Introduction Circadian rhythms are daily, 24-hour (h) oscillations in physiology and behavior such as food consumption, blood pressure, metabolism, body temperature, and locomotor activity ,. These rhythms are thought to give an adaptive advantage by allowing an organism to anticipate changes in the environment and regulate physiology accordingly. Moreover, disruptions of circadian rhythms contribute to numerous pathologies including metabolic and cardiovascular disorders, cancer, and aging –. A molecular and cellular clock composed of transcriptional feedback loops generates these oscillations . The central loci of the mammalian clock are two small clusters of hypothalamic neurons called the suprachiasmatic nuclei (SCN), which constitute the master pacemaker that orchestrates rhythmic patterns of behavior and physiology throughout the organism . Remarkably, most tissues in the body also contain autonomous circadian clocks that are necessary for the rhythmic expression of clock output genes  and capable of sustained oscillations outside of the body (e.g. ). These peripheral clocks are principally regulated by stimuli downstream from the SCN, and are entrained by the SCN via a number of different physiological signals such as glucocorticoid production, core body temperature, or cAMP input (e.g. ,). Rhythmic physiology is thought to manifest from the transcriptional output of core oscillator components. Consequently, studies have been performed in several model systems to identify rhythmically expressed genes in both central and peripheral tissues , –. One consistent observation is that the vast majority of circadian transcriptional output is tissue-, and not locus-, specific, implying that both local and systemic cues heavily influence circadian output. In order to more fully understand the mechanism by which local and systemic signals translate into rhythms of physiology and behavior, a detailed understanding of the circadian transcriptome is necessary. To address this question, we have developed a high resolution temporal profiling experimental design in which samples are taken every hour for 48 hours and subjected to rigorous statistical analysis. This approach has the capacity to identify rhythmic output genes with precision and accuracy. We applied this method to the study of gene transcription in the liver , an organ system that receives and integrates systemic cues, as well as synchronized NIH3T3 and U2OS cells, conventional models of the autonomous cellular oscillator ,. Here we report the identification of thousands of circadian transcripts in the mouse liver. Surprisingly, using identical statistical methods dramatically fewer cycling transcripts were identified from two models of the autonomous circadian clock, NIH3T3 and U2OS cells. In addition, we found hundreds of transcripts in the liver that cycle at the second and third harmonic of circadian oscillations. Like circadian genes, these ultradian rhythms are severely dampened in ex vivo hepatocytes. Moreover, these rhythms are shifted in a restricted feeding paradigm, demonstrating their responsiveness to systemic cues. Results Wildtype C57BL/6J mice were entrained to a 12 h light, 12 h dark (LD 12∶12) environment before being released into constant darkness. Starting 18 h after the first subjective day (CT18), liver samples from 3–5 mice per time point were collected every hour for 48 h. In parallel, we collected a 48 h time course from two different cellular models of the circadian clock in order to study circadian output in the absence of systemic, circadian cues. After synchronization by forskolin shock, NIH3T3 cells were sampled every hour for 48 h, starting 20 h after synchronization. Likewise, a human osteosarcoma cell line, U2OS, was synchronized with dexamethasone and samples were collected every hour for 48 h, starting 24 h after shock. To confirm that these cells were properly synchronized, parallel cell cultures were transfected either transiently (NIH3T3) or stably (U2OS) with a circadian reporter gene, Bmal1:luciferase (Bmal1:Luc) and imaged every 10 minutes for several days to validate synchronization and rhythmicity (Figure S1). Total RNA was purified from these samples and Affymetrix arrays were used to assess global gene expression. To account for mode failure, two different statistical algorithms were then used to identify rhythmically expressed transcripts as previously described . The first algorithm, COSOPT , measures the goodness-of-fit between experimental data and a series of cosine curves with varying phases and period lengths. p-values are then calculated by scrambling the experimental data and re-fitting it to cosine curves in order to determine the probability that the observed data matches a cosine curve by chance alone. The second algorithm, Fisher's G-test , uses Fourier transforms to systematically screen experimental data for sinusoidal components. The probability (and thus, the significance) of any observed periodicity can then be tested using Fisher's g-statistic. Importantly, neither algorithm is sensitive to amplitude nor are they intrinsically biased towards any single period length, and they work with different underlying principles minimizing the risk of mode failure. These tests were corrected for multiple comparisons post hoc using the method described by Storey and colleagues,. Briefly, by examining the distribution of p-values from a given data set, an estimate of the proportion that are truly non-rhythmic can be derived. Using this approach to model the rate of false-discoveries, the p-value for each transcript, which estimates the frequency that a truly null observation will be labeled as significant, can be converted to a more stringent q-value which instead estimates the frequency that significant observations are truly non-rhythmic. At a false discovery rate  of 20 and 10 and 7 and 20 and 10 and 7 and 20 and 10 and 7 and 20 and 10 and 7 and <9 for 8 h genes) closely fit these data (median p-values of 0.001 and 0.002, respectively). Note in particular the logarithmic scale of the y-axis in both panels. (0.20 MB TIF) Click here for additional data file. Figure S3 Quantitative PCR validation of 12 h rhythmic transcription. A second collection of liver samples was performed and qPCR was used to assess the levels of endogenous mRNA. Blue traces represent microarray profiles from the original tissue collection and were plotted on the left axis; red traces represent fold changes observed in qPCR from the second tissue collection and were plotted on the right axis. Both core clock genes (A, B), 12 h genes (C–H), and 8 h genes (I–J) showed a close correlation between experiments. For additional quantitative PCR validation also see Figure 7A–D. (1.41 MB TIF) Click here for additional data file. Figure S4 A subset of 12 h genes from the liver revert to 24 h periodicity in different tissues. qPCR analysis was used to assess the RNA profile in multiple tissues of two genes, Hspa5 (A, C, E, G, I) and Armet (B, D, F, H, J), which cycle with 12 h rhythms in the liver. Although these genes do not show 12 h periodicity outside of the liver (unlike Hspa1b, Figure 3), in several tissues they show robust circadian rhythms (e.g., within the Kidney and Heart). The original liver microarray traces for Hspa5 (K) and Armet (L) (previously shown in Figure S3) have been reprinted here to ease comparisons between experiments. (0.89 MB TIF) Click here for additional data file. Figure S5 Relative phasing of core clock genes in liver, pituitary and NIH3T3 and U2OS cells. The timing of peak-expression of core clock genes in the liver (A), pituitary (B), NIH3T3 cells (C), and U2OS cells (D) was estimated by visual inspection and plotted on a circular phase map. (5.48 MB TIF) Click here for additional data file. Figure S6 Ingenuity pathway analysis of subcircadian genes. Rhythmic genes identified by COSOPT and Fisher's G-test at a false-discovery rate of <0.05 were analyzed using Ingenuity pathway analysis. The path designer tool was used to identify networks of rhythmic genes involved in cell division and cancer (A), protein secretion/ER stress response (B), NF-kB signaling (C) and lipid metabolism (D). Genes in red cycle with 24 h periods, genes in yellow cycle with 12 h periods, and genes in green cycle with 8 h periods. (1.52 MB TIF) Click here for additional data file. Figure S7 Circadian transcripts oscillate with modestly higher amplitudes than either 12 or 8 h genes. The amplitude of cycling transcripts was estimated by calculating the peak to trough ratio ( = percentile[0.95 , x]/percentile[0.05 , x]) and plotted as a histogram. For 24, 12, and 8 h genes, the majority of cycling transcripts had amplitudes less than 4-fold (A–C); however, circadian transcripts showed a significantly larger proportion of genes with amplitudes greater than 10-fold (A). (0.38 MB TIF) Click here for additional data file. Figure S8 Examples of ‘harmonics’ in 12 h genes. Microarray intensity is plotted against CT time for three genes which show ‘harmonics’ of circadian gene expression, Hsap1b (A), Dnaja1 (B), and Dsc2 (C). (0.56 MB TIF) Click here for additional data file. Figure S9 Amplitude comparison between liver and NIH3T3 cells. The amplitude of core clock genes was estimated by calculating the peak to trough ratio ( = percentile[0.95 , x]/percentile[0.05 , x]) and graphed alongside the amplitudes of the same genes in the liver. The differences in amplitude we observed were independent of microarray intensity between experiments as indicated by a comparison of the coefficient of variance (standard deviation/mean) for each probe (data not shown). (0.21 MB TIF) Click here for additional data file. Figure S10 12 h genes do not cycle in NIH3T3 cells. To validate the microarray profiling of NIH3T3 cells, a second time course of cycling 3T3 cells was collected every 2 h for 48 h. Quantitative PCR was used to measure the fold change of endogenous RNA which was plotted against CT time. Core circadian clock genes including Per 2 (A) and NR1D1 (B) oscillate with periods of approximately 24 h. In contrast, genes with 12-hour periods in the liver are arrhythmic (C–H). (0.68 MB TIF) Click here for additional data file. Figure S11 Comparison of Liver, Pituitary NIH3T3 and U2OS datasets. High temporal resolution profiling has been performed on samples from the liver, pituitary , U2OS and NIH3T3 cells. Cycling transcripts were detected in the liver and pituitary at a false-discovery rate of <0.05; rhythmic transcripts in U2OS and NIH3T3 cells were identified at a false-discovery rate of <0.1. The number of cycling genes common to each group was plotted as a Venn diagram. In (B), the number off cycling genes common to NIH3T3 cells (in black) and U2OS cells (in red) were plotted as a Venn diagram. (0.75 MB TIF) Click here for additional data file. Table S1 24 h cycling genes. Cycling genes were identified which had false-discovery rates less than 0.05 in both COSOPT and Fisher's G-test, as well as a COSOPT period length greater than 20 h and less than 30 h. (0.88 MB XLS) Click here for additional data file. Table S2 12 h cycling genes. Cycling genes were identified which had false-discovery rates less than 0.05 in both COSOPT and Fisher's G-test, as well as a COSOPT period length greater than 10 h and less than 14 h. (0.08 MB XLS) Click here for additional data file. Table S3 8 h cycling genes. Cycling genes were identified which had false-discovery rates less than 0.05 in both COSOPT and Fisher's G-test, as well as a COSOPT period length greater than 7 h and less than 9 h. (0.03 MB XLS) Click here for additional data file. Table S4 NIH3T3 cycling genes. Cycling genes were identified which had false-discovery rates less than 0.1 by COSOPT as well as a COSOPT period length greater than 20 h and less than 30 h. (0.02 MB XLS) Click here for additional data file. Table S5 U2OS cycling genes. Cycling genes were identified which had false-discovery rates less than 0.1 by COSOPT as well as a COSOPT period length greater than 20 h and less than 30 h. (0.02 MB XLS) Click here for additional data file. Table S6 GO annotation of 12 h genes. Cycling genes were identified which had false-discovery rates less than 0.05 in both COSOPT and Fisher's G-test, as well as a COSOPT period length greater than 10 h and less than 14 h. These genes were analyzed using Spotfire DecisionSite to identify over-represented GO annotation classes. (0.03 MB XLS) Click here for additional data file. Table S7 Ex vivo hepatocyte time course statistics. Primary hepatocytes were prepared from Per2-luciferase mice and shocked with dexamethasone to synchronize their circadian clocks. Starting four hours after dexamethasone shock, mRNA samples from these cells were collected every two hours for an entire day and quantitative PCR was used to assess the levels of endogenous mRNAs. Fisher's G-test and COSOPT were used to assess the likelihood that these traces were oscillating and estimate their period length. (0.02 MB XLS) Click here for additional data file. Table S8 Sampling the NIH3T3 and U2OS datasets at 4 h resolution yields similar results to previous profiling studies of circadian cell lines. In order to determine the importance of high resolution temporal sampling and false discovery-rate corrections, both the NIH3T3 dataset and the U2OS dataset were analyzed by COSOPT using one quarter of the time points to simulate 4 h sampling. At a p-value cutoff of <0.05, thousands of transcripts representing 5–10% of the genome were declared rhythmic by COSOPT, consistent with the results of previous studies. Increasing the sampling resolution to once every hour dramatically increased the number of genes with p-values<0.05. However, in both sampling conditions, very few of these genes demonstrated q-values<0.05, suggesting that the actual number of cycling transcripts in circadian cell lines is considerably lower than previously thought. (0.02 MB XLS) Click here for additional data file.