CN103366100A - Method for filtering SNP (Single Nucleotide Polymorphism) unrelated to complex diseases from whole-genome - Google Patents

Abstract

A method for filtering SNPs irrelevant to complex diseases from the whole genome, which is used for the research of pathogenic mechanism, early diagnosis and biopharmaceutical development of complex diseases. (1) Preprocessing and initialization of single nucleotide polymorphism SNP data. According to the principle that any gene variation in homologous chromosomal alleles can be treated equally to the disease, the SNP data is processed into data containing only 0, 1, 2, and 3. (2) Define the correlation measure. The association I(Y;X) between a SNP subset X and a disease Y is defined as the mutual information MI(Y;X) between X and Y. (3) Use the FGSA method to search for the candidate suspected pathogenic SNP group in the SNP collection. (4) According to the frequency-association priority criterion, in the set of SNP groups of candidate suspected pathogenic causes, select the SNP group whose frequency of occurrence exceeds the threshold. (5) Outputting the SNPs whose frequency is greater than the threshold at the top. This method can retain the SNPs corresponding to the pathogenic causes covered by other pathogenic causes, and lay the foundation for the subsequent discovery of pathogenic causes.

Description

A method for filtering SNPs irrelevant to complex diseases from the whole genome

technical field

The invention belongs to the technical field of data processing. Specifically, a method for filtering SNPs irrelevant to complex diseases from the whole genome single nucleotide polymorphism (Single Nucleotide Polymorphism, SNP) data is proposed, which can be used for pathogenicity of complex diseases Mechanism research, early diagnosis and biopharmaceutical development.

Background technique

Complex diseases are caused by multiple genetic factors and environmental factors, and their occurrence and development are affected by multiple genes in a complex network structure. Complex diseases are different from Mendelian genetic diseases. In most cases, there are often no major genes that are sufficient to cause disease, and the effect of a single gene on disease may be negligible or even non-existent, but these single gene combinations that may play a negligible role Together, their combined effects may be the cause of complex diseases. These characteristics have brought great difficulties to the discovery of pathogenic genes of complex diseases, and it is difficult to find pathogenic genes or related markers for the study of pathogenic mechanisms, early diagnosis and development of biological drugs for complex diseases. How to find out multiple causes of disease and which genes combine to become a cause of disease on a genome-wide scale are the main problems at present.

Solutions can be divided into two categories: direct methods and two-step methods. The direct method searches directly on the original SNP set, and can only process small and medium-scale data, and the scale of the processed data varies depending on the algorithm (for example, the processing scale of MDR, BEAM, etc. varies greatly). The two-step method is to first filter out those SNPs that are not related to the disease from the original SNP set, and then search in the remaining SNP set. The present invention involves the first step in a two-step process: SNP filtering.

The first step of most two-step methods is carried out exhaustively, that is, to find pairs of SNPs with high scores, and leave them for the second step (such as BOOST, AntEpiSeeker, etc.). The ability of this method to deal with large-scale data is extremely limited, so an ant colony algorithm is introduced to find a higher-order SNP combination (AntEpiSeeker) than possible, and a random forest is introduced to find a subset of SNPs with strong classification ability to obtain further The set of SNPs investigated. These filtering methods all have following deficiencies:

1. The order of SNPs that can be processed is very limited. For example, the amount of calculation for exhaustive enumeration is huge. Only two or two SNPs are exhausted to obtain their scores, so that only the second-order SNP interaction (that is, the interaction of two SNPs) can be retained, and the Higher order SNP interactions.

2. The scale of SNPs that can be processed is very limited. This is because the filtering process requires complex calculations. For example, the data scale processed by the random forest method is only about 100 SNPs. Relatively speaking, AntEpiSeeker can handle more SNPs (such as 5,000 SNPs ).

3. It cannot handle the multi-cause situation where one cause of disease is covered by other causes of disease.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the use of a two-step method to find out multiple causes of disease and which genes combine to become a cause of disease in the whole genome, and to invent a method to filter and complex Disease-independent SNP method, this method filters out those SNPs that are not related to the disease phenotype individually or in combination from the genome-wide SNP data, thereby retaining all high-order SNPs that may be associated with the disease individually or in combination Factors, which lay the foundation for further detection and identification of the cause of the disease-causing SNP.

Realize the technical scheme of the present invention, comprise the steps:

(1) Preprocessing and initialization of genome-wide SNP data

其中x_i∈{0，1，2，3}^d为SNP i对应位点的取值：对应位点上的两个等位基因当为纯合子AA时取1，纯合子aa时取2，杂合子Aa或aA时取3，当该数据缺失时取0；y_i∈{1，2}为样本x_i的类标，1表示疾病组，2表示对照组，N为SNP数据中样本的个数，d为数据中SNP的个数，并记所涉及的SNP的集合为Ω；According to the principle that the variation of any gene in the homologous chromosome alleles has the same impact on the disease, the SNP data is preprocessed as follows:

Where x _i ∈ {0, 1, 2, 3} ^d is the value of the corresponding site of SNP i: when the two alleles at the corresponding site are homozygous AA, it takes 1, and when it is homozygous aa, it takes 2, Take 3 for heterozygote Aa or aA, and take 0 when the data is missing; y _i ∈ {1, 2} is the class label of sample _xi , 1 means the disease group, 2 means the control group, N is the number of samples in the SNP data number, d is the number of SNPs in the data, and record the set of involved SNPs as Ω;

(2) Define the correlation measure

Each SNP subset may become a pathogenic factor through interaction, a factor with 1 SNP is called an l-order factor, and the association I(Y;X) between SNP factor X and disease Y is defined as X Mutual information MI(Y;X) with Y:

I(Y;X)＝H(Y)-H(Y|X) (1)

in $h\left(Y\right)=-\underset{\mathrm{the\; y}\∈\left\{\mathrm{0,1}\right\}}{\mathrm{\Σ}}p\left(\mathrm{the\; y}\right)\log p\left(\mathrm{the\; y}\right)$ and $h\left(Y\right|x)=-\underset{\mathrm{the\; y}\∈\left\{\mathrm{0,1}\right\}}{\mathrm{\Σ}}\underset{x{\∈\{\mathrm{1,2,3},\}}^{\left|x\right|}}{\mathrm{\Σ}}p\left(\mathrm{the\; y}\right|x)\log p\left(\mathrm{the\; y}\right|x)$ are entropy and conditional entropy, respectively;

(3) Use the Factor based Genetic Search Algorithm (FGSA) to search for the SNP group of candidate suspected pathogenic causes in the SNP set Ω;

(4) sort the SNPs according to the frequency-association priority criterion;

(5) Outputting the SNPs whose frequency is greater than the threshold at the top.

The remarkable effect that the present invention has compared with prior art:

The invention discloses a method for filtering SNPs irrelevant to complex diseases from the single nucleotide polymorphism SNPs in the whole genome - a factor based genetic search algorithm FGSA (factor based genetic search algorithm), which guarantees The search for SNP factors that are weakly associated with a disease and strongly associated with a disease in combination, its layer-by-layer peeling criterion ensures that when multiple pathogenic factors exist, other possible strong associations will not be covered up by the existence of a strong association factor The search of the factors, and the frequency-association priority criterion adopted ensures the relative stability of the solution obtained in this way, and has the following significant effects:

(1) This method can effectively filter SNPs that are not related to diseases from the whole genome, that is, filter out those SNPs that are not related to the disease phenotype individually or in combination, and make the number of remaining SNPs after filtering as small as possible .

(2) This method can retain the SNPs corresponding to the pathogenic causes covered by other pathogenic causes, thereby laying the foundation for the subsequent discovery of these pathogenic causes.

(3) This method can handle genome-wide SNP scale, such as SNP data of more than 10,000, such as AMD (loss of vision due to retinal damage) data, which contains 103611 SNPs, 96 case samples and 50 control sample. The AMD-related SNPs found by other methods are all in the SNP set filtered by this method (see the description of the experimental comparison effect for details).

Description of drawings

Fig. 1 is the flowchart of FGSA algorithm of the present invention;

FIG. 2 is a flowchart of the genetic algorithm in FIG. 1 .

Detailed ways

With reference to Fig. 1 and Fig. 2, method of the present invention is called FGSA method, and its specific implementation steps are as follows:

Step 1, preprocessing and initializing the SNP data.

其中x_i∈{0，1，2，3}^d为SNP i对应位点的取值：对应位点上的两个等位基因当为纯合子AA时取1，纯合子aa时取2，杂合子Aa或aA时取3，当对应位点上的等位基因数据缺失时取0；y_i∈{1，2}为样本x_i的类标，1表示疾病组，2表示对照组，N为SNP数据中样本的个数，d为数据中SNP的个数，仅含0，1，2，3的数据，其中0表示缺失数据，所涉及的SNP的集合记为Ω。(1.1) According to the principle that the variation of any gene in the homologous chromosome alleles can have the same impact on the disease, the SNP data is processed into

Where x _i ∈ {0, 1, 2, 3} ^d is the value of the corresponding site of SNP i: when the two alleles at the corresponding site are homozygous AA, it takes 1, and when it is homozygous aa, it takes 2, Take 3 for heterozygote Aa or aA, take 0 when the allele data on the corresponding locus is missing; y _i ∈ {1, 2} is the class label of sample _xi , 1 means the disease group, 2 means the control group, N is the number of samples in the SNP data, and d is the number of SNPs in the data, including only the data of 0, 1, 2, and 3, where 0 means missing data, and the set of SNPs involved is recorded as Ω.

Step 2, define the correlation measure.

(2.1) Each SNP subset may become a pathogenic factor, and a factor with l SNP is called l-order factor. The correlation I(Y; X) between a factor X and a disease Y is defined as the mutual information MI(Y; X) between X and Y, expressed as formula (1):

I(Y;X)＝H(Y)-H(Y|X) (1)

in $h\left(Y\right)=-\underset{\mathrm{the\; y}\∈\left\{\mathrm{0,1}\right\}}{\mathrm{\Σ}}p\left(\mathrm{the\; y}\right)\log p\left(\mathrm{the\; y}\right)$ and $h\left(Y\right|x)=-\underset{\mathrm{the\; y}\∈\left\{\mathrm{0,1}\right\}}{\mathrm{\Σ}}\underset{x{\∈\{\mathrm{1,2,3},\}}^{\left|x\right|}}{\mathrm{\Σ}}p\left(\mathrm{the\; y}\right|x)\log p\left(\mathrm{the\; y}\right|x)$ are entropy and conditional entropy, respectively.

Step 3, use the FGSA method to search for the SNP group that is suspected to be the cause of the disease in the SNP set Ω.

(3.1) Set genetic algorithm parameters and FGSA related parameters

Set the number l of SNPs in the SNP group to be found. Set genetic algorithm parameters, including population size N _l , crossover probability P _c , mutation probability P _m , iteration number Iter _l ; set FGSA related parameters, including repeated search times Num _l , and the number of l-order interactions to be found each time M _l ;

(3.2) Initialize the number of the first-order interactions to be found k=1; set the SNP set S ₁ =Φ related to the suspected cause of disease; set the SNP set to be investigated as Ω ^* =Ω;

(3.3) Randomly initialize the population: Randomly generate l different effective SNP numbers from the SNPs in Ω ^* , constituting an l-order factor as an individual. A total of N _l individuals are generated to form a population;

(3.4) Calculation of fitness: For each individual in the population, calculate the mutual information according to formula (1) as the fitness of the individual;

(3.5) Selection operation: According to the fitness value of each individual in the population, use roulette and elite strategy to select _N1 individuals;

(3.6) Crossover operation: Randomly select two individuals from N _l individuals, randomly select the intersection point, and for these two individuals, exchange the part behind the intersection point with the crossover probability p _c to form two new individuals;

(3.7) Mutation operation: For each individual, randomly generate an effective SNP number, and replace any SNP number in this individual according to the mutation probability p _m ;

(3.8) Generate the next generation population: all individuals obtained after the operation of (3.7) are used as the next generation population;

(3.9) If the number of iterations is less than Iter _l , then jump to (3.4);

(3.10) Take the individual with the maximum fitness in the population and mark it as s _k and add it to the SNP set S _k suspected to be related to the disease, and remove this individual from the SNP set Ω ^* , that is

(3.11) Repeat (3.3)～(3.10) M _l times, each time add an individual to S _k , and remove this individual from the data, after M _l repetitions, get S _k ;

(3.12) Reset Ω ^* = Ω, repeat (3.2) ~ (3.11) a total of Num _l times, so as to obtain the SNP set ${\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HSA00000921376300011.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,S_2,...,S_{{\mathrm{Num}}_l};$

(3.13) Output the SNP set containing each l-order interaction

Step 4, calculate the frequency of each SNP in v.

(4.1) Calculation of frequency: the frequency of occurrence of the SNP of the suspected cause of disease found in step 3 as the frequency of the SNP;

(4.2) The SNPs were sorted according to the frequency-association priority criterion, that is, the SNPs were sorted according to the principle of the highest frequency priority, and the single SNP with the same frequency associated with the disease, that is, the highest mutual information priority.

Step 5, outputting the top SNPs whose frequency is greater than the threshold.

Among them, the implementation of the genetic algorithm in (3.3)～(3.8) in step 3 is to randomly generate l different effective SNP numbers from the SNPs in Ω ^* to form an l-order factor as an individual, and through their crossover The genetic evolution search method of mutation to obtain better individuals embodies the characteristics of this method's factor-based genetic evolution SNP filtering, thus ensuring the search for single SNP factors that are weakly associated with the disease and combined to be strongly associated with the disease; (3.9) in step 3 adds the individual with the greatest fitness in the population to the SNP set suspected to be related to the disease, and removes this individual from the data to realize the layer-by-layer peeling off of the pathogenic factors, so as to ensure that When multiple pathogenic factors exist, the search for other factors that may also be strongly correlated will not be covered up by the existence of a strong correlation factor; and the frequency-correlation priority criterion in step 4 ensures that the solution obtained in this way is relatively accurate. stability.

The present invention will describe the effect of this method in more detail through the following experimental examples. These experimental examples are for illustrative purposes and are not intended to limit the scope of the invention.

In the following experiments, the parameters of this method are set as: N _l =10, P _c =0.9, P _m =0.25, Iter _l =5000, Num _l =20, M _l =8.

Experiment 1: Filtering of simulated data SNPs.

The simulation data is based on the real SNP data of the New York population, and biologists add 7 SNP groups known to be associated with complex diseases, and the association models of these 7 SNP groups and diseases are different. There are two groups of data: the first group contains 2000 samples, 100 SNPs, represented by SNP100; the second group contains 2000 samples, 2000 SNPs, represented by SNP2000. The data information is shown in Table 1.

Table 1 Experimental data

Experiments are carried out on these two groups of data, wherein the parameters (Iter ₂ , Iter ₃ ) of FGSA take values for the two groups of data respectively, which are (Iter ₂ , Iter ₃ )=(600, 1100), (1200, 1700), ( M ₂ , M ₃ ) are the same for the two sets of data, both are (M ₂ , M ₃ )=(8,5). The two sets of experiments carried out are only different in the frequency threshold: for the two sets of data, the threshold is taken as Th=3, and 2 and 2 are taken as Th=1.

The experimental results of the FGSA algorithm and several typical feature selection methods (Minimum Redundancy Maximum Relevance Method--mRMR, Maximum Entropy Method--ME, etc.) No results found, the compression rate is the ratio of the number of SNPs obtained after filtering to the total number of SNPs in the data, and the factor rate is the ratio of the number of real pathogenic factors contained in the filtered SNPs to the number of real pathogenic factors in the data.

In Table 2, a is a true positive, and the factor rate is defined as the percentage of the number of pathogenic factors contained in the filtered SNP to the total number of pathogenic factors contained in Ω before filtering; the compression ratio is the percentage of the number of pathogenic factors contained in Ω after filtering Percentage of total number of SNPs.

Table 2 Performance of FGSA-frequency algorithm and comparison with other algorithms

Table 3FGSA-Frequency Algorithm Performance (Frequency Threshold is the result of Th=1)

It can be seen from Table 2 and Table 3 that:

(a) When the compression rate of the FGSA method is basically the same, its factor rate is higher than that of other methods, which shows the effectiveness of the algorithm;

(b) The factor rate of mRMR is only 3/7, that is, the selected SNP set contains only 3 of the 7 pathogenic factors, while the factor rate of the FGSA method is 5/7~6/7, that is At least 5 or 6 are fully included, which clearly shows the effectiveness of the FGSA method;

(c) The compression rate increases with the SNP size N, and the larger the N, the higher the compression rate, reaching 97% in the case of 2000 SNPs, indicating that the FGSA method is more effective for the genome-wide SNPs;

(d) When there are too many SNPs in the factor, the curse of dimensionality will appear, which is one of the important reasons why the pathogenic factors with a length of 5 cannot be completely selected by these methods.

Experiment 2: SNP filtering of real AMD data.

AMD is a medical condition affecting older adults who can experience loss of vision in the optic center due to damage to the retina. The AMD data (see Table 4) contains 103611 SNPs, 96 case samples and 50 control samples, with 0.811% missing data. The sample values are 0, 1, 2, 3, where 0 means that the data is missing. AMD data is often used in SNP association analysis, and many methods have been used on AMD data and some related genes have been obtained. Table 5 shows the SNPs found by other methods (such as BOOST, AntEpiSeeker, epiMODE, BEAM, HapForest, Single-Marker, DASSO-MB, etc.).

Table 4 AMD data

Number of SNPs Case sample number Number of Control Samples Missing data ratio AMD data 103611 96 50 0.822%

Table 5 The list of SNPs found by the FGSA method and also found by other methods

It can be seen from Table 5:

(a) The SNPs found by FGSA have a high overlap with those found by other methods, indicating the effectiveness of the FGSA method.

(b) FGSA also gives SNPs that are not found by other methods but have a high frequency, including numbers 19405, 6693, 56674, 80178, 76784, 92627, 46516, 88957, 42568, 51958, 41808, 47428 The SNPs, whose frequencies are 35, 26, 26, 25, 24, 24, 22, 21, 20, 15, 11, 9, do not rule out the possibility that they are associated with disease and, in particular, do not rule out true pathogenicity The factors are on the set of SNPs in Table 5 and the above-mentioned SNPs, or a subset of some of the SNPs.

Claims (1)

1. A method for filtering SNPs unrelated to complex diseases from the whole genome, comprising the steps of: Step 1, preprocessing and initialization of genome-wide SNP data According to the principle that the variation of any gene in the homologous chromosome alleles has the same impact on the disease, the SNP data is preprocessed into:

Where x _i ∈ {0, 1, 2, 3} ^d is the value of the corresponding site of SNP i: when the two alleles at the corresponding site are homozygous AA, it takes 1, and when it is homozygous aa, it takes 2, Take 3 for heterozygote Aa or aA, take 0 when the allele data on the corresponding locus is missing; y _i ∈ {1, 2} is the class label of sample _xi , 1 means the disease group, 2 means the control group, N is the number of samples in the SNP data, d is the number of SNPs in the data, and record the set of SNPs involved as Ω; Step 2, define the relevance measure The association I(Y; X) between a SNP factor X and a disease Y is defined as the mutual information MI(Y; X) between X and Y, expressed as formula (1): I(Y;X)＝H(Y)-H(Y|X) (1) In the formula $h\left(Y\right)=-\underset{\mathrm{the\; y}\&Element;\left\{\mathrm{0,1}\right\}}{\mathrm{\&Sigma;}}p\left(\mathrm{the\; y}\right)\log p\left(\mathrm{the\; y}\right)$ is the entropy, $h\left(Y\right|x)=-\underset{\mathrm{the\; y}\&Element;\left\{\mathrm{0,1}\right\}}{\mathrm{\&Sigma;}}\underset{x{\&Element;\{\mathrm{1,2,3},\}}^{\left|x\right|}}{\mathrm{\&Sigma;}}p\left(\mathrm{the\; y}\right|x)\log p\left(\mathrm{the\; y}\right|x)$ is the conditional entropy; Step 3, use the factor-based genetic search method "FGSA" to search for the SNP group in the SNP set Ω that is suspected to be the cause of the disease, so as to filter the SNPs that are not related to complex diseases; (3.1) Set genetic algorithm parameters and FGSA related parameters Set the number l of SNPs in the SNP group to be found. Set genetic algorithm parameters, including population size Nl, crossover probability P _c , mutation probability p _m , iteration number Iter _l ; set FGSA related parameters, including repeated search times Num _l , and the number of l-order interactions to be found each time M _l ; (3.2) Initialize the number of the first-order interactions to be found k=1; set the SNP set S ₁ =Φ related to the suspected cause of disease; set the SNP set to be investigated as Ω ^* =Ω; (3.3) Randomly initialize the population: Randomly generate l different effective SNP numbers from the SNP in Ω ^* , form an l-order factor as an individual, and generate a total of N _l individuals to form the population; (3.4) Calculation of fitness: For each individual in the population, calculate the mutual information according to formula (1) as the fitness of the individual; (3.5) Selection operation: According to the fitness value of each individual in the population, use roulette and elite strategy to select _N1 individuals; (3.6) Crossover operation: Randomly select two individuals from _N1 individuals, randomly select the intersection point, and exchange the part behind the intersection point for these two individuals with the crossover probability p _c to form two new individuals; (3.7) Mutation operation: For each individual, randomly generate an effective SNP number, and replace any SNP number in this individual according to the mutation probability p _m ; (3.8) Generate the population of the next generation: all individuals obtained after the operation of step (3.7) are used as the population of the next generation; (3.9) If the number of iterations is less than Iter _l , then jump to step (3.4); (3.10) Take the individual with the maximum fitness in the population and mark it as s _k and add it to the SNP set S _k suspected to be related to the disease, and remove this individual from the SNP set Ω ^* , that is

(3.11) Repeat steps (3.3)～(3.10) M _l times, add an individual to S _k each time, and remove this individual from the data, and obtain S _k through M _l repetitions; (3.12) Reset Ω ^* = Ω, repeat (3.2) ~ (3.11) Num _l times, get the SNP set $S_1,S_2,...,S_{{\mathrm{Num}}_l};$ (3.13) Output the SNP set containing each l-order interaction

Step 4, calculate the frequency of each SNP in v (4.1) Calculation of the frequency of each SNP: the number of occurrences of the SNP of the suspected cause of disease found in step (3) is used as the frequency of the SNP; (4.2) SNPs are sorted according to the frequency-association priority criterion, that is, the SNPs are sorted according to the principle of high frequency priority, and when the frequency is the same, the single SNP is associated with the disease, that is, the priority principle of mutual information is large; Step 5, outputting the top SNPs whose frequency is greater than the threshold.

Images