Population‐Based Identification of Clonal Hematopoiesis Using Peripheral Blood Whole‐Genome Sequencing in Japan

ABSTRACT

Clonal hematopoiesis of indeterminate potential (CHIP) is an age‐associated condition whose population‐level characteristics and clinical concordance in Japan remain incompletely defined. We analyzed whole‐genome sequencing data from 49,982 participants in the Tohoku Medical Megabank (TMM) cohort to characterize the prevalence, mutational spectrum, age dependency, and clinical relevance of CHIP. CHIP mutations were detected in 1234 individuals (2.5%), comprising 1413 variants across 1015 loci. DNMT3A and TET2 were the most frequently mutated genes, whereas ASXL1 and PPM1D mutations were less prevalent than reported in Western cohorts. Four recurrent CHIP candidate loci (ZNF318, SMC3, CBL, and GNAS) were identified, most of which were predicted to be oncogenic. Comparison with TOPMed and UK Biobank datasets demonstrated overall concordance in variant prevalence, with differences observed for selected driver mutations. dN/dS analysis indicated positive selection in the majority of CHIP‐associated genes. Concordance with clonal hematopoiesis (CH) mutations detected in cancer patients was evaluated using data from the MONSTAR‐SCREEN‐2 study with paired leukocyte sequencing. When focusing on genes in which CHIP mutations have been reported, 66% were confirmed as CH, yielding a positive predictive value of 66% and a negative predictive value of 76%. CHIP prevalence increased significantly with age, exceeding 5% in individuals aged ≥ 70 years, with an estimated 3.4% increase per additional year of age. Higher variant allele frequency mutations showed a slightly stronger age association. These findings define the landscape of CHIP in the Japanese population and clarify its relationship to CH detected by cancer‐focused liquid biopsy assays.

Whole‐genome sequencing of ~50,000 individuals from the Tohoku Medical Megabank (TMM) cohort defines the prevalence, mutational landscape, and age dependency of clonal hematopoiesis in the Japanese population, demonstrating overall concordance with clonal hematopoiesis detected in cancer patients. CHIP: clonal hematopoiesis of indeterminate potential.

Abbreviations

AD

allele depth

CGP

comprehensive genomic profiling

CH

clonal hematopoiesis

CHIP

clonal hematopoiesis of indeterminate potential

CRAM

Compressed Reference‐oriented Alignment Map

ctDNA

circulating tumor DNA

DP

depth

FATHMM

Functional Analysis through Hidden Markov Models

GATK

Genome Analysis Toolkit

GO

Gene Ontology

HSC

hematopoietic stem cell

LBx

liquid biopsy

NPV

negative predictive value

PPV

positive predictive value

SNV

single‐nucleotide variant

TMM

the Tohoku Medical Megabank

TOPMed

the Trans‐Omics for Precision Medicine

UKBB

the UK Biobank

VAF

variant allele frequency

WGS

whole genome sequencing

1. Introduction

Clonal hematopoiesis (CH) refers to the clonal expansion of blood cells driven by genetic alterations in hematopoietic stem cells (HSCs) [1]. CH is caused by random somatic mutations in the genome of HSCs that self‐replicate throughout the host's life. CH can potentially progress to hematologic malignancies, including myelodysplastic syndrome, myeloproliferative neoplasm, and acute myeloid leukemia. CH can also be driven by both small‐scale variants (e.g., single‐nucleotide variants and short insertions/deletions) and large‐scale variants (e.g., chromosomal gains/losses and uniparental disomy) [1, 2].

CH occurring in individuals without any hematologic abnormalities is referred to as clonal hematopoiesis with indeterminate potential (CHIP), which is commonly defined by the presence of small variants with a variant allele frequency (VAF) of ≥ 2% [3]. The definition of CHIP has recently expanded to encompass 91 genes, including 17 genes newly identified based on whole‐genome sequencing (WGS) of 200,618 individuals in the UK Biobank (UKBB) [4].

CHIP can complicate the interpretation of results of comprehensive genomic profiling (CGP) analyses. Because both CHIP mutations and tumor‐derived mutations are often detected at relatively low VAFs, some CHIP mutations could mistakenly be interpreted as tumor‐derived mutations. Moreover, incidentally detected germline mutations with low VAFs can be difficult to distinguish from CHIP or somatic mutations. These issues are particularly prominent in specimens exhibiting a high proportion of peripheral blood–derived DNA, especially liquid biopsies (LBx). Chan and colleagues performed simultaneous sequencing of tumor tissue, peripheral blood, and cell‐free DNA from patients with solid tumors [5]. They concluded that tumor‐derived mutations accounted for only 53% of all detected variants, whereas up to 17% of those in LBx specimens were classified as CHIP mutations. This result indicates that a nonnegligible proportion of CHIP mutations may be inadvertently included in CGP results. Therefore, it is essential to be aware of specific CHIP mutations when interpreting the results of CGP analyses.

The Trans‐Omics for Precision Medicine (TOPMed) project is a large nationwide program in the United States that involved high‐coverage (30×) WGS of peripheral blood samples obtained from patients with non‐cancer comorbidities in 51 observational studies. Bick and colleagues identified 4938 CHIP mutations in 4229 (4.33%) individuals among 97,691 participants [6]. Although CHIP was reported notably less prevalent in East Asians (odds ratio 0.56 after adjusting for age, genetic similarity, sex, and smoking in multivariate analysis) in the TOPMed project [6], Terao et al., detected mosaic chromosomal alterations in 15.6% of the Japanese population, with a prevalence reaching nearly 30% among men in their 80s [7]. The true prevalence of CHIP in the Japanese population remains to be validated using more appropriate datasets.

The Tohoku Medical Megabank (TMM) Project generated high‐quality WGS data for Japanese residents. In this study, we utilized the TMM WGS dataset to create a database of CHIP mutations (i.e., Japanese CHIP panel) to support specialists involved in CGP analyses. We also evaluated LBx data from patients with solid tumors enrolled in the MONSTAR‐SCREEN‐2 study to validate the accuracy of CHIP mutation detection using this database.

2. Materials and Methods

2.1. TMM Project

The TMM Project was designed to facilitate the development of personalized medicine and healthcare management [8]. Over 150,000 residents of Miyagi and Iwate prefectures of Japan participated in either of the prospective population‐based cohort studies, designated the TMM Community‐based Cohort Study [9] and the TMM Birth and Three‐Generation Cohort Study [10]. Numerous biospecimens and extensive health information and genomic and multi‐omics data from participants have been incorporated into an integrated biobank and made available to researchers [11, 12, 13, 14]. The TMM Community‐based Cohort Study and the TMM Birth and Three‐Generation Cohort Study generated high‐coverage (30×) WGS data from analyses of peripheral and umbilical cord blood samples. The WGS data were mapped using GRCh38 as the reference genome and did not include the Y chromosome. As of November 2022, when the present study was initiated, the repository dataset was the largest available, including 50,316 individuals (i.e., 50 K WGS repository). This dataset included approximately 12,000 (24%) participants genetically related to other participants. A wide variety of metadata pertaining to the participants were also available, including age, sex, self‐reported medical history, diet, body weight, smoking habits, depressive mood, and blood and urine testing results, which were referenced in analyzing the background characteristics of the participants.

Since the purpose of this study was to detect CHIP‐associated mutations, we analyzed 49,982 individuals after excluding 11 individuals with a history of hematologic diseases and 323 infants whose cord blood samples had been submitted from the 50,316 participants with available data.

2.2. Defining CHIP Mutations

To filter out sequencing errors and germline variants from the WGS data and to identify somatic mutations that fulfill the biological characteristics of CHIP, we performed the analyses in the following order according to previously published methods [15].

2.2.1. Variant Call

Referring the Compressed Reference‐oriented Alignment Map (CRAM) files for 49,982 participants with available WGS data, somatic variant calling was performed by applying the Genome Analysis Toolkit (GATK) Mutect2 pipeline. Subsequently, hard‐filtering was carried out using the GATK VariantFiltration pipeline. The genome reference version was GRCh38. The germline resource was set to the allele frequency panel based on short‐read WGS analysis of 54,000 Japanese individuals (available from jMorp: https://jmorp.megabank.tohoku.ac.jp/downloads). The panel of normal was set to the publicly available panel generated from 1000 Genomes Project samples (available at gs://gatk‐best‐practices/somatic‐hg38/1000g_pon.hg38.vcf.gz through “gsutil” command with an option “‐m cp”). We used GATK version 4.5.0.0, HTSJDK version 4.1.0, and Picard version 3.1.1, OpenJDK version 22.0.2, Python version 4.3.2, and R version 3.9.12. A total of 4,370,386 variant loci were called and passed the filtering across all exonic regions, of which 32,519 loci were located in 90 genes in which CHIP mutations have been reported.

A distinct analytical strategy was applied to detect somatic variants in the U2AF1 gene because an erroneous segmental duplication in the GRCh38 reference genome can result in a mapping quality score of zero during FASTQ alignment [16]. Weinstock et al. have released a program on GitHub (https://github.com/weinstockj/pileup_region) to accurately detect the read counts of U2AF1 gene mutations, and we also applied their script in this study.

2.2.2. Filtering Based on Sequencing Depth and Allele Count

We initially performed basic filtering to exclude variants with insufficient sequencing coverage. Specifically, we discarded variants with a total read depth (DP) below 20, a minimum allele depth (AD) for the alternate allele below 3, or lacking support from both forward and reverse sequencing reads. In addition, variants with a variant allele frequency (VAF) below the conventional 2% threshold for defining CHIP were excluded.

2.2.3. Excluding Sequencing Artifacts

Variants observed in ≥ 10 individuals in the TMM (0.02%) were evaluated for their potential to represent recurrent sequencing artifacts. To assess this, we tested the association of each variant group with two well‐established correlates of CHIP—age [17] and a common genetic variant in the TERT promoter (rs7705526) [6, 18]. Variants that showed no association with either age or rs7705526—that is, those with a coefficient p ≥ 0.10 in the logistic regression analysis—were considered sequencing artifacts and excluded from the dataset.

2.2.4. Excluding Germline Variants

To exclude possible germline variants, we performed a binomial test for all variants. This test evaluates whether the observed number of altered reads at a given site differs statistically from half of the total read count, as expected for heterozygous germline variants. Variants with a p ≥ 0.01 were considered potential germline variants and were excluded from the dataset. However, as previously reported, specific TET2 variants (NM_001127208: p.H1904R, p.I1873T, and p.T1884A) were likely to represent CHIP clones with large VAF; therefore, these variants were exempted from the binomial test–based filtering.

2.2.5. Variant Annotation

Variant annotation was performed using TransVar v2.4.0 [19], and accession numbers were based on canonical UCSC RefSeq transcripts.

2.2.6. Defining CHIP Mutations and Candidates

Finally, among the variants that passed all of the above filtering steps, we defined CHIP mutations exclusively as those that were concordant with previously reported CHIP mutations. In addition, variants that did not meet the established definitions were categorized as “CHIP candidates” only if they were recurrent variants observed in at least 10 individuals (0.02%), showed possible positive associations with age and the TERT promoter SNP (p < 0.10), and were predicted by functional annotation to have a high likelihood of causing transcript‐level functional disruption.

2.3. Reference of Known CHIP Mutations

In this study, we referred to two previously published studies based on data from the TOPMed project [6] and the UKBB [4] to define known CHIP mutations. In the former study, 2531 loci of CHIP mutations are listed in Table S3 of the corresponding publication [6], while in the latter, 2358 loci are reported in Table S5 of the paper [4]. These datasets were used as reference sets for known CHIP mutations in our analysis.

2.4. Accuracy Testing Data: MONSTAR‐SCREEN‐2 Study

The MONSTAR‐SCREEN‐2 study (UMIN000036749) is a multicenter collaborative research project focused on profiling and monitoring cancer‐related genetic abnormalities in circulating tumor DNA (ctDNA) in patients with unresectable solid malignancies. This study involved longitudinal and pan‐cancer analysis of ctDNA obtained from blood samples in these patients. Blood genomic profiling was performed by Caris Assure (https://www.carismolecularintelligence.com /profiling menu/mi profile international/international/). Caris Assure is a next‐generation sequencing assay, validated under the College of American Pathologists and Clinical Laboratory Improvement Amendments standards, that integrates whole‐exome and whole‐transcriptome analyses. The assay enables comprehensive detection of single‐nucleotide variants, insertions and deletions, gene fusions, copy‐number alterations, microsatellite instability, tumor mutational burden, and HLA genotypes in both plasma and buffy coat samples (capturing clonal CH and germline variants) from patients with advanced cancer. The workflow simultaneously prepares sequencing libraries for both cell‐free DNA and cell‐free RNA within a single run using a customized hybridization/capture methodology. To identify potential germline variants and CH‐associated variants, nucleic acids from matched buffy coat samples were sequenced separately. Buffy coat and plasma samples were processed in matched batch sizes to ensure equivalent sequencing depth and comparable limits of detection across sample types [20]. In this study, 2768 cases were recruited between March 19, 2021, and March 31, 2024, among which Caris Assure was utilized for 2093 cases. This dataset is particularly well suited for validating the accuracy of the CHIP detection algorithm. Incidentally, it is generally customary to refer to detected CHIP mutations in patients with cancers as “CH” to distinguish them from those identified in individuals with no malignancy.

To fairly assess the accuracy for the less prevalent variants, the percentages of CH, tumor‐derived, and germline variants were calculated for each locus, and the sum of these percentages was compared. That is, the sum of the percentages of CH divided by the number of all loci was calculated as the positive rate, whereas the sum of the other percentages divided by the total was calculated as the negative rate.

2.5. dN/dS Ratio

The dN/dS ratio is defined as the number of non‐synonymous mutations per number of non‐synonymous sites in each gene region, divided by the number of synonymous mutations per number of synonymous sites within the same region [21], (non‐)synonymous sites are defined as specific coding loci in which any single nucleotide substitution would lead to a (non‐)synonymous mutation. For example, any substitution in the third position of codon UCA (− > serine) always results in a synonymous mutation (i.e., UCU, UCG, UCC‐ > serine), which is called a synonymous site. A dN/dS ratio > 1 suggests that a gene is under positive selection for non‐synonymous mutations, whereas a ratio of < 1 indicates negative selection.

The procedure for calculating the dN/dS ratio for each gene region was as follows. First, all exonic nucleotide sequences of each gene region were referenced to determine the corresponding codons, from which synonymous and non‐synonymous sites were unambiguously defined. Next, all single‐nucleotide variants occurring at synonymous and non‐synonymous sites in the TMM dataset were counted on a per‐gene basis. The dN/dS ratio was then calculated by dividing the observed number of non‐synonymous mutations by the total number of non‐synonymous sites (dN), and similarly, the corresponding calculation was made for synonymous mutations (dS), after which the ratio of these two values was determined (dN/dS). Notably, if the same site was mutated in multiple individuals, it was counted only once. GRCh38 was used as the reference nucleotide sequence, and exon regions were defined based on the Twist Exome Dx Panel (https://www.twistbioscience.com/resources/data‐files/twist‐exome‐dx‐panel‐bed‐files).

3. Results

3.1. Basic Statistics of the TMM Participants

The median age of the participants was 44 [33–65] years, and 29.1% were male. The age distribution exhibited a bimodal pattern, with the first peak occurring in the early 30s and the second in the mid‐60s (Figure S1). Twenty‐eight percent of the participants were current smokers. Common medical issues in the internal medicine category for which > 10% of participants had a history included hypertension (25%), gastritis (21%), allergic rhinitis (21%), dyslipidemia (15%), iron‐deficiency anemia (13%), and gastro‐esophageal regurgitation (10%). Other significant internal morbidities included asthma (6.2%), type 2 diabetes mellitus (4.8%), angina pectoris (2.9%), and stroke (1.5%). The most frequently reported malignancy was breast cancer (2.0%), followed by gastric cancer (1.6%) and colorectal cancer (1.6%) (Table 1).

TABLE 1.

Background medical and life‐style information of all participants.

Item	The number of individuals with available data	The number of affected individuals	Proportion or median [IQR]
Basic Information
Age	49,982	—	43 [32–64]
Sex (proportion of male)	49,982	14,626	29%
Frequency of drinking a	12,908	—	Score 3 [1–4]
Proportion of current smoker	9649	2687	28%
Medical History of Non‐Malignant Disease
Hypertension	22,114	5557	25%
Gastritis	20,384	4312	21%
Allergic rhinitis	21,034	4314	21%
Dyslipidemia	20,653	3137	15%
Iron‐deficiency anemia	20,626	2620	13%
GERD	20,558	2080	10%
Asthma	20,670	1285	6.2%
Fatty liver	20,384	1165	5.7%
Type 2 DM	20,664	998	4.8%
Angina pectoris	20,767	608	2.9%
Hyperuricemia	20,333	576	2.8%
Stroke	20,759	319	1.5%
Heart failure	20,276	64	0.3%
COPD	20,131	51	0.3%
Medical History of Cancer
Breast cancer	20,361	404	2.0%
Gastric cancer	20,768	333	1.6%
Colorectal cancer	20,521	327	1.6%
Lung cancer	20,432	97	0.5%
Kidney cancer	20,206	44	0.2%
Brain cancer	20,161	43	0.2%
Skin cancer	20,160	40	0.2%
Pancreatic cancer	20,242	16	0.1%
Liver cancer	20,269	13	0.1%

Abbreviations: COPD, chronic obstructive pulmonary disease; DM, diabetes mellitus; GERD, gastroesophageal reflux disease; IQR, interquartile range.

^a Score 1: everyday, 2: 5–6 days/week, 3: 3–4 days/week, 4: 1–2 days/week.

3.2. CHIP Mutations Detected in TMM Genomic Data

Analysis of WGS data from 49,982 eligible participants identified 1413 CHIP mutations across 1015 loci, as well as 90 CHIP candidates spanning four loci (Figure 1 and Table S1). Variants were defined as CHIP candidates if they were not established CHIP mutations but were recurrent (detected in ≥ 10 individuals, 0.02%), showed possible positive associations with age and the TERT promoter SNP (p < 0.10), and were predicted to have a high likelihood of transcript‐level functional disruption based on functional annotation (Materials and Methods—Defining CHIP mutations). Variants excluded during filtering were compiled into a separate table, along with the reasons for their exclusion (Table S2). The total prevalence of CHIP was 2.5% (3.2% when CHIP candidates were included). DNMT3A (414 variants/266 loci) was the most frequently mutated gene, followed by TET2 (266 variants/229 loci), and ZRSR2 (98 variants/62 loci) (Figure 2). The prevalence of ASXL1 and PPM1D mutations were relatively lower than that reported in previous studies (ASXL1: 0.076% in TMM vs. 0.37% in TOPMed vs. 0.15% in UKBB; PPM1D: 0.04% vs. 0.18% vs. 0.05%, respectively). The CHIP candidates were detected at four loci, ZNF318 (NM_014345) p.L948V, SMC3 (NM_005445) p.E1226D, CBL (NM_005188) p.F827V, and GNAS (NM_080425) p.F611C. Using the cancer variant predictor of the Functional Analysis through Hidden Markov Models (FATHMM) v2.3, these mutations were classified as oncogenic with high scores (> 0.70), except for CBL p.F877V with a moderate score (0.47).

FIGURE 1.

Workflow for defining CHIP mutations from whole‐genome sequencing data. Somatic variants were identified from WGS data using the GATK Mutect2 pipeline, followed by hard filtering. Variants were sequentially filtered based on sequencing depth, allele count, and a minimum VAF threshold of 2%. Recurrent sequencing artifacts were excluded based on their lack of association with age and the TERT promoter SNP (rs7705526), and potential germline variants were removed using binomial testing, with known high‐VAF TET2 CHIP variants exempted. A specialized approach was applied for U2AF1 due to reference genome artifacts. Final CHIP mutations were defined as variants concordant with previously reported CHIP mutations, while additional recurrent, biologically plausible variants were classified as CHIP candidates.

FIGURE 2.

Frequency of CHIP mutations by gene. Blue bars indicate previously reported CHIP mutations (1413 variants across 1015 loci in 62 genes), while orange bars represent CHIP candidates proposed in this study (i.e., ZNF318 p.L948V [21 variants], SMC3 p.E126D [20 variants], CBL p.F827V [37 variants], and GNAS p.F611C [12 variants]). DNMT3A was the most frequently affected gene (414 variants across 266 loci), followed by TET2 (266 variants across 229 loci) and ZRSR2 (98 variants across 62 loci).

We compared the prevalence of CHIP mutations identified in this study with those reported in previous large‐scale studies (TOPMed and the UK Biobank) (Figure S2). For the 360 loci shared across these datasets, the prevalence was generally comparable for most variants. However, a small number of variants, including JAK2 p.V617F, DNMT3A p.R882C, and SRSF2 p.P95H, showed markedly higher prevalence in TOPMed, whereas their prevalence in the UK Biobank was relatively closer to that observed in our study.

To assess selective pressure on CHIP‐associated genes, dN/dS ratios were calculated for protein‐coding genes. Among the 61 genes (out of 62) for which dN/dS can be computed, most showed evidence of positive selection (dN/dS > 1) (Figure S3). The three genes including CHIP candidates (CBL, ZNF318, and GNAS) also exhibited dN/dS ratios greater than 1, except for the SMC3 gene (0.95).

3.3. Concordance Between CHIP Mutations in General Population and CH Mutations in Cancer Patients

Concordance between CHIP mutations detected in the TMM cohort and CH mutations defined in cancer patients was assessed using data from the MONSTAR‐SCREEN‐2 study, which employed the Caris Assure LBx assay, and parallel sequencing of peripheral blood leukocytes was also conducted. Consequently, in addition to the cell‐free DNA sequencing results, each detected variant was annotated with its origin—whether tumor‐derived, CH, or germline. Because the genomic data from the MONSTAR‐SCREEN‐2 trial likely contained mutations exclusively observed in patients with solid tumors, the analysis was limited to pathogenic mutations in the 62 CHIP genes which harbored CHIP mutations detected in this study. Among the 1494 participants in the MONSTAR‐SCREEN‐2 study for whom LBx data were available, pathogenic mutations in CHIP genes were detected in 1062 individuals (71.1%), encompassing 1747 variants across 797 loci in 27 genes. Of the 95 loci matched to the CHIP mutations in the TMM, 66% of the loci were also identified as CH (true positives) by Caris Assure (Figure 3). Conversely, among the remaining 702 loci, 76% of the loci were judged to be tumor‐derived or germline. Thus, the positive predictive value (PPV) and negative predictive value (NPV) of our panel were 66% and 76%, respectively.

FIGURE 3.

Performance of CHIP detection using LBx data. Among 1494 MONSTAR‐SCREEN‐2 participants with available LBx data, pathogenic mutations in CHIP genes were detected in 1062 individuals (71.1%), comprising 1747 variants across 797 loci in 27 genes. Comparison with TMM‐defined CHIP mutations showed a positive predictive value of 66% and a negative predictive value of 76% for Caris Assure.

A total of 582 CH mutations detected by Caris Assure were distributed across 31 genes. Among these, CH mutations in DNMT3A, GNAS, and PPM1D were frequently represented in our CHIP panel, whereas those in KRAS, RUNX1, and CREBBP were detected exclusively by Caris Assure. To characterize these gene sets, we divided them into two groups: Group A, comprising genes with at least partial overlap with the CHIP panel, and Group B, comprising genes with no overlap. Gene Ontology (GO) enrichment analysis was performed for each group. Group A showed enrichment limited to epigenetic regulation of gene expression, whereas Group B exhibited broader enrichment across developmental, signaling, immune, metabolic, and chromatin‐related processes (Figure S4).

3.4. Correlation Between CHIP Prevalence and Age

Among the 49,982 participants, 1234 participants (2.5%) harbored at least one CHIP mutation (not including candidate), as defined in this study. Using the age‐stratified prevalence derived from the TMM data, we recalculated the prevalence by applying it to the age distribution of the Japanese population as of October 2024, as reported by the Statistics Bureau of Japan (https://www.stat.go.jp/data/jinsui/2024np/index.html), yielding an estimated prevalence of 3.2%. The prevalence of CHIP exhibited a clear age‐dependent increase, exceeding 5% in individuals aged 70 years or older (Figure 4A). A generalized linear model (GLM) with a Poisson distribution and logit link function indicated a statistically significant positive correlation between age and CHIP prevalence, with a hazard ratio of 1.034 (95% confidence interval: 1.030–1.037), corresponding to an approximately 3.4% increase in CHIP prevalence with each additional year of age. When the analysis was stratified by variant allele frequency into mutations with VAF 2%–10% versus those with VAF ≥ 10%, the latter group showed a slightly stronger age‐associated increase than the former (hazard ratio, 1.023 [95% CI, 1.018–1.029] vs. 1.039 [95% CI, 1.035–1.043]) (Figure 4B,C). When restricting the analysis to the CHIP candidates spanning four loci, their prevalence also showed an association with age, similar to that observed for CHIP mutations (Figure 4D–F).

FIGURE 4.

Age‐dependent prevalence of CHIP in the Japanese population. Among 49,982 participants 1234 individuals (2.5%) harbored at least one CHIP mutation. CHIP prevalence increased markedly with age, exceeding 5% in individuals aged ≥ 70 years. A generalized linear model demonstrated a significant positive association between age and CHIP prevalence, corresponding to an approximately 3.4% increase per year (A). Stratification by variant allele frequency showed a slightly stronger age‐associated increase for mutations with VAF ≥ 10% (C) compared with those with VAF 2%–10% (B). When restricting the analysis to the CHIP candidates spanning four loci (i.e., ZNF318 p.L948V [21 variants], SMC3 p.E126D [20 variants], CBL p.F827V [37 variants], and GNAS p.F611C [12 variants]), their prevalence also showed an association with age, similar to that observed for CHIP mutations (D: All CHIP candidates, E: Those with VAF 2%–10%, F: Those with VAF ≥ 10%).

4. Discussion

4.1. Characteristics of CHIP Mutations in the Japanese Population

The prevalence of CHIP mutations (including candidates) in this study was 2.7% and was estimated 3.5% in the whole Japanese population, which was roughly comparable with those in previous reports (4.3% in the TOPMed Project [6] and 2.0% in the UKBB [4]). In contrast, a study analyzing single‐nucleotide variants (SNVs)/indels and copy number alterations associated with clonal hematopoiesis using genomic data from BioBank Japan [22], a large‐scale registry study in Japan, reported that SNVs/indels—corresponding to CHIP mutations—were detected in 27% of participants aged 60 years or older [23]. In that study, variants called from 23 genes (ASXL1, CBL, CEBPA, DDX41, DNMT3A, ETV6, EZH2, GATA2, GNAS, GNB1, IDH1, IDH2, JAK2, KRAS, MYD88, NRAS, PPM1D, RUNX1, SF3B1, SRSF2, TET2, TP53, and U2AF1) were filtered using a study‐specific approach distinct from previously reported CHIP definitions and included variants with low VAFs of 0.5%–2% as well as variants with high VAFs exceeding 50%. Thus, the definition of CHIP mutations in that study differs from that used in the present analysis. For reference, among variants in these 23 genes that passed our filtering criteria regardless of CHIP definitions, 3191 variants (19%) were detected in participants aged 60 years or older. Although there are differences in filtering strategies and VAF ranges between the studies, we believe that both describe largely similar biological phenomena.

In addition to known CHIP mutations, we identified CHIP candidates on four loci. They are recurrently found (≥ 10 individuals) in the TMM participants and showed possible association with both age and the SNP at TERT promoter region (rs7705526); none of them are registered in the gnomAD database. Except for a CBL mutation, all CHIP candidates were classified as “oncogenic” based on their FATHMM scores, suggesting that they may induce functional impairment of the corresponding transcripts and exert biologically meaningful effects. Although the FATHMM score for CBL p.F827V was moderate, the amino acid region spanning residues 648–906 of the CBL protein is known to be critical for interaction with CD2‐associated protein, an adapter protein between membrane proteins and the actin cytoskeleton [24]. Therefore, the possibility that a missense variant within this region may cause functional impairment cannot be excluded.

4.2. Sensitivity of Variant Detection Using WGS

Given that the WGS coverage in this study was 30× and that variants with an allele depth of fewer than three reads were excluded, a large number of variants with a VAF below 10% may have been missed. As shown in Extended Data Figure ten of a previous report by Bick et al. [6], sequencing coverage has a substantial impact on variant detection sensitivity. Compared with sequencing at sufficient depth (400×), sequencing at 30× coverage may fail to detect nearly all variants with VAFs of 2%–5% and approximately 87% of variants with VAFs of 5%–10%, resulting in an estimated coverage of only ~53% of potential CHIP mutations. Conversely, CHIP variants with VAFs ≥ 10% can be reliably detected even at 30× coverage; therefore, the CHIP mutations presented in this study may be enriched for variants with stronger tendencies toward clonal expansion.

4.3. Performance in Detecting CH Mutations in Cancer Patients

In an external validation using genomic data from solid tumor patients enrolled in the MONSTAR‐SCREEN‐2 study (Figure 3), the PPV and NPV for detecting CH variants were 66% and 76%, respectively. Notably, no germline variants were mistakenly classified as CH. However, only a relatively small number of variants (95 of 797 loci) matched those in the CHIP panel, indicating a limitation in sensitivity. One possible explanation is that MONSTAR‐SCREEN‐2 is based on WES, and differences in sequencing coverage may therefore contribute to this observation. Alternatively, it is also possible that CHIP in the general population and clonal hematopoiesis in patients with cancer have distinct biological characteristics.

We conducted GO enrichment analysis of genes with CH variants detected by Caris Assure, dividing them into Group A (overlapping with our CHIP panel) and Group B (non‐overlapping). In Group A, enrichment was limited to epigenetic regulation of gene expression, indicating that CH in this setting may primarily involve global transcriptional regulation without prominent engagement of specific developmental or signaling pathways. This restricted enrichment profile suggests that CH in the general population is associated with relatively constrained biological alterations. By contrast, CH genes in Group B demonstrated enrichment across a broad range of biological processes, including development, cell differentiation, signal transduction, immune‐related pathways, and responses to environmental stressors such as hypoxia. Enrichment of chromatin remodeling and histone acetylation pathways further points to more complex epigenetic involvement. Taken together, these results indicate that although epigenetic regulation represents a common feature of CH, the spectrum of associated biological processes appears more extensive in cancer patients than in the general population. These differences may reflect distinct biological contexts in which CH arises or expands, rather than direct causal effects, and warrant further investigation to clarify their functional and clinical implications.

In this study, we identified CHIP mutations from whole‐genome sequencing data of 49,982 participants in the TMM project and demonstrated a CHIP prevalence of 2.7%, corresponding to 3.5% when standardized to the national population. In addition, we proposed four CHIP candidates that appeared to be statistically and biologically significant. Finally, we evaluated the performance of this list of CHIP mutations in CGP and showed that the positive and negative predictive values were approximately two‐thirds and three‐quarters, respectively. Although we acknowledge the inherent limitations in detection sensitivity associated with WGS, we expect that our results will serve as a reference for CHIP mutations in the Japanese population.

Author Contributions

SungGi Chi: conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, writing – original draft, writing – review and editing. Ikuko N. Motoike: data curation, project administration, writing – review and editing. Takao Fujisawa: project administration, writing – review and editing. Yumi Yamaguchi‐Kabata: project administration, writing – review and editing. Takahiro Nobukuni: project administration, writing – review and editing. Kengo Kinoshita: project administration, writing – review and editing. Yoshiaki Nakamura: project administration, writing – review and editing. Hideaki Bando: project administration, writing – review and editing. Takayuki Yoshino: project administration, writing – review and editing. Kinuko Ohneda: project administration, writing – review and editing. Katsuya Tsuchihara: project administration, writing – review and editing. Yosuke Minami: project administration, writing – review and editing. Riu Yamashita: conceptualization, project administration, writing – review and editing.

Funding

This work was supported in part by a Scientific Research from the Japan Society for the Promotion of Science (23K14637).

Ethics Statement

Approval of the research protocol by an Institutional Reviewer Board: This study was approved by the Institutional Review Board of the National Cancer Center (2022–105).

Consent

The genomic and cohort data used in this study were obtained previously as part of the TMM Project. Participants in the TMM Project were given the opportunity to refuse the use of individual data in this study by providing opt‐out information through the official websites of the National Cancer Center and the Tohoku Medical Megabank Organization.

Conflicts of Interest

Takao Fujisawa received honoraria from Amelieff Co. Ltd. and AstraZeneca. Hideaki Bando reports research funding from GlaxoSmithKline and honoraria from Ono Pharmaceutical, Taiho Pharmaceutical, and Guardant Health. Yoshiaki Nakamura reports advisory role from Guardant Health Pte Ltd., Natera Inc., Exact Sciences Corporation, Tempus AI Inc., Pfizer Inc., and Shionogi & Co. Ltd. Inc.; speakers' bureau from MSD K.K., Merck Biopharma Co. Ltd., Daiichi Sankyo Co. Ltd., Chugai Pharmaceutical Co. Ltd., and, Guardant Health Japan Corp. Takayuki Yoshino reports honoraria from Chugai Pharmaceutical, Takeda Pharmaceutical, Merck Biopharma, Ono Pharmaceutical, and MSD K.K.; consulting fee from Sumitomo Corp, and Indivumed; and research funding from Amgen, Bristol‐Myers Squibb, Caris MPI, Chugai Pharmaceutical, Daiichi Sankyo, Eisai, Exact Sciences, FALCO Biosystems, Medical & Biological Laboratories, Merus N.V., Miyarisan Pharmaceutical, Molecular Health GmbH, MSD, Natera Inc., Nippon Boehringer Ingelheim, Ono Pharmaceutical, Pfizer Japan, Roche Diagnostics, Sanofi, Sysmex, Taiho Pharmaceutical, and Takeda Pharmaceutical. Katsuya Tsuchihara has received research funding from Ono Pharmaceutical Industries Ltd.; Canon Medical Systems; Takeda Pharmaceutical Company Ltd.; and MICIN. He is an editorial board member (associate editor) of Cancer Science. Yosuke Minami has received lecture fees from Pfizer, Bristol‐Myers Squibb, Novartis Pharma, Daiichi Sankyo, and Astellas, and research funding from Bristol‐Myers, Chugai, Novartis Pharma, Takeda, Ono Pharmaceutical, and CMIC. The other authors declare no conflicts of interest.

Supporting information

Figure S1: A histogram of the number of participants by age at the time of TMM registration. There were two peaks, at 31 years old (n = 1802) and 65 years old (n = 1658).

Figure S2: Scatter plot comparing the prevalence of CHIP mutations between TMM and previously published datasets (TOPMed and the UK Biobank). For the 83 loci shared across the datasets, the prevalence was generally comparable for most variants. However, a subset of variants, including JAK2 p.V617F, DNMT3A p.R882C, and SRSF2 p.P95H, showed markedly higher prevalence in TOPMed, whereas the prevalence of the same variants in the UK Biobank was closer to that observed in the present study. The solid black line indicates the line of identity (equal prevalence on the x‐ and y‐axes), and the dashed lines represent the 95% confidence intervals (95% CI).

Figure S3: Figure showing the dN/dS ratios (log scale) for 61 genes for which the dN/dS ratio could be calculated among the 62 genes harboring CHIP variants. All but 7 genes (PPM1D, SMC3, ZESR2, BRCC3, YLPM1, RAD21, and BRAF) exhibited dN/dS ratios > 1, with deviations of up to sixfold. The seven genes including CHIP candidates (CBL, SMC3, ZNF318, and GNAS) also showed ratios > 1, except for SMC3 gene (0.95).

Figure S4: Group A, comprising genes with at least partial overlap with the CHIP panel, and Group B, comprising genes with no overlap. Group A showed enrichment limited to epigenetic regulation of gene expression, whereas Group exhibited broader enrichment across developmental, signaling, immune, metabolic, and chromatin‐related processes, indicating extensive cellular reprogramming. GO: gene ontology, TP: true positive (detected in both Caris Assure and the CHIP panel), FN: false negative (detected in Caris Assure but not in the CHIP panel), FDR: false discovery rate.

Table S1: List of CHIP mutations and candidates.

Table S2: List of excluded variants. page/prs: a coefficient p‐value of age/rs7705526 in the logistic regression model.