Ethics statement

All methods were performed following relevant guidelines and regulations and reported under ARRIVE guidelines. All experimental protocols and sample collections were approved by the ethical committee and conducted in strict accordance with the guidelines for the ethics of animal research established by the Wildlife Research Center of Kyoto University (approval number WRC-2022-010A). The samples collected from Asian elephants were collected and used with the agreement of each zoo.

Sample collection and DNA extraction

Blood samples (n = 53) were collected from 2004–2022 across 11 zoos in Japan from 25 captive Asian elephants (S1 Table) during regular veterinary health examinations. As of 2021, there are only 31 zoos in Japan that house 81 captive Asian elephant individuals [48]. Most of the zoos in Japan undergo husbandry training to collect blood samples during health examinations from captive Asian elephants to minimise invasive sampling. However, not all elephants are trained, and even well-trained individuals suddenly resist getting their blood drawn. Thus, to minimise such invasive sampling, our sample collection focused on 25 well-trained individuals who were not reluctant to collect blood at the time. Of the 25 individuals, three were male and 22 were female. As blood samples from individuals were collected multiple times over the years, samples ranging from a few months to 65 years of age were provided by zoos. Known birthdates or estimated birth years were informed by zoo staff taken from those recorded in the studbooks. Individuals had known birthdates if they were born in a captive facility in Japan or a range country (usually from Myanmar, Laos, or Thailand), while those recorded with an estimated birth year were wild-born imported individuals. Age for estimated birth years was most likely done by the broker at the time of importation or by zoo staff on arrival, based on individual morphometric measurements. Our sample set had a larger proportion of known birthdates than estimated birth years (six estimated individuals and 19 known individuals). Individuals with estimated birth years were born before the late 1970s. This reflects the management history of the population as the implementation of captive breeding as a population management tool was not fully implemented until the 1980s.

All samples were stored in a −80°C or −20°C freezer until use. DNA from the collected blood samples was extracted using the DNeasy Blood & Tissue Kit (QIAGEN, GmbH, Hilden, Germany) according to the manufacturer’s protocol. The DNA concentration was determined using a Qubit 4 Fluorometer with the Qubit dsDNA Assay Kit (Thermo Fisher Scientific, San Jose, CA, USA) using 1 μL of input DNA and was stored at −20°C until further use.

Standard DNA

A standard DNA was prepared to calculate the methylation status of samples from MS-HRM [28,41]. The negative (fully unmethylated) control standard DNA was synthesised using a whole genome amplification treatment on one DNA sample using the REPLI-g Mini Kit (QIAGEN, GmbH, Hilden, Germany). The positive (fully methylated) control standard DNA was obtained by fully methylating the above-synthesised DNA sample with CpG methyltransferase (M.Sssl, New England Biolabs, Beverly, MA, USA). The obtained standard DNA (0% and 100% methylated) was then purified using a High Pure PCR Product Purification Kit (Roche Molecular Systems, Pleasanton, CA, USA).

Bisulphite conversion

All blood DNA samples, and purified standard DNA (0% and 100%) were bisulphite-converted using the EZ DNA Methylation-Gold Kit (Zymo Research, Irvine, CA, USA) according to the manufacturer’s protocol. The DNA input used to prepare bisulphite-converted DNA samples was 300 ng. In comparison, a higher DNA input of 500 ng was used to prepare bisulphite-converted standard control DNA, considering that methylation ratios were needed to be prepared from 0% and 100% bisulphite-converted standard control DNA. The concentration of bisulphite-converted DNA was then measured with a Qubit 4 Fluorometer using the Qubit ssDNA Assay Kit (Thermo Fisher Scientific, San Jose, CA, USA) and subsequently adjusted to 5 ng/μL.

Identification of age-related epigenetic markers and primer designs

Genes with age-related epigenetic changes were identified via a literature search using databases such as Google Scholar and PubMed. Search terms included ‘age estimation’ or ‘age determination’ and ‘DNA methylation’ or ‘epigenetics’, and further focused on literature related to ‘mammals’, ‘MS-HRM’, and ‘pyrosequencing’ so primers can easily be modified for amplification of bisulphite-converted DNA. The study on elephant epigenetics using microarrays [47] was included in the search for comparative purposes. We selected target genomic locations adjacent to two genes namely RALYL (RALY RNA-binding protein-like) and TET2 (Tet methylcytosine dioxygenase 2). The methylation levels of CpG sites adjacent to these genes change with age. RALYL has previously been used to estimate the age of domestic dogs [49], cats, and snow leopards [28], and it is hypermethylated in cancer tissues [50]. TET2 has been studied in cetacean species [36,37,39], bats [30], and for its role in cancer development, where TET2 mutations were associated with cancer tissue hypermethylation [51,52], which leads to epigenetic age acceleration in humans [53]. We identified homologous sequences against the genomic regions of Asian elephants (GCA_014332765.1) [54] containing the target CpG sites using the Standard Nucleotide Basic Local Alignment Search Tool (BLASTN) provided by the National Center for Biotechnology Information (NCBI) [55]. PCR primers for amplifying Asian elephant sequences were designed using Methyl Primer Express v1.0 (Thermo Fisher Scientific, San Jose, CA, USA) and/or manually when required (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Primer information and polymerase chain reaction (PCR) conditions.

https://doi.org/10.1371/journal.pone.0294994.t001

Other candidate genes were selected during the literature search. However, amplification and dissociation curves for these genes were unsatisfactory following real-time PCR (RT-PCR) (Thermal Cycler Dice Real Time System TP800, Takara Bio Inc, Kusatsu, Shiga, Japan), and they were disregarded. Results remained unsatisfactory even when PCR conditions were modified several times (S1 Table in S1 File). Even when RT-PCR results were satisfactory, MS-HRM results would be negative or show no correlation between methylation and chronological age (S1 Fig in S2 File).

Methylation-sensitive high-resolution melting (MS-HRM)

To evaluate the applicability and PCR conditions for RALYL and TET2, RT-PCR was performed using the Thermal Cycler Dice System TP800 (Takara Bio Inc, Kusatsu, Shiga, Japan) before using MS-HRM. The amplification plots and dissociation curves were observed to confirm that the PCR conditions amplified the target region. The PCR conditions are listed in Table 1. In MS-HRM, bisulphite-converted DNA samples were PCR-amplified, followed by melting analysis. This follows the methods of Hamano et al. [41] and Qi et al. [28], for which procedures were modified accordingly (Table 1). PCR amplification was carried out using a Roche LightCycler 480 II (Roche Molecular Systems, Branchburg, NJ, USA) equipped with Gene Scanning Software (version 1.5.1.62 SP2) in a 25 μL total volume containing 1xEpiTect HRM PCR Master Mix (EpiTect HRM PCR kit; QIAGEN GmbH, Hilden, Germany), 750 nM of each primer, and 2 μL of template DNA (5 ng/μL of bisulphite-converted DNA). In addition to the samples, standard curves with known methylation ratios (0%, 5%, 10%, 15%, 25%, 40%, 50%, 75%, and 100%) were included in the assay plates and were later used to calculate the methylation rate of each RALYL and TET2 genes per sample. The ratios were prepared by mixing 0% and 100% methylated standard DNA in appropriate ratios to create a standard series. All reactions were performed in duplicate.

The chosen detection format was set to SYBR Green I/HRM dye. Following the initial PCR activation step at 95°C for 5 min to activate the polymerase. The 3-step cycling PCR conditions used were the same as those listed in Table 1. Following PCR amplification, HRM analysis was initiated by cooling the samples to 65°C for 1 s and then heated to 95°C at a ramp rate of 0.02°C/s. The continuous fluorescence at this step allows for data acquisition, where the fluorescence intensity was measured at 25 acquisitions per second during the entire process. When HRM analysis was performed, Gene Scanning Software first normalised raw melt curves to compare samples. In this normalising process, after MS-HRM was performed, the pre- and post-melt temperature regions were set to 67–68°C and 82–83°C for RALYL, respectively. For TET2, these were set to 66–67°C and 80–81°C, respectively. The threshold was set to zero for both genes, as the shape of the melting curve was important for the analysis of the overall methylation status of the amplicon. If the threshold was not set to zero and the temperature shift process was not performed, the shape of the melting curve would be distorted [41]. A difference curve was then derived from the first derivative of the melting curves after setting the data to the 0% methylated standard sample as a baseline. The maximum absolute value of the relative signal difference from the difference curves was defined as the ‘Df value’ for each sample.

Methylation rate calculation

PCR bias often occurs when amplifying bisulphite-treated DNA, as unmethylated DNA tends to amplify more efficiently than in methylated DNA [56,57]. A standard curve for each target site should be obtained to analyse methylation profiles accurately before measuring unknown methylated samples with MS-HRM [41]. From the MS-HRM analysis, standard curves were prepared for each experiment to calculate the methylation status of samples using Eq 1.

The Df values of the standard series were plotted for both RALYL and TET2, and a non-linear regression model developed by Warnecke et al. [56] (modified by Hamano et al. [41]) was used as follows: (1) where ‘M’ represents the methylation rate (percentage DNA methylation of each region of interest), ‘Dfmax’ is the Df value of the 100% methylated standard sample, and ‘a’ is the coefficient. Conducting the regression model and calculating the estimated value ‘a’ were performed using the ‘nls’ command in R version 4.2.0 [58]. The methylation rates (M) of the samples for RALYL and TET2 were calculated by substituting the Df value and ‘a’ value into the equation above.

Correlation between methylation rate and chronological age

For each target region, the correlation between methylation rate and chronological age was analysed using Pearson’s product-moment correlation coefficients by considering the p-value (p <0.05 was considered significant) and cor value. Figures were created using the package ggplot2 [59].

Age estimation model development and validation

To construct the age estimation model, a support vector regression (SVR) was built using the ‘e1071’ package [60]. SVR is suggested to be a robust choice that has high estimation accuracy with low-level overfit [61]. The SVR model parameters were optimised using the ‘tune.svm’ command with the optimisable parameters ‘cost’ and ‘epsilon’, and the fixed setting ‘type = eps-regression, kernel = radial, gamma = 0.5 or 1’. The optimised parameters are summarised in the S2 Table in S1 File. As some individuals were sampled more than once, cross-validation of the SVR model was performed using leave-one-individual-out cross-validation (LOIOCV). When the dataset used for model building and evaluation is small, the LOIOCV approach is recommended for model evaluation as it maximises the size of the training set in each iteration and further tests the robustness of the model to validate the overfitting of the optimised model. Moreover, LOIOCV takes into account individuals sampled multiple times and ensures that an individual is used as a training set while all other individuals are used as the testing set. This is repeated as many times as the number of individuals. To evaluate the deviation source in the model, a linear regression model with the age-estimation difference (predicted age–chronological age) as the response variable was constructed using the ‘lm’ command, and model selection was conducted using the ‘MuMIn’ package. The prediction accuracy of the regression model was assessed using the coefficient of determination (R²) and the mean absolute error (MAE) after LOIOCV analysis. Each age-responsive gene using chronological age was used as the independent known variable, while the predicted age for each region at different age-responsive genes was the dependent variable. To compare the relative predictive accuracy of each data type, the R² and p-value between predicted and chronological age were compared using simple linear regressions for each of the genetic regions and when both regions were combined into the model.

To evaluate whether sex influences the deviation of the age estimation model, generalised linear regressions were generated. Here, we denoted ‘Δage residual’ as the residual resulting from a linear model when regressing predicted age on chronological age, after adjusting chronological age as a covariate. The practical utility of the model was then further assessed by classifying individuals into appropriate age classes: calf (<1 year), juvenile (1–5 years), subadult (5–15 years), and adult (>15 years) based on Sukumar [62] as it could provide valuable information on population viability [63,64]. The predictive power of achieving the correct age classification was assessed by calculating the kappa values using the ‘vcd’ package, where values closer to 1 indicate good predictive power [65,66]. To increase the confidence, a one-way analysis of variance (ANOVA) was used, which was then further analysed using a post-hoc test with the ‘TukeyHSD’ command and ‘multcompView’ package.

Assessment of within-individual change: Methylation-based longitudinal age prediction

As blood samples of certain individuals were collected over time (S1 Table), individuals with at least two samples were chosen to further validate the model’s predictive accuracy by assessing the methylation-based longitudinal age prediction. A binomial test was conducted using the ‘binom.test’ command to evaluate if the age estimation model would correctly predict that samples collected from an individual at a later date were indeed from an older sample. A simple regression analysis was performed to evaluate if within-individual change comparisons would affect the overall model (the relationship between predicted age and chronological age).

Standard curves

For MS-HRM analysis, a standard curve was prepared for each genetic region to calculate the methylation status of the samples, according to Eq (1) (refer to ‘Methylation rate calculation’ in the ‘Materials and methods’ section). The standard curves of RALYL and TET2 for Asian elephants are shown in Fig 2. The standard line of TET2 was less linear compared to that in RALYL, likely due to PCR bias. Both genetic regions could clearly detect methylation rate differences between 40% and 100% by Df values. In comparison, those under 10% were hard to detect. TET2 showed a higher PCR bias and did not fit the regression line well (Fig 2). However, as most of the samples from TET2 were distributed between 10% and 35%, it appeared to yield predictions close to the standard curve. Standard curves influenced by PCR bias were seen in a previous study with other genetic regions [41]. The estimated ‘a’ value using Eq (1) for RALYL and TET2 is summarised in the S3 Table in S1 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Standard curves of RALYL and TET2.

The standard curves describe the relationship between methylation rate and Df value for (A) RALYL and (B) TET2. The Df value represents the raw fluorescence data from methylation-sensitive high-resolution melting (MS-HRM).

https://doi.org/10.1371/journal.pone.0294994.g002

Correlation between methylation rate and age

The methylation rates of RALYL (cor = 0.52, p <0.001) and TET2 (cor = −0.60, p <0.001) in Asian elephants were significantly correlated with chronological age (Fig 3). RALYL sites revealed that increasing age was associated with increased methylation, whereas TET2 sites suggested that increasing age was associated with decreased methylation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Correlation between methylation rate and chronological age.

Correlation between methylation rate and chronological age for (A) RALYL (cor = 0.52, p <0.001) and (B) TET2 (cor = −0.60, p <0.001) in Asian elephants.

https://doi.org/10.1371/journal.pone.0294994.g003

Age estimation model

The age estimation model was constructed based on the SVR model and performances were compared after LOIOCV analysis (Table 2 and S4 Table in S1 File). The model with both genetic regions was the best-performing model and was chosen as the final age estimation model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary and comparison of model output.

R² and mean absolute error (MAE) values for each age estimation model after leave-one-individual-out cross-validation (LOIOCV).

https://doi.org/10.1371/journal.pone.0294994.t002

As the methylation rate of both genetic regions was correlated with age, both regions were used as explanatory variables in the final age estimation model. The accuracy of the model for combining RALYL and TET2 was assessed using the equation generated from the regression analysis with all samples to generate the SVR model (Fig 4A). The regression of R² = 0.82 (p <0.001) indicated that most of the response can be attributed to age. The overall estimation accuracy before LOIOCV was 5.03 years by calculating the MAE (Fig 4A). The overall estimation accuracy of the final age estimation model after LOIOCV was estimated at 7.36 years (R² = 0.74, p <0.001) (Fig 4B and Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Final age estimation model.

Analysis of estimation accuracy and model performance on the final age estimation model combining RALYL and TET2. (A) SVR model before leave-one-individual-out cross-validation (LOIOCV) analysis (R² = 0.82, p <0.001). 95% confidence limits are shown on the regression line (grey). (B) Results after LOIOCV analysis. The black line represents a y = x diagonal line, and the region between the grey dash lines represents the mean absolute error (MAE) range, which was 7.36 years (R² = 0.74, p <0.001).

https://doi.org/10.1371/journal.pone.0294994.g004

Influence of sex on the age estimation model

Linear regression analyses were used to identify the factors that affect Δage residuals in the final age estimation model. There was no significant influence of sex on the Δage residuals thus, not affecting the model (p = 0.99) (Fig 5 and S5 Table in S1 File).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Influence of sex on the final age estimation model.

Δage residuals, defined as the residual from regressing predicted age on chronological age was calculated. Chronological age was adjusted as a covariate. (A) Compares the distribution of Δage residuals between females and males. The boxplots show group medians (solid line), inter-quartile range (box outline) and spread of data with outliers (whiskers) for each group. (B) The relationship between Δage residuals and chronological age. Linear regression analysis indicated that sex did not affect Δage residuals significantly (p = 0.99).

https://doi.org/10.1371/journal.pone.0294994.g005

Further age estimation model validation

The final model tended to slightly overestimate or underestimate the age of individuals. However, when the data was split into appropriate age classes of calf, juvenile, subadult, and adult, the final age estimation model developed from both genes showed that it could differentiate certain age classes (one-way ANOVA: F = 16.02, p <0.001) with a kappa value of 0.620 (0.510–0.731, p <0.001) (Fig 6), which indicates a good level of prediction for practical purposes. However, there were still difficulties in differentiating between the younger age categories, such as calf-juvenile and juvenile-subadult (Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Age group validation of the final age estimation model.

Box plots represent the predicted age of captive Asian elephants categorised into four categories of known age: calf (<1 year), juvenile (1–5 years), subadult (5–15 years), and adult (>15 years). The boxplots show group medians (solid line), inter-quartile range (box outline) and spread of data with outliers (whiskers) for each group. The final model developed by combining both genes, RALYL and TET2, could differentiate age classes according to the one-way ANOVA (F = 25.14, p <0.001) with a good level of prediction (kappa value = 0.620, p <0.001). A kappa value closer to 1 indicates good predictive power. Further analysis with the Tukey-Kramer post-hoc test indicates a statistically significant difference between calf-subadult (*p <0.05), calf-adult (****p <0.0001), juvenile-adult (**p < 0.01), and subadult-adult (****p <0.0001).

https://doi.org/10.1371/journal.pone.0294994.g006

Assessment of within-individual change: Methylation-based longitudinal age prediction

We compared the within-individual change from the same individual sampled more than once. Individuals who were sampled multiple times showed similar trends in methylation level changes with age (Fig 7). With a subset of 28 individuals for which two or more samples were collected at different times (S1 Table), the final age estimation model with both RALYL and TET2, predicted that ages were older for the later-collected samples (p = 0.03) (Fig 7), with 71% (20/28) cases that correctly predicted to be from an older aged sample. Moreover, the model, independently and in combination, consistently showed a significant relationship between predicted age and chronological age during within-individual change over time (R² = 0.86, p <0.001; Fig 7).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Within-individual changes in DNA methylation with age.

Age tracking for the final age estimation model. Predictions for individuals containing at least two blood samples collected over time in the data set (S1 Table). The final model was able to correctly predict 71% (20/28) cases of which samples collected from an individual at a later date were from an older sample. The dashed line represents the overall simple regression analysis between predicted and chronological age. There was a consistent significant relationship between predicted age and chronological age, even with samples collected over time (R² = 0.86, p <0.001).

https://doi.org/10.1371/journal.pone.0294994.g007

Age estimation based on two age-responsive genes using MS-HRM

The estimation accuracy and validation of the model improved when the methylation status for both genetic regions, RALYL and TET2, were combined into the age estimation model instead of separating the regions as independent models (S1 and S2 Files). This was consistent with previous findings where increasing the number of target sites enhanced age prediction accuracy [41,68]. However, the estimation accuracy in this study was lower than in a previous study, which constructed an age estimation model for Asian elephants using a different molecular technique; a DNA microarray chip with a mean absolute error (MAE) of 3.41 years [47]. The relatively low estimation accuracy in this study in comparison to the DNA microarray chip method could be caused by the limited number of genetic regions used (two in comparison to 37 regions). As mentioned previously, MS-HRM has its advantages regarding the time and cost required for the analysis. However, less information is obtained with MS-HRM compared to that in the microarray chip, which measures individual CpG methylation levels within each gene region [41,69]. Knowing the methylation status of each CpG site could be useful, as combinations of CpG sites with the best relationship with age are assumed to increase the accuracy of the age estimation model [70,71]. However, our model has the potential to provide an alternative option to current age estimation methods as we further tested the validation of the final age estimation model (Figs 6 and 7) and revealed that combining both genetic regions could provide a practical value to estimate age using MS-HRM. The preferable number of target sites is suggested to be three for practical reasons [72]. Thus, incorporating at least one more candidate genetic region to establish a more accurate model should be investigated for future applications.

The models tended to overestimate younger individuals and underestimate older individuals (Fig 4). Similarly, previous studies that have attempted to construct an age estimation model using DNA methylation have demonstrated difficulties in accurately predicting chronological age with a small error, especially for older individuals [73,74]. However, other studies have reported little prediction error in younger individuals and observed a similar increase in prediction error as chronological age increases [41]. Change in prediction error across age could be due to younger individuals having a tendency to age at a faster rate. Moreover, samples of older individuals are often difficult to obtain, which could have caused the increase in prediction error in older ages. Although the exact reasons for less precise age prediction among younger and older individuals remain unclear, it is suggested that this might be attributable to individual differences in the rate of methylation change that accumulate with age [41] due to methylation saturation, in which CpG probes reach 0% or 100% methylation later in life, in addition to other age-related processes [74].

In this study, candidate age-related epigenetic genes successfully used in other species were selected to assess methylation rates for Asian elephants. TET2 is a well-known epigenetic age-related gene that has been used as a region to estimate chronological age across several mammalian species [30,36,37,39]. In the MS-HRM analysis using felids, RALYL showed significant correlations with chronological age [28]. Using the same genes, the age estimation model developed in this study showed significant correlations between the methylation rate of each gene and chronological age. However, the specificity may not be as high compared to that in methylation microarrays for the abovementioned reasons and as it quantifies methylation across the genome. This again accounts for the lower estimation accuracy in this study compared to that in the microarray technique. For these reasons, in MS-HRM analysis, careful primer design is required. The primers used in this study were specific to Asian elephants, as they were carefully designed using homogenous gene regions searched against the readily available Asian elephant genome [54]. In addition, TET2 is an evolutionary conserved gene that belongs to the TET family in mammals [75] and plays a key role in the epigenetic process across several mammals. In contrast, though not well studied, RALYL is suggested to be conserved in chimpanzees, rhesus monkeys, dogs, cows, and mice [76]. This suggests that TET2 and RALYL could be conserved in other elephant species, however, primers need to be specific and designed with care when using MS-HRM.

As mentioned in the introduction, MS-HRM is an easy-to-use, traditional polymerase reaction. However, like any other PCR reaction, MS-HRM has the issue of PCR bias, where methylated templates are less effectively amplified than in unmethylated templates. Due to this PCR bias, differences in methylation rates are hard to detect, especially in lower-methylated regions [41]. To avoid this, sequences without any PCR bias need to be identified, or primers that reverse such PCR bias need to be designed. In addition, if CpG sites in a CpG island were intended to be analysed, designing primers might become difficult as they should include as few CpG sites as possible [41]. These issues were observed in the ZFHX3 gene (S1 Table in S1 File), a candidate gene referred to in the microarray study [47], as the genetic sequence included many CpG sites. This caused difficulties in primer design and could have impacted the bisulphite conversion efficiency. These complications using MS-HRM analysis were observed in most of the genetic regions identified in the microarray study, as it quantified individual CpG sites for each genetic region across the genome.

The use of captive known-age individuals is required to construct the basis of the age estimation model because the equation of methylation status needs to be generated to construct a reliable model. In the case of wild individuals, information on chronological age is expected to be unknown (depending on the area). Then, with the unknown sample, the expected possible deviation range needs to be considered in the equation of the constructed model. An additional inspection to check if the DNA methylation values fall within the expected range would be needed [77]. Studies to test the applicability of the age estimation model have been assessed in certain species and had no significant difference between the methylation rates of captive and wild individuals [78].

Comparison with other species and other chronological age estimation methods

When we compared the age estimation error (MAD/MAE) to the average life span from previous studies that used MS-HRM to predict age, humans had an accuracy rate of 90.36%, considering its MAD of 7.71 years to an average life span of 80 years [41]. A study using domestic cats showed an accuracy rate of 78.70%, with a MAD of 3.83 years and an average life span of 18 years [28]. In comparison to these previous studies, this study constructed an age estimation model with an accuracy rate of 87.73%, considering the average life span of Asian elephants to be 60 years. When compared to the current methods to estimate age in elephants, dental methods have an accuracy rate of 73.3–90% (error of −6 to +16 years) [13,79] and rumbles with 70% [80]. Age estimation accuracy/error rate of dung measurements was not mentioned. However, it was suggested that due to the high variability in dung measurements in each age class, there were limitations in predicting an age structure for young elephants and those older than 20 years [17]. In addition, the most popular method of estimating elephant age is morphological observations such as shoulder height. However, we could not compare as the accuracy/error rate was not mentioned [9]. Moreover, age estimation by morphological observations could be heavily based on experience. If the observer is inexperienced, the accuracy and precision of the estimated age could be reduced. The only methylation studies on Asian elephants were investigated using methylation microarrays, with a higher accuracy rate of 3.41 years (MAE) compared to our study [47]. However, microarrays require a minimum concentration of 50 ng/μL, with a minimum input of 200–300 ng of bisulphite converted DNA [81]. Considering these requirements, this study only needed 5 ng/μL of bisulphite-converted DNA. Although a DNA input of 300–500 ng was used as a starting material, MS-HRM can measure DNA inputs of 20 ng/gene, while the manufacturer’s protocol for bisulphite conversion states that samples with at least 500 pg could be used. In comparison, another single gene technique similar to MS-HRM; pyrosequencing, requires at least 150 ng of DNA input. However, due to low amplification efficiency, it is recommended that the standard control DNA input should be greater than the sample [46]. With the use of MS-HRM, low DNA input will be beneficial for precious samples, especially for endangered species like Asian elephants. Thus, in comparison to other age estimation methods, MS-HRM has the potential to be used as an alternative technique to estimate age in Asian elephants, considering its advantages of being able to measure age genetically from a wide range of methylation status within the genetic region of interest in a labour-, time-, and cost-effective manner, preventing any observation bias from current age estimation methods.

Age estimation model limitations

A sample size of at least 70 or more is suggested [82], as it could decrease the MAE, standard deviation, and relative error, leading to improved accuracy. The small sample size reflects the difficulty of obtaining samples across Japan due to the limited number of individuals in captivity. Furthermore, husbandry and management should be considered within zoos across Japan, as mentioned in the methods, along with the longevity and life-history status of the elephants. Due to the limited number of individuals, our model lacked equal representation between age groups, especially between 15–30 and 40–50 years. However, one of the advantages of this study was the comparison of individual changes across individuals sampled multiple times over the years, with the final age estimation model showing increasing methylation levels with increasing age within the same sampled individual.

In this study, sex did not affect the final age estimation model. This finding indicates that no sex difference existed in age-dependent DNA methylation changes or individual variability. Previous studies have contradicting results on sex-specific age-related DNA methylation changes. Among humans, a few studies have indicated sex-specific age-related DNA methylation changes [83], while others indicate no significant difference [43]. Several studies have suggested that model accuracy improved when separating males and females in a study from felines [28], while others have suggested that separating sex does not significantly improve the age estimation model [36,78]. Notably, this study had small sample sizes for males (n = 10) compared to females (n = 43) which could have considerably impacted whether the model was indeed age-dependent or not. The small sample size was reflected by the small captive male populations of approximately only 20 individuals across Japan, which hinders sample collection. Further validation of the age estimation model on an independent data set with a balanced sex ratio is required before establishing this method as a widespread tool. Increasing samples from more known-age individuals, especially in the age range between 15–30 and 40–50 years, and increased representation of sex could foster further development and refinement of the model. However, it should be noted that other age estimation models with low sample sizes (<70 samples) still produced an age estimation model with small errors [30,33,37,84]. Furthermore, highly accurate or precise models may not be required for all applications [6,33], as it ultimately comes to how the model is being used and for what purpose.

Future perspectives

Approximately 82% of the variation observed in DNA methylation patterns among Asian elephants in the constructed model could be explained by the difference in chronological age. The remainder of the variation (18%) could potentially be due to other factors, such as stress exposures in captivity and/or genetic variability. Recent findings suggest that DNA methylation of certain genetic regions can be influenced by biological, social, and environmental factors [85]. In contrast, others are linked to genetically programmed ageing and are less influenced by such factors. For example, multiple studies have shown that DNA methylation pattern can be altered by disease status (e.g., cancer, cardiovascular disease, diabetes, and Alzheimer’s disease), demographic and lifestyle characteristics (e.g., race, smoking, and alcohol consumption), diet, and environmental exposures (e.g., sun exposure and air pollution) in humans [85–87]. In non-human animals, these aspects are not as well investigated. However, there are studies on the effects of social status in naked mole rats (Heterocephalus glaber) and baboons (Papio) [88,89], hibernation in ground squirrels (Ictidomys tridecemlineatus), brown bats (Eptesicus fuscus), and yellow-bellied marmots (Marmota flaviventer) [90–92], and temperature change in zebra finches (Taeniopygia guttata) [93] on DNA methylation patterns. Thus, different tendencies in DNA methylation patterns between individuals could be affected by different factors. However, in this study, other than sex, other biological, social, or environmental factors were not considered. To assess the methylation of certain genes as a useful marker, further studies must evaluate and examine the differences between chronological age and biological age, and the factors that cause them.

Epigenetic age estimation models have been well investigated using blood and tissue samples. High age correlation has consistently been observed in heterogeneous tissues (such as blood, brain, buccal epithelium, saliva, liver, lung, and kidney), with blood, brain, and saliva samples demonstrating slightly higher accuracy [18,94]. However, most of these studies are based on methylation arrays, and using brain, liver, lung, or kidney samples is not idealistic to estimate age for future conservation purposes. To our knowledge, only studies on blood and saliva samples use MS-HRM to estimate age [28,41,46]. In Hamano et al., the model that used saliva samples demonstrated a slightly higher accuracy with a difference of 1.46 years (MAD) than in the model developed with blood samples [41,46]. However, we assume this is most likely due to the larger sample size. As our model was constructed using blood samples, the accuracy might not significantly differ from other tissues for the abovementioned reasons unless the sample size and number of genetic regions increase. However, using the age estimation model for conservation efforts constructed in this study is currently limited as it can only be applied to blood samples. DNA methylation is suggested to be chemically stable compared to other age-related biomarkers as it has been assayed from historic bone samples or ancient DNA from 38,000–60,000 years ago [95,96]. This could suggest that methylation biomarkers may be more suitable for age estimation tools in comparison to other age estimation techniques. MS-HRM studies using faecal samples have only been performed in humans [97]. No studies have been made to assess the use of faecal samples to estimate age through methylation data. Although model accuracy may not be as high as in heterogeneous tissues due to DNA quality, developing a non-invasive molecular technique to estimate age from degraded DNA, such as faecal samples would still be valuable. Extending age estimation techniques to non-invasive samples, such as faecal samples would enable researchers to collect samples in the field without disturbing the animals, avoid long-term visual surveys, and limit distress to a minimum. When developing an age estimation model from different types of samples, it is advised not to apply an age estimation model developed, for example, saliva samples to blood samples or mixed samples as methylation profiles can change [46]. Nevertheless, it would still be worthwhile to explore the development of age estimation models using DNA methylation from low or non-invasive samples, such as hair or faecal samples using MS-HRM for future use in conservation efforts.