Fig 1.
Graphical representation of the methodology.
Workflow to build an age estimation model using methylation-sensitive high-resolution melting (MS-HRM).
Table 1.
Primer information and polymerase chain reaction (PCR) conditions.
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).
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.
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).
Fig 4.
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).
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).
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).
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).