Testing for batch effect through age predictors

Transcriptome profiling has been shown really useful in the understanding of the aging process. To date, transcriptomic data is the second most abundant omics data type following genomics. To deconvolute the relationship between transcriptomic changes and aging one needs to conduct an analysis on the comprehensive dataset. At the same time, biological aging clocks constructed for clinical use needs to robustly predict new data without any further retraining. In this paper, we develop a transcriptomic deep-learned age predictor. Deep neural networks (DNN) are trained and tested on more than 6 000 blood gene expression samples from 17 datasets. We apply methods based on output derivatives of DNN to rank input genes by their importance in age prediction and reduce the dimensional of the data. We also show that batch effect in transcriptome datasets of healthy humans is indeed significant, but the existing normalization techniques, while removing technical variation quite effectively, also remove age-associated changes. So robust methods of age prediction are needed.