Information processing method, information processing device, program, and method for predicting risk of developing Alzheimer's disease

Abstract

An information processing method according to the present invention includes: a step 1 for detecting a first SNP, which is a mutation of an Alzheimer's disease-related gene, in a genomic DNA sample derived from a subject; and a step 2 for determining, from the first SNP, whether the subject will develop Alzheimer's disease by using a machine learning model that was trained on the basis of a plurality a training datasets labeled with information relating to the development of Alzheimer's disease for a second SNP, which is a mutation of the Alzheimer's disease-related gene detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease.

Description

Information processing method, information processing device, and program

[0001] The present invention relates to an information processing method, an information processing device, and a program. This application claims priority to U.S. Patent No. 63 / 174,500, provisionally filed in the United States on April 13, 2021, the contents of which are incorporated herein by reference.

[0002] Alzheimer's disease (AD) is the most common neurodegenerative disease that causes dementia, with an estimated 30 million people worldwide suffering from it as of 2010. With the arrival of a super-aging society, the number of AD patients will continue to increase, and unless a fundamental treatment is found, it is estimated that the number will reach 60 million by 2030 and exceed 100 million by 2050.

[0003] With the recent advances in biological and genetic research, molecular biological understanding of the pathology of dementia, particularly AD, has advanced. Specifically, genome-wide association studies (GWAS) have investigated the genetic background associated with various human traits and identified more than 50 loci as AD-related genes (see, for example, Non-Patent Document 1). However, at present, only limited symptomatic treatments are available for AD. The background to this is that AD is thought to develop due to the combined action of multiple genes (polygenes), and there is no effective approach to exploring the genetic causes of the pathology, particularly for sporadic AD without a family history, which accounts for 95% of AD patients.

[0004] Sims R et al., “The multiplex model of the genetics of Alzheimer's disease.”, Nature Neuroscience, Vol. 23, pp. 311-322, 2020.

[0005] The present invention has been made in view of the above circumstances, and provides an information processing method, an information processing device, and a program that can predict the risk of developing AD in a subject.

[0006] As a result of intensive research to achieve the above-mentioned object, the inventors produced cerebral cortical neurons using iPS cell cohorts established from sporadic AD patients. Next, using the amyloid beta (Aβ) 42 / 40 ratio (Aβ42 / 40 ratio), one of the pathological indicators of AD, as a phenotype, GWAS (cell GWAS) was performed using the cerebral cortical neurons to search for genetic loci associated with the Aβ42 / 40 ratio. Furthermore, the inventors discovered that the identified genetic loci associated with the Aβ42 / 40 ratio can be used as a polygene dataset to predict a subject's risk of developing AD, leading to the completion of the present invention.

[0007] In this specification, the above-mentioned technique, i.e., a new technology that creates cerebral cortical neurons from iPS cells, decomposes the complex AD pathology into phenotypes (pathological traits) for each cell type and pathology, and reconstructs the actual pathology of AD from the underlying genetic data, is referred to as "Cellular Dissection of Polygenicity (CDiP)" technology.

[0008] That is, the present invention includes the following aspects: (1) An information processing method comprising: Step 1: detecting a first SNP, which is a mutation in an Alzheimer's disease-associated gene, in a genomic DNA sample derived from a subject; and Step 2: for a second SNP, which is a mutation in the Alzheimer's disease-associated gene, detected in a genomic DNA sample derived from a patient with Alzheimer's disease, using a machine learning model trained on multiple training datasets labeled with information related to the onset of Alzheimer's disease to determine whether the subject will develop Alzheimer's disease. (2) The information processing method according to (1), wherein the machine learning model is a random forest including multiple classifiers, and each classifier is trained using a specific training dataset selected from the multiple training datasets on the basis of attribute information and principal component information in genetic information of the patient with Alzheimer's disease. (3) The information processing method according to (1) or (2), wherein the plurality of training datasets include a dataset in which information regarding the onset of Alzheimer's disease is labeled for the imputed genotype of the second SNP estimated from the second SNP using genotype imputation. (4) The information processing method according to any one of (1) to (3), wherein the mutation is one or more of the mutations listed in Tables 1-1 to 1-77.

[0009]

[0010]

[0011]

[0012]

[0013]

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] (5) The information processing method according to (4), wherein the mutations further include one or more mutations listed in Tables 2-1 to 2-9.

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] (6) An information processing device comprising: a detection unit that detects a first SNP, which is a mutation in an Alzheimer's disease-related gene, in a genomic DNA sample derived from a subject; and a determination unit that determines whether or not the subject will develop Alzheimer's disease from a second SNP, which is a mutation in the Alzheimer's disease-related gene, detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease, using a machine learning model trained on multiple training datasets labeled with information related to the onset of Alzheimer's disease.

[0097] (7) A program for causing a computer to execute the following steps: Step 1: detecting a first SNP, which is a mutation in an Alzheimer's disease-associated gene, in a genomic DNA sample derived from a subject; and Step 2: determining whether or not the subject will develop Alzheimer's disease, from a second SNP, which is a mutation in the Alzheimer's disease-associated gene, detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease, using a machine learning model trained on multiple training datasets labeled with information related to the onset of Alzheimer's disease.

[0098] (8) A method for predicting the risk of developing Alzheimer's disease, using the information processing method according to any one of (1) to (5).

[0099] The information processing method, information processing device, and program of the above aspects can predict the risk of AD onset in a subject. Predicting the risk of AD onset can contribute to improving the effectiveness of AD prevention or treatment. Furthermore, it can provide a new therapeutic target for AD.

[0100] 1 is a diagram illustrating an example of the configuration of an information processing device 100 according to the first embodiment. It is a flowchart illustrating a series of runtime processes performed by a processing unit 120 according to the first embodiment. It is a diagram illustrating an example of a prediction model MDL according to the first embodiment. It is a flowchart illustrating a series of training processes performed by a processing unit 120 according to the first embodiment. It is a graph showing the amount of Aβ40 corresponding to apolipoprotein E (APOE) ε4 genotype in Example 1. The n number of patients with the APOE3 / 3 genotype was 44, the n number of patients with the APOE3 / 4 genotype was 44, and the n number of patients with the APOE4 / 4 genotype was 14. "N.S." indicates no significant difference (P>0.05). It is a graph showing the amount of Aβ42 corresponding to an APOE ε4 genotype in Example 1. The n number of patients with the APOE3 / 3 genotype was 44, the n number of patients with the APOE3 / 4 genotype was 44, and the n number of patients with the APOE4 / 4 genotype was 14. 4 / 4。 English: "N.S." indicates that there is no significant difference (P>0.05). Graph showing the Aβ42 / 40 ratio corresponding to the APOE ε4 genotype in Example 1. The n number of patients with the APOE 3 / 3 genotype = 44, the n number of patients with the APOE 3 / 4 genotype = 44, and the n number of patients with the APOE 4 / 4 genotype = 14. Graph showing the protein concentration of iPS cell-derived neurons corresponding to the APOE ε4 genotype in Example 1. The n number of patients with the APOE 3 / 3 genotype = 44, the n number of patients with the APOE 3 / 4 genotype = 44, and the n number of patients with the APOE 4 / 4 genotype = 14. "N.S." indicates that there is no significant difference (P>0.05). Graph showing the Aβ40 amount corresponding to the age at onset of cognitive impairment in Example 1 (N=102). 1 is a scatter plot (N=102) showing the amount of Aβ42 corresponding to the age at onset of cognitive impairment in Example 1. FIG. 2 is a scatter plot (N=102) showing the Aβ42 / 40 ratio corresponding to the age at onset of cognitive impairment in Example 1. FIG. 3 is a plot showing the amount of Aβ40 in the culture supernatant of iPS cell-derived neurons corresponding to gender in Example 1. The number of male patients was 36, and the number of female patients was 66. FIG. 4 is a plot showing the amount of Aβ42 in the culture supernatant of iPS cell-derived neurons corresponding to gender in Example 1.The n number of male patients was 36, and the n number of female patients was 66. This is a plot showing the Aβ42 / 40 ratio in the culture supernatant of iPS cell-derived neurons according to gender in Example 1. The n number of male patients was 36, and the n number of female patients was 66. This is a Manhattan plot of a genome-wide association study of polygenic cellular analysis (CDiP) to identify loci associated with the Aβ42 / 40 ratio without considering the APOE ε4 genotype in Example 1. The x-axis shows chromosomes, and the y-axis shows -log10 (p-value) of all SNPs tested. The top line indicates the Bonferroni-corrected significance threshold (p<5×10). －８ ) shows the Manhattan plot of the genome-wide association study of CDiP to identify loci associated with the Aβ42 / 40 ratio considering the APOE ε4 genotype in Example 1. The x-axis shows chromosomes, and the y-axis shows the -log10 (p-value) of all SNPs tested. The top line shows the Bonferroni-corrected significance threshold (p<5×10 －８ ) is shown. This is a graph showing the results of pathway analysis of 24 genes identified by CDiP using the Aβ42 / 40 ratio in Example 1. The horizontal axis shows p-values. This is a plot showing the phosphorylated tau (p231-tau) / total tau ratio corresponding to the APOEε4 genotype in Example 1. The n number of patients with the APOE3 / 3 genotype was 44, the n number of patients with the APOE3 / 4 genotype was 44, and the n number of patients with the APOE4 / 4 genotype was 14. "N.S." indicates no significant difference (P>0.05). This is a plot showing the p231-tau / total tau ratio in the culture supernatant of iPS cell-derived neurons corresponding to gender in Example 1. The n number of male patients was 36, and the n number of female patients was 66. This is a scatter plot (N=102) showing the p231-tau / total tau ratio corresponding to the age at onset of cognitive impairment in Example 1. Figure 1 shows a Manhattan plot of the genome-wide association study of CDiP to identify loci associated with the p231-tau / total tau ratio considering the APOE ε4 genotype in Example 1. The x-axis shows chromosomes, and the y-axis shows the -log10 (p-value) of all SNPs tested. The top line indicates the Bonferroni-corrected significance threshold (p<5x10 －８) is a Manhattan plot of the genome-wide association study of CDiP to identify loci associated with the p231-tau / total tau ratio without considering the APOE ε4 genotype in Example 1. The x-axis shows chromosomes, and the y-axis shows the -log10 (p-value) of all SNPs tested. The top line indicates the Bonferroni-corrected significance threshold (p<5×10 －８) shows the change in the Aβ42 / 40 ratio due to the knockdown of the identified genes in Example 1. The x-axis shows the level of change in the Aβ42 / 40 ratio compared to the non-siRNA treated control. Values ​​are shown as mean ± standard deviation. *: p<0.05, **: p<0.01, ***: p<0.005, ****: p<0.001. Graph showing the change in the amount of Aβ40 due to the knockdown of the identified genes in Example 1. The x-axis shows the level of change in the amount of Aβ40 compared to the non-siRNA treated control. Values ​​are shown as mean ± standard deviation. *: p<0.05, **: p<0.01, ***: p<0.005, ****: p<0.001. Graph showing the change in the amount of Aβ42 due to the knockdown of the identified genes in Example 1. The x-axis indicates the level of change in Aβ42 levels compared to non-siRNA-treated controls. Values ​​are shown as mean ± standard deviation. * p<0.05, ** p<0.01, *** p<0.005, and **** p<0.001. This graph shows changes in protein levels due to knockdown of identified genes in Example 1. The x-axis indicates the level of change in protein levels compared to non-siRNA-treated controls. Values ​​are shown as mean ± standard deviation. * p<0.05, ** p<0.01, *** p<0.005, and **** p<0.001. This graph shows a comparison of the neuronal expression levels of genes in which siRNA altered the Aβ42 / Aβ40 ratio between brains of Alzheimer's disease patients and non-demented control patients in Example 1. 1 is a graph comparing the neuronal expression levels of genes in which siRNA reduced Aβ42 levels between brains with Alzheimer's disease and brains of non-demented controls in Example 1. 2 is a diagram showing genes that are more highly expressed in AD brains and are potential therapeutic targets, identified using single-cell-based transcriptome data from six AD brains and six control brains that provide transcriptome data for individual cell types in Example 1. 3 is a boxplot showing the relationship between Aβ-positive and Aβ-negative patients and age at onset in Example 2. The n number of Aβ-positive patients was 15, and the n number of Aβ-negative patients was 4.1 is a plot showing the relationship between Aβ-positive patients and Aβ-negative patients in Example 2 and the amount of Aβ40 in the culture supernatant of cerebral cortical neurons induced from human iPS cells. The n number of Aβ-positive patients was 15, and the n number of Aβ-negative patients was 4. FIG. 1 is a plot showing the relationship between Aβ-positive patients and Aβ-negative patients in Example 2 and the amount of Aβ42 in the culture supernatant of cerebral cortical neurons induced from human iPS cells. The n number of Aβ-positive patients was 15, and the n number of Aβ-negative patients was 4. FIG. 1 is a plot showing the relationship between Aβ-positive patients and Aβ-negative patients in Example 2 and the Aβ42 / 40 ratio in the culture supernatant of cerebral cortical neurons induced from human iPS cells. The n number of Aβ-positive patients was 15, and the n number of Aβ-negative patients was 4. 1 is a graph showing the results of predicting brain Aβ deposition using covariates (age, sex, and genotype of the APOE-ε4 allele) (left graph) or the covariates and a genotype set identified by CDiP (right graph) in Example 2.

[0034] FIG. 1 is a graph showing the results of predicting the amount of Aβ (1-42) in cerebrospinal fluid (CSF) using covariates (age, sex, and genotype of the APOE-ε4 allele) (left graph) or the covariates and a genotype set identified by CDiP (right graph) in Example 2.

[0035] FIG. 1 is a graph showing the results of predicting the amount of total tau (t-tau) in CSF using covariates (age, sex, and genotype of the APOE-ε4 allele) (left graph) or the covariates and a genotype set identified by CDiP (right graph) in Example 2. 10 is a graph showing the results of predicting the amount of phosphorylated tau (p-tau) in CSF using covariates (age, sex, and genotype of the APOE-ε4 allele) in Example 2 (left graph), or the covariates and a genotype set identified by CDiP (right graph).

[0101] An embodiment of the present invention will be described. The following embodiment is an example for explaining the present invention, and is not intended to limit the present invention to only this embodiment. The present invention can be implemented in various forms without departing from the gist of the present invention.

[0102] <Mutations in Alzheimer's Disease (AD)-Related Genes> The inventors conducted GWAS (cell GWAS) using cerebral cortical neurons induced from iPS cells established from sporadic AD patients, with the Aβ42 / 40 ratio, which is one of the indicators of AD pathology, as a phenotype, and as shown in the Examples described below, they found that, among mutations in AD-related genes, the mutations listed in Tables 1-1 to 1-77 above are mutations associated with the Aβ42 / 40 ratio. Furthermore, as shown in the Examples below, when analyzing mutations in AD-related genes containing one or more mutations listed in Tables 1-1 to 1-77 above, and the APOE4 genotype believed to be involved in age, sex, and Aβ accumulation in the brain, compared with when analyzing only the APOE4 genotype believed to be involved in age, sex, and Aβ accumulation in the brain without analyzing mutations in AD-related genes containing one or more mutations listed in Tables 1-1 to 1-77 above, the AUC score, which is one of the indicators of prediction accuracy, is higher when analyzing mutations in AD-related genes containing one or more mutations listed in Tables 1-1 to 1-77 above, and the APOE4 genotype believed to be involved in age, sex, and Aβ accumulation in the brain. Specifically, using SNP information of sporadic AD patients, it is possible to predict whether intracerebral Aβ accumulation will occur with an accuracy of AUC = 0.76 ± 0.050, and using SNP information of sporadic AD patients, it is possible to predict whether abnormal test values ​​of Aβ will occur in cerebrospinal fluid with an accuracy of AUC = 0.73 ± 0.059. These intracerebral Aβ accumulation and abnormal test values ​​of Aβ in cerebrospinal fluid are almost consistent with clinical AD diagnosis. Therefore, the prediction of AD onset risk using the information processing device and information processing method of this embodiment can be performed with an accuracy of about 0.7 (more specifically, about 0.73 to 0.76). In sporadic AD, rather than familial AD (genetic AD), high-precision prediction with an AUC in the above numerical range has not been possible with previous methods. In contrast, the information processing device and information processing method of this embodiment can provide a method for determining risk with high accuracy or high predictive power by analyzing a SNP set including one or more mutations listed in Tables 1-1 to 1-77 above to determine the risk of AD.In other words, the information processing device and information processing method of this embodiment can be said to be an apparatus and method for predicting the risk of developing AD. Furthermore, the information processing device and information processing method of this embodiment can predict the risk of developing AD in subjects, including subjects suspected of having sporadic AD with no family history.

[0103] Furthermore, for diseases such as AD that are caused by multiple genes, stratification of patients into subtypes based on common characteristics is extremely significant in the prevention and treatment of the disease, as effective means for prevention or treatment may differ depending on the subtype. As shown in the examples described below, the information processing device and information processing method of this embodiment can also contribute to stratification of AD, which can also contribute to precision medicine.

[0104] As used herein, "risk of Alzheimer's disease (AD)" refers to the possibility of developing Alzheimer's disease, such as susceptibility or resistance to AD. "Assessing risk" includes, for example, outputting the possibility of developing AD now or in the future as several levels, or outputting it as a numerical value. Determining the risk of AD includes evaluating genetic factors or genetic susceptibility to the disease, such as whether a person is prone to or resistant to AD.

[0105] It should be noted that when assessing the risk of AD, it does not matter whether the subject whose risk of AD is being assessed actually suffers from (has developed) AD at the time of assessing the risk of AD.

[0106] One or more of the mutations listed in Tables 1-1 to 1-77 above can be used, but the mutations listed in Tables 1-1 to 1-77 above are SNPs whose association with AD has not been recognized in the past. That is, AD is thought to develop due to the combined action of polygenes, and by analyzing two or more mutations listed in Tables 1-1 to 1-77 above as a set of SNPs rather than individually analyzing the mutations listed in Tables 1-1 to 1-77 above, the risk of AD can be determined with higher accuracy. Therefore, it is preferable to use an SNP set including all the mutations listed in Table 4 below (mutations that are particularly highly associated with AD from the perspective of the Aβ42 / 40 ratio among the mutations listed in Tables 1-1 to 1-77 above), and among the mutations listed in Tables 3-1 to 3-77 below, a SNP set containing a p-value of 5 × 10 －８ It is more preferable to use a SNP set containing mutations that are less than 100% of the total number of mutations, and it is even more preferable to use a SNP set containing all of the mutations listed in Tables 3-1 to 3-77.

[0107] Furthermore, as shown in the Examples below, the inventors have found that among mutations in AD-related genes, the mutations listed in Tables 2-1 to 2-9 above are mutations associated with the phosphorylated tau / total tau ratio. Therefore, although it is possible to further include one or more of the mutations listed in Tables 2-1 to 2-9 above, since AD ​​is thought to develop due to the combined action of polygenes, in addition to the mutations associated with the Aβ42 / 40 ratio, it is preferable to use a SNP set that further includes the mutations listed in Table 6 below among the mutations associated with the phosphorylated tau / total tau ratio (among the mutations listed in Tables 2-1 to 2-9 above, mutations that are particularly highly associated with AD from the perspective of the phosphorylated tau / total tau ratio), and it is more preferable to use a SNP set that further includes all of the mutations listed in Tables 2-1 to 2-9 above.

[0108] For each SNP listed in each table in this specification, the rs number, the chromosome number on which each SNP is located (in the case of a sex chromosome, this is indicated by X or Y), and the location of each SNP on the chromosome are listed. Information on the base sequence and disease associated with each SNP can be obtained, for example, by searching the NCBI SNP Database based on the rs number. This information can be accessed through the database and is incorporated herein by reference. The chromosomal location of each SNP corresponds to assembly genome version GRCh37.

[0109] As shown in each table, each SNP can be identified by referring to the base sequence identified by the rs number, but if the rs number described herein is merged with another rs number and a new rs number is assigned, the relevant rs number in this specification includes the rs number after the merger and the other rs numbers that are merged. Also, if the rs number described herein is a number assigned by merging multiple rs numbers, the relevant rs number in this specification includes the other original rs numbers.

[0110] Furthermore, the base sequence indicated by each rs number for the SNP is shown as a specific base sequence by referring to a database such as the NCBI SNP Database, but the base sequence in the part other than the corresponding SNP in the base sequence may be changed due to racial differences, etc.

[0111] In the information processing method, information processing device, and program of this embodiment, the race and sex of the subject are not limited.

[0112] An information processing method, an information processing device, and a program to which the present invention is applied will be described below with reference to the accompanying drawings.

[0113] <First embodiment> [Overall configuration] Fig. 1 is a diagram showing an example of the configuration of an information processing device 100 according to the first embodiment. As shown in Fig. 1, the information processing device 100 includes, for example, a detection unit 110, a processing unit 120, and a storage unit 130.

[0114] (Detection Unit) The detection unit 110 detects SNPs (hereinafter referred to as first SNPs) that are mutations in Alzheimer's disease (AD)-related genes in a genomic DNA sample derived from a subject (step 1).

[0115] The genomic DNA sample derived from a subject can be cells or tissue collected from the subject's living body, and is not particularly limited as long as it contains nucleated cells, but examples include blood, cerebrospinal fluid, lymph, hair, etc. Among these, blood is preferably used because of its low invasiveness, and examples of nucleated cells derived from blood include peripheral blood mononuclear cells, etc. For SNP detection, genomic DNA isolated from these samples by conventional methods may be used directly, or the isolated genomic DNA may be amplified and the amplified genomic DNA may be used.

[0116] The method for extracting genomic DNA is not particularly limited, and extraction can be performed using a known method. Examples include the phenol / chloroform method and the cetyltrimethylammonium bromide (CTAB) method. DNA extraction can also be performed using a commercially available kit. Examples of such kits include the Wizard Genomic DNA Purification Kit (manufactured by Promega).

[0117] The detection unit 110 is configured with a device used in typical genetic polymorphism analysis, such as a DNA microarray, a conventional sequencer or a next-generation sequencer (NGS; Next Generation Sequencer), or a nucleic acid amplification device such as a polymerase chain reaction (PCR) device.

[0118] SNPs can be detected using known SNP detection methods using the above-mentioned devices, such as direct sequencing, PCR, restriction fragment length polymorphism (RFLP), hybridization, TaqMan (registered trademark) PCR (hereinafter, the term "registered trademark" will be omitted), mass spectrometry, and the like.

[0119] Direct sequencing is performed by cloning a region containing the SNP into a vector or amplifying it by PCR and determining the nucleotide sequence of the region. Cloning can be performed by screening a cDNA library using an appropriate probe. Alternatively, cloning can be performed by amplifying the SNP by PCR using appropriate primers and ligating it into an appropriate vector. Subcloning into another vector is also possible, but is not limited to these. Examples of vectors that can be used include commercially available plasmid vectors, viral vectors, artificial chromosome vectors, and cosmid vectors, such as pBlue-Script SK(+) (Stratagene), pGEM-T (Promega), pAmp (Gibco-BRL), p-Direct (Clontech), and pCR2.1-TOPO (Invitrogene). Known methods can be used to determine the nucleotide sequence, including, but not limited to, manual sequencing using radioactive marker nucleotides and automated sequencing using dye terminators. Based on the base sequence thus obtained, it is determined whether or not the sample contains the SNP.

[0120] PCR is performed using oligonucleotide primers (hereinafter sometimes referred to as "SNP detection primers") that hybridize only to sequences containing SNPs. Since multiple SNPs exist, the SNP detection primers may be primers capable of detecting all SNPs alone, or two or more primers capable of detecting each SNP may be used in combination. The sample DNA is amplified using these primers. If the SNP detection primers produce a PCR product, the sample contains the SNP. If no PCR product is produced, it indicates that the sample does not contain the SNP.

[0121] In the RFLP method, a region containing the SNP in a sample is first amplified by PCR. The PCR product is then cleaved with a restriction enzyme appropriate for the region containing the SNP. The PCR products digested with the restriction enzymes are separated by gel electrophoresis and visualized by ethidium bromide staining. The fragment lengths can be compared with molecular weight markers and, as a control, with the PCR product that has not been treated with the restriction enzyme, to detect the presence of the SNP in the sample.

[0122] Hybridization is a method for determining the presence or absence of SNPs in a sample based on the ability of DNA in the sample to hybridize with complementary DNA molecules (e.g., oligonucleotide probes). This hybridization method can be performed using various hybridization and detection techniques, such as colony hybridization, plaque hybridization, and Southern blot. For detailed procedures for hybridization, see "Molecular Cloning, A Laboratory Manual 3rd ed." (Cold Spring Harbor Press (2001); especially Sections 6-7), "Current Protocols in Molecular Biology" (John Wiley & Sons (1987-1997); especially Sections 6.3-6.4), and "DNA Cloning 1: Core Techniques, A Practical Approach 2nd ed." (Oxford University (1995); for hybridization conditions, see Section 2.10 in particular. Furthermore, hybridization can also be detected using a DNA chip. In this method, an oligonucleotide probe specific to the SNP is designed and attached to a solid support. Then, DNA in the sample is brought into contact with the DNA chip to detect hybridization.

[0123] The TaqMan PCR method uses a SNP-specific TaqMan probe and Taq polymerase to simultaneously detect an SNP and amplify a region containing the SNP. The TaqMan probe is an oligonucleotide of approximately 20 bases, labeled with a fluorescent substance at the 5' end and a quencher at the 3' end, and is designed to hybridize to a target SNP site. Taq polymerase has 5'->3' nuclease activity. When a region containing the SNP site is amplified using PCR primers designed to amplify the region containing the SNP site in the presence of these TaqMan probes and Taq polymerase, the TaqMan probe hybridizes to the target SNP site in template DNA in parallel with the amplification. When the extension reaction from the forward primer reaches the TaqMan probe hybridized to the template, the 5' to 3' nuclease activity of Taq polymerase cleaves the fluorescent substance attached to the 5' end of the TaqMan probe. As a result, the released fluorescent substance is no longer affected by the quencher and emits fluorescence. SNP detection is possible by measuring the fluorescence intensity.

[0124] Examples of methods using mass spectrometry include SNP typing methods that utilize MALDI-TOF / MS in combination with primer extension. This method allows for high-throughput analysis and involves the following steps: 1) PCR, 2) purification of the PCR product, 3) primer extension reaction, 4) purification of the extension product, 5) mass spectrometry, and 6) genotype determination. First, a region containing the target SNP site is amplified from genomic DNA by PCR. PCR primers are designed so that they do not overlap with the bases at the SNP site. The primers are then purified by enzymatic removal using exonuclease and shrimp alkaline phosphatase or by ethanol precipitation. Next, a primer extension reaction is performed using a genotyping primer designed so that its 3' end is directly adjacent to the SNP site. The PCR product is denatured at high temperature, and excess genotyping primer is added and annealed. Adding ddNTPs and DNA polymerase to the reaction system and thermal cycling produces oligomers one base longer than the genotyping primer. The oligomers produced by this extension reaction vary depending on the allele, depending on the design of the genotyping primer. The purified extension reaction products are analyzed by mass spectrometry.

[0125] Other detection methods include a high-throughput SNP typing method that applies single-molecule fluorescence analysis. For example, the MF20 / 10S (manufactured by Olympus) is a system that employs this method. Specifically, it uses a confocal laser optical system and a highly sensitive photodetector to measure and analyze the translational diffusion time at the single-molecule level of fluorescently labeled primers amplified by PCR using complementary and non-complementary primers in an ultra-small volume of approximately 1 femtoliter (1 / 1000 trillionth of a liter).

[0126] Another high-throughput typing method is the DNA chip method, which consists of a substrate on which a variety of DNA probes are immobilized in an array, and labeled DNA samples are hybridized on the chip, and fluorescent signals from the probes are detected.

[0127] An example of an SNP typing method using a gene amplification method other than PCR is the Snipper method. This method is an SNP typing method that applies the RCA (rolling circle amplification) method, which is a DNA amplification method in which a circular single-stranded DNA is used as a template and a DNA polymerase moves along it to synthesize a complementary strand of DNA. The probe is an oligo DNA of 80 to 90 bases in length, and contains sequences of 10 to 20 bases in length at both ends that are complementary to the 5'-end and 3'-end vicinity of the target SNP site, respectively, and is designed to anneal to the target DNA to form a circular DNA. In addition, the 3'-end of the probe is designed to have a sequence complementary to the target SNP site. If the 3'-end of the probe is completely complementary to the target SNP site, the probe will be circularized, but if the 3'-end of the probe is mismatched, the probe will not be circularized. The probe also has a backbone sequence of 40 to 50 bases in length, and contains sequences complementary to two types of RCA amplification primers.

[0128] Other examples of SNP typing methods that utilize gene amplification methods other than the PCR method include typing methods that utilize the UCAN method and the LAMP method.

[0129] The UCAN method is an application of the ICAN method, an isothermal gene amplification method developed by Takara Bio. The UCAN method uses a DNA-RNA-DNA chimeric oligonucleotide (DRD) as a primer precursor. This DRD primer precursor has modified DNA at the 3' end to prevent replication of the template DNA by DNA polymerase, and is designed so that the RNA portion binds to the SNP site. When this DRD primer precursor is incubated with a template, the coexisting RNase H cleaves the RNA portion of the paired DRD primer only if the DRD primer and template are perfectly matched. This removes the modified DNA from the 3' end of the primer, creating a new one, allowing the extension reaction by DNA polymerase to proceed and amplify the template DNA. On the other hand, if the DRD primer and template DNA do not match, RNase H does not cleave the DRD primer, and DNA amplification does not occur. After the perfectly matched DRD primer precursor is cleaved by RNase H, the amplification reaction proceeds via the ICAN reaction mechanism.

[0130] The LAMP method is a gene isothermal amplification method developed by Eiken Chemical. Six regions of a target gene (F3c, F2c, and F1c from the 3' end, and B3, B2, and B1 from the 5' end) are defined, and amplification is performed using four primers (FIP primer, F3 primer, BIP primer, and B3 primer) for these six regions. For typing purposes, only the target SNP site (one base) is required between F1 and B1, and the FIP and BIP primers are designed so that one base of the SNP is located at their 5' ends. In the absence of an SNP, a DNA synthesis reaction occurs from the dumbbell structure, which is the starting structure for the LAMP method, and the amplification reaction proceeds continuously. In the presence of an SNP, the DNA synthesis reaction from the dumbbell structure does not occur, and the amplification reaction does not proceed.

[0131] The Invader method does not use nucleic acid amplification, but instead uses two types of non-fluorescently labeled probes (allele probe, invader probe), one type of fluorescently labeled probe (FRET probe), and the endonuclease Cleavase. The allele probe has a sequence complementary to the template DNA from the SNP site to the 3' end, and a sequence unrelated to the template DNA called a flap at the 5' end of the probe. The invader probe has a sequence complementary to the SNP site to the 5' end of the template DNA, and the base at the portion corresponding to the SNP site can be any base. The FRET probe has a sequence complementary to the flap sequence at the 3' end. One 5' end is labeled with a fluorescent dye and a quencher, but the FRET probe is designed to form a double strand within the molecule and is usually quenched. When these are reacted with template DNA, when the allele probe forms a double strand with the template DNA, the 3' end (any base portion) of the invader probe invades the SNP site. Cleavase recognizes the structure into which the base has invaded and cleaves the flap portion of the allele probe. Next, when this released flap binds to the complementary sequence of the FRET probe, the 3' end of the flap invades the intramolecular double-stranded portion of the FRET probe. As in the case of the allele probe and invader probe, Cleavase recognizes the structure in which the base of the flap has invaded the FRET probe and cleaves the fluorescent dye of the FRET probe. The fluorescent dye separates from the quencher, generating fluorescence. If the allele probe does not match the template DNA, the specific structure recognized by Cleavase is not formed, and the flap is not cleaved.

[0132] When using primers to detect SNPs, they are designed to be primers suited to the region to be amplified and the typing method. For example, it is preferable that the region can be completely amplified, and the sequence can be designed based on the sequences near both ends of the region. Primer design techniques are well known in the art, and primers that can be used in the method of this embodiment are designed to satisfy conditions for specific annealing, for example, to have a length and base composition (melting temperature) that enable specific annealing. The length of the region to be amplified is not limited as long as it does not interfere with typing, and may be increased or decreased as appropriate depending on the detection method. Furthermore, while a portion of the amplified region contains an SNP site, there is no restriction on the position of this site within the amplified region, and it may be positioned at an appropriate position according to the detection method (typing method). Therefore, when designing primers, the positional relationship between the primer and the SNP site can be freely designed according to the detection method, and primers can be designed taking into account the characteristics of the typing method as long as they hybridize to a region containing the SNP to be detected (e.g., a continuous region of 50 to 500 bases in length). The length required for a primer to function is preferably 10 to 100 bases, more preferably 15 to 50 bases, and even more preferably 15 to 30 bases. Furthermore, during design, it is preferable to confirm the melting temperature (Tm) of the primer, which is the temperature at which 50% of any nucleic acid strand forms a hybrid with its complementary strand. In order for the template DNA and the primer to form a double strand and anneal, the annealing temperature must be optimized. However, a temperature that is too low is undesirable because it can cause nonspecific reactions. Known primer design software can be used to confirm the Tm.

[0133] When a probe is used to detect SNP, it is designed so that the probe recognizes the SNP site. In the probe design, the SNP site may be recognized at any location within the probe according to the typing method, and may be recognized at the end of the probe depending on the typing method. When a polynucleotide for SNP detection is used as the probe, the length of the base sequence complementary to genomic DNA is usually 15 to 200 bases, preferably 15 to 100 bases, more preferably 15 to 50 bases, but may be longer or shorter depending on the typing method.

[0134] (Processing Unit) The processing unit 120 uses a machine learning model trained on multiple training datasets labeled with information related to the onset of AD for SNPs, which are mutations in AD-related genes detected in genomic DNA samples from patients with AD, to determine whether or not a subject will develop AD from the SNPs detected by the detection unit 110 (i.e., "first SNPs") (Step 2). Hereinafter, the SNPs included in the training datasets will be referred to as "second SNPs."

[0135] The processing unit 120 includes, for example, an acquisition unit 121 , a feature conversion unit 122 , a determination unit 123 , an output control unit 124 , and a learning unit 125 .

[0136] The components of the processing unit 120 are realized by a processor such as a central processing unit (CPU) or a graphics processing unit (GPU) executing a program stored in the storage unit 130. Some or all of the components of the processing unit 120 may be realized by hardware (circuitry) such as a large-scale integration (LSI), an application-specific integrated circuit (ASIC), or a field-programmable gate array (FPGA), or may be realized by a combination of software and hardware.

[0137] (Storage Unit) The storage unit 130 is realized by a storage device such as a hard disk drive (HDD), a flash memory, an electrically erasable programmable read-only memory (EEPROM), a read-only memory (ROM), or a random access memory (RAM). The storage unit 130 stores various programs such as firmware and application programs as well as model information 131. The model information 131 will be described later.

[0138] [Runtime Processing Flow] A series of runtime processing flows by the processing unit 120 according to the first embodiment will be described below with reference to a flowchart. Runtime refers to a state in which an already trained prediction model MDL is used. Fig. 2A is a flowchart showing a series of runtime processing flows by the processing unit 120 according to the first embodiment. The processing of this flowchart may be repeated at a predetermined interval, for example.

[0139] First, the acquisition unit 121 acquires detection data of a first SNP, which is a mutation in an Alzheimer's disease-related gene, in a genomic DNA sample derived from a subject from the detection unit 110 (step S100). The acquired detection data of the first SNP can also be referred to as genotype data of the AD-related gene possessed by the subject (hereinafter, sometimes referred to as a "first SNP set" or "subject's genotype data").

[0140] Next, the feature conversion unit 122 converts the data of the first SNP acquired by the acquisition unit 121 into features that can be input to a model (step S101). The feature here is, for example, a parameter indicating whether the genotype data of the subject for each SNP is homozygous (AA), homozygous (BB), or heterozygous (AB). Genotypes are generally expressed using nucleotides such as "GG" indicating that both SNPs on homologous chromosomes are G (guanine) or "AG" indicating that one SNP is G (guanine) and the other is A (adenine). Therefore, the genotype data of the subject is converted into parameters that can be input to a model using a second SNP (genotype data of AD-related genes possessed by AD patients (hereinafter, sometimes referred to as a "second SNP set" or "genotype data of AD patients")), which is a mutation in an AD-related gene detected in a genomic DNA sample derived from a patient with AD. However, if the model is such that such conversion to parameters is not necessary, then the conversion is not required.

[0141] The subject's genotype data can be converted into features by, for example, assigning a value to the subject's genotype data for each SNP included in the second SNP set. For example, for each SNP, a value (e.g., 0, 1, or 2) is assigned to the SNP depending on whether the subject's genotype data corresponds to homozygous type (AA), homozygous type (BB), or heterozygous type (AB). This allows the subject's genotype data to be converted into features. Note that, in the following, an example will be described in which the value assigned to each SNP is 0, 1, or 2, but the values ​​assigned to SNPs are not limited to the three values ​​0, 1, and 2.

[0142] The value associated with the zygosity can be determined for each SNP. For example, a certain SNP may be associated with a value of 2 when the subject's genotype data is homozygous (AA), a value of 1 when the subject's genotype data is heterozygous (AB), and a value of 0 when the subject's genotype data is homozygous (BB). Another SNP may be associated with a value of 2 when the subject's genotype data is heterozygous (AB), a value of 1 when the subject's genotype data is homozygous (AA), and a value of 0 when the subject's genotype data is homozygous (BB). Alternatively, a value of 2 may be associated with a subject's genotype data when the subject's genotype data is homozygous (BB), a value of 1 when the subject's genotype data is heterozygous (AB), and a value of 0 when the subject's genotype data is homozygous (AA).

[0143] As described above, the genotype data of the subject can be converted into features. The values ​​used for the correspondence in this conversion to features can be determined arbitrarily. For example, for each SNP, a value of 2 can be associated with a genotype highly associated with AD, and a value of 1 or 0 can be associated with a genotype low associated with AD.

[0144] Next, the determination unit 123 inputs the feature converted from the data of the first SNP by the feature conversion unit 122 into the prediction model MDL indicated by the model information 131 (step S102).

[0145] The model information 131 is information (a program or a data structure) that defines a prediction model MDL for determining the risk of AD from the genotype data of a subject.

[0146] The prediction model MDL can be realized by various other models, such as a logistic regression model, a multilayer perceptron, a neural network such as a convolutional neural network (CNN) or a recurrent neural network (RNN), a support vector machine using an arbitrary kernel function such as a Gaussian kernel, a random forest modeled as a regression tree, a model using multiple regression analysis or a hidden Markov model, a statistical model, or a probabilistic model. A model that combines various models to perform a comprehensive judgment can also be adopted. For example, the prediction model MDL may be a random forest including multiple classifiers. The following description assumes that the prediction model MDL is a random forest, as an example.

[0147] FIG. 2B is a diagram showing an example of a prediction model MDL according to the first embodiment. The prediction model MDL includes, for example, N classifiers ML-1 to ML-N. Each classifier WL is a weak learner that has been trained in advance to output a score indicating the likelihood that a subject will develop AD as a likelihood or probability when a feature converted from data on the first SNP is input. The classifiers WL are in a parallel relationship with each other. This method of combining multiple weak learners to generate a single learning model is called ensemble learning.

[0148] For example, the prediction model MDL normalizes the score of each classifier WL, which is a weak learner, and outputs the normalized score. The normalization of the score is shown in Equation (1).

[0149]

[0150] where S represents the normalized score and si represents the score of the i-th classifier WL. ｉand S are, for example, two-dimensional vectors (= [P1, P2]) whose elements are the probability P1 of developing AD and the probability P2 of not developing AD. As shown in formula (1), the prediction model MDL may normalize the score by dividing the sum of the scores of all classifiers WL by N, which is the total number of classifiers WL. By using ensemble learning in this way, it is possible to improve the prediction accuracy of AD onset for unknown (unlearned) data that was not used in the training described below.

[0151] 2B, the prediction model MDL is a combination of N classifiers WL, but is not limited to this. For example, the prediction model MDL may be a single classifier WL.

[0152] Returning to the description of the flowchart in Fig. 2A, the determination unit 123 then determines whether the score (normalized score) output by the prediction model MDL is equal to or greater than a threshold value (step S103).

[0153] If the score is greater than or equal to the threshold, the judgment unit 123 judges that the subject has a high probability of developing AD (step S104), and if the score is less than the threshold, the judgment unit 123 judges that the subject has a low probability of developing AD (step S105).

[0154] Next, the output control unit 124 outputs the determination result by the determination unit 123 (e.g., information indicating the likelihood of developing Alzheimer's disease) (step S106). For example, the output control unit 124 may transmit the determination result to an external terminal device (not shown) via a communication interface. The communication interface may be, for example, a network card such as a network interface card (NIC) or a wireless communication module. Furthermore, if the information processing device 100 is equipped with a display (not shown), the output control unit 124 may display the determination result on the display.

[0155] [Training Processing Flow] The flow of a series of training processes by the processing unit 120 according to the first embodiment will be described below with reference to a flowchart. Training is a state in which a prediction model MDL to be used at runtime is trained. Fig. 2C is a flowchart showing the flow of a series of training processes by the processing unit 120 according to the first embodiment.

[0156] First, the learning unit 120 generates a training dataset for training the prediction model MDL (step S200). For example, the training dataset is a dataset in which genotype data of AD-related genes possessed by healthy individuals or AD patients are labeled with information regarding the onset of AD in the subject (e.g., AD onset positive or negative). If the second SNP set includes some unknown SNPs, the attributed genotypes can be estimated using genotype imputation.

[0157] For example, genotype data of AD-related genes possessed by healthy individuals or AD patients is obtained. The genotype data of AD-related genes possessed by healthy individuals is labeled with information indicating that the individual does not have AD (e.g., a score of 0.0), and the genotype data of AD-related genes possessed by AD patients is labeled with information indicating that the individual will develop AD (e.g., a score of 1.0).

[0158] In this way, when the genotype data of multiple labeled AD-related genes is generated as a training dataset, the feature conversion unit 122 converts the genotype data of the AD-related genes included in the training dataset into features (step S201).

[0159] Next, the learning unit 125 divides the multiple features converted from the genotype data of the AD-related genes in the training dataset by the feature conversion unit 122 into training features (training samples) and validation features (test samples), and inputs the training features to the i-th classifier WL-i among the N classifiers WL included as weak learners in the prediction model MDL (step S202).

[0160] The learning unit 125 may use principal component analysis to separate multiple features converted from the genotype data of AD-related genes in the training dataset, i.e., population features, into training features (training samples) and validation features (test samples). For example, the learning unit may select training features (training samples) from the population features based on principal component information among the attribute information and genetic information of healthy individuals or AD patients who are the providers of the AD-related genes. The attribute information of healthy individuals or AD patients may include, for example, information such as age and gender. The genetic information of healthy individuals or AD patients may include, for example, information on whether or not they have the APOE ε4 genotype, as well as other information. The training features (training samples) selected based on the principal components are an example of a "specific training dataset."

[0161] Next, the learning unit 125 outputs the output result, i.e., the score s ｉ (step S203).

[0162] Next, the learning unit 125 calculates the score s obtained from the i-th classifier WL-i. ｉ and the error (also called loss) between the score labeled with the training feature (step S204).

[0163] Next, the learning unit 125 determines the parameters of the i-th classifier WL-i so as to reduce the error (step S205).

[0164] Next, the learning unit 125 determines whether the learning for the i-th classifier WL-i has been repeated a predetermined number of times E (step S206). If the predetermined number of times E has not been reached, the process returns to S202, and the same training features as those used for learning in the previous process are input to the i-th classifier WL-i, thereby repeating the learning of the i-th classifier WL-i. At this time, the learning unit 125 stores the parameters updated by the learning in the storage unit 130, and inputs the training features to the i-th classifier WL-i whose parameters have been initialized. As a result, E classifiers WL-i with different parameters are generated by the time the learning for the i-th classifier WL-i has been repeated a predetermined number of times E.

[0165] On the other hand, when the learning for the i-th classifier WL-i has been performed a predetermined number of times E, the learning unit 125 inputs the verification features to each of the E i-th classifiers WL-i (step S207).

[0166] Next, the learning unit 125 selects the classifier WL-i with the highest prediction accuracy from among the E i-th classifiers WL-i (step S208). For example, the learning unit 125 selects the score s obtained when the validation features are input from among the E i-th classifiers WL-i. ｉ The classifier WL-i with the smallest error between the score of the training data and the classifier WL-i is selected.

[0167] Next, the learning unit 125 determines whether or not all of the N classifiers WL included as weak learners in the prediction model MDL have been learned (step S209). If the N classifiers WL have not yet been learned, the process returns to S202, and the (i+1)-th classifier WL-(i+1) is learned based on the training features.

[0168] On the other hand, when the learning unit 125 has completed learning for all of the N classifiers WL, the processing of this flowchart ends.

[0169] According to the first embodiment described above, the information processing device 100 inputs the features converted from the genotype data of the AD-related genes possessed by the subject into a prediction model MDL trained based on a training dataset in which the features converted from the genotype data of the AD-related genes are labeled with information indicating whether the onset of AD is positive or negative, and predicts that the subject will develop AD based on the output result of the prediction model MDL to which the features have been input, thereby making it possible to accurately predict whether the subject will develop AD in the future.

[0170] Generally, it is difficult to predict AD risk in subjects who have not yet developed AD, such as infants and young people. In contrast, in this embodiment, a prediction model MDL including multiple classifiers WL realized by a machine learning model is used, so it is expected that a weight indicating that a specific SNP in the genotype data of AD-related genes has a positive correlation with AD risk can be calculated as a score. As a result, it is possible to early predict AD risk in subjects who have not yet developed AD, such as infants and young people.

[0171] Furthermore, according to the first embodiment described above, since ensemble learning is performed on a plurality of classifiers WL included as weak learners in the prediction model MDL, it is possible to generate a prediction model MDL with high prediction accuracy.

[0172] <Modification of First Embodiment> A modification of the first embodiment will be described below. In the above-described first embodiment, the training dataset is described as data in which genotype data of AD-related genes possessed by healthy individuals or AD patients are labeled with a score indicating whether or not the individual will develop AD. However, this is not limited to this. For example, the training dataset may be data in which genotype data of AD-related genes possessed by healthy individuals or AD patients are labeled with the age of onset of AD in addition to the above-mentioned score. Using such a training dataset, when genotype data of AD-related genes is input, the learning unit 125 trains the prediction model MDL so that it outputs a three-dimensional vector (= [P1, P2, t]) whose elements are the probability P1 of developing AD, the probability P2 of not developing AD, and the age t of AD onset. The determination unit 123 predicts the age at which the subject will develop AD based on the t element of the vector output by the prediction model MDL.

[0173] Furthermore, the label is not limited to a score indicating the presence or absence of AD or the age at onset, but may also include attributes of the subject who is the source of the genotype data of AD-related genes. Attributes may include various information such as gender, weight, height, lifestyle habits, the presence or absence of illness, and family medical history. Genetic information of known AD-related genotypes, such as the APOE ε4 genotype, may also be included. By training the prediction model MDL using genotype data of AD-related genes associated with such labels, a prediction model MDL with even higher prediction accuracy can be generated. As a result, by inputting the subject's attributes in addition to the genotype data of AD-related genes into the prediction model MDL at runtime, the subject's future risk of AD can be predicted with even greater accuracy.

[0174] The determination results obtained as described above are also used to assist AD specialists in diagnosing AD. That is, the information processing device and information processing method of this embodiment can also be said to be an AD diagnosis support device and diagnosis support method.

[0175] <Other Embodiments> In one embodiment, the present invention provides a processor configured to execute the instructions of the information processing method described above, specifically: detecting a first SNP, which is a mutation in an Alzheimer's disease-associated gene, in a genomic DNA sample derived from a subject; and determining, for a second SNP, which is a mutation in the Alzheimer's disease-associated gene detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease, from the first SNP using a machine learning model trained on multiple training datasets labeled with information related to the onset of Alzheimer's disease.

[0176] EXAMPLES The present invention will be described in more detail below with reference to examples and comparative examples, but the present invention is not limited to these examples.

[0177] Methods: Establishment of iPS cell cohorts from peripheral blood cells of sporadic AD patients. This study was approved by the Institutional Review Board of the iPS Cell Research and Application Center, Kyoto University (approval numbers: CiRA19-05 and CiRA20-14). For the establishment of iPS cells from human peripheral blood mononuclear cells (PBMCs), PBMCs from Alzheimer's disease (AD) patients were collected in accordance with a research project approved by the Institutional Review Board of the Graduate School of Medicine, Kyoto University (approval numbers: R0091, G259, and G0722). Written informed consent was obtained from all participants in this study. Human cDNAs for reprogramming factors (SOX2, KLF4, OCT4, L-MYC, LIN28, and dominant-negative p53) were introduced into human PBMCs using episomal vectors. A few days after transduction, PBMCs were collected and replated on dishes coated with laminin 511-E8 fragment (iMatrix 511, Nippi). The next day, the medium was replaced with StemFit AK03. Thereafter, the medium was replaced every two days. Twenty days after transduction, iPS cell colonies were picked. iPS cells established from PBMCs were expanded and cultured for neural differentiation.

[0178] [Cerebral cortical neurons derived from human iPS cells] A robust and rapid differentiation method was established using direct transformation technology. Human neurogenin 2 (NGN2) cDNA under a tetracycline-inducible promoter (tetO) was transformed into iPS cells using the piggyBac transposon system and Lipofectamine LTX (Thermo Fisher Scientific). A vector containing tetO::NGN2 was used. After antibiotic selection with G418 bisulfate (Nacalai Tesque), colonies were selected, and subclones that could efficiently differentiate into neurons by inducing transient expression of NGN2 were selected with a purity of >96% MAP2 / DAPI.

[0179] Karyotyping and Genotyping Karyotyping was performed by our laboratory or by LSI Medience. Genotyping of single nucleotide polymorphisms was performed by PCR amplification of genomic DNA and direct sequencing (3100 Genetic Analyzer; Thermo Fisher). The APOE gene was amplified by PCR (forward primer: TCCAAGGAGCTGCAGGCGGCGCA (SEQ ID NO: 1); reverse primer: ACAGAATTCGCCCCGGCCTGGTACACTG (SEQ ID NO: 2)). The PCR product was digested with HhaI at 37°C for 2 hours, followed by electrophoresis to analyze band sizes.

[0180] Immunocytochemical Staining: Cells were fixed with 4% v / v paraformaldehyde (pH 7.4) at room temperature (RT, approximately 25°C) and permeabilized with 0.2% v / v Triton X-100 in PBST. To suppress nonspecific binding, cells were blocked with BlockingONE histo (Nacalai Tesque) for 60 minutes at RT. Cells were incubated with primary antibodies overnight at 4°C and then labeled with fluorescently tagged secondary antibodies. Nuclei were labeled using DAPI (Thermo Fisher). Cell images were acquired using a high-content confocal microscope, IN Cell Analyzer 6000 (GE Healthcare). The following primary antibodies were used for immunocytochemical staining: NANOG (1:100 dilution; Abcam, ab80892), TRA1-60 (1:400 dilution; CST#4746, Danvers, MA), MAP2 (1:4,000 dilution; Abcam, ab5392), SATB2 (1:400 dilution; Abcam, EPNCIR130A ab92446), Alexa488-conjugated antibody (1:400 dilution; Thermo Fisher, A11029), Alexa488-conjugated antibody (1:400 dilution; Thermo Fisher, A11039), and Alexa594-conjugated antibody (1:400 dilution; Thermo Fisher, A11040). A21207).

[0181] On day 10, the RIPA-soluble fraction of total protein was extracted from differentiated neurons. 30 μL of RIPA buffer was added and the cells were cultured in a 96-well plate. The supernatant was centrifuged at 12,000 g for 30 minutes to collect the supernatant. The protein concentration of the supernatant was measured using the Pierce BCA Protein Assay Kit (Thermo Fisher) according to the kit's manual.

[0182] [Pathway analysis of identified genes] 230 identified genes (p<5×10) were analyzed using commercially available IngenuityPathwayAnalysis (IPA, manufactured by QIAGEN, https: / / www.qiagenbioinformatics.com / ) software. －５ ) path analysis was performed and the top network was analyzed.

[0183] Amyloid-β (Aβ) Electrochemiluminescence Assay: On day 8, all media was replaced with 100 μL of fresh media. Conditioned media was collected on day 10 for further analysis. Aβ species in the media were measured for extracellular human Aβ using the Human (6E10) Aβ3-Plex kit (Meso Scale Discovery). For Aβ species, the assay used the 6E10 antibody to capture Aβ peptides and different SULFO-TAG-labeled C-terminal-specific anti-Aβ antibodies for electrochemiluminescence detection using a Sector Imager 2400 (Meso Scale Discovery). Quantified Aβ values ​​(N = 2 wells per clone) were adjusted using the total neuronal protein concentration to minimize noise due to variations in cell number and allow comparisons between conditions.

[0184] Tau species in RIPA lysates extracted from iPS cell-derived neurons were measured using the Phospho(Thr231) / Total Tau Kit (Meso Scale Discovery) according to the kit's instructions. Quantified tau values ​​(N = 2 wells per clone) were adjusted using the total protein concentration of neurons to compare conditions while minimizing noise due to altered cell numbers.

[0185] Genome-Wide Association Study (GWAS) for SNP Genotyping and Polygenic Cellular Analysis of AD Patients: All 102 AD patient samples were genotyped using the Infinium OmniExpressExome-8 v1.4 BeadChip according to the kit's manual (Illumina). To separate algorithmic issues from data format issues, all genotype data were normalized to the forward strand GRCh37.p13 direction generated by variant calling from WGS data. GenomeStudio (Illumina) and quality control (Hardy-Weinberg equilibrium: p>1.0 × 10) were used. －６After genotyping using linkage disequilibrium-based variant pruning (r2<0.8, window size: 100 kb, step size: 5), genotypes were imputed with minimac4 using the 1,000 Genomes Project Phase 3 as the reference panel. 7,349,481 SNPs exceeded the post-imputation quality threshold (r2≥0.3, minor allele frequency≥0.01). Linear associations between SNPs and the Aβ42 / 40 accumulation rate in iPSC-derived neurons were analyzed with plink1.9, and age at onset, sex, and genotype of the APOE-ε4 allele were included as covariates in the linear regression model. p<5x10 －５ was set as the suggestive level, and p<5×10 －８ The significance level for the association analysis was set at . No statistical methods were used to predetermine the sample size, but the sample size is similar to that reported in previous publications.

[0186] Prediction of clinical data from the ADNI dataset. Results of Aβ42 / 40 accumulation rates in cerebral cortical neurons were processed by LD-based aggregation (r2 > 0.2, window size = 1 Mb) using PLINK1.9. Among independent SNPs, the threshold level recommended for genome-wide analysis (p < 5 × 10) was used. －５ ) were used as variables in the prediction model. The SNP genotype matrix of the selected 102 AD patient samples, originally composed of 0, 1, or 2, was normalized and analyzed by principal component analysis (PCA).

[0187] Genotypes were collected for samples from the Alzheimer's Disease Neuroimaging Initiative (ADNI) 1 / GO / 2 dataset (Illumina Omni 2.5MB BeadChip). Quality control and imputation were performed on the genotype data under the same conditions. The imputed genotypes for 10,121,962 SNPs were filtered by 496 SNPs obtained from genome-wide analyses. Genotypes for SNPs listed in the CDiP list but not in the ADNI dataset were imputed with the average genotype of AD patients. The phenotypes of the ADNI samples were then predicted from the genotypes. Samples were predicted as plausible AD (positive) or not (negative). Samples were independently classified as positive / negative according to four criteria based on the results reported in the ADNI database. First, the standardized uptake value ratio (SUVR, reference: cerebellar reference region) from AV45 PET data (>1.1, threshold for positivity); second, Aβ(1-42) in CSF (<977 pg / mL, threshold for positivity); third, t-tau / Aβ(1-42) in CSF (>0.27, threshold for positivity); and fourth, p-tau / Aβ(1-42) in CSF (>0.025, threshold for positivity). All reported results were obtained from the baseline ADNIMERGE dataset. Samples with both genotype and phenotype data were included in the study (SUVR AV45: n=512; CSF Aβ(1-42), t-tau / Aβ(1-42), p-tau / Aβ(1-42): n=581). The genotype vectors of the ADNI samples were mapped onto the principal component space derived from the genotype matrix of the in-hospital AD patients. 10-fold cross-validation was performed.

[0188] The ADNI sample was divided into a training sample and a test sample. A random forest classifier (100 estimators) was trained on the training sample, and the target variable (positive / negative AD-like condition) was predicted from the top three PCs of the genotype matrix and covariates (age, sex, and APOE-ε4 genotype). Prediction performance was evaluated by the area under the curve (AUC) of the receiver operating characteristic (ROC) curve obtained from the prediction of the test sample. Prediction performance was compared with the case when the target variable was predicted only from the covariates. The significance of the improvement in AUC was tested using the Wilcoxon signed-rank test (significance threshold: p<0.05). The target variable corresponds to the "information on the onset of Alzheimer's disease" described above.

[0189] Target Gene Knockdown: Cells were seeded into 6-well plates at an initial density of 3,000,000 cells per well on day 5. Twenty-four hours after seeding (day 6), the medium was replaced with neurobasal medium containing 1 μM AccellSMARTpool siRNA (Horizon Discovery). To maximize the effect of Acell siRNA, iPSC-derived neurons were cultured for 72 hours from day 6 to day 9. Seventy-two hours after siRNA addition (day 9), the culture medium was replaced with fresh neurobasal medium containing 1 μM AccellSMARTpool siRNA or 1 μM JNJ-40418677 (Sigma-Aldrich). Cells were harvested on day 11 and analyzed for Aβ phenotype.

[0190] Analysis of rare variants associated with AD development. Whole-exome sequencing was performed on 407 blood-derived genomic DNA samples obtained from 255 AD patients and 152 cognitively normal controls participating in the ADNI project in Japan. Exon sequences were enriched by hybridization using the Agilent SureSelect Human All Exon Kit (V6) and sequenced on an Illumina HiSeq4000 using paired-end read chemistry. Short-read sequences of the target regions were mapped to the human reference genome (hg38) using BWA-MEM version 0.7.15-r1140 with default settings. Subsequent analysis (read processing, variant calling, and variant filtering) was performed according to the recommendations of the GATK4 best practice, followed by variant annotation using snpEff version 4.3t. Among all variants identified by whole-exome sequencing, we focused on nonsynonymous, nonsense, splice-site, insertion, or deletion variants. This was further refined to variants with an MAF < 0.05 in public databases using the following tools: ExAC release 0.3 (http: / / exac.broadinstitute.org / ), gnomAD release 2.1.1 for exomes and r.3.0 for genomes (https: / / gnomad.broadinstitute.org / ), HGVD version 2.3 (http: / / www.hgvd.genome.med.kyoto-u.ac.jp / ), and TfoMMo version 8.3KJPN (https: / / jmorp.megabank.tohoku.ac.jp). Gene-based association analysis of variants was performed using Burdentest in the R package seqMeta version 1.6.7 using J-ADNI (n = 407) and ADNI (n = 479) exome data.

[0191] Data Availability: Data used in this study were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). ADNI was launched in 2003 as a public-private partnership led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI was to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessments could be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer's disease (AD). SNP array data are available at the National Bioscience Database Center (NBDC) (https: / / humandbs.biosciencedbc.jp / en / , Study ID: hum0314.v1).

[0192] Code availability: All data management and analysis code is archived online on GitHub (https: / / github.com / HaruhisaInoue / iSNPs4ADNIpred). All other code is publicly available on the inventors' websites.

[0193] Statistics and Reproducibility: Statistical analyses were performed as follows, except for the prediction of clinical data from the ADNI dataset and the analysis of rare variants associated with AD onset. All data are presented as mean ± SD. To confirm reproducibility, two to three experimental replicates were performed. Data variance was assumed to be normal, but this was not formally tested. Comparisons of means between three or more groups were performed by one-way analysis of variance (ANOVA) followed by Tukey's multiple comparison test or Uncorrected Fisher's LSD (post-hoc test using GraphPad Prism 7.0 software (GraphPad)). A p value of less than 0.05 was considered significant.

[0194] Example 1 In this study, genome-wide analysis was performed using Aβ released from cerebral cortical neurons derived from iPS cells of an AD cohort as a pathological signature, followed by CDiP to elucidate the complex pathological mechanisms in a neuron-specific manner.

[0195] 1. Generation of iPS cells and analysis of Aβ phenotype in iPS cell-derived neurons. First, to analyze AD pathology in neurons, we generated iPS cells with normal karyotypes from a sporadic AD (SAD) cohort (N = 102). The iPS cells demonstrated the in vitro ability to generate all three germ layers and expressed X-inactive specific transcripts (XISTs), similar to human ESCs.

[0196] All iPS cell clones were directly differentiated into cerebral cortical neurons by forced expression of the human NGN2 gene. In this differentiation protocol, exogenous NGN2 expression was fully suppressed from day 8 onward, and the Aβ phenotype remained constant from day 8 to day 14. The complex pathology of AD is composed of various types of molecules or biological events, such as Aβ and tau, which may be candidates for GWAS signatures. Because Aβ is the triggering event in the initiation of the long-term pathological cascade of AD and leads to dementia, Aβ was selected as the pathological signature of cerebral cortical neurons. Aβ40 and Aβ42 were quantified as protective and toxic Aβ, respectively, and the Aβ42 / 40 ratio was quantified in the culture supernatant of SAD cerebral cortical neurons.

[0197] The APP and PSEN1 genes, which play a central role in the Aβ production pathway, are known to influence neural development and the tendency for neuronal differentiation from human iPS cells. Therefore, when evaluating Aβ expression between iPS cells from different patients, it is important to maintain uniform purity of neuronal differentiation and normalize for variations in the number of neurons per well. The direct differentiation method used in this study yielded uniform, highly pure cerebral cortical neurons. However, the stress of direct differentiation from day 0 to day 5 caused variations in neuronal density between patients, and this variability affected the amount of Aβ. To normalize for variations in the number of neurons per well, the total protein concentration extracted from neurons across the entire well was used. This is because changes in protein concentration linearly reflect the number of neurons per well across different independent neurons or patients.

[0198] To confirm the correlation with genomic information, we analyzed the correlation between APOE genotypes, which represent the strongest genetic risk for AD, and Aβ species. The APOE ε4 genotype was moderately correlated with the Aβ42 / 40 ratio (Figure 3C), as evidenced by other modalities, but not with Aβ levels (Figures 3A and 3B) or protein concentrations (Figure 3D).

[0199] Previous reports using genetic engineering techniques have also shown that the APOE4 allele influences the Aβ phenotype of iPSC-derived neurons with the same genetic background. However, the change in Aβ phenotype in the different SAD populations with APOE3 / 3 versus 4 / 4 (1.09-fold change in this study) was smaller than the change observed with genomic correction (approximately 1.2- or 2-fold change in previous reports) (Figure 3C).

[0200] We also analyzed the correlation between the quantified Aβ phenotype in cerebral cortical neurons and clinical conditions such as age at onset and gender. The amount and ratio of Aβ species did not correlate with age at onset (Figures 4A, 4B, and 4C) or gender (Figures 4D, 4E, and 4F).

[0201] These results indicated that the Aβ phenotype of SAD was influenced by the diverse polygenic architecture of SAD. Therefore, we performed a genome-wide analysis of the pathological characteristics of AD using Aβ in neurons of the cerebral cortex of SAD.

[0202] 2. Genome-Wide Analysis To understand the polygenic nature of Aβ, we performed a genome-wide analysis using the Aβ42 / 40 ratio in cerebral cortical neurons as a pathological signature. Statistical analysis was adjusted for APOE status, and a false discovery rate for multiple testing was applied. Overall, the results showed no significant deviation from that expected by chance (λ = 0.9659), implying no evidence of bias or inflation of the test statistic due to population structure. To estimate the influence of APOE genotype, CDiP was first performed without adjusting for APOE genotype (Figure 5A). As a result, the p-value for rs429358 (T / C, APOE ε4 locus) was 0.794, which was not statistically significant. Although APOE ε4 is associated with a high risk of clinical AD, CDiP demonstrated that the Aβ42 / 40 ratio in single-cell cultures of iPSC-derived neurons is primarily influenced by a complex set of genes other than APOE ε4 alone. Therefore, we performed CDiP adjusting for APOE genotype (Figure 5B) and compared the genotypes of the 24 SNPs and associated loci (p-value < 5 × 10 －８ " or "p value < 5 × 10 －５ We identified a genetic locus (containing more than 10 SNPs, where the SNPs are "Aβ42 / 40") that is associated with an altered Aβ42 / 40 ratio (Figure 5C and Tables 3-1 to 3-77). In Tables 3-1 to 3-77, "chr" means chromosome, "BETA" means partial regression coefficient, and "SE" means standard error. The same meanings are used in the following tables.

[0203]

[0204]

[0205]

[0206]

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214]

[0215]

[0216]

[0217]

[0218]

[0219]

[0220]

[0221]

[0222]

[0223]

[0224]

[0225]

[0226]

[0227]

[0228]

[0229]

[0230]

[0231]

[0232]

[0233]

[0234]

[0235]

[0236]

[0237]

[0238]

[0239]

[0240]

[0241]

[0242]

[0243]

[0244]

[0245]

[0246]

[0247]

[0248]

[0249]

[0250]

[0251]

[0252]

[0253]

[0254]

[0255]

[0256]

[0257]

[0258]

[0259]

[0260]

[0261]

[0262]

[0263]

[0264]

[0265]

[0266]

[0267]

[0268]

[0269]

[0270]

[0271]

[0272]

[0273]

[0274]

[0275]

[0276]

[0277]

[0278]

[0279]

[0280] Genome-wide association analysis identified rs34033747 on chromosome 11, an intronic SNP in the DENN domain containing 2B (DENND2B), as the most highly correlated SNP (p-value = 1.91 × 10 －９) (Table 4). In Table 4, the left side of "Allele" refers to the basic allele and the right side refers to the minor allele (mutant allele), and "Gene ID" refers to the ID assigned to each gene defined by NCBI. The same meaning is used in the following tables.

[0281]

[0282] Five loci and associated genes, including CUL1, QRFP, CTNNA3, DAB1, and DCC, were known to be associated with Aβ production. Furthermore, eight loci and associated genes, including MAGI1, TMTC1, TRPM1, KCNMA1, DAB1, CPXM2, ROBO2, and ANO3, have been reported as AD-related loci or clinical biomarkers in clinical GWAS. Twelve loci and associated genes were novel Aβ- or AD-related genes (Tables 5-1 to 5-2). In Tables 5-1 to 5-2, "EOAD" refers to early-onset Alzheimer's disease, "LOAD" refers to late-onset Alzheimer's disease, "CNV" refers to copy number variation, and "OR" refers to odds ratio. In addition, "yes" in the "Brain" section means high expression in the brain, "low" means low expression in the brain, and "nd" means that there is no data in the GTEx portal (https: / / gtexportal.org / home / ). In addition, the "Brain cell-type" section lists the top three cell types that showed high gene expression in the Brain RNA-Seq portal (https: / / www.brainrnaseq.org / ). The same meanings are used in the following tables.

[0283]

[0284]

[0285] Furthermore, most of the identified genes are expressed in the brain (GTEx Portal, https: / / gtexportal.org / home / ), and 19 genes are highly expressed in neurons (Brain RNA-Seq Portal, https: / / www.brainrnaseq.org / ) (Table 5-1 to Table 5-2). Unbiased pathway analysis identified the "calcium signaling pathway" as the top canonical pathway (p-value = 2.51 × 10 －５ ) (Figure 5C). These networks are known to alter Aβ metabolism. These results demonstrate that the SNPs and associated genes identified by the polygene architecture analysis described above contribute to the Aβ42 / 40 ratio in cortical neurons and Alzheimer's disease as cell-type-specific features of AD pathology.

[0286] Furthermore, because p231-tau, which is tau phosphorylated at the 231st threonine from the N-terminus, is a highly sensitive marker for diagnosing or tracking the progression of AD, the p231-tau / total tau ratio (p231-tau ratio) was quantified and applied to CDiP. The APOE ε4 genotype, gender, and age at onset of AD were not correlated with the p231-tau ratio (Figures 6A, 6B, and 6C). CDiP was performed using the p231-tau ratio as a trait, with or without adjustment for APOE genotype (Figures 6D and 6E), to identify SNPs and associated loci (p-value <5 × 10). －５ ) (Table 6 and Tables 7-1 to 7-9). The SNP with the lowest p-value was rs68888116 (p=1.24×10-6) in the TNFAIP8 gene locus, an inflammation-related molecule (Table 6). In Table 6, "insATCT" stands for ATCTCAG(A). １２ This means that TTCTCTATCT (SEQ ID NO: 3) has been inserted.

[0287]

[0288]

[0289]

[0290]

[0291]

[0292]

[0293]

[0294]

[0295]

[0296]

[0297] 3. Identification of Therapeutic Target Genes To demonstrate a direct interaction between the Aβ phenotype and the 24 genes identified by CDiP, we quantified Aβ species upon knockdown of the identified genes (Figures 7A, 7B, 7C, and 7D). Suppression of the expression of amyloid precursor protein (APP) or β-site APP cleaving enzyme 1 (BACE1), key components of Aβ production, reduced Aβ levels as expected (Figures 7B and 7C). Knockdown of eight of the 24 genes identified by CDiP significantly altered the Aβ42 / 40 ratio (Figure 7A). We focused in particular on the top three target genes with the greatest reduction in the Aβ42 / 40 ratio: CTNNA3, ANO3, and CSMD1. Regarding Aβ levels, knockdown of 23 of the 24 genes identified by CDiP altered the levels of Aβ42 or Aβ40 (Figures 7B and 7C). Because changes in neuronal density must affect Aβ42 levels, we quantified protein levels after siRNA treatment before selecting genes to focus on (Figure 7D). As previously reported, knockdown of QRFPR, INFLR1, ZNRF2, ROBO2, DCC, and APP confirmed a decrease in total protein levels. Therefore, we excluded ZNRF2, INFLR1, DCC, and APP from our candidate genes affecting Aβ42 levels. We then focused on the top three target genes with the greatest reduction in Aβ42 levels: ZFPM2, TMTC1, and KCNMA1.

[0298] To narrow down potential targets for knockout therapy, it is necessary to select genes with elevated expression in neurons in the brains of AD patients. To investigate the expression status of genes of interest in neurons of AD patients, we utilized single-cell-based transcriptome data from the cerebral cortex of six AD patient brains and six control brains. This provides transcriptome data for individual cell types, including neurons, astrocytes, oligodendrocyte progenitor cells, oligodendrocytes, microglia, and endothelial cells. Specifically, we plotted the average expression of focused genes, such as CTNNA3, ANO3, and CSMD1 against the Aβ42 / 40 ratio, and ZFPM2, TMTC1, and KCNMA1 against Aβ42 in neurons (Figures 7E and 7F). The results revealed that the expression of CTNNA3, ANO3, and KCNMA1 was higher in the brains of AD patients.

[0299] From these results, we concluded that CTNNA3 and ANO3 for the Aβ42 / 40 ratio, and KCNMA1 for the Aβ42 amount, may be potential therapeutic targets for AD (FIG. 7G).

[0300] The protein encoded by CTNNA3 plays a role in cell-cell adhesion, and mutations in CTNNA3 cause familial arrhythmogenic right ventricular dysplasia, which is caused by improper handling of electrolytes such as potassium and calcium. The protein encoded by KCNMA1 consists of a voltage- and calcium-sensitive potassium channel (KCa1.1), which regulates smooth muscle tone and neuronal excitability. KCa1.1 is known as a target of cromolyn and, interestingly, is being tested in phase III trials for AD. The protein encoded by ANO3 has been reported to function in endoplasmic reticulum-dependent calcium signaling, and ANO3 mutations cause familial dystonia type 24 through abnormal neuronal excitability. Based on these results, the identified therapeutic targets may be involved in calcium handling and excitability, important pathways for Aβ regulation.

[0301] In summary, we performed a genome-wide analysis of the complex cell types of AD by analyzing cortical neurons and defining neuron-specific Aβ and tau phenotypes as pathological hallmarks of AD. As a result, CDiP revealed a set of genotypes that partially contribute to the polygenic architecture underlying the disease pathogenesis of AD.

[0302] Of the 24 genes identified above, 11 are novel AD-related genes whose relationship to AD has not been reported until now. The reason these genes have not been discovered until now is thought to be because various confounding factors have become noise in clinical GWAS alone.

[0303] Example 2 (Prediction of AD Onset Using Polygene Datasets Obtained from Cell GWAS) Next, we evaluated the similarity of real-world data with an in vitro dataset consisting of PET imaging of Aβ deposition in the brains of patients who provided PBMCs for iPS cell generation in this study. We analyzed the correlation between quantified Aβ phenotypes in cerebral cortical neurons and brain Aβ deposition measured by Pittsburgh Compound-B (PiB)-PET imaging. However, neither age at onset nor Aβ phenotype correlated with brain Aβ deposition (Figures 8A, 8B, 8C, and 8D). These results confirmed that simple quantified disease phenotypes without genetic information cannot reflect real-world data. Therefore, we investigated whether these genotype sets could be used to predict real-world big data from an independent AD cohort.

[0304] We utilized the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, which includes genome-wide genotypes, brain Aβ deposition (AV45-PET), Aβ42 levels in cerebrospinal fluid (CSF), total tau (t-tau) levels in CSF, and phosphorylated tau (p-tau) levels in CSF. We first attempted to predict positive brain Aβ deposition using covariates consisting of age, sex, and APOE-ε4 allele genotype alone, or a set of covariates and genotypes. We established a machine learning model to predict positive brain Aβ deposition using covariates consisting of age, sex, and APOE-ε4 allele genotype alone, or a set of covariates and specified genotypes. We used the trained model to predict brain Aβ deposition, and compared the area under the curve (AUC) between the two different models. The AUC (AUC = 0.76) of the covariate and genotype set was statistically higher than the AUC (AUC = 0.66) of the covariate alone (Figure 9A). Similarly, the covariate and genotype set predicted the reduction of CSF Aβ42 levels with significantly higher accuracy than the covariate alone (Figure 9B). However, when predicting CSF t-tau levels or CSF p-tau levels, there was no significant difference between the AUC of the covariate and the covariate and genotype sets (Figures 9C and 9D). These results demonstrate that the genotype set identified by CDiP can be used to predict actual clinical data of AD. Furthermore, using SNP information from sporadic AD patients, we were able to predict whether brain Aβ accumulation would occur with an accuracy of AUC = 0.76 ± 0.050 (Figure 9A). Using SNP information from sporadic AD patients, it was possible to predict whether abnormal test values ​​for Aβ in cerebrospinal fluid would occur with an accuracy of AUC = 0.73 ± 0.059 (Figure 9B). Because these Aβ accumulations in the brain and abnormal test values ​​for Aβ in cerebrospinal fluid are nearly consistent with clinical AD diagnoses, predictions using the information processing method of this embodiment can also be extrapolated to have an accuracy of approximately 0.73 to 0.76 inclusive of AUC.

[0305] [Example 3] (Discovery of rare variants by cell GWAS) To further confirm the applicability of the system to actual clinical data, we investigated whether the identified gene set formed SAD. We investigated the association of genes identified as rare variants in this study. These genes are known to be minor factors in the development of AD, albeit at low frequency.

[0306] Using genome-wide exome data from the Japan Alzheimer's Disease Neuroimaging Initiative (J-ADNI), rare variants at identified loci were investigated. By examining exome data from healthy donors (N = 152) and patients with SAD (N = 255), rare variants at 24 loci associated with the Aβ42 / 40 ratio were investigated. Rare variants in KCNMA1 showed an association with AD (p = 0.032, odds ratio = 1.45) (Table 8). In Table 8, "p" represents the p-value from the summation test, "se" represents the approximate standard error associated with the effect of genotype, "cmafTotal" represents the cumulative minor allele frequency of the gene, "cmafUsed" represents the cumulative minor allele frequency of the SNPs used in the analysis, "nsnpsTotal" represents the number of SNPs in the gene, "nsnpsUsed" represents the number of SNPs used in the analysis, and "nmiss" represents the number of missing SNPs. For genes with a single SNP, "nmiss" represents the number of individuals who did not contribute to the analysis because the test did not report the results for that SNP. For genes with multiple SNPs, the results are summed across the entire gene. This is used in the following tables as well.

[0307]

[0308] To confirm the reproducibility of rare variants in different cohorts and different ethnicities, a meta-analysis was performed to investigate rare variants at 24 loci, and rare variants at the KCNMA1 locus were again identified by meta-analysis of J-ADNI and US-ADNI (p = 0.010, odds ratio = 1.49) (Table 9).

[0309]

[0310] These results indicate that the identified gene set can be applied to elucidate predisposition factors for the onset of SAD. Furthermore, understanding the genetic background of AD obtained through cell GWAS will enable the extraction of specific populations based on genetic information from clinically homogeneous AD patient groups, leading to more detailed classification (i.e., stratification) of AD.

[0311] Discussion: This study identified risk SNPs, the genes in which the SNPs are located, and molecular pathways that affect Aβ production in cerebral cortical neurons. Indeed, five of the 24 genes identified by CDiP, namely TMTC1, CTNNA3, KCNMA1, CPXM2, and ANO3, were consistent with reported results from clinical genome-wide studies based on clinical data associated with disease onset or brain Aβ deposition (see Tables 5-1 and 5-2 above). This advantage may be due to the fact that we used a homogenous population of cerebral cortical neurons, the main cell type that serves as a resource for Aβ production.

[0312] The newly identified genes in this study (Table 4) may not only play pivotal roles in the pathogenesis of AD but also represent potential biomarkers and candidate therapeutic targets. To expand the system presented here, CDiP can be applied to other types of neuronal phenotypes in AD pathology to identify the unique genetic background of each trait, such as synapse loss, neuronal death, drug response, and vulnerability to environmental stress. Furthermore, novel combinations of variable cell types, such as glial cells and cell type-specific pathologies, may reveal new genetic architectures of molecular pathologies that have been hidden by clinical GWAS.

[0313] Recent studies have emphasized the concept that AD is the sum of pathologies across diverse cell types. Based on a similar idea as the present study, single-nucleus transcriptomes from autopsy AD brains have provided information on gene expression across various cell types. However, such approaches based on autopsy brain samples provide a late-stage snapshot of AD pathology, which has been changing for decades. In contrast, CDiP can investigate isolated AD pathology with cell-type specificity and can also model the baseline state without confounding factors that can contribute to genome-wide studies. A limitation of CDiP is that it is based on 2D monolayer cultures consisting of a single cell type. To understand the cellular interactions between various cell types, the combination of CDiP and single-nucleus transcriptomes from autopsy brains of AD patients may be two of the most important tools for investigating the polygenic nature of AD presented in the present study (Figures 8B and 8C).

[0314] Furthermore, neuronal CDiP identified rare variants and potential therapeutic targets associated with the Aβ phenotype, whereas SNPs associated with the tau phenotype showed more moderate statistically significant correlations. This difference between Aβ and tau suggests that Aβ pathology is primarily neuronal polygenic, whereas tau pathology may be comprised solely of multiple non-neuronal cell types.

[0315] Indeed, previous reports have shown that inflammatory conditions and brain networks with microglia and astrocytes accelerate tau pathology. Furthermore, clinical evidence suggests that APOE regulates tau pathology independently of Aβ pathology. CDiP using neurons may suggest one aspect of the discontinuity between Aβ and tau pathology. In the future, it is expected that an integrated and comprehensive understanding of the genetic background obtained through these cell-type-specific analytical approaches will lead to a better understanding of the complex pathogenesis of AD.

[0316] In this study, CDiP predicted real-world AD data, stratified rare variant-associated AD, and identified CTNNA3, ANO3, and KCNMA1 as potential therapeutic targets. CDiP serves as a screening tool to correlate pathological phenotypes with hidden genotypes. Meanwhile, it is also important to accumulate evidence using various modalities, such as mouse models and patient specimens, to adapt to the actual AD pathology, which is composed of various cell types and will take decades to mature. CDiP provides clues for understanding complex pathologies, which consist of the sum of polygenic and phenotypic traits in disease target cells, paving the way for precision medicine.

[0317] According to the information processing method, information processing device, and program of this embodiment, it is possible to predict the risk of developing AD in a subject.

[0318] 100...information processing device, 110...detection unit, 120...processing unit, 121...acquisition unit, 122...feature amount conversion unit, 123...determination unit, 124...output control unit, 125...learning unit, 130...storage unit, 131...model information

Claims

1. A method for processing information using a computer, comprising: detecting a first SNP, which is a mutation in an Alzheimer's disease associated gene, in a genomic DNA sample from a subject; and determining whether or not the subject will develop Alzheimer's disease from the first SNP, using a machine learning model trained on a plurality of training datasets labeled with information regarding the onset of Alzheimer's disease, for a second SNP, which is a mutation in the Alzheimer's disease-related gene detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease; The mutations include two or more mutations described in Tables 1-1 to 1-77 below. Information processing methods. 【Table 1-1】 【Table 1-2】 【Table 1-3】 【Table 1-4】 【Table 1-5】 【Table 1-6】 【Table 1-7】 【Table 1-8】 【Table 1-9】 【Table 1-10】 【Table 1-11】 【Table 1-12】 【Table 1-13】 【Table 1-14】 【Table 1-15】 【Table 1-16】 【Table 1-17】 【Table 1-18】 【Table 1-19】 【Table 1-20】 【Table 1-21】 【Table 1-22】 【Table 1-23】 【Table 1-24】 【Table 1-25】 【Table 1-26】 【Table 1-27】 【Table 1-28】 【Table 1-29】 【Table 1-30】 【Table 1-31】 【Table 1-32】 【Table 1-33】 【Table 1-34】 【Table 1-35】 【Table 1-36】 【Table 1-37】 【Table 1-38】 【Table 1-39】 【Table 1-40】 【Table 1-41】 【Table 1-42】 【Table 1-43】 【Table 1-44】 【Table 1-45】 【Table 1-46】 【Table 1-47】 【Table 1-48】 【Table 1-49】 【Table 1-50】 【Table 1-51】 【Table 1-52】 【Table 1-53】 【Table 1-54】 【Table 1-55】 【Table 1-56】 【Table 1-57】 【Table 1-58】 【Table 1-59】 【Table 1-60】 【Table 1-61】 【Table 1-62】 【Table 1-63】 【Table 1-64】 【Table 1-65】 【Table 1-66】 【Table 1-67】 【Table 1-68】 【Table 1-69】 【Table 1-70】 【Table 1-71】 【Table 1-72】 【Table 1-73】 【Table 1-74】 【Table 1-75】 【Table 1-76】 【Table 1-77】

2. The mutation includes a mutation having a p-value of less than 5 x 10-8 among the multiple mutations listed in Tables 3-1 to 3-77 below. The information processing method according to claim 1 . 【Table 3-1】 【Table 3-2】 【Table 3-3】 【Table 3-4】 【Table 3-5】 【Table 3-6】 【Table 3-7】 【Table 3-8】 【Table 3-9】 【Table 3-10】 【Table 3-11】 【Table 3-12】 【Table 3-13】 【Table 3-14】 【Table 3-15】 【Table 3-16】 【Table 3-17】 【Table 3-18】 【Table 3-19】 【Table 3-20】 【Table 3-21】 【Table 3-22】 【Table 3-23】 【Table 3-24】 【Table 3-25】 【Table 3-26】 【Table 3-27】 【Table 3-28】 【Table 3-29】 【Table 3-30】 【Table 3-31】 【Table 3-32】 【Table 3-33】 【Table 3-34】 【Table 3-35】 【Table 3-36】 【Table 3-37】 【Table 3-38】 【Table 3-39】 【Table 3-40】 【Table 3-41】 【Table 3-42】 【Table 3-43】 【Table 3-44】 【Table 3-45】 【Table 3-46】 【Table 3-47】 【Table 3-48】 【Table 3-49】 【Table 3-50】 【Table 3-51】 【Table 3-52】 【Table 3-53】 【Table 3-54】 【Table 3-55】 【Table 3-56】 【Table 3-57】 【Table 3-58】 【Table 3-59】 【Table 3-60】 【Table 3-61】 【Table 3-62】 【Table 3-63】 【Table 3-64】 【Table 3-65】 【Table 3-66】 【Table 3-67】 【Table 3-68】 【Table 3-69】 【Table 3-70】 【Table 3-71】 【Table 3-72】 【Table 3-73】 【Table 3-74】 【Table 3-75】 【Table 3-76】 【Table 3-77】

3. The mutations include all of the mutations listed in Table 4 below. The information processing method according to claim 1 . 【Table 4】

4. The mutations include all of the mutations described in Tables 3-1 to 3-77 of claim 2. The information processing method according to claim 1 .

5. The machine learning model is a random forest including a plurality of classifiers; Each classifier is trained using a specific training dataset selected from the plurality of training datasets based on principal component information in the attribute information and genetic information of the patients who have developed Alzheimer's disease. The information processing method according to claim 1 .

6. The plurality of training datasets include a dataset in which information on the onset of Alzheimer's disease is labeled for the imputed genotype of the second SNP estimated from the second SNP using genotype imputation.

3. The information processing method according to claim 1 or 2.

7. The mutations further include one or more of the mutations described in Tables 2-1 to 2-9 below. The information processing method according to claim 1 . 【Table 2-1】 【Table 2-2】 【Table 2-3】 【Table 2-4】 【Table 2-5】 【Table 2-6】 【Table 2-7】 【Table 2-8】 【Table 2-9】

8. The mutations include all of the mutations described in Tables 2-1 to 2-9 of claim 7. The information processing method according to claim 1 .

9. The mutations include those listed in Table 6 below. The information processing method according to claim 1 . 【Table 6】

10. a detection unit that detects a first SNP, which is a mutation in an Alzheimer's disease-related gene, in a genomic DNA sample derived from a subject; a determination unit that determines whether or not the subject will develop Alzheimer's disease from the first SNP, using a machine learning model trained on a plurality of training data sets labeled with information on the onset of Alzheimer's disease, for a second SNP, which is a mutation in the Alzheimer's disease-related gene detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease, The mutations include two or more mutations listed in Tables 1-1 to 1-77 of claim 1. Information processing device.

11. A program for causing a computer to execute the program, detecting a first SNP, which is a mutation in an Alzheimer's disease associated gene, in a genomic DNA sample from a subject; and determining whether or not the subject will develop Alzheimer's disease from the first SNP, using a machine learning model trained on a plurality of training datasets labeled with information regarding the onset of Alzheimer's disease, for a second SNP, which is a mutation in the Alzheimer's disease-related gene detected in a genomic DNA sample derived from a patient who has developed Alzheimer's disease; The mutations include two or more mutations listed in Tables 1-1 to 1-77 of claim 1. program.

12. A method for predicting the risk of developing Alzheimer's disease, comprising using the information processing method according to claim 1.

Images