BAGE: a Bayesian framework for age prediction based on PBMC gene expression data.

<h4>Background</h4>Estimating chronological age from biological data is increasingly important for clinical and research studies of aging and age related diseases. In regards to transcriptomic data, it has been proven that chronological age prediction widely depends on sample type. While single-cell RNA-seq based clocks using PBMC single-cell data have been reported, to our knowledge no publicly available bulk RNA-seq age predictor has been trained specifically on peripheral blood mononuclear cells (PBMCs). To address this gap, we aggregated 16 publicly available PBMCs bulk transcriptomic datasets comprising 174 healthy individuals (ages 4-81 years, sex balanced) and implement BAGE (Bayesian framework for age prediction from gene expression data). BAGE implements a Bayesian linear mixed model that incorporates dataset as a random intercept to capture batch effects and technical heterogeneity.<h4>Results</h4>In this study we evaluated multiple BAGE implementations under the leave-one-out cross validation strategy. The evaluated models vary in age parametrization, counts transformation and variable selection approaches. Best model performance achieved a coefficient of determination ([Formula: see text]) of 0.86 and mean absolute error (MAE) of 5.5, outperforming elastic net model and RNAAgeCalc tool on our PBMC data. This performance was achieved with square root parametrization of age and the resulting predictors constitute a concise consensus signature of 70 genes, stable across datasets. Via over-representation analysis the signature points to biological pathways related to natural killer cells.<h4>Conclusions</h4>Explicitly modeling study-level heterogeneity and using a signature specific to PBMC improved predictive accuracy relative to an elastic net base-line and the RNAAgeCalc multi-tissue calculator. The BAGE framework is adaptable to larger, heterogeneous cohorts and is readily extensible for integration with additional omics layers. Subject to external and longitudinal validation, the selected gene set could provide interpretable biomarkers of immune aging.