Independent validation of the VISAGE enhanced tool for age prediction in Koreans and analysis of platform transition

Abstract

Age prediction based on DNA methylation is emerging as a pivotal topic in forensic genetics, providing crucial information for criminal investigations. The Europe-based VISible Attributes Through GEnomics (VISAGE) Consortium recently reported a series of age prediction models by analyzing blood, buccal cells, and bone DNA using the Massively Parallel Sequencing (MPS) method, demonstrating high prediction accuracy with Mean Absolute Errors (MAE) of less than 4 years. To explore population-specific differences, we replicated the VISAGE MPS-based analysis using 300 blood and 150 buccal cell DNA samples from Korean individuals. Since the VISAGE models' equations are not publicly available, we constructed new models using multiple linear regression at the same CpG sites. Despite differences in the ranking of age-correlated CpG markers, our newly developed blood age model for Koreans exhibited a comparable MAE (3.4 years) to the VISAGE blood model (MAE of 3.2 years), whereas our buccal cell model demonstrated a slightly higher MAE (4.3 years) compared to the VISAGE buccal cell model (MAE of 3.7 years). Furthermore, to facilitate broader application across platforms, we developed additional models using the SBE method and conducted a comparative analysis of methylation levels with those obtained by the MPS method. This comparison revealed significant inter-platform differences (p-value <0.05) in methylation levels across all CpG sites included in VISAGE blood and buccal cell models. To address this issue, we developed platform-independent models by calibrating methylation levels using 11 control DNAs with varying methylation ratios (0 %-100 %). Although the calibrated data still exhibited statistically significant differences between platforms, the platform-independent model for blood achieved high prediction accuracy, with an MAE of 3.6 years. However, the buccal cell model displayed lower accuracy and still showed inter-platform variations, indicating that the calibration method is affected differently by CpG sites. Additionally, we optimized age prediction models for Koreans by employing various regression methods to the Korean MPS dataset. Both blood and buccal cell models achieved high age prediction accuracies, with MAEs below 3.4 years and 4.2 years, respectively. In conclusion, our study underscores the importance of considering both biological and technical factors in developing reliable DNA methylation-based forensic models.