3.7. Identification of circadian genes

GRO-seq datasets from multiple time points are combined in order to identify oscillating gene expression patterns. Taking our previous study for example [22], 8 GRO-seq datasets were collected from samples taken every 3 h throughout the 24-hour cycle. Previous studies indicated that about ten thousand genes are actively transcribed in mouse liver at a given time [29,31], based on which an expression threshold of 0.5 rpkm was selected, and inactive genes whose maximum transcription at all time points was lower than the threshold were excluded from downstream analysis. To make oscillation amplitudes comparable across different genes, the absolute transcription level (in rpkm) was converted to relative transcription which is defined as log2 fold change to the average transcription of the day (Fig 4).

A. Time course of transcription (in rpktm) of 3 genes. B. Input of JTK_cycle, showing relative transcription of each time point, which is derived from the log2 fold change to the mean of all time points. C. Example of JTK_cycle output. Predicted oscillating phases were adjusted according to the time of the first data point.

Cycling transcription of active genes can be detected by an R package JTK_cycle, which uses non-parametric algorithm to predict circadian expression patterns [26]. This program takes as input a matrix containing time-course transcriptions and outputs predicted cycling phase, amplitude, period, and P-value for each gene (Fig.4). Genes with P-value<0.05, amplitude>0.5, and period between 21 and 27 hrs were considered as oscillating genes. Note that JTK_cycle takes the first data point as time 0, thus the real oscillation phase needs to be adjusted with the offset of the actual starting time (Fig. 4). Sensitivity of JTK_cycle largely relies on the number of time points in the dataset, thus duplication or concatenation of datasets might be necessary for thorough discovery of cycling targets [29]. To eliminate false positive targets, one can apply multiple prediction programs, such as COSOPT, CirWave, and ARSER, and select commonly detected hits [32].