The present invention relates to methods for determining a copy number of a genetic region of interest.
There have been many important advances in the understanding of the hereditary susceptibility to cancer and other diseases. Identification of mutations associated with hereditary cancer syndromes and other diseases can lead to reduction in morbidity and mortality through targeted risk management options. The traditional approach for germline testing has been to test for a mutation in a single gene or a limited panel of genes using Sanger sequencing. With advances in next-generation sequencing technology and bioinformatics analysis, testing of multiple genes simultaneously (panel-based testing) at a cost comparable to traditional testing is possible. Panel based testing can provide better accuracy compared to traditional methods as well as improved diagnostic yield with analytical concordance between results from next generation sequencing (“NGS”) and the traditional Sanger method for detection of small mutations, such as single nucleotide variants, small deletions, and small insertions.
Despite advancements in NGS technologies in the past years, NGS panels possess analytical limitations which arise from sample preparation, sequencing, mapping, GC content of the targets, target size, and sequence complexity. These factors affect the relationship between read depth and copy number, which is key to copy number variant calls, and as a result the accuracy of the use of NGS techniques for use detecting copy number variants. Such limitations make it challenging for NGS technologies to be used for detection of copy number variants (CNV), such as exon-level copy number variations, larger insertion variants or deletion variants, or rearrangements. Scientific research has suggested that many cancers and complex diseases, such as schizophrenia, are related, at least in part, to copy number variants. Thus, higher accuracy, and accounting for the effect on noise in relating sequencing depth to copy number is especially desirable. To address this concern, some laboratories complement NGS with microarrays, which introduce their own level of complexity and bias to the call. Copy number variants provide valuable information needed for better understanding and characterization of the hereditary susceptibility to cancer and other disease. As such, a method of detecting CNVs with high accuracy is desirable.
Generally, the performance of genetic variant screen is assessed for concordance with known reference samples. Intrinsic variation in sequencing data coupled with insufficient quality control (QC) measures can imperil high CNV-calling accuracy. Assessment of the screen can be performed using a large number of positive controls with known genetic variants, and a performance statistic (such as sensitivity or specificity) for the screen can be determined. However, when a large number of positive controls are unavailable, such as for controls with rare genetic variant events, the performance of the genetic variant calling algorithm (i.e., a “caller”) or assay cannot be accurately assessed. While large numbers of positive controls having single nucleotide variants (SNVs) are commonly available, positive control samples having copy number variants are less frequent.
The disclosures of all publications, patents, and patent applications referred to herein are each hereby incorporated by reference in their entireties. To the extent that any reference incorporated by reference conflicts with the instant disclosure, the instant disclosure shall control.
Disclosed herein are methods of assessing a sample-specific performance of a copy number variant model, a method for determining a copy number of an interrogated segment within a region of interest, and a method for determining a copy number variant abnormality within a region of interest.
In some embodiments, a method of assessing the sample-specific performance of a copy number variant caller comprising a copy number variant model, comprising: parameterizing the copy number variant model based on real numbers of sequencing reads mapped to segments within a region of interest, from a test sample, to determine one or more copy number variant model parameters; generating a plurality of synthetic copy number variants, each synthetic copy number variant comprising a synthetic number of copies of one or more of the segments, wherein each synthetic number of copies is represented by a synthetic number of sequencing reads based on a real number of sequencing reads for a corresponding segment from the test sample; calling a number of copies of the one or more segments for the synthetic copy number variants using the copy number variant model, and the one or more determined copy number variant model parameters; determining a sample-specific performance statistic for the copy number variant caller based on differences between the called number of copies and the synthetic number of copies in the synthetic copy number variants; and assessing a sample-specific performance of the copy number variant caller based on the sample-specific performance statistic.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the synthetic number of sequencing reads for the one or more segments is generated by increasing, decreasing, or maintaining the real number of sequencing reads for the corresponding segments from the test sample in proportion to a predetermined number of copies of the one or more segments. In some embodiments, the predetermined number of copies is an integer number of copies. In some embodiments, the predetermined number of copies is a non-integer number of copies.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the synthetic number of sequencing reads is generated by sampling a binomial distribution with a success probability equal to m/x and a number of trials equal to the real number of sequencing reads at the corresponding segment from the test sample, wherein m is the synthetic number of copies of the segment in the synthetic copy number variant, and x is an assumed number of copies of the corresponding segment from the test sample.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the synthetic number of sequencing reads is generated by: sampling a number of sequencing reads as a negative binomial distribution with a success probability equal to m/x and a number of successes equal to the real number of sequencing reads at the corresponding segment from the test sample, wherein m is the synthetic number of copies of the segment in the synthetic copy number variant, and x is an assumed number of copies of the corresponding segment from the test sample, and adding the sampled number of sequencing reads to the real number of sequencing reads for the corresponding segment from the test sample. In some embodiments, the synthetic number of sequencing reads is generated by sampling a number of sequencing reads as an expectation of the negative binomial distribution.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the copy number variant model is a hidden Markov model. In some embodiments, the hidden Markov model comprises: (i) one or more hidden states comprising a copy number corresponding to an interrogated segment or a plurality of sub-segments within the interrogated segment; (ii) an observation state comprising the real or synthetic number of sequencing reads for the interrogated segment; (iii) a copy number likelihood model based on an expected number of real or synthetic sequencing reads for the interrogated segment. In some embodiments, the method comprises determining the copy number likelihood model. In some embodiments, parameterizing the hidden Markov model comprises adjusting the copy number likelihood model to fit the real number of sequencing reads mapped to the interrogated segment, from the test sample. In some embodiments, the copy number likelihood model comprises a distribution for two or more copy number states. In some embodiments, the copy number likelihood model comprises a negative binomial distribution, wherein the negative binomial distribution is not a Poisson distribution. In some embodiments, the expected number of real or synthetic sequencing reads is based on an average number of mapped sequencing reads at a segment corresponding to the interrogated segment across a plurality of samples, and an average number of mapped sequencing reads across the segments within the test sample, wherein the average number of mapped sequencing reads at the segment corresponding to the interrogated segment across the plurality of samples or the average number of mapped sequencing reads across the plurality of segments within the test sample is a normalized average. In some embodiments, the copy number likelihood model is adjusted to account for the presence of GC content bias. In some embodiments, the hidden Markov model comprises a transition probability of the copy number of the interrogated segment for a given copy number of a spatially adjacent segment. In some embodiments, the hidden Markov model comprises a plurality of transition probabilities of the copy number of a sub-segment in the plurality of sub-segments within the interrogated segment for a given copy number of a spatially adjacent sub-segment. In some embodiments, the transition probability accounts for an average length of a copy number variant. In some embodiments, the transition probability accounts for a prior probability of a copy number variant at the interrogated segment or a spatially adjacent segment. In some embodiments, the average length of a copy number variant or the probability of a copy number variant at the interrogated segment is determined based on observations in a human population.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, parameterizing the copy number variant model comprises accounting for one or more spurious capture probes. In some embodiments, accounting for one or more spurious capture probes comprises weighting the one or more observation states in the plurality of observation states with a spurious capture probe indicator. In some embodiments, the spurious capture probe indicator is determined using a Bernoulli process. In some embodiments, accounting for one or more of the capture probes being spurious comprises using expectation-maximization. In some embodiments, if a capture probe is determined to be spurious, sequencing reads derived from that capture probe is disregarded in the copy number variant model.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, parameterizing of the copy number variant model comprises accounting for noise in the number of mapped sequencing reads.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the copy number variant model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more copy number variant model parameters.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the copy number variant model is parameterized by solving a trust region Newton conjugate gradient algorithm.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the copy number variant model is iteratively parameterized using expectation-maximization.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the method comprises mapping the real sequencing reads from the test sample to the segments within the region of interest, and determining the real numbers of sequencing reads mapped to the segments.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the test sample is enriched using one or more direct targeted sequencing capture probes.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the method comprises calling a copy number of the one or more segments for the test sample.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the segments comprise spatially adjacent segments.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the sample-specific performance statistic is a limit of detection, sensitivity, specificity, precision, recall, accuracy, positive predictive value, or negative predictive value.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the sample-specific performance statistic is sensitivity or accuracy.
In some embodiments of the method of assessing the sample-specific performance of the copy number variant caller, the method comprises failing the test sample if the sample-specific performance of the copy number variant model is below a desired performance threshold.
Also described herein is a method for determining a copy number of an interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to the interrogated segment, wherein the test sequencing library is enriched using one or more direct targeted sequencing capture probes; (b) determining a number of sequencing reads mapped to the interrogated segment; (c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment; (d) building a hidden Markov model comprising: (i) one or more hidden states comprising a copy number corresponding to the interrogated segment or a plurality of sub-segments within the interrogated segment, (ii) an observation state comprising the number of sequencing reads mapped to the interrogated segment; and (iii) the copy number likelihood model; (e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and (f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Further described herein is a method for determining a copy number of an interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises the interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes; (b) determining a number of sequencing reads mapped to each spatially adjacent segment; (c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment; (d) building a hidden Markov model comprising: (i) a plurality of hidden states comprising a copy number for each of the spatially adjacent segments or a plurality of sub-segments within each of the spatially adjacent segments, (ii) a plurality of observation states comprising the number of sequencing reads mapped to each spatially adjacent segment, and (iii) the copy number likelihood model for each spatially adjacent segment; (e) parameterizing the hidden Markov model, comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and (f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Also described herein is a method for determining a copy number variant abnormality within a region of interest, comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to an interrogated segment within the region of interest, wherein the test sequencing library is enriched using one or more direct targeted sequencing capture probes; (b) determining a number of sequencing reads mapped to the interrogated segment; (c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment; (d) building a hidden Markov model comprising: (i) one or more hidden states comprising a copy number corresponding to the interrogated segment or a plurality of sub-segments within the interrogated segment, (ii) an observation state comprising the number of sequencing reads mapped to the interrogated segment; and (iii) the copy number likelihood model; (e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and (f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model; (g) determining a copy number variant abnormality based on the most probable copy number of the interrogated segment.
Further described herein is a method for determining a copy number variant abnormality within a region of interest, comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises an interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes; (b) determining a number of sequencing reads mapped to each spatially adjacent segment; (c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment; (d) building a hidden Markov model comprising: (i) a plurality of hidden states comprising a copy number for each of the spatially adjacent segments or a plurality of sub-segments within each of the spatially adjacent segments, (ii) a plurality of observation states comprising the number of sequencing reads mapped to each spatially adjacent segment, and (iii) the copy number likelihood model for each spatially adjacent segment; (e) parameterizing the hidden Markov model comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and (f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model; (g) determining a copy number variant abnormality based on the most probable copy number of the interrogated segment.
Also described herein is a method for determining a copy number of an interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to the interrogated segment, wherein the test sequencing library is enriched using one or more capture probes; (b) determining a number of sequencing reads mapped to the interrogated segment; (c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment; (d) building a hidden Markov model comprising: (i) one or more hidden states comprising a copy number corresponding to the interrogated segment or a plurality of sub-segments within the interrogated segment, (ii) an observation state comprising the number of sequencing reads mapped to the interrogated segment; and (iii) the copy number likelihood model; (e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment and accounting for one or more spurious capture probes, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and (f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Further described herein is a method for determining a copy number of an interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises the interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes; (b) determining a number of sequencing reads mapped to each spatially adjacent segment; (c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment; (d) building a hidden Markov model comprising: (i) a plurality of hidden states comprising a copy number for each of the spatially adjacent segments or a plurality of sub-segments within each of the spatially adjacent segments, (ii) a plurality of observation states comprising the number of sequencing reads mapped to each spatially adjacent segment, and (iii) the copy number likelihood model for each spatially adjacent segment; (e) parameterizing the hidden Markov model comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment and accounting for one or more spurious capture probes, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and (f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
In some embodiments of the methods described above, the one or more parameters of the copy number likelihood model comprises a dispersion of a number of mapped sequencing reads for the segment (di), an average number of mapped sequencing reads for the segment (μi), a dispersion of a number of mapped sequencing reads for the segments within the test sequencing library (dj), or an average number of mapped sequencing reads for the segments within the test sequencing library (μj).
In some embodiments of the methods described above, the method further comprises determining a most probable copy number of a section within the region of interest, wherein the section comprises a plurality of spatially adjacent segments comprising the interrogated segment.
In some embodiments of the methods described above, the copy number likelihood model comprises a distribution for two or more copy number states.
In some embodiments of the methods described above, the copy number likelihood model comprises a negative binomial distribution, wherein the negative binomial distribution is not a Poisson distribution.
In some embodiments of the methods described above, the expected number of sequencing reads is based on an average number of mapped sequencing reads at a corresponding segment across a plurality of sequencing libraries and an average number of mapped sequencing reads across a plurality of segments of interest within the test sequencing library, wherein the average number of mapped sequencing reads at a corresponding segment across a plurality of sequencing libraries or the average number of mapped sequencing reads across a plurality of segments of interest within the test sequencing library is a normalized average.
In some embodiments of the methods described above, the copy number likelihood model is adjusted to account for the presence of GC content bias. In some embodiments, the adjustment depends on the GC content of the capture probe corresponding to the interrogated segment or the GC content of the interrogated segment.
In some embodiments of the methods described above, the hidden Markov model comprises a transition probability of the copy number of the interrogated segment for a given copy number of a spatially adjacent segment. In some embodiments, the transition probability accounts for an average length of a copy number variant. In some embodiments, the transition probability accounts for a prior probability of a copy number variant at the interrogated segment or a spatially adjacent segment. In some embodiments, the average length of a copy number variant or the probability of a copy number variant at the interrogated segment are determined based on observations in a human population.
In some embodiments of the methods described above, the hidden Markov model comprises a plurality of transition probabilities of the copy number of a sub-segment in the plurality of sub-segments within the interrogated segment for a given copy number of a spatially adjacent sub-segment. In some embodiments, the transition probability accounts for an average length of a copy number variant. In some embodiments, the transition probability accounts for a prior probability of a copy number variant at the interrogated segment or a spatially adjacent segment. In some embodiments, the average length of a copy number variant or the probability of a copy number variant at the interrogated segment are determined based on observations in a human population.
In some embodiments of the methods described above, parameterizing the hidden Markov model comprises accounting for one or more spurious capture probes. In some embodiments, accounting for one or more spurious capture probes comprises weighting the one or more observation states in the plurality of observation states with a spurious capture probe indicator. In some embodiments, the spurious capture probe indicator is determined using a Bernoulli process. In some embodiments, accounting for one or more of the capture probes being spurious comprises using expectation-maximization. In some embodiments, if a capture probe is determined to be spurious, the likelihood information from that capture probe is disregarded in the copy number likelihood model.
In some embodiments of the methods described above, parameterizing of the hidden Markov model comprises accounting for noise in the number of mapped sequencing reads.
In some embodiments of the methods described above, accounting for noise in the number of mapped sequencing reads comprises adjusting the copy number likelihood model. In some embodiments adjusting the copy number likelihood model to account for the noise comprises an expectation-maximization step. In some embodiments, the expectation-maximization step comprises weighing a level of noise in the number of mapped sequencing reads from the test sequencing library. In some embodiments, the most probable copy number of the interrogated segment is not called if the noise in the number of mapped sequencing reads is above a predetermined threshold.
In some embodiments of the methods described above, sequencing reads from overlapping capture probes are merged.
In some embodiments of the methods described above, a Viterbi algorithm, a Quasi-Newton solver, or a Markov chain Monte Carlo is used to determine the most probable copy number of the interrogated segment.
In some embodiments of the methods described above, the method further comprises determining a confidence of the most probable copy number of the segment.
In some embodiments of the methods described above, the one or more parameters of the copy number likelihood model comprises a dispersion of a number of mapped sequencing reads for the segment (di), an average number of mapped sequencing reads for the segment (μi), a dispersion of a number of mapped sequencing reads for the segments within the test sequencing library (dj), or an average number of mapped sequencing reads for the segments within the test sequencing library (μj).
In some embodiments of the methods described above, the analytic first derivative gradient and second derivative analytical Hessian of the one or more parameters in the copy number likelihood model is solved using a trust region Newton conjugate gradient algorithm.
Also described herein is a computer system comprising a computer-readable medium comprising instructions for carrying out any one of the methods described above.
The methods described herein allow for accurate determination of the number of copies of an interrogated segment of a genome, such as a gene or gene segment. In some aspects, quality of the copy number variant caller is controlled by generating a quality control metric (such as sensitivity) on a sample-by-sample basis. Accurate copy number calling allows for better diagnosis of certain genetic anomalies, which aid in making important medical decisions.
Copy number variant callers can be used to screen a test sequencing library of copy number variants at one or more segments within a region of interest. These callers operate by building a copy number variant model, such as a hidden Markov model (HMM), which is parameterized for a test sample to develop one or more copy number variant (CNV) model parameters. The CNV model parameters can vary depending on sequencing depth, sample noise, capture probe efficiency, and/or other artifacts that arise during sequencing of the test sequencing library.
Synthetic copy number variants can be generated to assess the performance of a copy number variant caller (or copy number variant model used by the caller). The caller is used to call copy numbers at one or more segments within the synthetic copy number variant, and a performance statistic can be determined that provides an assessment of the caller. Parameterization of the copy number variant model is computationally intensive. Therefore, sample-specific performance assessment by parametrizing the CNV model for each synthetic copy number variant is not practical, particularly when the caller is used to screen a large number of samples. However, as described herein, the copy number variant model can be parameterized using sequencing reads from the test sample to determine sample-specific CNV model parameters. The synthetic copy number variants can be generated based on the sequencing reads from the test sample and, because the CNV model parameters are specific to the test sample and the synthetic copy number variants are generated based on the test sample, the determined sample-specific CNV model parameters can be used by the caller to call a copy number for a segment of the synthetic copy number variant without re-parameterizing the model. Accordingly, the methods described herein save substantial computing power while generating a reliable performance statistic for the assessment of the CNV model.
The copy number variant model, such as a hidden Markov model (HMM) can be parameterized using an analytic first derivative gradient and a second derivative Hessian of one or more parameters of a copy number likelihood model. In some embodiments, the first derivative gradient and a second derivative Hessian is solved using a trust region Newton conjugate gradient algorithm. An expectation-maximization (EM) step can be used to determine the copy number variant model parameters, which can include a plurality of optimization loops. In some embodiments, the EM parameterizes the CNV model to maximize the log-likelihood weighted by an expected copy number call.
Certain methods include the use of a hidden Markov model (HMM) to determine a most probable copy number for an interrogated segment of a test sequencing library. In some embodiments, the test sequencing library has been enriched using direct targeted sequencing (DTS) methods. DTS methods provide high resolution targeting of interrogated sequences, and the HMM caller described herein is substantially benefitted by the large amount of collected data for copy number calling. To further increase the accuracy of the HMM caller, sequencing depth artifacts that can arise from direct targeted sequencing methods can be accounted for. Such sequencing depth artifacts may include, for example, GC bias correction and determination of spurious probes. Additionally, the methods described herein provide for accurate copy number calling when the sequencing reads are produced from a noisy sequencing library.
Sequencing libraries derived from patient samples can be sequenced to obtain a number of sequencing reads. The copy number of a segment is related to the sequencing depth (that is, the number of sequencing reads or a normalized number of sequencing reads) at that segment. The present disclosure describes a method of using the sequencing depth at the segment to determine the presence of a copy number state at the segment. The sequencing depth may be obtained by determining the sequencing reads mapped to that segment. The sequencing depth may be obtained by determining the sequencing reads mapped to a capture probe corresponding to that segment. The method takes into consideration several factors associated with the sequencing technology to optimize the call so that it is more accurate.
Determining the mapped number of sequencing reads for a segment depends at least in part, to the actual copy number state of a segment. The vast majority of genetic regions in mammals are diploid and as such it is expected that, generally, there will be two copies of a genetic segment; however, this might not always be the case. For example, some regions of the genome are not diploid due to their location (being located on the Y chromosome for example). Other regions of the genome lose their diploidy as a result of functional specialization of some cells, such as immune cells, that result in genomic re-arrangements. However, regardless of these deviations from the norm, the copy number state of most genomic regions is expected to be two, and a deviation from a copy number state of two is expected to be reflected in the number of mapped sequencing reads.
Mapping sequencing reads to a segment can be preceded by one or more upstream steps, such as sample preparation, including fragmentation, formation of the sequencing library (for example, by ligating sequencing adapters to the nucleic acid molecules in the sequencing library), and sequencing the sequencing library. Noise in the sequencing depth at any of these upstream steps can introduce noise to the number of sequencing reads. Furthermore, the various capture probes in a capture probe library may not behave identically. For example, certain segments within the region of interest may not allow for ideal capture probe design, which can lead to spurious capture probes. Thus, using the determined number of mapped sequencing reads to determine the copy number state of a segment is less direct than recognizing the existing dependency between the copy number state of a segment and the determined number of mapped sequencing reads at the segment. The methods of the present invention allow for a copy number call of an interrogated segment within the region of interest to be made using a hidden Markov model, which is parameterized and optimized to account for the dependency between the number of mapped sequencing reads and the copy number state of the segment. The hidden Markov model can also account for various sources and levels of confounders. This method allows for a particularly effective and efficient process for determining copy numbers of interrogated segments or sub-segments within a region of interest, and for determining a copy number variant abnormality within the region of interest.
In some embodiments of the present invention, the sequencing library is enriched for the region of interest using direct targeted sequencing. Direct targeted sequencing uses a capture probe library comprising a plurality of capture probes that hybridize to nucleic acid molecules in the sequencing library. The capture probes are designed to hybridize to segments within the region of interest, and each capture probe has a corresponding segment. The region of interest is therefore determined by the capture probes used to enrich the sequencing library. The capture probes are extended using the nucleic acid molecules hybridized to the capture probe as a template. The extended capture probe can then be sequenced to obtain the sequence a portion (that is, the portion corresponding to the segment from the region of interest) of the nucleic acid molecule. Because the sequence of the capture probe itself is determined, the segment corresponding to the capture probe begins following the terminus of the capture probe. In some embodiments, the extended capture probe is amplified to obtain additional copies. Amplification of the extended capture probe can also introduce artifacts in the sequencing depth, which can be normalized as described herein. U.S. Pat. No. 9,309,556, entitled “Direct Capture, Amplification and Sequencing of Target DNA using Immobilized Primers”; U.S. Pat. No. 9,092,401, entitled “System and Method for Detecting Genetic Variation”; U.S. Patent App. No. 2014/0024541, entitled “Methods and Compositions for High-throughput Screening”; Myllykangas et al. “Efficient targeted resequencing of human germline and cancer genomes by oligonucleotide-selective sequencing.” Nat Biotechnol. 29(11):1024-7 (2011); and Hopmans et al., “A programmable method for massively parallel targeted sequencing.” Nucleic Acids Res. 42(10):e88 (2014) describe embodiments of direct targeted sequencing. Direct targeted sequencing need not be performed using surface-based methods, but can also be performed in solution.
In some embodiments, the sequencing library is enriched for the region of interest using methods other than direct targeted sequencing. For example, the sequencing library can be enriched using hybrid capture techniques, which include combining the sequencing library with a capture probe library to hybridize the capture probes with nucleic acid molecules in the sequencing library. The hybridized nucleic acid molecules can then be isolated from the rest of the sequencing library (for example, by using biotinylated capture probes and using streptavidin beads to separate the hybridized molecules). The nucleic acid molecules in the enriched sequencing library can then be sequenced. Because the nucleic acid molecules from the sequencing library are directly sequenced (as opposed to direct targeted sequencing methods), the capture probes do not necessarily correspond with specific segments within the region of interest. Instead, the sequencing depth at any given base within the region of interest can be determined by the number of sequencing reads at that base.
Provided herein are definitions, explanations, examples and descriptions that allow one skilled in the art to understand the scope of the methods provided, and enable one skilled in the art to practice the invention. It is to be understood that one, some or all of the properties of the various embodiments described herein may be combined to form other embodiments of the present invention. The section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.
As used herein, the singular forms “a,” “an,” and “the” include the plural reference unless the context clearly dictates otherwise.
Reference to “about” or “approximately” a value or parameter herein includes (and describes) variations that are directed to that value or parameter per se. For example, description referring to “about X” includes description of “X.”
The term “average” as used herein refers to either a mean or a median, or any value used to approximate the mean or the median, unless the context clearly indicates otherwise.
A “capture probe” refers to a DNA molecule or an RNA molecule which hybridizes to a nucleic acid molecule present in a sequencing library having segments with complementary sequences or sufficiently complementary sequences to allow for hybridization under normal hybridization condition.
“Copy number likelihood” refers to the likelihood of a copy number state at a segment or sub-segment of interest.
“Copy number likelihood model” refers to a statistical model used to determine a copy number likelihood given a number of mapped sequencing reads at that segment. The copy number likelihood model includes a statistical distribution for each copy number state covered by the model, and each distribution reflects the probability that the copy number state is correct for a given number of mapped sequencing reads.
“Copy number variant” or “CNV” refers to a deviation in the copy number state from a wild type. A “wild-type” as used herein refers to a predetermined copy number state for a particular segment that is considered normal. The determination of what is “wild-type” can be made based on human, mammal or other animal population data. The determination of what a “wild-type” is can also be made based on reference runs, internal experiments and data generated from such experiments.
A “direct targeted sequencing capture probe” is a capture probe that is used to enrich a sequence from a sequencing library using direct targeted sequencing.
An “interrogated segment” refers to a segment within a region of interest for which a copy number variant model is used to determine the copy number state. The interrogated segment can be divided into sub-segments that may be as small as one base pair, but no longer than the length of the interrogated segment.
A “noisy sequencing library” or “noise” from a sequencing library refers to a sequencing library that generates poor data across one or more capture probes.
A “number of sequencing reads” as used herein refers to an absolute number of sequencing reads or a normalized number of sequencing reads.
A “real sample” refers to a nucleic acid sequence or sequencing reads originating from a nucleic acid sequence that originates from a physical sample subjected to genetic sequencing without the sequence, sequencing reads, or number of sequencing reads being altered. A “real reference sample” refers to a real sample that is compared to a synthetic sample (e.g., a synthetic copy number variant) by the genetic variant caller.
A “real sequencing read” refers to a sequencing read that originates from a real sample without alteration of the sequence. A “number of real sequencing reads” refers to an absolute number of real sequencing reads or a normalized number of sequencing reads, but does not refer to a number of sequencing reads that has be altered to reflect an increase in a number of copies of any segment or region of interest.
A “segment” refers to a nucleotide chain comprising two or more bases. A segment can be sub-divided into one or more “sub-segments.” A “sub-segment” can be as small as one nucleotide but not longer than the segment in which is it is located. A region of interest can be divided into one or more segments. The segments can be, but need not be, contiguous. Therefore a region of interest can optionally include non-contiguous sub-regions. The segments can be of the same length or of different lengths. Two or more segments within a region of interest can be grouped to make a section within the region of interest. The segments that make up a section within the region of interest may be, but need not be contiguous.
A “spurious capture probe” refers to a capture probe that generates artifacts in the number of sequencing reads that are unrelated to copy number. The artifacts can be due to sub-par sequencing reads, inconsistent sequencing reads, sequencing reads of length that fall below a predetermined level, a number of sequencing reads that fall below a predetermined level, or displays poor quality when compared to other capture probes.
“Spatially adjacent segments” refer to a set of sequential segments that are located within the same chromosome, but need not be contiguous. That is, two spatially adjacent segments can be separated by a number of intervening nucleotides but not by intervening segments outside of the set of spatially adjacent segments. The copy number of any intervening nucleotides if the two spatially adjacent segments are not contiguous may be inferred through the hidden Markov model. “Spatially adjacent capture probes,” including “spatially adjacent direct targeted sequencing capture probes,” refer to capture probes that correspond to the spatially adjacent segments.
The term “synthetic copy number variant” refers to an artificial sample generated using real sequencing reads or a real number of sequencing reads from a real sample with an increase or decrease in a number of copies of one or more segments within a region of interest relative to the real sample.
A “synthetic number of copies” refers to a number of copies of a segment within a region of interest in the synthetic copy number variant, and can be an increase, decrease, or the same as the number of copies relative to the real sample. As the synthetic copy number variant need not have an altered number of copies in each segment and may include a wild-type number of copies of one or more segments, the synthetic number of copies for one or more segments of the synthetic copy number variant may be the same as the real number of copies of the segment.
A “synthetic number of sequencing reads” refers to a number of sequencing reads that is used to represent a synthetic number of copies of a segment within a region of interest. The synthetic number of sequencing reads for a segment may be increased, decreased, or maintained relative to the real number of sequencing reads for the corresponding segment.
It is understood that aspects and variations of the invention described herein include “consisting” and/or “consisting essentially of” aspects and variations.
Where a range of values is provided, it is to be understood that each intervening value between the upper and lower limit of that range, and any other stated or intervening value in that stated range, is encompassed within the scope of the present disclosure. Where the stated range includes upper or lower limits, ranges excluding either of those included limits are also included in the present disclosure.
The present disclosure provides methods to determine the copy number of an interrogated segment (or sub-segment of the interrogated segment) of a region of interest, or a copy number variant abnormality within a region of interest, based on the determined number of mapped sequencing reads at that segment. The methods include determining a copy number likelihood model based on an expected number of mapped sequencing reads for one or more copy number states. A first derivative gradient and second derivative Hessian of one or more parameters of the copy number likelihood model, along with Expectation-Maximization (EM), can be used to enable latent parameter estimation and optimization of the model. The first derivative gradient and second derivative Hessian can be solved using, for example, a trust region Newton conjugate gradient algorithm. Several additional steps and adjustments to the model can be used to account for other factors that affect the relationship between a copy number and the number of mapped sequencing reads. The information is used to parameterize the hidden Markov model which can then be used to determine a most probable copy number state at an interrogated segment. Methods for building the copy number likelihood model, Expectation-Maximization implementation, adjustment to the model to account for multiple factors, parameterization of the hidden Markov model as well as methods of solving the various steps and the model overall, are provided generally below.
In brief, the methods for determining a copy number of a segment or sub-segment can include (1) determining the number of sequencing reads mapped to an interrogated segment; (2) building and parameterizing a hidden Markov model by determining a copy number likelihood model; and (3) determining a most probable copy number of the interrogated segment (or a sub-segment of the interrogated segment) using the parameterized hidden Markov model. The hidden Markov model is parameterized using a first derivative gradient and second derivative Hessian of one or more parameters of a copy number likelihood model, along with Expectation-Maximization (EM), which may be solved using a trust region Newton conjugate gradient algorithm. In some embodiments, the methods provided herein also include steps to refine the model by accounting for confounding effects that may arise during the process.
In some embodiments of the methods described herein, a hidden Markov model is used to determine the most probable copy number state of a segment. The hidden Markov model can include: a hidden layer which comprises the copy number state of a segment of interest; an observation layer, which comprises the mapped number of sequencing reads; a transition probabilities between the copy number state in the hidden layer and mapped number of sequencing reads (probability inter-layers); and a transition probabilities of a copy number state of a segment given the copy number state of a preceding adjacent segment (probability intra-hidden layer).
In some embodiments, the methods described herein include mapping a plurality of sequencing reads generated from a test sequencing library to one or more segments, such as an interrogated segment. In some embodiments, the methods described herein can include mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of segments (which may be spatially adjacent), wherein the plurality of segments includes the interrogated segment. The sequencing library is enriched for a region of interest, such as by direct targeted sequencing. The mapped sequencing reads can be counted to determine a number of sequencing reads that are mapped to the interrogated segment or the spatially adjacent segments.
In some embodiments, the segments are located within the same chromosome. In some embodiments, the segments are located within the same chromosome region. In some embodiments, the segments are located within the same gene. In some embodiments, the segments are located within the same region of interest. In some embodiments, the segments are located within the same portion within the region of interest.
The sequencing library can be sequenced to generate the plurality of sequencing reads, which can be mapped to a region of interest. The sequencing library includes a plurality of nucleic acid fragments, which can be isolated from bodily fluids such as blood, plasma, saliva, urine or from tissue or cultured cells. The nucleic acid fragments can be from an animal. The nucleic acid fragments can be from a mammal, for example, from a human. In a preferred embodiment the test sequencing library includes a plurality of nucleic acid fragments isolated from a patient. The nucleic acid molecules in the sequencing library can be ligated to sequencing adapters, which may aid alignment in certain sequencing methods. For example the adapters may be indexed, and the indexing may be used to aid alignment of the sequences. The sequencing library can be enriched (such as through direct targeted sequencing) for the region of interest, either before or after ligating the nucleic acid molecules to the sequencing adapters.
The nucleic acid fragments in the test sequencing library may be RNA or DNA nucleic acid fragments. The nucleic acid fragments may be cell-free DNA. In some embodiments, the cell-free DNA comprises fetal cell-free DNA. In some embodiments, the cell-free DNA comprises circulating tumor cell-free DNA.
The nucleic acid fragments in the sequencing library include the region of interest. The region of interest can be a full genome or any portion of the genome. In some embodiments the region of interest comprises one or more chromosomes. In some embodiments the region of interest comprises one or more genes of interest (such as 2 or more, 3 or more, 4 or more, 5 or more, about 10 or more, about 15 or more, about 20 or more, about 30 or more, about 40 or more, about 50 or more, about 75 or more, about 100 or more, about 150 or more, about 200 or more, about 250 or more genes, about 300 or more, about 350 or more, about 400 or more, about 450 or more, about 500 or more, about 550 or more, about 600 or more, about 650 or more, about 700 or more, about 750 or more, about 800 or more, about 850 or more, about 900 or more, about 950 or more, or about 1000 or more). The one or more genes of interest may be any gene associated with a disease. The one or more genes of interest may include any gene associated with a hereditary disease. The one or more genes of interest may include a gene associated with a form of cancer, such as a hereditary cancer. In some embodiments, the region of interest comprises one or more exons (such as 2 or more, 3 or more, 4 or more, 5 or more, 10 or more, 15 or more, 20 or more, 30 or more, 40 or more, 50 or more, 75 or more, 100 or more, 150 or more, 200 or more, or 250 or more, 500 or more, 1000 or more, or 2000 or more exons). In some embodiments, the region of interest comprises a gene or a portion of a gene, an exon, or a portion of an exon, selected from the group consisting of APC, ATM, BARD1, BMPR1A, BRCA1, BRCA2, BRIPI, CDH1, CDK4, CDKN2A, CHEK2, EPCAM, GREM1, MEN1, MLH1, MRE11A, MSH2, MSH6, MUTYH, NBN, PALB2, PMS2, POLD1, POLE, PTEN, RAD50, RAD51C, RAD51D, RET, SDHA, SDHB, SDHC, SMAD4, STK11, TP53, VHL, PEX10, MTHFR, ALPL, HMGCL, DHDDS, PPT1, MPL, MMACHC, POMGNT1, CPT2, ALG6, RPE65, ACADM, DPYD, AGL, SLC35A3, DBT, PHGDH, CTSK, NTRK1, NPHS2, LAMC2, LAMB3, USH2A, PHYH, ERCC6, PCDH15, LIPA, HOGA1, OAT, TH, HBB, SMPD1, TPP1, KCNJ11, ABCC8, USH1C, RAG2, RAPSN, TMEM216, PYGM, BBS1, PC, TCIRG1, CPT1A, DHCR7, MYO7A, MED17, PTS, SLC37A4, HYLS1, PFKM, BBS10, GNPTAB, PAH, MMAB, ACADS, PUS1, GJB2, GJB6, SGCG, SACS, ATP7B, CLN5, PCCA, TGM1, ZFYVE26, VSX2, NPC2, GALC, SERPINA1, VRK1, TECPR2, SLC12A6, IVD, CAPN3, CLN6, NR2E3, HEXA, MPI, FAH, MESP2, BLM, GNPTG, MEFV, PMM2, CLN3, BBS2, TAT, CYBA, FANCA, VPS53, ASPA, CTNS, ACADVL, ALDH3A2, PEX12, NAGLU, G6PC, SGCA, MKS1, DNAI2, GALK1, GAA, SGSH, NPC1, LAMA3, LOXHD1, MCOLN1, MAN2B1, GCDH, NPHS1, BCKDHA, OPA3, FKRP, HADHA, LRPPRC, FAM161A, ATP6V1B1, DYSF, ALMS1, NEB, CERKL, CPS1, BCS1L, CYP27A1, COL4A4, COL4A3, AGXT, NDUFAF5, ADA, RTEL1, HLCS, CBS, AIRE, TRMU, MLC1, TYMP, ARSA, SUMF1, XPC, BTD, GLB1, AMT, GBE1, HGD, PCCB, HPS3, CLRN1, BCHE, IDUA, EVC2, EVC, SEPSECS, SGCB, MTTP, BBS12, MMAA, AGA, F11, NDUFS6, DNAH5, NDUFS4, ERCC8, HEXB, HSD17B4, SLC22A5, SLC26A2, SGCD, PROP1, ADAMTS2, PEX6, MUT, PKHD1, EYS, SLC17A5, BCKDHB, RARS2, LAMA2, ARG1, PEX7, ASL, PEX1, SAMD9, ASNS, SLC26A4, DLD, CFTR, CLN8, STAR, HGSNAT, TTPA, PEX2, CNGB3, VPS13B, CYP11B1, CYP11B2, GLDC, DNAI1, GALT, RMRP, GNE, GRHPR, VPS13A, FANCC, XPA, ALDOB, FKTN, IKBKAP, ASS1, RS1, NR0B1, DMD, OTC, IL2RG, ATP7A, CHM, GLA, COL4A5, IDS, MTM1, ABCD1, or a combination thereof.
The region of interest can be divided into a plurality of segments. Each segment can be further divided into sub-segments. A sub-segment may be 1 or more nucleotides in length. The segments within the region of interest may be but need not be contiguous. For example, in some embodiments, the region of interest comprises 1 or more non-contiguous segments, 2 or more non-contiguous segments, 3 or more non-contiguous segments, 4 or more non-contiguous segments, 5 or more non-contiguous segments, 10 or more non-contiguous segments, 25 or more non-contiguous segments, 50 or more non-contiguous segments, 100 or more non-contiguous segments, 150 or more non-contiguous segments, 200 or more non-contiguous segments, 250 or more non-contiguous segments, 300 or more non-contiguous segments, 350 or more non-contiguous segments, 400 or more non-contiguous segments, 450 or more non-contiguous segments, 500 or more non-contiguous segments, 550 or more non-contiguous segments, 600 or more non-contiguous segments, 650 or more non-contiguous segments, 700 or more non-contiguous segments, 750 or more non-contiguous segments, 800 or more non-contiguous segments, 850 or more non-contiguous segments, 900 or more non-contiguous segments, 950 or more non-contiguous segments, or 1000 non-contiguous segments. In some embodiments, each of the non-contiguous segments comprises 1 or more contiguous bases, 2 or more contiguous bases, 3 or more contiguous bases, 4 or more contiguous bases, or 5 or more contiguous bases. For example, in some embodiments each of the non-contiguous segments comprises 1 to about 20 contiguous bases (such as 1 to about 10 contiguous bases, or about 1 to about 5 contiguous bases). In some embodiments the region of interest comprises 1 or more contiguous segments, 2 or more contiguous segments, 3 or more contiguous segments, 4 or more contiguous segments, 5 or more contiguous segments, 10 or more contiguous segments, 25 or more contiguous segments, 50 or more contiguous segments, 100 or more contiguous segments, 150 or more contiguous segments, 200 or more contiguous segments, 250 or more contiguous segments, 300 or more contiguous segments, 350 or more contiguous segments, 400 or more contiguous segments, 450 or more contiguous segments, 500 or more contiguous segments, 550 or more contiguous segments, 600 or more contiguous segments, 650 or more contiguous segments, 700 or more contiguous segments, 750 or more contiguous segments, 800 or more contiguous segments, 850 or more contiguous segments, 900 or more contiguous segments, 950 or more contiguous segments, or 1000 contiguous segments. In some embodiments, each of the contiguous segments comprises 1 or more contiguous bases, 2 or more contiguous bases, 3 or more contiguous bases, 4 or more contiguous bases, or 5 or more contiguous bases. For example, in some embodiments each of the non-contiguous segments comprises 1 to about 20 contiguous bases (such as 1 to about 10 contiguous bases, or about 1 to about 5 contiguous bases). In some embodiments the region of interest comprises a combination of non-contiguous and contiguous segments. In some embodiments, the region of interest comprises only one segment. In some embodiments the region of interest comprises at least one segment. In some embodiments the region of interest comprises at least two segments. In some embodiments the region of interest comprises at least two segments which are adjacent. In some embodiments one segment within a first region of interest may be adjacent to a segment in a second region of interest adjacent to the first region of interest.
The region of interest may be enriched with one or more capture probes. The reference location of the capture probes with respect to a region of interest is known. For example, the capture probes comprise a reference sequence that corresponds to pre-determined probe coordinates. In some embodiments the region of interest is divided into segments based on the location of capture probes (that is the capture probe corresponds with a segment). The capture probe comprises the reference sequence that corresponds to the probe coordinates. For example, the first nucleotide of a segment may coincide with the first nucleotide of a sequence hybridizing to the 3′end of a capture probe. In some embodiments the first nucleotide of a segment coincides with the first nucleotide of a sequence hybridizing to the 5′ end of a capture probe. In some embodiments the region of interest comprises two spatially adjacent segments. A segment within a region of interest may be divided into sub-segments. A sub-segment may be as small as one nucleotide can be as long as the segment. Sub-segments may overlap. For example a first sub-segment may be the first nucleotide of a segment plus one downstream nucleotide. A second sub-segment may comprise the first sub-segment plus an additional downstream nucleotide. In some embodiments, a segment of n nucleotides of length comprises n−1 sub-segments, wherein each subsequent sub-segment is 1 nucleotide longer than the previous. In some embodiments a segment of n nucleotides of length comprises n sub-segments, wherein each sub-segment is 1 nucleotide in length.
The region of interest comprises at least one interrogated segment. The interrogated segment is a segment for which it is desirable to know the copy number. The copy number state of an interrogated segment is an unknown state and solving the hidden Markov model determines the most probable copy number of an interrogated segment. Like other segments, an interrogated segment may be divided into sub-segments. In some embodiments the first nucleotide of an interrogated segment coincides with the first nucleotide of a sequence hybridizing to the 5′ end of a capture probe. In some embodiments the first nucleotide of an interrogated segment coincides with the first nucleotide of a sequence hybridizing to the 3′end of a capture probe. In some embodiments the interrogated segment comprises the sequence spanning two spatially adjacent capture probes. In a preferred embodiment the interrogated segment comprises the nucleotide sequence between two adjacent capture probes, with the first nucleotide of the sequence being the first nucleotide hybridizing to the 5′ end or the 3′ end of the capture probe and the last nucleotide of the segment being contiguous to the first nucleotide hybridizing to the 5′ end or the 3′ end of a spatially adjacent probe.
The test sequencing library can be sequenced using next generation sequencing to generate the sequencing reads. Next generation sequencing technologies are well known in the art. The test sequencing library can be sequenced using a high-throughput sequencer, such as an Illumina HiSeq2500, Illumina HiSeq3000, Illumina HiSeq4000, Illumina HiSeqX, Roche 454, PacBio Sequel System PacBio RS II, or Life Technologies Ion Proton sequencing systems can also be used. Other methods of sequencing are known in the art.
In some embodiments, the sequencing library is enriched with one or more capture probes by direct targeted sequencing. In direct targeted sequencing, capture probes hybridize specific target regions of nucleic acid molecules from within a sequencing library. This method enables enrichment of target regions and allows subsequent sequencing efforts to focus on relevant genomic regions or transcripts of interest. Enriching the target regions with capture probes for the region of interest allows for more efficient high throughput sequencing of the region of interest. The efficiency keeps the overall costs of sequencing test sequencing libraries down while maintaining or increasing the sensitivity and specificity of a diagnostic test or screen. The capture probes can be selected based on the region of interest such that those nucleic acid molecules in the sequencing library containing a portion of the region of interest hybridize to the capture probes and can be enriched, whereas those nucleic acid molecules in the sequencing library that do not contain a portion of the region of interest do not hybridize to the capture probes and are not enriched.
In direct targeted sequencing, capture probes that hybridize to a target sequence adjacent to the corresponding segment within the region of interest are combined with the sequencing library, thereby hybridizing the capture probes to the nucleic acid molecules comprising to the target sequence. In direct targeted sequencing methods, the capture probe is extended using the nucleic acid molecule as a template, and the extended capture probe is sequenced. Since the extended capture probe (or amplified copies of the extended capture probe) itself is sequenced, the sequence of the capture probe is not interpreted as the sequence arising from the test sequencing library, although it can be used to aid sequence alignment.
Other methods for enriching sequencing libraries using capture probes are generally known in the art, and can include hybrid capture techniques (e.g., using biotinylated capture probes), and PCR amplification using capture probes as PCR primers.
In some embodiments, hybrid capture techniques are used to enrich the region of interest by combining capture probes that are substantially complementary to a portion of the region of interest with the sequencing library, thereby hybridizing the capture probes to nucleic acid molecules comprising the portion of the region of interest. The nucleic acid molecules that hybridize to the capture probes can be isolated from non-hybridized nucleic acid molecules (for example, by pull-down methods). The hybridized complex can be denatured and the enriched nucleic acid molecules from the sequencing library can be sequenced. In some embodiments, the enriched nucleic acid molecules are re-enriched in a second (or more) round of hybridization to the capture probes, isolation and denaturation before being sequenced. Optionally, the nucleic acid molecules in the sequencing library can be amplified (for example, by PCR) either before or after enrichment.
In some embodiments, one or more of the capture probes are attached to an additional oligonucleotide (such as a primer binding site or other specialized nucleic acid segment). In some embodiments, the capture probes in the capture probe library are DNA oligonucleotides, RNA oligonucleotides, or a mixture of DNA oligonucleotides and RNA oligonucleotides. In some embodiments the capture probes are about 10-100 bases in length. In some embodiments the capture probes are about 20-60 bases in length. In some embodiments the capture probes are about 30-50 bases in length. In some embodiments the capture probes are 40 bases in length.
The number of capture probes in the capture probe library can depend on the size of the region of interest, as a larger region of interest generally requires a larger number of capture probes for adequate coverage. In some embodiments, the capture probe library comprises about 10 or more unique capture probes (such as about 50 or more, about 100 or more, about 250 or more, about 500 or more, about 1000 or more, about 2500 or more, about 5000 or more, about 10,000 or more, about 25,000 or more, about 50,000 or more, about 100,000 or more, or about 200,000 or more) unique capture probes.
Sequencing the enriched sequencing library generates a plurality of sequencing reads. In order to determine the sequencing depth for a segment or sub-segment, the number of sequencing reads mapped to that segment is determined. Sequencing reads can be mapped, for example, by aligning the sequencing reads (or a portion of the sequencing reads) to a reference sequence, or by assigning the sequencing read to a segment based on a portion of the sequencing read.
In some embodiments, the sequencing reads are mapped by aligning the sequencing reads (or a portion of the sequencing reads) to a reference sequence. For example, sequencing reads resulting from direct targeted sequencing can include a capture probe portion (that is, the portion of the sequencing read that is attributable to the capture probe itself) and a segment portion (that is, the portion of the sequencing read that is attributable to the segment targeted by and associated with the capture probe). In some embodiments, the segment portion is aligned with the reference sequence, the capture probe portion is aligned with the reference sequence, or the capture probe portion and the segment portion are aligned with the reference sequence. The reference sequence includes the region of interest pre-dived into segments. Therefore, the sequencing reads aligned to the reference sequence can be aligned to a corresponding segment, and the aligned sequencing reads are assigned or “mapped” to that segment.
In some embodiments, the sequencing reads are mapped by assigning the sequencing read to a segment based on a portion of the sequencing read. In such an embodiment, it is not necessary to align the sequencing read to a reference sequence. Because the capture probes each correspond with a segment, and the corresponding segment is known by the design of the capture probe, sequencing reads that contain a sequence of the capture probe (or its complement) can be assigned (or “mapped”) to the corresponding segment.
In some embodiments, the sequencing depth may be obtained by determining the sequencing reads mapped to that segment. In some embodiments the sequencing depth may be obtained by determining the sequencing reads mapped to a capture probe corresponding to that segment.
In some embodiments, two or more capture probes overlap (that is, the capture probes can hybridize to overlapping sequences within the region of interest). The two or more capture probes may overlap by about 0%-10%, about 10-20%, about 20%-30%, about 30%-40%, about 40%-50%, about 50%-60%, about 60%-70%, about 70%-80%, about 80%-90%, or about 90%-99% of the length of the probe. In some embodiments, two or more capture probes overlap 100%. In some embodiments, the number of sequencing attributable to two or more capture probes correlate with each other. Overlapping or correlated capture probes can be accounted for by merging (i.e., adding together) the number of sequencing reads attributed to the overlapping or correlated capture probes.
Once the plurality of sequencing reads are mapped to the interrogated segment or a plurality of spatially adjacent segments (including the interrogated segment), the number of sequencing reads mapped to the interrogated segment or the spatially adjacent segments (including the interrogated segment) can be determined by counting the number of sequencing reads that have been assigned to the segment.
The copy number likelihood model may be any statistical model that can be used to determine the likelihood of observing a number of sequencing reads mapped at a segment given the copy number state of the segment. An initial copy number likelihood model refers to the model where the parameters for the model have been defined, but before optimizing it. In a preferred embodiment the copy number likelihood model includes one or more likelihood distributions for an expected number of mapped sequencing reads given a copy number state. That is, each likelihood distribution corresponds to a copy number state. For example the copy number likelihood model may comprise a likelihood distribution of an expected number of sequencing reads given a copy number state of 1, a likelihood distribution of an expected number of sequencing reads given a copy number state of 2, a likelihood distribution of an expected number of sequencing reads given a copy number state of 3, and a likelihood distribution of an expected number of sequencing reads given a copy number state of 4. The copy number likelihood model need not comprise a likelihood distribution for each possible copy number state, but comprises at least one likelihood distribution. Similarly, the copy number likelihood model may comprise distributions for copy number states greater than 4, such as a copy number state of 5, of 6, of 7 or of 8. In some embodiments the distributions comprised in the copy number likelihood model are Poison distributions. In some embodiments the distributions comprised in the copy number likelihood model are binomial distributions. In some embodiments the copy number likelihood model comprises negative binomial distributions. For example, in some embodiments the copy number likelihood model comprises one or more negative binomial distributions (or one or more negative binomial distributions, wherein the negative binomial distribution is not a Poisson distribution) for expected mapped sequencing reads for interrogated segment i in test sequencing library j for copy number states ci,j.
The likelihood distribution of the copy number likelihood model can be further characterized by a mean (μ) and a dispersion (d). The mean and the dispersion of the likelihood distribution are optimized by using a determined expected number of sequencing reads, at segment i (that is, using the same capture probe) by sequencing the test sequencing library j at a plurality of segments (that is, using a capture probe library) and by setting a copy number state at the segment i for sequencing library j. The expected number of sequencing reads is based on at least three factors: the average number of mapped sequencing reads for the segment across a plurality of sequencing libraries, the average number of mapped sequencing reads for the test sequencing library across a plurality of segments, and the local copy number state of the segment. The mean of the distribution can be set as: μ=ci,jμiμj
wherein μi is the average number of mapped sequencing reads for segment i across Ns sequencing libraries, μj is the average number of mapped sequencing reads for the test sequencing library j across Np segments, and ci,j is the copy number state at segment i for test sequencing library j and ki,j is the determined number of sequencing reads at segment i for test sequencing library j; wherein μi and/or μj are normalized.
Formally:
The copy number likelihood model is set by determining distributions from an expected number of sequencing reads for different copy number states then maximized for a most probable ci,j given the number of actual mapped sequencing reads at the segment.
For the majority of genes the expected copy number (i.e., “wild-type”) is assumed to be two (i.e., diploid). This is not necessarily always the case. For example, for genes on the Y chromosome the expected copy number (i.e., “wild-type”) should be assumed to be 1. Considering this relationship, in some embodiments the copy number likelihood distribution for any given copy number state is centered at an average
wherein μi is the average number of mapped sequencing reads for segment i across Ns sequencing libraries, μj is the average number of mapped sequencing reads for the test sequencing library j across Np segments, and c is the number of copies for the given copy number likelihood distribution, wherein μi and/or μj is a normalized average number of mapped sequencing reads. The number of mapped sequencing reads for segment i in a given sequencing library can be normalized by dividing the number of mapped sequencing reads at segment i within the sequencing library by the average number of mapped sequencing reads across Np segments within that sequencing library.
The copy number likelihood distribution also includes a dispersion (d), estimated for segment i as:
wherein σi2 is the variance of the number of mapped sequencing reads for the plurality of sequencing libraries. As further described herein, the dispersion of the copy number likelihood distribution can include components for the both the segment i (that is, the dispersion from the noise due to the capture probe at segment i) and across segments in the test sequencing library j.
The copy number likelihood distribution can be a Poisson distribution, a binomial distribution, a negative binomial distribution (such as a generalized Poisson negative binomial distribution or a negative binomial distribution the is not a Poisson distribution), or any other suitable distribution. It has been found that a negative binomial distribution, wherein the negative binomial distribution is not a Poisson distribution is particularly useful for determining the copy number likelihood distributions.
A hidden Markov model allows for the determination of a most probable copy number (a hidden state) from the number of mapped sequencing reads (an observation state). Generally, there are four main parameters in the hidden Markov model: one or more hidden states, one or more observation states, one or more emission probabilities from the hidden states to the observation states, and the transition probabilities between the hidden states. Provided herein are methods of building the hidden Markov model and parameterizing the hidden Markov model. Also, provided herein are methods of training the hidden Markov model using an incomplete data set. Also provided herein are methods of optimizing the hidden Markov model by optimizing parameters in the hidden Markov model to account for variables that affect the emission probabilities between the hidden states and the observation states. Specifically, provided below are methods and explanations on the layers of the hidden Markov model; the transition probabilities of the Markov model; the copy number likelihood model; using expectation-maximization to parameterize the hidden Markov model; adjusting the hidden Markov model to account for a number of latent variables; solving the hidden Markov model.
An exemplary hidden Markov model that can be used with the disclosed methods is illustrated in
In some embodiments a hidden Markov model comprises only one hidden state and a corresponding observation state. In some embodiments the hidden state corresponds to the copy number state of a segment and the observation state corresponds to the mapped number of sequencing reads at that segment. In some embodiments the hidden Markov model comprises a plurality of hidden states and a plurality of observation states. In some embodiments the plurality of the hidden states corresponds to the copy number states at a plurality of segments and the plurality of observation states corresponds to the number of mapped sequence reads at the plurality of segments. In some embodiments, each segment within a region of interest corresponds to a capture probe for the region of interest. In some embodiments, two adjacent hidden states correspond to two spatially adjacent segments within the region of interest.
The segments may be divided in sub-segments, as previously described herein. In some embodiments the hidden states correspond to the copy number of the sub-segments. The sub-segments do not include a mapped number of sequencing reads independent of the mapped number of sequencing reads for the parent segment (that is, the segment to which the sub-segment is a member). In some embodiments, the mapped number of sequencing reads for the segment is attributed to each sub-segment within the segment. In some embodiments, the sub-segment includes a hidden state (i.e., a copy number), but the mapped number of sequencing reads is only attributed to the first sub-segment of the segment. This is illustrated in
The copy number state of a segment is related to the number of sequencing reads mapped to that location. Determining a copy number state of a segment or sub-segment (which can be denoted as ci,j) given a number of mapped sequencing reads (which can be denoted as ki,j) for segment (or sub-segment) i in test sequencing library j allows for calling of a copy number of that segment or sub-segment. The probability for a given copy number state being the correct copy number depends at least on the number of mapped sequencing reads. In Bayesian statistics, the posterior probability of ci,j given ki,j (that is, p(ci,j|ki,j)) can be determined using a copy number likelihood distribution. While posterior probability is a probability of a parameter given some data; a likelihood model is the probability of the data, given the parameter. In this case, the posterior probability is the probability of the copy number state of a segment or sub-segment given the number of sequencing reads mapped at that segment or sub-segment (that is, p(ci,j|ki,j)), whereas the copy number likelihood model is the likelihood of observing a number of sequencing reads mapped at a segment given the copy number state of the segment (that is, p(ki,j|ci,j)). Since p(ci,j|ki,j) cannot be directly determined, the copy number likelihood model p(ki,j|ci,j) can be used to parameterize the hidden Markov model, which can be used to solve for the posterior probability p(ci,j|ki,j). The following discusses the copy number likelihood model as a negative binomial distribution, but it is understood that the similar aspects would apply for other distribution forms. In some embodiments, the copy number likelihood model can be defined as:
p(ki,j|ci,j)=NegBinom(ki,j|μc,i,j=ci,jμiμj;d=di)
wherein ki,j is the number of mapped sequencing reads at segment i for the test sequencing library j.
The negative binomial distribution is parameterized to best fit the data. In its simplest form the copy number likelihood model is a negative binomial model. However, depending on the data generated, a different type of distribution may fit the data better and may be better suited. The general aspects of this invention would apply to models comprising different statistical distributions.
The transition probability for a copy number of a segment or sub-segment depends, in part, on the copy number state of a spatially adjacent segment or sub-segment. Lengths and frequencies of copy number variants can also impact the transition probabilities.
In some embodiments, the transition probability can be predetermined or fixed. In a preferred embodiment, the transition probability is variable. For example, the transition probability can be formally represented by the following stochastic transition matrix assuming a hidden copy number state limited to 0, 1, 2, 3, or 4 copies (assuming a wildtype copy number of 2):
wherein Ci is the copy number state of a first segment or first sub-segment; Ci+1 is the copy number state of the second segment or second sub-segment that is spatially adjacent to the first segment or first sub-segment; and rab represents the transition probability from a first copy number state a to a second copy number state b. For example, a can be a copy number state of 3 and b can be a copy number state of 2. The first segment can be the interrogated segment (or the first sub-segment can be a sub-segment of the interrogated segment). Although the above stochastic transition matrix assumed 0, 1, 2, 3, or 4 copies, it is understood that a stochastic transition matrix can be used for any number of copies.
Copy number variants have an average length, and copy numbers that are longer or shorter than this length are less likely than copy numbers at the average length. In some embodiments the transition probability (or transition probabilities) account for an average length of a copy number variant. The average length of the copy number variant can be based on observations from a historical population (e.g., a historical human population). The historical population is a historical population of sequencing libraries for which a copy number variant has been called. Larger historical populations can result in more accurate average copy number variant lengths. In some embodiments, the historical population comprises about 1000 or more sequencing libraries (such as about 5000 or more, about 10,000 or more, about 25,000 or more, about 50,000 or more, about 100,000 or more, about 250,000 or more, or about 500,000 or more sequencing libraries). The average length of a copy number variant is predetermined. In some embodiments, the average length of a copy number variant is about 3000 to about 1000 bases (such as about 4000 to about 8000 bases, about 5000 to about 7000 bases, about 5500 bases to about 6500 bases, or about 6200 bases). Accounting for the average length of a copy number, the transitions in the stochastic transition matrix which is used to calculate the per base transition probability (or sub-segment transition probability can be set as:
wherein |
CNV is the average length of a copy number variant.
The transition probabilities can also account for the probability of a copy number variant at the interrogated segment given the copy number state at a spatially adjacent segment. Certain portions of the genome may include “hot spots” of genetic variation, including copy number variation. Hot spots, refers to regions in the genome which display a high propensity for mutations of all kinds. This might be due to structural makeup of the region, or functional aspects of the region, which make it more prone to mutations. The probability of a copy number variant at any given segment (such as an interrogated segment or a spatially adjacent segment) can be based on observations from a historical population (e.g., a historical human population). The historical population is a historical population of sequencing libraries for which a copy number variant has been called. Larger historical populations can result in more accurate copy number variant probabilities. In some embodiments, the historical population comprises about 1000 or more sequencing libraries (such as about 5000 or more, about 10,000 or more, about 25,000 or more, about 50,000 or more, about 100,000 or more, about 250,000 or more, or about 500,000 or more sequencing libraries). To account for the probability of a copy number variant at the interrogated segment or a spatially adjacent segment, the transitions in the stochastic transition matrix can be set as:
wherein pCNV is the probability of a copy number variant, and l
CNV is the length of an average copy number variant. Since r01=r12=r32=r43 the relationship described holds true for all copy numbers.
In some embodiments, the hidden Markov model comprises one transition probability of a copy number state of a segment or of a sub-segment. In some embodiments, the hidden Markov model comprises a plurality of transition probabilities of a copy number state of a segment or of a sub-segment. In some embodiments, the transition probability of a copy number state given the copy number state of an adjacent preceding segment is dependent on length of a copy number variant. In some embodiments, the length of the copy number variant is specific for that particular region of the genome. In some embodiments, the length of a copy number variant is the average length of a copy number variant across the genome.
In some embodiments the transition probability of a copy number state given the copy number state of an adjacent preceding segment is dependent on the probability of observing a copy number variant. In some embodiments the probability of observing a copy number variant is specific for that particular region of the genome. In some embodiments the probability of observing a copy number variant is the average probability of observing a copy number variant across the genome.
Parameterizing the hidden Markov model and Determining a Most Probable Copy Number
As described above, the hidden Markov model includes (i) one or more hidden states comprising a copy number corresponding to the one or more segments or sub-segments (including at least the interrogated segment or a sub-segment of the interrogated segment), (ii) one or more observation states comprising the number of sequencing reads mapped to the one or more segments, and (iii) the copy number likelihood model. The copy number likelihood model is used to describe the probability of observing an observation state for a given hidden state (that is, p(ki,j|ci,j)). The hidden Markov model also includes a transition probability between the hidden states, which can be fixed or variable as described above.
The hidden Markov model is initiated using the copy number likelihood model. The hidden Markov model can also be initiated by assuming the copy number state (i.e., the hidden state) to have a wild-type number of copies (for example, two copies), which can be used to back-calculate the transitions (r) for determining the transition probabilities. The copy number likelihood model is based on the expected number of sequencing reads mapped to the segment, as explained above, but the copy number likelihood model can be adjusted to fit the determined number of sequencing reads mapped to the segment (i.e., the observed states), for example by allowing the mean μc,i,j and dispersion di for each copy number likelihood distribution in the copy number likelihood model to float when parameterizing the hidden Markov model. The transition probabilities, if variable, can also be adjusted during parameterization of the hidden Markov model.
Parameterization of the hidden Markov model includes adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the segment (e.g., the interrogated segment or the spatially adjacent segments). In some embodiments, the copy number likelihood model is optimized to fit the determined number of sequencing reads mapped to the segment (e.g., the interrogated segment or the spatially adjacent segments). The copy number likelihood model is “optimized” after a plurality of adjustment rounds to best fit the observed states. In some embodiments, parameterization of the hidden Markov model includes adjusting (or optimizing) the transition probabilities. The hidden Markov model can be parameterized by optimizing the copy number likelihood model using an analytic first derivative gradient and a second derivative Hessian of one or more parameters in the copy number likelihood model, which may be solved, for example, using a trust region Newton conjugate gradient algorithm. Exemplary parameters of the copy number like likelihood model that are optimized can include one or more of a dispersion of a number of mapped sequencing reads for the segment (di), an average number of mapped sequencing reads for the segment CA), a dispersion of a number of mapped sequencing reads for the segments within the test sequencing library (dj), or an average number of mapped sequencing reads for the segments within the test sequencing library (μj). An Expectation-Maximization (EM) algorithm can then be used to optimize the parameters in one or more iterations of applying the hidden Markov model to determine a most probable copy number of the segment (for example, using a Viterbi algorithm, a Quasi-Newton solver, or a Markov chain Monte Carlo, along with a Baum-Welch algorithm) and re-parameterizing the hidden Markov model.
For example, expectation-maximization (EM) may be used to adjust (or optimize) the copy number likelihood model (based on the expected number of sequencing reads) and/or one or more additional model parameters to find a maximized expected sequencing reads mapped to the segment (that is, an adjusted μc,i,j) and an adjusted dispersion for that segment (that is, an adjusted di). That is, so that the probability of an expected number of sequencing reads at an interrogated segment is maximized for a given copy number state at that segment.
Generally, expectation-maximization (EM) can be used to estimate latent, or unknown, parameters despite incomplete data. The EM algorithm can iteratively alternate between the expectation “E” step which selects a most likely copy number likelihood distribution from the copy number likelihood model given the determined number of sequencing reads mapped to the segment (such that the most probable copy number can be determined), and a maximization “M” step, which re-estimates the copy number likelihood model parameters (i.e., μc,i,j and di). The Maximization step assumes a fixed probabilistic model and number of sequencing reads, and then finds the copy number state that would, when applied to the model, result into a highest probability for the actual number of mapped sequencing reads out of all other possible copy numbers. An EM process can be applied at different parameters of the HMM, for example it can consider the transitions (r) between the hidden states if applicable, using the expectations generated in the “E” step. Simplistically, the EM is used to maximize the model so that we find for which ci,j are we most likely to see the number of mapped sequencing reads that we observed. Formally, a Viterbi algorithm can determine the maximum likelihood for the copy number likelihood model as:
In some embodiments, a Baum-Welch algorithm is used for the expectation step of the EM process, which determines the expected probability of a copy number call for the segment. The Baum-Welch algorithm uses a posterior probability α(ci|k[0,i]), which is the probability of a copy number state at segment i for a given number of mapped sequencing reads at segment i, and a likelihood β(k[i,l]|ci), which is the probability of the number of mapped sequencing reads for the downstream spatially adjacent segments i to I for a given copy number state at segment i. The Baum-Welch algorithm can be solved using methods known by a person of skill in the art.
The parameterized hidden Markov model can be used to determine a most probable copy number of the interrogated segment or a sub-segment of the interrogated segment during the Maximization step. The most probable copy number of the interrogated segment can be determined using any useful algorithm known in the art, such as a Viterbi algorithm, a Quasi-Newton solver, or a Markov chain Monte Carlo.
GC content of a segment of the region of interest or a capture probe corresponding to the segment can affect the number of sequencing reads mapped to the segment, for example due to differences in hybridization efficiency of the capture probe. Thus, depending on GC content, a capture probe may have strong effects on the number of sequencing reads mapped to a segment, irrespective of the copy number state at that segment. This GC content bias is well known and described in the art. In some embodiments of the methods described herein, the GC content bias is accounted for when determining a copy number of the segment. The GC content bias correction can be useful in any method of determining a copy number variant, and need not be used solely with direct targeted sequencing. For example, in some embodiments, GC content bias is corrected when determining a copy number of a segment in a region of interest, wherein the sequencing library is enriched using hybrid capture techniques. Additionally, the methods for correcting GC content bias need not be limited to methods using a hidden Markov model to determine a copy number, but the GC content bias can be corrected for any method that includes the use of a copy number likelihood model.
In some embodiments, a number of sequencing reads (such as the expected number of sequencing reads used to determine the copy number likelihood model) for any given segment is corrected for GC content by multiplying the number of sequencing reads by a GC bias correction factor. The GC bias correction factor is specific for the given segment and for the test sequencing library. That is, the GC bias correction factor is uniquely determined for the segment and the test sequencing library, and the GC bias correction factor must be re-determined for a different segment and for each different test sequencing library.
The number of sequencing reads mapped to a given segment (which may include the interrogated segments) can be normalized by dividing the number of mapped sequencing at that segment by the average number of mapped sequencing reads for a plurality of segments enriched from the test sequencing library. The normalized number of sequencing reads for each segment within the plurality of segments can be plotted against the GC content at that segment. The data points can then be fit using a second order correction:
g
i,j
=a+b(GC)+c(GC)2
wherein gi,j is a GC-bias correction factor specific for segment i of test sequencing library j for the plurality of segments, (GC) is the GC content, and a, b, and c are constants determined by the second order fit.
The GC bias correction factor can therefore be determined by fitting a second order function to a plurality of data points, wherein the data points each comprises a normalized number of sequencing reads mapped to a segment and the GC content of that segment, and wherein the plurality of data points represent a plurality of segments enriched by the capture probes in the test sequencing library; and defining the GC bias correction factor to be the normalized number of sequencing reads determined by the second order function for the GC content of the segment.
The copy number likelihood model can be adjusted to account for the presence of GC content bias in a similar manner. That is, the expected number of sequencing reads used as a basis for the copy number likelihood model can be adjusted to account for the presence of GC content. For example, the average of the copy number likelihood distribution in the model can be adjusted such that:
μc,i,j=ci,jμiμjgi,j
Further, the copy number likelihood model can be formalized as:
p(ki,j|ci,j)=NegBinom(ki,j|μc,i,j=ci,jμiμjgi,j,d)
wherein ki,j refers to the number of sequencing reads at segment i in test library j, wherein d is di, dj or di,j.
In some embodiments, there is a method for determining a copy number of an interrogated segment or a sub-segment of the interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to a segment within a region of interest, wherein the test sequencing library is enriched using a capture probe; (b) determining a number of sequencing reads mapped to the segment; (c) determining a copy number likelihood model for the segment based on an expected number of mapped sequencing reads at the segment, wherein the expected number of mapped sequencing reads is corrected for GC content of the segment; and (d) determining a most probable copy number of the interrogated segment based on the copy number likelihood model. The most probable copy number of the interrogated segment can be determined based on the copy number likelihood model using the hidden Markov model described herein, or can be done by any other method known in the art. For example, the most probable copy number can be determined based on the maximum copy number probability of each region based on a capture probe for that region. In another example, the most probable copy number can be determined using a brute force segmentation approach.
In some embodiments, there is a method for determining a copy number of an interrogated segment or a sub-segment of the interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to the interrogated segment, wherein the test sequencing library is enriched using one or more capture probes; (b) determining a number of sequencing reads mapped to the interrogated segment; (c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment, wherein the expected number of mapped sequencing reads is corrected for GC content of the interrogated segment; (d) building a hidden Markov model comprising: (i) one or more hidden states comprising a copy number corresponding to the interrogated segment or a plurality of sub-segments within the interrogated segment, (ii) an observation state comprising the number of sequencing reads mapped to the interrogated segment; and (iii) the copy number likelihood model; (e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment; and (f) determining a most probable copy number of the interrogated segment or a sub-segment of the interrogated segment based on the parameterized hidden Markov model.
In some embodiments there is a method for determining a copy number of an interrogated segment or a sub-segment of the interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises the interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent capture probes; (b) determining a number of sequencing reads mapped to each spatially adjacent segment; (c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment, wherein the expected number of mapped sequencing reads is corrected for GC content of the spatially adjacent segment; (d) building a hidden Markov model comprising: (i) a plurality of hidden states comprising a copy number for each of the spatially adjacent segments or a plurality of sub-segments within each of the spatially adjacent segments, (ii) a plurality of observation states comprising the number of sequencing reads mapped to each spatially adjacent segment, and (iii) the copy number likelihood model for each spatially adjacent segment; (e) parameterizing the hidden Markov model comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment; and (f) determining a most probable copy number of the interrogated segment or a sub-segment of the interrogated segment based on the parameterized hidden Markov model.
Certain capture probes used to enrich a segment within the region of interest can produce spurious results. For example, the number of sequencing reads generated by a spurious capture probe may not be consistent with the copy number of a corresponding segment, either by under or over enriching the segment. These spurious results can occur, for example, due to capture probe design or sequence variants (e.g., SNPs) within the sequence the capture probe was designed to hybridize to. Spurious capture probes affect number of mapped sequencing reads and can artificially confound the copy number likelihood model and parameters. It is therefore desirable to account for spurious capture probes. Spurious capture probes need not be direct targeted sequencing capture probes, and similar methods can be applied to capture probes used to enrich a test sequencing library (such as by hybrid capture techniques). The determination of whether a capture probe is a spurious capture probe can be made using EM. For example, the determination of whether the capture probe is spurious can be made during the expectation step, and when the probability of a capture probe being spurious changes during EM iterations, so will the maximization step, which in turn determines the most likely copy number state of the segment which now takes into consideration the spuriosity of the capture probe. If a capture probe is determined to be a spurious capture probe, the probability of the number of mapped sequencing reads for a segment for a copy number state is set to 1 during the expectation-maximization process. By setting it a constant it effectively allows the model to disregard the spurious capture probe as it provides no additional information and is thus not taken into consideration as the model is parameterized. Determination of the spuriousness of the capture probe can be iterative, for example by determining whether the capture probe is spurious after a number of EM cycles.
In some embodiments a Bernoulli process is used to determine the probability that a given capture probe is spurious. The Bernoulli process can be applied to some or all of the capture probes. That is, for each capture probe its spuriosity is independently determined. For capture probe i, an indicator variable bi is introduced where 1 means that the capture probe t is spurious and 0 means that the capture probe is not spurious.
b
i∈{0,1}
By using this indicator, it is possible to account for spurious capture probes by adjusting the copy number likelihood model. If a capture probe is determined to be spurious, the probability of a number of mapped sequencing reads for the corresponding segment for any given copy number is set to 1. If the capture probe is non-spurious, the copy number likelihood distributions in the copy number likelihood model are unchanged. Formally:
The indicator on the observed states of the hidden Markov model is illustrated in
The spuriousness of capture probes may depend on the test sequencing library. That is, some test sequencing libraries may be more prone to spurious capture probes than other test sequencing libraries. In some embodiments whether a test sequencing library is prone to spurious capture probes is determined based on test sequencing library priors. In some embodiments determining whether a test sequencing library will be prone to a particular probe being spurious depends on a general prior.
p(bi|πj)=πjb
As the Bernoulli distribution limits bi to be either 0 or 1, the above probability is set to πj (when bi=1), or 1−πi (when bi=0).
Given the determined number of sequencing reads mapped to spatially adjacent segments (or spatially adjacent sub-segments) 0 to I, the probability of capture probe i being spurious can be derived to:
Given the expectation of the indicator bi, the test sequencing library prior πj can be determined as:
πj=bi
k
In some embodiments, the most probable copy number of the interrogated segment or the one or more sub-segments of the interrogated segment is not called if the capture probe associated with the interrogated segment is determined to be spurious. In some embodiments, the most probable copy number of the interrogated segment or the one or more sub-segments of the interrogated segment is not called if the probability of a capture probe i being spurious (that is, p(bi|k[0,1])) is above a predetermined threshold (such as about 0.1 or more, about 0.2 or more, about 0.3 or more, about 0.4 or more, or about 0.5 or more).
In some embodiments, there is a method for determining a copy number of an interrogated segment or a sub-segment of the interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to the interrogated segment, wherein the test sequencing library is enriched using one or more capture probes; (b) determining a number of sequencing reads mapped to the interrogated segment; (c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment; (d) building a hidden Markov model comprising: (i) one or more hidden states comprising a copy number corresponding to the interrogated segment or a plurality of sub-segments within the interrogated segment, (ii) an observation state comprising the number of sequencing reads mapped to the interrogated segment; and (iii) the copy number likelihood model; (e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment and accounting for one or more spurious capture probes; and (f) determining a most probable copy number of the interrogated segment or a sub-segment of the interrogated segment based on the parameterized hidden Markov model.
In some embodiments, there is a method for determining a copy number of an interrogated segment or a sub-segment of the interrogated segment within a region of interest comprising: (a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises the interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes; (b) determining a number of sequencing reads mapped to each spatially adjacent segment; (c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment; (d) building a hidden Markov model comprising: (i) a plurality of hidden states comprising a copy number for each of the spatially adjacent segments or a plurality of sub-segments within each of the spatially adjacent segments, (ii) a plurality of observation states comprising the number of sequencing reads mapped to each spatially adjacent segment, and (iii) the copy number likelihood model for each spatially adjacent segment; (e) parameterizing the hidden Markov model comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment and accounting for one or more spurious capture probes; and (f) determining a most probable copy number of the interrogated segment or a sub-segment of the interrogated segment based on the parameterized hidden Markov model.
During preparation of test sequencing libraries, several steps can result in the nucleic acid of the test sequencing library to be more prone to “noise” across multiple capture probes. This results in inconsistent data and a high number of false positives
In some embodiments, parameterizing the hidden Markov model comprises accounting for noise in the number of mapped sequencing reads. In some embodiments, accounting for noise in the number of mapped sequencing reads comprises adjusting the copy number likelihood model. For example, parameterizing the hidden Markov model can include an expectation-maximization step, and accounting for the noise can occur during the expectation-maximization step.
The dispersion d of the copy number likelihood distribution in the copy number likelihood model was discussed above. When only the dispersion due to the capture probe (i.e., at the segment) is considered, d=di. The dispersion of the copy number likelihood distribution can also be used to account for noise across the segments in the test sequencing library j. The dispersion due to account for noise across the segments in the test sequencing library and the noise due to the capture probe can be determined through arithmetic combination, for example by multiplying or adding the dispersion due to the sequencing library noise and the dispersion due to the capture probe noise. For example, in some embodiments, the dispersion of the copy number likelihood distribution can formally be considered as:
d=d
i
*d
j
The dispersion due to the sequencing library noise and the dispersion to the capture probe noise can, in some embodiments, be combined through addition, for example:
d=e
d
+e
d
Parameterization of the hidden Markov model adjusts the copy number likelihood model, including the dispersion of the copy number likelihood distributions with the model. Thus, both components of the dispersion d (that is, di and dj) can be adjusted during parameterization of the hidden Markov model, for example using an expectation maximization algorithm. In some embodiments, an analytic first derivative gradient and second derivative Hessian of the dispersion of a number of mapped sequencing reads for the segments within the test sequencing library (dj) is used to account for noise. In some embodiments, a quasi-Newton method can be used to account for the noise during the maximization step. In particular, the expectation step asks to maximize the following
In the equation, l({right arrow over (μ)}, {right arrow over (d)}) represents the expected logarithmic likelihood given all the data and the current parameters of the model. TSL stands for test sequencing library and cpt probes refers to capture probes. The mean {right arrow over (π)} can be approximated by using a double normalization, which accounts for both the median sequencing depth across segments within a test sequencing library and the median sequencing depth of a plurality of test sequencing library across the same segment. In some embodiments, to find the dispersion {right arrow over (d)} that can maximize this function a quasi-Newtonian method is used. The quasi-Newton method sets the partial derivative of this function with respect to {right arrow over (d)} to 0. Since the test sequencing library and the capture probe shape are independent, it is equivalent to setting the partial derivative of each type to 0.
Once the parameters of the distribution are set, the parameterized hidden Markov model can be used to determine the most probable copy number state of the segment.
In certain aspects, the methods described herein are used to assess the sample-specific performance of a copy number variant screen or copy number variant model. Synthetic copy number variants are generated in silico using the real sequencing reads from the test sample. Therefore, the synthetic copy number variants are sample-specific. A copy number variant model is parameterized using the real number of sequencing reads mapped to segments within the region of interest for the test sample to determine copy number variant model parameters. Since the synthetic copy number variants are based on the test sample, and the determined copy number variant model parameters are sample specific, the determined sample-specific copy number variant model parameters are used by the copy number variant caller to call copy numbers of the segments within the synthetic copy number variants.
The synthetic copy number variant includes a synthetic number of copies of one or more segments with the region of interest, which is represented by a synthetic number of sequencing reads from one or more segments within the region of interest. In some embodiments, the synthetic number of sequencing reads is obtained by adjusting a number of sequencing reads of the one or more segments within the region of interest from the test sample. The adjustment is made in proportion to the synthetic number of copies. In some embodiments, the synthetic number of sequencing reads is obtained by direct manipulation of a database comprising sequencing reads of the one or more segments within the region of interest from a real sample, for example by random deletion or duplication of sequencing reads within the database. In some embodiments, the synthetic number of sequencing reads is generated by sampling a distribution (such as a binomial distribution or a negative binomial distribution). A plurality of synthetic copy number variants can be generated, for example based on a plurality of test samples or reference samples.
A synthetic number of copies of one or more segments within the region of interest present in the synthetic copy number variant is called using the copy number variant caller. In some embodiments, the caller compares the synthetic number of sequencing reads from the one or more segments in the synthetic copy number variant to the number of sequencing reads from the one or more segments in a real reference sample with a known number of copies of the segments. The caller can use, for example, a hidden-Markov model (HMM) described herein to determine the number of copies of the segments in the synthetic copy number variant. The real reference sample is preferably a different real sample that the real sample used as a basis for generating the synthetic copy number variants.
The copy number variant caller uses the synthetic copy number variants and the determined copy number variant model parameters, as shown in
A performance statistic for the copy number variant screen can be determined to assess the sample-specific performance of the copy number variant screen based on the differences between the called number of copies and the synthetic number of copies in the synthetic copy number variants. Since a plurality of synthetic copy number variants are generated and called by the caller, the performance statistic reflects the performance of the screen in the context of the synthetic variants. Thus, a greater diversity of synthetic copy number variants (which can be based on a plurality of real samples) provides a more accurate performance statistic characterizing the performance of the copy number variant model.
In some embodiments, a method of assessing the sample-specific performance of a copy number variant model, comprising: parameterizing the copy number variant model based on real numbers of sequencing reads mapped to segments within a region of interest, from a test sample, to determine one or more copy number variant model parameters; generating a plurality of synthetic copy number variants, each synthetic copy number variant comprising a synthetic number of copies of one or more of the segments, wherein each synthetic number of copies is represented by a synthetic number of sequencing reads based on a real number of sequencing reads for a corresponding segment from the test sample; calling a number of copies of the one or more segments for the synthetic copy number variants using the copy number variant model, and the one or more determined copy number variant model parameters; determining a sample-specific performance statistic for the copy number variant model based on differences between the called number of copies and the synthetic number of copies in the synthetic copy number variants; and assessing a sample-specific performance of the copy number variant model based on the sample-specific performance statistic. In some embodiments, the sample-specific performance statistic is a limit of detection, sensitivity, specificity, precision, recall, accuracy, positive predictive value, or negative predictive value.
In some embodiments, the copy number variant caller uses a hidden Markov Model to call the copy number in the synthetic copy number variant. As copy number variants of any given segment are relatively rare, it can be assumed that the test sample does not have a copy number variant at any given segment. Even if the test sample does have a copy number variant, the test sample will include sufficient non-variant (i.e., wild-type) segments such that the assessment method is reliable.
For purposes of generating the synthetic copy number variants, the test sample is assumed to be wild-type for a number of copies of the segments with the region of interest, and the number of sequencing reads can be assumed to form a negative binomial distribution with an average (mean or median) and a variance. The variance of the distribution can arise, for example, from noise during enrichment or sequencing of the segments. The distribution of sequencing reads from a population of synthetic copy number variants preferably resembles an expected negative binomial distribution of sequencing reads from a theoretical population of real copy number variants that are equivalently processed and therefore have the same copy number variant model parameters.
In some embodiments, a plurality of synthetic copy number variants comprising a synthetic number of copies of one or more segments represented by a synthetic number of sequencing reads from one or more segments within the region of interest is generated. The synthetic number of sequencing reads for each of the one or more segments can be generated by increasing, decreasing, or maintaining a number of real sequencing reads from the one or more segments within a region of interest from the test sample. For example, if a first number of real sequencing reads corresponds to a first segment in a region of interest, and a second number of real sequencing reads corresponds to a second segment in the region of interest, and the test sample is assumed or expected to have two copies of the region of interest, a synthetic copy number variant having three copies of the region of interest can be generated by generating a first synthetic number of sequencing reads corresponding to the first segment by increasing the first number of real sequencing reads to reflect three copies of the first segment, and generating a second synthetic number of sequencing reads corresponding to the second segment by increasing the second number of real sequencing reads to reflect three copies of the second segment. Since the synthetic number of sequencing reads corresponding to the first segment and the second segment are increased to reflect three copies, the synthetic copy number variant has three copies of the region of interest having the first segment and the second segment. In some embodiments, the synthetic number of sequencing reads are generated by multiplying the number of real sequencing reads by a factor (such as 1.5 to increase the copy number from two to three, or 0.5 to decrease the copy number from two to one). In some embodiments, the synthetic number of sequencing reads are generated by adding (or subtracting) a number of sequencing reads (such as 50% of the average number of real sequencing reads corresponding to all segments within the region of interest) to the number of real sequencing reads. In some embodiments, the number of sequencing reads are normalized (for example, as described below) such that a single copy of a region of interest is represented by a normalized number of sequencing reads (e.g., 0.5), and two copies of a region of interest are represented by a normalized number of sequencing reads (e.g., 1). Thus, in some embodiments, a number of normalized sequencing reads (such as 0.5) are added to the normalized number of sequencing reads to increase the number of copies in the synthetic copy number variant, and a number of normalized sequencing reads (such as 0.5) are subtracted to the normalized number of sequencing reads to decrease the number of copies in the synthetic copy number variant. Preferably, the number of real sequencing reads are increased or decreased to generate the synthetic number of sequencing reads to represent a synthetic copy number variant with a predetermined number (which may be an integer number or a non-integer number) of copies of the segment (such as 1 or more, 2 or more, 3 or more, 4 or more, or 5 or more copies of the segment).
In some embodiments, a synthetic number of sequencing reads is generated by adding or subtracting a number of sequencing reads from a number of sequencing reads from a test sample to generate a synthetic copy number variant. A synthetic copy number variant comprising a duplication is generated by adding the number of sequencing reads, and a synthetic copy number variant comprising a deletion event is generated by deleting a number of sequencing reads. The number of sequencing reads added or subtracted from the number of number of sequencing reads from the test sample is based, in part, on how many duplication or deletion events are simulated in the synthetic copy number variant. In some embodiments, a synthetic number of sequencing reads for a synthetic copy number variant comprising n copies of a region of interest (or segment thereof) more (or less) than an assumed (e.g., wild-type) number of copies x in a test sample is determined by adding (or subtracting)
times an average (e.g., mean or median) number of sequencing reads from a plurality of test samples for that region of interest (or segment thereof) to (or from) the number of sequencing reads for that region of interest (or segment thereof) from a test sample. For example, for a synthetic copy number variant comprising a deletion (i.e., having n fewer copies of a region of interest or segment thereof than an assumed number of copies x in a test sample), the synthetic number of sequencing reads kx−n for the synthetic copy number variant is determined as
wherein ki,j refers to the number of sequencing reads at region of interest (or segment) j for test sample i, and μ refers to an average (mean or median) number of sequencing reads, which can be, for example, an average number of sequencing reads from all segments within the test sample (i.e., μi), an average number of sequencing reads at region of interest (or segment) j across a plurality of test samples (i.e., μi), or a normalized (or double normalized) average number of sequencing reads that is an average number of sequencing reads for a region of interest (or segment) j for a plurality of test samples, wherein the number of sequencing reads for each test sample has been normalized across the test sample (i.e., μiμj). By way of example, a synthetic copy number variant having one copy of a segment j can be determined based on a number of sequencing reads from a test sample i assumed to have two copies of the segment can be determined as
In some embodiments, a synthetic copy number variant comprising a duplication (i.e., having n additional copies of a region of interest or segment thereof than an assumed number of copies x in a test sample), the synthetic number of sequencing reads kx+n for the synthetic copy number variant is determined as
By way of example, a synthetic copy number variant having three copies of a segment j can be determined based on a number of sequencing reads from a test sample i assumed to have two copies of the segment can be determined as
In some embodiments, a synthetic number of sequencing reads for a synthetic copy number variant comprising m copies of a region of interest (or segment thereof) can be generated based on a number of sequencing reads of that region of interests (or segment thereof) comprising x copies of the region of interest (or segment thereof) according to
For example, a synthetic number of sequencing reads for a synthetic copy number variant with three copies of a region of interest (or segment thereof) can be generated based on a number of sequencing reads from a test sample having two copies of the region of interest (or segment thereof) according to:
In some embodiments, a synthetic copy number variant with one copy of a region of interest (or segment thereof) can be generated based on a number of sequencing reads from a test sample having two copies of the region of interest (or segment thereof) according to:
In some embodiments, the synthetic number of sequencing reads for a synthetic copy number variant comprising m copies of a segment is generated from a number of sequencing reads from a test sample with an assumed (e.g., wild-type) number of copies x by multiplying the number of sequencing reads from the test sample by
That is, a synthetic number or sequencing reads ki,jm for a synthetic copy number variant can be determined based on a number of sequencing reads ki,jx according to:
For example, a synthetic number of sequencing reads for a synthetic copy number variant having three copies of a region of interest (or segment thereof) can be generated based on a number of sequencing reads from a test sample assumed to have two copies of the region of interest (or segment thereof) according to:
A synthetic number of sequencing reads for a synthetic copy number variant having one copy of a region of interest (or segment thereof) can be generated based on a number of sequencing reads from a test sample assumed to have two copies of the region of interest (or segment thereof) according to:
In some embodiments, a fudge factor is included when determining the synthetic number of sequencing reads, which can be used to more closely model the variance of a plurality of synthetic numbers of sequencing reads (i.e., a plurality of synthetic copy number variants) to the variance of a plurality of test samples used as a basis for the plurality of synthetic copy number variants. The fudge factor can be derived from the increase or decrease in variance expected for a Poisson distribution when changing the average number of sequencing reads.
In some embodiments, the synthetic number of sequencing reads for a synthetic copy number variant is determined by sampling a binomial distribution or a negative binomial distribution of reals sequencing reads from a test sample. For example, for synthetic copy number deletion variant having m copies of a region of interest (or segment thereof), the synthetic number of sequencing reads can be generated by sampling from a binomial distribution of a real number of sequencing reads from a test sample having x copies of the region of interest (or segment thereof) with a success probability equal to
and a number of trials equal to the number of real sequencing reads. That is, for a synthetic copy number deletion variant,
For example, for a synthetic copy number variant having one copy of a region of interest (or segment thereof), the synthetic number of sequencing reads can be generated by sampling from a binomial distribution of a real number of sequencing reads from a test sample having two copies of the region of interest (or segment thereof) with a success probability equal to ½ and a number of trails equal to the number of real sequencing reads. That is, ki,j1=Binomial(n=ki,j2,p=½).
In some embodiments, the synthetic number of sequencing reads for a synthetic copy number duplication variant having m number copies of a region of interest (or segment thereof) is generated by sampling from a negative binomial distribution, wherein a number of successes is equal to the real number of sequencing reads from a test sample having an assumed x number of copies of the region of interest (or segment thereof), and the probability of success is equal to
and adding an expectation of the sampled negative binomial distribution to the real number of sequencing reads. That is, for a synthetic copy number duplication variant,
For example, a synthetic number of sequencing reads for a synthetic copy number duplication variant having three copies of a region of interest (or segment thereof) can be generated by sampling from a negative binomial distribution, wherein a number of successes is equal to the real number of sequencing reads from a test sample having an assumed two number of copies of the region of interest (or segment thereof), and the probability of success is equal to ⅔, and adding an expectation of the sampled negative binomial distribution to the real number of sequencing reads. That is, ki,j3=ki,j2+E[NegBinom(n=ki,j2,p=⅔)]. In some embodiments, a fudge factor is included when determining the synthetic number of sequencing reads, which can be used to more closely model the variance of a plurality of synthetic numbers of sequencing reads (i.e., a plurality of synthetic copy number variants) to the variance of a plurality or test samples used as a basis for the plurality of synthetic copy number variants. The fudge factor can be determined empirically. For example, the fudge factor can be determined by comparing the distribution of sequencing reads from an X chromosome in males (which have a single copy of the X chromosome) to the distribution of sequencing reads from the X chromosome in females (which have two copies of the X chromosome) that have a simulated deletion of a single X chromosome (thus having a simulated one copy of the X chromosome). The fudge factor can be adjusted such that the observed one copy males are compared to simulated one copy females. For example, the synthetic number of sequencing reads can be determined according to:
wherein
and β is the fudge factor. In one example,
wherein α=NegBinom(n=ki,j2,p=⅔).
The copy number variant caller can call a number of copies of one or more segments within the region of interest for each synthetic copy number variant in the plurality of synthetic copy number variants. The number of copies of the one or more segments in each synthetic copy number variant is known, as the number of copies of the segments in the synthetic copy number variant are represented by the synthetic number of sequencing reads, which were generated by adjusting the number of real sequencing reads from the test sample to a desired number of copies of the one or more segments. The called number of copies can be compared to the number of copies in each of the synthetic copy number variant in the plurality of synthetic copy number variants to determine a performance statistic for the copy number variant model. The performance statistic can be, for example, sensitivity, specificity, precision, recall, accuracy, positive predictive value, negative predictive value, or any other metric of concordance.
The performance statistic indicates the performance of the copy number variant screen or model. For example, a high number of true positives and a low number of false negatives for a copy number variant model is preferable. Thus, the performance statistic can be used to assess the performance of the copy number variant model. In some embodiments, a predetermined threshold for the performance statistic can be selected. In some embodiments, if the performance statistic is below the predetermined threshold, the test sample can be re-analyzed and/or a new set of sequencing reads can be generated for the test sample.
In some embodiments, the methods described herein are implemented by a program executed on a computer system.
At least some values based on the results of the above-described processes can be saved for subsequent use. Additionally, a non-transitory computer-readable medium can be used to store (e.g., tangibly embody) one or more computer programs for performing any one of the above-described processes by means of a computer (e.g., the one or more central processing units (“CPU”) 1108 can execute the stored one or more computer programs (or instructions) to perform the above-described processes). The computer program may be written, for example, in a general-purpose programming language (e.g., Pascal, C, C++, Java, Python, JSON, R, etc.) or some specialized application-specific language.
In some embodiments, the summary statistic is reported (for example, to a patient, a doctor, a caregiver, or a regulator). In some embodiments, the summary statistic is displayed, for example on a monitor.
Various exemplary embodiments are described herein. Reference is made to these examples in a non-limiting sense. They are provided to illustrate more broadly applicable aspects of the disclosed technology. Various changes may be made and equivalents may be substituted without departing from the true spirit and scope of the various embodiments. In addition, many modifications may be made to adapt a particular situation, material, composition of matter, process, process act(s) or step(s) to the objective(s), spirit or scope of the various embodiments. Further, as will be appreciated by those with skill in the art, each of the individual variations described and illustrated herein has discrete components and features that may be readily separated from or combined with the features of any of the other several embodiments without departing from the scope or spirit of the various embodiments. All such modifications are intended to be within the scope of claims associated with this disclosure.
The following embodiments are exemplary and are not intended to limit the present invention.
Embodiment 1. A method of assessing the sample-specific performance of a copy number variant caller comprising a copy number variant model, comprising:
parameterizing the copy number variant model based on real numbers of sequencing reads mapped to segments within a region of interest, from a test sample, to determine one or more copy number variant model parameters;
generating a plurality of synthetic copy number variants, each synthetic copy number variant comprising a synthetic number of copies of one or more of the segments, wherein each synthetic number of copies is represented by a synthetic number of sequencing reads based on a real number of sequencing reads for a corresponding segment from the test sample;
calling a number of copies of the one or more segments for the synthetic copy number variants using the copy number variant model, and the one or more determined copy number variant model parameters;
determining a sample-specific performance statistic for the copy number variant caller based on differences between the called number of copies and the synthetic number of copies in the synthetic copy number variants; and
assessing a sample-specific performance of the copy number variant caller based on the sample-specific performance statistic.
Embodiment 2. The method of embodiment 1, wherein the synthetic number of sequencing reads for the one or more segments is generated by increasing, decreasing, or maintaining the real number of sequencing reads for the corresponding segments from the test sample in proportion to a predetermined number of copies of the one or more segments.
Embodiment 3. The method of embodiment 2, wherein the predetermined number of copies is an integer number of copies.
Embodiment 4. The method of embodiment 2, wherein the predetermined number of copies is a non-integer number of copies.
Embodiment 5. The method of any one of embodiments 1-4, wherein the synthetic number of sequencing reads is generated by sampling a binomial distribution with a success probability equal to m/x and a number of trials equal to the real number of sequencing reads at the corresponding segment from the test sample, wherein m is the synthetic number of copies of the segment in the synthetic copy number variant, and x is an assumed number of copies of the corresponding segment from the test sample.
Embodiment 6. The method of any one of embodiments 1-5, wherein the synthetic number of sequencing reads is generated by:
sampling a number of sequencing reads as a negative binomial distribution with a success probability equal to m/x and a number of successes equal to the real number of sequencing reads at the corresponding segment from the test sample, wherein m is the synthetic number of copies of the segment in the synthetic copy number variant, and x is an assumed number of copies of the corresponding segment from the test sample, and
adding the sampled number of sequencing reads to the real number of sequencing reads for the corresponding segment from the test sample.
Embodiment 7. The method of embodiment 6, wherein the synthetic number of sequencing reads is generated by sampling a number of sequencing reads as an expectation of the negative binomial distribution.=
Embodiment 8. The method of any one of embodiments 1-7, wherein the copy number variant model is a hidden Markov model.
Embodiment 9. The method of embodiment 8, wherein the hidden Markov model comprises:
(i) one or more hidden states comprising a copy number corresponding to an interrogated segment or a plurality of sub-segments within the interrogated segment;
(ii) an observation state comprising the real or synthetic number of sequencing reads for the interrogated segment;
(iii) a copy number likelihood model based on an expected number of real or synthetic sequencing reads for the interrogated segment.
Embodiment 10. The method of embodiment 9, comprising determining the copy number likelihood model.
Embodiment 11. The method of embodiment 9 or 10, wherein parameterizing the hidden Markov model comprises adjusting the copy number likelihood model to fit the real number of sequencing reads mapped to the interrogated segment, from the test sample.
Embodiment 12. The method of any one of embodiments 9-11, wherein the copy number likelihood model comprises a distribution for two or more copy number states.
Embodiment 13. The method of any one of embodiments 9-12, wherein the copy number likelihood model comprises a negative binomial distribution, wherein the negative binomial distribution is not a Poisson distribution.
Embodiment 14. The method of any one of embodiments 9-13, wherein the expected number of real or synthetic sequencing reads is based on an average number of mapped sequencing reads at a segment corresponding to the interrogated segment across a plurality of samples, and an average number of mapped sequencing reads across the segments within the test sample, wherein the average number of mapped sequencing reads at the segment corresponding to the interrogated segment across the plurality of samples or the average number of mapped sequencing reads across the plurality of segments within the test sample is a normalized average.
Embodiment 15. The method of any one of embodiments 9-14, wherein the copy number likelihood model is adjusted to account for the presence of GC content bias.
Embodiment 16. The method of any one of embodiments 9-15, wherein the hidden Markov model comprises a transition probability of the copy number of the interrogated segment for a given copy number of a spatially adjacent segment.
Embodiment 17. The method of any one of embodiments 9-15, wherein the hidden Markov model comprises a plurality of transition probabilities of the copy number of a sub-segment in the plurality of sub-segments within the interrogated segment for a given copy number of a spatially adjacent sub-segment.
Embodiment 18. The method of embodiment 16 or 17, wherein the transition probability accounts for an average length of a copy number variant.
Embodiment 19. The method of any one of embodiments 16-18, wherein the transition probability accounts for a prior probability of a copy number variant at the interrogated segment or a spatially adjacent segment.
Embodiment 20. The method of embodiment 18 or 19, wherein the average length of a copy number variant or the probability of a copy number variant at the interrogated segment is determined based on observations in a human population.
Embodiment 21. The method of any one of embodiments 1-20, wherein parameterizing the copy number variant model comprises accounting for one or more spurious capture probes.
Embodiment 22. The method of embodiment 21, wherein accounting for one or more spurious capture probes comprises weighting the one or more observation states in the plurality of observation states with a spurious capture probe indicator.
Embodiment 23. The method of embodiment 22, wherein the spurious capture probe indicator is determined using a Bernoulli process.
Embodiment 24. The method of embodiment 22 or 23, wherein accounting for one or more of the capture probes being spurious comprises using expectation-maximization.
Embodiment 25. The method of any one of embodiments 21-24, wherein if a capture probe is determined to be spurious, sequencing reads derived from that capture probe is disregarded in the copy number variant model.
Embodiment 26. The method of any one of embodiments 1-25, wherein the parameterizing of the copy number variant model comprises accounting for noise in the number of mapped sequencing reads.
Embodiment 27. The method of any one of embodiments 1-26, wherein the copy number variant model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more copy number variant model parameters.
Embodiment 28. The method of any one of embodiments 1-27, wherein the copy number variant model is parameterized by solving a trust region Newton conjugate gradient algorithm.
Embodiment 29. The method of any one of embodiments 1-28, wherein the copy number variant model is iteratively parameterized using expectation-maximization.
Embodiment 30. The method of any one of embodiments 1-29, comprising mapping the real sequencing reads from the test sample to the segments within the region of interest, and determining the real numbers of sequencing reads mapped to the segments.
Embodiment 31. The method of any one of embodiments 1-30, wherein the test sample is enriched using one or more direct targeted sequencing capture probes.
Embodiment 32. The method of any one of embodiments 1-31, comprising calling a copy number of the one or more segments for the test sample.
Embodiment 33. The method of any one of embodiments 1-32, wherein the segments comprise spatially adjacent segments.
Embodiment 34. The method of any one of embodiments 1-33, wherein the sample-specific performance statistic is a limit of detection, sensitivity, specificity, precision, recall, accuracy, positive predictive value, or negative predictive value.
Embodiment 35. The method of any one of embodiments 1-34, wherein the sample-specific performance statistic is sensitivity or accuracy.
Embodiment 36. The method of any one of embodiments 1-35, comprising failing the test sample if the sample-specific performance of the copy number variant model is below a desired performance threshold.
Embodiment 37. A method for determining a copy number of an interrogated segment within a region of interest comprising:
(a) mapping a plurality of sequencing reads generated from a test sequencing library to the interrogated segment, wherein the test sequencing library is enriched using one or more direct targeted sequencing capture probes;
(b) determining a number of sequencing reads mapped to the interrogated segment;
(c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment;
(d) building a hidden Markov model comprising:
(e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and
(f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Embodiment 38. A method for determining a copy number of an interrogated segment within a region of interest comprising:
(a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises the interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes;
(b) determining a number of sequencing reads mapped to each spatially adjacent segment;
(c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment;
(d) building a hidden Markov model comprising:
(e) parameterizing the hidden Markov model, comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and
(f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Embodiment 39. The method of embodiment 37 or 38, wherein the one or more parameters of the copy number likelihood model comprises a dispersion of a number of mapped sequencing reads for the segment (di), an average number of mapped sequencing reads for the segment (μt), a dispersion of a number of mapped sequencing reads for the segments within the test sequencing library (dj), or an average number of mapped sequencing reads for the segments within the test sequencing library (μj).
Embodiment 40. The method of any one of embodiments 37-39, further comprising determining a most probable copy number of a section within the region of interest, wherein the section comprises a plurality of spatially adjacent segments comprising the interrogated segment.
Embodiment 41. The method of any one of embodiments 37-40, wherein the copy number likelihood model comprises a distribution for two or more copy number states.
Embodiment 42. The method of any one of embodiments 37-41, wherein the copy number likelihood model comprises a negative binomial distribution, wherein the negative binomial distribution is not a Poisson distribution.
Embodiment 43. The method of any one of embodiments 37-42, wherein the expected number of sequencing reads is based on an average number of mapped sequencing reads at a corresponding segment across a plurality of sequencing libraries and an average number of mapped sequencing reads across a plurality of segments of interest within the test sequencing library, wherein the average number of mapped sequencing reads at a corresponding segment across a plurality of sequencing libraries or the average number of mapped sequencing reads across a plurality of segments of interest within the test sequencing library is a normalized average.
Embodiment 44. The method of any one of embodiments 37-43, wherein the copy number likelihood model is adjusted to account for the presence of GC content bias.
Embodiment 45. The method of embodiment 44, wherein the adjustment depends on the GC content of the capture probe corresponding to the interrogated segment or the GC content of the interrogated segment.
Embodiment 46. The method of any one of embodiments 37-45, wherein the hidden Markov model comprises a transition probability of the copy number of the interrogated segment for a given copy number of a spatially adjacent segment.
Embodiment 47. The method of any one of embodiments 37-45, wherein the hidden Markov model comprises a plurality of transition probabilities of the copy number of a sub-segment in the plurality of sub-segments within the interrogated segment for a given copy number of a spatially adjacent sub-segment.
Embodiment 48. The method of embodiment 46 or 47, wherein the transition probability accounts for an average length of a copy number variant.
Embodiment 49. The method of any one of embodiments 46-48, wherein the transition probability accounts for a prior probability of a copy number variant at the interrogated segment or a spatially adjacent segment.
Embodiment 50. The method of embodiment 48 or 49, wherein the average length of a copy number variant or the probability of a copy number variant at the interrogated segment are determined based on observations in a human population.
Embodiment 51. The method of any one of embodiments 37-50, wherein parameterizing the hidden Markov model comprises accounting for one or more spurious capture probes.
Embodiment 52. The method of embodiment 51, wherein accounting for one or more spurious capture probes comprises weighting the one or more observation states in the plurality of observation states with a spurious capture probe indicator.
Embodiment 53. The method of embodiment 52, wherein the spurious capture probe indicator is determined using a Bernoulli process.
Embodiment 54. The method of embodiment 52 or 53, wherein accounting for one or more of the capture probes being spurious comprises using expectation-maximization.
Embodiment 55. The method of any one of embodiments 52-54, wherein if a capture probe is determined to be spurious, the likelihood information from that capture probe is disregarded in the copy number likelihood model.
Embodiment 56. The method of any one of embodiments 37-55, wherein the parameterizing of the hidden Markov model comprises accounting for noise in the number of mapped sequencing reads.
Embodiment 57. The method of any one of embodiments 37-56, wherein accounting for noise in the number of mapped sequencing reads comprises adjusting the copy number likelihood model.
Embodiment 58. The method of embodiment 57, wherein adjusting the copy number likelihood model to account for the noise comprises an expectation-maximization step.
Embodiment 59. The method of embodiment 58, wherein the expectation-maximization step comprises weighing a level of noise in the number of mapped sequencing reads from the test sequencing library.
Embodiment 60. The method of any one of embodiments 56-59, wherein the most probable copy number of the interrogated segment is not called if the noise in the number of mapped sequencing reads is above a predetermined threshold.
Embodiment 61. The method of any one of embodiments 37-60, wherein sequencing reads from overlapping capture probes are merged.
Embodiment 62. The method of any one of embodiments 37-61, wherein a Viterbi algorithm, a Quasi-Newton solver, or a Markov chain Monte Carlo is used to determine the most probable copy number of the interrogated segment.
Embodiment 63. The method of any one of embodiments 37-62, further comprising determining a confidence of the most probable copy number of the segment.
Embodiment 64. A method for determining a copy number variant abnormality within a region of interest, comprising:
(a) mapping a plurality of sequencing reads generated from a test sequencing library to an interrogated segment within the region of interest, wherein the test sequencing library is enriched using one or more direct targeted sequencing capture probes;
(b) determining a number of sequencing reads mapped to the interrogated segment;
(c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment;
(d) building a hidden Markov model comprising:
(e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and
(f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model;
(g) determining a copy number variant abnormality based on the most probable copy number of the interrogated segment.
Embodiment 65. A method for determining a copy number variant abnormality within a region of interest, comprising:
(a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises an interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes;
(b) determining a number of sequencing reads mapped to each spatially adjacent segment;
(c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment;
(d) building a hidden Markov model comprising:
(e) parameterizing the hidden Markov model comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and
(f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model;
(g) determining a copy number variant abnormality based on the most probable copy number of the interrogated segment.
Embodiment 66. A method for determining a copy number of an interrogated segment within a region of interest comprising:
(a) mapping a plurality of sequencing reads generated from a test sequencing library to the interrogated segment, wherein the test sequencing library is enriched using one or more capture probes;
(b) determining a number of sequencing reads mapped to the interrogated segment;
(c) determining a copy number likelihood model based on an expected number of sequencing reads mapped to the interrogated segment;
(d) building a hidden Markov model comprising:
(e) parameterizing the hidden Markov model by adjusting the copy number likelihood model to fit the determined number of sequencing reads mapped to the interrogated segment and accounting for one or more spurious capture probes, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and
(f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Embodiment 67. A method for determining a copy number of an interrogated segment within a region of interest comprising:
(a) mapping a plurality of sequencing reads generated from a test sequencing library to a plurality of spatially adjacent segments, wherein the plurality of spatially adjacent segments comprises the interrogated segment, and wherein the test sequencing library is enriched using a plurality of spatially adjacent direct targeted sequencing capture probes;
(b) determining a number of sequencing reads mapped to each spatially adjacent segment;
(c) determining a copy number likelihood model for each spatially adjacent segment based on an expected number of mapped sequencing reads at the spatially adjacent segment;
(d) building a hidden Markov model comprising:
(e) parameterizing the hidden Markov model comprising adjusting each copy number likelihood model to fit the determined number of sequencing reads mapped to each spatially adjacent segment and accounting for one or more spurious capture probes, wherein the hidden Markov model is parameterized using an analytic first derivative gradient and second derivative Hessian of one or more parameters in the copy number likelihood model; and
(f) determining a most probable copy number of the interrogated segment based on the parameterized hidden Markov model.
Embodiment 68. The method of any one of embodiments 64-67, wherein the one or more parameters of the copy number likelihood model comprises a dispersion of a number of mapped sequencing reads for the segment (di), an average number of mapped sequencing reads for the segment (μi), a dispersion of a number of mapped sequencing reads for the segments within the test sequencing library (dj), or an average number of mapped sequencing reads for the segments within the test sequencing library (μj).
Embodiment 69. The method of any one of embodiments 37-68, wherein the analytic first derivative gradient and second derivative analytical Hessian of the one or more parameters in the copy number likelihood model is solved using a trust region Newton conjugate gradient algorithm
Embodiment 70. A computer system comprising a computer-readable medium comprising instructions for carrying out the method of any one of embodiments 1-68.
Biological samples from blood or saliva were sequenced using an Illumina platform HiSeq2500 after direct targeted sequencing enrichment across a panel of 178 genes. Batches of 46 samples were analyzed, with the batches containing different proportion of saliva and blood samples. Saliva samples generally produce noisier sequencing results and can affect the sensitivity other samples in the same flowcell batch. The numbers of sequencing read from segments within each sample were used to parameterize a hidden Markov model for the segments, generate 400 synthetic copy number variants, and calling a number of copies of segments within the synthetic copy number variants for each sample using the hidden Markov model caller. The hidden Markov model included (i) hidden states of for the copy number for a given segment, (ii) an observation state with the synthetic number of sequencing reads for the given segment, and (iii) a copy number likelihood model based on the number of synthetic reads for the given segment. The sensitivity for each test sample was determined using the called number of copies for the segments within the synthetic variants and the actual number of copies within the synthetic variants.
The copy number variant call analysis was conducted using two different hidden Markov model callers. In the reference hidden Markov model caller, sample noise (i.e., the dispersion due to noise within the test sequencing library) and spurious capture probe noise was ignored. In the test hidden Markov model, noise within the test sequencing library was accounted for in the copy number likelihood model by including a parameter for dispersion due to noise across the segments in the sequencing library (di) in the dispersion parameter according to: d=ed
The determined sensitivities for each sample are shown in
The present application is a continuation of International Patent Application No. PCT/US2019/034998, filed May 31, 2019, which claims priority to U.S. Provisional Application No. 62/681,517, filed Jun. 6, 2018, and to U.S. Provisional Application No. 62/733,842, filed Sep. 20, 2018, each of which is hereby incorporated in its entirety including all tables, figures, and claims.
Number | Date | Country | |
---|---|---|---|
62681517 | Jun 2018 | US | |
62733842 | Sep 2018 | US |
Number | Date | Country | |
---|---|---|---|
Parent | PCT/US2019/034998 | May 2019 | US |
Child | 17111272 | US |