Bayesian neural networks for the optimisation of biological clocks in humans

DNA methylation is related to aging. Some researchers, such as Horvath or Levine have managed to create epigenetic and biological clocks that predict chronological age using methylation data. These authors used Elastic Net methodology to build age predictors that had a high prediction accuracy. In this article, we propose to improve their performance by incorporating an additional step using neural networks trained with Bayesian learning. We shown that this approach outperforms the results obtained when using Horvaths method, neural networks directly, or when using other training algorithms, such as Levenberg-Marquardts algorithm. The R-squared value obtained when using our proposed approach in empirical (out-of sample) data was 0.934, compared to 0.914 when using a different training algorithm (Levenberg Marquard), or 0.910 when applying the neural network directly (e.g. without first reducing the dimensionality of the data). The results were also tested in independent datasets that were not used during the training phase. Our method obtained better R-squared values and RMSE than Horvaths and Levines method in 8 independent datasets. We demonstrate that building an age predictor using a Bayesian based algorithm provides accurate age predictions. This method is implemented in an R function, which is available through a package created for predicting purposes and is applicable to methylation data. This will help to elucidate the role of DNA methylation age in complex diseases or traits related to aging.