Quantitative circadian variation across 191 Swedish Arabidopsis accessions
We used delayed fluorescence imaging to characterize circadian rhythms in groups of 14 day old seedlings entrained in 12h L 12h D cycles at 22°C and assayed under free-running conditions of constant light (L:L) at 22°C. Accession means and standard errors (SE) adjusted for experimental effects were obtained using linear mixed models for period, amplitude and RAE and a circular regression model for phase (see Extended Methods). We found a 4.42h difference between the mean period of the fastest (21.28h, SE=0.329) and slowest (25.70h, SE=0.355) accessions tested (Figure 1A). The mean phase of peak DF intensity for each accession occurred over the dark half of the cycle (12-24/0h) with a huge difference (10.32h) from the earliest peaking (12.21h, SE=0.58) to the latest (22.54h, SE=0.20) (Figure 1B). This variation in period and phase is consistent with data previously reported from a global collection of Arabidopsis accessions(Michael et al., 2003). We also measured the robustness of these rhythms by looking at their RAE values. The most rhythmic (lowest RAE) mean was 0.31 (SE=0.023) and the least rhythmic was 0.44 (SE=0.024) representing approximately 20% of the possible range of this trait in this study (Figure 1C). Accession means for circadian phenotypes can be viewed in Supplementary File 1.
To assess the effect of genotype on each trait we performed a likelihood ratio test against a model which omitted genotype as a variance component. Including genotype in the mixed model had highly significant effects on period (χ2 (1 df)=455.57,p< 0.0001) and RAE (χ2 (1)= 65.97,p< 0.0001). Including genotype in the circular regression model also had a strongly significant effect on phase (χ2 (191)=407, p< 0.0001). Amplitude was analysed as a log10 transformation (see Extended Methods) and there was no significant effect of genotype on Log10Amplitude (χ2 (1)= 0.19, p= 0.6). With this in mind, we dropped Amplitude as a trait of interest for the temperature and mutant screening experiments. REML output tables (Supplementary Tables 1-3), model checking graphs (Supplementary Figures 1-3) and log-likelihood results (Supplementary Tables 4-7) are available to view.
We observed a strong correlation between the period, RAE and phase of each accession (Supplementary Figure 4), especially between period and phase (Adjusted R2 =0.43, p< 0.0001). We are unaware of any previously published correlation between natural variation in period and RAE. In this study, we found a highly significant negative relationship with longer periods having lower RAE scores (Adjusted R2 =0.1, p< 0.0001).
We chose 10 accessions from the tails of the distributions of mean period, phase and RAE to create “phenotypic tail” groups representing the extremes of each trait. Some accessions represented two or more tail groups and could be split into two master groups of: long period, dusk phased, low RAE accessions and short period, dawn phased, high RAE accessions as explained in Figure 1D. These phenotypic tail groups were used for the temperature experiments described later in this paper (accessions in each tail are listed in Supplementary Tables 8-10).
Our next question was whether other quantified traits co-varied with our circadian data. We used several previously published datasets to examine possible phenotypic relationships between our traits and flowering time (Y. Li et al., 2010; Sasaki et al., 2015), seed dormancy (Kerdaffrec et al., 2016) and freezing tolerance (Horton et al., 2016) (Supplementary Table 11). Only flowering time data from Li et al. (2010) showed a moderate correlation with period (p<0.01, R2 =0.28) and phase (p<0.01, R2 =0.26). This flowering time data is based on flowering times recorded in conditions of variable day-length reflecting annual conditions of a natural Swedish or Spanish climate (Y. Li et al., 2010). However, no significant correlations were found between flowering times recorded under long-days under either 10°C or 16°C by Horton et al. (2016), indicating that the relationship is highly dependent on the exact experimental conditions used.
We then investigated whether circadian traits varied according to the original location of the accessions. Period was found to be significantly correlated with longitude and latitude (p< 0.001), however a considerable amount of variation remained unexplained (Adjusted R2 <0.1) (Supplementary Figure 5). Despite this weak overall association with latitude, we observed an obvious segregation of accessions from the phenotypic tails of the period distribution, with the shortest period accessions in Northern and mid-latitude regions and the longest period accessions in the south (Figure 2A). No significant correlation with altitude was observed for these accessions.
Following on from this, we asked the question whether population sub-structure could better explain the distribution of period phenotypes. We re-classified the population into three groups split across the axes of the first two principal components (which together explain over 15% of the total genetic variation) (Figure 2D and E). PC.B broadly contained Northern accessions (N=54, red points in Figure 2D). The Southern accessions split into two sub-groups existing within the same geographic region; PC.A (N=111) and PC.C (N=26). We found that period is significantly longer in group PC.C (Figure 2F) (F( 2,188)=8.39, TukeyHSD for PC.A and PC.C p< 0.01, One-way ANOVA). PC.A had mean periods similar to those obtained from Northern accessions (PC.B, Figure 2F). PC.C accessions also have significantly earlier phase peaks than those in PC.A (One-way ANOVA, F(2,188)=6.9149, TukeyHSD for PC.A and PC.C p< 0.001).
We removed the effects of PC groups from period length variation and found that the effect of latitude on the remaining period variation was much less significant (REML Model: Period ~ (1 | PC.blocks) + Latitude, applying Satterthwaite’s method. F(1,7.23) = 8.87, p >0.01). This indicates that period variation with latitude is due to the different locations of sub-populations, rather than due to adaption following a latitudinal cline. Accessions in the long period tail group described in Figure 1A belonged to both PC.A (N=6) and PC.C (N=4), suggesting that long-period traits may have been preserved in both sub-populations in this area.