The genetics of human aging: predicting age and age-related diseases by deep mining high dimensional biomarker data

Recent research efforts have shown compelling evidence of DNA methylation alterations in aging and age-related disease. The traditional formula of DNA methylation aging suffers from multiple hypothesis testing due to the interacting, high dimensional, and non-linear nature of the data. Neural network analyses have shown effectiveness on biological age prediction for its ability in learning interacting and nonlinear relationships. However, the high dimensionality of DNA methylation data often results in overfitting and poor generalization in neural networks. To solve this problem, we developed a neural network model that selects input features based on their correlations with biological age. We compared it with the traditional statistical regressions and other dimension reduction models in neural networks, such as neural networks with LASSO and elastic net regularizations and the dropout neural networks. The results showed that our model decreased the age prediction error to 2.7 years, outperforming all other models. In addition, we studied age acceleration in two age-related diseases (Down Syndrome and Schizophrenia). Our model is able to confirm age acceleration in Down Syndrome with a much smaller variance comparing to existing studies and find extrinsic epigenetic age acceleration (EEAA) in Schizophrenia, which is a weak pattern that other models couldn't detect. Our research is one of the first to adapt neural network algorithms in biological aging prediction. It can be applied to a wide range of high-dimensional biomarker data, and ultimately improve understanding of the aging process and benefit public health.