Modeling transcriptomic age using knowledge-primed artificial neural networks

摘要

a Schematic of the artificial neural network architecture. Gene expression data is fed to the input layer, which is connected to the following hidden layer through gene-specific edges that are constructed based on pathway affiliation. In the following hidden layers, information is processed by the network in a pathway-centric manner culminating into a final linear pathway layer with one neuron per pathway, which also serves as an auxiliary output to monitor pathway aging states. Finally, the information from all pathway neurons is aggregated in the main output neuron, which generates the age prediction. b Ensemble setup. To improve the stability and accuracy of the final model, an ensemble model was constructed from individually trained networks by joining the separate models to the common input and output layers.