Adaptive feature selection framework for DNA methylation-based age prediction

Aging process is one of the main unsolved problems of modern biology that affects almost all living species, resulting from multiple interactions of genetics and environmental factors. Numerous studies have shown that DNA methylation changes are one of the most sustainable biomarkers in predicting biological age that has complex relationship with chronological age. This point shows the importance of selecting age-related CpG-sites. Most feature selection methods that have been proposed in this field are problem-dependent techniques for finding important age-related CpG-sites. However, in this study, we propose a general-purpose framework that is problem independent. This adaptive framework is proposed to find the best sequence of feature selection methods and the number of features that selected in each step according to the used dataset. To evaluate our proposed framework, we used two groups of DNA methylation dataset related to blood tissue and non-blood tissues from healthy samples. The results of our adaptive framework have been compared with four studies in terms of mean absolute deviation (MAD) and correlation (R-2) separately on blood and non-blood datasets. Our framework achieved MAD of 3.9 years and 5.33 years on the blood and non-blood test datasets, respectively. Also, a correlation (R-2) of 95.24% and 91.92% between chronological age and DNAm has been reported on the blood and non-blood test datasets, respectively. The experimental results show that our proposed framework was able to adaptively find the best feature selection method appropriate to the data that has an acceptable performance compared to other studies.