WO2014098479A1 - Computer implemented method for analyzing genomic mutation or epigenetic mutation - Google Patents

Abstract

The present invention relates to a novel mutation analysis protocol which can obtain a clinically meaningful prediction result by more accurately analyzing diverse variations (for example, genetic variations) found in an organic body. The present invention can accurately predict susceptibility to certain traits of one individual. The prediction accuracy of the present invention indicates accuracy which is increased several times when compared to random prediction, and this prediction degree can determine a state of health of an individual or population, thereby exhibiting a much improved prediction accuracy.

Description

 【Specification】

[Name of invention]

 Computer-implemented method for analyzing genome or epigenetic variation

 The present invention relates to a computer im lemented method for analyzing genome variation or epigenetic variation of an organism, a computer-readable storage medium and a system therefor. [Background technology]

Individual whole genome (WG) sequence information has revolutionized the understanding of systematic characterization genome variation in the human genome (ES Lander (2011) .Initial impact of the sequencing of the human genome.Nature 470 (7333). : 187-197). Most parts of human genomes are known to have identical sequences, and only small parts have genome variations such as monobasic polymorphism (SNP), variable length indels, copy number variations, and variable length repeats or inversions. (M. Snyder, et al. (2010) .Personal genome sequencing ^' -current approaches and challenges.Genes Dev 24 (5): 423-431). Of these, most of the mutations are SNPs, which have been identified at 3M genome locations with minor allele frequencies of more than 5% (International HapMap Consortium (2005) .Ha lotype map of the human genome. Nature 437 (7063). 1299-1320; The International HapMap Consortium (2007) .A second generation human haplotype map of over 3.1 million SNPs.Nature 449 (7164): 851-861).

These non-genome factors are thought to be linked to phenotypic variations. One of the objectives of research in the genome-wide association study (GAS) is to treat diseases such as cancer, chronic diseases, neurological and infectious diseases. To assess the genome factor for the disease suscept ibi lity of an individual. Intensive studies to determine the association between the SNP genotype and cancer phenotype revealed about 100 genome-sensitive loci for 16 cancers, some of which were associated with many loci, and others of only a few. Fletcher & RS Houlston (2010) .Architecture of inherited susceptibility to common cancer.Nat Rev Cancer 10 (5): 353-361), but few specific causal loci were identified ( G. Gibson (2011) .Rare and common variants: twenty argument s. Nat Rev Genet 13 (2): 135-145). Furthermore, some researchers have criticized that GWAS-defined loci did not explain the high familial risk of most cancers (TA Manolio, et al. (2009) .Finding the missing heritabi lity of complex diseases.Nature 461 ( 7265): 747-753). Thus, the results obtained by current analytical methods and their interpretations for predicting disease sensitivity are so vague that they do not have clinical utility for individuals or populations. Throughout this specification, many papers and patent documents are referenced and their citations are indicated. The disclosures of cited papers and patent documents are incorporated herein by reference in their entirety, so that the level of the technical field to which the present invention belongs and the contents of the present invention are more clearly explained.

[Content of invention]

 Problem to be solved

The present inventors have tried to solve the above-mentioned problems of the prior art. As a result, the inventors have developed a novel variation analysis protocol that can more accurately analyze various variations (eg, genetic variations) found in an organism to obtain clinically meaningful predictive results. The present invention is an approach similar to "word" and "word frequency profile" in natural language analysis, in which two distinct concepts, variation syntax (VAR-S; SNP), are applied. In case of application, we propose a method of expressing and analyzing the organizational characteristics of an individual using SNP syntax: SNP-S) and VAR-S feature frequency profile (FFP).

In addition, ^"the inventors mutation analysis that is to improve the accuracy of the transformation decision of the objects and characters at least two kinds of descriptions for variations applied to each of at least two types of class prediction algorithm, reasoning appropriate the results from this application Another protocol has been developed for the algorithm.

 Accordingly, an object of the present invention is to provide a computer im lemented method for analyzing genomic variat ion or epigenomic variation of an organism.

 It is another object of the present invention to provide computer-readable instructions embodying instructions instructing a computer processor to perform steps for analyzing genomic variation or epigenomic variation of an organism. (computer—readable) storage media.

 It is still another object of the present invention to provide a system for analyzing genomic variation or epigenomic variation of an organism.

 Other objects and advantages of the present invention will become apparent from the following detailed description, claims and drawings.

[Task solution]

 The analysis method using the variation syntax (VAR-S) which is the 1st invention of this invention is as follows:

According to an aspect of the invention, the invention provides a computer implemented method for analyzing genomic variation or epigenomic variation of an organism comprising the following steps: (a) constructing a linked string of the variants;

(b) constructing a variation syntax (VAR-S) of a specific length by applying a sliding window of a specific length along the entire length of the linking string; .

 (C) counting all possible features in the particular length variation syntax and assembling them into feature frequency profiles (FFPs); And

 (d) determining the distance between the FFPs or classifying the FFPs.

 The present inventors have tried to solve the above-mentioned problems of the prior art. As a result, the inventors have developed a novel variation analysis protocol that allows for more accurate analysis of various variations (eg, genetic variations) found in an organism to yield clinically meaningful predictive results. The present invention is similar in approach to "words" and "word frequency profiles" in natural language analysis, with two distinct concepts: variation syntax (VAR-S; SNP syntax: SNP-S) when applied to SNPs. And a method of expressing and analyzing organizational characteristics of an individual using a feature frequency profile (FFP) 'of VAR-S.

More specifically, according to embodiments of the SNP and cancer susceptibility prediction of the present invention, the inventors of the present invention describe "words" and "word frequency profiles" (CD Manning & H. Schuetze (1999). Foundations of Statistical Natural Language In a similar manner to The MIT Press, 1 edn), two inventive ideas, "SNP syntax (SNP syntax: SNP-S)" and "Feature Frequency Profile (SFP-S)" We present for the first time a novel method of expressing the systematic features of an individual's WG SNPs, and as a proof of concept, cancer from the TCGACThe Cancer Genome Altas database. The FFPs of SNP-Ss analyzed from the WG SNP genotype of the patient's blood samples are compared to predict quantitative genetic sensitivity for eight types of cancer.

 The rapid accumulation of individual genome sequences predicts the discovery of new genetic properties, which will continue to revise, expand and diversify our models of cancer and other diseases. In recognition of this continuous development model for disease, we present a general and comprehensive model for cancer susceptibility. In particular, our study is open to as many alternatives as possible, and the results of this study will identify evidence supporting the following assumptions or hypotheses:

 1. Each genetic susceptibility allele to the cancer itself can have minor detrimental phenotypic consequences, pose minor hereditary risks, have variable penetrance, and in small portions of population Appear, and generally cause little "fatal" results. Thus, each genetic susceptible allele can occur at minor frequencies in populations (ET Cirulli & DB Goldstein (2010) .Uncovering the roles of rare variants in common disease through whole-genome sequencing.Nat Rev Genet 11 ( 6): 415-425);

 2. One particular cancer type has many subtypes, and in a single individual, each subtype is caused by many genes, with varying degrees of effectiveness (M. Ger linger, et. (2012) .Intratumor Heterogeneity and Branched Evolution Revealed by Mult i region Sequencing.N Engl J Med 366 (10): 883-892), multiple inheritances (genetic code sequences and nongenic ciphers) caused by various genetic variations Sequence) alleles make one individual sensitive to cancer, most of which can be represented as minor alleles in the genome of one individual;

3. Among the different cancer subtypes in one particular cancer and even among different cancer subtypes in a single tumor, there is a heterogeneity of cancer cells with different somatic mutations (M. Ger linger, et al. (2012). Intratumor Heterogeneity and Branched Evolution Revealed by Multi region Sequenc ing. N Engl J Med 366 (10): 883—892); And,

 4. Each set of genetic susceptibility alleles, together with its own and / or non-genetic events (e.g., events caused by pathogens, radiation, compounds, environmental factors, etc.), can trigger cancer development followed by driver mutations. Followed by one or more consecutive waves of clonal expansion that depend on one or more consecutive acquisition of pj Stephens, et al. (2012) .The landscape of cancer genes and mutational processes in breast cancer.Nature.486 (7403) ): 400-404). Thus, the correlation between cancer driver alleles and cancer sensitive alleles may appear, may be direct or potent, or not.

 Genome mutations to be analyzed in the present invention include various variations found in an organism, preferably SNPCsingle nucleotide polymorphisms in nucleotide sequences), deletions, insertions or repeat variations; Or epigenomic variation. Examples of epigenetic variations include DNA methylation or histone modifications. Most preferably, the mutation to be analyzed in the present invention is SNP.

 In the present invention, the mutation to be analyzed is a mutation present in a nucleotide sequence, and the nucleotide sequence is a sequence on one chromosome, a sequence on a plurality of chromosomes, or a whole genome sequence, more preferably a whole genome (WG). )

 Most preferably, the mutations to be analyzed in the present invention are SNPs in the entire genome sequence.

 According to a preferred embodiment of the present invention, step (a) is carried out by assigning a code to each of the variants to build an associative string of the codes.

More preferably, step (a) is performed by assigning a code to each genotype or haplotype of the SNP to construct a linked string of the codes. For example, when analyzing the SNP of the human genome, There may be ten possible SNP genotypes, and each SNP genotype may be assigned an alphabetic code to construct a linking string for the SNP (see Table 3).

 According to a preferred embodiment of the invention, the variants to be analyzed in the present invention are SNPs, the SNPs are (i) 5% or less (more preferably, 4% or less, more preferably 3% or less, even more preferred Preferably from the group consisting of (ii) Hardy Weinberg Equilibrium test and (Π) folate-effect test. SNPs that are QC (Quality Control) by at least one method selected. This sample QC allows the analysis results of the present invention to be more accurate.

 According to a preferred embodiment of the invention, the method further comprises the step of determining the optimal length of the variant syntax before step (b). Determination of the optimal length of variant syntax (eg, SNP-S) can be made in a variety of ways.

 According to an embodiment of the invention, determination of the optimal length of variant syntax (eg, SNP-S) is carried out empirically determined to a length that exhibits the highest accuracy for the phenotype of the organism. In this case, the phenotype is preferably a disease (eg cancer).

 Alternatively, the determination of the optimal length of the variance syntax may be performed by selecting the optimal length in the convergence section in the tree topology fabricated using the Robinson-Foulds distance (see reference). Gregory E. Sims, et al. (2009) .Whole-genome phylogeny of mammals: Evolutionlay information in gene ic and nongenic regions.PNAS USA 106 (40): 17077-17082).

If the optimal length of the variant syntax (eg SNP-S) is already known or the optimal length has been determined, step (b) is carried out using a sliding window having the determined optimal length. According to a preferred embodiment of the present invention, the variation to be analyzed in the present invention is SNP and the optimum length of the SNP density (density) of 1 million SNPs / genome 6-14 (more preferably 8-12 , Most preferably 10). If the density of the SNP increases, the optimum length also increases. According to a preferred embodiment of the present invention, the FFPs whose length is determined in step (d) are for rare VAR-S (eg SNP-S) of a certain length and are filtered-in. The regression VAR-S of a particular length (eg, SNP-S) is 20% or less (more preferably, 5% or less, even more preferably 3% or less, most likely in a population including a mutation of the analyte). Preferably a rare VAR 'S (eg, SNP-S) filtered-in with a low frequency of 2% or less).

 According to a preferred embodiment of the present invention, the distance between the FFPs in step (d) can be obtained by applying various distance functions, for example, Jensen-Shannon (JS) divergence, Euclidean distance function ， Cosine distance function ， Minkowski distance function and Pearson linear correlation ^ can be obtained, and most preferably, the distance between FFPs with JS (Jensen-Shannon) divergence Get

For example, using JS (Jensen-Shannon) divergence, the distance between two FFPs (P _/ , Q _; ) is calculated according to the following equation:

 [Equation 1]

 Braided pigtails

JSt (Pi, Qi) = 2 1, +-REiQ _t , M _t ) ^ 難 ι = ^ M ¾ 酵

In the above equation, M / is the average FFP of P and Q _/ , and RE is the relative entropy.

Specific embodiments of the present invention will be described with reference to the following examples, in which one individual has the SNP-Ss FFP and the smallest JS (Jensen-Shannon) divergence. If the other subject is a breast cancer patient, the subject may be determined to have high susceivability to breast cancer.

 The pairwise all-against-all distances thus obtained are stored in the distance matrix. In the distance matrix, identical objects have a distance of zero and different dissimilar objects have a large distance.

 The distance relationship between FFPs can be visualized in various ways (eg, nearest-neighbor connection map or systematic tree) (see FIG. 3).

 The method of classifying FFPs may be performed using a class prediction algorithm, for example, as a support vector machine (SVM).

 According to a preferred embodiment of the present invention, the organism to be analyzed in the present invention is an animal, a plant, a fungus, a yeast, a bacterium or a protist. Animals that can be analyzed by the present invention include, but are not limited to, mammals, strata, reptiles, and birds. Preferably, the animals analyzed by the present invention include humans, mice, rats, cattle, pigs, horses, sheep, rabbits, goats, birds, fish, and stratum.

Plants that can be analyzed by the present invention include, but are not limited to, monocotyledonous plants, dicotyledonous plants and algae. Preferably, the plants analyzed by the present invention are food crops, including rice, wheat, barley, corn, soybean potatoes, wheat, palm, oats and sorghum; Vegetable crops including arabidopsis, cabbage, radish, peppers, strawberries, tomatoes, watermelons, cucumbers, cabbages, melons, pumpkins, green onions, onions, and carrots; Special crops including ginseng, tobacco, cotton, sesame, sugar cane, sugar beet, perilla, peanuts and rapeseed; Fruit trees including apples, pears, jujube peaches, lambs, grapes, chisels, persimmons, plums, apricots and bananas; Flowers, including roses, gladiolus, gerberas, carnations, chrysanthemums, lilies, and yulips; And fodder crops, including lygras, red clover, orchardgrass, alphaalpha, pescue and perennial lygragrass. Examples of such bacteria include eubacteria and akibacteria. Examples of milk bacteria that can be analyzed by the present invention include, but are not limited to, Escherichia coli, Thermus thermophi 1 ics, Bacillus subti lis, Bacillus st earo thermoph ilus, Salmonella typhimuriu, Pseudo onas, Streptomyces, Staphylococcus, Lactobacillus, Lactococcus and Streptococcus ^■ It doesn't work. Examples of achibacteria that can be analyzed by the present invention include Methanococcus jannaschi i (Mj), Methanosarcina azei (Mm), Methanobacterium thermoautotroph icu (Mt), Methanococcus maripaludis, Methanopyrus kandleri, Halobacterium, Archaeoglobus fulgidus (rocfh, ArocPus, ArocPh, ), Pyrobaculimi aerophi lum, Pyrococcus abyss /, Sulfolobus solfataricus (Ss), Sulfolobus tokodaii, Aeuropyrum pernix (Ap), Thermoplasina acidophi lum and Thermoplasma volcanium.

 Protists that can be analyzed by the present invention include, but are not limited to, algae, Plasmodium, Phytophthora, slime molds, protozoans It is not.

 According to a preferred embodiment of the invention, the mutation of the analyte is a mutation associated with the traits of the organism and the method of the invention is used to predict sensitivity to the traits of the organism.

According to a preferred embodiment of the present invention, the traits to be analyzed in the present invention are bad traits (adverse traits), diseases, disorders, conditions or symptoms. For example, the disease, disease, condition or symptoms may include cancer, tumor, chronic disease, infectious disease, neurological disease, metabolic disease, immune disease, inflammatory disease, cardiovascular disease, respiratory disease, bone disease, thyroid disease, otolaryngology Ophthalmic diseases, dermatological diseases, dental diseases, endocrine diseases, gastrointestinal diseases, hereditary diseases, musculoskeletal disorders, arthritis, obesity and hyperlipidemia. More preferably, the trait to be analyzed in the present invention is a cancer disease. For example, _susce ptibility to a cancer disease of one subject can be quantitatively predicted by the present invention.

 According to a preferred embodiment of the present invention, the trait to be analyzed in the present invention is an advantegeous trait, which is growth rate, yield or quality.

 According to a preferred embodiment of the present invention, the mutation to be analyzed in the present invention is a mutation associated with the therapeutic responsiveness of the organism and the method is used to predict the therapeutic responsiveness of the organism. A representative example of treatment response is drug responsiveness. Respondents, non-respondents, and reverse respondents with respect to a particular drug can be determined by the present invention.

 According to one embodiment of the present invention, the mutations analyzed by the present invention can be used for multi iclass cancer classification. The most important thing in the treatment of cancer is the accurate diagnosis or information about the cancer of the patient. Multiclass cancer classification is required for such accurate diagnosis (Ramaswamy S, et al. (2001). Multi class cancer diagnosis using tumor gene expression signatures. PNAS USA 98 (26): 15149-54). The present invention can be used for this multiclass cancer classification.

According to another aspect of the invention, the invention provides a computer-readable storage medium embodied with instructions instructing a computer processor to perform the following steps: (a Constructing a linked string of the variants; (b) constructing a variation syntax (VAR-S) of a specific length by applying a sliding window of a specific length along the entire length of the linking string; (c) counting all possible features in the particular length variation syntax and assembling them into feature frequency profiles (FFPs) step; And (d) determining the distance between the FFPs or classifying the FFPs.

 According to another aspect of the invention, the invention provides a system for analyzing genomic variation or epigenomic variation of an organism comprising:

 (a) a computer processor; And

 (b) the computer-readable storage medium of the present invention coupled with the processor.

 The storage medium and system of the present invention are for carrying out the method of the present invention as described above, and the contents in common between the two are omitted in order to avoid excessive complexity of the present specification.

 The storage medium of the present invention is not particularly limited, and various storage media known in the art, for example, CD-R, CD-ROM, DVD, data signals contained in carrier waves, flash memory, floppy disks, hard drives, Magnetic tapes, MINIDISC, nonvolatile memory cards, EEPR0M, optical disks, optical storage media, RAM, ROM, system memory, and web servers.

 The system of the present invention can be built in a variety of ways. For example, the system of the present invention may be built with a multiprocessor computer array, a web server and a multi-user / interactive system.

The system of the present invention may include various elements, for example, to construct a variant (e.g. SNP) information storage database, a processor to create an associative string of variants, to construct a variant syntax (e.g. SNP-S). Processor, processor to determine optimal length of variant syntax (e.g. SNP-S _/ ), FFP generator, processor to perform distance determination between FFPs, processor to create distance matrix and processor to visualize distance matrix Can be built to include The detailed description of the integrated approach of the second inventor of the invention is as follows:

According to another aspect of the invention, the invention provides a computer implemented method for analyzing genomic variation or epigenomic ^' variation of an organism comprising the following steps:

 (a) constructing at least two kinds of descriptors for said mutations;

 (b) applying at least two kinds of class prediction algorithms to each of the at least two kinds of descriptors to analyze the genome variation or epigenetic variation of the organism to obtain at least four kinds of prediction results; And

 (c) finally predicting the trait of the organism by applying at least four kinds of prediction results obtained in step (b) to an inference algorithm.

 The basic strategy of the present invention is to apply a variation of a particular entity by applying at least two kinds of descriptors for the variation to each of at least two kinds of class prediction algorithms, and applying the results from these applications to the appropriate ^ inference algorithm. Assay (eg, susceptibility to specific traits).

 Genome mutations to be analyzed in the present invention include various variations found in organisms, preferably SNPCsingle nucleotide polymorphisms in nucleotide sequences), deletions, insertions or repeat variations; Or epigenomic variation.

In the present invention, the mutation to be analyzed is a mutation present in a nucleotide sequence, and the nucleotide sequence is a sequence on one chromosome, a sequence on a plurality of chromosomes, or a whole genome sequence, more preferably a whole genome (WG). )to be. Most preferably, the mutations to be analyzed in the present invention are SNPs in the entire genome sequence.

 According to the invention, at least two kinds of descriptors for the mutation are constructed. Preferably, the at least two kinds of descriptors for the variation are: (i) the profile of the variations (eg, the profile of ordered SNPs) assuming that each variation is independent of the neighbor's variation and (ii) a particular length Profiles of the above-described variant syntax (VAR-S) (eg SNP syntax) that are associated variants. The use of syntax (VAR-S) (eg SNP syntax) as one of two descriptors is due to the fact that each variation (eg SNP) location is not independent and is connected to neighbors to varying degrees.

 According to a preferred embodiment of the invention, step (a) is carried out by constructing a string of codes by assigning a code to each of the variants.

 More preferably, step (a) is performed by assigning a code to each genotype of the SNP to construct a string of the codes.

 According to a preferred embodiment of the invention, the analytes are SNPs, wherein the SNPs are (i) 5% or less (more preferably 4% or less, even more preferably 3% or less, even more preferably Is a minimum selected from the group consisting of removal of SNPs exhibiting an allele frequency of 2% or less, most preferably 1¾ or less), (ii) Hardy Weinberg Equilibrium test and (ii) Plate-effect test SNPs QCX Quality controlled by one method. As such, sample QC allows the analytical results of the present invention to be more accurate.

Since the specific analysis method using the variance syntax is the same as described above, the storage medium and the system of the present invention are for carrying out the method of the present invention described above, and the common content between the two is to avoid excessive complexity of the present specification. For that reason, the description is omitted. The class prediction algorithm applied in step (b) of the present invention includes various algorithms known in the art, for example, the ^ r-nearest neighbor O-nearest neighbor (N) algorithm (Bremner et al. al., (2005). “Output—sensit ive algorithms for computing nearest—neighbor decision boundaries.” Discrete and Computational Geometry 33 (4): 593-604), support vector machine (SVM) algorithms (Theodor idis). S & Koutroumbas K (2009) Pattern recognition, compound covar iate predictor (Emura T, et al., (2012) Survival Predict ion Based on Compound Covar iate under Cox Proportional Hazard Models. PLoS ONE 7 ( 10): e47627), Linear Discriminant Analysis (LDA) (McLachlan, GJ (2004) .Discriminant Analysis and Statistical Pattern Recognition.Wieley Interscience) , GJ (2004) Discriminant Analysis and Statistical Pattern Recognition.Wieley Interscience.

According to a preferred embodiment of the present invention, the class prediction algorithm applied in step (b) is an r-nearest neighbor (ANN) algorithm _. And support vector machine (SVM) algorithms.

 The nearest neighbor analysis algorithm searches the k nearest neighbors of the test subject. In the analysis algorithm, all pairwise "distances" are calculated between the descriptor of one entity and the descriptor of each entity. ANNs are then selected for the test subject and predicted whether the subject is sensitive to the most common traits among the A Ns.

The support vector machine (SVM) algorithm is a fractional classification method that identifies the most likely class to which the test subject belongs. In the SVM analysis algorithm, the SVM is trained to correct the correct trait of one individual in each of all binary traits. Be aware. Finally, having the maximum selection of all pair classifications by SVM predicts the susceptibility of the test subject to the most likely trait.

 According to one embodiment of the invention, the descriptor for the variation is a VAR-S and a profile, the class prediction algorithm is the nearest neighbor algorithm, and step (b) is less than 20% in (b-1) population A small step of selecting rare VAR-S found at low frequency of; (b-2) sub-steps to normalize to the total number of regression VAR-S; (b-3) constructing a Jensen-Shannon (JS) divergence matrix using the profile of the rare VAR-S; And (b-4) selecting a k-nearest neighbor for the organism using the JS divergence matrix. This embodiment is abbreviated as KNN / VAR-S (KNN / SNP-S, when applied to SNP).

The KNN / SNP-S is described in more detail as follows: A vector of SNP-Ss for all members of the training set is obtained, followed by a feature selection step. At this stage, the syntax shared by any percentage of population is removed (filtered out) and the remainder (filtered in) is used for analysis. It is then normalized to the total number of rare SNP syntax of the subject. Finally, we use the rare SNP syntax to build the JS divergence matrix between all members. The reason why JS divergence was chosen to measure the distance of the descriptor is that it is more predictable than other conventional methods such as allele sharing. Measure the paired JS distance for every entity, each entity, then select or vote classes among the top nearest entities and select the one with the highest count. For the best accuracy of trait (eg cancer susceptibility) estimation for training datasets, optimize SNP-S for optimal length, 1, f parameter for low frequency selection, and parameter k. In the testing phase, the same / and / optimal parameters are used. The JS distance vector between the subject and the training sample is then measured and the test subjects are predicted through the same selection process in the training phase with the optimal parameters. According to one embodiment of the invention, the descriptor for the variation is a profile of the variation, the class prediction algorithm is a support vector machine (SVM) algorithm, and step (b) is (b-1) 10 ^'2 to 1 Substep selecting mutations with a low P-value of ⁶ ; (b-2) substeps of performing SVM on each of all binary traits; And (iii) substeps classified according to a max-win voting scheme. This embodiment is abbreviated SVM / VAR (SVM / SNP, if applied to SNP).

A more detailed description of SVM / SNP is as follows: SVM can be implemented in a variety of ways, for example in the One-Versus—One (OVO) method. The 0V0 method produces an n (n−l) / 2 classifier for each pair of two classes and takes the class with the highest election from n (nl) / 2 predictions for the test sample. To implement the 0V0 SVM method, LIBSVM from Chang et al is used (Chang CC & Lin CJ (2011) LIBSVM: A Library for Support Vector Machines. Acm T Intel Syst Tec 2 (3)). Choose the adial basis function (RBF) for the kernel function. To build a binary classifier of SNP descriptors, the SNPs are filtered out for a given P-value threshold (p) to select the associated SNPs between the two classes. It is recommended that cutoffs less than 1 (Γ ⁶⁾ should not be applied, since some classifiers do not leave SNPs after filtering by the association test. Encoding each genotype is done in the case of ambiguous predictions (ie multiple best elections). Repeat the poles in the set of highest elected classes until the tie breaks, and train the SVM to recognize the correct trait of one individual in each of the two binary traits. The most likely trait of the test subject is predicted to have the maximum election of all pair classifications by SVM.

According to one embodiment of the present invention, The description of the variation and the profile of the variation, the class prediction algorithm ^ closest neighbor ^algorithm, said, step (b) is 20% of the (b-1) papul illustration. Found at a lower frequency of Small steps of screening for rare mutations; (b-2) substep normalizing to the total number of rare mutations; (b_3) constructing a JS divergence matrix using the profile of the rare variant; And (b-4) selecting a k-nearest neighbor NN) for the organism using the JS divergence matrix. This embodiment is abbreviated as KNN / VAR (N / SNP, if applied to SNP). Implement KNN / SNP by replacing the SNP-S descriptor with SNP and following the same procedure as KNN / SNP-S.

According to one embodiment of the present invention, The description of the variation and the VAR-S profile, the class prediction algorithm is a support vector machine (SVM) algorithm, wherein step (b) is a (bl) 10- ² to 10 — Small step of selecting VAR-S with low P-value of ⁶ ; (b-2) substeps of performing SVM on each of all binary traits; And (iii) substeps classified according to a max-win voting scheme. This embodiment is abbreviated as SVM / VAR— S (SVM / SNP-S, if applied to SNP). Instead of SNP, SVM / SNP is performed using SNP-S and SVM / VAR-S is performed. For SVM / VAR-S, additional parameters for the optimal length of SNP-S are used.

 According to the present invention, at least four kinds of prediction results obtained in step (b) are applied to an inference algorithm to finally predict traits of an organism or an individual whose trait is not determined.

 According to one embodiment of the invention, the inference algorithm used in step (c) comprises a Bayesian inference algorithm and a voting scheme, most preferably a Bayesian inference algorithm. to be.

분모 sil^ ^ fi ^)^ 정규화 상수이기 때문에， 분모는 빠지게 된다. 각 방법의 예측 결정은 서로 내재적으로 독립적이기 때문에, 체인규칙을 적용한다 (Zhang H (2005) Exploring conditions for the optimality of Naive bayes . Int J Pattern Recogn 19(2)： 183-198)： Specific examples of the steps applied to the Bayesian inference algorithm are as follows: In order to classify individuals into one of nine phenotypes (eight cancer classes and health traits), each phenotype is represented by the total initials of each trait. Label it with the first character. Bayesian inference of the prediction results of the four methods is used. These methods have the following shorthand: KNN / SNP-S, KNN / SNP, SVM / SNP-S, SVM / SNP. The methods are mathematically represented by nP and m ⁴ , respectively. For each test subject i, the highest post-probability trait conditioned on the predictions obtained from the training method is selected, which can be formulated as PCsi ^ ii fiJ ^) ^. Where s _/ is the predicted trait of the individual / and how? Represent the individual / and traits predicted by. By Bayesian theorem, Bayes theorem can be expressed as:

Denominator sil ^ ^ fi ^) ^ Since this is a normalization constant, the denominator is omitted. Since the prediction decisions of each method are inherently independent of each other, we apply the chain rule (Zhang H (2005) Exploring conditions for the optimality of Naive bayes. Int J Pattern Recogn 19 (2): 183-198):

t 瞧 = argmaxteT J ^ i P (M ^ _f / s / = t) ^ Pfs, = i) where P ^' ilsft) and P ( _Si = t) are obtained from the observations during each of the four methods. It can be inferred empirically by maximal likelihood estimation. For example, among the total BRCA samples in the training set, P (정 C / s 尸 B) can be estimated by identifying some of the true BRAC individuals estimated to be C0AD by the ^ N / SNP-S method. P ( _Si = t) corresponds to a portion of a sample of the traits of all trained individuals, which is the same for each of the 9 traits (since they use the same sample size for each trait). Through this inference process, it is determined (classified) what kind of trait the individual has not been determined.

According to one embodiment of the present invention, transfection of the organism is a disease (diseases), disease (disorders), conditions (conditions), symptoms (symptoms) or the value ^"fee (therapy) reactivity (responsiveness). More preferably, the trait to be analyzed in the present invention is a cancer disease. For example, susceptibility to cancer disease in one subject can be quantitatively predicted by the present invention.

 According to a preferred embodiment of the present invention, the trait to be analyzed in the present invention is an advantegeous trait, which is growth rate, yield or quality.

 According to a preferred embodiment of the present invention, the variant to be analyzed in the present invention is a variation associated with the therapeutic responsiveness of the organism and the method is used to predict the therapeutic response of the organism. A representative example of therapeutic responsiveness is drug responsiveness. Respondents, non-respondents, and reverse respondents for a particular drug can be determined by the present invention. According to another aspect of the invention, the invention provides a computer-readable storage medium embodied with instructions instructing a computer processor to perform the following steps: (a Constructing at least two kinds of descriptors for the mutations; (b) applying at least two kinds of class prediction algorithms to each of the at least two kinds of descriptors to analyze the genome variation or epigenetic variation of the organism to obtain at least four kinds of prediction results; And (c) obtaining in step (b) the final prediction of the trait of the organism by applying at least four kinds of prediction results to an inference algorithm.

According to another aspect of the invention, the invention provides a system for analyzing genomic variation or epigenomic variation of an organism comprising: (a) a computer processor; And (b) the computer-readable storage medium coupled with the processor.

【effect]

 The features and advantages of the present invention are summarized as follows:

 (a) The present invention is similar to comparing two texts with words of natural language, and through this method provides a systematic feature frequency profile (FFP) for various variations (eg SNPs) found in an individual.

 (b) The present invention also determines the distance between FFPs to accurately predict susceptibility to certain traits of an individual.

 (c) According to the present invention, cancer sensitivity of an individual can be predicted with an accuracy of 47 to 76% even when the sample size is small. However, this accuracy can be increased by increasing the size of the SNP genotype data, and can be further increased by classifying in advance.

 (d) The prediction accuracy of the second invention represents several times increased accuracy compared to the random prediction, and the degree of such prediction is highly improved prediction accuracy as it is possible to determine the health state of the individual or the population.

[Brief Description of Drawings]

 1 is a diagram of the process of a method for assessing the sensitivity of eight cancer types. This method includes preprocessing of SNP data, such as sample conditioning screening and genotyping coding, filtering of common SNP syntax and profiling of SNP syntax frequencies (FFPs of SNP-Ss), and the nearest neighbors of the smallest branches ("distances"). There are several processes, such as calculating the distance between paired FFPs to identify the sister.

2 is a graph showing the accuracy assessment of cancer sensitivity to length (/) and percentage filtering—phosphorus of SNP-S. Increase the performance of risk assessment for multiclass cancers while increasing length (/) and reducing percentage filtering-in Measured. For example, 2% filtering means maintaining SNP-S / s generated at less than 2% of population. This process retains only "regressive" SNP-S _/ s present in populations below 2%. The gray line represents the baseline accuracy and "No syntax" means the accuracy assessment by comparison of the entire SNPs as a non-associated feature, ie not using the FFP of the SNP-Ss. For the control curve (dark red) to compare with the performance of the SNP-S method, individual markers were profiled, in which case each had 10 characteristics of coding SNPs and counted 1 for homozygosity. The count is 0.5 for heterojunctions and 0 for other cases.With the profiled data, the nearest neighbors are identified at Jensen-Shanan distance by applying the same procedures of SNP syntax method from common feature filtering.

3 is a nearest-neighbor connection map. The nearest neighbors of 594 individuals (66 in each of the eight cancers and 66 controls) were identified. Each individual is represented by a rare SNP-S ₁₀ s FFP (2% filtered-in), and the nearest neighbor of one individual is defined as another individual with FFP with the smallest Jensen-Sha 皿 on divergence (distance) from the first. do. Types of arms are listed on the outside of the outer circle, indicated by different colors in the inner circle, and the interior of each curve of the circle connects two entities with nearest-neighbor ("sister") correlations. The color of the curve is the same as the cancer type of the nearest-neighbor pound by the search member (also the color of the small sal in the outer circle). The nearest neighbor pound may or may not be the same cancer type ("true" sister) or not ("error" sister). If the search and found objects are interchangeable, the curve is represented by a thick line. Of these, the curves for all error sister correlations are shown in dark gray. The color scheme is as follows: CEU, red; BRCA, orange; C0AD, bright orange; HNSC, yellow; KIRC, green; LGG, light blue; 0 V, blue; READ, dark blue; UCEC, purple. This map was created using circos. 4 is a genome mapping of a sensitive marker allele on chromosome 3. The density of the sensitive marker allele is indicated by heat-maps on colored circle tracks for each cancer type (from inside to outside, CEU, red; BRCA, orange; C0AD, light orange; HNSC, yellow; KIRC , Green; LGG, light blue; 0V, blue; READ, dark blue; UCEC, purple), high density areas are indicated in dark colors. The outermost track shows the cytoband of chromosome 3, and the labeled light blue tick marker indicates the location of the known cancer gene. Cytoband tracks The blue short arches represent individual cytobands with one or more GS hits of known arms in Caucasus population. The green short bar on the next inner track shows the genetic code site and the next inner circle shows the density of the published SNPs. This map was created using circos.

Figure 5 maps the sensitive marker alleles near the locus of two known cancer genes (BRCA2 and TP53). Each marker allele is represented by a cluster of circles (SNPs making up a specific SNP-S ₁₀ allele) in the color of the cancer type. The X-axis represents the physical location of chromosome 17 or 13 where TP53 and BRCA2 are found, respectively, and the Y-axis divides different marker alleles for different cancer types: CEU, red; BRCA, orange; C0AD, bright orange; HNSC, yellow; KIRC, green; LGG, light blue; 0 V, blue; READ, dark blue; UCEC, purple. Sensitive marker alleles do not overlap with TP53 or BRCA2 (each represented by a dashed vertical line). Recombination ratio is indicated by blue spikes and the other genes around the two genes are indicated in the lower box of each figure. This picture was produced using Lo isZoom i / J.

6 shows QC (Quality Control) results. The graph shows the overall accuracy of different datasets from different QC criteria as a function of filtering threshold (left: dataset with HapMap control, right: dataset without HapMap control). Two filters namely HWE and plate effect ΊΉΜ1 (dataset used in this study), HapMap data with TCGA and MAF>0.05; THM5, same as THM1 except MAF>0.05; THMO, same as THM1 except without MAF filtering; Same as THM0 except without THMOR, HE and plate effect tests; TCGA data of TM5, two filters, MAF> 0.05 with HWE and plate effect; Same as TM5 except TMl, MAF>0.01; Same as TM1 except without TMO, MAF filter; Same as TM0 except TM0R, two filters ie E and no plate effect.

 7 shows accuracy vs. Show sample size and number of features. The accuracy of the (left) sensitivity assessment increases as the characteristic sample size increases. BRCA, 0V, and UCEC were selected for analysis and excluded other characteristics because of the limited sample size. The number of (right) characteristics is 3, 6, and 9 (BRCA, C0AD, and CEU have three characteristics; BRCA, COAD, HNSC, KIRC, 0V, and CEU have six characteristics; and BRCA, C0AD, HNSC, HNSC, KIRC, Increasing 0V, REDA, UCEC, and CEU to 9 characteristics decreases accuracy. Each feature dataset size was fixed at 66 individuals.

8 specifies a region of the genome covered by sister-specific features. The number of known cancer genes near the sensitivity marker SNP-S ₁₀ s for each cancer is counted while increasing the distance threshold between the cancer gene and SNP-S ₁₀ s. For each cancer of BRCA and C0AD, known cancer genes were downloaded from 0MIM. The total number of known cancer genes at 0MIM for each cancer is indicated by dotted lines of brown (BRCA) and green (C0AD). OMIM, Online Mendelian Inheritance in Man.

9 is a schematic diagram of the method of the present invention for assessing genome sensitivity to eight major cancer and health traits. The method of the present invention comprises SNP data preprocessing, including sample control screening and genotype encoding, selection of low ^ value SNPs and low frequency SNP syntax, application of two different analysis algorithms and the results of the four methods. It includes the final prediction step of integrating. 10A shows the optimization of the parameters used in the process of applying the k—nearest neighbor algorithm to the profile of the SNP-syntax. 10B shows the optimization of parameters used in applying the k-nearest neighbor algorithm to the profile of SNPs.

 Figure 10c shows the optimization of the parameters used in the process of applying the SVM algorithm to the profile of the SNPs.

 Figure 10d shows the optimization of the parameters used in the process of applying the SVM algorithm to the profile of the SNP-Ss.

 11A-11C show the 9-class prediction results of the test set for each of three cancer classes, BRCA (FIG. 11A), 0V (FIG. Lib) and UCEC (FIG. 11C). Prediction results of four methods and Bayesian inference on 50 test subjects for each of the three cancer classes are shown. The dotted horizontal line represents the random prediction, and the tick marks on each bar represent the standard error for the prediction result measured by resampling 50 test subjects 10 times.

 12 shows the confidence in the prediction of an individual. Average accuracy vs. Posterior probability threshold.

[Specific contents to carry out invention]

Hereinafter, the present invention will be described in more detail with reference to Examples. These examples are only for illustrating the present invention more specifically, it will be apparent to those skilled in the art that the scope of the present invention is not limited by these examples in accordance with the gist of the present invention. . Example

I. First Invention: Method Research Materials and Methods Using SNP-Syntax

Database and sampling

A total of 2192 SNP array results and relevant clinical information were approved by the NCBI dbGap for general study use, and from April 2, 2012 to April 4, 2012, NIH (National Downloaded from the Institute of Health's TCGA. The SNP genotype data for these patients' blood tested by the Broad Institute were downloaded. The patients were mostly white (some outliers from other ancestors were removed during the sample conditioning phase). All markers were run on Affymetrix 6.0 SNP chips ^. Typed. As a control for their cancer, the HapMap Project's CEl Caucasians from Utah, USA) population data was used because it is considered to be the most representative data of all white individuals to date. To reduce the loss of SNP information in the process of integrating two datasets with different marker sets, the C165's 165 SNP array results (typed as Affymetrix 6.0 SNP) were downloaded from the HapMap ftp website. The data was genotyped using Affymetrix Power Tools with default parameter settings and discarded samples reported to have low sample quality from the website (see Table 1).

 TABLE 1

 Sample Quality Control of Affymetrix 6.0 SNP Genotype Data (as of April 4, 2012)

 Sample QC

 Before QC After QC

 Study characteristics / Koho E

 Male / female "Sr ᄇ male / female total

B CA 6/694 700 5/511 516

COAD 179/159 338 101/86 187

HNSC 106/38 144 95/34 129

KI C 47/30 77 43/25 68

TCGA LGG 38/36 74 34/32 66

OV 0/427 427 .22 / 379 401

READ 72/59 131 54/41 95

UCEC 0/301 301 0/237 237

HapMap CEU 80/85 165 31/38 69 Total 528/1829 2357 385/1383 1768 The Cancer Genome Atlas project initiated by the National Institute of Health (NIH); Breast Invasive Carcinoma (BRCA); Colon Adenocarcinoma (COAD); Head and Neck Squamous Cell Carcinoma (HNSC); Idney Renal Clear Cell Carcinoma (KIRC); Brain Lower grade glioma; Ovarian Serous Cystadenocarcinoma (OV); Rectum adenocarcinoma (EAD); Uterine Corpus Endometrioid Carcinoma (UCEC); Haptype Map Project (HapMap); Caucasians from Utah, USA; European-American (EA); PI_HAT, P NK parameter determined when two entities are related; Sample quality control + genetic association test (PIJHAT <0.2) + removal of self-released wheat for EA individuals. Sample quality control

To study how quality control (QC) affects the overall test, different QCs from lenient to stringent were applied and tested (Tables 1 and 2, FIG. 1). ). PLINK (S. Purcell, et al. (2007). PLINK: A Tool Set for Whole-Genome Association and Population-Based Linkage Analyses. Aw J Hum Genet 81 (3): 559-575) Initial results were obtained from the dataset under conditions of; SNPs with a minor allele frequency of less than 1% were removed and the Hardy Weinberg Equilibrium test was applied to each marker in CEU individuals> 1 × 10 1 ⁶ ). Association test between one plate and other plates (with more than 30 samples) for all plates within cancer characteristics .> To the 1 X 10- ⁸⁾ embodiment, the plate-effective test was applied (See: DG Clayton, et al (2005) Population structure, differential bias and genomic control in a large-scale, case-control association study Nat Genet 37 (11): 1243-1246). For SNPs that passed QC, S. Ripke, et al. (2011) Genome-wide association study identifies five new schizophrenia loci. Nat Genet 43 (10): 969-976) (PI—HAT <0.2). In addition, Pedigree information was used to remove all relevant entities from the CEU. For example, you removed two objects from the trio and one from the duo. Finally, we integrate the sample and add it to the post-QC characteristic dataset. SNPs were jointed to obtain genotypes for 714649 non-repetitive autologous SNPs for 1768 individuals (see Tables 1 and 2).

 Table 2

SNP QC

QC Standard QC Post Data Set

 Study HWE Platen Effect

 Name MAF Common SNPs

 (> 1E-07) (> 1E-08)

 TM0R TCGA X X X 867266

TM0 TCGA X X 0 845 025

TM1 TCGA 0.01 X 0 732 705

TM5 TCGA 0.05 X 0 625 702

 TCGA X X X

THM0 867087

 HapMap X 0 X

 TCGA X X 0

 THM0 844889

 HapMap X 0 X

 TCGA 0.01 X 0

 THM1 714649

 HapMap 0.01 0 X

 TCGA 0.05 X 0

 THM5 616722

 HapMap 0.05 0 X

 Single Nucleotide Polymorphism (SNP); Quality Control (QC); Minor Allele Frequency (MAF); Hardy-Weinberg

Equilibrium); TCGAf The Cancer Genome Atlas project initiated by the National Institute of Health (NIH)); Haptype Map project of worldwide human populations (HapMap); Caucasians from Utah, USA; THMl (dataset used in this study), TCGA and HapMap data of MAF> 0.01 with two filters of HWE and plate effect; THM5, same as THMl except MAF>0.05; THM0, same as THM1 except without MAF filtering; Same as THM0 except no THM0R, HWE and plate effect tests; TCGA data with two filters, TM5, HWE and Plate effect, with MAF>0.05; Same as TM5 except TM1, MAF>0.01; TM0, same as TM1 except without MAF filtering; Same as TM0 except there are no two filters, TM0R, HWE and plate effect; X indicates that the dataset did not have an associated QC, and 0 is the opposite. SNP code conversion

 In order to avoid artificial errors that could arise from computer paging methods using a small sample size, only experimentally obtained genotype information was used, not computer-generated haplotype information. Each of the ten possible SNP genotypes was converted to one of the ten alphabets listed in Table 3. TABLE 3

Characteristic frequency profiling of SNP syntax (SNP-S)

From the converted genome-wide SNP genotype data, a vector of feature counts was constructed and the features were all possible continuous (associated) SNPs, SNP syntax (SNP-Ss) of fixed length in the subject of study. The vector, feature frequency profile (FFP) for an individual represents the systematic characteristics of the individual's WG SNPs, which slide a fixed length window along the entire length of the individual genome's SNP strings and all possible features (SNP-Ss in this case) are constructed (see GE Sims, et al. (2009). Alignment-free genome comparison with feature frequency profiles (FFP) and optimal resolutions. Proc Natl Acad Sci USA 106 (8): 2677-2682). The optimal feature length for profiling is determined to be the length that shows the highest accuracy in calculating cancer sensitivity. In SNP-Ss of one individual, the optimal length was 10 (FIG. 2). Since each SNP has genotype information without its exact chromosomal allele (Haplotype) order, in the SNP syntax. The number of heterozygotes determines the possibility of the presence of haplotype information in the polymorphic context. Thus, each haplotype occurs in the same syntax Under the premise of having a likelihood, the "count" of the occurrence of SNP-S is inversely proportional to the number of possible haplotypes represented by SNP-S. Although the study dataset does not include a missing genotype, when this case exists, it includes all features arising from the combination of possible genotypes in the untyped marker and their counts are extended by the missing. By dividing by the total number of SNP-Ss, it can be processed easily. The following equation represents the count in all cases:

Equation 1

¾-r ^¾ * ¾¾ 零^' mi

In the above equation, _x is the count of SNP-S (count in fraction), i is the number of heterozygotes in SNP-S, and j is the number of missing markers in the SNP-S sequence (see Table ⁴ ).

 Table 4

백분율 필터링 -인， 정규화 및 Jensen-Shannon 발산 매트릭스 모든 멤버들의 SNP-S₁₀ 카운트의 FFPs를 얻은 다음， 이 파풀레이션의 몇 퍼센트에 의해 공유되는 신택스를 제거 (필터-아웃) 하고 나머지를 분석을 위하여 보관한다. 이어， 각각의 잔여 희귀 (rare) 신택스의 카운트를 위하여 희귀 SNP 신택스의 총 개수에 의해 정규화 한다. 최종적으로, 백분율 필터링 -인에 의해 생성된 희귀 SNP 신택스를 이용하여 모든 멤버들 사이의 Jensen-Shannon(JS) 발산 매트릭스를 구축하였다. 최근접. 이웃 ( "시스터" )의 동정 및 민감성의 정확도

Percent Filtering-In, Normalization, and Jensen-Shannon Dispersion Matrix Obtain FFPs of SNP-S ₁₀ counts for all members, then remove (filter out) the syntax shared by a few percent of this population and analyze the remainder. keep it. Then, it is normalized by the total number of rare SNP syntax for counting each residual rare syntax. Finally, a rare SNP syntax generated by percentage filtering-in was used to construct a Jensen-Shannon (JS) divergence matrix between all members. Nearest. Accuracy of identification and sensitivity of neighbors ("sisters")

 After pairwise JS distances are measured for all individuals, the "sister" of each member with the shortest JS distance is identified, and it is checked whether the sister pairs correspond to the same arm type. Count the number of correct assignments (same cancer type) and divide it by the number of all members, measure the overall accuracy, and similarly calculate the cancer type-specific accuracy, and the number exactly assigned to one specific cancer type Measured by dividing by the total number of members in the category. Sister-specific marker alleles and cancer-specific marker alleles

For true sister pairs in one particular cancer group, all common rare SNP-SlOs were selected between the two sister members of the sister pair without being found in other cancer type members. The alleles thus selected were named as sister-specific marker alleles for the sister pairs. Similarly, the cancer-specific marker allele was named the sum of the non-repeating sister-specific marker alleles of all true sisters in the cancer type. Thus, each sister-specific marker Alleles can be considered "classified" from cancer-specific marker alleles (multiple classification models)

 The results of this study can be divided into five sections: Section I presents two ideas of invention: SNP-S and FFP, and the overall systematic features of one individual's genome WG SNPs are assigned to FFP of SNP—Ss. How it can be represented; Section II shows the process of empirically identifying the optimal length of SNP-Ss and the optimal filtering level to reveal "rare" SNP-Ss to best practice the method of the present invention; Section III details the sensitivity predictions; Section IV summarizes the verification results for the approach of the present invention; And section V shows the genetic location of susceptible SNP—S alleles for known cancer genes, recently identified cancer-related SNPs and other genetic characteristics from G Ss.

I. Frequency profile of SNP syntax as representative of systematic features of WG SNPs

 The method of the present invention for comparing the systematic features of any two individuals' WG SNPs comprises four steps:

(1) Linked WG SNP Strings: The present invention starts with the most general description of the systematic features of an individual's WG SNPs, which is similar to the description in the natural language booklet (CD Manning & H. Schuetze (1999). Language Processing.The MIT Press, I edn), a very important difference is that the associated WG SNPs are treated as natural language text without spaces between words. Thus, an individual's WG SNPs are represented by a single linked string of SNPs arranged in the genome of each individual, with each SNP genotype being one of ten alphabetic codes representing the ten possible genotypes of the SNP under the assumption of genotype two alleles. (See Table 3). (2) SNP syntax: SNP syntax (SNP-S) is defined as a short ordered string of SNPs of a given specific length, which plays a role similar to a "word" of a certain length in natural language text. All possible SNP-S of a given body length (/) for one genome is obtained by sliding a window of length 1 along the total length of the SNP string of the genome. Thus, SNP-S identifies not only the systematic features of SNPs caused by various genetic mutations, but also those known to exist in WG SNPs such as associative imbalances. Halotype map of the human genome.Nature 437 (7063): 1299-1320; The International HapMap Consortium (2007) .A second generation human haplotype map of over 3.1 million SNPs.Nature 449 (7164): 851-861). Thus, in the manner of describing WG SNPs, each SNP-S can be represented by a genetic allele.

Because the number of SNP-Ss of all possible lengths is very large (about 10 ¹² for one SNP string of length 10 ⁶ ) and the mathematical calculations required for comparison of these sizes are difficult, Only SNP-S of length was used. The use of optimal lengths to greatly reduce the computer burden has been described in our previous paper (GE Sims, et al. (2009) .Al ignment® free genome comparison with feature frequency profiles (FFP) and optimal resolutions.Proc Natl Acad Sci USA 106 (8): 2677-2682). Most SNP-Ss with a length of six or more are unique in one genome and occur rarely in the population of genomes.

(3) SFP-Ss (FFP, GE Sims, et al. (2009) .Al ignment—free genome comparison with feature frequency profiles (FFP) and optimal resolutions.Proc Natl Acad Sci USA 106 (8): 2677-2682): represented by an FFP vector consisting of the frequency of all possible features of all aspects of the systematic features of SNPs in one genome of an individual, wherein the features are SNP- S. This approach, which represents the systematic aspect of the string of "alphabet", is similar to that of the WG sequence of organisms performed in the alignment-depletion comparison of the WG (or entire proteum) sequences of different organisms by the FFP method (GE Sims , et al. (2009) Alignment—free genome comparison with feature frequency profiles (FFP) and optimal ^' resolutions.Proc Natl Acad Sci USA 106 (8): 2677-2682; SR. Jun, et al. (2010). Whole-proteome phylogeny of prokaryotes by feature frequency profiles: An alignment-free method with optimal feature resolution.Proc Natl Acad Sci USA 107 (1): 133-138).

 (4) Jensen-Shannon Emissions: Alignment—As in the lack comparison, the “distance” between any two FFPs is represented by Jensen-Shannon Emissions (J. Lin (1991). Divergence measures based on the Shannon entropy.IEEE Trans Inf Theory 37 (1): 145-151) creates a boundary with a threshold between 0 and 1. II. Optimal Parameters of "Regression" SNP-Ss for Predicting Cancer Susceptibility

 FFP of SNP-Ss is suitable for showing the overall systematic characteristics of WG SNPs. However, as mentioned above, FFP of "rare" SNP-Ss is more suitable for disease-specific systematic features. Therefore, the following two assumptions were used as a useful criterion in empirical search for optimal length regression SNP-Ss for cancer sensitivity studies:

 1. Each SNP-S (sensitive allele) associated with cancer sensitivity appears under regression in the population of such cancer types; And

2. The number of sets of genome-sensitive alleles for one particular cancer type may be large, but we treat this as limiting. Thus, given the definite sample size for each cancer type, for the FFP of a rare SNP-Ss of one cancer individual, it is very similar to the population of the same cancer type than that of the other cancer type or the control group. sister" ) There will be one or more other individuals with FFP (s) (this assumption is proven correct in the study described below).

 For simplicity, our empirical approach to finding the "rare" SNP—Ss of optimal length consists of the following steps (see Figure 1):

 1. Build an assembled dataset from the same number of WG SNPs (derived from the blood sample of the individual) of each of the eight types of cancer and non-cancer individuals (control) from the TCGA database and the HpaMap database (see reference). Table 1); Sample quality control is performed to minimize the human error that may occur in each database and the human error that occurs when integrating different data sets (JT Leek, et al. (2010). Tackling the widespread and critical impact of batch effects in high-throughput data; Nat Rev Genet 11 (10): 733—739).

 2. Construct an FFP vector consisting of the frequencies of all possible SNP-Ss for each individual for a particular length of SNP-S syntax, and repeat this process for different lengths of syntax.

 3. Remove (filter out) all common SNP-Ss (and their frequencies) that are equal to or greater than a certain number of populations and repeat this process for the% filter-in level range.

 4. Find all "sisters" (member's sister is defined as another member with the smallest Jensen-Shannon divergence between them). If two members of a sister pair are of the same arm type, they form a "true" sister pair, and if they are of a different arm type, they form a "false" sister pair.

 5. Calculate accuracy as ¾ of all true sister pairs for all arm types of all sister pairs.

6. Accuracy (%) vs. Graph the SNP-S length and% filtering-in to find the optimal "rare: empirically find the optimal length for SNP-Ss and determine the% filtering-in level. In assessing the sensitivity of 8 cancers The rare SNP-Ss filter-in level, which can be optimally performed, was determined at the 2% level and the syntax length 10 SNP-S ₁₀ was determined empirically (FIG. 2). This rare SNP-S ₁₀ s was 2% or less of all samples. However, the filtering removed only 14% or less of SNP-Sios for FFP profiling for nine groups of 66 members each, leaving about 44 million SNP-S ₁₀ s alleles.

 7. Perform a “validation” test using SNPs of individuals not selected from the dataset used to find the optimal feature length of 10 and 2% filtering-in.

III. Analysis using optimal rare SNP-Ss

 Table 5 summarizes the prediction accuracy for the genetic sensitivity determined by the method of the present invention.

 Table 5

 Prediction characteristics

 ^ ¾ ᅳ accuracy

BRCA COAD HNSC IRC LGG OV READ UCEC CEU

 size (%)

53.0%

B CA 35 4 8 2 4 2 2 7 2 66

 (55.6%)

COAD 3 38 4 11 2 2 2 4 0 66 57.6%

HNSC 7 2 47 3 2 0 3 2 0 66 71.2%

KI C 1 7 3 38 1 3 5 6 2 66 57.6% Actual LGG 4 4 4 3 40 4 1 6 0 66 60.6% Characteristics

 47.0%

OV 6 5 3 3 4 31 3 6 5 66

 (43%)

READ 6 1 2 3 3 0 50 1 0 66 75.8%

 53.0%

UCEC 3 3 3 4 7 4 6 35 1 66

 (66.5%)

CEU 0 0 0 0 0 0 0 0 66 66 100%

Very good results were obtained with the regression SNP-S _/ at filtering-in and / = 10.

IV. Validation of the Method of the Invention

In the method of the present invention, a dataset of 66 samples each of eight cancer types and controls was used to optimize two parameters, the length and percent filtering of SNP-Ss— (2%). Because of the small SNP data available in public databases, we chose three cancers (BRCA, 0V and UCEC), which are slightly more data. In each of the three cancers, 66 new samples were randomly selected that were not included in the dataset used in the optimization process, and two parameters were used to calculate sensitivity accuracy for the cancer type. This process was repeated 10 times for each of the three cancers and the average of the accuracy was calculated. The average accuracy for this verification test was 55.6% (43.93-71.21%), 43.0% (34.84-48.48%) and 66.5% (57.5 78.78%) for BRCA, 0V and UCEC, respectively. These values are quite similar to the values in the first sample. ^'

V. Identification of the Genome Location of Susceptible Alleles

내용을 보여준다. To localize the genome region covered by the SNP-S ₁₀ susceptibility allele for one cancer type, it appears only in membership of that cancer type and is common among one or more truth-sister pairs but in other cancer types An undetected SNP-S ₁₀ (referred to as "sensitive SNP-S marker allele" or hereinafter "sensitive marker allele") was identified. These were then analyzed at three levels: (1) overall observation of all marker alleles in the entire genome, (2) intermediate level observation of the position of the marker allele on one chromosome, and (3) several known cancer genes. Close-up observation of the position of the marker allele for. Table 6 shows the quartiles for the entire genome

Show the contents.

 Table 6

Quartile

 type :

 Q0 Q1 Q2 Q3 Q4 Sister-specific Markers

 1136 1394 1603 2152 11160 Allele

 In axon 26 55 69 96 476 intron 23 59 76 102 503 b

 Other

 SNPs 4321 6793 7967 10693.5 56458 In Sites

 Total 4374 6911.5 8098 10898 57437 Non-cancer disease 19 39 48 65.5 278 Covered cancer 0 4 5 7 27 Gene Other 275 501.5 597 825.5 3719

Total 298 546.5 650 895 4018 In the first column, the quartiles for the number of sisters 5 ^-^ ₀ for each sister were analyzed. For each set of SNP-S _10s in each of the sisters, the number of non-overlapping SNPs was measured and classified into three categories of axons, introns and other sites without transcripts. The number of genes covered by a set of sisters for each sister is listed in the last column and the genes were defined as sites spanning 5 kb upstream / downstream of transcript initiation / termination. Gene annotation data was downloaded from the Gene Track of the UCSC Table Browser on the Human Genome Build 19, and disease genes were downloaded from the GAD (Genetic Association Studies of Complex Diseases and Disorders) track of the UCSC Table Browser, and the cancer gene was cancer of the Wellcome Trust Sangerlnstitute. Annotated from Gene Census.

 Table 6 shows the following:

 (1) Most of the sensitive marker alleles do not overlap with the genetic code region. Even when overlapped, not all alleles in this case overlap with known cancer genes.

 (2) There are many marker alleles per sister-pair, averaged approximately

1,600. 4 shows various features known on the chromosome 3 of 8 cancer sensitivity marker alleles on chromosome 3 (many cancer genes have been identified) (eg, location of known cancer genes, SNP density, genetic code site and cancer sensitivity). Shows the result of mapping relative to the cytoband) where the GWAS hit is found. The following general observations were made:

 (3) There are many sensitive marker alleles for each cancer type, but they form very overlapping and distinct clusters, indicating that the sensitive marker alleles are not random "noise";

 (4) consistent with the observation of (1) above, there is no significant correlation between the cancer gene and the marker allele;

 (5) No strong correlation was observed between the position of the marker allele and the cytobands associated with cancer sensitive SNP hits obtained at G Ss.

 Most of the sensitive marker alleles are found in the non-genetic code region of the genome, but Figure 5 shows two regions of the genetic code region where the marker allele is mapped near the genes (a) TP53 and (b), which are well known cancer genes. An example is shown. All marker alleles of both cancers (BRCA and C0AD), whose most cancer genes are recorded in the 0MIM database, were examined. The following experimental results were obtained:

 (6) Sensitivity marker alleles for the two cancer types do not overlap with the two gene positions. Sensitivity markers. Alleles overlap with other nearby genes;

 (7) However, about 50% of all known cancer genes for the two cancer types overlap with the SNP positions of the marker alleles of BRCA and C0AD, respectively, or 20 kb of the SNP positions of the marker alleles of BRCA and C0AD, respectively; Found in the 50 kb range (see FIG. 8);

(8) The location of marker alleles relative to known associated SNPs identified by GWASs for the two cancer types (BRCA and C0AD) were analyzed. According to the analysis results, 8 out of 13 BRCA-associated SNPs are labeled with their respective markers. Within 5 kb of the allele and three of the six C0AD-associated SNPs are within 20 kb of each marker allele. In both cancer types, all associated SNPs are found within 500 kb. Summary and Discussion

summary

 The present inventors introduce the concept of SNP syntax (SNP-S) and the feature frequency profile of this syntax to provide a method for analyzing the systematic characteristics of WG SNPs of an individual. Subsequently, multiclass cancer susceptibility was evaluated by comparing the FFP of the rare SNP-Ss of each individual with the FFP of the control individual and those with eight main cancers. Although the amount of SNP data currently available in the TCGA database is small, the present invention predicts genetic susceptibility to eight major cancers with an accuracy in the range of about 47-76%, depending on the type of cancer. This accuracy will increase as the size of the sample for each cancer type increases, and the increase in sample size will be readily obtainable by current sequencing techniques.

The findings of the present invention support the "multiple assortment model" for cancer susceptibility:

 1. The individual's susceptibility to cancer is associated with a set of many regressive SNP syntaxes (sister specific marker alleles) present in the non-genetic code region of the genome (Table 6);

 2. There are many such sets of cancer susceptibility in one population (average about 34 sister pairs), and most of the marker alleles of one sister pair are different from the marker alleles of the other sister pair in the same cancer type;

3. In one type of cancer, Sister specific ^markers, each set of alleles can be expected from a cancer-specific marker allele "classified", which all unusual unique Sister for one type of cancer ever It is a collection of marker alleles. Discussion

Sample size VS. accuracy

 Increasing the number of cancer types degrades the performance of the present invention, while increasing the size of the dataset for each cancer type improves the performance of the present invention (Fig. 7). Substantial increase in sample size compensates for performance degradation as the number of cancer types increases. In addition, if the sample size of each cancer subtype is quite large, the present invention can be used to predict sensitivity to individual subtypes of one cancer type. Accuracy Limits for Sensitivity Prediction

 The present invention predicts genetic susceptibility to eight major cancers with an accuracy in the range of about 47-76%, depending on the type of cancer. Although increasing the sample size for one cancer type increases accuracy (Figure 7), it does not reach 100%. Not all genetic susceptibility to one cancer type triggers cancer, and in most cases, the occurrence of cancer requires one or more triggering events that are non-genetic.

"Error" sister

 Error sisters can be misclassified by a number of factors:

1) Because of the small sample size, true sisters may not be found in the dataset, which will be found as the sample size increases in the future. 2) Genetic sensitivity is similar but non-genetic factors can be used to compose different cancer types and sisters. 3) the error sister may be a genotype sister for a similar but similar cancer phenotype, and 4) a systematic error due to an incorrect genotype call (caused by experimental or computer bias) 5) population stratification (stratification) was not taken into account, and 6) any markers caused by the selection of the "optimal" filtering threshold with limited samples. Loss of allele. If the sample size increases substantially, some error sisters may turn out to be true sisters. Or, instead of looking for a sister, there is another way to classify VAR-synthax. For example, SVM can be applied. Full genome sequence vs. SNP vs. Tagging Common SNPs

 Since the available SNP densities may limit the accuracy of the present invention, increasing the number of SNPs using all SNPs loaded in sequence rather than the tagging common SNPs of the high-speed platform used in this study may improve the sensitivity prediction of the present invention. Substantially improve. Using WG sequences other than SNPs or tagging consensus SNPs will further improve the performance of the present invention. Number of susceptible marker alleles

A very large number of sensitive marker alleles (approximately 40,000-118,000, depending on the type of cancer) were classified into an average of about 34 sister pairs per cancer type, with each sister pair having about 1,600 sensitive marker alleles. This is equivalent to about 0.05% of all SNP-S ₁₀ s in one individual. Most of the alleles are very overlapping and clustered, indicating that the alleles are not "noise". For example, FIG. 4 shows that there are about 21 clusters (range 5-33) per cancer type on chromosome 3. Population Stratification and Other Factors

We have limited our study to individuals with European ancestors in the United States with racial information, but there may be potential population substructures in the samples of this study (AL Price, et al. (2008). of European Americans in Genetic Association Studies.PL ^' oSGenet D-e236), which indicates a biased sister in the same sub-population. W 201

There is a possibility to provide. Similarly, there may be other hiding variables that may systematically affect the results of the present invention.

 "Forced calls" of genome-wide data, absence of epigenetic information, errors in the human reference genome sequence (2004) .Finishing the euchromatic sequence of the human genome.Nature 431 (7011): 931-945) and genotype call errors (N. Rabbee & TP Speed (2006) .A genotype calling algorithm for Affymetrix SNP arrays. Β / οϊη format ics 22 (1): 7-12). There may be a number of other factors. Improvements in both of these factors will improve the prediction accuracy of the present invention. Other cancer phenotypes, experimental placement and other systemic biases (S. Turner, et al. (2011) .Quality control procedures for genome-wide association studies.Curr Protoc Hum Genet Chapterl; DG Clayton, et al. (2005). structure, differential bias and genomic control in a large-scale, case-control association study.Additional quality control required for SNP genotypes obtained for Nat Genet 37 (11): 1243-1246) may improve prediction accuracy. will be. Other possible gender

At the Population level or at the individual level, the present invention can provide substantial information: quantitatively predicting the size of the population with high genetic susceptibility to cancer is essential in establishing cancer prevention policies and cost control strategies. This is very useful information. Similarly, predicting the genetic susceptibility of an individual provides motivation for prevention and for early early diagnosis. Other applications in which the present invention may be applied include the study of genetic susceptibility to other diseases such as chronic diseases, infectious diseases and neurological diseases. In addition, if there is genome data for a stratified sample, the present invention may also be useful in determining the sensitivity and therapeutic benefit of a patient to a particular treatment. It may be applied to assess the patient's sensitivity to clinical trials that may increase the likelihood of efficacy and reduce the risk of adverse events.

II. Second invention: integrated approach

Research method

Sample and Zino Typing

 Sampling and genotyping were performed in the same manner as "database and sampling" described in the first invention described above. Sample quality control

Several instances of non-European ancestors were removed from the TCGA to ensure that no stratification errors occurred. All markers were typed with Affymetrix 6.0 SNP chips. The dataset used in this study was obtained using PUNK under the following conditions: Given that the platform used by the inventors was designed to be suitable for the type high polymorphic position, it had a minor allele frequency of less than 1%. SNPs were removed regarded as noise, and apply the Hardy Weinberg Equilibrium test for each of the marker object in the CEU Go> 10 ^{~ 6).} In addition, the association test between one plate and other plates (having more than 30 samples) for all plates within the cancer characteristics? > 1 X 10 ^{_ 8} ), the plate-effect test was applied. For SNPs that passed QC, self-published Caucasian individuals in the US were extracted from TCGA data and subjected to genetic relevance testing (PI_HAT <0.2). In addition, the Pedigree information was used to remove all relevant entities from the CEU. Finally, samples were integrated and SNPs were jointed against post—QC characteristic datasets to obtain genotypes for non-repeating 714649 SNPs of autosomal to 1741 individuals (Table 7).

[Table 7]

Encoding of Descriptor Elements

 Two encodings were used: (i) to increase classification performance, each SNP genotype was converted to a number of 0, 1 or 2 depending on the number of minor alleles in that genotype; (ii) For effective profiling of SNP-Ss, each SNP of the SNP-S descriptor was converted to one of 10 alphabets (see Table 3).

Optimum Length of SNP-Ss

Since the number of SNP-Ss of all possible lengths is too large (about 10 ¹² for one SNP string of length 10 ⁶ ), the mathematical calculations required to compare FSPs of this size are difficult, and the Only SNP-S of optimal "length was used. Using the optimal length ⁱ to give greatly ease the burden computer is described in the previous study of the present inventors. 4 ways

 The four methods presented by the inventors are summarized in FIG. 9. The details are as follows:

1) kNN / SNPS method: k-nearest neighbor (kNN) algorithm for SNP syntax (SNPSs) _,

Vectors of SNP-Ss for all members of the training set were obtained and then the feature screening step proceeded. In this step, the syntax shared by some percentage of the population is removed (filtered out) and the residue (filtered in) is used for analysis. It was then normalized to the total number of rare SNP syntax of the subject. Finally, a rare SNP syntax was used to construct a Jensen-Shannon (JS) divergence matrix among all members. The reason why JS divergence was chosen to measure the distance of the descriptor is that it is more predictable than other conventional methods such as allele sharing. The paired JS distances were measured for all individuals, each individual, then voted 9 classes from the top k nearest individuals and selected the one with the highest count. In the case of equivalence, the class having the shortest average distance from the target entity among the class entities in the upper k was selected. Accuracy was measured using the correct guess assignments for all members. For the best accuracy of cancer sensitivity estimation for the training dataset, the optimal length, 1, f parameter for low frequency selection, and parameter k for SNP-S were optimized. Optimal parameter values were 8, 1, and 40 for!, F, and k, respectively (FIG. 10A, Table 8A). In the testing phase, the same / and / optimal parameters were used. The JS distance vector between the subject and the training sample was then measured. The test subjects were predicted through the same selection process in the training phase with the optimal k parameter.

2) kNN / SNP Method: k-Nearest Neighbor (kNN) Algorithm for SNPs

 SNP-S descriptors were replaced with SNPs and KNN was remodeled in the same manner as in 1) above. Different from SNP-S, each SNP was converted to the numeric form of 0, 1 and 2, depending on the count of minor alleles in the genotype. In the SNP, the / and ^: parameters (see Figure 10b, Table 8b) were trained. Optimal values for / and the parameters were 15% and 200, respectively.

 3) SVM / SNP Method: Support Vector Machine (SVM) for SNPs

SVM is a supervised classification method, originally designed for building binary classifiers, and later used to build multiple classifiers in various ways. We used the One-Versus-One (OVO) scheme because it is empirically known to be superior to other methods (Duan KB & Keerthi SS (2005) Which is the best multiclass SVM method? An empirical study. Lect Notes Comput Sc 3541: 278-285. The 0V0 method generates a Γ · classifier for each pair of n classes and takes the class with the highest election from / predictions for the test sample. In order to implement the 0V0 SVM method, LIBSVM by Chang et al. (Chang CC & Lin CJ (2011) LIBSVM: A Library for Support Vector Machines. Acm T Intel Syst Tec 2 (3)). We chose RBF (Radial Basis Function) for the kernel function, because it is known to be superior to other functions. To build a binary classifier of SNP descriptors, SNPs were filtered out for a given P-value threshold?) To select the associated SNPs between the two classes (see FIG. 10C). In order to examine the optimal cut-off, the range was from ⁶ to 10- ^{10_ 3 (Wei Z, et al} (2009) From Disease Association to Risk Assessment:.. An Optimistic View from Genome-Wide Association Studies on Type 1 Diabetes Plos Genet 5 (10)). No cutoff less than 1 (Γ ⁶⁾ was applied and no classifier leaves SNPs after filtering by the associated test. During training, ribs from the dataset of 66 samples (594 individuals in total) for each cancer. We evaluated the performance of the 0V0 SVM prediction through leave-one-out cross-validation, based on the parameters trained from the rest of the dataset by rib-one-out cross-validation. _the prediction of how the performance was evaluated for ^"random sample is repeated for the process in all cases, collecting the results of the class (type of cancer) assigned then made marks on the uncertainty matrix (contingency matrix) (see: Table 8c In the case of ambiguous predictions (ie, multiple best elections), the poll is repeated in the set of best elected classes until the tie is broken. One prediction result was best when the ^ value cutoff value was 1 × 10 − ⁵ (see FIG. 10C).

4) SW / SNPS Method: Support Vector Machine (SVM) of SNPSs

Another predictive model was constructed using SVM using SNP-S instead of SNP (see FIG. 10D, Table 8 (1). Additional parameters for optimal length of SNP-S (explored and optimized during training) Except is included, the overall pipeline of the method is the same as 3) above. ^^ value greater optimal length for optimal values, and SNP-S on-off value is a 10- ⁵ and 2, respectively. Bayes i an inference of multipole prediction algorithm

In order to classify individuals into one of the nine phenotypes represented by T = {B, C, H, K, 0, L, R, U, N, each phenotype was labeled with the first letter of the full initial of each trait. . Bayesian inference of the prediction results of the four methods was used. These methods have the following abbreviations: KNN / SNPS, KNN / SNP, SVM / SNPS, SVM / SNP. The methods are mathematically represented by nf, π ?, m ⁴ respectively. For each test subject i, the traits with the highest postconditioning were selected for the predicted results from the training method, which can be formulated PCsjlHfiJ ^ fi ^ —. In the above formula, s _/ is the predicted trait of the subject i, i is the trait of the subject / predicted by the method. By Bayesian theorem we can write:

분모 Ps j^ fi ^J^)는 정규화 상수이다. 각 방법의 예측 결정은 서로 내재적으로 독립적이기 때문에, 체인규칙을 적용한다 (Zhang H (2005) Ex loring conditions for the optimality of Naive bayes . Int J Pattern Recogn 19(2)： 183-198)：

Denominator Ps j ^ fi ^ J ^) is the normalization constant. Since the prediction decisions of each method are inherently independent of each other, we apply the chain rule (Zhang H (2005) Ex loring conditions for the optimality of Naive bayes. Int J Pattern Recogn 19 (2): 183-198):

t 蒙 = argmax _teT IT P (! W _} / s,-= t) x P (s _} = t)

In the above equation, PCMils ^ t) and P ( _Si = t) can be empirically inferred by maximum likelihood estimation from what was observed during each of the four methods. For example, among the total BRCA samples in the training set, P0 = Cls # B) can be estimated by identifying a portion of the true BRAC individuals assumed to be C0AD by the / SNP-S method. Ys 尸 ^ is the trait t of all trained individuals. W 201

This corresponds to a portion of the sample, which is the same for each of the nine traits (because they use the same sample size for each trait). Multi-class Cancer Prediction in Men

 The methods of the invention classify individuals into multiple cancer types, including three female-specific cancers and three common cancers. In order to prevent men from being predicted to be one of the female cancers, male subjects were classified into five general cancers excluding breast cancer, ovarian cancer and endometrial cancer. Results

In identifying specific aspects of this system (eg cancer susceptibility) in connection with complex systems such as genomes, two decisions are required: of the many methods, which method is the appropriate description of the system, and which Can analytical methods be applied to the descriptors to provide useful information about this aspect. We used two analysis algorithms, four methods, applied to two different descriptors of the individual genome. The two descriptors of the individual genome are: (i) the profile of ordered SNPs (each SNP is assumed to be independent of its neighbors), and (ii) the profile of the SNP syntax (SNP-S is a specific length of linkage, Defined as ordered SNPs). Using SNP-S as one of two descriptors reflects the observation that each SNP location is not independent and is connected to neighbors to varying degrees. The use of experimentally obtained genotypes instead of computer inferred haplotypes is due to the fact that haplotypes are unreliable, in particular unreliable for the regression frequency SNPs of unrelated individuals on which the methods of the present invention are constructed ( Fan HC, Wang J, Potanina A, & Quake SR (2011) Who 1 e-genome molecular ha lotyping of single cells.Nature biotechnology 29 (1): 51—57). The individual genome SNP-Ss is created by sliding a window of a certain length along the entire length of the total genome SNPs. Descriptor elements (SNP or For SNP-S), factors that increase the sensitivity of different cancer types are selected: SNPs or SNP-Ss with "very low Kal" or "rare frequency" depending on the analysis algorithm used.

 Two common analysis algorithms used were: (i) the nearest neighbor NN) analysis algorithm and (ii) the support vector machine (SVM) analysis (Theodoridis S & Koutroumbas K (2009) Pattern recognition). The former searches for the k nearest neighbors of the test subject and the latter identifies the most likely class to which the test subject belongs.

In the analysis algorithm, all pairwise "distances" are calculated between the descriptor of one entity and the descriptor of each entity. The NNs for the test subject are then selected and the subject is predicted whether it is sensitive to the most common traits among the A Ns (if there is more than one of the most likely traits, see the method above). In the SVM analysis algorithm, SVM is trained to recognize the correct trait of one individual in each of all binary traits. Finally, having the maximum selection of all pair classifications by SVM predicts the susceptibility of the test subject to the most likely trait. The final prediction of the sensitivity of the test subject is estimated based on Bayesian inference from the four prediction results. For female subjects, multiclass susceptibility was estimated for nine classes (eight joint cancer classes and one health trait), and for male subjects, predictions were made for six classes except three female-specific cancer classes. Was carried out.

All data used in this study was obtained from public databases (The Cancer Genome Atlas (TCGA) and HapMap). Details of data selection, sampling methods, sample control procedures and other details are described in the above experimental methods, and the numbers before and after sample control are listed in Table 7. The dataset was divided into two groups: a training set for optimization of the parameters for each method and a test set for independent verification of the methods. The maximum size of the sample for each trait in the training set was limited to the minimum sample size (66) of one trait of TCGA. To prevent artificially skewed predictions resulting from inappropriate sample sizes for each trait, 66 individuals were randomly equally extracted from each trait group. Lack of TCGA Due to the sample, a test set for all nine phenotypes could not be constructed. Instead, 50 individuals (individuals not used in the training set) were tested for each of the three traits (BRCA, 0V and UCEC) and the procedure was repeated 10 times. 9 is a workflow of the four methods. With the optimization parameters for each method, the performance for the training set (Tables 8A-8D) and the performance for the test set (FIGS. 11A-11C) were conducted.

Each test subject was taken from 594 practice populations to assess the statistical accuracy of sensitivity predictions for each trait. Table 8a is the result of ANN / SNP-S and the remaining three methods and results are described in Tables 8b-8d. Table 8a

Trainingperformanceof / cNNalgorithm applied to profiles of SNP-Ss |-

Predicted traits

B CA ^ 3¾ 10 fi 11 5 ^ Of 66 _* -485% ^ 纖 ffi m ： 細 爾^■ a 5¾ 66¾ Ι5, ΰβ

 %% n¾: 2A 2. 藤 '83.3%'

 , '3m m 騰-m 2i.職 鶴 纖

. ¾ i _f -52 ^ ¾ ^: 3 ^ 66 _* 78.8% ^ 嫌 ft: 51 m as m 2¾ mm m ^:扁 霸 麵 ¾ 誇^: ¾¾:

w ^■ m .11- '7, 纖 鹏;縮 醫

6 疆 ¾ 66 ^ 66 ¾ιδ¾

Sum 594, Overall 66.0%,

 ζ 画 ¾¾ «¾¾? 1¾ 觀 M4 6a

 m a¾ ： 飾 囊 幽 fe

! INS 56 * Ik

 滅 聽 r 龜 議

Fort e 6bbreviations.refer.to Tab 1 iegend.

Training performs nce of SVM algorithm applied to profiles of SNPs ..

Predicted tra t-

 Sura 594? Overall 64.1% ^

Fcr: he abbra-iations, re¾r to T.ble I ^' legeridg

Predictedjrai ; E。 ⁸

Predictedjrai

I 13 ^ li ^; S 47.0%; v .;

 25- 疆 顧 赚 &

 iii a¾ 涵 涵

 if * i fo i¾ si? 66 ...

LGG ^ ¾ 37 ·> 1 66 »·

. ' ^:觀_: ¾ w; ¾ 66Kr; g7.0¾

^' ᅳ>% 2¾. I 6¾

W (Mg: 簾 (k ^'麵 fe 66ί

 CEU ft OS- fife 66e 66 * 翻

^Sum594 "Oveall 51.0%,

¾iHei6greviatiors, refer to Table 1 legend ,?

a-

Predictions were made with significantly higher accuracy than random predictions (eg, in the 扁 / SNP-S method of Table 8a, the true positive rate was 66% and the false positive rate was 33%); (iii) the single of the four methods does not show the best performance in predicting the sensitivity of all traits; (iv) There were no false positive results for healthy traits and there were some cancer individuals classified as belonging to the healthy group, but no healthy group was found in any cancer group.

 Due to the small sample size of any cancer in the dataset obtained from TCGA, no test set for all nine traits could be constructed. Therefore, we randomly selected 100 new samples from three groups: Breath Invasive Carcinoma (BRCA), 0varian Serous Cytadenoc arci noma (0V) and Uter ine Corpus Endometrioid Carcinoma (UCEC). In each method, multiclass accuracy for the test subjects was calculated using parameters optimized in the training set. Resamples of 50 individuals (randomly selected from 100 test samples) were repeated 10 times for each of the three cancer classes. 11A-C show the results of the four methods above with statistical spreading from multiple sampling. The results of the test set can be summarized as follows; (i) For each cancer class, three of the four methods predicted the test set with significantly better accuracy than random prediction; (ii) Individual genome variants of BRCA and 0V (strictly depicted in terms of SNPs or SNP-Ss) are more interrelated than the rest of the cancer types and are slightly less than the descriptors in 0V and UCEC. There was a similar connection between them.

For each test subject, the certainty of the prediction was estimated by the posterior probability of Bayesian inference (FIG. 12). The results for the three classes show no prediction by the maximum posterior probability below 0.3. In the case of BRCA, 30% of the test subjects had a high uncertainty call defined as a test subject with a maximum posterior probability of 0.9 or higher, which was 83.3% accurate, Accuracy increased by 25.3%. Taken together, (i) the multiclass accuracy of predictions for cancer sensitivity based on a combination of four methods was several times higher than the 11% accuracy of random predictions; (ii) the quality of predictions made to make health-determinations for individuals or populations, which may be improved if more data is available in the future; (iii) Descriptors of the two cancers, BRCA and 0V, were more similar to each other than the other traits. 0V and UCEC also showed similarities, but less than BRCA and 0V.

 As described above, a specific part of the present invention has been described in detail. For those skilled in the art, the specific technology is merely a preferred embodiment, and it is obvious that the scope of the present invention is not limited thereto. Therefore, the substantial scope of the present invention will be defined by the appended claims and equivalents thereof.

Claims

[Patent Claims]  [Claim 1]  Computer implemented method for analyzing genomic variation or epigenomic variation of an organism comprising the following steps:  (a) constructing a linked string of the variants; (b) constructing a variation syntax (VAR-S) of a particular length by applying a sliding window of a particular length along the entire length of the linking string;  (c) counting all possible features in the particular length variation syntax and assembling them into feature frequency profiles (FFPs); And  (d) determining the distance between the FFPs or classifying the FFPs. [Claim 2]  The method of claim 1, wherein the genome mutation is a single nucleotide polymorphism (SNP), deletion, insertion or repeat mutation in a nucleotide sequence. [Claim 3]  3. The method of claim 2, wherein said mutation is SNP. [Claim 4] The method according to claim 2, wherein the nucleotide sequence is a sequence on one chromosome, a sequence on a plurality of chromosomes or the entire genome sequence. [Claim 5]  The method of claim 4, wherein the nucleotide sequence is an entire genome sequence. [Claim 6]  2. The method of claim 1, wherein step (a) is performed by assigning a code to each of the variants to construct a linking string of the code. [Claim 7]  The method of claim 3, wherein the step (a) is performed by assigning a code to each genotype or haplotype of the SNP to construct a linked string of the codes. [Claim 8]  The method of claim 1, wherein the mutations are SNPs, wherein the SNPs are (i) removal of SNPs exhibiting an allele frequency of 5% or less, (Π) Hardy Weinberg Equilibrium test and (ii) plated. A method characterized in that the SNPs are QC (Quality Control) by at least one method selected from the group consisting of effect tests. [Claim 9]  2. The method of claim 1, further comprising determining an optimal length of the variant syntax prior to step (b). [Claim 10] 10. The method of claim 9, wherein the determination of the optimal length is empirically determined with a length that exhibits the highest accuracy with respect to the phenotype of the organism. [Claim 11]  The method of claim 10, wherein the phenotype of the organism is a disease. [Claim 12]  10. The method of claim 9, wherein step (b) is performed using a sliding window having the determined optimal length. [Claim 13]  10. The method of claim 9, wherein the mutation is SNP, the density of the SNP is 1 million SNPs / genome, and the optimal length is 6-14. [Claim 14]  14. The method of claim 13, wherein the optimum length is 8-12. [Claim 15]  15. The method of claim 14, wherein the optimum length is ten. [Claim 16] 2. The method of claim 1, wherein the FFPs whose length is determined in step (d) are filtered-in for a rare VAR-S of a particular length, wherein the rare VAR- of that particular length is filtered. S contains the variation of the subject of analysis A method characterized in that it is a rare VAR— S filtered-in frequency less than 20% in population. [Claim 17]  17. The method of claim 16, wherein said filtering-in is at or below 5%. [Claim 18]  18. The method of claim 17, wherein said filtering-in is at or below 3%. [Claim 19]  The method of claim 1, wherein the distance in step (d) is obtained by applying a distance function. [Claim 20]  The method of claim 1, wherein the classification of step (d) is performed using a class prediction algorithm. [Claim 21]  The method of claim 1, wherein the organism is an animal, plant, fungus, yeast, bacteria or protist. [Claim 22] The method of claim 1, wherein the mutation is a variation associated with traits of the organism and the method is used to predict sensitivity to traits of the organism. [Claim 23]  23. The method of claim 22, wherein said trait is diseases, disorders, conditions or symptoms. [Claim 24]  The method of claim 23, wherein the trait is a cancer disease. [Claim 25]  23. The method of claim 22, wherein said trait is growth rate, yield, or quality. [Claim 26]  The method of claim 1, wherein the variation is a variation associated with therapeutic responsiveness of the organism and the method is used to predict the therapeutic responsiveness of the organism. [Claim 27] Computer-readable storage media containing instructions instructing the computer processor to perform the following steps: (a) Genomic variation or epigenomic variation Constructing a linked string of); (b) constructing a variation syntax (VAR-S) of a specific length by applying a sliding window of a specific length along the entire length of the linking string; (c) counting all possible features in the particular length variation syntax and assembling them into feature frequency profiles (FFPs); And (d) determining the distance between the FFPs or classifying the FFPs. [Claim 28]  Systems for analyzing genomic variation or epigenomic variation in organisms, including:  (a) a computer processor; And  (b) the computer-readable storage medium of claim 27 coupled with the processor, [Claim 29]  Computer implemented method for analyzing genomic variation or epigenomic variation of an organism comprising the following steps:  (a) constructing at least two kinds of descriptors for said mutations;  (b) applying at least two kinds of class prediction algorithms to each of the at least two kinds of descriptors to analyze the genome variation or epigenetic variation of the organism to obtain at least four kinds of prediction results; And  (c) finally predicting the trait of the organism by applying at least four kinds of prediction results obtained in step (b) to an inference algorithm. [Claim 30]  30. The method of claim 29, wherein said genome variant is a single nucleotide polymorphism (SNP), deletion, insertion, or repeat variant in a nucleotide sequence. [Claim 31] 31. The method of claim 30, wherein said mutation is an SNP. [Claim 32]  30. The method according to claim 29, wherein at least two kinds of descriptors for the variation are (0) a profile of variations that each variation is assumed to be independent of neighboring variation and (ii) associated variation of a particular length. Wherein the variation of the term comprises a profile of the syntax (VAR-S). [Claim 33]  30. The method of claim 29, wherein the variants are SNPs, wherein the SNPs are (i) removal of SNPs exhibiting an allele frequency of 5% or less, (ii) Hardy Weinberg Equilibrium test and (ii) plate- A method characterized in that the SNPs are QC (Quality Control) by at least one method selected from the group consisting of effect tests. [Claim 34] 31. The method of claim 30, wherein the at least two kinds of class prediction algorithms include a k-nearest neighbor algorithm and a support vector machine (SVM) algorithm. ^' [Claim 35] 33. The method of claim 32, wherein the descriptor for the variation is a profile of VAR-S and the class prediction algorithm is a nearest neighbor algorithm, and step (b) is less than 20% lower in (b-1) population. A small step of selecting regression VAR-S found at a frequency; (b-2) substep normalizing to the total number of regression VAR—S; (b-3) constructing a Jensen-Shannon (JS) divergence matrix using the profile of the regression VAR-S; And (b_4) selecting a k-nearest neighbor NN) for the organism using the JS divergence matrix. [Claim 36] 33. The method of claim 32, wherein the descriptor for the variation is a profile of a VAR-S, the class prediction algorithm is a support vector machine (SVM) algorithm, and step (b) comprises (b-1) 10_ ² to 10— Selecting a VAR-S having a low P-value of ⁶ ; (b-2) substeps of performing SVM on each of all binary traits; And (iii) a substep of sorting according to a max-win voting scheme. [Claim 37]  33. The method of claim 32, wherein the descriptor for the variation is a profile of the variation, the class prediction algorithm is the nearest neighbor algorithm, and step (b) is found with a frequency of less than 20% in (b-1) populations. Selecting a regression variation to be generated; (b-2) substep normalizing to the total number of rare mutations; (b-3) constructing a JS divergence matrix using the profile of the rare variant; And (b-4) selecting a k-nearest neighbor NN) for the organism using the JS divergence matrix. [Claim 38] 33. The method of claim 32, wherein the descriptor for the variation is a profile of the variation, the class prediction algorithm is a support vector machine (SVM) algorithm, and step (b) comprises (b-1) 1 to ² to 1 (Γ ^6). A substep of screening for variants with low P-values of (b-2) a substep of performing SVM for each of all binary traits; and (iii) max—win voting The method comprises the step of classifying according to the) method. [Claim 39] 30. The method of claim 29, wherein the inference algorithm is a Bayesian inference algorithm. [Claim 40]  30. The method of claim 29, wherein the trait of said organism is diseases, disorders, conditions, symptoms, or treatment responsiveness. [Claim 41]  41. The method of claim 40, wherein said trait is a cancer disease. [Claim 42]  30. The method of claim 29, wherein said trait is growth rate, yield, or quality. [Claim 43]  A computer-readable storage medium containing instructions instructing a computer processor to perform the following steps: (a) at least two types of descriptors for the mutation; Constructing; (b) applying at least two kinds of class prediction algorithms to each of the at least two kinds of descriptors to analyze the genome variation or epigenetic variation of the organism to obtain at least four kinds of prediction results; And (c) finally predicting traits of the organism by applying at least four kinds of prediction results obtained in step (b) to an inference algorithm. [Claim 44] Systems for analyzing genomic variation or epigenomic variation in organisms, including: (a) a computer processor; And  (b) The computer-readable storage medium of claim 43 coupled with the processor.