Information processing apparatus, method, and program

Abstract

To provide an information processing device capable of constructing a conversion model for performing data expansion in DNA methylation rate data having different attributes.SOLUTION: The learning device includes a pairing unit 1 that pairs first data, which is DNA methylation rate data of a plurality of samples acquired in a first attribute, with second data, which is DNA methylation rate data of a plurality of samples acquired in a second attribute, the first data and the second data being determined to have the same or similar trait of a sample, and a learning unit 2 that constructs a conversion model that outputs data on a side of the second attribute of the paired data as input by learning using the paired data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing device, method, and program for performing processing related to DNA methylation rates. [Background technology]

[0002] In the field of molecular biology, there are examples of conventional data augmentation techniques, such as those described in Non-Patent Document 1, in which genome-wide gene expression data is used with GAN (generative adversarial network) to compensate for data with many missing values ​​in single-cell analysis. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Marouf M, Machart P, Bansal V, Kilian C, Magruder DS, Krebs CF, Bonn S.Realistic in silico generation and augmentation of single-cell RNA-seq data using generative adversarial networks. Nat Commun. 2020 Jan 9;11(1):166. doi:10.1038 / s41467-019-14018-z. PMID: 31919373; PMCID: PMC6952370. Summary of the Invention [Problem to be solved by the invention]

[0004] However, because conventional techniques perform data expansion by complementing defects in single-cell analysis, they assume that the target cells are single. In other words, conventional techniques perform data expansion based on the assumption that the target cells are homogeneous, for example, originating from the same individual and the same tissue.

[0005] For this reason, in conventional technology, when the tissues from which the data was obtained were different, such as data obtained from saliva and data obtained from blood, it was impossible to perform data extension of blood-derived data based on saliva-derived data, or conversely, to perform data extension of saliva-derived data based on blood-derived data.

[0006] Here, DNA methylation patterns are a specific example of data for which differences in tissue origin are important, and expanding this data can provide various benefits, as described below.

[0007] Genome-wide differences in DNA methylation are largely due to tissue origin, and when estimating age or specific traits from DNA methylation patterns, differences in tissue can lead to reduced accuracy. For example, methylation patterns from blood, which is somewhat invasive, are relatively easily obtained through blood tests and other methods, resulting in abundant data and the development of models capable of accurately estimating specific traits. However, methylation patterns from saliva, which is less invasive, are relatively scarce, resulting in a paucity of data and the lack of models capable of accurately estimating these traits. In such cases, by expanding saliva-derived data to build models with similar accuracy to blood data, it is possible to estimate specific traits using saliva-derived data. This approach has wide application in healthcare, where obtaining information from minimally invasive tissues is preferred.

[0008] Similarly, in fields outside of healthcare, for example, in forensic medicine, where information must be obtained from limited physical evidence, if the tissue obtained differs from the tissue corresponding to an existing estimation model, the existing model cannot be applied. However, if data expansion between different tissues can be achieved, it will be possible to reconstruct the existing model for different tissues.

[0009] On the other hand, if data augmentation is not used, it is necessary to obtain real data, and since genome-wide data incurs a certain cost for acquisition, it is extremely costly to collect enough data to build accurate models for each tissue.

[0010] As described above, data augmentation between different tissues has various applications and advantages, but was not possible with conventional technology. Furthermore, while the above explanations have used data augmentation between different tissues when the tissues from which methylation patterns are obtained are different, data augmentation between different attributes can also have various advantages when the attributes of the subjects from which methylation patterns are obtained are different (for example, not just tissues, but cell types, or attributes such as the sex, age, and lifestyle of the subjects from which the patterns are obtained), but such data augmentation was not possible with conventional technology.

[0011] In view of the above-mentioned problems of the conventional technology, an object of the present invention is to provide an information processing device, method, and program capable of constructing a conversion model for data expansion in DNA methylation rate data with different attributes. [Means for solving the problem]

[0012] To achieve the above object, the present invention is an information processing device comprising: a pairing unit that pairs first data, which is DNA methylation rate data of multiple samples acquired according to a first attribute, with second data, which is DNA methylation rate data of multiple samples acquired according to a second attribute, when the samples are determined to have the same or similar traits; and a learning unit that constructs a conversion model that inputs data on the first attribute side of the paired data and outputs data on the second attribute side of the paired data by learning using the paired data. The present invention is also characterized as a method or a program corresponding to the information processing device. [Effects of the Invention]

[0013] According to the present invention, by performing pairing between data and learning, it is possible to construct a conversion model for inputting first attribute data and outputting second attribute data. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a functional block diagram of an information processing device according to an embodiment. [Figure 2] FIG. 1 is a diagram showing the properties of DNA methylation rates. [Figure 3] FIG. 10 is a diagram illustrating a pairing method. [Figure 4] FIG. 10 is a diagram illustrating a case where a denoising autoencoder is used. [Figure 5] FIG. 10 is a schematic diagram for explaining a second embodiment. [Figure 6] FIG. 1 is a diagram illustrating a hardware configuration of a typical computer. DETAILED DESCRIPTION OF THE INVENTION

[0015] 1 is a functional block diagram of an information processing device according to one embodiment. As shown in the figure, the information processing device 10 includes a pairing unit 1, a learning unit 2, a conversion unit 3, a model construction unit 4, and a model verification unit 5, and performs processing in the order of the illustrated functional blocks. Each of these functional blocks will be described in detail below.

[0016] The pairing unit 1 reads the DNA methylation patterns of multiple samples obtained in the first attribute (referred to as "first data") and the DNA methylation patterns of multiple samples obtained in the second attribute (referred to as "second data") as input data, performs pairing between the first data and the second data as much as possible, and outputs the results to the learning unit 2.

[0017] As shown in FIG. 1, the first data input to the pairing unit 1 is also input to the conversion unit 3 at the subsequent stage.

[0018] In the following description of this embodiment, the first data and the second data are assumed to be in the following state. The first attribute and the second attribute are attributes resulting from the fact that the tissues from which the DNA methylation patterns were measured are tissue A and tissue B, which are different from each other. ●Tissue A is, for example, blood, and the first data for tissue A contains a large amount of DNA methylation measurement results (a data set of DNA methylation patterns) labeled with traits, making it possible to train a trait prediction model with a certain level of accuracy. ●Tissue B is, for example, saliva, and the second data for tissue B contains only a small amount of DNA methylation measurement results with trait labels, which is insufficient for training a trait prediction model. ●There is no data collected from a common sample (subject, individual) in Organization A and Organization B.

[0019] In the above situation, by carrying out the processing from pairing 1 to model construction unit 4 of this embodiment, the second data of organization B, which has an insufficient amount of data, is expanded with the first data of organization A, which has an abundant amount of data, thereby making it possible to construct a trait prediction model that ensures a predetermined accuracy even for the second data.

[0020] Both the first and second data are DNA methylation rate data. As is well known, DNA methylation rate data can be obtained by simultaneously measuring methylation information for hundreds of regions (CpG sites) on the genome using microarrays or high-speed parallel sequencers. When targeting a cell population, the methylation rate can be expressed as a continuous value in the range of 0-1 (0% to 100%). Furthermore, information obtained from single-cell resolution data can also be treated as a continuous value for a convoluted cell population.

[0021] For example, if there are 1000 CpG sites (genomic regions) being processed (CpG1, CpG2, CpG3, ..., CpG999, GpG1000), the DNA methylation rate data will be a 1000-dimensional vector (r(1), r(2), r(3), ..., r(999), r(1000)), where the methylation rate value r(i) is given in the range of 0-1 for each CpGi (i = 1, 2, 3, ..., 999, 1000) of the 1000 genomic regions.

[0022] As is known, DNA methylation rates vary across different regions of the genome, and although they are generally similar, they tend to vary by tissue and sex, and also vary between individuals depending on factors such as age and environmental exposure.

[0023] As shown in Figure 2, the above-described properties of DNA methylation rates are schematically illustrated. Because of these properties, it is possible to construct a model that predicts biological age, etc., using DNA methylation rates as input. However, if a model already constructed using abundant first data is applied directly to second data (for example, if DNA methylation rates derived from saliva, rather than blood, are used as input for a biological age prediction model based on blood-derived DNA methylation rates), the prediction accuracy will decrease. Therefore, it is necessary to reconstruct the prediction model for the sparse second data. This embodiment makes this reconstruction possible.

[0024] In the pairing unit 1, as described above, samples with the same or similar traits can be paired between each sample of the first data and each sample of the second data. FIG. 3 is a diagram schematically showing the pairing method, and as the trait, one or more arbitrary traits can be used which are given to the first data and the second data in advance. Specifically, for example, as an arbitrary numerical value related to a person's health, the cumulative smoking amount or the like can be used. The criterion for determining that the traits are the same or similar may be one given in advance according to the specific data content of the trait. For example, when the value of a specific trait of a sample s1 of the first data is c1 and the value of a specific trait of a sample s2 of the second data is c2, the following threshold determination can be used to determine identity or similarity. ● If |c1 - c2| ≤ TH1, it is determined that the traits of s1 and s2 are the same. ● If TH1 < |c1 - c2| ≤ TH2, it is determined that the traits of s1 and s2 are similar. ● If |c1 - c2| > TH2, it is determined that the traits of s1 and s2 are neither the same nor similar. Here, TH1 and TH2 are given as determination thresholds with 0 ≤ TH1 < TH2.

[0025] Note that the trait values may be discrete or continuous numerical values, or may be composed of any scale level (nominal scale, ordinal scale, interval scale, ratio scale) in statistical variables including cases where they are selected from predetermined response items such as questionnaires. They may also be composed of an enumeration of multiple types of trait values in the form of a vector or the like. Even when a trait is given by an item other than a numerical value, identity or similarity can be determined according to a predetermined criterion in the same manner as when using threshold determination for numerical values as described above.

[0026] 3, example EX1 is a schematic diagram showing an example of pairing samples s1 and s2 whose traits c1 and c2 are identical or similar when there is no data of the same sample (same person) in the first data and the second data, or when a common ID for identifying each sample cannot be used between the first data and the second data. Note that pairing may be performed only when the sample is determined to be identical, and not when the sample is determined to be similar, or the sample may be paired when the sample is determined to be similar (identity determination is also included as a special case of similarity determination).

[0027] As shown in Example EX2, in addition to determining whether traits c1 and c2 are identical or similar, if attribute information of samples s1 and s2 can be referenced in the first and second data prepared in advance, samples whose attributes are determined to be identical or similar may be paired. Attributes of samples s1 and s2 may include gender, age, and lifestyle evaluation values ​​based on health checkup data, etc. The traits and attributes of samples may have items in common.

[0028] Although the situation described above is that "there is no data collected from a common sample (subject, individual) between organizations A and B," if all or part of the first data and second data share a common sample (if a common ID is assigned to all or part of the first data and second data, and it can be determined that they are a common sample), they may be paired as corresponding to the same trait. This is shown as example EX3 in Figure 3, and if the first data and second data contain data of the same person (hereinafter "person α"), the first data of person α and the second data of person α may be paired.

[0029] The learning unit 2 uses the pairing results from the pairing unit 1 as learning data to learn a conversion model that inputs the first data and outputs the second data, and outputs the model to the conversion unit 3. That is, when the pairing results from the pairing unit 1 are N pairs in total, and the methylation rates A(i) (i=1, 2, ..., N) of the first data are paired with the methylation rates B(i) (i=1, 2, ..., N) of the second data as pairs {A(i), B(i)}, the learning unit 2 learns a function f as a conversion model that inputs A(i) and outputs B(i). B(i)=f(A(i))

[0030] Note that since the pair {A(i), B(i)} is training data, when the function f obtained as a result of actual training is used, it is expected that the value of f(A(i)) will be close to B(i), but it is not necessarily a perfect match.

[0031] For this learning, any machine learning or deep learning technique can be used, such as a fully connected neural network or a Transformer, which is a neural network for sequence conversion, but in one embodiment, a denoising autoencoder technique, i.e., an autoencoder technique with noise removal, may also be used. Figure 4 is a diagram for explaining the use of a denoising autoencoder, showing the configuration of an autoencoder in Example EX41 and the configuration of a denoising autoencoder in Example EX42.

[0032] As shown in Example EX41, an autoencoder is configured to process a high-dimensional input x using an encoder to obtain a reduced-dimensional intermediate output z as a latent representation, and then process this latent representation z using a decoder to obtain a high-dimensional output y. The input x and output y are provided as training data to learn the parameters of the encoder and decoder, which are composed of a neural network or the like. As shown in Example EX42, a denoising autoencoder is similar in that the input x and output y are provided as training data to learn the parameters of the encoder and decoder, which are composed of a neural network or the like, but by applying a predetermined noise superposition process to the input x and then training the autoencoder, a noise removal effect can also be expected in the learning results.

[0033] In this embodiment, input A(i), which is methylation rate data of a first attribute, is considered to be methylation rate data B(i) of a second attribute, which has a different attribute, with noise n superimposed thereon due to the difference in attributes. Therefore, the denoising autoencoder is trained to take A(i) as input and output B(i). Therefore, it is sufficient to perform training with the autoencoder configuration shown in Example EX41, without having to perform the noise superimposition process shown in Example EX42 again. B(i)=A(i)+n

[0034] The conversion unit 3 applies the conversion model (function f) obtained in the learning unit 2 to each sample of the first data, thereby obtaining methylation rate data of the second attribute as a result of converting the methylation rate data of the first attribute, and outputs this to the model construction unit 4. By this conversion, (if the first and second attributes are differences due to their origin in tissues A and B) the methylation rate data of tissue A, for which there is a large amount of data, is converted into the methylation rate data of tissue B, thereby realizing data expansion of the methylation rate data of tissue B, for which there is only a small amount of data, and it is expected that a highly accurate prediction model for tissue B will be constructed in the next model construction unit 4.

[0035] The first data to be converted by the conversion unit 3 may be all or part of the data used for pairing in the pairing unit 1 (including data that was not paired), or it may be first data that is separate from the data handled by the pairing unit 1.

[0036] The model construction unit 4 constructs a model that predicts the traits of corresponding samples from the methylation rate data of the second attribute by learning using the data-extended second data obtained by the conversion unit 3, and outputs the model to the model verification unit 5. Note that the data used for model construction in the model construction unit 4 may not only be the data-extended second data obtained by the conversion unit 3, but also the second data used when pairing in the pairing unit 1 (as actual data, not obtained by data extension processing).

[0037] It is preferable that the trait predicted by the model constructed in the model construction unit 4 is the same as the trait paired by the pairing unit 1 based on its identity or similarity judgment, but it may also be a different trait (if it is composed of multiple traits, all or part of them are different traits).

[0038] The second data expanded by the conversion is linked to the target trait value of the corresponding sample in the first data before the conversion as a known value. Therefore, the target trait value corresponding to the expanded second data, which needs to be referenced when building a model, can be the target trait value linked in the first data of the corresponding sample before the conversion.

[0039] The model constructed by the model construction unit 4 may be any type of existing model. For example, if the trait is biological age, a known biological age prediction model, which the applicant also mentions in JP 2022-142021 A, may be constructed.

[0040] In other words, biological age prediction models can calculate biological age by analyzing modification rates at specific sites in the epigenome, using biological age as an indicator of the functional capacity and degree of aging of biological tissues rather than chronological age. Models for calculating biological age are generally trained using a multiple linear regression model or a regression model (e.g., elastic net) with a regularization term added to prevent overfitting, with biological age obtained from various aging indicators as the objective variable and epigenetic modification rate data (generally, data on the modification level of modified sites) as the explanatory variables. Here, the epigenetic modification rate data may be the level of DNA methylation or DNA hydroxymethylation. For example, when DNA methylation rates are used as epigenetic modification rate data, biological age F can be calculated using a model such as the following equation, using multiple DNA methylation sites (CpG regions) that correlate with age (biological age). F=b0+b1·CpG1+…+b n CpG n

[0041] In the above formula, b k (k=0,1,...n) is the coefficient in the prediction model, and CpG k (k=1, 2, ..., n) is the methylation rate of the kth CpG region in the prediction model. Note that if only some of the CpG regions (k=1, 2, ..., n) are correlated with biological age in the prediction model, the methylation rate of the uncorrelated CpG regions is k Coefficient b for k Ga b k = 0. In general, when any prediction model is used, the methylation rate can be taken as the modification level of the modified site.

[0042] The model verification unit 5 verifies the accuracy of the model obtained by the model construction unit 4. Any existing method may be used as the verification method, and for example, a model may be used to predict a trait for verification methylation rate data of the second attribute to which the correct answer for the trait has been assigned, and an evaluation result of the model accuracy may be obtained based on the degree of agreement between the prediction result and the correct answer.

[0043] As described above, according to the embodiment of the present invention, a conversion model from the first data to the second data is learned based on the results of pairing the first data, which is methylation rate data of a first attribute, with the second data, which is methylation rate data of a second attribute, and the conversion model can be used to expand the second data. This enables various uses, such as the construction of a trait prediction model using the second data, even when the amount of second data is small or the acquisition cost is high.

[0044] That is, the cost of obtaining DNA methylation patterns (and expression data, etc.) is high, and DNA methylation patterns (and expression data) vary significantly between tissues due to differences in cellular composition, etc. On the other hand, latent expressions that are not derived from cellular composition contain trait information about the subject. Therefore, the tissue-converted genome-wide DNA methylation patterns obtained in this embodiment can be used as training data when estimating target traits from DNA methylation patterns. This makes it possible to build a model that predicts target traits from the tissue of interest even when the amount of DNA methylation data for the tissue of interest is insufficient.

[0045] Various supplementary points, additions, alternatives, etc. will be described below.

[0046] (1) According to an embodiment of the present invention, for example, the aforementioned biological age assessment can be easily performed, and as an application thereof, for example, advice information can be generated that encourages behavioral changes to maintain and improve health, which can contribute to Goal 3 of the United Nations-led Sustainable Development Goals (SDGs) “Ensure healthy lives and promote well-being for all at all ages.”

[0047] (2) Examples of applications in which the embodiments of the present invention are suitable include the following. ● Test panel design service for target traits (latent characteristics) ● Estimation service for traits correlated with the degree of DNA methylation, such as biological age ● Software that uses DNA methylation rate measurement data from a small number of samples taken widely on a genome-wide scale and data on target traits as input to reconstruct models of other tissues constructed from a large number of samples.

[0048] (3) The difference between the first attribute and the second attribute that distinguishes the first data from the second data can be configured by all or part of the following instead of and / or in addition to those listed in the above explanatory examples, and embodiments of the present invention can be applied. ● Differences in somatic tissues from which DNA methylation data are obtained ● Differences in cell types from which DNA methylation data are obtained ● Genotype differences of samples from which DNA methylation data are obtained ● Gender differences in samples from which DNA methylation data are obtained ● Differences in the age of samples from which DNA methylation data are obtained

[0049] (4) In the above explanation, the sample from which DNA methylation rate data is obtained is assumed to be human (human), but it may also be other living organisms, such as pets such as dogs and cats.

[0050] (5) In order to enable the pairing unit 1 to obtain multiple pairing results for the second data, overlapping may be allowed when pairing with the first data. That is, for one sample of the second data, there may be multiple samples paired with it in the first data. Conversely, for one sample of the first data, there may be multiple samples paired with it in the second data.

[0051] (6) For the sake of explanation, the above embodiment of the information processing device 10 will be referred to as the first embodiment. As a second embodiment relating to the criteria used for pairing and the processing until the final result is obtained, the information processing device 10 may use the estimated specific trait as the pairing criterion by the following machine learning approach procedure, and repeat the process from pairing to model construction and trait estimation based on the constructed prediction until convergence is achieved. Here, as an explanatory example, the premise situation of the first data and the second data is as follows. There is a large amount of primary data, and a highly accurate prediction model has been constructed that predicts biological age as an example of a specific trait. By using this prediction model, each sample of the primary data is associated with a biological age. (This premise is realized in step 1 described below.) ●Since only a small amount of secondary data exists and a predictive model that can predict biological age with high accuracy has not been constructed, biological age is not associated with each sample. ●All samples in the first and second data are linked by chronological age (actual age).

[0052] <Step 1> A biological age prediction model is constructed using the first data, and the biological ages of all samples in the first data are predicted. <Step 2> The pairing unit 1 performs pairing based on the chronological age of the first data sample and the chronological age of the second data sample. <Step 3> The conversion model is learned by the processing of the learning unit 2 using the pairing result of step 2 (from the second repetition onwards, the pairing result of the previous step 6). <Step 4> The first data is converted into second data by the processing of the conversion unit 3 using the conversion model of step 3, and provisional extended data is obtained. <Step 5> Using the original second data (not resulting from data extension that became pairing candidates in step 2) and the second data after conversion in step 4, a prediction model of biological age, which is an example of a specific trait, is constructed by processing in the model construction unit 4, and the constructed prediction model is applied to each sample of the original second data to obtain an estimate of the biological age of each sample.

[0053] <Step 6> Based on the estimated biological age of the original second data obtained by applying the prediction model in Step 5 and the biological age estimated with high accuracy for the first data in Step 1, the pairing unit 1 performs pairing between the first data and the second data. <Step 7> If the pairing result in step 6 is the same as the previous time, the process ends. If it is different, the process returns to step 3 and repeats steps 3 → 4 → 5 → 6 → 7 in the same way until the end is determined.

[0054] The above steps can be performed regardless of whether or not the original data has trait labels. In steps 2 and 6 of the second embodiment, the only difference is that the criteria used by the pairing unit 1 for pairing are different from those in the first embodiment, and the pairing determination is performed based on whether the criteria are identical or similar, as in the first embodiment.

[0055] 5 is a diagram showing the progress of the process according to the second embodiment, and shows that the following first to fifth processes (the first time also involving the zeroth process) are repeated until a convergence determination is obtained. As shown in FIG. 5, in the second embodiment, the following is repeatedly executed until convergence is achieved. <0th Process> Provisional pairing is performed by the pairing unit 1. (The first pairing is performed using chronological age. Note that this 0th process is omitted from FIG. 5.) <First process> A provisional transformation model is learned by the learning unit 2 using the provisional pairing results. (In the first repetition, the result of the 0th process is used as the provisional pairing, and in the second and subsequent repetitions, the pairing result based on the estimation result of the previous 3rd process, i.e., the pairing result of the previous 4th process, is used as the provisional pairing result.) <Second Process> The first data is converted by the conversion unit 3 using the provisional conversion model to obtain converted second data. <Third process> Using the converted second data and the original second data, the model construction unit 4 constructs a provisional prediction model of the biological age in the second data, and uses the prediction model to obtain an estimate of the biological age for each sample of the original second data. <Fourth process> The pairing unit 1 performs provisional pairing between the first data and the second data based on the estimated biological age for each sample of the second data using a provisional prediction model and the biological age predicted and assigned as a label for each sample of the first data using a prediction model previously constructed with high accuracy. <Fifth Process> Determine whether the pairing of the fourth process has converged, and if it has converged, end the process, but if it has not converged, return to the first process and repeat. In the first iteration, determine whether the pairing result of the zeroth process matches the pairing result of the fourth process, and in the second and subsequent iterations, determine whether the pairing result of the current fourth process matches the pairing result of the previous fourth process.

[0056] If the pairing results in the fifth process match, the results are considered to have converged at that point, and the prediction model and estimated biological age values ​​for the second data obtained at that point in the repetition are used as the final results. The match determination may be based on a perfect match, or a match of, for example, 90% or more based on a threshold determination may be used as the final result. A predetermined upper threshold may be set for the number of repetitions, and if the fifth process reaches the upper threshold number of times that the determination is non-convergent, the estimated value at that point may be used as the final result (assuming that a certain level of accuracy may not necessarily be ensured), or a result may be obtained indicating that the second embodiment should be repeated after increasing the number of samples of the second data used.

[0057] (7) FIG. 6 is a diagram showing an example of the hardware configuration of a general computer device 70. The information processing device 100 can be realized as one or more computer devices 70 having such a configuration. When the information processing device 100 is realized using two or more computer devices 70, information required for processing may be transmitted and received via a network. The computer device 70 includes a CPU (Central Processing Unit) 71 that executes predetermined instructions, a GPU (Graphics Processing Unit) 72 as a dedicated processor that executes some or all of the CPU 71's execution instructions in place of or in cooperation with the CPU 71, a RAM 73 as a main storage device that provides a work area for the CPU 71 (and GPU 72), a ROM 74 as an auxiliary storage device, a communication interface 75, a display 76, an input interface 77 that accepts user input via a mouse, keyboard, touch panel, etc., and a bus BS for transmitting and receiving data among these.

[0058] Each functional unit of the information processing device 100 can be realized by a CPU 71 and / or a GPU 72 that reads a predetermined program corresponding to the function of each unit from a ROM 74 and executes it. Both the CPU 71 and the GPU 72 are types of arithmetic units (processors). Here, when display-related processing is performed, a display 76 also operates in conjunction with the CPU 71 and the GPU 72, and when communication-related processing related to data transmission and reception is performed, a communication interface 75 also operates in conjunction with the CPU 71 and the GPU 72. [Explanation of symbols]

[0059] 10...information processing device, 1...pairing unit, 2...learning unit, 3...conversion unit, 4...model construction unit, 5...model verification unit

Claims

1. a pairing unit that pairs first data, which is DNA methylation rate data of multiple samples acquired according to a first attribute, with second data, which is DNA methylation rate data of multiple samples acquired according to a second attribute, when the samples are determined to have the same or similar traits; a learning unit that constructs a conversion model by learning using the paired data, in which data on the first attribute side of the paired data is input and data on the second attribute side of the paired data is output.

2. The difference between the first attribute and the second attribute is Differences in the somatic tissues from which DNA methylation data are obtained; Differences in the cell types from which DNA methylation data are obtained Genotypic differences in the samples from which DNA methylation data are obtained; Differences in the sex of the samples from which DNA methylation data are obtained, or 2. The information processing device according to claim 1, wherein the information processing device is configured as at least one of the following: a difference in the age of the sample from which the DNA methylation rate data is obtained;

3. The information processing device described in claim 1, characterized in that when at least some of the multiple samples constituting the first data are assigned sample identifiers and at least some of the multiple samples constituting the second data are assigned sample identifiers, the pairing unit pairs samples with the same identifiers as being determined to have the same traits of the samples.

4. The information processing device according to claim 1 , wherein the traits include those related to health, such as cumulative smoking amount.

5. The information processing device according to claim 4, characterized in that the pairing unit, in addition to determining that the traits of the samples are identical or similar, also pairs samples that are determined to have identical or similar attributes such as age and gender.

6. The information processing device according to claim 1 , wherein the learning unit learns the conversion model using deep learning or machine learning.

7. The information processing device according to claim 1, characterized in that the learning unit learns the transformation model using an autoencoder by treating the data of the paired data having the first attribute as having noise superimposed thereon compared to the data of the second attribute, the noise being caused by a difference in attributes.

8. The information processing device according to claim 1, further comprising a conversion unit that applies the conversion model to DNA methylation rate data of at least one sample acquired according to the first attribute to convert the data, thereby obtaining DNA methylation rate data of the sample as if acquired according to the second attribute.

9. The information processing device according to claim 8, further comprising a model construction unit that constructs a prediction model that predicts a target trait from the DNA methylation rate data of a sample by learning using the DNA methylation rate data of multiple samples converted by the conversion unit and a target trait of the sample that is given in advance for each sample.

10. the prediction model is constructed in advance as a first prediction model for the first data, and a target trait is assigned as a label to each sample of the first data using the first prediction model; the prediction model is not constructed in advance for the second data, and a target trait is not assigned as a label to each sample of the second data; A trait different from the target trait is assigned as a label to each sample of the first data and the second data, The pairing unit executes a zeroth process of performing provisional pairing using the different traits, the learning unit executes a first process of constructing a provisional transformation model using the provisionally paired data; the conversion unit executes a second process of converting each sample of the one data into data of the second attribute using the provisional conversion model; the model construction unit constructs the prediction model using the data converted from the first data by the conversion model and the second data, and executes a third process of applying the prediction model to each sample of the second data to estimate a target trait of each sample; the pairing unit executes a fourth process of provisionally pairing the second data and the first data based on the target trait estimated in the second data by the third process and the target trait previously assigned to the first data; 10. The information processing apparatus according to claim 9, wherein a fifth process is executed to determine whether the pairing result by the pairing unit has converged, and the first process to the fifth process are repeated until the convergence determination is obtained.

11. 11. The information processing device according to claim 10, wherein in the fifth process, convergence is determined based on whether or not the pairing results in the zero process and the first fourth process match in the first iteration, and from the second iteration onwards, convergence is determined based on whether or not the pairing results in the fourth process of this iteration match in the fourth process of the previous iteration.

12. a pairing step of pairing first data, which is DNA methylation rate data of multiple samples acquired according to a first attribute, with second data, which is DNA methylation rate data of multiple samples acquired according to a second attribute, when the samples are determined to have the same or similar traits; a learning step of constructing a transformation model that uses the paired data to input data on the first attribute side of the paired data and outputs data on the second attribute side, the learning step being carried out by a computer.

13. a pairing unit that pairs first data, which is DNA methylation rate data of multiple samples acquired according to a first attribute, with second data, which is DNA methylation rate data of multiple samples acquired according to a second attribute, when the samples are determined to have the same or similar traits; a learning unit that constructs a conversion model that uses the paired data to input data on the first attribute side of the paired data and outputs data on the second attribute side by learning using the paired data.

Images