Sub-cluster extraction device, sub-cluster extraction method, and program

The subcluster extraction method addresses the challenge of discovering hierarchical structures in high-dimensional data by iteratively analyzing feature data through principal component analysis and clustering, effectively extracting and refining clusters.

JP7595936B2Active Publication Date: 2024-12-09UNIV OF TSUKUBA
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2021109128
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2021-06-30
Publication Date
2024-12-09
Estimated Expiration
2041-06-30

AI Technical Summary

Technical Problem

Conventional cluster analysis methods struggle to discover hierarchical cluster structures in high-dimensional data, particularly in single-cell gene expression data, due to variations within clusters and lack of consideration for hierarchical structures.

Method used

A subcluster extraction method involving feature data acquisition, principal component analysis, feature selection, clustering, and subgroup selection, with iterative processes to identify and refine clusters based on sample space principal components.

Benefits of technology

The method effectively extracts hierarchical cluster structures by removing noise features and identifying subclusters, enhancing the discovery of deep hierarchical structures in high-dimensional data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007595936000001
    Figure 0007595936000001
  • Figure 0007595936000002
    Figure 0007595936000002
  • Figure 0007595936000003
    Figure 0007595936000003
Patent Text Reader

Abstract

To extract a hierarchical cluster structure.SOLUTION: A sub-cluster extraction apparatus includes: a feature quantity data acquisition unit which executes feature quantity data acquisition processing to acquire feature quantity data including, for each multiple samples, a set of values of feature quantities for samples; a feature quantity selection unit which executes feature quantity selection processing to select a feature quantity from among the feature quantities included in the feature quantity data; a clustering unit which executes clustering processing to classify the samples into clusters on the basis of the selected feature quantity; a sub-group selection unit which executes sub-group selection processing to select a cluster, as a sub-group, from among the classified clusters; and a repeating unit which repeats processing to execute the feature quantity data acquisition processing, the feature quantity selection processing, the clustering processing, and the sub-group selection processing, on the samples included in the selected sub-group.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a sub-cluster extraction device, a sub-cluster extraction method, and a program. [Background technology]

[0002] Cluster analysis, which aims to extract cluster structures in the background of data, plays an important role in various data analyses, including medical and life science data analyses. In real data, it is assumed that there is a hierarchical cluster structure in the background, but cluster structures in deep layers are difficult to discover with normal cluster analysis that uses all feature quantities.

[0003] In high-dimensional data such as single-cell gene expression data, which has become possible with recent advances in measurement equipment, large variations due to sample-to-sample variations and state change periods can appear even within the same cluster. For data with such characteristics, cluster analysis based on conventional statistical tests does not perform sufficiently well.

[0004] In addition, it is important to select features by removing features that have a common pattern across the entire sample group in order to extract a hierarchical cluster structure. A method for selecting features based only on feature statistics has been known as a conventional and widely used feature selection method (Non-Patent Document 1). In addition, the M3drop method is known as a feature selection method that uses the sparsity of features (Non-Patent Document 2). [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] "Computaional Biology", April 13, 2015, Volume 33, Issue 5, p.495-502 [Non-Patent Document 2] "Bioinformatics", December 24, 2018, Volume 35, Issue 16, 2019, p.2865-2867 [Non-Patent Document 3] “Gene Expression Omnibus”, [online], National Center for Biotechnology Information, [Retrieved April 16, 2021], Internet〈URL:https: / / www.ncbi.nlm.nih.gov / geo / 〉 [Non-Patent Document 4] “IGSR: The International Genome Sample Resource”, [online], EMBL-EBI, [accessed April 16, 2021], Internet〈URL: http: / / www.1000genomes.org〉 Summary of the Invention [Problem to be solved by the invention]

[0006] The method of selecting features from only feature statistics as described in Non-Patent Document 1 does not take into account fluctuations within a cluster. Also, the M3drop method as described in Non-Patent Document 2 does not take into account a hierarchical cluster structure. There is a demand for the ability to extract a hierarchical cluster structure.

[0007] The present invention has been made in consideration of the above-mentioned points, and provides a sub-cluster extraction device, a sub-cluster extraction method, and a program capable of extracting a hierarchical cluster structure. [Means for solving the problem]

[0008] The present invention has been made to solve the above-mentioned problems, and one aspect of the present invention is a feature data acquisition unit that executes a feature data acquisition process for acquiring feature data including a set of values ​​of a plurality of feature amounts for each of a plurality of samples, and a feature data acquisition method for acquiring the feature data acquired by the feature data acquisition unit. a principal component analysis is performed on a sample space which is a collection of a set of values ​​for each of the plurality of samples of the feature quantity, and the feature quantity data is calculated based on the result of the principal component analysis. The plurality of Speciala feature selection unit that executes a feature selection process which is a process of selecting features from among features; a clustering unit that executes a clustering process which is a process of classifying the samples into a plurality of clusters based on the features selected by the feature selection unit; a subgroup selection unit that executes a subgroup selection process which is a process of selecting, as a subgroup, a cluster from the plurality of clusters classified by the clustering unit; and a repetition unit that repeatedly executes a process of causing the feature data acquisition unit to execute the feature data acquisition process for samples included in the subgroup selected by the subgroup selection unit, causing the feature selection unit to execute the feature selection process, causing the clustering process to be executed by the clustering unit, and causing the subgroup selection unit to execute the subgroup selection process.

[0010] In one aspect of the present invention, in the sub-cluster extraction device, the sub-group selection section selects clusters as sub-groups from among the plurality of clusters on the basis of a predetermined feature amount.

[0011] According to another aspect of the present invention, there is provided a feature amount data acquisition step of acquiring feature amount data including a set of values ​​of a plurality of feature amounts for each of a plurality of samples, and a principal component analysis is performed on a sample space which is a collection of a set of values ​​for each of the plurality of samples of the feature quantity, and the feature quantity data is calculated based on the result of the principal component analysis. The plurality of Special a feature selection step of selecting a feature from among the features; a clustering step of classifying the samples into a plurality of clusters based on the feature selected by the feature selection step; a subgroup selection step of selecting a cluster as a subgroup from the plurality of clusters classified by the clustering step; and a repeat step of repeatedly executing the feature data acquisition step, the feature selection step, the clustering step, and the subgroup selection step for samples included in the subgroup selected by the subgroup selection step. The computer runs This is a sub-cluster extraction method.

[0012] Also, one aspect of the present invention is a method for making a computer include a feature amount data acquisition step of acquiring feature amount data including a set of values ​​of a plurality of feature amounts for each of a plurality of samples, and a method for acquiring the feature amount data acquired by the feature amount data acquisition step. a principal component analysis is performed on a sample space which is a collection of a set of values ​​for each of the plurality of samples of the feature quantity, and the feature quantity data is calculated based on the result of the principal component analysis. The plurality of Special a feature selection step of selecting a feature from among the features; a clustering step of classifying the samples into a plurality of clusters based on the feature selected by the feature selection step; a subgroup selection step of selecting, as a subgroup, a cluster from the plurality of clusters classified by the clustering step; and a repetition step of repeatedly executing the feature data acquisition step, the feature selection step, the clustering step, and the subgroup selection step for samples included in the subgroup selected by the subgroup selection step. Effect of the Invention

[0013] According to the present invention, a hierarchical cluster structure can be extracted. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 is a diagram illustrating an example of a configuration of a sub-cluster extraction system according to an embodiment of the present invention. [Diagram 2] FIG. 11 is a diagram illustrating an example of a sub-cluster extraction process according to the embodiment of the present invention. [Diagram 3] FIG. 11 is a diagram illustrating an example of a feature amount selection process according to the embodiment of the present invention. [Figure 4] FIG. 2 is a diagram showing an example of artificial data which is feature amount data according to the first embodiment of the present invention. [Diagram 5] FIG. 2 is a diagram showing an example of a dense portion of artificial data according to the first embodiment of the present invention; [Figure 6] FIG. 11 is a diagram showing an example of a visualized result of a clustering result according to the first embodiment of the present invention. [Figure 7] FIG. 13 is a diagram showing an example of a plot when the markers used in the visualized clustering result according to the first embodiment of the present invention are changed to labels (true clusters). [Figure 8] FIG. 11 is a diagram showing an example of a visualized result of the clustering process in the second iterative process according to the first embodiment of the present invention. [Figure 9] FIG. 11 is a diagram showing an example of a plot in which the markers used in the visualized clustering results from the clustering process in the second iterative process in the first embodiment of the present invention are changed to labels (true clusters). [Figure 10] FIG. 11 is a diagram showing an example of a visualized result of a clustering result according to the second embodiment of the present invention. [Figure 11] FIG. 13 is a diagram showing an example of a visualized result of the clustering process in the second iterative process according to the second embodiment of the present invention. [Figure 12] FIG. 13 is a diagram showing an example of a visualized result of the clustering result according to the third embodiment of the present invention. [Figure 13] FIG. 13 is a diagram showing an example of a visualized result of the clustering process in the second iterative process according to the third embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] (Embodiment) Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. Fig. 1 is a diagram showing an example of the configuration of a sub-cluster extraction system 1 according to this embodiment. The sub-cluster extraction system 1 extracts a cluster structure that exists behind high-dimensional feature data from the high-dimensional feature data. In the process of extracting the cluster structure, the sub-cluster extraction system 1 selects a feature from among a plurality of feature values.

[0016] In the following description, when a plurality of samples included in a cluster are further classified into a plurality of clusters, each of the plurality of clusters is referred to as a subcluster. All of the samples included in a subcluster are also referred to as a subgroup. A subcluster is also referred to as a subgroup. Therefore, a subgroup (subcluster) is each of a set of samples contained in one or more of a plurality of clusters obtained by classifying a plurality of samples into clusters, and further classifying the samples contained in the cluster into clusters.

[0017] Classifying multiple samples into clusters is also called extracting a cluster structure. Extracting a cluster structure is also called simply extracting clusters. When sub-clusters are extracted from a cluster, the cluster is said to have a hierarchical cluster structure. Selecting a feature from among a plurality of feature values ​​is also called feature removal or feature selection.

[0018] In the process of selecting features, the sub-cluster extraction system 1 performs principal component analysis in the sample space on the high-dimensional feature data, and selects features based on the sample space principal components that contribute to cluster separation of samples. In conventional multivariate analysis, analysis such as principal component analysis is performed on the feature space. In contrast, the sub-cluster extraction system 1 extracts sub-clusters while selecting features by capturing the relationships between multiple features in the sample space.

[0019] A sample space is a collection of feature values ​​for each of a plurality of samples of the feature, and is visualized by plotting points corresponding to the feature values ​​in a space having dimensions corresponding to each of the samples. It should be noted that the feature space used in conventional analysis is a collection of sets of feature values ​​for a plurality of samples.

[0020] [Functional configuration of sub-cluster extraction system 1] The sub-cluster extraction system 1 includes a sub-cluster extraction device 10, a feature amount data supply unit 20, and a presentation unit 30. The feature data supply unit 20 supplies high-dimensional feature data to the sub-cluster extraction device 10. The high-dimensional feature data is data including a set of values ​​of a plurality of feature quantities for each of a plurality of samples. Here, the dimension of a feature quantity means the number of feature quantities. High dimension means that the number of feature quantities is equal to or greater than a predetermined number (e.g., several thousand). In the following description, the high-dimensional feature data is simply referred to as feature data D. Note that the number of feature quantities included in the feature data D may be equal to or less than a predetermined number, and may be, for example, several to several hundred.

[0021] The feature data D is, for example, data in the form of a two-dimensional array consisting of rows and columns in which multiple feature values ​​are stored for each sample. In this array, for example, the rows correspond to samples and the columns correspond to features. Therefore, in the feature data D, for example, the cell in the i-th row and j-th column stores the value of the j-th feature of the i-th sample. The feature may be either a feature represented as a categorical variable or a feature represented by a numerical value. Hereinafter, a feature represented as a categorical variable is also referred to as a categorical feature, and a feature represented using a numerical value is also referred to as a numerical feature.

[0022] The feature data supply unit 20 may be, for example, an information storage device such as a server, or a human interface device such as a keyboard, tablet, or scanner.

[0023] The sub-cluster extraction device 10 includes a feature data acquisition unit 100, a preprocessing unit 101, a numerical feature normalization unit 102, a feature selection unit 103, a clustering unit 104, a subgroup selection unit 105, a repeating unit 106, and an output unit 107. The sub-cluster extraction device 10 is, for example, a personal computer (PC). Each functional unit included in the sub-cluster extraction device 10 is realized by a CPU (Central Processing Unit) reading a program from a ROM (Read Only Memory) and executing processing.

[0024] The feature amount data acquiring unit 100 acquires the feature amount data D supplied by the feature amount data supplying unit 20. The process of acquiring the feature amount data D is referred to as a feature amount data acquiring process. The preprocessing unit 101 performs preprocessing on the feature amount data D. A specific example of the preprocessing will be described later. The numerical feature quantity conversion / normalization unit 102 converts the pre-processed feature quantity data D into a numerical feature quantity and performs normalization processing. A specific example of the numerical feature quantity conversion / normalization processing will be described later.

[0025] The feature selection unit 103 selects a feature from among the multiple feature amounts included in the feature data D acquired by the feature data acquisition unit 100. In this embodiment, the feature selection unit 103 selects a feature from among the multiple feature amounts based on the sample space principal components. The process of selecting a feature from among the multiple feature amounts included in the feature data D is called a feature selection process.

[0026] The clustering unit 104 classifies the samples into a plurality of clusters based on the feature amounts selected by the feature amount selection unit 103. The process of classifying the samples into a plurality of clusters based on the feature amounts selected by the feature amount selection unit 103 is called a clustering process.

[0027] The subgroup selection unit 105 selects clusters as subgroups from among the multiple clusters classified by the clustering unit 104. The process of selecting clusters as subgroups from among the multiple clusters classified by the clustering unit 104 is called a subgroup selection process.

[0028] The repetition unit 106 repeatedly executes the processes of causing the feature data acquisition unit 100 to execute feature data acquisition processing for samples included in the subgroup selected by the subgroup selection unit 105, causing the feature selection unit 103 to execute feature selection processing, causing the clustering unit 104 to execute clustering processing, and causing the subgroup selection unit 105 to execute subgroup selection processing.

[0029] The output unit 107 outputs the subgroup extraction result R to the presentation unit 30. The subgroup extraction result R is information indicating a subgroup selected from the samples.

[0030] The presentation unit 30 presents the subgroup extraction result R output from the output unit 107 included in the sub-cluster extraction device 10 by presentation means such as display or printing. The presentation unit 30 is, for example, a display or a printer. The presentation unit 30 may be a storage device such as a network server. In this case, the presentation unit 30 stores the subgroup extraction result R output from the output unit 107, and supplies the stored subgroup extraction result R to another device.

[0031] [Operation of the sub-cluster extraction device 10] Next, a sub-cluster extraction process in which the sub-cluster extraction device 10 extracts sub-clusters will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example of the sub-cluster extraction process according to the present embodiment.

[0032] Step S10: The feature amount data acquiring unit 100 acquires the feature amount data D supplied by the feature amount data supplying unit 20. The feature amount data acquiring unit 100 supplies the acquired feature amount data D to the pre-processing unit 101.

[0033] Step S20: The preprocessing unit 101 performs preprocessing on the feature data D supplied from the feature data acquisition unit 100. If a feature value corresponding to a sample included in the feature data D is missing, the preprocessing unit 101 interpolates the value by missing value processing. Alternatively, if a feature value corresponding to a sample included in the feature data D is missing, the preprocessing unit 101 removes the sample from the feature data D. Through these processes, the feature data D becomes data with no missing feature values.

[0034] In addition, the preprocessing unit 101 reduces the dimensions of the features included in the feature data D based on a feature reduction method according to the field to which the sub-cluster extraction system 1 is applied. For example, in the field of single-cell gene expression analysis, the preprocessing function FindVariableFeatures provided by Seurat is used. The preprocessing unit 101 supplies the preprocessed feature amount data D to a numerical feature amount normalization unit 102 .

[0035] Step S30: The numerical feature normalization unit 102 performs a process of converting the preprocessed feature data D into a numerical feature and normalization. In the process of converting the numerical feature, the numerical feature normalization unit 102 converts categorical features into numerical features for the preprocessed feature data D. The numerical feature normalization unit 102 uses, for example, one-hot encoding or label encoding in the process of converting categorical features into numerical features. The numerical feature normalization unit 102 also performs a normalization process on the feature data D. The numerical feature quantity normalization unit 102 supplies the feature quantity data D that has been subjected to the numerical feature quantity and normalization processes to a feature quantity selection unit 103.

[0036] Step S40: The feature selection unit 103 performs a feature selection process on the feature data D supplied from the numerical feature normalization unit 102. The feature selection unit 103 uses the feature selection process to remove features that are common to all samples included in the feature data D. The feature selection unit 103 uses the feature selection process to remove features that are common to all samples so that M features remain. The feature selection unit 103 selects the remaining M features. M is a predetermined number and is specified in advance by the user.

[0037] Details of the feature selection process will now be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the feature selection process according to this embodiment. In this embodiment, as an example, the feature selection unit 103 removes features common to all samples included in the feature data D based on a principal component analysis in the sample space. The processes of steps S110 to S130 shown in FIG. 3 are executed as the process of step S40 shown in FIG.

[0038] In the following description, principal components obtained by principal component analysis in the sample space are referred to as sample space principal components P. The sample space principal components P include as many principal components as the number of dimensions of the sample space, and are referred to as the first principal component, second principal component, etc.

[0039] Step S110: The feature quantity selection unit 103 performs principal component analysis in the sample space on the feature quantity data D supplied from the numerical feature quantity normalization unit 102. The feature quantity selection unit 103 generates sample space principal components P as a result of the principal component analysis.

[0040] Step S120: The feature selection unit 103 determines whether or not there is a distortion in the distribution of the sample space principal components P supplied from the feature selection unit 103. The feature selection unit 103 performs this determination on the principal components included in the sample space principal components P in order from the first principal component.

[0041] In this embodiment, the skewness of the distribution of the sample space principal components P refers to, for example, a deviation of the distribution from a normal distribution. The feature selection unit 103 performs the determination based on skewness, for example. The feature selection unit 103 determines the deviation of the distribution of the sample space principal components P from a normal distribution. When the skewness of the distribution of the sample space principal components P deviates from 0 by a predetermined value, the feature selection unit 103 determines that the distribution is skewed.

[0042] The feature quantity selection unit 103 may make a determination based on kurtosis instead of skewness. The feature quantity selection unit 103 may make a determination based on an arithmetic mean or a standard deviation. The feature quantity selection unit 103 may also make a determination based on a combination of one or more of the arithmetic mean, standard deviation, skewness, and kurtosis.

[0043] In the present embodiment, an example has been described in which the feature selection unit 103 determines that the distortion of the distribution of the sample space principal component P is a deviation of the distribution from a normal distribution, but the present invention is not limited to this. The feature selection unit 103 may make the determination based on the similarity between the distribution of the sample space principal component P and an asymmetric distribution. In this case, when the distribution of the sample space principal component P and the asymmetric distribution are not similar, the feature selection unit 103 determines that the distribution of the sample space principal component P is not distorted. An asymmetric distribution is, for example, a distribution that is not linearly symmetric about a central value.

[0044] The feature selection unit 103 supplies the result of the determination of the distortion of the distribution of the sample space principal components P to the feature selection unit 103. In the feature selection process shown in Fig. 3, it is assumed, as an example, that the first principal component to the Nth principal component of the sample space principal components P are determined to have distortion. In other words, it is assumed that the N+1th and subsequent principal components of the sample space principal components P are determined to have no distortion.

[0045] Step S130: The feature selection unit 103 selects a feature from among the multiple features included in the feature data D, based on the sample space principal components P supplied from the feature selection unit 103 and the determination result supplied from the feature selection unit 103. Here, the feature selection unit 103 selects a feature that is a large distance from the origin of the sample space for the sample space principal components P that are determined to have no distortion in their distribution.

[0046] The feature quantity selection unit 103 selects feature quantities that are far from the origin of the sample space by using K components from the N+1th principal component to the N+Kth principal component that are determined to have no distribution distortion. In other words, the feature quantity selection unit 103 selects feature quantities that are farther from the origin of the sample space than a predetermined distance in a K-dimensional subspace corresponding to the N+1th principal component to the N+Kth principal component of the sample space. Here, the feature quantity selection unit 103 selects M feature quantities in descending order of distance from the origin of the sample space.

[0047] K is an integer equal to or greater than 0. M is an integer equal to or greater than 1. The values ​​of K and M are supplied from the feature data supply unit 20 to the sub-cluster extraction device 10 together with the feature data D. For example, the values ​​of K and M are values ​​specified by a user and supplied from the feature data supply unit 20. Note that the feature selection unit 103 may use, as the values ​​of K and M, predetermined values ​​based on a cluster structure expected for the features depending on, for example, the field to which the sub-cluster extraction system 1 is applied.

[0048] The distance of the feature from the origin is, for example, the Euclidean distance. However, a distance other than the Euclidean distance may be used as the distance of the feature from the origin. The feature amount selecting unit 103 may select feature amounts whose distance from the origin of the sample space is greater than a predetermined distance, without setting an upper limit on the number of feature amounts to be selected.

[0049] Here, features that are a short distance from the origin in the sample space are considered to be features that do not contribute to cluster separation of samples and are equivalent to noise. The distribution of the features tends to follow a normal distribution. The sub-cluster extraction device 10 removes features that are equivalent to noise from the feature data D by selecting features that are a long distance from the origin based on the sample space principal components P.

[0050] The feature amount selection unit 103 supplies a feature amount selection result indicating the selected M feature amounts to the clustering unit 104 .

[0051] As described above, the feature selection unit 103 performs principal component analysis on the feature data D in a sample space, which is a collection of multiple feature values ​​of multiple feature samples, and selects a feature from among the multiple feature values ​​based on the results of the principal component analysis.

[0052] 3, an example has been described in which the feature selection unit 103 selects features based on principal components determined to have no distribution distortion among the sample space principal components P, but the present invention is not limited to this. For example, the feature selection unit 103 may remove a predetermined number of principal components starting from the first principal component from the sample space principal components P. That is, in this embodiment, the number N is determined based on the distribution distortion among the sample space principal components P, but a predetermined number may be used as the number N.

[0053] In the feature selection process shown in Fig. 3, an example has been described in which features common to all samples included in the feature data D are removed based on principal component analysis in the sample space, but the present invention is not limited to this. As the feature selection process, the feature selection unit 103 may remove features based on feature statistics, for example, using the feature selection method described in Non-Patent Document 1. In addition, as the feature selection process, the feature selection unit 103 may remove features based on sparsity of features, for example, using the feature selection method described in Non-Patent Document 2.

[0054] Furthermore, in the feature selection process, the feature selection unit 103 may remove at least one feature from among the features included in the feature data D as a feature having a pattern common to all the samples included in the feature data D. For example, the feature selection unit 103 may remove any one or more feature(s) whose statistics, such as variance, for the samples included in the feature data D are equal to or less than a predetermined value.

[0055] Returning to FIG. 2, the sub-cluster extraction process will be continued. Step S50: The clustering unit 104 classifies the samples into a plurality of clusters based on the M features selected by the feature selection unit 103. The clustering unit 104 performs a clustering process based on a known clustering method. As the known clustering method, either a hierarchical clustering method or a non-hierarchical clustering method may be used.

[0056] Step S60: The subgroup selection unit 105 selects a cluster as a subgroup from among the multiple clusters classified by the clustering unit 104. In this way, the subgroup selection unit 105 selects a subset (subgroup) from the entire sample set included in the feature amount data D. The subgroup selection unit 105 may select multiple subgroups.

[0057] In this embodiment, the subgroup selection unit 105 selects clusters as subgroups from among a plurality of clusters based on a predetermined feature. For example, the subgroup selection unit 105 selects one or more clusters corresponding to one or more predetermined feature. The one or more predetermined feature is set by a user specifying in advance one or more feature in which the user is interested.

[0058] The subgroup selection unit 105, for example, calculates statistics for each of the one or more predetermined features for each cluster, and selects a cluster in which each of the calculated statistics has a value greater than a reference value. The statistics is, for example, an average value. The reference value is determined for each of the one or more predetermined features. In the following description, the predetermined feature is also referred to as a marker feature.

[0059] Note that the subgroup selection unit 105 may arbitrarily select a subgroup from among the multiple clusters classified by the clustering unit 104. Arbitrarily selecting a subgroup includes, for example, randomly selecting a subgroup from among the multiple clusters. Moreover, the subgroup selection unit 105 may select a subgroup designated by the user.

[0060] Step S70: The repetition unit 106 determines whether or not a termination condition is satisfied. Here, the termination condition is, for example, that the number of repetitions, which is the number of times the processes from step S10 to step S60 have been repeated since the start of the sub-cluster extraction process, is equal to or greater than a predetermined number. The predetermined number is a number equal to or greater than 2. The predetermined number may be specified in advance by the user, or may be stored in advance in the sub-cluster extraction device 10. Here, the process from step S10 to step S60 is called a repetitive process. The repetitive process is a unit of processing that is repeatedly executed in the sub-cluster extraction process.

[0061] The termination condition may be that a predetermined numerical value is equal to or less than a threshold value. For example, the termination condition may be that the number of features included in the feature data D at the time when the process of step S60 is completed is equal to or less than a threshold value.

[0062] If the repetition section 106 determines that the end condition is satisfied (step S70; YES), the sub-cluster extraction device 10 executes the process of step S80. On the other hand, when it is determined that the termination condition is not satisfied (step S70; NO), the repetition unit 106 executes the repetition process for the samples included in one or more subgroups selected in the subgroup selection process as new samples as a whole. That is, the repetition unit 106 makes the feature data acquisition unit 100, the preprocessing unit 101, the numerical feature normalization unit 102, the feature selection unit 103, the clustering unit 104, and the subgroup selection unit 105 execute the processes from step S10 to step S60 again.

[0063] Here, the repetition unit 106 supplies, as new feature amount data D, to the feature amount data acquisition unit 100, feature amount data for samples included in one or more subgroups selected in the last executed subgroup selection process (step S60).

[0064] The repetition unit 106 selects features to be included in the new feature data D as follows. For example, the repetition unit 106 causes the new feature data D to include the feature selected by the last executed feature selection process (step S30). Furthermore, the repetition unit 106 may include, for example, a feature selected by a selection method other than the feature selection process in the new feature data D. As the selection method other than the feature selection process, for example, a known feature selection method such as a filter method or a wrapper method may be used. Alternatively, the repetition unit 106 may include the features included in the current feature data D as they are in the new feature data D without reducing the number of features.

[0065] The repetition unit 106 repeatedly executes the repetition process until a termination condition is satisfied. That is, the repetition unit 106 repeatedly executes the process of causing the feature data acquisition unit 100 to execute feature data acquisition processing for samples included in the subgroup selected by the subgroup selection unit 105, causing the feature selection unit 103 to execute feature selection processing, causing the clustering unit 104 to execute clustering processing, and causing the subgroup selection unit 105 to execute subgroup selection processing, for the samples included in the subgroup selected by the subgroup selection unit 105.

[0066] Step S80: The output unit 107 outputs the subgroup extraction result R to the presentation unit 30. The output unit 107 outputs, as the subgroup extraction result R, information indicating one or more clusters obtained in the last executed clustering process (step S50). With this, the sub-cluster extraction device 10 ends the sub-cluster extraction process.

[0067] In the sub-cluster extraction process shown in FIG. 2, the pre-processing (step S20) and the numerical feature conversion and normalization processes (step S30) may be omitted depending on the values ​​and / or types of features included in the feature data D. In that case, depending on the value and / or type of feature included in the feature data D, the pre-processing unit 101 and the numerical feature normalization unit 102 may be omitted from the configuration of the sub-cluster extraction device 10 shown in FIG.

[0068] 2, in the first iteration of the sub-cluster extraction process, the clustering process (step S50) may be omitted if information for dividing samples included in the feature data D into clusters is provided. For example, if the feature data D contains information (such as a label) indicating a cluster for each sample from the beginning, the clustering process (step S50) may be omitted even if the clustering process is not executed, because the samples included in the feature data D have already been divided into multiple clusters. In the second and subsequent iterative processes, a clustering process is required to further classify the samples included in the subgroups selected in the last subgroup selection process (step S60) into clusters.

[0069] An example in which the sub-cluster extraction system 1 according to this embodiment is applied will be described below. [First Example] In the first embodiment, artificial data D1, which is artificially generated data, is used as the feature data D. FIG. 4 is a diagram showing an example of the artificial data D1, which is the feature data D according to this embodiment. The artificial data D1 stores 4600 feature values ​​for each of 1500 samples. In the artificial data D1, six cluster structures are assumed, and the distribution of the features is given a sparse and dense distribution. As shown in FIG. 4, the dense portion is given as a plurality of rectangles in the range from the 1st to the 700th feature. Background noise is given to the portion other than the dense portion, which is the portion other than the plurality of rectangles.

[0070] FIG. 5 is a diagram showing an example of a dense portion of the artificial data D1 according to this embodiment. In FIG. 5, the range of the 1st to 700th feature amounts of the artificial data D1 shown in FIG. 4 is enlarged. In FIG. 5, the boundaries of (true) clusters assumed in this embodiment are shown using dashed lines. In the following description, the clusters are referred to as "Label 1", "Label 2", ..., "Label 6" in ascending order of sample numbers. In order to distinguish them from clusters extracted by the sub-cluster extraction device 10, the name "label" is used for the true clusters. It can be seen that a common sparse-dense pattern is given to each sample in the same label.

[0071] The details of the sub-cluster extraction process performed by the sub-cluster extraction device 10 of this embodiment will be described below in association with the above-mentioned process of FIG. In the feature amount data acquisition process of step S10, the feature amount data acquisition unit 100 acquires, as the feature amount data D, the artificial data D1 described above. The pre-processing in step S20 is omitted because the artificial data D1 has no missing features.

[0072] In the process of converting the feature quantities into numerical values ​​and normalization in step S30, the feature quantities included in the artificial data D1 in this embodiment are all numerical feature quantities, so that conversion into numerical feature quantities is not necessary. The numerical feature quantity conversion normalization unit 102 converts the feature quantity values ​​by normalization processing so that the average value of the feature quantity becomes 0 and the standard deviation becomes 1.

[0073] In the feature selection process of step S40, the feature selection unit 103 selects 200 feature amounts based on a principal component analysis in the sample space.

[0074] In the clustering process in step S50, clustering was performed on the 200 features selected by the feature selection process. As the clustering process, the clustering unit 104 performed clustering based on the Louvain method for the principal components obtained by the principal component analysis of the feature space. As a result of the clustering process, five clusters were obtained. Note that in the Louvain method, the number of clusters obtained is automatically determined in the algorithm.

[0075] The visualization of the clustering results is shown in Figure 6. Figure 6 shows plots corresponding to each sample of the results of applying two-dimensional unsupervised dimensionality reduction (UMAP) in the feature space for the principal components. In Figure 6, the clustering results are visualized by associating the types of markers showing the plots with the clusters of the Louvain method on the two-dimensional plane obtained by the dimensionality reduction.

[0076] Moreover, Fig. 7 shows plots when the markers used in Fig. 6 are changed to labels (true clusters). Comparing Fig. 6 and Fig. 7, the plots corresponding to "Label 5" and "Label 6" in Fig. 7 are grouped into one cluster, "Cluster 4", by the Louvain method in Fig. 6. Therefore, it can be considered that there is a further hierarchical cluster structure under "Cluster 4". It can be seen that the cluster structure was not extracted in the first sub-cluster extraction process.

[0077] In the subgroup selection process in step S60, the subgroup selection unit 105 first selects a marker feature. The marker feature is designated in advance as a feature in which the user is interested. In this embodiment, it is assumed that the 157th feature out of the 200 features is selected as the marker feature. The subgroup selection unit 105 calculates the average value of the marker features in each cluster, and selects 300 samples belonging to "Cluster 4" for which the average value is 0.3 or more as a subgroup.

[0078] The repetition unit 106 again executes the repetition process from step S10 to step S60. The repetition unit 106 supplies the artificial data D1 for the 300 samples belonging to "Cluster 4" selected in the first repetition process to the feature data acquisition unit 100 as new feature data D. Here, the repetition unit 106 causes all the features included in the original artificial data D1 before the start of the sub-cluster extraction process to be included in the new feature data D.

[0079] In the second iterative process, the pre-processing of step S20 is also unnecessary and is therefore omitted. In the numerical feature conversion and normalization process of step S30, the numerical feature conversion normalization unit 102 converts the feature values ​​by normalization processing so that the average value of the feature values ​​becomes 0 and the standard deviation becomes 1, as in the first iteration of the process. In the feature selection process of step S40, the feature selection unit 103 selected 300 features based on principal component analysis in the sample space.

[0080] A visualization of the clustering results from the clustering process in the second iteration (step S50) is shown in Figure 8. The numbers in the legend indicate the cluster numbers using the Louvain method. In Figure 6, which shows the visualization of the clustering results from the clustering process in the first iteration, the group of samples that was a single mass known as "Cluster 4" is separated into two clusters in Figure 8. It can be seen that a cluster structure with a deeper hierarchy has been obtained that was not obtained in the first iteration.

[0081] Figure 9 shows the plot when the markers used in Figure 8 are changed to labels (true clusters). The numbers in the legend indicate the label numbers. Figure 9 shows that "Label 5" and "Label 6" are regarded as different clusters by the Louvain method.

[0082] [Second Example] In the second example, human gene expression data D2 published in Non-Patent Document 3 is used as the feature data D. The gene expression data D2 stores 6713 feature values ​​for each of 4296 samples.

[0083] Gene expression data is data in which each feature is the expression level of a specific (single) gene. In gene expression data, samples correspond to cells. In gene expression data, sample groups are classified into abnormal cell groups and normal cell groups, and are used in research to discover genes that are specifically over- or under-expressed in abnormal cell groups.

[0084] The details of the sub-cluster extraction process performed by the sub-cluster extraction device 10 of this embodiment will be described below in association with the above-mentioned process of FIG. In the feature amount data acquisition process of step S10, the feature amount data acquisition unit 100 acquires, as the feature amount data D, the gene expression data D2 described above.

[0085] In the preprocessing of step S20, the preprocessing unit 101 performs preprocessing on 6713 features included in the gene expression data D2. In the preprocessing of step S20, a preprocessing function FindVariableFeatures provided by Seurat is applied. Through the preprocessing, the preprocessing unit 101 reduces the 6713 features included in the gene expression data D2 to 2000 features.

[0086] In the process of converting the feature quantities into numerical values ​​and normalization in step S30, the feature quantities included in the artificial data D1 in this embodiment are all numerical feature quantities, so that conversion into numerical feature quantities is not necessary. The numerical feature quantity conversion normalization unit 102 converts the feature quantity values ​​by normalization processing so that the average value of the feature quantity becomes 0 and the standard deviation becomes 1.

[0087] In the feature selection process of step S40, the feature selection unit 103 selects 200 feature amounts based on a principal component analysis in the sample space.

[0088] In the clustering process in step S50, clustering was performed on the 200 features selected by the feature selection process. As the clustering process, the clustering unit 104 performed clustering based on the Louvain method for the principal components obtained by the principal component analysis of the feature space. As a result of the clustering process, 10 clusters were obtained. As described above, in the Louvain method, the number of clusters obtained is automatically determined in the algorithm.

[0089] The visualization of the clustering results is shown in Figure 10. Figure 10 shows plots corresponding to each sample of the results of applying two-dimensional unsupervised dimensionality reduction (UMAP) in the feature space for the principal components. In Figure 10, the clustering results are visualized by associating the types of markers showing the plots with the clusters of the Louvain method on the two-dimensional plane obtained by the dimensionality reduction.

[0090] In the subgroup selection process of step S60, the subgroup selection unit 105 first selects a marker feature. The marker feature is designated in advance as a feature of interest to the user. In this embodiment, it is assumed that the feature (gene) "PSAP" is selected as the marker feature from among 200 features. The subgroup selection unit 105 calculates the average value of the marker feature in each cluster, and selects samples belonging to clusters "2, 4, 6, 7" where the average value is 0.3 or more as subgroups.

[0091] The repetition unit 106 again executes the repetition process from step S10 to step S60. The repetition unit 106 supplies the gene expression data D2 for the samples belonging to the clusters "2, 4, 6, 7" selected in the first repetition process to the feature data acquisition unit 100 as new feature data D. Here, the repetition unit 106 causes all the features included in the original gene expression data D2 before the start of the sub-cluster extraction process to be included in the new feature data D.

[0092] In the numerical feature conversion and normalization process of step S30, the numerical feature conversion normalization unit 102 converts the feature values ​​by normalization processing so that the average value of the feature values ​​becomes 0 and the standard deviation becomes 1, as in the first iteration of the process. In the feature selection process of step S40, the feature selection unit 103 selected 200 features based on principal component analysis in the sample space.

[0093] A visualization of the clustering results from the clustering process in the second iteration (step S50) is shown in Fig. 11. The numbers in the legend indicate the cluster numbers using the Louvain method. In Fig. 10, which shows the visualization of the clustering results from the clustering process in the first iteration, the group of samples that was once a single cluster known as cluster "2, 4, 6, 7" is separated into five clusters in Fig. 11. It can be seen that a deep hierarchical cluster structure has been obtained that was not obtained in the first iteration.

[0094] [Third Example] In the third embodiment, genotype data D3 based on whole genome sequencing published in Non-Patent Document 4 is used as the feature data D. The genotype data D3 stores 20 million feature values ​​for each of 2500 samples.

[0095] Genotype data is data that represents the difference between the bases of each gene locus and the reference genome. Genotype data is used in research to classify samples (for example, humans) into disease groups and non-disease groups and to discover genetic mutations that appear specifically in the disease group. The genotype data D3 used in this embodiment is not data for two groups related to diseases, and in this embodiment, an analysis is performed focusing on the genetic origin of ancestors for the results of unsupervised dimensionality reduction.

[0096] The details of the sub-cluster extraction process performed by the sub-cluster extraction device 10 of this embodiment will be described below in association with the above-mentioned process of FIG. In the feature amount data acquisition process of step S10, the feature amount data acquisition unit 100 acquires, as the feature amount data D, the above-mentioned genotype data D3.

[0097] In the preprocessing of step S20, the preprocessing unit 101 performs preprocessing on 20 million features included in the genotype data D3. Here, the preprocessing unit 101 removes features that are missing in 20 percent or more of the samples. The preprocessing unit 101 removes samples that are missing in 20 percent or more of the features. The preprocessing unit 101 removes features that are missing in 2 percent or more of the samples. The preprocessing unit 101 removes samples that are missing in 2 percent or more of the features. The preprocessing unit 101 removes features with a minor allele frequency of 5 percent or less.

[0098] In the process of converting the feature values ​​to numerical features and normalizing the feature values ​​in step S30, since all the features included in the genotype data D3 are categorical features called genotypes, the numerical feature conversion normalization unit 102 converts the categorical features into numerical features using label encoding. Note that in this embodiment, the feature values ​​are not normalized.

[0099] In the feature selection process of step S40, the feature selection unit 103 selects 50,000 features based on principal component analysis in the sample space. The clustering process in step S50 is omitted in this embodiment.

[0100] A visualization result of the clustering result is shown in FIG. 12. FIG. 12 shows plots corresponding to each sample as a result of applying two-dimensional unsupervised dimensionality reduction on the feature space for the principal components. FIG. 12 shows the clustering result visualized on a two-dimensional plane obtained by the dimensionality reduction, using markers indicating race. Here, the color of each marker indicates a population such as European ancestry, and the shape of the marker indicates the subpopulation therein. Unlike the above-mentioned Examples 1 and 2, in this Example, the color and shape of the markers use information given in the original data set, genotype data D3. As shown in Figure 12, the plots form clusters for each population, and it can be seen that a two-dimensional space capturing the characteristics of the samples has been obtained.

[0101] In the subgroup selection process of step S60, the subgroup selection unit 105 selects the subgroup CL1 of AFR (African descent, orange) at the top left in FIG.

[0102] The repetition unit 106 again executes the repetition process, which is the process from step S10 to step 60. The repetition unit 106 supplies the genotype data D3 for the samples belonging to the subgroup CL1 selected in the first repetition process to the feature data acquisition unit 100 as new feature data D. Here, the repetition unit 106 causes the new feature data D to include only the features selected in the feature selection process (step S40).

[0103] In the second iterative process, the process of converting the data into numerical feature quantities and normalization in step S30 is omitted. In the feature selection process of step S40, the feature selection unit 103 selected 50 features based on principal component analysis in the sample space.

[0104] A visualization of the clustering results from the clustering process (step S50) in the second iterative process is shown in Fig. 13. The shape of the markers indicates each sub-cluster. In Fig. 12, which shows the visualized results of the clustering process in the first iteration, the group of samples that was a single mass as subgroup CL1 is separated into five clusters in Fig. 13. It can be seen that a deep hierarchical cluster structure that was not obtained in the first iteration has been obtained.

[0105] [summary] As described above, the sub-cluster extraction device 10 according to this embodiment includes the feature data acquisition unit 100, the feature selection unit 103, the clustering unit 104, the subgroup selection unit 105, and the repeat unit . The feature amount data acquiring unit 100 executes a feature amount data acquiring process which is a process for acquiring feature amount data D including a set of values ​​of a plurality of feature amounts for each of a plurality of samples. The feature amount selection unit 103 executes a feature amount selection process, which is a process of selecting a feature amount from among a plurality of feature amounts included in the feature amount data D acquired by the feature amount data acquisition unit 100. The clustering unit 104 executes a clustering process, which is a process of classifying samples into a plurality of clusters based on the features selected by the feature selection unit 103 . The subgroup selection unit 105 executes a subgroup selection process for selecting clusters as subgroups from among the multiple clusters classified by the clustering unit 104 . The repetition unit 106 repeatedly executes a process (in this embodiment, a repetition process) of causing the feature data acquisition unit 100 to execute a feature data acquisition process for samples included in a subgroup selected by the subgroup selection unit 105, causing the feature selection unit 103 to execute a feature selection process, causing the clustering unit 104 to execute a clustering process, and having the subgroup selection unit 105 execute a subgroup selection process.

[0106] With this configuration, the sub-cluster extraction device 10 according to this embodiment can remove features that are common to all of the samples and identify subgroups that are the basis of sub-clusters, thereby extracting a hierarchical cluster structure. A hierarchical cluster structure is also called a cluster structure in a deep hierarchy.

[0107] Moreover, in the sub-cluster extraction device 10 according to this embodiment, the feature selection unit 103 performs principal component analysis on the feature data in a sample space, which is a collection of multiple feature values ​​for each of multiple sample features, and selects a feature from among the multiple feature values ​​based on the results of the principal component analysis.

[0108] With this configuration, the sub-cluster extraction device 10 according to the present embodiment can remove features that become noise and select main features based on the results of principal component analysis in sample space, so that features that become noise can be removed more efficiently than in the case where principal component analysis in sample space is not used when removing features that act commonly across a plurality of samples.

[0109] Furthermore, in the sub-cluster extraction device 10 according to this embodiment, the subgroup selection unit 105 selects clusters as subgroups from among a plurality of clusters on the basis of a predetermined feature amount (marker feature amount).

[0110] With this configuration, the sub-cluster extraction device 10 according to this embodiment can extract sub-clusters based on a predetermined feature amount (marker feature amount), and can therefore detect sub-groups in which the predetermined feature amount (marker feature amount) is active.

[0111] In addition, a part of the sub-cluster extraction device 10 in the above-mentioned embodiment, for example, the feature data acquisition unit 100, the preprocessing unit 101, the numerical feature normalization unit 102, the feature selection unit 103, the clustering unit 104, the subgroup selection unit 105, the repeat unit 106, and the output unit 107 may be realized by a computer. In that case, a program for realizing this control function may be recorded in a computer-readable recording medium, and the program recorded in the recording medium may be read into a computer system and executed to realize the control function. In addition, the "computer system" here refers to a computer system built into the sub-cluster extraction device 10, and includes hardware such as an OS and peripheral devices. In addition, the "computer-readable recording medium" refers to a portable medium such as a flexible disk, an optical magnetic disk, a ROM, a CD-ROM, and a storage device such as a hard disk built into a computer system. Furthermore, the "computer-readable recording medium" may include a device that dynamically holds a program for a short time, such as a communication line when a program is transmitted via a network such as the Internet or a communication line such as a telephone line, and a device that holds a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. Furthermore, the above program may be for realizing part of the functions described above, and may further be capable of realizing the functions described above in combination with a program already recorded in the computer system. In addition, a part or all of the sub-cluster extraction device 10 in the above-mentioned embodiment may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each functional block of the sub-cluster extraction device 10 may be individually processed, or a part or all of the blocks may be integrated and processed. The integrated circuit method is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. In addition, when an integrated circuit technology that replaces LSI appears due to the progress of semiconductor technology, an integrated circuit based on that technology may be used.

[0112] Although one embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design changes, etc. are possible within the scope that does not deviate from the gist of the present invention. [Explanation of symbols]

[0113] 10...Sub-cluster extraction device, 100...Feature amount data acquisition unit, 103...Feature amount selection unit, 104...Clustering unit, 105...Subgroup selection unit, 106...Repetition unit, D...Feature amount data

Claims

1. a feature data acquisition unit that executes a feature data acquisition process for acquiring feature data including a set of values ​​of a plurality of feature amounts for each of a plurality of samples; a feature selection unit that performs a principal component analysis on the feature data acquired by the feature data acquisition unit in a sample space that is a collection of a plurality of feature values ​​of the plurality of samples of the feature data, and selects a feature from among the plurality of feature data based on a result of the principal component analysis; a clustering unit that executes a clustering process for classifying the samples into a plurality of clusters based on the feature amounts selected by the feature amount selection unit; a subgroup selection unit that executes a subgroup selection process for selecting a cluster as a subgroup from the plurality of clusters classified by the clustering unit; a repetition unit that repeatedly executes a process of causing the feature data acquisition unit to execute the feature data acquisition process for samples included in the subgroup selected by the subgroup selection unit, causing the feature selection unit to execute the feature selection process, causing the clustering process to be executed by the clustering unit, and causing the subgroup selection unit to execute the subgroup selection process; A sub-cluster extraction device comprising:

2. The subgroup selection unit selects a cluster as a subgroup from among the plurality of clusters based on a predetermined feature amount. The sub-cluster extracting apparatus according to claim 1 .

3. a feature data acquisition step of acquiring feature data including a set of values ​​of a plurality of feature amounts for each of the plurality of samples; a feature amount selection step of performing a principal component analysis on the feature amount data acquired by the feature amount data acquisition step in a sample space that is a collection of a set of values ​​for each of the plurality of samples of the feature amount, and selecting a feature amount from among the plurality of feature amounts included in the feature amount data based on a result of the principal component analysis; a clustering step of classifying the samples into a plurality of clusters based on the features selected by the feature selection step; a subgroup selection step of selecting a cluster as a subgroup from the plurality of clusters classified by the clustering step; a repeating step of repeatedly executing the feature data acquisition step, the feature selection step, the clustering step, and the subgroup selection step for samples included in the subgroup selected by the subgroup selection step; The sub-cluster extraction method is performed by a computer.

4. On the computer, a feature data acquisition step of acquiring feature data including a set of values ​​of a plurality of feature amounts for each of the plurality of samples; a feature amount selection step of performing a principal component analysis on the feature amount data acquired by the feature amount data acquisition step in a sample space that is a collection of a set of values ​​for each of the plurality of samples of the feature amount, and selecting a feature amount from among the plurality of feature amounts included in the feature amount data based on a result of the principal component analysis; a clustering step of classifying the samples into a plurality of clusters based on the features selected by the feature selection step; a subgroup selection step of selecting a cluster as a subgroup from the plurality of clusters classified by the clustering step; a repeating step of repeatedly executing the feature data acquisition step, the feature selection step, the clustering step, and the subgroup selection step for samples included in the subgroup selected by the subgroup selection step; A program for executing the above.

Citation Information

Patent Citations

  • Method and device for cluster classification

    JP1997265529A

  • Device and method for preparing similar pattern category discrimination dictionary

    JP1999203415A

  • Method and device for inspection

    JP2006098155A

  • Sample attribution class estimation method, sample attribution class estimation program, and attribution class estimation apparatus

    WO2010016109A1