Nanoclustering of a plant transcription factor enables strong yet specific DNA binding Authors: Kasper Arfman, Bas P.J. Janssen, Robin Romein, Sophie van den Boom, Maarten van der Woude, Lucas Jansen, Maartje Rademaker, Jorge Hernández-García, João Jacob Ramalho, Melissa Dipp-Álvarez, Martijn de Roij, Jasper Lamers, Johannes Hohlbein, Jan Willem Borst, Dolf Weijers, Carlo P.M. van Mierlo, Joris Sprakel Published: 2026 ABSTRACT Transcription factors (TFs) are traditionally depicted as mono- or oligomeric units that recognize and bind specific DNA sequences to regulate transcription. Emerging evidence suggests that many TFs can undergo liquid phase separation, resulting in condensates that bind DNA through a different mechanism. For the auxin response factors (ARFs), canonical plant TFs, evidence for both scenarios exists, but which of these scenarios is operational in the plant nucleus is unclear. Here, we demonstrate using MpARF2 of Marchantia polymorpha that a third scenario is operational: MpARF2 forms nanoscopic clusters under physiological conditions. Nanoclusters combine DNA binding features that cannot be accessed by the other two scenarios: high-affinity, switch-like, and sequence-specific. Our results suggest nanoclustering as a mechanism that equips TFs with the DNA binding properties necessary for transcriptional regulation. INTRODUCTION Transcription factors (TFs) are proteins that bind to specific DNA sequences to regulate gene expression (1–3). The classical picture of TF binding is that of a monomeric or small oligomeric structure that binds to a specific DNA sequence through structural recognition of a DNA binding domain (DBD) (4–6) to activate or repress the transcription of a particular gene or set of genes (7). The other domains in a TF are typically not associated with DNA binding but act as modules for homo- or hetero-interactions, for example, to recruit accessory proteins needed for transcriptional regulation (8). A contrasting picture has recently begun to emerge: Many TFs are found to form clusters (i.e., higher-order assemblies) that exhibit enormous multivalency in their binding to DNA (9–19). These clusters are often described as biomolecular condensates formed through liquid-liquid phase separation (LLPS) (20, 21). The formation of multimeric—in some cases, even microscopic—structures is associated with the presence of intrinsically disordered regions (IDRs) (22–25) and can have a drastic impact on the DNA binding mechanism (26–28). While the monomeric picture relies on structural recognition between the DBD and a regulatory element in the DNA, liquid condensates bind DNA through a physical wetting process (26). Although functions for the formation of large TF structures, such as condensates, have been proposed (29), their physiological relevance and precise biological function remain elusive. These recent findings have led to a dichotomy. Studies that emphasize the classical picture, often from the viewpoint of the precise targeting of a DBD, do not suggest a need for the formation of higher-order structures. However, studies focusing on higher-order TF structures are beginning to highlight their prevalence but leave their function and relevance unanswered. Specifically, the in vivo occurrence of TF condensates is open to debate, as many observations rely on nonphysiological conditions, such as strong overexpression of the TF, to report on a process that is by nature strongly concentration dependent. To reconcile these perspectives, two key questions must be addressed: (i) what structural form do TFs take under native conditions in the cell nucleus, and (ii) by what mechanism do these structures bind genomic DNA? Here, we explore these questions in the context of the auxin response factors (ARFs), a canonical family of plant TFs (30–33). TFs in the ARF family, which are deeply conserved across all land plants and even their algal sister clades (34), regulate gene expression in response to the plant hormone auxin. In addition to a DBD, which encodes sequence-specific binding to auxin responsive elements (AuxREs) (5), many ARF proteins contain an intrinsically disordered middle region (MR) and a C-terminal Phox and Bem1 (PB1) multimerization domain. The DBD is reported to form dimers in solution and on DNA (35), and disrupting DBD-DBD interactions between ARFs results in strong defects (5), indicating that these interactions are essential for ARF function. The MR and PB1 domains both appear to play a role in the formation of higher-order structures (14). For ARFs, evidence exists to support both pictures outlined above. In line with the classical picture, binding of ARFs to AuxREs is specific (36) and appears in previous work to occur in a dimeric configuration (5, 35, 37). However, these conclusions are based on measurements performed on short double-stranded DNA (dsDNA) oligomers with two AuxREs in tandem (5, 37–39). These oligomers are designed to accommodate dimers, but their short length precludes the binding of higher-valency structures. Moreover, in a biological context, ARFs may interact with DNA regions flanking the recognition motif; however, the use of oligomeric DNA artificially eliminates the possibility of such nonspecific interactions. Therefore, oligomers cannot be used to determine whether the DNA binding structure is larger than a dimer. In contrast, ARFs can also form microscopic condensates in vivo (14, 40), with each condensate containing many protein molecules. By analogy to other condensate-forming TFs, this could result in completely different DNA binding mechanisms and affinities (26). To date, ARF condensates have only been observed in plants that either overexpress the ARF protein or contain a transgene expressing a fluorescently tagged copy of the ARF gene from a nonnative genomic location (14, 40). Several studies show that the formation of higher-order assemblies of other TFs is highly dependent on the degree of overexpression, with droplets only appearing when the overexpression is sufficiently strong (41–43). It is therefore important to consider expression levels that are as close as possible to the native ones to understand the physiological relevance of ARF condensates in transcriptional regulation. We use the repressor MpARF2 from the liverwort Marchantia polymorpha as a model (44–47) to explore the formation of higher-order assemblies of this TF and how these assemblies affect DNA binding. We find that, at near-native nuclear concentrations, MpARF2 forms nanoscopic clusters, both in vivo and in vitro. These nanoclusters are notably larger than a protein dimer but much smaller than the previously observed condensates and form a thermodynamic state distinct from phase-separated condensates. Although MpARF2 forms condensates in vitro, these appear only at protein concentrations well above the physiological range, and while condensates can bind DNA, their binding is sequence independent when a sufficiently long DNA template is used, which is incompatible with the biological function of TFs. Instead, we show how nanoclusters, which feature a size-limiting mechanism similar to micelles, turn the gradual and low-affinity binding of monomeric ARFs into a high-affinity switch-like binding transition in the physiological range, while retaining strong sequence specificity. We conclude that, at physiological concentrations, MpARF2 monomers and dimers may not bind DNA, and condensates do not exist, but that micelle-like nanoclusters are the relevant regulatory entity that equip ARFs with DNA binding properties beneficial for transcriptional regulation. RESULTS MpARF2 forms nanoclusters in vivo To explore the assembly state of our TF, MpARF2 (fig. S1), we first realize that the association of proteins is strongly concentration dependent. Hence, to arrive at an understanding of ARF assembly and DNA binding that is physiologically meaningful, we first need to define the native biological context by studying the appearance of MpARF2 in its cellular environment (Fig. 1A). MpARF2 forms nanoclusters at native expression levels in vivo. (A) M. polymorpha knock-in plant expressing MpARF2-mNeonGreen at native levels. (B) Confocal fluorescence image of an entire gemma showing the localization of MpARF2 in nuclei. Scale bar, 50 μm. (C) Deconvoluted confocal image of a single nucleus with nanoclusters under native expression of MpARF2. Scale bar, 1 μm. (D) mNeonGreen with a nuclear localization signal (NLS) shows no clustering when expressed under the same promoter as MpARF2. (E) The MpARF2 concentration in the cell nuclei (n = 225) is estimated at 2.2 ± 0.9 nM (mean ± SD, filled symbol) by means of fluorescence intensity calibration with purified mNeonGreen solutions of known concentration (open symbols). a.u., arbitrary unit. We make use of the recently developed genomic MpARF2-mNeonGreen knock-in lines (47). In these lines, a fluorescent mNeonGreen protein is fused to the C terminus of MpARF2 in its native genomic locus. Because increased or decreased MpARF2 levels show dramatic developmental defects (47), and MpARF2-mNeonGreen knock-in plants do not show developmental defects, we can infer (i) that the fusion protein is biologically active and (ii) that it accumulates at close-to native concentrations. We determined the transcription levels of MpARF2 in wild-type plants and the knock-in line using quantitative polymerase chain reaction (qPCR) and found that the knock-in line has approximately six times higher gene expression (fig. S2). Although this does not necessarily reflect protein levels directly, it suggests that protein levels in the knock-in line may be elevated relative to native conditions but remain within the same order of magnitude and lack any developmental phenotype (47). We imaged gemmae, the asexual propagules of M. polymorpha, in the knock-in line pMpARF2::MpARF2-mNeonGreen (Fig. 1B). Using confocal deconvolution microscopy, we determined that MpARF2 is present in the cell nuclei as small punctae against a background of low fluorescence intensity (Fig. 1C). In the deconvolved images, the punctae appear as diffraction-limited objects, implying that they are of nanoscopic dimensions (i.e., below 200 nm in diameter), and we observe that their fluorescence intensity is weak. To confirm that these distinct nanostructures are the result of TF assembly, we also imaged plants expressing the same fluorescent protein with a nuclear localization signal (i.e., pMpARF2::NLS-mNeonGreen). This nucleosolic mNeonGreen, which is not fused to a TF, yields a homogeneous image (Fig. 1D). We introduce the term nanocluster to refer to a nanoscopic protein assembly without implying any specific molecular identity. The small size and low brightness of the nanoclusters suggests that they contain a rather limited number of MpARF2 molecules. These assemblies are larger than monomers or dimers yet much smaller than the liquid-like droplets that are observed in overexpression lines of ARFs (40) or other TFs (48–50). The nanoclusters show limited mobility in the nucleus (movie S1), suggesting that they are anchored to immobile matrix components, likely genomic DNA. To perform meaningful in vitro experiments, an order-of-magnitude estimate of the nuclear MpARF2 concentration is required. To this end, we calibrated the optical response of our confocal fluorescence microscope by imaging solutions of purified mNeonGreen at known concentrations, yielding a linear calibration curve (Fig. 1E, open symbols). Using identical experimental settings, we then imaged nuclei in M. polymorpha gemma cells (Fig. 1B). From the average fluorescence intensity of 225 nuclei, we estimate the nuclear MpARF2 concentration in the knock-in line to be 2.2 ± 0.9 nM (Fig. 1E, closed symbol). This measurement provides an estimate of the order of magnitude of nuclear MpARF2 levels but should not be taken as an exact value. Optical artifacts, such as tissue scattering or fluorophore quenching, can lead to measurement inaccuracies. Moreover, protein levels likely vary across tissue types and developmental stages. In addition, as mentioned above, the knock-in line, while providing the closest-to-native construct possible, has higher MpARF2 transcript levels relative to wild-type plants; hence, it is possible that native concentrations are even lower than 2.2 nM. MpARF2 forms nanoclusters in vitro at physiological concentrations To explore the phase behavior of MpARF2, we purified recombinant MpARF2. We observed that large protein assemblies formed during purification, which were difficult to separate from monomeric protein (Fig. 2A). These observations indicate that MpARF2 has a strong intrinsic tendency to form larger structures. This tendency became more apparent when we performed dynamic light scattering (DLS) on the protein immediately after its elution from a size exclusion chromatography (SEC) column. The protein fraction that eluted as a monomer was diluted to 50 nM in the same buffer, and, within minutes, the solution reached a state consisting of a mixture of monomers and distinct nanoclusters of ~150 nm in diameter (Fig. 2B, solid black line). MpARF2 forms nanoclusters and condensates in vitro. (A) SEC chromatogram as the final step in the purification of MpARF2, after removal of the affinity/solubility tag MBP; mNeonGreen absorbs at 506 nm. (B) DLS size distributions of the MpARF2 monomer fraction after sequential salt treatments: starting with 1 M NaCl (solid black line) down to 150 mM (dashed line) and readjusting to 1 M (colored-line time series). (C and D) Confocal fluorescence images of MpARF2 nanoclusters (C) and condensates (D). Scale bars, 1 μm. (E) Centrifugation assay in which the protein concentration in the supernatant after centrifugation, csup, is measured as a function of the total protein concentration. The onset of condensate formation is estimated at csat ≈ 28 nM. (F) Assembly sizes, determined by DLS, show that condensates (100 nM) grow substantially, whereas nanoclusters at 20 nM remain nanoscopic in size. (G) DLS size distributions for domain-deletion variants of MpARF2. The MR is the main driver of clustering, the PB1 enhances clustering, and the DBD confers a defined cluster size. Of note, the size distribution is distinctly bimodal, which indicates that nanoclusters have a well-defined size, while intermediate-sized assemblies (i.e., between monomers and nanoclusters) are disfavored. The size distribution at high salt shows a small peak for monomers and a larger peak for nanoclusters. However, the DLS results are presented as intensity-weighted rather than number-weighted distributions. Because in light scattering, the scattered intensity is proportional to the diameter of the object to the sixth power, these distributions are heavily biased toward larger objects. Based on our data, we estimate that at high salt, the number ratio between monomers and nanoclusters is 108:1. Hence, at high salt, the solution is primarily monomeric. However, lowering the salinity to a more physiological level (150 mM NaCl) at 50 nM MpARF2 induced nanocluster formation to such an extent that the monomer population could no longer be detected by DLS (Fig. 2B, dashed line). To assess whether nanocluster formation is reversible, we induced clustering again by diluting the monomeric SEC fraction of MpARF2 in a low salt concentration (150 mM NaCl, 100 nM MpARF2), but this time, we subsequently diluted the sample back into high salt (1 M NaCl, 50 nM MpARF2) before the measurement. This revealed a gradual and partial dissociation of nanoclusters, but the signal of monomers was not recovered (Fig. 2B). The DLS experiments were performed at MpARF2 concentrations (50 nM) significantly higher than the established physiological range, since monomer signals could not be detected at much lower protein concentrations. To validate these findings at more physiologically relevant protein levels, we also performed fluorescence correlation spectroscopy (FCS). With this technique, we were able to determine the size distribution of MpARF2 at a concentration of 7 nM, comparable to the levels found in the nucleus of our knock-in lines. These measurements also revealed a clear coexistence of monomers and nanoclusters (fig. S3), which supports the physiological relevance of our observations. Because nanoclusters form rapidly and reversibly even at low concentrations, these data suggest that MpARF2 exists in a dynamic equilibrium between nanoclusters and low-order oligomers under physiologically relevant conditions. This implies that MpARF2 cannot be maintained as a pure monomer in vitro, but it also means that such a state is unlikely to occur naturally in vivo. Small oligomers and nanoclusters coexist almost unavoidably in solution, and this behavior mirrors our in vivo observations of small clusters superimposed on a diffuse background of weak fluorescence (Fig. 1C). Confocal microscopy on purified MpARF2 solutions confirms that at physiological protein concentrations, in the low nanomolar range, nanoclusters appear as diffraction-limited spots (Fig. 2C), which is consistent with our in vivo observations at native protein levels (Fig. 1C). However, at much higher concentrations, MpARF2 forms microscopic condensates, recognizable by their spherical shape and their ability to coalesce (Fig. 2D and movie S2), similar in appearance to what has been observed for AtARF19 in Arabidopsis thaliana (14, 40). Moreover, this is also consistent with the prediction by the FuzDrop model (51) that MpARF2 shows a 99.8% probability of spontaneous LLPS (fig. S4). MpARF2 nanoclusters are distinct from condensates AtARF19 has been shown to form microscopic condensates in plant cells at high expression levels (14), and its domain architecture is homologous to that of MpARF2. This raises the question whether the nanoclusters we observe both in vivo and in vitro are small condensates or a different type of structure. To determine the protein concentration at which condensates form, we performed a centrifugation assay that allows condensates to be gravitationally separated from nanoclusters and small oligomers. After centrifugation, we measured the protein concentration remaining in the supernatant, csup, as a function of the total protein concentration prior to centrifugation. In the absence of condensate formation, csup equals the total protein concentration. Thermodynamics dictates that condensates form as a first-order phase transition, the onset of which is marked by a threshold concentration, csat, at which the solution has become saturated with dissolved protein (52). We determine the formation of condensates as a distinct departure from the linear trend at csat ≈ 28 nM (Fig. 2E). Nanoclusters remain in the supernatant after centrifugation under subsaturated conditions, whereas they are depleted from solution once concentrations exceed csat (fig. S5). This suggests that nanoclusters phase-transition into condensates at csat and thus represent a distinct type of assembly. For condensates formed by LLPS, the thermodynamic equilibrium state would be one single condensate that minimizes the interfacial area between the two phases. Therefore, condensate droplets are expected to continuously grow by coalescence and Ostwald ripening (53). Using fluorescence imaging, we found that MpARF2 solutions above csat exhibit continuous condensate growth over time. However, below csat, where we exclusively observe nanoclusters, we detect no significant increase in fluorescence intensity over 24 hours (fig. S6). Our DLS data suggest that some nanocluster growth occurs in the subsaturated concentration regime, but also indicate that this growth takes place early (i.e., within 3 hours) and that no further detectable growth takes place during the remaining 21 hours (Fig. 2F). Such behavior is not to be expected for condensates, but is consistent with a process settling into an equilibrium state. Therefore, we conclude that the nanoclusters we observe below csat are not condensates. If these MpARF2 nanoclusters are not condensates, then what are they? To address this question, we investigated the nanoclustering mechanism using domain deletions. A complete loss of clustering was obtained for MpARF2 in which the intrinsically disordered MR was removed (ΔMR) (Fig. 2G, orange line). From this, it would appear that the MR plays a central role in nanocluster formation. In support of this hypothesis, we find that MR deletion increases the threshold concentration for condensation csat from 28 nM to 1.5 μM (fig. S7). Although these experiments reveal that the MR is required for clustering, these experiments cannot distinguish whether the MR serves mainly as a linker or whether it acts independently as a driver for self-assembly. To test this, we purified MpARF2-MR and investigated its phase behavior. Purified MR forms condensates on its own with csat ≈ 200 nM (fig. S8) and forms nanoclusters in subsaturated solutions as low as 15 nM (fig. S9). Therefore, we conclude that the MR domain is necessary and sufficient for nanoclusters to form at biologically meaningful protein concentrations. The Phox and Bem1 (PB1) domain mediates multimerization via self-complementary charged interactions (54). Deleting PB1 significantly reduces clustering, although some nanoclusters were still observed (Fig. 2G, purple line). Thus, PB1 appears to enhance the clustering that is driven by the MR. By providing additional multivalent interactions, PB1 likely creates local spots of increased MpARF2 concentration where cluster formation is nucleated. Based on these observations, one may conclude that MR and PB1 are responsible for nanocluster formation, with the DBD primarily effectuating DNA binding. However, our results suggest something different. Deleting the DBD domain leads to very strong clustering, exceeding that of the full-length protein, but these clusters lacked a defined size (Fig. 2G, yellow line). This finding could indicate that the DBD acts as a growth-limiting module: a behavior strongly reminiscent of micellization, where the combination of a phase separation tendency with a growth-limiting mechanism leads to the formation of nanoscopic structures of well-defined size, typically containing tens to hundreds of monomers. The notion of a micellar structure is consistent with our observations of a bimodal size distribution, as micelles coexist with monomers under thermodynamic equilibrium. The growth-limiting mechanism in micelles typically originates from the formation of an outward-facing corona, such as the charges on surfactant molecules. In this view, DBDs might form an outer shell that stabilizes the micelle-like nanoclusters. Supporting this idea, AlphaFold-Multimer predicts that MpARF2 pentamers (the largest supported multimer in this model) adopt a conical shape, with DBDs at the base and MRs at the tip (fig. S10). An arrangement of outward-facing DBDs could also be beneficial for the DNA binding function of TF nanoclusters. Another defining feature of micelles is the presence of a critical micelle concentration (CMC). The CMC marks a concentration threshold below which micelles are absent, and above which micelles form abruptly with their number increasing linearly with concentration. To test whether MpARF2 follows this description, we quantified cluster abundance as a function of protein concentration using single-particle tracking (SPT). MpARF2 nanoclustering commences abruptly at a critical concentration of ~0.5 nM, and the number of clusters scales linearly with concentration beyond this threshold (fig. S11). Together, these observations demonstrate that MpARF2 nanoclusters in subsaturated solutions have many of the defining features of a micelle. Higher-order MpARF2 assemblies gain DNA binding affinity but lose specificity Above, we have established that MpARF2 forms nanoclusters under physiological protein and salt concentrations, both in vitro and in vivo. However, it remains unclear how the formation of these structures affects DNA binding. Through single-molecule studies using immobilized oligos, the dissociation constant, Kd, of MpARF2-DBD to short oligomeric DNA with a tandem AuxRE was previously estimated at 61 nM (55). Because the nuclear concentration of MpARF2 that we measure is much lower, it appears that under physiological conditions in the nucleus, virtually no monomer or dimer would be bound to DNA. However, by featuring many DBDs in one entity, nanoclusters have the potential to interact with DNA with a high valency. In multivalent interactions, weak binding affinities of monomers are exponentially enhanced. AuxREs are sparse in genomic DNA. Hence, the number of specific DNA interactions that may form is likely not limited by cluster size, but rather by the prevalence of recognition elements in the DNA that are within reach of a single cluster. Clustering should therefore not substantially increase the number of specific DNA contacts that can form but greatly enhance the likelihood of nonspecific interactions. The DBDs of ARFs are highly sequence-specific (38). To achieve such specificity, nonspecific DNA interactions must be very weak, potentially negligible, compared with binding to an AuxRE. Hence, nonspecific interactions have little effect on monomeric DNA binding. However, for multimers, given that the binding affinity scales exponentially with valency, many nonspecific interactions may collectively contribute significantly to DNA binding. To test this hypothesis, we conducted electrophoretic mobility shift assay (EMSA) experiments in which protein-DNA mixtures are separated by gel electrophoresis, allowing free and bound states to be spatially resolved and quantified using fluorescence imaging (Fig. 3A and fig. S12). The experimentally determined bound fraction of DNA, θDNA, plotted as a function of protein concentration, produces a binding curve from which an effective dissociation constant, Kd∗, of the interaction may be estimated. The true Kd may be overestimated in EMSA because the system is not at equilibrium during electrophoresis: Bound protein-DNA complexes can dissociate while the gel is running, but reassociation is limited as the unbound species become physically separated. EMSA was originally developed to study the lac repressor TF (56), which does not contain an IDR. However, it is predicted that ~90% of TFs contain IDRs (57), and the presence of an IDR is one of the hallmark features of phase-separating proteins (58). Protein clustering may complicate the EMSA experiment, as large protein assemblies bound to DNA may not be able to penetrate the nanoporous gel and thus become trapped in the loading well. Fluorescence signals coming from wells may therefore be unavoidable, but excluding them from analysis, as is common practice, would introduce a bias. Clustering and DNA accessibility define DNA binding. (A) Result of an EMSA assay to quantify binding of MpARF2-DBD to oligomeric dsDNA, showing both DNA (green) and protein fluorescence (magenta, log color scale). (B) DNA binding curves of full-length MpARF2 (blue) and its DBD (gray) to 19-bp DNA containing the IR7 recognition sequence (filled symbols) or lacking the TGTCNN motif (open symbols). (C) Approach for obtaining long DNA templates containing IR7 motifs surrounded by nonspecific DNA. A. thaliana gDNA fragments without TGTCNN motifs were inserted into a pUC19 backbone, with IR7 being introduced via primer overhangs. The resulting plasmid was linearized and biotinylated for single-molecule studies (3IR7kb12) or PCR amplified for EMSA (IR7b1500). Created in BioRender: K. Arfman (2026); https://BioRender.com/nu8rczr. (D) MpARF2 binding to 1500-bp DNA containing IR7 (filled) or a nonspecific variant lacking TGTCNN (NS, open symbols), showing that sequence preference is lost. Inset: Concentration-averaged fractions of bound protein, θARF, showing that the DBD retains sequence specificity for 1500-bp DNA, whereas MpARF2 does not. ***P < 0.001; not significant (ns), P > 0.05; Welch’s t test. To assess DNA binding in the absence of extensive clustering, we first performed EMSA on MpARF2-DBD with 19–base pair (bp) oligonucleotides containing two AuxREs in inverted repeat with 7-bp spacing (IR7b19; Fig. 3A). MpARF2-DBD exhibits a moderate affinity for IR7b19 with an effective Kd∗ ≈ 250 nM. Moreover, we find that this interaction is sequence-specific: MpARF2 shows virtually no binding to a nonspecific DNA template of the same length (NSb19) that lacks the TGTCNN recognition sequence (Fig. 3B, gray; and fig. S13). Hence, classical TF binding behavior is observed when clustering is suppressed. To study how TF clustering can affect DNA binding, we repeated the experiment using full-length MpARF2. As before, DNA binding to oligomeric DNA was highly sequence-specific. However, consistent with our earlier observations on AtARF2 (37), the full-length protein has a substantially higher affinity (Kd∗ ≈ 50 nM) compared to MpARF2-DBD, already revealing how DNA binding thermodynamics can be controlled by multimerization (Fig. 3B, blue; and fig. S14). The use of short oligos is standard practice in EMSA and in the study of interactions between ARFs and DNA. However, we suspected that the use of rather short DNA oligos could impose an experimental bias, since TF clusters have the capacity to occupy many more binding sites than a short DNA template can provide. In fact, the oligonucleotides most commonly used in previous studies are just long enough to include a tandem AuxRE, such as IR7 (19 bp), and can accommodate, at most, two DBDs (59). In their native cellular context, ARF proteins interact with genomic DNA of which accessible regions vary in length but are often much longer than 19 bp (60). To explore the scenario of a longer stretch of accessible DNA, we engineered 1500-bp DNA fragments of two types: one version containing a single IR7 AuxRE in the center (IR7b1500), and a nonspecific version (NSb1500) free of the recognition motif (Fig. 3C and table S1). Although TGTCGG and TGTCTC are considered high-affinity targets, ARFs have a moderate affinity for any TGTCNN variant (5, 45). Therefore, we obtained a genomic DNA motif that contains no TGTC to use as our nonspecific DNA backbone. MpARF2-DBD binds 1500-bp DNA in the same way it binds 19-bp oligomers: Visual inspection of the gels reveals that binding remains sequence-specific and relatively weak (fig. S15). However, this experiment could not be analyzed using conventional EMSA quantification, because the 1500-bp DNA is roughly 20 times heavier than the DBD molecule. Consequently, monomeric or dimeric binding does not cause a measurable DNA mobility shift. To quantify DNA binding, we computed the bound fraction of protein (instead of DNA), θARF, averaged across protein concentration to improve statistics. This analysis confirms that MpARF2-DBD remains sequence-specific even for long DNA templates. For full-length MpARF2, however, extending the DNA template from 19 to 1500 bp resulted in a complete loss of sequence specificity (Fig. 3D, inset). We interpret this as follows: Condensates exhibit so much residual (nonspecific) affinity to any form of DNA (even DNA that contains no TGTC sequences at all) that specificity is effectively lost. This is based on the concept of multivalency: Although the nonspecific binding affinity of a single ARF to non–TGTC-containing DNA is very weak, such nonspecific interactions are exponentially strengthened in a condensate by the vast number of MpARF2 proteins in the droplet, resulting in very strong and nonspecific DNA binding. Like the DBD, full-length MpARF2 is much lighter than the 1500-bp DNA template, yet, in contrast to the DBD-only experiments, binding of MpARF2 caused the DNA signal to shift completely into the wells (fig. S16). This pronounced shift indicates that, whereas MpARF2-DBD binds as low-order assemblies (likely dimers), full-length MpARF2 binds 1500-bp DNA exclusively in assemblies of at least tens to hundreds of protein molecules. The shift observed only for full-length MpARF2 allowed the binding curve to be constructed from θDNA as before. The transition of this curve coincides with csat, suggesting that condensates, rather than nanoclusters, are responsible for the DNA binding observed here (Fig. 3D). In the condensate regime, the length of DNA thus imposes a trade-off between the affinity and specificity of DNA binding. However, this is problematic: Sequence specificity is a crucial trait of TFs, so our results seem to be at odds with biological reality. Our EMSAs highlight the DNA binding behavior of condensates formed at relatively high TF concentrations, but what about nanoclusters present at significantly lower and biologically relevant protein concentrations? We would like to emphasize that EMSA may fall short in detecting DNA binding of nanoclusters due to limitations related to stoichiometry and sensitivity. In an EMSA, the amount of DNA used determines the maximum observable binding fraction. For example, at a 100-fold excess of DNA, the proportion of bound DNA can at most be 0.01 if all protein molecules are bound to DNA in a monomeric state. Clustering amplifies this stoichiometry effect even further: Because clusters contain many individual MpARF2 molecules, the concentration of clusters can only be a small fraction of the total MpARF2 concentration. Nanoclusters are a low-abundance species (ccluster≪cprotein) present only at the low end of our protein concentration range (cprotein≪cDNA). Therefore, their binding to DNA can be expected to have a negligible effect on the total proportion of bound DNA. To restore this stoichiometry imbalance, the DNA content would have to be drastically reduced. However, doing so would compromise the detectability of the already faint DNA bands. In short, EMSA is not a suitable technique to probe the low nanomolar binding of TF clusters due to intrinsic technical limitations. Hence, an alternative approach is required that combines high sensitivity with the ability to operate at low DNA concentrations. Nanoclusters offer strong, sequence-specific, switch-like DNA binding at physiological levels To investigate DNA binding of MpARF2 nanoclusters, we used a single-molecule imaging approach that integrates microfluidics with total internal reflection fluorescence (TIRF) microscopy. In this assay, custom-engineered DNA molecules of 12 kb in length, containing three equally spaced IR7 sites (3IR7kb12; Fig. 3C), are tethered to the surface of a passivated microfluidics chip through biotin-neutravidin interactions. The DNA was stained with SYTOX Orange (SxO), allowing individual DNA molecules to be visualized under mild flow. This flow stretches the DNA coils into linear strands, facilitating the observation of protein binding (Fig. 4A). For these experiments, MpARF2 proteins, either full-length or domain deletions, were covalently labeled with ATTO633 with a degree of labeling of ~1 dye molecule per protein. The proteins were introduced into the flow cell, already containing the surface-bound DNA; protein and DNA channels were imaged using alternating laser excitation (61). We used protein concentrations between 1 and 10 nM, a physiological concentration range well below csat, to prevent condensates from being present. This approach allows tens to hundreds of individual DNA molecules to be monitored simultaneously within a single field of view (FOV), providing a real-time visualization of MpARF2 binding (Fig. 4B). DNA binding of nanoclusters at physiological concentrations is strong, specific and switch-like. (A) Illustration of the single-molecule imaging assay, which combines microfluidics, surface immobilization and TIRF microscopy. Created in BioRender: K. Arfman (2026); https://BioRender.com/ts5uu4a. (B) Typical FOV showing individual flow-aligned DNA strands (green) with MpARF2 nanoclusters (magenta) bound to the DNA templates. [MpARF2] = 1 nM. Scale bar, 5 μm. Inset: Filtered ROI. (C) Kymograph of a nanocluster binding to DNA, illustrating the high temporal stability of a nanocluster once bound, with virtually no lateral movement. (D) Binding isotherm of MpARF2 nanoclusters (blue) and MpARF2-DBD (gray), determined from the area fraction of protein covering the DNA, ϕ. Inset: Binary-filtered ROIs used in coverage analysis. (E) Spatial distribution of binding events along individual DNA strands (solid) compared to the predicted IR7 positions (dashed). Inset: Enrichment of protein on nonspecific (green) and IR7-containing (orange) DNA templates, measured as the ratio of protein intensity on DNA compared to the surrounding background. ***P< 0.001, Welch’s t test. Upon introducing full-length MpARF2, we observe small protein clusters, significantly brighter than a single fluorophore, attaching to the DNA. This reveals that nanoclusters bind DNA at physiological protein concentrations (Fig. 4B). Although some nanoclusters also bind to the passivated glass surface, we discriminate these from DNA-bound nanoclusters by their positional fluctuations that follow the waving motion of the DNA strands to which they are bound (movie S3). DNA binding is observed at MpARF2 concentrations as low as 1 nM, suggesting that nanoclusters have a strong affinity for DNA, well below reported Kd values for binding of monomeric/dimeric species to short oligomers (55). Due to the high affinity and multivalent interaction, nanocluster binding events are stable and typically remain bound for minutes without noticeable motion of the nanoclusters along the DNA (Fig. 4C). While nanoclusters form high-affinity and time-stable bonds with DNA, monomeric or dimeric ARFs bind much weaker and exhibit fast dissociation dynamics (59). Above, we reported that in the absence of DNA, clusters of full-length MpARF2 form spontaneously in solution, but that this relies on the presence of a PB1 domain that acts as a concentrator. Here, our observations further support that a concentrator is needed. For full-length MpARF2, DNA binding occurred instantaneously (<100 ms), suggesting that nanoclusters preform in solution rather than assembling on DNA. In contrast, the ΔPB1 protein variant typically first binds to DNA and then progressively grows into a cluster, rather than forming a cluster in bulk solution (fig. S17). The EMSA data provide complementary evidence: At ~10 nM, ΔPB1 can bind DNA as small oligomers without producing a mobility shift, whereas larger protein clusters dominate at elevated protein concentrations (fig. S18). Together, this suggests that when the PB1 domain is present, clustering occurs primarily in bulk. In the absence of this domain, clustering in the bulk solution is suppressed, consistent with the hypothesized role of the PB1 domain as a concentrator. However, by acting as a scaffold, DNA can take over the role of the PB1 as a concentrator, thereby forcing cluster formation to occur after binding. This picture is fully consistent with the concept of templated assembly, where multiple binding transitions (MR-MR, PB1-PB1, and DNA-DBD), acting synchronously, orchestrate DNA binding (62). At 1 nM MpARF2, binding events are rare, and the number of nanoclusters present in solution is limited. By contrast, at 10 nM, DNA rapidly gets covered by MpARF2 nanoclusters, which are also more abundant in solution. This hints at a narrow binding transition between 1 and 10 nM. To better resolve this binding transition, we conducted our assay on a 48-kb λ-DNA template containing numerous AuxREs and systematically varied the MpARF2 concentration inside the same microfluidics chip. Between measurements, the flow cell was thoroughly washed with buffer to minimize any unwanted time-dependent effects. We quantified the coverage of the DNA with MpARF2, ϕ, and observed an unexpectedly sharp, almost binary, switch-like binding transition only 2 nM wide (Fig. 4D, blue line). This is notable, as the binding curves for ARF monomers and dimers span orders of magnitude (37). To confirm that this switch-like transition is caused by the assembly of MpARF2 into nanoclusters, we also measured the binding isotherm for MpARF2-DBD, which does not form nanoclusters, and found a complete absence of a binding transition in the same concentration range (Fig. 4D, gray line). In addition, the DNA coverage for the full-length MpARF2 reaches a saturation plateau before the DNA is fully covered; this is exactly what one would expect for a sequence-specific DNA binder to DNA that contains both specific and nonspecific regions. In our EMSA assays, we observed that the formation of large protein assemblies (i.e., condensates) led to a loss of sequence specificity in DNA binding (Fig. 3D). We hypothesized that, by keeping their size limited, nanoclusters retain their sequence specificity, while still benefiting from the enhanced affinity provided by clustering. To test this hypothesis, we generated a nonspecific DNA variant (NSkb12; table S1) that lacks the IR7 motif and contains minimal TGTCNN elements. Aside from the absence of the three IR7 motifs, the sequence of NSkb12 is identical to that of 3IR7kb12. We performed single-molecule experiments on both DNA variants using 10 nM MpARF2 and compared the extent of DNA binding. We first select regions of interest (ROIs) that each contains a single DNA strand. For each ROI, DNA fluorescence was used as a mask to distinguish protein signal on the DNA from that outside of it. DNA binding was quantified as the signal enrichment on the DNA (⟨Iin⟩/⟨Iout⟩). Conducting this assay on NSkb12 DNA revealed that nanoclusters can occasionally interact with nonspecific DNA (fig. S19); however, such interactions were rare, showing only a modest enrichment of ~1.5-fold. By contrast, DNA containing three IR7 motifs binds MpARF2 much more strongly, with the enrichment increasing to approximately seven. This confirms our hypothesis and reveals that nanoclusters, while binding cooperatively, retain strong sequence specificity (Fig. 4E, inset). At low nanomolar concentrations within the physiological range, the PB1 domain is necessary for nanocluster formation and consequently for nanocluster-based DNA binding, as shown by the EMSA data in fig. S18. At physiological concentrations, ΔPB1 therefore binds DNA as monomers or small oligomers, because the PB1 domain lowers the threshold for nanocluster formation by acting as a concentrator. The DNA itself, however, has a similar function by locally concentrating the protein. As a result, ΔPB1 (at 10 nM) binds DNA through the formation of nanoclusters, which retain their sequence specificity (fig. S20). For full-length MpARF2, sequence specificity was pronounced enough that we could resolve it within a single DNA molecule. When stretched by weak hydrodynamic flow, the three IR7 binding sites on a single 3IR7kb12 DNA strand are physically separated by ~1 μm, which is just enough to resolve them optically. We selected 46 ROIs, gathered across 1 to 10 nM protein concentrations, in which DNA binding is clearly visible, and computed the spatial distribution of MpARF2 along the contour of the DNA strand, averaged over time and across ROIs. Three preferred binding sites emerged, corresponding almost perfectly with the predicted IR7 motif positions within this DNA construct (Fig. 4E). For the DNA template without recognition elements, overall binding is significantly reduced, since only nonspecific binding remains that shows no clear spatial pattern (fig. S21). This confirms that AuxRE sites can spatially direct nanocluster-driven ARF binding. DISCUSSION The results presented above provide a picture of how the plant TF MpARF2 forms nanoscopic assemblies at biologically relevant protein concentrations. These nanoclusters, estimated to contain tens to several hundreds of protein molecules, exhibit many of the characteristics of a micelle, including a well-defined size corresponding to a state of thermodynamic equilibrium. This distinctly separates them from condensates, which are macroscopic droplets containing thousands to possibly millions of proteins and are thermodynamically unstable; they tend to grow until a single macroscopic condensate remains. MpARF2 also forms these phase-separated condensates in vitro, but not at concentrations we encountered in plant nuclei. The nanoclusters of this ARF arise from a balanced interplay between the different protein domains. The MR, high in intrinsic disorder, is the main driver of nanocluster formation, while the PB1 interaction domain appears to function as a concentrator that promotes protein clustering, and the DBD appears to function as a growth limiter. This is also exactly what is needed for micelle formation: a driving force for phase separation that is limited to a nanoscopic size by a growth limitation mechanism. The nuclear MpARF2 concentration range we observed in vivo lies within a regime where neither monomers/dimers nor condensates are capable of efficiently binding DNA. The measured MpARF2 levels are below the saturation threshold required for condensate formation, ruling out condensates as the dominant DNA binding species. Although monomers and dimers are likely present in the nucleus, their weak affinities would preclude stable DNA interactions at physiological protein concentrations. Instead, our data support a model in which nanoclusters represent the functional DNA binding form of MpARF2. These assemblies exhibit high-affinity, sequence-specific, and stable interactions with DNA, and their binding displays a switch-like dependence on protein concentration, with the transition of this switch occurring within the physiological MpARF2 range. Monomers/dimers, nanoclusters, and condensates are notably different in their DNA binding, but which DNA binding traits are actually favorable from a biological perspective? To address this question, it is helpful to consider the conceptual requirements for TFs to accurately regulate gene expression. Their binding must be specific to their target genes, i.e., they must exhibit sequence specificity. At the same time, their binding is ideally switch-like, so distinct DNA-bound and DNA-unbound states can be achieved. A high affinity is favorable for reasons of protein economy and such that the binding is temporally stable, so that a transcriptional state can be maintained without large fluctuations. Last, there must be options to toggle this switch-like state. Therefore, the binding must be tunable via a signaling process. From a physical chemistry perspective, it is evident that monomeric binding cannot meet all needs: Although monomers can offer excellent sequence specificity, it is challenging to create interactions that are both high-affinity and tunable. Moreover, the binding isotherms of monomers will be gradual rather than resembling a switch. Macroscopic condensates formed by LLPS, on the other hand, can meet the constraints of switch-like and high-affinity binding, but have little sequence specificity because the background affinity for nonspecific DNA is dramatically amplified by the enormous multivalency of a droplet containing thousands of proteins. Our data demonstrate that clustering, by amplifying weak, nonspecific interactions, inherently poses a trade-off between DNA binding affinity and specificity. Without clustering, DNA binding is specific but weak. The formation of large assemblies, on the other hand, leads to strong but nonspecific binding. Nanoclusters, situated in the middle of these extremes, provide an optimal balance. By coupling phase separation with a growth-limiting process, nanoclusters achieve high-affinity switch-like binding at concentrations within the physiological range, while sequence specificity is retained. It is therefore not surprising that the formation of nanoscopic assemblies is beginning to emerge as a conserved and universal mechanism for DNA and RNA binding proteins (63–68). However, our data do not fully address how nanoclusters can be toggled between bound and unbound states. Since we observe an almost binary switch-like transition from unbound to bound, changes in TF concentration could be used as a switching mechanism. Mechanisms exist for the targeted degradation of ARFs (44, 47). However, two additional regulatory mechanisms would be possible based on our data. First, we identify the accessibility of DNA as a key factor in DNA binding. Through multivalent interactions, nanoclusters can bind strongly to DNA if many binding sites are available. However, when shorter stretches of DNA are available for binding, this collectivity is lost causing clusters to engage DNA in a way that resembles monomeric binding. In a biological context, the length of accessible DNA may vary from spanning just a few base pairs to possibly kilobases (60). From an experimental point of view, this means that a DNA template of fixed, limited length likely provides a meaningful but incomplete description of DNA binding. Biologically, this suggests that cells can tune the interactions between nanoclusters and DNA not only by changing the concentration of TFs, but also by regulating the accessibility of DNA (69). Second, DNA binding may be regulated through PB1 interactions. Our results indicate that the PB1 domain enhances nanocluster formation by promoting specific interactions between ARFs. Thus, we speculate that PB1 domains on other proteins may, through competition, suppress the formation of ARF nanoclusters and their subsequent DNA binding. Notably, the key binding partner of ARFs in the nuclear auxin pathway, the auxin-regulated Aux/IAA, contains a PB1 domain. In the absence of auxin, Aux/IAA binds the ARF through the PB1 interaction, preventing it from binding to DNA. When auxin is introduced, the Aux/IAA protein is degraded, which may enable ARF to form nanoclusters and perhaps bind its targets through a different mode. Based on our results, this picture can be interpreted as follows: Without auxin, the PB1 interaction of Aux/IAA competes with PB1 interactions between ARFs, suppressing nanocluster formation. In the presence of auxin, this PB1 competition between Aux/IAA and ARFs is suppressed, releasing ARFs for mutual PB1 association, which then switches on nanoclustering and thus the activation of DNA binding. While this remains speculative at this stage, if true, it could provide a mechanistic understanding of the regulatory role of TF nanoclusters as transcriptional switches. MATERIALS AND METHODS Confocal microscopy in vivo The M. polymorpha MpARF2-mNeonGreen knock-in line used for in vivo imaging of MpARF2 was previously described in (47). Plants were grown on ½ strength Gamborg B5 medium at 22°C with 40 μmol photons m−2 s−1 of continuous white light. To estimate MpARF2 levels in vivo, dormant gemmae from 1-month-old plants were mounted on liquid Gamborg B5 medium and imaged using a Leica Stellaris 8 inverted confocal microscope. mNeonGreen fluorescence was detected in photon-counting mode with time-gated acquisition to suppress autofluorescence. A 100× objective was used to obtain high-resolution recordings of nuclei. To estimate native MpARF2 concentrations in gemmae, z-stacks were taken using a 20× objective. The mean fluorescence intensity of 225 nuclei, in eight independent gemmae, around the notch of the gemmae was measured, corrected for background, and converted to absolute concentrations by calibration. Calibration was performed by measuring purified mNeonGreen solutions (0 to 16 nM) in TE7.5 buffer (20 mM tris, 1 mM EDTA, and 150 mM NaCl, pH 7.5) under identical imaging conditions. For deconvolution confocal microscopy experiments, nanoclusters were visualized in 2-week-old MpARF2-mNeonGreen knock-in lines using a Leica Stellaris 8 inverted confocal microscope in the Leica Lightning (deconvolution) module. An HC PL APO CS2 86×/1.2 water objective was used. Samples were excited using 20% white light laser at 488 nm, and emission was detected between 494 and 550 nm. Four frames and three lines were accumulated. We recorded a single nucleus using 20 to 30 z-stacks, with a pixel voxel size of 23 nm (x and y) by 215 nm (z). These were the optimal settings for deconvolution with the Lightning module. Deconvolution was performed with the default Lightning settings. MpARF2::NLS-mNG was used as a negative control using the same settings. Measurements were repeated several times in independent gemmae to verify the consistency of the results. RNA isolation, cDNA synthesis, and MpARF2 expression analysis by reverse transcription qPCR Dormant gemmae from MpARF2 knock-in line and wild-type M. polymorpha strain Tak-1 were collected from 1-month-old plants and immediately frozen and stored at −80°C. Total RNA was purified from ground frozen tissue with the RNeasy plant mini kit (QIAGEN). Deoxyribonuclease (DNase) treatment was performed on-column using RNase-free DNase I (QIAGEN) following the manufacturer’s instructions. RNA samples were quantified using a NanoDrop spectrophotometer and checked via agarose electrophoresis. cDNA was prepared from 1 μg of total RNA using the iScript Select cDNA Synthesis Kit (Bio-Rad) in a 20-μl reaction using oligo(dT) primer. qPCR was performed using the iQ SYBR Green Supermix (Bio-Rad) in a CFX384 Touch Real-Time PCR thermocycler (Bio-Rad). Reactions were prepared in 10-μl volume containing 5 μl of 2× iQ SYBR Green Supermix, 500 nM forward and reverse primers, and 2.5 μl of cDNA. The qPCR experiment consisted of an initial denaturation step of 3 m at 95°C, followed by 39 cycles of a denaturation step of 10 s at 95°C, and elongation for 30 s at 55°C. A melt curve analysis was included after the run and consisted of an initial denaturing step for 10 s at 95°C and a gradient ranging from 55° to 95°C in 0.5°C increments. Three biological replicates per genotype and two technical replicates per reaction were assessed. MpEF1a and MpSAND expression was used for normalization. Gene expression analysis was performed using the CFX Maestro software (Bio-Rad) on normalized expression (ΔΔCq) mode, followed by a two-tailed unpaired t test. Cloning expression plasmids of MpARF2 variants To simplify cloning, we first generated pET_MBP-3C-MCS-TEV-mNG-His, a pET plasmid containing a multiple cloning site (MCS), flanked by regions encoding a 3C-cleavable maltose-binding protein (MBP; N-terminal), and a TEV protease (tobacco etch virus nuclear-inclusion-a endopeptidase) cleavable mNeonGreen (mNG) and His-tag (C-terminal). This was achieved through HiFi assembly (NEB) of linearized pET28a(+) plasmid, MBP, and mNG (table S2; pET28, MBP, and mNG) along with an MCS dsDNA oligo (table S1; MCS). This plasmid was linearized using restriction digestion or PCR (table S2, pET_MCS), and MpARF2 variants were inserted as described in table S3. The sequence of MBP was amplified from pET_His-MBP-TEV-LIC (Addgene no. 29656), and MpARF2 and mNG coding sequences were amplified from pMON_35S::MpARF2-mNG. Protein purification All media were supplemented with kanamycin (50 μg/ml) and chloramphenicol (10 μg/ml). Unless mentioned otherwise, cells were grown at 37°C, 180 rpm. All chromatography steps were performed at room temperature (RT). Rosetta (DE3) Escherichia coli (Novagen) harboring pET vectors expressing the MBP-3C-MpARF-TEV-mNG-His fusion proteins was first grown overnight in 50 ml of Luria Broth. This preculture was then diluted 1:100 into 3 liter of Terrific Broth (TB). The newly inoculated culture was then grown to optical density 0.6 to 0.8, after which 0.3 mM isopropyl-β-d-thiogalactopyranoside was added, the temperature was lowered to 20°C, shaking speed was lowered to 160 rpm, and growth was continued overnight (18 to 20 hours). Cells were harvested by centrifugation at 5000g, weighed (typically 30 g), and resuspended in lysis buffer (20 mM tris, 1 M NaCl, 10% glycerol, 0.1% NP-40, 2 mM MgCl2, 10 mg DNAse I, and cOmplete Protease Inhibitor, pH 8) at a ratio of 3 ml of buffer per 1 g of cell pellet. Cells were lysed by sonicating the suspension in a Q500 sonicator (Qsonica) equipped with a 0.5-inch (12.7-mm) probe at 80% amplitude for 2.5 min (5 s on/20 s off). Cell-free extract (CFE) was generated by centrifuging the lysate at 50,000g for 1 hour at 4°C and collecting the supernatant by gently decanting. To the CFE, 10 ml of Nuvia Ni-IMAC resin (Bio-Rad) was added, and the mixture was shaken at RT at 120 rpm for 1 hour. The mixture was then transferred to a column, and the resin was collected by pumping from the bottom using a peristaltic pump. After removing the CFE, the resin was washed with 10 column volumes (CVs) of Ni Wash Buffer (20 mM tris, 1 M NaCl, 10% glycerol, 1 mM tris(2-carboxyethyl)phosphine, 20 mM imidazole, and 0.01% Tween 20, pH 8), after which the protein was eluted with 2.5 CVs of Ni Elution Buffer (Ni Wash Buffer + 480 mM imidazole). The eluted protein was then passed over a Superdex 200PG column with bed dimensions of 2.6 cm by 100 cm, equilibrated in SEC buffer no. 1. Monomer fractions were pooled, 5 mM dithiothreitol (DTT) was added, and the mixture was incubated for 30 min. The pooled fractions were then loaded onto a 5-ml MBPTrap (Cytiva) equilibrated with SEC buffer no. 1, after which the column was washed with 2 CVs of SEC buffer no. 1, and the protein was eluted with 2 CVs of amylose elution buffer (SEC buffer no. 1 + 20 mM maltose). We collected fractions containing the highest concentrations as estimated by A280 (usually ~3 ml), and the concentration was calculated by measuring A506, using the ε506 of mNG (116,000 M−1 cm−1). Protein purity was assessed through SDS–polyacrylamide gel electrophoresis and Coomassie staining. Purified protein was snap-frozen using liquid nitrogen in 100-μl aliquots and stored at −80°C until use for experiments. For centrifugation assays, confocal microscopy, and DLS, a fourth column is run on the day of the experiment. Up to 500 μl of MBP-tagged protein was thawed, to which HRV 3C protease (Sigma-Aldrich) was added at a ratio of 1 μg of 3C per 1 nmol of protein. After incubating for 2 hours at RT, the mixture was passed over a Superdex 200 Increase 10/300 GL (Cytiva) equilibrated with SEC buffer no. 2 (20 mM tris, 500 mM NaCl, 1 mM EDTA, 1 mM DTT, and 0.001% Tween 20, pH 7.5), and monomer fractions were collected (retention volume ~10 ml). If the concentration of the protein was insufficient, monomer fractions were pooled and concentrated using Amicon Ultra-0.5 10k MWCO Centrifugal Filters (Millipore). Concentration was then calculated by measuring A506 on a NanoDrop One (Thermo Fisher Scientific) in triplicate. The protein was stored in a Protein Lo-Bind microcentrifuge tube at RT, for up to 1 hour before use. For DLS, SEC was run in 1 M instead of 500 mM NaCl. Centrifugation assay Forty-microliter assays were prepared in a Protein Lo-Bind 1.5-ml microcentrifuge tube using a Hamilton syringe. Proteins were prediluted in high-salt buffer (20 mM tris, 500 mM NaCl, 1 mM EDTA, 1 mM DTT, and 0.001% Tween 20, pH 7.5) and brought to desired concentrations by 4× dilution in salt-free buffer (20 mM tris and 0.001% Tween 20, pH 7.5). Samples were incubated for 15 min at RT and centrifuged for 1 min at 21,000g at RT. After centrifugation, 25 μl of supernatant was collected without disturbing the pellet and diluted in 75 μl of assay buffer (20 mM tris, 0.25 mM EDTA, 125 mM NaCl, 0.25 mM DTT, and 0.001% Tween 20, pH 7.5). The protein concentration of diluted supernatant was determined from fluorescence intensity using a Synergy H1 Plate Reader (Bio-Tek), using MBP-tagged protein at known concentrations for calibration. The csup of the original sample was then calculated by correcting for the 4× dilution. Confocal microscopy in vitro Confocal microscopy samples were prepared as described for centrifugation assays, with a total volume of 20 μl. Immediately after preparation, 5 μl of the sample was transferred onto a microscopy slide fitted with a Secureseal 8-wells imaging spacer (Grace Bio-Labs). A coverslip was added, and the slide was inverted and incubated for 15 min to allow the condensates to settle. Coverslip surfaces were imaged on a Leica SP8 equipped with single-molecule detection using a 63× water-immersion objective. mNG fluorescence was excited with a pulsed white light laser at 488 nm, and emission was collected from 515 to 545 nm with hybrid detectors. Samples for coarsening were prepared with a total volume of 40 μl and imaged in a μ-slide 18-well glass coverslip (Ibidi). Solutions were then imaged at 15 min and 24 hours after sample preparation. Between measurements, the coverslip was sealed with parafilm and incubated in the dark at RT. Dynamic light scattering Monomeric protein obtained after SEC was diluted in high-salt buffer (20 mM tris, 1 mM EDTA, and 1 M NaCl, pH 7.5, sterile filtered). Protein variants were studied at 1 μM to provide enough signal to fit broad distributions. The experiments in which the salt dependence and reversibility of clustering were tested (Fig. 2B) were performed at 50 nM MpARF2. High-salt samples were obtained by diluting the protein in high-salt buffer. Samples with 150 mM NaCl were prepared by prediluting the protein in high-salt buffer, such that the sample could be brought to the desired concentration by diluting in salt-free buffer (20 mM tris and 1 mM EDTA, pH 7.5, sterile filtered). The reversibility of clustering was assessed by first inducing MpARF2 clustering at 100 nM in 150 mM NaCl, as described above, followed by diluting the protein back to 1 M NaCl (and 50 nM MpARF2) using the above-described high-salt buffer supplemented to 2 M NaCl. Measurements were performed using disposable microcuvettes (Brand; ≥100 nM) or quartz cuvettes (Hellma; <100 nM) on a Malvern Zetasizer Nano with an acquisition time of 15 min (three to five cycles). All DLS data presented are shown as intensity-weighted size distributions. Fluorescence correlation spectroscopy An 18-wells plate (Ibidi) was pretreated with bovine serum albumin (BSA; 1 g/liter) in buffer (20 mM tris, 1 mM EDTA, and 150 mM NaCl) and incubated for 5 min to prevent surface adsorption. The solution was removed, and plates were allowed to air-dry prior to use. A MpAF2 aliquot was thawed and centrifuged at 21,000g for 30 min before being prediluted in high-salt buffer and subsequently adjusted to a final concentration of 7 nM MpARF2 and 150 mM NaCl using salt-free buffer. FCS measurements were performed on a Leica Stellaris 8 inverted confocal microscope equipped with a PicoHarp module for time-correlated single-photon counting. Measurements were conducted using a 60× water immersion objective and a confocal pinhole size of 1 Airy unit. System calibration was performed by measuring a 100 nM sample of rhodamine 110 (10 cycles, 30 s per cycle). Autocorrelation curves were fitted using a triplet-state single-component diffusion model to determine the structural parameter of the confocal volume, a, and the characteristic diffusion time of rhodamine 110, τR110. Subsequently, MpARF2 was measured on the same plate and with the same acquisition settings. Autocorrelation data were analyzed using a maximum entropy model implemented in Python, yielding a distribution of diffusion times, τ. These diffusion times were subsequently converted into a size distribution, d, usingDR110τR110=kBT3πηdτ(1)where η is the viscosity of the medium and DR110 is the diffusion coefficient of rhodamine 110, with a value of 270 μm2/s (70). DNA engineering 3IR7kb12 and NSkb12 DNA were obtained by building plasmids, consisting of one pUC19-derived backbone edited to contain minimal TGTC (table S1; UC19′) and three unique TGTC-free “blocks” amplified from A. thaliana genomic DNA. All assemblies were done using HiFi. Blocks were made from pairs of ~1500-bp TGTC-free fragments (table S2; A1, A2, B1, B2, C1, and C2) and assembled into block plasmids (pBlockA, pBlockB, and pBlockC). UC19′ was purchased as a gBlock (Integrated DNA Technologies) and assembled into a backbone plasmid (table S4; pUC19′) using genomic filler DNA (table S2; filler) that lacks TGTC but harbors a SapI site capable of producing a GGG-overhang. The extended backbone (table S2; pUC19′) and blocks were amplified and assembled into p3IR7kb12 (table S2; A_IR7, B_IR7, and C_IR7) and pASkb12 (table S2; A_NS, B_NS, and C_NS) plasmids. Plasmids were propagated in E. coli, strain: DH5α. Sequences were verified by nanopore sequencing (PlasmidSaurus). Plasmids were linearized using SapI and biotinylated using DNA polymerase Klenow fragment. Standard protocols were used, except for the use of biotinylated deoxycytidine triphosphate (dCTP) instead of regular dCTP in the Klenow buffer. Custom Python software was used to scan the Arabidopsis genome for regions to amplify. IR7b1500 and NSb1500 were PCR amplified from corresponding 12-kb plasmids (table S2; b1500) and purified using gel electrophoresis and cleanup. IR7b19 and NSb19 (table S1) were obtained by annealing Cy5-labeled oligos in a 95°C heat bath, which was gradually cooled overnight. λ-DNA (NEB) templates were biotinylated in batches of 0.2 pmol by annealing primers (table S2; CosR-bio, 2 pmol; CosL, 20 pmol) at 65°C followed by ligation (NEB, T4 Ligase) for 2 hours at RT, as described in (71). Biotinylated λ-DNA was separated from unreacted oligos using centrifugal spin filters (Amicon Ultra-0.5 100K) according to the supplier’s protocol. Single-particle tracking MpARF2 aliquots were thawed and centrifuged at 21,000g for 30 min and then diluted to desired concentrations in tracking buffer [20 mM tris, 1 mM EDTA, 150 mM NaCl, and 10% (w/w) polyethylene glycol, molecular weight 1000 (PEG-1000), pH 7.5]. SPT was performed on a Nikon Ti-2E inverted microscope equipped with a Teledyne Kinetix sCMOS camera and a 100× silicon oil immersion objective. Recordings of 301 frames were acquired in duplicate at 50 fps. Tracking analysis was carried out in Python using the TrackPy package. Trajectories shorter than 10 frames were excluded from analysis. The number of clusters was quantified as the average number of particles detected per FOV (208 μm by 208 μm). Electrophoretic mobility shift assay Agarose gels (15 cm by 15 cm) were prepared in 0.5× tris-borate EDTA (TBE), using two 15-well combs: one at the top and the other in the middle of the gel. Agarose (2%) was used for 19-bp DNA and 0.5% agarose for other DNA lengths. Frozen MBP-MpARF2-mNG-His aliquots (1 μM, 100 μl) were thawed and cleaved with 3C protease (1 μg protease per 1 nmol ARF) for 2 hours at RT to cleave the MBP tag. Thirteen binding reactions (40 μl) containing DNA (50 ng) and protein (10 to 1000 nM, logarithmic, for DBD or 7.5 to 97.5 nM, linear, otherwise) were prepared in binding buffer (20 mM Hepes, 100 mM tris, 2.5% glycerol, and 150 mM NaCl, pH 7.5) along with negative controls for protein and DNA, and equilibrated for 30 min at RT. The protein concentration of the control equals that of the middle lane. The samples were loaded using 10× loading dye and ran for 90 min at 3 V/cm in 0.5× TBE at 4°C. Nineteen–base pair DNA templates contained a Cy5 fluorescent label. Gels using unlabeled 1500-bp DNA were poststained using 50 ml SxO (50 nM in Milli-Q) for 30 min on a rotating platform. Gels were imaged with a Typhoon gel scanner (Cytiva), detecting protein (Cy2 filter set) and DNA (Cy3 or Cy5 filter set depending on the stain) fluorescence. Analysis of gel image data was performed using custom Python software. For each gel, the bands in the DNA and protein fluorescence channels were annotated by specifying their center positions and classifying them as free, bound, or present in the wells. Bands of the same type align horizontally on the gel and are hereafter referred to as rows. Subsequently, the annotated band centers were used to define ROIs that were used for intensity integration. Because each gel contained a lane with only DNA or protein, each row contained at least one empty ROI, which was used for background correction. The integrated band intensities, I, were corrected by subtracting the corresponding background intensity. Although DNA on its own did not produce detectable signals in the wells, the addition of protein occasionally caused the DNA to end up in the wells. Well signals were therefore included in all analyses. The bound fraction, θ, was calculated asθ=1−Ifree′Ifree′+Ibound′+Iwell′(2)where I′ represents the background-corrected, integrated band intensity. The above-described analysis is visualized in fig. S12. Pretreatment of coverslips for single-molecule imaging Microfluidic devices were assembled and surface-treated in batches of eight. A clean glass staining jar (BR472700, Sigma-Aldrich) was filled with 70 ml of spectroscopy-grade acetone. Wearing powder-free gloves, we rinsed glass coverslips (24 mm by 60 mm, no. 1.5) with Milli-Q water and isopropanol to remove particulates and organic residues. Coverslips were then dried using compressed air and treated with plasma (Harrick PDC-32g) for 1 min. After plasma cleaning, the coverslips were immediately placed in acetone and transferred to a fume hood. Wearing gloves, we added 1.4 ml of VECTABOND (VectorLabs) and mixed the solution by pipetting up and down without introducing bubbles. After incubating for 15 min, VECTABOND solution was discarded appropriately. Coverslips were rinsed once with acetone and five times with Milli-Q. Slides were dried gently with compressed air and temporarily kept in falcon tubes. Microfluidic chip assembly Microfluidic chips were assembled from a pretreated coverslip, a double-sided tape spacer, and a polydimethylsiloxane (PDMS) lid. Double-sided tape (Hi-Bond, HB397F-25), protected by foil on both sides, was laser-cut into 25 mm–by–20 mm rectangular spacers containing a 10 mm–by–1 mm channel. To obtain protective backing on both sides of the tape, two strips were stacked front-to-back, and the exposed layer of tape was peeled off. Rectangular blocks of 25 mm by 20 mm by 5 mm were cast from PDMS (10:1 PDMS:resin) and provided with an inlet and outlet using a biopsy punch. Punctured PDMS pieces were rinsed with water and ethanol, air dried, and treated with plasma for 15 s. Chips were assembled by attaching tape to the cover glass and sticking the PDMS lid on top. Chips were firmly pressed to ensure a tight seal. Passivation of microfluidic chips Four milligrams of biotin-PEG-SC (MW 5000, Laysan Bio) and 32 mg of mPEG-SVA (MW 5000, Laysan Bio) were dissolved in 160 μl of Mops buffer (50 mM, pH 7.5). The solution was vortexed for 1 min and centrifuged for 1 min at 21,000g. Twenty microliters of the PEG solution was pipetted directly into the chips, and the inlet and outlet were covered with parafilm to prevent air from entering. The chips were incubated for 3 hours at RT in a humidity box, which had been prepared by partially filling an empty pipette tip box with water. A secondary PEGylation was performed to further improve passivation. A solution of 25 mM MS4-PEG (Thermo Fisher Scientific) in sodium bicarbonate buffer (0.1 M, pH 8.5) was prepared and stored at −20°C in 500-μl aliquots. After thawing, an aliquot was centrifuged for 1 min at 21,000g before injecting 50 μl into each chip. The chips were incubated overnight at 4°C in a humidity box. Protein labeling for single-molecule imaging Frozen MBP-MpARF2-mNG-His aliquots (100 μl, 1 μM) were thawed and brought to an amine-free labeling buffer (20 parts PBS, adjusted to 1 M NaCl; 1 part 0.2 M sodium bicarbonate, pH 9) using a desalting spin column (Zeba Spin, catalog no. 89882) according to the supplier’s protocol. Protein concentration was measured using a NanoDrop, and a 10× molar excess of ATTO633-NHS was introduced and incubated protected from light for 1 hour at RT. Degree of labeling was measured according to the ATTO-NHS labeling protocol, which involves separating free and unreacted dye via a gravity-flow desalting column (Bio-Rad; 10DG). Aliquots of 10 pmol labeled protein were flash-frozen in 1.5 ml of Lo-Bind tubes (Eppendorf) and stored at −80°C. Single-molecule imaging Microfluidic chips were mounted on a custom built TIRF microscope (72) and connected to a distributor valve (MUX-D-12, Elveflow) using Tygon tubing. Three distributor valve positions were used to connect (i) a 1.5-ml tube for samples, (ii) a 15-ml tube for buffers, and (iii) a plug to interrupt flow. DNA tethering was achieved by sequentially flushing in various 1-ml solutions with incubating steps in between. First, chips were washed with TE7.5 buffer (10 mM tris, 1 mM EDTA, and 150 mM NaCl, pH 7.5) to provide a clean starting point. Next, NeutrAvidin (0.05 g/liter in TE7.5) was incubated for 5 min and flushed out using blocking buffer [Pluronic F108 (1 g/liter), BSA (1 g/liter) in TE7.5 buffer]. Last, biotinylated DNA (1 pM in blocking buffer) was introduced and incubated for 1 hour. Tethered DNA was visualized, and excess DNA was removed by flushing with imaging buffer (50 nM SxO in blocking buffer). Frozen ATTO633-labeled protein aliquots were thawed, cleaved with 3C protease (2 hours, RT, 1 μg/nmol sample), and centrifuged for 1 min at 21,000g to remove large aggregates. Samples were adjusted to contain 50 nM SxO and 150 mM NaCl. Recordings (Andor Zyla 4.2+ sCMOS) of alternating laser excitation (561 nm, 642 nm, 50-ms interval) were started after which flow was switched on to flush in protein using gravity at an approximate flow rate of 100 μl/min for 5 to 10 min. Single-molecule imaging—Data analysis FOVs were manually marked to indicate ROIs and flow direction. A convolutional median filter (101 by 101 pixels) was applied to correct for uneven illumination before extracting ROI data as individual files for further analysis. For each ROI, DNA was detected by applying a three-dimensional Gaussian blur followed by binarization and a series of morphological operations (erosion, dilation, skeletonization, and pruning) that yield an unbranched chain one pixel in width. This chain was expanded to create a DNA mask, which was applied to binarized ROIs (for coverage analysis) and to blurred ROIs (for specificity analysis). Filtered ROIs were inspected and excluded in case DNA detection failed. DNA coverage, ϕ, was quantified from binary ROI data asϕ=⟨AARFADNA⟩t(3)where A represents the area of the signal in pixels. Sequence specificity was quantified as the enrichment of protein on the DNA compared to the environment (⟨Iin⟩ / ⟨Iout⟩), averaged over time and across ROIs. The DNA intensity profile along the x axis was obtained by averaging the protein signal over the y axis, time, and across ROIs.