Age prediction is a key component of forensic DNA phenotyping. DNA methylation has emerged as one of the most reliable and accurate biomarkers for age prediction due to its strong correlation with chronological age across populations. In this study, genome-wide DNA methylation sequencing was performed on peripheral blood samples from 147 unrelated healthy Han Chinese individuals. After quality control, 2,703,896 CpG sites were retained. A combined correlation, defined as the product of Pearson and Spearman correlation coefficients between methylation level and chronological age, was used to select the top 10,000 candidate age-related CpG sites. Elastic net regression was then applied for feature selection and initial model construction, yielding 51 candidate sites. Based on this feature set, four models, including elastic net, random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost), were constructed. Model comparison demonstrated that the elastic net model achieved the best overall performance, with a median absolute deviation (MAD) of 3.48 years and an R² value of 0.95 in the test set, indicating high accuracy and strong generalization ability. In contrast, RF, SVM, and XGBoost exhibited notable performance degradation on the test set, suggesting overfitting under the current sample size and feature structure. To further refine the feature set, a SHAP-based recursive feature elimination (SHAP-RFE) strategy combined with nested cross-validation was employed. Using the one-standard-error rule, a final set of 9 core CpG sites was identified. The elastic net model built on these 9 sites achieved a mean R² of 0.88 and a mean MAD of 4.11 years, demonstrating that high predictive performance can be maintained with substantially reduced model complexity. These findings establish a framework for screening age-related DNA methylation sites and demonstrate that a parsimonious panel of 9 CpG sites can achieve reliable age prediction in a Chinese Han population, providing a valuable reference for forensic age prediction.

Authors

• Sitong Liu (Institute of Forensic Science, Fudan University; School of Forensic Medicine and Science, Fudan University, China)
• Suhua Zhang (Institute of Forensic Science, Fudan University; School of Forensic Medicine and Science, Fudan University, China)
• Chengtao Li (Institute of Forensic Science, Fudan University; School of Forensic Medicine and Science, Fudan University, China)
• Anqi Chen (Institute of Forensic Science, Fudan University; School of Forensic Medicine and Science, Fudan University, China)