Less-invasive age estimation using hair based on DNA methylation in brown bears Authors: Shiori Nakamura, Jumpei Yamazaki, Naoya Matsumoto, Kyogo Hagino, Hideyuki Sakamoto, Masami Yamanaka, Masanao Nakanishi, Mina Jimbo, Yojiro Yanagawa, Hideyuki Ito, Toshio Tsubota, Michito Shimozuru Published: 2025 ABSTRACT AbstractInformation on chronological age is essential for exploring the life history, conservation, and management of wildlife. Recently, DNA methylation-based methods using blood or skin have been established as alternatives to the traditional tooth-based method in bear species. However, the collection of these tissues is limited to captured or dead individuals. In the present study, we established the first hair-based age estimation model based on DNA methylation levels in brown bears, aiming for future application to less-invasively obtained hair of wild individuals. We performed bisulfite pyrosequencing and measured the methylation levels of hair root DNA. The methylation levels of cytosine-phosphate-guanine sites adjacent to the genes VGF, KCNK12, and ELOVL2 were found to be correlated with age. The best age estimation model used three cytosine-phosphate-guanine sites adjacent to two genes, VGF and KCNK12, with a mean absolute error of 3.2 years and median absolute error of 2.2 years after leave-one-out cross-validation. Our method is innovative because of the simplicity of sampling and the lack of requirement to capture bears. If this method can be widely applied to hair samples obtained in the field, the age structure of wild populations can be understood, contributing to ecological research, conservation, and management of bear species.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-27455-2.Keywords: Epigenetic clock, Age estimation, Brown bear, DNA methylation, Wildlife management, AgingSubject terms: Biological techniques, Ecology, Ecology, Genetics, Molecular biology, Zoology Introduction Age information is essential for wildlife research to examine the life history of animals (including sexual maturity, reproductive success, disease dynamics, and risk of mortality) and contributes to wildlife conservation and management by providing estimates of age structure and population dynamics. Although the simplest way to obtain age information is to ascertain the year of birth by direct observation or camera-trap-based surveys1–3, these methods require a long-term effort and considerable human resources; therefore, the target species and fields available for investigation are limited. Various methods have been developed to determine the age of wild animals. The most common method, especially for terrestrial and marine mammals, is based on counting the cementum annuli or growth layer groups of the teeth that form annually4,5. Other methods include tooth eruption patterns, wear, skeletal features, physical characteristics, otoliths, and scales6–8. However, most of them are only applicable to dead or immobilized animals and are often difficult to implement in live animals. Methods based on samples that can be obtained noninvasively, such as hair and feces, have also been attempted. For example, the measurement of steroid hormone levels in hair can distinguish the age classes of brown bears (Ursus arctos)9, and fecal near-infrared reflectance spectroscopy can be applied to discriminate the age classes of giant pandas (Ailuropoda melanoleuca)10. Over the past decade, another age estimation method based on DNA methylation levels was developed (i.e., the epigenetic clock11). DNA methylation is a type of epigenetic modification. In vertebrates, it predominantly occurs at cytosine-phosphate-guanine (CpG) sites, forming 5-methylcytosine by transferring a methyl group to the C5 position of cytosine12. DNA methylation regulates gene expression, and its function has been suggested to vary with genomic context13,14. DNA methylation levels can change with age15, which has enabled its use as an indicator for age estimation, first demonstrated in humans (Homo sapiens)16,17. Epigenetic clocks are now available in various animals18 and have already been used to examine population structure19. DNA methylation is not static among tissues; the number of age-related differentially methylated positions and rate of change with age differ among tissues20. DNA methylation also varies among animal species, and its association with phylogenetic relations and lifespan has attracted attention from the perspective of comparative epigenomics21–23. Consequently, various species- and tissue-specific epigenetic clocks have been developed (skin of humpback whales [Megaptera novaeangliae]24; blood of mice [Mus musculus]25; blood of chimpanzees [Pan troglodytes]26; blood of long-lived seabirds [Ardenna tenuirostris]27; blood of Asian elephants [Elephas maximus]28). By contrast, some epigenetic clocks are shared among closely related species29–33 and different tissues (mice)34. Indeed, epigenetic clocks have been established using samples of 59 tissue types across 185 mammalian species, analyzed by mammalian methylation microarrays (HorvathMammalMethylChip4035) using hundreds of CpG methylation levels36. Such genome-wide deep sequencing approaches provide robust and large epigenetic age predictors; however, highly accurate age estimation can also be achieved with a limited number of CpGs using site-specific analysis37. Among the eight bear species worldwide, six (e.g., polar bears [Ursus maritimus]) are listed as vulnerable on the International Union for Conservation of Nature Red List38 owing to population declines caused by habitat reduction, poaching, and other factors. While attempts are being made to restore the population, human–bear conflicts can hinder the acceptance of conservation initiatives39. Addressing these challenges requires the implementation of population management strategies grounded in a comprehensive understanding of population dynamics, and the age at harvest has already been incorporated into population dynamics models40,41. If it were possible to estimate the age of living bears, this would contribute significantly to the improvement of population dynamics models. In bear species, the traditional age estimation method uses the cementum annuli of teeth42–44. However, this method has several disadvantages: (1) it requires pulling teeth and is invasive to live animals; (2) the observer requires sufficient experience to ensure accurate and precise age estimation45; and (3) accuracy declines in older bears46. Recently, age estimation methods based on DNA methylation levels have been developed for bear species, using blood, skin, and fibroblast samples33,47,48. A critical limitation of these methods is the requirement to obtain samples from either captured or dead individuals, which complicates their application in wildlife studies. Reports on epigenetic clocks using noninvasive or less-invasive samples, including human49 and cattle (Bos taurus)50 hair, as well as Indo-Pacific bottlenose dolphin (Tursiops aduncus)3 and mouse51 feces, are limited. In the present study, we aimed to establish an epigenetic clock using hair from brown bears. In brown bear research, hair samples that can be obtained noninvasively have been used for various purposes, including identifying species, estimating population size, determining genetic diversity, and detecting diet items52,53. In particular, the technique of using bait or scent lures to attract brown bears and collecting hairs snagged in barbed wire placed around them (hereafter simply referred to as “hair-traps”) is commonly used2,54,55. This study aimed to establish an age estimation method based on DNA methylation in brown bears using hair. DNA was extracted from hair roots collected from both captive and wild brown bears of known age. We targeted 39 CpGs in 12 distinct genomic locations whose methylation levels correlated with age in studies using blood DNA, given that age-related markers have been shown to function among different tissues56. Subsequently, we sought to identify the CpG sites in hair-derived DNA that exhibited a strong correlation between DNA methylation levels and age. Finally, we attempted to establish an epigenetic clock that could be implemented with fewer CpG sites (i.e., with a limited amount of DNA), with the aim of application in DNA-scarce sources, such as hairs obtained from hair-traps. Materials and methods Study area, animals, and hair sampling Captive bears Hair samples were obtained from 35 brown bears (18 males and 17 females) maintained at Noboribetsu Bear Park (Noboribetsu, Hokkaido, Japan). Their ages ranged from 1 to 33 years. Detailed information (bear ID, sex, birth year, sampling date, and age at sampling) is provided in Supplementary Table S_M1. Age at sampling was determined based on the assumption that all bears were born on February 1, on the grounds that the bears were born between mid-January and early February57–59. All hair samples were plucked under anesthesia, and the location was the upper back, where wild bears were thought to leave hair when they rubbed their backs against trees. The bears were anesthetized by intramuscular administration of xylazine HCl (1 mg/kg; Selactar; Bayer) and a 1:1 mixture of zolazepam HCl and tiletamine HCl (2.0–4.0 mg/kg; Zoletil 100; Virbac), administered remotely using a blow dart. After sampling was completed, atipamezole HCl (150 µg/kg; Atipame; Kyoritsu Co., Ltd.) was injected intramuscularly to facilitate recovery from anesthesia. The collected samples were stored at − 30 °C until genomic DNA extraction. Please refer to Nakamura et al.47 for details regarding the housing conditions at this facility. Wild bears Hair samples were obtained from 12 female brown bears living in the Rusha area (44°11′N, 145°11′E) of the Shiretoko Peninsula, eastern Hokkaido, Japan. Their ages ranged from 4 to 27 years. In this area, long-term continuous bear monitoring surveys have been conducted since 1997, enabling age determination using visual and DNA-based identification1. Detailed information (bear ID, sex, birth year, sampling date, and age at sampling) is provided in Supplementary Table S_M1. Age at sampling was determined based on the assumption that all bears were born on February 1, similar to captive bears. The bears were anesthetized by intramuscular administration of medetomidine HCl (75 µg/kg; Dorbene Vet; Kyoritsu Co., Ltd.) and a 1:1 mixture of zolazepam HCl and tiletamine HCl (5.5 mg/kg; Zoletil 100; Virbac), with dosages calculated based on estimated body weight. The drugs were administered remotely using an air rifle. After hair samples were plucked from the upper back, atipamezole HCl (375 µg/kg; Atipame; Kyoritsu Co., Ltd.) was injected intramuscularly to facilitate recovery from anesthesia. The collected samples were stored at − 30 °C until genomic DNA extraction. Genomic DNA extraction and bisulfite conversion Genomic DNA was extracted from 10 hair roots using a DNA Extractor FM Kit (FUJIFILM Wako Pure Chemical Corporation, Osaka, Japan). Extraction was performed according to the manufacturer’s protocol, and the DNA was finally eluted in 50 µL of AE Buffer. The extracted DNA was stored at − 30 °C until bisulfite conversion. Twenty microliters of DNA were subjected to bisulfite conversion using the EZ DNA Methylation-Gold Kit (Zymo Research, Irvine, CA, USA), according to the manufacturer’s instructions. After final elution in 10 µL of Elution Buffer supplied with the kit, 30 µL of TE Buffer was added (i.e., 40 µL in total). Jimbo et al.60 reported that guard hair bulbs change seasonally in morphology and are classified into three types (white sphere type, black hook type and white hook type). Bear guard hair grows year-round, except during hibernation, and is shed once a year in the summer. Newly grown and growing hairs are considered to be of the black hook type (around June to August), then change to the white hook type, continue to grow (around September to October), and then change to the white sphere type (from around November to shedding in the following summer) after growth stops until the hair sheds. Hair of the white sphere type is present, except for a brief period in autumn, around September and October. Most of the samples we obtained were of the white sphere type, and we did not obtain sufficient numbers for the other two types. Therefore, we only used hair bulbs of the white sphere type and excluded the other two types from the analysis. Target genomic locations, polymerase chain reaction (PCR), and pyrosequencing Target genomic locations were adjacent to 12 genes, GSE1, VGF, SLC12A5, SCGN, KCNK12, OTUD7A, BCL6B, POU4F2, ELOVL2, RALYL, KISS1R, and CAPS2 as in our previous study on the epigenetic clock in blood samples from brown bears47. DNA methylation levels of CpG sites adjacent to these genes have been reported to change with age in humans61–64, dogs65,66, and cats67. The DNA methylation levels of the target CpGs were analyzed using PCR followed by pyrosequencing, in accordance with our previous reports33,47,66. PCR and subsequent pyrosequencing were performed in duplicate. PCR was conducted in two steps using TaKaRa EpiTaq HS (Takara Bio Inc., Shiga, Japan). Each PCR product was electrophoresed on a 2% agarose gel to confirm that the target was successfully amplified. Pyrosequencing was performed using PyroMark Q48 software (Qiagen, Venlo, Netherlands) with PyroMark Q48 Advanced Reagents (Qiagen), according to the manufacturer’s instructions. The mean methylation levels obtained from duplicate reactions of the two-step PCR and pyrosequencing were used to establish age estimation models. Details, including primer sequences and amplification conditions, are available in our previous study47. Initially, as a screening step, each target CpG site was analyzed across nine samples to identify the CpG sites whose DNA methylation levels were strongly correlated with age. These samples were selected from captive bears, and the number of individuals was balanced according to age and sex. From 12 candidate genomic locations, those with R² values greater than 0.8 and a range of methylation level changes greater than 20% were selected for further analysis. Finally, we analyzed the DNA methylation levels of the selected genomic locations for the remaining hair samples and calculated Pearson’s product-moment correlation coefficients and p-values. Age Estimation model and model validation Building age estimation models was performed similarly to the method described in our previous studies using blood samples from bears33,47. We used only one sample per animal to avoid overfitting the model to methylation changes in specific individuals sampled multiple times. Based on the DNA methylation levels obtained, three age estimation models were generated using R ver. 4.2.268. Single regression, elastic net regression, and support vector regression (SVR) models were built using the R command “lm” and R packages “glmnet”69,70 and “e1071,”71 respectively. Elastic net regression models, a type of penalized regression, have frequently been used for age estimation11,29,36,72, whereas SVR models have been reported to achieve high estimation accuracy33,47,67,73. Unlike elastic net regression, SVR does not perform feature selection, which necessitates careful consideration of predictor combinations. We constructed SVR models using feature sets in which the less important features identified by the elastic net regression were sequentially removed. In addition, SVR models were constructed using CpGs from each genomic location. Age and DNA methylation levels were standardized before integration into the models. Leave-one-out cross-validation (LOOCV) was performed to validate all models. To evaluate whether age, sex, or growth environment (i.e., captive or wild condition) affected the deviation of the age estimation model, linear regression and generalized linear regression were generated with Δage (predicted age − chronological age) and |Δage| (absolute difference between predicted age and chronological age) as dependent variables and the three factors and their interactions as explanatory variables. Model construction and selection were performed using the R command “lm” and R packages “MuMIn”74 and “lme4.”75. Results Selection of age-related genomic locations Following the screening of nine samples, three genomic locations— adjacent to VGF, KCNK12, and ELOVL2—were selected from twelve candidates based on R² values exceeding 0.8 and methylation level changes greater than 20%. All the results are shown in Supplementary Figures S_R1_1–11. Correlation between DNA methylation level and chronological age For each genomic location, multiple CpG sites showed significant correlations between DNA methylation levels and age. The three CpGs adjacent to VGF, KCNK12, and ELOVL2 are graphically presented in Supplementary Figures S_R2_1, 2, and 3, and the corresponding correlation coefficients and p-values for each CpG are provided in Supplementary Table S_R1. The CpGs that exhibited the strongest correlations for each genomic location are shown in Fig. 1. Scatter plots of chronological age (year) versus DNA methylation level (%) at the cytosine-phosphate-guanine (CpG) sites that showed the strongest correlation at each selected genomic location. Each plot represents a CpG site adjacent to VGF (a), KCNK12 (b), or ELOVL2 (c). Age Estimation model We constructed 15 age estimation models (three single regression models, one elastic net regression model, and 11 SVR models) based on the DNA methylation levels of CpGs adjacent to VGF (VGF-1, -2, and - 3), KCNK12 (KCNK12-1, -2, and - 3), and ELOVL2 (ELOVL2-1, -2, and - 3). Their performance can be confirmed from the mean absolute error (MAE), median absolute error (MedAE), and root mean square error values for each model in Table 1. Table 1Values of mean absolute error (MAE), median absolute error (MedAE), and root mean square error (RMSE) for each age estimation model.modelCpG sitesMAEMedAERMSESingle regressionVGF-34.138 3.127 5.353KCNK12-3 3.978 3.270 5.010 ELOVL2-35.5254.3597.278Elastic net regressionAll CpGs 3.627 3.375 4.630 Support vector regressionAll CpGs3.3982.9164.409VGF-1, -2, -3, KCNK12-1, -2, -3, ELOVL2-2, -33.3422.9144.329VGF-1, -2, -3, KCNK12-1, -2, -3, ELOVL2-33.2552.511 4.138 VGF-1, -3, KCNK12-1, -2, -3, ELOVL2-33.3592.6474.241VGF-1, -3, KCNK12-2, -3, ELOVL2-33.3672.7184.246VGF-1, -3, KCNK12-2, -33.6343.0154.533VGF-1, -3, KCNK12-3 3.192 2.188 4.238VGF-3, KCNK12-33.5132.7664.468VGF-1, -2, -33.4622.4395.020KCNK12-1, -2, -33.8453.2144.909ELOVL2-1, -2, -35.0483.8647.260All values were rounded to the fourth decimal place.Each bold value denotes the minimum value for each regression model. Values of mean absolute error (MAE), median absolute error (MedAE), and root mean square error (RMSE) for each age estimation model. All values were rounded to the fourth decimal place. Each bold value denotes the minimum value for each regression model. Among the single regression models, the model using the methylation level of KCNK12-3 showed the best performance; the MAE value after LOOCV was 3.978 years (Fig. 2a). The formula for age estimation was as follows: Scatter plots of predicted age (year) versus chronological age (year) for the age estimation models after leave-one-out cross-validation (LOOCV). The solid line represents predicted age = chronological age. The distance between the solid and dotted lines indicates the mean absolute error (MAE) of each model after LOOCV. Panels show the single regression model (a), elastic net regression model (b), and best support vector regression (SVR) model among all models (c). predicted age = (− 1.3183e-11 + 0.80980 × standardized methylation level of KCNK12-3) × 8.2771 (standard deviation of training data) + 12.149 (mean of training data). In the elastic net regression model, the MAE after LOOCV was 3.627 years (Fig. 2b). The formula for age estimation was as follows: predicted age = (1.2403e-11 + 0.13467 × standardized methylation level of VGF-1 + 0.097753 × standardized methylation level of VGF-2 + 0.16030 × standardized methylation level of VGF-3 + 0.12121 × standardized methylation level of KCNK12-1 + 0.13121 × standardized methylation level of KCNK12-2 + 0.15645 × standardized methylation level of KCNK12-3 + 0.026854 × standardized methylation level of ELOVL2-2 + 0.13107 × standardized methylation level of ELOVL2-3) × 8.2771 + 12.149. Among the SVR models, the model using the methylation levels of VGF-1, VGF-3, and KCNK12-3 showed the best performance; the MAE after LOOCV was 3.192 (Fig. 2c). The parameter details for the elastic net regression and SVR models are listed in Supplementary Table S_R2. Linear regression analysis showed that one explanatory variable significantly affected Δage in the best model (i.e., the SVR model using VGF-1, -3, and KCNK12-3). When Δage was used as the dependent variable, the optimal regression model included age and growth environment as explanatory variables (adjusted R² = 0.215) (Table 2; Fig. 3). Among the examined variables, age had a significant negative effect on Δage, while sex, growth environment [i.e., captive or wild] and their interactions showed no significant influence. Generalized linear regression analysis further revealed that none of the explanatory variables (i.e., age, sex, and growth environment, or their interactions) significantly affected |Δage| in the best-fitting model (Table 2). The R script used to build the models, estimate age, and evaluate the factors influencing model deviation is available in the Supplementary File. Table 2Coefficients and p-values from the linear regression analysis of Δage and generalized linear regression analysis of |Δage| in the best-performing model (SVR model using three CpGs: VGF-1, VGF-3, and KCNK12-3).Estimatep-valueΔage(Intercept)0.070090.33846Age-0.21261 0.00169 Growth environment (wild)-0.20370000.1647|Δage|(Intercept)-1.041–Note: Bold values denote statistical significance (p < 0.05). Coefficients and p-values from the linear regression analysis of Δage and generalized linear regression analysis of |Δage| in the best-performing model (SVR model using three CpGs: VGF-1, VGF-3, and KCNK12-3). Note: Bold values denote statistical significance (p < 0.05). Scatter plot of Δage (year) versus chronological age (year) in the best-performing model. Discussion In this study, we identified three genomic locations where DNA methylation levels correlated with age using hair-derived DNA and established age estimation models for captive and wild brown bears. To the best of our knowledge, this is the third report of age estimation using hair-derived DNA and first involving wild animals49,50. One advantage of this method is the simplicity of sampling: hair is much easier to collect than teeth or blood. Collecting blood from live wild bears requires anesthesia and appropriate sampling techniques. Blood can also be collected from the thoracic cavity, heart, or major vessels after removing the sternum or ribs; however, this procedure is generally labor-intensive and prone to coagulation. By contrast, hair can be collected from hunted or opportunistically discovered carcasses, as it requires no special equipment, such as needles or syringes, and can be performed by non-specialists. Our results show that the values of MAE (3.192) and MedAE (2.188) are satisfactory given the lifespan of brown bears, which is typically 20–30 years76, with a maximum of 40 years (AnAge: The animal aging and longevity database77). Moreover, the MAE and MedAE values were not substantially inferior to those reported in previous studies using less invasive samples. For instance, in the previous study using human hair DNA, the MedAE was 4.15 years49. In cattle tail hair, with a lifespan of approximately 20 years (AnAge: The animal aging and longevity database77), MAE values were 1.4 years for individuals younger than 3 years, 1.5 years for those aged 3–10 years, and 9 years for individuals over 10 years50. For human nails, the MAE values were 5.48 years for combined fingernails and toenails and 5.61 years for toenails alone78,79. Additionally, there was no difference in the prediction error of our model between wild and captive bears, although the number of analyzed samples was limited. This finding suggests that our method can be applied to wild populations. Until now, hair collected via hair-traps has primarily been used to identify individuals and determine sex for population estimation in wild bear studies2,80. The current method adds valuable biological information, i.e., age, enabling the estimation of sex and age structure within a population. This advancement will significantly improve our understanding of population dynamics and support the development of more scientifically informed management and conservation strategies. Furthermore, it may broaden the scope of ecological research using hair, such as monitoring age-related dietary shifts through stable isotope analysis. Therefore, this method holds potential for advancing the field of wildlife ecology. However, there is room for improving accuracy. The lower accuracy of hair-based age estimation in brown bears, compared to blood-based methods (MAE = 1.30, MedAE = 1.00)47, may be attributed to suboptimal selection of genomic targets. The same genomic regions used in previous blood DNA analyses were applied here. Compared to blood-based studies, data on epigenetic age markers in hair are scarce, and methylation changes associated with aging vary across species, making target selection challenging. Age-related methylation changes in the same genomic region often differ among tissues. Of the four genomic locations that showed strong correlations between methylation levels and age in bear blood (SLC12A5, POU4F2, VGF, and SCGN), only one (VGF) was included in the hair-based model, and its methylation levels differed between blood- and hair-derived DNA (Supplementary Figures S_R3_1–11). Among genes identified as epigenetic age markers for hair in humans (LAG3, SCGN, ELOVL2, KLF14, C1orf132, SLC12A5, GRIA2, and PDE4C)49 and cattle (LHX1, ISL1, NKX6-1, SIM1, and TLX3; Shown here are only the five most significant markers among 43)50, only ELOVL2 was found in brown bear hair. To improve the accuracy of hair-based age estimation models, it is necessary to employ genome-wide approaches to identify suitable target regions in hair DNA. Linear regression analysis of Δage suggested that the ages of young bears tended to be overestimated, whereas the ages of older bears were underestimated. However, generalized linear regression analysis of |Δage| indicated that the magnitude of prediction error was not significantly associated with age. Many previous studies have reported that age estimation models based on DNA methylation often underestimate the ages of older individuals30,32,72,81. One possible explanation is that epigenetic aging follows a logarithmic trajectory across the lifespan; that is, DNA methylation changes occur rapidly in early life but decelerate with age82. Alternatively, given that the offspring of long-lived individuals appear to age more slowly than expected in human studies83,84, it is possible that individuals who survive to older ages possess biological mechanisms that retard age-related methylation changes. If the relation between DNA methylation and age at the CpG sites used in our model is nonlinear, this may explain why SVR, which accommodates nonlinear patterns, outperformed elastic net regression, which assumes linearity. Nonetheless, the SVR model only partially mitigated the underestimation of older individuals (Fig. 2). Teschendorff and Horvath (2025)85 have suggested that nonlinear methods, such as Gaussian processes, probabilistic kernel density-based approaches, or deep learning could better capture complex and previously unrecognized aging trajectories—particularly when integrated with explainable artificial intelligence tools—thereby improving biological age prediction and advancing our understanding of epigenetic aging. In contrast to prior studies that reported increasing prediction error with age86,87, our findings did not show such a trend, and reports of age overestimation in younger individuals remain limited. Future investigations are needed to elucidate the factors contributing to errors in DNA methylation-based age estimation. The genes located adjacent to the CpG sites that demonstrated strong age-related methylation changes in this study were VGF (VGF Nerve Growth Factor Inducible), KCNK12 (Potassium Two Pore Domain Channel Subfamily K Member 12), and ELOVL2 (ELOVL fatty acid elongase 2). Correlations between blood DNA methylation levels and age have been previously reported for VGF in humans, dogs, and brown bears47,62,65; for KCNK12 in humans and dogs61,65; and for ELOVL2 in humans, mice, and cats63,67,88. ELOVL2 is recognized as a robust epigenetic age marker in humans20. These three genomic locations are not tissue-specific, suggesting that hair roots may share common age-related epigenetic markers with other tissues. Although age-associated CpG sites do not necessarily influence gene expression, previous studies have reported that genes linked to these CpG sites are related to lifespan and age-related diseases. Notably, VGF has been associated with Alzheimer’s disease, whereas ELOVL2 has been linked to visual function89; deletion of ELOVL2 results in early-onset hair loss in mice90. As the epigenetics of hair remains an emerging research area91,92, further studies are required to clarify the potential relations between gene expression and the age-related methylation changes identified in this study. The limitations and future challenges of this study are as follows: (1) as previously noted, the selection of target genomic regions lacked sufficient rationale. Comprehensive genome-wide analyses may address this limitation. (2) Sex distribution in aged bears was skewed toward females, which may have influenced the results. In captive bears, sampling from aged bears (> 24 years old) was performed only when anesthesia was required for treatment purposes, which led to the absence of male samples during the study period. In wild bears, males born to resident females leave their birthplace (i.e., the study area) at around two years of age93, which makes it impossible to sample hair from aged males. (3) Our analysis was restricted to hair bulbs of the white sphere type, owing to the limited availability of the other types (i.e., white hook and black hook). To effectively apply the current age estimation method in field research, hair should be collected between spring and early August, when white sphere types are predominant60. Incorporating additional hair types in future analyses could enhance model’s applicability. (4) While we used 10 follicles per sample, further validation is necessary to determine whether accuracy is maintained when fewer follicles are used. Previous studies have not clearly defined the number of hairs required for DNA extraction, with reported values ranging from five to 3049,50. (5) It is essential to investigate whether hair storage conditions affect DNA methylation levels. This is particularly relevant for samples stored at varying temperatures over extended periods or those collected from hair-traps left in the field for several weeks. Hairs left outdoors for long periods of time showed a decrease in DNA amplification success94, suggesting that environmental exposure (e.g., rain, UV, and high humidity) contribute to DNA degradation. For hair samples collected outdoors, prior verification of DNA quality and quantity would ensure more robust and reliable results. (6) The body hair collected in field settings may not originate from the upper back, which was the source used in this study. Nevertheless, no significant differences in DNA methylation have been reported among various body hair sites—such as the calf, pubic area, armpit, and scalp—in humans49. The aforementioned issues are required to be addressed before our method can be applied to field-collected hair samples, such as those obtained using hair-traps. In conclusion, we developed the first hair-based age estimation model for brown bears, achieving an MAE of 3.192 years following LOOCV. In bear research, DNA obtained from hair-traps has traditionally been used to identify individuals and estimate population size. If our method can be applied to past or future hair-trap samples, it may enable the reconstruction of population age structures in wild bears. Although our hair-based model is less accurate than blood-based models and remains limited in estimating the age of single individuals, it offers a simpler, less invasive sample collection method and shows promise in determining population-level age distributions. Moreover, DNA methylation-based age estimation using blood in brown bears has proven applicable to other bear species33. Our hair-based method therefore holds substantial potential for cross-species application. While further accuracy improvements are necessary, age estimation using hair is expected to contribute meaningfully to bear conservation and management. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements We would like to express our sincere thanks to all staff at Noboribetsu Bear Park for providing bear hair and sample information. We wish to thank Hatsusaburo Ose and all the members of the Shiretoko Fishery Productive Association for their kind support. We are deeply grateful to all the members of the Shiretoko Nature Foundation for their generous support. We also thank everyone involved in sample collection. We would like to express our sincere gratitude to Dr. Miho Inoue-Murayama from Wildlife Research Center, Kyoto University for her support and helpful advice. Finally, we thank Editage (www.editage.jp) for English language editing. Author contributions S.N. designed the study, performed laboratory work, and constructed each age estimation model. S.N., N.M., K.H., H.S., M.Y., M.N., M.J., Y.Y., and M.S. were involved in sample collection. J.Y. and H.I. supported with the technical aspects of the experiment. S.N. and M.S. wrote the article with inputs from J.Y., H.I., and T.T. All authors reviewed the article. Funding This study was supported by funding from the Japan Society for the Promotion of Science (JSPS) (https://www.jsps.go.jp/english/e-grants/index.html) KAKENHI grant numbers JP19K06833, JP23K05312, JP25H01002, JP24KJ0304, and JP22K14910, Grant for Basic Science Research Projects from The Sumitomo Foundation (grant no. 200561), a grants-in-aid of The Inui Memorial Trust for Research on Animal Science, Japan, JST SPRING (grant no. JPMJSP2119), and World-leading Innovative and Smart Education (WISE) Program from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) (grant no. 1801). Data availability The data obtained from pyrosequencing analyses are available in Dryad at: 10.5061/dryad.h44j0zpzc. Declarations Competing interests The authors declare no competing interests. Ethical approval All procedures involved in sample collection from animals were conducted in accordance with the Guidelines for Animal Care and Use, Hokkaido University, and were approved by the Animal Care and Use Committee of the Graduate School of Veterinary Medicine, Hokkaido University (Permit Number: 1152, 15009, 17005, 18–0083, 19–0021, 20–0146, and 23 − 0014). In addition, all methods were carried out in compliance with ARRIVE guidelines. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement The data obtained from pyrosequencing analyses are available in Dryad at: 10.5061/dryad.h44j0zpzc.