JP5242568B2 - Clustering method, program and apparatus - Google Patents

Description

  The present invention relates to a clustering method, program, and apparatus for clustering a plurality of data.

  Analyzing various phenomena and characteristics peculiar to samples derived from the same or different living bodies enables biologically and medically important evaluations. For example, in gene polymorphism analysis, the speed of the polymorphism discrimination reaction varies from sample to sample depending on the concentration of the sample and the presence or absence of an inhibitor. For this reason, in the gene polymorphism analysis, numerical data (polymorphism data) having a wide distribution is obtained.

  When the obtained numerical data is identified, the operator may perform clustering by visually observing a scatter diagram of the numerical data. However, when an operator identifies numerical data, the identification result may differ depending on the operator.

  Under such circumstances, various attempts have been made to automatically identify numerical data. For example, Patent Document 1 below discloses a technique that uses a statistical technique for a signal from a sample in gene polymorphism analysis. However, with this technique, numerical data corresponding to a low-frequency genetic polymorphism such that there are only a few samples in hundreds of samples is not statistically meaningful, and it is difficult to handle such numerical data. There was a problem.

  Therefore, a technique of incorporating a genetic statistical technique into a statistical technique in gene polymorphism analysis has also been disclosed (see, for example, Patent Document 2). In this technique, the reliability of numerical data obtained by genetic polymorphism analysis is genetically evaluated using the Hardy-Weinberg equilibrium.

JP 2004-272350 A JP 2006-107396 A

  However, when performing genetic polymorphism analysis incorporating a genetic statistical method, it is necessary to perform random sampling. For this reason, data obtained by biased sampling such as family samples and patient samples are not suitable for genetic statistical analysis. In addition, when a statistical method is used, a reliable statistic is not obtained when the polymorphism frequency is low, and the determination may be wrong.

  The present invention has been made in view of the above, and provides a clustering method, a program, and an apparatus capable of accurately performing numerical data clustering related to a sample regardless of how the sample is selected. Objective.

  In order to solve the above-described problems and achieve the object, the clustering method according to the present invention is such that a computer provided with storage means for storing a plurality of multidimensional numerical data stores one or more of the plurality of multidimensional numerical data. A data conversion step of reading the plurality of multidimensional numerical data from the storage means and converting each of the read multidimensional numerical data into lower dimensional numerical data A probability density function generating step for generating a plurality of probability density functions that give data existence probabilities corresponding to each of the numerical data converted in the data conversion step, and a plurality of probability density functions generated in the probability density function generating step A reliability distribution function that numerically determines the reliability of the plurality of multidimensional numerical data by taking a linear sum of A reliability distribution function generation step for generating a multi-dimensional numerical data cluster based on the reliability distribution function generated in the reliability distribution function generation step And

  Further, in the clustering method according to the present invention, in the above invention, the data conversion step uses a ratio value of different components of the multidimensional numerical data, so that each of the plurality of multidimensional numerical data has a dimension. It is characterized by being converted into numerical data one dimension lower.

  The clustering method according to the present invention is characterized in that, in the above invention, the number of dimensions of the multi-dimensional numerical data is 2, and the sum of the one-dimensional numerical data after the conversion in the data conversion step is 1. To do.

  In the clustering method according to the present invention, in the above invention, the probability density function generated in the probability density function generation step is a Gaussian function, and the average of the Gaussian function is calculated for each dimension of the two-dimensional numerical data of interest. The variance of the Gaussian function is 2 in a predetermined range from the two-dimensional numerical data of interest and the two-dimensional numerical data on a two-dimensional plane that gives a distribution of a plurality of two-dimensional numerical data. It is characterized by being determined using a distance from the dimension numerical data.

  The clustering method according to the present invention is characterized in that, in the above invention, the two-dimensional numerical data is detection data of an allele of a single nucleotide polymorphism, and the data converted in the data conversion step is an allyl concentration. And

  Further, in the clustering method according to the present invention, in the above invention, the cluster dividing step includes a reliability distribution function differentiation step for differentiating the reliability distribution function with respect to numerical data after being converted in the data conversion step, A minimum value calculation step for calculating a minimum value of the reliability distribution function from a value differentiated in the reliability distribution function differentiation step, and a minimum value feature for calculating a minimum value feature characterizing the minimum value calculated in the minimum value calculation step A quantity calculation step, and a cluster division position setting step for setting a cluster division position in a space in which the multidimensional numerical data is distributed using the minimum value feature quantity calculated in the minimum value feature quantity calculation step. And

  Further, the clustering method according to the present invention is characterized in that, in the above invention, the method further includes a cluster division result output step for outputting a cluster division result in the cluster division step.

  A clustering program according to the present invention causes the computer to execute a clustering method according to any one of the above inventions.

  A clustering apparatus according to the present invention is a clustering apparatus that divides a plurality of multidimensional numerical data into one or a plurality of clusters, and a storage unit that stores the plurality of multidimensional numerical data; and the plurality of multidimensional numerical data Data conversion means for converting each of the read multi-dimensional numerical data into lower-dimensional numerical data, and data corresponding to each of the numerical data converted by the data conversion means A function that generates a plurality of probability density functions that give probabilities and generates a reliability distribution function that numerically defines the reliability of the plurality of multidimensional numerical data by taking a linear sum of the generated plurality of probability density functions And a clustering unit that performs cluster division of the plurality of multidimensional numerical data based on a reliability distribution function generated by the generation unit and the function generation unit. Characterized by comprising a static splitting means.

  According to the present invention, each of a plurality of multidimensional numerical data is converted into lower dimensional numerical data, and a plurality of probability density functions that give data existence probabilities corresponding to each of the converted numerical data are generated, Generate a reliability distribution function that numerically defines the reliability of multiple multidimensional numerical data by taking the linear sum of multiple probability density functions, and then cluster multiple multidimensional numerical data based on this reliability distribution function By performing the division, it is possible to realize processing independent of the characteristics of the numerical data. Therefore, it is possible to accurately perform clustering of numerical data related to the sample regardless of how the sample is selected.

FIG. 1 is a block diagram showing a configuration of a clustering apparatus according to an embodiment of the present invention. FIG. 2 is a schematic diagram illustrating a schematic configuration of the measurement apparatus. FIG. 3 is a flowchart showing an outline of processing of the clustering method according to the embodiment of the present invention. FIG. 4 is a flowchart showing details of the data conversion process. FIG. 5 is a diagram illustrating a generation example (first example) of the reliability distribution function. FIG. 6 is a diagram illustrating a generation example (second example) of the reliability distribution function. FIG. 7 is a flowchart showing details of the cluster division processing. FIG. 8 is a diagram schematically showing the valley width and depth of the reliability distribution function. FIG. 9 is a diagram illustrating a display output example (first example) of the cluster division result. FIG. 10 is a diagram illustrating a display output example (second example) of the cluster division result.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Clustering apparatus 2,141 Transmission / reception part 3 Input part 4,104 Control part 5 Storage part 6 Output part 41 Data conversion part 42 Function generation part 43 Cluster division part 51 Measurement data storage part 52 Conversion data storage part 53 Function storage part 54 Cluster Division result storage unit 101 Measuring device 102 Microarray 103 Fluorescence detector M1, M2 Mountain V Valley

  The best mode for carrying out the present invention (hereinafter referred to as “embodiment”) will be described below with reference to the accompanying drawings. FIG. 1 is a diagram showing a configuration of a clustering apparatus according to an embodiment of the present invention. The clustering apparatus 1 shown in FIG. 1 is an apparatus that clusters a plurality of measurement data (numerical data) transmitted from the measurement apparatus 101, and is realized using a computer.

  The clustering device 1 is realized by a transmission / reception unit 2 that transmits / receives data to / from the measurement device 101, a keyboard, a mouse, and the like. A control unit 4 that controls the operation of the clustering apparatus 1, a storage unit 5 that stores information including measurement data and calculation results in the control unit 4, and a clustering result of measurement data obtained by calculation in the control unit 4. And an output unit 6 that outputs information including the output information.

  The control unit 4 includes a data conversion unit 41 that performs predetermined conversion on measurement data input from the measurement apparatus 101, a function generation unit 42 that generates a predetermined function using data converted by the data conversion unit 41, a function And a cluster division unit 43 that divides the measurement data into clusters using the function generated by the generation unit 42. The control unit 4 is realized using a CPU (Central Processing Unit) having an arithmetic function and a control function. The data conversion unit 41 constitutes at least part of the data conversion means. The function generation unit 42 constitutes at least a part of the function generation unit, and the cluster division unit 43 constitutes at least a part of the cluster division unit.

  The storage unit 5 includes a measurement data storage unit 51 that stores measurement data, a conversion data storage unit 52 that stores data obtained by converting measurement data, and a function storage unit 53 that stores functions generated using the converted data. A cluster division result storage unit 54 for storing the cluster division result for the measurement data. Such a storage unit 5 is a ROM (Read Only Memory) that stores in advance a clustering program according to the present embodiment, a program for starting a predetermined OS, and the like, and temporarily stores information used when the control unit 4 performs calculations. This is realized by using a RAM (Random Access Memory) or the like that stores the data, and constitutes at least a part of the storage means. The storage unit 5 may include an external storage device such as a hard disk.

  The output unit 6 has a function of generating an image based on a control signal from the control unit 4 and displaying the generated image, and is realized using a display such as liquid crystal, plasma, and organic EL.

  FIG. 2 is a schematic diagram illustrating a schematic configuration of the measurement apparatus 101. The measuring device 101 is a device that performs SNP typing to detect a single nucleotide polymorphism (SNP) of a gene. A microarray 102 in which a plurality of spots SP are formed on a substrate, and each spot of the microarray 102 A fluorescence detector 103 having a laser light source for irradiating the SP with laser light as excitation light and having a photomultiplier tube for detecting the intensity of fluorescence generated by the irradiated laser light, And a control unit 104 that controls the operation.

  The control unit 104 includes a transmission / reception unit 141 that transmits / receives information including measurement data to / from the clustering apparatus 1.

  A gene (probe) having a sequence complementary to a gene corresponding to a specific SNP of a sample is spotted on the spot SP of the microarray 102. Such a probe includes a probe for internal control serving as a reference for measurement data, and is spotted on a predetermined spot SP. Hereinafter, the probe for internal control is referred to as “hybrid control”. In addition, a spot SP in which probes other than the hybrid control are arranged is described as SPmn by labeling the SNP with m and labeling the sample with n (where m and n are natural numbers).

  When SNP typing is performed by the measuring apparatus 101, the allyl (allelic) cDNA generated from a sample (n) is expressed by fluorescent dyes Cy3 (indicated by white circles (◯) in FIG. 2) and Cy5 (indicated by black circles (●) in FIG. The labeled DNAs (tags) labeled in () are hybridized with the probes spotted on the spots SPmn of the microarray 102. Thereafter, the fluorescence detector 103 detects the intensity (signal luminance) of the fluorescence generated by the hybridization. When the fluorescence detector 103 detects a fluorescence signal corresponding to one of the fluorescent dyes, the allele of the SNP has homozygosity (○○ and ●● in FIG. 2). On the other hand, when the fluorescence detector 103 detects fluorescence signals corresponding to the two fluorescent dyes, the allele of the SNP has heterozygosity (（in FIG. 2).

  The control unit 104 uses the fluorescence generated from the spot SP detected by the fluorescence detector 103 to calculate the signal luminances corresponding to the fluorescent dyes Cy3 and Cy5, which are two-dimensional numerical data as measurement data, and this calculation The signal luminance thus transmitted is transmitted to the clustering apparatus 1 via the transmission / reception unit 141.

  Note that the details of the SNP typing method performed by the measuring apparatus 101 are essentially the same as the methods described in the following documents. N. Nishida, T. Tanabe, K. Hashido, K. Hirayasu, M. Takasu, A. Suyama, K. Tokunaga, "DigiTag assay for multipulex single nucleotide polymorphism typing with high success rate", Anal Biochem. 346 (2005) 281-288; N. Nishida, T. Tanabe, M. Takasu, A. Suyama, K. Tokunaga, "Further development of multipulex single nucleotide polymorphism typing method, the DigiTag2 assay", Anal Biochem. 364 (2007) 78-85 .

FIG. 3 is a flowchart showing an outline of processing of the clustering method according to the present embodiment. In the present embodiment, the following two points (1-1) and (1-2) are assumed regarding the reliability of measurement data when clustering is performed.
(1-1) Measurement data with high signal luminance is highly reliable.
(1-2) In the figure showing the distribution of measurement data of signal luminance, measurement data in which measurement data of another sample is distributed in the vicinity has high reliability.

  Under the above-mentioned assumption, the data conversion unit 41 of the clustering apparatus 1 reads the fluorescence signal luminance as measurement data stored in the measurement data storage unit 51, and converts the read signal luminance according to a predetermined rule ( Step S1).

In general, the signal luminance is the concentration of the sample introduced into the system, the concentration of the probe spotted on the microarray 102, the intensity of the laser light emitted by the fluorescence detector 103, and the sensitivity of the photomultiplier tube that the fluorescence detector 103 has. Due to the influence of the Therefore, in order to eliminate the influence of the measurement system described above, the data conversion unit 41 uses the signal luminance of the hybrid control as a reference, and the signal luminance of each spot SPmn is an amount that does not depend on the probe spotting amount or the labeled DNA concentration. Convert to In performing this data conversion processing, the following three points (2-1) to (2-3) are assumed.
(2-1) The signal luminance is proportional to the amount of labeled DNA spotted on the microarray 102.
(2-2) The luminous efficiency of the fluorescent dye depends only on the fluorescent dye and does not depend on the DNA sequence.
(2-3) The molar ratio of the hybrid control labeled DNAs labeled with the fluorescent dyes Cy3 and Cy5, respectively, is 1: 1.

と定義される。ここで、ｄは対応する蛍光色素の発光効率、ＳはスポットＳＰのプローブ点着量に比例する係数、Ｃは標識ＤＮＡの濃度である。以下、発光効率ｄについては、蛍光色素Ｃｙ３の発光効率をｄ_Cy3とし、蛍光色素Ｃｙ５の発光効率をｄ_Cy5とする。Based on the above assumptions (2-1) to (2-3), the signal intensity I of the fluorescence generated from the microarray 102 is

Is defined. Here, d is the luminous efficiency of the corresponding fluorescent dye, S is a coefficient proportional to the amount of spot spot spot SP, and C is the concentration of the labeled DNA. Hereinafter, regarding the luminous efficiency d, the luminous efficiency of the fluorescent dye Cy3 is d _Cy3 and the luminous efficiency of the fluorescent dye Cy5 is d _Cy5 .

と表される。ここで、標識ＤＮＡ濃度Ｃの添字中のＥＤ−１、ＥＤ−２は、蛍光色素Ｃｙ３、Ｃｙ５によってそれぞれ標識される標識ＤＮＡを識別するためのものである。A specific calculation performed by the data conversion unit 41 in step S1 will be described with reference to a flowchart shown in FIG. The data conversion unit 41 calculates the ratio of the light emission efficiency d _Cy3 of the fluorescent dye _Cy3 and the light emission efficiency d _Cy5 of the fluorescent dye Cy5 (light emission efficiency ratio) d _Cy3 / d _Cy5 to the signal luminance I _HybriContCy3 for each fluorescent dye of the hybrid control. _Obtained from I _HybriContCy5 (step S11). The signal intensities I _HybriContCy3 and I _HybriContCy5 for each fluorescent dye are _expressed by the following equation (1).

It is expressed. Here, ED-1 and ED-2 in the subscript of the labeled DNA concentration C are for identifying the labeled DNAs respectively labeled with the fluorescent dyes Cy3 and Cy5.

が得られ、蛍光色素Ｃｙ３と蛍光色素Ｃｙ５との発光効率比ｄ_Cy3／ｄ_Cy5が、測定データであるハイブリコントロールのシグナル輝度の比Ｉ_HybriContCy3／Ｉ_HybriContCy5を用いて表される。From the above assumption (2-3), the labeled DNA concentrations C _{HybriContED-1} and C _{HybriContED-2} are equal (C _{HybriContED-1} = C _{HybriContED-2} ). Therefore,

Is obtained, luminous efficiency ratio d _Cy3 / d _Cy5 with a fluorescent dye Cy3 and the fluorescent dye Cy5 are represented using the ratio I _HybriContCy3 / I _HybriContCy5 signal intensity of hybridization controls the measurement data.

と定義される。ここで、Ｓ_SNPmSAMPLEnはスポットＳＰmnのプローブ点着量、Ｃ_{SNPmSAMPLEnED-1}、Ｃ_{SNPmSAMPLEnED-2}は、スポットＳＰmnの蛍光色素Ｃｙ３、Ｃｙ５でそれぞれ標識した標識ＤＮＡの濃度である。Subsequently, the data conversion unit 41 corrects the signal luminance of the spot SPmn using the luminous efficiency ratio d _Cy3 / d _Cy5 of the corresponding fluorescent dye (step S12). The signal brightness I _{SNPmSAMPLEnCy3} and I _{SNPmSAMPLEnCy5} of the spot SPmn is

Is defined. Here, S _SNPmSAMPLEn is the spot spot amount of spot SPmn, and C _{SNPmSAMPLEnED-1} and C _{SNPmSAMPLEnED-2} are the concentrations of labeled DNAs labeled with fluorescent dyes Cy3 and Cy5 of spot SPmn, respectively.

と補正する。The data conversion unit 41 uses the luminous efficiencies d _Cy3 and d _Cy5 of the fluorescent dyes Cy3 and _Cy5 to obtain the signal luminances I _{SNPmSAMPLEnCy3} and I _{SNPmSAMPLEnCy5} .

And correct.

と再定義する。Subsequently, the data conversion unit 41 _redefines the sum of the corrected signal luminances I ′ _{SNPmSAMPLEnCy3} and I ′ _{SNPmSAMPLEnCy5} as a coefficient S ′ _SNPmSAMPLEn that is proportional to the probe spot amount of the spot SPmn (step S13). That is, the data conversion unit 41 calculates a coefficient proportional to the amount of probe spot landing on the spot SPmn.

And redefine.

と定義される。ここで、式（１０）、（１１）は、式（５）〜（９）を用いて導出される。例えば、式（１０）は、次のように変形することによって導出される。

この導出において、２番目の等号では式（９）を代入し、３番目の等号では式（７）、（８）を代入し、最後の等号では式（５）、（６）を代入した。Thereafter, the data conversion unit 41 uses the corrected signal luminance I ′ _{SNPmSAMPLEnCy3} and I ′ _{SNPmSAMPLEnCy5} and the labeled DNA concentration C _{HybriContED−} defined by using the coefficient S ′ _SNPmSAMPLEn proportional to the re-defined probe _spotting amount. ₁ , the correction value of C _{HybriContED-2} is calculated, and is written and stored in the conversion data storage unit 52 (step S14). Correction values C ′ _{SNPmSAMPLEnCy3} and C ′ _{SNPmSAMPLEnCy5} of the labeled DNA concentration calculated in step S14 are:

Is defined. Here, Expressions (10) and (11) are derived using Expressions (5) to (9). For example, equation (10) is derived by transforming as follows.

In this derivation, equation (9) is substituted for the second equal sign, equations (7) and (8) are substituted for the third equal sign, and equations (5) and (6) are substituted for the last equal sign. Substituted.

The corrected labeled DNA concentrations C ′ _{SNPmSAMPLEnCy3} and C ′ _{SNPmSAMPLEnCy5} are values normalized by the sum of the labeled DNA concentrations before correction, and correspond to the concentrations of allyl labeled with the fluorescent dyes Cy3 and Cy5, respectively. In this way, the data conversion unit 41 converts the two measurement data I _{SNPmSAMPLEnCy3} and I _{SNPmSAMPLEnCy5} at the spot SPmn into a one-dimensional amount that does not depend on the probe spot deposition amount or sample concentration of the spot SPmn.

を確率密度関数として生成し、関数記憶部５３に書き込んで記憶する。Next, the function generation unit 42 generates a probability density function that gives a data existence probability that the measurement data distributed in one dimension in step S1 described above exists as a true value (step S2). Specifically, the function generation unit 42 _reads the measurement data C ′ _{SNPmSAMPLEnCy3} and C ′ _{SNPmSAMPLEnCy5} converted into one dimension from the conversion data storage unit 52, and gives a normal distribution centered on the measurement points of each data.

Is generated as a probability density function, and is written and stored in the function storage unit 53.

と定義される量である。In equation (12), the coefficient I _SNPmSAMPLEn corresponding to the area of the Gaussian function is

It is an amount defined as

と定義される量である。ここで、Δθ^(k) _SNPmSAMPLEnは、注目しているサンプルとそのサンプルの近傍に位置するサンプルとの２次元平面（Ｉ_{SNPmSAMPLECy3}，Ｉ_{SNPmSAMPLECy5}）における角度差である。また、定数ａ^(k) _SNPxSAMPLEyは距離の平滑化に関わる数であり、１より大きい値として適宜定められる。なお、式（１４）では、近傍の３つのサンプルまでの角度差Δθ^(k) _SNPmSAMPLEnを用いて代表距離ｄ_SNPmSAMPLEnを算出している。In addition, the constant d _SNPmSAMPLEn corresponding to the variance of the Gaussian function in Expression (12) is from a sample of interest to a sample within a predetermined range on the two-dimensional plane (I _{SNPmSAMPLECy3} , I _{SNPmSAMPLECy5} ) indicating the signal luminance distribution. Is a quantity (representative distance) determined using the distance of

It is an amount defined as Here, Δθ ^(k) _SNPmSAMPLEn is an angle difference in a two-dimensional plane (I _{SNPmSAMPLECy3} , I _{SNPmSAMPLECy5} ) between the sample of interest and a sample located in the vicinity of the sample. The constant a ^(k) _SNPxSAMPLEy is a number related to distance smoothing and is appropriately determined as a value larger than 1. In Expression (14), the representative distance d _SNPmSAMPLEn is calculated using the angle difference Δθ ^(k) _SNPmSAMPLEn to three neighboring samples.

Further, r _SNPmSAMPLEn in the formula (12) is a concentration ratio C ′ _{SNPmSAMPLEnCy3} / C ′ _{SNPmSAMPLEnCy5} of the labeled DNA after correction for each fluorescent dye.

を算出し、関数記憶部５３に書き込んで記憶する（ステップＳ３）。図５および図６は、異なるサンプル、ＳＮＰの組み合わせに対する信頼性分布関数の生成例を示す図である。これらの図に示す信頼性分布関数Ｇ_SNPm（ｘ）においては、各サンプルに対するガウス関数ｇ_SNPmSAMPLEn（ｘ）が足し合わされ、複数の山と谷のピークが現れている。Thereafter, the function generator 42 is a function defined as the sum of the probability density functions of all samples for the same SNP as a reliability distribution function that numerically defines the reliability of the measurement data.

Is written and stored in the function storage unit 53 (step S3). FIG. 5 and FIG. 6 are diagrams showing examples of generation of reliability distribution functions for combinations of different samples and SNPs. In the reliability distribution function G _SNPm (x) shown in these figures, the Gaussian function g _SNPmSAMPLEn (x) for each sample is added, and a plurality of peaks and valleys appear.

Next, the cluster dividing unit 43 divides the cluster on the two-dimensional plane based on the reliability distribution function G _SNPm (x) read from the function storage unit 53 for each SNP (step S4). Details of the cluster division process will be described below with reference to the flowchart of FIG. First, the cluster dividing unit 43 obtains a numerical differentiation with respect to x of the reliability distribution function G _SNPm (x) (step S41).

Thereafter, the cluster dividing unit 43 calculates the minimum value of the reliability distribution function G _SNPm (x) using the result of Step S41 (Step S42).

Subsequently, the cluster dividing unit 43 calculates a minimum value feature amount that characterizes the minimum value calculated in step S42 (step S43). The minimum value feature here is the width and depth of the valley when the minimum value of the reliability distribution function G _SNPm (x) is the valley bottom. FIG. 8 is a diagram schematically showing the valley width and depth of the reliability distribution function G _SNPm (x). The width w of the valley V shown in the figure is the horizontal distance between the peaks of the mountains M1 and M2 adjacent to each other across the valley V. The depth p of the valley V is the average value (p1 + p2) of the height p1 of the mountain M1 seen from the valley bottom (minimum position) of the valley V and the height p2 of the mountain M2 seen from the valley bottom of the valley V. / 2.

を谷Ｖごとに求める。その後、クラスタ分割部４３は、式（１６）にしたがって求めた全ての谷Ｖの評価関数Ｑ_Vの中で所定の閾値Ｑ_thを超えているものの中から上位２つまでをクラスタ分割点として抽出する。なお、式（１６）で定義される谷Ｖの評価関数Ｑ_Vはあくまでも一例に過ぎず、谷Ｖの幅ｗや深さｐを用いて定義される関数であれば、式（１６）以外の関数でもかまわない。Next, the cluster division unit 43 sets a cluster division position using the minimum value feature amount calculated in step S43 (step S44). Specifically, the cluster dividing unit 43 determines the evaluation function defined using the width w and depth p of the valley V and the constant b.

For each valley V. After that, the cluster dividing unit 43 extracts, from the evaluation functions Q _V of all the valleys V obtained according to the equation (16), the upper two of the evaluation functions Q _V exceeding the predetermined threshold value Q _th as cluster dividing points. To do. Note that the evaluation function Q _V of the valley V defined by Expression (16) is merely an example, and any function other than Expression (16) can be used as long as the function is defined using the width w and depth p of the valley V. It can be a function.

Thereafter, the output unit 6 outputs the cluster division result in step S4 (step S5). FIGS. 9 and 10 are diagrams showing display output examples of cluster division results (divided into three clusters Cr1 to Cr3) for combinations of different samples and SNPs. 9 shows the result of clustering using the reliability distribution function G _SNPm (x) shown in FIG. FIG. 10 shows the result of clustering using the reliability distribution function G _SNPm (x) shown in FIG. 9 and 10, the three divided clusters Cr1 to Cr3 correspond to different polymorphic data.

Note that the cluster dividing unit 43 divides the cluster into two when there is only one valley V whose evaluation function Q _V exceeds the threshold value Q _th . In this case, the cluster dividing unit 43 belongs to a region between the straight line and the vertical axis or a straight line whose valley V is located on a two-dimensional plane with a straight line having the same value on the vertical axis and the horizontal axis as a boundary. And the type of polymorphic data to which the cluster belongs is determined.

  According to the clustering method described above, the cluster is divided without using a statistical value or a genetic statistical index. Therefore, the statistic regarding the SNP typing result and the reliability of the genetic statistical index are determined. Can be increased.

  According to the embodiment of the present invention described above, each of a plurality of multidimensional numerical data is converted into lower-dimensional numerical data, and a plurality of data existence probabilities corresponding to each of the converted numerical data are provided. Generate a probability density function, and generate a reliability distribution function that numerically defines the reliability of multiple multi-dimensional numerical data by taking the linear sum of the plurality of probability density functions. Based on this reliability distribution function Since a plurality of multidimensional numerical data are divided into clusters, processing independent of the characteristics of the numerical data can be realized. Therefore, it is possible to accurately perform clustering of numerical data related to the sample regardless of how the sample is selected.

  In addition, according to the present embodiment, when performing cluster division based on the reliability distribution function, the minimum value of the reliability distribution function and the minimum value feature amount (valley width and depth) that characterizes the minimum value are calculated. Since the cluster division position is set using the evaluation function defined using the calculated minimum value feature amount, the cluster division is not performed at a position that does not satisfy the predetermined condition. Therefore, it is not necessary to perform extra class ring for a data set in which the number of clusters to be divided is smaller than a normal data set.

  In the above-described embodiment, the SNP allele detection data is used as the multidimensional numerical data. However, the present invention is also applicable to a method of fractionating (or classifying) multidimensional numerical data. In addition, the present invention can be effectively applied to biological measurements in which a large number of numerical data varies.

  In the above-described embodiment, analysis is performed based on fluorescence signals from various samples immobilized or immobilized on the microarray. However, the present invention is not limited to beads other than the microarray or affinity columns. Widely applicable to Solid Phase Assays.

  Further, in the present invention, as a method not based on the solid phase assay, MAS (Magic Angle Spinning) using a mass spectrometry tag having a different molecular weight without using an optical label such as fluorescence as an identification tag. Such a classification method may be applied.

  In the present invention, in addition to fluorescence, a label related to light emission (chemiluminescence or bioluminescence), light absorption (colorimetric or turbidity), scattered light, or polarization may be applied as an optical label. Furthermore, depending on the object, electromagnetic energy such as radiation, magnetism, atomic force, electron beam, electromagnetic ultrasonic wave (EMAT), or the like may be used as a label.

  The present invention is also suitable for cell-based assays in which shape parameters such as various blood cells and somatic cells are optically or electromagnetically imaged and digitized by image analysis.

  In the clustering method according to the present invention, multidimensional numerical data can generally be converted into numerical data of a lower dimension in the data conversion step.

  Note that the clustering apparatus according to the present invention may be configured to be connected to the measurement apparatus through a communication network configured by an appropriate combination of the Internet, an intranet, a fixed telephone network, a mobile phone network, a leased line network, and the like.

  Further, the clustering program according to the present invention can be recorded on a computer-readable recording medium such as a hard disk, a flexible disk, a CD-ROM, a DVD-ROM, a flash memory, or an MO disk and widely distributed.

  Thus, the present invention can include various embodiments and the like not described herein, and various design changes and the like can be made without departing from the technical idea specified by the claims. It is possible to apply.

  The clustering method, program and apparatus according to the present invention are suitable for analyzing various phenomena and characteristics peculiar to samples derived from the same or different organisms, and particularly suitable for gene polymorphism analysis.

Claims (9)

A computer comprising storage means for storing a plurality of multidimensional numerical data is a clustering method for dividing the plurality of multidimensional numerical data into one or a plurality of clusters,
A data conversion step of reading the plurality of multidimensional numerical data from the storage means, and converting each of the read multidimensional numerical data into lower dimensional numerical data;
Probability density function generation step for generating a plurality of probability density functions giving data existence probabilities corresponding to each of the numerical data converted in the data conversion step;
A reliability distribution function generation step for generating a reliability distribution function that numerically defines the reliability of the plurality of multidimensional numerical data by taking a linear sum of the plurality of probability density functions generated in the probability density function generation step; ,
A cluster division step of performing cluster division of the plurality of multidimensional numerical data based on the reliability distribution function generated in the reliability distribution function generation step;
A clustering method characterized by comprising: The data conversion step includes
The clustering method according to claim 1, wherein each of the plurality of multidimensional numerical data is converted into numerical data whose dimension is one dimension lower by using a ratio value of different components of the multidimensional numerical data. . The number of dimensions of the multidimensional numerical data is 2,
The clustering method according to claim 1 or 2, wherein the sum of the one-dimensional numerical data after conversion in the data conversion step is 1. The probability density function generated in the probability density function generating step is a Gaussian function,
The average of the Gaussian function is determined by the ratio of each dimension of the two-dimensional numerical data of interest,
The variance of the Gaussian function is the distance between the focused two-dimensional numerical data and the two-dimensional numerical data within a predetermined range from the two-dimensional numerical data on a two-dimensional plane that gives a distribution of a plurality of two-dimensional numerical data. The clustering method according to claim 3, wherein the clustering method is defined using The two-dimensional numerical data is detection data for an allele of a single nucleotide polymorphism,
The clustering method according to claim 3 or 4, wherein the data converted in the data conversion step is an allyl concentration. The cluster dividing step includes:
A reliability distribution function differentiation step for differentiating the reliability distribution function with respect to the numerical data after being converted in the data conversion step;
A minimum value calculating step for calculating a minimum value of the reliability distribution function from the value differentiated in the reliability distribution function differentiation step;
A minimum value feature quantity calculating step for calculating a minimum value feature quantity characterizing the minimum value calculated in the minimum value calculation step;
A cluster division position setting step for setting a cluster division position in a space in which the multidimensional numerical data is distributed using the minimum value feature quantity calculated in the minimum value feature quantity calculation step;
The clustering method according to claim 1, wherein the clustering method includes:   The clustering method according to claim 1, further comprising a cluster division result output step for outputting a cluster division result in the cluster division step.   A clustering program that causes the computer to execute the clustering method according to claim 1. A clustering device that divides a plurality of multidimensional numerical data into one or a plurality of clusters,
Storage means for storing the plurality of multidimensional numerical data;
Data conversion means for reading the plurality of multidimensional numerical data from the storage means, and converting each of the read multidimensional numerical data into lower-dimensional numerical data;
A plurality of probability density functions that give data existence probabilities corresponding to each of the numerical data converted by the data converting means are generated, and the plurality of multidimensional numerical data are obtained by taking a linear sum of the generated probability density functions. Function generating means for generating a reliability distribution function that numerically defines the reliability of
Cluster dividing means for performing cluster division of the plurality of multidimensional numerical data based on the reliability distribution function generated by the function generating means;
A clustering apparatus characterized by comprising:

Images