We used RNAseq to perform a transcriptomic analysis of the effects of PGE2 on the TNF response in monocytes. Primary human monocytes were stimulated with PGE2 (280 nM) and/or TNF (20 ng/ml) and harvested 3 or 24 hr after stimulation for RNAseq analysis; the experimental design is depicted in Figure 1A. As expected, co-stimulation of monocytes with PGE2+TNF induced expression of genes enriched in inflammatory and NF-κB pathways, largely reflecting a TNF response (Figure 1B). Clustering of genes upregulated after 3 hr of stimulation (n=747; fold change >2, false discovery rate [FDR] < 0.05) based upon pattern of expression confirmed that PGE2 suppresses expression of a subset of TNF-inducible genes including TNF itself and interferon-stimulated genes CXCL9, CXCL10, and CXCL11 (Figure 1C, group III, n=86 suppressed genes). Surprisingly, PGE2 augmented TNF-mediated induction of a substantially larger number of genes (Figure 1C, groups I and V, 395 co-stimulated genes). Pathway analysis revealed that genes costimulated by PGE2+TNF were similarly enriched in inflammatory pathways (e.g., response to lipopolysaccharide, neutrophil migration), although the enriched pathways were partially distinct (Figure 1D). We also observed two groups of genes that were induced by PGE2 alone, which included canonical cAMP signaling and CREB target genes (Zhang et al., 2005) such as HBEGF and CREM (Figure 1C, groups II and IV). RT-qPCR analysis confirmed strong costimulation of IL1B, IL1A, and CSF3, induction of HBEGF by PGE2 alone, and suppression of TNF-induced expression of TNF and CXCL10 by PGE2 (Figure 1E). Similar results on the regulation of TNF-mediated gene expression by PGE2 were observed using mouse bone marrow-derived macrophages (BMDMs) (Figure 1—figure supplement 1). Subsets of inducible genes showed distinct kinetics of induction with transient expression at 3 hr versus sustained expression over the 24 hr stimulation period. Costimulation of STAT4 was confirmed at the protein level (Figure 1F). Analysis of RNAseq performed at the 24 hr time point to determine the effects of PGE2 on the late phase TNF response showed similar results and additionally revealed that PGE2 inhibits TNF-induced expression of cholesterol pathway genes, which we had previously shown to be induced by TNF with delayed kinetics and to play a role in promoting IFN responses (Kusnadi et al., 2019; Figure 1—figure supplement 2). Collectively, these results indicate that PGE2 has a dichotomous effect on TNF-induced gene expression, superinducing select key inflammatory genes such as IL1B, while suppressing expression of distinct inflammatory genes including TNF and interferon-stimulated genes in line with previous reports (Chen et al., 2012; Cilenti et al., 2021; Gerlo et al., 2011; Kawahara et al., 2015; Luan et al., 2014; MacKenzie et al., 2013; Sundberg et al., 2014; Tsuge et al., 2019).

The above-described results were similar to results we had previously observed using TNF-stimulated monocyte–synovial fibroblast cocultures (Donlin et al., 2014; Kuo et al., 2019), except that PGE2-inducible genes such as HBEGF were strongly induced by PGE2 alone, and expression was not further increased by TNF (Figure 1C, groups II and IV). These differences may reflect delayed kinetics of induction of fibroblast-derived PGE2 in cocultures or from the modulating activity of various factors and cytokines produced by activated fibroblasts in these cultures. To assess the extent to which TNF-stimulated monocyte-fibroblast cross talk could be recapitulated by stimulation of monocytes with purified PGE2 and TNF, we compared the current (TNF+PGE2)-costimulated gene set with the TNF/fibroblast-costimulated gene set previously reported (Donlin et al., 2014; Kuo et al., 2019). Strikingly, (TNF+PGE2)-costimulated genes recapitulated 61% of fibroblast-augmented TNF-inducible genes (Figure 2A, hypergeometric p=8.36e-128), which is in accord with our previous conclusions that PGE2 mediates a substantial fraction of the fibroblast effect on monocytes (Kuo et al., 2019). Since TNF-stimulated fibroblast-cocultured monocytes modeled the phenotype of RA synovial macrophage cluster 1 (C1, also termed HBEGF+ or IL1B+ macrophages) (Kuo et al., 2019; Zhang et al., 2023; Zhang et al., 2019), we next tested whether (TNF+PGE2)-stimulated monocytes expressed the defining genes of cluster 1 (128 genes, as defined in Kuo et al., 2019). Strikingly, (TNF+PGE2)-costimulated genes (Figure 2B, orange+yellow) highly significantly mimicked the RA C1 phenotype (p<10e-9 by Monte Carlo simulation) and recapitulated the C1 phenotype more completely than TNF/fibroblast-costimulated genes (Figure 2B, orange+red). The (TNF+PGE2)-costimulated genes were parts of pathogenic gene modules related to IL-1-NF-κB and Jak-STAT signaling, neutrophil chemotaxis, and the Notch pathway that has been recently implicated in RA pathogenesis (Wei et al., 2020; Zack et al., 2024; Figure 2C, blue font = C1 genes). Induction of neutrophil chemokines such as CXCL1/2/3/5 that recruit neutrophils to inflamed tissues complements induction of CSF3 that mobilizes neutrophils from the bone marrow. Additional examples of induction of RA macrophage C1 genes by PGE2 and TNF are depicted in Figure 2D. These results link (TNF+PGE2)-costimulated macrophages to a pathogenic RA macrophage phenotype and reveal that, instead of global suppressive effects in human MonoMacs, PGE2 promotes induction of genes important in various pathogenic inflammatory pathways.

We wished to test whether macrophage subsets expressing TP-costimulated genes could be detected in other types of inflammatory arthritis and thus investigated synovial macrophages from patients with ICI-arthritis. While immune checkpoint inhibitors are effective in the treatment of malignancies, ICI can lead to autoimmune side effects including inflammatory arthritis (Dougan et al., 2021). Approximately 4% of ICI-treated patients can develop autoimmune inflammatory arthritis that persists despite cessation of ICI therapy (Braaten et al., 2020), and understanding mechanisms and pathways that activate immune cells in ICI-arthritis may yield insights into the pathogenesis of spontaneous autoimmune arthritides (Cappelli et al., 2020). We performed single-cell RNA sequencing analysis of mononuclear cells FACS-sorted from synovial fluids of five ICI-arthritis patients and from two synovial tissues from one ICI-arthritis patient who underwent bilateral knee replacements. These patients had received anti-PD1 antibodies as monotherapy or in combination with anti-CTLA4 therapy (clinical characteristics of patients are presented in Supplementary file 1). Following subclustering of monocytes and macrophages, we defined eight clusters from 14,110 cells (Figure 2E). Cluster 4, accounting for 15% of cells, selectively showed elevated expression of (TNF+PGE2)-costimulated genes and known PGE2 targets, including HBEGF, IL1B, CXCL2/3/8, CREM, and PLAUR (Figure 2F, genes marked in red). Cluster 4 macrophages are distinct from those in the other clusters, such as clusters 0 and 1 that express interferon-stimulated and different inflammatory genes. These results suggest that costimulation of gene expression by TNF and PGE2 can occur in ICI-arthritis.

We next wished to analyze the relationship of ICI-arthritis myeloid cells to previously defined RA pathogenic synovial MonoMac subsets. To this end, we retrieved the RA myeloid cell scRNAseq dataset from Zhang et al., 2023. We utilized the same 15 myeloid cell clusters defined in the original study (Zhang et al., 2023) and generated an essentially similar uniform manifold approximation and projection (UMAP) plot that displayed the myeloid clusters including proposed pathogenic IL1B+ FCN1+ HBEGF+ (termed ‘inflammatory’), STAT1+ CXCL10 + (express high IFN signature), and SPP1+ (express ‘wound healing’ phenotype) monocytic subsets (Figure 3A). Reference mapping (Hao et al., 2024; Lotfollahi et al., 2024) showed that the majority of ICI-arthritis myeloid cells mapped onto four RA myeloid clusters: IL1B + FCN1 + HBEGF+, STAT1+ CXCL10+, SPP1+, and MERTK+ HBEGF+ clusters (Figure 3B). The correspondence of ICI-arthritis cells to these specific RA clusters was corroborated by an independent computational approach involving integration, reclustering, and within-cluster dataset overlap analysis of the two datasets (not shown). These results support the similarity of certain ICI-arthritis and RA myeloid cell types, but this interpretation is subject to several caveats (see ‘Discussion’). We then investigated the expression of TP gene sets in RA and ICI-arthritis myeloid clusters by calculating gene set activity scores in each cell using the AUCell package (Aibar et al., 2017) as described in ‘Materials and methods’, and visualizing gene expression on UMAP plots. The gene sets analyzed were the most synergistically induced TP genes from clusters I and V (as defined in Figure 1C), and TP genes with proposed pathogenic functions similar to the gene groups shown in Figure 2C; this gene list was comprised of IL1A, IL1B, NLRP3, CD274, RELB, CXCL1/2/3/5, STAT4, PTGS2, PTGES, REL, CEBPB, IL2RA, DLL1, JAG1, NOTCH4, CREM, HBEGF, PLAUR, NR4A1/2/3, and KDM6B. Analysis of RA myeloid cells showed that expression of the TP genes aligned most closely with the IL1B+ FCN1+ HBEGF+ cluster, although gene expression extended into subsets of adjacent SPP1+, MERTK+ LYVE1-, and DC2 clusters (Figure 3C). Notably, the TP signature was absent from the IFN signature-expressing STAT1+ CXCL10+ cells. We then analyzed the expression of TP genes in ICI-arthritis myeloid cells (projected onto RA cell clusters as shown in Figure 3B). TP gene expression in ICI-arthritis cells was apparent in the same general and contiguous area of the projected UMAP and spanned the same clusters as expression in RA cells, except TP genes were expressed in a smaller fraction of cells in some of the RA-defined cell clusters (Figure 3D). Expression of TP genes in a central and contiguous region of the UMAP is suggestive of a functional myeloid cell subset associated with TP gene-related activity that is partially distinct from clusters previously identified in Zhang et al., 2023. Collectively, these results highlight the correspondence of ICI-arthritis myeloid cells to RA monocytic cells in defined pathogenic RA cell clusters and identify myeloid cell subsets that show evidence of co-stimulation by TNF and PGE2 in RA and ICI-arthritis.

We wished to test whether both the pro- and anti-inflammatory effects of PGE2 on TNF-induced gene expression were mediated by cAMP signaling. We addressed this question using selective agonists of PGE2 receptors EP2 and EP4 that signal predominantly via cAMP, and also cell membrane-permeable dibutyryl-cAMP that directly activates this pathway. Selective agonists of EP2 and EP4 augmented TNF-induced expression of IL1B, IL1A, STAT4, and CXCL1, while suppressing induction of the ISG CXCL10 (Figure 4A). Dibutyryl-cAMP showed similar results (Figure 4B). These results implicate cAMP signaling in both anti- and pro-inflammatory effects in this experimental system.

We then performed ATACseq analysis of chromatin accessibility (Figure 5—figure supplement 1A) to gain insight into how PGE2 costimulates TNF-induced gene expression. In accord with different signaling pathways activated by these factors, TNF and PGE2 induced mostly distinct ATACseq peaks (Figure 5A); indeed, out of 15,190 induced ATACseq peaks (FDR < 0.05, fold induction >2), only 1547 peaks (10.2%) were commonly induced by individual stimulation by TNF or PGE2 (Figure 5A, columns 4+7). As expected, TF binding motifs most significantly enriched under TNF-induced ATACseq peaks corresponded to AP-1 and NF-κB binding sites; additional enriched motifs included RUNX and IRF1 sites, the latter of which is in accord with the role of IRF1 in the delayed TNF-induced autocrine IFN response (Yarilina et al., 2008; Figure 5B, left). Similar to TNF, PGE2 induced peaks enriched for AP-1 motifs and additionally induced peaks with highly significant enrichment of CEBPD and NR4A1 motifs (Figure 5B, middle), which is in accord with PGE2-mediated induction of expression of TFs belonging to these three TF families (Altarejos and Montminy, 2011; Zhang et al., 2005). In line with the largely distinct open chromatin regions (OCRs) induced by TNF and PGE2, GREAT analysis of genes associated with these peaks showed enrichment of genes in largely distinct pathways, with TNF-induced peaks associated with genes in immune and cytokine pathways, whereas PGE2-induced peaks were associated with genes in cell matrix interactions and chemotaxis (Figure 5—figure supplement 1B and C).

To examine how these disparate factors could interact to costimulate gene expression, we extended our analysis to peaks that were co-induced by both TNF and PGE2 (Figure 5A, column 4; n=1531) and peaks that were induced only under conditions of costimulation (Figure 5B, column 2, n=3337; termed T+P unique). TNF and PGE2 cooperated to increase chromatin accessibility at co-induced peaks (Figure 5C, upper), and the most significantly enriched motifs were AP-1 (Fra1, also known as FOSL1) and CEBPE (Figure 5C, lower). These co-induced peaks were associated with immune and inflammatory response genes (Figure 5—figure supplement 1D). The peaks that were uniquely induced by PGE2+TNF were most significantly enriched for AP-1, CEBP, and ISRE/IRF motifs, and also for GAPBA, KLF7, and PBX1 binding sites (Figure 5B, right, and data not shown), and were associated with immune response and cytokine genes (Figure 5—figure supplement 1E). Collectively, the results suggest that TNF and PGE2 signals converge on gene regulatory elements associated with inflammatory genes to increase chromatin accessibility and induce de novo enhancers and highlight a potential role for AP-1 and CEBP TFs in costimulating TNF-induced gene expression.

We then used ChromVAR (Schep et al., 2017) to assess induction of TF activity by PGE2, TNF, or their combination (Figure 5D). This analysis reinforced the notion that, relative to TNF alone, the PGE2+TNF costimulation condition was characterized by increased AP-1 and CEBP activity, and also NR4A1 activity. We then performed a TOBIAS analysis that uses footprinting to measure actual occupancy of TF motifs within ATACseq peaks (Bentsen et al., 2020). TNF stimulation induced the most significant occupancy of NF-κB and AP-1 motifs, whereas PGE2 stimulation induced AP-1 and CEBP occupancy (Figure 5—figure supplement 1F). A direct pairwise comparison of T+P to T revealed the most significant increases in AP-1 and CEBP occupancy (Figure 5E). Overall, all three complementary approaches show that, relative to TNF alone, PGE2 costimulation boosts AP-1 and adds CEBP activation, and these factors can cooperate to increase chromatin accessibility at commonly targeted sites and to induce de novo enhancer formation.

In inflamed RA synovium, macrophages that express a ‘TP signature’ are also exposed to IFN-γ (Kuo et al., 2019; Orange et al., 2018; Zhang et al., 2023; Zhang et al., 2021; Zhang et al., 2019), which not only induces ISGs but also augments TLR- and TNF-induced expression of various inflammatory NF-κB target genes including TNF, IL6, and IL12 family members (Mishra and Ivashkiv, 2024). We next examined the effects of IFN-γ on the expression of (TNF+PGE2)-costimulated inflammatory genes. Surprisingly, IFN-γ essentially completely abolished induction of inflammatory genes IL1B and CXCL2, and the Notch target gene HEY1 (Figure 6A); as a specificity control, induction of STAT4 was minimally affected. We next tested whether IFN-γ broadly suppressed (T+P)-costimulated genes by analyzing the effects of IFN-γ on induction of subsets of these genes in our RNAseq dataset, as defined in Figure 6—figure supplement 1A. Interestingly, IFN-γ suppressed expression of only approximately 27% of genes that were most clearly synergistically induced by (T+P) (Figure 6B; this group included IL-1 pathway genes NLRP3 and PTGS2). Similarly, IFN-γ suppressed induction of only a subset of additional genes that were costimulated by (T+P) (Figure 6—figure supplement 1B and C; these suppressed genes included IL-1 family members IL1B and IL1A, neutrophil chemokine genes CXCL1/2/3/5/8, and canonical cAMP signaling targets such as EREG and NR4A1/3). Thus, IFN-γ suppressed expression of genes in the key RA pathogenic pathways described in Figure 2. This opposition between IFN-γ and PGE2 was also apparent in the regulation of a subset of TNF-induced genes that were suppressed by PGE2, whose expression was substantially higher under IFN-γ-stimulated conditions (Figure 6C and D). Genes more highly expressed under IFN-γ-stimulated conditions included inflammatory genes such as TNF and ISGs such as CXCL10. Collectively, the results support a biology whereby PGE2 and IFN-γ at least in part oppose each other by regulating TNF-induced gene expression in different directions and suggest that IFN-γ inhibits select inflammatory pathways while promoting others.

To obtain a broader view of the effects of IFN-γ on the entire TP-induced gene response, we performed a general linear model with between-treatment interaction contrast using our RNAseq dataset (Figure 7A and B). The conclusion that IFN-γ opposes a large component of the (TNF+PGE2) response was supported by the identification of three positive interaction clusters (ICs) comprised of genes regulated in opposing directions when IFN-γ was added to (T+P) (Figure 7A, ICs IC2, IC3, and IC6 where red and blue lines diverge). Genes whose suppression by TP was reversed by IFN-γ (IC2) were enriched in IFN pathways and regulated by IRF TFs, implicating IRFs in opposing negative regulation of gene expression by (T+P). In contrast, genes that were induced by (T+P) but strongly suppressed when IFN-γ was added to the TP condition were enriched in Myc targets (IC6), suggesting counter-regulation of metabolic pathways in these terminally differentiated and nonproliferating cells. Interestingly, IFN-γ minimally affected the (T+P)-induced NF-κB response (IC4). Genes that comprise the two negative ICs are upregulated (IC1) or downregulated (IC5) by IFN-γ or (T+P), but in combination IFN-γ did not provide an additional effect.

To explore the regulatory drivers of ICs, we selected all interacting differentially expressed TFs (|logFC|>1 in the interaction contrast) and built a protein association network using stringDB (Figure 7C). The size of a node reflects the number of unique functional connections between a node and other nodes in the network and edges connect associated nodes. Notably, TFs were grouped together with their targets in several ICs, for example, IC2 contains IRF2 and IRF8 and their targets; IC4 contains NF-κB1,2 and RELB and their targets; IC6 contains Myc and its targets (Figure 7B and C). Although RNA levels of Jun are regulated only weakly, its ultimate activity and binding site repertoire are affected by relative levels of dimerization partners and interacting proteins (Chang et al., 2018). For example, BATF2 (IC2) and BATF (IC6) were up- and down-regulated by the combination of IFN-γ and (T+P) relative to (T+P), which potentially affects relative abundance of JUN/BATF/IRF ternary complexes. Similarly, CEBPA, CEBPE (IC3), and MYC (IC6) are regulated in a reciprocal fashion: IFN-γ relieves a (T+P) inhibition of CEBPA and CEBPE yet also concomitantly decreases MYC mRNA levels. A recent report describing negative regulation of Myc by CEBPA/E (Theilgaard-Mönch et al., 2022) suggests that a similar mechanism is involved in IFN-γ-dependent inhibition of Myc and its targets. Overall, the interaction analysis indicates that IFN-γ does not alter the core NF-κB response (which is activated primarily by TNF) and suggests that instead IFN-γ regulates the PGE2 response, at least in part by modulating PGE2-induced AP-1 and CEBP factors.

We tested this notion that IFN-γ regulates the PGE2 response using combined transcriptomic and chromatin accessibility analysis. Analysis of our RNAseq data (Figure 8—figure supplement 1) revealed two clusters of genes C2 and C5 that were induced by PGE2 but not by IFN-γ (Figure 8A); the great majority of these PGE2-inducible genes were inhibited by IFN-γ, and C2 genes were more strongly inhibited than C5 genes. Strikingly, IFN-γ suppressed induction of canonical PGE2-cAMP target genes such as CREM, HBEGF, HIF1A, PLAUR, KDM6B, AREG, VEGFA, and genes in RA pathogenic gene modules and expressed in RA synovial macrophage cluster 1 (corresponding to cluster 1 in Kuo et al., 2019 and Figure 2), such as IL-1 and Notch pathway genes and neutrophil chemokines CXCL1/2/3/5/8 (Figure 8A and B, Figure 8—figure supplement 1B). Thus, IFN-γ broadly suppresses the PGE2 response, including genes important in RA pathogenesis.

We then analyzed our ATACseq data to gain insight into mechanisms underlying negative regulation of PGE2 responses by IFN-γ (Figure 8—figure supplement 1C, Figure 8C). IFN-γ suppressed 1765 out of 2411 (72%) of PGE2-induced ATACseq peaks (Figure 8C, column G2 depicts suppressed peaks). These IFN-γ-suppressed peaks were most significantly enriched in AP-1 and CEBP motifs (Figure 8D; the most significantly enriched motif TGAGTCA closely resembles an AP-1 site). ChromVAR analysis confirmed that PGE2 induced AP-1 activity, and this activity was suppressed by IFN-γ (Figure 8E). Additionally, NR4A1/2, NFE2L2, and MAF activity was induced by PGE2 and suppressed by IFN-γ. In contrast to the HOMER de novo motif analysis described above that focused on IFN-γ-suppressed peaks, ChromVAR analysis, which takes into account all peaks, showed that both PGE2 and IFN-γ induced CEBP activity. Remarkably, at the mRNA expression level, PGE2 and IFN-γ both induced and repressed different members of the CEBP TF family, with coordinate induction of CEBPB, but opposing regulation of CEBPA, CEBPD, and CEBPE (Figure 8F). TOBIAS footprinting analysis showed that both PGE2 (see Figure 5E) and IFN-γ induced CEBP occupancy (Figure 8G); the latter confirms findings in our previous report that used a different independent dataset (Mishra and Ivashkiv, 2024). Direct comparison of IFN-γ+PGE2 relative to PGE2 alone using TOBIAS showed striking suppression of AP-1 occupancy and further supported IFN-γ-mediated suppression of NR4A2, NFE2L2, and MAF (Figure 8H). Collectively, these results show that IFN-γ inhibits PGE2 responses by suppressing AP-1 and a PGE2-induced transcriptional program mediated by AP-1, NR4A1/2, NFE2L2, and MAF TFs, and suggest that IFN-γ and PGE2 costimulation remodels the composition and genomic binding profile of CEBP proteins to be distinct from that induced by each stimulus alone.

Deidentified buffy coats were purchased from the New York Blood Center following a protocol (# 2016-958) approved by the Hospital for Special Surgery Institutional Review Board. Peripheral blood mononuclear cells (PBMCs) were isolated using density gradient centrifugation with Lymphoprep (Accurate Chemical), and monocytes were purified with anti-CD14 magnetic beads from PBMCs immediately after isolation as recommended by the manufacturer (Miltenyi Biotec). Monocytes were cultured at 37°C, 5% CO₂ in RPMI-1640 medium (Invitrogen) supplemented with 10% heat-inactivated defined fetal bovine serum (FBS) (HyClone Fisher), penicillin-streptomycin (Invitrogen), L-glutamine (Invitrogen), and 20 ng/ml human macrophage-colony stimulating factor (M-CSF).

Animal experiments were approved by the Weill Cornell Medicine IACUC Committee (protocol # 2015-0055). Male C57BL/6J mice at 6–8 weeks old were purchased from the Jackson Laboratories and housed under specific pathogen-free conditions. Bone marrow cells were harvested after euthanasia by CO₂ asphyxiation and cultured in RPMI-1640 medium (Invitrogen) supplemented with 10% heat-inactivated defined FBS (HyClone Fisher), penicillin-streptomycin (Invitrogen), L-glutamine (Invitrogen), and 20 ng/ml mouse M-CSF.

Total RNA was isolated using the RNeasy Mini Kit (QIAGEN) following the manufacturer’s instructions. Reverse transcription of RNA into complementary DNA (cDNA) was performed using the RevertAid RT Reverse Transcription Kit (Thermo Fisher Scientific) according to the manufacturer’s protocol, and the resulting cDNA was used for downstream analysis. For quantitative real-time PCR (qPCR), Fast SYBR Green Master Mix (Applied Biosystems) and a QuantStudio5 Real-time PCR system (Applied Biosystems) were used. CT values obtained from qPCR were normalized to the housekeeping gene GAPDH. The primers used for qPCR were: CXCL10: F: 5′-ATTTGCTGCCTTATCTTTCTG-3′ R: 5′-TCTCACCCTTCTTTTTCATTGTAG-3′; TNF F: 5′-AATAGGCTGTTCCCATGTAGC-3′ R: 5′-AGAGGCTCAGCAATGAGTGA-3′; GAPDH: F: 5′- ATCAAGAAGGTGGTGAAGCA-3′ R: 5′-GTCGCTGTTGAAGTCAGAGGA-3′; IL1B: F: 5′- TTCGACACATGGGATAACGAGG; R: 5′-TTTTTGCTGTGAGTCCCGGAG; CSF3: F: 5′-CCAGGAGAAGCTGGTGAGTG; R: 5′-GAAAAGGCCGCTATGGAGTT; HBEGF: 5′- F:AGGAGCACGGGAAAAGAAAG; R: 5′-CTCAGCCCATGACACCTCTC; IL1A: F: 5′-AGTAGCAACCAACGGGAAGG; R: 5′-AAGGTGCTGACCTAGGCTTG; STAT4: F: 5′-GAGACCAGCTCATTGCCTGT; R: 5′-CAATGTGGCAGGTGGAGGAT. Relative expression of target genes was calculated using the ΔCt method, where ΔCt represents the difference in threshold cycle values between the target gene and GAPDH. The results are presented as a percentage of GAPDH expression (100/2^ΔCt). Data points were not omitted from analysis.

Primary human monocytes were washed with cold PBS after indicated treatments and harvested in 50 ul cold lysis buffer containing Tris-HCl pH 7.4, NaCl, EDTA, Triton X-100, Na₃VO₄, phosSTOP EASYPACK, Pefabloc, and EDTA-free complete protease inhibitor cocktail. After a 10 min ice incubation, cell debris was pelleted at 16,000×g at 4°C for 10 min. The soluble protein fraction was combined with 4×Laemmli sample buffer containing 2-mercaptoethanol and subjected to SDS-PAGE (electrophoresis) on 4–12% Bis-Tris gels. Following the transfer of gels to polyvinylidene difluoride membranes, membranes were blocked in 5% (w/v) bovine serum albumin in TBS with Tween-20 (TBST) at room temperature for at least 1 hr. Incubation with primary antibodies (diluted 1:1000 in blocking buffer) was performed overnight at 4°C. The p38 antibodies were from Cell Signaling Technology (cat# 9212) and Stat4 antibodies (C46B10 rabbit mAb) were from Cell Signaling Technology (cat# 2653). Membranes were washed three times with TBST and probed with anti-rabbit IgG secondary antibodies conjugated to horseradish peroxidase (diluted 1:5000 in blocking buffer) for 1 hr at room temperature. Enhanced chemiluminescent substrates (ECL western blotting reagents (PerkinElmer, cat# NEL105001EA)) were used for detection, followed by visualization on autoradiography film (Thomas Scientific, cat# E3018). Restore PLUS western blotting stripping buffer (Thermo Fisher Scientific) was applied for membranes requiring sequential probing with different primary antibodies.

Libraries for sequencing were prepared using mRNA that was enriched from total RNA using NEBNext Poly(A) mRNA Magnetic Isolation Module, and enriched mRNA was used as an input for the NEBNext Ultra II RNA Library Prep Kit (New England Biolabs [NEB]), following the manufacturer’s instructions. Quality of all RNA and library preparations was evaluated with BioAnalyser 2100 (Agilent). Libraries were sequenced by the Genomic Resources Core Facility at Weill Cornell Medicine, obtaining 50 bp single-end or paired-end reads to a depth of ~15–20 million reads per sample. Read quality was assessed and adapters trimmed using fastp. Reads were then mapped to the human genome (hg38) using STAR aligner and reads in exons were counted against Gencode v37 with Featurecount. Differential gene expression analysis was performed in R with edgeR using the quasi-likelihood framework. Only genes with expression levels exceeding 4 counts per million reads in at least one group were used for downstream analysis. Benjamini–Hochberg FDR procedure was used to correct for multiple testing. Genes were categorized as upregulated if log2FC ≥1 and FDR ≤ 0.05 threshold was satisfied, downregulated if log2FC ≤ –1 and FDR ≤ 0.05. Heatmaps were obtained using the Morpheus web application and replotted using the R package pheatmap. To analyze the regulatory factors involved in IFN-γ - (T+P) interactions, we selected all TFs based on Lambert et al., 2018 from the interacting clusters (Figure 6) and used them to build a protein associating network with StringDB (Szklarczyk et al., 2019) limiting the associations to the highest interaction score (0.9) and visualized with R tidygraph framework using circular layout. RNAseq experiments were performed using different blood donors: (1) monocytes stimulated by vehicle control or T, P, or TP for 3 or 24 hr (three donors). (2) Monocytes primed with or without IFN-γ and then stimulated with vehicle control, T, P, or TP for 3 hr (three different donors). Aliquots of cells from this experiment were used for ATACseq.

One million monocytes were lysed using cold lysis buffer (10 mM Tris-HCl, pH 7.4, 10 mM NaCl, 3 mM MgCl₂, and 0.1% IGEPAL CA-630), and nuclei were immediately pelleted at 500×g for 10 min in a refrigerated centrifuge. The pellet was resuspended in a transposase reaction mix consisting of 25 µl 2×TD buffer, 2.5 µl transposase (Illumina), and 22.5 µl nuclease-free water. The transposition reaction was carried out for 30 min at 37°C. Following transposition, the sample was purified using a MinElute PCR Purification kit. Library fragments were amplified using 1×NEB next PCR master mix and 1.25 M custom Nextera PCR primers, with subsequent purification using a QIAGEN PCR cleanup kit, yielding a final library concentration of ∼30 nM in 20 µl. Libraries were amplified for a total of 10–13 cycles and subjected to high-throughput sequencing at the Genomic Resources Core Facility at Weill Cornell Medicine with 50 bp paired-end reads. Data on ATAC-seq experiments were derived from three independent experiments with different blood donors.

For ATACseq data analysis, read alignments were performed against the GRCh38/hg38 reference human genome. Peak calling was conducted using MACS2 with the following parameters: "macs2 callpeak -f BAMPE -g hs -q 0.01 --nomodel --shift 37 --extsize 76 ". A master consensus peak set was generated by merging the resulting peak files for each treatment condition, followed by merging peaks within 50 bp of each other. Quantification of peaks to compare global ATACseq signal changes in the BAM files was conducted using the NCBI/BAMscale program. Raw count matrices were obtained utilizing the BAMscale program. Subsequent analysis utilized the HSS Genomic Center Bioinformatic Core’s ATACseq analysis pipeline (https://gitlab.com/hssgenomics/Shiny-ATAC; Oliver, 2023) for peak filtering, annotation relative to genomic features, differential peak analysis, and enrichment of signal around specific motifs using ChromVAR (Schep et al., 2017). Footprint analyses were performed using TOBIAS (Bentsen et al., 2020) according to the user manual. Briefly, we filtered the JASPAR2022-CORE_vertebrates TF list based on the TF expressed in our RNAseq data and used this as TF input for motif footprinting. De novo TF motif enrichment analysis was carried out using the motif finder program findMotifsGenome in the HOMER package, focusing on the given peaks. Peak sequences were compared to random genomic fragments of the same size and normalized G+C content to identify enriched motifs in the targeted sequences.

All research using patient samples at the Hospital for Special Surgery adhered to approved protocols (IRB protocols 2017-1898 and 2017-1871) with informed consent as required; samples investigated in this study correspond to a subset of those subjected to a distinct analysis in Wang et al., 2023. Synovial fluids were collected from ICI-arthritis as discarded fluid during patient visits. Tissue samples were collected during arthroplasty surgery. Mononuclear cells were isolated from synovial fluid by density centrifugation using Ficoll-Paque Plus (GE Healthcare), preserved in Cryostor CS10 (Stemcell Technologies), and placed in liquid nitrogen for long-term storage. Tissues received post-surgery were processed into fragments of 2–3 mm³, preserved in Cryostor CS10, and placed in liquid nitrogen for long-term storage.

For sample preparation for single-cell RNA sequencing, tissues were thawed into pre-warmed media at 37°C. The media composition was RPMI 1640 (Corning) containing 10% defined FBS (Cytiva) and 1% of 200 mM L-glutamine (Gibco). The tissues were then washed twice in RPMI alone for 5 min and chopped into fine pieces using a blade. Finely cut tissue was subjected to 30 min in a digestion buffer composed of RPMI, Liberase TL (100 μg/ml; Roche) and DNaseI (100 μg/ml; Roche). The fragments were digested in 5 ml polystyrene tubes (5 ml/sample of digestion buffer), securely placed in a MACSmix tube rotator (Miltenyi Biotec) in an incubator at 37°C with 5% CO₂. The tissue preparation was filtered through a 70 µM cell strainer (BD) and mashed using the back of a syringe plunger. The eluate containing the cells was washed with media and centrifuged at 1500 rpm for 4 min at 4°C. The pellet was resuspended in media and filtered again through a 40 µM cell strainer (BD). Cells in the eluate were then counted and used in downstream experiments. Fluid samples were thawed from cold storage into pre-warmed media at 37°C and then processed similarly to tissue samples. Live mononuclear cells were then FACS-sorted from five synovial fluid and two synovial tissue samples. Cells were sorted on a three-laser BD FACS Aria Fusion cell sorter at the Flow Cytometry Core Facility at Weill Cornell Medicine (WCM). Intact cells were gated according to forward scatter and side scatter area (FSC-A and SSC-A). Doublets were excluded by serial FSC-H/FSC-W and SSC-H/SSC-W gates (H, height; W, width). Non-viable cells were excluded based on DAPI uptake, sorted through a 70 μM nozzle at 70 523 psi. Flow cytometric quantification of cell populations was performed using FlowJo v.10.0.7.

3′ gene expression (GEX) libraries of live synovial mononuclear cells were prepared using Chromium Single Cell 3′ v3 Kit reagents and protocols provided by 10X Genomics. The pooled libraries at 10 nM concentration were sequenced using NovaSeq6000 S2 Flow Cell using the Illumina platform at Genomics Research Core Facility at WCM. 10× FASTQ files were processed with the Cellranger count 4.0 pipeline with default parameters. Reads were aligned to the human reference sequence GRCh38. Seurat package (v.4.0.0) was used to perform unbiased clustering. The 3′ GEX dataset of sorted live synovial cells had QC performed to also remove cells with less than 150 genes, more than 7500 genes, or >25% mitochondrial gene expression, resulting in a total of 44,959 cells and 26,424 genes. This dataset was then log-normalized using a scale factor of 10,000. Potential confounders such as percent mitochondrial gene expression and number of UMI per cell were regressed out during scaling (mean of 0 and variance of 1) for 3′ GEX dataset.

Principal component analysis was used with the top 2000 highly variable genes. Elbow plot was used to determine the statistically significant principal components of 17 PCs for follow-up analysis. Harmony (v1.0) was performed to improve integration and correct for batch effects on our samples, with parameters of max.iter.cluster=30, and max.iter.harmony=20 and sample as the only covariate. Eleven clusters for the 3′ dataset at 0.2 resolution were found, and their identity was annotated based on the expression of differentially expressed genes (DEGs) using the FindAllMarkers function using default parameters. Additional grouping of the 3′ GEX dataset of select clusters at resolution 0.2 was used to define only monocytes and macrophages to create a new data object. These clusters contained 14,110 cells after going through linear dimensional reduction of the top 5 PCs with Harmony as described above. Eight clusters were defined at 0.3 resolution for the final analysis.

To compare the ICI-arthritis scRNAseq data to a previously reported RA CITEseq dataset, we first retrieved the RA myeloid cell data from Zhang et al., 2023; available on Synapse syn52297840. Reference mapping was performed using the uwot model (https://jlmelville.github.io/uwot/authors.html) provided in syn52297840 with Seurat’s MapQuery function as described (Hao et al., 2024). For analysis of TP gene expression in RA and ICI-arthritis myeloid cells, gene set activity scores were calculated using the Bioconductor tool AUCell (v1.24.0) in R (v4.3.2) on the scRNA-seq datasets. AUCell ranks genes based on their expression in each cell and computes the area under the curve (AUC) to quantify the enrichment of predefined gene sets. The threshold for determining active gene sets (corresponding to the pink and red dots in Figure 3) was automatically set by the AUCell algorithm using the default settings (AUCell_buildRankings and AUCell_calcAUC functions), which identifies the ranking threshold based on the distribution of expression ranks across cells (Aibar et al., 2017).

GraphPad Prism was used for all statistical analysis. Information about the specific tests used and number of independent experiments is provided in the figure legends.

Sequencing data from this study have been deposited at GEO (RNAseq: GSE272019; ATACseq: GSE272017) and will be publicly available from the date of publication. The scRNAseq data object is available at https://gitlab.com/hssgenomics/icia_myeloid.