A variegated model of transcription factor function in the immune system

Abstract

Specific combinations of transcription factors (TFs) control the gene expression programs that underly specialized immune responses. Previous models of TF function in immunocytes had restricted each TF to a single functional categorization (e.g., lineage-defining vs signal-dependent TFs) within one cell-type. Synthesizing recent results, we instead propose a variegated model of immunological TF function, where many TFs play flexible and different roles across distinct cell states, contributing to cell phenotypic diversity. We discuss evidence in support of this variegated model, describe contextual inputs that enable TF diversification, and look to the future to imagine warranted experimental and computational tools to build quantitative and predictive models of immunocyte gene regulatory networks.

Sense from complexity: untangling immunocyte transcription factor networks

Cells of the immune system must mount diverse and highly specialized responses to foreign antigens, inflammatory triggers, and deviations from organ homeostasis. These responses are mediated by coordinated gene expression programs, each encompassing synergistic functionalities (homing, interaction ligands/receptors, effector molecules). Such programs are controlled by the action of transcription factors (TFs), which in turn act on cis-regulatory elements (CREs), promoters, and enhancers. How TFs select their targets, are deployed in response to specific signals, or are organized combinatorially, are questions of fundamental importance for the understanding and control of immunity. Given the thousands of transcriptional regulators encoded in mammalian genomes and the intricate packing of DNA in a nucleus, the complexity of the problem has led some to characterize efforts to truly comprehend TF function as futile [1]. Nevertheless, astounding advances in data generation capabilities and computational tools that are well-suited to discover specific patterns within enormously complex data, have enabled novel experimental strategies. These novel approaches are illuminating how TFs contribute to orchestrating immune responses and prompt a rethinking of previous conceptual frameworks. Here, we argue that a conceptual limitation has been the reductionist restriction of TFs to single functional categories. Instead, recent results support a view in which a given TF can play variegated roles across different cell states and genomic loci. Such variegation in dozens of TFs can contribute to the fine diversity of phenotypic states that individual cells adopt within a defined cell-type.

We begin with a brief discussion of existing paradigms of TF function in the immune system, review recent studies supporting the variegated model of TF function, and conclude with forward-looking key questions for future investigation. We also speculate on new potential tools that would be needed to tackle such queries.

Existing models of transcription factor function

Current paradigms tend to focus on the classification of TFs into individual functional categories. One framework, the “collaborative-hierarchical” model, distinguishes “lineage-defining TFs” (LDTFs), those with pioneering ability to displace nucleosomes and to open lineage-specific regulatory elements, from “signal-dependent TFs” (SDTFs), which bind to elements primed by these LDTFs or collaborate with LDTFs to open latent or de novo enhancers [2–4]. Under this theoretical model, SDTFs lie below LDTFs in the hierarchy of TF function, with combinations of LDTFs selecting cell type-specific elements. LDTFs overlap with “master TFs,” classically defined as factors which are both necessary and sufficient to bring about cell type-specific regulatory programs [5–8]. SDTFs can also be split into primary and secondary responder TFs, to account for the successive waves of gene activation and repression that occur after activation of any cell-type [4]. Pre-synthesized TFs which control primary responses are separated from secondary responder TFs which depend on other TFs for their expression and synthesis [4]. Importantly, by the nature of their categorization, such paradigms restrict each TF to a single functional role across all contexts and treat all individual cells within a cell-type as identical.

Such TF hierarchies are not new and were highlighted a long time ago in models of Drosophila melanogaster development [9,10]. However, these models face several limitations. First, while conceptually convenient by reducing complexity, simplified models of linear operation of LDTFs followed by SDTFs do not reflect the convolutions of in vivo environments, where cells are simultaneously exposed to gradients of multiple inputs (cell-cell interactions, soluble mediators, metabolites). For instance, no in vivo situation exists in which macrophages are solely activated by LPS, or T cells by CD3/CD28 engagement. Would LDTFs and their functions really remain impervious to such influences?

Second, such reductionist models ignore the many modalities through which any one TF might operate. From a theoretical standpoint, one can draw out many strategies by which TFs regulate transcription, beyond the simple bind/activate or bind/repress modes that we usually have in mind (Figure 1). There is no reason to assume that a regulator would be restricted to any one of these, especially if one considers that many TFs contain long stretches of intrinsically disordered regions (IDRs), which confer functional flexibility and allow them to adapt to diverse co-factors [11]. For example, the long N-terminal region of FoxP3 is an IDR, and interacts changeably with many potential cofactors [12,13]; the TAZ1 domain of the transcriptional co-activators CBP/p300 interacts with multiple, diverse TFs with IDRs, such as CITED2 and STAT2 [11,14,15]. Furthermore, systematic deconstructions of human regulatory networks have shown that TFs are highly combinatorial in their function, with extensive network connections as well as feedback and feed-forward loops controlling regulatory activities [16].

Figure 1:

Beyond simple binding and transactivation: the diversity of TF regulatory mechanisms. Several of these mechanisms are compiled and adapted from [23,93]. GRN, gene regulatory networks.

Third, TF categorizations are not absolute. In one study of mouse fibroblasts, analysis of genetic variants affecting ATAC-sequencing (seq) and H3K4me1/H3K27Ac chromatin immunoprecipitation (ChIP)-seq, overlaid with AP-1 ChIP-seq, showed that enhancer selection depended on AP-1 factors, which are typically classified as SDTFs, such that SDTFs in one context could be LDTFs in another [17]. Similarly, TCF1 (encoded in mice by Tcf7) has many guises: it is a pioneer factor in the determination of the entire T cell lineage, a late SDTF at the effector/memory fork of activated CD8⁺ T cells, and a factor whose repression enables Regulatory T cell (Treg) identity [18,19]. The master TF FOXA1 is considered a canonical pioneer in mammalian endodermal lineage specification, and HNF4A a non-pioneer factor [20,21]. However, upon ectopic overexpression in a lymphoblastoid cell line, both factors were able to open previously inaccessible regions (assessed by ATAC-seq and CUT&TAG) and induce endodermal gene expression programs [22,23]. Furthermore, HNF4A does act as a lineage-determining pioneer factor in hepatocyte differentiation [24] and in a subpopulation of thymic mimetic cells [25], further demonstrating the flexibility of TF pioneering capabilities.

Finally, many apparent “master TFs” may actually be markers, but not drivers, of specific programs. For example, FoxP3 has lost its throne as the master regulator of the Treg lineage. While it crucially contributes to Treg function, FoxP3 is neither necessary nor sufficient to establish Treg identity: Treg-like cells develop without it, and Treg-specific programs contain both FoxP3-independent and -dependent modules [26–32]. Moreover, while RORγ is classically considered the master regulator of Th17 cell differentiation, RORγ-deficiency has been shown to lead to only modest changes on the chromatin landscape (by p300 and H3K4Me3 ChIP-seq) of differentiating mouse Th17 cells. RORγ’s lineage-defining effects instead depend on priming by BATF, IRF4, and STAT3, which are normally considered SDTFs [33,34]. Furthermore, a comparison of RORγ target binding (by ChIP-seq) and accessibility (by ATAC-seq) in a pan-mouse immune system atlas found RORγ’s targets to be highly variable across different cell-types [35]. Thus, the domain of master regulators may not equally reach across all cell conditions, furthering the limitations of absolute classification of TFs into single functional categories.

TF function: variegated, not monomorphic

Whether in response to distinct external stimuli, or across the continua of cell-to-cell variation, the synthesis of recent studies clarifies that the same TFs can perform distinct roles across diverse conditions (we will not enter here into the thorny debate on the boundaries between cell-types and - states [36–38]). However, we contend that this variegated model of TF function depicted in Figure 2 applies across all models of mammalian cell-type configurations.

Figure 2: Variegated Model of TF function.

Under this model, instead of each TF having uniform, monomorphic activity across all phenotypic cell states, TFs can have flexible, variegated functions across different settings.

Recent studies illustrate how TF functions can vary in response to extracellular stimuli. One such study leveraged natural genetic variation in F1 hybrids (see Box 1) of different mouse strains to uncover the causal determinants of chromatin accessibility (readout by ATAC-seq) in naïve and activated CD4⁺ and CD8⁺ T cells, following acute lymphocytic choriomeningitis virus (LCMV) infection (Armstrong strain for acute infection) [39]. A small number of TF families, including Ets, Runx, and TCF/Lef factors, were the major drivers of chromatin accessibility in both naïve and activated conditions. Based on TF binding data (measured by CUT&RUN), Ets1, Runx1, and TCF1 collectively occupied upwards of 90% of accessible chromatin in naïve T cells. The overwhelming contribution of only three TF families was surprising given the large number of TFs expressed in these cells. However, each TF family responded differently to CD4+ and CD8+ T cell activation. Specifically, while Ets1 acted as a “housekeeper,” by binding primarily to regions that remained unchanged in terms of accessibility (e.g., promoters of housekeeping genes) following activation, Runx1 appeared to amplify activation-dependent changes because its binding coincided with those sites that most changed in accessibility relative to naïve cells. Conversely, TCF1-bound regions became less accessible during the response to infection and sometimes became newly occupied by activation-induced TFs. This led the authors to suggest that TCF1 may act as a “placeholder” which maintains accessibility in naïve cells. Thus, three TF families that occupied most of the accessible T cell genome exhibited diverse binding behaviors, which would not have been captured by motif enrichment analyses and would have been refractory to knockout/perturbation approaches (difficult to interpret given their widespread functions). Of note, in a study of NK cells activated during acute Toxoplasma gondii infection in mice [40], STATs were the primary drivers of new enhancers (based on analysis of STAT motifs and binding, in addition to p300 deposition and ATAC-seq in STAT-deficient cells; also, ChIP-seq analyses showed a redeployment of T-bet (a presumed LDTF) to loci containing STAT but not T-bet motifs. This observation suggested that TF crosstalk can reverse the traditional roles of LDTFs and SDTFs because TFs canonically considered to be SDTFs could control the activity of a TF otherwise classified as an LDTF; this further illustrates the variegation of TF function in response to distinct external influences.

Box 1: Using natural genetic variation to identify chromatin regulators genome-wide.

Identifying how TFs globally control immunocyte chromatin state requires genome-wide functional experiments. In the past, a number of studies collectively identified the major important players in immunocyte chromatin states by knockout of individual TFs (although surprises may still be lurking). These were continued more systematically at high-throughput with RNA interference (RNAi) or CRISPR-Cas9 screens [78–80], which also offer-cell-type resolution (e.g. Perturb-seq or equivalent [81–83]) However, the inherent issues with gene knockouts (see Box 2), or the issues long recognized to plague TF over-expression experiments [84], may limit the ability of reverse-genetics approaches to identify the major regulatory nodes at a systems level. In recent years, an increasing number of investigators have taken advantage of natural genetic variation across different inbred mouse strains [17,32,39,50,85–87] (similar exploitation can be made with datasets from humans, which are heterozygous at most loci, though interpretation is much more difficult since phasing and linkage disequilibrium structure introduce distinct complications, unlike inbred mouse crosses where the continuous sequence of each chromosome is known). In particular, F1 hybrids between distant strains offer the opportunity to identify causal regulators genome-wide [17,32,39,50,85–87]. The principle of the experiment is to compare within the same cell, with a common concentration of trans-regulators, the consequence of a low-affinity variant in a TF binding site in one allele compared to a high-affinity allele. Performing this comparison across all such binding sites or motif instances is equivalent to a genome-wide mutagenesis screen, powered by the 5–20 million single nucleotide polymorphisms (SNPs) that can distinguish any two inbred strains (and tend to be enriched at cis-regulatory elements) [88]. Such studies can help delineate which sequence elements are required for chromatin features at sites of interest, genome-wide. This strategy adds a layer of information beyond simple motif enrichment and enables causal inference of TF contributions at scale, while maintaining physiologic TF concentrations [89,90].

Variegated TF function extends beyond binary comparisons of activated vs resting immunocytes. Single-cell approaches have enabled the study of TF functions across the full complexity of in vivo cell continua elicited by combinatorial gradients of external signals. Treg cells offer a good model for this question given that their inherent reactivity to self, combined with their ability to recognize non-self, results in a broad range of cell states at baseline in lymphoid and non-lymphoid tissues [41]. These diverse states reflect their equally diverse functions in preventing autoimmunity and maintaining organismal homeostasis [41–43]. Recent work from our lab involved using single-cell ATAC-seq (scATAC-seq) to identify the TFs controlling the diversity of mouse Tregs [32]. Using a machine learning approach (topic modeling) [44,45] to extract groups (topics) of open chromatin regions (OCRs) with covarying patterns of activity across single cells, the state of each cell could be decomposed into combinations of variably active chromatin programs, reflecting the integration performed by each cell in vivo. Analyses leveraging allele-specific single cell (sc)ATAC-seq of Tregs from B6/Cast F1 hybrid mice identified the causal regulators of each of these programs, as also validated by Treg-specific TF knockouts. Some of the major TF families that have been previously reported [39] (e.g., Ets, TCF/Lef factors) contributed to the accessibility of many Treg chromatin programs. However, many other TFs (including NFkB, GATA, AP-1, or Nuclear Receptor family members) exhibited highly focal effects, affecting the accessibility of only a narrow selection of chromatin programs, themselves active in specific subpopulations. These selective impacts on chromatin program accessibility occurred despite enrichment of corresponding TF motifs across a broader range of Treg programs. Supporting the model of variegated TF activity, this finding suggested that while TFs may be present across many contexts, their functional effects are more restricted, reflecting a dependence on cell state. Further underscoring the point of differential TF effects across varied cell states, analysis of FoxP3-dependent OCRs from comparisons of FoxP3 KO and WT cells across the single cell space showed that FoxP3 could act as a repressor of accessibility in some chromatin programs but an activator of accessibility in others. Thus, TF function may be highly conditional on cell state, in some cases even reversing its direction in different groups of OCRs and different cell populations. We posit that variegated TF effects, dependent on cell state, may manifest as overlapping gradients of activity across the single cell chromatin space.

Determinants of TF variegation

If TF function varies across cell states, what determines context-specificity? Much as numerous mechanisms can explain the activity of a given TF (Figure 1), TF interactions can arise through several modalities (Figure 3). These diverse modalities of TF-TF interactions might then account for high variability across cells, and for the overlapping gradients of TF function that are observed in single-cell datasets. Moreover, combinatorial integration of several TFs is a long-appreciated mechanism to achieve specificity, going back to the first descriptions of regulatory control by TFs [2,3,9,10,16,46]. Recent examples include distinct compositions of lineage-defining TFs at B cell functional vs nonfunctional enhancers [47], co-localization of bZIP and T-box motifs specifically at sites with increased Runx1 occupancy in activated T cells [39], and activating or repressive FoxP3 regulatory effects mediated by distinct FoxP3-cofactor ensembles [13]. Thus, state-specific TF function might be conditioned by the availability of partner TFs; possible mechanistic interactions are schematized in Figure 3.

Figure 3: Theoretical mechanisms of TF-TF interactions and crosstalk: how one TF can modify the activity of another.

The schematics depict how a secondary TF (“TF2”) or cofactor can modify the activity of a first TF (“TF1”), directly or indirectly.

Persistent expression of a TF is not always essential to modulate the action of other factors: in so-called “hit & run” mechanisms, regions opened earlier during differentiation can remain active in the absence of continued expression of the pioneer factor. For example, in mature Tregs, FoxP3 binds to loci opened first by other Forkhead factors (e.g., FoxO1) earlier in mouse thymic T cell differentiation [48]. Nucleosome remodeling can account for pioneer events around lineage commitment that open loci to subsequent TF action, but remodeling is also involved at later stages. For instance, in response to activation signals in mouse fibroblasts, AP-1 family TFs can recruit nucleosome remodelers of the BAF complex to mediate enhancer selection, which then enable further function at serum-responsive genes [17]. Similarly, during mouse T cell activation, most sites bound by Ets1 in naïve CD4⁺ and CD8⁺ T cells remain unchanged in accessibility, but a small subset of Ets1-bound sites undergo a dramatic increase in accessibility [39]. These sites lack canonical Ets1 motifs and colocalize with a distinct group of activation-induced TFs as well as Brg1, a component of the SWI/SNF chromatin remodeling complex [39]. Thus, the specific activity of these Ets1-bound sites depends on both chromatin remodeling and a specific complement of TFs. Notably, despite the influence of chromatin remodelers, motifs from other TFs that are bound at these locations still have a positive effect on chromatin accessibility based on F1 genetic variation analysis [39]; this suggests that TFs are not simply recruited to accessible sites but also independently contribute to chromatin states at their target enhancers.

From another angle, TF-TF interactions need not operate directly at TF binding sites: changes in cell state can modulate TF activities remotely as well. For instance, in PU.1 overexpression experiments and in surveys across mouse early T cell development, ChIP-seq experiments show that PU.1 (whose main function is as a LDTF for myeloid and B cells) can redirect Satb1 and Runx1 binding to lower affinity target regions within PU.1-bound sites via TF ‘theft’’ [49]. Because PU.1 expression is limited to a specific time window prior to final T lineage commitment, this process provides stage-specific control of otherwise ‘constitutively’-bound TFs [49]. Similarly, upstream regulators can provide contextual information by regulating the expression of other TFs. For example, a recent study found that FoxP3 repressed the closely related conventional T cell (Tconv) lineage program by downregulating the expression of Tcf7 (encoding TCF1 – a positive regulator of Tconv-like chromatin accessibility) [50]. How much of Treg identity is truly determined by this mechanism is debated, however [51] (discussion in preprint article). Thus, whether by re-directing TF binding or modulating TF expression, one TF can affect the activity of another without directly binding to the same genomic position.

Another mechanism by which TFs can provide contextual information is by modifying chromatin state via the epigenetic recruitment of methyltransferases and/or demethylases to control DNA CpG methylation (the forms and role of DNA methylation have been recently discussed [52–54]). In a recent methylation profiling study across six lymphoid (B, CD4⁺ T, CD8⁺ T, NK cell) and myeloid (monocyte, granulocyte) immunologic cell types in healthy humans, lineage-specific TFs bound to cell type-specific hypomethylated regions [55]. Hypomethylated sites bound by lineage-defining TFs in one cell type were correspondingly methylated in other cell types. This is relevant as generally, methylation can directly inhibit TF binding [56], thus providing negative regulatory control of TF function. However, TF function does not exclusively occur subsequent to chromatin remodeling: TFs can themselves regulate both chromatin remodeling and methylation [57]. For example, PU.1—previously reported to direct site-specific methylation [58]—has been reported to bind to methylated regions in each of the profiled human cell types that were discussed [55], consistent with the hypothesis that PU.1 might partially control lineage specification via methylation of lineage-inappropriate loci. Notably, based on nucleosome binding arrays, PU.1 can also recruit nucleosome remodelers via its intrinsically disordered region [23,59]. Therefore, the dichotomy that is often drawn between sequence-specific TF binding and chromatin mechanisms is better described as a feedback loop whereby TFs can both modify and respond to changes in chromatin to diversify their regulatory roles. Each chromatin-modifying process provides an additional layer of regulatory information, which can act at different timescales.

A Gene Regulation “Wish List”

As recent studies suggest, if the regulatory landscape is a ballet of dozens of interacting TFs, all varying along cross-cutting gradients of activity in the space of individual cell phenotypes, should we also give up on truly understanding immunological gene regulation? There is hope, perhaps, in that the convoluted architectures recently uncovered do show structure and are reproducible between individuals. In truth, it is an exciting time to be studying immune gene regulation, and the vistas of the genome’s activity that we observe, even if bewildering, are awe-inspiring. What technologies might we wish for to answer open questions regarding chromatin landscapes and transcriptional regulation? Here, we select a “wish list” (by no means exhaustive) of such questions and possible solutions, some already emerging, others still far-off.

Can we move beyond binary descriptions of TF binding towards quantitative measurements of in vivo genome-wide affinities, and understand residence times of TFs at specific loci? Investigators must currently make a tradeoff between using genomic TF-binding assays to gain descriptive information on TF binding and genomic localization in vivo, or make quantitative measurements of TF biophysical parameters in vitro (affinities, dissociation constants) under conditions that bear little resemblance to the physico-chemical milieu of a cell’s nucleus. Two recent studies (one in preprint) [60,61] measured genome-wide ATAC-seq or TF binding across a range of TF concentrations, either by tunable expression or by varying the amount of TF spiked into each reaction. The changes in accessibility or binding signals across a range of TF concentrations enabled derivation of per-locus TF affinity parameters (e.g., K_d, Hill coefficients), providing an initial proof-of-principle towards genome-wide kinetic and affinity measurements, but which will require robust validation [60,61]. It will also be important to ascertain true co-binding of two TFs (not just similar localization in parallel immunoprecipitation experiments). Beyond, one could envision imaging solutions that read out the interplay of different TFs at a given enhancer in real-time to grasp binding and interaction parameters at and within individual regulatory elements. One challenge is adapting such solutions to be tractable in primary cells (rather than tumor cell lines) and to make them scalable to assaying multiple TFs at once.

Moreover, what are the direct targets, and what are the binding preferences of each TF in vivo? Currently, most summaries of TF binding rely on position-weight matrices, which are derived from in vitro binding assays or ChIP-seq experiments. However, such descriptions reflect only major patterns in binding, and may not capture the full range of in vivo binding preferences of a given TF. Also, an open question is how faithfully ChIP-seq represents true in vivo binding [62,63]. In addition, new tools will be needed to disentangle primary from secondary TF effects. Instead of TF gene knockout experiments, inducible TF protein degradation, such as by degrons [64], PROTACs [65], or other chemical means [66], might enable an improved assessment of TF control in a time-resolved manner because acute depletion would allow studying immediate direct dependencies without the compensatory secondary effects of brutal knockouts (Box 2). New developments are warranted to sample changes over time in response to TF reduction to quantitatively measure biochemical kinetic parameters at TF targets [67]. Recent work has also developed CRISPR-Cas9-based approaches to identify upstream cis-regulatory elements and trans-regulatory factors controlling the expression of key immune response genes [68,69]. These will be important for mapping complete networks of gene expression responses that both impact and are impacted by a given TF.

Box 2: Assessing TF function (TF knockouts are not always the answer).

Genetic knockouts have generally been considered the gold standard for discerning the causal chain of TF actions. However, this approach is not without limitations, which are worth considering in the design of experiments to assess TF function. First, both direct and indirect effects manifest in TF knockouts. As transcriptional regulation is mediated by complex, inter-connected networks, removing one node from the network may lead to compensatory changes or propagation of downstream effects, which may mask the direct function of the TF in question. Inducible TF KOs or more acute chemical degradation strategies [66] may provide one solution to this problem. These may still be limited by need for development of reagents to cover all proteins of interest, portability of such systems to in vivo contexts, and specificity and efficiency of protein degradation [89]. However, evaluating function over time may be another necessary component for isolating the most proximal effects. The appropriate timescale may vary depending on whether the TF is pre-synthesized or induced only in response to some stimulus [4]. Second, TF KOs provide information outside the physiologic range of TF function [89]. TFs operate within tightly controlled concentrations [8] with non-linear responses to changes in TF occupancy. Instead of treating TF function as a binary variable, reading TF function across a range of finely-controlled TF concentrations [60] may better reflect in vivo mechanisms, moving from qualitative effects to quantitative parameters (e.g. affinities and dissociation constants). Leveraging variation in cis (see Box 1) may be another complementary strategy in this regard. Finally, parsing the relative contributions of different members of the same TF family or of TF paralogs remains an important challenge. Redundancy increases the rate of false negatives in TF KO experiments. This can partially be overcome by coupling knockout and overexpression studies [91] and by multi-family member perturbations [92] but the caveats of non-physiologic concentrations and ability to scale perturbations remain.

How can we develop a predictive and deeply granular understanding of immunocyte gene regulatory networks? Tackling this challenge will require the integration of all the questions and approaches discussed above. New advances in multi-modal data integration are beginning to reconstruct gene regulatory networks from single cell genomics data [70–72]. These will need to be further developed in conjunction with experimental perturbations of putative regulators and kinetic measurements with or without immunologically relevant stimuli; in turn, this may enable movement towards mechanistically grounded predictions of network responses to varied inputs. As changing every possible configuration of such regulatory networks is impossible, model-guided experimentation will thus be required to guide future investigation.

Finally, as illustrated with advances in image recognition, or protein structure prediction by AlphaFold, advances in Artificial Intelligence (AI) may become increasingly important in rendering the complexity of immune TF networks tractable. Because of their inherent ability to deal with highly complex inputs, deep learning and other novel AI approaches have already begun to tackle the genome’s cis-regulatory grammar at the sequence level with spectacular results [73–75] in several cases: identifying new modes of TF cooperativity and binding for mammalian pluripotency TFs [74], engineering new promoter sequences [76], and capturing decades of mouse immunocyte gene regulation biology within a single model [77]. Such models may prove crucial to identify, in quantitative terms, multiple modes of TF function. Deep models might also uncover multiple types of sequence features that are predictive of TF binding or motif interactions, and enable in silico predictions of regulatory changes [74,75]. Development of “transparent” models that provide not only predictive capability but also provide interpretable views into the underlying biological logic will be especially useful in informing our understanding of how the complexity of such networks is constructed from its constituent parts [75], how they are affected by genetic variation, and enable their targeted modification.

Concluding remarks

Immunology has provided a fertile ground for uncovering basic mechanisms of TF function. New results make clear that TFs are not monomorphic but flexible in their functional capabilities, even within a cell-type. TFs can further condition the function of other factors in response to external inputs, creating an interconnected loop between environmental stimuli, signaling cascades, TF regulation, and chromatin states. It is possible that the variegation is more marked in immunocytes, which must flexibly and rapidly adapt to changing environments, than in neurons which operate more in fixed locations. The next frontier (see Outstanding Questions) will be to combine advances in genome-wide measurements of gene regulatory function with computational advances in analyzing these functions, aiming to arrive at a complete, predictive understanding of immunocyte gene regulatory networks.

Outstanding Questions.

How can we move beyond binary descriptions of TF binding towards quantitative measurements of in vivo genome-wide affinities, and understand residence times of TFs at specific loci?

What are the direct targets of each TF?

What are the true binding preference(s) of each TF in vivo?

How does TF binding vary across single cells, or in a single cell over time?

Which sites are truly co-bound by multiple TFs and what are the relative dynamics of binding of each and interaction, at single-regulatory element resolution?

Do single TFs, or even TF pairs, really exist, or they mostly exist within multi-molecular complexes?

How can we develop a predictive and complete understanding of gene regulatory networks, for even one immunocyte?

How can new computational methods, such as deep learning and other artificial intelligence tools, be best deployed to further understand immunologic gene regulation?

Highlights.

Existing models restrict transcription factors (TF) to individual functional categories, and these apply across all contexts within a cell type.

Recent results suggest that instead of having a single, fixed role, TFs can play varied functions across different settings, which may give rise to diversity in phenotypic cell states. We refer to this as the “variegated model of TF function.”

This variation in function may be mediated by direct or indirect TF-TF crosstalk, or by interplay between TFs and chromatin state.

Descriptions of the many ways TFs can affect transcription emphasize that classic bind-activate (or bind-repress) scenarios vastly oversimply TF interplay and quantitative effects.

Alternatives to TF knockouts, such as natural cis-genetic variation, acute TF degradation, or continuous tuning of TF concentrations may better capture direct TF targets.

New measurement tools and computational methods will be required to develop a quantitative and predictive understanding of TF-mediated gene regulatory networks in immunity.

Significance.

Unlike previous models of transcription factor (TF) function in immunity which assigned a single role to each TF across all contexts within a cell type, recent results suggest that TFs can flexibly play different roles across distinct cell states, with implications for an understanding and control of immunocyte gene regulatory networks.

GLOSSARY

AlphaFold

Predictive model developed by DeepMind to predict folded protein structure from amino acid sequences.

Assay for Transposase-Accessible Chromatin Using Sequencing (ATAC-seq)

Assay for genome-wide chromatin accessibility based on the Tn5 transposase-mediated integration of next generation sequencing adapters selectively within nucleosome-free, accessible chromatin regions.

Chromatin Immunoprecipitation sequencing (ChIP-seq)

Assay to survey protein-DNA binding genome-wide by using protein-DNA cross-linking followed by next generation sequencing of DNA fragments pulled down using antibody-based immunoprecipitation.

cis-regulatory element

Non-coding DNA locus which regulates the expression of nearby genes on the same chromosome (in cis).

Cleavage Under Targets and Tagmentation (CUT&TAG)/ Cleavage Under Targets & Release Using Nuclease (CUT&RUN)

Two different techniques for mapping protein DNA-binding sites using antibody-targeted DNA cleavage using either Tn5 transposase (CUT&TAG) or micrococcal nuclease (MNase; CUT&RUN)

Deep learning

Subfield of machine learning using deep neural networks to extract increasingly complex features from datasets. Given enough data, such models make highly accurate predictions across a range of tasks in computer science and genomics.

Degron

Protein tag enabling inducible degradation of the targeted protein.

DNA CpG methylation

Methylation of DNA nucleotides at sites with consecutive cytosine and guanosine bases. Methylated DNA tends to be epigenetically repressed/inactive.

Enhancer

cis-regulatory element, often regulated by TF-binding and function, which increases the probability of transcription from a gene in cis. Can regulate genes over large distances.

F1 Hybrid

First generation offspring of a cross between genetically different parents, with one chromosome coming from each parent. Crosses between parents from two different mouse strains are useful to assess allele-specific gene expression in relation to strain-specific sequence variants in regulatory elements.

H3K27Ac

Histone code post-translational modification, acetylation of Histone 3, Lysine 27; indicative of active enhancer elements.

H3K4Me1

Histone code post-translational modification, mono-methylation of Histone 3 at Lysine 4; indicative of enhancer elements.

H3K4Me3

Histone code post-translational modification, tri-methylation of Histone 3, Lysine 4; indicative of active promoter elements.

Intrinsically-disordered region (IDR)

Protein region without a fixed, stable, secondary structure which can adopt many flexible conformations.

Lineage-defining TF

TF capable of opening lineage-specific regulatory elements to initiate a lineage-specific gene expression program.

Master TF

TF whose expression is both necessary and sufficient to initiate a cell-type-specific gene expression program.

Motif enrichment analysis

Bioinformatic analysis that quantifies statistical enrichment of TF binding motifs within a set of genomic loci, compared to a background set of regions.

Nucleosome remodeling

Process of changing histone-DNA interactions to displace or move nucleosomes; often involves the activity of ATP-dependent nucleosome remodeling complexes.

Nucleosome

Unit of chromatin packaging in eukaryotic nuclei, consisting of DNA wrapped around a histone octamer.

P300

Transcriptional co-activator protein with acetyltransferase activity. Enriched at active enhancers.

Perturb-seq

Functional genomics technique coupling CRISPR-Cas9-mediated gene editing with single-cell RNA-seq readouts to perform pooled genetic screens; the identity and transcriptional effect of each perturbation can be read at the single-cell level.

Pioneer factor

Capable of directly evicting nucleosomes to open chromatin regions.

Position-weight matrix

Mathematical summary of DNA binding preferences for a TF; consists of a per-position weight assigned to each possible DNA base; derived from in vitro DNA-binding arrays or in vivo DNA-binding experiments and used to scan sequences for the presence or absence of specific DNA-binding motifs.

Promoter

Non-coding DNA region located immediately at the transcription start site of a gene (typically −100 to +20 bp) where initiation complexes are assembled to initiate transcription.

Proteolysis-targeting chimera (PROTAC)

Small molecules that bind proteins of interest and bridge them to E3 ubiquitin ligases to mediate proteasome-mediated protein degradation.

Regulatory T cell

Subset of CD4⁺ T cells expressing the transcription factor FoxP3; responsible for suppressing immune responses and maintaining tissue homeostasis.

Signal-dependent TF

Binds to or regulates chromatin regions opened by LDTFs. Operates in response to specific environmental signals.

Single-cell ATAC-seq

ATAC-seq profiled at the single-cell level, where genome-wide ATAC-seq chromatin accessibility profiles can be assigned to single cells of origin.

Th17 cell

Subset of CD4⁺ T cells expressing the transcription factor Rorγ and producing the cytokine IL-17.

Thymic mimetic cells

Epithelial cells in the thymic medulla which express lineage-defining TFs and mirror gene expression and chromatin programs of a wide variety of somatic cells (myocytes, keratinocytes, enterocytes, etc); enable tissue-specific tolerance of developing T cells.

Topic modeling

Machine learning technique, initially developed for document text analysis; uncovers features with common patterns in data. In the scATAC-seq context, it can be used to group open chromatin regions with similar patterns of accessibility across single cells into “topics”: basically, co-regulated programs.

Transcription factor (TF)

Sequence-specific DNA-binding protein regulating transcription.

Treg-like cells

Treg “wannabes”; cells that differentiate in FoxP3-deficient mice and IPEX patients, and that have many features of Treg identity, including a transcriptionally active FOXP3 locus. In mice, they are easily identified in the context of FoxP3 null mutant GFP reporter mice, where GFP+ Treg-like cells meet the conditions for Foxp3 expression but do not produce functional FoxP3 protein.