Development of an improved SVMp model for age prediction: exploring sequencing data from the 11-CpG panel

This study aimed to improve age prediction in a Danish cohort of 150 individuals by in-depth analyses of DNA methylation sequencing data. In previous work, 11 well-known age-associated CpGs located on 10 PCR amplicons associated with 10 distinct genes were selected, and a support vector machine model with a polynomial function (SVMp) was developed for age estimation in blood. The 10 amplicons were PCR-amplified after bisulphite conversion and the methylation levels were assessed by massively parallel sequencing. The 10 amplicons contained a total of 73 CpG sites, of which 63 CpG sites were shown to be medium to highly associated with age in this work (Pearson correlation coefficient, |r| > 0.3). More than 70 000 SVMp models with various combinations of the 63 CpGs were tested to identify the model with the lowest mean absolute error (MAE) and root mean square error (RMSE) in training and test datasets from the Danish cohort. A final model with 34 CpGs (SVMp-34CpG) was selected. The SVMp-34CpG model showed considerable improvement over many other forensic prediction models and achieved mean absolute errors of 1.64 years and 1.76 years for the training and test datasets, respectively. The SVMp-34CpG model indicated robust generalization and minimal prediction bias across the 18–50 years age range (MAE: 1.47 years), although slightly higher errors were observed for individuals aged 50–70 years (MAE: 2.62 years). This was most likely due to epigenetic drift, together with the small number of individuals analysed in this age group. Compared to the previous 11-CpG model, the SVMp-34CpG model showed superior accuracy in the Danish cohort by predicting the age of 91% of the individuals within 4 years of their actual age, versus 68% for the previous model.