Cluster size determines internal structure of transcription factories in human cells

TL;DR

This study investigates how transcription factories assemble in human nuclei and whether clusters are specialized (demixed) or mixed in composition. It employs a multicolor bead-spring chromatin model with diffusing TF:pol complexes that bind to cognate coloured transcription units (TUs) and weakly to others, enabling bridging-induced clustering. A central result is a robust, cluster-size–dependent crossover: small clusters are typically demixed while large clusters become mixed, quantified by the demixing coefficient $\theta_{\rm dem}$ around the threshold $\sim 10$ TFs per cluster. The framework, extended with DHS-driven chromatin maps and HiP-HoP dynamics, yields better agreement with GRO-seq and Hi-C data and offers predictions for transcriptional networks and responses to genome edits.

Abstract

Transcription is a fundamental cellular process, and the first step of gene expression. In human cells, it depends on the binding to chromatin of various proteins, including RNA polymerases and numerous transcription factors (TFs). Observations indicate that these proteins tend to form macromolecular clusters, known as transcription factories, whose morphology and composition is still debated. While some microscopy experiments have revealed the presence of specialised factories, composed of similar TFs transcribing families of related genes, sequencing experiments suggest instead that mixed clusters may be prevalent, as a panoply of different TFs binds promiscuously the same chromatin region. The mechanisms underlying the formation of specialised or mixed factories remain elusive. With the aim of finding such mechanisms, here we develop a chromatin polymer model mimicking the chromatin binding-unbinding dynamics of different types of complexes of TFs. Surprisingly, both specialised (i.e., demixed) and mixed clusters spontaneously emerge, and which of the two types forms depends mainly on cluster size. The mechanism promoting mixing is the presence of non-specific interactions between chromatin and proteins, which become increasingly important as clusters become larger. This result, that we observe both in simple polymer models and more realistic ones for human chromosomes, reconciles the apparently contrasting experimental results obtained. Additionally, we show how the introduction of different types of TFs strongly affects the emergence of transcriptional networks, providing a pathway to investigate transcriptional changes following gene editing or naturally occurring mutations.

Figures (8)

• Figure 1: Toy model, with TUs coloured randomly (the random string). $\textbf{(A)}$ Overview. (i) Yellow, red, and green TFs ($25$ of each colour) bind strongly (when in an on state) to $100$ TUs beads of the same colour in a string of $3000$ beads (representing $3$ Mb), and weakly to blue beads. TU beads are positioned regularly and coloured randomly, as indicated in one region of the string. TFs switch between off and on states at rates $\alpha_{off}=10^{-5}~\tau_B^{-1}$ and $\alpha_{on}=\alpha_{off/4}$ ($\tau_B$ Brownian time, which one can map to $0.6-6~10^-3~s$, see SI). (ii) The sequence of bars reflects the random sequence of yellow, red, and green TUs (blue beads not shown). (B) Snapshot of a typical conformation obtained after a simulation (TFs not shown). Inset: enlargement of boxed area. TU beads of the same colour tend to cluster and organize blue beads into loops. $\textbf{(C)}$ Bridging-induced phase separation drives clustering and looping. Local concentrations of red, yellow, and green TUs and TFs might appear early during the simulation (blue beads not shown). Red TF $1$ -- which is multivalent -- has bound to two red TUs and so forms a molecular bridge that stabilizes a loop; when it dissociates it is likely to re-bind to one of the nearby red TUs. As red TU $2$ diffuses through the local concentration, it is also likely to be caught. Consequently, positive feedback drives growth of the red cluster (until limited by molecular crowding). Similarly, the yellow and green clusters grow as yellow TF $3$ and green TF $4$ are captured. $\textbf{(D)}$ Bar heights give transcriptional activities of each TU in the string (average of $100$ runs each lasting $8~10^{5}\tau_B$). A TU bead is considered to be active whilst within $2.24\sigma\sim 6.7 \times 10^{-8}m$ of a TF:pol complex of similar colour. Dashed boxes: regions giving the 3 clusters in the inset in $\textbf{(B)}$. $\textbf{(E)}$ Pearson correlation matrix for the activity of all TUs in the string. TU bead number (from low to high) is reported on axes, with pixel colour giving the Pearson value for each bead pair (bar on right). Bottom: reproduction of pattern shown in $\textbf{(A,ii)}$. Boxes: regions giving the 3 clusters in the inset in $\textbf{(B)}$.
• Figure 2: Simulating effects of mutations. Yellow TU beads $1920$, $1950$, $1980$, $2010$, $2040$ and $2070$ in the random string have the highest transcriptional activity. $1$-$4$ of these beads are now mutated by recolouring them red. $\textbf{(A)}$ The sequence of bars reflects the sequence of yellow, red, and green TUs in random strings with $1$, $2$ and $4$ mutations (blue beads not shown). Black boxes highlight mutant locations. $\textbf{(B)}$ Typical snapshots of conformations with $\textbf{(i)}$ one, and $\textbf{(ii)}$ 4 mutations. $\textbf{(C)}$ Transcriptional-activity profiles of mutants (averages over $100$ runs, each lasting $8~10^{5}\tau_B$). Bars are coloured according to TU colour. Black boxes: activities of mutated TUs. $\textbf{(D)}$ Activities (+/- SDs) of wild-type (yellow) and different mutants. $3$ mutations: TUs $1950$, $1980$ and $2010$ mutated from yellow to red. $\textbf{(E)}$ Typical kymographs for $\textbf{(i)}$ wild-type, corresponding to the same original string presented in Figure 1, and $\textbf{(ii)}$$4$-mutant cases, in which 4 yellow TUs have been mutated to red. Each row reports the transcriptional state of a TU during one simulation. Black pixels denote inactivity, and others activity; pixels colour reflects TU colour. Blue boxes: region containing mutations. $\textbf{(F)}$ Pearson correlation matrices for wild-type and $4$-mutant cases. Black boxes: regions containing mutations (mutations also change patterns far from boxes).
• Figure 3: Reducing the concentration of yellow TFs reduces the transcriptional activity of most yellow TUs while enhancing the activities of some red TUs.(A) Overview. Simulations are run using the random string with the concentration of yellow TFs reduced by $30\%$, and activities determined (means from $100$ runs each lasting $8~10^{5}\tau_B$). (B) Activity profile. Dashed boxes: activities fall in the region containing the biggest cluster of yellow TUs seen with $100\%$ TFs, as those of an adjacent red cluster increase. (C) Differences in activity induced by reducing the concentration of yellow TFs. This plot is obtained by subtracting the transcriptional activity of the wild-type, Figure \ref{['fig1']}D, from that of the current system in panel B. (D) Pearson correlation difference matrix. This plot is obtained by subtracting the Pearson correlation matrix of the wild-type, Figure \ref{['fig1']}E, from that of the current system. Boxes: regions giving the $3$ clusters from $\textbf{Figure 1B, inset}$.
• Figure 4: Clustering similar TUs in 1D genomic space increases transcriptional activity.(A) Simulations involve toy strings with patterns (dashed boxes) repeated $1$ or $6$ times. Activity profiles plus Pearson correlation matrices are determined ($100$ runs, each lasting $8~10^{5}\tau_B$). (B) The $6$-pattern yields a higher mean transcriptional activity (arrow highlights difference between the two means). (C) The $6$-pattern yields higher positive correlations between TUs within each pattern, and higher negative correlations between each repeat.
• Figure 5: TU transcriptional networks and demixing. Simulations are run using the toy models indicated, and complete correlation networks (qualitatively reminiscent of gene regulatory networks) constructed from Pearson correlation matrices. (A) Simplified network given by the random string. TUs from first (bead $30$) to last (bead $3000$) are shown as peripheral nodes (coloured according to TU); black and dashed grey edges denote statistically-significant positive and negative correlations, respectively (above a threshold of $0.2$, corresponding to a $p$-value $\sim5~10^{-2}$). The complete network consists of $n=100$ individual TUs, so that there are $n_c=\binom{100}{2}=4950$ pairs of TUs couples; we find $990$ black and $742$ gray edges. Since $p$-value$\cdot n_c=223$, most interactions (edges) are statistically significant. Networks shown here only correlations (i) between red TUs, and (ii) between red and green TUs. (ii) (B) Average Pearson correlation (shading shows +/-SD, and is usually less than line/spot thickness) as a function of genomic separation for the (i) random, (ii) $6$-, and (iii) $1$-pattern cases. Correlation values at fixed genomic distance are taken from super-/sub-diagonals of Pearson matrices. Red dots give mean correlation between TUs of the same color ($3$ possible combinations), and blue dots those between TUs of different colors ($4$ possible combinations). Cartoons depict contents of typical clusters to give a pictorial representation of mixing degree (as this determines correlation patterns); see SI for exact values of $\theta_{\rm dem}$.