Of numbers and movement – understanding transcription factor pathogenesis by advanced microscopy

Transcription factors (TFs) are life-sustaining and, therefore, the subject of intensive research. By regulating gene expression, TFs control a plethora of developmental and physiological processes, and their abnormal function commonly leads to various developmental defects and diseases in humans. Normal TF function often depends on gene dosage, which can be altered by copy-number variation or loss-offunction mutations. This explains why TF haploinsufficiency (HI) can lead to disease. Since aberrant TF numbers frequently result in pathogenic abnormalities of gene expression, quantitative analyses of TFs are a priority in the field. In vitro single-molecule methodologies have significantly aided the identification of links between TF gene dosage and transcriptional outcomes. Additionally, advances in quantitative microscopy have contributed mechanistic insights into normal and aberrant TF function. However, to understand TF biology, TF-chromatin interactions must be characterised in vivo, in a tissue-specific manner and in the context of both normal and altered TF numbers. Here, we summarise the advanced microscopy methodologies most frequently used to link TF abundance to function and dissect the molecular mechanisms underlying TF HIs. Increased application of advanced single-molecule and super-resolution microscopy modalities will improve our understanding of how TF HIs


Introduction
Precise tissue-and cell-specific regulation of gene expression is required for development and homeostasis. Transcription factors (TFs) tightly control transcriptional programmes in a cell-and tissuedependent manner. General TFs and the core transcriptional machinery bind to gene promoters, whereas specific activating and repressing TFs bind to gene regulatory elementsi.e. transcriptional enhancers or silencersand interact with promoter-bound complexes to control transcription (Horikoshi et al., 1988). This interaction is thought to occur through the Mediator complex, a multi-subunit protein complex that recognises enhancer-bound TF complexes and signals to RNA polymerase II to transcribe genes. The Mediator complex was first discovered in yeast (Kim et al., 1994); in humans, the basic Mediator complexes generally enhance transcription (Andrey et al., 2013;Bergiers et al., 2018;Ernst et al., 2011;Golan-Lagziel et al., 2018;Heintzman et al., 2009;Noordermeer et al., 2011;van Bömmel et al., 2018). Thus, cell type-specific gene expression is mainly influenced by the ability of TFs to bind their target sites within promoters or enhancers, which, in turn, relies on the chromatin state of these sites (Ernst et al., 2011).
The main TF families are classified according to their DNAbinding domains. These include the helix-turn-helix (HTH), zinc-finger (ZNF), basic helix-loop-helix (bHLH), basic leucine zipper (bZIP) and nuclear hormone receptor binding domains (Table 1), each featuring distinct mechanisms of sequence-specific DNA recognition and binding (Badis et al., 2009;Lambert et al., 2018;Wei et al., 2010). Furthermore, TFs often work in unique combinations, for example during development and differentiation (Bergiers et al., 2018). Thus, different cell types can be distinguished based on their TF repertoires. Thereby, the clustering of TF-binding sites in enhancer sequences favours functional synergies of TFs that are co-expressed in the same tissue (Golan-Lagziel et al., 2018).
The aberrant function of TFs can have profound effects on development and disease. Of particular interest is how reduced levels of some TFs may influence their function and result in haploinsufficient phenotypes. Haploinsufficiency (HI) is the inability of a gene to rely on only one (of its two) alleles to exhibit its normal function. Therefore, HIs generally result in loss-of-function (LOF), often abnormal, phenotypes. It follows that HIs are properties of genes and, although 'dosage-sensitive genes' is a broader term than 'HI genes', henceforth, for simplicity, we use the two terms interchangeably.
LOF mutations or deletions of HI genes are inherited in a dominant manner, resulting in phenotypic alterations or abnormalities caused by the insufficiency of one intact allele to confer full gene functionality. Computational models aim to predict the tolerance of a system at the cell, tissue or organismal level to a single functional copy of a gene. Such models have mainly focussed on predicting the probability of genes being LOF-intolerant (pLI) (Lek et al., 2016), i.e. rating genes with a score of n≥0.9 as intolerant and with a score of n≤0.1 as tolerant to LOF mutations (Lek et al., 2016). Bioinformatics approaches and machine learning have identified 7841 HI genes in the human genome (Shihab et al., 2017), linking them to a number of dominantly inherited HI diseases. Key examples are eye disorders, such as aniridia (Jordan et al., 1992), keratitis (Mirzayans et al., 1995) and ocular colobomas (Sanyanusin et al., 1995;Williamson et al., 2014), as well as multiple cranial, facial and limb diseasesincluding synpolydactyly (Muragaki et al., 1996), schizencephaly (Brunelli et al., 1996), craniosynostosis of Adelaide type (Hollway et al., 1995;Jabs et al., 1993) and Greig cephalopolysyndactyly syndrome (Hui and Joyner, 1993; reviewed by Dang et al., 2008). Disease phenotypes have been attributed to ∼300 HI genes and range from primary immunodeficiencies (Franco-Jarava et al., 2018;Torgerson and Ochs, 2014) to ribosomopathies (Kyritsis et al., 2019) and cancer (Inoue and Fry, 2017;Jennings et al., 2012;Largaespada, 2001). Interestingly, a significant proportion of dosage-sensitive genes encode TFs, with higher relative numbers of these TF genes exhibiting HI compared to other human genes (Ni et al., 2019).
Over 1600 human TF genes have been grouped into gene families (Lambert et al., 2018), allowing for in-depth analysis of HI incidence in specific subsets of TFs (Table 1). Overall, 122 TFs from 32 gene families were identified and designated as the most reliable dosage-sensitive (MRDS) genes (Ni et al., 2019). This categorisation of TF families has attempted to explain the sensitivity or insensitivityof different TFs to dosage and found HI to be predominantly correlated with small TF families. Such small TF familiescomprising fewer than eleven members, e.g. the Grainyhead and C2H2-ZNF/homeodomain familiesare present in the MRDS dataset, whereas the largest subgroup of the zinc-finger TF family Krüppel-associated box (KRAB) domaincontaining zinc finger proteins (KZFPs), is found to be dosageinsensitive and is not present in the MRDS dataset (Ni et al., 2019). This observation is yet to be explained, with evolutionary pressure being a plausible cause for limiting TF family size (Ni et al., 2019). Why other evolutionary mechanisms, such as gene duplication and divergence (Gehring et al., 2009), have not been deployed to expand these small families of TFs and alleviate HI remains unknown. Additional mechanisms that could account for such bias may relate to the regulatory, activating or repressing, behaviour of such TFs, or to the numbers of downstream-regulated genes. Thus, more work is needed to identify possible structural or functional similarities between dosage-sensitive TFs to further pinpoint the underlying disease-causing mechanisms of HIs. This Review summarises how the advanced quantitative microscopy methodologies aid in the study of biophysical and functional properties of TFs, such that TF dosagesensitivity at the molecular level may be linked to development and disease. These TFs contain four α-helices and an N-terminal flexible arm. The third α-helix forms base-specific contacts with the major DNA groove, the N-terminal flexible arm interacts with the minor DNA groove in a sequence-specific manner. The turn between the first and second helix interacts with the DNA backbone.
Homodimeric receptors: glucocorticoid (NR3C1), mineralocorticoid (NR3C2), progesterone (NR3C3), estrogen (ESR1, ESR2) receptors; androgen receptor isoforms A and B (AR-A and AR-B), triiodothyronine receptor (THRA/THRB) Heterodimeric receptors: retinoid X receptors alpha, beta and gamma (RXRA, RXRB, RXRG), vitamin D receptor (VDR) (Chandra et al., 2008;Giguere et al., 1987;Green et al., 1986;Hollenberg et al., 1985;Lippman et al., 1973;Luisi et al., 1991;Mangelsdorf et al., 1990;Miesfeld et al., 1986;Petkovich et al., 1987) Factors resulting in HIs Insufficient TF protein levels are the common denominator of haploinsufficient TF genes. In these cases, the abnormal phenotypes are triggered in a dosage-dependent manner (Gidekel et al., 2003;Gilchrist and Nijhout, 2001;Nishimura et al., 2001;Takeuchi et al., 2011). At the molecular level, altered TF numbers impair their functionality by interfering with their dynamic behaviour in cells and, thereby altering their transcriptional programmes (Bottani and Veitia, 2017;Veitia, 2003). We examine the following molecular and functional properties of TFs, which may underlie TF HIs in a tissue-specific manner (for a schematic representation, see Fig. 1): a. absolute TF numbers; b. stoichiometry of TFs and their interacting proteins; c. availability and binding kinetics to their TF target sites; d. stochastic nature of gene transcription; e. biophysical properties, such as their ability to form condensates. How these parameters vary simultaneously between or within cells, tissues or entire organisms adds to the sheer complexity in characterising the causes of dosage sensitivity and HI in humans (Elowitz et al., 2002;Veitia, 2002). HI may manifest in a range of related phenotypes in humans. Moreover, it is gradually becoming clear that not one single molecular mechanism is responsible for TF HIs; rather, parameters that determine TF function or dysfunction act in concert.
Studying the mechanisms of TF concentration and function in TF HIs is becoming increasingly possible thanks to the continuous advancement of microscopy methods. Table 2 briefly outlines these methodologies and their potential applications to investigate TF biology, and selected methodologies are depicted in Fig. 2.

Absolute TF numbers
Most TFs are expressed and, typically, act in a cell type-or tissuespecific manner. RNA sequencing (RNA-seq) analyses of human TFs found over a third of them to be enriched in specific tissues (Lambert et al., 2018;Uhlen et al., 2015). Spatiotemporally precise TF levels are key to driving cell specification and differentiation. Particularly during early development, different levels of stemness TFs control the fate of individual cells. For example, in mouse blastocysts, the distribution of SOX2 and POU5F1 (hereafter, referred to as OCT4), two of the key stemness TFs, confers an initial differentiation bias at the 4-cell stage (Goolam et al., 2016;Kaur et al., 2013;White et al., 2016). Later in embryonic development, differential expression of NANOG and GATA6 in a 'salt-andpepper' pattern in the mouse embryonic inner cell mass controls the segregation of the epiblast from primitive endoderm lineages (Chazaud et al., 2006). Stemness TF levels continue to control the pluripotency of different stem cell populations in development (Adachi et al., 2013). For instance, the intrinsically fluctuating amounts of OCT4 and SOX2 contribute to the lineage commitment of mouse embryonic stem cells (ESCs) during differentiation (Strebinger et al., 2019). When Pou5f1 and Sox2 become downregulated, mouse ESCs can differentiate and activate the trophoblast stem cell TF networks required for normal embryonic development (Adachi et al., 2013;Masui et al., 2007;Niwa et al., 2000Niwa et al., , 2005. This continued dependence on concentration for the precise function of TFs during development explains why aberrant TF gene dosage can have detrimental effects on development and physiology. Changes to the required concentration of TFs at different points in development can elicit diverse effects in gene regulation and, thus, result in different pathogenic phenotypes. At the subnuclear level, local concentrations of TFs affect gene expression by controlling the kinetics of TF binding to chromatin, recruitment of RNAPolII and transcriptional bursting (Nelson et al., 2004;Senecal et al., 2014). In mouse ESCs, the clustering of SOX2bound enhancers modulates the search for local SOX2 target sites and facilitates gene transcription (Liu et al., 2014). RNAPolII activity at the POU5F1 enhancer is regulated by the accumulation of   (Ehrenberg and Rigler, 1974;Elson, 2001;Elson and Magde, 1974;Magde et al., 1972). While temporal autocorrelation analysis is originally proposed, and most often applied, FCS records photon arrival times, which can also be examined using Photon Counting Histogram (PCH) analysis (Chen et al., 1999b). This method can quantitatively characterise in live specimens (live cells in culture, ex vivo tissue/ organ and the whole organism) TF concentration, mobility and the extent of TF binding to chromatin through changes in TF mobility (Chen et al., 2014b;Ehrenberg and Rigler, 1974;Mazza et al., 2012;Mir et al., 2017;Normanno et al., 2015;Papadopoulos et al., 2010;Vukojevic et al., 2010;White et al., 2016;Yao et al., 2006). It can also provide information about TF stoichiometry, i.e. homooligomerisation by brightness analysis (Gaglia et al., 2013).
FCS allows the detection of TF molecules with the ultimate, single-molecule sensitivity. It is quantitative and non-destructive, enabling measurements in live specimens. Due to its high temporal resolution, FCS is suitable for the study of fast dynamical processes, e.g. diffusion of small molecules and fast reaction kinetics. FCS measurements are best performed when the system is in a steady-state. This is advantageous for the study of reaction kinetics, as one does not need to establish initial concentration gradients (Elson, 2013 Fluorescence cross-correlation spectroscopy (FCCS) Interacting proteins are each labelled with spectrally distinct fluorophores. Separate detectors enable simultaneous recording of temporal fluctuations of fluorescence intensity from both fluorophores. Temporal autocorrelation analysis is performed for each signal separately. In addition, covariance in fluorescence intensity fluctuations from the two different fluorophores is analysed to assess the codiffusion of the investigated molecules (Bacia and Schwille, 2007;Krieger et al., 2014;Langowski, 2017;Rika and Binkert, 1989;Schwille et al., 1997). This method can be used to study the efficiency and stoichiometry of TF complex formation, such as hetero-/homo-dimers, as well as the chromatin-binding behaviour of these complexes (Bacia and Schwille, 2007;Krieger et al., 2014;Langowski, 2017;Papadopoulos et al., 2015;Reho et al., 2020).
Co-diffusion, which is detected by FCCS, indicates the presence of molecular interactions with no requirement for the proximity and orientation of individual fluorophores. FCCS provides information about both the concentration of free and bound TF molecules without the need to separate the two fractions. The TF-chromatin binding equilibrium dissociation constant (K D ) can be measured in live specimens. False positive results may be obtained in the cases of signal bleed-through (cross-talk) from one detector to the other. False negative results may be obtained if the system alignment is not optimal. Conventional, single-point FCCS is not an imaging methodology.

Massively parallel FCS (mpFCS)
Simultaneous recording of fluorescence intensity fluctuations at multiple points is achieved using spinning disc confocal microscopy (SDCM) and electron-multiplying charge-coupled device (EM-CCD) camera (Needleman et al., 2009) or a diffractive optical element (DOE) and a single-photon avalanche photodiode (SPAD) camera (Krmpot et al., 2019;Papadopoulos et al., 2015) to generate multiple observation volume elements. TF concentration and diffusion across a whole nucleus or among cells in the same tissue can be mapped. This allows location-specific measurement of the variability in TF concentration and chromatin-binding behaviour (Capoulade et al., 2011;Krmpot et al., 2019;Papadopoulos et al., 2015). All advantages as for FCS, with the addition of spatial information. Quantitative confocal imaging without scanning is enabled (Krmpot et al., 2019). Study of fast dynamical processes is made possible due to the high temporal resolution of the SPAD camera with a temporal resolution of (∼20 µs) (Krmpot et al., 2019), which is more than an order of magnitude better than for the best EM-CCD cameras.
mpFCS has the same disadvantages as FCS.
Additionally, it has reduced sensitivity, because outof-focus light, originating from molecules in remote focal planes above/below the focal plane, can pass through adjacent pinholes and increase the background signal. The disadvantage of SPAD matrix detectors compared to EM-CCDs stems mainly from the variability in dark count rates between individual detectors, which, if not accounted for, gives a non-uniform signal-to-noiseratio over an image frame.

Selective plane illumination microscopy (SPIM)-FCS
Unlike conventional fluorescence microscopy where both illumination and fluorescence are collected along the same axis, the collection objective in SPIM is placed perpendicularly to the excitation path. A thin While applications of SPIM-FCS for the study of TFs are still scarce, the dual-colour variant SPIM-FCCS was used to map the mobility and interactions of FOS-eGFP This allows 3D imaging of biological samples with a high frame rate and micron-scale spatial resolution in the axial direction.
Sample preparation is more complex than for confocal microscopy and requires special holders to enable the use of two objectives. Structural inhomogeneities in the sample produce 'shadows'

Advantages
Disadvantages light sheet (0.5-1.5 µm thick) is created using a cylindrical lens to focus light in one direction only (the sheet thickness), while leaving it to spread in the other two axes (the sheet width). Fluorescence intensity fluctuations are recorded from a plane in the sample using fast scientific complementary metal oxide semiconductor (sCMOS) or EM-CCD cameras (Capoulade et al., 2011;Krieger et al., 2015;Krieger et al., 2014;Langowski, 2017;Wohland et al., 2010). and JUN-mRFP1 transcription factors in live cells (Pernus and Langowski, 2015).
behind the obstructions, creating a light sheet of uneven intensity. Non-uniformity of light sheet thickness over longer distances limits the area from which a uniform signal can be recorded.
Image correlation spectroscopy (ICS) and image cross-correlation spectroscopy (ICCS) ICS methodologies extract quantitative information about molecular concentration, mobility and oligomerisation by calculating a correlation function from fluorescence intensity fluctuations as a function of space across an image acquired by confocal laser scanning microscopy (CLSM) or across space and time from a CLSM imaging time series (Digman et al., 2005;Mets et al., 2020;Wiseman, 2013;Wiseman, 2015). The dual-colour variant of ICS, Image crosscorrelation spectroscopy (ICCS) can quantify molecular interactions from the analysis of twochannel CLSM images (Comeau et al., 2008), whereas pair correlation function (pCF) analysis can be used to study long-range intracellular molecular diffusion, detect the presence of barriers to diffusion and produce a map of transport of chemically inert molecules in live cell nuclei (Hinde et al., 2010). Intra-and inter-cellular analysis of TF movement and in vivo flow was characterised in a multicellular organism (Clark et al., 2016). Additionally, ligandinduced TF translocation and diffusion maps have been obtained (Mikuni et al., 2017).
Quantitative characterization of TF concentration, mobility and homo-oligomerisation can be achieved using well-established techniques and nonspecialised microscope systems.
ICS and ICCS perform well only on relatively homogenous samples, since the presence of regions that are significantly brighter than the average fluorescent species biases the correlation function (Hendrix et al., 2016). Large number of pixels, n>64, need to be considered for an accurate spatial correlation analysis.

Single-particle tracking/single-molecule tracking (SPT/SMT)
Tracks individual fluorescently-labelled particles (molecules, vesicles, virions or other molecular complexes) in order to obtain information on their dynamic behaviour over time. By tracking individual TF molecules in a specified/ observation area/volume, the trajectories are visualised and displacements of TFs can be measured.
SPT visualises the displacement of individual molecules. Because only a small subset of TF molecules are fluorescent in the sample, the total TF concentration is not a limiting factor for the success of the measurements. The spatial and temporal resolutions of this method are limited by the resolution and image acquisition speed of the microscopy setup. Analysis of TF displacement relies on the diffusion and kinetics model chosen to fit the data, such as the number of different diffusion behaviours (Mazza et al., 2012).

Fluorescence recovery after photobleaching (FRAP)
Within a small selected region of the sample, intense laser light quickly renders the fluorescent molecule unable to fluoresce. The rate of fluorescence recovery in the photobleached region is recorded using a low-intensity laser beam to excite the notbleached molecules, and fitted with suitable models to derive the diffusion and/or the off-rates (Axelrod, 1981;White and Stelzer, 1999).
FRAP allows the measurement of TF diffusion and chromatin-binding behaviour and is used to derive the mobile (freely diffusing) versus immobile, e.g.
FRAP is used to determine the freely diffusing or chromatin-binding behaviour of TF molecules, even at higher concentrations (that are not accessible by FCS) and can, therefore, be used to investigate TF diffusion even in highly expressing cells and within condensates. FRAP is easy to perform and easy to understand intuitively. The imaging instrumentation used may be too slow for the recording of rapidly diffusing molecules.

Förster resonance energy transfer (FRET)
Interacting molecules are fluorescently labelled with distinct fluorophores selected so that the emission spectrum of one of the fluorophores, called a donor, overlaps with the absorption spectrum of the other fluorophore, called an acceptor. The transfer of energy from an excited donor to an acceptor molecule leads to acceptor molecule excitation and subsequent emission of fluorescence. This allows us to measure the physical proximity (≤10 nm) between the molecules (Forster, 1946).
This method -although strictly speaking being a proximity assay -is often used to study binding and complex-formation between different TFs, such as the formation of hetero-and homo-dimers, and interactions with chromatin (Evers et al., 2007;Mivelaz et al., 2020;Musselman et al., 2013;Robertson et al., 2018;Vancraenenbroeck and Hofmann, 2018;Wang et al., 2006;reviewed by Day et al., 2001). FRET has also been used to study ternary TF complexes (Demarco et al., 2006;Shyu et al., 2008;Sun et al., 2010). However, FRET is indispensable for the study of conformational changes, in which case the donor and acceptor molecules reside on the same TF molecule (Vamosi et al., 2008).
FRET imaging is the first super-resolution fluorescence microscopy technique, as it allows imaging of distances <10 nm. FRET is simple to configure on wide-field/CLSM fluorescence microscopes that are available in most labs. The spectral crosstalk between donor and acceptor fluorophores makes it difficult to obtain unambiguous FRET results. Considerable control experiments and image processing are needed to subtract crosstalk components.

Bimolecular fluorescence complementation (BiFC)
Interacting molecules are tagged with complementary, symmetrically/asymmetrically split fluorescent protein halves that are, on their own, not fluorescent. Formation of complexes leads to assembly of fluorescent proteins; interactions are inferred from the reconstitution of fluorescence (Hu et al., 2002;Hu and Kerppola, 2003).
The hetero-and homodimers are visualised whereas the monomeric partners are not.
False-positive signal can be obtained due to reconstitution of complementary fluorescent protein halves, without interactions between the investigated molecules. BiFC is not be best suited to study reversible binding since the reconstituted fluorescent protein does not dissociate easily.
Confocal spinning disc (CSD) microscopy Multipoint excitation is obtained by using a pair of rotating disks. The collector disk contains microlenses, whereas the pinhole disk contains pinholes. This way, each micro-lens has an associated coaligned pinhole, which is axially positioned at the focal plane of the micro-lens. As the discs spin, an array of focused laser beams scans across the specimens (Hayashi and Okada, 2015;Jonkman and Brown, 2015;Petran et al., 1986). CSD is used for TF imaging and localisation, and it can be combined with other methodologies to study TF mobility and binding dynamics (Hayashi and Okada, 2015;Jonkman and Brown, 2015;Petran et al., 1986;Schueder et al., 2017;Zhan et al., 2014).
CSD provides higher temporal resolution as compared to classic point scanning confocal microscopy, thus allowing for imaging of fast dynamic processes in live specimens. As in mpFCS, out-of-focus light, originating from molecules in remote focal planes above/below the focal plane, can pass through adjacent pinholes and increase the background signal.

2-/multi-photon excitation (2PE/MPE) microscopy
Out-of-focus excitation of fluorophores in the axial direction is significantly reduced, compared to that of single-photon excitation, by using two (or more) lower-energy photons (Denk et al., 1990; reviewed by So et al., 2000). This is used for imaging and localisation of TFs and it can be combined with methods to measure TF mobility and chromatin-binding dynamics (Denk et al., 1990;Ferguson et al., 2012;Guan et al., 2015;Yao et al., 2006).
The use of multi-photon excitation enhances penetration depth, thus facilitating imaging deeper within tissues. Background fluorescence is reduced by decreasing out-of-focus excitation.
Lateral spatial resolution is somewhat worse than in single-photon excitation, as light with a longer wavelength is used to excite the fluorescent molecules.
TIRF similarly allows for less phototoxicity and less photobleaching, and high signal-to-noise ratio. The limited penetration depth of the evanescent wave only permits imaging and studying of fluorescent proteins at the cell surface, the plasma membrane at the adhesive surface or in vitro.
As a result, structures located deeper within cells can usually not be visualised. Therefore, for cells bearing nuclei deeper within the cell, the study of TF behaviour can be challenging.

Structured illumination microscopy (SIM, 3D-SIM)
A super-resolution fluorescence imaging technique where a movable diffraction grating is placed in the path of the laser beam and the sample is excited with a known, spatially structured pattern of light. Several images -the number of which depends on width and spacing of the pattern lines -are acquired by changing the angle and orientation of the pattern, processed and mathematically deconvolved to yield a super-resolution image (Gustafsson, 2000;Gustafsson, 2005). This method is used for TF imaging, localisation and the measurement of TF molecular movement and dynamic binding to chromatin, in combination with other methodologies (Gustafsson, 2000;Gustafsson, 2005;Patel et al., 2013;Penrad-Mobayed et al., 2018).
SIM is based on wide-field microscopy. By altering the pattern frequency, the axial resolution can be adjusted. The lateral resolution in SIM is ∼100 nm; and the axial resolution is Fluorophores used for conventional microscopy can be used for SIM imaging. Fast processes can be studied in live cells and dual-colour SIM imaging is possible.
Errors in the grating and mismatch of refractive indexes can lead to errors in image reconstruction.
SIM is very sensitive to out of focus light and the signal is not clear cells overexpressing fluorescently tagged TF molecules cells are being imaged. Image reconstruction is slow, requiring ∼1-30 s of processing time.

Stimulated emission depletion (STED) and 3D STED
An excitation laser beam with a Gaussian intensity profile is well-aligned with a red-shifted STED laser beam, characterised by a 'doughnut-shaped' intensity profile that is close-to-zero at the centre and of high intensity at the circumference. These superimposed laser beams are raster-scanned across the sample. The Gaussian beam excites fluorescent molecules. However, before the excited molecules emit a fluorescence photon, the STED laser stimulates them to return to ground state by emitting a photon with exactly the same energy as the STED laser. This, effectively, reduces the probability of fluorescence emission and only molecules located in the very centre of the STED beam -which are not efficiently stimulated to return to the ground state, since there the STED beam intensity is low -can emit fluorescence. This, in turn, increases the resolution past the diffraction limit (Donnert et al., 2007;Hell and Wichmann, 1994;Willig et al., 2007). This method can localise TFs with a precision that is, for measurements in live cells, up to an order of magnitude better than confocal microscopy (50-100 nm) (Donnert et al., 2007;Hell and Wichmann, 1994;Heller et al., 2013;Li et al., 2019;Willig et al., 2007). In combination with other methodologies, e.g. STED-FCS (Eggeling et al., 2009;Kastrup et al., 2005), TF movement and dynamic binding to chromatin can be characterised.
STED, in its original realisation, greatly improves the lateral resolution, leaving the axial resolution unchanged. Axial resolution and lateral resolution are improved in 3D-STED. STED images are acquired directly, and image reconstruction and processing are not required, although processing is often applied to enhance the contrast. STED microscopy imaging can be performed in live cells. The power of the STED laser beam determines the spatial resolution, and laser power of 5-200 MW/ cm 2 is needed to achieve the highest possible spatial resolution. This, in turn, may lead to photobleaching and phototoxicity, when measurements are performed in live cells. Only some fluorophores are sufficiently photostable and compatible with the STED lasers available in STED microscopes. STED microscopy is a raster scanning methodology. The temporal resolution of STED microscopy imaging is, therefore, low.
Stochastic optical reconstruction microscopy (STORM), direct STORM (dSTORM), photoactivated localisation microscopy (PALM) and fluorescence PALM (fPALM) This group of super-resolution fluorescence microscopy methodologies relies on the use of photo-switchable fluorophores that can undergo reversible changes in their intrinsic properties through the use of light as an external stimulus. At the beginning of the imaging process, all fluorophores are in the so-called 'off' state, which is nonfluorescing (dark). A small subset of molecules residing at longer intermolecular distances than These methods can locate TFs with a precision in the tens of nanometres range (∼50 nm) (PALM (Cai et al., 2019;Gebhardt et al., 2013;Hansen et al., 2017;Izeddin et al., 2014;Liu et al., 2014); fPALM (Akamatsu et al., 2017;Spacek et al., 2017); dSTORM (Gao et al., 2017;Hansen et al., 2017).
PALM and STORM allow for the highest spatial resolution of all optical microscopy methodologies discussed and are quantitative, as they are able to detect single molecules and count their numbers.
Multi-colour imaging is also possible.
These methods require particular fluorescent labelling, e.g. photo-activatable and photoswitchable fluorophores or dyes. Reconstruction of the images necessitates many on-off cycles (>10,000), which limits the application of these methodologies to slow processes. Image processing is time consuming. Very high sample labelling density is needed -the Nyquist-Shannon  (Betzig et al., 2006); fPALM ] or organic dyes [STORM (Churchman et al., 2005;Rust et al., 2006); dSTORM (Heilemann et al., 2008)] are being used, the localization accuracy of a single molecule depends on the number of photons being detected from that molecule.
sampling frequency of a signal that will not distort its underlying information, should be double the frequency of its highest frequency component. According to this theorem a dye molecule should be present at least every 10 nm to achieve an image resolution of 20 nm. Thus, immuno-based labelling approaches may result in a spotty image due to the size (∼≥20-30 nm) of the primary and secondary antibodies.
Atomic force microscopy (AFM) AFM is the collective term for an ever growing number of techniques characterised by the use of a mechanical probe to perform force measurement, topographic imaging or controlled sample manipulation (e.g. local stimulation of cells). In the most common AFM realization, a cantilever moving over the sample is deflected by changes at the sample surface. This allows the measurement of topology and forces due to binding (Binnig et al., 1986;Butt et al., 1990;Dufrene et al., 2017;Scheuring et al., 2001).
Visualisation and manipulation, e.g. mechanical dissection, of cell surfaces, subcellular compartments or chromatin. TF-chromatin binding can be visualised and the chromatin-binding kinetics measured in vitro. DNA is immobilised, e.g. using DNA origami scaffolding, and the properties of TF-DNA complex formation can be studied (Kierzek et al., 2001;Sanchez et al., 2011;Yamamoto et al., 2014).
AFM is characterised by a high spatial resolution at the nanometre level and a high signal-to-noise ratio. It neither requires fluorescent labelling of the studied TFs nor sample fixation. Assessment of multiple physical, chemical and biological parameters is possible (Binnig et al., 1986;Butt et al., 1990;Kierzek et al., 2001;Sanchez et al., 2011;Scheuring et al., 2001;Shaffer et al., 2017;Yamamoto et al., 2014). AFM is restricted to surfaces. Measurement artefacts from the cantilever tip or the applied force upon contacting the sample are common. The temporal resolution of AFM imaging is low -in the secondsto-minutes range (Dufrene et al., 2017).
SOX2 and the chromatin reader BRD4, their chromatin-binding dynamics, as well as their interactions with the Mediator complex and the elongation controller P-TEFb, a cyclin-dependent kinase consisting of CDK9 and one of several cyclin subunits . The subnuclear localisation of TFs may also vary between cell cycle stages and many TFs (including SOX2) become enriched on mitotic chromosomes, a phenomenon referred to as 'mitotic bookmarking' (Caravaca et al., 2013;Dufourt et al., 2018;Kadauke et al., 2012;Young et al., 2007). This local enrichment is mediated by the active nuclear import of TFs (Teves et al., 2016). TF binding persists through cell divisions, thus conferring transcriptional 'memory'. At the cellular level, the concentration of TFs may provide cells with a molecular 'readout' of their relative positions in a tissue. For example, in Drosophila embryos, the chromatinbinding activity of Bicoid (Bcd), a maternally provided TF and key morphogen, which determines embryonic anteroposterior polarity, remains unaltered despite its low levels within posterior nuclei. This is achieved when Bcd binds hubs together with the pioneer TF Zelda (Zld), which results in increased local concentrations . Thus, while distinct concentrations of TFs may elicit differential cell fates, cells can deploy alternative mechanisms to preferentially counteract the effect of different TF concentrations when needed.
Although TF function has been extensively studied in flies, the latter are a rather inappropriate model to study TF HIs, as the majority of heterozygous mutations do not cause observable developmental defects. This may be explained by the fact that the Drosophila genome is ∼20 times smaller than the human genome, but the average concentration of TFs regulating this smaller genome is not proportionately decreased. Additionally, extensive endoreplication, which takes place in several fly tissues, produces polyploid cells during development, which are expected to be able to better buffer low TF gene dosage than diploid cells (Sher et al., 2013). However, gene regulation by TFs has been studied very extensively in this organism, therefore, we discuss crucial findings from Drosophila. Taken together, the discussed findings support the notion that the absolute number of TFs is an important determinant affecting gene expression by means of various mechanisms.
The importance of TF numbers and their subnuclear localisation for normal function are an enticing subject of research. The absolute number of TFs can be measured in living or fixed cells by fluorescence correlation spectroscopy (FCS) or super-resolution microscopy (SRM), respectively (Table 2 and Fig. 2). FCS allows the measurement of TF numbers and their molecular mobility with high temporal resolution (Ehrenberg and Rigler, 1974;Papadopoulos et al., 2010;Vukojevic et al., 2010). Additionally, the spatial variability of TF numbers within or between cell nuclei can be simultaneously analysed by employing massively parallel FCS (mpFCS) that uses an array of detectors coupled to diffractive optical elements (EM-CCDs) (

2015)
, or by selective plane illumination microscopy (SPIM) coupled to fast cameras, such as electron-multiplying charge-coupled devices (EM-CCDs) (Capoulade et al., 2011;Krieger et al., 2015Krieger et al., , 2014Wohland et al., 2010). FCS has been used to study the dynamic behaviour of TF binding to chromatin, such as in the case of the formation of the Bcd gradient during early Drosophila development (Abu-Arish et al., 2010), the TF-chromatin binding dynamics of SOX2-OCT4 (Chen et al., 2014b), TetR (Normanno et al., 2015), the specific and non-specific binding of the Hox TF Sex combs reduced (Scr) Vukojevic et al., 2010) and the variable dynamic binding behaviour of the TF MYC (Rosales et al., 2013). FCS has also been employed to study the variability in TF concentration and how this leads to the acquisition of differential developmental fates as, for example, in the case of the TF Senseless in the Drosophila wing imaginal discs (Giri et al., 2020), FOS and JUN in HeLa cells (Szalóki et al., 2015) or the transcriptional co-activator Yorkie (the ortholog of the human YAP transcriptional co-activator) in different subcellular compartments of the developing Drosophila airways . Taken together, these studies show how diverse FCS methodologies can provide information on TF dynamic behaviour in live cells. SRM methodologies, however, provide higher spatial resolution to visualise the distribution of subnuclear TFs and, therefore, complement FCS. Despite lacking the temporal and quantitative information, which FCS allows to derive by studying live samples, SRM methodologies provide excellent spatial information, which FCS methodologies cannot. Single-molecule spatial resolution may be achieved by limiting the number of fluorescent light-emitting molecules through selective illumination and/or sparse excitation (see Table 2 and Fig. 2). In these cases, TFs can be located − with high precisionon chromatin, whereas the concentration of TFs can still be inferred semi-quantitatively, by comparing their fluorescence intensity levels with those derived from a sample of known TF concentration (Lasker et al., 2020;Li et al., 2019;Reisser et al., 2018;Wollman and Leake, 2015). Furthermore, the quantification of TF transcripts can provide information about relative gene expression levels and, in some cases, may be used as a proxy to quantify TF abundance in a population of cells. Such relative measures of TF messages, their target transcriptomes and/or cell-to-cell heterogeneities of transcript numbers can be obtained by detection of RNA in single-cells. These measures scale with, albeit do not necessarily predict, TF variability at protein level and include single-molecule fluorescence in situ hybridisation (smFISH) and single-cell RNA sequencing (scRNA-seq). smFISH has been used to measure levels of FOXO mRNA (Blice-Baum et al., 2019), NKX2-2 mRNA distribution in mouse pancreatic islet cells (Cui et al., 2018) and OCT4/SOX1/T-BRACHYURY transcription in differentiating mouse ESCs (Lanctôt, 2015). A main advantage of scRNA-seq is that it can distinguish between cells on the basis of their expressed mRNA repertoire (transcriptome profiling), and can measure the levels of individual TF messages, such as of OCT4/ SOX2 in developing mouse embryos (Goolam et al., 2016), or PAX6 in a subclass of mouse cortical cells (Zeisel et al., 2015). The combination of microscopy-based and transcriptomics modalities allows researchers to precisely quantify the cell-to-cell variability of TF concentration at the protein or mRNA level, as well as to assess how such variability affects their target transcriptome. Overall, one can link the absolute amount of TFs to their function in transcriptional regulation. To understand how aberrant levels of TFs may lead to HI phenotypes, the absolute number of TFs needs to be efficiently measured in single cells and between different cells in a tissue. Therefore, quantitative methodologies need continuous advancement to become routinely accessible and accommodate a continuously increasing demand for precision.

Stoichiometry of TF complexes
TFs rarely act alone. They predominantly bind target enhancers jointly with other TFs or cofactors to regulate the transcription of their target genes (Golan-Lagziel et al., 2018;Veitia, 2002). Hence, the formation of TF complexes depends on TF numbers, a phenomenon that appears to be conserved in evolution (Papp et al., 2003;Sopko et al., 2006;Veitia, 2002Veitia, , 2003. A study by Ni et al. showed that dosage-sensitive TFswhich often drive HI phenotypesinteract with a total of 851 proteins, 25% of which are other TFs (Ni et al., 2019). Dosage-insensitive TFs, by contrast, interact only with 263 proteins, of which 17% are other TFs (Ni et al., 2019). Whether such TF-TF interactions are required for TF function and whether HIs of some TF genes are partially caused by abnormal TF stoichiometry within complexes remain to be shown. However, this disparity is striking and might shed light into the widespread downstream transcriptional effects triggered by reduced levels of individual TFs in HIs.
Many TF complexes display cooperative binding to chromatin, rendering the maintenance of stable cell type-specific transcriptional programmes reasonably sensitive to TF numbers (Bartman et al., 2016;Fukaya et al., 2016). This might explain why either a reduction or increase in individual TF protein levels can have deleterious effects on normal cellular function (Chen et al., 2008;Cox et al., 2010;Gao et al., 2012;Papp et al., 2003;Veitia, 2002;Wang et al., 2006;Wuebben et al., 2012). An extensively studied example of this is regulation of the ESC state by distinct TF combinations and levels (Chen et al., 2008;Cox et al., 2010;Gao et al., 2012;Wuebben et al., 2012). A second example is the cooperative binding between PAX6 and SOX2 during optic field induction. Here, both Pax6 overexpression (e.g. through a third-copy allele) and heterozygous pathogenic mutations result in ocular developmental defects that are similar between mice (Ouyang et al., 2006;Schedl et al., 1996) and humans (Gerth-Kahlert et al., 2013;Hall et al., 2019)see Fig. 3 for developmental abnormalities in the eye, caused by heterozygous PAX6 and SOX2 mutations. Such strong dependence on TF abundance, in which either increased or decreased TF numbers trigger a similar abnormality, is particularly interesting and may hold true for other, less well-studied TF genes. In addition to cooperative binding, TFs may recruitor assist the binding ofother factors to chromatin. 'Pioneer' TFs, such as the HIassociated forkheadbox A1 and 2 (FOXA1/2) and SPI1 (also known as PU.1), bind 'closed' chromatin, making it accessible to non-pioneer TFs that act either alone or through interactions with chromatin remodelers (Barozzi et al., 2014;Cirillo et al., 2002;Heinz et al., 2010;Li et al., 2012;Zaret, 2018).
The functional impairment of protein complexes in disease has been recently summarised by Bergendahl et al., 2019, who discuss possible mechanisms by which disease-causing mutations alter protein structure. This, in turn, may result in aberrant complex formation by affecting the protein stoichiometry or by inhibiting other protein interactions, thus causing abnormal cell and tissue functions, ultimately leading to disease. Therefore, the number of TFs may influence the function of other TFs when transcriptional regulation depends on TF-cofactor complexes.
Single-molecule analysis of the stoichiometry of TF complexes can be performed in both living or fixed cells. The proximity of molecules, as an implicit measure of complex formation, can be analysed by Förster resonance energy transfer (FRET) (Dikovskaya et al., 2019;Szalóki et al., 2015) or bimolecular fluorescence complementation (BiFC) (Hu and Kerppola, 2003;Moustaqil et al., 2018;Papadopoulos et al., 2015) (see Table 2 and Fig. 2). However, steric hindrance and the need for correct three-dimensional orientation of TF complexes are main limiting factors of these methods, and may cause false-negative results. FRET, FCS and combined SPIMfluorescence cross-correlation spectroscopy (SPIM-FCCS) have been used in combination to show that FOS forms homodimers that are also capable to bind chromatin when JUN is not present in equal concentrations to FOS (Szalóki et al., 2015). As the formation and transcriptional regulatory activity of FOS homodimers were suggested to play a role in oncogenesis, the stoichiometry of TF complexes and their activity have been investigated by combining different quantitative approaches in the context of FOS overexpression (Szalóki et al., 2015). TF complexes at specific nuclear compartments or gene loci, which are rendered visible in BiFC or undergo energy transfer in FRET, may also be imaged by using SRM (Kwon et al., 2017;Yamamoto et al., 2014) (see Table 2 and Fig. 2). Furthermore, one can analyse how the stoichiometry of TF complexes changes during development or varies across cells in the same tissue. This is possible by using FCS, as in studies regarding the homodimerization of Scr on chromatin (Papadopoulos et al., 2015;Rosales et al., 2013;Skouloudaki et al., 2019;Szalóki et al., 2015;Vukojevic et al., 2010) andFOS (Szalóki et al., 2015). FCS has also been applied to study the formation of heterocomplexes between MYC and the transcriptional regulator MAX (Rosales et al., 2013). Oligomerisation of the mitochondrial TF TFAM on chromatin was similarly investigated using fast-scanning stimulated emission depletion (STED) microscopy (Heller et al., 2013). Moreover, TF complexes have been studied by using dual labelling strategies. For example, Savatier and co-workers used twophoton FCCS to investigate the interactions between alpha and beta estrogen receptors (ESR1 and ESR2, respectively) or between ESRs and TIF2 (Savatier et al., 2010). The colocalisation of glucocorticoid receptor (NR3C1) and its interacting protein GRIP1, as well as of the transcriptional activators BMAL (aslo known as ARNTL) and CLOCK was similarly visualised by using dual-colour singleparticle tracking (SPT) (Gebhardt et al., 2013). Thus, TF complexes can be quantitatively analysed in time and in space to provide molecular insights into how their interactions and concentrations affect entire regulatory networks. To uncover the roles of TFs in HI, their interactome needs to be thoroughly investigated and quantitative microscopy methods should be used to characterise the potentiially abnormal TF function based upon aberrant TF stoichiometries.

Target site number, accessibility and binding
The cell-and tissue-specific sensitivity to TF levels also depend on the number of TF target-binding sites on chromatin and their accessibilitywhich might be subject to developmental control, as well as how TF molecules dynamically move in nuclei and undergo various interactions with chromatin, until they find and bind to their target sites. On average, the number of TF molecules per cell is between thousands and tens-of-thousands (Simicevic et al., 2013). This is roughly one order of magnitude higher than the number of specific TF binding sites, which is on average several hundreds to a few thousands (Ganapathi et al., 2011;Johnson et al., 2007;Palii et al., 2011;Robertson et al., 2007). Therefore, the number of binding sites of a given TF in a certain cell type cannot always explain why reduced TF amounts may not suffice for proper TF function. Additionally, genes controlled by a single TF may be more sensitive to molecular numbers than those controlled by two or more TFs, making the correlation between the 'number of TF molecules' and the 'number of TF-binding sites' far from straightforward. The complexity increases when we consider the existence of mechanisms, which overcome the dependence of transcriptional control on the stoichiometry of TF molecules to binding sites, i.e. by allowing the expression of important developmental genes to also depend on collaborating TFs. This is achieved by the evolutionary 'addition' of such 'collaborating' TF-binding sites into enhancers. For instance, the presence of both Zld-and Dorsal (Dl)-binding sites in an enhancer of the short gastrulation (sog) gene guarantees the potentiation of sog transcription in the early the Drosophila embryo, such that all cells express the same amount of sogeven in positions where the Dl morphogen TF concentration is low Yamada et al., 2019). In this case, additional Zld-binding sites convert sog expression from 'analog'i.e. proportional to Dl concentration along the Dl dorsoventral gradientto 'digital', whereby Zld guarantees efficient sog expression as long as Zld-binding sites are present in the sog enhancer Yamada et al., 2019). Thus, studies from the fly shed light onto how some target genes maintain robust expression, even upon limiting amounts of the TFs that regulate them. Such collaborating and 'potentiating' TFs may themselves be expressed in a tissuespecific manner. In this case, expression of the same target gene across different tissues or of different target genes in the same tissue both depend on the amount of TF. This might explain why certain TFs result in HIs in one tissue but not in another. Generally, the number of TF-binding sites may strongly influence the expression of specific genes in the same tissue or of the same gene in some tissues only and not in others, making gene expression less, or more, sensitive to TF numbers. Therefore, TF HIs that result in developmental abnormalities or disease frequently manifest in a subset of tissues, at distinct developmental stages or are due to the mis-regulation of only some of their target genes. The binding affinity of TFs on their DNA-binding sites is another important factor required for normal gene regulation (Levine, 2010). This chiefly depends on the binding site sequence and defines how 'strongly' a TF binds to this site. Low-affinity binding sites are, thus, less likely to be bound when TF concentrations are low and, therefore, may require higher numbers of TFs and/or cofactors for transcriptional regulation (Arthur et al., 2017;Tsai et al., 2017). This may serve as an additional control mechanism for differential gene expression in space and time. In Drosophila, the Hox TF Ultrabithorax can bind low-affinity target sites on certain shavenbaby enhancers when complexed with the Hox cofactors Extradenticle and Homothorax. This confers region-specific control for the formation of epidermal denticle belts in the embryonic cuticle (Crocker et al., 2015;Tsai et al., 2017). While low-affinity binding sites render gene expression more sensitive to TF levels, their existence can be important for safeguarding the specificity of TF binding. This is of particular importance for the specification, development and differentiation of body structures. Additionally, it is an important mechanism to control and exploit the function of TFs that exhibit very similar binding behaviours, such as Hox TFs. As a result, the transcriptional output of genes with multiple low-affinity binding sites can be dosage dependent (Crocker et al., 2015;Driever and Nüsslein-Volhard, 1988;Giorgetti et al., 2010;Lorberbaum and Barolo, 2013;Ochoa-Espinosa et al., 2005;Ramos and Barolo, 2013;Stewart-Ornstein et al., 2013;Struhl et al., 1989;Tsai et al., 2017). This, in turn, enables cells to acquire distinct developmental fates based on their position along a TF concentration gradient (Driever and Nüsslein-Volhard, 1988;Lorberbaum and Barolo, 2013;Ochoa-Espinosa et al., 2005;Ramos and Barolo, 2013;Struhl et al., 1989). For instance, the homeodomain TF CUX1, a tumour suppressor, binds distal target enhancers in cultured human cells in concert with co-activators and cohesin. Loss of one CUX1 allele abolishes CUX1 binding, and genes that contain multiple low-affinity CUX1-binding sites become mis-regulated (Arthur et al., 2017). Furthermore, Louphrasitthiphol et al. (2020) recently investigated the binding between microphthalmia-associated transcription factor (MITF) and chromatin, and suggested that low-affinity binding sites also act as a TF 'reservoir' in the genome. Chromatin immunoprecipitation sequencing (ChIP-seq) and single-molecule tracking (SMT) were used to show that the binding affinity between MITF and chromatin is increased by MITF acetylation, and that an acetylation-mutant MITF exhibits increased numbers of transient interactions with chromatin at low-affinity binding sites (Louphrasitthiphol et al., 2020). The authors suggested that low-affinity sites act as a reservoir for nonacetylated MITF, which is released upon acetylation in order to bind high-affinity sites (Louphrasitthiphol et al., 2020). Although both high-and low-affinity binding sites can favour TF specificity and facilitate differential gene expression, the number of TF-binding sites in enhancers, as well as their affinity, may also strongly affect the sensitivity of gene expression relative to TFs and, therefore, may result in TF HI.
Yet, cell-or tissue-specific transcriptional responsiveness of genes relative to the level of certain TFs may stem from the differential accessibility of their binding sites, and the set of TFs and cofactors targeting them. Both processes contribute to the kinetics of TF target site search and binding (Liu et al., 2014;Tsai et al., 2017). To find their specific binding sites, TFs 'search' chromatin by diffusion and transient, non-specific electrostatic interactions (Chen et al., 2014b;Elf et al., 2007;Liu et al., 2014;Louphrasitthiphol et al., 2020;Papadopoulos et al., 2010;Raccaud et al., 2019;Slutsky and Mirny, 2004;Voss et al., 2011;Vukojevic et al., 2010). Specific binding to their cognate sites depends on the strength of weak dipole interactions, such as hydrogen bonds and Van der Waals forces, which are exercised between amino acid (aa) residues of the TF and the DNA bases. Thus, the balance between non-specific and specific chromatin interactions determines the amount of TF molecules available to bind their cognate sitesan important determinant of how efficiently TFs find these specific binding sites.
The accessibility of binding sites depends on chromatin conformation and nuclear organisation, and may modulate the ability of TFs to efficiently search for and find their target sites. Chromatin organisation within cell nuclei is dynamic, cell typespecific and has been extensively studied with respect to its effect on gene expression. The degree of chromatin compaction affects TF binding and gene expression (Akhtar et al., 2013). Additionally, topologically associated domains (TADs), i.e. regions of selfinteracting chromatin, are thought to regulate gene expression by establishing chromatin contacts within and between TADs, as well as by generating transcriptional microenvironments (Dixon et al., 2015;Tsai et al., 2017;reviewed by Gonzalez-Sandoval and Gasser, 2016). For example, Tsai and collaborators suggested that, in Drosophila, clustering of enhancers with low-affinity binding sites permits the formation of high local concentrations of the TF Ultrabithorax, which potentiates its binding interactions with chromatin (Tsai et al., 2017). TADs are also dynamic and cell type-specific, and have been studied with regard to differentiation (Kaur et al., 2013;Plachta et al., 2011;White et al., 2016), reprogramming (Beagan et al., 2016) and their role in HOX gene expression during limb development (Gebhardt et al., 2013;Langowski, 2017;Paakinaho et al., 2017;Rehó et al., 2020;Savatier et al., 2010;Vámosi et al., 2008). Nucleosome positioning further influences chromatin accessibility and, thereby, TF binding to gene regulatory elements (Chen et al., 2014a,b;Gebhardt et al., 2013;Hansen et al., 2017;Li et al., 2019;Mir et al., 2017;Normanno et al., 2015;Tirosh and Barkai, 2008;White et al., 2016;Wollman et al., 2019;Yao et al., 2006;Zhao et al., 2017). At the whole-genome level, histone modification signatures dictate chromatin accessibility and are subjected to developmental control (Bell et al., 2010). As such, the effect of binding site accessibility on TF-chromatin binding kinetics may trigger differential gene expression and differentiation biases between seemingly identical cells, even when the TFs mainly responsible for such processes are expressed in all cells at the same levels. In mouse preimplantation embryos, OCT4 exhibits differential chromatin-binding kinetics as early as at the 8-cell stage, which is one of the first determinants of differentiation bias towards inner cell mass or trophectoderm (Plachta et al., 2011). Moreover, the chromatin-binding kinetics of SOX2 and OCT4 can be used to predict cell fate (White et al., 2016). In pluripotent cells, OCT4 and SOX2 bind chromatin more stably than in extraembryonic cells (Kaur et al., 2013). Long-lived SOX2-chromatin binding is regulated by methylation of histone H3 at arginine 26 (H3R26), such that reduction in H3R26 methylation decreases the lifetime of SOX2-bound complexes. As a result, SOX2 target expression declines, and so do the numbers of pluripotent cells (White et al., 2016). These are excellent examples of how the kinetics of TF binding and the epigenetic landscape can influence development. Therefore, the dependence of gene expression on TF concentrations in a cell-and tissue-specific manner also relies on the number, affinity, accessibility and chromatin-binding kinetics of TF-binding sites.
It follows that quantitative analyses of TF-chromatin interaction kinetics and studies of cognate site configuration provide a deeper insight into gene-and cell-specific sensitivity to TF levels. FCS and SPT are used to obtain the fractions of freely diffusing TF molecules that seem to facilitate cognate site search, as well as of TF molecules bound to chromatin (see Fig. 2 and Table 2). For example, FCS studies of Scr in flies suggest that slowly diffusing TFs predominantly engage in rapid electrostatic interactions with chromatin, thereby reflecting the molecular movement TF molecules exhibit while searching for their specific DNA-binding sites Vukojevic et al., 2010). Interestingly, the diffusion of MYC and P-TEFbas measured by SPTsuggests that a binding site search may be TF-and target site-specific (Izeddin et al., 2014). By using SPT, enhancer clustering has been suggested to facilitate target-site search and binding (Liu et al., 2014). Chen and collaborators combined multifocus SMT and FCS to study how mutant and wild-type SOX2 and OCT4 in mouse ESCs display differential binding dynamics on enhanceosomes, as well as the effects chromatin modifications have on TF molecules searching for their cognate DNA sites (Chen et al., 2014b). Once bound, the DNA-residence time of a TF reflects its binding affinity. SPT is better suited to measure long-livedpresumably specific -TF-DNA interactions. Monitoring of these interactions does not require very high temporal resolution and photobleaching can, thus, be avoided by time-lapse imaging experiments with longer 'dark' intervals between rounds of image acquisition. For example, by using reflected light-sheet microscopy (RLSM)-SPT, Gebhardt and colleagues compared the DNAresidence time of NR3C1 and of ESR monomers and dimers (Gebhardt et al., 2013). Mazza and collaborators combined SPT with FCS and FRAP to identify different proportions and residence times of mutant p53 (Mazza et al., 2012). FRAP can also be used to characterise TF-chromatin interactions by investigating their kinetic on-and off-chromatin rates, for instance, how TFs bind mitotic chromosomes. Here, the TFs exhibit high on-rates and reduced mobility during interphase, resulting in a more efficient search of TFs for binding sites (Raccaud et al., 2019). Furthermore, application of FRAP and SPT showed that SOX2-chromatin binding behaviour during mitosis is more dynamic than during interphase (Teves et al., 2016). FRAP experiments also identified that mutations in the high-mobility group domain of SOX2 largely abolished its clustering at the POU5F1 enhancer, again suggesting that TFs can engage in different modes of chromatin-binding behaviour . In addition, the chromatin-binding behaviour of TF heterocomplexes with other TFs has been studied by FCCS; particularly in cases where TF heterodimerisation is known to be required for DNA binding, such as for the dimers between FOS and JUN (Langowski, 2017), retinoic acid receptor (RAR) and retinoid X receptor (RXR) (Rehó et al., 2020), and ESR1/2 and TIF2 (Savatier et al., 2010). FRET has similarly been used to study chromatin binding of TF dimers and complexes, in some studies in combination with FCS or FCCS. Dimerisation and chromatin-bound intermediates of the basic helix-loop-helix leucine zipper (bHLH-LZ) domains of the TFs MYC and MAX were investigated by single-molecule FRET (smFRET) and FCS (Vancraenenbroeck and Hofmann, 2018). FCCS and FRET have also been used to study the chromatin-binding dynamics and intermolecular distance between FOS and JUN in homo-and heterodimers (Szalóki et al., 2015;Vámosi et al., 2008). Additionally, ternary complexes between FOS, JUN and the NFκB subunit p65 have been investigated by BiFC-FRET (Shyu et al., 2008).
To address these challenges, CRISPR/dCas9-and TALE-based labeling of specific DNA sequences can be used in live cells to visualise TF dynamics. However, the low signal-to-noise ratio has so far restricted this application to repetitive sequences (Chen et al., 2013;Knight et al., 2015;Ma et al., 2013;Thanisch et al., 2014). In a recent study, Li et al. targeted dCas9 to distal enhancers of the Pou5f1 and Nanog genes, to link clusters of BRD4 and enhancers with active transcription . However, the authors did not address the dynamic behaviour that might underlie such interactions. Specific conformations of TF-binding sites can be analysed using SRM in combination with chromatin markers associated with accessibility and active transcription (Cai et al., 2019). The kinetics and variability of TF-chromatin interactions can be analysed by quantitative microscopy at high temporal or spatial resolution. These methodologies allow to link the intercellular variability of TFchromatin interactions and gene expression to gene-and cell-specific target site availability, andfrom thereto further understand the role of TF-binding dynamics in the context of HIs.

Stochastic gene expression
Although highly significant, TF-chromatin interactions alone do not sufficiently explain the significance of TF stoichiometry underlying HIs. Aberrant TF numbers can also affect gene expression because this in itself is an inherently stochastic process. Transcription is a 'noisy' process and, for a plethora of genes, it occurs in bursts, resulting in temporally fluctuating levels of mRNA (Bothma et al., 2014;Chubb et al., 2006;Fukaya et al., 2016;Larsson et al., 2019;Lee et al., 2019;Rodriguez et al., 2019). Because the frequency of transcriptional bursts depends on TF concentration (Senecal et al., 2014), when TF numbers are low, transcriptional noise can become a limiting factor for the normal expression of some genes and may result in abnormal function. The aforementioned stochasticity further results in temporally and spatially variable gene expression within a cell, and elicits heterogeneity between cells in both prokaryotic (Kierzek et al., 2001;Ozbudak et al., 2002;Wolf and Arkin, 2002) and eukaryotic organisms (Blake et al., 2003;Raser and O'Shea, 2004). In mammals, gene expression noise occurs both during development (Abranches et al., 2014;Mohammed et al., 2017;Olsson et al., 2016;Trapnell et al., 2014) and in disease (Avraham et al., 2015;Shaffer et al., 2017;Tirosh et al., 2016).
In bacteria, auto-activation of the competence TF ComK confers resistance to environmental stress by transcribing genes involved in the uptake of DNA (Maamar and Dubnau, 2005;Smits et al., 2005;Süel et al., 2006). Perturbing ComK levels by upregulating the rate of transcription and downregulating translation decreased the noise of ComK-dependent gene expression and reduced overall competency. This observation underlines that a subset of cells may exploit this cellto-cell variabilityintroduced by transcriptional noiseto confer a group advantage (Maamar et al., 2007;Ozbudak et al., 2002;Süel et al., 2007;Thattai and van Oudenaarden, 2001). Albeit shown in a simple system, this indicates that stochasticity can result in transcriptional heterogeneity and increase the level of 'fitness' in a population. Similarly, in multicellular organisms, perturbations in TF numbers may amplify the cell-to-cell variability of transcriptional programmes. For example, in humans, deletions of one allele of the TF-encoding gene NKX3.1 have been described in ∼40% of prostate cancers (Bova et al., 1993;Bowen et al., 2000;Macoska et al., 1995). Mouse models recapitulating this deletion display NKX3.1 HI, aberrant target gene expression and progressive prostate diseases (Abdulkadir et al., 2002;Magee et al., 2003). Transcriptional 'bursting-like' behaviour and the resulting cell-to-cell variability have been specifically associated with HIs of tumour suppressor genes (Kemkemer et al., 2002). The inactivation of a single allele of NF1 results in neurofibromas in humans (Fahsold et al., 2000;Hoffmeyer et al., 1998). Furthermore, cultured melanocytes derived from patients with a single functional NF1 allele exhibited variable dendritic outgrowth, owing to increased stochastic transcriptional noise compared to normal melanocytes (Kemkemer et al., 2002). Overall, such variabilities can lead to excessive divergence in gene expression, even among cells of the same lineage that feature very similar transcriptional programmes. Such paradigms underscore the complexity of TF HIs. That is, gene transcription might only become intolerably impaired in a subset of functionally identical cellsyet still lead to abnormal organ function and disease. Therefore, although it is established that low TF concentrations significantly influence 'bursty' gene expression, more studies will be needed to identify which target genes become the most severely impaired.
The inherently stochastic nature of transcription, discussed above, can be quantified by using single-molecule and single-cell methodologies and, thus, be linked to TF HIs. The dynamics of transcription can be observed in vivo by incorporating labelled nucleotides to monitor the production of nascent transcripts (Morisaki et al., 2014) or it can be visualised at specific loci by smFISH (Hsu et al., 2017;Kochan et al., 2015;Mehta et al., 2018;Titlow et al., 2018). Tagging the untranslated regions of endogenous genes with stem-loop-encoding sequences of bacteriophages MS2 or PP7, fused to a fluorescent reporter, allows the visualisation of nascent RNA and, thus, is suitable to track the expression of genes of interest (Bothma et al., 2014;Katz et al., 2018). Intercellular variability of transcription can be quantified by smFISH and scRNA-seq (Beach et al., 1999;Bertrand et al., 1998;Guo et al., 2017;Halpern et al., 2017;Hocine et al., 2013;Treutlein et al., 2016;Yan et al., 2013). Further, advances in scRNA-seq have allowed the analysis of allele-specific 'bursting' to be used as a means of identifying imbalances of gene expression (Borel et al., 2015;Chen et al., 2016;Deng et al., 2014;Faddah et al., 2013;Jiang et al., 2017;Kim and Marioni, 2013;Marks et al., 2015). In other studies, the cellular transcriptional activity has been inferred by examining localisation and dynamics of the transcriptional machinery (Cho et al., 2018;Cisse et al., 2013;Li et al., 2019;Steurer et al., 2018). From this, it becomes clear that to link TF behaviour to gene expression at high temporal and spatial resolution, studies often have to utilise both live-and fixed-cell methods to obtain complementary information. For example, the local concentration and dynamics of SOX2, CDK9, BRD4 and MED22 were measured by 3D STED and FRAP, and nascent POU5F1 transcripts were visualised by MS2-MCP labelling . This correlation between local concentrations of TFs and gene expression leads to the final TF HI-influencing mechanism explored in this Review: the formation of condensates.

Control of TF function through the formation of condensates
In recent years, phase separation of proteins has gained substantial attention in biological research. The formation of biomolecular condensates has been implicated in a plethora of cellular functions. These are as diverse as membrane-less organelles, such as the nucleolus (Feric et al., 2016), as well as the normal and abnormal variants of widely studied proteins, such as FUS, G3BP1, TDP43 (officially known as TARDBP) and BRD4 (Han et al., 2020;McGurk et al., 2018;Niaki et al., 2020;Riback et al., 2020;Yang et al., 2020) which are implicated in disease pathogenesis. However, our understanding of the biological meaning of condensate formation, particularly by TFs, remains incomplete. As such, phase separation of TFs, aiding nuclear compartmentalisation and gene regulation, is a relatively novel concept .
Phase separation of proteins depends on physico-chemical conditions, such as protein concentration, charge, 3D structure and cellular pH (Taratuta et al., 1990;Wang et al., 2018). The thermodynamics of phase-separated systems predicts that protein concentration inside condensates is higher than in the surrounding dilute phase (Klosin et al., 2020). Condensates may possess liquid, gel or solid-like properties but the biological consequences of such phase-separated entitiesand whether they are beneficial or notremain ill-defined (Brangwynne et al., 2009;Conicella et al., 2016;Kato et al., 2012;Lin et al., 2015;Strom et al., 2017). Condensates exhibit rapid protein exchange with the surrounding dilute phase, movement within the dense phase, as well as fusion and fission phenomena during their formation and maturation (Brangwynne et al., 2009;Handwerger et al., 2005;Phair and Misteli, 2000;Strasser et al., 2008). Although the exact biophysical mechanism of condensate formation remains elusive, weak multivalent interactions between intrinsically disordered regions seem to be the main driver (Dzuricky et al., 2020;Kato et al., 2012;Lin et al., 2017;Sabari et al., 2018;Wei et al., 2017).
As previously discussed, the intranuclear concentration of some TFs substantially influences gene regulation and function. Since condensate formation buffers concentration and functionally compartmentalises the nucleus (Klosin et al., 2020), condensates can control local TF concentrations. Super-enhancers are clusters of enhancers that accumulate components of the transcriptional machinery. They are thought to harness the formation of condensates to favour biochemical reactions through local increases in TF concentration and the formation of compartmentalised reaction/ diffusion networks, without excluding additional mechanisms. In this way, the molecular crowding brings regulatory sequences and promoters into close proximity (Hnisz et al., 2013), and such reversible local reaction/diffusion networks favour gene expression. Such a process can depend on nucleation events triggered by physiological (e.g. NF-κB in inflammation; Nair et al., 2019), developmental (e.g. Prospero in neural differentiation; Liu et al., 2020) or molecular processes (e.g. depletion of proteins such as the Mediator complex and BRD4; Sabari et al., 2018). The high-density assemblies of transcriptional machinery components at enhancers is, by definition, substantially assisted by the formation of condensates. This is because TFs, e.g. SOX2, OCT4 and NANOG , co-activators, e.g. BRD4 and MED1  and RNAPolII (Boehning et al., 2018;Cho et al., 2018) all form condensates. However, to what extent this depends on condensate formation alone or on clustered DNA-binding interactions, remains under investigation . So far, studies of the function of TF condensates in the regulation of gene expression remain scarce and have mostly focussed on linking condensate formation to transcriptional output Sabari et al., 2018), examining silencing of genes, such as prospero in Drosophila (Liu et al., 2020), and investigating diseases, such as HOXD13-associated synpolydactyly (Basu et al., 2020). Alternatively, condensate formation has been recently proposed to be involved in buffering the concentration of TFs in the dilute phase (Klosin et al., 2020), thereby dampening the variability of TF concentration. Although the universality of such functions remains to be experimentally confirmed, it is intriguing that many TFs depend on a stringent regulation of their local concentration for normal function. Examples include PAX6/SOX2 during ocular development (Matsushima et al., 2011), OCT4 in the preimplantation embryo (Gerovska and Araúzo-Bravo, 2019) and NANOG during blastocyst formation (Bessonnard et al., 2014). Therefore, formation of condensates and how they control TF variability across cells warrant further investigation in the context of HIs.
Understanding the biological function of TF condensate formation requires the characterisation of their constituents, their subcellular location and their dynamic behaviour over time. Fast scanning microscopysuch as confocal spinning-disk (CSD) microscopy and LLSMis used to observe cellular localisation and co-partitioning of proteins into condensates (Chong et al., 2018;Sabari et al., 2018), and can be quantified by SRM (Cai et al., 2019;Chong et al., 2018;Sabari et al., 2018) (see Table 2 and Fig. 2).
The mobility of TFs within condensates, and the exchange of molecules between condensates and the environment, have been measured by FRAP, FCS and SPT. Condensates formed by transcriptional machinery components, such as BRD4 and MED1 Cai et al., 2019;Chong et al., 2018;Gibson et al., 2019;Guillén-Boixet et al., 2020;Klosin et al., 2020;Liu et al., 2020;Sabari et al., 2018;Teves et al., 2016;Zamudio et al., 2019), and TF condensates, such as OCT4 , have been also studied by FRAP. Additionally, the formation of TF condensates driven by low complexity domains has been studied by FRAP and SPT (Chong et al., 2018). Stress-induced condensates, e.g. YAP, have also been studied by FRAP (Cai et al., 2019), and the formation and maturation of G3BP1-related stress granules have recently been investigated by FRAP and FCS (Guillén-Boixet et al., 2020).
It will be interesting to further investigate the biological functions of TF condensates through a combination of single-molecule approaches to link TF numbers and their dynamic behaviour to gene expression and HIs.

Conclusions
Here, we have outlined how TF abundance and dynamic interactions with chromatin are required for normal development and how abnormalities in such regulation can result in disease.
We have discussed the key factors and cellular functions that control TF abundance in cells and tissues. TFs undergo complex kinetic interactions with chromatin; therefore, investigating TF numbers and molecular movement at the cellular and tissue levels is essential. Studied paradigms of dosage-sensitivity and HI emphasise how recent quantitative microscopy advances help researchers to investigate such complex TF interactions and their malfunction in disease. To date, such information on TFs and TF-cofactor complexes can be obtained with high spatiotemporal resolution. When these methodologies are applied to multiple cell types and disease models, they help to understand the molecular underpinnings of HI-associated diseases. However, a remaining challenge is to simultaneously study the behaviour of collaborating TFs and TF-complexes in live cells and tissues. This shortcoming is likely to result from the limitations of fluorescence microscopy to faithfully investigate several differently labelled proteins at the same time. Mass spectroscopy, RNA-seq and ChIP-seq methods do provide much broader, albeit static, datasets of TF binding and protein and mRNA abundances in cells but lack the dynamic information of live systems. A further challenge is to combine temporal and spatial super-resolution methods so that the mobility and chromatin-binding dynamics of individual TFs can be investigated at specific genomic loci. As discussed, this is currently limited to repetitive loci, which can be studied in tandem using the same fluorescent probe. Non-repetitive DNA sequences require a large number of sequence-specific probes to sufficiently increase the signal-to-noise ratio for fluorescence imaging (Chen et al., 2013;Knight et al., 2015;Ma et al., 2013;Thanisch et al., 2014).
Nevertheless, when studying the function of fluorescently labelled TFs, microscopy methodologies have become sensitive enough to detect molecules at low, physiologically relevant concentrations and by using low excitation power, which largely preserves the normal cell function. In fact, smaller and brighter fluorescent tags have been identified (Govindan et al., 2018;Oliinyk et al., 2019) but tagging genes non-disruptively does necessitate functional, i.e. genetic, validation of the endogenous behaviour of TFs. Additionally, the generation of further tissue and live animal models will explain the aberrant function(s) of TFs at the molecular level, in a biologically relevant context. For example, heterozygous LOF mutations of the widely expressed transcriptional co-activator YAP1 cause diverse defects in patients, such as ocular abnormalities or craniofacial/intellectual disabilities, suggesting that the observed phenotypes are the result of tissue-specific HIs of YAP1 (Williamson et al., 2014). The generation of animal models to study the underlying mechanism of these defects will facilitate understanding the phenotypic variability observed in human patients by means of quantitative microscopy methodologies.
From a translational perspective, new animal models will enable the detection of early developmental transcriptional and signalling defects that normally lead to disease onset and progression. This will be possible by combining high spatial and temporal resolution microscopyas discussed in this Reviewwith in vitro quantification of TF transcriptomes and interactomes. To date, most studies on TF function have addressed the impact of their mutations in disease only qualitatively. The information on intracellular TF concentrations and chromatin-binding behaviour is scarce, and may have been underappreciated. Until now, computational and bioinformatics analyses, coupled to machine learning and highthroughput screening, have been predominantly deployed to predict TF binding sites (Elmas et al., 2017;Shen et al., 2018) or TF-binding behaviour based on DNA structure (Zhou et al., 2015). Further approaches need to be developed to investigate the levels and molecular behaviour of TFs, and to predict transcriptional outcomes. To this end, the ample availability of current and future single-cell transcriptome datasets should be exploited to study how TF dosage affects gene transcription. This will enable us to understand the underlying mechanisms and the causal relationships between abnormal TF numbers, impaired TF behaviour and the manifestation of disease.