Natural variation in autumn expression is the major adaptive determinant distinguishing Arabidopsis FLC haplotypes

In Arabidopsis thaliana, winter is registered during vernalization through the temperature-dependent repression and epigenetic silencing of floral repressor FLOWERING LOCUS C (FLC). Natural Arabidopsis accessions show considerable variation in vernalization. However, which aspect of the FLC repression mechanism is most important for adaptation to different environments is unclear. By analysing FLC dynamics in natural variants and mutants throughout winter in three field sites, we find that autumnal FLC expression, rather than epigenetic silencing, is the major variable conferred by the distinct Arabidopsis FLChaplotypes. This variation influences flowering responses of Arabidopsis accessions resulting in an interplay between promotion and delay of flowering in different climates to balance survival and, through a post-vernalization effect, reproductive output. These data reveal how expression variation through non-coding cis variation at FLC has enabled Arabidopsis accessions to adapt to different climatic conditions and year-on-year fluctuations.


Introduction
Developmental transitions in plants are aligned with specific seasons to maximise reproductive success. A major seasonal cue used to time the transition to flowering is temperature (Hancock et al., 2011;Andrés and Coupland, 2012;Manzano-Piedras et al., 2014). Many plants overwinter in a vegetative form and flower in spring. This alignment of the floral transition with spring is achieved through a process called vernalization, in which flowering is accelerated by previous exposure to prolonged winter cold. In Arabidopsis thaliana and its relatives, FLOWERING LOCUS C (FLC), a MADS-box transcription factor, delays the transition to flowering in warm temperatures until its expression is downregulated by exposure to cold (Michaels and Amasino, 1999;Sheldon et al., 1999). On return to warm, spring-like conditions, FLC remains epigenetically silenced, permitting expression of FT and floral meristem identity genes that trigger the floral transition (Helliwell et al., 2006;Searle et al., 2006).
FLC expression and memory is influenced by natural non-coding cis polymorphism at the locus (Lempe et al., 2005;Shindo et al., 2006;Coustham et al., 2012;Li et al., 2015;Qüesta et al., 2020). Different Arabidopsis accessions need different lengths of cold to epigenetically silence FLC, and this requirement is determined by non-coding SNPs at FLC (Shindo et al., 2006;Coustham et al., 2012). Indeed, within the Arabideae tribe, cis regulation of FLC orthologues contributes to perennial-vs.-annual flowering differences (Kiefer et al., 2017). Within the worldwide Arabidopsis thaliana population, non-coding variation defines a small number of FLC haplotypes, which confer quantitatively different flowering time responses to different lengths of vernalization in constant cold in the laboratory (Li et al., 2014). These haplotypes appear to have been maintained in the A. thaliana population due to their contributions to life history diversity (Sánchez-Bermejo et al., 2012;Li et al., 2014;Méndez-Vigo et al., 2016), and they show different geographic distributions, suggesting that they are adapted to different climates (Méndez-Vigo et al., 2011;Li et al., 2014).
Recent work has shown the importance of analysing flowering in natural field conditions ( Figure 1A), 'in natura' (Kudoh, 2016). Such studies have highlighted how the influence of genes of major effect, including FLC, changes depending on the field environment (Wilczek et al., 2009;Fournier-Level et al., 2011;Duncan et al., 2015;Å gren et al., 2017;Taylor et al., 2019). In particular, the investigation of RNA levels in natura has identified genetic mechanisms in the field that were previously unknown from laboratory observations Song et al., 2018;Nagano et al., 2019). For example, field work in Sweden and the UK with the vernalization reference genotype Col FRI SF2 and mutants in VIN3 (vin3-4) has shown that initial VIN3independent FLC downregulation requires cold temperatures, whereas VIN3-mediated epigenetic silencing requires the absence of daily temperatures above 15˚C . Autumn conditions greatly influence when epigenetic silencing initiates, sometimes leading to temporal separation of the two phases, which in laboratory constant-condition experiments and some natural environments occur concurrently .
Here, we exploit field studies in three climatically distinct locations, over two years, to define the roles of FLC cis polymorphism and known genetic regulators of FLC in VIN3-dependent and -independent phases of the vernalization response under natural conditions ( Figure 2A). We hypothesised that accessions bearing different FLC haplotypes would show different flowering time responses to winter environments, and that this would be due, at least in part, to variation in the downregulation and silencing of FLC itself, arising from the cis polymorphism at FLC. Moreover, we expected variation in the FLC epigenetic silencing mechanism would play a particularly important role, as this mechanism has previously been shown to vary in natural accessions (Coustham et al., 2012;Wollenberg and Amasino, 2012;Duncan et al., 2015).
Our results demonstrate how in the field the major FLC haplotypes, differing only through noncoding variation, have different starting FLC levels and rates of response to autumn cold, but show more similar epigenetic silencing rates in winter. This results in effective repression of FLC by midwinter and post-winter flowering in most years, across haplotypes and climates. Our experiments also reveal effects of the FLC haplotypes on prevention of precocious flowering in warm years, and on reproductive success after winter via effects on branching and silique number. By studying gene expression across years and climates, we have been able to dissect how non-coding cis variation at FLC modulates flowering time and fitness in response to different natural fluctuating environments.  Li et al., 2014. The lab genotype Col FRI was also used in this study as a vernalization-requiring reference. (B-D) Temperature profiles experienced by plants at the three field sites, North Sweden -Ramsta (B), South Sweden -Ullstorp (C) and Norwich, UK (D) (Source data 1, as from Hepworth et al., 2018 andAntoniou-Kourounioti et al., 2018). Flowers above temperature profile indicate the median time of bolting of each of the natural accessions and of Col FRI (legend at bottom left corner). Black dots below temperature profile indicate the timepoints when plant material was collected for expression analysis. Black arrows below temperature profiles indicate time of transfer to greenhouse with long-day, warm conditions to assess degree of vernalization based on bolting time. The online version of this article includes the following figure supplement(s) for figure 1:     and equivalent times in VIN3 and temperature profiles (as in Figure 1D). Expression data were normalised to the control sample for 2014-5 (see Materials and methods). N = 6 except where samples Figure 2 continued on next page

Results
Field experiments Li et al., 2014 identified 20 FLC haplotypes across a worldwide panel of Arabidopsis accessions defined predominantly through non-coding SNPs. These haplotypes conferred different flowering time responses to vernalization at constant 5˚C, with the accessions carrying them clustered broadly into 'Rapid Vernalizing' (RV) and 'Slow Vernalizing' (SV) types. To investigate their function in field conditions, we selected accessions to represent each of the five most populous haplotypes, representing more than 60% of tested accessions, as well as a further accession, Lö v-1, for which there is evidence of local adaptation to the climate in the region of our North Sweden field site (Duncan et al., 2015;Qüesta et al., 2020). To compare these haplotypes in a common genetic background, we exploited extant and developed new near-isogenic lines (NILs) in which the FLC haplotype from each accession had been repeatedly backcrossed to Col FRI SF2 ('Col FRI'; Duncan et al., 2015;Li et al., 2015). We confirmed that these selected accessions and NILs show variation in their rate of response to time in constant cold in the laboratory and also found that they varied in branch response to vernalization, an effect mediated by FLC ( Figure 1-figure supplements 1-2, Huang et al., 2013;Lazaro et al., 2018). These genotypes were tested across two years and three field sites, with the exception of three NIL lines, which were generated during the experiments ( Figures 1A and 2B-J). The three field sites in Norwich, UK, in Ullstorp, Sweden ('South Sweden') and in Ramsta, Sweden ('North Sweden') were chosen to represent climates with different winters. Norwich has a temperate oceanic climate and a mild winter, Ullstorp has a warm-summer continental climate with a winter in which temperatures are often below freezing, and Ramsta is subarctic, usually with snow cover during winter (hence the 0˚C flatline in temperatures in Figure 1B; Beck et al., 2018). The Swedish sites are also close to the source sites of several of the tested accessions; Vå r2-6 and Ull2-5 in Skå ne, near or at Ullstorp; Lö v-1 near Ramsta ( Figure 1A). In the second year of experimentation, we also included the vin3-1 mutant in the Col FRI background. The experiments ran from August/September 2014 until spring 2015 and again from August/September 2016 to the spring of 2017, with sowing times adjusted to each site (earlier in the colder climates). In the first year, two plantings were performed in North Sweden, two weeks apart. The temperatures that the plants experienced are shown in Figure 1B-D. We measured the levels of spliced and unspliced FLC, and mRNA levels of VIN3, to follow the progress of vernalization in the field. The transition to flowering (bolting) was assessed both in the field and by transfers to warm inductive conditions. were lost to death or degradation (see Materials and methods and Source data 2). Error bars show standard error of the mean (s.e.m). The initial measurement in the field (Starting levels), the rate of downregulation before induction of VIN3 expression (VIN3-independent, estimated from the slope of the fitted line) and the rate of downregulation after VIN3 induction (VIN3-dependent) are the three features that were analysed and compared for each genotype and treatment in the next panels, based on the data of

Natural variation in different phases of FLC silencing in the field
Across all the genotypes we tested and all seven field experiments, as expected, FLC expression reduced over weeks in response to autumn and winter temperatures, and VIN3 was upregulated. Previously, we had noted that in 2014-5 in Norwich, substantial VIN3 upregulation did not occur until~58 days after sowing, although temperature conditions were suitable for VIN3-independent FLC downregulation for most of this time . In the following field season, this pattern occurred again, with VIN3 upregulation delayed until 48 days after sowing . For the Col FRI reference, we had found that we could fit two separate exponential decay curves to FLC; the first for the initial, slow, VIN3-independent phase and the second for the faster, VIN3-dependent phase of the downregulation (Figure 2A; Hepworth et al., 2018). Thus three features of the FLC profile contribute to the level of FLC at any time: firstly the 'Starting level' of FLC before vernalization, secondly the rate of downregulation in the initial VIN3independent phase, and thirdly the rate of VIN3-dependent downregulation ( Figure 2A). This pattern was consistent across the accessions and NILs, allowing us to investigate the effect of natural variation on different aspects of FLC regulation. The time of upregulation of VIN3, and thus the time of switching from the VIN3-independent to the VIN3-dependent shutdown, also affects FLC levels, but this was very similar between genotypes at the same site.
In Norwich 2014-5, the RV FLC accession Edi-0 (haplotype group RV1) behaved very similarly to Col FRI (RV2). In comparison, the SV accessions (Lö v-1, Ull2-5, Bro1-6 and Var2-6) show higher levels of FLC throughout the winter (Figure 2, Figure 2-figure supplement 1). Ull2-5 started with similar levels of FLC as Col FRI, but slower rates of downregulation in both the VIN3-independent and VIN3-dependent phases generated higher levels of FLC throughout autumn and winter. For both Var2-6 and its NIL, an apparently slower VIN3-independent rate of downregulation contributed to a higher FLC level, although the difference in rate is not significant. The VIN3-dependent phase was also slower in Var2-6 and Lö v-1 ( Figure 2).
In Sweden 2014-5, at both sites the VIN3-independent and -dependent phases occurred concurrently (  . Instead, most of the natural variation in FLC levels throughout winter in Sweden was generated by differences in the early expression level. In North Sweden again both RV accessions had similar Starting levels whereas the SV Swedish accessions were higher to different degrees. However, in South Sweden Lö v-1 started with similar levels to Col FRI. In the first year of field experiments we observed that a high proportion of the FLC shutdown occurred early in autumn, and furthermore, a lot of the variation in total FLC between lines was during that early phase. To maximise this period, and to extend the time that the plants spent in warmer autumn conditions, so mimicing potential climate change effects, we sowed the second field experiment earlier. For the 2016-7 experiment, we sowed plants in North Sweden and Norwich two weeks earlier and South Sweden three weeks earlier than for the 2014-5 season ( Figures 1A and 3, Figure 3-figure supplements 1-2). The differences in daylight length due to this earlier sowing must be considered in direct comparisons between years, but the timing of both sowing dates was within the natural germination range. In North Sweden, this early sowing was followed by an usually warm autumn: September temperatures at the nearest weather station, Lungö , averaged 11.04˚C between 1990-2016, whereas they were 12.60˚C in 2016, the fourth highest year during that period (Swedish Meteorological and Hydrological Institute, 2020, www.smhi.se/data/meteorologi/laddaner-meteorologiska-observationer). This produced a delay in VIN3 induction similar to that seen in both years in Norwich ( Figure 1A; Antoniou-Kourounioti et al., 2018). However, the first stage of shutdown was still more rapid in North Sweden than in Norwich, despite higher average temperatures in North Sweden ( Figure 1A), so that both VIN3-independent and VIN3-dependent phases had similar rates, resulting in the appearance of a single decline. As in the previous experiment in Sweden, these rates of downregulation were generally similar between different genotypes, with higher FLC levels in SV accessions and the Var NIL due to higher starting levels (   In Norwich, the patterns seemed to be largely repeated, though our statistical power was lower in the 2016-7 experiment due to fewer timepoints. Bro1-6, Lö v-1, Var 2-6 and their NILs appeared to have higher starting FLC levels, though of these only Bro1-6 was significant in our analysis. Ull2-5, Var2-6, Edi-0 and Lö v NIL2 showed slower VIN3-independent downregulation and Ull2-5 and Var2-6 also again showed slower VIN3-dependent downregulation ( Figure 3). Other than a slower VIN3independent rate in Norwich in Edi-0, the RV haplotypes behaved similarly to each other in 2016-7.
The slower rate of the epigenetic silencing phase in Var2-6 was consistent across years in Norwich (Figures 2 and 3). Unlike the other accessions, this change in the VIN3-dependent phase in Var2-6 was consistently mirrored in lower levels of VIN3 ( Nagel et al., 2015). Therefore, variation in the circadian clock may underlie some of the difference in FLC regulation in Var2-6.

Chromatin modifiers control FLC regulation in the field
To understand the contribution of other trans factors to temperature registration at the different phases of FLC shutdown in the field, we exploited our contained site in Norwich to investigate mutants in genes known to affect FLC levels before cold or in response to cold ( Figure 4).
As previously reported , vin3 mutants (vin3-4, vin3-1) did not show the increase in downregulation rate that marks the later epigenetic phase of silencing ( Figure 4C,F). This effect was also seen for other mutants defective in epigenetic memory; vrn1-4; vrn2-1 (defective in the Su(z)12 homologue component of PRC2); vrn5-8 (defective in the VIN3-related protein VRN5); and lhp1-3, in accordance with this phase being required for epigenetic silencing. Loss of the H3K36 methyltransferase sdg8 generated the same effect, correlated with its delayed upregulation of VIN3 (Kim et al., 2010;Jean Finnegan et al., 2011).
In the initial, VIN3-independent silencing phase, both Col fri (Col-0) and sdg8 FRI were hyperresponsive, consistent with a role for SDG8 with FRI in the establishment of the high FLC expression state (Hyun et al., 2017;Li et al., 2018). Mutants in the autonomous pathway, fca-9, fld-4 and fve-3, which upregulate FLC expression in the absence of FRI, had no significant effect on VIN3-independent silencing, but behaved mostly like Col FRI, although fve-3 reduced the rate of downregulation non-significantly in both years. The mutant with the most reduced VIN3-independent response was the B3-binding transcription factor VAL1, required for PRC2 action at FLC ( Figure 4E; Qüesta et al., 2016;Yuan et al., 2016), followed closely by the vrn mutants, vrn1-4 and vrn5-8, and a non-significant but consistent trend for vrn2-1. Conversely, the lhp1-3 and vin3 mutants do not show clear impairment in the early phase of downregulation. Overall, it seems that chromatin modifiers are required for both phases of downregulation in the field.

Autumnal expression of FLC is the major variable in vernalization response
To identify which of the three features of FLC regulation is the major variable that defines the different haplotypes in different climates, we estimated the coefficient of variation for the rates of shutdown and starting levels for all the natural accessions, NILs, and vin3-1 where available (from     2). In Sweden, the starting levels are significantly more variable than the shutdown rates (p-val-ue=2.8Á10 -8 for the first year, Figure 5A), but in Norwich, the VIN3-independent shutdown rate is more variable than the starting levels and VIN3-dependent rate (p-value=2.1Á10 -5 for 2014, p-val-ue=5.0Á10 -4 for 2016, Figure 5B). Combining Norwich and Sweden data, the early shutdown rate is again most variable (p-value=1.9Á10 -10 for 2014, p-value=2.0Á10 -5 for 2016, Figure 5C) hand, there was no significant difference in the variability of the starting levels ( Figure 5D) between the different field sites and years, and similarly for the shutdown rates ( Figure 5E-G). What we describe as the starting level was measured after some days in the field, so it is not equivalent to a non-vernalized control. Some FLC shutdown, most likely VIN3-independent, will have occurred at that time. Therefore, it seems that the combination of these two determinants (starting levels and rate of VIN3-independent shutdown), which together we call 'autumnal expression', provides most of the variation in FLC levels between these accessions and NILs.
Vernalization is saturated before midwinter in the field Having characterised how the haplotypes vary in their FLC dynamic response to field conditions, we investigated whether this variation had an effect on the floral transition in the field. In the 2014-5 season, all accessions and NILs bolted relatively synchronously in spring in Norwich and Sweden, although there were small but significant differences in both cases. In Norwich, the flowering time difference between the NILs and Col FRI was small, and in Sweden not significant, suggesting that most of the variation seen between the accessions did not derive from the FLC haplotype ( Figure 6-figure supplements 1 and 2). There was also very little difference in branch production in the NILs and accessions in Norwich ( Figure 6-figure supplement 1), which was surprisingly low, despite the known role of FLC in suppressing branching (Huang et al., 2013;Lazaro et al., 2018). Nevertheless, we confirmed that this effect does occur in the accessions and NILs in controlled condition experiments, in which longer vernalization times resulted in increased numbers of branches ( Figure 1-figure supplement 1).
The observed flowering times suggested that the vernalization requirement is normally saturated during the winter in the field. To test this hypothesis, in 2016-7 we transferred plants from field conditions to heated, long-day conditions at different stages, and scored the length of time until floral buds were visible at the shoot apex (bolting). There was wide variation and substantial delay to bolting within the genotypes for plants transferred to the warm greenhouses early in autumn, at all sites, indicating that vernalization requirement had not been saturated at this point ( Figure 6A,C,E). Indeed, in both Swedish sites many plants did not flower within the experimental time (136 days after transfer for South, and 104 days for the North) when transferred 25 days after sowing. Some genotypes (particularly Lö v-1) required more than 46 days in the field before most individuals were competent to flower in warm, long-day conditions ( Figure 6-figure supplement 3). There was also clear variation between genotypes in their time to flowering in the earlier transfers. However, as winter progressed, time to bolt after transfer reduced quantitatively and became more uniform for plants within and between each genotype ( Figure 6B,D,F), although vernalization saturated at different rates in different genotypes ( Figure 6-figure supplements 3-4). All accessions and NILs bolted broadly synchronously and rapidly after removal on 21st December in Norwich, in North Sweden NILs and accessions were almost synchronous by 24th November, in line with previous findings (Duncan et al., 2015), and in South Sweden FRI-containing lines were broadly synchronous by the last transfer on 17th December ( Figure 6, Figure 6-figure supplements 3-4). Therefore, in current climates almost all Arabidopsis plants have probably saturated their requirement for vernalization before midwinter, ensuring that flowering time is not delayed the following spring.

Relationship of FLC expression and timing of flowering
We tested whether the observed effects on flowering time were linked to the varying levels of FLC in transferred plants. Across genotypes, in Norwich and North Sweden 2016-7, the time to the floral transition after transfer from the field correlated closely with FLC expression at the time of transfer, as expected (Norwich, R 2 = 0.68, p<0.001, North Sweden, R 2 = 0.85, p<0.001, linear regression Figure 6G,H). Accessions, NILs and mutants with high starting levels and slower downregulation rates generally bolted later and took longer to saturate their vernalization requirement, indicating that the variation in FLC dynamics in autumn has a phenotypic effect.
However, each accession differed in the relationship between FLC levels at transfer and subsequent bolting time ( Figure 6-figure supplements 5-6). For example, for Lö v-1 and the Lö v NIL1, the time to bolt for a given level of FLC at transfer is longer than for Col FRI, but for Edi-0, it is shorter ( Figure 6-figure supplements 5-6). This analysis allowed us to extract a further feature of FLC regulation (Figure 2A). We named this feature the FLC-post-vern value (m in  bolting-per-FLC-unit), representing the slope of the regression between FLC spliced RNA levels at the time of transfer (x-axis) to the subsequent floral transition time for these transferred plants (yaxis, as for Figure 6G and H). This means that, for a given level of FLC at the time of transfer, a genotype with a higher m 'post-vern' value flowers later than one with a lower m value. We calculated this value individually for each genotype from both the Norwich and the North Sweden transfers (Figure 6-figure supplements 5-6), using the equation given in the legend for Table 1. There was substantial variation between the different glasshouses in the estimates for the accessions ( Table 1), suggesting that there is a large genotype-by-environment effect on this parameter due to non-FLC effects. Moreover, for Col-0 and Lö v-1, estimations of m were unreliable (p>0.1) due to flowering occurring rapidly even without vernalization (as FLC values are too low to cause measurable differences in flowering: Col-0) or not at all (as in North Sweden FLC values were       NA -estimate only based on two data points, so no standard error is calculable. n.d. -no data. No value given for Col-0 as initial FLC value is too low for useful estimation. too high to allow flowering and its timing to be measured within this experiment: Lö v-1). Practically, therefore, although any genotype with functional FLC should have a theoretical post-vern value, it can only be measured for high-FLC genotypes under conditions in which they will nevertheless flower.
The m post-vern value also varied between the NILs and the common Col FRI parent within each site, suggesting that this relationship is also influenced by cis variation at FLC. Within the lines in the common Col FRI background, the estimates correlated more closely across the two glasshouse environments, with the exception of the Ull NIL. Therefore, within the Col FRI background, we can quantify the effect of FLC levels on bolting time in warm, long-photoperiod environments, and find that the FLC haplotypes have different effects after vernalization, as well as different responses to vernalization. This effect may be linked to reactivation of FLC expression in the warm, as the vin3-1 and Lö v-1 NIL1, both of which have reactivation phenotypes (Sung and Amasino, 2004;Duncan et al., 2015;Qüesta et al., 2020), have high 'post-vern' values.  Figure 6). Therefore, although we observed natural variation in FLC levels in autumn ( Figure 5), at most sites in most years the conditions were such that this did not result in flowering time variation. However, in the warm North Sweden autumn 2016-7 many of the plants with the Col background (NILs, Col FRI, and vin3-1 FRI) transitioned to flowering early, by the 18th November, before winter and snowfall ( Figure 7A). This precocious bolting was rare in the SV accessions and was much reduced even in the RV accession Edi-0. Over all the genotypes, the percentage of plants transitioning to flowering before winter negatively correlated with genotype FLC expression on 5th October, one day after the first recorded bolting for plants in the field (p<0.001, Generalised Linear Models (GLMs) for binomial data, Figure 7B). This was before substantial upregulation of VIN3, indicating that this variation in FLC must derive from the combination of the starting levels and VIN3-independent shutdown. Within the Col FRI SF2 background the Var and Lö v FLC haplotypes, the NILs with the highest FLC levels during autumn (Figure 3-figure supplement 1), substantially reduced precocious bolting (p=0.005, binomial proportions test, Figure 7A). Thus, variation in FLC starting expression and shutdown (collectively called autumnal expression) results in variation in the alignment of flowering with the seasons.

Variation in FLC expression correlates with fitness in the field
Across all genotypes, plants that bolted precociously were not observed to set seed before winter.
In this year (2016) and at this site (N.Sweden), mortality was high, compared to other years and sites, and plants that bolted precociously were less likely to survive the winter (42% of bolting plants survived, whereas 67% of non-bolting plants survived to spring, p<0.001, binomial proportions test, Figure 7-figure supplement 2A). Therefore, high FLC levels in October correlated with higher survival to seed set in the field (p<0.003, GLM for binomial data, Figure 7-figure supplement 2B).
We therefore investigated whether FLC was linked to total genotype fitness, as measured by silique number, and found a complex pattern ( Figure 7C). Despite many of the NILs having a high level of precocious flowering, several NILs produced more seed than the locally-adapted accessions ( Figure 7C, Figure 7-figure supplement 2C).
To better understand why some NILs had higher fecundity than others, we investigated whether FLC might have an additional effect on silique production. In the Col FRI background plants, there was no overall penalty in average silique number for individual surviving plants that bolted before winter compared to those that transitioned afterwards ( Figure 7C, Figure 7-figure supplement 2D), suggesting that the differences were not due to an FLC effect on flowering time. Saturating vernalization increased seed set in the 2016 Norwich transfer experiments, as plants that were transferred before saturation of vernalization requirement produced lower and more variable amounts of seed ( Figure 6-figure supplement 7). However, in the field in North Sweden 2017 silique production occurred only after winter, after saturation of vernalization requirement ( Figure 6F). We noted   that there was large variation in branching within the Col FRI background plants, and branching is known to be affected by FLC (Figure 1-figure supplements 1-2; Huang et al., 2013;Lazaro et al., 2018). We asked whether the likelihood of FLC reactivation after winter might affect branching and thereby silique production. To estimate this, we used as proxy for likely FLC reactivation the relationship between FLC and bolting time in the warm, the 'FLC-post-vern value' as derived from Figure 6-figure supplements 5-6 ( Table 1). This did not seem to affect survival to seed set or date of bolting of the survivors ( As expected, survival was the principle factor explaining total silique production (Figure 7-figure supplement 2I). However, we found that although the post-vern value of FLC did not well explain total genotype fitness on its own ( Figure 7F), when combined with survival to seed set both factors explained a large part of the variation in genotype fitness in Col FRI background (linear model adjusted R 2 = 0.94, p-value=0.001 for model, post-vern p-value=0.006, percentage survival to silique set p-value=0.002, Figure 7F, Figure 7-figure supplement 2I).

Discussion
Investigation of the dynamics of key floral integrators in the field has recently led to important insights into their molecular response to natural environments Song et al., 2018). Here, we integrated experiments on the expression of FLC in field conditions with investigation on how natural genetic variation and induced mutation interact in distinct climates to affect phenotype and fitness. In our previous work, we found that, as expected from laboratory studies, FLC was repressed and VIN3 was induced by winter cold Hepworth et al., 2018). However, in 2014-5, winter conditions in Norwich generated a subtly different response to those in Sweden, activating the VIN3-independent transcriptional shutdown before the VIN3-dependent epigenetic pathway (delay in VIN3 onset and slow then fast shutdown of FLC in Norwich, Figure 2-figure supplement 1). In the following season, this pattern was triggered again, not only in Norwich, but also in North Sweden, implying that this is a common occurrence that plants must adjust to across climates, although with different frequencies at different locations ( Figure 3-figure supplement 1).
We find that variation in FLC mRNA levels in both accessions and NILs is largely due to differences in both the starting FLC level and the VIN3-independent rate of shutdown (collectively called autumnal expression), which vary widely between accessions but not between field sites or years ( Figure 5). This variation in rate of response, as well as absolute levels of FLC, partly explains why FLC levels measured at any one time may not correlate with the final flowering time phenotype in different accessions (Sasaki et al., 2018).
In Norwich, we investigated the mechanisms of the responses. Mutants of the autonomous pathway, which increase FLC expression in fri plants in laboratory conditions (Ausín et al., 2004;Liu et al., 2010;Wu et al., 2016), generally behaved in a similar manner to Col FRI in terms of vernalization response (Figure 4). However, the reduced response of the fve mutants in the VIN3- (R 2 = 0.59, p=0.0058). (C) Total number of siliques produced per genotype, showing contribution from plants that bolted before winter and plants that bolted after. Within the Col FRI genetic background there was no overall penalty in average silique number for surviving plants bolting before vs. after winter (92 and 77 per plant respectively, not significant in Mann-Whitney U test). (D) Mean silique production in plants surviving to set seed positively correlated to their mean rosette branch production for Col FRI genetic background genotypes (NILs and vin3-4; R 2 = 0.56, p-value=0.002). (E) Rosette branch production of Col FRI genotypes surviving to set seed is strongly negatively correlated with the FLC post-vern value for that genotype as from Table 1 (R 2 =   independent phase, though not significant in each year, may be worth further investigation, as FVE has been implicated in intermittent cold-sensing through histone deacetylation at FLC, independently of vernalization (Kim et al., 2004;Jung et al., 2013).
All three factors that set FLC levels in the field (starting levels, VIN3-independent and VIN3dependent phases) require chromatin modifiers for their correct function. The VIN3-independent phase involves chromatin-modifier regulation of transcription, with the VIN3-dependent phase using a similar set of modifiers. These chromatin regulators are rarely found in genome-wide-association studies for flowering time, likely because their alteration would have pleiotropic effects. In the natural accessions we find that the VIN3-dependent (epigenetic) phase is indeed the least variable (Figure 5). This may explain why so much variation in vernalization maps to FLC itself (Lempe et al., 2005;Sánchez-Bermejo et al., 2012;Dittmar et al., 2014;Sasaki et al., 2015;Bloomer and Dean, 2017;Sasaki et al., 2018). Nevertheless, we did observe variation in VIN3 expression in Var2-6, which seems to derive from differences in circadian regulation (Figure 3-figure supplement 3), and QTLs over the VIN3 region have previously been found to affect flowering (Dittmar et al., 2014;Méndez-Vigo et al., 2016;Å gren et al., 2017). Variation in VIN3 expression in the other accessions we tested was more complex than the general downregulation in Var2-6. Given the intricate regulation of VIN3 induction , identifying the source of this difference will require detailed investigation.
The importance of FLC variation in adaptation for natural populations is well established (Méndez-Vigo et al., 2011;Sánchez-Bermejo et al., 2012;Dittmar et al., 2014;Li et al., 2014;Duncan et al., 2015;Å gren et al., 2017;Bloomer and Dean, 2017). However, we found that even in a challengingly warm year, vernalization requirement saturated before midwinter across three climates ( Figure 6). This response had been seen previously in North Sweden in the locally-adapted Lö v-1 accession as a reaction to the extreme winters (Duncan et al., 2015). However, we find that this is a general response in all our tested accessions. The principle exceptions were the vin3 and vrn mutants, which had notably delayed flowering in both years at the Norwich site. The consequences of incomplete vernalization are severe -in our 2016-7 transfer experiments this led to delayed flowering and reduced fecundity ( Figure 6, Figure 6-figure supplement 7), and several field experiments have found a general trend towards earlier flowering promoting fitness in environments with warmer winters (Å grena et al., 2013;Fournier-Level et al., 2013;Grillo et al., 2013;Dittmar et al., 2014;Å gren et al., 2017;Exposito-Alonso et al., 2018). In 2014-5, the high synchronicity of flowering in all the accessions and NILs after winter and the fact that flowering aligned with spring across our field sites ( Figure 6-figure supplement 2), suggests that in most plants vernalization has adapted to reduce the delaying effect of FLC on flowering as early as possible.
Nevertheless, we observed that the Swedish, cold-winter-adapted accessions and FLC haplotypes associated with slow vernalization expressed the highest FLC levels throughout autumn and early winter (Figures 2 and 3, Figure 2-figure supplement 1, Figure 3-figure supplement 1). In North Sweden in 2016-7, we captured behaviour for a warmer growing season that revealed an adaptive role for these high FLC expression levels. There was a relatively warm September that year with a long period with less precipitation (Swedish Meteorological and Hydrological Institute, 2020, www.smhi.se/data/meteorologi/ladda-ner-meteorologiska-observationer), which required us to water the plants to avoid drought stress. This, in combination with the earlier sowing (warmer and longer days) meant that the equivalent period of growing time for the plants (relative to their germination time) was more inductive to growth and flowering. This allowed us to capture a different behaviour. Plants are exposed to both types of the years we recorded, since the timing of both sowing dates was within the natural germination range, although in natural environments seed dormancy has a strong influence on phenology and is highly responsive to temperature and microclimate, and affected by FLC itself (Springthorpe and Penfield, 2015;Kerdaffrec et al., 2016;Marcer et al., 2018). In our experiments, in plants with a haplotype characteristic of a slow-vernalizing accession, including the locally-adapted Swedish accessions, the higher level of FLC expression protected against precocious flowering and its consequent reduced survival (Figure 7, Figure 7-figure supplement 2). These rare, but highly selective occurrences, may play a role in local adaptation, although we do not know the reason behind the high mortality in this year. However, that the effects of these haplotypes are only revealed occasionally is a logical consequence of the fact that the flowering time genes can have strong or weak effects depending on the environment (Wilczek et al., 2009;Sasaki et al., 2015;Burghardt et al., 2016;Fournier-Level et al., 2016;Sasaki et al., 2018;Taylor et al., 2019).
FLC also controls fecundity as well as survival (Å grena et al., 2013). Li et al., 2014 found that FLC haplotypes from SV accessions produced lower seed weight compared to those from RV accessions in non-saturating vernalization conditions, though the mechanism for this effect is unknown. In the field, silique production is closely linked to branch production (Figure 7; Taylor et al., 2019). In Arabidopsis, saturation of vernalization requirement is known to increase flowering branch production, particularly rosette branch production, and this effect is linked to FLC (Figure 1-figure supplements 1-2; Huang et al., 2013;de Jong et al., 2019). As expected, FLC levels in autumn did not correlate with branch production in spring (Figure 7-figure supplement 2H), and in Norwich 2014-5, there was little variation in branch production. However, we found that under the conditions in the field in North Sweden, branch production was negatively related to a factor we named 'FLC post-vern', which encoded the relationship between FLC levels and subsequent flowering in warm controlled conditions (Figure 7). Variation in this factor probably derives from regulatory differences at FLC post-cold, such as the reactivation phenotype of the Lö v-1 FLC allele (Shindo et al., 2006;Coustham et al., 2012;Duncan et al., 2015;Qüesta et al., 2020), allowing FLC repression of flowering to saturate at the shoot apical and axillary meristems at different rates, a phenomenon that occurs in the perennial relative Arabis alpina (Wang et al., 2009;Lazaro et al., 2018). Why the field conditions in Sweden, but not Norwich, revealed this effect is not yet clear.
In summary, our detailed analysis of the different phases of FLC silencing through winters in distinct climates, over multiple years, has shed light on adaptive aspects of the vernalization mechanism. Autumnal FLC expression, made up from FLC starting levels and early phases of FLC silencing, is the major determinant for variation in FLC levels during vernalization ( Figure 5). FLC downregulation aligns flowering response to spring and non-coding variation affects this alignment in different climatic conditions and year-on-year fluctuations in natural temperature profiles. In a changing climate, understanding the complex genotype-by-environment interactions that govern timing mechanisms will become ever more important.

Plant materials
Sources of previously described mutant lines and transgenics are presented in Supplementary file 4. Requests for materials should be addressed to Caroline Dean.

NILs
All near-isogenic lines were produced by six rounds of backcrossing to the Col FRI parent, selecting for the introgressed FLC in each generation, before one round of selfing and selection of homozygous families.

Experimental conditions Field experiments
Field experiments have been described previously (2016; Hepworth et al., 2018Hepworth et al., (2014). Briefly, seeds were stratified at 4˚C for three days. For gene expression measurements, for all field sites and sowing dates and timepoints within them, six replicate tray-cells were sown using a block-randomised design within 5.7 cm 28 cell trays (Pö ppelman, Lohne, Germany), and where each replicate included material from at least three plants. For flowering time, plants were thinned to a single plant per cell in 3.9 cm 66 cell trays (Pö ppelman). Trays were watered when necessary.
In Norwich, trays were placed on benches in an unheated, unlit glasshouse, and bedded in vermiculite. For expression, plants were randomised within six single-replicate sample-sets, which were then randomised using Research Randomiser (Urbaniak and Plous, 2015) in a three complete block design lengthwise along the greenhouse, adjusting to ensure each of the two replicates per block were on different benches. For flowering time, plants were block-randomised per tray and per block.
In Sweden, trays were germinated and grown outside under plastic covers for two weeks at Mid Sweden University, Sundsvall (North Sweden) or Lund University (South Sweden), due to the difficulty of sowing directly into the field and to ensure sufficient germination for analysis, based on previous experiments. Trays were then moved to the experiment sites and dug into the soil. The experiment site in the North was at Ramsta (62˚50.988´N, 18˚11.570´E), and in the South at Ullstorp (560 6.6721´N, 13˚94.4655´E). Expression sample-sets were randomised in three blocks. Flowering time plants were in a completely randomised design across two   For all expression analysis except the 48 hr sampling (Figure 3-figure supplement 3), six replicates per timepoint per genotype were chosen in order to allow sufficient number of samples for statistical analysis while allowing for losses in the field and to allow duplication within randomisation blocks. Where resulting samples are smaller, this is due to experimental or processing loss (e.g., death of plants in the field, degradation due to poor sample quality or processing, see RNA extraction and QPCR). For Figure 3-figure supplement 3 and replicates per timepoint per genotype were chosen due to space constraints. Each expression sample (single replicate) was of at least three plants pooled. For flowering time, 12 plants per genotype per transfer condition were chosen to provide replication across blocks and trays while remaining within size constraints. For field flowering in North Sweden 2016-7, 36 plants per genotype were sown to allow for losses, although in the event these were more substantial than anticipated. Temperature was recorded at plant level at each site with TinyTag Plus two dataloggers (Gemini Data Loggers (UK) Ltd, Chichester, UK). Bolting was scored when flower buds were visible at the shoot apical meristem. For the North Sweden 2016-7 field experiment, plants were scored for survival and flowering in the field from planting to December 2016, and then from March to May 2017. Plants were harvested and scored for branching and silique production after the end of flowering, in July 2017.

Branching analysis
; Seeds were sown on soil, stratified after sowing for three days at~4-5˚C, and transferred to a Norwich long-day glasshouse set to 18˚C/15˚C, 16/8 hr light/dark conditions for 7 days before being returned to vernalization conditions (a 4˚C growth chamber under short day, low light conditions; 8/ 16 hr light/dark) for 12, 8, 4, and 0 weeks. Sowing was staggered so that after vernalization all plants were transferred to glasshouse conditions simultaneously. Plants were scored for their flowering time, total branch number, cauline branch number and rosette branch number. In all cases, plants were randomised into blocks and at least three replicate plants for each accession/cultivar per treatment were scored for their flowering and branching phenotypes. Primary rosette and cauline branch number were scored at senescence.

RNA extraction and QPCR
RNA extraction and QPCR for field experiments were performed as described in Hepworth et al., 2018 andAntoniou-Kourounioti et al., 2018. Field data were unified across sites and timepoints within the yearly datasets and normalised to a synthetic control sample, as described in Hepworth et al., 2018, to give the 'normalised concentration' reported in the results. QPCR results were analysed using LinReg (Ruijter et al., 2009), and normalised to the geometric means of At5g25760 ('PP2A') and At1g13320 ('UBC') control genes (Czechowski et al., 2005;Yang et al., 2014).
QPCR samples that showed high Cp values (UBC Cp >28 for LinReg analysis) of the control genes, indicating possible degradation, were excluded if test amplicon (FLC, VIN3) results were also anomalous (criteria: absent, or varying from non-flagged samples by an order of magnitude). Any measurements where amplicon Cp values varied by more than 0.6 were also excluded if insufficient sample was available for a repeat.
Primers used are described in Supplementary file 5.

Statistics
The rates of FLC shutdown were estimated using linear regression, as shown in Figure 2A. For Norwich 2014-5 and for all sites in 2016-7, a separate rate was calculated for timepoints before the point of induction of VIN3 (Norwich 2014:~58 days, Norwich 2016:~48 days, South Sweden 2016: 35 days, North Sweden:~46 days) and for those after. For North and South Sweden 2014-5, when the VIN3 induction was not delayed, all timepoints were combined, excluding measurements after 155 days. These timepoints were during or after the snow, when temperatures started to increase and reactivation was observed in some genotypes.
For the analysis presented in Supplementary file 1 and Figures 2-4, the Starting level comparisons and the Shutdown rate comparisons were done separately, using the following methods. For the starting levels, ANOVA was performed in the R (R Development Core Team, 2018) statistical language using the lm function, followed by a Dunnett post-hoc test comparing all genotypes against the control Col FRI. For the shutdown rates, the lmer function from the lmerTest package in R was used to perform the same comparison on the genotype-timepoint interaction, controlling for blocks as random factors and after setting the mean of the timepoints to 0 on the x-axis, using the default Satterthwaite's method for the post-hoc t-tests.
We used the R package cvequality (Version 0.1.3; Marwick and Krishnamoorthy, 2019) with the asymptotic test (Feltz and Miller, 1996) to assess differences between the coefficients of variation of different groups as described in the text and Supplementary file 2. The significance limit was adjusted to control the false discovery rate using the Benjamini-Hochberg procedure with a false discovery rate of 0.05 (Supplementary file 2, Supplementary file 6). This analysis was performed including all the natural accession and NILs combined (Figure 4), but also for all the accessions separately and for all the NILs separately (including Col FRI in both cases, but not Col-0 or mutants). The different analyses do not change our overall conclusions and all three are reported in Supplementary file 2.
Statistical analysis for flowering and branching data (Supplementary file 3, Figure 6, Figure 6figure supplements 1, 2, 3 and 4 and Figure 7-figure supplement 1) were performed on the plants that flowered during the experiment using the same procedure as for the Starting levels described above.
For multiple regression on field data, R was used to obtain minimal adequate models using linear regression (lm function), except when n > 10 for count data, for which general linear models (GLM, glm function) using Poisson error distributions were used (total numbers of siliques), or for proportion data for which GLMs with binomial errors were used (survival, bolting before winter).
helpful suggestions on the initial submission of the manuscript. Finally the authors would like to reiterate their appreciation of all those from the Dean, Howard, Holm, Sä ll and Irwin groups who helped in cold, heat, wind, rain and laboratory with the field studies, plus Ingalill Thorsell at Drakamö llan Gå rdshotell for making the lab retreats during the field studies so enjoyable. The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.  . Transparent reporting form

Data availability
All data generated or analysed during this study are included in the manuscript and supporting files. Source data files have been provided for all figures.