IRF1 cooperates with ISGF3 or GAF to form innate immune de novo enhancers in macrophages

Abstract

Macrophages exposed to immune stimuli reprogram their epigenomes to alter their subsequent functions. Exposure to bacterial lipopolysaccharide (LPS) causes widespread nucleosome remodeling and the formation of thousands of de novo enhancers. We dissected the regulatory logic by which the network of interferon regulatory factors (IRFs) catalyzes the opening of chromatin and the formation of de novo enhancers. We found that LPS-activated IRF3 is primarily required for de novo enhancer formation indirectly through activation of the type I interferon (IFN)–induced ISGF3. However, ISGF3 generally required collaboration with IRF1, particularly where chromatin was less accessible. At these locations, IRF1 was required for the initial opening of chromatin, with ISGF3 extending accessibility and promoting the deposition of H3K4me1, marking poised enhancers. Because IRF1 expression is NF-κB–dependent, IRF-regulated enhancers require activation of both the IRF3 and NF-κB branches of the innate immune signaling network. However, type II IFN (IFN-γ), typically produced by T cells, may also induce IRF1 expression through the STAT1 homodimer GAF. We showed that upon IFN-γ stimulation, IRF1 was also responsible for opening inaccessible chromatin sites that could then be exploited by GAF to form de novo enhancers. Our results reveal how combinatorial logic gates of IRF1-ISGF3 or IRF1-GAF restrict immune epigenomic memory formation to macrophages exposed to pathogens or IFN-γ–secreting T cells, but not bystander macrophages exposed transiently to type I IFN.

INTRODUCTION

Macrophages are key components of the innate immune system that adapt their functions to microenvironmental context and in response to prior exposure to cytokines or pathogen components (1). The latter has been described as innate immune memory resulting in trained immunity (2), expanding the well-established concepts of macrophage polarization (3) and endotoxin-induced tolerance (4). There are three broad mechanisms underlying innate immune training: stimulus-induced adaptation of signaling pathways (for example, through modulation of receptor expression or induction of positive or negative signal transducers) (5), metabolic reprogramming (6), or epigenetic reprogramming by the formation of de novo enhancers (7). Latent enhancers are chromatinized genomic regions that are opened in response to a specific stimulus to increase their accessibility. When they gain mono-methylation of lysine 4 of histone H3 (H3K4me1), a marker of poised enhancers (8), they are considered de novo enhancers. Acetylation of lysine 27 of histone H3 (H3K27ac) indicates transcriptional activity and therefore marks active enhancers (9). The persistence of histone modifications even after removal of the initial activation suggests epigenetic immune memory that reprograms the macrophage’s subsequent stimulus responses (10).

Lineage-determining transcription factors (LDTFs), such as PU.1, are critical for the establishment and maintenance of macrophage lineage–specific enhancers (9, 11). In contrast, the formation of de novo enhancers in response to cytokines or pathogen-associated molecular patterns (PAMPs) is mediated by stimulus-dependent transcription factors (SDTFs), such as NF-κB (8, 12–14). The underlying mechanisms of de novo enhancer formation involve two distinguishable phases. First, SDTFs bind to cognate sequences within nucleosomal DNA, which produces nucleosomal opening and increased chromatin accessibility within a matter of minutes, potentially in cooperation with chromatin-remodeling enzymes, such as SWI/SNF (SWItch/Sucrose Non-Fermentable), FACT (Facilitates Chromatin Transcription), and RNA Polymerase II (RNAPol II) (12, 15). Second, SDTFs may contribute to the recruitment of LDTFs (for example, PU.1) to maintain open chromatin and chromatin-modifying enzymes to catalyze the deposition of H3K4me1 and H3K27ac within hours after stimulation (8, 11, 12).

Studies have shown that the formation of de novo enhancers is stimulus-specific (8, 14). NF-κB opens chromatin (14, 16) and induces hundreds of de novo enhancers, but only when it is activated with nonoscillatory dynamics (14). This dynamic requirement ensures that enhancer formation is restricted to MyD88-mediated signals emanating from bacterial PAMPs, such as lipopolysaccharide (LPS), but not, paracrine tumor necrosis factor (TNF), which activates NF-κB with oscillatory dynamics. A second set of hundreds of de novo enhancers are associated with interferon-sensitive response elements (ISREs), the cognate motif for the family of interferon regulatory factors (IRFs). Of the nine IRF family members, IRF1, IRF3, and IRF9 (which is part of the ISGF3 complex) are relevant for stimulus responses in murine M-CSF–differentiated bone marrow–derived macrophages (BMDMs) (17). A key impediment to dissecting the functional specificity of IRFs is their largely overlapping binding specificity for the ISRE GAAANNGAAACT and their interdependent activation mechanisms (18, 19). Whereas the first PAMP-responsive IRF is IRF3, its induction of IFN-β stimulates activation of the transcription factor ISGF3 (IRF9) through the interferon receptor IFNAR, which dominates the subsequent gene expression response (19). However, which IRF family member is responsible for the induction of ISRE-associated de novo enhancers is unknown.

Here, we combined a genetic approach with biochemical characterization of the IRF signaling network and epigenomic profiling to delineate the signaling roles of IRF family members in de novo enhancer formation during the innate immune response. We found that IRF3 acts directly only at a minority of de novo enhancers, but its major role is indirect through the induction of IFN-β induction and consequent ISGF3 activation. However, ISGF3 requires the coordinated function of IRF1, which is activated by NF-κB in cells responding to pathogen exposure. IRF1 is also induced by type II IFN, and in this context, it cooperates with IFN-γ-activated factor (GAF) to stimulate formation of de novo enhancers. We conclude that although IRF1 is a versatile, chromatin-remodeling SDTF, it must function combinatorially with other SDTFs (ISGF3 or GAF) to ensure that long-lasting epigenome remodeling is restricted and does not occur in all cells that gain antiviral protection from paracrine type I IFN.

RESULTS

Many LPS-induced de novo enhancers are associated with an ISRE and type I IFN signaling

Our previous studies revealed that de novo enhancers induced by endotoxin LPS in BMDMs and associated with ISRE motifs were abolished by the combined deficiency of IFNAR and IRF3 (Ifnar^−/−Irf3^−/− BMDMs), which abrogates both IRF3 and ISGF3 activity (Cheng et al. 2021). To dissect the contributions of IRF3 and ISGF3 in de novo enhancer formation (Fig. 1A), we stimulated BMDMs generated from WT, Irf3^−/−Ifnar^−/−, Ifnar^−/−, and Irf3^−/− mice with LPS (100 ng/ml) for 8 hours. We performed chromatin immunoprecipitation sequencing (ChIP-seq) with a validated antibody against H3K4me1 and identified 4800 de novo enhancer regions by applying a cutoff for the false discovery rate (FDR) of < 0.05 and for the log₂ fold change (LFC) of > 0.5 of triplicate data upon stimulation of WT cells with LPS. More than 90% of these H3K4me1 regions were in intergenic and intronic regions, whereas a minority of locations were found in TSS or exonic regions (fig. S1A). When considering knockout data, these 4800 de novo enhancers clustered into two major groups by unsupervised k-means clustering (Fig. 1B): cluster 1 (C1) was enriched for the IRF DNA-binding motif known as interferon stimulated response element (ISRE), whereas cluster 2 (C2) was enriched for NF-κB motifs. Consistent with our previous findings (14), we observed a significant loss of C1 de novo ISRE enhancer formation in Irf3^−/−Ifnar^−/− cells (Fig. 1B). Both Ifnar^−/− and Irf3^−/− single knockouts also showed deficiencies in the formation of C1 de novo ISRE enhancers (Fig. 1B), suggesting that both IRF3 and ISGF3 are required for the formation of ISRE enhancers.

Fig. 1. IRF3 and ISGF3 stimulate the formation of long-lasting de novo enhancers.

(A) Schematic of the LPS signaling network, including secondary ISGF3 activation by IFN-β. (B) Heatmap of the z-scored H3K4me1 ChIP-seq data for 4800 locations induced [log₂ fold change (LFC) > 0.5, FDR < 0.05] by 8 hours of stimulation with LPS (100 ng/ml). Clusters generated by unsupervised k-means were subjected to de novo motif analysis. Right: The top enriched sequences. Each column represents a single mouse, with n = three WT mice. (C to E) Box plots of the log₂ RPKM normalized ChIP-seq data (GSE38377) for (C) H3K27ac, (D) RNA Pol II, and (E) PU.1 in the LPS-induced C1 or C2 de novo enhancers or promoter regions of LPS-inducible genes (LFC > 0.5). Data are from a single mouse per antibody. (F) Box plots of the log₂ RPKM normalized H3K4me1 ChIP-seq on different time intervals after removal of LPS after 8 hours of stimulation. Data are from a single mouse per timepoint. (G) Boxplots of the log₂ RPKM ATAC-seq signals at C1 or C2 enhancer locations in WT, Ifnar^−/−, and Irf3^−/− (GSE234914) BMDMs treated for 3 hours with LPS or Lipie A (100 ng/ml). Data are the average of two to three mice per condition. (H) Percentage of H3K4me1 peaks overlapping with the IRF3 (GSE67357) or IRF9 (GSE115435) ChIP-seq peaks on C1 or C2 enhancer locations. Data are from two mice per condition.

To further characterize these de novo enhancers, we examined publicly available ChIP-seq datasets of H3K27ac, PU.1, and RNA Polymerase II (RNA Pol II) on LPS-stimulated BMDMs (8). We found that most de novo ISRE enhancers acquire the active enhancer mark H3K27ac at the 4-hour timepoint (Fig. 1C), but this signal is transient, as it decreases from a median log₂ RPKM value of 2.7 at 4 hours to 1.4 at the 24-hour timepoint (Fig. 1C). Similar observations pertain to the NF-κB–regulated enhancers of cluster C2. In contrast, promoter regions of LPS-inducible genes (LFC >0.5, at 8 hours) had median log₂ RPKM values of 4.2 before stimulation (Fig. 1C), indicating that many promoters were primed in naïve macrophages. Indeed, RNA Pol II occupancy at the basal state was higher in promoters (median log₂ RPKM = 5.8) than C1 (median log₂ RPKM = 3.1) or C2 enhancers (median log₂ RPKM = 3.3), but the enzyme was recruited to both IRF- and NF-κB–associated enhancers at 4 hours (Fig. 1D). We next examined the binding patterns of PU.1, the macrophage LDTF that establishes macrophage enhancers (8, 20, 21). We found that PU.1 binding was induced substantially at both the C1 and C2 enhancer locations within 4 hours of LPS stimulation and persisted for at least 24 hours (Fig. 1E). In contrast, PU.1 signals at promoters of LPS-inducible genes already high before stimulation and barely inducible. Together, these results suggest that in contrast to LPS-inducible promoters that are primed with RNA Pol II, PU.1, and H3K27ac, latent de novo enhancer regions are in a more inactive state, from which they must be activated by stimulus-induced SDTFs that initiate chromatin remodeling.

To investigate the longevity of the IRF- and NF-κB–associated enhancers, we removed the LPS after an 8-hour stimulation and performed H3K4me1 ChIP-seq in a subsequent time course. Contrary to the transient H3K27ac epigenetic changes, the H3K4me1 marks remained largely unchanged for at least 6 days after stimulation in both the C1 and C2 de novo enhancers (Fig. 1F). Indeed, only 738 enhancers decreased more than 0.5 LFC, compared to 4062 enhancers that remained unchanged. We did not find a difference in sequence composition between these enhancers. The longer-lasting temporal dynamics of H3K4me1 suggests that these enhancers remain in a “poised” state that enables rapid activation in response to subsequent stimulation, consistent with previous findings (7, 8, 14).

Our results suggest that in response to LPS exposure, the activation dynamics of de novo ISRE enhancers are similar to those previously described for NF-κB–associated enhancers, and that both IRF3 and ISGF3 are required for their formation. To further address the relative contributions of IRF3 and ISGF3, we investigated their respective roles in opening chromatin at LPS-induced enhancer locations with an ATAC-seq dataset from WT and Ifnar^−/− BMDMs (LPS-treated) and WT and Irf3^−/− BMDMs (Lipid A–treated) (16). Knockouts of either factor (Ifnar^−/− or Irf3^−/−) resulted in deficiency in chromatin accessibility within 2 hours of stimulation with LPS of C1 but not C2 enhancers (Fig. 1G). However, when we analyzed ChIP-seq datasets for IRF3 in Lipid A–treated BMDMs (22) or for IRF9 in IFN-β–treated BMDMs (23), we found that whereas IRF9 was frequently found on C1 de novo enhancers (58%), IRF3 binding to C1 and C2 enhancers was similarly low (~22%) (Fig. 1H). These results suggest that IRF3 may have not only direct but also indirect roles in de novo ISRE enhancer formation. An indirect role is also suggested by the observation that new protein synthesis is required for many ISRE-associated chromatin-opening events and that IRF3 is directly involved in a small number of locations (16).

ISGF3 is required but not sufficient to form most ISRE de novo enhancers

We considered that IRF3 may indirectly generate de novo enhancers through IFN-β production that activates ISGF3. To further understand the interplay between IRF3 and ISGF3 activity, we performed Western blotting analysis of nuclear extracts from Irf3^−/− BMDMs. We observed a deficiency in the activation of the ISGF3 subunits STAT1 and STAT2 in response to LPS (Fig. 2A), consistent with the role of IRF3 in IFN-β production (24). Thus, deficiencies in enhancer formation observed in IRF3-deficient cells may be due to defects in ISGF3 activation. To determine whether IRF3 was directly involved in enhancer formation, we rescued ISGF3 activation in Irf3^−/− BMDMs by co-stimulating LPS-treated cells with IFN-β (0.3 or 10 U/ml) after 1 hour of the LPS stimulation time course (Fig. 2A). Co-stimulation with IFN-β at 10 U/ml resulted in similar amounts of nuclear pSTAT1 and pSTAT2 as those seen LPS-treated WT cells. We therefore performed H3K4me1 ChIP-seq with Irf3^−/− BMDMs stimulated with LPS alone or with IFN-β (Fig. 2, B and C). We observed a 57% median increase in H3K4me1 signal in de novo ISRE enhancers with the addition of IFN-β (p < 0.001) (Fig. 2B), and a smaller difference in the NF-κB–regulated enhancers (Fig. 2C).

Fig. 2. ISGF3 is required but not sufficient for inducing most ISRE enhancers.

(A) Western blotting analysis of pSTAT1, pSTAT2, and the loading control p84 in nuclear extracts from WT and Irf3^−/− BMDMs stimulated with LPS, with the indicated addition of IFN-β at the one-hour timepoint. Blots are representative of two independent experiments. (B and C) Violin plots of the log₂ RPKM counts of H3K4me1 ChIP-seq for (B) C1 or (C) C2 enhancer locations. Line indicates the median of distribution. Data are from three WT mice and a single Irf3^−/− mouse. (D) Scatterplots indicating the log₂ RPKM H3K4me1 ChIP-seq counts from LPS-stimulated WT BMDMs (x-axis) versus Irf3^−/− (y-axis) (top) or additionally supplemented with IFN-β (bottom). Colors indicate three different groups determined by FC cutoff thresholds of WT compared to Irf3^−/− (WT/Irf3^−/−); C1.1 (purple) LFC > 0.5 in LPS and LFC > 0.5 in LPS+IFN-β; C1.2 (orange) LFC > 0.5 in LPS and LFC < 0.5 in LPS+IFN-β; C1.3 (gray) LFC <0.5 in LPS. (E) Heatmap of the z-scored H3K4me1 ChIP-seq data from LPS-stimulated WT, Ifnar^−/−, and Irf3^−/− BMDMs or additionally supplemented with IFN-β. Each column represents a single mouse. (F) Boxplots of the log₂ RPKM counts of ATAC-seq (GSE234914) signal overlapping with C1.1, C1.2, or C1.3 enhancer locations in Lipid A–stimulated WT or Irf3^−/− BMDMs. Data are the average of two mice per condition. (G) Boxplots of the log₂ RPKM counts of the IRF3 ChIP-seq (GSE67357) signal for Lipid A–treated BMDMs overlapping with the C1.1, C1.2, or C1.3 enhancer locations. Data are the average of two mice per condition. (H) Violin plots of the H3K4me1 LFC values for WT BMDMs stimulated with IFN-β (1 U/ml) or LPS (100 ng/ml). Line indicates median of distribution. Data are from a single IFN-b–treated mouse or the average of three LPS-treated mice. Statistical significance was determined by Wilcoxon rank-sum test. *P < 0.05, **P < 0.001, ***P < 0.0001. (I) Heatmap of the z-scored H3K4me1 CUT&Tag (C&T) data for 1548 locations induced (LFC > 2, FDR < 0.01) in human macrophages after 8 hours of stimulation with LPS (100 ng/ml), Pam3CSK (P3K; 100 ng/ml), or IFN-β (10 U/ml). Clusters generated by unsupervised k-means were subjected to de novo motif analysis. Each column represents a single mouse. Right: The top enriched sequence for each cluster. (J) Violin plots of the log₂ RPKM counts of H3K4me1 C&T signals for Cluster 1 locations. Data are an average of two mice per condition.

We further analyzed the data to distinguish between de novo enhancers on the basis of their involvement with IRF3 by classifying the C1 enhancers into three different groups based on WT-to-Irf3^−/− log fold change (LFC) differences (WT/Irf3^−/−); C1.1 enhancers were deficient in Irf3^−/− BMDMs and were not rescued by IFN-β (LFC >0.5 in LPS and LFC >0.5 in LPS+IFNβ); C1.2 enhancers were rescued by IFN-β (LFC >0.5 in LPS and LFC <0.5 in LPS+IFNβ); and C1.3 enhancers showed only a moderate decrease in Irf3^−/− BMDMs (LFC <0.5 in LPS) (Fig. 2D). Of the 1090 de novo enhancers substantially affected by IRF3-deficiency, we found that 870 were rescued by the addition of IFN-β (C1.2), whereas only 220 locations remained deficient in enhancer formation even in the presence of ISGF3 activation (C1.1) (Fig. 2E). These data indicate that the IRF3-ISGF3 axis plays an important role in the formation of enhancers, though a minority of regions may be regulated directly by IRF3.

To determine the regulatory control of chromatin opening, we examined ATAC-seq data from endotoxin-treated BMDMs (16). We observed that IRF3-regulated de novo enhancers (C1.1) were more defective in chromatin opening in Irf3^−/− BMDMs than those that were rescued by IFN-β–induced ISGF3 (Fig. 2F). In the basal state of WT macrophages, C1.1 locations showed slightly less accessibility than C1.2 locations and statistically significantly less accessibility than C1.3 locations (fig. S1B). IRF3 binding after 2 hours of stimulation with Lipid A was significantly greater at the IRF3-dependent locations (C1.1) than at the C1.2 or C1.3 enhancers (Fig. 2G). These results identify C1.1 locations as being directly IRF3-regulated.

However, because most C1 enhancers were dependent on ISGF3 and not IRF3, we asked whether ISGF3 alone was sufficient to stimulate the formation of these enhancers. We stimulated WT macrophages with IFN-β at a concentration that fully activates ISGF3 but not IRF3. We found that IFN-β had little effect on the C1.1 and C1.3 locations (median LFCs of 0.16 and 0.14, respectively), and had only a slightly greater effect on C1.2 locations (median LFC of 0.25) (Fig. 2H). In contrast, LPS induced C1.1 and C1.2 enhancers by about 1.0 median LFC. These results demonstrate that ISGF3 is necessary but not sufficient to produce most LPS-induced de novo ISRE enhancers.

We next extended the study to human macrophages. We produced macrophages from peripheral blood mononuclear cells with M-CSF and stimulated them with LPS and Pam3CSK (P3K) for 8 hours to identify NF-κB–driven and ISRE-containing de novo enhancers by CUT&Tag (C&T). K-means clustering revealed two major groups, with cluster 1 (1098 locations) being unresponsive to P3K and containing an ISRE as the most prominent motif, and cluster 2 (450 locations) being responsive to P3K and containing an NF-κB motif as the most prominent motif (Fig. 2I). We then asked whether IFN-β could induce the de novo ISRE enhancers. We chose a concentration of IFN-β (10 U/ml) that was ten-fold greater than the saturating dose for expression of the canonical ISGF3 target gene ISG15 (fig. S2). Even at this concentration, we found that IFN-β only partially induced de novo ISRE enhancers compared to LPS (Fig. 2, I and J). These results support the notion that ISGF3 is insufficient for complete formation of de novo ISRE enhancers in human macrophages, consistent with our results from experiments with BMDMs (Fig. 2H).

Formation of de novo IRF-associated enhancers requires the combinatorial activity of IRF1 and ISGF3

We hypothesized that the IRF family member IRF1 may act in concert with ISGF3 to produce ISRE de novo enhancers. IRF1 has primarily been studied in the context of type II interferon (IFN-γ) responses (8, 25, 26), but its expression was also induced by LPS-activated NF-κB (Fig. 3A). Furthermore, type I IFN–activated ISGF3 did not induce IRF1 expression, nor was ISGF3 activity reduced because of the lack of IRF1 activity (fig. S3A). These observations lead us to hypothesize that ISGF3 and IRF1 may collaborate to regulate de novo ISRE enhancer formation.

Fig. 3. ISRE de novo enhancer formation requires both IRF1 and ISGF3.

(A) Schematic diagram of the LPS signaling network indicating control of IRF1 production by NF-κB. (B) Heatmap of the z-scored H3K4me1 ChIP-seq data after 8 hours of stimulation of WT, Ifnar^−/−, and Irf1^−/− BMDMs with LPS. LPS-induced enhancers (n = 4800) are grouped into those that are IFNAR-dependent (Group 1; LFC > 0.5, FDR <0.05; WT/Ifnar^−/−), IRF1-dependent (Group 2; LFC > 0.5, FDR < 0.05; WT/Irf1^−/−), dependent on both factors (Group 3; LFC > 0.5, FDR < 0.01; WT/Ifnar^−/− and WT/Irf1^−/−), or independent of both factors (Group 4; FDR > 0.8 in WT/Ifnar^−/− and WT/Irf1^−/−). Each column represents a single mouse. (C) Boxplots of the log₂ RPKM counts of H3K4me1 ChIP-seq for the groups determined in (B). Data are the average of two to three mice per condition. (D) Percentage of the H3K4me1 peaks overlapping with binding events of IRF1 (LPS stimulation) (GSE56123, data from a single mouse) or IRF9 (IFN-β stimulation) (GSE115435, data from two mice). (E) H3K4me1 genome browser tracks of representative de novo enhancer regions of G1, G2, G3, and G4. Statistical significance was determined by Wilcoxon rank-sum test. *P < 0.05, **P < 0.001, ***P < 0.0001.

To test this hypothesis, we stimulated BMDMs of Irf1^−/− mice with LPS and performed H3K4me1 ChIP-seq. Principal component analysis (PCA) of the previously identified set of 4800 LPS-induced enhancers showed that Irf1^−/− BMDMs responded similarly to Irf3^−/− and Ifnar^−/− BMDMs (fig. S3B). To explore the relative contribution of these two factors, we used knockout data to compare to wildtype (WT/KO) and classified the LPS-induced enhancers into those that are ISGF3-dominant (Group 1; LFC >0.5, FDR <0.05 in Ifnar^−/−), IRF1-dominant (Group 2; LFC >0.5, FDR <0.05 in Irf1^−/−), IRF1- and ISGF3-dependent (Group 3; LFC >0.5, FDR <0.05 in both Irf1^−/− and Ifnar^−/−), or IRF1- and ISGF3-independent as a control group (Group 4; FDR >0.8) (Fig. 3, B to D). Whereas the WT H3K4me1 median RPKM values averaged 2.3 to 2.4 log₂ RPKM in each of the four groups, median RPKM values for Ifnar^−/− cells were 1.4 and 1.5 in Groups 1 and 3, respectively, and in Irf1^−/− cells, they were 1.5 in both Groups 2 and 3 (p <0.001 for all comparisons) (Fig. 3C). In the control group, no effect was observed in either knockout cell (Fig. 3, C and D). The H3K4me1 counts in the basal state of all genotypes were highly correlated (>0.78 spearman coefficient) for the four groups, suggesting that the poised enhancer landscape in naïve macrophages is not substantially affected by the absence of ISGF3 or IRF1 (fig. S3C). Additionally, the IRF1- and ISGF3-dependent de novo enhancers were largely distinct from the earlier described IRF3-dependent de novo enhancers, because only 9% of Group 1, 12% of Group 2, and 6% of Group 3 enhancer regions overlapped with the enhancer locations of cluster C1.1 (fig. S3D), and no deficiency in IRF1 expression was observed in Irf3^−/− BMDMs (fig. S3E).

TF motif enrichment analysis revealed only subtle differences in the connecting nucleotides between the half-sites of the ISRE: CT was most frequent in Groups 1 and 3 vs. GT in Group 2. In addition, the 3’ end of the motif in Group 1 was less well-defined (fig. S3F). These differences may partially contribute to differential IRF binding (19, 27). We examined available IRF1 and IRF9 ChIP-seq datasets of BMDMs stimulated with LPS or IFN-β, respectively (23, 28) and determined the overlap of the IRF1 and IRF9 binding locations with the de novo enhancer locations (Fig. 3D). We found slightly higher rates of IRF1 binding at IRF1-dominant de novo enhancers (95%) than ISGF3-dominant de novo enhancers (76%) and slightly higher rates of IRF9 binding at ISGF3-dominant (75%) than IRF1-dominant de novo enhancers (67%). We observed similarly frequent binding of IRF1 and IRF9 in the IRF1- and ISGF3-dependent Group 3 (>85%) and substantially less binding to locations within the IRF1- and ISGF3-independent control group (Fig. 3, D and E). The relative binding positions of IRF1 and IRF9 in relation to the H3K4me1 peaks followed a normal distribution centered at the midpoint of the H3K4me1 peak (fig. S3G), whereas the average peak width of IRF1 or IRF9 was slightly greater than 500 bp, demonstrating a high level of specificity in the analyzed region (fig. S3H). Furthermore, IRF1 and IRF9 binding co-localized in > 60% of Group 1 and Group 2 locations and >80% in Group 3 enhancer locations (fig. S3I). In contrast, only 10% of Group 4 locations had both IRF1 and IRF9 co-localized at the same locations. Together, these results suggest that although subtle differences in ISRE motif variants and IRF1 and IRF9 binding rates could be identified between the groups, these appeared to be insufficient to explain the differential factor requirement in generating ISRE de novo enhancers.

IRF1 and ISGF3 have sequential roles in the formation of ISRE de novo enhancers

To assess the mechanistic roles of IRF1 and ISGF3 in opening chromatin, we performed ATAC-seq on BMDMs stimulated with LPS (100 ng/ml) for 0 to 4 hours. We then found the overlap between ATAC-seq peaks and LPS-induced de novo enhancer regions. Using the same four groups determined earlier (Fig. 3), we observed similar overall trends of IRF1- vs. ISGF3-dependency in chromatin opening (Fig. 4, A and B, and fig. S4, A to D). Furthermore, upon close examination of individual time points, we found that the deficiency in chromatin opening in Ifnar^−/− BMDMs was not significant until 2 hours, whereas Irf1^−/− BMDMs diverged from WT cells within 1 hour of stimulation (Fig. 4, and C, fig. S4, A to D). The temporal specificity in Irf1 vs. Ifnar requirement was more prominent for the highly IRF1-dependent groups (Groups 2 and 3) than for the ISGF3-dominant group (Group 1).

Fig. 4. Early vs. late temporal roles of IRF1 and ISGF3 in chromatin opening at ISRE de novo enhancer locations.

(A) Heatmap of the z-scored ATAC-seq signal from the peaks that overlap with de novo enhancer regions (Fig. 2.4C) after LPS stimulation for the indicated times. Each column represents a single mouse. (B) Boxplot of the log₂ RPKM counts of ATAC-seq signal in group 3 locations, which are dependent on both ISGF3 and IRF1. No deficiency is seen in Ifnar^−/− at 1-hour timepoint. Data are from one mouse per genotype. (C) PCA plots of the ATAC-seq signal from groups 1, 2, and 3, at the indicated timepoints and genotypes. (D) Spearman correlation analysis of the knockout H3K4me1 signal as a percentage of WT and the knockout ATAC-seq signal as a percentage of WT. Loss of H4K4me1 signal in knockouts is generally mirrored by loss of ATAC-seq signal, but not at the 1-hour timepoint for Ifnar^−/−. (E to G) Violin plots of RPKM counts of ATAC-seq, PU.1 ChIP-seq, or RNA Pol II ChIP-seq (GSE38377) in the basal state at regions overlapping with G1–G4 enhancer locations. Line indicates the median of the distribution. Each blot represents a single mouse. Statistical significance was determined by Wilcoxon rank-sum test. *P < 0.05, **P < 0.001, ***P < 0.0001.

Next, we quantitatively compared the relationship between chromatin opening, as assessed by ATAC-seq, and the formation of de novo enhancers, as determined by ChIP-seq (Fig. 4D). As a reference point of comparison, we calculated the fraction of H3K4me1 signal loss in Ifnar^−/− and Irf1^−/− cells relative to WT cells at 8 hours for each location. We also calculated the fraction of ATAC-seq signal loss in Ifnar^−/− and Irf1^−/− cells relative to WT cells for each location at all three timepoints. We then used spearman coefficients to determine whether loss of the ATAC-seq signal correlated with loss of the ChIP-seq signal. We found that for both genotypes, at the 4-hour time point, losses of ATAC-seq signal correlated with losses of ChIP-seq signal, with ρ >0.4 for all groups of locations (Fig. 4D). However, at early time points, in Ifnar^−/− BMDMs the ATAC-seq signal did not mirror the loss of the later ChIP-seq signal; this was especially evident at the one-hour time point, where the correlation coefficient was near zero for all groups of locations (Fig. 4D). In contrast, the loss of ATAC-seq signal in Irf1^−/− cells correlated with the loss of the subsequent ChIP-seq signal even at the 1-hour time point. These results suggest that IRF1 plays a critical role in the initial steps of opening chromatin, whereas ISGF3 is important in subsequent steps, with both factors being required for the formation of de novo enhancers.

If IRF1 plays a greater role in initiating chromatin opening at early timepoints, we hypothesized that compared to IRF1-indepdent enhances, IRF1-dependent enhancers may have less chromatin accessibility under basal conditions. To characterize the basal chromatin state, we investigated available BMDM datasets and found that Group 2 and Group 3, which are highly IRF1-dependent, showed lower ATAC-seq signals and less PU.1 and RNAPol II binding than Group 1, which contains locations that have a less strict requirement for IRF1 (Fig. 4, E to G). Together, these results suggest that IRF1 plays a particularly critical role at locations where chromatin is tightly compacted and devoid of factors associated with enhancer priming and basal transcriptional activity.

IRF1 functions in concert with ISGF3 or GAF to produce de novo enhancers

Having shown that IRF1 plays a critical role in chromatin remodeling downstream of LPS, we wondered whether IRF1 plays a similar role when induced by IFN-γ–activated GAF (Fig. 5A). Stimulation of macrophages with IFN-γ leads to broad chromatin remodeling (8, 14, 29), but it is not clear which SDTFs are responsible. Previous studies have suggested, by ChIP-seq analyses, that the ISGF3 components IRF9 and STAT2 aid IFN-γ–induced transcriptional activation (23) but not chromatin opening (29). To directly assess whether ISGF3 activity was induced in BMDMs upon stimulation with IFN-γ, we performed electrophoretic mobility shift assay (EMSA) with probes for GAF and ISGF3. Our data indicate that whereas GAF was activated at the lowest concentration of IFN-γ (3 ng/ml), ISGF3 was not activated even by the greatest concentration (100 ng/ml) (fig. S5A). In addition, nuclear Western blots did not detect pSTAT2 upon stimulation by IFN-γ, whereas STAT1 and IRF1 were highly activated (fig. S5B).

Fig. 5. IRF1 cooperates with ISGF3 or GAF to stimulate IFN-γ–induced de novo enhancer formation.

(A) Schematic diagram of the LPS and IFNγ signaling pathways. (B) Heatmap of the z-scored H3K4me1 ChIP-seq data showing 2231 regions induced after 8 hours of IFNγ (100 ng/ml) stimulation [LFC > 0.5, FDR < 0.01]. IRF1-dependent and independent clusters determined by FDR < 0.05, LFC > 0.5 compared to WT. Each column represents a single mouse. (C to E) Violin plots of RPKM counts in the basal state of ATAC-seq (D), PU.1 ChIP-seq (E), and RNA Pol II ChIP-seq (GSE38377) in the IFNγ-induced enhancer locations. Line indicates the median of the distribution. (F) Heatmap of the z-scored H3K4me1 ChIP-seq data showing 2279 regions induced by LPS or IFNγ stimulation and dependent on Ifnar (for LPS) or Irf1 (for LPS and IFNγ). Regions were clustered as LPS-inducible [LFC > 0.5, FDR < 0.05] (LPS-dominant), IFNγ-inducible [LFC > 0.5, FDR < 0.01] (IFNγ-dominant), or inducible by both (common). Each column represents a single mouse. (G) Venn diagram of IRF1 (IFNγ stimulation) (GSE77886), STAT1 (IFNγ stimulation) (GSE115435, GSE33913) and IRF9 (IFNβ stimulation) (GSE115435) ChIP-seq peaks in the “common” de novo enhancer locations. Locations derived from one or two mice. (H) Boxplot of log₂ RPKM counts of H3K4me1 ChIP-seq for the “common” locations in WT, Ifnar^−/−, or Irf1^−/− BMDMs. Averages of two or three mice per condition are shown. (I) Boxplot of log₂ RPKM counts of H3K27ac ChIP-seq in basal conditions or after four hours of IFNγ (100 ng/ml) stimulation (GSE38377). Colors indicate locations with the indicated transcription factor binding. Data are from one mouse per condition. Statistical significance was determined by Wilcoxon rank-sum Test. *P < 0.05, **P < 0.001, ***P < 0.0001.

To assess the role of IRF1 in IFN-γ–induced de novo enhancer formation, we performed H3K4me1 ChIP-seq on WT and Irf1^−/− BMDMs stimulated with IFN-γ (100 ng/ml). We identified 2231 de novo enhancer regions by applying a cut-off of FDR <0.01 and LFC >0.5 on duplicate data from WT cells. Of these, 1820 IFN-γ–induced de novo enhancers were IRF1-dependent (FDR <0.01, LFC >0.5), whereas 411 appeared to be IRF1-independent (Fig. 5B). Indeed, motif enrichment analysis revealed that the top motif for IRF1-dependent enhancers was “IRF1,” whereas that for IRF1-independent enhancers was “STAT1” (fig. S5C). Similar to the LPS-induced enhancers, we also observed that the basal chromatin state in WT BMDMs was less accessible in the IRF1-dependent group than in the IRF1-independent group (Figs. 4E and 5C). Furthermore, analysis of PU.1 and Pol II ChIP-seq data from BMDMs (8) revealed that under basal conditions, the genomic regions of IRF1-dependent de novo enhancers have reduced PU.1 and RNA Pol II binding compared to the IRF1-independent enhancers (Fig. 5, D and E). These results indicate that GAF remodels chromatin without IRF1 only at a minority of locations where the chromatin state is less compacted and where there is a greater extent of PU.1 and RNA Pol II binding; however, IRF1 is required for most IFN-γ–induced de novo enhancers, and these locations contain IRF-cognate ISREs.

We asked whether IRF1 cooperates with ISGF3 (when activated by LPS) or GAF (when activated by IFN-γ) at the same enhancer locations. Of the 1820 IFN-γ—induced, IRF1-dependent de novo enhancers and the 1026 LPS-induced, IRF1- or ISGF3-dependent de novo enhancers (Groups 1 to 3; Fig. 3B), 567 locations (25%) passed the significance threshold as being induced by both LPS and IFN-y with available datasets (Fig. 5F). However, it was apparent that for the 459 LPS-dominant de novo enhancers (20%), stimulation with IFN-γ also led to a degree of activation, as did LPS for the 1258 IFN-γ–dominant de novo enhancers (55%). Focusing on the common LPS- and IFN-γ–induced locations, de novo motif analysis revealed an IRF1 binding consensus sequence (fig. S5D), which we had also identified in the ISGF3-and-IRF1-dependent Group 3 (fig. S3C).

We also explored the differences between LPS- and IFN-γ–induced de novo enhancers in human macrophages. H3K4me1 C&T on stimulated human macrophages revealed that there were 256 de novo enhancers induced by both LPS and IFN-γ, whereas 842 were predominantly induced by LPS and 391 predominantly by IFN-γ (fig. S5E). Whereas human macrophages derived from PBMCs appeared to produce fewer IFN-γ–induced enhancer than did mouse BMDMs, a marked overlap in type I and type II IFN–dependent locations was observed in both macrophage preparations. Motif enrichment analysis also confirmed that IRF1 motifs were not only present in LPS-induced de novo enhancers, but also those in common, and to a lesser extent those in the IFN-γ–dominant group, suggesting collaboration between IFN-γ–activated STAT1 (through GAF) and IRF1.

Using publicly available ChIP-seq data from mouse macrophages (23, 30, 31), we found that of the 567 common de novo mouse enhancers, 354 showed STAT1 (GAF) binding in response to IFN-γ and 452 showed IRF9 (ISGF3) binding in response to LPS (Fig. 5G and fig. S5F). Their formation with either stimulus was highly dependent on IRF1 (Fig. 5H), indicating that IRF1 cooperates with either ISGF3 or GAF, depending on the stimulus, to induce de novo ISRE enhancer formation. Closer inspection of the Irf1^−/− cell data revealed a stronger deficiency when IFN-γ was the stimulus (Fig. 5H). To further understand the cooperativity mechanism between GAF and IRF1 in these regions, we divided the common enhancer locations into those that had only IRF1 or both IRF1 and STAT1 ChIP-seq signals and compared their H3K27ac amounts (Fig. 5I). We observed greater H3K27ac abundance in response to either LPS or IFN-γ in the regions that had both IRF1 and STAT1 binding compared to those that had IRF1 only, suggesting that ISGF3 and GAF binding to these enhancer regions promoted the recruitment of enzymes that activate enhancers. Furthermore, among the LPS and IFN-γcommon enhancer locations that had both IRF1 and STAT1 ChIP-seq signals, 57% of the IFN-γ–induced STAT1 peaks were associated with an ISRE motif and only 16% with a GAS motif (fig. S5G). In comparison, the STAT1 peaks in the IRF1-independent, IFN-γ–induced de novo enhancer locations were enriched by GAS motifs (43%), rather than ISRE motifs (26%). Together, these results suggest that IRF1 has the ability to recruit GAF to ISREs to form de novo enhancers in response to IFN-γ, for example by binding directly to GAF, as previously suggested (32–34), in contrast to the sequential action of IRF1 and ISGF3 on ISREs in response to LPS.

Genes near de novo enhancers show potentiated expression in response to subsequent immune challenge

Next, we examined whether IRF1-dependent enhancers altered macrophage transcriptional responses to a subsequent challenge. We stimulated WT BMDMs with IFN-γ for 8 hours, removed the stimulus, and let the cells rest for 64 hours. We then challenged the rested BMDMs with LPS (0.1 ng/ml) and collected samples at 0, 1.5, and 3 hours for RNA-seq (Fig. 6A). To assess the effect of IFN-γ training on the LPS response, we first identified genes of interest as those whose expression was induced at LFC >0.5 at least one timepoint upon LPS challenge in PBS or IFN-γ–trained conditions. For the resulting 1337 genes, we calculated the effect of IFN-γ at the three-hour timepoint when compared to the PBS control and divided these fold changes into 10 bins (fig. S6A). In general, bins 1 to 5 were enriched for genes with GO terms related to metabolic and growth processes, whereas genes in bin 6 and higher were enriched for “Response to external biotic stimulus” or “response to other organism” (fig. S6B). Bin 10 had the highest enrichment of GO terms, and “Response to interferon-beta” was statistically significant only in this group. These results suggest that IFN-γ tolerizes the induction of genes encoding factors involved in metabolic pathways, wheras it potentiates the expression of genes encoding factors in inflammatory pathways and responses to innate immune challenges.

Fig. 6. De novo enhancers direct gene expression responses to a subsequent immune challenge.

(A) Experimental scheme for induction of innate immune memory. (B) Heatmap of the z-scored RNA-seq data from WT or Irf1^−/− BMDMs trained with IFNγ and rechallenged with LPS, showing the nearest expressed genes to IRF1-dependent IFNγ de novo enhancers (Fig. 4B). K-means clustering reveals three clusters, where cluster 3 is potentiated by IFNγ training. Each column represents a single mouse. (C) Heatmap showing the most highly enriched GO terms for the three clusters in (B). (D) Paired dot plot showing the IFNγ training effect (log₂FC) on LPS-inducible cluster 3 genes (LFC > 0.5), comparing WT (purple) and Irf1^−/− (green) genotypes. (E) Line plots of representative cluster 3 genes showing the effect of IFNγ training in WT and Irf1^−/− BMDMs. Data are from one mouse per condition. (F) Heatmap of the z-scored RNA-seq data from BMDMs trained with LPS (100 ng/ml) and rechallenged with LPS (0.1 ng/ml), showing genes within 100kb of enhancers that are LPS-inducible and IRF1- or IFNAR-dependent (Fig. 2B; Groups 1, 2, 3). K-means clustering reveals two clusters, where cluster 2 is potentiated by LPS training. Each column represents a single mouse. (G) Heatmap showing the most highly enriched GO terms for the two clusters in (G). Statistical significance was determined by Wilcoxon rank-sum test. *P < 0.05, **P < 0.001, ***P < 0.0001.

Next, we assessed the expression of genes closest to the previously identified 1820 IRF-dependent, IFN-γ–induced de novo enhancers (Fig. 5B) and identified 791 unique genes (Fig. 6B). We clustered them based on their expression response to LPS by the k-means algorithm into three groups (I1 to I3) (Fig. 6B). We found that Cluster 1 (I1) genes were not induced by LPS, and there was very little enrichment of GO terms (Fig. 6C). Cluster 2 (I2) genes were diminished by IFN-γ training and were enriched for GO terms related to metabolic pathways. Cluster 3 (I3) genes showed potentiated LPS-responsiveness after IFN-γ training. The top GO terms for cluster 3 were “response to external biotic stimulus,” “innate immune response,” and “cellular response to Type II IFN” (Fig. 6C).

We then tested the IRF1-dependency of these genes. We found that for Cluster 3 genes, the potentiation effect by IFN-γ was abrogated in Irf1^−/− BMDMs. When looking at the LPS-inducible genes of Cluster 3 (LFC >0.5), the potentiation effect of IFN-γ was reduced in Irf1^−/− cells when viewed in a pairwise comparison of fold changes (p <0.001; Fig. 6D). Genes in Cluster 3 included Ifit3 and Mx1, which encode antiviral effectors (Fig. 6E). Together, these results suggest that IRF1-dependent enhancers are associated with the potentiation of nearby genes that are also IRF1-dependent.

We next asked whether LPS-induced, IRF-dependent enhancer formation correlated with the potentiation of gene expression responses. Although it is well established that LPS treatment leads to an overall state of tolerance, whereby cells respond to a second stimulation with lower inflammatory gene expression (4), genes related to tissue repair and antimicrobial effectors may be expressed to a greater extent (35). We trained macrophages with LPS (100 ng/ml) for 8 hours and then challenged them with LPS (0.1 ng/ml) after 72 hours. We found 1192 genes that were within 100 kb of LPS-induced IRF-dependent enhancers. We filtered for those genes that were inducible at at least one timepoint of secondary LPS exposure either in PBS or LPS-trained macrophages (LFC >0.5) and found 225 genes that clustered in two distinct clusters (L1 and 2) by the k-means algorithm (Fig. 6F). Cluster 1 (L1) was characterized by tolerized genes that were enriched for GO terms such as “Regulation of gene expression” and “Regulation of cytokine production”. Furthermore, cluster 2 (L2) was characterized by potentiated genes that were strongly enriched for the GO terms “response to external biotic stimulus,” “response to other organism,” and “response to IFNβ” (Fig. 6H). Additionally, we observed a decrease in the potentiation of L2 genes upon LPS training in both Ifnar^−/− and Irf1^−/− macrophages (fig. S6C). These results suggest that IRF-regulated enhancers potentiate the expression of a subset of LPS-responsive genes in a manner that is dependent on IRF1 and ISGF3.

A pathway map for ISRE de novo enhancers

Together, these findings identify the regulatory logics of IRF family members and their collaborative relationships with STAT transcription factors that are responsible for the formation of de novo enhancers induced by LPS and type II IFN (IFN-γ). Our data support a stepwise model of enhancer formation (Fig. 6A), which has four distinguishable steps: (i) compacted chromatin that has low accessibility as determined by ATAC-seq analysis must first be opened before (ii) nucleosomes are fully displaced. Nucleosome displacement is a pre-requisite for (iii) recruiting RNA polymerase and histone-modifying enzymes to activate the de novo enhancer (H3K27Ac marks), before (iv) the enhancer assumes a poised state with H3K4me1 marks (but no H3K27ac marks) within a few hours, but that is long-lasting for the week-long duration of the experiment. We found that de novo enhancers may be generated from either highly compacted, low-accessibility latent enhancer regions or moderate-accessibility latent enhancer regions. The former show a strict IRF1 requirement to provide the initial chromatin-opening that may then be extended by the collaborating factors ISGF3 or GAF. The latter show a less strict requirement for IRF1, with GAF or ISGF3 being able to extend chromatin accessibility and induce enhancer marks.

This model informs a pathway map of distinct classes of de novo enhancers (Fig. 6B). De novo enhancers are distinguished by their association with the DNA-binding motifs of SDTFs, their stimulus-specific inducibility, their genetic requirement for SDTF family members, and their chromatin compactness vs. accessibility in the basal, naïve state. For example, of the 4800 LPS-induced de novo enhancers in our experiments, 2688 are induced by NF-κB, 220 by IRF3, and 1892 by a combination of IRF1 and ISGF3, which function sequentially. The degree of IRF1 requirement appears to be determined by their chromatin accessibility in the basal state. Similarly, of the 2200 I IFN-γ–induced de novo enhancers in our experiments, 1820 show a strict IRF1 requirement and low accessibility, induced by the combinatorial IRF1-GAF action through ISRE sites, whereas 411 appear to be induced solely by GAF from a moderate accessibility state through GAS sites. Low-accessibility ISRE de novo enhancer therefore require the combined action of IRF1 with either ISGF3 (in response to LPS) or GAF (in response to IFN-γ). In contrast, NF-κB–induced de novo enhancers require a non-oscillatory activity that emanates from MyD88 mediated stimuli (14). The biological implication of both is that during an innate immune response, de novo enhancers are formed only in cells directly exposed to pathogen and not in bystander macrophages that are exposed to paracrine cytokines, such as TNF (14) or type I IFN, which are secreted by the primary responders.

DISCUSSION

After pathogen infection, it is critical that the host mounts an appropriate immune response quickly. Innate immune memory in macrophages is a mechanism by which a response to future infections may be fine-tuned and rendered more rapid or efficacious. The fitness rationale may be that conditions that involve pathogen exposure and immune activity may be predictive of further pathogen exposure and the need to mount immune responses. Our results provide evidence for IRF1 as a key epigenomic reprogramming factor in macrophages, but one that must cooperate with STATs to establish de novo enhancers. During the innate immune response to PAMPs, the type I IFN-induced transcription factor ISGF3, which consists of STAT1, STAT2, and IRF9, extends IRF1-opened latent enhancer regions, whereas in the context of type II interferon (IFN-g), typically provided by activated T cells, the transcription factor GAF (a STAT1 homodimer) may combine with IRF1 to induce overlapping sets of de novo enhancers.

Our studies involved newly generated epigenomic profiling datasets from genetic knockouts defective in specific IRFs to classify ISRE de novo enhancers and then leveraged datasets from a number of leading laboratories in the field to characterize these enhancers further. The results paint a picture that is consistent given the numerous subtle differences in experimental protocols among different laboratories over an almost ten-year period and supports the robustness and reliability of our conclusions in mouse macrophages. Our results with human macrophages suggest that key conclusions are consistent; however, further molecular characterization in human cells may be warranted.

We found that latent enhancer regions are first opened to increase accessibility, as revealed by ATAC-seq, before recruiting the pioneer factor PU.1 and RNA polymerase II, acquiring histone modifications of de novo enhancers, and being transiently activated (H3K27Ac). After a few hours, they lose the activation mark but remain in a long-lasting poised state characterized by the H3K4me1 modification. The innate immune gene expression response of macrophages by LPS is initiated by IRF3 and amplified by ISGF3. Previous investigations determined that IRF3, although required for IFN-β production, appears to make few direct contributions to the large innate immune gene expression program (19). One outlier is the gene Ccl5, which requires a nucleosome remodeling event for full activation (Tong et al, 2016). Our studies extend this observation by identifying more than 200 genomic locations at which IRF3 is required for nucleosome remodeling and thus de novo enhancer formation. However, although this is a large number, it is dwarfed by the more than 1000 locations that show de novo enhancer formation in a manner that involves ISGF3 but not IRF3. Instead, ISGF3 cooperates with NF-κB–induced IRF1 to establish de novo enhancers. What sequence or topological chromatin features render a location IRF3-regulated versus IRF1-ISGF3–regulated remains unclear. Our studies were unable to identify distinct ISRE sequence variants or other chromatin hallmarks. We recognize that IRF7 might be compensating for the loss of IRF3 in Irf3^−/− BMDMs, expanding the number of de novo enhancers that are formed by IRF3 and IRF7; however, given the requirement for IRF1, we expect that the combined action of IRF3 and IRF7 stimulates enhancer formation only at a minority of locations.

The cooperation between IRF1 and ISGF3 appears to be through sequential actions that are both required for the formation of de novo enhancers. Even though ISGF3 is activated within 1 hour of LPS stimulation, it is not functionally required until 2 hours, whereas IRF1-deficiency results in defects in chromatin opening by 1 hour. The engagement of two stimulus-induced factors in a sequence of required steps results in a logical AND gate of combinatorial synergy, even if the factors do not interact or bind each other. Other examples of this logic include the stimulus-induced expression of many inflammatory cytokine genes that depend on NF-κB–dependent transcriptional synthesis of mRNA and p38-dependent extension of the mRNA half-life, thereby forming a logical AND gate despite acting at different, albeit sequential biochemical steps and in different subcellular compartments (Cheng et al 2017).

In the context of IFN-γ signaling, where ISGF3 inducibility was not detected, we demonstrated the cooperative action of IRF1 with GAF in forming de novo enhancers. Previous studies have shown the formation of a complex between STAT1 and IRF1 (32–34) or direct contact between STAT1 and IRF1 through chromatin looping (25, 36). We showed that the IRF1 requirement was more stringent in macrophages stimulated with IFN-γ than in those stimulated with LPS, while showing induced STAT1 binding. In addition, we found that STAT1 and GAF were co-localized with IRF1-binding locations that are enriched in ISRE sites. These observations are consistent with a model of direct cooperation between IRF1 and GAF cooperation, as opposed to the sequential model of action of the IRF1-ISGF3 pair. Furthermore, we observed greater amounts of H3K27ac in enhancer regions that were bound by ISGF3 (in response to LPS) or GAF (in response to IFN-γ). Other studies have previously reported that in the IFN-γ response, STAT1 binds to sites that are already occupied by IRF1, inducing epigenomic activation of H3K27ac (34). These results suggest that whereas IRF1 is required to initiate chromatin opening, GAF is needed to recruit enzymes that deposit epigenetic marks.

The biochemical characteristics of IRF1 and ISGF3-GAF may determine which biochemical steps each catalyzes. We note, for example, that the smaller size of IRF1 (3 7kDa; potentially functioning as a monomer) (37) may enable it to function as a pioneer factor in opening nucleosomal DNA at genomic regions where chromatin is tightly compacted, whereas the larger ISGF3 complex may stabilize and extend partially accessible DNA locations. Indeed, the prominent STAT1 activation domain may enable particularly efficient recruitment of histone-modifying enzymes and RNA polymerase to establish and transiently activate the de novo enhancer (34, 38, 39).

Whereas our analysis of IFN-γ–induced de novo enhancers identified and focused on IRF1-independent enhancers that contain GAS motifs, and IRF1-dependent enhancers that contain ISREs and largely overlap with LPS-induced enhancers, there may also be de novo enhancers that require both IRF1 and GAF that function through their respective ISRE and GAS sites. The combinatorial versatility is expanded by the fact that both ISRE and GAS elements are composed of GAAA half-sites, arranged in the former as direct repeats and in the latter as palindromes, which means that three half-sites may be sufficient for both IRF3 and GAF to function independently or sequentially.

Because IRF1 production must be stimulus-induced, either through PAMP-stimulated NF-κB activity or IFN-γ–stimulated GAF activity, the combinatorial requirement of IRF1 and a partner SDTF ensures that de novo enhancers are not formed in all cells that are exposed to paracrine type I IFN. Type I IFN–exposed cells are warned of a nearby infection and induce antiviral and other innate immune genes, but will not undergo substantial epigenomic reprogramming. In contrast, exposure to pathogen (which activates NF-κB through PRRs) or T helper 1 (T_H1) CD4⁺ T cells (which secrete IFN-γ) will reprogram the epigenome of macrophages. Such specificity through a combinatorial control mechanism mirrors the specificity of NF-κB–regulated enhancers, which are also only induced by PAMPs (14) or when conditioning with IFN-γ potentiates NF-κB activation (40). Thus, two fundamental gene regulatory mechanisms that govern stimulus-specific gene expression, combinatorial and dynamic control (41, 42), also govern the stimulus-specificity of innate immune de novo enhancer formation through the IRF-STAT and NF-κB signaling pathways, respectively.

MATERIALS AND METHODS

Animals, cell culture, and stimuli

Wild-type and specific gene-deficient C57BL/6 mice were housed and handled according to guidelines established by the UCLA Animal Research Committee under protocols ARC-2014–110 and ARC-2014–126. Bone marrow was isolated and cells were grown as previously described (43). Cells were stimulated on day 7 with 100 ng/mL LPS (L6529–1MG, Sigma-Aldrich), 1 U/mL IFNβ (12401-1, pbl assay science), or 100 ng/mL IFNγ (485-MI R&D Systems) for the times indicated in the figure legends. Human blood from de-identified human subjects was obtained from the UCLA CFAR Centralized Laboratory Support Core, according to IRB 11–000443. PBMCs were isolated by Ficoll (Cytiva) gradient centrifugation. Monocytes were purified from PBMCs with human CD14 microbeads (130-050-201, Miltenyi) according to the manufacturer’s protocol. Monocytes were plated on 6-well plates at density of 1.2×10⁶cells/well and cultured for 7 days in 3 mL of RPMI supplemented with 10% fetal bovine serum (FBS, Omega), penicillin-streptomycin, and 25 ng/mL human MCSF (300-25, peprotech). MCSF was refreshed on day 5 of culture by restoring the concentration to 25 ng/mL (assuming that all MCSF was depleted). On day 7, mature macrophages were stimulated with 100 ng/mL LPS (L6529-1MG, Sigma-Aldrich), 100 ng/mL human IFN-γ (100-2, peprotech), 10 U/mL human IFN-β (11415, PBL assay science), or 100 ng/mL Pam3CSK4 (tlrl-pms, invivogen) for 8 hours.

Biochemical analysis

Nuclear extracts were collected as previously described (43). For Western blotting analysis, the following antibodies were used: rabbit anti-IRF1 (Santa Cruz sc640), mouse anti-pSTAT1 (Santa Cruz sc136229), rabbit anti-pSTAT2 (Sigma-Aldrich 07–224), mouse anti-IRF9 (Millipore-Sigma MABS1920), and rabbit anti-p84 (Abcam ab131268), followed by mouse anti-rabbit IgG-HRP (Cell Signaling 7074) or anti-mouse IgG-HRP (Cell Signaling 7076). EMSA was performed as previously described (43, 44). For the ISRE consensus sequence we used 5’-GATCCTCGGGAAAGGGAAACCTAAACTGAAGCC-3’ and 5’- GGCTTCAGTTTAGGTTTCCCTTTCCCGAGGATC-3’; for the GAS consensus sequence we used 5’-TACAACAGCCTGATTTCCCCGAAATGACGC-3’ and 5’- GCGTCATTTCGGGGAAATCAGGCTGTTGTA-3’; and for the NFY consensus sequence we used 5’-GATTTTTTCCTGATTGGTTAAA-3’ and 5’- ACTTTTAACCAATCAGGAAAAA-3’ as a loading control.

ChIP-seq analysis

Chromatin immunoprecipitation was performed as previously described (14). ChIP-seq libraries were prepared with the NEBNext Ultra II DNA Library Prep Kit (New England Biolabs E7645). Libraries were single-end sequenced with a length of 50 bp on an Illumina HiSeq 3000. Reads were processed and aligned to the mouse genome (mm10) as previously described (14). MACS (MACS2/MACS3) (45) was used to call peaks at 1% FDR. We generated two reference peak files by merging the peaks in the LPS or IFN-γ (+unstimulated control) conditions in WT cells. We used these genomic locations to count the fragments in the WT and knockout samples for each stimulus condition with deeptools multiBamSummary (46). We used edgeR (47) to determine the significantly induced regions by applying a cutoff of FDR <0.05 and LFC >0.5 (LPS) or FDR <0.01 and LFC >0.5 (IFN-γ) compared to the unstimulated condition in WT cells. For the IRF1- or IFNAR-dependent groups (Fig. 2), significant peaks were identified by applying a cutoff of FDR <0.05 and LFC >0.5 in WT vs. Irf1^−/− or Ifnar^−/− conditions, and the control group was identified by an FDR >0.8 in WT vs. Irf1^−/− and Ifnar^−/− in the LPS inducible peaks. IRF1-dependent or -independent groups (Fig. 4) were defined by applying a cutoff of FDR <0.05 and LFC >0.5 comparing duplicates of IFN-γ–stimulated WT and Irf1^−/ samples. Analysis of de novo transcription factor motif enrichment was performed with the findMotifsGenome function in the HOMER suite (21), using all detected peaks in the WT as background. Data were visualized with ggplot2 or the pheatmap packages in R. The following ChIP-seq datasets from BMDMs were obtained from the Gene Expression Omnibus: H3K27ac (LPS or IFN-γ stimulation; GSE38377), PU.1 (LPS or IFN-γ stimulation; GSE38377), RNA Polymerase II (LPS stimulation; GSE38377), IRF3 (Lipid A stimulation; GSE99895), IRF9 (IFN-β stimulation; GSE77886), IRF1 (LPS stimulation; GSE56123), IRF1 (IFN-γ stimulation; GSE77886), STAT1 (IFN-γ stimulation; GSE115435 and GSE33913) (8, 22, 23, 28, 30, 31). Raw datasets were aligned against mm10 as previously described (14). MACS (MACS3) (45) was used to call peaks at 1% FDR. A merged file was obtained for each transcription factor, and overlaps with the stimulus-specific H3K4me1 peaks were determined with the intersect function of the Bedtools package (48).

CUT&Tag analysis

Mature stimulated and unstimulated macrophages were lifted from plates with 0.5 mM EDTA in PBS and gentle scraping. Nuclear isolation and tagmentation was performed with CUTANA CUT&Tag Kit (Epicypher) according to the manufacturer’s protocol and as previously described (49) with anti-H3K4me1 antibody (abcam ab8895). Libraries were sequenced with paired-end, 50-bp reads on an Illumina NovaSeq X Plus. Reads were processed as described for ChiP-seq with the exception that the reads were aligned to the human hg38 genome. CUT&Tag peaks were called with MACS3 version 3.0.0b1 with standard options except -f BAMPE and -q 0.01. Differential peaks were identified as described for ChIP-seq with the exception that LFC 2.0 and FDR 0.01 were used for all conditions. Motif analysis was performed as described earlier with the HOMER suite using the entire genome for background.

ATAC-seq analysis

ATAC performed as previously described (14). Libraries were prepared with the Nextera DNA Library Preparation Kit (Illumina, FC-121), and single-end sequenced (50-bp) on an Illumina HiSeq 3000. Sequenced reads were processed and aligned to the mouse genome (mm10) as previously described (14). MACS (MACS2) (45) was used to call peaks at 1% FDR. The peaks for all the ATAC-seq samples were used to generate a single reference peak file, and the number of reads that fell into each peak was counted with deeptools multiBamSummary (46). The overlap between the ATAC-seq and ChIP-seq peaks was determined with the intersect function of the Bedtools package (48). Reads were normalized by RPKM. Data were visualized with ggplot2 or the pheatmap packages in R. The Lipid A–treated WT and Irf3^−/− BMDM datasets were obtained from the GEO (GSE234914),

RNA-seq analysis

BMDMs were lysed with TRIzol reagent (Life Technologies), and total RNA was purified with the DIRECTzol RNA miniprep kit (Zymo Research). RNA samples were submitted to BGI Genomics for selection of polyadenylated RNA and paired-end library preparation. Samples were sequenced on the DNBSEQ Technology platform (100 bp). Raw data were filtered for adaptor sequences or low-quality sequences with SOAPnuke. Reads were aligned to the mm10 genome with STAR (50). Aligned reads were processed as previously described (14). Data were normalized by TPM. Genes with a TPM >5 under at least two conditions were selected. The LPS-inducible genes were determined by applying a cutoff of LFC > 0.5 in at least one timepoint in the PBS, IFN-γ, or LPS conditions. The closest genes to the IFN-γ enhancers and the LPS-inducible genes within ± 100 kilobase were based on linear proximity to the transcription start sites (TSSs). Data were visualized with ggplot2 or the pheatmap packages in R.

Fig. 7. A pathway map for the formation of innate immune de novo enhancers.

(A) Model schematic of the de novo enhancer formation pathway. The first steps involve chromatin remodeling to increase accessibility. Locations with low accessibility in naïve macrophages require IRF3 or IRF1. Once moderate accessibility is achieved, ISGF3 or GAF may extend it. The subsequent steps involve transient enhancer activation by ISGF3 or GAF, as evidenced by PU.1 and RNA Pol II recruitment, and H3K27Ac modification, and then the establishment of a durable poised state whose hallmark is H3K4me1 marks. (B) Pathway summary of innate immune de novo enhancers. From top to bottom, transcription factors induced by TNF, LPS, IFN-β, or IFN-γ, function through their indicated cognate DNA-binding motifs to induce the formation of the indicated number of de novo enhancers determined in this study, subject to the indicated regulatory requirements. Note that many of the LPS- and IFN-γ–induced locations overlap. Whereas the stimulus-specificity of NF-κB–regulated enhancers is achieved by dynamic control principles, the stimulus-specificity of IRF-regulated enhancers is achieved by combinatorial control principles. However, the biological implication is the same: only directly pathogen-exposed cells, not bystanders, induce innate immune de novo enhancers on a large scale.

Data and materials availability:

All data needed to evaluate the conclusions in the paper are present in the paper or Supplementary Materials. All epigenomic profiling data reported in this manuscript have been deposited to GEO and are publicly available: GSE241821 for mouse and GSE270507 for human.