Coordination of rhythmic RNA synthesis and degradation orchestrates 24- and 12-h RNA expression patterns in mouse fibroblasts

Significance

Circadian RNA expression is essential to regulate a plethora of downstream rhythmic processes. Both transcriptional and posttranscriptional mechanisms are important to drive rhythmic RNA expression; however, the extent to which each process contributes to this rhythmicity remains controversial. By monitoring RNA dynamics during a circadian cycle in mouse fibroblasts, we found that rhythmic RNA synthesis is the primary contributor of 24-h RNA rhythms, while rhythmic degradation is more important for 12-h RNA rhythms. Interestingly, core clock RNAs are regulated by multiple rhythmic processes and have the highest amplitude of synthesis and degradation, presumably critical to drive robust cell-autonomous rhythmicity. Overall, our study yields invaluable insights into the temporal dynamics of both 24- and 12-h RNA rhythms.

Abstract

Circadian RNA expression is essential to ultimately regulate a plethora of downstream rhythmic biochemical, physiological, and behavioral processes. Both transcriptional and posttranscriptional mechanisms are considered important to drive rhythmic RNA expression; however, the extent to which each regulatory process contributes to the rhythmic RNA expression remains controversial. To systematically address this, we monitored RNA dynamics using metabolic RNA labeling technology during a circadian cycle in mouse fibroblasts. We find that rhythmic RNA synthesis is the primary contributor of 24-h RNA rhythms, while rhythmic degradation is more important for 12-h RNA rhythms. These rhythms were predominantly regulated by Bmal1 and/or the core clock mechanism, and the interplay between rhythmic synthesis and degradation has a significant impact in shaping rhythmic RNA expression patterns. Interestingly, core clock RNAs are regulated by multiple rhythmic processes and have the highest amplitude of synthesis and degradation, presumably critical to sustain robust rhythmicity of cell-autonomous circadian rhythms. Our study yields invaluable insights into the temporal dynamics of both 24- and 12-h RNA rhythms in mouse fibroblasts.

Circadian rhythms are the internal timing mechanisms that allow organisms to anticipate and respond to daily changes in the environment, and most organisms on Earth exhibit circadian rhythms in their behavior, physiology, and biochemical processes (1, 2). Identification of rhythmically expressed genes has been an active area of research in recent decades, as they are believed to ultimately drive the diverse range of rhythmic biological processes. Circadian transcriptome studies have demonstrated that rhythmic RNA expression is pervasive in various organisms and their rhythmic patterns must be tightly regulated to ensure correct period, phase, and amplitude (reviewed in ref. 3).

For a molecule, such as RNA or protein, to exhibit rhythmic expression, its synthesis, degradation, or a combination of both must be rhythmic. In addition, the average half-life of the molecule must be short enough (<10 h) (4–6). The mechanisms of driving rhythmic RNA synthesis (i.e., transcription) are well established in mice. The “core” mammalian clock mechanism, composed of interlocking transcription–translation feedback loops, is considered the primary force of generating cell-autonomous circadian rhythms and rhythmic gene expression (7). Many of the core clock genes, such as Arntl (or Bmal1), Clock, Ror, and Nr1d (or Rev-Erb), encode transcription factors and activate rhythmic RNA transcription by recognizing target DNA motifs in the promoter sequence of the target genes (7, 8). These studies uncovered that approximately 5 to 40% of transcriptomes are rhythmic in the mouse depending on the tissue and more than 50% of transcripts are rhythmic in at least one tissue (9–12). Previous studies estimated the extent of rhythmic RNA synthesis in driving rhythmic RNA expression; however, there has been significant variation between studies, ranging from 20 to 85% in the mouse liver (5, 13–15).

Compared to RNA synthesis, however, much less is known about how rhythmic RNA degradation is regulated and how much rhythmic RNA degradation contributes to driving and sustaining rhythmic RNA expression. Several "posttranscriptional" mechanisms (e.g., miRNA, poly(A) tail length, RNA methylation, and RNA-binding protein, etc.) regulate rhythmic RNA expression, and many of them directly influence mRNA stability (16–28). This suggests that rhythmic RNA degradation is an important posttranscriptional mechanism driving rhythmic RNA expression. In support of this, RNA degradation of core clock genes, such as Period2 (Per2) and Cryptochrome1 (Cry1), is faster when their level is declining compared to when it is accumulating (21, 22). A recent study combining transcriptomic data with mathematical estimation also predicted that at least 35% of rhythmically expressed mRNA are rhythmically degraded (5).

Given that rhythmic degradation is energetically cost-effective to drive RNA rhythms (5) and the regulatory mechanism for RNA degradation is more diverse (29), we hypothesized that rhythmic RNA degradation is one of the major posttranscriptional mechanisms regulating rhythmic RNA expression. To test this, we analyzed RNA dynamics using a metabolic labeling approach throughout the circadian cycle on a genome-wide scale in NIH3T3 cells. We also evaluated the impact of Bmal1 or the core clock machinery by knocking down Bmal1, as Bmal1 is one of the core circadian clock genes critical for cell-autonomous rhythmicity (30). Deciphering the dynamics of rhythmic RNA expression deepens our understandings of how the core circadian clock machinery orchestrates RNA rhythms and ultimately regulates rhythmic downstream processes.

Results

Monitoring Rhythmic RNA Dynamics in Mouse Fibroblasts using Metabolic Labeling.

To monitor RNA dynamics over a circadian cycle, we used a metabolic labeling approach, as it allows us to simultaneously measure the synthesis and degradation kinetics from a single sample and minimally interferes with cell metabolism and physiology (29, 31). We used the Bmal1-luc cell line, in which a luciferase reporter gene driven by Bmal1 promoter is stably expressed in NIH3T3 cells (32), allowing us to monitor the rhythmicity of the cells in real time. We also knocked down Bmal1, one of the core clock genes critical for cell-autonomous rhythmicity (30), to evaluate whether the rhythmicity is driven by Bmal1 or the core clock machinery (Fig. 1A). We synchronized cells by adding 50% horse serum to the culture media (33) and labeled newly synthesized RNAs with 4-thiouridine (4sU) every 2 h for 2 h starting at T22 (i.e., 22 h after the end of serum shock) (Fig. 1A). We opted to use "pulse-in" procedure, as this caused little or no disruption to cellular rhythmicity (SI Appendix, Fig. S1A). We extracted total RNA at 12 time points from T24 to T46, corresponding to one circadian cycle (Fig. 1A). The 4sU-labeled RNAs were subsequently biotinylated and pulled-down with a pull-down efficiency of 50 to 75% in all samples (SI Appendix, Fig. S1B), comparable to other reports (34, 35). We used higher time resolution (2-h interval) with fewer biological replicas (n = 1) given the cost, potential technical variability between batches, and most importantly the community guidelines for genome-scale analysis of biological rhythms (36). We then subjected both total and newly synthesized RNA fractions to transcriptomic analyses (RNA-seq and 4sU-seq, respectively) (Fig. 1A). As expected, reads from RNA-seq were mapped predominantly to exons, while those from 4sU-seq were mapped more to introns and intergenic regions both in control and Bmal1 KD cells (Fig. 1B). Approximately 10% increase of intron reads as well as 5% increase in intergenic reads in 4sU-seq compared to RNA-seq (Fig. 1B) are comparable to previous studies (37, 38), given the concentration (400 µM) and labeling duration (2 h) of 4sU in our study. We used INSPEcT (39), an R package, to derive the rates of RNA synthesis, processing, and degradation and to quantify the levels of pre- and mature RNAs (Fig. 1C) at each time point, as it was built for a dynamic system and hence can handle time-series data and can computationally normalize the pull-down efficiency between each time point (39).

Fig. 1.

Rhythmic 4sU-seq and RNA-seq analysis in NIH3T3 cells. (A) Experimental design. Cells were plated 24 h before lentiviral transduction (72 h) containing scrambled shRNA (shCtrl) or shRNA targeting Bmal1 (shBmal1) (40, 41). Cells were then incubated with 50% horse serum for 2 h to synchronize cellular circadian rhythms. Twenty-two hours after the serum shock, 400 mM 4sU was added to the culture media for 2 h to label newly synthesized RNAs of the first sample. At the end of the 2-h incubation, RNAs were harvested. We repeated this procedure every 2 h starting from T22 until T44. Newly synthesized RNAs were then isolated from total RNAs, and both total and newly synthesized RNAs were subjected to RNA-seq analysis. (B) Percentage of read counts mapped to exon (dark purple), intron (light purple), or intergenic region (gray). The average from all time points in each dataset was used. (C) Schematic representation of the lifecycle of an RNA.

We first performed extensive quality checks of our dataset. First, successful knock-down (KD) of Bmal1 was verified both by Bmal1-luc bioluminescence output and the levels of Bmal1 RNA in both RNA-seq and 4sU-seq (SI Appendix, Fig. S1 C–F). Second, we verified that the RNA expression level was very strongly correlated between RNA-seq and 4sU-seq data at each time point (Pearson correlation coefficient > 0.95) in all samples (SI Appendix, Fig. S2). Third, the level of mature RNA correlated with the RNA synthesis and degradation rates, either positively or negatively, as expected (SI Appendix, Fig. S3). Fourth, we compared our dataset to the previous 4sU-seq study in NIH3T3 cells (42) and found a strong correlation for mature RNA level (Pearson r = 0.672) and synthesis rate (Pearson r = 0.578) between the two datasets (SI Appendix, Fig. S4 A and B). The correlation for RNA half-lives was still strong (Pearson r = 0.611), although the median half-life of our dataset was shorter than that of the previous report (this study: 2.27 h, Schwanhäusser: 10.04 h) (SI Appendix, Fig. S4 C and D). This could be due to 1) the difference between static (Schwanhäusser) vs. dynamic systems (this study), 2) the effect of serum shock in our dataset, or 3) the effect of viral transduction in our dataset, although absolute RNA half-lives are often inconsistent between datasets even with the same method and the same cells (29). Nevertheless, the distribution of RNA half-lives in our dataset was similar to what was estimated from the mouse liver over a circadian cycle and free of serum shock or viral transduction (43). Overall, these data demonstrate that our datasets are of high quality.

Bmal1 or the Core Clock Machinery Drives Rhythmic RNA Synthesis but not Degradation.

INSPEcT calculates the rates of synthesis, processing, and degradation at each time point using four variables: pre-RNA level (or intronic RNA-seq RPKM), total RNA level (or exonic RNA-seq RPKM), pre-RNA level in the nascent fraction (or intronic 4sU-seq RPKM), and total RNA level in the nascent fraction (or exonic 4sU-seq RPKM), while also accounting for the 4sU incubation time (which is 2 h in our case) (SI Appendix, Fig. S5 and Datasets S1 and S2). After the analysis with INSPEcT, we obtained a total of 11,313 RNAs that met our analytical criteria. We then used MetaCycle (44) (20 < τ < 28) to statistically determine the rhythmicity of the rates of RNA synthesis, processing, and degradation and the levels of pre- and mature RNAs. All the rhythmicity statistics and parameters shown in this study were calculated by MetaCycle. We did not use a stringent statistical threshold to determine the rhythmicity because circadian transcriptome output in cell cultures is not as robust as those from tissues (45). We identified 653, 272, 336, 685, and 6 RNAs whose synthesis rate, pre-RNA level, processing rate, mature RNA level, and degradation rate, respectively, were rhythmic in control cells (B.H. q < 0.25) (Fig. 2A and Dataset S1). Profiles of RNAs with five lowest B.H. q values for each process are shown as examples (SI Appendix, Figs. S6–S8). Rhythmic mature RNAs included most of the core clock genes, such as Arntl (or Bmal1), Period (Per)1-3, Nr1d1-2 (or Rev-erbα/β), and Dbp (Dataset S1). The number of rhythmic mature RNAs was also comparable with a previous study using NIH3T3 cells with a similar statistical threshold and experimental design (SI Appendix, Fig. S1G), and 23 (or 3.4%), 92 (13.4%), and 111 (16.2%) RNAs were commonly rhythmic (B.H. q < 0.25) between our dataset and Hughes dataset (n = 1, one 24-h cycle, 1-h interval), Hughes dataset (n = 1, one 24-h cycle, 2-h interval), or Schick dataset (n = 2, one 24-h cycle, 4-h interval), respectively. These overlapping RNAs include both core clock genes (Nr1d2, Cry1, Per2, and Per3) as well as clock-controlled genes (Mocs2, Lss, Tomm20, Zfas1, etc.). The number of overlapping RNAs is not very high, presumably due to the differences in the statistical and experimental conditions (e.g., rhythmicity threshold, microarray vs RNA-seq, etc.). However, the results from circadian transcriptomic studies are often inconsistent, and only a handful of genes were found to be commonly rhythmic between different experiments, even in cases when the exact same tissue was examined (46). Globally, the rhythmic mature RNA expression had a bimodal phase distribution and was highly enriched around T26 and T38 (Fig. 2 B and C). A similar phase distribution pattern was observed for pre-RNAs (Fig. 2 B and C), while both RNA synthesis and processing had a single peak at around T38 (Fig. 2 B and C). As predicted, we found that DNA recognition motifs for core clock proteins were enriched among those with rhythmic synthesis, whereas an RNA recognition motif for RBM38, a regulator of alternative splicing (47, 48), was highly enriched among RNAs with rhythmic processing (SI Appendix, Fig. S9). Nevertheless, the biological significance of this enrichment or rhythmic RNA processing is unclear, as processing includes multiple steps, such as (potentially cotranscriptional) splicing, capping, polyadenylation, and nuclear export, and the processing time is much shorter than a circadian cycle (median: 12.8 min in control cells) (Fig. 2H: blue). Rhythmic degradation of RNAs was not robust in Bmal1-luc NIH3T3 cells (Fig. 2 A–C) (5, 6).

Fig. 2.

Rhythmicity of RNA dynamics in NIH3T3 cells. A total of 11,313 genes were analyzed in all datasets. (A and D) Number of RNAs whose RNA synthesis (green), processing (yellow), or degradation (red) as well as unprocessed (gray) or mature RNA (blue) levels were rhythmic in control (A) or Bmal1 KD (D) cells with various statistical thresholds (B.H. q < 0.25, 0.15, or 0.05). (B and E) Phase-sorted heatmap of RNAs with rhythmic RNA synthesis (green), unprocessed RNA levels (gray), processing (yellow), mature RNA levels (blue), and degradation (red) in control (B) or Bmal1 KD (E) cells with the rhythmicity cutoff as B.H. q < 0.25. Color represents Z-scores, in which red is high while blue is low. Each line represents one RNA. (C and F) Circular histograms representing the peak phase distribution of rhythmic RNA synthesis (green), processing (yellow), or degradation (red) as well as pre- (gray) or mature RNA (blue) levels. Each bin represents 2 h. Number of genes are represented by the radius, and the outer radius represented 170 (synthesis), 60 (pre-RNA), 120 (processing), 120 (mature RNA), and 3 (degradation) RNAs in control and 30 (synthesis), 3 (pre-RNA), 100 (processing), 6 (mature RNA), and 3 (degradation) RNAs in Bmal1 KD cells. (G) Distribution of relative amplitude (rAMP) for rhythmic RNAs in control (blue) and Bmal1 KD (orange) cells. Boxplots represent two quartiles ± 1.5 interquartile range from the median (midline). *; P < 0.05, N.S.; P > 0.05 (two-tailed Student’s t test). (H and I) Distribution of RNA processing time (H) and half-life (I) in control (blue) and Bmal1 KD (orange) cells. The average of all time points was used.

We also analyzed the effect of Bmal1 and found that the number of rhythmic RNAs and processes was much lower in Bmal1 KD cells. We identified 117, 7, 373, 23, and 7 RNAs whose synthesis rate, pre-RNA level, processing rate, mature RNA level, and degradation rate, respectively, were rhythmic in Bmal1 KD cells (B.H. q < 0.25) (Fig. 2D and Dataset S2). The phase distribution patterns were similar between control and Bmal1 KD cells, except for processing (Fig. 2 C and F). Relative amplitudes (rAMPs) of synthesis were comparable between control and Bmal1 KD cells, while that of processing and degradation were higher in Bmal1 KD cells (Fig. 2G). Nevertheless, there was little or no overlap between control and Bmal1 KD cells in any of the processes or RNA levels, and some RNAs even gained rhythmicity in Bmal1 KD cells (Datasets S1 and S2 and SI Appendix, Tables S1–S4). The number of rhythmic RNAs as well as the degree of overlap between the two conditions increased in all processes when the rhythmicity threshold was loosened, except for degradation in which there were no overlapping rhythmic RNAs regardless of rhythmicity threshold (SI Appendix, Tables S1–S4). We also found 13 RNAs whose mature RNA levels were rhythmic exclusively under the Bmal1 KD condition. These could be false positives or statistical noise without biological relevance, given that our statistical threshold for rhythmicity is not too stringent. It could also be due to a compensatory response for the loss of rhythmicity, or an emergence of rhythmicity usually masked by the Bmal1-driven core clock mechanism in the control condition. Unexpectedly, both processing time (median WT: 12.8 min, KD: 8.7 min, P = 2.2*10⁻¹⁶ two-sided Mann–Whitney U test) and RNA half-life (median WT: 1.99 h, KD: 1.44 h, P = 2.2*10⁻¹⁶ two-sided Mann–Whitney U test) were shorter in Bmal1 KD cells (Fig. 2 H and I), indicating that Bmal1 and/or the core clock machinery regulates RNA processing and degradation either directly or indirectly. Overall, these data indicate that Bmal1 or the core clock machinery drives rhythmic RNA synthesis but not degradation.

Rhythmic RNA Expression Is Primarily Regulated by Rhythmic RNA Synthesis in NIH3T3 Cells.

We next focused on the 685 rhythmic mature RNAs (Fig. 2 A–C) and analyzed which of the three processes contribute most to the rhythmicity of their mature RNA levels. Because the Venn diagram method underestimates the number of rhythmic transcripts overlapping in different datasets (49), and all the current statistical algorithms to compare rhythmicity are applicable only to transcriptomic data and not between RNA levels and process rates [DODR (43), LimoRhyde 1/2 (50, 51), and CompareRhythms (49)], we instead calculated adjusted B.H. q values for rhythmicity of process rates using only the set of 685 rhythmic mature RNAs (50, 51) (Dataset S3).

By using the adjusted B.H. q < 0.25 as a new statistical threshold, we found that 402 RNAs had at least one rhythmic process (Fig. 3A). The most significant contribution was rhythmic RNA synthesis, as 389 (57%) mature RNAs had rhythmic RNA synthesis (Fig. 3 A and B). Both processing and degradation only had a modest contribution (Fig. 3 A and B). The rhythmicity of both mature RNA levels and RNA synthesis were significantly dampened in Bmal1 KD cells, although not completely abolished (Fig. 3 B–D and Dataset S3), supporting the role of Bmal1 as a transcriptional activator. A principal component analysis demonstrated that consecutive samples for RNA synthesis, but not processing nor degradation, were placed in a spiral configuration in control, but not in Bmal1 KD cells (Fig. 3E). These data indicate that RNA synthesis is the primary contributor of RNA rhythms and Bmal1 or the core clock machinery plays a major role in driving rhythmic RNA synthesis.

Fig. 3.

Rhythmic RNA synthesis is the major driver for rhythmic RNA expression in NIH3T3 cells. 685 rhythmic mature RNAs in control cells were analyzed in all datasets. (A) Venn diagram shows the number of rhythmic RNAs in synthesis (green), processing (yellow), and degradation (red) was rhythmic (B.H. q < 0.25). (B) The percentage of RNAs with statistically significant synthesis, processing, or degradation rhythms (B.H. q < 0.25) among the 685 rhythmic mature RNAs in control cells. (C) Distribution of adjusted B.H. q values for the 685 rhythmic RNAs in each dataset. Box plots represent two quartiles ± 1.5 interquartile range from the median (midline). ***; P < 1.0 × 10⁻¹³ (two-tailed Mann–Whiney U test). (D) Phase-sorted heatmap of 685 rhythmic mature RNAs (blue) for their RNA synthesis (green), processing (yellow), and degradation (red) in control (Top) and Bmal1 KD (Bottom) cells. Color represents Z-scores, in which red is high while blue is low. Each line represents one RNA. (E) Principal component analyses of the 685 rhythmic RNAs in each dataset in control (Top) or Bmal1 KD (Bottom) cells. The first two components in each dataset are plotted against each other. Arrows represent the progression of time from T24 to T46. Variation of each principal component in control cells: Mature RNA (PC1—49%, PC2—29%), Synthesis (PC1—40%, PC2—22%), Processing (PC1—28%, PC2—16%), and Degradation (PC1—23%, PC2—18%); or in Bmal1 KD cells: Mature RNA (PC1—39%, PC2—24%), Synthesis (PC1—38%, PC2—26%), Processing (PC1—33%, PC2—16%), and Degradation (PC1—33%, PC2—24%).

The Effect of Interplay between Rhythmic Synthesis and Degradation on RNA Rhythms.

Theoretical studies predicted how RNA dynamics can provide flexibility in both the phase and amplitude of rhythmic RNA expression (5, 6, 52). Since our dataset provides additional information for RNA dynamics that cannot be retrieved from RNA-seq or 4sU-seq data alone, they present an ideal opportunity to test these predictions. To this end, we again focused on the 685 rhythmic mature RNAs (Fig. 3) and used phase and rAMP values, even though some were not statistically rhythmic (B.H. q > 0.25).

The peak phase of rhythmic mature RNA expression was predicted to be anywhere throughout the 24-h cycle when the amplitude of RNA degradation rhythms is greater than that of RNA synthesis rhythms (6). This is largely true, as while 71% of rhythmic RNAs peaked within −3/+3 h of the peak RNA synthesis when the amplitude of the synthesis rate was higher than that of degradation rate, the RNA peak phase was more widely distributed when the amplitude of the degradation rate was higher than that of synthesis rate (P = 5.33 × 10⁻⁸, Watson–Wheeler test) (Fig. 4A). The peak phases of mature RNA were also predicted to be within 6 h of that of RNA synthesis if the RNA synthesis is rhythmic, but degradation is constant (6, 52). We only found one RNA that matches this prediction, in which both RNA synthesis and degradation were rhythmic and had more than 6 h of phase difference between RNA synthesis and mature RNA level (Abcd4: ΔPhase [mat-syn] = −10.08 h). When RNA degradation is rhythmic (Fig. 4B, red), the phase difference between mature RNA levels and RNA synthesis were clustered into two time zones: close to 0 h (cluster 1) or between −6 and −9 h (cluster 2), whereas 65% of rhythmic RNAs had their phase peak within −1 and + 6 h from that of RNA synthesis when RNA synthesis was rhythmic (Fig. 4B, green). Interestingly, all but one in cluster 1 had their RNA synthesis rhythmic (Dataset S3), while none in cluster 2 had rhythmic RNA synthesis. We also found that the effect of RNA half-lives on the phase difference between RNA synthesis and mature RNA levels was minimal (Fig. 4C), presumably because half-lives of all RNAs were less than 10 h in our dataset (Fig. 2I).

Fig. 4.

Effects of rhythmic RNA synthesis and degradation on the phase and amplitude of RNA rhythms. 685 rhythmic mature RNAs in control cells were analyzed in all datasets. (A and B) Phase difference between mature RNA expression and RNA synthesis (ΔPhase [mat-syn] in h) when the relative amplitude of RNA synthesis (rAMP [syn]) is higher than that of RNA degradation (rAMP [deg]) (green) or the relative amplitude of RNA degradation (rAMP [deg]) is higher than that of RNA synthesis (rAMP [syn]) (red) (A) or when RNA synthesis is rhythmic (green), degradation is rhythmic (red), and neither is rhythmic (gray) (B.H. q < 0.25) (B). ***; P < 0.001, *; P < 0.05 (Watson–Wheeler test). (C) Relationship between RNA half-lives (average of all time points) and phase difference between mature RNA expression and RNA synthesis (ΔPhase [mat-syn]). (D) Relationship between the phase difference between mature RNA levels and RNA synthesis (ΔPhase [mat-syn]) and the phase difference between RNA synthesis and degradation (ΔPhase [deg-syn]). (E) Phase difference (h) between mature RNA levels and RNA synthesis (light gray), RNA synthesis and degradation (dark gray), and mature RNA and RNA degradation (black). (F) Peak phases of RNA synthesis (green), mature RNA (blue), and degradation (red). (G and H) Relationship between the relative amplitude of mature RNA expression (rAMP [mat]) and the phase difference between RNA synthesis and degradation (ΔPhase [deg-syn]) (G) or RNA half-lives (H). (I and J) Correlation between the level of mature RNA and relative amplitude of RNA degradation rate (rAMP [deg]) (I) or synthesis rate (rAMP [syn]) (J). All values are log2 transformed. (K and L) Correlation between the relative amplitude of mature RNA (rAMP [mat]) and that of RNA synthesis (K) or degradation rate (L). r: Pearson correlation coefficient. Colored dots represent those that are statistically rhythmic (B.H. q < 0.25) for synthesis (J and K) or degradation (I and L), whereas gray dots are statistically not rhythmic.

We also noticed from the heatmap (Fig. 3D) that RNA degradation appeared to have a rhythmic pattern in control cells, even though the amplitude is low and most RNAs do not reach the statistical threshold for rhythmicity (Fig. 3 B–D). Interestingly, the peak phase of RNA degradation was approximately 6.02 h (median) delayed from that of RNA synthesis while the phase difference between RNA synthesis and mature RNA level was only 0.10 h (median) (Fig. 4 D–F). The antiphasic relationship between synthesis and degradation was predicted to lead to the highest amplitude (6); however, in our study, the highest amplitude of RNA rhythm was achieved when the phase difference between RNA synthesis and degradation was 8.52 h (Dbp: meta2d_phase_[syn] = 22.56 and meta2d_phase_[deg] = 7.08) and a majority of rhythmic RNAs had their relative amplitude around the median (=0.11) (Fig. 4G, gray dotted line). Higher relative amplitude of mature RNA levels was also observed when RNA half-lives were shorter in general as predicted (6), although the RNA half-life of the highest amplitude RNA (Dbp: 2.17 h) was not the shortest among all (Fig. 4H). Because the RNA half-life is significantly shorter and its range is narrower than previously reported (38, 42, 53–55), the effect of half-lives on the phase or amplitude of rhythmic RNA expression is minimal in our dataset.

Contrary to the theoretical prediction, an increased amplitude of RNA degradation did not lead to higher mean RNA levels, even though RNA half-lives are less than 10 h (6). Instead, we found that the amplitude of RNA degradation negatively correlated with the mean RNA levels (Fig. 4I). The amplitude of RNA synthesis did not have any impact on the level of mean RNA levels, either (Fig. 4J). Regardless, the amplitude of RNA rhythm is higher when that of synthesis or degradation is also higher (Fig. 4 K and L), as predicted (6). Most remarkably, most of the core clock genes with rhythmic RNA expression had high amplitude in RNA synthesis and/or degradation (Fig. 4 K and L), presumably critical for sustaining robust rhythmic RNA expression. Overall, our data provided experimental support for some of the theoretical predictions, but not all, suggesting that some parameters are biologically constrained.

Most Core Clock Genes Are Regulated by Multiple Rhythmic Processes.

Some RNAs were regulated by more than one rhythmic process, and 15, 19, and 2 RNAs were regulated by a combination of rhythmic synthesis and degradation, synthesis and processing, or all three processes, respectively (Fig. 3A). Interestingly, most core-clock genes were regulated by more than one rhythmic process (Nr1d1 and Per3: synthesis, processing, and degradation; Cry1 and Per2: synthesis and degradation; Nr1d2: synthesis and processing, Bmal1: synthesis only) (Fig. 5 and Dataset S3), suggesting that multiple layers of regulation are necessary to sustain robust rhythmic RNA expression of core clock RNAs. Rhythmicity of each process was significantly damped under the Bmal1 KD condition, although some were still rhythmic at the statistically significant level (Fig. 5).

Fig. 5.

Multiple rhythmic processes regulate most core clock gene expression. Time-series data of mature RNA levels (blue), RNA synthesis rate (green), processing rate (yellow), and degradation rate (red) in control and Bmal1 KD cells. In the merged plots (Bottom), all values were normalized by dividing values at each time point by the average across all time points for each gene. Asterisks represent adjusted B.H. q < 0.25.

Puzzlingly, some RNAs did not have any rhythmic processes that could account for the rhythmicity of the mature RNA levels (Fig. 3A). These RNAs are generally less robust and have low amplitude (SI Appendix, Fig. S10 A and B). Since the threshold to define the rhythmicity of synthesis and degradation was arbitrary (B.H. q < 0.25), we tested whether all the RNA rhythms could be explained by at least one rhythmic process by setting a different statistical threshold for rhythmicity. However, some rhythmic RNAs have adjusted B.H. q value = 1 for all three processes (Dataset S3). Conversely, several RNAs had rhythmic processes, but not rhythmic mature RNAs (Dataset S1), and in fact, 35 RNAs had two processes being rhythmic (synthesis and processing) and 345 and 280 RNAs had one process being rhythmic (synthesis and processing, respectively) without RNA rhythms (Dataset S1). Their arrhythmic RNA expression cannot be explained solely by their RNA half-lives because the median RNA half-lives of nonrhythmic RNAs are shorter than that of rhythmic RNAs and they are also comparable regardless of whether one or more processes are rhythmic or not (SI Appendix, Fig. S10C). Given that our statistical threshold for rhythmicity is not too stringent (B.H. q < 0.25), some of the rhythms (or lack thereof) may be due to statistical noise or false positives. In addition, the rhythmicity mismatch between RNA levels and kinetic processes may derive from a limitation in statistically comparing the rhythmicity between different types of datasets.

Twelve-Hour RNA Rhythms Are Primarily Regulated by Rhythmic RNA Degradation in NIH3T3 Cells.

In addition to ~24-h RNA rhythms, ultradian RNA rhythms with a period of approximately 12 or 8 h have been observed both in mouse tissues and cultured cells (45, 56–58) and are considered important for metabolism and stress response to dawn/dusk transition (45, 56–58). The regulatory mechanisms driving ultradian gene expression, however, have remained largely elusive. Using our transcriptomic dataset with high temporal resolution, we analyzed RNA dynamics of ultradian RNA expression. By preassigning the period range in MetaCycle, we identified 161 and 293 mature RNAs with a period of approximately 12 (10 < τ < 14) or 8 (6.7 < τ < 9.3) h, respectively (B.H. q < 0.25) (Datasets S4 and S5). This number is most likely underestimated, as our method precludes RNA rhythms superimposed by those with different periods including 24 h (57, 58). These 8- or 12-h RNA rhythms were not as robust as 24-h RNA rhythms (Fig. 6 A and B). The median RNA half-lives of 8-h, but not 12-h, rhythmic RNAs were slightly shorter than that of 24-h rhythmic RNAs (Fig. 6C) although their distribution was not that much different from each other (Fig. 6D). We did not observe any RNAs that are overlapping between 24-h and 12- or 8-h rhythms (Datasets S1, S4, and S5).

Fig. 6.

Comparison between 8-, 12-, and 24-h RNA rhythms. (A) Distribution of B.H. q values, (B) relative amplitude (rAMP), and (C) RNA half-lives (h) of RNAs with 24-h (n = 685), 12-h (n = 161), and 8-h (n = 293) rhythms (B.H. q < 0.25). Box plots represent two quartiles ± 1.5 interquartile range from the median (midline). Tests for differences among groups: (A) Kruskal–Wallis test, P = 6.4 × 10⁻²⁸; (B) Kruskal–Wallis test, P = 4.5 × 10⁻²³; and (C) one-way ANOVA, P = 0.0003. Asterisks indicate results of post hoc pairwise comparisons (A and B: Mann–Whitney U test with Bonferroni correction. C: Tukey HSD test): ***; P < 0.01, N.S.; P > 0.05. (D) Distribution of RNA half-life for rhythmic RNAs with 24-h (median: 2.20 h), 12-h (median: 1.97 h), and 8-h (median: 1.94 h) periods.

We also detected hundreds of RNAs undergoing rhythmic synthesis, processing, and degradation with a period of 12 h in control cells (SI Appendix, Fig. S11A). Interestingly, 12-h rhythms of RNA synthesis and mature RNA levels were dampened in Bmal1 KD cells (SI Appendix, Fig. S11B and Datasets S4 and S6). This is a stark contrast to the mouse liver and mouse embryonic fibroblasts (MEFs), in which Bmal1 or the core circadian clock mechanisms have little or no effect on the 12-h RNA rhythms (58). Xbp1s, a basic region leucine zipper transcription factor, is expressed with the ~12-h rhythmicity and predominantly regulates 12-h gene expression in the mouse liver or MEFs (58). Curiously, however, the expression of Xbp1s was not rhythmic in Bmal1-luc NIH3T3 cells (SI Appendix, Fig. S11C). These data suggest that 12-h RNA rhythms in NIH3T3 cells are driven by a different mechanism and primarily regulated by Bmal1 itself or the circadian core clock machinery.

We next focused on the 161 12-h rhythmic RNAs and assessed which RNA processes contribute to drive 12-h RNA rhythms. By using an adjusted B.H. q value, we found that only 12% (or 19/161 RNAs) of the 12-h RNA rhythms had rhythmic RNA synthesis, whereas 67% (or 108/161 RNA) had rhythmic RNA degradation (Fig. 7 A and B and Dataset S7). Twelve-hour RNA synthesis rhythms were less robust than 24-h synthesis rhythms (Fig. 7C), while the relative amplitudes of RNA processing and degradation rhythms were comparable between 24 and 12 h (Fig. 7D). In addition, 12-h RNA synthesis rhythms were less robust than 12-h degradation rhythms [mean B.H. q value: 0.804 (synthesis) vs. 0.239 (degradation), P < 2.20 × 10⁻¹⁶, Mann–Whitney U test] (Fig. 7C), and the relative amplitude of 12-h degradation rhythms was higher than that of synthesis rhythms [mean relative amplitude: 0.045 (synthesis) vs. 0.061 (degradation), P = 2.26 × 10⁻⁷, two-tailed Student’s t test] (Fig. 7D). Unlike 24-h RNA rhythms, there was no discernable phase relationship between synthesis, mature RNA level, and degradation for 12-h rhythms (Fig. 7 E and F). Interestingly, the relative amplitude of 12-h RNA rhythms correlated with that of 12-h RNA degradation (Fig. 7H), but not with that of 12-h RNA synthesis (Fig. 7G). These data collectively suggest that in NIH3T3 cells, 12-h ultradian RNA rhythms are primarily driven by rhythmic degradation and regulated by Bmal1 or the circadian core clock machinery.

Fig. 7.

Twelve-hour RNA rhythms are primarily driven by rhythmic degradation. 161 rhythmic mature RNAs with a period of 12 h in control cells were analyzed. (A) Venn diagram showing the number of RNAs whose RNA synthesis (green), processing (yellow), and degradation (red) was rhythmic (B.H. q < 0.25). (B) The percentage of RNAs with statistically significant synthesis, processing, or degradation rhythms (B.H. q < 0.25) among the 161 rhythmic mature RNAs with a period of 12 h in control (blue) and Bmal1 KD (orange) cells. (C) Distribution of adjusted B.H. q values for the 685 24-h (Left) or 161 12-h rhythmic RNAs (Right) in control cells. (D) Distribution of relative amplitude (rAMP) for the 685 24-h or 161 12-h rhythmic RNAs in control cells. Box plots represent two quartiles ± 1.5 interquartile range from the median (midline). ***; P-value < 0.001, N.S.; P > 0.05 (C: two-tailed Mann–Whitney U test, D: two-tailed Student’s t test). (E) Peak phases of RNA synthesis rate (green), mature RNA levels (blue), and degradation rate (red). (F) Phase difference (h) between mature RNA levels and RNA synthesis (light gray), RNA synthesis and degradation (dark gray), and mature RNA and RNA degradation (black). (G and H) Correlation between the relative amplitude of mature RNA and RNA synthesis rate (G) or degradation rate (H). All values are log2 transformed. r: Pearson correlation coefficient. Colored dots represent those that are statistically rhythmic (B.H. q < 0.25), whereas gray dots are statistically not rhythmic.

Discussion

In this study, we used a metabolic labeling approach and monitored RNA dynamics in NIH3T3 cells around the circadian cycle. We found that a) rhythmic RNA synthesis is the major contributor in driving ~24-h rhythmic RNA expression (Figs. 2 and 3), whereas rhythmic RNA degradation is important for 12-h RNA rhythms (Fig. 7), b) these rhythms were largely abolished in Bmal1 KD cells (Figs. 2 and 3 and SI Appendix, Fig. S11), and c) rhythmic expression of core clock genes is higher in amplitude and regulated by multiple rhythmic processes (Figs. 4 K and L, and 5). One of the limitations of our study is that we only had one biological replica (each taken from a single distinct cell culture dish) and samples were taken over one circadian cycle, reducing the statistical power of our analysis to some extent. We also used in vitro cell culture system, as metabolic labeling methodology is not currently applicable in vivo, and detected a lower number of rhythmic RNAs compared to mouse tissues (9, 45). Nevertheless, our data from NIH3T3 cells closely align with other circadian transcriptome and 4sU-seq studies (42, 53) and provide critical RNA dynamics information to fully understand how the core circadian clock machinery shapes RNA rhythms.

Previous studies have assessed the effect of RNA synthesis (i.e., transcription) in regulating rhythmic RNA expression. In some studies, the level of pre-RNA (i.e., intronic levels in RNA-seq) was used as a proxy for RNA synthesis (5, 13–15, 59), while newer studies inferred RNA synthesis from directly measuring the level of nascent or actively transcribed RNAs (13, 60, 61). We therefore compared the RNA synthesis rate with the levels of newly synthesized RNAs in our dataset. As INSPEcT calculates RNA synthesis rate from the exonic levels in 4sU-seq (39), these two variables were linearly correlated (r = 1) (SI Appendix, Fig. S12A). There was also a very strong correlation between the RNA synthesis rate and the levels of newly synthesized RNAs (r = 0.964 with 4sU-seq [RPKM_gene] and r = 0.837 with 4sU-seq [TPM_gene]) (SI Appendix, Fig. S12 B and C). In support of this, the phase of RNA synthesis and that of the newly synthesized RNA levels aligned well with each other (SI Appendix, Fig. S12G) and the variation of their phase difference was small (ΔPhase_{(Syn-4sUseq[RPKMgene]} = 0.38 ± 1.54 h, ΔPhase_{(Syn-4sUseq[TPMgene])} = 0.03 ± 3.07 h) (SI Appendix, Fig. S12H). The correlation was slightly weaker between the RNA synthesis rate and the pre-RNA level (i.e., intronic levels from RNA-seq) (r = 0.717) (SI Appendix, Fig. S12D), and the phase difference was more variable (ΔPhase_{(Syn-RNAseq[RPKMintron])} = −0.17 ± 4.57 h) (SI Appendix, Fig. S12 G and H). The pre-RNA level also correlated strongly to that of newly synthesized RNAs (r = 0.765 with 4sU-seq [RPKM_gene] and r = 0.810 with 4sU-seq [TPM_gene]), but at a lesser degree compared to RNA synthesis rate (SI Appendix, Fig. S12 E and F). The difference in peak phase was also more variable between the levels of pre-RNA and newly synthesized RNAs (ΔPhase_{(PreRNA-4sUseq[RPKMgene]} = 0.17 ± 4.41 h and ΔPhase_{(PreRNA-4sUseq[TPMgene])} = 0.12 ± 4.27 h) (SI Appendix, Fig. S12 G and H). In fact, the correlation between pre-RNA and mature RNA was stronger (r > 0.96) and their phase difference was less than that between pre-RNA and RNA synthesis (ΔPhase_{(PreRNA-MatureRNA)} = −0.03 ± 4.01 h) (SI Appendix, Figs. S2 and S12 G and H). These data collectively demonstrate that the RNA synthesis rate correlates more strongly to the level of newly synthesized RNAs (i.e., 4sU-seq [RPKM_gene] and 4sU-seq [TPM_gene]) compared to the pre-RNA levels.

One of the surprising results was that rhythmic RNA degradation only had a modest contribution to 24-h RNA rhythms at least in NIH3T3 cells (Figs. 2 and 3). Even though several mechanisms, such as miRNA, RNA-binding proteins, and poly(A) tail length, have been implicated in regulating rhythmic RNA degradation (3), our data indicate that these mechanisms don’t have a strong impact, or are not rhythmically regulated at least in NIH3T3 cells. Interestingly, however, RNA degradation appeared to be weakly rhythmic (Fig. 3D, Top Right), and this rhythmicity was abolished in Bmal1 KD cells (Fig. 3D, Bottom Right). How can rhythmic degradation, albeit weak, be controlled by Bmal1 or the core clock machinery? Because the amplitude of RNA degradation rhythms is generally much lower than those of RNA synthesis (Fig. 3D) and we don’t see the global effect of rhythmic degradation (Fig. 2), we propose that the weak RNA degradation rhythm is regulated by a transcript-specific manner, and potentially coupled with rhythmic RNA synthesis. One possible mechanism is transcript buffering, a cellular mechanism to maintain the homeostasis of transcript levels. The rates of RNA synthesis and degradation can be kinetically coupled by establishing a feedback loop between RNA synthesis in the nucleus to degradation of their corresponding RNAs in the cytoplasm (62–69). This mechanism has been well established in yeast and recently demonstrated in human cells (70, 71); however, all the studies were done in a steady-state condition and its regulatory mechanisms, particularly under a dynamic condition, remain elusive (reviewed in ref. 63). Another possibility is the rhythms in poly(A) tail length, as has also been suggested to be coupled with rhythmic RNA synthesis (19). Indeed, several rhythmic mature RNAs detected in our dataset exhibited rhythmic poly(A) tail length in the mouse liver (19). It will be of future interest to uncouple RNA synthesis from degradation and test whether degradation is still rhythmic even in the absence of rhythmic RNA synthesis.

Another surprising finding was that both 24- and 12-h rhythms were regulated by Bmal1 or the core clock mechanism in Bmal1-luc NIH3T3 cells (Figs. 2 and 7 and SI Appendix, Fig. S11). Interestingly, our motif enrichment analysis demonstrated that DNA recognition motifs for the core clock proteins are enriched among RNA with rhythmic synthesis for both 24- and 12-h cycles (SI Appendix, Fig. S9). These data indicate that Bmal1 or the core clock mechanism can drive not only 24-h but also 12-h RNA rhythms; however, it still remains unclear why some RNAs have 24-h rhythms while others have 12-h rhythms. It is plausible that additional motifs only enriched among RNAs with 12-h rhythm are responsible for driving either 12 or 24 h with antiphasic RNAs rhythms, although this possibility needs to be experimentally tested in the future. It is also puzzling why the effect of rhythmic degradation is stronger for 12-h rhythms compared to 24-h rhythms. Because we did not observe any correlation between the phase of synthesis and degradation for 12-h rhythms (Fig. 7E), unlike the 24-h rhythms (Fig. 4F), regulatory mechanisms of 12-h degradation rhythms are most likely independent from their synthesis rhythms. Motif enrichment analysis did not reveal any specific motifs that could be responsible for driving 12-h degradation rhythms (SI Appendix, Fig. S9); however, this could be due to the low number of RNAs with 12-h rhythms (Fig. 7) or RNA degradation is regulated independent of RNA sequences or outside of the 3′ untranslated region.

Our study also highlighted the robust control of rhythmic core clock RNA expression (Figs. 4 and 5). Given that core clock genes are rhythmic in many more tissues compared to clock-controlled genes (9, 10), it is reasonable that their RNA rhythmicity is regulated by multiple steps to ensure that cell-autonomous rhythmicity with high amplitude can be sustained even in the presence of environmental noise. What makes their amplitude high? Because circadian cistrome datasets for various core clock proteins as well as different forms of RNA polymerase II are already available in the mouse liver (14, 59, 72), analyses of their dynamics will likely provide an important clue as to what drives the high amplitude of core clock gene expression.

Overall, our study further elucidates the extent by which transcriptional and posttranscriptional regulatory steps drive rhythmic RNA expression. It would be of great interest to expand our study and explore the dynamics of rhythmic protein expression in the future to fully understand how the core circadian clock machinery regulates gene expression rhythms and ultimately rhythmic downstream processes.

Materials and Methods

Cell Culture and Lentiviral Transduction.

NIH3T3-derived Bmal1-luc (31) and HEK293/T17 cells were cultured with Dulbecco’s Modified Eagle Medium (DMEM) (Gibco: Cat #11965118) supplemented with 10% fetal bovine serum (Corning, Cat#MT35010CV) at 37 °C with 5.0% CO₂. We followed a standard procedure to prepare lentiviruses containing short-hairpin (sh) RNAs (69). Briefly, viral particles were prepared in HEK293/T17 cells by transfecting pLP-1, pLP-2, and pLP-VSVG plasmids (ThermoFisher) along with pLL3.7GW vector containing shRNAs for control (target sequence: CAACAAGATGAAGAAGAG CACCA) or Bmal1 (target sequence: GGAAGGATCAAGAATGCAA) using polyethylenimine (Polysciences Inc., #23966) (70). Forty-eight hours after the DNA transfection, culture medium containing viral particles was collected and filtered by 0.22 µm PDVF membrane (Millipore, Cat#SLGV013SL). The media were further ultracentrifuged at 75,000g for 2 h at 10 °C before being reconstituted to the same volume of fresh cell culture media. Bmal1-luc NIH3T3 cells were then transduced with polybrene (Millipore, Cat#TR-1003-G), and media containing viruses were replaced with fresh media 24 h after transduction. Transfection and transduction efficiency was visually inspected by EGFP signals that is encoded in the pLL3.7GW vector.

4sU Labeling and RNA Pull-Down.

Seventy-two hours after the viral transduction, cells were synchronized with DMEM containing 50% horse serum for 2 h. We chose to use 50% horse serum shock because our pilot study demonstrated that 50% horse serum shock induced more robust rhythmicity in Bmal1 and Per2 RNA expression compared to 100 nM dexamethasone in Bmal1-luc NIH3T3 cells (SI Appendix, Fig. S13; RNA levels measured by RT-qPCR as described in ref. 73). Twenty-four hours after the onset of 50% horse serum shock, Bmal1-luc NIH3T3 cells were labeled with 400 μM of 4-thiouridine (4sU) (Cayman Chemical, Cat#16373) for 2 h. Subsequently, RNAs were extracted with TRIzol (Invitrogen, Cat# 15596018) according to the manufacturer’s instructions and then treated with DNase I (ThermoFisher, AM2239) at 37 °C for 1 h. After repurifying RNAs using TRIzol, 1 µg of yeast spike-in RNA (gift from Dr. Silke Hauf: Virginia Tech) labeled with 4-thiouracil (Neta Scientific, CMX-21484-1G) for 1 h and 300 fmol 4sU-labeled DNA/RNA hybrid spike-in control (ATTTAGGTGACACTATAGGATCCTCTAGAGTCGACCTTCTCCCTATAGTGAGTC G​T​A​TTAGCA[4sU]CAG) were added to each RNA sample to normalize the RNA pull-down efficiency (see below). 4sU-labeled RNA was then biotinylated with 1 mg/mL EZ-link HPDP-Biotin (ThermoFisher, cat: 21341) for 90 min at room temperature as previously described (34, 74, 75). Thirty ug RNA (approximately 50% for shCtrl and 60-70% for shBmal1 samples) was subsequently subjected to streptavidin pull-down, while the remainder was set aside as “total RNA.” Streptavidin magnetic beads (ThermoFisher cat: 65001) were prewashed with 0.1 M NaOH, 0.1 M NaCl, and then washing buffer (10 mM Tris-HCl pH 7.4, 10 mM EDTA, and 1 M NaCl) to remove RNase. Beads were then incubated with RNA samples in washing buffer for 45 min at room temperature with rotation. Beads were washed twice with washing buffer for 5 min, and all the flow through (“preexisting”) fractions from each wash were separated from the bead-bound fraction using a DynaMag-2 magnetic rack (Life Technologies, Cat#12321D) and pooled together. After the second wash, RNAs bound to the beads were recovered first by incubating the beads with 0.1 M DTT for 10 min at room temperature. The remaining beads were further incubated with prewarmed 0.1 M DTT at 65 °C for 10 min, and all the RNAs were collected in a single tube (“newly synthesized”). RNAs from both preexisting and newly synthesized fractions were precipitated with ethanol with 1 M NaCl and 20 µg glycogen (Ambion; Cat#9510) and reconstituted in diethyl pyrocarbonate (DEPC)-treated deionized and distilled water (DDW).

RNA Pull-Down Efficiency Calculation.

We calculated the pull-down efficiency using 4sU-labeled DNA/RNA hybrid spike-in control, as was previously described (34). Briefly, RNAs from both pull-down and flow-through fractions were purified by TRIzol and responded in the same volume of DEPC-treated DDW. An equal volume of RNA solution from both fractions was used to generate cDNA separately using the High-Capacity cDNA Reverse Transcriptase Kit (Applied Biosystems, REF4368813) according to the manufacturer’s instruction. We quantified the relative amount of cDNA in each fraction using QuantStudio6 (Life Tech) with PowerSYBR Green PCR Master Mix (Applied Biosystems, REF4367659). We determined the pull-down efficiency by calculating the ratio of the level of spike-in control between the pull-down fraction and the combination of pull-down and flow-through fractions).

Transcriptomic Analyses.

Approximately 100 to 200 ng of total and newly synthesized RNAs from each time point was subjected to transcriptomic analysis (Azenta Biosciences). RNA quality of each sample was verified by the RNA integrity number (RIN) > 9.5 measured by Tapestation 4200 (Agilent). Additional DNase I treatment was performed before proceeding to the library preparation per recommendation from Azenta. Paired-end, strand-specific, and rRNA-depleted sequencing library was then prepared using the QIAGEN FastSelect rRNA HMR Kit (Qiagen, cat#334378) and NEB Directional Ultra II RNA library prep kit (NEB, E7760) by following the manufacturer’s recommendations (NEB, cat# E7765). Qubit 2.0 Fluorometer (ThermoFisher Scientific) and quantitative PCR (KAPA Biosystems) method were used to quantify each library before multiplexing, and the distribution of reads lengths was measured by the Agilent Tapestation 4200 (Agilent Technologies). Libraries were then sequenced by Illumina Hi-seq.

Bioinformatical Analyses.

Azenta used their in-house quality control bioinformatical pipeline to return FASTQ files with a quality score of >Q30 on average. We further used TrimGalore (76) to remove adapter sequences and remove reads of low quality (<Q20), which we found to be about 0.2% per file. Reads were mapped with STARv2.7.7a (77) using options --outFilterScoreMinOverLread 0.3 and --sjdbOverhang 100. Reads mapped to the NR003278.3 (Mus musculus 18S rRNA) and NR003279.1 (Mus musculus 28S rRNA) were eliminated first. The remaining reads were then mapped to the mouse mm10 genome (GENECODE: GRCm38.p6.genome.fa), and read counts against gene (transcript per million or TPM), exon, and intron were independently quantified with HOMER (V4.11.1) (78) using gencode.vM25.annotation.gtf. using option condenseGenes. Transcripts with TPM < 0.25 (as an average of all time points and conditions) were removed for all downstream analysis. Reads per kilobase per million mapped (RPKM) was also calculated using HOMER (v4.11) to quantify the reads mapped to exon or intron separately with the –exon or –intron option for both RNA-seq and 4sU-seq (Datasets S8–S15). We supplied these four input datafiles to INSPEcT (v1.22.0) (37) using Rstudio (v4.1.0). INSPEcT first uses DESeq2 (v1.32.0) to calculate variances (which is zero in our case as our dataset only has one biological replica) and arranges samples in a matrix to be analyzed. The DESeq2 output was then applied to INSPEcT using the ratesFirstGuess options: 2-h labeling time, degDuringPulse = FALSE. This generates a total of four output files (synthesis, pre-RNA, processing, and degradation). We supplied the complete script to run the INSPEcT analysis as an R Markdown document (SI Appendix). INSPEcT eliminates single-exon genes as it requires both exonic and intronic levels to return these output files, as well as genes for which the intron reads are more abundant than the exon reads.

Other Computational Analyses.

Read count distribution.

Read counts for the intron and exon were from HOMER (v4.11) after mapping raw reads to the mouse transcripts (GRCm38.p6, gencode.vM25.annotation.gtf) using STAR (v2.7.7a). Unmapped reads were then mapped to the mouse genome (GRCm38.p6, gencode.vM25.annotation.gtf) using STAR (v2.7.7a), and the reads uniquely mapping to the genome but not the transcript sequences were determined to be intergenic. The average of all time points within each dataset was then calculated.

Rhythmicity analysis.

We used MetaCycle (v1.2.0, CRAN R package) (44) using cycMethod = c(“ARS”, “JTK”, “LS”) and used meta2d B.H. q values to define rhythmicity. All the rhythmicity statistics and parameters shown in this study were calculated by MetaCycle. To detect rhythmicity with different periods, we set the period range to 20–28, 10–14, and 6.7–9.3, for 24-h, 12-h, and 8-h periods, respectively. These ranges reflect a 16.6% variation of the period. In our analysis, time points were set to 24–46 based on our experimental design (Fig. 1A). We also included an R markdown document that describes the codes to run MetaCycle (SI Appendix).

Sequence data visualization (GBiB).

The bam files generated by STAR (see above) were sorted and indexed by Samtools (version 1.11) (79) and converted into bigWig files by DeepTools (version 3.5.0) (80) with the option -filterRNAstrand in order to visualize strand specifically. The files were uploaded to UCSC Genome Browser using the Genome Browser in a Box (GBiB) (81) to visualize mapped reads.

Phase-sorted heatmaps.

Mature RNA expression data were first log₂ transformed, and Z-scores for each RNA profile were manually calculated. The data were then sorted by meta2d_phase, and heatmaps were generated by gplots heatmap.2 function (82) using Rstudio (v4.1.0) with no additional scaling options.

PCA analysis.

Principle component analysis was performed using the prcomp function in the R (1.3.2) (83) package “stats” (v4.1.0) with scaling = TRUE option.

Motif enrichment analysis.

Motif enrichment analysis was performed using Simple Enrichment Analysis (SEA) (v5.5.4) with HOCOMOCO Mouse v11 (DNA) or Ray2013 Mus musculus (RNA) databases (84). Both input and background sequences were retrieved from the UCSC Genome Browser annotation track database using gencode.vM25.annotation.gtf. We used the promoter DNA sequence defined as the 1,000 bp upstream of the transcription start site, the entire pre-RNA sequences including both exons and introns, and the 3′ untranslated region RNA sequences, to identify motifs enriched among those with rhythmic synthesis, processing, and degradation, respectively. For 24-h rhythms, we used 653 and 336 RNAs with rhythmic synthesis and processing (B.H. q < 0.25), respectively, and 685 RNA with rhythmic mature RNAs (Dataset S3) for degradation as only six RNAs were rhythmically degraded. For 12-h rhythms, we used 642 and 469 RNAs with rhythmic synthesis and processing (B.H. q <0.25), respectively, and 161 RNA with rhythmic mature RNAs for degradation as only five RNAs were rhythmically degraded. As background, we used all 11,313 genes that met our analytical criteria (Datasets S1 and S2) for synthesis and degradation, whereas shuffled RNA sequences were used for processing. We included all the isoforms included in gencode.vM25.annotation.gtf for a gene for the analysis. SEA reports motifs with statistical significance better than the given enrichment E-value (in our case E-values < 10).

Statistical Analyses.

The statistical tests were performed using Rstudio (v4.1.0). The default R stats package v4.1.0 was used for the Pearson correlation test, Student’s t test, Mann–Whitney U test, Kruskal–Wallis test, one-way ANOVA, and Tukey’s Honestly Significant Different (HSD) test, while the “car” package (v3.1) was used for the Levene test. The “circular” package was used for the Watson–Wheeler test (v0.4-95) with Rstudio (v.4.2.1).

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Dataset S02 (XLSX)

Dataset S03 (XLSX)

Dataset S04 (XLSX)

Dataset S05 (XLSX)

Dataset S06 (XLSX)

Dataset S07 (XLSX)

Dataset S08 (CSV)

Dataset S09 (CSV)

Dataset S10 (CSV)

Dataset S11 (CSV)

Dataset S12 (CSV)

Dataset S13 (CSV)

Dataset S14 (CSV)

Dataset S15 (CSV)

Data, Materials, and Software Availability

The data reported in this paper have been deposited in the Gene Expression Omnibus (GEO) database, www.ncbi.nlm.nih.gov/geo (accession no. GSE253826) (85).