SlimVar: rapid in vivo single-molecule tracking of chromatin regulators in plants

Epigenetic regulation maintains gene expression patterns over many rounds of cell division in higher organisms. However, visualization of factors regulating epigenetic switches in vivo is limited by the challenge of imaging cells deep in living tissue, with molecular sensitivity and rapid sampling. We report an easy-to-implement method called Variable-angle Slimfield microscopy (SlimVar), which by simple modification of an inverted optical microscope, enables single-molecule tracking of fluorescent reporters in Arabidopsis thaliana. Using SlimVar, we imaged stepwise photobleaching of chromatin-protein assemblies in individual nuclei, 30 µm deep in root tips through multiple cell layers. We find that two homologous proteins key to the epigenetic switch at FLOWERING LOCUS C (FLC) —cold-induced VERNALIZATION INSENSITIVE3 (VIN3) and constitutively expressed VERNALIZATION 5 (VRN5)—exhibit dynamic nuclear assemblies during FLC silencing. Upon cold exposure, these assemblies increase in stoichiometry by up to 100% to a median of ∼20 molecules. Larger VRN5 assemblies preferentially co-localize with an FLC lacO transgenic reporter during prolonged cold, persisting after return to warm conditions. Our findings support a hybrid model of epigenetic memory in which nucleation of histone trimethylation is assisted by dynamic protein assemblies over extended durations. SlimVar therefore has potential to offer molecular insights into proteins expressed at physiological levels in a range of tissues.


Introduction
Understanding the mechanistic basis of epigenetic silencing remains a major question, with implications for cell differentiation and ageing through to crop security [1].A conserved epigenetic mechanism across eukaryotes is Polycomb-mediated silencing [2], which involves the histone modification trimethylated histone H3 lysine 27 (H3K27me3).H3K27me3 initially accumulates at a nucleation site and then spreads across a locus to stably maintain a silenced state through many rounds of cell division [3], [4].A gene where Polycomb Repressive Complex 2 (PRC2) silencing has been well studied is Arabidopsis FLOWERING LOCUS C (FLC) [5].FLC encodes a repressor of flowering and is epigenetically silenced during a process called vernalization by the prolonged cold of winter; repression of a repressor thus enables flowering in spring [6], [7], [8].Cold exposure increases the proportion of FLC alleles that have epigenetically switched from ON to OFF states through nucleation of H3K27me3 at an intragenic site [9], [10].Upon return to warm conditions, H3K27me3 spreads across the locus to give long-term stable silencing [11].The relative stability of the nucleated state through cell division in the cold at FLC has raised an interesting question as to the mechanism of epigenetic inheritance.Each nucleation event involves only three nucleosomes, too few to survive random replicative dilution through the classic Polycomb 'read-write' mechanism [8], [12], [13].Metastable protein assemblies have been proposed to explain the inheritance of the nucleated state [14].PRC2 accessory proteins are thought to restore the stochastic loss of histone marks by stimulating Polycomb Repressive Complex 2 (PRC2) activity [15].While individual proteins may only interact transiently with nucleation factors and the locus, an assembly (with the appropriate positive cooperativity) is predicted to become dynamically self-sustaining above a threshold number of recruited proteins [14].In this framework, a sufficiently large assembly could therefore act as a binary memory element working together with established H3K27me3 machinery at a given locus.Two PRC2 accessory proteins required for stable cold-induced silencing at FLC, VERNALIZATION INSENSITIVE3 (VIN3) and VERNALIZATION 5 (VRN5, also known as VIN3-LIKE 1 / VIL1), expressed in shoot and root tips are clear candidates for this form of memory storage [16].VIN3 and VRN5 (collectively called VEL proteins) associate with the PRC2 complex and with the FLC nucleation region that accumulates H3K27me3 specifically during cold conditions [8], [17].At <15 o C, VIN3 expression gradually rises over several weeks, while in warm conditions >15 o C, expression decays rapidly within ~4h [6], [7], [17].VIN3 and its assemblies could therefore in principle report the duration of cold conditions during winter to promote an epigenetic switch.However, due to its instability in the warm, one or more additional factors-a promising candidate being VRN5-are required at the FLC nucleation region to explain persistent memory of silencing following a change to warm conditions.Both VIN3 and VRN5 contain complex plant homeodomains (PHDs) [16] which do not interact directly with histone tails [20], FNIII domains and C-terminal VEL domains [18], [19].Of these three domains common to VIN3 and VRN5, the VEL domain mediates head-to-tail interactions.VEL proteins that oligomerize and even form phase-separated droplets under transient overexpression are directly associated with stable silencing at FLC [19].A key question is then: do physiological levels of VIN3 and VRN5 oligomerize sufficiently in the vicinity of FLC during vernalization, to fulfil this model?Here, we describe development of an efficient new imaging modality, Slimfield Variable Angle (SlimVar) and apply it to rapidly track VEL proteins in live plant tissue.Investigation of nuclear oligomers has been constrained by an inability to detect individual, rapidly diffusing protein molecules deep within plant tissue.Traditional confocal [21], structured illumination [22] and lightsheet [23] methods lack the required combination of sensitivity and speed for single-molecule tracking in live plants, which remains a technical challenge beyond the first cell membrane [24], [25], [26], [27], [28], [29], [30], [31].Although complex and expensive super-resolution methods including lattice lightsheet [32], [33] and MINFLUX [34], [35] are capable of deeper imaging, to date none have been adequately optimised in plants.In addition, these typically require both specialised hardware and photoactivatable / photoswitchable fluorescent proteins [36], [37] or dyes [38], [39].Our method, SlimVar, instead leverages common fluorophore fusions in existing transgenic plants, and an accessible microscope platform.SlimVar adapts Slimfield microscopy [40], [41], [42] with a grazing angle of incidence optimal for observing single nuclei up to ~30 µm deep in root tissue, while tracking single fluorescent protein molecules at 40 nm resolution.Whereas normal Slimfield has been limited to simple unicellular samples or the surfaces of ex vivo tissue sections [43], SlimVar enhances the image contrast to enable dynamic spatial localization in complex multicellular samples.The use of root tips confers optical advantages of lower autofluorescence and more regular, less refractive layers than other plant tissues.This single-molecule sensitivity can be combined with stepwise photobleaching analysis [40] to quantify the number of molecules in any observed oligomeric assemblies.This sensitivity simplifies use of transgenic plant lines expressing from as few as a single gene copy and enables imaging of physiological states of low-abundance nuclear proteins.
We find that both VIN3 and VRN5 proteins form assemblies in cell nuclei, composed of consistent dimeric subunits.In lines expressing from single-digit exogenous copies of VIN3 and VRN5, the median assembly comprises up to ~20 molecules of each protein, in agreement with the required stoichiometry of protein memory elements predicted from modelling [14].We also use an FLC-lacO/LacI-YFP transgenic reporter to localize the FLC locus and characterise its mobility relative to VEL proteins.Finally, we demonstrate dualcolour SlimVar, which directly shows VRN5 assemblies present at FLC.We find that larger VRN5 assemblies preferentially colocalize with FLC after long cold exposure and after return to warm conditions, and that this interaction between individual larger oligomers of VRN5 and FLC is dynamic on a sub-second timescale.

SlimVar resolves dynamic single particles of VIN3 and VRN5 in plant nuclei
As an initial benchmark, we used traditional confocal imaging with a typical sampling time of 35 s per frame to examine VIN3 and VRN5 protein localization, during and after cold treatment (Fig. 1b, ColFRI negative control Supplementary Fig. 1).We utilised lines with VRN5-EYFP/vrn5-8 JU223 [8], [16] or VIN3-EGFP/vin3-4 FRI [8], [44] (referred to as VRN5-YFP and VIN3-GFP respectively).To investigate the effects of protein expression and to exploit the greater imaging contrast and lower autofluorescence associated with yellow and red fluorophores, we later also characterised new VIN3-SYFP2/ColFRI line (Supplementary Figs. 2 and 3) and VRN5-mScarlet-I/vrn5-8 FRI lines (Supplementary Fig. 4).These lines include active FRI alleles so require effective vernalization for flowering [17].After identifying nuclei from transmitted light images, we found that both VIN3 and VRN5 exhibited high intensity but largely diffuse fluorescence localized to the nucleoplasm.While VRN5 was detectable well above background levels in all nuclei (N=241) at all timepoints, VIN3 was only discernible at diurnal maxima during the cold period itself.Its level decreased at subsequent timepoint after this, being undetectable within one week after return to warm conditions, as reported previously [7].During cold, the VIN3 signal per cell was initially greatest in the vicinity of the meristem and epidermis, before becoming brighter in all cells after further cold exposure.We then optimised Airyscan, an enhanced form of confocal laser scanning microscopy which uses a point detector array to obtain faster frame sampling times down to 60 ms for individual root tip nuclei [45].Images showed a marginally more granular spatial patterning for VIN3 and VRN5 than standard confocal, hinting at the presence of distinct foci within the diffusive fluorescence (Fig. 1c).We then implemented SlimVar (Fig. 1a), capable of single-molecule fluorescent protein detection at sub-ms levels, and found the same nucleoplasmic localization, but instead of diffuse fluorescence we observed multiple individual distinct fluorescent foci (Fig. 1d) consistent with a greater sensitivity and sampling speed that eradicates motion blur from mobile protein assemblies.SlimVar employs a very narrow field of oblique-angle epifluorescence illumination (similar to variable angle epifluorescence microscopy (VAEM) [47] and highly inclined illumination (HILO) [48], [49]) and high localized excitation intensities to enable millisecond sampling.To achieve this, the excitation laser is focused into the back focal plane of a high numerical aperture immersion objective lens at a precise lateral displacement from the optic axis.To mitigate refractive index mismatch and facilitate deeper tissue imaging up to 30 µm from the coverslip surface, we performed optimisation at a representative depth of 20 µm in a live root sample (Methods).The resulting beam is sufficient to encapsulate individual cell nuclei between 4-16 µm wide but with minimal excitation of the remaining >70% of the cell volume.It is ideally aligned for the target root tip cells in the epidermis, cortex and endodermis overlaying the stem cell niche (Fig. 1a), with associated reduction in aberration, backscatter and out-of-focus fluorescence excitation of intermediate cell layers, including the lateral root cap.The net result is an improvement in signal-to-noise for imaging VIN3 and VRN5 fluorescent reporters by a factor of ~3 (Methods), enabling independent single-molecule detection consistent with control samples of purified fluorescent protein (Supplementary Fig. 5).

Vernalization induces upregulation and self-assembly of VIN3 and VRN5
We integrated the pixel intensities in SlimVar images to quantitatively determine the nuclear abundance of VIN3 and VRN5 (Methods).The counts per nucleus were normalised by the characteristic intensity due to a single fluorescent protein, determined by singlemolecule photobleaching steps in situ [40] (Supplementary Fig. 5).We define the nuclear protein copy number as the difference in mean integrated intensity (Fig. 2) between the test population and the negative control, ColFRI, and identify this with the total number of labelled molecules of either VIN3 or VRN5 in a typical nucleus.The nuclear protein copy number (number of labelled molecules excluding any associated with autofluorescence) is then calculated as the net excess above the mean level of the negative control ColFRI (horizontal line), which represents autofluorescence background.Bar, box and whisker denote median, interquartile range (IQR) and 1.5× IQR respectively; cross: mean ± sem.The abundance of VIN3 protein was insignificant prior to vernalization (Brunner-Munzel (BM) test vs ColFRI, N=66, p=0.11 |ns: not significant at adjusted p<0.01).However, the VIN3 nuclear protein copy number increases sharply to ~28,000 ± 3,700 molecules above control after 2 weeks cold (N=64, p=0.0031 |*), and peaks at ~44,000 ± 4,700 after 6 weeks cold (N=83, p<0.001 |**).Following transfer to warm conditions, VIN3-GFP levels reduce to ~3,200 ± 1,600 molecules within 7 days (N=74, p=0.04 |ns).VRN5 levels increase between non-vernalized and post-vernalized timepoints from ~110,000 ± 23,000 to ~190,000 ± 37,000 molecules (N=94, p=0.0089|*).b) Distributions of the number of tracks per nucleus (binned in columns of width 2 for clarity).The number of VRN5 tracks increases initially (NV: 20.8 ± 1.9 up to 26.8 ± 1.6 tracks per nucleus at 2 weeks cold; N=86, p=0.0054|*) but stops increasing after 6 weeks' cold (27.0 ± 1.5 and 26.2 ± 2.6 tracks per nucleus at 6 weeks cold and 14 days post-cold respectively; N=94, p=0.80|ns); c) Collated distributions of stoichiometry (number of labelled molecules per nuclear assembly) of individual tracks at different timepoints before, during and after vernalization; nt: no tracks detected.
We explored the effect of transgene copy number on the abundance of the VEL proteins.We generated a homozygous single transgene copy line of VIN3-SYFP2, though not in a deletion background, meaning endogenous VIN3 is also present at a similar level (Supplementary Fig 2a .).The expression of exogenous VIN3 still follows the expected pattern (Supplementary Fig 1 .)but at much lower levels, reflecting the reduction to a single transgene copy of VIN3.In both VIN3 fusions the nuclear protein copy number exhibits an increase from two to six weeks (Fig 2a ., Supplementary Fig 3a .),although the distributions partially overlap.The two-week level is between half and two-thirds of the six-week level that approaches full vernalization.VRN5 was highly abundant at all timepoints, with levels an order of magnitude greater than VIN3 in the VIN3-labelled lines (Fig 2a).The total amount of VRN5 approximately doubles in response to full vernalization and persists after return to warm conditions.Nuclear protein copy numbers translate to nucleoplasmic concentrations of ~100 nM-1 μM for VIN3 and 1-10 μM for VRN5 (Methods).When applied to the cell cytoplasm, the high sensitivity of SlimVar was also able to establish an upper bound to the fluorescent signal for both VIN3 and VRN5 marginally above the ColFRI negative control levels, equivalent to a concentration approximately four orders of magnitude smaller than those measured in the nucleus.
Using SlimVar, highly motile fluorescent foci for both VIN3 and VRN5 could be tracked for up to ~20 consecutive image frames before photobleaching of nuclear contents occurred (Supplementary Videos 1, 2).Using bespoke tracking software (Methods) we found that these nucleoplasm-localized protein assemblies were largely excluded from the nucleolus, evident as dark regions of effective diameter 3-6 µm (Fig. 1c).In larger nuclei, the centres of nucleoli also appeared to exhibit weak VIN3 and VRN5 localization (Fig. 1b,c).We detected 10-40 tracks per nucleus while no tracks were detected either in the negative control or for VIN3 at pre-or post-vernalized timepoints (Fig. 2b).The number density of VRN5 tracks increased by about a third over the vernalization time course (0.55 ± 0.05 and 0.71 ± 0.05 μm -2 before cold and 14 days post-cold respectively: N=62, p=0.0042|*).Specifically, the mean number of VRN5 tracks per nucleus increased after the onset of vernalization and this increase was maintained after return to warm (Fig. 2b).We characterised the average characteristic brightness of a single molecule of the corresponding fluorescent protein (Methods) to determine the bleach-corrected stoichiometry of each track, with which we identified the number of labelled VIN3 or VRN5 molecules in each detected assembly.We found that both VIN3 and VRN5 exhibited broad stoichiometry distributions from a few molecules up to several tens of molecules for individual assemblies (Fig. 2c) and that the average stoichiometry increased with time duration spent in the cold during vernalization.For VIN3, the mean stoichiometry was 12.0 ± 0.4 molecules at V2W, increasing to 18.6 ± 0.5 molecules at V6W (N=1,988, p<0.001|**).For VRN5, assemblies were found to be well developed prior to vernalization (mean 18.5 ± 0.6 molecules at NV).However, there was an increase in stoichiometry during vernalization which persisted after the return to warm conditions (mean of 24.4 ± 0.9 molecules at V6W+T14; N=1,626, p<0.001|**).The greatest change occurred during the intermediate stages of vernalization between V2W and V6W (17.4 ± 0.7 to 23.4 ± 0.6 molecules, N=1,928, p<0.001|**).About ~1% of the nuclear protein copy number was detected in tracks at the fixed focal z-position, which compares well with our estimate, based on the theoretical depth-of-field, that ~4% of the mean nuclear volume is in sharp focus in each frame.

Multimolecular assemblies of VIN3 and VRN5 contain multiples of two molecules
The stoichiometry distributions each show a series of periodic peaks (Fig. 3) that are revealed when represented as a kernel density estimate, a method which objectifies the equivalent histogram bin width used.If the assemblies represented have a common oligomeric structure, the characteristic peak spacing is then equivalent to the number of molecules associated with a physical subunit of the assembly [50].We developed a new analysis method to discriminate this periodicity using the most common nearest-neighbour peak intervals.We verified it using realistic statistical simulations (Methods) and experimental data from standard LacI tetramers in vivo [51] (Supplementary Figs 6 and 7).Both VIN3 and VRN5 exhibit neighbouring peaks in their stoichiometry distributions which are separated by two molecules (Fig 3 insets).Performing identical SlimVar imaging on the VIN3-SYFP2 line indicated comparable levels of VIN3 nuclear protein copy number when corrected for the unlabelled VIN3 present using qPCR (Methods), but indicated stoichiometry values, and a characteristic peak spacing of stoichiometry, of approximately half that observed in the VIN3-GFP line (Supplementary Fig. 2), suggesting that the measured 2-molecule periodicity is unlikely to be an effect driven by dimerization of the fluorescent protein tag as reported in bacterial studies [52].Insets: the number of molecules in this subunit can be estimated from the most common spacing between neighbouring peaks in each stoichiometry distribution.The threshold above which a null (aperiodic) distribution can be rejected is the 95 th percentile fraction of intervals (grey trace) output from simulated random stoichiometry (Methods).The most common interval is given by the modal kernel density estimate ± sem above the null threshold (VIN3-GFP: V2W, 1.9 ± 0.3; V6W, 2.2 ± 0.3.VRN5-YFP: NV, 1.9 ± 0.4; V2W, 2.2 ± 0.4; V6W, 2.0 ± 0.3; V6W+T14, 2.0 ± 0.4).The periodic unit in each of these cases is consistent only with an assembly subunit of 2 molecules of either VIN3-GFP or VRN5-YFP.

Mobility of larger VIN3 and VRN5 assemblies matches that of FLC during cold exposure
We estimated the microscopic diffusion coefficient D for each detected track by calculating the gradient to the initial portion of its corresponding mean square displacement (see Methods).We found that VIN3 assemblies became significantly less mobile between two and six weeks of cold exposure (Fig. 4b, Supplementary Figure 3e).VRN5 assemblies exhibited a similar ~20% decrease in diffusivity during the same central stage of vernalization (Fig. 4c).For comparison with the diffusivity of FLC loci, we performed SlimVar imaging of a FLC-lacO/LacI-YFP line with 120 lacO copies integrated downstream of the FLC transgene [53].To obtain a qualitative indication of the proportion of VIN3 and VRN5 assemblies which might be bound to FLC and their stoichiometry, we analysed just tracks whose diffusion coefficient was comparable to that of a typical FLC locus (DFLC = 0.20 µm 2 s -1 , Supplementary Figure 6) within individual track measurement error (± 0.07 µm 2 s -1 ) such that D lies in the range 0.13 -0.27 µm 2 s -1 (Fig. 4a).This simple method of diffusivity matching [54] is particularly helpful when only single-label lines are available.The subset of VIN3 or VRN5 assemblies with diffusivity consistent with FLC (Table 1) have stoichiometries distributed similarly to the full cohort of VIN3 or VRN5 tracks (Fig. 3c) and show similar increases in stoichiometry over the course of vernalization for both VIN3 and VRN5.The median stoichiometry of each protein using this diffusivity matching was in the range 10-20 molecules per assembly, increasing with vernalization.In keeping with the increase in stoichiometry and lower diffusivity, the proportion of VIN3 assemblies that show slow FLC-like diffusion increases from around 11-15% during early stages of vernalization up to around 17-18% (Table 1).The fraction of FLC-like VRN5 assemblies present in the nucleus is already ~18% prior to vernalization and maintains this level throughout.However, given the vastly greater concentration of VEL protein assemblies compared to FLC loci, this proportion is unlikely to be representative of VEL localizations at FLC, which if present, would be a minority observable only by direct colocalization.

VRN5 assemblies of greater stoichiometry are enriched at FLC during and after vernalization
To directly track VRN5 at the FLC locus, we generated transgenic plants co-expressing VRN5 (fused to mScarlet-I [55], mScI) and FLC-lacO (via LacI-YFP) for dual-colour SlimVar.Our test for colocalization of VRN5 at FLC first required a reliable method to identify FLC without perturbing the mScarlet-I reporter for VRN5.For each nucleus, we first performed a rapid zstack using 514 nm laser excitation to screen for and localize bright, low mobility LacI-YFP foci consistent with FLC transgenes (Fig. 5a).At a chosen z-position containing one or more of these FLC candidates, we then tracked VRN5-mScI and LacI-YFP in a time course using alternating laser excitation (Methods); most images were dominated by the presence of unbound LacI-YFP foci, however, we measured a distinct subset bound to FLC (Fig. 5b) with a frequency of 2.3 ± 1.4 (mean ± sd) per nucleus in the meristem, matching the expectation of 2 FLC loci per nucleus within experimental error [53].This rose to 3.3 ± 1.7 per nucleus in cells toward the transition zone, consistent with more nuclei exhibiting additional pairs of FLC loci under genomic endoreduplication.Approximately 40% of detected FLC loci were colocalized with VRN5-mScI assemblies, though this proportion was constant across all vernalization times.This led us to question: might a putative memory element be conditional on the properties of the colocalized VRN5 assemblies, such as stoichiometry?Both the number and stoichiometry of VRN5 assemblies were considerably lower in this single-copy VRN5-mScI line than in the multiple-copy VRN5-YFP line (Fig. 2c).Like VRN5-YFP, however, these VRN5 assemblies showed stoichiometries that increased with vernalization independent of colocalization at FLC (Fig. 5d, grey).There was a proportionally far greater increase in stoichiometry of VRN5 assemblies colocalized at FLC, particularly after long cold exposure and after the return to warm (Fig. 5d, magenta).At timepoints prior to and after short cold exposure, the mobility of the colocalized VRN5 assemblies exceeded that typical of FLC loci by a factor of ~3, indicating extremely short-lived binding relative to the ~10 ms measurement timescale.After longer cold, the mobility of colocalized VRN5 more closely matched the typical diffusion rate of FLC loci, potentially pointing to a tighter interaction (Fig. 5e).Notably, the fraction of VRN5 not colocalized at FLC remained at the higher diffusivity independent of vernalization.An interesting, unexpected observation was the high rate of turnover; the colocalization dwell time of individual VRN5 assemblies at FLC remained transient, albeit increasing with vernalization, limited to a typical dwell time of ~100 ms.
In Lövkvist et al., 2021 [14], an assembly size of 17 was predicted as the best fit to observations of nucleating silencing marks, potentially reducing further in the case of maximal positive cooperativity.Although we do not know whether VIN3 assemblies colocalize specifically with FLC, and if so, the size of such an assembly, a large minority of observed VIN3 and VRN5 assemblies with sufficiently low mobility to interact with FLC clearly do exceed this size threshold during and after vernalization.In the case of VRN5, we show this is exceeded by a small minority of assemblies directly colocalized with FLC only after vernalization, though at a single gene copy, the maximal distinction with uncolocalized assemblies occurs at even lower thresholds of as low as 6-10 molecules.This finding suggests that if any lower bound size is required for positive feedback to become active in these colocalized assemblies, it must be very low.

Discussion
Here, we have developed a new optical microscopy pipeline-SlimVar-and applied it to image cell nuclei in Arabidopsis root tips.SlimVar enables direct detection of rapidly diffusing molecular assemblies in comparatively deep, multicellular samples, by sampling faster than the typical motion of nuclear assemblies, and by its optimising its sensitivity to single fluorescent proteins.Also, SlimVar does not require complex protocols for chemically conjugating target biomolecules.In animal tissues, studies using HILO [45] [56], lightsheet [57] [58] or lattice lightsheet [32] have previously demonstrated single-molecule tracking to a standard of 50 ms sampling at ~300 µm depth, or 10 ms at ~30 µm depth using dyes, which is comparable to SlimVar with fluorescent proteins.However, in plant tissues, single particle tracking at molecular sensitivity has been demonstrated with TIRF [59], [60], or VAEM [28], [30], [61], [62] only in the vicinity of a surface cell layer.SlimVar therefore advances the ability to track and count single-molecular assemblies in plants, and potentially in a range of tissues, to that of more complex existing microscopy technologies.
In achieving this speed, SlimVar trades off some of its 3D capability; the detection of foci is restricted to a limited depth of field much smaller than the nucleus.Nonetheless, it is capable of z-stacks as used here for systematic FLC detection, and it should be possible to extend oblique angle [48] or light sheet-based approaches to achieve rapid volumetric scans [63] and extended depth of field [64].The implementation of optimal index-matching and photon-efficient adaptive optics [33], [65] could also improve the range of accessible depths while also mitigating the required adjustments for each depth.Single-molecule experimental schemes related to SlimVar such as multiple-colour or photoconvertible labelling on the same target [66], may provide promising future avenues to test this turnover experimentally.
Using SlimVar, we demonstrated oligomeric assemblies of the PRC2 accessory proteins VIN3 and VRN5.VIN3 and VRN5 assemblies increase in stoichiometry above a demonstrable threshold during prolonged cold.Both sets of assemblies exhibit mobility signatures similar to FLC gene loci, with a subset of larger VRN5 assemblies clearly demonstrating FLC colocalization.The larger assemblies are therefore prime candidates to contribute to memory element function predicted in a hybrid model coupling protein self-assembly and histone modification [14].Our observation that adding more protein subunits preferentially results in a larger average assembly stoichiometry, rather than a greater number of assemblies, is itself consistent with a model of positive cooperativity.Taken together, these findings support the view that VIN3 and VRN5 assemblies mediate epigenetic memory over the extended vernalization cycle of several weeks (Fig 5f).
The model [14] makes no predictions as to the underpinning factors and mechanismsneither for protein self-assembly, nor dynamic exchange with the surrounding nucleoplasm.
The head-to-tail polymerization via the VEL domain [19] is likely involved, but the predominantly transient interactions between most individual VEL protein assemblies and FLC may suggest that the physical feedback processes are more complex than currently understood.Our current interpretation is that the VRN5 assemblies enriched at FLC, of sufficient size to satisfy the model, are most likely simple oligomers.Yet, we contemplate whether the very largest of these (~100 VRN5) are instead small dynamic phase separated condensates, related to those observed during transient overexpression of VEL proteins [16], [19], [67].If so, the collective, multivalent interactions characteristic of condensates might offer a longer, or otherwise more effective, residence time at FLC than we observed for typical molecular assemblies, and therefore a disproportionate contribution to epigenetic memory.SlimVar is an excellent tool capable and primed to further investigate these rare mechanistic events in vivo.
While pioneered for the Arabidopsis FLC system, these protein-mediated feedback processes may underpin Polycomb-based epigenetic memory common to all eukaryotic systems.Our study here, using highly interdisciplinary approaches, demonstrates the value of developing new high precision bioimaging technologies to tackle outstanding biological questions in epigenetic processing and memory storage.Although optimised for applications in Arabidopsis root tips, SlimVar is comparatively straightforward and costeffective to implement with minor modifications to a standard inverted optical microscope, and as such may have future applications in democratizing functional bioimaging research at the molecular scale inside a range of living tissues.

Plant material and growth conditions
The VIN3-GFP and VRN5-YFP lines have previously been described [8], [16].To generate the VIN3-SYFP2 line, the GFP of the pENTR pVIN3::VIN3-GFP construct [68] was replaced by SYFP2 by seamless mega-primer cloning.VIN3-SYFP2 was cloned to the SLJ destination vector, a derivative of SLJ755I5 [69] and transformed to Agrobacterium tumefaciens C58 by triparental mating.The transgenic VIN3-SYFP2 plant in ColFRI background was generated by floral dipping with Agrobacterium.To generate pENTR pVRN5::VRN5-mScarlet-I, the SYFP2 of VRN5-SYFP2 [67] was replaced by mScarlet-I by seamless mega-primer cloning.The VRN5-mScarlet-I was cloned to the SLJ destination vector, a derivative of SLJ6991 [69], transformed to Agrobacterium tumefaciens C58 and subsequently transferred into vrn5-8 FRI mutants as described above.All primers used in this study are listed in Supplementary Table 1.Transgene copy number was determined in T1 or T2 transformants by IDna Genetics (Norwich Research Park).To generate the plant co-expressing VRN5-mScarlet-I and FLC-lacO/LacI-YFP, the VRN5-mScarlet-I line was crossed into the FLC-lacO/lacI-YFP line [53] and was selected by antibiotics.
All seeds were surface-sterilized and sown on 100 mm growth plates containing Murashige and Skoog (MS) medium (Duchefa) with 1 wt.% agar (Difco Bacto) without sucrose.The plates were sealed with Micropore tape (3M) and kept at 4°C in the dark for 2-3 days to stratify the seeds.Plates were racked vertically in the growth chamber in warm conditions (16 h light/ 8 h dark with constant 22 ± 2°C, 60 ± 10% Relative Humidity (RH)) for 7 days.Non-vernalized timepoints (NV) were then imaged on the final day.For all other timepoints, plants were grown in warm conditions for 7 days as above and then were transferred to cold conditions (8h light / 16h dark, 5 ± 1°C, 50 ± 30 % RH) to vernalize for either 2 or 6 weeks (V2W and V6W respectively).Following vernalization, a subset of plates was returned to warm conditions for an additional period of either 7 or 14 days (V6W+T7 or V6W+T14).Plates for warm timepoints (NV, V6W+T7, V6W+T14) were handled at room temperature, while those imaged at cold timepoints (V2W and V6W) were transferred on ice to a 4°C cold room for slide preparation to avoid temperature spikes affecting the fragile vernalized state.All VIN3 imaging was performed in daylight hours 4-7 to align with the diurnal maximum in VIN3 expression [7].At least 3 independent vernalization courses were grown for each line and timepoint (Supplementary Table 2).
Immediately prior to imaging the FLC-lacO/LacI-YFP lines (single colour or dually labelled with VRN5-mScarlet-I), LacI-YFP was induced by placing opened growth plates next to a bath of 0.5% EtOH at 25⁰C in an airtight container for 2.5 h.This resulted in an optimal amount of LacI-YFP expression in the meristem for SlimVar imaging without spot-bleaching, below the near-saturated induction level (1.5-2% EtOH for 1.5 h [53]) used for confocal study (Supplementary Figure 6a).

RNA expression analysis
Total RNA was extracted using the phenol method as previously described [8], [70].Genomic DNA was removed from the TURBO DNA-free kit (Invitrogen, AM1907) before reverse transcription with SuperScript IV reverse transcriptase (Invitrogen, 18090050) and gene-specific primers.Quantitative PCR analysis was performed on a LightCycler480 II (Roche).Target gene expression was normalized by PP2A (AT1G13320) and UBC (AT5G25760).All primers used in this study are listed in Supplementary Table 1.

Preparation of samples for imaging
Identical slides were prepared for confocal, Airyscan or SlimVar imaging [45].Briefly, GeneFrames (Thermo Scientific, AB0578) were affixed to standard slides (VWR) and filled with MS medium plus 1 wt.% agarose to produce agar pads.Where necessary due to seedling size, the terminal >10mm of the primary root of each plant was excised using a razor.Root tips were laid on each agar pad with tweezers.Liquid MS media was applied to exclude air and each slide was sealed with a plasma-cleaned #1.5 coverslip (VWR).Each slide was imaged within <1h.

Confocal and Airyscan imaging
Confocal imaging was performed on a Zeiss LSM880 microscope equipped with argon ion laser and Plan-Apochromat 63×/1.40NA oil objective lens (Zeiss).Samples were illuminated at 488 or 514 nm (GFP or YFP/SYFP2 channels) respectively, and the emission detected in wavebands 490-550 nm or 518-550 nm respectively [45].The root tip confocal images (Fig 1b, Supplementary Figure 1, Supplementary Figure 6a-d) were acquired as z-stacks over ≤3 z-slices of 1000 × 1000 pixels at 0.6× zoom factor, at 1.5 μm intervals using 20 mW excitation power.Slices were postprocessed in FIJI/ImageJ [71] with a 2D median filter (0.2 μm filter size) to suppress noise before performing a mean z-projection.

SlimVar imaging
The SlimVar platform was adapted from a Slimfield microscope consisting of an objectivelens-based total internal reflection fluorescence setup, custom-built from benchtop optical components (Thorlabs).A general scheme of Slimfield microscopy is available [40], [41], [72] with key terms defined in Table 2 [73].
With increasing acquisition depth, refractive index mismatch between immersion oil and aqueous sample is a key challenge [74] due to spherical aberration and excitation beam deviations.This can be largely avoided with expensive water or silicone immersion objective lenses specialised for TIRF, but to demonstrate affordable and accessible performance, we used a widely available oil immersion lens: NA 1.49 Apo TIRF 100× oil immersion objective lens (Nikon).Single molecule sensitivity was afforded by a fast sCMOS camera (Teledyne Photometrics Prime95B, 12-bit 'Sensitivity' gain).A 580 nm longpass beamsplitter enabled simultaneous detection of GFP (525/50 nm centre/bandwidth emission filter) or YFP (550/25 nm) in a green/yellow channel, and mScarletI (594/25 nm) in a red channel.The total magnification was ~200× giving an oversampled pixel width of 53 nm.
Continuous wave lasers (Coherent OBIS) delivered Gaussian beams (TEM00) at 488 nm, 514 nm and 561 nm that were circularised by quarter waveplate [75], de-expanded, then focused at the objective back aperture to collimate them through the sample.SlimVar featured an oblique beam angle similar to HILO [48], PaTCH [76] and VAEM [48], [49].To optimise background contrast for Arabidopsis root tips, a four-step calibration was performed: i) the second convex lens in the expansion telescope was mounted on a lateral translation stage with a high-precision micrometer; this shift (up to 2.3 mm) generated an equivalent lateral displacement of the beam at the 6 mm diameter objective back aperture, thereby tilting the beam away from the optic axis at the sample.The beam delivery angle was calibrated for lens micrometer position following [78] and set to 50° ± 3° from normal incidence in oil, corresponding to 60° ± 5° in water by Snell's Law, with minimal coupling into the evanescent field at the coverslip.ii) a field stop was placed in the conjugate plane upstream of the first telescope lens.Rather than thinning the beam with rectangular slits to further suppress background [49], we chose to maintain a circular beam; this illuminated the full depth of each nucleus for representative estimates of nuclear protein copy and efficient screening of FLC loci.The stop was tightened to crop the beam (Gaussian 1/e 2 diameter of 17 μm) to between 4-9 μm cross-section.The field then approximated an ellipse of 1 × 3 1/2 the chosen beam crosssection, or 4-16 µm diameter, uniformly illuminated at a power density of 1-5 kW cm -2 , which matched the saturation excitation intensity of the fluorescent proteins.
iii) a pair of mirrors were used to incline the beam ~5 mrad from the optic axis at the objective back aperture, to compensate the beam's mismatch-dependent lateral deflection of ~11 µm at the 20 µm calibration depth.iv) the objective correction collar was adjusted to compensate spherical aberration at the calibration depth for a minimal FWHM of detected foci ~170 nm (Fig. 1d) and a localization precision [79] of 40 nm (Supplementary Fig. 8) within live roots.
The net increase in median signal-to-noise ratio (when normalised by the square root of exposure time × excitation power) due to the above beam angle optimisation was in the range 2.5-3.3 (n = 500 foci).
Nuclei were identified in brightfield to find best focus-typically the nucleolar midbodywithout any photobleaching, and were captured for manual segmentation.Fluorescence acquisition settings in green, yellow and/or red channels were pre-optimised to avoid initial camera saturation and to ensure detection of individual tracks of molecular brightness for GFP, SYFP2 / YFP, and mScarlet-I respectively.Three exposure times were used: i) a 'representative' 20 ms providing a low-bias detection of all particles down to single molecules, ii) 'fast' 2 ms to ensure robustness of tracking fidelity and mobility measurements of only the brightest assemblies, and iii) a 'balanced' 10 ms, also capable of single molecule detection.Fast acquisitions captured the subset of assemblies with fourfold (upper quartile) mean stoichiometry at a tripled detection rate due to the faster sampling bandwidth.Results shown derive from data collected with the representative (VRN5-YFP and VIN3-GFP lines) or balanced mode (VIN3-SYFP2, FLC-lacO/LacI-YFP and VRN5-mScarlet-I lines) unless otherwise stated.These settings were fixed to minimize systematic variation in the characteristic single molecule brightness (Supplementary Fig. 5).
The region of interest spanned ≤300 rows (16 µm) giving a readout time of 2.7 ms per frame, and sampling rates of 44-217 fps.Lasers were triggered in each frame by the camera in 'All Rows' mode to provide global shuttering without extraneous photobleaching.
For dual colour colocalization, 514 nm and 561 nm lasers were excited in alternating frames (ALEX) to facilitate distinct signals for each reporter.To capture photoblinking, the number of frames in each sequence was set to >50× the photobleaching decay constant.

Metric / Object Definition
Focus (foci) A spot-like local intensity maximum in a single frame, which corresponds to a localized group of labelled molecules.Associated properties include spatial location, total intensity, and signal-to-noise ratio.Track A set of foci in adjacent frames which are spatially close enough to form a trajectory.

Characteristic singlemolecule brightness
The average number of pixel counts in foci associated with a single fluorescent reporter molecule (e.g.GFP), under a fixed imaging condition.Equivalent to the modal step size in intensity for tracks in the final stage of photobleaching.

Integrated nuclear intensity
The total fluorescence intensity of an entire nuclear segment in pixel counts, normalised by the characteristic single molecule brightness.

Nuclear protein copy number
The average number of molecules in a nucleus, as estimated from the increase in integrated nuclear intensity above negative control.

Stoichiometry
The number of labelled molecules in a track, as estimated by dividing the track's initial intensity by the characteristic single-molecule brightness.

Periodicity
The number of labelled molecules in a repeat unit within tracked objects, as estimated by the consistent stoichiometry intervals between nearestneighbour peaks in the stoichiometry distribution.Diffusion coefficient Measure of the random microscopic motion of a track based on the increase in its (mean-squared) displacement over time.
Table 2. Definitions of analysis metrics for single particle tracking.

Protein tracking and intensity analysis
Fluorescent foci in the SlimVar microscope's depth of field were extracted using ADEMScode software [42] with a Gaussian masking algorithm and linked into tracks.The centroid of each focus was localized to super-resolved, subpixel precision [79], typically 40 nm (Supplementary Fig. 8) consistent with best Slimfield performance [80].
The probability of random overlap between similar foci (leading to spurious summation of intensities) was estimated by modelling the probability of nearest neighbour distances in a random Poisson process falling within the widefield localization precision [72].For the track densities in this study, the estimated fraction of randomly overlapping tracks is no greater than 10%.Analyses were restricted to the image region of effectively uniform laser illumination (80% ± 9% s.d.peak intensity) no greater than 190 × 300 pixels (10 × 16 μm).
Fluorescence sequences were cropped to specify individual nuclei, using masks manually segmented from brightfield images.The 2D diffusion coefficient of each track, , was estimated according to a random walk model as a quarter of the rate of increase of the mean-squared displacement with lag time [81].The intensity of each focus was calculated by integrating the pixel value intensity within 5 pixels of each focus and subtracting a background level averaged over the remainder of a 17 × 17-pixel window.The initial intensity of each track was determined by backward extrapolation of the intensities of the first 5 foci to a virgin timepoint prior to photobleaching.The characteristic single molecule brightness corresponding to each fluorescent reporter was determined based on the Chung-Kennedy-filtered [82] terminal intensity of tracks in each acquisition (Supplementary Fig. 5).

Stoichiometry periodicity analysis
To calculate stoichiometry periodicity, first the stoichiometries of all tracks within each nucleus were aggregated across nuclei in a dataset (genotype and vernalization status), then represented as a kernel density distribution.The upper limit for the kernel width was the empirical range on the characteristic single molecule brightness of 0.6 molecules [40]; to avoid oversmoothing, we used the standard deviation of 0.3 molecules.Peaks in this distribution were detected using the MATLAB findpeaks function, and the intervals between nearest neighbour peaks were calculated.The uncertainty in peak-to-peak interval was estimated as the single molecule uncertainty of 0.6 molecules multiplied by the square root of the mean stoichiometry, divided by the square root of the number of interpolated intervals [76].To suppress noise and spurious intra-peak sampling, all peak intervals smaller than the interval uncertainty were discarded.A second kernel density estimate was calculated over the remaining peak intervals, with the interval uncertainty as the kernel width.This curve describes the distribution of peak intervals in the stoichiometry as shown in Fig. 3 (insets) and Supplementary Fig. 3d.The modal value of this interval distribution was reported as the predominant periodicity of assemblies in each dataset.This method of estimating periodicity was verified as independent of the mean stoichiometry using simulated positive control data drawn from noisy Poisson-distributed multiples of an oligomeric ground truth [73].This analysis reproduced the expectation that the minimum number of tracks required for sufficient peak sampling, and therefore the limit of periodicity detection, scales with the square root of the mean stoichiometry.To demonstrate a negative control, 100 aperiodic sets of 10 4 stoichiometry values, uniformly distributed at random between 1-30 molecules, were generated using the randperm MATLAB function and processed to generate a set of 100 independent interval distribution curves each corresponding to null periodicity.The 95 th percentile fraction of peak intervals was calculated at each interval size to generate a null curve, below which test data could no longer be considered periodic.The uncertainty in the reported modal peak interval was estimated as the standard error on the mean (sem) of the peak intervals falling within the range above the null threshold line.The periodicity analysis was validated in vivo using the gold standard of tetrameric LacI-YFP [51] detected in the FLC-lacO/LacI-YFP lines (Supplementary Fig. 7b).
Raw estimates for the total number of molecules in each nucleus were extracted with an ImageJ macro following [42].The pixel values were integrated within each nuclear segment, then normalized by the characteristic single molecule brightness to give an integrated nuclear intensity in molecular equivalents (Fig. 2a).These values did not account for the additive contribution from autofluorescence background, which we estimated using the corresponding unlabelled control line, ColFRI.The nuclear protein copy number of each labelled dataset was refined to exclude autofluorescence by taking the difference between mean integrated nuclear intensities of the labelled dataset and unlabelled control, adjusted in proportion to the ratio of mean areas of nuclear segments.The negative control was much brighter in green acquisitions (15,000 ± 1500 GFP equivalents vs. 3,800 ± 400 YFP/SYFP2; mean ± sem, N=33, N=27 respectively).The nucleoplasmic concentrations were estimated by dividing each nuclear protein copy number by the mean nuclear volume (assuming prolate spheroidal nuclei aligned in the image plane) and multiplying by Avogadro's number.

Colocalization imaging
For each nucleus, a z-stack was performed, first with brightfield imaging to ensure alignment, then with 514 nm wavelength SlimVar excitation at 10 ms/frame exposure to track FLC loci via LacI-YFP.This z-range extended from the highest to lowest surface of each nucleus with respect to the coverslip surface, and was divided into equally spaced intervals of 280-360 nm.During each stack, the z-position of the image frame (denoted I*) containing the brightest LacI-YFP focus was noted and subsequently revisited (Fig. 5a).We then performed a dual-colour SlimVar acquisition, 10 ms exposure time alternating between 561 nm and 514 nm excitation wavelengths.The total duration of the fluorescent z-stack and alternating excitation acquisitions for each nucleus was ≤15 s.We estimate that the maximum displacement of FLC loci during the 15 s period was within 180 nm, less than our optical resolution limit and consistent with previous observations [83] (Supplementary Fig. 8d), thus in effect immobile over this timescale.
During post-processing, the two colour channels were spatially registered to sub-pixel precision using affine transforms generated from SlimVar images of 200 nm diameter fluorescent beads in vitro.Then both yellow and red channel image sequences were tracked independently (Fig 5b) to generate lists of LacI-YFP and VRN5-mScI tracks respectively.To account for FLC candidates observed in the z-stack but photobleached prior to the alternating acquisition, a copy of the alternating sequence was generated in which each yellow channel image simply comprised I*.This copy was also tracked, and its list of LacI-YFP tracks appended to the list from the original sequence.To exclude false positive detections of FLC due to free LacI-YFP, we then selected only slow and bright yellow tracks.Our earlier observations suggest a maximum of 8 FLC loci in the entire nucleus (Supplementary Fig. 6).On that basis, we selected the LacI-YFP tracks whose diffusivity was ≤ DFLC (Fig 4a) and from these, retained the 8 brightest (or all if fewer than 8) tracks in each nucleus.These selected FLC tracks, typically 1-4 per nucleus (Fig 5b, white circles), were run through colocalization analysis with the corresponding VRN5-mScI tracks (Fig 5b, magenta traces) using a previously reported algorithm [72].
Briefly, VRN5 and FLC tracks were deemed colocalized if they met an intensity overlap condition [84] of at least 50% (effectively a lateral distance of ~3 pixels or one widefield localization precision) and remained within a distance of 7 pixels (twice the widefield localization precision) for ≥3 frames.The high numerical aperture and short depth of field ensured an axial precision better than <220 nm FWHM for all colocalizations.The likelihood of false positive overlaps between VRN5 tracks, and the likelihood of false positive colocalizations (an FLC locus being colocalized with a VRN5 assembly by random chance) were both <5%, based on the average initial number density of tracks in each frame (5.2 VRN5 and 3.3 candidate FLC) distributed in the nucleoplasm under random point statistics [85].

Statistics and samples
All pairwise comparisons used the non-parametric Brunner-Munzel test.Sample size and significance are indicated alongside each result.Investigators were not blinded and each acquisition was independent.We predetermined a target sample size of >24 cells total per line per condition, sufficiently powered to detect changes of <1 s.d. in each of the five test variables (number of tracks, nuclear protein copy number, stoichiometry, periodicity, diffusivity) at a Bonferroni-adjusted significance level of α = 0.05/5 = 0.01.We planned biological replicates of >3 independent cycles of growth and vernalization, with >3 roots per cycle and >3 cells per root.In the dual line, >10 cells per condition sufficed for estimates of track number and stoichiometry disaggregated by FLC colocalization.Technical replicates were identified with tracks detected within each nuclear segment.Actual numbers of replicates analysed are detailed in Supplementary Table 2.

Supplementary Figure 2. RNA expression analysis and characterization of VIN3-SYFP2.
Expression level of VIN3 (A), fluorescent protein (B) in VIN3-GFP, VIN3-SYFP2 and ColFRI seedlings.(C) Expression level of spliced FLC in seedlings after 2 weeks and 6 weeks vernalized cold seedlings relative to non-vernalized seedlings.(A-C) Data are presented as mean ± sem (N=9) from three biologically independent experiments.(D) Based on the assumption that protein levels scale proportionally with RNA levels, we derive a correction factor to infer the total number of VIN3 proteins in the yellow line from the detected number of VIN3-SYFP2 proteins.This factor can be estimated by taking the ratio of mRNA expression levels of VIN3 to SYFP2, then normalising this by the ratio of VIN3 to GFP expression in the green line, which lacks endogenous VIN3.The error in the correction factor is propagated directly from the sem of the independent measurements in panels A-B.
Given the uniform rate of increase of VIN3 expression in all lines regardless of labelling, the correction factor is expected to be constant over time.Taking an inverse-variance-weighted average over the three timepoints, the correction factor used in this work is 2.0 ± 0.7, which is consistent with the expected genotype of the yellow line: a single exogenous gene copy of VIN3-SYFP2 and a single native gene copy of VIN3.

Figure 1 .
Figure 1.SlimVar resolves single particle dynamics of VIN3 and VRN5 during cold exposure of root tips.a) The optical scheme for single-or dual-colour excitation with beam delivery at a steep angle into a live root sitting on an agarose pad.The intersection of the focal plane and the excitation beam defines a thin detection volume, closely matched to the nuclear dimensions up to 30 µm deep into live tissue.High numerical aperture, high laser intensity and a photon-efficient sensor all contribute to the speed and sensitivity required to detect single molecule dynamics.b) Projections of confocal z-stacks of live whole root tips of VIN3-GFP and VRN5-YFP after 6 weeks of cold; insets (interpolated) show detail of the labelled protein signals, which are consistently localized to the nucleoplasm, but not the nucleolus.The nuclear patterning of VIN3 and VRN5 was typically either round or lens-shaped depending on cell type and cell wall attachment, with a median length 7.8 μm (interquartile range IQR: 5.7-10.3μm, N=571), and median aspect ratio of 1.16 (IQR: 1.06-2.10),comparable to other nuclear reporters [46].Acquisition time is 35s.c,d) Airyscan images of VRN5-YFP and VIN3-GFP after 2 weeks' cold indicating heterogeneity spatial distribution; c) maximum intensity projection of three z-slices, averaged over three consecutive timepoints; d) sensitivity to the dynamics over the same three timepoints represented by contrast between median (cyan) and standard deviation (magenta) of the pixel values over time, the latter indicating faster detected motion of VRN5 compared to VIN3 assemblies.e-h) SlimVar microscopy of VRN5-YFP before vernalization; e) brightfield, f) initial frame in YFP channel prior to photobleaching; g) early stage photobleaching revealing distinct assemblies (mean projection of frames 4-6); h) SlimVar resolves assemblies of differential mobility on 20 ms timescales shown here as distinct slow-(cyan) and fast-moving (magenta) objects, represented by median and standard deviation projections over three frames.i) Automated tracking analysis (Methods) identifies assemblies to super-resolved localization precision and estimates the corresponding stoichiometry and mobility.Individual tracks are shown as arrows with one vertex per timepoint.

Figure 2 .
Figure 2. Cold exposure causes VIN3 and VRN5 to form assemblies each of larger, more broadly distributed stoichiometry, and a relatively modest increase in the number of assemblies.a) Distributions of integrated nuclear intensity (total number of labelled molecules per nucleus prior to correcting for autofluorescence) of individual nuclei aggregated across cell type at different timepoints before, during and after vernalization, for exogeneous VIN3-GFP and VRN5-YFP.The nuclear protein copy number (number of labelled molecules excluding any associated with autofluorescence) is then calculated as the net excess above the mean level of the negative control ColFRI (horizontal line), which represents autofluorescence background.Bar, box and whisker denote median, interquartile range (IQR) and 1.5× IQR respectively; cross: mean ± sem.The abundance of VIN3 protein was insignificant prior to vernalization (Brunner-Munzel (BM) test vs ColFRI, N=66, p=0.11 |ns: not significant at adjusted p<0.01).However, the VIN3 nuclear protein copy number increases sharply to ~28,000 ± 3,700 molecules above control after 2 weeks cold (N=64, p=0.0031 |*), and peaks at ~44,000 ± 4,700 after 6 weeks cold (N=83, p<0.001 |**).Following transfer to warm conditions, VIN3-GFP levels reduce to ~3,200 ± 1,600 molecules within 7 days (N=74, p=0.04 |ns).VRN5 levels increase between non-vernalized and post-vernalized timepoints from ~110,000 ± 23,000 to ~190,000 ± 37,000 molecules (N=94, p=0.0089|*).b) Distributions of the number of tracks per nucleus (binned in columns of width 2 for clarity).The number of VRN5 tracks increases initially (NV: 20.8 ± 1.9 up to 26.8 ± 1.6 tracks per nucleus at 2 weeks cold; N=86, p=0.0054|*) but stops increasing after 6 weeks' cold (27.0 ± 1.5 and 26.2 ± 2.6 tracks per nucleus at 6 weeks cold and 14 days post-cold respectively; N=94, p=0.80|ns); c) Collated distributions of stoichiometry (number of labelled molecules per nuclear assembly) of individual tracks at different timepoints before, during and after vernalization; nt: no tracks detected.

Figure 3 .
Figure 3. VIN3 and VRN5 assemblies exhibit a two-molecule spacing in their stoichiometry distributions.The number of labelled molecules in each assembly shows consistent periodicity of a) VRN5-YFP and b) VIN3-GFP.The fixed kernel width of 0.6 molecules corresponds to the rms error in the observation of a single molecule.Insets: the number of molecules in this subunit can be estimated from the most common spacing between neighbouring peaks in each stoichiometry distribution.The threshold above which a null (aperiodic) distribution can be rejected is the 95 th percentile fraction of intervals (grey trace) output from simulated random stoichiometry (Methods).The most common interval is given by the modal kernel density estimate ± sem above the null threshold (VIN3-GFP: V2W, 1.9 ± 0.3; V6W, 2.2 ± 0.3.VRN5-YFP: NV, 1.9 ± 0.4; V2W, 2.2 ± 0.4; V6W, 2.0 ± 0.3; V6W+T14, 2.0 ± 0.4).The periodic unit in each of these cases is consistent only with an assembly subunit of 2 molecules of either VIN3-GFP or VRN5-YFP.
LacI-YFP tracks are shown in Fig 5b (yellow traces)
The characteristic single molecule brightness was 172 ± 15 (counts, mean ± sem) for VIN3-GFP, 126 ± 16 for VRN5-YFP and LacI-YFP, 131 ± 25 for VIN3-SYFP2 and 140 ± 18 for VRN5-mScarlet-I.These values were used for internal calibration of each track stoichiometry (the initial track intensity divided by this single molecule brightness) and each dataset periodicity and nuclear protein copy number.For comparison, in vitro values for characteristic single molecule brightness were obtained from E. coli recombinant fluorescent proteins at a coverslip surface under the same SlimVar imaging conditions.