Data processing device and method, program, and recording medium

Abstract

Provided are a data processing device, a data processing method, a program, and a recording medium that make it possible to effectively reduce noise mixed in with feature amounts when a data scale differs between feature amounts. The data processing device processes first data having two or more feature amounts and including noise in each feature amount by means of a processor. The processor acquires the first data, classifies the two or more feature amounts into clusters using the first data, generates second data in which the scale of the feature amounts is normalized using the first data, creates new feature amounts by aggregating the classified feature amounts belonging to the same clusters in the second data using classified cluster classification information, and generates third data having the created new feature amounts.

Description

Data processing device and method, program, and recording medium

[0001] The present invention relates to a data processing device and method, a program, and a recording medium, and more particularly to a technique for reducing noise mixed into data.

[0002] All data in the world is mixed with noise. This makes it difficult to detect the signals you want to detect and has a negative impact on the analysis that uses the data and the accuracy of predictions made by machine learning models. Therefore, research is underway into methods to reduce and remove noise.

[0003] A commonly used method for noise reduction is to average the features contained in the original data and create new data having new features with less noise.

[0004] This method has problems such as insufficient noise reduction when there is a difference in the magnitude of noise between the features to be averaged, and the possibility that the amount of information in the original features may be significantly reduced.

[0005] Omics data is an example of data in which noise in multiple feature quantities is not uniform. Omics data will be described below.

[0006] First of all, the genes present in a living organism are collectively called the genome, and data obtained by measuring the genome is called genomic data. Furthermore, by relating this concept to a series of phenomena occurring within the organism, it is possible to expand on this concept, and in addition to the genome, other terms have been proposed, such as the epigenome, which is a collective term for acquired regulatory factors that do not involve changes in gene sequence; the transcriptome, which is a collective term for gene transcription products; and the proteome, which is a collective term for proteins. In recent years, the idea of ​​generalizing these "general terms for biological substances that exist at the same level in life phenomena" and calling them omics / omics data has emerged.

[0007] When analyzing data in various industrial fields, noise is likely to be a problem in many situations, so the development of methods that can reduce noise would be of great benefit to each industrial field.

[0008] Patent Literature 1 describes an analysis device that uses analysis target data including a plurality of feature quantities to construct a regression model in which a response variable, which is one of the plurality of feature quantities, is obtained from an explanatory variable corresponding to one of the plurality of feature quantities, calculates the similarity between the explanatory variable and each of the plurality of feature quantities other than the feature quantities corresponding to the response variable and the explanatory variable, and extracts similar feature quantities whose similarity is higher than a predetermined value.

[0009] Non-Patent Document 1 describes taking a weighted average when aggregating feature amounts to reduce noise.

[0010] JP 2018-151883 A

[0011] "Denoising large-scale biological data using network filters" Kavran and Clauset BMC Bioinformatics (2021) 22:157<https: / / doi.org / 10.1186 / s12859-021-04075-x>

[0012] The invention described in Patent Document 1 does not take measures against noise, and therefore is not expected to be effective for data with a low signal-to-noise ratio (S / N ratio).

[0013] The noise removal method described in Non-Patent Document 1 does not take into account cases where the data scales of features are different. Therefore, when aggregating features with different data scales, there is a high possibility that the weighting of the information of the feature with the larger data scale will be greater, which may result in the loss of important information.

[0014] The present invention has been made in consideration of the above circumstances, and aims to provide a data processing device, method, program, and recording medium that can effectively reduce noise mixed into feature quantities when the data scales between feature quantities are different.

[0015] In order to achieve the above object, a first aspect of the invention is a data processing device that processes first data, which has two or more features and each feature contains noise, using a processor. The processor acquires the first data, uses the first data to classify the two or more features into clusters, uses the first data to generate second data in which the scales of the features are normalized, uses cluster classification information to aggregate features that belong to the same cluster into which the second data has been classified to create new features, and generates third data having the newly created features.

[0016] According to a first aspect of the present invention, first data having two or more feature quantities, each of which contains noise, are used to generate second data from the first data by normalizing the scale of the feature quantities. The two or more feature quantities are classified into clusters, and new feature quantities are created by aggregating feature quantities of the second data belonging to the same cluster using the classification cluster classification information, and third data is generated having the new feature quantities. Because the new feature quantities of the third data thus generated are generated using second data in which the data scale between feature quantities is normalized and by aggregating feature quantities belonging to the same cluster, even if the data scale between feature quantities of the first data is different, the information of each feature quantity of the first data is effectively reflected and noise is reduced.

[0017] In the data processing device according to the second aspect of the present invention, in the first aspect, it is preferable that the normalization of the features is performed by normalizing the scale of the features belonging to the same cluster using the cluster classification information. In order to aggregate the features belonging to the same cluster, it is only necessary to normalize the data scale between the features belonging to the same cluster, and it is possible to realize normalization that is not affected by the data scale of the features belonging to different clusters.

[0018] In a data processing device according to a third aspect of the present invention, in the first or second aspect, it is preferable that, when aggregating features belonging to the same cluster, the processor aggregates the features using the magnitude of noise in the features belonging to the same cluster.

[0019] A data processing device according to a fourth aspect of the present invention is any of the first to third aspects, wherein the processor, when aggregating feature quantities belonging to the same cluster, preferably aggregates the feature quantities by taking a weighted average based on the inverse of the magnitude of noise in the feature quantities belonging to the same cluster. This makes it possible to achieve sufficient noise reduction while retaining the characteristics of the feature quantities even when the noise in each feature quantity is not uniform, and enables more precise weighted averaging based on the magnitude of the noise.

[0020] In a data processing device according to a fifth aspect of the present invention, in any one of the first to fourth aspects, it is preferable that the noise be defined by a coefficient of variation, variance, or standard deviation of measured values ​​of a plurality of samples whose true values ​​of the feature are known.

[0021] In the data processing device according to a sixth aspect of the present invention, in any one of the first to fifth aspects, the samples for defining noise are preferably samples with known degrees of methylation or samples from a control group.

[0022] In a data processing device according to a seventh aspect of the present invention, in any of the first to sixth aspects, it is preferable that the processor calculates a similarity between two or more features using the first data, and classifies the features into clusters using a cluster determination threshold and the calculated similarity.

[0023] In a data processing device according to an eighth aspect of the present invention, in the seventh aspect, when first data is input, the processor uses a first machine learning model that classifies features into clusters, and it is preferable that the first machine learning model optimizes the cluster determination threshold by using a cluster determination threshold as a hyperparameter of the machine learning task and feeding back task accuracy.

[0024] A data processing device according to a ninth aspect of the present invention is the data processing device of the seventh or eighth aspect, wherein the first data is composed of a plurality of samples, and the first data for calculating the similarity between feature quantities is preferably a sample of a processing group.

[0025] A tenth aspect of the present invention is a data processing device according to any of the seventh to ninth aspects, wherein the first data is composed of a plurality of samples, and the processor uses the first data to determine whether the samples are outliers for each feature, and further classifies the clusters into two or more small clusters based on the similarities between two or more feature values ​​using the first data excluding the samples determined to be outliers for each cluster classified based on the similarity between the features, and preferably uses the first data excluding the samples and the cluster classification information of the classified small clusters instead of the first data and the cluster classification information.

[0026] According to the tenth aspect of the present invention, by clustering that takes outliers into consideration and aggregating features based on the clusters, it is possible to reduce redundancy of features. That is, by using clustering that is robust against outliers, only features with truly high similarity are aggregated, so only redundancy is reduced and the amount of information is not significantly reduced.

[0027] In an eleventh aspect of the present invention, in any one of the first to tenth aspects, the data processing device is preferably configured such that the first data is composed of a plurality of samples, and the processor uses the first data to determine whether the samples are outliers for each feature, and uses the first data and the outlier determination results to exclude from the first data any samples determined to be outliers for one or more feature.

[0028] A data processing device according to a twelfth aspect of the present invention is preferably the data processing device of the eleventh aspect, wherein the outlier determination of a sample is performed using the Smirnoff-Grubbs test, the k-nearest neighbor method, or the Local Outlier Factor and the deviation from the characteristic value of the feature.

[0029] A data processing device according to a thirteenth aspect of the present invention is the eleventh or twelfth aspect, wherein the processor uses a second machine learning model that, when inputted with first data, outputs an outlier determination result or outputs the first data excluding samples determined to be outliers, and it is preferable that the second machine learning model uses a threshold indicating the degree of deviation of the feature value from the characteristic value as a hyperparameter of the machine learning task, and optimizes the threshold indicating the degree of deviation by feeding back the task accuracy.

[0030] A fourteenth aspect of the present invention is the data processing device of any one of the eleventh to thirteenth aspects, wherein the first data on which outlier determination is performed is preferably a sample of a processing group.

[0031] A fifteenth aspect of the present invention relates to a data processing device according to any one of the first to fourteenth aspects, wherein the first data is composed of a plurality of samples, and the processor, when normalizing the scales of features belonging to the same cluster, preferably normalizes the scales of the first features of the plurality of samples so that the scales of the first features of the plurality of samples and the scales of second features different from the first features of the plurality of samples are aligned.

[0032] A data processing device according to a 16th aspect of the present invention is any of the first to fifteenth aspects, wherein the first data is composed of a plurality of samples, and the processor, when normalizing the scales of features belonging to the same cluster, preferably normalizes the scales so that the median or average of the scales of the first feature of the plurality of samples matches the median or average of the scales of a second feature of the plurality of samples that is different from the first feature.

[0033] A seventeenth aspect of the present invention is directed to a data processing device according to the sixteenth aspect, wherein the samples for which the median or mean is calculated are preferably samples from a control group.

[0034] In the data processing device according to an eighteenth aspect of the present invention, in any one of the first to seventeenth aspects, it is preferable that the first data is omics data, DNA methylation data, or cell-free DNA methylation data in blood.

[0035] A data processing device according to a 19th aspect of the present invention is any of the 1st to 18th aspects, wherein the first data includes a control group and a treatment group, and the control group and the treatment group preferably have different features.

[0036] In a data processing device according to a twentieth aspect of the present invention, in any of the first to nineteenth aspects, it is preferable that the processor replaces the newly created features with the original features of the first data to create third data with a reduced number of features.

[0037] In a data processing device according to a 21st aspect of the present invention, in any one of the 7th to 10th aspects, it is preferable that the processor performs a variable transformation on the first data before calculating the similarity between the features of the first data.

[0038] A data processing device according to a twenty-second aspect of the present invention is the twenty-first aspect, wherein the variable transformation of the first data is preferably a logarithmic transformation or a square root transformation, which can mitigate the influence of outliers.

[0039] A twenty-third aspect of the invention is a data processing method for processing first data, which has two or more feature quantities and each feature quantity contains noise, by a processor, the data processing method including the steps of: acquiring the first data by the processor; classifying the two or more feature quantities into clusters using the first data by the processor; generating second data by the processor by normalizing the scales of the feature quantities by the processor; and using cluster classification information to create new feature quantities by aggregating feature quantities belonging to the same cluster into which the second data has been classified, and generating third data having the newly created feature quantities.

[0040] A twenty-fourth aspect of the present invention is a data processing method according to the twenty-third aspect, wherein the normalization of the feature amounts is preferably performed by normalizing the scale of feature amounts belonging to the same cluster using cluster classification information.

[0041] In a data processing method according to a 25th aspect of the present invention, in the 23rd or 24th aspect, when aggregating features belonging to the same cluster, the processor preferably aggregates the features using the magnitude of noise in the features belonging to the same cluster.

[0042] A twenty-sixth aspect of the invention is a program for causing a computer to execute the data processing method of any one of the twenty-third to twenty-fifth aspects.

[0043] A twenty-seventh aspect of the invention is a non-transitory computer-readable recording medium on which the program of the twenty-sixth aspect is recorded.

[0044] According to the present invention, even when the data scales of the feature quantities of the original data are different, it is possible to effectively reflect the information of each feature quantity of the original data and reduce noise.

[0045] FIG. 1 is a graph showing the relationship between the true methylation degree values ​​of various methylation markers and the measured methylation degree values ​​(measured values ​​with noise width). FIG. 2(A) shows DNA derived from normal cells, FIG. 2(B) shows DNA derived from cancer cells with abnormal DNA methylation, and FIG. 2(C) is a diagram conceptually illustrating the ratio of DNA derived from normal cells to DNA derived from cancer cells in blood. FIG. 3 is a conceptual diagram of a chart in which the vertical axis represents the number of samples and the horizontal axis represents the number of methylation markers (number of features). FIG. 4 is a diagram showing one method for reducing noise and feature redundancy. FIG. 5 is a diagram showing a method for deleting features that have a small contribution to the target variable. FIG. 6 is a block diagram showing an embodiment of the hardware configuration of a data processing device according to the present invention. FIG. 7 is a functional block diagram showing a first embodiment of a data processing device according to the present invention. FIG. 8 is a functional block diagram showing a second embodiment of a data processing device according to the present invention. FIG. 9 is a functional block diagram showing a third embodiment of a data processing device according to the present invention. FIG. 10 is a conceptual diagram of data obtained when DNA methylation data samples are subjected to a process of aggregating features belonging to the same cluster using the noise magnitude of each feature. FIG. 11 is a functional block diagram showing a fourth embodiment of a data processing device according to the present invention. FIG. 12 is a diagram showing methylation markers of DNA methylation data in a marker feature space. FIG. 13 is a functional block diagram showing a fifth embodiment of a data processing device according to the present invention. FIG. 14 is a diagram showing specific numerical examples when normalizing the scale of features A, B, and C belonging to the same cluster and when weighting averages of features A, B, and C for noise reduction. FIG. 15 is a functional block diagram showing a sixth embodiment of a data processing device according to the present invention. FIG. 16 is a conceptual diagram showing the case where samples determined to be outliers by outlier determination are excluded from the original data. FIG. 17 is a diagram showing an embodiment of data processing such as outlier testing for feature m. FIG. 18 is a diagram showing a cluster determination threshold th r , th r ', the p-value threshold th for outlier detection p , absolute value threshold th d FIG. 19 is a diagram illustrating two-stage clustering. FIG. 20 is a diagram illustrating cell-free DNA methylation data consisting of N samples and M features. FIG. 21 is a diagram illustrating a cluster set obtained when the cell-free DNA methylation data shown in FIG. 20 is cluster-classified. FIG. 22 is a diagram illustrating a cluster set obtained when the features belonging to each cluster shown in FIG. 21 are reclustered. FIG. 23 is a diagram illustrating weights for each feature belonging to the cluster with cluster number 0 shown in FIG. 22. FIG. 24 is a diagram illustrating the scaling ratios for each feature belonging to cluster C0′ with cluster number 0 shown in FIG. 22. FIG. 25 is a diagram illustrating some of the data having new features for each cluster obtained by weighted averaging. FIG. 26 is a diagram illustrating other data having new features for each cluster obtained by weighted averaging. FIG. 27 is a diagram illustrating data obtained by quantile transformation of the features shown in FIG. 25. Fig. 28 is a chart showing data obtained by quantile transforming the feature values ​​shown in Fig. 26. Fig. 29 shows the cluster determination threshold th r , th r ' , p-value threshold th p , absolute value threshold th d Fig. 30 is a chart showing quantile-transformed data of isolated features that cannot be classified into the 11 clusters shown in Fig. 22 from the cell-free DNA methylation data shown in Fig. 20 .

[0046] Hereinafter, preferred embodiments of a data processing device, method, program, and recording medium according to the present invention will be described with reference to the accompanying drawings.

[0047] [Source of the Invention] When developing a system that supports early cancer diagnosis based on DNA (deoxyribonucleic acid) methylation information in blood, the following problems arise.

[0048] (1) Influence of Noise (Low S / N Ratio) The measured value of the methylation level of the methylation marker (feature amount) of a sample contains noise.

[0049] FIG. 1 is a graph showing the relationship between the true methylation degree values ​​of various methylation markers and the measured methylation degree values ​​(measured values ​​with noise width).

[0050] As shown in Figure 1, for example, measurements of methylation markers with a methylation level of 0% and 1% cannot be distinguished due to noise. This is because the S / N ratio of these measurements is too low. Data with such a low S / N ratio has the problem of adversely affecting the analysis using the data and the prediction accuracy of machine learning models.

[0051] Figure 2(A) shows DNA derived from normal cells, and Figure 2(B) shows DNA derived from cancer cells in which abnormal DNA methylation has occurred. Figure 2(C) is a conceptual diagram showing the ratio of DNA derived from normal cells to DNA derived from cancer cells in the blood. As shown in Figure 2(C), the ratio of DNA derived from cancer cells that leaks into the blood is expected to be extremely low, at around 0.1% to 1% of the total.

[0052] (2) Impact of excessive number of markers (high redundancy) As shown in Figure 2, the proportion of DNA derived from cancer cells is small, and it takes a lot of effort to obtain a large number of samples. However, when the number of samples is small, a lack of redundancy in features results in AI (artificial intelligence) that is robust to unknown data.

[0053] FIG. 3 is a conceptual diagram of a chart in which the vertical axis represents the number of samples and the horizontal axis represents the number of methylation markers (number of features). In the example shown in FIG. 3, the number of samples is 280 and the number of methylation markers is 700.

[0054] Data with 700 markers for 280 samples as shown in Figure 3 does not allow for sufficient learning due to the excessive number of markers (resulting in an AI that is not robust).

[0055] Therefore, a method that can (1) reduce noise and (2) reduce feature redundancy is desirable even with a low S / N ratio and a small amount of data.When considering use as part of a machine learning model, it is desirable to adjust it so that optimal noise reduction and feature redundancy reduction can be performed according to the purpose of learning.

[0056] FIG. 4 is a diagram showing one technique for reducing noise and feature redundancy.

[0057] In Figure 4, for data with two or more feature quantities, the similarity between the feature quantities is calculated, and clustering is performed so that the average similarity between the feature quantities within a cluster is maximized. Clustering can be performed using an index called Modularity.

[0058] In the example shown in FIG. 4A, six feature amounts are classified into a cluster having three feature amounts on the left side and a cluster having three feature amounts on the right side based on the similarity between the feature amounts.

[0059] Then, the feature amounts within the same cluster are simply averaged to generate new data (new data) having new feature amounts (new feature amounts) as shown in FIG. 4(B).

[0060] The new data has reduced noise in the new features, and the number of new features (number of markers) has been reduced to reduce feature redundancy.

[0061] On the other hand, since the scales of the new features of the new data are not consistent among the features belonging to the same cluster, the weighting of the features with larger scales increases, which may result in the loss of important information.In addition, the new features of the new data are averaged regardless of the noise level of each feature belonging to the same cluster, which is not considered to be efficient noise reduction.

[0062] Another method for reducing feature redundancy is to delete features that have a small contribution to the objective variable, as shown in FIG.

[0063] The six feature amounts shown in FIG. 5A are sorted according to the magnitude of their contribution to the objective variable.

[0064] The four features shown in Fig. 5(B) are obtained by deleting the two lowest-ranked features that have a small contribution to the objective variable from the six features shown in Fig. 5(A). This reduces the number of features and reduces feature redundancy, but deleting features that have a small contribution to the objective variable reduces the amount of information, and deleting highly independent features has a negative impact on task prediction (generalization) performance.

[0065] [Hardware Configuration of Data Processing Device] FIG. 6 is a block diagram showing an embodiment of the hardware configuration of a data processing device according to the present invention.

[0066] The data processing device 100 shown in FIG. 6 is configured by, for example, a personal computer, a workstation, or the like, and includes a processor 110 , a memory 120 , a display 130 , an input / output interface 140 , and an operation unit 150 .

[0067] The processor 110 is composed of a CPU (Central Processing Unit) and the like, and controls all the components of the data processing device 100, and executes a data processing program to effectively reflect information on the features of the data to be processed, while performing noise reduction processing, etc. Details of the various processes performed by the processor 110 will be described later.

[0068] The memory 120 includes a flash memory, a read-only memory (ROM), a random access memory (RAM), a hard disk drive, etc. The flash memory, the ROM, or the hard disk drive is a non-volatile memory that stores an operating system, various programs including the data processing program according to the present invention, etc. The non-volatile memory (storage device) such as the flash memory or the hard disk drive also stores data to be processed (first data), processed data (second data, third data), an index quantifying the magnitude of noise in each feature of the data, thresholds used for various judgments, etc.

[0069] The RAM functions as a work area for processing by the processor 110. It also temporarily stores various programs stored in flash memory, etc. The processor 110 may incorporate part of the memory 120 (RAM).

[0070] The display 130 is also used as part of a GUI (Graphical User Interface) when accepting user input from the operation unit 150, and in addition to displaying a screen for operating the data processing device 100, it can also display thresholds set by the user, processed data, etc.

[0071] The input / output interface 140 includes a connection unit connectable to an external device, a communication unit connectable to a network, etc. As the connection unit connectable to an external device, a Universal Serial Bus (USB), a High-Definition Multimedia Interface (HDMI) (HDMI is a registered trademark), etc. can be applied.

[0072] The data processing device 100 can acquire necessary information from external storage devices or the cloud via the input / output interface 140, in addition to the data to be processed stored in the memory 120. Furthermore, the processor 110 can store the information acquired via the input / output interface 140 in the memory 120.

[0073] The operation unit 150 includes a pointing device such as a mouse, a keyboard, and the like, and functions as part of a GUI that uses the display screen of the display unit 130 to accept instruction inputs by user operations.

[0074] The display 130 and the operation unit 150 are not essential components of the data processing device according to the present invention.

[0075] First Embodiment FIG. 7 is a functional block diagram showing a first embodiment of a data processing device according to the present invention, and shows a data processing method by the processor 110 of the data processing device 100 shown in FIG.

[0076] In FIG. 7, the processor 110 acquires first data (hereinafter also referred to as "original data") having two or more feature amounts from the memory 120 (step S11).

[0077] The original data is matrix data consisting of N samples×M feature quantities, and is data to be processed before data processing by the data processing device 100 .

[0078] Next, the processor 110 classifies the M feature quantities of the acquired original data into clusters (step S12). The clustering of the M feature quantities is performed by grouping together feature quantities with high similarity.

[0079] The processor 110 also normalizes the scale of the feature quantities of the acquired original data (step S13). The type of data for which scaling is defined is preferably a control group. The processor 110 normalizes the scale of each feature quantity of the original data by dividing the value of the feature quantity by the representative value (median) of the data of the control group. This allows the median of all feature quantities to be 1.

[0080] Next, the processor 110 aggregates the feature quantities of the data (second data) obtained by normalizing the scale of the feature quantities of the original data and belonging to the same cluster classified in step S12 (step S14). The feature quantities belonging to the same cluster can be aggregated into a single feature quantity (new feature quantity) by, for example, calculating the average.

[0081] The processor 110 can generate third data (hereinafter referred to as "new data") with a reduced number of features by replacing the features of the original data with the newly created new features with low noise (step S15). The new data aggregates features belonging to the same cluster, reducing noise contained in the new features and reducing the number of features (feature redundancy) compared to the original data, resulting in matrix data consisting of N samples × M' new features, where M' is the number of features after data processing.

[0082] The processor 110 stores the generated new data in the memory 120 for use, or outputs the data to the outside via the input / output interface 140 .

[0083] According to the first embodiment, new data having new features with reduced noise and feature redundancy can be generated. Furthermore, since the new data has reduced noise and feature redundancy compared to the original data, it is suitable for analysis using data. Furthermore, by using the new data as a dataset for training a machine learning model, the prediction accuracy of the machine learning model can be improved.

[0084] <Second embodiment> Figure 8 is a functional block diagram showing a second embodiment of a data processing device according to the present invention, and shows a data processing method by the processor 110 of the data processing device 100 shown in Figure 6. In Figure 8, parts that are common to the first embodiment shown in Figure 7 are given the same step numbers, and detailed descriptions thereof will be omitted.

[0085] The second embodiment shown in FIG. 8 differs from the first embodiment shown in FIG. 7 in that steps S23 and S24 are performed instead of steps S13 and S14.

[0086] In step S23, the scales of the feature quantities belonging to the same cluster are normalized using the original data and the cluster classification information obtained by cluster classification in step S12.

[0087] Here, when normalizing (scaling) the scales of features belonging to the same cluster, the median or average of the scales of the feature (first feature) of multiple samples is normalized to match the median or average of the scales of the feature (second feature) different from the first feature of multiple samples. Specifically, processing is performed on a cluster-by-cluster basis as shown below.

[0088] The number of features belonging to a cluster is n g Here, g is the cluster number (g=0, 1, ..., G-1). Also, the parameter i g The feature number (i g =0,1,…,n g-1), and the median or average value of the sample values ​​of each feature value belonging to the same cluster is m(x ig ) in the same cluster. ig ) is the maximum value in CM g ≡max(m(x ig ))

[0089] The processor 110 is a CM g Using this, feature values ​​x belonging to the same cluster are ig is scaled by performing the calculation shown in Equation 1.

[0090]

[0091] In step S23, the processor 110 performs the above scaling operation to align the scales of the features belonging to the same cluster, thereby generating second data in which the scales of the features belonging to the same cluster are standardized.

[0092] In step S24, similar to step S14 shown in FIG. 7, feature quantities belonging to the same cluster are aggregated to generate new feature quantities. However, this step differs in that feature quantities of the data (second data) whose scale has been normalized in step S23 are aggregated to generate new feature quantities.

[0093] <Third embodiment> Figure 9 is a functional block diagram showing a third embodiment of a data processing device according to the present invention, and shows a data processing method by the processor 110 of the data processing device 100 shown in Figure 6. In Figure 9, parts that are common to the second embodiment shown in Figure 8 are given the same step numbers, and detailed descriptions thereof will be omitted.

[0094] The third embodiment shown in FIG. 9 differs from the second embodiment shown in FIG. 8 in that steps S34 and S35 are performed instead of step S24.

[0095] In step S34, similar to step S24 shown in FIG. 8, features belonging to the same cluster are aggregated to generate new features, but this step differs in that new features are generated by aggregation using the magnitude of noise in features belonging to the same cluster.

[0096] The noise contained in the feature is defined by the variance, standard deviation, or coefficient of variation of the observed values ​​using a plurality of samples whose true values ​​of the feature are known. The samples for defining the noise are preferably samples whose methylation degree is known or samples from a control group.

[0097] By quantifying noise using synthetic oligo DNA data for which the true value of the methylation degree is known, it is possible to obtain a noise index that accurately represents the bias and variance properties of the measurement system. Therefore, if the measurement range of the target methylation degree can be estimated in advance, it is also possible to quantify the noise within that measurement range, making it possible to reflect measurement characteristics specific to each CpG site and each methylation degree in the noise.

[0098] The index obtained by quantifying the noise magnitude of the feature amount as described above is stored in the memory 120 as noise information.

[0099] When aggregating the features belonging to the same cluster, the processor 110 obtains noise information indicating the magnitude of noise in the features belonging to the same cluster from the memory 120 separately from the original data (step S35).

[0100] The processor 110 aggregates the features belonging to the same cluster using the magnitude of noise of each feature (step S34). Specifically, when aggregating the features belonging to the same cluster, the processor 110 aggregates the features belonging to the same cluster by taking a weighted average based on the inverse of the magnitude of noise of the features belonging to the same cluster.

[0101] Weight w when weighting the average of features belonging to the same cluster ig is defined as the inverse of the noise level contained in the feature, and the weight w ig is preferably stored in the memory 120 as noise information. ig The subscripts of the parameter i g is the feature number (i g =0,1,…,n g -1).

[0102] The processor 110 uses the weights w ig The features in the same cluster are aggregated into one feature by weighting the average using x. ig Then, the aggregated feature x g can be expressed by the following equation:

[0103]

[0104] The processor 110 calculates the new feature x with less noise by performing the calculation of the above [Equation 2]. g , and the feature value x within the same cluster before data processing is calculated. ig The calculated new feature x g Replace with.

[0105] The processor 110 generates N×M′ matrix data (new data) through the data processing consisting of the above procedures, and stores it in the memory 120 or outputs it to the outside via the input / output interface 140 .

[0106] The new data allows for simultaneous noise reduction and feature redundancy reduction. In addition, because it employs aggregation using a weighted average based on quantified noise, even if the noise in each feature is not uniform, it is possible to achieve sufficient noise reduction while retaining the characteristics of the feature.

[0107] FIG. 10 is a conceptual diagram of data obtained by performing a process on DNA methylation data samples in which feature amounts belonging to the same cluster are aggregated using the magnitude of noise in each feature amount.

[0108] The noise level is estimated for each marker belonging to the same cluster after cluster classification of DNA methylation data. In the case of DNA methylation data, it can be estimated from the variation in the methylation level of healthy individuals who are assumed to be unmethylated, and can be obtained as noise information.

[0109] Then, by performing weighted averaging using the magnitude of noise, the contribution of features with low noise increases, enabling effective noise reduction and reducing feature redundancy.

[0110] Fourth Embodiment FIG. 11 is a functional block diagram showing a fourth embodiment of a data processing device according to the present invention, and shows a data processing method by the processor 110 of the data processing device 100 shown in FIG.

[0111] In FIG. 11, the processor 110 acquires original data having two or more feature amounts from the memory 120 (step S11).

[0112] Next, the processor 110 calculates the similarity between two or more feature amounts of the original data (step S42).

[0113] As an index showing the similarity between the feature quantities, an example will be shown here in which the distance between the feature quantities is quantified using the Pearson correlation coefficient.

[0114] In step S42, a correlation coefficient r indicating the similarity between the feature quantities i and j is calculated. i,j (j=0,1,…,N-1(j≠i))

[0115] Next, the processor 110 reads the cluster determination threshold th from the memory 120. r (step S44), and the cluster determination threshold th r The feature amounts are classified into clusters according to the above (step S43).

[0116] In step S43, the correlation coefficient r kg , lg and the cluster determination threshold th r From this, we have a set C that satisfies the following conditions: g Ask for.

[0117] C g =[k g ,l g |∀k g ,∀l g ,r kg , lg ≧th r ] Set C obtained by the above clustering process g The features are grouped according to the

[0118] The following process is performed for each cluster. g (∈C g) where g is the cluster number (g=0, 1, ..., G-1).

[0119] By using the clustering procedure described above, it is possible to specify a distance criterion for classifying data as belonging to the same cluster, and to allow the same data to be included in multiple clusters.

[0120] FIG. 12 is a diagram showing the methylation markers of DNA methylation data in a marker feature space.

[0121] In Figure 12, the cross marks indicate the position of each methylation marker in the marker feature space. The clustering results in three circled marker sets (clustering), and there are also markers that are not included in any cluster.

[0122] There is a trade-off between noise reduction and the reduction in information volume due to feature aggregation, and the cluster determination threshold th r It is possible to optimize it by appropriately setting

[0123] 11 , the processor 110 extracts (acquires) a representative feature from the features belonging to the same cluster (step S45). Examples of the representative feature may include, but are not limited to, the median of the features belonging to the same cluster, a feature corresponding to a predetermined or user-specified percentile value, a feature with the highest S / N ratio among the features belonging to the same cluster, or an average value (including the weighted average value described above) of the features belonging to the same cluster.

[0124] The processor 110 can generate new data having new features with low redundancy through the data processing procedures described above (step S46). For the new data, the features that represent the same cluster are extracted (adopted) from among the features that belong to that cluster, resulting in a reduction in the number of features (feature redundancy) compared to the original data.

[0125] Fifth Embodiment FIG. 13 is a functional block diagram showing a fifth embodiment of a data processing device according to the present invention, and shows a data processing method by the processor 110 of the data processing device 100 shown in FIG.

[0126] The data processing device of the fifth embodiment shown in Fig. 13 is an integration of the third embodiment shown in Fig. 9 and the fourth embodiment shown in Fig. 11. Therefore, in Fig. 13, the same step numbers are assigned to parts common to the third and fourth embodiments, and detailed descriptions thereof will be omitted.

[0127] In FIG. 13 , the data processing in steps S11, S42, S43, and S44 is the same as in the fourth embodiment, and the processor 110 calculates a correlation coefficient indicating the similarity between feature quantities and classifies the feature quantities into clusters based on the correlation coefficient and a cluster determination threshold (steps S43 and S44).

[0128] The data processing in the subsequent steps S23, S34, and S35 is the same as that in the third embodiment, and the processor 110 normalizes the scales of the features belonging to the same cluster (step S23), and then aggregates the normalized features using the noise magnitude of each feature (step S34). Note that when aggregating the features belonging to the same cluster, the processor 110 aggregates them by taking a weighted average based on the inverse of the noise magnitude of the features belonging to the same cluster.

[0129] This makes it possible to generate new data having new features with reduced noise and feature redundancy.

[0130] FIG. 14 is a diagram showing specific numerical examples when normalizing the scale of feature amounts A, B, and C belonging to the same cluster and calculating a weighted average of feature amounts A, B, and C for noise reduction.

[0131] In FIG. 14, the feature values ​​0.45, 0.2, and 0.08 indicate the median values ​​of feature values ​​A, B, and C of a plurality of samples, respectively.

[0132] The correction (median) indicates the correction factor (1, 2.35, and 5.625) used to multiply the median of feature quantities A, B, and C of multiple samples to 0.45, with 0.45 (feature quantity A) being the maximum of these medians (0.45, 0.2, and 0.08). That is, the correction factor for feature quantity A as the reference is 1.0, the correction factor for feature quantity B is 0.45 / 0.2 = 2.25, and the correction factor for feature quantity C is 0.45 / 0.08 = 5.265.

[0133] Next, the weights corresponding to feature quantities A, B, and C indicate how many times the original value should be weighted when the above correction factor and the reciprocal of the coefficient of variation of the measured values ​​of multiple samples whose true feature values ​​are known, which defines the noise, are combined. In this case, the weighting factor for feature quantity A is 2.0 x 1.0 = 2.0 times, for feature quantity B it is 10.0 x 2.25 = 22.5 times, and for feature quantity C it is 10.0 x 5.265 = 52.65 times. Therefore, feature quantities A, B, and C of the original data are aggregated by the weighted averages of the above 2.0 times, 22.5 times, and 52.65 times (weights).

[0134] Sixth Embodiment FIG. 15 is a functional block diagram showing a sixth embodiment of a data processing device according to the present invention, and shows a data processing method by the processor 110 of the data processing device 100 shown in FIG.

[0135] 15, the processor 110 acquires data (original data) having two or more features from the memory 120 (step S11). This step S11 is similar to step S11 shown in Fig. 7, and the original data acquired in step S11 is matrix data consisting of N samples x M features, and is data to be processed before data processing by the data processing device 100.

[0136] Next, the processor 110 uses the original data to determine whether the samples are outliers for each feature amount, and detects outlier samples for each feature amount (step S62).

[0137] The N-dimensional vector focusing on the feature quantity m of the original data is x m (m=0, 1, . . . , M−1) The processor 110 performs outlier detection for each feature amount m in the following procedure.

[0138] First, for the feature m, we perform a variable transformation f such as a square root transformation or a logarithmic transformation. t The effect of outliers is mitigated by the following. m t Let x m t can be expressed by the following equation: [Equation 3] x m t = f t (x m )

[0139] Next, the processor 110 performs an outlier test using the Smirnoff-Grubbs test. However, the method of outlier test is not limited to this. The p value of the sample i obtained at this time is expressed as p m i (i=0,1,...,N-1). The p-value is an abbreviation for the realized value of probability.

[0140] Next, the processor 110 calculates the absolute value d of the difference between the characteristic value, such as the mean, median, or any percentile value, and the sample i. m i However, the characteristic values ​​are not limited to these.

[0141] Finally, the processor 110 sets a threshold th of the p-value for determining an outlier. p and the absolute value threshold th d and are obtained from the memory 120, and p m i >th p , or d m i >th d If the above condition is satisfied, sample i is determined to be an outlier in feature m.

[0142] Absolute value threshold th d may be set to an arbitrary value or may be set to a statistically significant number such as three times the standard deviation. m It is defined as:

[0143] Here, the sample set is treated as an outlier and is excluded from all feature quantities. When calculating the similarity index between feature quantities, the above processing is performed so that the samples used match between feature quantities.

[0144] FIG. 16 is an image diagram showing a case where a sample determined to be an outlier by the outlier determination is excluded from the original data.

[0145] Figure 16(A) shows the methylation levels of all samples for features (markers) 1 and 2, and Figure 16(B) shows the methylation levels of samples for markers 1 and 2 after excluding samples determined to be outliers.

[0146] 16 , the outlier determination is performed using, for example, the Smirnoff-Grubbs test, which is a common method for outlier determination. The Smirnoff-Grubbs test selects a sample whose absolute value from the median or mean is the largest in feature m, performs an outlier test on that sample, and if it is determined to be an outlier, removes that sample from feature m and performs the outlier test again, repeating this process.

[0147] The outlier test is not limited to the Smirnoff-Grubbs test, and for example, the k-nearest neighbor method or the local outlier factor can be applied.

[0148] FIG. 17 is a diagram showing an embodiment of data processing such as outlier testing for the feature m.

[0149] The graph in FIG. 17 shows the feature value m of N samples, the absolute value d indicating the deviation from the median, m i , and parameter values ​​(thresholds) used to determine outliers are shown.

[0150] In this example, the deviation from the median (absolute value d m i ) threshold th d is set to 0.3, and the Smirnoff-Grubbs test p-value (p m i ) threshold th p is set to 0.99.

[0151] In FIG. 17, the absolute value d of sample number Nm N is 0.67, and the p-value p m N is 0.999. The absolute value d of sample number N m N (=0.67) is the threshold value th d Since the p-value d of sample number N exceeds 0.3, it is detected as an outlier. p N (=0.999) is also the threshold th p (=0.99) and is therefore tested as an outlier.

[0152] The outlier test is performed again on the remaining samples excluding sample N of sample number N. The absolute value d of sample number (N-1) m N-1 is 0.28, and the p-value p m N-1 is 0.992. The absolute value d of sample number (N-1) m N-1 (=0.28) is the threshold value th d (=0.3), but the p-value p N-1 (=0.992) is the threshold value p (=0.99), it is tested as an outlier.

[0153] The above outlier testing is continued until no outliers are detected.

[0154] Returning to FIG. 15, when the processor 110 detects outliers for each feature amount in step S62, it removes from the data samples that are determined to be outliers for one or more feature amounts (step S63).

[0155] Next, the processor 110 performs cluster classification of the features using the data from which the outlier samples have been excluded (step S64). Although the data used for the cluster classification here is different, it can be performed in the same manner as step S12 of the first embodiment shown in Fig. 7 or steps S42 to S44 of the fourth embodiment shown in Fig. 11.

[0156] Although not shown, the features classified into clusters in step S64 are aggregated for each cluster as in the first to fourth embodiments, or a representative feature is extracted for each cluster as in the fifth embodiment.

[0157] The processor 110 generates new data having new features with low redundancy using new features newly created by aggregating for each cluster or representative features extracted for each cluster (step S65).

[0158] According to the sixth embodiment, by using data from which samples determined to be outliers have been excluded, clustering that is not affected by some samples with significantly different feature values ​​is possible.

[0159] <Optimization of each threshold> FIG. 18 shows the optimization of the cluster determination threshold th r , th r ' , the p-value threshold th for outlier detection p , absolute value threshold th d FIG. 10 is a conceptual diagram showing the implementation of optimization.

[0160] When applied to a machine learning model for classification or regression, each threshold th r , th r ' , th p , th d However, the purpose of machine learning tasks is not limited to this.

[0161] The hyperparameters that set the behavior of the machine learning algorithm include the thresholds th r , th r ' , th p , th d By introducing this, the evaluation index becomes a function with arguments such as each threshold and learning model parameters. By learning to optimize this evaluation index, it is possible to find each threshold optimized for each task.

[0162] When the processor 110 receives the original data, it uses a machine learning model (first machine learning model) that classifies the feature quantities into clusters. The first machine learning model uses a cluster determination threshold th r is used as a hyperparameter for the machine learning task, and the cluster determination threshold th r Optimize the threshold th r ' is the cluster determination threshold when clustering is performed again for each cluster.

[0163] Furthermore, the processor 110 uses a machine learning model (second machine learning model) that, when input with original data, outputs an outlier determination result or outputs data excluding samples determined to be outliers, and the second machine learning model uses a threshold th that indicates the degree of deviation from the characteristic value (median, average, etc.) of the feature amount. d , threshold th for Smirnoff-Grubbs test p-value p is used as a hyperparameter for the machine learning task, and the task accuracy is fed back to determine the threshold th d , the threshold value for p-value th p Optimize.

[0164] Seventh Embodiment A data processing device according to a seventh embodiment of the present invention performs data processing including two-stage clustering.

[0165] FIG. 19 is a conceptual diagram showing two-stage clustering.

[0166] FIG. 19A is a diagram showing the first stage of clustering, and FIG. 19B is a diagram showing the second stage of clustering.

[0167] A seventh embodiment including two-stage clustering will now be described.

[0168] First, clustering of feature amounts is performed (FIG. 19(A)).

[0169] The processor 110 acquires original data consisting of matrix data made up of N samples and M features, and calculates a correlation coefficient r indicating the similarity between features i and j as shown in the fourth embodiment of FIG. 11. i,j, and the cluster determination threshold th r A set of highly similar features C based on g (g=0,1,…,G-1) and classify the features into clusters.

[0170] Next, the processor 110 converts the N-dimensional vector focusing on the feature quantity m of the original data into x m (m=0, 1, ..., M-1), for each feature m, an outlier determination is performed, and the set of samples determined to be outliers by feature m is called O. m It is defined as:

[0171] Next, outlier removal and reclustering are performed (FIG. 19(B)).

[0172] Sample set is an outlier, and set C g (g=0, 1, ..., G-1) where M g is C g is a set of features contained in

[0173] Then, the cluster C with outliers removed is g For each cluster, a cluster C is created based on the similarity of two or more features. g Clustering is then performed again. g It is possible to further classify into two or more small clusters within the cluster. In the next clustering, a cluster determination threshold th is used to determine whether or not to aggregate into the same cluster. r ' is the cluster determination threshold th r may be used, or a different value may be used.

[0174] By following the above procedure and performing outlier sample determination within each cluster after the first stage of clustering, the number of samples excluded by outlier determination can be reduced, making it possible to calculate similarity indices with higher accuracy.

[0175] Thereafter, noise reduction and feature redundancy reduction are performed in the same manner as in the first embodiment.

[0176] In the above embodiment, the Pearson correlation coefficient is used as an index showing the similarity between feature quantities. However, this is not limiting, and in addition to indices that can be quantitatively determined from data, such as the Spearman rank correlation coefficient or the Mahalanobis distance, known properties in the technical field to which the data belongs may also be used as an index of similarity.

[0177] Furthermore, by using the distance between CpG sites (base pairs), distance taking into account tertiary structure, or measurement reliability determined by the measurement bias and variance of each CpG site as a similarity index for features, it is possible to aggregate CpG sites with similar biological properties under conditions that are in line with the actual measurement system. In this way, by performing weighting that does not rely on training data, the degree of freedom in learning can be reduced, which has the effect of ensuring the stability of learning even with small amounts of data.

[0178] [Data to be analyzed] <Biological information> The data to be analyzed is biological information, more preferably omics data.

[0179] Omics data refers to data obtained by comprehensively analyzing specific functional molecules in living organisms in fields such as genomics, epigenomics, transcriptomics, and proteomics. However, the original data is not limited to this. For example, omics data includes gene expression data, microRNA (ribonucleic acid) expression data, SNP (single nucleotide polymorphism) analysis data, aCGH (array-based comparative genomic hybridization) analysis data, DNA methylation data, cell-free DNA methylation data in blood, histone modification data, nuclear chromatin higher-order structure data, ubiquitinated protein data, and glycosylation data such as lectin arrays.

[0180] The reason for the high expected effectiveness of omics data is that there is measurement-induced noise and deficiencies that occur when measuring the target data, as well as biological noise and subclasses that depend on individual differences in each sample. This means that noise is not uniform across each feature, and that the same label has many subclasses. In addition, because features tend to be highly similar, the dataset has high feature redundancy, resulting in a large number of features that can be clustered. In other words, there are more features that can be noise-reduced than with other data. Another problem with omics data is that the amount of data is limited.

[0181] This characteristic can be a major barrier when utilizing biometric information for industrial applications. The reason for this is explained using a prediction task using supervised learning as an example. When a machine learning model is trained using small amounts of noisy data, there is a possibility that random biases in the noise will be recognized as characteristics of the data. Furthermore, if feature redundancy is high, there is a possibility that the learning will not be completed successfully, resulting in an unlearned state. In this case, it becomes difficult to make predictions for data not included in the training data, making the method unusable for industrial applications. Therefore, the present invention performs the above-described data processing on the original data to be analyzed, thereby sufficiently reducing noise and feature redundancy while generating data with minimal reduction in information volume. This makes it possible to build a highly accurate machine learning model even when using small amounts of data.

[0182] The data to be analyzed can be DNA methylation data included in omics data.

[0183] When the methylation level of CpG sites is considered as a feature, DNA methylation data has the property of being highly redundant when multiple CpG sites are used as features, since the type of feature is unified as DNA methylation. However, CHH sequences or CHG sequences may also be used. Furthermore, unlike mutations that can be directly measured using nanopores, methylation level measurement requires steps such as bisulfite conversion, making the measurement process complex. Therefore, measurement noise is likely to be introduced. Due to these properties, DNA methylation data is expected to be more effective in reducing noise and feature redundancy using the present invention.

[0184] Furthermore, data with higher expected effects are obtained from cell-free DNA methylation data in blood. In addition to the effects of DNA methylation data, the following effects can be expected.

[0185] When measuring the methylation of cell-free DNA in blood, DNA fragmentation can result in extremely low or high methylation levels, or even make the methylation level unmeasurable. This phenomenon can be considered a case of extremely high noise. It also tends to result in a high number of subclasses. Furthermore, because blood contains DNA derived from cells throughout the body, it is expected that the proportion of DNA derived from target cells is extremely low, such as 0.1% or 1%. Given these characteristics, this data is more susceptible to noise and a high number of subclasses than DNA methylation data in general, making it an area in which the present invention is particularly effective.

[0186] <Data with a control group and a treatment group> If the data to be analyzed contains a control group and a treatment group, and the control group and the treatment group have different features, the following effects can also be expected.

[0187] By quantifying noise and scaling features within a cluster using control data, we can effectively reduce noise without requiring external data. This is because, when the target signal is used to distinguish between affected individuals (treated group) and controls, the signal in the control group is expected to be a constant value, and the variability of the observed values ​​in the control group samples can be interpreted as the noise level. By scaling this noise to account for measurement bias and biological bias, this method can be applied even when sufficient measurement data for noise quantification is not available.

[0188] Furthermore, by performing outlier detection and similarity calculation using only data from affected individuals (samples from the treatment group), it is possible to detect outlier samples and cluster features specifically for detecting affected individuals.

[0189] [Example] Figures 20 to 30 are diagrams showing an example of a data processing device and method according to the present invention, and show a case where the present invention is applied to cell-free DNA methylation data as the data to be analyzed.

[0190] 20 is a diagram showing cell-free DNA methylation data consisting of N samples and M features. The number of samples, N, is 166, and the number of features (markers), M, is 60.

[0191] Since the total data size of the 166×60 matrix data is large, data that has been appropriately thinned out is displayed in FIG.

[0192] In addition, in Figure 20, the first column is information indicating whether the sample is from a healthy individual "Normal_cfDNA" or a colorectal cancer patient "Tumor_COADREAD", and the second column and beyond indicate the "methylation level" of each sample (1 to N) for each feature (maker_0 to maker_59).

[0193] FIG. 21 is a diagram showing a cluster set when the cell-free DNA methylation data shown in FIG. 20 is cluster-classified.

[0194] The cell-free DNA methylation data shown in Figure 20 is subjected to square root transformation to mitigate the influence of outliers. Next, the similarity index between each feature is calculated. Here, the distance between features is quantified using the Pearson correlation coefficient. The similarity index calculated in this way is applied to the cluster determination threshold th r (=0.85) as set C g Set C g is shown in FIG.

[0195] The first column in FIG. 21 indicates the cluster number of each cluster, and the second column indicates the feature number of the feature belonging to each cluster indicated by the cluster number.

[0196] As shown in FIG. 21, out of the 60 features (features indicated by feature numbers 0 to 59) corresponding to maker_0 to maker_59, a total of 43 features belong to each cluster.

[0197] Fig. 22 is a diagram showing a cluster set obtained when the feature quantities belonging to each cluster shown in Fig. 21 are re-clustered. Note that Fig. 22 is a diagram in a format similar to that shown in Fig. 21.

[0198] Cluster C using the Smirnoff-Grubbs test g In this example, an outlier test is performed for each p-value threshold th p (=0.999), absolute value threshold th d (=0.3).

[0199] p-value threshold th p , absolute value threshold th d After performing outlier processing using g Clustering is performed again for each cluster. r ' is changed to 0.7. The set C′ after reclustering is shown in FIG.

[0200] FIG. 23 is a table showing the weights for the feature quantities belonging to the cluster with cluster number 0 shown in FIG.

[0201] In this embodiment, the weight used to average each feature is defined as the inverse of the noise.i0 An example is shown in FIG.

[0202] FIG. 24 is a table showing the enlargement and reduction ratios of the features belonging to the cluster C0′ with the cluster number 0 shown in FIG.

[0203] In this embodiment, scaling is performed using the median in order to make the feature values ​​uniform. The enlargement and reduction ratios at the time of scaling in cluster C0' are shown in FIG.

[0204] 25 and 26 are charts showing data having new features for each cluster based on weighted averages.

[0205] That is, cluster C g For each ', weight w ig Using this, cluster C g 'Feature x ig The feature amounts aggregated for each of the 11 clusters shown in FIG. 22 are shown in FIGS. 25 and 26.

[0206] FIG. 25 shows feature amounts aggregated for five clusters with cluster numbers g (=0, 1, 2, 3, 4), and FIG. 26 shows feature amounts aggregated for six clusters with cluster numbers g (=5, 6, 7, 8, 9, 10).

[0207] 27 and 28 are charts showing data obtained by subjecting the feature amounts shown in FIGS. 25 and 26 to quantile transformation, respectively.

[0208] It is preferable to perform normalization such as quantile transformation or standardization on the data to be input into the machine learning model.

[0209] FIG. 29 shows the cluster determination threshold th for the data to be analyzed. r , th r ' , p-value threshold th p , absolute value threshold th d 10 is a diagram showing an example of a data-driven parameter search.

[0210] Specifically, an example of parameter search is shown for the quantile-converted cell-free DNA methylation data shown in Figure 20 and the data shown in Figures 27 and 28, when logistic regression is used in a machine learning model to classify healthy individuals / colon cancer patients.

[0211] The results of the ROC-AUC for the values ​​of each parameter and the classification performance of healthy subjects / colon cancer patients are shown in Figure 29. The highest ROC (Area Under the Curve)-AUC (Area Under the Curve) was found for Experiment ID: 3, and the data included r =0.9,th r ' =0.85,th p =0.999,th d A threshold value of 0.4 is preferably applied to the present invention.

[0212] Figure 30 is a chart showing quantile-transformed data for isolated features (25 features) that are not classified into the 11 clusters shown in Figure 22 from the cell-free DNA methylation data shown in Figure 20.

[0213] Since the 166×25 matrix data is large, data that has been appropriately thinned out is displayed in FIG.

[0214] Then, new data is generated that has a total of 36 features, consisting of the 11 new features aggregated into 11 cluster units shown in FIGS. 27 and 28 and the 25 isolated features shown in FIG. 30. This reduces the number of features in the new data from 60 in the original data to 36, thereby achieving a reduction in redundancy.

[0215] [Others] The present invention is not limited to any one of the first to seventh embodiments described above, but also includes configurations in which the first to seventh embodiments are appropriately combined.

[0216] The hardware structure that executes various controls of the data processing device according to the present invention is the following various processors: The various processors include a CPU (Central Processing Unit), which is a general-purpose processor that executes software (programs) and functions as various control units, a programmable logic device (PLD), such as an FPGA (Field Programmable Gate Array), whose circuit configuration can be changed after manufacture, and a dedicated electrical circuit, such as an ASIC (Application Specific Integrated Circuit), which is a processor with a circuit configuration designed specifically for executing specific processing.

[0217] A single processing unit may be configured with one of these various processors, or may be configured with two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). Multiple control units may also be configured with a single processor. Examples of multiple control units configured with a single processor include: a first configuration, as typified by client or server computers, in which a single processor is configured with a combination of one or more CPUs and software, and this processor functions as multiple control units; and a second configuration, as typified by system-on-chip (SoC), in which a processor is used to realize the functions of an entire system including multiple control units on a single IC (Integrated Circuit) chip. In this way, the various control units are configured with one or more of the above-mentioned various processors as a hardware structure.

[0218] Furthermore, the present invention includes a program that, when installed on a computer, causes the computer to execute the data processing method of the present invention, and a non-transitory, computer-readable, non-volatile storage medium on which this program is recorded.

[0219] Furthermore, it goes without saying that the present invention is not limited to the above-described embodiment, and various modifications are possible within the scope of the present invention.

[0220] REFERENCE SIGNS LIST 100... Data processing device 110... Processor 120... Memory 130... Display 140... Input / output interface 150... Operation unit S11 to S15, S23, S24, S34, S35, S42 to S46, S62 to S65... Steps

Claims

1. A data processing apparatus that processes first data having two or more feature amounts and in which noise is included in each feature amount by a processor, wherein the processor: acquires the first data; classifies the two or more feature amounts into clusters using the first data; generates second data in which the scale of the feature amounts is normalized using the first data; aggregates the feature amounts belonging to the same classified cluster of the second data using the classified cluster classification information to create a new feature amount, and generates third data having the created new feature amount. Data processing apparatus.

2. The normalization of the feature amounts normalizes the scales of the feature amounts belonging to the same cluster using the classified cluster classification information. The data processing apparatus according to claim 1.

3. When aggregating the feature amounts belonging to the same cluster, the processor aggregates them using the magnitude of the noise of the feature amounts belonging to the same cluster. The data processing apparatus according to claim 1 or 2.

4. When aggregating the feature amounts belonging to the same cluster, the processor aggregates them by taking a weighted average using the reciprocal of the magnitude of the noise of the feature amounts belonging to the same cluster. The data processing apparatus according to claim 1.

5. The noise is defined by the coefficient of variation, variance, or standard deviation of the measured values of a plurality of samples whose true values of the feature amounts are known. The data processing apparatus according to claim 1.

6. The samples for defining the noise are samples with a known degree of methylation or samples of a control group. The data processing apparatus according to claim 5.

7. The processor: calculates the similarity between two or more of the feature amounts using the first data; classifies the feature amounts into clusters using a cluster determination threshold value and the calculated similarity. The data processing apparatus according to claim 1.

8. When the first data is input, the processor uses a first machine learning model that classifies the feature amounts into clusters, wherein the first machine learning model uses the cluster determination threshold value as a hyperparameter of a machine learning task and optimizes the cluster determination threshold value by feeding back the task accuracy. The data processing apparatus according to claim 7.

9. The first data is composed of a plurality of samples, and the first data for calculating the similarity between the feature amounts is a sample of a processing group. The data processing apparatus according to claim 7.

10. The first data is composed of a plurality of samples, and the processor uses the first data to perform outlier determination of the samples for each feature amount, and based on the similarity of the feature amounts, in units of clusters classified, uses the first data with the samples determined as outliers excluded to further classify into two or more smaller clusters within the cluster, and uses the first data with the samples excluded and the cluster classification information of the classified smaller clusters instead of the first data and the cluster classification information. The data processing apparatus according to claim 7.

11. The first data is composed of a plurality of samples, and the processor uses the first data to perform outlier determination of the samples for each feature amount, and uses the first data and the outlier determination result to exclude the samples determined as outliers from the first data with one or more feature amounts. The data processing apparatus according to claim 1.

12. The outlier determination of the sample uses the Smirnov-Grubbs test, the k-nearest neighbor method, or Local Outlier Factor, and uses the deviation from the characteristic value of the feature amount to determine outliers. The data processing apparatus according to claim 11.

13. When the processor inputs the first data, it uses a second machine learning model that outputs the outlier determination result or outputs the first data with the samples determined as outliers excluded. The second machine learning model uses the threshold indicating the degree of deviation from the characteristic value of the feature amount as a hyperparameter of the machine learning task, and optimizes the threshold indicating the degree of deviation by feeding back the task accuracy. The data processing apparatus according to claim 11.

14. The first data for which the outlier determination is performed is a sample of a processing group. The data processing apparatus according to claim 11.

15. The first data is composed of a plurality of samples. When the processor normalizes the scales of the feature quantities belonging to the same cluster, the processor normalizes them so that the scale of the first feature quantity of the plurality of samples and the scale of a second feature quantity different from the first feature quantity of the plurality of samples are aligned. The data processing apparatus according to claim 1.

16. The first data is composed of a plurality of samples. When the processor normalizes the scales of the feature quantities belonging to the same cluster, the processor normalizes them so as to match the median or average value of the scales of the first feature quantity of the plurality of samples with the median or average value of the scales of a second feature quantity different from the first feature quantity of the plurality of samples. The data processing apparatus according to claim 1.

17. The samples for which the median or the average value is calculated are samples of a control group. The data processing apparatus according to claim 16.

18. The first data is omics data, DNA methylation data, or cell-free DNA methylation data in blood. The data processing apparatus according to claim 1.

19. There exist a control group and a treatment group for the first data, and the control group and the treatment group have different feature quantities. The data processing apparatus according to claim 1.

20. The processor creates the third data with the number of feature quantities reduced by replacing the newly created feature quantity with the original feature quantity of the first data. The data processing apparatus according to claim 1.

21. The processor performs variable transformation of the first data before calculating the similarity between the feature quantities of the first data. The data processing apparatus according to claim 7.

22. The variable transformation of the first data is logarithmic transformation or square root transformation. The data processing apparatus according to claim 21.

23. A data processing method for processing first data having two or more feature amounts and including noise in each feature amount by a processor, the method comprising: a step in which the processor acquires the first data; a step in which the processor classifies the two or more feature amounts into clusters using the first data; a step in which the processor generates second data in which the scale of the feature amounts is normalized using the first data; and a step in which the processor creates new feature amounts by aggregating the feature amounts belonging to the same classified cluster of the second data using the classified cluster classification information, and generates third data having the created new feature amounts.

24. The data processing method according to claim 23, wherein the normalization of the feature amounts normalizes the scales of the feature amounts belonging to the same cluster using the classified cluster classification information.

25. The data processing method according to claim 23 or 24, wherein when the processor aggregates the feature amounts belonging to the same cluster, the processor aggregates the feature amounts using the magnitude of the noise of the feature amounts belonging to the same cluster.

26. A program for causing a computer to execute the data processing method according to claim 23 or 24.

27. A non-transitory and computer-readable recording medium on which the program according to claim 26 is recorded.

Images