

Fig. 3 shows a flow chart of a method 300 for comparing a sample genome (SG) with a constitutive genome (CG) using a reference genome (RG) according to embodiments of the present invention.

Fig. 4 is a table 400 showing the number of cancer-associated single nucleotide mutations correctly detected using the frequency of different amounts thereof as a criterion for classifying a mutation as present in a sample according to embodiments of the present invention, when the fractional concentration of tumor-derived DNA in the sample is set to 10%.

Fig. 5 shows a table showing the expected number of false positive loci and the expected number of mutations detected when the fractional concentration of tumor-derived DNA in the sample is taken to be 5%.

- 2 048297

Fig. 6A is a diagram 600 showing the detection frequency of cancer-associated mutations in plasma with fractional concentrations of tumor-derived DNA in plasma of 10% and 20%, and the use of incidences (r) of four and six as criteria for identifying potential cancer-associated mutations.

Fig. 6B is a diagram 650 showing the expected number of nucleotide positions that were misclassified as containing a nucleotide substitution using r criteria of 4, 5, 6, and 7, as a function of sequencing depth.

Fig. 7A is a diagram 700 showing the number of true cancer-associated mutated sites and false positive sites for different sequencing depths when the fractional concentration of tumor-derived DNA in the sample is set to 5%.

Fig. 7B is a diagram 750 showing the predicted number of false positive sites when performing a whole genome (WG) and all exons analysis.

Fig. 8 is a table 800 showing the results for 4 patients with hepatocellular carcinoma (HCC) before and after treatment, including fractional concentrations of tumor-derived DNA in plasma according to embodiments of the present invention.

Fig. 9 is a table 900 showing the detection of HCC-associated single nucleotide variations (SNVs) in 16 healthy control subjects according to embodiments of the present invention.

Fig. 10A shows a graph of the distribution of read sequence densities of a tumor sample from a patient with HCC according to embodiments of the present invention.

Fig. 10B shows a graph of the distribution of 1050 z-scores for all fragments in the plasma of a patient with HCC according to embodiments of the present invention.

Fig. 11 shows a graph of the distribution of 1100 z-scores for plasma from a patient with HCC according to embodiments of the present invention.

Fig. 12 is a flow chart of a method 1200 for determining a fractional concentration of tumor DNA in a biological sample containing extracellular DNA, according to embodiments of the present invention.

Fig. 13A shows Table 1300 of the analysis of mutations in the plasma of a patient with ovarian cancer and breast cancer at the time of diagnosis according to embodiments of the present invention.

Fig. 13B shows Table 1350 of the analysis of mutations in the plasma of a patient with bilateral ovarian cancer and breast cancer after tumor resection according to embodiments of the present invention.

Fig. 14A is a table 1400 showing the detection of single nucleotide variations in plasma DNA for patient 1 with HCC (HCC1).

Fig. 14B is a table 1450 showing the detection of single nucleotide variations in plasma DNA for patient HKK2.

Fig. 15A is a table 1500 showing the detection of single nucleotide variations in plasma DNA for patient HKK3.

Fig. 15B is a table 1550 showing the detection of single nucleotide variations in plasma DNA for patient HKK4.

Fig. 16 is a Table 1600 showing the detection of single nucleotide variations in DNA in plasma for a patient with ovarian (and breast) cancer.

Fig. 17 is a table 1700 showing the predicted sensitivities of different frequency and sequencing depth requirements.

Fig. 18 shows Table 1800, which shows the predicted numbers of false positive loci for different threshold values and different sequencing depths.

Fig. 19 shows a tree diagram illustrating the number of mutations detected in different tumor locations.

Fig. 20 shows Table 2000, which shows the number of fragments carrying tumor-related mutations in a plasma sample before and after treatment.

Fig. 21 is a diagram 2100 showing the distributions of the plasma incidence of mutations detected in one tumor site and mutations detected in all four tumor sites.

Fig. 22 is a diagram 2200 showing the predicted distribution of the incidence of mutations in plasma originating from a heterogeneous tumor.

Fig. 23 shows the specificity of the implementation variants for 16 healthy control subjects recruited.

Fig. 24 is a flow chart of a method 2400 for analyzing heterogeneity of one or more tumors of a subject according to embodiments of the present invention.

Fig. 25 is a block diagram of a typical computer system 2500 useful with the system and methods according to embodiments of the present invention.

- 3 048297

Definitions

In this application, the term locus or its plural form loci represents the position or address of a fragment of nucleotides (or base pairs) of any length that may contain variations in different genomes. A fragment is a section of a predetermined length in a genome. A plurality of fragments may have the same first length (resolution), while a different plurality may have the same second length. In one embodiment, the fragments do not overlap with each other.

The term random sequencing in this application refers to sequencing in which the nucleic acid fragments to be sequenced are not specifically identified or predetermined prior to the sequencing procedure. Sequence-specific primers to target specific gene loci are not required. The term universal sequencing refers to such sequencing in which the sequencing process can begin from any fragment. In one embodiment, adapters are added to the end of the fragment and sequencing primers are coupled to the adapters. In this way, any fragment can be sequenced using the same primer and thus the sequencing can be random.

The term sequence marker (also referred to as a readable sequence) as used herein refers to a string of nucleotides sequenced from any portion or all of a nucleic acid molecule. For example, a sequenced marker may be a short string of nucleotides (e.g., ~30) sequenced from a nucleic acid fragment, a short string of nucleotides from both ends of a nucleic acid fragment, or sequencing of an entire nucleic acid fragment that is present in a biological sample. A nucleic acid fragment is any portion of a longer nucleic acid molecule. A fragment (e.g., a gene) may exist separately (i.e., unlinked) from other portions of the longer nucleic acid molecule.

In this application, the constitutive genome (also referred to as CG) consists of consensus nucleotides at loci within the genome and can therefore be considered a consensus sequence. The CG may encompass the entire genome of a subject (e.g., the human genome) or only portions of the genome. The constitutive genome (CG) may be obtained from cellular DNA as well as from cell-free DNA (e.g., that which can be found in plasma). Ideally, the consensus nucleotides should indicate that the locus is homozygous for one allele or heterozygous for two alleles. A heterozygous locus typically includes two alleles that contain a genetic polymorphism. As an example, a criterion for determining whether a locus is heterozygous may be a threshold for each of the two alleles to appear in at least a predetermined percentage of sequence reads (e.g., 30% or 40%) aligned to the locus. If a single homozygous nucleotide occurs at a sufficient percentage (e.g., 70% or more), then the locus can be considered homozygous in the GC. Although the genome of one healthy cell may differ from that of another healthy cell due to random mutations that occur spontaneously during cell division, the GC should not change when such a consensus is used. Genomic rearrangements may occur in the genomes of some cells, for example, in B and T lymphocytes, such as rearrangements affecting antibody and T cell receptor genes. Such large-scale differences would still affect a relatively small subset of the total population of nucleated cells in the blood, and thus such rearrangements would not affect the definition of the constitutive genome if sufficient sampling (e.g., sequencing depth) of blood cells were available.

Other cell types, including buccal cells, skin cells, hair follicles, or biopsies of various normal body tissues, can also serve as sources of CG.

The term constitutive DNA refers to any source of DNA that reflects the genetic makeup with which a subject is born. Examples of constitutive samples of a subject from which constitutive DNA can be obtained include healthy blood cell DNA, buccal cell DNA, and hair root DNA. DNA from these healthy cells constitutes the CG of the subject. It can be determined that the cells are healthy in a variety of ways, such as when the individual is known to be cancer-free, or the sample can be obtained from tissue that is unlikely to contain cancerous or pre-cancerous cells (e.g., hair root DNA when liver cancer is suspected). As another example, a plasma sample can be obtained when the patient is cancer-free, and the detected constitutive DNA compared to the results for a subsequently obtained plasma sample (e.g., one year or later). In another embodiment, a single biological sample containing <50% tumor DNA can be used to detect the constitutive genome and tumor-associated genetic changes. In such a sample, the concentrations of tumor-associated single nucleotide mutations will be lower than those for each allele of the heterozygous single nucleotide polymorphisms of the SNPs in the CG. Such a sample may be the same as the biological sample used for the genome determination of the sample described below.

The term biological sample in this application refers to any sample that is collected from a subject (e.g., a human, an individual suffering from cancer, an individual suspected of having

- 4 048297 cancer, or in other organisms), and includes one or more extracellular nucleic acid molecule(s) of interest. The biological sample may include extracellular DNA, some of which may have originated from healthy cells and some of which may have originated from tumor cells. For example, tumor DNA may be found in blood or other fluids, such as urine, pleural fluid, ascitic fluid, peritoneal fluid, saliva, tears, or cerebrospinal fluid. A non-fluid example is a stool sample, which may be mixed with diarrheal fluid. For some of these samples, the biological sample may be obtained non-invasively. In some embodiments, the biological sample may be used as a constitutive sample.

The term sample genome (also called SG) is a set of sequence reads that are aligned with positions in a genome (e.g., the human genome). The sample genome (SG) is not a consensus sequence, but includes nucleotides that can only appear at a sufficient number of sequence reads (e.g., at thresholds of at least 2 or 3 or more). If an allele appears a sufficient number of times and is not part of the CG (i.e., is not part of the consensus sequence), then this allele can denote a single nucleotide mutation (also called OHM). Using the present invention, other types of mutations can also be detected, such as mutations involving two or more nucleotides (such as those affecting the number of tandem repeats in a microsatellite, or such as a simple tandem repeat polymorphism), chromosomal transposition (which can be intrachromosomal or interchromosomal) and sequence inversion.

The term reference genome (also called RG) refers to a haploid or diploid genome to which reads from a biological sample and a constitutive sample can be aligned and compared. Each locus in a haploid genome can only contain one nucleotide. Heterozygous loci can be found in a diploid genome, where a locus contains two alleles, where either allele can provide a match for alignment with that locus.

The term cancer grade may refer to any of the following: the presence of cancer, the stage of cancer, the size of the tumor, and/or other measures of the severity of cancer. Cancer grade may be a quantity or other characteristic. The grade may be zero. Cancer grade also includes pre-malignant or precancerous conditions (statuses) associated with mutations or sets of mutations. Cancer grade may be used in a variety of ways. For example, screening may test for cancer in someone who has not previously been diagnosed with cancer. Assessment may test for cancer in someone who has been diagnosed with cancer. Detection may mean screening, or it may mean checking to see if someone who has signs suggestive of cancer (such as symptoms or other positive tests) has cancer.

Detailed description of the invention

Implementation options are provided for detecting cancer by analyzing a biological sample (e.g., a plasma/serum sample) that is not taken directly from a tumor and that contains extracellular nucleic acids. Extracellular nucleic acids can originate from various types of tissue throughout the body. Thus, a broad analysis can be performed to detect various cancers.

Genetic aberrations (including single nucleotide mutations, deletions, amplifications and rearrangements) accumulate in tumor cells during the development of cancer diseases. In embodiments, massively parallel sequencing can be used to detect and quantify single nucleotide mutations (OHMs), also called single nucleotide variations (SNVs), in biological fluids (e.g., plasma, serum, saliva, ascitic fluid, pleural fluid and cerebrospinal fluid) to detect and monitor cancer diseases. Analysis of the amount of OHMs (or other types of mutations) can provide a mechanism for detecting early stages of cancer as part of screening tests. In various embodiments, care is taken to distinguish sequencing errors and to distinguish spontaneous mutations occurring in healthy cells (e.g., by requiring the detection of several OHMs, such as at least 3, 4 or 5, at a particular locus).

In some embodiments, non-invasive methods are also provided for analyzing tumor heterogeneity, which may include cells from the same tumor (i.e., intra-tumor heterogeneity) or cells from different tumors (either from the same site or from different sites) within the body. For example, the clonal structure of such tumor heterogeneity can be non-invasively analyzed, including assessing the relative abundance of tumor cells containing each mutation. Mutations that are present in higher relative concentrations are present in more malignant cells in the body, such as cells that arose earlier in the tumorigenesis process compared to other malignant cells also present in the body (Welch JS et al. Cell 2012; 150: 264-278). Such mutations, due to their higher relative abundance, are expected to provide greater diagnostic sensitivity for the detection of cancer DNA than those with lower relative abundance. Periodically

- 5 048297 logical monitoring of changes in the relative prevalence of mutations will provide the possibility of noninvasive monitoring of changes in the clonal composition of tumors, either spontaneous as the disease progresses or in response to treatment. Such information will be useful for assessing prognosis or for early detection of tumor resistance to treatment.

I. Introduction.

Mutations may arise during cell division due to errors in DNA replication and/or DNA repair. One type of such mutation involves a change in single nucleotides, which may occur in multiple sequences from different regions of the genome. Cancers are generally believed to arise from the clonal expansion of a single cancer cell that has gained a growth advantage. Such clonal expansion will result in the accumulation of mutations (e.g., single nucleotide mutations) in all cancer cells that have evolved from the original cancer cell. Such secondary tumor cells will have the same set of mutations (e.g., single nucleotide mutations). This application describes that cancer-associated single nucleotide mutations are found in the plasma/serum of patients with cancer.

Some embodiments can efficiently screen for all mutations in a biological sample (e.g., plasma or serum). Because the number of mutations is not fixed (hundreds, thousands, or millions of cancer-associated mutations may be detected from different tumor cell subpopulations), embodiments can provide greater sensitivity than techniques that detect specific mutations. Multiple mutations can be used to detect cancer.

In order to perform such screening of a plurality or all mutations, embodiments may search (e.g., randomly search) for genetic changes in a biological sample (e.g., in physiological fluids, including plasma and serum), which may contain tumor-derived DNA. Using a sample such as plasma eliminates the need for an invasive tumor or cancer biopsy. In addition, since the screening may cover the entire genome or large portions of the genome, such screening is not limited to any enumerated and known mutations, but may use any existing mutation. Moreover, since the number of mutations is summed up over the entire genome or large portions of the genome, higher sensitivity may be achieved.

However, there are polymorphic sites in the human genome, including single nucleotide polymorphisms (SNPs), which should not be considered mutations. The embodiments can determine whether the genetic changes that have been detected are likely to be cancer-associated mutations or whether they represent polymorphisms in the genome. For example, in the context of distinguishing between cancer-associated mutations and polymorphisms in the genome, the embodiments allow the determination of a constitutive genome that may include polymorphisms.

Constitutive genome (CG) polymorphisms can be limited to polymorphisms that are represented at a sufficiently high percentage (e.g., 30-40%) in the sequencing results.

Sequences obtained from a biological sample can then be aligned with the housekeeping genome to detect variations that represent single nucleotide mutations (OHMs) or other types of mutations. These OHMs will represent variations that are not included in the known polymorphisms and can thus be noted as being associated with cancer and not part of the housekeeping genome. A healthy individual may have some OHMs due to random mutations in healthy cells, such as those that arise during cell division, but an individual with cancer will have many more.

For example, in an individual suffering from cancer, the amount of OHMs detected in a physiological fluid will be higher than the amount of polymorphisms present in the constituent genome of the same individual. The amounts of variation detected in a physiological fluid sample containing tumor-derived DNA can be compared with a DNA sample containing predominantly constituent DNA. In one embodiment, the term predominantly will mean greater than 90%. In another preferred embodiment, the term predominantly will mean greater than 95, 97%, 98%, or 99%. If the amount of variation in a physiological fluid exceeds that in a sample with predominantly constituent DNA, there is an increased likelihood that the physiological fluid may contain tumor-derived DNA.

One method that can be used to randomly search for variations in DNA samples is random or shotgun sequencing (e.g., using massively parallel sequencing). Any massively parallel sequencing platform can be used, including ligation sequencing platforms (e.g., Life Technologies SOLiD Platform), Ion Torrent/Ion Proton, semiconductor sequencing, Roche 454, single-molecule sequencing platforms (e.g., Helicos, Pacific Biosciences, and Nanopore). In addition, it is known that sequencing errors can occur, which can be misinterpreted as variations in the constitutive DNA or as mutations,

- 6 048297 from tumor DNA. Thus, to improve the specificity of our proposed approach, the probability of sequencing error or other components of analytical errors can be taken into account, for example, by using an appropriate sequencing depth along with the requirement to detect at least a certain number (e.g. 2 or 3) of alleles in a locus to consider it as containing OHM.

This application describes that in embodiments, evidence of the presence of tumor-derived DNA in a biological sample (e.g., a physiological fluid) can be provided when the number of incidentally detected genetic variations present in the sample exceeds that expected for constitutive DNA and for variations that may be incidentally detected due to analytical errors (e.g., sequencing errors). The obtained information can be used for screening, diagnosis, prognosis, and monitoring of cancer diseases. In the following sections, we describe analytical steps that can be used to detect single nucleotide mutations in plasma/serum or other samples (e.g., physiological fluids). Physiological fluids can include plasma, serum, cerebrospinal fluid, pleural fluid, ascites fluid, nipple secretions, saliva, bronchoalveolar lavage fluid, sputum, tears, sweat, and urine. In addition to physiological fluids, this method can also be applied to a stool sample, as the latter has been shown to contain tumor DNA from colorectal cancer (Berger BM, Ahlquist DA. Pathology 2012; 44: 80-88).

II. Basic screening method.

Fig. 1 is a flow chart of a method 100 for detecting cancer or a pre-cancerous change in a subject according to embodiments of the present invention. Embodiments analyze extracellular DNA in a biological sample from a subject to detect variations in extracellular DNA likely to originate from a tumor. The analysis may use the subject's constituent genome to account for polymorphisms that belong to healthy cells and may account for sequencing errors. The method 100 and any of the methods described herein may be implemented in whole or in part by a computer system that includes one or more processors.

At step 110, a constituent genome of the subject is obtained. The constituent genome (CG) can be determined from the constituent DNA of the subject under study. In various embodiments, the CG can be read from memory or determined in an active manner, such as by analyzing read sequences of the constituent DNA that can be present in cells from a sample that includes extracellular DNA. For example, if a non-hematological malignancy is suspected, blood cells can be analyzed to determine the constituent DNA of the subject.

In various embodiments, the analysis of the constitutive DNA can be performed using massively parallel sequencing, chip hybridization, probe-based solution hybridization, ligation-based assays, primer extension reaction assays, and mass spectrometry. In one embodiment, the CG can be determined at one point in the life of a subject, such as at birth or even prenatally (which can be done using fetal cells or using cell-free DNA fragments, see US Publication No. 2011/0105353), and then compared to it with results from bodily fluids or other samples obtained at other points in the life of the subject. In this way, the CG can be simply read from computer memory. The constitutive genome can be read as a list of loci at which the constitutive genome differs from the reference genome.

At step 120, one or more sequence markers are obtained for each of a plurality of DNA fragments in a biological sample of a subject, if the biological sample contains extracellular DNA. In one embodiment, said one or more sequence markers are obtained as a result of random sequencing of DNA fragments in the biological sample. More than one sequence marker may be obtained by performing paired-end sequencing. One marker will correspond to each end of a DNA fragment.

An analysis can be performed for the presence of genetic changes in cell-free DNA in a sample (e.g., plasma, serum, or other biological fluid). The cell-free DNA can be analyzed using the same assay platform that was used to analyze the constitutive DNA. Alternatively, a different assay platform can be used. For example, the cell-free DNA sample can be sequenced using massively parallel sequencing, or parts of the genome can be captured or enriched before massively parallel sequencing. When enrichment is used, for example, liquid-phase or solid-phase capture of selected parts of the genome can be used. The captured DNA can then be massively parallel sequenced.

At step 130, the genomic positions of the sequence markers are determined. In one embodiment, the sequence markers are aligned with a reference genome that is obtained from one or more other subjects. In another embodiment, the genomic sequence markers

- 7 048297 is aligned with the constitutive genome of the subject under study. The alignment can be performed using techniques known to a person skilled in the art, for example, using the basic local alignment search tool (BLAST).

In step 140, a first number of loci are determined in which at least N sequence markers contain a sequence variant compared to the housekeeping genome (WG). N is equal to or greater than two. As discussed in more detail below, sequencing errors as well as somatic mutations that occur randomly in cells (e.g., as a result of cell division) can be eliminated by requiring N to be equal to 2, 3, 4, 5 or greater. Loci that meet one or more specified criteria can be considered a mutation (variant) or mutant loci (altered loci), while a locus that contains a variant but does not meet said one or more criteria (e.g., only one altered sequence marker) is referred to as a potential or putative mutation. Only one nucleotide or multiple nucleotides in a sequence can be altered.

N can be defined as a percentage of all markers for a locus rather than an absolute value. For example, an altered locus can be detected when a certain fractional concentration of tumor DNA resulting from variant reads is equal to or greater than 10% (or some other percentage). In other words, if the coverage of a locus is 200 reads, at least 10 reads in which the variant allele is detected can be required as a criterion for determining that a given variant is a mutation. 10 reads of the variant allele and 190 reads of the wild-type allele mean that the fractional concentration of tumor DNA is 10% (2 x 10/(10+190)).

In one embodiment, sequence markers (collectively referred to as the sample genome) can be compared directly to the CG to determine variants. In another embodiment, the sample genome (SG) is compared to the CG using a reference genome (RG) to determine variants. For example, both the CG and SG can be compared to the RG to determine the corresponding numbers (e.g., sets) of loci at which variants are detected, and then the difference can be calculated to obtain a first number of loci. The first number can be obtained simply as a number, or it can correspond to a certain set of loci, which can then be further analyzed to determine a parameter from the sequence markers at the first loci.

In one embodiment, the results of sequencing of the housekeeping DNA and the DNA in the plasma are compared to determine whether a single nucleotide mutation is present in the DNA in the plasma. The regions in which the housekeeping DNA is homozygous may be analyzed. For the purpose of illustration, assume that the genotype at a certain locus of the housekeeping DNA is homozygous and is AA. Then, the presence of an allele other than A in the plasma would indicate the potential presence of a single nucleotide mutation (OHM) at that locus. The loci in which the potential presence of OHM is detected may constitute the first number of loci in step 140.

In one embodiment, it may be useful to target parts of the genome that are known to be particularly susceptible to mutations in a particular type of cancer or in a particular subpopulation. With respect to the latter aspect, embodiments may search for types of mutations that are particularly prevalent in a particular population group, such as mutations that are particularly prevalent in subjects who are carriers of the hepatitis B virus (for liver cancer) or the human papillomavirus (for cervical cancer), or who have a genetic predisposition to somatic mutations, or in subjects with germline mutations in a DNA mismatch repair gene. The method will also be useful for screening for the presence of mutations in ovarian and breast cancers in subjects with mutations in the BRCA1 or BRCA2 genes. The method will also be useful for screening for the presence of mutations in colorectal cancer in subjects with mutations in the APC gene.

In step 150, the parameter is determined based on the count of sequence markers with a variant sequence in the first loci. In one example, the parameter is the first number of loci in which at least N DNA fragments contain a sequence variant in the locus compared to the constitutive genome. Thus, the count can be used to simply ensure that more than N copies of a particular variant are detected in the locus, which is determined before it is included in the first number. In another embodiment, the parameter can be or include the total number of sequence markers with a sequence in the first loci that is altered relative to the constitutive genome.

At step 160, the parameter for the subject is compared to a threshold value (e.g., obtained from one or more other subjects) to determine a classification of the cancer level in the subject. Examples of cancer level include the presence of cancer or a pre-malignant condition in the subject, or an increased likelihood of developing cancer. In one embodiment, the threshold value can be determined in a sample previously obtained from the subject.

- 8 048297

In another embodiment, it may be determined that one or more other subjects do not have cancer or have a low risk of developing cancer. Thus, the threshold value may be a normal value, a normal range, or indicate a statistically significant deviation from a normal value or range. For example, the number of mutations, compared to the CG of a particular subject, detected in the plasma of subjects who do not have cancer or who have a low risk of developing cancer may be used as a normal range to determine whether the number of mutations detected in the subject being tested is normal. In another embodiment, it may be known that other subjects have cancer and, thus, a similar number of mutations may indicate the presence of cancer.

In one embodiment, the other subjects may be selected to have clinical properties that match those of the test subject, such as gender, age, diet, smoking habits, drug history, previous illness, family history, genotypes at selected genomic loci, status with respect to viral infections (e.g., hepatitis B or C virus infections, or human papillomavirus, or human immunodeficiency virus, or Epstein-Barr virus) or infections with other infectious agents (such as bacteria (e.g., Helicobacter pylori) and parasites (e.g., Clonorchis sinensis), etc.). For example, subjects who are carriers of hepatitis B or C virus have an increased risk of developing hepatocellular carcinoma. Thus, test subjects who have a similar number of mutations or mutation pattern to a carrier of hepatitis B or C can be considered to have an increased risk of developing hepatocellular carcinoma. On the other hand, a patient with hepatitis B or C who is found to have a higher number of mutations than another patient with hepatitis can be correctly classified as having a higher cancer rate because the appropriate baseline condition is used (i.e., comparison with another patient with hepatitis). Similarly, subjects who are carriers of human papillomavirus infection have an increased risk of cervical cancer and head and neck cancer. Epstein-Barr virus infection has been associated with nasopharyngeal carcinoma, gastric cancer, Hodgkin lymphoma, and non-Hodgkin lymphoma. Helicobacter pylori infection has been associated with gastric cancer. Clonorchis sinensis infection has been associated with cholangiocarcinoma.

Monitoring changes in mutation counts at different time points can be used to monitor cancer progression and response to treatment. Such monitoring can also be used to confirm progression of a pre-malignant condition or a change in the risk that a subject will develop cancer.

The number of sequence markers in which variations are detected can also be used for monitoring. For example, the fractional concentration of read sequences of a variant at a locus can be used. In one embodiment, an increase in the fractional concentrations of tumor-associated genetic aberrations in samples during serial monitoring can indicate disease progression or impending relapse. Similarly, a decrease in the fractional concentrations of tumor-associated genetic aberrations in samples during serial monitoring can indicate a response to treatment and/or remission and/or a favorable prognosis.

III. Identification of genomes.

The various genomes discussed above are explained in more detail below. For example, the reference genome, the constitutive genome, and the sample genome are discussed.

A. Reference genome.

The reference genome (RG) refers to the haploid or diploid genome of a subject or to a population consensus. The RG is known and can therefore be used to compare reads to it when sequencing new patient genomes. The reads for a patient sample can be aligned and compared to determine variations in these reads compared to the RG. In a haploid genome, only one nucleotide is present at each locus, and therefore each locus can be considered hemizygous. In a diploid genome, heterozygous loci can be found, where such loci contain two alleles, and when aligned to a given locus, a match can be with either allele.

The reference genome may be the same in a population of subjects. Such a same reference genome can be used to determine in healthy subjects an appropriate threshold to use for classifying a patient (e.g. as having or not having cancer). However, different reference genomes can be used for different populations, e.g. for different ethnic groups or even for different families.

B. Constitutive genome.

The constitutive genome (CG) of a subject (e.g., a human or other diploid organism) refers to the diploid genome of the subject. At heterozygous loci, the CG can be precisely determined to have the first allele belonging to the first haplotype and a distinct second allele belonging to the second haplotype. It should be noted that the structures of the two haplotypes that span the two heterozygous loci do not necessarily have to be known, i.e. which allele of one heterozygous locus is in that

- 9 048297 same haplotype as the allele of the other heterozygous locus. Simply having two alleles at each heterozygous locus may be enough.

The CG may differ from the EG by polymorphisms. For example, a locus in the EG may be homozygous for T, while in the CG it may be heterozygous: T/A. Thus, for the CG, a variation at this locus will be detected. The CG may also differ from the EG by inherited mutations (for example, those inherited by family members) or de novo mutations (those that occur in the embryo but are not present in its parents). An inherited mutation is usually called a germline mutation. Some such mutations are associated with a predisposition to cancer, such as a mutation in BRCA1, which is inherited by family members. Such mutations are distinct from somatic mutations, which may occur as a result of cell division throughout the life of the subject and may cause a cell and its daughter cells to become cancerous.

The purpose of defining the CG is to separate such germline and de novo mutations from sample genome (SGM) mutations to define somatic mutations. The number of somatic mutations in the SGM can then be used to estimate the likelihood of a subject developing cancer. These somatic mutations can be further screened to remove sequencing errors and potentially remove somatic mutations that are rare (e.g., a variant found in only one sequence read), as these somatic mutations are unlikely to be associated with cancer.

In one embodiment, the GC may be determined using cells (DNA from the buffy coat). However, the GC may also be determined using cell-free DNA (e.g., from plasma or serum). In a sample type in which the majority of cells are non-malignant, such as the buffy coat from a healthy subject, the majority of the genome, or consensus genome, is the GC. In the GC, each genomic locus consists of a DNA sequence that is shared by the majority of cells in the sample tissue. The sequencing depth should be sufficient to detect heterozygous sites in the constitutive genome.

As another example, plasma can be used as a constitutive sample for determining the CG. For example, in cases where the tumor DNA content in the plasma is less than 50% and the OHMs are heterozygous, e.g., the mutation is an insertion of a new allele, the concentration of the new allele can be less than 25%. While the concentration of heterozygous alleles of the SNPs in the CG should be up to about 50%. In this way, a somatic mutation and a polymorphism of the CG can be distinguished. In one embodiment, a suitable threshold value for distinguishing a somatic mutation from a polymorphism can be in the range of 30-40%, when using plasma or other mixtures with a significant concentration of tumor. Measuring the concentration of tumor DNA can be useful to ensure that the concentration of tumor DNA in the plasma is less than 50%. Examples of determining the concentration of tumor DNA are described in this application.

C. Genome of the sample.

The sample genome (SG) is not simply the haploid or diploid genome, as in the case of EG and CG. SG is the set of sequence reads of the sample and may include: sequence reads of constitutive DNA that correspond to the CG, sequence reads of tumor DNA, sequence reads of healthy cells that have random mutations compared to the CG (e.g., due to mutations that arise during cell division), and sequencing errors. Various parameters can be used to precisely control which sequence reads are included in the SG. For example, requiring that an allele appear in at least 5 sequence reads will reduce the sequencing errors present in the SG and also eliminate the portion of sequence reads associated with random mutations.

As an example, assume that the subject is healthy, i.e., does not have cancer. For the purpose of illustration, there is DNA from 1000 cells in 1 ml of plasma (i.e., 1000 genome equivalents of DNA) obtained from the subject. The DNA in plasma typically consists of DNA fragments of approximately 150 bp in size. Since the human genome is 3 x 10 ⁹ bp, there will be approximately 2 x 10 ⁷ DNA fragments per haploid genome. Since the human genome is diploid, there will be approximately 4 x 10 ⁷ DNA fragments in 1 ml of plasma.

Since millions to billions of cells release DNA into the plasma at any given time, and DNA fragments from these cells will mix in the bloodstream, 4 x 10 ⁷ DNA fragments could have come from 4 x 10 ⁷ different cells. If these cells are not of recent origin (as opposed to distant origin, such as from an original zygote) from a single clone (i.e., they do not share a common recent ancestral cell), then it is statistically likely that no mutation will occur more than once among these fragments.

On the other hand, if among the 1000 genome equivalents per ml of DNA in the plasma there is a certain percentage of cells that share a common recent ancestral cell (i.e. they are clonally related

- 10 048297 to each other), it can be seen that mutations from a given clone will be predominantly represented in DNA in plasma (e.g., representing a clonal mutational profile in plasma). Such clonally related cells may be cancer cells or cells that will soon become cancerous but are not yet cancerous (i.e., preneoplastic). Thus, requiring a mutation to occur more than once allows for the elimination of such natural variations in the mutations detected in a sample, allowing for the retention of mutations associated with cancer cells or preneoplastic cells, thereby enabling detection, especially early detection, of cancer or precancerous conditions.

One approximation has been that after each cell division, the genome will accumulate on average one mutation. A previous study showed that the majority of DNA in plasma originates from hematopoietic cells (Lui YY, et al. Clin Chem 2002: 48: 421-427). It has been estimated that hematopoietic stem cells replicate every 25 to 50 weeks (Catlin SN, et al. Blood 2011; 117: 4460-4466). Thus, as a simplified approximation, a healthy subject by age 40 will accumulate approximately 40 to 80 mutations per hematopoietic stem cell.

If there are 1000 genome equivalents in 1 ml of plasma from a given individual, and if each of these cells is derived from a single hematopoietic stem cell, one would expect to detect between 40,000 and 80,000 mutations among 4 x ^{10 10} DNA fragments (i.e. 4 x 10 ⁷ DNA fragments per genome and 1000 genome equivalents per ml of plasma). However, since each mutation will only occur once, each mutation may still be below the detection limit (e.g. if the cutoff N is greater than 1), and these mutations can therefore be screened out, allowing the analysis to focus on mutations that are more likely to be cancerous. The cutoff can be any value (integer or non-integer) greater than one, and can vary for different loci and fragments. Sequencing depth and fractional concentration of tumor DNA may also influence the sensitivity of mutation detection (e.g., percentage of detected mutations) from cancer cells or preneoplastic cells.

IV. Direct comparison of GO with CG.

In some embodiments, it is possible to determine the nucleotide positions at which KG is homozygous, but at a smaller number of molecules (i.e., in tumor DNA) in GO those positions are heterozygous. With a high sequencing depth at a position (e.g., greater than 50x coverage), it is possible to detect the presence of one or two alleles at that position in a mixture of normal and cancer cell DNA. If two alleles are detected, then either (1) KG is heterozygous, or (2) KG is homozygous but GO is heterozygous. These two options can be distinguished by looking at the relative amounts of the major and minor alleles. In the former, the amounts of reads of the two alleles will be similar; but in the latter, there will be a large difference in the amounts of read variants of each. This comparison of the relative amounts of allele read variants in a test sample represents one embodiment for comparing sequence markers to the constituent genome. It can be determined that the first loci from method 100 are loci in which the number of alleles is below the upper threshold (the threshold corresponding to polymorphisms in the CG) and above the lower threshold (the threshold corresponding to errors and somatic mutations occurring at a low enough frequency not to be associated with the cancerous condition). Thus, the constitutive genome and the first loci can be determined simultaneously.

In another embodiment, the process of identifying mutations may first determine the CG and then determine loci that contain a sufficient number of mutations compared to the CG. The CG may be determined from a constitutive sample that differs from the sample being tested.

Fig. 2 shows a flow chart of a method 200 for directly comparing a sample genome (SG) with a constituent genome (CG) according to embodiments of the present invention. In block 210, a constituent genome of a subject is obtained. The constitutive genome can be obtained, for example, from a sample obtained earlier, or from a constituent sample that is obtained and analyzed immediately before performing the method 200.

In block 220, one or more sequence markers are obtained for each of a plurality of DNA fragments in a biological sample of a subject. Sequencing can be performed using various techniques mentioned in this application. The sequence markers represent an estimate of what the sequence of the fragment is believed to be. But one or more bases in the sequence marker may be erroneously determined.

In block 230, at least a portion of the sequence markers are aligned with the constitutive genome. The alignment may take into account that the CG may be heterozygous at different loci. The alignment will not require an exact match to detect variants.

In block 240, sequence markers are detected that contain sequence changes at the locus compared to the constitutive genome. It is possible that more than one variant of a sequence marker will be detected. Variants can be tracked for each locus and for each sequence marker. A variant can be any allele that is not represented in the CG. For example, the CG may be heterozygous for A/T and the variant may be G or C.

- 11 048297

In block 250, for each locus with a variant, the computer system may determine a corresponding first number of sequence markers that align with the locus and contain a sequence change at the locus. In this way, each locus may be associated with a counted number of variants observed at the locus. Typically, a locus will have fewer variants than sequence markers that correspond to a CG, for example, because the tumor DNA concentration is less than 50%. However, in some samples, the tumor DNA concentration may be greater than 50%.

In block 260, the parameter is determined based on the corresponding first quantities. If, in one embodiment, the corresponding quantity is greater than a threshold value (e.g., greater than two), then the corresponding quantity can be added to the sum that represents the parameter or is used to determine the parameter. In another embodiment, the number of loci for which the corresponding quantity is greater than the threshold value is used as the parameter.

In block 270, the parameter is compared with a threshold value to classify the cancer level. As described above, the threshold value can be determined by analyzing samples from other subjects. Depending on whether these other subjects are healthy or suffer from cancerous conditions, a classification can be determined. For example, if other subjects are diagnosed with stage 4 cancer, then if the analyzed parameter was close (e.g., within a certain range) to the value of the parameter obtained for the other subjects, then it is possible to classify the cancer stage of the studied subject as stage 4 cancer. However, if the specified parameter crosses the boundaries of the threshold value (i.e., is greater than or less than the threshold value, depending on the method for determining the parameter), then it is possible to classify the cancer stage as less than stage 4. A similar analysis can be performed when the other subjects do not suffer from cancer.

Multiple thresholds may be used to determine the classification, with each threshold being determined for a different group of subjects. Each group of subjects may have the same cancer level. The parameter of interest may thus be compared to the values for each group of subjects, resulting in a match for one of the groups or a range. For example, the parameter may be approximately the same as the parameter obtained for subjects who are at the precancer stage or at stage 2 cancer. As another example, the parameter of interest may fall within a range that may possibly fit several different cancer levels. Thus, the classification may include more than one cancer level.

V. Use of the reference genome.

The genomic sequences of both the constitutive DNA and the DNA from the biological sample can be compared to a human reference genome. If more changes are detected in the plasma sample than in the constitutive DNA, compared to the reference genome, then there is an increased probability of cancer. In one embodiment, homozygous loci in the reference genome are examined. The numbers of heterozygous loci in the constitutive DNA and in the DNA from the biological sample are compared. If the number of heterozygous sites detected in the DNA from the biological sample exceeds that in the constitutive DNA, then there is an increased probability of cancer.

It is also possible to limit the analysis to loci that are homozygous in the CG. OHM can also be determined for heterozygous loci, but this will usually require the formation of a third variant. In other words, if the heterozygous locus is A/T, then the new third variant will be either C or G. Identifying OHM at homozygous loci is usually easier.

The degree to which the number of heterozygous loci in a biological DNA sample is increased relative to the constitutive DNA may suggest the presence of a cancer or pre-malignant condition when compared to the level of change observed in healthy subjects. For example, if the degree of increase in the number of such sites exceeds that observed in healthy subjects by some threshold, then the results may be considered to be indicative of the presence of a cancer or pre-malignant condition. In one embodiment, the distribution of mutations in subjects without cancer is determined, and some number of standard deviations (e.g., 2 or 3 standard deviations) may be used as a threshold.

In one embodiment, at least a certain number of variants at a locus may be required before the locus is taken into account. In another embodiment, even results obtained from a single substitution are analyzed. For example, if the total number of variations (errors + true mutations or polymorphisms) observed in plasma is statistically significantly higher than in the constitutive DNA, then this indicates the presence of cancer.

Fig. 3 shows a flow chart of a method 300 for comparing a sample genome (SG) with a constitutive genome (CG) using a reference genome (RG) according to embodiments of the present invention. The method 300 assumes that the RG has already been obtained and that sequence markers in the biological sample have already been obtained.

In block 310, at least a portion of the sequence markers are aligned with a reference genome. The alignment may allow for mismatches, as variations need to be detected. The reference genome may be obtained from a population similar to the subject. The aligned markers

- 12 048297 sequences actually include the sample genome (GO).

In block 320, a first quantity (A) of potential variants, such as single nucleotide mutations (OHMs), is identified. Potential OHMs are loci in which a nucleotide is detected in the GO sequence marker that is different from that in the EG. Other criteria may be used, such as the number of sequence markers in which the variation is detected must be greater than a threshold value, and whether the locus in the EG is homozygous. The set of potential OHMs may be represented as a set A, where certain loci are detected and tracked by storing the loci in memory. Certain loci may be detected, or the number of such OHMs may simply be determined.

In block 330, the constituent genome is determined by aligning sequence markers obtained by sequencing DNA fragments from the constituent sample with a reference genome. This step may have been performed at any earlier time and using a constituent sample obtained at any earlier time. The CG may simply be read from memory if the alignment has been performed earlier. In one embodiment, the constituent sample may be blood cells.

In block 340, a second number (B) of loci are detected in which the aligned KG sequence marker contains a variant (e.g., OHM) at the locus compared to the reference genome. If a set of loci is specifically monitored, then B may be a set rather than just a number.

In block 350, set B is subtracted from set A to determine variants (OHMs) that are present in the sample genome but not in the CG. In one embodiment, the set of OHMs may be limited to nucleotide positions at which the CG is homozygous. To accomplish this selection, certain loci at which the CG is homozygous may be designated as set C. In another embodiment, a locus is not counted in the first count A or the second count B if the CG is not homozygous at that locus. In another embodiment, any known polymorphism (e.g., due to its presence in a SNP database) may be screened out.

In one embodiment, the subtraction in block 350 may simply be a subtraction of numbers, and thus specific potential OHMs are not removed, but simply subtracted by their value. In another embodiment, the subtraction yields the difference between set A and set B (e.g., if set B is a subset of set A) to determine specific OHMs that are not represented in set B. In Boolean variables, this may be expressed as [A AND NOT(B)]. The resulting set of detected variants may be denoted by C. The specified parameter may be defined as a number C, or determined from set C.

In some embodiments, the nature of the mutations may be taken into account and different classes of mutations may be assigned a certain weight. For example, mutations that are commonly associated with cancer may be assigned a higher weight (also referred to as a significance level in terms of relative locus weights). Such mutations may be found in tumor-associated mutation databases, such as the Catalogue of Somatic Mutations in Cancer (COSMIC) (www.sanger.ac.uk/genetics/CGP/cosmic/). As another example, mutations associated with non-synonymous changes may be assigned a higher weight.

Thus, the first quantity A can be defined as a weighted sum, in which the weight of the number of markers with a variant at one locus can be different from that of the number of markers at another locus. The first quantity A can reflect such a weighted sum. A similar calculation can be made with B, and thus the quantity C and the specified parameter can reflect such a weighting. In another embodiment, the weights are taken into account when determining the set C of specific loci. For example, a weighted sum can be determined for the number of loci of the set C. Such weights can be used in other methods described in this application.

Accordingly, the parameter that is compared with the threshold value to determine the classification of the cancer level may be the number of loci in which the variation is detected in the GO and CG compared to the EG. In other embodiments, the total number of DNA fragments (which is determined using sequence markers) containing the variation may be counted. In other embodiments, such quantities may be used in another formula to obtain the said parameter.

In one embodiment, the concentration of the variant at each locus may be a parameter and may be compared to a threshold. Such a threshold may be used to determine whether a locus is a potential altered locus (in addition to a threshold of a certain number of sequence reads in which the variant was detected), and then score the locus. The concentration may also be used as a weighting factor in the OHM sum.

VI. Reducing the number of false positives.

Results using threshold values.

It was mentioned above that single nucleotide mutations can be observed in a large number of extracellular DNA fragments (e.g., circulating plasma DNA) from a large region of the genome (e.g., the entire genome) or from multiple regions of the genome to improve the sensitivity of the data.

- 13 048297 approach. However, analytical errors such as sequencing errors can affect the feasibility, accuracy, and specificity of this approach. Here, we use a massively parallel sequencing platform as an example to illustrate the importance of sequencing errors. The sequencing error rate of the Illumina sequencing-by-synthesis platform is approximately 0.1% to 0.3% per nucleotide sequenced (Minoche et al. Genome Biol 2011; 12:R112). Any massively parallel sequencing platform can be used, including a ligation-by-sequencing platform (e.g., the SOLiD platform from Life Technologies), the Ion Torrent/Ion Proton platform, the semiconductor sequencing platform, Roche 454, and single-molecule sequencing platforms (e.g., Helicos, Pacific Biosciences, and Nanopore).

An earlier study on hepatocellular carcinoma showed that there are approximately 3,000 single nucleotide mutations (SNPs) in the entire cancer genome (Tao Y et al. 2011 Proc Natl Acad Sci USA; 108: 12042–12047). Assuming that only 10% of all DNA in the circulation originated from tumor cells and that we sequenced plasma DNA with an average sequencing depth of 1x haploid genome coverage, we would encounter 9 million (3x109 x0.3%) SNPs due to sequencing errors. However, most of the SNPs are expected to occur on only one of the two homologous chromosomes. At a sequencing depth of 1x haploid genome coverage in a sample with 100% tumor DNA, we would expect to detect only half of the 3,000 mutations, i.e., 100% of the SNPs in the plasma were sequencing. 1500 mutations. When sequencing a plasma sample containing 10% tumor-derived DNA per haploid genome coverage, we expect to detect only 150 (1500x10%) cancer-associated single nucleotide mutations. Thus, the signal-to-noise ratio for detecting cancer-associated mutations is 1 in 60,000. This very low signal-to-noise ratio suggests that the reliability of this approach to discriminate between normal and cancer patients will be very low if we simply use all single nucleotide changes in a biological sample (e.g., plasma) as a parameter.

It is expected that with the advancement of sequencing technologies, the rate of sequencing errors will steadily decrease. It is also possible to analyze the same sample using more than one sequencing platform and, by comparing the sequencing results from different platforms, pinpoint the sequence reads likely to contain sequencing errors. Another approach is to analyze two samples taken at different time points from the same subject. However, such approaches are time-consuming.

In one embodiment, one method for increasing the signal-to-noise ratio in detecting single nucleotide mutations in the plasma of patients with cancer is to only consider a mutation that occurs multiple times in a sample. In selected sequencing platforms, sequencing errors involving certain nucleotide substitutions may be the most common and will affect the sequencing results of the test sample and the constituent DNA sample of both the test subject and the control subjects. However, sequencing errors typically occur randomly.

The probability of sequencing error decreases exponentially when the same substitution at the same nucleotide position is observed in multiple DNA fragments. On the other hand, the probability of detecting a true cancer-associated mutational change in a sample is affected by the sequencing depth and the fractional concentration of tumor DNA in the sample. The probability of detecting a mutation in multiple DNA fragments will increase with increasing sequencing depth and fractional concentration of tumor DNA. In various embodiments using samples containing extracellular tumor DNA (such as plasma), the fractional concentration may be 5%, 10%, 20%, and 30%. In one embodiment, the fractional concentration is less than 50%.

Fig. 4 is a table 400 showing the number of cancer-associated single nucleotide mutations correctly detected using the frequency of different amounts as a criterion for classifying a mutation as present in a sample according to embodiments of the present invention. Also shown are the numbers of nucleotide positions at which a mutation was erroneously detected due to sequencing error based on the same classification criteria. The sequencing error rate is assumed to be 0.1% (Minoche et al. Genome Bio 2011, 12:R112). The fractional concentration of tumor-derived DNA in the sample is assumed to be 10%.

Fig. 4 shows that the ratio between the number of cancer-associated mutations detected in plasma and the number of false positive signals will increase exponentially with the number of observations of the same substitution in the sample in which the mutation is detected, when the fractional concentration of tumor-derived DNA in the sample is taken to be 10%. In other words, both the sensitivity and specificity of cancer-associated mutation detection will improve. In addition, the sensitivity of cancer-associated mutation detection will be

- 14 048297 influence the sequencing depth. With 100-fold coverage sequencing of the haploid genome, 2205 (73.5%) of 3000 mutations can be detected, even with the criterion of occurrence of a particular mutation in at least 4 DNA fragments in the sample. Other values of the minimum number of fragments can be used, such as 3, 5, 8, 10, and more than 10.

Fig. 5 shows Table 500, which shows the expected number of false positive loci and the expected number of mutations detected when the fractional concentration of tumor-derived DNA in the sample is assumed to be 5%. At a lower fractional concentration of tumor-derived DNA in the sample, a higher sequencing depth will be required to achieve the same sensitivity for detecting cancer-associated mutations. A more stringent criterion will also be required to maintain specificity. For example, a criterion of occurrence of a single mutation in a sample in at least 5 DNA fragments would need to be used, instead of a criterion of occurrence in at least 4 fragments in the case of a tumor DNA fraction of 10%. Tables 400 and 500 provide guidance on selecting a threshold value based on the fold coverage and the concentration of tumor DNA that can be inferred or measured as described in this application.

Another advantage of using more than a single single nucleotide substitution detection criteria to define a mutation is that it is expected to minimize the detection of false positives that result from single nucleotide changes in non-malignant tissues. Since nucleotide changes can occur during mitosis of normal cells, each healthy cell in the body may contain multiple single nucleotide changes. These changes could potentially result in false positives. However, the changes will be released from the cell into the plasma/serum after cell death. Although different normal cells are expected to carry different sets of mutations, mutations that occur in a single cell are unlikely to be present in multiple copies in the plasma/serum. In contrast, multiple copies of mutations from tumor cells are expected to be found in the plasma/serum because tumor growth is clonal in nature. Thus, many cells from a single clone will die and release DNA with characteristic mutations typical for these clones.

In one embodiment, targeted enrichment for certain regions of the genome may be performed prior to sequencing. This targeted enrichment step may increase the depth of sequencing of regions of interest with the same total number of sequencing runs. In another embodiment, a round of sequencing may first be performed with a relatively low depth of sequencing. The material may then be enriched for regions in which at least one single nucleotide substitution is detected, and then a second round of sequencing may be performed with a higher fold coverage. A multiple occurrence criterion may then be applied to the results of the targeted enrichment sequencing to determine the mutation.

VII. Dynamic thresholds.

As described above, a threshold N can be applied to the number of reads in which a variant (potential mutation) is detected to determine whether a locus can be considered to contain a mutation (e.g., contains OHM) that should be considered. Applying such a threshold can reduce the number of false positives. The discussion below suggests methods for selecting a threshold for different loci. In the following embodiments, we assume that there is one predominant cancer clone. A similar analysis can be performed for cases involving multiple clones of cancer cells releasing different amounts of tumor DNA into the plasma.

A. Number of cancer-associated mutations detected in plasma.

The number of cancer-associated mutations detected in plasma can be influenced by many parameters, such as: (1) the number of mutations in tumor tissue (NT) - the total number of mutations present in tumor tissue represents the maximum number of tumor-associated mutations detectable in patient plasma; (2) fractional concentration of tumor-derived DNA in plasma (f) - the higher the fractional concentration of tumor-derived DNA in plasma, the higher the probability of detecting tumor-associated mutations in plasma will be; (3) sequencing depth (D) - sequencing depth refers to the number of times the sequence is read when sequencing a region. For example, an average sequencing depth of 10x means that each nucleotide in the sequenced region is covered by an average of 10 sequence reads. The probability of detecting a cancer-associated mutation will increase as the sequencing depth increases; and (4) the minimum number of times a nucleotide substitution is detected in plasma to determine that it is a potential cancer-associated mutation (r), which is a threshold used to distinguish sequencing errors from true cancer-associated mutations.

In one embodiment, a Poisson distribution is used to predict the number of cancer-associated mutations detected in plasma. Assuming that a mutation is present at some nucleotide position on one of two homologous chromosomes, at a depth

- 15 048297 sequencing D, the expected number of times the mutation is present in plasma (M _P ) is calculated as:

M _P = Dxf/2.

The probability of detecting a mutation in plasma (Pb) at a particular mutation site is calculated as:

Pb = 1 - Σ Poisson _R ) = 0 where r (cutoff value) is the number of times a nucleotide substitution is detected in plasma to determine that it represents a potential tumor-associated mutation;

Poisson(|. M _P ) is the Poisson distribution of the probability of occurrence/times given the average number of MPs.

The total number of cancer-associated mutations expected to be found in plasma (N _P ) can be calculated as:

N _P =NTxPb.

where NT represents the number of mutations present in the tumor tissue.

The following graphs show the percentages of tumor-associated mutations expected to be found in plasma using different frequency (r) criteria for defining a potential mutation and different sequencing depths.

Fig. 6A is a diagram 600 showing the detection rate of cancer-associated mutations in plasma with fractional concentrations of tumor-derived DNA of 10% and 20% and using frequencies (r) of four and six as criteria for identifying potential cancer-associated mutations. For the same r, a higher fractional concentration of tumor-derived DNA in plasma will result in the detection of more cancer-associated mutations in plasma. For the same fractional concentration of tumor-derived DNA in plasma, a higher r will result in the detection of fewer mutations.

B. The number of false positive single nucleotide changes detected due to error.

Single nucleotide changes in plasma DNA in sequencing results may occur due to sequencing and alignment errors. The number of nucleotide positions with false-positive single nucleotide changes can be predicted mathematically based on the binomial distribution. Parameters affecting the number of false-positive sites (N _FP ) may include: (1) sequencing error rate (E) - the sequencing error rate is defined as the proportion of nucleotides incorrectly identified during sequencing; (2) sequencing depth (D) - the greater the sequencing depth, the greater the number of nucleotide positions at which a sequencing error will be detected; (3) the minimum number of detections of the same nucleotide substitution to determine that it is a potential cancer-associated mutation (r); and (4) the total number of nucleotide positions within the fragment of interest (NI).

The occurrence of mutations is generally considered to be a random process. Therefore, as the frequency criteria for defining a potential mutation increase, the number of false positive nucleotide positions will decrease exponentially. In some of the existing sequencing platforms, certain sequence environments are more susceptible to sequencing errors. Examples of such sequence environments include the GGC motif, homopolymers (e.g. AAAAAAAA), and simple repeats (e.g. AT AT AT AT AT).

Such sequence environments will inherently increase the number of false single nucleotide substitutions or insertions/deletions (Nakamura K et al. Nucleic Acids Res 2011;39.e90 and Minoche AE et al. Genome Biol 2011; 12. R112). In addition, repeat sequences such as homopolymers and simple repeats will introduce alignment ambiguity into the calculations and hence lead to false positive results for single nucleotide variations.

The larger the region of interest, the greater the number of false positive nucleotide positions that will be observed. If we want to find mutations in the entire genome, then the region of interest will be the entire genome and the number of nucleotides to be examined will be 3 billion. On the other hand, if we focus on exons, then the region of interest will consist of nucleotides encoding exons, the number of which is approximately 45 million.

The number of false positive nucleotide positions due to sequencing errors can be determined using the following calculations. The probability (P _Er ) that the same nucleotide substitution will occur at the same position due to sequencing errors can be calculated as:

- 16 048297

P _Sr = C(D,r)E where C(D,r) is the number of possible combinations of a sample of r elements from the total number of elements D;

r is the frequency value for determining the potential mutation;

D represents the sequencing depth; and

E represents the sequencing error rate.

C(D,r) can be calculated as:

D!

The number of nucleotide positions (NFP) false positive for mutations can be calculated as: A'f-p = where NI is the total number of nucleotide positions in the region of interest.

Fig. 6B is a plot 650 showing the expected number of nucleotide positions that were misclassified as containing a nucleotide substitution using r criteria of 4, 5, 6, and 7 as a function of the sequencing depth. In this calculation, the region of interest is the entire genome (3 billion nucleotide positions). The sequencing error rate is set to 0.3% of the sequenced nucleotides. It can be seen that the r value has a significant impact on false positives. However, it can be seen from Fig. 6A that a higher r value also reduces the number of detected mutations, at least until significantly higher sequencing depths are used.

B. Selecting the minimum occurrence (r).

It was discussed above that the number of true cancer-associated mutated sites and false-positive sites due to sequencing errors will increase with increasing sequencing depth. However, the rates of their increase will be different. Therefore, the choice of sequencing depth and r value can be used to maximize the detection of true cancer-associated mutations while keeping the number of false-positive sites low.

Fig. 7A is a diagram 700 showing the number of true cancer-associated mutated sites and false positive sites for different sequencing depths. The total number of cancer-associated mutations in the tumor tissue is assumed to be 3000 and the fractional concentration of tumor-derived DNA in plasma is assumed to be 10%. The sequencing error rate is assumed to be 0.3%. In the figure legend, TRS denotes true positive sites in which the corresponding mutation is present in the tumor tissue, FP denotes false positive sites in which an inappropriate mutation is present in the tumor tissue and nucleotide changes present in the sequencing results are due to sequencing errors.

From the 700 plot, we can see that at 110X sequencing depth, we will detect approximately 1410 true cancer-associated mutations if we use a minimum frequency of 6 (r=6) as a criterion to define a potential mutation site in plasma. Using this criterion, we will only detect approximately 20 false positive sites. If we use a minimum frequency of 7 (r=7) as a criterion to define a potential mutation, the number of cancer-associated mutations that can be detected decreases by 470 (to approximately 940). Therefore, the criterion r=6 will allow for more sensitive detection of cancer-associated mutations in plasma.

On the other hand, at 200x depth of sequencing, the number of detected true cancer-associated mutations will be approximately 2800 and 2600 if we use the minimum frequency (r) criteria of 6 and 7, respectively, to define potential mutations. Using these two r values, we obtain the numbers of false positive sites of approximately 740 and 20, respectively. Therefore, at 200x depth of sequencing, using the more stringent criterion of r=7 to define a potential mutation will significantly reduce the number of false positive sites without significantly affecting the sensitivity of detecting true cancer-associated mutations.

C. Dynamic threshold value for sequencing results for detection of potential mutations in plasma.

The depth of sequencing of each nucleotide in the region of interest will vary. If we apply a fixed threshold for the frequency of a nucleotide substitution to define a potential mutation in plasma, nucleotides that are covered by a greater number of sequence reads (i.e., sequenced to a greater depth) will be more likely to be detected.

- 17 048297 probability of erroneous detection of a variation therein due to a sequencing error, in the absence of such a change in the tumor tissue, compared to nucleotides that are sequenced at a lower depth. In one embodiment, to overcome this problem, a dynamic threshold value r is applied to different nucleotide positions, in accordance with the actual sequencing depth of a particular nucleotide position and in accordance with a desired upper limit of the probability of detecting false positive variations.

In one embodiment, the maximum allowable false positive variation rate may be fixed as 1 in 1.5x108 nucleotide positions. With this maximum allowable false positive variation rate, the total number of false positive sites that will be detected in the entire genome will be less than 20. The r value for different sequencing depths can be determined from the curves shown in Fig. 6B, and these thresholds are shown in Table 1. In other embodiments, other different maximum allowable false positive variation rates may be used, such as 1 in 3x108, 1 in 108, or 1 in ^6x107 . The corresponding total number of false positive sites will be less than 10, 30, and 50, respectively.

Table 1. Minimum number of times a nucleotide substitution is detected in plasma to determine that it is a potential mutation (r), for different sequencing depths of a particular nucleotide position. The maximum false-positive variation rate was recorded as 1 in 1.5x108 nucleotides.

Table 1

Sequencing depth of a specific nucleotide position Minimum number of times a nucleotide substitution is detected in plasma DNA sequencing results to determine that it is a potential mutation (g) <50 5 50-110 6 111-200 7 201-310 8 311-450 9 451 -620 10 621 - 800 AND

D. Targeted enrichment sequencing.

Fig. 7A shows that a higher sequencing depth can result in better sensitivity for detecting cancer-associated mutations while keeping the number of false positive sites low due to the ability to use a higher r value. For example, at a sequencing depth of 110x, 1410 true cancer-associated mutations can be detected in plasma using an r value of 6, whereas the number of true cancer-associated mutations detected would be 2600 if the sequencing depth were increased to 200x and an r value of 7 were used. These two sets of results would yield an expected number of false positive sites of approximately 20.

Although whole genome sequencing to 200x depth is currently relatively expensive, one possible way to achieve such sequencing depth is to focus on a smaller region of interest. Targeted region analysis can be accomplished, for example, but not limited to, by using DNA or RNA primers to capture genomic regions of interest via hybridization. The captured regions are then precipitated, for example, by magnetic means, and then sequenced. Such targeted capture can be accomplished, for example, using the Agilent SureSelect Targeted Enrichment System, the Roche Nimblegen Targeted Enrichment System, and the Illumina Targeted Resequencing System. Another approach is to amplify the target regions using PCR and then sequence. In one embodiment, the region of interest is the exome. In such an embodiment, all exons of DNA in plasma can be targeted captured, and the exonic region-enriched DNA from the plasma can then be sequenced.

In addition to the potential for greater sequencing depth, focusing on specific regions rather than analyzing the entire genome will significantly reduce the number of nucleotide positions in the search region and result in fewer false positive sites with the same sequencing error rate.

Fig. 7B is a plot 750 showing the predicted number of false positive sites when performing whole genome (WG) and all exons analysis. Two different r values of 5 and 6 were used for each type of analysis. At 200x depth of sequencing, if r=5 is used to detect mutations in plasma, the predicted number of false positive sites is approximately 23,000 and 230 for the whole genome and all exons, respectively.

- 18 048297

If r=6 is used to detect mutations in plasma, the predicted number of false positive sites is 750 and 7, respectively. Therefore, limiting the number of nucleotides in the region of interest can significantly reduce the number of false positive sites when analyzing for mutations in plasma DNA.

With exon-captured or even exome-captured sequencing, the number of nucleotides in the search region is reduced. Therefore, even if we allow a higher false-positive variation rate for detecting cancer-associated mutations, the absolute number of false-positive sites can be kept relatively low. Allowing a higher false-positive variation rate will allow a less stringent minimum frequency criterion (r) to be applied for detecting single-nucleotide variations in plasma. This will lead to higher sensitivity in detecting true cancer-associated mutations.

In one embodiment, we may use a maximum false positive rate of 1.5 x 10 ⁶ . With this false positive rate, the total number of false positive sites within the target exons will be only 20. The r values for various depths of read using a maximum false positive rate of 1.5 x 10 ⁶ are shown in Table 2. In other embodiments, different maximum false positive rates may be used, such as 1 in 3 x 10 ⁶ , 1 in 10 ⁶ , or 1 in 6 x 10 ⁵ . The corresponding total number of false positive sites will be less than 10, 30, and 50, respectively. In one embodiment, different weights may be assigned to different classes of mutations, as described above.

Table 2. Minimum number of times a nucleotide substitution is present in plasma to be considered a potential mutation (r), for different sequencing depths of a given nucleotide position. The maximum false-positive variation rate was found to be 1 in 1.5 x 10 ⁶ nucleotides.

Table 2

Sequencing depth of a specific nucleotide position Minimum number of times a nucleotide substitution is detected in plasma DNA sequencing results to determine that it is a potential mutation (g) <50 4 50-125 5 126-235 6 236-380 7 381 -560 8 561 -760 9

VIII. Cancer detection.

It was mentioned above that the number of sequence markers at altered loci can be used to determine in various ways a parameter that is compared to a threshold to classify the cancer level. The fractional concentration of variant reads compared to all reads at a locus or at multiple loci is another parameter that can be used for this purpose. Some examples of calculating this parameter and threshold are given below.

A. Definition of the parameter.

If the HA is homozygous at a particular locus for the first allele and the variant allele is present in a biological sample (e.g., plasma), the fractional concentration can be calculated as 2p/(p+q), where p is the number of sequence markers containing the variant allele and q is the number of sequence markers containing the first allele of the HA. This formula assumes that only one of the tumor haplotypes contains the variant, which will generally be the case. Thus, for each homozygous locus, the fractional concentration can be calculated. The fractional concentrations can be averaged. In another embodiment, the number p can include the number of sequence markers for all loci and, similarly to the number q, determine the fractional concentration. An example is now described.

A genome-wide assay was performed to identify tumor-related single nucleotide variants (SNVs) in the plasma of four patients with HCC. We sequenced tumor and buffy coat DNA at mean depths of 29.5-fold (range, 27-fold to 33-fold) and 43-fold (range, 39-fold to 46-fold) haploid genome coverage, respectively. Massively parallel sequencing (MPS) results from tumor and buffy coat DNA were compared for each of the four patients with HCC, identifying SNVs present in

- 19,048,297 tumor DNA, but not in the buffy coat DNA, using a rigorous bioinformatics algorithm. This algorithm requires that a putative NVB be present in at least a threshold number of sequenced tumor DNA fragments (i.e., in the corresponding sequenced marker) in order to be classified as a true NVB. The threshold number is determined by taking into account the sequencing depth of a particular nucleotide and the sequencing error rate, such as those described in this application.

Fig. 8 is a table 800 showing the results for 4 patients with HCC before and after treatment, including fractional concentrations of tumor-derived DNA in plasma according to embodiments of the present invention. The number of tumor-associated NVs ranged from 1334 to 3171 in the 4 cases of HCC. A list of the ratios of such NVs that were detected in the plasma before and after treatment is provided. Before treatment, 15-94% of tumor-associated NVs were detected in the plasma. After treatment, their percentage was 1.5-5.5%. Thus, the number of NVs detected does correlate with the cancer grade. Thus, it was shown that the number of NVs can be used as a parameter for classifying the cancer grade.

Fractional concentrations of tumor-derived DNA in plasma were determined by fractional counting of mutant sequences compared to the total number of sequences (i.e. mutant plus wild type). The formula for this calculation is 2p/(p+q), where 2 takes into account that only one haplotype is mutated in the tumor. The obtained fractional concentrations correlated well with those determined by genome-wide analysis of total allelic loss (GAAL) (Chan KC et al. Clin Chem 2013; 59:211–24) and decreased after surgery. Thus, the fractional concentration was shown to be a useful parameter for determining the cancer grade.

The fractional concentration obtained by analyzing the NVB can express the tumor burden. A patient with a cancer with a higher tumor burden (e.g., with a higher detected fractional concentration) will have an increased frequency of somatic mutations compared to a patient with a lower tumor burden. Thus, embodiments can also be used for prognosis. In general, patients with cancers with a higher tumor burden have a worse prognosis than patients with a lower tumor burden. Thus, the former group will have an increased risk of dying from said disease. In some embodiments, if the absolute concentration of DNA in a biological sample, such as plasma, can be determined (e.g., using real-time PCR or fluorometry), then the absolute concentration of tumor-associated genetic aberrations can be determined and used for clinical detection and/or monitoring and/or prognosis.

B. Determining the threshold.

Table 800 can be used to determine a threshold. As mentioned above, the amount of NBC and the fractional concentration determined by the NBC assay correlate with the cancer grade. The threshold can be determined on an individual basis. For example, the pre-treatment value can be used to determine the threshold. In various embodiments, the threshold can be an absolute change from the pre-treatment value. A suitable threshold can be a 50% decrease in the amount of NBC or the fractional concentration. Such a threshold would classify each case in Table 800 as a lower cancer grade. It should be noted that such a threshold can be dependent on the sequencing depth.

In one embodiment, the threshold may be applied to all samples and may or may not take into account the pre-treatment values of a given parameter. For example, a threshold of 100 NHBs may be used to classify a subject as having no cancer or having a low cancer level. This threshold of 100 NHBs is satisfied by each of the four cases in Table 800. If fractional concentration is used as the parameter, at a threshold of 1.0%, HCC1 through HCC3 would be classified as having virtually no cancer, and at a second threshold of 1.5%, HCC4 would be classified as having a low cancer level. Thus, more than one threshold may be used to obtain more than two classifications.

To illustrate other possible cutoff values, we analyzed the presence of tumor-associated OHBs in the plasma of healthy controls. In healthy subjects, multiple measurements can be made to determine the range of variation expected in a biological sample compared to the constitutive genome.

Fig. 9 is a Table 900 showing the detection of HCC-related NVs in 16 healthy controls according to embodiments of the present invention. Table 900 can be used to evaluate the specificity of the NV assay approach. The 16 healthy controls are listed in different rows. The columns list the NVs detected for individual HCC patients and show the numbers of reads at the altered loci containing the variant allele and the number of reads with the wild-type allele (i.e., the allele from CG). For example, for HCC1 in control C01, there were 40 variant reads at such altered loci but 31,261 reads of the wild-type allele. In the latter row,

- 20 048297 lonke shows the summed fractional concentration for all NBCs for patients with HCC1. Since HCC-related NBCs are specific to patients with HCC, the present HCC-related NBCs represent false positives. If the cutoff values described in this application are applied to these identified sequence variants, all such false positives will be filtered out.

The presence of a small number of these putative tumor-associated mutations in the plasma of 16 healthy controls represents the stochastic noise of the method, and is likely due to sequencing errors. The median fractional concentration estimated for this noise was 0.38%. These values represent the range for healthy subjects. Thus, the cutoff value for classification of a zero cancer level for HCC may be approximately 0.5%, since the highest fractional concentration was 0.43%. Thus, if all cancer cells were removed from a patient with HCC, these low fractional concentrations would be expected to be present.

Returning again to Table 800, if 0.5% is used as the threshold for a cancer level of zero, then the results for plasma from patients with HCC1 and HCC3 after treatment would be classified as zero based on the NBC analysis. HCC2 would be classified as one level higher than zero. HCC4 would also be classified as one level higher than zero, or even higher, but still relatively low compared to the pretreatment samples.

In one embodiment, where the parameter is the number of altered loci, the threshold may be zero (i.e., one altered locus may indicate a non-zero cancer level). However, for many parameters (e.g., depth), the threshold will be higher, such as an absolute value of 5 or 10. In one embodiment, where the individual is followed after treatment, the threshold may be a certain percentage of DATs (directly detectable tumors) detected in the sample. If the threshold for the number of variant reads needed at a locus is high enough, the presence of only one altered locus may indicate a non-zero cancer level.

Thus, quantitative analysis of variations (e.g., single nucleotide variations) in DNA from a biological sample (e.g., plasma) can be used to diagnose, monitor, and prognosticate cancer. The number of single nucleotide variations detected in the plasma of a test subject can be compared with that in a group of healthy subjects to detect cancer. Detection of single nucleotide variations in the plasma of healthy subjects can be due to sequencing errors, non-clonal mutations in blood cells, and other organs. It has been shown that cells from normal healthy subjects can carry a small number of mutations (Conrad DF et al. Nat Genet 2011;43:712-4), as shown in Table 900. Thus, the total number of single nucleotide variations detected in the plasma of a group of apparently healthy subjects can be used as a reference range to determine whether an abnormally high number of single nucleotide variations corresponding to a non-zero level of cancer is present in the plasma of a test patient.

Healthy subjects used to establish a reference range can be matched by age and gender to match the subjects being studied. An earlier study showed that the number of mutations in somatic cells increases with age (Cheung NK et al., JAMA 2012; 307:1062-71). Thus, the accumulation of clones of cells as we age is normal, even if they are relatively benign most of the time or if they take a very long time to become clinically significant. In one embodiment, reference levels can be obtained for different groups of subjects, such as those with different ages, genders, ethnicities, and other parameters (e.g., smoking status, hepatitis, alcoholism, drug history).

The reference range may vary depending on the threshold used (i.e. the presence of a required number of altered sequence markers at the locus), as well as the expected level of false positive variations and other variables (e.g. age). Thus, the reference range can be defined for a specific set of one or more criteria, and the same criteria will be used to determine the parameter for the sample. The specified parameter can then be compared with the reference range, since both were determined using the same criteria.

It was mentioned above that in embodiments, several threshold values can be used to determine the cancer level. For example, using the first level, it is possible to determine the absence of signs of cancer for parameters below the threshold and at least the first level of cancer, which can represent a pre-neoplastic level. Other levels can correspond to different stages of cancer.

C. Dependence of experimental variables.

Sequencing depth may be important for determining the minimum detection threshold for a minor (e.g., tumor) genome. For example, if a sequencing depth of 10 haploid genomes is used, the minimum concentration of tumor DNA that can be detected is

- 21 048297 even with the sequencing technology without any errors is ¹ / ₅ , i.e. 20%. On the other hand, if a sequencing depth of 100 haploid genomes is used, the minimum concentration will drop to 2%. This analysis applies to the case where only one mutated locus is analyzed. However, if a larger number of mutated loci are analyzed, the minimum tumor DNA concentration may be lower and is determined using the binomial probability function. For example, if the sequencing depth is 10-fold and the fractional tumor DNA concentration is 20%, the probability of detecting a mutation is 10%. However, if there are 10 mutations, the probability of detecting at least one mutation will be 1-(1-10%) ¹⁰ = 65%.

There are several implications of increasing the sequencing depth. The greater the sequencing depth, the more sequencing errors will be observed, see Figs. 4 and 5. However, with greater sequencing depth, it will be easier to distinguish sequencing errors from mutations due to clonal expansion of a subpopulation of cells (e.g., cancer cells), since sequencing errors will occur randomly in the genome and mutations will occur at the same position for a given cell population.

The greater the sequencing depth, the more mutations will be detected in healthy cells. However, if there is no clonal expansion of these healthy cells and their mutation profiles are different, then mutations in these healthy cells can be distinguished from mutations by their frequencies in plasma (e.g., by applying a threshold N for the required number of sequence reads in which a mutation is detected, e.g., N should be 2, 3, 4, 5, or more).

It was mentioned above that the threshold may depend on the number of mutations in healthy cells that will clonally expand and therefore may not be distinguishable from other mechanisms. Such variability, which may be expected, may be obtained by analyzing healthy subjects. Since clonal expansion occurs over time, the variability observed in healthy subjects may be affected by the patient's age, and thus the threshold may be age dependent. D. Combinations of Targeted Approaches.

In some embodiments, random sequencing can be used in combination with targeted approaches. For example, a plasma sample can be randomly sequenced upon admission to a patient with cancer. Copy number and NVD abnormalities in the DNA sequencing results from the plasma can be analyzed. Regions that show abnormalities (e.g., amplification/deletion or high NVD density) can be targeted for serial monitoring. Monitoring can be performed over a period of time or performed immediately after random sequencing, essentially as a single procedure. For targeted analysis, liquid-phase hybridization-based capture approaches can be successfully used to enrich plasma DNA for noninvasive prenatal diagnosis (Liao GJ et al. Clin Chem 2011;57:92-101). Such techniques are mentioned above. Thus, targeted and random approaches can be used in combination for cancer detection and monitoring.

Thus, targeted sequencing can be performed on loci where potential mutations have been identified using the untargeted whole genome approach mentioned above. Such targeted sequencing can be performed using solution hybridization or solid phase hybridization techniques (e.g., using Agilent SureSelect, NimbleGen Sequence Capture, or the Illumina Targeted Resequencing System) followed by massively parallel sequencing. Another approach is to perform amplification (e.g., PCR-based systems) for targeted sequencing (Forshew T et al. Sci Transl Med 2012; 4: 135ra68).

IX. Fractional concentration.

The fractional concentration of tumor DNA can be used to determine a cutoff value for the amount of variation required at a locus before a mutation is detected at the locus. For example, if the fractional concentration is known to be relatively high, a high cutoff value can be used to filter out more false positives, since it is known that a relatively high number of variant reads must be present for a true NVD. On the other hand, if the fractional concentration is found to be low, a lower cutoff value may be needed to ensure that some NVDs are not missed. In this case, the fractional concentration would be determined using a method different from the NVD assay, where it is used as a parameter.

Various techniques can be used to determine the fractional concentration, some of which are described in this application. These techniques can be used to determine the fractional concentration of tumor-derived DNA in a mixture, such as in a biopsy containing a mixture of tumor cells and non-malignant cells, or in a plasma sample from a patient with cancer containing DNA released from tumor cells and DNA released from non-malignant cells.

- 22 048297

A. GAAL.

Genome-wide analysis of total allele loss (GAAL) allows analysis of loci that have lost heterozygosity (Chan KC et al. Clin Chem 2013; 59:211-24). For a heterozygous site of the housekeeping genome (GH) in a tumor, a locus is often found in which one of the alleles has been deleted. Thus, the reads for such a locus will reveal more of one allele than the other, with the difference being proportional to the fractional concentration of tumor DNA in the sample. An example of such a calculation is given below.

DNA isolated from buffy coats and tumor tissues of patients with HCC was genotyped using the Affymetrix Genome-Wide Human SNP Array 6.0 system. Results obtained for the microarray were processed using Affymetrix Genotyping Console version 4.1. Genotyping analysis and determination of single nucleotide polymorphisms (SNPs) were performed using the Birdseed algorithm version 2. Genotyping results for buffy coats and tumor tissues were used to determine regions of loss of heterozygosity (LOH) and to perform copy number analysis. Copy number analysis was performed using the default parameters of the Affymetrix Genotyping Console, with a minimum genomic fragment size of 100 bp and a minimum number of genetic markers within a fragment equal to 5.

Regions with LOH were detected as regions with 1 copy in tumor tissue and 2 copies in the buffy coat, and SNPs in these regions were heterozygous in the buffy coat but homozygous in tumor tissue. For a genomic region in which LOH was detected in tumor tissue, SNP alleles that were present in the buffy coat but absent or at a lower intensity in tumor tissues were considered to be alleles on the deleted fragment of the chromosome fragment. Alleles that were present in both the buffy coat and tumor tissue were considered to be derived from the undeleted fragment of the chromosome fragment. For all chromosome fragments with one copy loss in the tumor, the total number of read sequences with and without allelic loss was calculated. The difference between these two values was used to calculate the fractional concentration of tumor-derived DNA (F _GAAL ) in the sample using the following equation:

_ N6es del - Ыdel

Fgaal N6es div where N _div is the total number of reads without allele loss and N _div is the total number of reads with allele loss.

B. Assessment using genomic representation.

The problem with the GAAL method is that it defines specific loci (i.e., loci where loss of heterozygosity (LOH) regions are detected) and only uses read sequences aligned to these loci. This requirement may lead to additional steps and thus higher costs. An implementation is described here that uses only copy number, e.g., read density.

Chromosomal aberrations, such as amplifications and deletions, are frequently observed in the genomes of cancer cells. Chromosomal aberrations observed in cancer tissues typically occur at subchromosomal regions, and such aberrations can be shorter than 1 megabase. In addition, cancer-associated chromosomal aberrations are heterogeneous among patients, and therefore different fragments may be affected in different patients. Also, tens, hundreds, or even thousands of copy number variations are often found in the genome of a cancer cell. Each of these factors makes it difficult to determine the concentration of tumor DNA.

Embodiments include analysis of quantitative changes resulting from tumor-associated chromosomal aberrations. In one embodiment, DNA samples comprising DNA derived from cancer cells and normal cells are sequenced using massively parallel sequencing, such as using the Illumina HiSeq2000 sequencing platform. The resulting DNA may be cell-free DNA from plasma or another suitable biological sample.

Chromosome regions that are amplified in tumor tissues are more likely to be sequenced, and regions that are deleted in tumor tissues are less likely to be sequenced. As a result, the density of sequence reads that align to amplified regions will be increased, and the density of sequence reads that align to deleted regions will be decreased. The degree of variation is proportional to the fractional concentration of tumor-derived DNA in the DNA mixture. The higher the proportion of tumor-derived DNA, the greater the variation that will be caused by chromosomal aberrations.

1. Evaluation of a sample with high tumor concentration.

DNA was isolated from tumor tissues of four patients with hepatocellular carcinoma. DNA was fragmented using the Covaria sonic DNA disruption system and sequenced using

- 23 048297 described Illumina HiSeq2000 platform (Chan KC et al. Clin Chem 2013; 59:211–24). Read sequences were aligned to the human reference genome (hg18). The genome was then divided into 1-Mb fragments (regions) and the density of read sequences was calculated for each fragment after correction for GC bias as described (Chen EZ et al. PLoS One. 2011; 6:e21791).

After alignment of the read sequences with the reference genome, the read density for the various regions can be calculated. In one embodiment, the read density is a ratio defined as the number of read sequences mapped to a particular fragment (e.g., a 1-megabase region) divided by the total number of read sequences that can be aligned to the reference genome (e.g., to a unique position in the reference genome). Fragments that overlap with chromosomal regions amplified in tumor tissue are expected to have higher read densities than fragments without such overlaps. On the other hand, fragments that overlap with chromosomal regions containing a deletion are expected to have lower read densities than fragments without such overlaps. The magnitude of the difference in the densities of read sequences between regions with and without chromosomal aberrations is mainly influenced by the proportion of tumor-derived DNA in the sample and the degree of amplification/deletion in tumor cells.

Various statistical models can be used to identify fragments with read densities corresponding to different types of chromosomal aberrations. In one embodiment, a normal mixture model can be used (McLachlan G and Peel D. Multvariate normal mixtures. In Finite mixture models 2004: pp. 81-116. John Wiley & Sons Press). Other statistical models can also be used, such as a binomial mixture model and a Poisson regression model (McLachlan G and Peel D. Mixtures with non-normal components, Finite mixture models 2004: pp. 135-174. John Wiley & Sons Press).

The read density of a fragment can be normalized by applying the read density of the same fragment determined by buffy coat DNA sequencing. The read densities of different fragments can be affected by the sequence context of a particular chromosomal region, and thus normalization allows for more precise identification of regions with aberrations. For example, the mappability (which refers to the probability of aligning a sequence to its original position) of different chromosomal regions can differ. In addition, copy number polymorphism (i.e., copy number variations) will also affect the read densities of fragments. Therefore, normalization to buffy coat DNA can potentially minimize the amount of variation due to differences in sequence context between different chromosomal regions.

Fig. 10A shows a graph 1000 of the distribution of read densities of a tumor sample from a patient with HCC according to embodiments of the present invention. The tumor tissue was obtained after surgical resection from a patient with HCC. The x-axis represents the log ₂ ratio (R) of the read density of the tumor tissue to that of the patient's buffy coat. The y-axis represents the number of fragments.

The peaks can be fitted to a distribution curve to show regions with deletion, amplification, and no chromosomal aberrations using a mixture model of normal distributions. In one embodiment, the number of peaks can be determined using the Akaike information criterion (AIC) for all different likelihood values. The central peak with log ₂ R=0 (i.e., R=1) represents regions without any chromosomal aberrations. The left peak (relative to the central one) represents regions with a loss of one copy. The right peak (relative to the central one) represents regions with a gain of one copy.

The fractional concentration of tumor DNA can be reflected by the distance between the peaks representing the amplified and deleted regions. The greater the said distance, the higher the fractional concentration of tumor DNA in the sample. The fractional concentration of tumor DNA in the sample can be determined using the genomic representation approach denoted as FGR, using the following equation: F _gr = R right - _{R left} , where R _right is the R value for the right peak and R _left is the R value for the left peak. The largest difference will be represented by 1, which corresponds to 100%. The fractional concentration of tumor DNA in the tumor sample obtained from a patient with HCC was estimated to be 66%, where the R _right and R _left values were 1.376 and 0.712, respectively.

To confirm the obtained result, the genome-wide analysis of total allele loss (GAAL) method was also used to independently determine the fractional concentration of tumor DNA ratio (Chan KC et al. Clin Chem 2013; 59:211-24). Table 3 shows the fractional concentrations of tumor-derived DNA in tumor tissues of four patients with HCC determined using the genomic representation (FGR) and GAAL (FGAAL) approaches. The values determined by these two different approaches are in good agreement with each other.

- 24 048297

Table 3

Showing fractional concentration determined using GAAL and genomic representation (GR)

GKK tumor Fgaal Fgr 1 60.0% 66.5% 2 60.0% 61.4% 3 58.0% 58.9% 4 45.7% 42.2%

2. Evaluation of a sample with low tumor concentration.

The above analysis showed that our genomic representation method can be used to measure the fractional concentration of tumor DNA when more than 50% of the sample DNA is of tumor origin, i.e., when tumor DNA represents the dominant fraction. In a previous analysis, we showed that this method can also be applied to samples in which tumor DNA represents a minor fraction (i.e., less than 50%). Samples that may contain a minor fraction of tumor DNA include, but are not limited to, blood, plasma, serum, urine, pleural fluid, cerebrospinal fluid, tears, saliva, ascites fluid, and stool from patients with cancer. In some samples, the fractional concentration of tumor DNA may be 49%, 40%, 30%, 20%, 10%, 5%, 2%, 1%, 0.5%, 0.1%, or less.

For such samples, the peaks in the density of the variant read sequence representing the regions with amplification and deletion may not be as obvious as for the samples containing a relatively high concentration of tumor-derived DNA described above. In one embodiment, regions of chromosomal aberrations in cancer cells can be detected by comparison with reference samples that are known to be free of cancer DNA. For example, plasma from subjects without cancer can be used as a reference to determine the normative range of the variant read sequence densities for chromosomal regions. The read sequence density of the subject of interest can be compared with its value for the reference group. In one embodiment, the mean value and standard deviation (SD) of the read sequence density can be determined. For each fragment, the read sequence density of the subject of interest is compared with the mean value in the reference group to determine a z-score using the following formula:

(gRtsct - GRaTa.n) z - score =-----------------СОref, where GR _test is the density of the read sequence of a patient with cancer;

G Ra tal represents the average value of the density of the read sequence of the reference subjects and SD _etal represents the SD of the densities of the read sequences of the reference subjects.

Regions with a z-score <-3 indicate significant underrepresentation of read density for a particular fragment in a patient with cancer, suggesting the presence of a deletion in tumor tissue. Regions with a z-score >3 indicate significant overrepresentation of read density for a particular fragment in a patient with cancer, suggesting the presence of an amplification in tumor tissue.

A distribution of z-scores can then be constructed for all fragments to identify regions with different amounts of copy gains and losses, such as deletions of 1 or 2 copies of a chromosome; and amplifications with gains of 1, 2, 3, and 4 extra copies of a chromosome. In some cases, more than one chromosome or more than one region of a chromosome may be involved.

Fig. 10B shows a plot 1050 of the z-score distribution for all fragments in the plasma of a patient with HCC according to embodiments of the present invention. Peaks (from left to right) representing 1 copy loss, no copy change, 1 copy gain, and 2 copy gain were fitted to the z-score distribution. Regions with different types of chromosomal aberrations can then be identified, for example, by applying the normal mixture model described above.

The fractional concentration of cancer DNA in the sample (F) can then be inferred from the read densities of fragments in which one copy gain or one copy loss is detected. The fractional concentration determined for a particular fragment can be calculated using the equation:

G =

Il;|gRt9<:t — GRsTan)|x2)

SKateal

X 100%

-25 048297

This equation can also be represented as:

F = (|z ^_ index x С0зтал|)/((GRsran))) x 2 and can also be rewritten as:

F = Iz - indicator! X KB x 2 where KB is the coefficient of variation of the density measurement of the read sequence of reference subjects; and _{kv =} COetaLSyetal

In one embodiment, the results obtained for the fragments are combined. For example, the z-scores of the fragments in which the acquisition of 1 copy was detected can be averaged, or the F values obtained for them can be averaged. In another embodiment, the z-score value used to derive F is determined using a statistical model and is represented as the peaks shown in Fig. 10B and Fig. 11. For example, the z-score for the right peak can be used to determine the fractional concentration of regions in which the acquisition of 1 copy was detected.

In another embodiment, all fragments with a z-score <-3 and a z-score >3 can be assigned to regions with one copy loss and one copy gain, respectively, since these two types of chromosomal aberrations are the most common. This approximation is most useful when the number of fragments with chromosomal aberrations is relatively small and the fit to a normal distribution cannot be exact.

Fig. 11 shows a graph 1100 of the z-score distribution for plasma from a patient with HCC according to embodiments of the present invention. Although the number of fragments overlapping with chromosomal aberrations is relatively small, all fragments with a z-score <-3 and a z-score >3 were fitted to normal distributions of one copy loss and one copy gain, respectively.

Fractional concentrations of tumor-derived DNA in plasma of four patients with HCC were determined using the GAAL assay and the GR-based approach. The results are presented in Table 4. It is evident that the obtained fractional representation correlates well between the GAAL assay and the GR assay.

Table 4

Fractional plasma concentration of tumor DNA obtained by ________chromosomal aberration analysis________

Fractional concentration of tumor DNA in plasma Samples GAAL Analysis GR Analysis Case 11 4.3% 4.5% Case 13 5% 5.5% Case 23 52% 62% Case 27 7.6% 6.1%

C. Method for determining fractional concentration.

Fig. 12 is a flow chart of a method 1200 for determining a fractional concentration of tumor DNA in a biological sample containing extracellular DNA according to embodiments of the present invention. The method 1200 can be implemented using various embodiments, including the embodiments described above.

In block 1210, one or more sequence markers are obtained for each of a plurality of DNA fragments in the biological sample. Block 1210 may be performed as described in this application for other methods. For example, one end of a DNA fragment may be sequenced from a plasma sample. In another embodiment, both ends of the DNA fragment may be sequenced, which allows the length of the fragment to be estimated.

In block 1220, the positions of the sequence markers in the genome are determined. The positions in the genome can be determined, for example, as described in this application, by aligning the sequence markers with a reference genome. If both ends of the fragment are sequenced, then the paired markers can be aligned as a pair with the distance between the two markers being less than a certain distance, for example, 500 or 1000 bases.

In block 1230, for each of the plurality of genomic regions, a corresponding number of DNA fragments belonging to the given genomic region are determined using sequence markers located in the genome within the given genomic region. The genomic regions may be non-overlapping fragments of equal length in a reference genome. In one embodiment, the number of markers that are aligned with a fragment may be counted. Thus, a corresponding number of markers may be aligned with each fragment. A histogram may be constructed illustrating the frequency of alignment with fragments of a certain number of markers.

-26048297

Method 1200 can be performed for regions of the genome, each of which has the same length (for example, for fragments of 1 megabase), and the regions do not overlap. In other embodiments, different lengths can be used, which can be taken into account, and the regions can overlap.

In block 1240, the corresponding quantity is normalized to obtain the corresponding density. In one embodiment, normalizing the corresponding quantity to obtain the corresponding density includes using the same total number of aligned reference markers to determine the corresponding density and the reference density. In another embodiment, the corresponding quantity can be divided by the total number of aligned reference markers.

In block 1250, the corresponding density is compared with a reference density to detect a loss of 1 copy or a gain of 1 copy in a given region of the genome. In one embodiment, the difference between the corresponding density and the reference density is calculated (e.g., as part of determining a z-score) and compared with a threshold value. In various embodiments, the reference density can be obtained for a sample of healthy cells (e.g., for a buffy coat) or for corresponding quantities of the same cells (e.g., by obtaining a median or average value, assuming that most regions do not have copy losses or gains).

In block 1260, the first density is calculated from one or more corresponding densities in which a loss of 1 copy is detected, or from one or more corresponding densities in which a gain of 1 copy is detected. The first density may correspond to only one region of the genome, or it may be determined from the densities of several regions of the genome. For example, the first density may be calculated from the corresponding densities with a loss of 1 copy. The corresponding densities make it possible to measure the magnitude of the density difference that resulted from the deletion of the specified region in the tumor, allowing one to obtain the tumor concentration. Similarly, if the first density is related to the corresponding densities with a gain of 1 copy, then the magnitude of the density difference that resulted from the duplication of the given region in the tumor may be obtained. The sections above describe various examples of how densities of several regions may be used to determine an average density, which may be used to determine the first density.

In block 1270, the fractional concentration is calculated by comparing the first density with another density to obtain a difference. The resulting difference is normalized to a reference density, which can be done in block 1270. For example, the difference can be normalized to the reference density by dividing the difference by the reference density. In another embodiment, the difference can be normalized in the previously described blocks.

In one embodiment, the other density is a reference density, such as described above in section 2. Thus, calculating the fractional concentration may include multiplying the difference by two. In another embodiment, the other density is a second density calculated from the corresponding densities in which a loss of 1 copy was detected (where the first density is calculated using the corresponding densities in which a gain of 1 copy was detected), such as described above in section 1. In this case, the normalized difference can be determined by calculating a first ratio (e.g., R _right ) of the first density to the reference density and calculating a second ratio (R _left ) of the second density to the reference density, where the difference is between the first ratio and the second ratio. As described above, it is possible to determine the region of the genome in which a loss of 1 copy or a gain of 1 copy is detected by fitting the peaks to a distribution curve of a histogram of the corresponding densities.

Briefly, embodiments allow analyzing the genomic representation of DNA in plasma in various chromosomal regions to simultaneously determine whether a chromosomal region in tumor tissue is amplified or deleted, and, if said region is amplified or deleted, using this genomic representation to calculate the fractional concentration of tumor-derived DNA. In some embodiments, a mixture model of normal distributions is used to analyze the overall distribution of the genomic representation of various fragments to determine the genomic representation associated with various types of abnormalities, namely, gains of 1, 2, 3, or 4 copies and losses of 1 or 2 copies.

The described embodiments have several advantages over other methods, such as the genome-wide analysis of total allele loss (GAAL) approach (US patent application 13/308,473; Chan KC et al. Clin Chem 2013; 59:211-24) and tumor-associated single nucleotide mutation analysis (Forshew T et al. Sci Transl Med. 2012; 4:136ra68). All reads mapped to regions with chromosomal aberrations can be used to determine the read density of said region and are therefore informative regarding the fractional concentration of tumor DNA. On the other hand, in the GAAL analysis, only reads covering individual nucleotides that are heterozygous in an individual and located within the chromosomal region with a gain or loss of a chromosome will be informative. Similarly, in cancer-associated mutation analysis, only sequence reads covering mutations will be useful for calculating tumor DNA concentration. Therefore, embodiments may provide more cost-effective use of the results.

- 27,048,297 sequencing reads, as relatively fewer sequencing reads may be required compared to other approaches to achieve the same degree of confidence in estimating the fractional concentration of tumor-derived DNA.

X. Alternative methods.

In addition to using the number of times a particular mutation is observed in a sequence marker as a criterion for determining whether a locus contains a true mutation (which adjusts the positive predictive value), other techniques can be used instead of or in addition to the threshold to obtain a higher predictive value for identifying a cancer mutation. For example, bioinformatic filters of varying stringency can be used when processing sequencing results, such as by taking into account the quality score of the sequenced nucleotide. In one embodiment, DNA sequencers and sequencing chemistries with different sequencing error profiles can be used. Sequencers and chemistries with lower sequencing error rates will yield higher positive predictive values. It is also possible to resequence the same DNA fragment to increase the confidence in the sequencing. One possible strategy is the circular consensus sequencing strategy from Pacific Biosciences.

In another embodiment, information on the size of the sequenced fragments can be used to interpret the results obtained. Since tumor-derived DNA is shorter in plasma than non-tumor-derived DNA (see U.S. Patent Application No. 13/308,473), the positive predictive value for a shorter DNA fragment in plasma containing a potential tumor-derived mutation will be higher than that for a longer DNA fragment in plasma. Size data can be readily obtained by paired-end sequencing of DNA in plasma. Alternatively, DNA sequencers that read long molecules can be used, thereby allowing the full length of the DNA fragment in plasma to be obtained. It is also possible to size fractionate the DNA in the plasma sample prior to DNA sequencing. Examples of methods that can be used for size fractionation include gel electrophoresis, the use of a microfluidic approach (e.g., the LabChip XT Caliper system), and size exclusion spin columns.

In another embodiment, the fractional concentration of tumor-associated mutations in plasma in a patient with non-hematological cancer will be expected to be higher when focusing on shorter DNA fragments in plasma. In one embodiment, the fractional concentration of tumor-associated mutations in plasma DNA fragments from two or more different size distributions can be compared. A patient with non-hematological cancer will have higher fractional concentrations of tumor-associated mutations in shorter fragments compared to longer fragments.

In some embodiments, the sequencing results for two or more aliquots of the same blood sample, or for two or more blood samples obtained at the same time or at different time points, may be combined. The positive predictive value of potential tumor-associated mutations observed in more than one aliquot or in more than one sample will be higher. The positive predictive value will increase with the number of samples in which such a mutation is detected. Potential mutations that are present in plasma samples taken at different time points may be considered potential mutations.

XI. Examples.

The following are examples of methods and results, but they should not be considered as limiting the embodiments of the present invention.

A. Materials and methods.

Samples were collected from patients with hepatocellular carcinoma (HCC), chronic hepatitis B carriers, and a patient with simultaneous breast and ovarian cancer. All HCC patients from the Barcelona Clinic had liver cancer at stage A1 disease. Peripheral blood samples were collected from all participants in tubes containing EDTA. Tumor tissues from HCC patients were obtained during surgical resection of the cancer.

Peripheral blood samples were centrifuged at 1600 g for 10 min at 4°C. The resulting plasma was centrifuged again at 16,000 g for 10 min at 4°C and then stored at 80°C. Cell-free DNA molecules were isolated from 4.8 ml of plasma according to the protocol for DNA isolation from blood and biological fluids from the QIAamp DSP DNA Blood Mini kit (Qiagen). Plasma DNA for each case was concentrated using a SpeedVac concentrator (Savant DNA120; Thermo Scientific) to a final volume of 40 μl for subsequent DNA sequencing library preparation.

Genomic DNA was isolated from buffy coat samples of patients according to the protocol for DNA isolation from blood and biological fluid from the QIAamp DSP DNA Blood Mini kit. DNA was isolated from tumor tissues using the QIAamp DNA Mini kit (Qiagen).

Sequencing libraries of genomic DNA samples were constructed using the Paired kit

- 28 048297

End Sample Preparation Kit (Illumina), according to the manufacturer's instructions. Briefly, 1-5 micrograms of genomic DNA were first sheared into 200 bp fragments using a Covaris S220 Focused-ultrasonicator. Subsequently, the ends of the DNA molecules were filled in using T4 DNA polymerase and the Klenow fragment of DNA polymerase; then, T4 polynucleotide kinase was used to phosphorylate the 5' ends. A 3' overhang was generated using a Klenow fragment lacking 3'-5' exonuclease activity. Illumina oligonucleotide adapters were ligated to the sticky ends. The adapter-ligated DNA was amplified using 12 cycles of PCR. Because the DNA molecules in plasma were short fragments and the amounts of total DNA in plasma samples were relatively small, we skipped the fragmentation steps and used 15 cycles of PCR to construct DNA libraries from plasma samples.

An Agilent 2100 Bioanalyzer (Agilent Technologies) was used to check the quality and size of the adapter-ligated DNA libraries. DNA libraries were then analyzed using the KAPA library quantification kit (Kara Biosystems) following the manufacturer's instructions. Library DNA was diluted and hybridized to paired-end sequencing flow cells. DNA clusters were generated using the cBot Cluster Generation System (Illumina) with the TruSeq PE Cluster Generation Kit Version 2 (Illumina), followed by 512 or 762 cycles of sequencing on a HiSeq 2000 System (Illumina) with the TruSeq SBS Kit Version 2 (Illumina).

Paired-end sequencing results were analyzed with the Short Oligonucleotide Alignment Program 2 (SOAP2) in paired-end mode. For each paired-end sequence read, 50 bp or 75 bp from each end were aligned to the human reference genome without repeat masking (hg18). A mismatch of 2 nucleotides was allowed in each end alignment. The genomic coordinates of these potential 2-end alignments were then analyzed to determine whether any combination of the 2 ends could align to the same chromosome in the correct orientation with an insert size less than or equal to 600 bp and mapping to the same position in the human reference genome. Duplicate sequence reads were defined as paired-end sequence reads in which the start and end positions in the human genome of the inserted DNA molecule were identical; Duplicate read sequences were removed as described previously (Lo et al. Sci Transl Med 2010; 2:61ra91).

In some embodiments, pairs of tumor and housekeeping DNA samples were sequenced to identify tumor-associated single nucleotide variants (SNVs). In some embodiments, we focused on SNVs occurring at homozygous sites in the housekeeping DNA (in this example, the DNA was obtained from the buffy coat). In principle, any nucleotide variant detected in the tumor tissue sequencing results but not in the housekeeping DNA could represent a potential mutation (i.e., a SNV). Due to sequencing errors (0.1% - 0.3% of sequenced nucleotides), however, millions of false positives would be detected in the genome if a single occurrence of any nucleotide substitution in the tumor tissue sequencing results was considered a tumor-associated SNV. One way to reduce the number of false positives would be to establish a criterion for the repeated occurrence of the same nucleotide substitution in sequencing results for tumor tissue before calling a potential mutation a tumor-associated NVG.

Since the occurrence of sequencing errors is a stochastic process, the number of false positives due to sequencing errors will decrease exponentially with an increase in the required number of NV detections to consider a potential mutation as a tumor-associated NV. On the other hand, the number of false positives will increase with an increase in the sequencing depth. Such relationships can be predicted using Poisson and binomial distribution functions. In embodiments, a dynamic threshold of occurrence can be determined to consider an observed NV as tumor-associated. In embodiments, the actual coverage of a particular nucleotide in the sequencing results for the tumor, the sequencing error rate, the maximum acceptable level of false positive variations, and the desired sensitivity for detecting the mutation can be taken into account.

In some examples, we set very stringent criteria to reduce the number of false positives. For example, we may require that the mutation be completely absent from the housekeeping DNA sequencing results and that the sequencing depth at a particular nucleotide position be 20×. In some embodiments, we set a threshold to achieve a false positive rate of less than 10 ^-7 . In some examples, we also removed SNPs that were detected in centromeric, telomeric, and low-complexity regions to minimize false positives due to alignment artifacts. In addition, we also removed putative SNPs that mapped to known SNPs present in the dbSNP build 135 database.

- 29 048297

B. Before and after resection.

Fig. 13A shows a table 1300 of the analysis of mutations in the plasma of a patient with ovarian cancer and breast cancer at the time of diagnosis according to embodiments of the present invention. Here, we have shown an example of a patient with bilateral ovarian cancer and breast cancer. The results of DNA sequencing from the plasma were compared with the results of sequencing the patient's constitutive DNA (from the buffy coat). Single nucleotide changes that were present in the plasma but not in the constitutive DNA were considered potential mutations. Ovarian cancer samples were taken from the right and left sides of the patient from two locations on each side, i.e., a total of four tumor samples were obtained. Mutations detected in all four ovarian tumor tissues at four different locations were considered tumor mutations.

Plasma DNA sequencing detected more than 3.6 million single nucleotide changes at least once. Of these changes, only 2,064 were also detected in tumor tissues, and the positive predictive value was 0.06%. When applying the criterion of at least two detections in plasma, the number of potential mutations significantly decreased by 99.5% to 18,885. The number of tumor mutations decreased by only 3% to 2,003, and the positive predictive value increased to 11%.

When using the criterion of at least fivefold detection in plasma, only 2572 potential mutations were detected and among them, 1814 were mutations detected in all tumor tissues, thus the positive predictive value was 71%. Other criteria of the number of detections (e.g., 2, 3, 4, 6, 7, 8, 9, 10, etc.) can be used to detect potential mutations depending on the required sensitivity and positive predictive value. The higher the frequency value used as the criterion, the higher the positive predictive value will be and the lower the sensitivity.

Fig. 13B shows Table 1350 of the analysis of mutations in the plasma of a patient with bilateral ovarian cancer and breast cancer after tumor resection according to embodiments of the present invention. A tumor was surgically resected from the patient. A blood sample was taken one day after the resection of the ovarian and breast cancer tumors. The DNA in the plasma was then sequenced. In this example, only mutations from the ovarian cancer were analyzed. More than 3 million potential mutations were detected at least once in the plasma sample. However, when using the criterion of at least five detections, the number of potential mutations was reduced to 238. A significant reduction was observed when compared to the number of potential mutations in the sample taken at diagnosis when using the same criterion of detecting five mutations.

In one embodiment, the number of single nucleotide changes detected in plasma can be used as a parameter for detecting cancer, monitoring a patient with cancer, and making a prognosis for him. The occurrence of different amounts can be used as a criterion to achieve a desired sensitivity and specificity. A patient with a higher tumor burden, and therefore a worse prognosis, will be expected to have a higher mutational load observed in the plasma.

For such analysis, it is possible to establish a profile of the mutational load for different types of cancer. For monitoring purposes, it will be seen that the mutational load in the plasma of a patient who responds to treatment will decrease. If the tumor returns, for example, during a relapse, then an increase in the mutational load will be expected. Such monitoring will allow one to control the effectiveness of the chosen method of treatment of the patient and to detect the emergence of resistance to a particular treatment.

By analyzing specific mutations that can be observed in plasma DNA sequencing results, it is also possible to identify targets that will predict sensitivity (e.g., mutations in the epidermal growth factor receptor gene and response to tyrosine kinase inhibitor treatment) and resistance to a specific targeted therapy (e.g., KRAS mutations in colorectal cancer and resistance to panitumumab and cetuximab treatment), and can guide the planning of treatment regimens.

The example above was for bilateral ovarian cancer. One could also perform a similar analysis of breast cancer mutations and then track the mutations of both cancer types in the plasma. One could also use a similar strategy to track the mutations of the primary cancer and its metastasis or metastases.

Embodiments will be useful for screening for cancer in apparently healthy subjects or in subjects with certain risk factors (e.g., smoking status, virus carrier status (in subjects who are hepatitis B virus carriers, subjects infected with human papillomavirus)). The mutational load that can be observed in the plasma of such subjects will allow one to identify the risk that the subject will develop symptomatic cancer over a certain period of time. Thus, subjects with a large mutational load in their plasma would be expected to have an increased risk compared to those with a lower mutational load. Moreover, the time profile of such mutational load in plasma would also be a powerful indicator of risk. For example, if a subject is tested for mutational load in plasma once a year and if the mutational loads gradually increase

- 30 048297 then the subject should be referred for additional cancer screening, such as chest x-ray, ultrasound, computed tomography, magnetic resonance imaging, or positron emission tomography.

C. Dynamic thresholds for mutation detection in plasma DNA sequencing results.

Four patients with hepatocellular carcinoma (HCC) and one patient with ovarian and breast cancer were recruited for this study. For the latter patient, we focused on the analysis of ovarian cancer. Blood samples were collected from each patient before and after surgical resection of the tumors. Excised tumor tissues were also collected. DNA was isolated from tumor tissue, and white blood cells from the preoperative blood sample and preoperative and postoperative plasma samples were sequenced using the HiSeq2000 sequencing system (Illumina). Sequencing results were aligned to the human reference genome sequence (hg18) using the Short Oligonucleotide Alignment Package 2 (SOAP2) (Li R et al. Bioinformatics 2009; 25: 1966–1967). White blood cell DNA sequences were considered as constitutive DNA sequences for each study subject.

In this example, tumor-associated OHMs were first identified using plasma and CG DNA sequencing results, ignoring tumor tissues. The plasma results were then compared with tumor tissue sequencing results (as a gold standard) to establish the validity of the results. In this regard, the gold standard was obtained by comparing tumor tissue sequencing results with the constitutive sequence to identify mutations in tumor tissues. In this analysis, we focused on the nucleotide positions at which the constitutive DNA of the subject was homozygous.

1. Untargeted whole-genome analysis.

Sequencing depths for white blood cells, tumor tissues, and plasma DNA from each patient are shown in Table 5.

Table 5

Median sequencing depths of different _________________ samples for four cases of HCC_________________

Case Mean sequencing depth (fold) White blood cells Tumor tissue Plasma before surgery Plasma after surgery GKK1 39 29 23 24 GKK2 39 29 25 28 GKKZ 46 33 18 21 GKK4 46 27 20 23 Patient with ovarian cancer 44 53 37 28

Dynamic thresholds for the minimum frequencies for detecting mutations in plasma (r) shown in Table 1 were used to detect mutations in each patient's plasma. Since the depth of sequencing at each locus can vary, the threshold can vary, effectively making the threshold dependent on the total number of sequence reads at the locus. For example, although the median depth is less than 50 (Table 5), the depth of sequencing at individual loci can vary greatly and coverage can be >100x.

In addition to sequencing errors, another source of error will be alignment. To minimize this type of error, sequence reads containing a mutation were realigned to the reference genome using the Bowtie alignment program (Langmead B et al. Genome Biol 2009, 10:R25). Only sequence reads that could be aligned to a unique position in the reference genome using SOAP2 and Bowtie were used for subsequent mutation analysis in plasma. Other combinations of alignment software packages based on different algorithms can also be used.

To further minimize sequencing and alignment errors in the actual sequencing results, we applied two additional filtering algorithms to identify nucleotide positions where we identified single nucleotide variations in the sequence reads: (1) >70% of sequence reads containing mutations could be re-aligned to the same genomic coordinate using Bowtie with a mapping quality of >Q20 (i.e., <1% misalignment probability); (2) >70% of sequence reads containing

- 31,048,297 mutations were located outside 5 bp from both ends (i.e. 5' and 3' ends) of the read sequences. This filtering principle was established because sequencing errors were more widespread at both ends of the read sequences.

We also investigated factors influencing tumor detection without any tumor genomic information. One such parameter was the fractional concentration of tumor-derived DNA in plasma. This parameter can be regarded as another gold standard parameter and was determined for comparison with known tumor genomic information obtained using GAAL.

Table 6 shows the nucleotide variations detected in plasma before and after treatment. For HCC1, without any tumor genomic information, a total of 961 single nucleotide variations were detected. Among these nucleotide variations detected in plasma, 828 were cancer-associated mutations. After surgical resection of HCC, the total number of nucleotide variations decreased to 43, and none of them were cancer-associated mutations.

For comparison, the fractional concentration of tumor-derived DNA in the preoperative plasma sample was 53%, which was determined using the known tumor genomic information. GCC2, GCC3, and GCC4, without any tumor genomic information, had plasma single nucleotide variation counts ranging from 27 to 32 in preoperative plasma samples. These results are consistent with the mathematical prediction that at approximately 20x sequencing depth, a very low percentage of cancer-associated mutations can be detected in plasma, and most sequence variations detected in plasma were due to sequencing errors. There was no significant change in the number of sequence variations detected after tumor resection. For comparison, the fractional concentrations of tumor-derived DNA in plasma ranged from 2.1 to 5%, which were determined using the known tumor genomic information.

Table 6

Nucleotide variations detected in plasma

Plasma before surgery Plasma after surgery Fractional concentration of tumor DNA Total number of single nucleotide variations Number of detected cancer-associated mutations Fractional concentration of tumor DNA Total number of single nucleotide variations Number of detected cancer-associated mutations GKK1 53% 961 828 0.4% 43 0 GKK2 5% 32 0 0.6% 49 0 gkkz 2.1% 29 0 0.2% 32 0 GKK4 2.6% 27 0 1.3% 35 1 Patient with ovarian (and breast) cancer 46% 1718 1502 0.2% 2 0

2. Targeted exon enrichment.

It was discussed above that increasing the sequencing depth of a region of interest can increase both the sensitivity and specificity of detecting cancer-associated mutations in plasma and, therefore, increase the discriminatory power between cancer patients and non-cancer subjects. Although increasing the sequencing depth of the entire genome is still very expensive, one alternative would be to enrich for some fragments for sequencing. In one embodiment, targeted enrichment for selected exons or virtually the entire exome can be performed for sequencing. This approach can significantly increase the sequencing depth of a target region without increasing the total number of sequence reads.

DNA sequencing libraries from plasma of HCC patients and a patient with ovarian (and breast) cancer were captured using the Agilent SureSelect All Exon Targeted Exome Enrichment Kit. Exon-enriched sequencing libraries were then sequenced using the HiSeq 2000 Sequencing System.

The read sequences were aligned with the human reference genome (hg18). After

- 32,048,297 alignments were analyzed for the presence of single nucleotide variations in read sequences uniquely mapped to exons. Dynamic thresholds shown in Table 2 were used to detect single nucleotide variations in plasma during captured exome analysis.

Fig. 14A is a table showing the detection of single nucleotide variations in plasma DNA from HKK1. Without any tumor genome information, we detected a total of 57 single nucleotide variations in plasma from targeted sequencing. Upon subsequent validation of the tumor sequencing results, 55 were found to be true tumor-associated mutations. It was previously discussed that the fractional concentration of tumor-derived DNA in plasma was 53% before surgery. After tumor resection, no single nucleotide variations were detected in plasma from targeted sequencing. These results indicate that plasma single nucleotide variation abundance analysis can be used to monitor disease progression in cancer patients.

Fig. 14B is a table showing the detection of single nucleotide variations in plasma DNA of HKK2. Without any tumor genome information, we detected a total of 18 single nucleotide variations in the plasma targeted sequencing results. All of these mutations were detected in tumor tissues. It was discussed earlier that the fractional concentration of tumor-derived DNA in plasma was 5% before surgery. No single nucleotide variations were detected in plasma after tumor resection. Compared with HKK1, which had a higher fractional concentration of tumor-derived DNA in plasma, fewer single nucleotide variations were detected in plasma in HKK2. The obtained results suggest that the amount of single nucleotide variations in plasma can be used as a parameter that reflects the fractional concentration of tumor-derived DNA in plasma and, consequently, the tumor burden in a patient, as it was shown that the concentration of tumor-derived DNA in plasma was positively correlated with tumor burden (Chan KC et al. Clin Chem 2005; 51:2192-5).

Figure 15A is a table showing the detection of single nucleotide variations in plasma DNA from HKK3. Without any tumor genomic information, we did not observe any single nucleotide variations in the targeted sequencing results in plasma samples either before or after resection. This may likely be due to the relatively low fractional concentration (2.1%) of tumor-derived DNA in the plasma of this patient. We hypothesized that further increasing the sequencing depth would improve the sensitivity of detection of cancer-associated mutations in cases with low fractional concentration of tumor-derived DNA.

Fig. 15B is a table showing the detection of single nucleotide variations in plasma DNA of HKK4. Without any tumor genome information, we detected only 3 single nucleotide variations in the plasma targeted sequencing results. All of these mutations were detected in tumor tissues. Compared with HKK1 and HKK2, which had higher fractional concentrations of tumor-derived DNA in plasma, HKK4, which had a fractional concentration of tumor DNA in plasma of 2.6%, had fewer single nucleotide variations in plasma. These results suggest that the amount of single nucleotide variations in plasma can be used as a parameter that reflects the fractional concentration of tumor-derived DNA in plasma and the tumor burden of a patient.

Fig. 16 is a table 1600 showing the detection of single nucleotide variations in plasma DNA from a patient with ovarian (and breast) cancer. Without any information about the tumor genome, we identified a total of 64 single nucleotide variations in the plasma targeted sequencing results. Of these, 59 were detected in ovarian tumor tissues. The estimated fractional concentration of ovarian tumor-derived DNA in plasma was 46%. A significant reduction in the total number of single nucleotide variations in plasma was found after resection of ovarian cancer.

In addition to using the SureSelect Targeted Enrichment System (Agilent), we also used the Nimblegen SeqCap EZ Exome+UTR Targeted Enrichment System (Roche) to enrich sequences from exons for sequencing. The Nimblegen SeqCap System covers exonic regions of the genome as well as the 5' and 3' untranslated regions. Plasma samples from four pretreatment HCC patients, two healthy controls, and two non-cancer chronic hepatitis B carriers were analyzed (Table 7). In other embodiments, other targeted enrichment systems can be used, including but not limited to liquid-phase or solid-phase hybridization systems.

Table 7. Exome sequencing results from four HCC patients (HCC1-4) using the Nimblegen SeqCap EZ Exome+UTR Targeted Enrichment System for sequence capture. Pre-treatment HCC3 plasma sequencing analysis was suboptimal due to a higher percentage of PCR-duplicate reads.

- 33 048297

Table 7

Plasma before treatment Plasma after treatment Fractional concentration of tumor-derived DNA in plasma obtained by GAAL analysis Number of sequence variations detected in plasma that satisfy dynamic thresholds Number of sequence variations that overlap with mutations detected in the corresponding tumor tissue Number of sequence variations detected in plasma that satisfy dynamic thresholds Number of sequence variations that overlap with mutations detected in the corresponding tumor tissue GKK1 53% 69 64 1 1 GKK2 5% 51 47 3 0 gkkz 2.1% 0 0 1 0 GKK4 2.6% 8 7 0 0

In two chronic hepatitis B carriers and two healthy controls, one or fewer single nucleotide variations were detected that met the dynamic threshold criteria (Table 8). In three of the four HCC patients, the number of sequence variations detected in plasma that met the dynamic threshold requirement was at least 8. In HCC3, no SNPs were detected that met the dynamic threshold. The sequenced reads of this sample had an increased ratio of PCR-duplicated reads, resulting in fewer non-duplicated sequenced reads. A marked decrease in the number of SNPs detected in plasma was observed after surgical resection of the tumor.

Table 8. Exome sequencing results from 2 chronic hepatitis B carriers (HBV1 and HBV2) and 2 healthy controls (control1 and control2) using the Nimblegen SeqCap EZ Exome+UTR Targeted Enrichment System for sequence capture.

Table 8

Number of sequence variations detected in plasma that satisfy dynamic thresholds VGV1 0 VGV2 1 Control! 1 Control2 1

XII. Tumor heterogeneity.

Quantitative analysis of single nucleotide mutations in a biological specimen (e.g. plasma/serum) is also useful for analyzing tumor heterogeneity, both intratumoral and intertumoral heterogeneity. Intratumoral heterogeneity refers to the existence of multiple tumor cell clones within the same tumor. Intertumoral heterogeneity refers to the existence of multiple tumor cell clones in two or more tumors of the same histologic type but present in different locations (either in the same organs or in different organs). In some tumor types, the existence of tumor heterogeneity is a poor prognostic indicator (Yoon HH et al. J Clin Oncol 2012; 30: 3932-3938; Merlo LMF et al. Cancer Prev Res 2010; 3: 1388-1397). For some tumor types, the greater the degree of tumor heterogeneity, the higher the likelihood of tumor progression or the development of resistant clones after targeted treatment.

Although cancers are thought to arise from the clonal expansion of a single tumor cell, the growth and progression of cancer will result in the accumulation of new and different mutations in different parts of the cancer. For example, when a patient with cancer develops metastases, the tumor in the original organ and the metastatic tumor will have many of the same mutations. However, cancer cells from the two sites will also carry a unique set of mutations that are not present in the other site of the tumor. Mutations that are the same in the two sites are expected to be present in greater concentrations than mutations that are found in only one site of the tumor.

A. Example.

We analyzed the blood plasma of a patient with bilateral ovarian cancer and breast cancer. Both ovarian tumors were serous adenocarcinoma. The left one was 6 cm and the right one was 12 cm in the longest dimension. Several metastases were also present.

- 34 048297 in the colon and omentum. DNA isolated from leukocytes was sequenced using the Illumina sequencing-by-synthesis platform with an average of 44-fold coverage of the haploid genome. In addition, single nucleotide mutations in plasma were analyzed for nucleotide positions at which only one allele was present, i.e., the homozygous allele.

DNA was isolated from four different sites of the left and right tumors and sequenced using the Illumina sequencing platform. Two sites (sites A and B) were from the right tumor and the other two sites (sites C and D) were from the left tumor. Sites A and B were approximately 4 cm apart. The distance between sites C and D was also approximately 4 cm. Plasma samples were collected from the patient before and after surgical resection of the ovarian tumors. DNA was then isolated from the patient's plasma. The sequencing depths of tumors from sites A, B, C, and D and from the plasma samples are shown in Table 9.

Table 9

Sequencing depth of tumors from localizations A, B, C and D

Sample Number of primary reads in sequencing Number of read sequences Haploid Genome Coverage Multiplicity Constitutive DNA from buffy coat 1091250072 876269922 43.81 Tumor of the right ovary 1374495256 1067277229 53.36 Tumor of the right ovary 934518588 803007464 40,15 Tumor of the left ovary (localization C) 1313051122 1036643946 51.83 Tumor of the left ovary (localization 1159091833 974823207 48.74 Plasma sample collected prior to surgery 988697457 741982535 37,10 Plasma sample collected after surgery 957295879 564623127 28,23

In this example, to detect a single tumor-associated single nucleotide mutation, the nucleotide position was sequenced at least 20 times in the tumor tissue and 30 times in the constitutive DNA. In other embodiments, other sequencing depths may be used, such as 35, 40, 45, 50, 60, 70, 80, 90, 100, and >100 times. Reducing the cost of sequencing will allow sequencing to be performed much more readily at greater depths. The nucleotide position is homozygous in the constitutive DNA, while a nucleotide substitution is observed in the tumor tissue. The criterion for the occurrence of a nucleotide substitution in the tumor tissue depends on the total sequencing depth of a particular nucleotide position in the tumor tissue. For 20-30-fold nucleotide coverage, the occurrence of a nucleotide substitution (threshold value) will be at least five times. For 31-50x coverage, the occurrence of a nucleotide substitution is at least six times. For 51-70x coverage, the required occurrence is at least seven times. These criteria are derived from predictions using the Poisson distribution of the detection sensitivity of true mutations and the expected number of false positive loci.

Figure 17 shows Table 1700, which shows the predicted sensitivities of different frequency requirements and sequencing depths. The sensitivity will correspond to the number of true mutations detected at a particular coverage depth using a particular threshold. The higher the sequencing depth, the more likely it is that a mutation will be detected.

- 35 048297 yield a mutation for a given threshold, since more reads with the mutation will be obtained. For higher thresholds, the probability of detecting a mutation will be lower, since the criterion will be more stringent.

Fig. 18 is a table 1800 showing the predicted numbers of false positive loci for different thresholds and different sequencing depths. The number of false positives increases with increasing sequencing depth because more sequence reads are performed. However, for a threshold of five or more, no false positives are predicted, even at a sequencing depth of 70 times. In other embodiments, different occurrence criteria can be used to achieve the desired sensitivity and specificity.

Figure 19 is a tree diagram illustrating the number of mutations detected at different tumor sites. Mutations were determined by direct sequencing of the tumors. Site A contained 71 tumor-specific mutations and site B contained 122 site-specific mutations, even though they were only 4 cm apart. Ten mutations were observed simultaneously at sites A and B. Site C contained 168 tumor-specific mutations and site D contained 248 site-specific mutations, even though they were only 4 cm apart. Twelve mutations were observed simultaneously at sites C and D. Significant heterogeneity was found in the mutational profiles among the different tumor sites. For example, 248 mutations were detected only at site D but not at the other three tumor sites. A total of 2129 mutations were detected across all sites. Thus, many mutations were similar across tumors. Thus, there were only seven ONV groups. No differences were observed among these four regions with respect to copy number variations.

Fig. 20 is a table 2000 showing the number of fragments carrying tumor-derived mutations in a plasma sample before and after treatment. Also shown are the detected fractional concentrations of tumor-derived DNA carrying the corresponding mutations. The mutation category refers to the tumor site(s) in which the mutation was detected. For example, mutation category A refers to a mutation present only in site A, whereas mutation category ABCD refers to a mutation present in all four tumor sites.

Of the 2129 mutations that were present in all four tumor sites, 2105 (98.9%) were detected in at least one plasma DNA fragment. On the other hand, of the 609 mutations that were present in only one of the four tumor sites, only 77 (12.6%) were detected in at least one plasma DNA fragment. Therefore, quantitative analysis of single nucleotide mutations in plasma can be used to identify the relative prevalence of these mutations in tumor tissues. This information will be useful for studying cancer heterogeneity. In this example, a variation was considered a potential mutation if it was observed in the sequencing results at least once.

Fractional concentrations of circulating tumor DNA were determined within each NBC group. Fractional concentrations of tumor DNA in plasma before and after surgery, as determined by NBCs present in all 4 regions (i.e., group ABCD), were 46% and 0.18%, respectively. These percentages correlated well with those obtained by GAAL analysis: 46% and 0.66%. Mutations present in all 4 regions (i.e., group ABCD) made the largest fractional contribution to tumor-derived DNA in plasma.

Fractional tumor-derived DNA concentrations in preoperative plasma, as determined by OHBs from groups AB and CD, were 9.5 and 1.1%, respectively. These concentrations corresponded to the relative sizes of right and left ovarian tumors. Fractional tumor-derived DNA concentrations, as determined by region-unique OHBs (i.e., from groups A, B, C, and D), were generally low. These results suggest that the use of a whole-genome shotgun approach may provide a more representative picture of the overall tumor burden in a patient with cancer than the more traditional approach of targeting specific tumor-associated mutations. In the latter approach, if only a subset of tumor cells have the target mutations, important information regarding the risk of relapse or disease progression caused by tumor cells lacking the target mutations may be missed, or the emergence of a treatment-resistant clone may be missed.

Fig. 21 is a diagram 2100 that shows the distributions of the plasma frequencies of mutations detected at one tumor site and mutations detected at all four tumor sites. The bar diagram 2100 shows the results for two types of mutations: (1) mutations detected at only one site and (2) mutations detected at all four tumor sites. The horizontal axis shows the number of times a mutation

- 36 048297 were detected in plasma. The vertical axis shows the percentage of mutations that correspond to a specific value on the horizontal axis. For example, approximately 88% of type (1) mutations occurred in plasma only once. It can be seen that mutations that occurred in one location were detected predominantly once, and not more than four times. Mutations present in only one tumor location were detected in plasma much less frequently compared to mutations present in all four tumor locations.

One application of this method is that it allows clinicians to assess the mass of tumor cells that carry different classes of mutations. Some of these mutations could potentially be compensated for by targeted agents. Agents that target mutations that are present in a larger number of tumor cells are expected to have a pronounced therapeutic effect.

Fig. 22 is a diagram 2200 showing the predicted distribution of plasma occurrence of mutations originating from a heterogeneous tumor. The tumor includes two groups of mutations. One group of mutations is present in all tumor cells and the other group of mutations is present in only _1/4 of the tumor cells, based on the approximation that the two locations represent tumors from each ovary. The total fractional concentration of tumor-derived DNA in plasma was taken to be 40%. It was assumed that the plasma sample was sequenced with an average depth of 50 times per nucleotide position. According to this predicted distribution of plasma occurrence, mutations that were present in all tumor tissues could be distinguished from mutations that were present in only ^1/4 of the tumor cells by their plasma occurrence. For example, an occurrence of 6 times could be used as a threshold value. For mutations present in all tumor cells, 92.3% of the mutations will be present in plasma at least 6 times. In contrast, for mutations present in ^1/4 of the tumor cells, only 12.4% of the mutations will be present in plasma at least 6 times.

Fig. 23 is a table 2300 showing the specificity of the embodiments for 16 healthy control subjects. Their plasma DNA samples were sequenced with median coverage 30 times. Mutations that were present in the plasma of the ovarian cancer patient described above were detected in the plasma samples of these healthy subjects. Mutations present in the tumor of the ovarian cancer patient were very rarely detected in the sequencing results of the plasma of the healthy controls, and none of the mutation categories occurred at a fractional concentration of >1%. The results showed that this detection method is highly specific.

B. Method.

Fig. 24 is a flow chart of a method 2400 for analyzing heterogeneity of one or more tumors of a subject according to embodiments of the present invention. Some steps of the method 2400 can be performed as described herein.

At block 2410, a constituent genome of the subject is obtained. At block 2420, one or more sequence markers are obtained for each of a plurality of DNA fragments in a biological sample of the subject, wherein the biological sample contains extracellular DNA. At block 2430, the genomic positions of the sequence markers are determined. At block 2440, the sequence markers are compared to the constituent genome to determine a first number of first loci. At each of the first loci, the number of sequence markers with a variant sequence relative to the constituent genome is above a threshold value, where the threshold value is greater than one.

In block 2450, measures of heterogeneity of one or more tumors are calculated based on the corresponding first numbers from the set of first positions in the genome. In one aspect, said measures can provide a value that represents the number of mutations that are the same among tumors, compared to the number of mutations that are different among tumors. Here, different tumors can exist as a single entity, with different tumors within the entity, which can represent what is commonly referred to as intratumor heterogeneity. Said measure can also relate to whether mutations are present in one or more tumors, or whether mutations are present in many or most tumors. More than one measure of heterogeneity can be calculated.

In block 2460, the heterogeneity measure may be compared to a threshold value to determine a classification of the level of heterogeneity. One or more measures may be applied in various ways. For example, one or more heterogeneity measures may be applied to predict the likelihood of tumor progression. The higher the heterogeneity of some tumors, the higher the likelihood that they will progress, and the higher the likelihood that a resistant clone will emerge after treatment (e.g., targeted therapy).

C. Measures of tumor heterogeneity.

One example of a measure of heterogeneity is the number of concentration zones of different groups of mutations in plasma. For example, if a patient has two predominant tumor clones, and if these clones are present in different concentrations, then we would expect to see two different mutations with different concentrations in plasma. These different values can be calculated by determining the fractional concentration for different sets of mutations, where each set corresponds to one of

- 37,048,297 tumors.

Each of these concentrations can be called a concentration zone or concentration class. If a patient has more clones, then more concentration zones/classes will be observed. Thus, the more zones, the higher the heterogeneity. The number of concentration zones can be observed by plotting the fractional concentrations for different mutations. A histogram of the different concentrations can be constructed, with the different peaks corresponding to different tumors (or different clones of a single tumor). Mutations that are common in all or some tumors (or tumor clones) are likely to have a larger peak. These peaks can be analyzed to determine which smaller peaks combine to form a larger peak. A fitting procedure can be used, for example, similar to the fitting procedure in Figs. 10B and 11.

In one embodiment, the histogram is a graph in which the y-axis represents the number (e.g., number or proportion) of loci and the x-axis represents the fractional concentration. Mutations that are common in all or some tumors will result in a higher fractional concentration. The size of the peak will represent the number of loci that result in a particular fractional concentration. The relative size of the peaks at low and high concentration will reflect the degree of heterogeneity of the tumors (or tumor clones). A larger peak at high concentration reflects that most mutations are common in most or all tumors (or tumor clones), and indicates a lower degree of tumor heterogeneity. If the peak at low concentration is larger, then most mutations are common in several tumors (or tumor clones), indicating a higher degree of tumor heterogeneity.

The more peaks present, the more site-specific mutations are present. Each peak may correspond to a different set of mutations, where the set of mutations originates from a subset of tumors (e.g., from only one or two tumors, as described above). In the example in Fig. 19, there may be a total of 7 peaks: 4 peaks each for a single site that are likely to have the lowest concentration (depending on the relative size of the tumors), two peaks for site AB and site CD, and one peak for mutations that are the same in all sites.

The position of the peaks can also indicate the relative size of the tumors. A higher concentration will correlate with a larger tumor, as a larger tumor will release more tumor DNA into a sample, such as plasma. In this way, the mass of tumor cells carrying different classes of mutations can be estimated.

Another example of a measure of heterogeneity is the ratio of mutated sites with a relatively small number of variant reads (e.g., 4, 5, or 6) compared to the ratio of mutated sites with a relatively large number of variant reads (e.g., 9-13). Referring again to Fig. 22, for site-specific mutations, there are fewer variant reads (also resulting in a lower fractional concentration). For site-general mutations, there are more variant reads (also resulting in a higher fractional concentration). The first ratio, with 6 reads (the lower number), divided by the second ratio, with 10 reads (the higher number), expresses a measure of heterogeneity. If the ratio is small, there are few site-specific mutations and thus the level of heterogeneity is low. If the ratio is large (or at least larger than the values determined for known samples), the level of heterogeneity is higher.

D. Determination of threshold values.

Thresholds can be determined from subjects whose tumors have been biopsied (e.g., as described above) to directly determine the level of heterogeneity. This level can be determined in a variety of ways, such as using the ratio of site-specific mutations to site-common mutations. Biological samples (e.g., plasma samples) can then be analyzed to determine measures of heterogeneity, whereby the heterogeneity measure for biological samples can be related to the level of heterogeneity determined by direct analysis of tumor cells.

This procedure will calibrate the thresholds relative to the heterogeneity levels. If the heterogeneity measure in a given analysis falls between two thresholds, then the heterogeneity level can be estimated to be between the levels corresponding to the thresholds.

In one embodiment, a calibration curve may be calculated from heterogeneity levels determined for biopsies and a corresponding measure of heterogeneity determined for a plasma sample (or other sample). In such an example, the heterogeneity levels are numerical, where these numerical levels may correspond to different classifications. Different ranges of numerical levels may correspond to different diagnoses, such as different stages of cancer.

E. Method using fractional concentration from genomic representation.

Tumor heterogeneity can also be analyzed using fractional concentration, such as determined using embodiments of method 1200. Genomic regions,

- 38 048297 in which a loss of one copy is detected may originate from different tumors. Thus, the fractional concentration determined for different regions of the genome may differ depending on whether the amplification (or deletion for a loss of 1 copy) is present only in one tumor or in several tumors. Thus, the same measures of heterogeneity can be applied to the fractional concentrations determined using embodiments of the method 1200.

For example, one may find that one region of the genome corresponds to a loss of 1 copy, and one may determine the fractional concentration from the corresponding density in that region of the genome alone (the corresponding density may be used as the fractional concentration). One may construct a histogram from the different corresponding densities by counting the number of regions with different densities. If only one tumor, or one tumor clone, or one tumor metastasis was found to have a copy gain in a particular region, the density of that region will be less than the density of a region in which copy gain was found in several tumors, or several tumor clones, or several tumor metastases (i.e., the fractional concentration of tumor DNA in a region common to mutations will be greater than in a region with an individual mutation). Thus, the measures of heterogeneity described above may be applied to peaks detected from copy gains or losses in different regions, as well as to the fractional concentration at different locations in which the distribution of fractional concentrations was detected.

In one embodiment, if the corresponding densities are used to construct the histogram, a separation of gains and losses can be obtained. Regions in which a gain was detected can be analyzed separately by constructing a histogram only for gains, and a separate histogram can be constructed only for losses. When fractional concentration is used, the peaks of losses and gains can be analyzed together. For example, for fractional concentrations, the difference (e.g., as an absolute value) from a reference density can be used, and therefore, the fractional concentrations for gains and losses can contribute to the same peak.

XIII. Computer system.

In any computer systems mentioned in this application, any suitable number of subsystems may be used. Examples of such subsystems are shown in Fig. 25 in a computer device 2500. In some embodiments, the computer system includes a single computer device, where the subsystems may be components of the computer device. In other embodiments, the computer system may include multiple computer devices, each of which is a subsystem, with internal components.

The subsystems shown in Fig. 25 are interconnected by a system bus 2575. Additional subsystems such as a printer 2574, a keyboard 2578, a hard drive 2579, a monitor 2576 which is connected to a video adapter 2582, and other subsystems are shown. Peripheral devices and input/output (I/O) devices which are connected to the input/output controller 2571 may be connected to the computer system by any number of means known in the art, such as a serial input/output port 2577. For example, a serial input/output port 2577 or an external interface 2581 (e.g., an Ethernet local area network, a Wi-Fi wireless network, etc.) may be used to connect the computer system 2500 to a wide area network such as the Internet, to a mouse input device, or to a scanner. The interconnection via the system bus 2575 allows the central processor 2573 to exchange data with each subsystem and control the execution of instructions from the system memory 2572 or the hard disk 2579, as well as exchange information between the subsystems. The system memory 2572 and/or the hard disk 2579 may include computer-readable media. Any values mentioned in this application may be transferred from one component to another component and may be output to a user.

A computer system may include multiple identical components or subsystems, such as those connected to each other via an external interface 2581 or via an internal interface. In some embodiments, computer systems, subsystems, or devices may exchange information via a network. In such cases, one computer may be considered a client and another computer a server, where each may be part of the same computer system. Both a client and a server may include multiple systems, subsystems, or components.

It should be obvious that any of the embodiments of the present invention can be implemented in the form of logic circuits of a control device using hardware (for example, an application-specific integrated circuit or a field-programmable gate array) and/or using computer software, typically with a programmable processor in modular or integrated form. In this application, the processor includes a multi-core processor on the same integrated chip or several processing modules located on a single printed circuit board or combined into a network. Based on the description and idea proposed in this application, other ways and/or methods for implementing embodiments of the present invention using hardware and combinations of hardware and software will be known and understood by a person of ordinary skill in the art.

- 39 048297

Any software components or functions described in this application may be implemented as program code that will be executed by a processor using any suitable programming language, such as, for example, Java, C++ or Perl, using, for example, conventional or object-oriented techniques. The program code may be stored as a series of instructions or commands on machine-readable storage and/or transmission media, suitable media include random access memory (RAM), read-only memory (ROM), magnetic media such as a hard disk or floppy disk, or optical media such as a compact disc (CD) or DVD (digital versatile disk), flash memory and the like. Machine-readable media may be any combination of such storage or transmission devices.

Such programs may also be encoded and transmitted using carrier signals suitable for transmission over wired, optical and/or wireless networks corresponding to a variety of protocols, including the Internet. In this regard, computer-readable media according to an embodiment of the present invention may be obtained using an information signal encoded by such programs. Computer-readable media encoded with a program code may be packaged together with a compatible device or provided separately from other devices (e.g., via download via the Internet). Any such computer-readable media may be located within or on a single program product (e.g., a hard disk, a CD, or an entire computer system), and may be present within or on different program products within a system or network. The computer system may include a monitor, printer, or other suitable display device for providing any of the results mentioned in this application to a user.

Any of the methods described in this application can be implemented in whole or in part by a computer system including one or more processors that can be configured to perform the specified steps. Thus, embodiments can be directed to computer systems configured to perform the steps of any of the methods described in this application, possibly with different components performing the corresponding steps or a corresponding group of steps. Although the steps are numbered, the steps according to the methods described in this application can be performed simultaneously or in a different order. In addition, the blocks of these steps can be used with blocks of other steps from other methods. Also, the entire step or blocks of a step can be optional. In addition, any of the steps of any of the methods can be implemented using modules, circuits or other means for implementing these steps.

In particular, a method for detecting cancer or a pre-malignant change in a subject is proposed for implementation, wherein said method includes:

obtaining the constitutive genome of the subject;

obtaining one or more sequence markers for each of a plurality of DNA fragments in a biological sample of a subject, wherein said biological sample comprises cell-free DNA;

determination of the positions of sequence markers in the genome;

comparison of sequence markers with the constitutive genome to determine the first number of first loci, whereby:

in each of the first loci, the number of sequence markers with a variant sequence relative to the constitutive genome is above a threshold value, where the specified threshold value is greater than one;

determining a parameter based on the number of sequence markers with a variant sequence in the first loci; and comparing this parameter with a threshold value to determine the classifications of the level of cancer in the subject.

In this method, the threshold value may be determined from one or more samples from one or more other subjects.

The threshold value for a locus may depend on the total number of sequence markers that are located in a given locus of the genome.

For at least two of the first loci, different threshold values may be applied.

The method may further comprise dynamically determining a first threshold value for one of the first loci, wherein said one of the loci is located within the first region.

In this case, the first threshold value can be determined based on the sequencing depth of one of the first loci or based on the false positive rate, which depends on the sequencing error rate, the sequencing depth of the first region, and the number of nucleotide positions in the first region.

In the latter case, the first threshold value can be determined based on the number of true positive values in the first region. Moreover, it is not necessary to calculate the number of true positive values for the first threshold value based on the sequencing depth D and the fractional concentration f of tumor DNA in the biological sample. For example, to determine

- 40 048297 numbers of true positive values, we can apply the Poisson probability distribution according to the formula: where Pb is the probability of detecting true positive values, r is the first threshold value, and M _P = Dxf/2.

Alternatively, in the case of dynamically determining the first threshold for one of the first loci, the first threshold can be determined using any of the following criteria:

if the sequencing depth is less than 50, the first threshold is 5, if the sequencing depth is 50-110, the first threshold is 6, if the sequencing depth is 111-200, the first threshold is 7, if the sequencing depth is 201-310, the first threshold is 8, if the sequencing depth is 311-450, the first threshold is 9, if the sequencing depth is 451-620, the first threshold is 10, and if the sequencing depth is 621-800, the first threshold is 11.

The parameter specified in the description of the method may be a weighted sum of the first number of first loci, wherein the contribution of each of the first loci is weighted by the significance indicator assigned to the corresponding first loci. Alternatively, this parameter may include the sum of the sequence markers in which the variant sequence is detected in the first number of first loci. Moreover, for example, this sum may be a weighted sum, and wherein the first weight of one of the first loci differs from the second weight of the second of the first loci. As an option, the first weight may be greater than the second weight, and wherein one of the first loci may be associated with cancer, and the second of the first loci may not be associated with cancer.

In one embodiment, said parameter may be a first number of first loci.

In the above method, determining the position of a sequence marker in the genome may include:

aligning at least a portion of the sequence markers with a reference genome, wherein the alignment of the sequence marker allows for one or more mismatches between the sequence marker and the constituent genome. Moreover, the comparison of the sequence markers with the constituent genome may optionally include:

comparing the constitutive genome with the reference genome to determine a second number of second loci containing a variant compared to the reference genome;

determination based on the alignment of the third number of third loci, where:

in each of the third loci, the number of sequence markers with a variant sequence compared to the reference genome is above a threshold value; and calculating the difference between the third number and the second number to obtain the first number of the first loci.

Calculating the difference between the third quantity and the second quantity can allow one to identify specific loci from the first loci.

In this case, the definition of the specified parameter may include:

for each locus from the first number of first loci:

determining the number of sequence markers that align with the locus and contain the sequence variant at that locus; and determining a parameter based on the corresponding numbers.

In the above method, a constitutive genome can be obtained from a constitutive sample of a subject that contains more than 50% constitutive DNA.

In the above method, determining the position of a sequence marker in the genome may include:

aligning at least a portion of the sequence markers with the constituent genome, wherein the alignment of the sequence marker allows for one or more mismatches between the sequence marker and the constituent genome. Moreover, the comparison of the sequence markers with the constituent genome optionally includes:

determining, based on the alignment, sequence markers that contain a sequence variant at some position in the genome compared to the constituent genome of the subject;

for each position in the genome where a variant sequence is found:

determining a corresponding first number of sequence markers that align with the specified position in the genome and contain the sequence variant at the given position in the genome;

determination of the parameter based on the corresponding first quantities.

In this embodiment, determining the parameter based on the corresponding first quantities may include:

adding the corresponding first quantities to obtain a first sum; and applying the first sum to determine the parameter.

- 41 048297

Moreover, it is not necessary to use the first sum to determine the parameter; it may include: subtracting the number of positions in the genome in which the variant sequence was detected from the first sum.

Alternatively, the use of the first sum to determine the parameter may include: normalizing the first sum based on the number of aligned sequence markers. The described method may further include:

obtaining a constitutive sample from a subject that contains more than 90% constitutive DNA;

performing random sequencing of DNA fragments in the constitutive sample to obtain one or more second sequence markers for each of a plurality of DNA fragments in the constitutive sample;

aligning at least a portion of the second sequence markers with a reference genome, wherein the alignment of the second sequence marker allows for a mismatch between the second sequence marker and the constitutive genome at M or fewer positions in the genome, where M is an integer value equal to or greater than one; and constructing the constitutive genome based on the second sequence markers and the alignment.

In this case, the constitutive sample may not necessarily be a biological sample and construction of the constitutive genome may involve:

determination of a consensus sequence, which involves determination of a homozygous locus or a heterozygous locus containing two alleles; and application of the consensus sequence to the constitutive genome.

The described method may also additionally include:

obtaining a biological sample taken from a subject; and performing random sequencing of DNA fragments in the biological sample to obtain one or more sequence markers for each of a plurality of DNA fragments in the biological sample.

In the described method, one or more sequence markers may optionally be obtained by random sequencing of DNA fragments in a biological sample.

In the described method, the biological sample may be urine, pleural fluid, ascitic fluid, peritoneal fluid, saliva, cerebrospinal fluid, or a stool sample.

In the described method, the parameter may alternatively represent the fractional concentration of tumor-derived DNA.

Accordingly, a computer product is also proposed that includes a non-temporary machine-readable medium on which a set of instructions are stored that, when executed, control the implementation of the method described above by the computer system.

It is also proposed to implement a method for analyzing the heterogeneity of one or more tumors of a subject, wherein said method includes:

obtaining the constitutive genome of the subject;

obtaining one or more sequence markers for each of a plurality of DNA fragments in a biological sample of a subject, wherein said biological sample comprises cell-free DNA;

determination of the positions of sequence markers in the genome;

comparison of sequence markers with the constitutive genome to determine the first number of first loci, whereby:

at each of the first loci, the number of sequence markers with a variant sequence relative to the constitutive genome is above a threshold value, where said threshold value is greater than one; and calculating a measure of heterogeneity of one or more tumors based on the corresponding first quantities from the set of first positions in the genome.

This method may additionally include:

comparing a measure of heterogeneity with one or more threshold values to determine a classification of the level of heterogeneity.

In this case, one or more threshold values may be determined from one or more other subjects who have had a tumor biopsy and an analysis may be performed to determine mutations in the tumor biopsies to determine the level of heterogeneity, and a measure of heterogeneity of a biological sample containing cell-free DNA from one or more other subjects may be used to determine the threshold value.

In such an embodiment, comparing the heterogeneity measure to one or more threshold values may include:

inputting a measure of heterogeneity into a calibration function that outputs a level of heterogeneity based on the measure of heterogeneity.

Alternatively, a method for analyzing the heterogeneity of one or more tumors in a subject provides

- 42 048297 considers that one or more measures include the total number of first positions in the genome at which more than one DNA fragment aligns with the constitutive DNA, although it contains a variant sequence at that position in the genome.

As another embodiment, a method for analyzing heterogeneity of one or more tumors of a subject comprises calculating a plurality of heterogeneity measures, wherein calculating the heterogeneity measures comprises:

for each of the first loci, calculating the proportion of sequence markers containing the variant sequence;

constructing a histogram of the number of first loci for the values of the specified ratios; and identifying the number of peaks in the histogram.

In this case, the heterogeneity measure may correspond to the number of peaks detected or include the ratio of the heights of two peaks.

Each ratio may represent the fractional concentration of tumor DNA measured at a particular first locus.

As another option, the measure of heterogeneity may correspond to the ratio of a first proportion of first loci containing a first specified number of sequence markers with a variant sequence to a second proportion of first loci containing a second specified number of sequence markers with a variant sequence.

In this case, the first specific quantity may be less than the second specific quantity.

Moreover, it is not necessary that the first specific quantity represent the first range and the second specific quantity represent the second range, wherein the first range is lower than the second range.

Also, optionally, the first specific amount and the second specific amount may correspond to a fractional concentration or an absolute amount of sequence markers with a variant sequence.

In another variant, the measure of heterogeneity can be determined from the histogram of the number of first loci corresponding to each of the corresponding first quantities.

In this case, the set of first identified positions in the genome may include a first subset and a second subset, wherein the one or more measures may include a first histogram of corresponding first counts corresponding to the first subset and a second histogram of corresponding second counts corresponding to the second subset.

In another embodiment, the heterogeneity measure may include the proportion of first loci that include the corresponding first quantities above a certain value.

A method for determining the fractional concentration of tumor DNA in a biological sample containing extracellular DNA is also proposed for implementation, wherein said method includes:

obtaining one or more sequence markers for each of a plurality of DNA fragments in a biological sample;

determination of the positions of sequence markers in the genome;

for each of the many regions of the genome:

determination of the corresponding number of DNA fragments within a given region of the genome using sequence markers located in the genome within a given region of the genome;

normalizing the corresponding quantity to obtain the corresponding density; and comparing the corresponding density with a reference density to detect the loss of 1 copy or the gain of 1 copy in a given region of the genome;

calculating the first density from one or more corresponding densities in which a loss of 1 copy is detected, or from one or more corresponding densities in which a gain of 1 copy is detected; and calculating the fractional concentration by:

comparison of the first density with another density to obtain the difference, and the difference is normalized to the reference density.

This method may additionally include:

construction of a histogram of the number of genomic regions with different corresponding densities;

identifying a number of peaks in the histogram; and calculating a measure of heterogeneity based on the ratio of the number of genomic regions in one or more first peaks to the number of genomic regions in one or more second peaks.

It is not necessary to use fractional concentration values in the histogram, determined separately for each region of the genome when constructing the histogram.

Optionally, the densities of the first peaks may correspond to a first specific amount, and the densities of the second peaks may correspond to a second specific amount.

In this case, the first specific quantity may represent a first range, and the second specific quantity may represent a second range, provided that

-

Claims (54)

The first range is smaller than the second range. Alternatively, comparison of the corresponding density with the reference density to detect loss of 1 copy or gain of 1 copy in a given genomic region involves: calculating the difference between the corresponding density and the reference density; and comparing the difference with a threshold value. The difference can be normalized to the reference density by: dividing the difference by the reference density. Optionally, another density may represent a reference density. In this case, the calculation of the fractional concentration may additionally include multiplying the difference by two. In a method for analyzing heterogeneity of one or more tumors of a subject, a first density may be calculated from the corresponding densities that are determined to demonstrate a gain of 1 copy, then the other density may be a second density that is calculated from the corresponding densities that are determined to demonstrate a loss of 1 copy. In this case, the difference can be normalized to the reference density by: calculation of the first ratio of the first density and the reference density; calculating the second ratio of the second density and the reference density, wherein the difference is calculated between the first ratio and the second ratio. Alternatively, comparison of the corresponding density with the reference density to detect loss of 1 copy or gain of 1 copy in a given region of the genome may include: fitting the peaks to the distribution curve of the histogram of the corresponding densities, where the first density corresponds to the first peak and the second density corresponds to the second peak. In a method for analyzing heterogeneity of one or more tumors in a subject, all regions of the genome that show a statistically significant increase in the corresponding density compared to a reference density can be defined as demonstrating gain of 1 copy. Normalization of the corresponding quantity to obtain the corresponding density may involve the use of the same total number of aligned reference markers to determine the corresponding density and the reference density. Alternatively, normalizing the corresponding count to obtain the corresponding density may involve dividing the corresponding count by the total number of aligned reference markers. All regions of a set of genomic regions may have the same length. Alternatively, the genomic regions may not overlap. The specific details of particular embodiments may be combined in any suitable manner without departing from the spirit and scope of the embodiments of the present invention. However, other embodiments of the present invention may be directed to specific embodiments relating to each individual aspect, or to specific combinations of these individual aspects. The foregoing description of typical embodiments of the present invention is presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the present invention to the precise form disclosed, and many modifications and variations are possible in light of the teachings described above. The embodiments have been chosen and described in order to better explain the principles of the present invention and its practical application, in order to thereby enable others skilled in the art to better utilize the present invention in various embodiments and with various modifications that are suitable for the particular intended application. The use of the singular form is intended to mean one or more unless expressly stated otherwise. All patents, patent applications, publications and patent specifications mentioned in this application are incorporated by reference in their entirety for all purposes. None of these references are considered prior art. CLAUSE OF THE INVENTION 1. A method for determining, using a computer system, a cancer classification level in a subject, comprising obtaining a first plurality of read sequences for a first sample of the subject, wherein the first sample comprises a constitutive sample taken from the subject at a first point in time; identifying a first set of sequence variants by comparing the first set of read sequences with a reference genome; obtaining a second plurality of readable sequences for a second sample of the subject, wherein the second sample comprises a biological sample comprising a plurality of extracellular molecules - 44,048,297 DNA molecules taken from the subject at the second point in time; identifying a second set of sequence variants by comparing the second set of sequence reads with a reference genome; comparing the first set of sequence variants and the second set of sequence variants to identify sequence variants present in the second set of sequence variants but not present in the first set of sequence variants; determining a parameter based on sequence variants present in a second set of sequence variants but not present in the first set of sequence variants; and comparing the parameter to a threshold value to determine a cancer level classification in the subject, wherein the threshold value corresponds to a range of parameter values for subjects having the cancer level classification. 2. The method of claim 1, wherein identifying the first set of sequence variants or the second set of sequence variants comprises determining genomic positions for the first plurality of sequence reads obtained from the first sample or the second plurality of sequence reads obtained from the second sample; and detecting sequence variants in a subset of the sequence reads located at a plurality of loci, wherein the sequence variants at the plurality of loci are detected based on each of the subset of sequence reads comprising one or more nucleotides that differ from a reference genome at the plurality of loci. 3. The method according to claim 1 or 2, wherein the constitutive sample is a biological sample that contains less than 50% tumor DNA. 4. The method according to claim 3, wherein the constitutive sample is selected from the group consisting of blood cells, leukocytes, DNA of healthy blood cells, DNA of buccal cells, DNA of hair roots and DNA of the buffy coat. 5. The method according to any one of claims 1 to 4, wherein the first plurality of readable sequences are obtained by sequencing DNA molecules from a constitutive sample of the subject. 6. The method according to any one of claims 1 to 5, wherein the second sample is selected from one or more groups consisting of plasma, serum, urine, ascites fluid, peritoneal fluid, saliva, cerebrospinal fluid and a stool sample. 7. The method of claim 6, wherein the second plurality of readable sequences are obtained by sequencing extracellular DNA molecules from a second sample of the subject. 8. The method according to claim 7, wherein sequencing the extracellular DNA molecules from the second sample comprises targeted sequencing. 9. The method according to claim 8, wherein the targeted sequencing comprises using a solid-phase capture method, a solution hybridization method, or PCR amplification. 10. The method of claim 8, wherein the targeted sequencing comprises targeted exon enrichment. 11. The method of claim 8, wherein the targeted sequencing is directed at known tumor-associated mutations. 12. The method according to any one of claims 7 to 11, wherein sequencing the extracellular DNA molecules from the second sample of the subject comprises paired-end sequencing, random sequencing, or massively parallel sequencing. 13. The method according to any one of claims 1 to 12, wherein each of the one or more sequence variants comprises a single nucleotide variant. 14. The method according to any one of claims 1 to 13, wherein one or more sequence variants comprise one or more non-synonymous variants. 15. The method according to any one of claims 2 to 14, wherein the parameter is determined based on the concentration of one or more sequence variants in a plurality of loci. 16. The method according to any one of claims 2 to 14, wherein the parameter is the number of loci containing one or more sequence variants present in the second set of sequence variants but not present in the first set of sequence variants. 17. The method according to any one of claims 2 to 14, wherein determining the parameter comprises, for each locus of the plurality of loci, counting read sequences that align with the locus and contain a sequence variant at the locus; and determining the parameter based on the corresponding counts. 18. The method according to any one of claims 2 to 14, wherein the parameter is a weighted sum of the number of loci containing one or more sequence variants present in the second set of sequence variants but not present in the first set of sequence variants, wherein the contribution of each locus is weighted based on a significance score assigned to the locus. 19. The method of any one of claims 1-18, wherein the threshold value is determined based on one or more samples from one or more other subjects. 20. A method for determining, using a computer classification system, the level of cancer in a subject, - 45 048297 comprising, at a first point in time, determining the constituent genome of a subject, wherein this constituent genome is determined on the basis of a plurality of read sequences obtained for DNA molecules from a first sample of the subject, and wherein the first sample contains the constituent sample; comparing the constitutive genome with a reference genome and identifying a first set of loci, wherein each locus from the first set of loci has a sequence variant in the constitutive genome relative to the reference genome; at a second time, determining a genomic sample of the subject, wherein the genomic sample is determined based on a plurality of read sequences obtained for DNA fragments from a second sample from the subject, wherein the second sample comprises a biological sample including a plurality of extracellular DNA molecules; comparing the sample genome with a reference genome and identifying a second set of loci, wherein each locus of the second set of loci has a sequence variant in the sample genome relative to the reference genome; determining a parameter based on the number of loci that have a sequence variant in the second set of loci but not in the first set of loci; and comparing the parameter to a threshold value to determine a cancer level classification for the subject, wherein the threshold value corresponds to a range of parameter values for subjects having the cancer level classification. 21. The method of claim 20, wherein identifying the first set of loci or the second set of loci comprises determining genomic positions for a plurality of sequence reads obtained from the first sample or the second sample; and detecting sequence variants in a subset of the sequence reads located at the plurality of loci, wherein the sequence variants in the first set of loci or the second set of loci are detected based on each sequence read of the subset of sequence reads comprising one or more nucleotides that differ from a reference genome at one or more genomic positions. 22. The method according to claim 20 or 21, wherein the constitutive sample is a biological sample that contains less than 50% tumor DNA. 23. The method according to claim 22, wherein the constitutive sample is selected from the group consisting of blood cells, leukocytes, DNA of healthy blood cells, DNA of buccal cells, DNA of hair roots and DNA of buffy coat. 24. The method according to any one of claims 20 to 23, wherein the plurality of read sequences of the first sample are obtained by sequencing extracellular DNA molecules from a constitutive sample of the subject. 25. The method according to any one of claims 20 to 24, wherein the second sample is selected from one or more of the group consisting of plasma, serum, urine, ascites fluid, peritoneal fluid, saliva, cerebrospinal fluid and a stool sample. 26. The method of claim 20, wherein the plurality of readable sequences of the second sample are obtained by sequencing extracellular DNA molecules from the second sample of the subject. 27. The method according to claim 26, wherein sequencing the extracellular DNA molecules from the second sample comprises targeted sequencing. 28. The method according to claim 27, wherein the targeted sequencing comprises using a solid-phase capture method, a solution hybridization method, or PCR amplification. 29. The method of claim 27, wherein the targeted sequencing comprises targeted exon enrichment. 30. The method of claim 27, wherein the targeted sequencing is directed at known mutations associated with the tumor. 31. The method according to any one of claims 26-30, characterized in that sequencing extracellular DNA molecules from a biological sample of a subject comprises paired-end sequencing, random sequencing, or massively parallel sequencing. 32. The method according to any one of claims 20 to 31, wherein each of the one or more sequence variants comprises a single nucleotide variant. 33. The method according to any one of claims 20 to 32, wherein one or more sequence variants comprise one or more non-synonymous variants. 34. The method according to any one of claims 21 to 33, wherein the parameter is determined based on the concentration of one or more sequence variants at a plurality of loci. 35. The method according to any one of claims 21 to 33, wherein the parameter is the number of loci containing one or more sequence variants present in the second set of loci but not present in the first set of loci. 36. The method according to any one of claims 21 to 33, wherein determining the parameter includes, for each locus of the first set of loci, counting the read sequences that - 46 048297 are aligned to the locus and have a sequence variant at that locus; and parameter determination based on the corresponding counts. 37. The method of any one of claims 21 to 33, wherein the parameter is a weighted sum of the number of loci containing one or more sequence variants present in the second set of loci but not in the first set of loci, wherein the contribution of each locus is weighted based on a significance score assigned to the locus. 38. A method for monitoring, using a computer system, the progression and/or changes in one or more levels of cancer in a subject, comprising obtaining a first set of readable sequences for a first plurality of extracellular DNA molecules of the subject at a first point in time; detection of tumor mutations in the first set of read sequences; determining a first mutational load of the subject at a first point in time based on the identified tumor mutations in the first set of read sequences; obtaining a second set of readable sequences for a second plurality of extracellular DNA molecules of the subject at a second time; detection of tumor mutations in the second set of read sequences; determining a second mutational load of the subject at a second time point based on the identified tumor mutations in the second set of read sequences; comparing the second mutational load with the first mutational load; and monitoring progression and/or changes in one or more levels of cancer in the subject based on the comparison of the first mutational load and the second mutational load. 39. The method according to claim 38, wherein tumor mutations are identified in the first set of read sequences and/or in the second set of read sequences based on the alignment of the first set of read sequences and/or the second set of read sequences with a reference genome. 40. The method according to claim 38, wherein each of the one or more tumor-derived mutations comprises a single nucleotide variant. 41. The method of claim 38, wherein detecting tumor mutations further comprises obtaining readable sequences from one or more non-cancerous cells obtained from the subject; identifying non-tumor-associated mutations from sequence reads from one or more non-cancerous cells; and filtering out one or more non-tumor-associated mutations from the first set of sequence reads and/or the second set of sequence reads. 42. The method of claim 41, wherein non-neoplastic mutations are identified based on read sequences obtained from blood cells, leukocytes, buccal cells, hair follicles, or any combination thereof. 43. The method according to claim 42, wherein non-neoplastic mutations are identified based on read sequences obtained from hematopoietic cells. 44. The method according to any one of claims 38-43, wherein the first mutational load at a first time and/or the second mutational load at a second time is compared to a threshold, and wherein the predetermined relationship between the first mutational load and/or the second mutational load and the threshold indicates the presence of a certain level of cancer in the subject. 45. The method of any one of claims 38-43, further comprising identifying the level of cancer in the subject based at least in part on tumor-derived mutations present in extracellular DNA molecules at the first time point and/or the second time point. 46. The method according to any one of claims 38 to 43, wherein the first plurality of extracellular DNA molecules and the second plurality of extracellular DNA molecules are derived from blood plasma. 47. The method according to claim 38, wherein the first set of readable sequences and the second set of readable sequences are generated as a result of random sequencing of DNA molecules of the subject. 48. The method according to any one of paragraphs 38-43, characterized in that the first set of readable sequences is obtained as a result of targeted sequencing. 49. The method of claim 48, wherein the targeted sequencing comprises hybridization in solution. 50. The method of claim 49, wherein the targeted sequencing comprises targeted enrichment of the entire exome. 51. The method of claim 49, wherein the targeted sequencing comprises targeted exon enrichment. 52. The method of claim 49, wherein the targeted sequencing targets specific mutations associated with the tumor. 53. A machine-readable medium storing a series of instructions that, when executed, control a computer system to perform any of the methods according to paragraphs 1-52. 54. A computer system including processors programmed to implement any of the methods according to any of paragraphs 1-52. -