Peptidylarginine Deiminase IV Regulates Breast Cancer Stem Cells via a Novel Tumor Cell–Autonomous Suppressor Role

Breast cancer is a global problem accounting for almost a quarter of all cancers in women and despite therapeutic advances, 20% to 30% of treated patients still suffer from cancer relapse and metastasis (1). The cancer stem cell hypothesis has generated new insights into this problem. Many tumors consist of a phenotypic hierarchy of tumor cells, with a small subpopulation of cancer stem cells (CSC) at the apex, giving rise to proliferative but progressively differentiating nontumorigenic cells (2, 3). These rare CSCs are uniquely capable of the self-renewal necessary for tumor initiation and sustainability. CSCs have been identified in many tumors including breast cancer (4), and are thought to be the key to cancer recurrence, due to their intrinsically elevated resistance to a wide range of therapeutic approaches (3). Thus, a better understanding of the mechanisms regulating CSC dynamics may lead to the design of more effective therapeutics.

The hierarchical organization of tissues and tumors is maintained by differences in the epigenome (5, 6). Central to this process is the dynamic regulation of histones and other proteins by posttranslational modification, of which, methylation and acetylation are the most studied (7, 8). However, other posttranslation modifications are also emerging as important. Peptidylarginine deiminases (PADI) are a family of Ca²⁺-dependent enzymes that catalyze citrullination of arginine residues to modulate the activity of target proteins, with the citrullinome now comprising several hundred proteins (reviewed in ref. 9). Citrullination results in the loss of a positive charge, leading to altered biological activity and/or protein interactions. Among the five member PADI family, only PADI4 has a classic nuclear localization signal (10), suggesting a particularly important role in modulating activity of nuclear targets. Within the nucleus PADI4 targets histones (H1, H2A, H3, and H4) and modulates activity of some transcription factors (e.g., p53) through effects on nuclear proteins (e.g., ING4), as well as affecting targets outside the nucleus, including extracellular matrix (e.g., Col1), and intracellular proteins (e.g., vimentin; reviewed in ref. 9). In doing so, it regulates many biological processes, including apoptosis, neutrophil extracellular trap (NET) formation, and pluripotency (11–13). NET formation, a host defense mechanism, is the most dramatic manifestation of the effect of PADI4 citrullination. In this process, PADI4 hypercitrullinates histones in neutrophils, facilitating global chromatin decondensation and extrusion of a DNA web (the NET) to trap extracellular pathogens (14).

Posttranslational modification of histones is key to maintenance of the pluripotent stem cell state in embryonic stem cells, and of the stem phenotype in normal tissue and disease (8). Several recent studies have implicated citrullination of chromatin components by PADI4 in the regulation of stemness and differentiation. Indeed, PADI4 was first identified in human myeloid leukemia cells, where it was induced by differentiating agents such as retinoic acid (15). PADI4 suppressed proliferation of multipotent hematopoietic stem cells, and promoted their differentiation through effects on histone H3 and altered transcriptional regulation (16). Conversely in the early mouse embryo, citrullination of histone H1 by PADI4 displaced it from chromatin, resulting in global chromatin decondensation, thereby creating open chromatin necessary for reprogramming of pluripotency/stemness (13).

PADIs have been most extensively studied in the context of inflammation and autoimmunity (17). However, important roles for PADIs in tumorigenesis are emerging. Both PADI2 and PADI4 have been shown to have context-dependent roles in many different types of cancers (9). PADI2 promotes cell proliferation and invasiveness in gastric cancer, while suppressing cell growth and metastasis in liver cancer (18) and tumor cell invasion in breast cancer (19). In MCF7 breast cancer cells stimulated with estrogen, PADI4 transcriptionally represses estrogen receptor (ER) target genes via H4 citrullination (20), while activating the ELK1 oncogene under EGF simulation (21). In addition, dysregulation of PADI4 in breast cancer cells activated TGFβ signaling and induced an epithelial-to-mesenchymal transition (EMT), increasing the invasive phenotype of xenografted tumors (22).

The recent connections between PADI4 and stemness in other systems led us to hypothesize that PADI4 may play a role in breast cancer progression by regulating the CSC phenotype. Here, we use pharmacologic and genetic approaches in breast cancer models to generate new mechanistic insights into the role of PADI4 in breast cancer progression.

The MCF10CA1h and MCF10Ca1a cell lines were obtained from Dr. Fred Miller, Karmanos Cancer Institute (Detroit, MI). MCF7E (early passage MCF7 cells) were a gift from Dr. Michael Brattain, Eppley Institute (Omaha, NE). MDA-MB-231 LM2 cells were a gift from Dr. Joan Massague, Memorial Sloan Kettering Cancer Center (New York, NY). Additional breast cancer cell lines were all obtained from the ATCC. Cell lines not obtained directly from ATCC were authenticated by short terminal repeat (STR) analysis using 15 polymorphic markers and one gender marker (Laragen Inc.). All cell lines were routinely tested using Mycoplasma PCR Detection Kit (Sigma, MP0035) and shown to be Mycoplasma free. Cells were used for experiments within 2 to 12 passages after recovery from frozen stocks. Growth conditions are given in Supplementary Table S1. Proliferation and apoptosis were assessed using Click-iT imaging assays (Invitrogen). Invasion and migration were assessed by Transwell assays with or without Matrigel. Tumorsphere-forming activity was assessed in semisolid MethoCult medium (Stem Cell Technologies).

Cells were stably transduced with lentiviruses expressing PADI4 shRNAs for knockdown (Supplementary Table S2), or various PADI4 isoforms for overexpression, and were used after recovery from selection.

Cells were treated with 1 μmol/L PADI4 inhibitor GSK484 (Cayman Chemical) or its inactive analogue GSK106 for 3 to 10 days prior to assay, as indicated.

All animal experiments were conducted under protocol LC-070 approved by the Animal Care and Use Committee of the National Cancer Institute. To quantify the frequency of CSCs in vivo, serial dilutions of tumor cells (1,000–250,000 cells) were orthotopically implanted into immunodeficient mice. After 60 days, tumor incidence data were analyzed using extreme limiting dilution assay (ELDA) software (http://bioinf.wehi.edu.au/software/elda).

Cells were transduced with a modified version of the SORE6 fluorescent cancer stem cell reporter (23), followed by FACS for SORE6⁺ cells (CSC enriched) and SORE6- cells (CSC depleted).

To assess possible effects of PADI4 inhibition on the epigenome/transcriptome, breast cancer cells were treated with the GSK484 PADI4 inhibitor or inactive analogue GSK106 for 3 days and then harvested for chromatin immunoprecipitation sequencing (ChIP-seq) for multiple Histone H3 marks (see below), and for RNA sequencing (RNA-seq). Because poised promoters play a particularly important role in stem cell biology (24), H3K4me3 and H3K27me3 marks were used to identify loci containing bivalently marked peaks under the different treatment conditions, and the data were integrated with transcriptomic data to address whether PADI4 inhibition led to changes in expression of relevant stem/differentiation factors.

All ChIPs were done with SimpleChIP Enzymatic Chromatin IP Kit (Cell Signaling Technology). Briefly, cells (4 × 10⁶ cells/IP) treated with GSK484 or control GSK106 for 72 hours were fixed with formaldehyde and lysed, and chromatin was fragmented by partial digestion with micrococcal nuclease. For immunoprecipitation (IP), 5 to 10 μg of digested, cross-linked chromatin per IP was incubated with antibodies for total Histone H3, H3K4me3, H3K27me3, H3R2R8R17citr, H3R17me2, H3R2me2, or IgG. ChIPed DNA was purified and 5 ng was processed for NG-sequencing DNA library construction for all marks except H3R17me2 and H3R2me2, which were just analyzed by ChIP-qPCR. Sequencing libraries were run on Illumina NextSeq platform SR 75 bp cycle runs of 50 M reads/sample. The enrichment of specific DNA sequences during IP was validated by ChIP-qPCR. In some experiments, cells transduced with the SORE6 reporter were FACS-sorted into SORE6⁺ (CSC) and SORE6⁻ (non-CSC) compartments prior to ChIP-qPCR at targeted loci. Primers and antibodies used for ChIP are listed in Supplementary Tables S3 and S4, respectively. The fastq sequence reads were mapped to the human hg19 reference genome and peaks were identified using MACS2. Heatmaps were generated using deepTools2.

Total RNA was extracted using TRIzol Reagent (Ambion) and all RINs were 9 or above. For RNA-seq, sequencing libraries were generated according to Illumina's TruSeq Stranded mRNA Sample Preparation protocol and run on Illumina HiSeq platform PE 125 bp cycle runs of 200 M reads/sample. For microarray analysis, RNA samples were labeled and hybridized to the Clariom S Human Array (Affymetrix).

Experiments were analyzed for statistical significance using GraphPad Prism, version 8 (GraphPad Software Inc). Where indicated, the t test (2 samples) or Dunnett multiple comparison test (>2 samples) were used. A P value of 0.05 was considered statistically significant.

More detailed methods are available in the Supplementary Data and Methods.

ChIP-seq, RNA-seq, and microarray datasets relating to this study are available from GEO as SuperSeries GSE138506.

To address the prevalence of PADI4 in breast cancer, expression was assessed in a panel of breast cancer cell lines. qRT-PCR showed that PADI4 mRNA trended to slightly higher expression in luminal subtype cell lines when compared with basal subtype cell lines, with the notable exceptions of MDA-MB-231 LM2, a lung-tropic subclone of MDA-MB-231 (25), and MDA-MB-436 (Fig. 1A). A similar nonsignificant trend was seen in a larger publicly available dataset of 50 breast cancer cell lines (Supplementary Fig. S1A), while in human breast tumors in the TCGA cohort, there was no significant difference in PADI4 expression between basal and luminal subtypes (Supplementary Fig. S1B). In three lines examined in more detail, PADI4 mRNA expression was equal to or higher than the other three PADI family members (Fig. 1B). PADI4 protein was not readily detectable by Western blot analysis of total cell lysates, but analysis of nuclear extracts revealed multiple PADI4 protein bands in the range of 65 to 43 kDa (Fig. 1C and D). PADI4 full-length transcript is predicted to translate to a protein of 73 kDa. However, several bands have previously been detected, depending on the tissue of origin (https://www.proteinatlas.org/ENSG00000159339-PADI4/antibody#western_blot), and RefSeq predicts eight alternatively spliced PADI4 transcripts encoding seven protein isoforms (Supplementary Fig. S2A and S2B) of unknown functional significance. The protein bands marked by arrows in Fig. 1C and D were confirmed to be PADI4 by shRNA knockdown in three representative cell lines (see the next section). Consistent with the RNA data, luminal breast cancer cell lines generally expressed slightly higher levels of PADI4 protein than basal lines, although only MDA-MB-231 LM2 cells expressed high levels of full-length PADI4 (Fig. 1C and D). We selected luminal MCF10Ca1h, MCF7E and basal MDA-MB-231 LM2 models for further study, as these breast cancer cell lines had the highest PADI4 protein levels and captured the range of different protein isoforms expressed.

PADI4 function in the breast cancer models was addressed by shRNA knockdown. Three shRNAs reduced PADI4 protein expression by 60% to 70% (shown for sh79 and sh80 in Fig. 1E and F), and significantly decreased PADI4 mRNA expression (Fig. 1G). Citrullination of histone H3 at arginine R2, R8, R17, and R26 was reduced by 2- to 5-fold, confirming a reduction in PADI4 catalytic activity (Fig. 1H). Addressing biological responses, PADI4 knockdown had no effect on proliferation in vitro (Fig. 2A; Supplementary Fig. S3A), but enhanced cell survival in response to staurosporine, a potent inducer of apoptosis (Fig. 2A; Supplementary Fig. S3B), consistent with known inhibitory effects of PADI4 in cell survival (26, 27). No effect on apoptosis under basal or low serum conditions was seen (Supplementary Fig. S3C). PADI4 knockdown led to significantly increased cell invasion and migration (Fig. 2A; Supplementary Fig. S3D), as seen by others (22). Clonogenicity was also increased in all lines except MDA-MB-231 LM2 (Fig. 2A; Supplementary Fig. S3E). Thus, PADI4 suppresses multiple biological processes relevant to tumor progression.

The ability to form tumorspheres under anchorage-independent conditions is an in vitro surrogate for cancer stem cell activity (28). Tumorsphere-forming efficiency increased upon PADI4 knockdown in all cell lines tested (Fig. 2B and C), suggesting that PADI4 inhibition may increase the size of the CSC subpopulation. To demonstrate a dependency on PADI4 enzymatic activity, we inhibited PADI4 pharmacologically using GSK484, a citrullination inhibitor with >35-fold selectivity for PADI4 over other PADs, or its inactive analogue GSK106 (12), and observed a 3-fold reduction in Histone H3 citrullination after 3 days (Fig. 2D). Pretreatment with GSK484 for 3 days significantly increased tumorsphere formation (Fig. 2E) and invasion and migration (Fig 2F). Prolonged treatment (8–10 days) with GSK484 further increased tumorsphere formation to levels comparable with the genetic knockdown, while having no additional effect on invasion/migration (Supplementary Fig. S3F and S3G). Overall, these data suggest that endogenous PADI4 activity may reduce the size of the CSC population and that this effect is dependent on its citrullination activity.

To further explore the effect of PADI4 on CSCs, we overexpressed PADI4 in MCF10Ca1h cells. In our Western blots, we observed several endogenous PADI4 bands (Fig. 1C and D). Given that the PADI4 effect on the CSC population was dependent on its citrullination activity (Fig 2D and E), we overexpressed the five RefSeq-predicted protein isoforms of PADI4 that retained most or all of the C-terminal catalytic domain; full-length PADI4 cDNA (FL), and four shorter protein isoforms: X1, X2, X3, and X7 (see Supplementary Fig S2). We detected all five protein isoforms in the transduced MCF10Ca1h cells by Western blot analysis (Fig. 2G and H). All isoforms ran 2 to 8 kDa smaller than predicted (Supplementary Fig. S4A), suggesting some structural feature or posttranslational modification that reduces the apparent molecular weights. This observation is consistent with our inability to detect a band at 75 kDa corresponding to the predicted size of the full-length transcript endogenously in any breast cancer cell line. X3 and X7, which lack the predicted nuclear localization signal, were still found in the nucleus at levels higher than the endogenous protein (Supplementary Fig. S4B). All isoforms, including X2, which lacks two of the four predicted catalytic residues, gave up to a 4-fold increase in PADI4 catalytic activity over control, that was further enhanced following estradiol and calcium ionophore treatment (Fig. 2I; Supplementary Fig. S4C). Despite the wide range of protein expression, all isoforms showed similar maximal catalytic activity (varying <2-fold), suggesting that PADI4 protein levels may not be the major limiting factor for PADI4 activity. Indeed, all five PADI4 isoforms similarly suppressed tumorsphere formation in MCF10Ca1h cells (Fig. 2J). These results support the hypothesis that PADI4 may inhibit the CSC subpopulation, and show that this occurs with all predicted isoforms that retain the majority of the catalytic domain.

The gold standard assay for cancer stem cell activity is the ability to initiate tumorigenesis in vivo (29). To confirm the increase in CSC frequency following PADI4 knockdown, we performed ELDA in vivo in immunocompromised mice using four different breast cancer models. Groups of female virgin athymic nu/nu mice were transplanted with progressively lower tumor cell inocula and tumor incidence was assessed at day 60 (Table 1). The relative stem cell frequency was 2- to 7-fold higher following PADI4 inhibition in all four breast cancer models tested. These results confirm that endogenous PADI4 limits the size of the CSC population in breast tumors in vivo.

Earlier work had suggested that PADI4 could affect transcriptional programs of self-renewal and differentiation through histone citrullination and modification of the epigenome (15, 16). To address effects of PADI4 inhibition on the cancer cell epigenome, MCF10Ca1h cells were treated with the PADI4 inhibitor GSK484, or its inactive analogue (GSK106, hereafter “CON”) for 3 days, and then subjected to ChIP-seq analysis for histone H3, H3R2/8/17citr, H3K4me3 and H3K27me3, and RNA-seq. The short treatment timeframe was selected to highlight direct effects of PADI4 inhibition. Concurrent enrichment of the repressive H3K27me3 mark with an activating H3K4me3 mark identifies “bivalent domains” that are characteristic of poised promoters of many genes involved in lineage specification and differentiation in embryonic stem cells (24).

The heatmaps in Fig. 3A–D show relative ChIP-seq intensity for the GSK484-treated versus CON condition in ± 10 kb regions flanking the transcriptional start site (TSS) for genes encoding all RefSeq NM transcripts. The most prominent effect of PADI4 inhibition was a widespread increase in H3 binding to chromatin, particularly between the TSS and +400bp (Fig. 3A). Total cellular H3 protein levels were unaffected by PADI4 inhibition (Supplementary Fig. S5A), and PADI4 inhibition had little effect on genomic distribution of the other histone marks (Fig. 3B–D). The increased H3 loading was seen at the TSS (Fig. 3F), and at proximal enhancers within ± 4 kb of the TSS (Fig. 3G), but was much weaker at distal enhancers (Fig. 3H). The increased H3 binding on PADI4 inhibition is consistent with the expectation that blockade of arginine citrullination increases the net positive charge on H3, resulting in tighter binding to DNA. We noted that the H3 binding increase on PADI4 inhibition was negatively correlated with H3 binding in the control (Pearson r = −0.178, P = 2.35E−281; Supplementary Fig. S5B and S5C), suggesting the possibility that TSS sites associated with the more open (nucleosome-depleted) chromatin may be more accessible to binding of new H3. Interestingly, the heatmap appeared almost identical between H3.CON and H3R2/8/17citr.CON (Supplementary Fig. S5B and S5D), indicating that all H3 histones carried a similar level of citrullination across the region surrounding the TSS. Matched heatmaps for the other two histone marks and RNA-seq are given in Supplementary Fig. S5E–S5K. Surprisingly, there was essentially no change in levels of citrullinated H3 around the TSS following 3 days of PADI4 inhibition (Fig. 3B), suggesting that the citrullination modification is very stable, and that PADI4 inhibition leads to loading of unmarked H3 that dilutes but does not displace the pre-existing citrullinated H3. The functional consequence of the increased H3 loading at the TSS is unclear, asKit occurred to a similar extent at loci where transcription was increased or decreased by PADI4 inhibition (Supplementary Fig. S6A).

RNA expression across the transcriptome was negatively correlated with H3 binding near TSS (Pearson r = −0.127; P = 2.58E−143), and was positively correlated with the activation mark H3K4me3 (r = 0.17; P = 4.15E−262) but negatively correlated with the repressive H3K27me3 mark (r = −0.05; P = 4.52E−262) in the control state (Supplementary Figs. S5B–S5K and S6B–S6K). Thus, the major histone marks were behaving as expected with respect to transcriptional activity. However, there were no widespread changes in global H3K3me3, H3K27me3 occupancy or global RNA expression following PADI4 inhibition (Fig. 3C–E). Indeed, even at an individual level, surprisingly few genes were differentially expressed following the 3 days of PADI4 inhibition. Only 94 genes were affected by GSK484 and not control compound (P < 0.05) and of these, only 24 showed an absolute fold change ≥ 1.5 (Supplementary Table S5, Tab1). The minimal effect of GSK484 treatment on gene expression may reflect a difference in kinetics between PADI4 control of chromatin and its readout in gene expression. However, because we saw clear biological effects of PADI4 inhibition in this 3-day timeframe, it is also possible that the biological output is driven by the integrated effect of multiple small changes in gene expression, most of which are below our detection limit.

In relation to the original hypothesis that PADI4 might preferentially target poised promoters, we identified genetic loci that had both H3K4me3 and H3K27me3 marks within ± 4 kb of the TSS (Supplementary Fig. S7). Further analysis at the peak level identified 448 promoters that were bivalent in either GSK484 or control samples or both (Supplementary Fig. S8A; Supplementary Table S5, Tab2, 3). However, there was almost no overlap of these poised promoters with the differentially expressed gene list generated by RNA-seq on GSK484-treated cells, or with a larger gene list generated by microarray analysis following prolonged PADI4 inhibition by genetic knockdown (Supplementary Fig. S8B and S8C). Thus the observed biological effects of PADI4 inhibition could not be attributed to changes in state of poised promoters. Overall, PADI4 inhibition had a profound effect on histone H3 loading at TSS, but this altered epigenetic state had minimal effects on the transcriptome of the bulk tumor cell population, at least in the short term.

Our epigenetic and transcriptomic analyses of the bulk tumor cell population following PADI4 inhibition gave no clear clues as to the molecular mechanism underlying the effect of PADI4 on the CSCs. We therefore reasoned that PADI4 might be having effects specifically in the CSC compartment, that our bulk analyses did not have the sensitivity to detect. To test this hypothesis, we transduced MCF10Ca1h and MDA-MB-231 LM2 cells with the SORE6 cancer stem cell reporter (Fig. 4A; ref. 23) to facilitate separation of CSCs from non-CSCs. This reporter has been rigorously validated in these cell lines as enriching for cells with the expected CSC properties, including the ability to initiate and sustain tumorigenesis through multiple transplant generations in vivo (23). The SORE6 reporter identified a minority population of cells as CSCs in both cell lines (Fig. 4B), and treatment of the unsorted cultures with the GSK484 PADI4 inhibitor led to the expected increase in CSCs (Fig. 4C). Cells were treated with either the PADI4 inhibitor (GSK484) or inactive analog (GSK106) for 3 days, and then FACS sorted into CSCs (SORE6⁺) and non-CSCs (SORE6⁻) for transcriptomic and ChIP-qPCR analysis (Fig. 4D). qRT-PCR showed no significant differences in PADI4 mRNA expression between CSCs and non-CSCs in multiple cell lines (Supplementary Fig S9A). As expected, basal expression of stem cell transcription factors was higher in the CSC compartment of untreated cells, with NANOG, POU5F1 (OCT4), SOX9, and ZEB1 being higher in the CSC compartment in both cell lines, and MCF10Ca1h additionally showing higher expression of SNAI (Supplementary Fig. S9B), confirming enrichment of the stem phenotype. Pathway analysis of RNA-seq datasets for compartment-specific effects of PADI4 inhibition indicated an increase in invasion, migration, and metastasis and a reduction in cell death specifically in the non-CSCs and unsorted tumor cells (Fig. 4E), suggesting that the non-CSC compartment may be the main target of these previously described tumor suppressive effects of PADI4. In contrast, PADI4 inhibition was predicted to suppress cell-cycle progression and transcription specifically in the CSC compartment (Fig. 4E). These data suggest that PADI4 has distinct biological activities in the CSC and non-CSC compartments, despite similar levels of expression, with the compartment-specific effects possibly due to predicted differences in epigenetic context (5, 30).

Although we were unable to detect core stem cell transcription factors in our RNA-seq libraries, they were detectable in the same RNA samples by qRT-PCR. Importantly, we found that PADI4 inhibition upregulated NANOG and POU5F1/OCT4 mRNA specifically in the CSC population in both breast cancer models (Fig. 4F and G), while having no effect (POU5F1) or the opposite effect (NANOG) in the non-CSC compartment. This upregulation of NANOG and POU5F1 was not evident in bulk (unsorted) cells following short-term GSK484 PADI4 inhibitor treatment (Supplementary Fig. S9C) or long-term genetic PADI4 knockdown (Supplementary Fig. S9D). Indeed, PADI4 inhibition actually decreased NANOG expression in the bulk (unsorted) cell population in the MCF10Ca1h model, consistent with the observed effect on the non-CSCs. We speculate that PADI4 inhibition may enforce the stem phenotype in CSCs while enforcing a more differentiated phenotype in the non-CSCs. Other stem cell transcription factors were relatively unaffected by PADI4 inhibition in the CSC compartment (Supplementary Fig. S9E). Although we did not directly test the functional role of NANOG and POU5F1/OCT4 by the gold standard in vivo limiting dilution assay in the current study, NANOG and POU5F1/OCT4 have previously been shown to induce/maintain a stem phenotype in many breast cancer models (see e.g., refs. 31, 32). Importantly, knockdown of NANOG in the MCF7 and MDA-MB-231 models that we used in this study decreased the CD44⁺CD24⁻ fraction that is enriched for stem phenotype, as well as decreasing mammosphere formation (31). Thus, PADI4 may decrease the size of the CSC population by inhibiting transcription of these master stemness transcription factors specifically in the CSC compartment. Because of the relatively small size of the CSC compartment, this effect cannot readily be seen when analyzing the bulk tumor cell population.

Citrullination of histone arginines by PADI4 can counteract arginine methylation by protein arginine methyltransferases (PRMT), with variable effects depending on whether the specific arginine methylation is normally activating or repressive (20, 33, 34). A recent study found that PADI4 repressed CTCF transcription by keeping the activating H3R17me2a histone mark low, while stimulating IL6ST transcription by lowering the repressive H3R2me2a mark (35). We confirmed these effects of PADI4 at the CTCF and IL6ST loci in the MCF10Ca1h cells, and observed them to occur primarily in the non-CSC compartments (Supplementary Fig. S10A and S10B). No effect of PADI4 inhibition was seen on either mark at the GAPDH or MYOD1 promoters used as negative controls (Supplementary Fig. S10C and S10D. To address whether a similar mechanism might underlie the effect of PADI4 on POU5F1 and NANOG expression, we assessed H3K4me3, H3R17me2a, and H3R2me2a histone marks at the NANOG and POU5F1 promoters by ChIP-qPCR in sorted CSCs and non-CSCs from MCF10Ca1h or MDA-MB-231 LM2 cells treated with either the PADI4 inhibitor GSK484 or the inactive analogue GSK106. Normalizing to total H3, we found that PADI4 inhibition increased the activating H3R17me2a mark at the NANOG and POU5F1 promoters in CSCs in both cell lines (Fig. 4H and I). We also observed an enrichment of the activating H3K4me3 mark at these loci in the CSCs following PADI4 inhibition (Fig. 4J and K). No effect was seen on the repressive mark H3R2me2a at NANOG and POU5F1 promoters in the CSCs of either cell lines (Supplementary Fig. S10E and S10F). These data show that endogenous PADI4 downregulates NANOG and POU5F1 expression specifically in CSCs by blocking establishment of the activating H3R17me2a histone mark at their promoters, and thus suppressing the transcription of these master transcription factors of stemness.

Finally, we wished to address whether the tumor cell-autonomous suppressive effects of PADI4 might have clinical relevance. In this study, we found a poor correlation between PADI4 mRNA, protein and enzymatic activity, likely reflecting multiple layers of control over PADI4 activity. This observation complicates the interpretation of studies that seek to correlate PADI4 mRNA or protein levels directly with outcome in clinical cancer datasets. Furthermore, stromal PADI4 expression can have the opposite effect on cancer progression (36, 37). To overcome this problem, we sought to generate a gene expression signature that specifically reflects tumor cell–autonomous responses to PADI4, and is enriched for PADI4-driven tumor suppressor effects described by us and others.

In our RNA-seq analyses of GSK484-treated cells, we saw relatively little impact on gene expression following short-term (3 days) pharmacologic PADI4 inhibition, despite this treatment clearly reducing citrullination activity (Fig. 2D), and affecting migration, invasion, and tumorsphere formation (Fig. 2E and F). Given the stability of the citrullination mark, we reasoned that there might be a delay before readily detectable changes in the transcriptome are seen. We therefore performed microarray analysis of MCF10Ca1h cells with and without stable shPADI4 knockdown to look for long-term effects (Fig. 5A). shRNA knockdown reduced PADI4 mRNA levels by 5-fold (Fig. 1G) and citrullination activity by >2-fold (Fig. 1H). Consistent with the RNA-seq data, principal component analysis showed that PADI4 knockdown had a relatively small effect on the transcriptome as the samples did not clearly segregate by genotype (Supplementary Fig. S11). However, differential gene expression analysis using DESeq2 identified 181 genes that were affected specifically by PADI4 knockdown (P < 0.01, FC ≥ 1.5; Fig. 5A; Supplementary Table S5, Tab4). Also consistent with results from our smaller RNA-seq genesets (Fig. 4E), Ingenuity Pathway Analysis (IPA) showed the expected enrichment for functional annotations relating to cell migration and survival (Fig. 5B). IPA analysis further identified a connection between PADI4 and estrogen signaling as seen in the literature (20, 33), and suggested provocative links between PADI4 and inflammatory mediators such as TGFβ, IL1, and TNF (Fig. 5C). Hallmark gene set enrichment analysis identified only the EMT program as showing a weak enrichment on PADI4 knockdown (Fig. 5D), consistent with our observations of increased cancer stem activity and a previous study implicating PADI4 in suppressing EMT (22). Thus despite the relatively small number of affected genes and the low magnitude of the expression changes on PADI4 knockdown, the genelist nevertheless captures components of tumor suppressive effects of PADI4 on multiple biological processes, with processes affecting the bulk tumor cell population being most prominent.

Next, we generated a 20-gene PADI4 inhibition signature from the differentially expressed genes identified above (Supplementary Table S5, Tab5 for signature genes, and Supplementary Fig. S12 for functional annotation). Expression of the signature was lowest in luminal A subtype tumors, suggesting that these have the highest PADI4 tumor suppressor activity (Supplementary Fig. S13). In the METABRIC dataset of 2,000 patients with breast cancer (38), patients whose tumors showed high expression of the PADI4 inhibition signature had significantly worse overall survival compared with those with low expression (Fig. 5E), an effect that was more prominent in ER⁺ breast cancers (Fig. 5E). We saw similar results with the same signature using the GOBO tool for a metanalysis of independent breast cancer datasets (Supplementary Fig. S14A–S14C). These results suggest that PADI4 has tumor cell–autonomous suppressive effects that can affect outcome in human breast cancer, particularly within the ER⁺ subtype.

Epigenetic regulation plays an essential role in specifying cell phenotype in the differentiation hierarchies of normal and diseased tissue (39). Histones and their associated code of posttranslational modifications are major epigenetic modifiers that regulate gene expression (7). Citrullination catalyzed by PADIs is an understudied posttranslational modification that has recently been implicated in the regulation of stemness and differentiation in embryonic and hematopoietic stem cells (13, 15, 16). Here, we demonstrate for the first time that endogenous PADI4 suppresses breast cancer progression by limiting the size of the cancer stem cell population in multiple breast cancer models in vitro and in vivo. This effect is mediated through epigenetic regulation of NANOG and POU5F1 expression specifically in the CSC compartment.

Our data suggest an important role for PADI4 as a tumor suppressor in breast cancer. The existence of multiple PADI4 isoforms, and the poor correlation between expression levels and citrullination activity, pose challenges for interpreting studies that correlate PADI4 expression with outcome in human cancer cohorts. In one large study, PADI4 was shown to be overexpressed in many different human tumors, including breast cancer, suggesting it might play a tumor-promoting role (40). In contrast, our preclinical findings clearly demonstrate that endogenous PADI4 functionally restricts the size of the CSC population in multiple breast cancer models and suppresses tumor formation. We also confirmed the findings of others that PADI4 can suppress additional pro-tumorigenic activities such as cell survival, invasion, and migration (22, 27). The ability of PADI4 to promote differentiation of promyelocytic leukemia cells, and suppress EMT and the development of more invasive tumors in the MCF7 breast cancer model is also consistent with a tumor suppressor role (22, 41). Supporting the potential clinical relevance of PADI4-driven tumor suppression, our gene signature reflecting tumor cell-autonomous PADI4 activity was associated with good outcome in clinical breast cancer cohorts. In addition, the PADI4 locus is genetically altered in 8/816 sequenced cases of breast cancer in the TCGA cohort, with 1 amplification, 5 deep deletions, and 2 missense mutations of unknown significance. The deep deletions are also consistent with a potential role for PADI4 as a tumor suppressor.

Reconciling these disparate observations, we propose that PADI4 may have opposing effects on the tumor cell and on the tumor stroma in regulating tumor progression. Under physiologic conditions, PADI4 is mainly expressed in peripheral blood neutrophils to provide antibacterial innate immunity via the formation of NETs (14). Interestingly, metastatic breast cancer cells were recently shown to induce neutrophils to form metastasis-supporting NETs in the absence of infection, in a PADI4-dependent manner (42). Furthermore, NETs could awaken dormant metastatic breast cancer cells in mice (36) and facilitate premetastatic niche formation in ovarian cancer models (43). In a mechanistically distinct process, citrullination of the extracellular matrix by PADI4 secreted in tumor exosomes facilitated human liver colorectal cancer metastasis through effects on tumor cell motility and plasticity (44). These data strongly suggest that PADI4 in stromal cells, or secreted into the extracellular microenvironment by tumor cells, can promote tumor progression. Thus, it seems likely that the net effect of PADI4 in cancer progression depends on a context-dependent balance between tumor cell–autonomous suppressive effects and stromally mediated tumor promoting effects.

Given the importance of histones as PADI4 targets (10, 13, 33), we addressed whether PADI4 might regulate stemness through global effects on the epigenome via changes in histone citrullination. Our ChIP-seq analysis revealed a strikingly widespread effect of PADI4 activity in maintaining a reduced level of H3 around TSS. Because nucleosome occupancy is typically decreased upstream of transcriptionally active genes and increased in regulatory regions of silenced genes (45, 46), we speculate that PADI4 may maintain a state of elevated genomic responsiveness to transcriptional regulation by creating a more permissive and accessible genome (47). This finding is consistent with a previous study showing PADI4 enrichment near TSSs of actively transcribed genes in MCF7 breast cancer cells (21). However, we observed relatively little impact of PADI4 inhibition on the tumor cell transcriptome in the present study. This suggests either that there is a kinetic uncoupling of the effects of PADI4 inhibition on genomic organization and downstream transcriptional responses, potentially due to a requirement for additional regulatory signals, or that biological responses to PADI4 may be elicited by the integration of many small magnitude effects on the transcriptome that are difficult to detect over the noise. These findings merit further exploration, but did not explain our observed effects on the CSCs. Instead, using a reporter-based strategy to enrich for CSCs, we showed that PADI4 regulates breast CSCs via epigenetic downregulation of the master transcription factors of stemness, NANOG and POU5F1, specifically in the CSC compartment. Because CSCs are a relatively small subpopulation, this mechanism is not evident in analyses of the bulk tumor cell population. PADI4 represses the promoters of these key stemness genes by inhibiting establishment of the activating H3R17me2a histone mark, thus suppressing transcription of NANOG and POU5F1. A similar mechanism has been described for PADI4 suppression of CTCF transcription in erythroleukemia cells (35).

In conclusion, our findings reveal important insights into PADI4 function in breast cancer progression, suggesting a novel tumor-suppressive role involving epigenetic modulation of the tumor-initiating breast CSCs. PADI4 inhibitors are in preclinical development for treatment of autoimmune disease (48). Also, given compelling data showing a tumor-promoting role for PADI4 in the tumor stroma, PADI4 inhibition has been proposed as a cancer therapy (37, 49). The relative balance between the pathologic and homeostatic effects of PADI4 will determine the potential benefit of systemic PADI4 inhibition in different disease states. Further exploration of this complex biology will help guide clinical intervention with PADI4 as a target.

No potential conflicts of interest were disclosed.

Conception and design: N. Moshkovich, L.M. Wakefield

Development of methodology: N. Moshkovich, H.J. Ochoa, B. Tang, Y. Yang

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): N. Moshkovich, H.J. Ochoa, B. Tang, Y. Yang

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): N. Moshkovich, H.J. Ochoa, B. Tang, H.H. Yang, J. Huang, M.P. Lee, L.M. Wakefield

Writing, review, and/or revision of the manuscript: N. Moshkovich, H.J. Ochoa, B. Tang, J. Huang, M.P. Lee, L.M. Wakefield

Study supervision: L.M. Wakefield

This work was supported by funding from the Intramural Research Program of the National Cancer Institute, Center for Cancer Research, NIH project ZIA BC 005785. We acknowledge the expert technical assistance of Dr. Xiaolin Wu (Laboratory of Molecular Technology, Frederick National Laboratory for Cancer Research) with the microarray analysis, and Dr. Bao Tran (Center for Cancer Research Illumina Sequencing Facility) with the NextGen Sequencing, and Dr. Huaitian Liu with data deposition in the GEO repository.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.