Method and system for identifying genetic abnormalities in blood

Abstract

The present invention discloses a method for fetal genotyping, comprising receiving maternal genomic DNA (gDNA) data, maternal cell-free DNA (cfDNA) data and paternal gDNA data that are paired with a fetus, wherein the data are analyzed to identify a first set of multiple sites and a second set of multiple sites, wherein multiple different alleles of the parents are homozygous at the first set of multiple sites, and at the second set of multiple sites, at least one of the parents has a mutation; for each site of the first set, determining a probability that a corresponding portion of the maternal cell-free DNA data is obtained from the fetus, and classifying each site of the second set as fetal or maternal according to the multiple probabilities to perform genotyping on the fetus.

Description

[0001] Related Applications

[0002] This application claims the benefit of priority to U.S. Provisional Patent Application No. 62 / 726,337, filed on September 3, 2018, the entire contents of which are incorporated by reference into this specification.

[0003] Technical field and background technology

[0004] The present invention, in some embodiments, relates to bioinformatics, and more particularly, but not limited to, to a method and system for identifying genetic diseases in maternal blood.

[0005] Determining the copy number of multiple genetic sequences in a fetus has important diagnostic value. For example, in a dominant genetic disease, the presence of a single copy of an allele that causes the disease leads to the expression of the phenotype of the genetic disease. In contrast, in a recessive genetic disease, the presence of a single copy of the allele that causes the disease only causes the individual to become a carrier, but does not cause the expression of the phenotype of the genetic disease. In addition, multiple abnormal copy numbers of multiple genetic sequences, such as partial or complete aneuploidy of chromosome segments or entire chromosomes, often cause various genetic diseases. For example, trisomy 21 can cause Down syndrome (DS).

[0006] Currently, prenatal diagnosis is performed using conventional cytogenetic analysis (e.g., karyotyping) or DNA analysis, which requires obtaining fetal genetic material through amniocentesis, chorionic villus sampling, or cordocentesis. However, these are invasive procedures and are associated with risks, such as a risk of fetal loss.

[0007] The advent of multiple technologies that allow sequencing of entire genomes in a relatively short period of time and the discovery of circulating cell-free DNA (cfDNA) in the blood of the pregnant woman, including both maternal and fetal DNA, provides an opportunity for non-invasive analysis of fetal genetic material without the multiple risks associated with multiple invasive sampling methods. Multiple non-invasive prenatal diagnostic techniques are known in the art for identifying multiple specific chromosomal abnormalities and for identifying multiple single-gene diseases of paternal origin. Multiple non-invasive prenatal diagnostic techniques are also known for excluding multiple single-gene diseases of maternal origin, fetal sex determination, and fetal rhesus D genotyping. Summary of the invention

[0008] According to some embodiments of the present invention, the present invention provides a method for fetal genotyping, the method comprising the following steps: receiving maternal genomic DNA (genomic DNA, gDNA) data, maternal cell-free DNA (cell-free DNA, cfDNA) data and paternal genomic DNA data that are paired with a fetus; analyzing the data to identify a first set of multiple sites and a second set of multiple sites, at the first set of multiple sites, multiple different alleles of the parents are homozygous, and at the second set of multiple sites, at least one of the parents has a mutation; for each site of the first set, determining a probability that a corresponding portion of the maternal cell-free DNA data is obtained from the fetus; and classifying each site of the second set as fetal or maternal according to the multiple probabilities to perform genotyping on the fetus.

[0009] According to some embodiments of the present invention, the analyzing step includes: identifying a first group of multiple reads and a second group of multiple reads among multiple reads covering multiple sites of the first set, wherein the multiple reads of the first group only include multiple reads representing multiple paternal alleles but do not include other reads, and the multiple reads of the second group include all other reads, and wherein the step of determining the probability is based on multiple differences between the multiple reads in the first group and the multiple reads in the second group.

[0010] According to some embodiments of the invention, the step of determining the probability is based on at least one sequence alignment map (SAM) parameter.

[0011] According to some embodiments of the present invention, the sequence alignment map parameter is selected from the group consisting of observed template length, length-related fetal fraction, tight heterogeneous gap alignment report string, CIGAR string of paired segment, nucleotide sequence, nucleotide sequence of paired segment, read alignment flag indicating whether it is a paired read, read alignment flag indicating whether a read is correctly mapped to a paired read, read alignment flag indicating whether it is an unmapped read, read alignment flag indicating whether it is an unmapped paired segment, read alignment flag indicating whether it is a read on the reverse strand, read alignment flag indicating whether it is a paired read on the reverse strand, a read alignment mark of a paired segment, a read alignment mark indicating whether it is the first paired read, a read alignment mark indicating whether it is the second paired read, a read alignment mark indicating whether it is a non-primary alignment, a read alignment mark indicating whether it is a platform and / or a supplier quality check failed, a read alignment mark indicating whether it is a PCR or optical repeat, a read alignment mark indicating whether it is a supplementary alignment, a marker for paired segments, mapping quality, mapping quality of paired segments, multiple genomic coordinates of chromosomes, multiple genomic coordinates of absolute starting positions on chromosomes, staining a plurality of genomic coordinates of the absolute end position on the chromosome, a plurality of genomic coordinates of the start position normalized to the chromosome length, a plurality of genomic coordinates of the end position normalized to the chromosome length, a plurality of genomic coordinates of the paired segment, the number of G and C nucleotides divided by the read length, the number of G and C nucleotides divided by the read length in the read sequence of the paired segment, the ratio of A and / or C and / or G and / or T nucleotides in a nucleotide sequence of the read, the ratio of A and / or C and / or G and / or T nucleotides in the nucleotide sequence of the paired segment, information about the origin of the read or the paired segment of the read The information of a variation of a nucleotide sequence, including at least one feature appearing in Table A.1, the composition of Kmers in the nucleotide sequence, the composition of multiple Kmers in the nucleotide sequence of the partner, the nucleotide quality sequence, the nucleotide quality sequence of the paired segment, the average value and / or standard deviation and / or median of the nucleotide quality sequence, the average value and / or standard deviation and / or median of the nucleotide quality sequence, the Kmer composition in the nucleotide quality sequence, the Kmer composition in the nucleotide quality sequence of the paired segment, the number of multiple methylated nucleotides divided by the read length, and the group consisting of methylation at multiple specific positions.

[0012] According to some embodiments of the invention, the step of determining the probability is based on at least an observed template length.

[0013] According to some embodiments of the invention, the step of determining the probability is based on at least a fetal fraction associated with a length.

[0014] According to some embodiments of the invention, the step of determining the probability is based on at least one close heterogeneous gap alignment report string.

[0015] According to some embodiments of the invention, the step of determining the probability is based on at least a nucleotide sequence.

[0016] According to some embodiments of the invention, the step of determining the probability is based at least on the nucleotide sequence of the paired segments.

[0017] According to some embodiments of the present invention, the determination may be based on at least one read alignment mark, wherein the read alignment mark is selected from the group consisting of a read alignment mark indicating whether it is a paired read, a read alignment mark indicating whether a read is mapped correctly to a pair, a read alignment mark indicating whether it is an unmapped read, a read alignment mark indicating whether it is an unmapped paired segment, a read alignment mark indicating whether it is a read on the reverse strand, a read alignment mark indicating whether it is a paired segment on the reverse strand, a read alignment mark indicating whether it is the first paired read, a read alignment mark indicating whether it is the second paired read, a read alignment mark indicating whether it is a read that is not a primary alignment, a read alignment mark indicating whether it is a read that fails a platform and / or a supplier quality check, a read alignment mark indicating whether it is a PCR or optical repeat, and a read alignment mark indicating whether it is a supplementary alignment.

[0018] According to some embodiments of the present invention, the method further comprises the step of calculating a total fetal fraction, wherein the step of classifying further comprises using the total fetal fraction.

[0019] According to some embodiments of the present invention, the method further comprises calculating a total fetal fraction, and constructing a fetal size distribution and a maternal size distribution, wherein the classification step comprises binning the fetal size distribution, and calculating a fetal fraction for each segment size bin, and calculating a probability that the segment is a fetus for at least one site and at least one segment at the at least one site based on a fetal fraction of a corresponding segment size bin to which the segment belongs.

[0020] According to some embodiments of the invention, the step of classifying comprises applying a Bayesian method.

[0021] According to some embodiments of the invention, the Bayesian method comprises a plurality of prior probabilities calculated using sequencing data of at least one of the parents.

[0022] According to some embodiments of the present invention, the sequencing data includes whole genome sequencing (WGS).

[0023] According to some embodiments of the present invention, the sequencing data includes whole exome sequencing (WES).

[0024] According to some embodiments of the invention, the method further comprises using machine learning to recalibrate the output of the Bayesian method.

[0025] According to some embodiments of the invention, the step of classifying comprises applying a machine learning procedure to the plurality of identified loci to genotype the fetus.

[0026] According to some embodiments of the invention, the machine learning program comprises a deep learning program.

[0027] According to some embodiments of the invention, the method further comprises recalibrating the output of the machine learning.

[0028] According to one aspect of some embodiments of the present invention, a method for fetal genotyping is provided, the method comprising the following steps: receiving maternal genomic DNA (gDNA) data, maternal cell-free DNA data (cfDNA), and paternal genomic DNA data that are paired with a fetus; accessing a computer-readable medium storing a machine deep learning program, the machine deep learning program being trained to classify multiple sites in the data as fetal or maternal; providing the data to the method; and receiving an output from the program indicating multiple fetal sites in the data, thereby genotyping the fetus.

[0029] According to some embodiments of the present invention, the method is applied to multiple heterozygous sites of the maternal cell-free DNA data.

[0030] According to some embodiments of the present invention, the method is applied to multiple homozygous sites of the maternal cell-free DNA data.

[0031] According to some embodiments of the invention, the genotyping step comprises identifying fetal single-gene disorders (SGDs) originating from the paternal line.

[0032] According to some embodiments of the invention, the step of genotyping comprises identifying fetal single gene diseases (SGDs) of maternal origin.

[0033] According to some embodiments of the invention, the genotyping step comprises predicting inherited multiple insertions and multiple deletions.

[0034] According to some embodiments of the invention, the genotyping step comprises identifying a fetal chromosomal abnormality.

[0035] According to some embodiments of the invention, the method is applied to perform non-invasive prenatal diagnosis (NIPD) of a single gene disease.

[0036] According to some embodiments of the present invention, the method is applied to perform non-invasive prenatal diagnosis (NIPD) of a polygenic disease.

[0037] According to some embodiments of the present invention, the method is applied to perform non-invasive prenatal diagnosis (NIPD) of a genetic disease, wherein the genetic disease is selected from the group consisting of Bloom Syndrome, Canavan Disease, Cystic fibrosis, Familial Dysautonomi, Riley-Day syndrome, Fanconi Anemia (Group C), Gaucher Disease, Glycogen storage disease type 1a, Maple syrup urine disease, Mucolipidosis IV, Niemann-Pick Disease, Tay-Sachs disease, Beta thalassemia, Sickle cell anemia, Alpha thalassemia, and thalessemia, Beta thalessemia, Factor XI Deficiency, Friedreich's Ataxia, MCAD, Parkinson disease-juvenile, Connexin26, SMA, Rett syndrome, Phenylketonuria, Becker Muscular Dystrophy, Duchennes Muscular Dystrophy, Fragile X syndrome, Hemophilia A, Alzheimer dementia-early onset, Breast / Ovarian cancer, Colon cancer, Diabetes / Maturity-onset Diabetes of the Juvenile (MODY), Huntington disease disease), myotonic dystrophyMuscular Dystrophy, Parkinson Disease-early onset, Peutz-Jeghers syndrome, Polycystic Kidney Disease, and Torsion Dystonia.

[0038] According to some embodiments of the invention, the method is applied to perform non-invasive prenatal diagnosis (NIPD) for a genetic disease, and the method further comprises performing prenatal or postnatal treatment for the genetic disease.

[0039] According to some embodiments of the invention, the method is applied to perform non-invasive prenatal diagnosis (NIPD) for a congenital disease.

[0040] According to some embodiments of the present invention, the congenital disease is selected from the group consisting of malformation, neural tube defect, chromosomal abnormality, Down syndrome (or trisomy 21), Trisomy 18, spina bifida, cleft palate, Tay Sachs disease, sickle cell anemia, thalassemia, cystic fibrosis, Huntington's disease and fragile x syndrome.

[0041] According to some embodiments of the invention, the chromosomal abnormality is selected from Down syndrome (an extra chromosome 21), Turner syndrome (45X0) and Klinefelter's syndrome (a male has 2 X chromosomes).

[0042] According to some embodiments of the invention, the deformity comprises a limb deformity.

[0043] According to some embodiments of the present invention, the limb deformity is selected from the group consisting of amelia, ectrodactyly, phocomelia, polymelia, polydactyly, syndactyly, polysyndactyly, oligodactyly, brachydactyly, achondroplasia, congenitalaplasia or hypoplasia, amniotic band syndrome and cleidocranial dysostosis.

[0044] According to some embodiments of the invention, the malformation comprises a congenital malformation of the heart.

[0045] According to some embodiments of the present invention, the congenital malformation is selected from the group consisting of patent ductus arteriosus, atrial septal defect, ventricular septal defect and tetralogy of Fallot.

[0046] According to some embodiments of the invention, the malformation comprises a congenital malformation of the nervous system.

[0047] According to some embodiments of the present invention, the congenital malformation of the nervous system is selected from the group consisting of multiple neural tube defects (e.g., spina bifida, meningocele, meningomyelocele, encephalocele and anencephaly), Arnold-Chiari malformation, Dandy-Walker malformation, hydrocephalus, microencephaly, megencephaly, lissencephaly, polymicrogyria, holoprosencephaly and agenesis of the corpus callosum.

[0048] According to some embodiments of the invention, the malformation comprises a congenital malformation of the gastrointestinal system.

[0049] According to some embodiments of the invention, the congenital malformation of the gastrointestinal system is selected from the group consisting of stenosis, atresia and imperforate anus.

[0050] According to some embodiments of the invention, the method further comprises prenatal or postnatal treatment of the congenital disease.

[0051] According to some embodiments of the invention, the method further comprises performing prenatal or postnatal treatment of the deformity.

[0052] According to some embodiments of the invention, the method further comprises administering a prenatal or postnatal treatment selected from the group consisting of drug-based intervention, surgery, gene therapy, nutritional therapy, and combinations thereof.

[0053] According to some embodiments of the invention, the method further comprises performing a pregnancy termination.

[0054] According to some embodiments of the invention, the method further comprises obtaining fetal genetic material when the genotyping indicates a fetal abnormality or a genetic abnormality, and analyzing the fetal genetic material to determine whether at least the fetus has the abnormality or the genetic abnormality.

[0055] According to some embodiments of the present invention, a computer software product is provided, which includes a computer-readable medium, in which a plurality of program instructions are stored, and when the plurality of instructions are read by a data processor, the data processor receives maternal cell-free DNA data and paternal cell-free DNA data paired with a fetus, so as to execute the method as described above, and preferably as further described in detail below.

[0056] According to some embodiments of the present invention, a fetal genotyping system is provided, the fetal genotyping system comprising: an input circuit for receiving maternal cell-free DNA (cfDNA) data and paternal cell-free DNA (cfDNA) data paired with a fetus; and a data processor configured to analyze the data to identify multiple sites where the parents are homozygous, so as to perform the method as described above, and preferably as described in further detail below.

[0057] Unless otherwise defined, all technical terms and / or scientific terms used herein have the same meaning as those generally understood by those of ordinary skill in the art to which the invention belongs. Although embodiments of the invention may be implemented or tested by any method and material similar or equivalent to those described in embodiments of the invention, embodiments of the invention, the methods and / or materials listed are described below. In the event of a conflict, this patent specification and including definitions will prevail. In addition, materials, methods and embodiments are exemplary only and are not necessarily intended to be limiting.

[0058] The implementation of the method and / or system of the present invention involves performing or completing selected multiple tasks manually, automatically, or a combination thereof. In addition, actual instruments and devices according to multiple embodiments of the method and / or system of the present invention can use an operating system to implement several selected steps through hardware, software, or firmware, or a combination thereof.

[0059] For example, the hardware for performing selected tasks according to an embodiment of the present invention can be implemented as a chip or a circuit. As software, the selected tasks of the present invention can be implemented as multiple software instructions executed by any suitable operating system used by the computer. In an exemplary embodiment of the present invention, multiple selected one or more tasks of the method and / or system according to the present invention can be described as being executed by a data processor, such as: a computing platform for executing multiple instructions. Optionally, the data processor includes a volatile memory for storing multiple instructions and / or data, and / or a non-volatile memory, such as: a magnetic hard disk and / or removable media. Optionally, the present invention also provides a network connection, and the present invention also optionally provides a display and / or a user input device, such as: a keyboard or a mouse. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Some embodiments of the present invention are described herein by way of example only and with reference to a number of drawings. By describing in detail the specific references to the drawings, it should be emphasized that the details shown are only examples for the purpose of illustrating the embodiments of the present invention. Based on this, the drawings and description together enable those skilled in the art to clearly implement the embodiments of the present invention.

[0061] In the attached picture:

[0062] Figure 1 is a schematic diagram of a pipeline for non-invasive prenatal variation detection according to some embodiments of the present invention.

[0063] Figures 2A to 2D shows the relationship between the accuracy and multiple posterior probabilities obtained in multiple experiments conducted according to some embodiments of the present invention, Figures 2A to 2C Multiple results for families G1 to G4 in three categories of single nucleotide polymorphism (SNP) positions determined by parental genotypes and subsampled data for G1 ("multiple simulations") are shown. The x-axis shows the minimum threshold for the posterior probability of the multiple predictions, the threshold indicating the deterministic level of the included loci, and the y-axis indicates the accuracy among all sites where the posterior probability is above the threshold. The overall accuracy appears at the lowest point on the x-axis, where the lowest possible threshold for the maximum posterior probability is set (at Figure 2A and 2B In , each locus has two possible fetal genotypes, the value is 0.5; and Figure 2C , the value is 0.33, a multiple because all 3 fetal genotypes are possible), and the multiple counts of total loci for which the accuracy was calculated at each threshold are also shown.

[0064] Figures 3A to 3D Shown is the alignment of multiple insertion / deletion sites when a mother is heterozygous, as obtained in multiple experiments performed according to some embodiments of the present invention. Figures 3A to 3D The format is similar to Figures 2A to 2D , but for multiple indels. Accuracies between multiple loci are shown together with a posterior probability of being above the multiple thresholds appearing on the x-axis. Figure 3A The same three categories of genomic loci are shown in C3, and the total accuracy of each category is the accuracy of the leftmost point on the x-axis ( Figure 3A and 3B The middle is 0.5, Figure 3C The bottom of each sub-plot shows the counts of loci used for accuracy calculation at each threshold.

[0065] Figures 4A to 4C The results obtained from multiple experiments performed according to some embodiments of the present invention are shown. Figures 2A to 3D Heat maps of accuracy as a function of a sequencing depth and a fetal fraction at the three types of loci are shown.

[0066] Figures 5A to 5G The performance of first trimester cases relative to SNPs and indels as obtained in experiments performed according to some embodiments of the present invention is shown. Accuracies for loci are shown along with a posterior probability of being above the thresholds appearing on the x-axis. ... is shown according to the previously used loci categories. Figures 5A to 5C ) and multiple indels (Figures D to F) in the multiple families sequenced in the multiple experiments, the overall accuracy for each category is the accuracy of the leftmost point on the x-axis (0.5 in A to B and D to E, 0.3 in C and F), and the multiple total counts calculated from the accuracy for each threshold are also shown.

[0067] Figures 6A to 6F Receiver operating characteristic (ROC) curves of variant probability recalibration as obtained in multiple experiments performed according to some embodiments of the present invention are shown, and multiple ROC curves of the two test sets are shown for multiple same previously used locus categories before and after a machine learning-based variant recalibration step. For each curve, the micro-average area under the curve (AUC) and the overall accuracy (ACC) are also shown.

[0068] Figures 7A to 7G The calibration of the accuracy as obtained in multiple experiments performed according to some embodiments of the present invention is shown, showing the accuracy as a function of the predicted probability before and after variable recalibration. The "perfect calibration" line represents a situation where the probability assigned to a prediction perfectly describes the correct probability (in terms of accuracy), and the marginal distribution of loci in each bin is also shown.

[0069] Figures 8A to 8G Demonstrating the ability to perform machine learning based variant recalibration to filter variants, accuracies for multiple loci are shown along with a posterior probability of being above the thresholds on the x-axis. Findings are provided both before and after variant recalibration. The total accuracy for each category is the accuracy of the leftmost point on the x-axis, and the total counts calculated from the accuracy for each threshold are also shown.

[0070] Figures 9A to 9GShown are multiple cell-free DNA fragment length distributions obtained in multiple experiments performed according to some embodiments of the present invention, listing the length distributions of multiple fetal (blue) and maternal (green) fragments of multiple families included in the multiple experiments.

[0071] Fig.10 Multiple mutation validation Sanger results obtained in multiple experiments performed according to some embodiments of the present invention are shown, using Sanger to confirm multiple mutation families E1 and G5. The first row shows a control group homozygous with the reference allele (G), the second row shows the father, who is a carrier of the variant allele (T), and the third and fourth rows show the multiple results of the fetus, once using a forward primer and once using a reverse primer, and the fetus is homozygous with the variant allele.

[0072] like Fig.11 is a summary of a Bayesian model (in Fig.11 Hoobari in Fig.11 A diagram of the multiple differences between the two (referred to as Deep Hobari).

[0073] Figures 12A to 12P Results from a first phase of an experiment performed using a machine deep learning program according to some embodiments of the present invention are shown, showing accuracy and loss results for multiple SNPs in the sub-sampled dataset G1 using four tested networks.

[0074] Figures 13A to 13P Shown are multiple one-hot and group-hot encoding results of an experiment performed using a machine deep learning program according to some embodiments of the present invention.

[0075] Figures 14A to 14P Results of a first phase of an experiment conducted using a machine deep learning program according to some embodiments of the present invention when only genotypes in parent information are used are shown.

[0076] Fig.15 is a flow chart of a method suitable for fetal genotyping according to various exemplary embodiments of the present invention.

[0077] Fig.16 is a schematic diagram of a server-client computer configuration that may be used for fetal genotyping according to some embodiments of the present invention. DETAILED DESCRIPTION

[0078] The present invention, in some embodiments, relates to bioinformatics, and more particularly, but not limited to, to a method and system for identifying genetic diseases in maternal blood.

[0079] Before explaining at least one embodiment of the present invention in detail, it should be understood that the present invention is not necessarily limited to the details of the construction of the present invention, and the arrangement of components and / or methods described in the following description and / or drawings and / or embodiments, and the present invention can be implemented or applied in other embodiments or in various methods.

[0080] Fig.15 is a flow chart of a method suitable for fetal genotyping according to various exemplary embodiments of the present invention. It should be understood that, unless otherwise defined, the multiple operations described below may be performed simultaneously or sequentially in many combinations or execution orders. Specifically, the order of the multiple flow charts should not be considered restrictive. For example, two or more operations that appear in a particular order in the following description or the multiple flow charts may be performed in a different order (e.g., a reverse order) or substantially simultaneously. In addition, several operations described below are optional and may not be performed.

[0081] The multiple processing operations of the multiple embodiments of the present invention can be implemented in multiple forms, for example: the multiple processing operations can be implemented in a tangible medium, such as: a computer for performing the multiple operations, the multiple processing operations can be implemented on a computer-readable medium, and the computer-readable medium includes multiple computer-readable instructions for performing the multiple method operations. The multiple processing operations can also be implemented in an electronic device having multiple digital computer functions, and the digital computer functions are arranged to run the computer program on the tangible medium or execute the instructions on a computer-readable medium.

[0082] Multiple computer programs implementing the method according to some embodiments of the present invention can generally be distributed to multiple users on a distribution medium, such as but not limited to: CD-ROM, multiple flash memory devices, multiple flash drives, or in some embodiments, multiple drives that can be accessed through the Internet (for example: in a cloud environment) or through a cellular network for network communication. The multiple computer programs can be copied from the distribution medium to a hard disk or a similar intermediate storage medium. The multiple computer programs can be run by loading the multiple computer instructions from the distribution medium of the multiple computer instructions or the intermediate storage medium of the multiple computer instructions into the execution memory of the computer, and configuring the computer so that the computer operates according to the method of the present invention. Multiple computer programs implementing the method according to some embodiments of the present invention can also be executed by one or more data processors belonging to a cloud computing environment. All of the multiple operations are known to multiple technicians in the field of computer systems. The data used and / or provided by the method of the present multiple embodiments can be transmitted by means of network communication through the Internet, through a cellular network, or through any network type suitable for data transmission.

[0083] The method according to the preferred embodiments of the present invention can be embedded in a plurality of healthcare systems and can allow identification of a plurality of fetal abnormalities or diseases, such as but not limited to: a plurality of fetal single gene diseases (SGDs) originating from the father and / or the mother, a plurality of inherited insertions and deletions, a plurality of chromosomal abnormalities and / or a plurality of single gene or polygenic diseases, including but not limited to: Bloom's syndrome, Canavan disease, cystic fibrosis, familial autonomic dysfunction, Riley-Day syndrome, Fanconi anemia (Group C), Gaucher's disease, glycogen storage disease type 1a, maple syrup urine disease, mucolipid storage disease type IV, Niemann-Pick disease, Tay-Sachs syndrome, beta thalassemia, Sickle cell anemia, alpha thalassemia, beta thalassemia, factor XI deficiency, Friedreich's ataxia, fatty acid metabolism disorders (MCAD), Parkinson's disease - juvenile, connexin 26, spinal muscular atrophy (SMA), Rett syndrome, phenylketonuria, Becker muscular dystrophy, Duchenne muscular dystrophy, fragile X syndrome, hemophilia A, early-onset Alzheimer's disease, breast / ovarian cancer, colon cancer, diabetes / maturity-onset diabetes of the young (MODY), Huntington's disease, myotonic muscular dystrophy, early-onset Parkinson's disease, Pits-Jeghers syndrome, polycystic kidney disease, and torsion dystonia.

[0084] The method according to preferred embodiments of the present invention can allow identification of a fetal congenital disease, such as: neural tube defects, chromosomal abnormalities, Down syndrome (or trisomy 21), trisomy 18, spina bifida, cleft lip and palate, Tay-Sachs disease, sickle cell anemia, thalassemia, cystic fibrosis, Huntington's disease, fragile x syndrome, Turner syndrome (45X0) and Klinefelter syndrome (a male has 2 X chromosomes).

[0085] The method according to the preferred embodiments of the present invention may allow identification of a malformation, such as but not limited to: a limb malformation, such as ectromelia, embolism, phocomelia, polymelia, polydactyly, syndactyly, polysyndactyly, oligodactyly, brachydactyly, achondroplasia, congenital malformations or hypoplasia, amniotic band syndrome and cleidocranial dysplasia; a congenital malformation of the heart, such as patent ductus arteriosus, atrial septal defect, ventricular septal defect and tetralogy of Fallot; a congenital malformation of the nervous system, such as multiple neural tube defects (such as spina bifida, meningocele, meningomyelocele, encephalocele and anencephaly), Einard-Charlie syndrome, Dandy-Walker malformation, hydrocephalus, microcephaly, macrocephaly, flat brain disease, polymicrogyria, holoprosencephaly and corpus callosum dysgenesis; a congenital malformation of the gastrointestinal system, such as stenosis, atresia and anal atresia.

[0086] The method starts from step 10 and continues to step 11, at which a pair of DNA data related to a fetus is received, and the DNA data can be received from a computer-readable medium storing the DNA data by the method, and the DNA data preferably includes at least maternal genomic DNA (gDNA) data, maternal cell-free DNA (cfDNA) data, and paternal genomic DNA data. In some embodiments of the present invention, the DNA data only includes multiple genotypes, and has no information at the read level, or no information at the nucleotide level, or no information at both the read level and the nucleotide level.

[0087] As used in this specification, "genotype" refers to the diploid combination of multiple alleles at a given genetic locus or a set of related loci in a given parent or fetal cell. A homozygous object carries two copies of the same allele, while a heterozygous object carries two different alleles. In the simplest case of a locus with two alleles "A" and "a", three genotypes can be formed: A / A, A / a and a / a. There may be more than two allelic forms, and therefore, there may be more than three genotypes.

[0088] The plurality of genotypes of the parents may be determined by genotyping.

[0089] As used herein, "genotyping" a polymorphic allele of a gene(s) of a subject (or a DNA sample) refers to detecting which allele(s) or polymorphic form(s) of the gene(s) present in a subject (or a sample).

[0090] As used in this specification, the term "gene" or "recombinant gene" refers to a nucleic acid molecule including an open reading frame and containing at least one exon and one (optional) intron sequence. The term "intron" refers to a DNA sequence present in a given gene, which is spliced ​​during the mRNA maturation process.

[0091] Optionally, the method continues to 12, where metadata relating to one or both parents is received, the received metadata optionally and preferably comprising at least one, more preferably more than one, of the plurality of characteristics listed in Table A.1 of Annex 1.

[0092] The method proceeds to step 13, where the data is analyzed to identify a first set of loci at which the parent is homozygous for a plurality of different alleles. Optionally and preferably, the operation is performed for at least each locus of the maternal DNA data. Operation 13 may be performed in a plurality of individuals using a variation detection technique based on sequencing of DNA and alignment of the genome, applying a commercially available variation detector, and screening or selecting from the resulting variation list a plurality of the loci at which the parent is homozygous for a plurality of different alleles.

[0093] A number of sequence alignment techniques that may be used according to some embodiments of the present invention include, but are not limited to, Burrows Wheeler Aligner (BWA), ABA, ALE, AMAP, anon, BAli-Phy, base-to-base, BHAOS / DIALIGN, Bowtie, Bowtie 2, and the like. 2), ClustalW, CodonCode aligner, Comass, DECIPHER, DIALIGN-TX, DIALIGN-T, DNA alignment, DNA base sequence assembler, EDNA, FSA, Geneious, Kalign, MAFFT, MARNA, MAVID, MSA, MSAProbs, MULTALIN, Multi-LAGEN, MUSCLE, Opal, Pecan, Phylo, Praline, PicXAA, POA, Probalign, ProbCons, PROMALS3D, PRRN / PRRP, PSAlign, RevTrans, SAGA, SAM, Se-Al, STAR, STAR-Fusion, StatAlign, Stemloc, T-Coffee, UGENE, VectorFriends, and GLProbs

[0094] Exemplary variant detectors suitable for use in this embodiment include, but are not limited to, the Genome Analysis Toolkit (GATK) and Freebayes. For example, Freebayes may include an alignment of multiple text sequences based on multiple reads aligned to a specific target, rather than their exact alignment. GATK may include: (i) preprocessing; (ii) variant discovery, and (iii) detection set improvement. Preprocessing may include starting with raw sequence data (e.g., FASTQ or uBAM format) and generating multiple analysis-ready BAM files, the processing may include alignment with a reference genome and multiple data cleaning operations to correct multiple technical biases and make the data suitable for analysis, variant discovery may include starting with multiple analysis-ready BAM files and generating a detection set in VCF format, the processing may involve identifying multiple sites that show possible genomic variations in one or more individuals, and applying multiple filtering methods suitable for experimental design, detection set improvement may include starting and ending with a VCF detection set, and processing may involve using metadata to evaluate and improve the accuracy of genotyping, attach more information and evaluate the overall quality of the detection set.

[0095] Multiple variant detectors are also contemplated, such as, but not limited to: Platypus, VarScan, Bowtie analysis, MuTect, and / or SAMtools. For example: Bowtie analysis may include implementing a Burrows-Wheeler transformation for alignment. MuTect may include: (i) preprocessing; (ii) statistical analysis; and (iii) post-processing. Preprocessing may include an initial alignment of multiple sequence reads, statistical analysis may include using two Bayesian classifiers and a classifier that may detect whether a SNP is non-reference at a given site, for the multiple sites found to be non-reference, other classifiers may ensure that they are normal and do not carry the SNP, and post-processing may include removal of sequencing artifacts, short read alignment, and hybrid capture. SAMtools may include storing, manipulating, and aligning multiple sequence reads stored as multiple SAM files.

[0096] In various exemplary embodiments of the present invention, the method proceeds to step 14, in which, for each site of the first set, a probability of obtaining each portion of the maternal cell-free DNA data from the fetus is determined. Operation 14 is advantageous because operation 14 allows the method to obtain information about the multiple differences in the maternal DNA data between the reads from the fetus and the reads from the mother, which can be done by searching for multiple reads presenting paternal alleles in the first set, and defining each site covered by one or more of the multiple reads and preferably each site not covered by other reads as a site from the fetus. All sites from the first set that are from the fetus define a first group within the first set, and a plurality of remaining sites in the first set define a second group in the first set. The multiple differences between the features of the multiple reads covering the multiple sites in the first group and the features of the multiple reads covering the multiple sites in the second group are then used to determine the probability that a particular site in the maternal cell-free DNA data is derived from the fetus.

[0097] The method may also proceed to step 15, in which the data is analyzed to identify a second set of sites in which at least one of the parents has a mutation. Operation 15 may be performed using any of the commercially available techniques described above for operation 13, except that multiple sites other than those in which none of the parents has a mutation are filtered out of the output. The second set preferably includes at least one site that does not belong to the first set, and more preferably multiple sites. However, the first and second sets of sites need not necessarily be disjoint, since a site that is homozygous for different alleles of the parents may in principle present a mutation. However, since the probability of each site of the first set has been determined at step 14, the present embodiments contemplate a scheme in which the first and second sets are multiple disjoint sets.

[0098] The method proceeds to 16, and each site of the second set is classified as fetal or maternal based on the multiple probabilities obtained at 14, and then the genotype of each site classified as fetal can be extracted to genotype the fetus.

[0099] In some embodiments of the invention, the method proceeds to 17, wherein fetal genetic material is obtained and analyzed at 17. Preferably, 17 is performed when the genotyping at 16 indicates a fetal abnormality or a genetic abnormality, in which case the analysis 17 is intended to determine at least whether the fetus has the abnormality or genetic abnormality.

[0100] In some embodiments of the invention, the method proceeds to 18, and prenatal or postnatal treatment is administered in 18. For example, when the method identifies a fetal genetic disease in 17 or 18, the identified fetal genetic disease can be treated prenatally or postnatally. When the method identifies a congenital disease in 17 or 18, the identified congenital disease can be treated prenatally or postnatally. When the method finds a malformation in 17 or 18, the malformation can be treated prenatally or postnatally. According to some embodiments of the invention, other prenatal or postnatal treatments are contemplated, including but not limited to: drug-based interventions, surgery, gene therapy, nutritional therapy, and combinations thereof. In some embodiments of the invention, a pregnancy termination is performed.

[0101] The method ends at 19 .

[0102] The determination of the multiple probabilities in 14 and the classification 16 may be accomplished in more than one manner.

[0103] In some embodiments of the invention, the decision in 14 is based on one or more sequence alignment map (SAM) parameters. SAM is a well-known format for storing biological data (such as but not limited to nucleotide sequences) and defines various parameters (referred to as SAM parameters in this specification). A description of the specification of the SAM format can be found at www(dot)samtools(dot)github(dot)io / hts-specs / SAMv1(dot)pdf, the contents of which are incorporated herein by reference.

[0104] Representative examples of multiple SAM parameters applicable to the present embodiments include, but are not limited to, observed template length, fetal fraction associated with length, tight heterogeneous gap alignment report string, CIGAR string of paired segment, nucleotide sequence, nucleotide sequence of paired segment, read alignment flag indicating whether it is a paired read, read alignment flag indicating whether a read is mapped correctly to a paired read, read alignment flag indicating whether it is an unmapped read, read alignment flag indicating whether it is an unmapped paired segment, read alignment flag indicating whether it is a read on the reverse strand, read alignment flag indicating whether it is a read on the reverse strand a read alignment mark of a paired segment of a forward chain, a read alignment mark indicating whether it is the first paired read, a read alignment mark indicating whether it is the second paired read, a read alignment mark indicating whether it is a read that is not a primary alignment, a read alignment mark indicating whether it is a read that fails a platform and / or a supplier quality check, a read alignment mark indicating whether it is a PCR or optical repeat, a read alignment mark indicating whether it is a supplementary alignment, a marker for a paired segment, mapping quality, mapping quality of a paired segment, multiple genomic coordinates of a chromosome, and multiple genomic absolute starting positions on a chromosome coordinates, multiple genomic coordinates of the absolute end position on the chromosome, multiple genomic coordinates of the start position normalized by the chromosome length, multiple genomic coordinates of the end position normalized by the chromosome length, multiple genomic coordinates of the paired segment, the number of G and C nucleotides divided by the read length, the number of G and C nucleotides divided by the read length in the read sequence of the paired segment, the ratio of A and / or C and / or G and / or T nucleotides in a nucleotide sequence of the read, the ratio of A and / or C and / or G and / or T nucleotides in the nucleotide sequence of the paired segment, the read or the read Information of a variation originating from a paired segment, including at least one feature appearing in Table A.1, the composition of Kmers in the nucleotide sequence, the composition of multiple Kmers in the nucleotide sequence of the partner, the nucleotide quality sequence, the nucleotide quality sequence of the paired segment, the average and / or standard deviation and / or median of the nucleotide quality sequence, the average and / or standard deviation and / or median of the nucleotide quality sequence, the Kmer composition in the nucleotide quality sequence, the Kmer composition in the nucleotide quality sequence of the paired segment, the number of multiple methylated nucleotides divided by the read length, and the methylation of multiple specific positions.

[0105] In some embodiments of the invention, the probability in 14 is determined based on at least an observed template length, in some embodiments of the invention, the probability in 14 is determined based on at least a fetal fraction associated with a length. In some embodiments of the invention, the probability in 14 is determined based on at least a CIGAR string, in some embodiments of the invention, the probability in 14 is determined based on at least a nucleotide sequence, in some embodiments of the invention, the probability in 14 is determined based on at least a nucleotide sequence of a paired segment.

[0106] In some embodiments of the invention, the probability in 14 is determined based at least on at least one read alignment marker. Multiple representative examples of multiple read alignment marks applicable to the present embodiment include, but are not limited to: the decision may be based on at least one read alignment mark, the read alignment mark being selected from a read alignment mark indicating whether it is a paired read, a read alignment mark indicating whether a read is mapped with correct pairing, a read alignment mark indicating whether it is an unmapped read, a read alignment mark indicating whether it is an unmapped paired segment, a read alignment mark indicating whether it is a read on the reverse strand, a read alignment mark indicating whether it is a paired segment on multiple reverse strands, an alignment mark indicating whether it is the first paired read pair, an alignment mark indicating whether it is the second paired read, a read alignment mark indicating whether it is a read that is not a primary alignment, a read alignment mark indicating whether it is a read that fails a platform and / or a supplier quality check, a read alignment mark indicating whether it is a PCR or optical repeat, and a read alignment mark indicating whether it is a supplementary alignment.

[0107] The present embodiments also contemplate calculating a total fetal fraction. In the embodiments, the classification 16 is optionally and preferably based on the calculated total fetal fraction. Embodiments are also contemplated in which a fetal size distribution and a maternal size distribution are constructed. In the embodiments, the classification 16 optionally and preferably includes binning the fetal size distribution, calculating a fetal fraction for each fragment size bin, and calculating a probability that the fragment is fetal for at least one site and at least one fragment at at least one site based on a fetal fraction of a corresponding fragment size bin to which the fragment belongs. In some embodiments, optionally and preferably, the classification in 16 includes applying a Bayesian procedure, which optionally and preferably includes multiple prior probabilities calculated using sequencing data (e.g., genome sequencing or whole exome sequencing) of at least one of the parents.

[0108] In some embodiments of the invention, the determination of the plurality of probabilities in 14 and optionally and preferably also the classification in 16 comprises applying a machine learning procedure.

[0109] As used in this specification, the term "machine learning" refers to an implementation as a program of a computer program configured to derive multiple patterns, regularities or rules from previously collected data to produce an appropriate response to future data, or to describe the data in some meaningful way.

[0110] Representative examples of machine learning programs suitable for use in the embodiments include, but are not limited to, clustering, association rule algorithms, feature evaluation algorithms, subset selection algorithms, support vector machines, classification rules, cost-sensitive classifiers, voting algorithms, stacking algorithms, Bayesian algorithm networks, decision trees, neural networks, convolutional neural networks, example-based algorithms, linear modeling algorithms, k-nearest neighbor (KNN) analysis, ensemble learning algorithms, probability models, graphical models, logistic regression methods (including multinomial logistic regression methods), gradient ascent methods, singular value decomposition methods, and principal component analysis.

[0111] The following is an overview of some machine learning procedures suitable for use in this embodiment.

[0112] Support vector machines are algorithms based on statistical learning theory. According to some embodiments of the present invention, support vector machines (SVM) can be used for multiple classification purposes and / or for numerical prediction. A support vector machine for classification is referred to as a "support vector classifier" in this specification, and a support vector machine for numerical prediction is referred to as a "support vector regression" in this specification.

[0113] An SVM is typically characterized by a kernel function, the selection of which determines whether the resulting SVM provides classification, regression, or other functions. By applying the kernel function, the SVM pairs multiple input vectors into a high-dimensional feature space, in which a decision hypersurface (also called a separator) can be constructed to provide classification, regression, or other decision functions. In the simplest case, the surface is a hyperplane (also called a linear separator), but the use of multiple more complex separators is also envisioned, and multiple kernel functions can be applied, and the data points defining the hypersurface are called multiple support vectors.

[0114] The support vector classifier selects a separator where the distance between the separator and the closest data points is as large as possible, thereby separating feature vector points associated with objects in a given classification from feature vector points associated with objects outside the classification. For support vector regression, a high-dimensional pipeline with a radius of acceptable error is constructed that minimizes the error of the data set while maximizing the flatness of the associated curve or function. In other words, the pipeline is a shell around the fitted curve, defined by a set of data points closest to the curve or surface.

[0115] An advantage of a support vector machine is that once the multiple support vectors are determined, the multiple remaining observations can be deleted from the calculation, thereby greatly reducing the computational complexity of the problem. An SVM is generally divided into two phases: a training phase and a testing phase. During the training phase, a set of support vectors will be generated for executing the decision rules. In the testing phase, the decision rules will be used to determine multiple decisions. A support vector algorithm is a method for training an SVM. By executing the algorithm, a set of training parameters will be generated, including the multiple support vectors that characterize the SVM. A representative example of a support vector algorithm suitable for the present multiple embodiments includes but is not limited to sequential minimum optimization.

[0116] In KNN analysis, the affinity or closeness of multiple objects is determined, and the affinity is also called the distance in a feature space between multiple objects. Based on the multiple determined distances, the objects are clustered and an outlier is detected. Therefore, the KNN analysis is a technique for finding multiple distance-based outliers based on the distance of an object from the kth nearest neighbor in the feature space. Specifically, each object is ranked according to the distance of each object to the kth nearest neighbor of the object, and the farthest object is declared as a plurality of outliers. In some cases, multiple farthest objects are declared as multiple outliers. That is, for example: k neighbors and a specific distance, if there are no more than k objects at a specific distance equal to or less than the object, then an object is an outlier with respect to multiple parameters. The KNN analysis is a classification technique using supervised learning, presenting an item and comparing the item with a training set having two or more classifications, and the item is assigned to the most common category among the multiple k nearest neighbors of the item. That is, the distances to all items in the training set are calculated to find the nearest k, and then the majority classification is extracted from the k and assigned to the item.

[0117] Association rule algorithm is a technique for extracting multiple meaningful association patterns between multiple features.

[0118] In the context of machine learning, the term "association" refers to any relationship between features, not just relationships between features that predict a specific categorical or numerical value. Association includes, but is not limited to, finding association rules, finding patterns, performing feature evaluation, performing feature subset selection, developing predictive models, and understanding interactions between features.

[0119] The term “association rule” refers to a plurality of elements that frequently appear in the plurality of data sets, and the association rule includes but is not limited to a plurality of association patterns, a plurality of distinguishing patterns, a plurality of frequent patterns, a plurality of closed patterns, and a plurality of huge patterns.

[0120] A common main step of an association rule algorithm is to find a set of items or features that occur most frequently among all observations. Once the list is obtained, multiple rules can be extracted from the list.

[0121] The self-organizing map mentioned above is an unsupervised learning technique that is often used to visualize and analyze high-dimensional data. Several typical applications focus on the visualization of multiple central dependencies in the data on the map. The map generated by the algorithm can be used to speed up the identification of multiple association rules by other algorithms. The algorithm usually includes a grid of multiple processing units, called "neurons". Each neuron is associated with a feature vector called an observation. The map attempts to use a restricted set of models to represent all available observations with the best accuracy. At the same time, the multiple models become ordered on the grid so that multiple similar models are close to each other, while multiple different models are far away from each other. The program allows the identification and visualization of multiple dependencies or multiple associations between the multiple elements in the data.

[0122] A plurality of feature evaluation algorithms are for the ranking of a plurality of features or for the ranking after the plurality of features are selected based on the influence of the plurality of features.

[0123] Information gain is one of the machine learning methods applicable to feature evaluation, and the definition of information gain requires the definition of entropy, which is a measure of impurity in a set of training examples. The decrease in the entropy of a target feature that occurs by knowing multiple values ​​of a particular feature is called information gain. Information gain can be used as a parameter to determine the effectiveness of a feature in explaining the response to the treatment. According to some embodiments of the present invention, symmetric indeterminacy is an algorithm that can be used by a feature selection algorithm. Symmetric indeterminacy compensates for the biases of information gain towards features with more values ​​by normalizing multiple features to the range [0,1].

[0124] Subset selection algorithms rely on a combination of an evaluation algorithm and a search algorithm, and similar to multiple feature evaluation algorithms, multiple subset selection algorithms rank multiple subsets of multiple features. However, unlike multiple feature evaluation algorithms, a subset selection algorithm applicable to multiple embodiments of the present invention aims to select the subset of multiple features that have the greatest impact on the probability that each portion of DNA data is derived from a fetus, while taking into account the redundancy between the multiple features included in the subset. The benefits of feature subset selection include facilitating data visualization and understanding, reducing measurement and storage requirements, reducing training and utilization time, and eliminating scattered multiple features to improve classification.

[0125] The two basic approaches to multiple subset selection algorithms are the process of adding multiple features to a working subset (forward selection) and removing multiple features from the current subset (backward elimination). In machine learning, forward selection is different from the statistical process with the same name. The features to be added to the current subset in machine learning can be found by evaluating the performance of the current subset enhanced by a new feature using cross-validation. In forward selection, multiple subsets are constructed by adding each remaining feature to the current subset in turn, while evaluating the expected performance of each new subset using cross-validation. The features that result in the best performance will be retained and continued because they are added to the current subset. The search ends when none of the remaining available features improves the predictive power of the current subset, and the process has found a locally optimized feature set.

[0126] Backward elimination is implemented in a similar manner, using the backward elimination function, when further reducing the feature set does not improve the predictive power of the subset, the search is terminated. The present multiple embodiments contemplate search algorithms that search forward, backward, or in both directions. Representative examples of search algorithms suitable for the present embodiments include, but are not limited to, exhaustive search, greedy hill climbing, multiple random perturbations of multiple subsets, multiple packing algorithms, probabilistic race search, graph search, ranked race search, and Bayesian classifiers.

[0127] A decision tree is a decision support algorithm that forms a logical path of steps involved in considering the inputs to make a decision.

[0128] The term "decision tree" refers to any type of tree-based learning algorithm, including but not limited to model trees, classification trees, and regression trees.

[0129] A decision tree can be used to hierarchically classify the multiple data sets or their relationships, the decision tree having a tree structure including multiple branch nodes and multiple leaf nodes, each branch node specifies an attribute (split attribute) and a test (split test) to be performed on the value of the split attribute, and branches to other nodes for all possible results of the split test. The branch node that serves as the root of the decision tree is called the root node. Each leaf node can represent a classification (for example, whether a specific parameter affects the probability of obtaining each part of DNA data from a fetus) or a value (for example, the probability that each part of DNA data is derived from a fetus). The leaf node may also contain additional information about the represented classification, such as a credibility score that measures a credibility in the represented classification (i.e., the accuracy of the prediction).

[0130] A variety of regression techniques that may be used according to some embodiments of the present invention include, but are not limited to, linear regression, multiple regression, logistic regression, probit regression, ordinal logistic regression, ordinal probit regression, Poisson regression, negative binomial regression, multinomial logistic regression (MLR), and shortened regression.

[0131] A logarithmic regression or logarithmic regression is a type of regression analysis used to predict the outcome of a categorical dependent variable (a dependent variable can take on a finite number of values, the magnitude of which is meaningless, but the order of the magnitudes of which may or may not be meaningful). Logistic regression can also predict the probability of occurrence of each data point based on one or more predictor variables. Multiple logistic regressions also include a polynomial variant. The polynomial logistic regression model is a regression model that generalizes logistic regression by allowing more than two discrete outcomes. In other words, the polynomial logistic regression model is a model for predicting the multiple probabilities of multiple different possible outcomes of a categorically distributed dependent variable given a set of independent variables (which can be real values, binary values, categorical values, etc.). For multiple binary-valued variables, the Yuden index is often used to determine a boundary between the multiple associations from 0 to 1.

[0132] A Bayesian network is a model that represents multiple variables and conditional interdependencies between multiple variables. In a Bayesian network, multiple variables are represented as multiple nodes, and multiple nodes can be connected to each other by one or more links, a link indicating a relationship between two nodes. Multiple nodes usually have multiple corresponding conditional probability tables, which are used to determine the probability of a state of a node based on the states of other multiple nodes to which the node is connected. In some embodiments, a Bayesian optimal classifier algorithm is used to apply a maximum a posteriori hypothesis to a new record to predict the probability of the classification of the new record, and the multiple probabilities are calculated based on each of the other multiple hypotheses obtained from a training set, and the multiple probabilities are used as multiple weighting factors to make multiple future predictions of the probability that various parts of the DNA data are derived from a fetus. An algorithm suitable for a search for the optimal Bayesian network includes but is not limited to an algorithm based on a global score metric. In another method of constructing the network, a Markov blanket may be used, which isolates a node from the influence of any node outside the boundary of the Markov blanket, the Markov blanket consisting of the node's parent, its children, and its children's parents.

[0133] Example-based multiple techniques generate a new model for each example, rather than making multiple predictions based on multiple trees or multiple networks (at once) generated from a training set.

[0134] In the context of machine learning, the term "example" refers to an example in a dataset.

[0135] Example-based techniques typically store the entire data set in memory and build a model based on a set of records that are similar to the records being tested. For example, the similarity can be assessed by a nearest neighbor approach or a locally weighted approach, such as using multiple Euclidean distances. Once a set of records is selected, several different techniques (e.g., naive Bayes) can be used to build the final model.

[0136] Neural networks are a class of algorithms based on the concept of interconnected "neurons". In a typical neural network, neurons contain data values, each of which affects the value of a connected neuron based on connections with predetermined strengths and whether the sum of connections to each particular neuron meets a predetermined threshold. By determining appropriate connection strengths and thresholds (also known as a training process), a neural network can achieve effective recognition of images and characters. Typically, the neurons are divided into several layers to make the connections between groups more obvious and to obtain each calculation of the values. Each layer of the network may have a different number of neurons, and the neurons may or may not be related to specific qualities of the input data.

[0137] In one implementation known as a fully connected neural network, each of the plurality of neurons in a particular layer is connected and provides input values ​​to the plurality of neurons in the next layer, and the plurality of input values ​​are then summed and compared to a bias or threshold. If the value exceeds the threshold for a particular neuron, the neuron retains a positive value that can be used as input to a plurality of neurons in the next layer of neurons. The calculations continue throughout the layers of the neural network until the calculations reach a final layer. At this point, the output of the neural network routine can be read from the plurality of values ​​in the final layer. Unlike fully connected neural networks, convolutional neural networks operate by associating an array of multiple values ​​with each neuron rather than a single value, and the conversion of a neuron value in the subsequent layer is from multiplication to convolution.

[0138] The machine learning program used in accordance with some embodiments of the present invention is a trained machine learning program, optionally and preferably a deep learning program (e.g., a convolutional neural network), which provides an output that is nonlinearly related to the multiple parameters fed to the deep learning program.

[0139] It is expected that many related machine learning programs will be developed between the time of this application and the expiration of a patent life, and the scope of the term machine learning program is intended to include all similar prior art.

[0140] According to some embodiments of the present invention, a machine learning training program, optionally and preferably a deep learning program, can be trained by feeding DNA data of each of a group of subjects having a fetus with a genotype as described in further detail above, and optionally and preferably metadata, such as but not limited to: a convolutional neural network. When a convolutional neural network is used, the input of the network is usually (but not necessarily) in the form of multiple multidimensional tensors. For example: each tensor may include input data corresponding to a specific locus in the genome. In multiple preferred embodiments, the tensor is a stack (pileup) of multiple readings covering a candidate single nucleotide polymorphism (SNP) (e.g.: centered on the SNP), the first two dimensions of the tensor can, for example, correspond to the length of the stack and the number of reads, and one or more other dimensions of the tensor can correspond to the metadata of the SNP. In the training phase, each tensor is associated with a label, and the label contains the true fetal genotype found using an invasive test.

[0141] Once the data is fed, the machine learning training program generates a trained machine learning program that can then be used without retraining or feeding the machine learning program with multiple fetal genotypes found by an invasive test. The trained machine learning program can provide a probability that a portion of the maternal DNA data is from a fetus and / or can classify a locus with a mutation in at least one of the parents as fetal or maternal.

[0142] Preferably, the trained machine learning program is used as an end-to-end program, in which case the DNA data is fed to the trained machine learning program and an output indicating the probability that a site is obtained from a fetus, or more preferably an output can be received from the trained machine learning program indicating that multiple sites in at least one of the parents having a mutation are classified as fetal or maternal.

[0143] Alternatively, the trained machine learning program can be used to supplement other techniques, such as for recalibration. For example: In multiple embodiments where the classification is performed by a Bayesian program in 16, the trained machine learning program can be used to recalibrate the output of the Bayesian program.

[0144] It is further contemplated that in various embodiments, the output of a trained machine learning program is used to retrain the machine learning program, thereby allowing the accuracy to improve over time.

[0145] According to some embodiments of the present invention, fetal genotyping may be performed by a computer or a server-client computer configuration, as will now be referred to. Fig.16 As described.

[0146] Fig.16 A client computer 30 is shown having a hardware processor 32, which typically includes an input / output (I / O) circuit 34, a hardware central processing unit (CPU) 36 (e.g., a hardware microprocessor), and a hardware memory 38. The hardware memory 38 typically includes volatile memory and non-volatile memory. The CPU 36 communicates with the I / O circuit 34 and the memory 38. The client computer 30 preferably includes a user interface, such as a graphical user interface (GUI) 42, which communicates with the processor 32. The I / O circuit 34 preferably communicates information with the GUI 42 in the form of an appropriate structure. A server computer 50 is also shown, which may similarly include a hardware processor 52, an I / O circuit 54, a hardware CPU 56, and a hardware memory 58. The multiple I / O circuits 34 and 54 of the client 30 and server 50 computers preferably operate as multiple transceivers that communicate information with each other via a wired or wireless communication. For example, the computers of the client 30 and the server 50 may communicate via a network 40 such as a local area network (LAN), a wide area network (WAN), or the Internet. In some embodiments, the server computer 50 may be part of a cloud computing resource of a cloud computing facility that communicates with the client computer 30 via the network 40.

[0147] The GUI 42 and the processor 32 may be integrated in the same housing, or the GUI 42 and the processor 32 may be separate units that communicate with each other. The GUI 42 may optionally and preferably be part of a system including a dedicated CPU and I / O circuits (not shown) to allow the GUI 42 to communicate with the processor 32. The processor 32 issues graphical and textual output generated by the CPU 36 to the GUI 42. The processor 32 also receives signals from the GUI 42 related to control commands generated by the GUI 42 in response to user input. The GUI 42 may be of any type known in the art, such as, but not limited to, a keyboard and a display, a touch screen, etc. In preferred embodiments, the GUI 42 is a GUI of a mobile device, such as a smartphone, a tablet, a smart watch, etc. When the GUI 42 is a GUI of a mobile device, the CPU circuit of the mobile device may serve as the processor 32, and may optionally and preferably perform the method by executing code instructions.

[0148] The computers of the client 30 and the server 50 may further include one or more computer-readable storage media 44, 64, respectively. The plurality of media 44 and 64 are preferably a plurality of non-transitory storage media, the media 44 and 64 storing a plurality of computer code instructions for executing the methods of the present embodiments, and the plurality of processors 32 and 52 execute the plurality of code instructions. The plurality of code instructions may be executed by loading the corresponding code instructions into the corresponding execution memories 38 and 58 of the plurality of corresponding processors 32 and 52. The plurality of storage media 64 preferably also stores one or more databases, the databases including a database of olfactory signatures with a plurality of psychological annotations as described in further detail above.

[0149] In operation, the processor 32 of the client computer 30 receives the DNA data, and optionally and preferably receives the metadata as described above, which can be read, for example, from the storage medium 44. The processor 32 typically sends the data to the server computer 50 via the network 40. The medium 64 can store a program (e.g., a deep learning program and / or a Bayesian program) for genotyping a fetus as described above. The server computer 50 can access multiple media 64, feed the stored program with the data received from the client computer 30, and receive an output from the program indicating the multiple genotypes of the fetus or indicating the probability that a portion of the data (e.g., a mutation site) is derived from the fetus. The server computer 50 can also send the obtained output to the client computer 30, and the client computer 30 can display the information on the GUI 42.

[0150] Alternatively, the medium 44 may also store the program for genotyping the fetus as described in further detail above, in which case the entire method may be performed by the computer 30. Specifically, the computer 30 may store the DNA data received from the medium 44 in the memory 38, access the medium 44 to obtain the program for genotyping the fetus, feed the data stored in the memory 38 to the program, and receive from the program an output indicating the plurality of genotypes of the fetus or indicating the probability that a portion of the data (e.g., a mutation site) is derived from the fetus. The computer 30 may then display the information on the GUI 42.

[0151] The term "about" as used in this specification means ± 10%.

[0152] The word “exemplary” is used in this specification to mean “serving as an example, instance, or illustration.” Any implementation described as “exemplary” is not necessarily to be construed as preferred or advantageous over other implementations and / or to exclude the incorporation of features from other implementations.

[0153] The word “optionally” is used in this specification to mean “provided in some embodiments and not provided in other embodiments.” Any particular embodiment of the present invention may include a number of “optional” features, unless such features conflict.

[0154] The terms "comprises," "comprising," "includes," "including," "having" and their conjugations mean "including but not limited to."

[0155] The term "consisting of" means "including and limited to".

[0156] The term "consisting essentially of" means that the composition, method or structure may include additional ingredients, steps and / or components, but only when the additional ingredients, steps and / or components do not substantially change the basic and novel characteristics of the claimed composition, method or structure.

[0157] As used herein, the singular forms "a", "an" and "the" may include plural references unless the context clearly indicates otherwise. For example, the term "a compound" or "at least one compound" may include a plurality of compounds, including mixtures thereof.

[0158] Throughout this application, various embodiments of the present invention may be presented in the form of a range. It should be understood that the description in the form of a range is only for convenience and brevity and should not be construed as a hard limitation on the scope of the present invention. Therefore, it should be considered that the range description has specifically disclosed all possible sub-ranges and single numerical values ​​within the range. For example, it should be considered that the range description from 1 to 6 has specifically disclosed sub-ranges, such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., as well as single numbers within the numbered range, such as 1, 2, 3, 4, 5 and 6, which applies regardless of the range.

[0159] Whenever a numerical range is indicated herein, it is meant to include any cited numeral (fractional or integer) within the indicated range. The phrases: "a range between" a first indicated number and a second indicated number and "a range from" a first indicated number to" a second indicated number are interchangeable herein and are meant to include the first and second indicated numbers, and all fractions and integers therebetween.

[0160] The term "method" as used herein refers to manners, means, techniques and procedures for accomplishing a specific task, wherein the given task includes but is not limited to multiple manners, means, techniques and procedures that are known or can be easily developed by practitioners in the fields of chemistry, pharmacology, biology, biochemistry and medicine from known manners, means, techniques or steps.

[0161] As used herein, the term "treating" includes halting, substantially inhibiting, slowing or reversing the progression of a disorder, substantially ameliorating the clinical or psychological symptoms of a disorder or substantially preventing the appearance of the clinical or psychological symptoms of a disorder.

[0162] When reference is made to multiple specific sequence listings, the reference should be understood to also include multiple sequences that substantially correspond to the complementary sequences of the multiple specific sequence listings, including: multiple smaller sequence variations due to, for example, sequencing errors, cloning errors, or other changes resulting in base substitutions, base deletions, or base additions, provided that the frequency of the multiple variations is less than one in fifty nucleotides, less than one in one hundred nucleotides, alternatively, less than one in two hundred nucleotides, less than one in five hundred nucleotides, alternatively, in less than one in a thousand nucleotides, alternatively, less than one in five thousand nucleotides, alternatively, less than one in ten thousand nucleotides.

[0163] It should be understood that certain features of the invention, which for the sake of clarity are described in the context of multiple separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which for the sake of simplicity are described in the context of a single embodiment, may also be provided separately or in any suitable subcombination or in any other described implementation suitable for the invention. Some features described in the context of various embodiments should not be considered essential features of those embodiments unless the embodiment does not function without those components.

[0164] Various embodiments and aspects of the present invention as delineated hereinabove and as claimed in the claims section below find experimental support in the following multiple examples.

[0165] Multiple examples:

[0166] Reference is now made to the following examples which together with the above descriptions illustrate some embodiments of the invention in a non-limiting fashion.

[0167] Non-Invasive Parental Diagnosis (NIPD) is a test for multiple genetic variations in the fetus, which is based on multiple cell-free DNA (cfDNA) fragments circulating freely in plasma. During pregnancy, the cell-free DNA in the mother's plasma contains cell-free fetal DNA (cffDNA) originating from multiple placental cells. The content of cell-free fetal DNA (also called fetal fraction) accounts for about 10% of the cell-free DNA at the end of the first three months of pregnancy and increases in the following three months. Available NIPD is currently mainly used for multiple chromosomal abnormalities (e.g. Down syndrome).

[0168] NIPD of point mutations is a more challenging task because the amount of cell-free fetal DNA covering a given genomic position is considerably lower than that of an entire chromosome. In mutations where only the father is a carrier, the presence of the foreign (non-maternal) allele in the mother's plasma can be modeled as a binomial distribution; if enough DNA fragments in the mother's plasma contain the foreign allele, it can be inferred that the foreign allele is inherited. Therefore, the detection of a mutation in the fetus in the position is considered to be more straightforward. In mutations where only the mother is heterozygous, both alleles are present in large quantities in the mother's plasma regardless of the identification of the inherited allele, and a common approach to solving the problem is to measure the amount of each allele. If the multiple alleles have equal amounts, the fetus is considered to be a carrier of the mutation, i.e., a heterozygote. If there is an allelic imbalance, the fetus is homozygous with an excess of the allele.

[0169] Measuring subtle allelic imbalances requires an ultra-precise quantitative method, such as a digital polymerase chain reaction (PCR). The inventors realized that such methods require a custom design for each case and are therefore not suitable for testing a large number of mutations simultaneously. Another option is to use next-generation sequencing (NGS), which is a less accurate method but is performed with very deep coverage (i.e., each position is read multiple times) and uses special preparation protocols that can reduce artifacts. Few methods have been suggested for NGS-based whole-genome NIPD of multiple point mutations (Chan et al. 2016; Fan et al. 2012; Kitzman et al. 2012).

[0170] The inventors of the present invention have found that for multiple SGDs of maternal origin, sensitivity poses a challenge in that the detection is limited to one genetic disease at a time. Example 1 below provides a Bayesian procedure for the NIPD of multiple single-gene diseases, which is independent of inheritance mode and parental origin. Sequenced parental and fetal data are used to calibrate and validate the model. Example 1 shows that taking into account multiple differences in the fragment length distribution of cell-free DNA from the fetus and the mother leads to increased accuracy. The model of Example 1 is extended to the prediction of genetic insertions and deletions (indels). Example 1 also shows that using data from multiple previously analyzed families and recalibrating the multiple posterior probabilities of the model through a machine learning algorithm can correct a large number of genotyping errors. In Example 1, the method of the present embodiments was successfully applied to predict the inheritance of congenital chloride diarrhea in a pregnancy where the parents are multiple carriers of a mutation. This example confirms that next-generation sequencing (NGS) can be used for NIPD of multiple single-gene diseases at the same time, and the example shows that the method of the present embodiments can be used as a general framework for the NIPD of multiple SGDs.

[0171] Typically, NIPD can be achieved by analyzing cell-free DNA in the mother's plasma, which contains cell-free fetal DNA derived from the placenta. The main use of the cell-free fetal DNA is to identify multiple chromosomal abnormalities, such as trisomy (Down syndrome). Multiple other clinical applications are fetal sex determination and rhesus monkey D genotyping. Genetic diagnosis of SGD can be achieved in various ways, from epigenetic description and a linkage analysis, through multiple different laboratories such as polymerase chain reaction (PCR) for multiple known mutations and multiple DNA microarrays to Sanger sequencing for confirmation of multiple results and NGS for a more in-depth investigation. Two known tools are whole exome sequencing (WES) and whole genome sequencing (WGS) based on NGS. Although the cost of WGS is still high and the multiple meanings of multiple results of WGS have yet to be studied, WES covers approximately 2 to 3% of the genome and is relatively low in cost, so WES is still often used. However, WGS is more reliable than WES even for multiple exome variants, and WGS has become increasingly popular because of its ability to resolve multiple structural variants. WES of multiple infants suspected of having multiple genetic diseases is more likely to have an impact on medical care, and WES of DNA obtained by amniocentesis can help make such prenatal diagnoses in some cases.

[0172] The inventors of the present invention have found that although the application of NGS in the NIPD of multiple monogenic diseases has been proven to be feasible, some improvements can still be made. Unlike the identification of paternally transmitted alleles in cell-free DNA, which is considered to be straightforward, maternally transmitted alleles pose a greater challenge because both alleles are present in her plasma at multiple sites where the mother is heterozygous. In the multiple cases, conventional techniques cannot determine which of the multiple alleles is inherited. The current solution is to look for an allelic imbalance (a slightly higher number of alleles that appear when the fetus is homozygous). However, due to the small number of cell-free DNA and even fewer cell-free fetal DNA, the execution of the allelic imbalance is limited to multiple ultra-precise devices such as digital PCR. Furthermore, the method is not scalable;. When more genomic sites are tested, the method becomes less feasible and therefore, of little help. NGS can also be used, but NGS requires a very deep coverage of each site.

[0173] A technique using WES to provide deeper coverage is known [Fan, HC et al., Noninvasive prenatal measurement of the fetal genome, Nature 487, 320-324 (2012)]. In this technique, a high proportion of the fetal exons were reconstructed when deep WES (221x and 631x) was used in the second and third trimesters, respectively. Strict data filtering was applied before the analysis. In another attempt to genotype a fetus, cell-free DNA was sequenced to a deep coverage (270x) using WGS, and a sequential probability ratio test was performed for each locus where the mother was heterozygous and the parents were not haplotyped [Chan et al., Second-generation non-invasive fetal genomic analysis reveals de novo mutations, single-base parental inheritance, and multiple preferred DNA ends, Proceedings of the National Academy of Sciences, 201615800 (2016), Digital Object Identifier (doi): 10.1073 / pnas.1615800113]. Although the study showed good accuracy, the inventors found some limitations of the technology. First, the samples sequenced were from the third trimester, where the amount of cell-free DNA and the amount of cell-free fetal DNA in the cell-free DNA were high. Second, the method applied did not use the available information about the paternal inheritance. Third, it is unclear whether a sequential test has an advantage when genotyping a single position, since the data in this case are not cumulative. In addition, in the study, accuracy was from a relatively low of only 6.5 x 10 5 The results were calculated from the loci where the mother was heterozygous.

[0174] The inventors have discovered that a method that can help improve non-invasive fetal genotyping can rely on inherent differences in fetal and maternal characteristics, one of which is the actual cell-free DNA fragments and cell-free fetal DNA fragments. For example: fetal-derived fragments are generally shorter, and the pattern of the size distribution of the fetal-derived fragments indicates a relationship with the positioning of nucleosomes. Attempts were made to enrich cell-free fetal DNA by applying a hard threshold setting to handle chromosomal abnormalities using size differences. However, the inventors found that because the two size distributions overlap significantly, the threshold may result in a loss of relevant information.

[0175] Example 1 below shows that multiple small variations in the fetus can be accurately detected in a non-invasive manner by using deep NGS and an improved algorithm. Example 1 uses a Bayesian model to demonstrate that multiple small variations in the fetus can be accurately detected at multiple positions and in multiple small insertions and deletions (indels) even if the mother or both parents are multiple carriers of a mutation, all of which are considered more difficult to genotype. The prior probability is calculated using the parental genotypes according to simple Mendelian laws of inheritance, and the likelihood function is based on multiple supports for each cell-free DNA fragment in each possible fetal genotype. The multiple fragment lengths are also used in the likelihood function because the length distributions of the multiple fetal fragments and maternal fragments described previously are different. The multiple length distributions and the fetal fraction (i.e., the ratio of cell-free fetal DNA in the cell-free DNA) can all be empirically calculated based on a given cell-free DNA sample, and the Bayesian model is also assisted by a machine learning recalibration procedure.

[0176] Example 2 below presents an end-to-end method based on a deep learning program that can replace the Bayesian model and the machine learning recalibration program used in Example 1. One of the advantages of the end-to-end method described in Example 2 is that the end-to-end method can automatically model multiple systematic errors, which is useful in difficult-to-interpret data (e.g., low-depth sequencing). With the ease of use of the two programming methods and the decline in hardware prices, deep learning is rapidly becoming available.

[0177] Example 1:

[0178] The embodiment describes a technique for non-invasive prenatal variant detection, which the inventors call Hoobari. The Hoobari technique is based on a Bayesian algorithm that uses the information of each read individually, and the technique can be fine-tuned in a modular way without having to rewrite the entire model. An example of this is multiple probabilistic recalibration steps based on a machine learning, and the technique uses fragment size differences to improve fetal genotyping, which is particularly useful for multiple challenging loci where the mother is heterozygous. The technique shown in this example can be generalized to allow prediction of the inheritance of multiple small insertions and deletions (indels). The inventors demonstrated the ability of the technique to address the diagnosis based on NGS using deep WES (>600x) of two first-trimester cases and deep WGS (310x) of another first-trimester case, leading to a wide range of SGD NIPD.

[0179] Multiple Materials and Methods:

[0180] Considering that the mother is heterozygous for multiple single nucleotide polymorphisms (SNPs), it is difficult to determine whether a fragment in the multiple sites is fetal or maternal. Therefore, the probability of each fragment being fetal is calculated based on the size of the fragment, and the multiple probabilities are used in a Bayesian classification model.

[0181] Using multiple sites where the multiple parents are homozygous for multiple different alleles, two empirical size distributions are created: a fetal empirical size distribution and a maternal empirical size distribution. Among the multiple sites, a cell-free DNA fragment showing the paternal allele is likely to be fetal. The total fetal fraction is calculated, which is the fraction of cell-free fetal DNA in all maternal cell-free DNA. Then, a fetal fraction for each fragment size is calculated using all fragments with the same length. When the Bayesian model is applied to a site of interest in a specific genome, the fetal fraction divided by size is used to calculate the probability that each fragment in all of the multiple fragments covering the site is a fetus. In this way, shorter fragments generally have a higher fetal probability and strict size thresholds can be avoided.

[0182] like Figure 1Schematically illustrating the process of Hoobari for non-invasive prenatal variant detection, the code returns three posterior probabilities, one for each possible fetal genotype: homozygous with the reference allele (0 / 0), heterozygous (0 / 1), and homozygous with the variant allele (1 / 1). The predicted genotype for each locus is the genotype with the highest probability, and multiple fetal variants found in a pure fetal sample (e.g., amniotic fluid, chorionic villus, or umbilical cord blood) are used as multiple true variants.

[0183] The pipeline differs from conventional variant detection workflows in that the prior probabilities are calculated with high confidence using existing parental sequencing data obtained from an initial genotyping of the parents. In addition, cell-free DNA is an unbalanced mixture of two similar genomes. Therefore, the pipeline also differs from conventional variant detection workflows in the dedicated technique for calculating the likelihoods in the Bayesian model, which uses the cell-free DNA fragment length distribution, which is calculated in a preprocessing step.

[0184] A more detailed description of the study follows.

[0185] Handling multiple families from multiple previous studies:

[0186] The technology of this embodiment is tested on whole genome data of four family trios sequenced to different coverage depths in multiple previous studies, and the fetal fraction in the multiple cell-free DNA has changes. The multiple families are called G1, G2, G3 and G4. As Chan and KCA et al. and Kitzman and JO et al. above, non-invasive whole genome sequencing of human fetuses, Science Translational Medicine, 4, 137ra76 (2012), multiple samples from G1 to G2 families and G3 to G4 families were collected and sequenced.

[0187] Sample collection and DNA extraction:

[0188] With informed consent, multiple samples from each family were collected during week 11, and DNA was extracted from chorionic villus sampling (CVS) using the DNA tissue protocol of the MagNA pure compact nucleic acid isolation kit I-large volume (Roche Life Sciences). Peripheral maternal blood was collected using multiple 2-4 ethylenediaminetetraacetic acid (EDTA) tubes, and plasma was separated from the blood by centrifugation at 1600×g for 10 minutes at 4°C, and the plasma was then centrifuged again at 16,000×g for 10 minutes at room temperature to remove any residual cells. Cell-free DNA was extracted using the QIAamp circulating nucleic acid kit (Qiagen). Excess salts generated by purification of cell-free DNA were removed using Agencourt AMPure XP magnetic beads (Beckman Coulter) at a 2-fold ratio of the volume of cell-free DNA. Pure maternal DNA was extracted from multiple white blood cells in the mother's buffy coat using a procedure including (i) buffy coat separation and (ii) using the Gentra Pure Genetic Blood Kit (Qiagen) according to the manufacturer's multiple instructions. Pure paternal DNA was similarly collected and purified.

[0189] Library construction and sequencing:

[0190] Libraries were constructed for samples subjected to WGS using the TruSeq DNA PCR-free library construction kit (Illumina) according to the manufacturer's instructions and subsequently sequenced using the HiSeq X+ system (Illumina) with 151 base pair (bp) paired-end reads.

[0191] For samples that underwent WES, library construction was performed using the SureSelect V5 Exome Kit (Agilent) according to the manufacturer's instructions, enrichment was achieved by hybridization of the constructed genomic DNA with complementary RNA probes, and then sequencing was performed using a HiSeq 4000 (Illumina) with 101 base pair paired-end reads.

[0192] Multiple cell-free DNA samples were not fragmented during library construction and were sequenced in two steps: (1) using HiSeq 4000 (Illumina) with paired-end reads of 101 base pairs to achieve a required 50x coverage; (2) using NovaSeq (Illumina) with paired-end reads of 151 base pairs to achieve a required 950x coverage.

[0193] Aligned with the genome:

[0194] Multiple reads were aligned to the Genome Reference Consortium Human Build 37 (GRCh37 / hg19) using Burrows-Wheeler version 0.7.834 with multiple default parameters, multiple replicate reads obtained from PCR clonality or multiple optical replicates, and multiple reads paired to multiple positions were excluded from downstream analysis.

[0195] Variant detection in pure genome sequencing data:

[0196] Using Freebayes software version 1.1.0-3-g961e5f335 with default parameters, single nucleotide substitutions and small insertions and deletions were identified. First, Freebayes ran the aligned sequencing data of the parents together, and then ran it on the aligned data of the CVS sample using the variant sites identified in the parental genomes. Since reported variants were not filtered, all reported SNPs and insertions and deletions were retained for downstream analysis.

[0197] Preprocessing of cell-free DNA data:

[0198] Free Bayes was run on the cell-free DNA samples only at variant sites identified in the parental genomes, and the allele observed for each read as well as the insert size of the read were stored in a separate database using Hoobari.

[0199] Non-invasive fetal mutation detection:

[0200] Hobari was run using the parental variants and cell-free DNA pre-processing result database as input, and the output was a standard variant call format (VCF) file. The results were analyzed using several software dedicated to VCF operations (e.g., vcflib and vcftools).

[0201] Bayesian non-invasive genotyping:

[0202] At each locus of interest, a Bayesian calculation is applied. For each possible fetal genotype:

[0203]

[0204] Where G is the fetal genotype, and G i is the i-th possible fetal genotype among n possibilities. For multiple biallelic variants, G iThe reference allele (AA) may be homozygous, heterozygous (Aa) or homozygous (aa) of the variant allele, and P(G) is the prior probability of each genotype, calculated by Mendel's law. The data variable represents the multiple reads covering a site, and P(data|G) represents the likelihood function, which in this example is defined as the product of the likelihood of each read:

[0205]

[0206] First reading paragraph j The probability depends on the genotype of the fetus and is calculated using the maternal genotype and the fetal fraction. j | fetus) and P(r j | mother) are the multiple probabilities of a read segment observation supporting a specific allele, assuming that the read segment is fetal or maternal. This depends on the fetal genotype G tested i , the maternal genotype G M and the observed alleles. P(fetal) and P(maternal) are the multiple probabilities of observing a fetal or maternal read based only on the fetal score, without considering the alleles it supports. In order to use the size difference between fetal and maternal fragments, the fetal score for each read is calculated only from multiple reads with the same fragment size. For multiple reads that are not correctly paired or have a fragment size greater than 500, the total fetal score is used.

[0207] Confirmation of the mutation in the E1 family:

[0208] After the birth of the previous litter of an affected offspring, a deleterious nonsense mutation (NM_000111.2, c.559G>T) in the SLC26A3 gene was found in both parents using Sanger sequencing, which causes congenital chloride diarrhea (CCD), which is inherited in an autosomal recessive manner. Sanger sequencing of the CVS sample confirmed that the fetus sequenced in the study was homozygous for the mutation.

[0209] Multiple results:

[0210] Fetal genotyping using multiple fragment sizes

[0211] In each family, the genomic DNA of all individuals was sequenced to a coverage range of 25x to 60x, and the multiple cell-free DNA samples were sequenced to multiple depths of 270x, 195x, 78x, and 56x (see Table 1 below). The calculated total fetal fractions were 32.4%, 24.9%, 14%, and 8.7%, respectively. Therefore, the four cases were classified according to two coverage levels: high (G1 to G2) and low (G3 to G4).

[0212] Table 1: Summary of multiple samples from multiple previous studies

[0213]

[0214]

[0215] The method of this embodiment is performed for each family to test the calibration of the method algorithm using the relationship between the plurality of posterior probabilities and the accuracy, and to determine whether adding multiple segment sizes increases the accuracy.

[0216] To test the accuracy with and without fragment size, the method was performed once with the total fetal fraction as a fixed parameter for each fragment and then again using a fetal fraction that depended on the fragment size. In all four families, the method showed satisfactory calibration and the addition of fragment size information improved accuracy. Figure 2A -C shows that only the mother is heterozygous at multiple SNPs ( Figure 2A ), only the father is heterozygous ( Figure 2B ) or both parents are heterozygous ( Figure 2C ) of multiple SNPs, the relationship between the accuracy and the posterior probability for multiple families G1 to G4. The accuracy is shown as a function of the threshold of the maximum posterior probability, which represents the decisiveness level of the multiple predictions. A calibrated model should show a smooth rising curve. For multiple low coverage cases (multiple families G3 to G4, Figures 2A to 2C , see Figure 2D ), lower calibration was observed, but the use of the fragment sizes showed an improvement in calibration.

[0217] Non-invasive prenatal insertion and deletion detection:

[0218] The methods of the present embodiments are also used for parental indel detection. Again, the total fetal fraction is used as a fixed parameter and the method is then performed with the fetal fraction for each fragment size. Figures 3A to 3C ( Figure 3DThe calibration of the model at multiple insertion and deletion sites where the mother is heterozygous is shown in Figure 1 ) with multiple color codes in Figure 1 . The accuracy is shown as a function of the maximum a posteriori probability threshold for each site in multiple insertion sites that are only maternally heterozygous (and paternally homozygous) for the cases of G1 to G4 ( Figure 3A ) or only paternal heterozygotes (maternal homozygotes) ( Figure 3B ) or both parents are heterozygous ( Figure 3C ). Method executions with a fetal fraction per fragment size are shown in red, and method executions with a fixed fetal fraction are shown in blue.

[0219] At a fixed fetal fraction, the low coverage families showed lower accuracy and inconsistent calibration results, while the high coverage families showed higher accuracy and very good calibration results. However, when the fragment size information was used, a considerable improvement in both accuracy and calibration was achieved ( Figures 3A to 3C When the method was applied to multiple sites where the father was heterozygous for an indel and the mother was homozygous for the reference or the indel, the accuracy was higher compared to the maternally heterozygous indel sites ( Figures 3A to 3C ). With the size information, all families show a better calibration effect, while the accuracy in the high coverage families is improved.

[0220] Simulation data analysis:

[0221] The model of the present embodiments was examined for multiple robustness at multiple low fetal fractions with high sequencing depths using simulated data, and six triplicates of multiple cell-free DNA samples with multiple varying values ​​of fetal fraction were subsampled from the G1 family while maintaining their sequencing depth. Each fetal part was further classified according to multiple sequencing depths, and at the highest fetal fraction with the maximum depth, the results showed high accuracy for each locus category: 96.0% for loci that were only maternally heterozygous, 99.1% for loci that were only maternally heterozygous, and 91.1% for loci where both parents were heterozygous ( Figures 4A to 4C ). For fetal fraction values ​​that are more common in the first trimester (i.e., 10 to 15%), the accuracies for the same categories of maximum sequencing depth were 88.4 to 92.4%, 99.1%, and 82.8 to 87.1%, respectively.

[0222] Three first trimester cases using deep WES / WGS analysis

[0223] Three families were sequenced (see Table 2 below) and their data were input into the method of this example. In two cases, WES was used to obtain a deeper coverage of the sequenced region (approximately 2 to 3% of the genome). In the third case, deep WGS was used. In the El family, the parents were multiple carriers of a mutation that causes congenital chloride diarrhea (CCD), an autosomal recessive disease.

[0224] Table 2:

[0225] Aggregation of multiple sequencing samples

[0226] family personal sample Sequencing region Depth of coverage* E1 Mother Multiple white blood cells WES 99 Plasma (11 weeks) WES 735 Father Multiple white blood cells WES 93 Descendants Chorionic villi WES 97 E2 Mother Multiple white blood cells WGS 36 Plasma (11 weeks) WES 664 Father Multiple white blood cells WGS 35 Descendants Chorionic villi WGS 36 G5 Mother Multiple white blood cells WGS 38 Plasma (11 weeks) WGS 310 Father Multiple white blood cells WGS 41 Descendants Chorionic villi WGS 38

[0227] *Median, by target.

[0228] Similar to before, the methods of the present embodiments were tested on multiple SNPs and insertions and deletions at multiple sites where the mother was heterozygous, Figures 5A to 5F (See Figure 5G The multiple color codes in the figure show the multiple SNP sites of multiple families E1, E2 and G5 ( Figures 5A to 5C ) and multiple insertion and deletion sites ( Figures 5D to 5F ) calibration, showing the accuracy as a function of the threshold of the maximum a posteriori probability.

[0229] The fetal fractions calculated in the cell-free DNA samples of families E1, E2, and G5 were 15.8%, 12.8%, and 18.5%, respectively. Figures 5A to 5F The method was demonstrated to be well calibrated. Furthermore, even though the overall accuracy was limited, an improvement was achieved by adding fragment length information in all cases. In the E2 family, the calibration was also improved. The algorithm was generally better calibrated for multiple SNPs, although the results were lower in the E2 family where WGS was used to verify multiple variants. For multiple indels, the model had a good calibration in the E2 family, with the G5 family showing the highest results in multiple categories for all SNPs and indels.

[0230] Of the multiple predicted SNPs, the fetal genotype of the mutation in SLC26A3 was successfully predicted, showing that the fetus was homozygous for the mutated allele, a result that matched the WES of chorionic villus sampling (CVS), which was further verified using Sanger sequencing. The prediction was more established when information about multiple fragment lengths was included. At the relevant sites, the post-occurrence probability was 61% with length information and 56% without the length information. As expected, the mutation was not detected in the E2 family.

[0231] Recalibration of mutation probability based on machine learning

[0232] In the step, a newly trained machine learning model is used to improve the results for each newly processed sample, the features in the model being taken from the metadata available when the parents and the cell-free DNA were genotyped (see Table A.1 in Appendix 1). The G1 family with the highest sequencing depth and fetal score was selected as the first training set, the G2 family was randomly divided, and 75% of the variants were used as a validation set for different models trained on G1. The random forest algorithm was found to have the best results. It has been found that in maternal and double heterozygous SNPs, the training should be performed using loci within the same category, for example: a model designed to improve the results for maternal heterozygous SNPs will only be trained on maternal heterozygous SNPs. In paternal heterozygous SNPs, training should be performed using all loci where at least one of the parents is heterozygous. Indels should be trained for SNPs and indels in the same category.

[0233] The selected model was tested on the remaining 25% of G2, the same model was trained again on the combined families G1 and G2, and tested on the G5 family, i.e., on a different data set. The area under the receiver operating characteristic curve (AUC) improved in all classes of loci, and accuracy improved in almost all cases ( Figures 6A to 6F). A significant improvement was observed in a case where both parents were heterozygous for a locus that originally showed a lower accuracy. It was also found that after recalibration, the multiple predicted probabilities better represented the actual accuracy, i.e., when the posterior probability was 0.9, the expected accuracy was also ~0.9. In addition, the accuracy at each predicted probability threshold was significantly improved, indicating that filtering of loci can indeed be performed using recalibration rather than setting multiple strict thresholds for multiple specific features. For example: at the threshold of 0.7, the accuracy of the G5 family was improved to 98.3-99.8% higher than multiple SNPs from a total of 722,630 maternally heterozygous loci, 1,358,503 paternally heterozygous loci, and 358,114 double heterozygous loci. The accuracy of insertion and deletion prediction was improved to 94.3 to 97.1% from 42,726 loci that were only maternally heterozygous, 142,577 loci that were only paternally heterozygous, and 20,388 doubly heterozygous loci. It was also found that multiple features derived from Hoobari, such as the prior probability and the posterior probability, the likelihood, and the predicted genotype had the highest importance, followed by multiple features related to the allele balance in the cell-free DNA.

[0234] Figures 7A to 7F The calibration shows the accuracy of the model ( Figure 7G , multiple color codes in ).

[0235] Shown are the accuracy before and after variant recalibration as a function of the predicted probability. The "perfect calibration" line represents a situation where the probabilities assigned to a prediction perfectly describe the correct multiple probabilities (in terms of accuracy), and the marginal distribution of loci in each bin is shown in the lower pane of each graph. Fig. 7A Corresponding to multiple SNPs that are only heterozygous for maternal origin, Figure 7B Corresponding to multiple SNPs that are only heterozygous for paternal origin, Figure 7C For multiple SNPs where both parents are heterozygous, Fig.7D Corresponding to multiple insertions and deletions that are only heterozygous for maternal origin, Fig. 7E Corresponding to multiple heterozygous insertions and deletions of paternal origin only, Figure 7F Corresponding to multiple insertions and deletions for which both parents are heterozygous.

[0236] Figures 8A to 8F shows the performance of the model after recalibration ( Figure 8G multiple color codes in ), Figures 8A to 8F The ability to perform machine learning-based variant recalibration to filter multiple variants is demonstrated. Figures 2A to 2D, 3A to 3D and 5A to 5G, in this specification, multiple accuracies for loci with a posterior probability above the multiple thresholds of the x-axis are shown. Multiple findings are provided before and after the variation recalibration. The total accuracy for each category is the accuracy of the leftmost point on the x-axis. The total counts used to calculate the accuracy for each threshold are shown at the bottom of each illustration. Fig. 8A Corresponding to multiple SNPs that are only heterozygous for maternal origin, Figure 8B Corresponding to multiple SNPs that are only heterozygous for paternal origin, Figure 8C Corresponding to multiple parental heterozygous SNPs, Fig.8D Corresponding to multiple insertions and deletions that are only heterozygous for maternal origin, Fig. 8E Corresponding to multiple insertions and deletions that are only heterozygous for paternal origin, Fig.8F Corresponding to multiple parental heterozygous insertions and deletions.

[0237] Figures 9A to 9G Multiple density plots of fragment length distribution of multiple families G1, G2, G3, G4, E1, E2 and G5, respectively.

[0238] Fig.10 Multiple confirmation results of the mutation in the SLC26A3 gene of the E1 family obtained using Sanger sequencing are shown, the first row shows a control group that is homozygous with the reference allele (G), the second row shows that the father is a carrier of the variant allele (T), and the third and fourth rows show the multiple results of the fetus using a forward primer (third row) and a reverse primer (fourth row), the fetus is homozygous with the variant allele.

[0239] discuss:

[0240] In this example, non-invasive fetal genotyping is demonstrated using the methods of the multiple embodiments of the present invention, the example confirms that the multiple size differences between multiple fragments of maternal and fetal origin improve fetal genotyping based on cell-free DNA, the multiple results are prominent at multiple sites where the mother is heterozygous, which currently pose the greatest identification challenge, and the example also demonstrates the ability of the methods of the multiple embodiments of the present invention to predict multiple fetal insertions and deletions, which are the second most common type of variation and can be deleterious, especially when they affect the reading frame.

[0241] One advantage of the Bayesian procedure is that it is modular, in the sense that it allows one to add any available information. In the present case, fragment size information is used, but other features may be included in addition or in lieu thereof, including other features of the cell-free fetal DNA, such as clusters of preferred termination positions. Haplotype information is also contemplated. Cell-free DNA enriched for cell-free fetal DNA may also be processed by the method according to some embodiments of the invention.

[0242] The accuracy calculations shown in the analysis of this example are based on multiple unfiltered raw results, so the actual accuracy is higher than the accuracy reported in this specification. The accuracy also depends on multiple functions that are not optimized in this study. In this example, predictions are applied to all sites, even multiple sites with lower credibility, and all relevant information used in each site is returned. Then, multiple different annotations, multiple statistical tests, and multiple machine learning programs can be applied to filter out multiple low-credibility results, as shown in the machine learning recalibration step. The consistency of the Bayesian program with the acceptance process of variation detection is another advantage allowed by the Bayesian program because the multiple posterior probabilities constitute the calibration parameters for filtering.

[0243] This example demonstrates exome sequencing and genome sequencing. Compared to deep WGS data, when the methods of the embodiments are applied to deep WES data, the accuracy obtained at multiple positions where the mother is heterozygous is not high, and the results are obtained in an early and clinically relevant week of pregnancy, during which DNA amplification is required. With amplification-free WGS, this example shows that the method is accurate even in the early stages of pregnancy. In order to achieve high accuracy in multiple smaller genomic regions, other library constructions and multiple sequencing methods will be tested.

[0244] This example shows that the NIPD with a large number of SGDs can be obtained using available technologies such as NGS. The probability scale used in this example ensures that some percentage of the multiple prediction sites have a specific prediction. The multiple sites can be used for downstream analysis to discover multiple rare variants that have not yet been discovered. Due to the low cost of WES, according to some embodiments of the present invention, WES can be used for large population studies. Since a large and unified data set can be created using WES, the data set can be further analyzed to improve the model, so WES has advantages over conventional techniques.

[0245] Example 2:

[0246] This example shows that a deep learning program can be used to accurately detect multiple small variations in the fetus in a non-invasive manner. In some embodiments of the present invention, the deep learning program can replace the Bayesian model and the machine learning recalibration process used in Example 1, and can be used in an end-to-end framework. One of the advantages of the multiple embodiments is the deep learning program, which can automatically model multiple systematic errors. This is useful in difficult-to-interpret data (e.g., low-depth sequencing). The multiple inventors found that the end-to-end deep learning performs better than the Bayesian model. For example: the deep learning program can handle low fetal scores and sequencing coverage, which are the multiple main limiting factors in this task.

[0247] Unlike Bayesian models, the deep learning program can provide an approximation of an unknown interdependent likelihood function, which exploits the relationship between multiple reads covering a position in all three individuals and multiple adjacent nucleotides. The multiple inventors of the present invention have discovered that the use of deep learning programs can be applied to multiple features other than multiple lengths (e.g., the nucleotide sequence itself). The deep learning program can be applied to various sequencing platforms (Illumina, Nanopore, etc.) and multiple methods (WES, WGS, panels, etc.).

[0248] Fig.11 Summarizing the Bayesian model (in Fig.11 Hoobari in Fig.11 The multiple differences between the two are called Deep Hoobari.

[0249] Artificial Neural Networks:

[0250] The artificial neural network is optionally and preferably a convolutional neural network (CNN).

[0251] Feature extraction:

[0252] The data of the cell-free DNA and the parent's DNA can be represented using the multiple aligned reads and their metadata, such as: multiple base qualities (information at the nucleotide level) or multiple pairing qualities (information at the read level) of the multiple aligned reads or any one of the features listed in Table A.1 of Appendix 1. Alternatively or additionally, variation detection software (such as but not limited to: GATK's HaplotypeCaller, Google's DeepVariant or Freebayes) can be used to extract information related to multiple alleles that may be supported, the information is the level of the genomic locus, and the cell-free DNA, mother and father can use different levels of data. In some embodiments of the present invention, multiple data sources are reduced. For example: by reducing the coverage depth used and the cell-free fetus / or avoiding the use of the parent information.

[0253] Stacked tensor of cell-free DNA:

[0254] During the learning phase, all data corresponding to a particular locus in the genome are optionally and preferably represented by an input multidimensional tensor associated with a marker containing the true fetal genotype found using an invasive test, and the CNN receives a batch of input multidimensional tensors and multiple labels of the batches of input multidimensional tensors in each iteration (forward and backward propagation).

[0255] The tensor can be, for example, a stack of multiple reads covering a candidate SNP centered around the evaluated SNP. In the manner, the first two (e.g., width and height) of the tensor correspond to the length of the stack and the number of reads, and the multiple reads can be encoded in any technique, including but not limited to: single-hot encoding, group-hot encoding, etc. The third dimension (e.g., depth) of the tensor is the number of multiple input channels of the CNN, each channel providing metadata for the SNP. In some embodiments of the present invention, the information represented by the multiple channels is processed by multiple multiplication functions, wherein the outputs of the multiple multiplication functions constitute the multiple channels within the tensor.

[0256] Experimental design:

[0257] Multiple datasets:

[0258] Tables 3A and 3B below list the multiple datasets used in this study, which was designed to prevent overfitting of the prediction model. Four families were analyzed: families G1 to G2 contained high-quality data and had been sequenced in a previous study; family G5 included high-quality data, which was biologically challenging because it was sequenced from the first trimester; and family E1 was also challenging because it was sequenced using WES, which was shown to have more errors.

[0259] Table 3A:

[0260]

[0261] Table 3B:

[0262]

[0263] a) Chan et al., 2016; b) Kitzman et al., 2013; 1) Multiple biallelic variants found on multiple autosomes, with relaxed filtering of ≥5 parental and fetal depths and ≥20 cell-free DNA depths; 2) Whole-genome sequencing (WGS) performed with a PCR-free library construction protocol; 3) Median, target achieved; 4) Gestational age.

[0264] Phase 1:

[0265] The first phase of the study was to select the re-expression of the data, the structure of the tensors and the basic architecture of the network based on the dataset G1, which was divided into a training set (80% of the multiple variants) and a test set (20% of the multiple variants). The training was further divided into 90% training and 10% multiple validation sets. Since the number of multiple variants is still large, most of the experiments require further sampling of the multiple variants. Once a satisfactory architecture was selected, it was tested only once on the test set - the architecture was also tested in G2 to present an optimal situation where the multiple technical differences between multiple families were minimal (sequencing multiple families using the same method in the same laboratory).

[0266] Phase 2:

[0267] The datasets G1 and G2 were merged into a dataset divided into training, validation and test sets, and the architecture selected in stage 1 was trained and tested. Thereafter, the architecture was tested on dataset G5 to show a more challenging scenario, where there were technical differences between the families used for training and the families used for testing, which demonstrated the generalizability of the model.

[0268] Phase 3:

[0269] Dataset E1 is similarly partitioned, but with a larger test set to maintain a sufficient amount of variation. The models of the previous stages are used for transfer learning, where the trained models are further trained and fine-tuned. In a separate or alternative process, the architectures of the previous stages are trained based solely on the training set of E1.

[0270] Other experiments:

[0271] The multiple experiments in stages 1 to 3 can be repeated using multiple indels, and multiple further experiments can include (1) downsampling the sequencing depth and fetal fraction; (2) excluding data from the father; (3) excluding data from both parents; (4) reducing the required parental information to only the genotype, such that only the cell-free DNA requires NGS.

[0272] Multiple results:

[0273] In stage 1, the tensor was based on multiple aligned reads and GATK's HaplotypeCaller was used to perform a more accurate realignment and assign supported alleles to each read. Since HaplotypeCaller is not designed for multiple pregnancy cell-free DNA samples and treats the multiple samples as coming from one individual, the assignment of multiple alleles was assigned without detecting the genotype. A tensor based on the alignment of the multiple reads of the cell-free DNA and parents was found to match the accuracy of Hoobari. The results are shown in Figures 12A to 12P , for the training set ( Fig. 12A , 12B , 12E, 12F, 12I, 12J, 12M, 12N) and the validation set ( Fig. 12C , 12D , 12G, 12H, 12K, 12L, 12O, 12P) show a simple CNN (red), a CNN with multiple cell-free DNA reads sorted by fragment length (green), a CNN with cell-free DNA reads sorted by fragment length and parental reads sorted by supported alleles (blue), and an Inception version 3 CNN (yellow). Figures 12A to 12P As shown in the figure, for the training set ( Fig. 12A ); for the validation set ( Fig. 12C); divided into 3 groups of variations according to the parental genotype: only maternal heterozygosity (mat-het), only paternal heterozygosity (pat-het) or both parents are heterozygous (both), for the 3 possible fetal genotypes: homozygous with the reference allele (homref), the variant allele (homalt) and the heterozygote (het). The multiple results of the deletion ( Fig. 12B , 12D) is not comparable between the multiple models and is therefore not considered.

[0274] At positions where both parents were heterozygous, it was previously difficult to predict the genotype of the fetus, and the results showed a significant improvement compared to Hoobari. The results were improved when the parental reads were sorted by supported alleles and the cell-free DNA was sorted by fragment length (each read was from a fragment with a different length).

[0275] It should be understood that more complex tensors and multiple network architectures can be used. For example: in a one-hot encoded tensor, each nucleotide is represented by a combination of 1 and 0 representing its base quality, multiplied by the respective credibility. The bases A, C, G and T can be encoded on four channels, for example: encoded as 1000, 0100, 0010 and 0001 respectively. For example: if the credibility of the base A is 0.999, the first channel is assigned 1*0.999=0.999, and the other channels will show 0*0.999=0. Another example is a group hot encoding model, in which the hot encoded nucleotides are multiplied by a vector of all other features, and the multiple results of the multiple representations of the data are in Figures 13A to 13P As shown in FIG. 1 , multiple representations are shown as the single heat (blue) and group heat (pink). Figures 12A to 12P The same groups are shown to show accuracy and omissions. Figures 13A to 13P As shown, the one-hot and group-hot representations improve the ability of the CNN to learn the multiple features of the data.

[0276] Figures 14A to 14P A plurality of results obtained using a model in which the parent information includes only the genotypes of the parents without information at the read level or the nucleotide level are shown, wherein the plurality of results of the plurality of representations of the data are Figures 13A to 13P , multiple representations are single heat (blue), group heat (light blue), and for the case where the only parental information used is the multiple parental genotypes (homozygous or heterozygous with the reference or variant allele) (pink), as shown in Figures 12A to 12P Multiple identical groups are shown to illustrate accuracy and omissions.

[0277] Figures 14A to 14P We show that high accuracy can be achieved even with only the multiple parental genotypes, which means that it is not necessary to sequence the multiple parents using NGS and other technologies may suffice. However, the read and nucleotide-level information from NGS appears to improve the multiple results.

[0278] Attachment 1:

[0279] Table A.1:

[0280] Multiple features used in the multiple machine learning models:

[0281]

[0282]

[0283]

[0284]

[0285]

[0286]

[0287]

[0288]

[0289]

[0290]

[0291]

[0292] Although the present invention has been described in conjunction with its specific embodiments, it is apparent that many alternatives, modifications and variations are apparent to those skilled in the art. Therefore, the present invention is intended to encompass all such alternatives, modifications and variations that fall within the spirit and scope of the claims.

[0293] All publications, patents and patent applications mentioned in this specification are incorporated into this specification in their entirety as references, to the same extent as if each independent publication, patent or patent application was clearly and individually marked as being incorporated herein by reference. In addition, the citation or identification of any reference in this application should not be construed as an admission that the reference can be used as prior art of the present invention. The title part in this application is used herein to make this specification easy to understand and should not be construed as a necessary limitation.

[0294] In addition, any priority documents (multiple) of the present application are hereby incorporated by reference into this specification in their entirety.

[0295] Multiple references:

[0296] 1. Lo, YMD, et al., Digital PCR for the molecular detection of fetal chromosomal aneuploidy, Proceedings of the National Academy of Sciences of the United States of America 104, 13116-13121 (2007).

[0297] 2. Fan H.C., Blumenfeld Y.J., Chitkara U., Hudgins L., and Quake S.R. Noninvasive diagnosis of fetal aneuploidy by shotgun sequencing of DNA from maternal blood. Proceedings of the National Academy of Sciences of the United States of America, 105, 16266-16271 (2008).

[0298] 3. Hill M., Compton C., Lewis C., Skirton H., and Chitty L. S., Determination of fetal sex in pregnant women at risk for hemophilia: a qualitative study exploring clinical practices and attitudes among healthcare professionals in the UK, Hemophilia 18, 575-583 (2012).

[0299] 4. Lewis C., Hill M., Skirton H., and Chitty L.S. Non-invasive prenatal diagnosis for determining the sex of the fetus: multiple advantages and disadvantages from the perspective of users of the service. European Journal of Human Genetics 20, 1127-1133 (2012).

[0300] 5. Finning K., Martin P. and Daniels G. Prediction of fetal rhesus (Rh) D blood type using free fetal DNA in maternal plasma in a clinical service in the United Kingdom, Annals of the New York Academy of Sciences 1022, 119-123 (2004).

[0301] 6. Minon J.-M., Gerard C., Senterre J.-M., Schaaps J.-P., and Foidart J.-M. Routine fetal RHD genotyping with maternal plasma: a four-year experience in Belgium, Blood Transfusion (Paris) 48, 373–381 (2008).

[0302] 7. Mahdieh N. and Rabbani B., An overview of methods for detecting multiple mutations in multiple genetic diseases, Iranian Journal of Pediatrics 23, 375-388 (2013).

[0303] 8. Yang Y. et al., Clinical whole-genome sequencing for the diagnosis of Mendelian diseases, New England Journal of Medicine 369, 1502–1515 (2013).

[0304] 9. Isakov O., Perrone M., and Shomron N. Exome sequencing analysis: a guide to disease variant detection. Methods Biology, Clifton, NJ 1038, 137–158 (2013).

[0305] 10. Meng L. et al., Utility of exome sequencing in multiple infants admitted to multiple intensive care units: identification of multiple severe monogenic disorders and implications for medical management, JAMA Pediatrics 171, e173438–e173438 (2017).

[0306] 11. Mackie F.L., Carss K.J., Hillman S.C., Hurles M.E., and Kilby M.D. Exome sequencing of multiple fetuses with multiple structural malformations. J Clin Med 3, 747–762 (2014).

[0307] 12. Vora N.L et al., Prenatal exome sequencing of multiple fetal anomalies: multiple new opportunities and challenges, Medical Genetics 19, 1207 (2017).

[0308] 13. Kitzman J.O. et al., Non-invasive whole-genome sequencing of the human fetus, Science Translational Medicine 4, 137ra76 (2012).

[0309] 14. Fan H.C. et al., Non-invasive prenatal measurement of the fetal genome, Nature 487, 320-324 (2012).

[0310] 15. Hill M. et al. Noninvasive prenatal diagnosis for cystic fibrosis: testing multiple paternal mutations, exploring multiple patient preferences, and cost analysis. Prenatal Diagnosis 35, 950–958 (2015).

[0311] 16. Lun F. MF et al. Noninvasive prenatal diagnosis of multiple monogenic diseases by digit size selection and relative mutation dosage in maternal plasma DNA. Proceedings of the National Academy of Sciences of the United States of America 105, 19920-19925 (2008).

[0312] 17. Lam K.-WG. et al. Targeted massively parallel sequencing of maternal plasma for non-invasive prenatal diagnosis of multiple monogenic diseases: application to β-thalassemia. Clinical Chemistry 58, 1467–1475 (2012).

[0313] 18. Lo Y. MD et al., Maternal plasma DNA sequencing reveals the genome-wide genetic and mutational characteristics of the fetus, Science Translational Medicine 2, 61ra91-61ra91 (2010).

[0314] 19. Chen S. et al., Accurate and non-invasive fetal whole genome recovery assisted by haplotypes by maternal plasma sequencing, Genomic Medicine 5, 18 (2013).

[0315] 20. Chan K. C. A. et al., Second-generation noninvasive fetal genomic analysis reveals de novo mutations, single-base parental inheritance, and multiple preferred DNA ends, Proceedings of the National Academy of Sciences of the United States of America 201615800 (2016), Digital Object Identifier (doi): 10.1073 / pnas.1615800113

[0316] 21. Snyder M.W. et al., Noninvasive fetal genome sequencing: a primer, Prenatal Diagnosis 33, 547–554 (2013).

[0317] 22. Snyder M. W., Adey A., Kitzman J. O., and Shendure J., Haplotype-resolved genome sequencing: multiple experimental approaches and multiple applications, Nature Reviews Genetics 16, 344–358 (2015).

[0318] 23.Jenkins L.A., Deans Z.C., Lewis C., and Allen S. An approved noninvasive prenatal diagnosis service for multiple single gene disorders with multiple best practice recommendations, Prenatal Diagnosis (2017), doi: 10.1002 / pd.5197

[0319] 24. Chan K. CA et al. Multiple size distributions of maternal and fetal DNA in maternal plasma. Clinical Chemistry 50, 88-92 (2004).

[0320] 25. Fan H.C., Blumenfeld Y.J., Chitkara U., Hudgins L., and Quake S.R., Size distribution of cell-free DNA in fetal and maternal cells analyzed by paired-end sequencing, Clinical Chemistry, 56, 1279-1286 (2010).

[0321] 26. Yu S. CY et al., Multiple size-based molecular diagnostics for noninvasive prenatal testing using plasma DNA, Proceedings of the National Academy of Sciences of the United States of America 111, 8583–8588 (2014).

[0322] 27.Cirigliano V., Rueda L., Syngelaki A., and Nicolaides K.H., Performance of the neoBona assay: a novel paired-end massively parallel shotgun sequencing approach for aneuploidy screening based on cell-free DNA, Ultrasound in Obstetrics and Gynecology 49, 460–464 (2017).

[0323] 28. Sun K. et al., COFFEE: Uncontrolled non-invasive fetal chromosome testing using maternal plasma DNA, Prenatal Diagnosis 37, 336–340 (2017).

[0324] 29. Sillence K. Enrichment of cell-free fetal DNA (cffDNA) for non-invasive prenatal testing (NIPT): a comparison of multiple molecular techniques (2016).

[0325] 30. Mullaney J.M., Mills R.E., Pittard W.S., and Devine S.E. Multiple small insertions and deletions (INDELs) in the human genome. Human Molecular Genetics 19, R131-R136 (2010).

[0326] 31. Neuman J.A., Isakov O., and Shomron N. Analysis of insertions and deletions from deep sequencing data: evaluation of software for optimal detection. Brief introduction to bioinformatics 14, 46–55 (2013).

[0327] 32. Jiang Y., Turinsky A.L., and Brudno M., Indels and deletions: an estimate of insertion and deletion variation in the human genome and analysis of multiple factors that impede detection, Nucleic Acids Res. 43, 7217-7228 (2015).

[0328] 33. Hwang S., Kim E., Lee I., and Marcotte E.M., Systematic comparison of multiple variant detection pipelines using multiple gold standard personal exome variants, Scientific Reports 5, 17875 (2015).

[0329] 34. Li, H. and Durbin, R. Fast and accurate short-read alignment using the Burrows-Wheeler transform. Bioinformatics 25, 1754-1760 (2009).

[0330] 35. Garrison E. and Marth G., Haplotype-based variant detection based on short-read sequencing, ArXiv 12073907 Q-Bio (2012).

[0331] 36. Danecek P. et al., The variant detection format and multiple VCF tools (VCFtool), Bioinformatics 27, 2156-2158 (2011).

[0332] 37. Sequence Alignment / Map Format Specification, May 22, 2018, www.samtools.github.io / hts-specs / SAMv1.pdf.

[0333] 38. Chan KCA, Jiang P, Sun K, Cheng YKY, Tong YK, Cheng SH, Wong AIC, Hudecova I, Leung TY, Chiu RWK, et al. 2016. Second-generation noninvasive fetal genomic analysis reveals de novo mutations, single-base parental inheritance, and preferred multiple DNA ends. Proceedings of the National Academy of Sciences of the United States of America, 201615800.

[0334] 39. Fan HC, Gu W, Wang J, Blumenfeld YJ, El-Sayed YY, Quake SR. 2012. Non-invasive prenatal measurement of the fetal genome. Nature 487: 320–324.

[0335] 40. Kitzman JO, Snyder MW, Ventura M, Lewis AP, Qu R, Simmons LE, Gammill HS, Rubens CE, Santillan DA, Murray JC, et al., 2012, Noninvasive whole-genome sequencing of the human fetus. Science Translational Medicine 4:137ra76.

[0336] 41. Luo R, Sedlazeck FJ, Lam TW, Schatz M., 2018. Clairvoyante: a multitask convolutional deep neural network for variant detection in single-molecule sequencing. bioRxiv310458.

[0337] 42. Poplin R, Chang PC, Alexander D, Schwartz S, Colthurst T, Ku A, Newburger D, Dijamco J, Nguyen N, Afshar PT, et al. 2018. Creating a universal SNP and small insertion and deletion variant detector with deep neural networks. bioRxiv 092890.

[0338] 43. Torracinta R, Campagne F., 2016. Training multiple genotype detectors with multiple neural networks. bioRxiv097469.

Claims

1. A computer software product comprising: A computer-readable medium storing a plurality of program instructions, wherein when the plurality of program instructions are read by a data processor, the data processor receives maternal cell-free DNA data and paternal cell-free DNA data paired with a fetus, and executes a method for fetal genotyping, wherein the method for fetal genotyping comprises the following steps: Receiving maternal genomic DNA data, maternal cell-free DNA data, and paternal genomic DNA data that are paired with the fetus; analyzing the data to identify a first set of multiple sites at which the parents are homozygous for multiple different alleles and a second set of multiple sites at which at least one of the parents has a mutation, wherein the analyzing comprises identifying, among the multiple reads covering the first set of multiple sites, a first group of multiple reads and a second group of multiple reads, the first group of multiple reads including only multiple reads representing multiple non-maternal alleles but not other reads, and the second group of multiple reads including all other reads, and determining a probability is based on a plurality of differences between the first group of multiple reads and the second group of multiple reads; For each locus in the first set, determining a probability that a corresponding portion of the maternal cell-free DNA data was obtained from the fetus, wherein determining the probability is based at least on an observed template length or a fetal fraction related to length; and Each locus of the second set is classified as fetal or maternal according to the plurality of probabilities to genotype the fetus, wherein the classification is performed by applying a Bayesian method.

2. A fetal genotyping system, comprising: An input circuit receives maternal cell-free DNA data and paternal cell-free DNA data paired with the fetus; and a data processor configured to analyze the data to identify multiple sites where the parents are homozygous, so as to perform a method for fetal genotyping, wherein the method for fetal genotyping comprises the following steps: Receiving maternal genomic DNA data, maternal cell-free DNA data, and paternal genomic DNA data that are paired with the fetus; analyzing the data to identify a first set of multiple sites at which the parents are homozygous for multiple different alleles and a second set of multiple sites at which at least one of the parents has a mutation, wherein the analyzing comprises identifying, among the multiple reads covering the first set of multiple sites, a first group of multiple reads and a second group of multiple reads, the first group of multiple reads including only multiple reads representing multiple non-maternal alleles but not other reads, and the second group of multiple reads including all other reads, and determining a probability is based on a plurality of differences between the first group of multiple reads and the second group of multiple reads; For each locus in the first set, determining a probability that a corresponding portion of the maternal cell-free DNA data was obtained from the fetus, wherein determining the probability is based at least on an observed template length or a fetal fraction related to length; and Each locus of the second set is classified as fetal or maternal according to the plurality of probabilities to genotype the fetus, wherein the classification is performed by applying a Bayesian method.

3. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method further comprises the step of calculating a total fetal fraction, wherein the step of classifying further comprises using the total fetal fraction.

4. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method further comprises calculating a total fetal fraction and constructing a fetal size distribution and a maternal size distribution, wherein the step of classifying comprises binning the fetal size distribution and calculating a fetal fraction for each segment size bin, and calculating a probability that the segment is fetal for at least one site and at least one segment at the at least one site based on the fetal fraction of the corresponding segment size bin to which the segment belongs.

5. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The Bayesian method includes a plurality of prior probabilities calculated using sequencing data of at least one of the parents.

6. The computer software product or system according to claim 5, characterized in that: The sequencing data includes whole genome sequencing.

7. The computer software product or system according to claim 5, characterized in that: The sequencing data includes whole exome sequencing.

8. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method also includes using machine learning to recalibrate the output of the Bayesian method.

9. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The step of classifying includes applying a machine learning program to the plurality of loci to genotype the fetus.

10. The computer software product or system according to claim 9, characterized in that: The machine learning program includes a deep learning program.

11. The computer software product or system according to claim 9, characterized in that: The method also includes recalibrating the output of the machine learning.

12. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is applied to multiple heterozygous sites in the maternal cell-free DNA data.

13. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is applied to multiple homozygous sites in the maternal cell-free DNA data.

14. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The genotyping step includes identifying fetal monogenic diseases that originate from the paternal line.

15. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The genotyping step includes identifying fetal monogenic diseases of maternal origin.

16. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The genotyping step includes predicting inherited multiple insertions and multiple deletions.

17. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The genotyping step includes identifying fetal chromosomal abnormalities.

18. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is applied to non-invasive prenatal diagnosis of single gene diseases.

19. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is applied to non-invasive prenatal diagnosis of polygenic diseases.

20. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is applied to non-invasive prenatal diagnosis of genetic diseases, wherein the genetic diseases are selected from the group consisting of Bloom's syndrome, Canavan disease, cystic fibrosis, familial dysautonomia, Riley-Day syndrome, Fanconi anemia group C, Gaucher's disease, glycogen storage disease type 1a, maple syrup urine disease, mucolipid storage disease type IV, Niemann-Pick disease, Tay-Sachs syndrome, beta thalassemia, sickle cell anemia, alpha thalassemia, beta thalassemia, factor XI deficiency, Friedreich's ataxia, fatty acid metabolism disorders, Parkinson's disease-juvenile, spinal muscular atrophy, Rett syndrome, phenylketonuria, Becker muscular dystrophy, Duchenne muscular dystrophy, fragile X syndrome, hemophilia A, early-onset Alzheimer's disease, breast cancer / ovarian cancer, colon cancer, diabetes, Huntington's disease, myotonic muscular dystrophy, early-onset Parkinson's disease, Pits-Jeghers syndrome, polycystic kidney disease, and torsion dystonia.

21. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is applied to non-invasive prenatal diagnosis of genetic diseases, and the method also includes prenatal or postnatal treatment of the genetic diseases.

22. The computer software product according to claim 1 or the system according to claim 2, characterized in that: The method is used for non-invasive prenatal diagnosis of congenital diseases.

23. The computer software product or system according to claim 22, characterized in that: The congenital disease is selected from the group consisting of malformation, neural tube defect, chromosomal abnormality, spina bifida, and cleft lip and palate.

24. The computer software product or system according to claim 23, characterized in that: The chromosomal abnormality is selected from the group consisting of Down syndrome, Turner syndrome and Klinefelter syndrome.

25. The computer software product or system according to claim 23, characterized in that: The deformities include limb deformities.

26. The computer software product or system according to claim 25, characterized in that: The limb deformity is selected from the group consisting of ectromelia, embolism, phocomelia, polymelia, syndactyly, polysyndactyly, oligodactyly, brachydactyly, achondroplasia, congenital malformations or hypoplasia, amniotic band syndrome and cleidocranial dysplasia.

27. The computer software product or system according to claim 23, characterized in that: The malformations include congenital malformations of the heart.

28. The computer software product or system according to claim 27, characterized in that: The congenital malformation is selected from the group consisting of patent ductus arteriosus, atrial septal defect, ventricular septal defect and tetralogy of Fallot.

29. The computer software product or system of claim 23, wherein: The malformations include congenital malformations of the nervous system.

30. The computer software product or system of claim 29, wherein the congenital malformation of the nervous system is selected from the group consisting of multiple neural tube defects such as spina bifida, meningocele, meningomyelocele, encephalocele and anencephaly, Einard-Charlie syndrome, Dandy-Walker malformation, hydrocephalus, microcephaly, macrocephaly, lissencephaly, polymicrogyria, holoprosencephaly and agenesis of the corpus callosum.

31. The computer software product or system of claim 23, wherein: Such malformations include congenital malformations of the gastrointestinal system.

32. The computer software product or system according to claim 31, characterized in that: The congenital malformation of the gastrointestinal system is anal atresia.

Images