Histone H4 lysine 20 tri-methylation safeguards breast cancer lineage fidelity and unfolded protein response

Abstract

Triple-negative breast cancer (TNBC) has poor outcomes and few effective treatments, partly because of epigenetic heterogeneity. We show that the heterochromatin mark H4K20me3 is more variable than H3K9me3 across normal and malignant breast tissues, and that H4K20me3-enriched regions are linked to more open chromatin and active transcription. Pharmacologic inhibition of the H4K20me3 methyltransferases KMT5B and KMT5C with A-196 decreases H4K20me3 and increases H4K20me1 at loci involved in lineage identity, stress responses, and the cell cycle. A-196 suppresses tumor growth in vivo while promoting mesenchymal features and the unfolded protein response (UPR). Multi-omic profiling shows that mesenchymal TNBCs have the lowest H4K20me3 levels and the greatest sensitivity to UPR stress. In basal TNBC, KMT5B&5 C inhibition drives mesenchymal transition and elevated endoplasmic reticulum stress–associated growth arrest via E2F4. Combining A-196 with the HSP90, IRAK4 inhibitors, or STING agonist synergistically reduces TNBC tumor growth. These findings nominate KMT5B&5C as promising therapeutic targets in TNBC.

Triple-negative breast cancer (TNBC) is highly heterogeneous, complicating treatment. Here, the authors show that the H4K20me3 heterochromatin mark regulates lineage identity and the unfolded protein response in TNBC and suggest KMT5B and KMT5C as therapeutic targets.

Introduction

Chromatin is a dynamic scaffold that responds to external cues to regulate the various functions of DNA^1,2. Histone modifications are crucial in shaping chromatin states during normal development and in many diseases, including cancer^3–5. Although histone H3 modifications and their functions in cancer are well-characterized⁶, the functional relevance of histone H4 modifications in tumorigenesis remains less well understood.

Heterochromatin, a tightly packed form of DNA, is generally transcriptionally inert and plays a vital role in maintaining genomic integrity^7,8. These regions are generally gene-poor, highly enriched for repetitive elements⁹, and frequently found at telomeric and pericentromeric loci. Heterochromatic regions are enriched for trimethylated histone H3 lysine 9 (H3K9me3) and H4 lysine 20 (H4K20me3) marks^10,11. H3K9me3 is best known for its role in establishing and maintaining constitutive heterochromatin, which is essential for cellular identity^12,13. H4K20me3 has been shown to safeguard embryonic stem cell self-renewal¹⁴ and is recognized by DNMT1 to reinforce LINE-1 DNA methylation¹⁵. Recent comparative ChIP-seq profiling of H4K20me3 and H3K9me3 in the NCI-60 cancer cell line panel revealed distinct genomic binding patterns, although the uniquely marked regions were not characterized in detail¹⁶.

Breast cancer is the most common cancer in women worldwide¹⁷. Triple-negative breast cancer (TNBC), defined by the absence of estrogen receptor, progesterone receptor, and HER2 expression (ER-PR-HER2-), is often associated with therapeutic resistance and a high risk of distant metastasis, resulting in poor patient outcomes¹⁸. We recently described extensive epigenetic heterogeneity in TNBC¹⁹, and identified three major transcriptional and epigenetic subtypes: luminal, basal, and mesenchymal. Notably H4K20me3 emerged as the most variable histone mark in TNBC cell lines, and high heterogeneity was also observed in clinical TNBC samples. However, the functional relevance of H4K20me3 and the basis for its high variability in TNBC remain unclear.

Here, we present an integrated multi-omic and functional characterization of histone H4K20me3 in TNBC. Our findings reveal a distinct role for H4K20me3 in marking facultative-like heterochromatin and in repressing chromatin regions associated with mesenchymal cell states and the endoplasmic reticulum stress response.

Results

TNBC histone H4 modification landscape

To explore TNBC epigenetic heterogeneity, we carried out in-depth analysis of our previously published quantitative histone mass spectrometry data¹⁹. Principal component analysis revealed that histone marks do not segregate based on transcription subtypes (basal, luminal and mesenchymal) and histone modification variation is primarily driven by known histone marks associated with active (e.g., H3K27ac peptides) or repressive (e.g., H3K27me3 peptides) chromatin states (Fig. 1a and Supplementary Fig. 1a). H4K20me3 was the most variable modification across all cell lines and contributed to the primary histone variation confirming our prior finding¹⁹. Further analysis of H4K20me3 immunofluorescence data from TNBC clinical samples revealed that the variability in the percentage of H4K20me3-positive cells, quantified as the standard deviation across multiple representative regions, is significantly higher in early-stage TNBC (stages I–II) compared with advanced-stage disease (stages III-IV) (Fig. 1b, c), despite no differences in overall positive cell fraction (Supplementary Fig. 1b). To investigate putative mechanisms underlying H4K20me3 variability, we analyzed the expression of histone modifying enzymes in TNBC cell lines. We found that H4K20me3 levels positively correlated with its known methyltransferase KMT5C and with KMT2C (a H3K4 methyltransferase) and KDM5B (a H3K4 demethylase) (Supplementary Fig. 1c).

Fig. 1. Histone H4 modification landscape in TNBC.

a Pearson correlation (two-sided) R-values between histone peptides and first principal component versus peptide abundance variability across 34 TNBC cell lines. b Representative images of H4K20me3 immunofluorescence from two regions of TNBC tumors, representing 32 stage I and 11 stage III/IV tumors. Scale bar 50 μ. c Standard deviation of H4K20me3 intensities derived from three regions per tumor (n = 32 Stage I, n = 42 Stage II and n = 11 Stage III/IV tumors). Two-sided Mann–Whitney U test. d Pearson correlation (two-sided) R-values between H4 (n = 16) and all histone peptides (n = 57) across 34 TNBC cell lines. H4 acetylation peptides negatively correlated with H4K20me3 are highlighted. e Euclidean distances of ChIP-seq signal intensity between the indicated histone H4 modifications. f Average signal intensity of ChIP-seq from n = 2 biological replicates in each of the two cell lines in the indicated genomic regions (MDA-MB-468: n = 34,899, 76,296, 79,594, 113,109 regions for ATAC-seq, H3K27ac, H3K27me3 and H3K9me3; MDA-MB-231: n = 20,596, 81,509, 8492, 39,577 regions for ATAC-seq, H3K27ac, H3K27me3, and H3K9me3). Two-sided Mann–Whitney U test was used to compare each group to H4K20me3. g Genomic track view of indicated marks at the BICD2 gene locus. h Representative images of HA immunofluorescence in the indicated groups. Scale bar 100 μ. Experiment was performed once. i Immunoblot analysis in the indicated models of MDA-MB-231 cells. Top: HA in whole cell lysate; Middle: histone H4 modifications in HA immunoprecipitants and total H4 in 10% of the input. Bottom: HA in nucleosome extract. Tubulin and Coomassie blue staining were used as loading control. HA band intensity was quantified. The experiment was independently repeated three times with similar results. j Heatmap of log2-fold changes in ChIP-seq signal for H4 derivatives relative to WT across genomic bins differing between WT and H4K20A mutant (n = 39 gained and n = 228 lost bins). k Representative genomic regions with lost or gained integration in the H4K20A versus WT. l Genomic feature distribution of unchanged (NC), gained and lost bins between the H4K20A and WT. m Scatter plot of predicted motif frequency versus enrichment significance in regions with lost signal in the H4K20A mutant (n = 110 motifs); significant motifs (n = 43) with E-values < 10⁻¹⁵ are highlighted, with top hits labeled. n Aggregation plot illustrating HA ChIP-seq signal intensities for WT and mutant H4 derivatives at SMARCA5 binding sites in the MDA-MB-231 cell line. Box plots in (c, f) span the maxima, median and minima Whiskers extend a maximum of 1.5× interquartile range. Source data are provided as a Source Data file.

Given the significant variation observed in H4 modifications and their relatively understudied role in TNBC, we explored their epistatic relationships with other histone modifications. Across 34 TNBC cell lines, histone H4 peptides acetylated at residues K5, K8, and K12 negatively correlated with H4K20me3, while H4K16ac positively correlated with H4K20me3 (Fig. 1d). We also observed a strong negative correlation between H4K20me3 and H4K20me1, consistent with the opposing mono- and tri-methylation at the same residue (Fig. 1d). Correlations between H3 and H4 modifications were generally weaker than intra-tail counterparts. However, many active H3 acetylation marks (e.g., H3K9ac, H3K18ac, and H3K27ac) positively correlated with H4 acetylation at residues K5, K8 and K12, suggesting cooperative roles for these marks in modulating active chromatin states across cell lines (Fig. 1d). Additionally, these correlation patterns were conserved across different TNBC subtypes, with mesenchymal TNBC exhibiting the strongest anti-correlation between H4K20me3 and H4 acetylation at residues K5, K8, and K12 (Supplementary Fig. 1d).

To comprehensively investigate the interplay between H4 modifications at the chromatin occupancy level, we conducted ChIP-seq analyses for all major H4 modifications (i.e., H4K5/K8/K12/K16ac, H4K20me1, and H4K20me3) in MDA-MB-231 and MDA-MB-468 TNBC cell lines. Given the widespread distribution of H4 signals, we quantified their intensity in 500-bp genomic bins, normalized to input. Principal component analysis distinctly separated histone H4 acetylation from H4K20 methylation signal (Supplementary Fig. 1e). Euclidean distance analysis further underscored the high similarity between H4K5ac and H4K8ac, which were notably distinct from H4K20 methylation signal (Fig. 1e). Mapping H4 modification signal onto genomic regions in conjunction with other epigenetic features demonstrated a significant reduction of the H4K20me3 signal in open chromatin regions, such as ATAC and H3K27ac peaks, while showing enrichment in repressive chromatin regions, including H3K27me3 and H3K9me3 peaks (Fig. 1f,g). Notably, these genome-wide associations of H4K20me3 with other H3 modifications within the same cell line were not reflected in the overall abundance correlations across 34 TNBC cell lines (Fig. 1d). These observations suggest that H4K20me3 may interact with other epigenetic marks to establish distinct chromatin states, even though their overall abundance is independently regulated across different cell lines.

To elucidate the functional roles of H4 modifications in shaping chromatin states, we generated MDA-MB-231 cell lines stably expressing HA-tagged wild-type (WT) H4 or H4 variants with lysine-to-alanine point mutations at residues 5, 8, 12, 16, or 20. Immunofluorescence staining and immunoblot analysis of HA confirmed equivalent level of ectopic H4 variants expression across all models, with nuclear-restricted localization (Fig. 1h, i). The H4 modifications were absent in the HA immunoprecipitants of their corresponding mutants, although they still interacted with the modified endogenous H4 (Fig. 1i). Additionally, none of the mutants substantially affected modifications at neighboring lysine sites (Fig. 1i). The H4K20A mutant exhibited reduced level in nucleosome extracts compared to other variants, suggesting limited integration into the chromatin (Fig. 1i). This was corroborated by HA ChIP-seq profiling, which showed significant losses at 228 genomic bins exclusively in the H4K20A mutant-expressing cells, exemplified by the MIR513A2 locus (Fig. 1j, k). In contrast, gains were observed in 39 bins shared among H4K20A, H4K8A, and H4K12A mutants, including the MIR8078 locus (Fig. 1j, k). Focused analysis of the H4K20A-specific lost bins revealed that nearly all were located at distal intergenic regions (Fig. 1l) and largely overlapped (48.4%) with H3K9me3-marked heterochromatic regions (Supplementary Fig. 1f).

Motif enrichment analysis of these regions identified motifs for transcription factors associated with development (e.g., MNX1) and inflammatory responses (e.g., STAT5A), Zinc-finger proteins (ZNFs), and ISWI chromatin remodeler ATPases SMARCA1 and SMARCA5 (Fig. 1m). To validate this, we performed SMARCA5 ChIP-seq in MDA-MB-231 cells and projected HA-ChIP-seq signals from H4 WT and mutants onto SMARCA5 peaks. This analysis confirmed enrichment of H4 integration and a specific loss of H4K20A signal at SMARCA5 binding sites, in contrast to equivalent random genomic regions. (Fig. 1n and Supplementary Fig. 1g).

Divergent roles for H4K20me3 and H3K9me3 in the heterochromatin

H4K20me3 and H3K9me3 are thought to have similar and often redundant roles in heterochromatin formation^7,10 yet H3K9me3 was not variable in TNBC cell lines (Fig. 1a). To explore the relationship between H3K9me3 and H4K20me3 in clinical samples, we performed multicolor immunofluorescence in 10 normal breast tissues and 27 breast tumors (Fig. 2a and Supplementary Table 1). Quantification of immunofluorescence at the single-cell level demonstrated reduced fractions of H3K9me3 and H4K20me3 positive cells in breast tumors compared with normal breast tissue, regardless of tumor subtype (Supplementary Fig. 2a). The expression of the KMT5B H4K20me3 methyltransferase was lower in breast tumors in the TCGA cohort whereas KMT5C was elevated (Supplementary Fig. 2b) implying distinct roles for these enzymes in tumorigenesis. Cells were further categorized into four groups based on the median intensity of H4K20me3 and H3K9me3 signals, revealing a significant reduction in double-high cells and an increase in double-low cells in breast tumors compared to normal breast tissue (Fig. 2b and Supplementary Fig. 2c). Notably, H4K20me3^highH3K9me3^low epithelial cells were more prevalent in breast tumors than in normal breast tissues, suggesting a specific role for H4K20me3 in tumorigenesis. Diversity metrics, as inferred by Shannon’s equitability index, consistently showed significantly greater variation in H4K20me3 compared to H3K9me3 in both normal and tumor tissues (Fig. 2c and Supplementary Fig. 2d). Within the tumor cohort, double-high cells were less frequent in TNBC (Supplementary Fig. 2e), older (>45 years old) patients (Supplementary Fig. 2f), and in tumors positive for p53 by immunohistochemistry (Supplementary Fig. 2g), highlighting the potential functional significance of heterochromatin alterations in breast tumorigenesis and aging.

Fig. 2. Divergency of H4K20me3 and H3K9me3 in heterochromatin.

a Representative images of multicolor immunofluorescence for H3K9me3 and H4K20me3 in a TNBC and a normal breast tissue sample (out of 15 TNBC and 10 normal breast samples analyzed). DAPI was used to stain nuclei. Scale bar 100 μ and 50 μ (insets). b Bar plot depicting the fractions of epithelial cell subgroups categorized based on H4K20me3 and H3K9me3 signal in normal breast (n = 10) and breast cancer (n = 27) tissues. Statistical significance was determined using a two-sided Mann–Whitney U test. c Line plot comparing the matched Shannon’s diversity index for H3K9me3 (green) and H4K20me3 (red) signals from the same tissue across normal breast (n = 10), and HR + /HER2- (n = 5), HER2+ (n = 7), and TNBC (n = 15) tumor tissues. Two-sided Wilcoxon matched-pairs signed-rank test was used. d Scatter plot showing the differentially enriched 1 kb-sized genomic bins (FDR < 0.1) in three TNBC cell lines with distribution of their mean normalized counts. Differentially enriched bins are labeled, with bin counts indicated. e Box plots depicting average normalized intensity for ChIP-seq or ATAC-seq signal from n = 2 biological replicates in each of the three cell lines between H4K20me3-enriched (n = 67,438, 117 and 263 for SUM149, MDA-MB-231 and MDA-MB-468) and H3K9me3-enriched genomic bins (n = 66,522, 685 and 352 for SUM149, MDA-MB-231 and MDA-MB-468) and normalized mRNA counts for genes whose transcriptional start sites are within 100 kb of these bin sets (n = 7847 and 13,119 for SUM149; n = 161 and 88 for MDA-MB-231; n = 630 and 3508 for MDA-MB-468) in SUM149, MDA-MB-231 and MDA-MB-468 cell lines represented in each row. Two-sided Mann–Whitney U test was used. f Bar plot comparing gene density between H4K20me3-enriched and H3K9me3-enriched bins across three cell lines. g Stacked bar plots showing the genomic feature distribution for each bin set. h Box plot comparing average normalized ATAC-seq counts from 17 TNBC in the TCGA dataset between matched H3K9me3-enriched (n = 2123) and H4K20me3-enriched (n = 29,685) bins merged from the three cell lines. Two-sided Mann–Whitney U test was used. Box plots in (e, h) span the maxima (upper limit), median (center) and minima (lower limit). Whiskers extend a maximum of 1.5× interquartile range. Source data are provided as a Source Data file.

To explore relationships between H4K20me3 and H3K9me3, we conducted ChIP-seq analyses on MDA-MB-468, MDA-MB-231, and SUM149 TNBC cell lines and analyzed genomic bins marked by these two modifications. We observed a positive correlation in signal intensity between the two marks (Supplementary Fig. 2h). However, we also identified regions with significant enrichment of either H3K9me3 or H4K20me3 across all three cell lines with SUM149 exhibiting the most pronounced difference (Fig. 2d). Regions enriched for H3K9me3 consistently displayed high overall signal intensity, whereas those enriched for H4K20me3 predominantly showed moderate signal levels (Fig. 2d). To further characterize these distinct genomic features, we performed an integrated analysis with additional chromatin marks. Regions uniquely enriched for H4K20me3 demonstrated higher levels of active chromatin signals (e.g., ATAC-seq and H3K27ac) and lower levels of repressive signals (e.g., HP1γ and H3K27me3), except for H3K27me3 in MDA-MB-468 cells (Fig. 2e). Additionally, H4K20me3-enriched regions exhibited higher gene density (Fig. 2f), increased transcriptional activity (Fig. 2e), and a greater enrichment in promoters compared to H3K9me3-enriched regions (Fig. 2g). These observations were independently validated by reanalyzing previously reported paired H4K20me3 and H3K9me3 ChIP-seq data from Hs578T cells (Supplementary Fig. 2i–k). Furthermore, H4K20me3-enriched regions showed greater chromatin accessibility in TNBC tumors based on ATAC-seq from 17 TNBC tumors in the TCGA dataset²⁰ (Fig. 2h).

KMT5B and KMT5C blockade results in a shift to mesenchymal cell state

To investigate the impact of modulating H4K20me3 levels, we evaluated the KMT5B and KMT5C inhibitor A-196²¹ in three TNBC cell lines. A-196 was used at a concentration of 1 μM for in vitro experiments, a dosage previously established to achieve optimal inhibition of both KMT5B and KMT5C while still maintaining specificity²¹. Immunoblot analysis revealed that extended treatment with A-196 resulted in a consistent increase in H4K20me1 and a reduction in H4K20me3, without affecting other histone modifications tested (Fig. 3a). To further examine A-196-induced chromatin alterations, we conducted ChIP-seq for H4K20me1 and H4K20me3 in cells treated with either vehicle or A-196 for 5 days. Genomic-bin-based quantitative analysis indicated a predominant gain in H4K20me1 (89.9% of differential bins) and a loss of H4K20me3 (86.5% of differential bins) across all three cell lines (Fig. 3b). The majority of regions with reduced H4K20me3 (MDA-MB-468: 80.9%; MDA-MB-231: 95.1%; SUM149: 92.5%) overlapped with regions that exhibited increased H4K20me1 (Fig. 3c, d). A-196 treatment consistently diminished H4K20me3 signal intensity in H4K20me3-enriched regions in all three cell lines while it had variable effects on H3K9me3-enriched regions (Fig. 3e and Supplementary Fig. 2l). Consensus motifs enriched in the top 2000 bins characterized by increased H4K20me1 and decreased H4K20me3 signals demonstrated associations with multiple functional themes. The most prevalent motifs were those for transcription factors (TFs) regulating lineage and development, including key determinants of breast epithelial cell lineages, such as GATA3 (luminal), KLF5 (basal), and ZEB1 (mesenchymal) (Fig. 3f). This was followed by numerous TFs mediating interferon signaling (e.g., STAT1, IRF1)^22,23 and stress response (e.g., XBP1, FOXO1)^24,25, as well as cell cycle regulation (e.g., E2F4/6/7) and zinc finger proteins commonly linked to heterochromatin²⁶. As an orthogonal approach, we utilized predefined broad peaks derived from H4K20me3 ChIP-seq data and conducted through intersection analysis to identify peaks as downregulated or unchanged following A-196 treatment (Supplementary Fig. 2m). We identified 96 motifs uniquely enriched in A-196-downregulated H4K20me3 regions, with 85.4% (82/96) overlapping with motifs identified using the genomic-binning approach, (Supplementary Fig. 2n and Supplementary Table 2), thereby reinforcing confidence in the functional associations identified.

Fig. 3. KMT5B and KMT5C inhibition confers mesenchymal lineage shift.

a Immunoblot analysis of the indicated histone marks in three cell lines −/+ 1 μM A-196 for 5 days. Tubulin was used as loading control. The experiment was repeated three times with different batches of cells (i.e., biological triplicates) with similar results. b Numbers of significantly up- or downregulated H4K20me3 and H4K20me1 genomic bins (1 kb) by A-196 treatment (Padj <0.05) defined by two-sided Benjamini–Hochberg-adjusted moderated t test. c Venn diagram illustrating the overlap of A-196-induced significantly increased H4K20me1 bins and decreased H4K20me3 bins. d Genomic track view of the TMEM246 locus in MDA-MB-468 cell line. e Average H4K20me3 signal intensities −/+ A-196 treatment from n = 2 biological replicates in each of the three cell lines at H4K20me3-enriched genomic bins (n = 4230, 117 and 67,441 bins from MDA-MB-468, MDA-MB-231 and SUM149 cells). Two-sided Mann–Whitney U test. f Motifs enriched (E value < 10⁻⁵) in the top 2000 bins with increased H4K20me1 and decreased H4K20me3 after A-196 treatment across three cell lines. Motifs are grouped by function and ranked within each category; key motifs are highlighted in red. g, h Box plots showing g scaled H4K20me3-to-H4K20me1 ratios and h KMT5B and KMT5C expression across 34 TNBC cell lines classified into three subtypes (basal, n = 17; luminal n = 6; mesenchymal n = 11). Two-sided Mann–Whitney U test. i Pearson correlations (two-sided) between the average expression of KMT5B and KMT5C with mesenchymal signature enrichment scores among TNBC samples in the TCGA (n = 115) and METABRIC (n = 317) cohorts. j GSEA enrichment plot of mesenchymal and basal signatures in SUM149 cells −/+ A-196. Permutation test was used. k Heatmap showing log2-fold changes in mesenchymal and basal marker gene expression in SUM149 cells −/+ 1 μM A-196 for 5 days (n = 2 biological replicates). l Heatmap illustrating gained (n = 37,581) and lost ATAC-seq peaks (n = 7563) in SUM149 cells −/+ 1 μM A-196 for 5 days at a range of ± 2 kb of peak center. Analysis was performed with n = 2 biological replicates. m Normalized H4K20me3 and H4K20me1 read coverage from vehicle or 1 μM A-196 treated cells at gained ATAC-seq peaks by A-196 treatment in SUM149 cells (n = 37,581). Two-sided Mann–Whitney U test. n Aggregation plot showing ATAC-seq signal intensities at ±2 kb around TSS regions of mesenchymal genes in SUM149 cells −/+ 1 μM A-196 (n = 491 peaks). Box plots in e, g, h, m span the maxima, median and minima. Whiskers extend a maximum of 1.5× interquartile range. Source data are provided as a Source Data file.

The presence of lineage-specific transcription factor motifs within A-196-sensitive genomic regions, coupled with the high variability of H4K20me3 across TNBC cell lines, suggests that H4K20me3 may function as a regulator of cell state. We therefore investigated the correlation between H4K20me3 levels and TNBC subtypes. We found that both H4K20me3 levels and the H4K20me3/H4K20me1 ratio were lowest in mesenchymal TNBC and highest in luminal TNBC, with the opposite pattern observed for H4K20me1 (Fig. 3g and Supplementary Fig. 3a). These associations were independent of the H3K9me3 heterochromatin mark (Supplementary Fig. 3b). The H4K20me3 pattern notably paralleled the expression of its methyltransferase KMT5C (Fig. 3h). Transient knockdown of either KMT5B or KMT5C resulted in an increase in H4K20me1 and a corresponding decrease in H4K20me3 with dual knockdown of both genes producing a more pronounced shift implying overlapping functions (Supplementary Fig. 3c-e). Therefore, while both enzymes are required for the maintenance of H4K20me3, KMT5C might have higher regulatory flexibility. Notably, the average expression levels of KMT5B and KMT5C combined and the H4K20me1 methyltransferase SETD8 exhibited consistent and significant inverse correlation with mesenchymal signature enrichment scores in both the METABRIC²⁷ (n = 317) and TCGA²⁸ (n = 115) TNBC cohorts, whereas individually only KMT5B showed a significant inverse correlation in the METABRIC cohort (Fig. 3i and Supplementary Fig. 3f).

To explore the transcriptional effects of reduced H4K20me3, we conducted RNA-seq on SUM149 basal TNBC cell line following a 5-day treatment with A-196. This analysis revealed altered signaling pathways, including enhanced immune response pathways such complement and IFNγ signaling, alongside EMT features, and a reduction in TNFα signaling via NFκB and hypoxia pathways post-A-196 treatment (Supplementary Fig. 3g, h and Supplementary Data 1). Examination of TNBC subtype-specific gene signatures indicated a shift from basal to a mesenchymal state (Fig. 3j, Supplementary Fig. 3i and Supplementary Data 2), characterized by diminished expression of basal cytokeratins (e.g., KRT5, KRT17) and elevated expression of mesenchymal transcription factors (e.g., PRRX1, ZEB1) and their downstream targets (e.g., VCAN, HAS2) (Fig. 3k). The enrichment of mesenchymal features was not simply due to an increase in the expression of inflammatory cytokines present within the signatures (Supplementary Fig. 3j). Immunoblot analysis supported these findings, demonstrating reduced expression of luminal epithelial markers such as E-cadherin, claudin 1, as well as basal lineage markers CK5 and CK17, while showing increased expression of mesenchymal markers such as fibronectin, vimentin, hyaluronan synthase 2, tansgelin and the mesenchymal TNBC master transcriptional factor PRRX1¹⁹, and no change in luminal cytokeratin CK8 (Supplementary Fig. 3k). We also found that A-196 treatment induces the expression of vimentin and fibronectin while reducing the expression of luminal markers such as FOXA1 and TFF3 in the SUM185 and SUM229 luminal TNBC cell lines (Supplementary Fig. 3l) suggesting that H4K20me3 plays a role in maintaining lineage fidelity.

ATAC-seq analysis using the same treatment conditions revealed that inhibition of KMT5B and KMT5C led to remarkable chromatin opening at 37,581 sites, with a smaller subset of 7563 weakly opened regions becoming closed (Fig. 3l and Supplementary Data 1). We observed an increase in H4K20me1 and a decrease in H4K20me3 signal at these newly accessible ATAC sites (Fig. 3m), but not at unchanged or closed ATAC sites (Supplementary Fig. 3m), confirming a link between changes in H4K20 methylation and chromatin accessibility. Although A-196 treatment did not induce genome-wide changes in chromatin accessibility at promoter regions (Supplementary Fig. 3n, top panel), it markedly increased accessibility at transcription start sites (TSS) of mesenchymal marker genes such as ZEB1 (Fig. 3n and Supplementary Fig. 3o). We also noted a slight increase in chromatin accessibility at basal gene loci, suggesting that the reduction of basal genes expression is unlikely to be a direct result of chromatin-level changes (Supplementary Fig. 3n, bottom panel). Additionally, there was a notable increase in chromatin accessibility at ~7% of repetitive elements (Supplementary Fig. 3p), which may explain the activation of innate immune signaling pathways through retroelement transcription²⁹.

The effects of KMT5B and KMT5C inhibition on tumor growth

We subsequently investigated the effects of KMT5B and KMT5C blockade on cell growth. At A-196 concentrations below 10 µM, no significant changes in cell proliferation were observed across 11 human and murine TNBC cell lines. Only three cell lines demonstrated a reduction in viability below 50% at 100 µM, which is likely due to off-target toxicity (Supplementary Fig. 3q). Additionally, A-196 treatment did not affect sensitivity to commonly used chemotherapeutic drugs, such as doxorubicin and cisplatin, or targeted therapies like palbociclib and JQ1 (Supplementary Fig. 3r). A PRISM multiplexed cell line screen³⁰ of A-196 across 881 cancer cell lines revealed that KMT5B and KMT5C inhibition did not significantly impact the growth of most human cancer cell lines (AUC < 0.8), with exceptions including the SNU878 hepatocellular carcinoma and A704 renal cell carcinoma cell lines Supplementary Fig. 3s and Supplementary Data 3). Among breast cancer cell lines, only the BT549 mesenchymal TNBC line exhibited moderate growth suppression, and HER2+ breast cancer cell lines were relatively more sensitive to A-196 treatment than ER+ and TNBC lines (Supplementary Fig. 3t).

We further assessed the impact of A-196 on tumor growth in immunocompetent mice using the 4T1 and EMT6 mouse TNBC cell lines. Consistent with findings in human TNBC lines, A-196 treatment of these cells reduced H4K20me3 and increased H4K20me1 levels without affecting other H4 modifications or cellular proliferation (Fig. 4a and Supplementary Fig. 4a, b). However, A-196 administration in mice moderately suppressed primary tumor growth in both models (Fig. 4b and Supplementary Fig. 4c). To confirm the specificity of the observed effects to KMT5B and KMT5C, we transduced 4T1 and EMT6 cells with doxycycline-inducible shRNAs targeting mouse Kmt5b and Kmt5c and repeated the tumor growth assays. qPCR analysis confirmed efficient Kmt5c knockdown following doxycycline administration, while Kmt5b expression was below the detection limit due to its low basal expression in these models (Supplementary Fig. 4d). Despite this, shRNA-mediated knockdown specifically decreased H4K20me3 levels and concurrently increased H4K20me1 (Fig. 4c). Notably, doxycycline treatment consistently suppressed tumor growth in both the 4T1 and EMT6 shKmt5b&shKmt5c models, but not in the shLacZ controls (Fig. 4d and Supplementary Fig. 4e), thereby replicating the phenotype associated with A-196 treatment.

Fig. 4. The effects of KMT5B and KMT5C inhibition on tumor growth.

a, c Immunoblot analysis of the indicated histone modifications in 4T1 and EMT6 cell −/+1 μM A-196 (a) or doxycycline-inducible models treated −/+ 1 μg/ml doxycycline (Dox) for 5 days (c). Data represent three independent experiments performed with different batches of cells (i.e., biological triplicates) with similar results. b, d Tumor growth curves of indicated 4T1 and EMT6 cell models-derived tumors treated −/+ A-196 (b) or doxycyline-containing diets (d). Data are presented as mean ± SD with tumor numbers of each group labelled (n = 4, 6, 4, 5 tumors for 4T1 vehicle, 4T1 A-196, EMT6 vehicle, EMT6 A-196 groups; n = 8, 9, 10, 10 tumors for 4T1 shLacZ-Dox, shLacZ+Dox, shKmt5b&Kmt5c-Dox, shKmt5b&Kmt5c+Dox; n = 10, 9, 8, 10 tumors for EMT6 shLacZ-Dox, shLacZ+Dox, shKmt5b&Kmt5c-Dox, shKmt5b&Kmt5c+Dox). Two-way ANOVA (two-sided). e UMAP visualization of scRNA-seq data from 4T1 and EMT6 tumors −/+ A-196 treatment with n = 3 tumors collected from n = 3 mice (i.e., biological triplicates) per group, 174,390 cells total. f Dot plot showing enrichment of marker genes across the indicated cell types. The expression value is from SCTransform normalization and is unitless. g Percentage of each cell type normalized to the total number of cells in the indicated groups with n = 3 tumors collected from n = 3 mice (i.e., biological triplicates) per group (number of cells/group - 4T1 vehicle, n = 46,365; 4T1 A-196, n = 47,782; EMT6 vehicle n = 42,938; EMT6 A-196 n = 37,305). Two-sided Fisher’s exact test. h Top10 enriched pathways from the Hallmark collection (n = 50) in differentially expressed genes (FDR < 0.05, n = 83) between vehicle- and A-196-treated tumor epithelial cells, consistent across 4T1 and EMT6 models. i Correlation of A-196-induced changes of 50 Hallmark gene signatures module scores in tumor epithelial cells from 4T1 and EMT6 tumors. Consistently increased or decreased signatures are labeled in red and blue, respectively. j Immunoblot analysis of histone and UPR marks in tumor lysates from the indicated groups. Tubulin was used as loading control. Experiment was performed once due to limited tumor tissue, but each tumor is a biological replicate, thus, results reflect biological pentaplicates/quadruplicates. k Quantification of j after normalizing to tubulin control. Data are presented as mean ± SD, with n = 5 (4T1 Vehicle) or 4 (remaining groups) tumors per condition. Two-sided Mann–Whiteny U test. l Enrichment scores for seven stress signatures in cancer cells from 4T1 and EMT6 tumors with vehicle or A-196 treatment. m EMT and UPR module score Pearson correlations (two-sided) in tumor epithelial cells from scRNA-seq of 19 TNBC from two published cohorts^40,41. n Lollipop chart summarizing Pearson correlation (two-sided) R-values and −log10(P values) of EMT and UPR signature enrichment module scores of cancer cells from 19 TNBC samples shown in (m). Source data are provided as a Source Data file.

To investigate potential mediators of A-196-induced tumor growth suppression, we conducted single-cell RNA sequencing (scRNA-seq) across both models and treatment groups (n = 3 tumors per treatment per model). Major cell populations were identified using established cell type-specific markers, and their relative proportions were quantified (Fig. 4e–g and Supplementary Fig. 4f). A subtle yet significant reduction in the fraction of tumor epithelial cells and an increase in myeloid cell fractions were observed in tumors from A-196-treated mice across both models, with other cell populations remaining relatively stable (Fig. 4g). Further subclustering of cancer cells was performed for a more detailed analysis. UMAP plots of the cancer cell population, based on Louvain clustering, showed a homogenous distribution without treatment-specific subclusters (Supplementary Fig. 4g). Gene Set Enrichment Analysis (GSEA) of differentially expressed genes (DEGs) consistently altered by A-196 treatment in both models identified the unfolded protein response (UPR), MYC targets, TNFα signaling via NFκB, apoptosis, and epithelial-to-mesenchymal transition (EMT) as the most significantly enriched pathways (Fig. 4h, Supplementary Fig. 4h, and Supplementary Data 4). By employing an orthogonal approach to correlate A-196-altered Hallmark signature module scores in both 4T1 and EMT6 cancer cells, we consistently identified the upregulation of UPR and EMT pathways as top hits, alongside a reduction in reactive oxygen species (ROS) and inflammatory responses (Fig. 4i and Supplementary Data 4). Immunoblot analysis corroborated the impact of A-196 on H4K20 methylation and UPR response in tumors, as evidenced by the increased expression of multiple canonical UPR markers (Fig. 4j, k). Notably, the elevation of UPR markers ATF4 and CHOP was observed in both 4T1 and EMT6 tumor models following doxycycline-induced Kmt5b and Kmt5c knockdown (Supplementary Fig. 4i,j). To determine whether A-196 broadly influences responses to other stress in tumors, we evaluated seven widely recognized tumor-related stress gene signatures^31–37. Only UPR and heat shock signatures were consistently increased following A-196 treatment in both models (Fig. 4l). Conversely, bulk RNA-seq of 4T1 and EMT6 cells subjected to prolonged A-196 treatment in vitro showed no change in UPR or heat shock signatures (Supplementary Fig. 4k), but it reduced the enrichment levels of hypoxia, oxidative stress, fluid shear stress, and DNA replication stress signatures in the EMT6 model. A plausible explanation for the observed differences between in vivo and in vitro is that standard cell culture conditions lack sufficient basal stress to activate UPR target gene transcription. In contrast, the tumor microenvironment (TME) inherently subjects tumor cells to multiple stressors related to UPR, such as hypoxia³⁸ and interferons secreted by immune cells³⁹, making the UPR-enhancing effects of KMT5B and KMT5C blockade more pronounced in vivo. We subsequently investigated the changes in the TME associated with A-196 treatment. Cytokine profiling of tumor lysates revealed increased pro-inflammatory cytokines, such as CXCL1 and IL23, and a decrease in anti-inflammatory cytokines, including IL10 and CXCL5, in tumors from A-196-treated mice. In contrast, cell culture supernatants exhibited minimal changes, with IFN-γ being the only cytokine consistently elevated (Supplementary Fig. 4l). Flow cytometry analysis revealed no significant difference in total CD45⁺ leukocyte infiltration in tumors; however, there was an increased proportion of macrophages and a decrease in neutrophils in tumors from A-196-treated mice (Supplementary Fig. 5a–c). scRNA-seq analysis of myeloid cells further demonstrated a shift from monocytes to pro-inflammatory macrophage populations, including enrichment of Il1b^high M1-like macrophages, with Milo neighborhood analysis confirming this transition (Supplementary Fig. 5d–g). Macrophages from tumors of A-196–treated mice showed upregulation of inflammatory genes (Tnf, Cxcl1, Cxcl2) and activation of multiple inflammatory pathways (Supplementary Fig. 5h, i and Supplementary Data 4), consistent with elevated CXCL1 detected in tumor lysates. Despite the reduced abundance of neutrophils, activated neutrophils were enriched following A-196 treatment (Supplementary Fig. 5j–l), likely driven by macrophage-derived CXCL1. Additionally, A-196 induced a shift from myofibroblast cancer-associated fibroblasts (CAFs) to inflammatory CAFs in the 4T1 model (Supplementary Fig. 5m–o).

Elevated UPR in mesenchymal TNBC

Our scRNA-seq analysis revealed that UPR and EMT were the two concomitantly upregulated pathways in tumor epithelial cells by A-196 treatment. Furthermore, we observed a positive correlation between these two pathways in TNBC clinical samples, both in merged cancer cells and in individual TNBC tumors (18 out of 19), across two published scRNA-seq datasets^40,41 (Fig. 4m, n), suggesting a potential functional interplay between UPR and EMT. Thus, we investigated the possible functional link between UPR and mesenchymal features in TNBC, as well as the role of H4K20me3 in this context. Among the three major TNBC transcriptional subtypes we previously identified¹⁹, mesenchymal TNBC cell lines exhibited the highest baseline enrichment of the UPR signature compared to basal or luminal TNBC (Fig. 5a). Notably, UPR was one of the most enriched functional terms in mesenchymal TNBC, ranking just after EMT among the 50 Hallmark signatures (Supplementary Fig. 6a). At the single-cell level, scRNA-seq of individual mesenchymal TNBC cell lines, such as Hs578T, identified a subgroup of UPR^high cells that also displayed high mesenchymal characteristics, without correlation to luminal or basal traits (Supplementary Fig. 6b, c). Additionally, two mesenchymal TNBC patient-derived xenograft (PDX) models, IDC50 and T272¹⁹, showed the highest enrichment for UPR signature (Supplementary Fig. 6d).

Fig. 5. Escalated UPR in mesenchymal TNBC.

a Enrichment scores of Hallmark UPR signatures among three TNBC subtypes (basal, n = 17; luminal n = 6; mesenchymal n = 11). Two-sided Mann–Whitney U test. b Enrichment z-scores for the indicated motifs within H3K27ac peaks across 33 TNBC cell lines. Two-sided Mann–Whitney U test (mesenchymal versus non-mesenchymal cell lines). c Normalized H3K27ac signal intensities at ±10 kb of Hallmark UPR signature genes between the two groups (non-mesenchymal n = 22; mesenchymal n = 11). Two-sided Mann–Whitney U test. d Pearson correlation (two-sided) between mesenchymal scores and H3K27ac signal intensities at UPR gene loci in 33 TNBC cell lines. The linear regression line with 95% confidence interval is shown. e Left: Normalized PRRX1 signal intensities at ±10 kb of TSS for EMT (n = 200), UPR (n = 112), and random background regions (n = 200). Two-sided Mann–Whitney U test. Right: Average PRRX1 signal intensities near EMT and UPR genes compared to 20 random gene sets. Two-sided one-sample t test. f Genomic track views of PRRX1 and H3K27ac ChIP-seq in the indicated groups at STC2 and CHAC1 loci. g Enrichment scores for lineage markers in HCC3153 inducible PRRX1 models (n = 1 per condition). h Immunoblot analysis for the indicated markers in HCC3153 cells treated −/+ 1 μM tunicamycin for 12 h following −/+ doxycycline-induced PRRX1 for 5 days. Vinculin was used as a loading control. Experiment was performed three times using different batches of cells (i.e., biological triplicates) with similar results. i Enrichment scores for lineage markers in EpCAM^high and EpCAM^negative subpopulations of SUM149 paclitaxel-resistant derivatives (n = 2 biological replicates). j Immunoblot analysis of indicated markers in the same groups in i treated −/+ 1 μM tunicamycin for 12 h. Experiment was performed three times using different batches of cells (i.e., biological triplicates) with similar results. k, l Pearson correlation (two-sided) R values between histone peptides and Hallmark UPR signature enrichment in (k) 34 TNBC cell lines and (l) 101 breast tumors from CPTAC cohort⁴³. m Immunoblot analysis of the indicated markers in two basal TNBC cell lines treated −/+ 1 μM tunicamycin for 12 h following 5 days of pre-treatment −/+ 1 μM A-196. Tubulin was used as loading control. Experiment was performed three times using different batches of cells (i.e., biological triplicates) with similar results. Box plots in (a, c, e) span the maxima, median and minima. Whiskers extend a maximum of 1.5× interquartile range. Source data are provided as a Source Data file.

To investigate the potential causes of elevated UPR in mesenchymal TNBC, we first examined protein synthesis activity by analyzing key phosphorylated regulators of protein synthesis pathways across mesenchymal and other TNBC subtypes. However, we did not observe consistent differences (Supplementary Fig. 6e). Similarly, the tumor mutation burden was comparable among the three TNBC subtypes in both cell lines (n = 22, CCLE) and primary tumors (n = 102, TCGA) (Supplementary Fig. 6f), indicating that the elevated UPR was not attributable to the overall protein quantity or misfolded protein load due to mutations. Subsequently, we evaluated the enrichment of key UPR transcription factors (TFs) in TNBC cell lines based on our H3K27ac ChIP-seq data¹⁹. We discovered that the motifs for all four major UPR TFs (ATF4, ATF6, CHOP, and XBP1s) were significantly more enriched in H3K27ac peaks in mesenchymal cell lines compared to non-mesenchymal ones (Fig. 5b and Supplementary Fig. 6g). This enrichment was not due to a global increase in H3K27ac signal, as mesenchymal cells did not exhibit a higher total number of H3K27ac peaks (Supplementary Fig. 6h), and general transcription factors such as SP1 did not display similar enrichment patterns (Fig. 5b). Additionally, mesenchymal TNBC cells showed higher H3K27ac signal intensity at the TSS of canonical UPR genes (Fig. 5c), and both the H3K27ac signal of UPR genes and the enrichment of UPR TF motifs in H3K27ac-marked active chromatin were positively correlated with the degree of “mesenchymal-ness” (Fig. 5d and Supplementary Fig. 6i).

To investigate whether mesenchymal TNBC-driving transcription factors, such as PRRX1¹⁹, directly regulate UPR genes, we analyzed our PRRX1 ChIP-seq data from three mesenchymal TNBC cell lines: Hs578T, MDA-MB-157, and MDA-MB-436. Our analysis revealed significant PRRX1 genomic binding at UPR gene loci across all three cell lines, comparable to its binding at mesenchymal gene loci (Fig. 5e and Supplementary Fig. 6j). Building on our previous study, which demonstrated that ectopic doxycycline-induced PRRX1 overexpression in HCC3153 basal TNBC cells could initiate a mesenchymal program¹⁹, we further explored this model. RNA-seq lineage signature analysis confirmed an increase in mesenchymal features and a decrease in luminal and basal characteristics in HCC3153 cells upon PRRX1 induction (Fig. 5g). We also observed an elevated H3K27ac signal at PRRX1 binding sites (Supplementary Fig. 6k), including several PRRX1-associated UPR gene loci such as STC2 and CHAC1 (Fig. 5f). To test if PRRX1-induced mesenchymal shift grants UPR sensitivity, we treated the HCC3153 basal TNBC cell line with tunicamycin, with or without doxycycline to induce PRRX1 expression. PRRX1 overexpression substantially enhanced the expression of ATF4 and CHOP induced by tunicamycin treatment (Fig. 5h), along with an increase in H3K27ac, indicating a more active chromatin state. As an alternative model, we examined EpCAM^high basal and EpCAM^neg mesenchymal subpopulations in the paclitaxel-resistant SUM149 cell line derivative⁴² (Fig. 5i). Tunicamycin treatment resulted in a significantly greater increase in CHOP and XBP1s in mesenchymal EpCAM^neg cells compared to EpCAM^high cells, despite the absence of detectable differences in H3K27ac, H4K20me3, and PRRX1 levels (Fig. 5j). This suggests that broad mesenchymal characteristics, beyond the PRRX1-driven mechanism, may enhance cellular susceptibility to ER stress and UPR activation. Conversely, simultaneous knockdown of key mesenchymal transcriptional factors ZEB1 and PRRX1 in two mesenchymal TNBC cell lines, Hs578T and SUM1315, led to a diminished induction of UPR markers following tunicamycin treatment (Supplementary Fig. 6l).

KMT5 blockade enhances UPR sensitivity in basal TNBC

To elucidate the role of H4K20me3 in the mesenchymal state transition and UPR, we examined the correlation between H4K20me3 and UPR. An unbiased analysis of histone mass spectrometry data from both TNBC cell lines and the CPTAC patient cohort⁴³ consistently identified H4K20me3 as the top histone modification inversely correlated with UPR signature enrichment scores derived from matched RNA-seq data. In contrast, H4K20me1 and H4K20me0 exhibited strong positive correlations with UPR enrichment (Fig. 5k, l). This negative correlation was specific to H4K20me3 and did not indicate a general association with repressive chromatin mark, as H3K9me3 and H3K27me3 did not display similar correlations (Fig. 5k, l). Further analysis of ChIP-seq data at canonical UPR response gene loci in three cell lines revealed that the H4K20me3 signal at these loci inversely correlated with corresponding gene expression levels (Supplementary Fig. 6m), suggesting direct regulatory involvement. Consistent with these findings, immunoblot analysis demonstrated that pre-treatment with A-196, which reduced H4K20me3, led to increased induction of key UPR TFs, including CHOP, ATF4, and spliced XBP1 (XBP1s), in basal TNBC cell lines (SUM149 and HCC1806) but not in mesenchymal lines (Hs578T and SUM1315) (Fig. 5m and Supplementary Fig. 7a). Similarly, in the 4T1 mouse mammary tumor cell line, which exhibits more basal characteristics, A-196 treatment enhanced the tunicamycin-induced upregulation of UPR markers in a dose-dependent manner, whereas mesenchymal EMT6 cells did not show a differential response (Supplementary Fig. 7b). Notably, A-196 did not activate UPR in these mouse cell lines in vitro, but tumor lysates from A-196-treated mice showed elevated UPR signaling (Fig. 4j). This again suggests that the tumor microenvironment imposes stronger UPR-inducing stimuli on cancer cells compared to in vitro culture conditions. We also confirmed the enhancement of the tunicamycin-induced UPR response by A-196 in SUM149 cells by immunofluorescence analysis, which revealed a significantly increased CHOP signal intensity in the combination treatment group compared to tunicamycin alone, alongside with reduced H4K20me3 levels after A-196 treatment (Supplementary Fig. 7c, d).

To confirm that the observed effects were not due to off-target activity of A-196, we employed genetic ablation of KMT5B and KMT5C, since both enzymes are required for H4K20me3 maintenance (Supplementary Fig. S3c–e). Genetic deletion of KMT5C alone (siRNA), KMT5B alone (dox-inducible shRNA) or combined KMT5B and KMT5C knockdown (siRNA) yielded similar sensitization to UPR activation (Supplementary Fig. 7e–k). In contrast, reduction of H3K27me3 via the PRC2 inhibitor EED226, or depletion of H3K9me3 through siRNA targeting its methyltransferases (KMT1A and KMT1B), did not alter the tunicamycin response (Supplementary Fig. 7l–p).

KMT5B and KMT5C blockade strengthens ER stress-induced growth suppression via cell cycle arrest

We next investigated the phenotypic outcomes of KMT5B and KMT5C inhibition-mediated augmentation of UPR. Using our tunicamycin growth response data from 34 TNBC cell lines¹⁹, we identified a positive correlation between the H4K20me3/H4K20me1 ratio and the area under the curve (AUC) in a subset of TNBCs with high baseline UPR, as determined by RNA sequencing (Fig. 6a). This finding indicates that elevated H4K20me3 levels are associated with increased resistance to tunicamycin. Consistent with this observation, colony formation assays demonstrated that basal TNBC cells treated with A-196 showed increased sensitivity to tunicamycin-induced growth suppression (Fig. 6b, c). Moreover, the combination of A-196 and tunicamycin synergistically inhibited the growth of SUM149 xenografts compared to either treatment alone, although this was accompanied by significant toxicity primarily attributed to tunicamycin (Fig. 6d). Enhanced sensitivity to tunicamycin was also observed in SUM149 cells following dual KMT5B and KMT5C siRNA knockdown or individual KMT5C siRNA or KMT5B shRNA knockdown (Supplementary Fig. 8a–c), without significant effects on baseline cell viability (Supplementary Fig. 8d). Similarly, HCC3153 cells overexpressing PRRX1 exhibited increased sensitivity to tunicamycin suggesting a link between tunicamycin sensitivity and mesenchymal cell states (Supplementary Fig. 8e).

Fig. 6. KMT5B and KMT5C blockade augments ER stress-induced growth suppression.

a Pearson correlation (two-sided) between the scaled H4K20me3-to-H4K20me1 ratio and tunicamycin response area under the curve in TNBC cells separated by the median UPR enrichment score (n = 17 for UPR high and low subgroups). The linear regression line with 95% confidence interval is shown. b Representative images of colony growth assays from n = 3 independent experiments using the indicated doses of tunicamycin for 12 days following 5 days of pre-treatment −/+ 1 μM A-196. c Representative cell growth ratios from (b) normalized to vehicle controls, presented as mean ± SD (n = 3 technical replicates) from n = 3 independent repeats (i.e., biological triplicates) showing similar results. The IC₅₀ values and their standard deviations were calculated by combining the results from three independent experiments. d SUM149 xenograft tumor growth curves under the indicated treatments, presented as mean ± SD with tumor number per group labelled (n = 10 for vehicle and A-196, n = 8 for Tunicamycin, n = 9 for combination). Pairwise two-way ANOVA (two-sided). e Enrichment scores of the indicated UPR signatures normalized to vehicle controls −/+ A-196 pre-treatment. f Intersection of DEGs induced by tunicamycin, with and without A-196 pre-treatment. g BETA association between A-196-induced open chromatin and A-196-unique tunicamycin-induced DEGs. One-sided Kolmogorov–Smirnov test. h Significantly enriched (FDR < 0.05) Hallmark signatures in the A-196-unique tunicamycin-induced DEGs (n = 73 up genes; n = 98 down genes). i Immunoblot analysis of the cell cycle markers under the specified conditions with n = 3 independent repeats with similar results. Tubulin was used as loading control. j Representative scatter plots (from n = 3 independent experiments) of GFP–Geminin versus mCherry–Cdt1 intensities with cell-cycle phase gating in SUM149 Fucci cells treated with 48-h tunicamycin following 5-day pre-treatment −/+ 1 μM A-196. k Fractions in each cell cycle phase from (j), presented as mean ± SEM merged from three independent experiments (i.e., biological triplicates). Two-sided ordinary one-way ANOVA. l Enrichment significance of regulators (n = 1316) in shared and A-196-unique tunicamycin-induced DEGs derived from n = 2 biological replicates; red dots mark regulators whose motifs enriched in A-196-induced open chromatin. One-sided in silico deletion, regression-based regulatory potential model from LISA. m BETA association of ATF4 and E2F4 binding sites in A-196 + tunicamycin group with A-196-unique tunicamycin-induced DEGs. One-sided Kolmogorov–Smirnov test. n Signal intensity of indicated groups within ±2 kb of TSS regions for A-196-unique tunicamycin-induced DEGs. o Genomic track view of indicated signals at the RFC3 locus. Source data are provided as a Source Data file.

To further explore the functional changes induced by tunicamycin following the inhibition of KMT5B and KMT5C, we conducted transcriptomic profiling of SUM149 cells treated with tunicamycin, both with and without A-196 pre-treatment. Our findings revealed that cells pre-treated with A-196 displayed amplified tunicamycin-induced ER stress signatures (Fig. 6e), including the upregulation of canonical ER stress genes such as HERPUD1 and HSPA5 (Supplementary Fig. 8f). Differential gene expression analysis revealed an increase in both tunicamycin-induced and repressed genes in the presence of A-196 (Fig. 6f, Supplementary Fig. 8g, and Supplementary Data 1). Genes uniquely induced by tunicamycin in the presence of A-196 were significantly associated with A-196-induced open chromatin regions, as demonstrated by Binding and Expression Target Analysis (BETA)⁴⁴ (Fig. 6g). This suggests that the additional effects triggered by tunicamycin were primarily due to enhanced chromatin accessibility following the removal of H4K20me3-marked heterochromatin.

Subsequently, we performed gene set enrichment analysis (GSEA) on the A-196-unique tunicamycin response genes (referred to as the A-196-unique program). This analysis revealed a strong enrichment of cell cycle-related signatures, particularly E2F targets and the G2/M checkpoint, within the downregulated gene set. Conversely, the upregulated genes were enriched in pathways such as the p53 pathway, unfolded protein response (UPR), and TNFα/NFκB signaling (Fig. 6h). In contrast, canonical tunicamycin-induced signatures, including UPR, p53 pathway, and apoptosis, were significantly enriched among DEGs common to tunicamycin treatment regardless of A-196 exposure (Supplementary Fig. 8h). We hypothesized that the excessive suppression of cell cycle genes contributed to the observed growth suppression phenotype. Detailed analysis of genes showing increased inhibition under tunicamycin exposure revealed their roles in DNA replication, including critical components such as DNA helicase DNA2 and replication factor RFC3 (Supplementary Fig. 8i). A targeted examination of cell cycle phase-specific signatures indicated that tunicamycin’s additional inhibitory effects were most significant in G1/S and S-phase-related gene sets (Supplementary Fig. 8j). Immunoblot analysis demonstrated reduced levels of cell cycle proteins including cyclins, BACH1, ESCO2, RFC3, and phospho-RB (Ser780) alongside elevated expression of the p21 CDK inhibitor in the A-196-treated and even more pronouncedly in the combination treatment groups (Fig. 6i). Using the SUM149 Fucci (Fluorescent Ubiquitination-based Cell Cycle Indicator) model we previously established⁴⁵, we confirmed that tunicamycin treatment induced a more pronounced G1 phase arrest in the A-196-treated group (Fig. 6j, k).

We next sought to identify H4K20me3-associated suppressors of genes regulating the G1/S transition. Utilizing LISA⁴⁶, we predicted regulators of both shared and A-196-unique tunicamycin-induced DEGs and integrated these predictions with motif enrichment data from A-196-induced open chromatin. The ATF4 UPR regulator TF emerged as the top enriched driver of the shared program, while the E2F4 cell cycle regulator was identified as the primary regulator of the A-196-unique program (Fig. 6l). Additional E2F family members, E2F1 and E2F7, were also highly ranked as regulators of the A-196-unique and shared programs, respectively. Both E2F7 and E2F4, which dimerize with E2F1, were enriched in A-196-induced open chromatin regions (Supplementary Data 1) and identified as TFs associated with H4K20me3-edited genomic loci (Fig. 3f). This suggests that E2F4 may act as a de novo TF recruited upon the depletion of H4K20me3-marked repressive chromatin. To validate this hypothesis, we conducted ChIP-seq for ATF4 and E2F4 under the same treatment conditions (i.e., with or without tunicamycin and A-196 pre-treatment in SUM149 cells). ATF4 binding sites were significantly associated with the shared program in both tunicamycin-treated groups but showed no relevance to the A-196-unique program (Fig. 6m, top panel, and Supplementary Fig. 8k). In contrast, E2F4 binding sites demonstrated a strong association with repressed genes within the A-196-unique program (Fig. 6m, bottom panel). Signal projection near A-196-unique program genes confirmed that ATF4 genomic binding was irrelevant, whereas E2F4 exhibited enhanced binding exclusively in the A-196-plus-tunicamycin group (Fig. 6n). This is exemplified by the genomic track view at the RFC3 locus, which also showed increased chromatin accessibility following A-196 treatment, as revealed by ATAC-seq (Fig. 6o).

Combinational treatment strategies with KMT5B and KMT5C inhibition

Our data demonstrated a promising synergistic growth inhibitory effect with the combination of A-196 and tunicamycin in TNBC cell lines. However, the significant in vivo toxicity associated with tunicamycin, as observed by us and others, precludes its use in therapeutic applications. To identify less toxic alternatives, we evaluated the combination of A-196 with other ER stress inducers, including BiP inhibitor HA15⁴⁷ and HSP90 inhibitor 17-AAG. We observed a similar synergistic suppression of cell growth when combining 17-AAG or HA15 with 5 days of A-196 pre-treatment in SUM149 cells (Supplementary Fig. 9a–d). Immunoblot analysis confirmed enhanced induction of the ER stress markers along with a reduction in phospho-Rb (pRb) after 17AAG stimulation for 12 h in SUM149 cells pre-treated with A-196 for 5 days (Supplementary Fig. 9e). HA15 also showed similar effects of UPR response enhancement, while did not change pRb levels likely due to the limited (12 h) stimulation (Supplementary Fig. 9f). Given that 17-AAG has previously been evaluated in clinical trials and requires much lower doses to induce UPR compared to HA15, we chose to test its combination with A-196 in vivo. Notably, Hsp90b, the gene encoding HSP90 proteins, was one of the consistent upregulated genes by A-196 treatment in 4T1 and EMT6 tumor epithelial cells (Supplementary Data 4). In a SUM149 xenograft model, the combination of A-196 and 17AAG significantly reduced tumor growth compared to either treatment alone, both of which elicited moderate tumor suppression individually (Fig. 7a, b). Importantly, neither treatment caused significant weight loss in mice, indicating reasonable tolerability, at least for this short-term treatment period (Supplementary Fig. 9g).

Fig. 7. Combination therapy with KMT5B and KMT5C inhibitor.

a SUM149 xenograft tumor growth curves under the indicated treatments, presented as mean ± SD. Tumor numbers were n = 10 tumors from 5 mice (2 tumors per mouse) for the Vehicle, A-196, and 17-AAG groups, and n = 9 tumors from 5 mice for the A-196 + 17-AAG group (2 tumors per mouse except for one mouse bearing a single tumor). Two-sided pairwise two-way ANOVA. The vehicle and A-196 data were generated from the same experiment as Fig. 6d. b Individual tumor weights from (a), with data presented as mean ± SD with tumor numbers labelled. Two-sided ordinary one-way ANOVA. c Enrichment scores for indicated pathways of each tumor across the four treatment groups (n = 4 from Vehicle and A-196; n = 3 for 17-AAG and A-196 + 17-AAG). Two-sided Kruskal–Wallis test (combination versus vehicle group). d Heatmap representing the union of DEGs for each treatment group compared to the vehicle group (n = 4 tumors from Vehicle and A-196; n = 3 tumors for 17-AAG and A-196 + 17-AAG, each tumor was derived from a different animal). e Significantly enriched (FDR < 0.05) Hallmark signatures using one-sided Fisher’s exact test in Enrichr from indicated groups, as identified in (d). f Representative images of immunofluorescence for EpCAM and phospho-histone H3-Ser10 in the indicated tumors, representing n = 6 tumors per condition (i.e., biological sextuples). Scale bar 100 μ. g Quantification of the percentage of phospho-histone H3-positive epithelial cells, presented as mean ± SD (n = 6 tumors per condition). Two-sided Kruskal–Wallis test. h, i Tumor growth curves of EMT6 syngeneic model with indicated treatments. The two plots shared the identical vehicle and A-196 treatment group. Data are presented as mean ± SD with numbers labelled for each group (n = 9, 8, 9, 10, 8, 10 tumors for vehicle, A-196, ADUS100, A-196 + ADUS100, Emavusertib, A-196+Emavusertib groups respectively). Pairwise two-way ANOVA (two-sided). j Individual tumor weights from the experiment described in (h, i), with data presented as mean ± SD with tumor numbers labelled (n = 9,9,9,10,7,10 tumors for vehicle, A-196, ADUS100, A-196 + ADUS100, Emavusertib, A-196+Emavusertib groups respectively). Two-sided Kruskal–Wallis test. k Fractions of monocytes, neutrophils and dendritic cells normalized to CD45⁺ cells and MHC-II^high macrophages normalized to total macrophages across the indicated treatment groups. Data are presented as mean ± SD with numbers labelled for each group (n = 7, 7, 9, 5, 6, 8 tumors for vehicle, A-196, ADUS100, A-196 + ADUS100, Emavusertib, A-196+Emavusertib groups respectively). Two-sided Kruskal–Wallis test. l Pearson correlation (two-sided) between tumor weights and the indicated immune cell fractions as shown in (k). Treatment groups were labelled (n = 41 tumors in total). The linear regression line with 95% confidence interval is shown. m Schematic model of the proposed mechanism based on our findings. Created in BioRender. Li, Z. (2026) https://BioRender.com/8hnikm0. Source data are provided as a Source Data file.

To elucidate the downstream pathways contributing to the observed synergism, we conducted bulk RNA-seq on 3–4 tumors per treatment group. Principal component analysis demonstrated distinct segregation among the groups, with the combination treatment exhibiting a trend towards greater transcriptomic divergence from the vehicle group (Supplementary Fig. 9h). Targeted pathway enrichment analysis consistently indicated that tumors treated with the combination therapy displayed enhanced ER stress signatures and a reduction in G1 and G1/S cell cycle signatures (Fig. 7c). Furthermore, we noted a general upregulation of interferon response signatures in both the A-196 and combination groups (Fig. 7c), which was consistent with scRNA-seq profiling in mouse tumors. Analysis of differentially expressed genes revealed 500 and 608 genes uniquely up- and downregulated in the combination group, respectively, with fewer genes shared between combination and 17-AAG treatment groups (Fig. 7d and Supplementary Data 5). Gene Set Enrichment Analysis of different subgroups of DEGs highlighted the enrichment of inflammatory/immune response signatures (e.g., TNFα signaling via NFκB and inflammatory response) and EMT pathways in the upregulated genes of the combination group. Conversely, cell cycle-related signatures (e.g., G2-M checkpoint, E2F targets, and mitotic spindle) were suppressed in the combination group (Fig. 7e), consistent with our in vitro findings. Immunofluorescence staining of the G2/M marker phospho-histone H3 (Ser10) further confirmed a significant reduction in cell proliferation uniquely in the combination treatment group (Fig. 7f,g). In light of the increased inflammatory signatures observed in the tumors, we performed a human cytokine array analysis on xenograft tumor lysates, which revealed elevated levels of pro-fibrinolytic (e.g., PAI1 encoded by the SERPINE1 gene) and pro-inflammatory cytokines (e.g., ICAM1, IL33, and IL1A) following A-196 and 17AAG treatment, with the strongest induction seen in the combination-treated tumors (Supplementary Fig. 9i), corroborated by the RNA-seq data (Supplementary Fig. 9j).

Based on our observation that A-196 treatment induces a pro-inflammatory tumor microenvironment in both 4T1 and EMT6 syngeneic models, as well as in immunodeficient NSG mice, we hypothesized that inhibition of KMT5B and KMT5C could potentiate immune-based therapies. To test this hypothesis, we used the EMT6 model to assess the therapeutic potential of A-196 in combination with standard treatments commonly used for TNBC patients, such as paclitaxel and anti-PD-1 immunotherapy. Over a 12-day treatment period, the combination of A-196 and paclitaxel demonstrated superior tumor growth inhibition compared to either agent alone (Supplementary Fig. 9k, l). However, no synergistic effect was observed when A-196 was combined with anti-PD-1, or in the triple combination of A-196, anti-PD-1, and paclitaxel. Consequently, we explored the combination of A-196 with other actionable immune modulators.

STING agonism has emerged as a promising strategy to enhance anti-tumor immunity across various cancer types⁴⁸. In breast cancer, STING activation has been shown to synergize with PARP inhibitors in BRCA1-deficient models, promoting cytotoxic T cell infiltration and activating the unfolded protein response^49,50. Separately, myeloid-specific IRAK4 inhibition has recently been reported to reduce immunosuppressive macrophage populations and stimulate anti-tumor immunity by alleviating tumor-associated inflammation in hematologic malignancies and pancreatic ductal adenocarcinoma^51,52. Based on these findings, we hypothesized that the therapeutic efficacy of both STING agonists and IRAK4 inhibitors⁵³ might be enhanced within the tumor microenvironment altered by KMT5B and KMT5C inhibition. Therefore, we tested the STING agonist ADUS100 and the IRAK4 inhibitor Emavusertib in combination with A-196 in the EMT6 syngeneic breast cancer model.

During a 12-day treatment period, both combinations demonstrated significant synergy. Notably, ADU-S100 plus A-196 induced substantial tumor regression (Fig. 7h–j). To evaluate alterations in the immune landscape, we performed polychromatic flow cytometry on dissociated tumors. Although total CD45⁺ leukocyte levels remained consistent across all treatments (Supplementary Fig. 9m), both combination therapies led to an overall increase in monocytes and dendritic cells, alongside a reduction in neutrophils compared to single-agent treatments (Fig. 7k). While the total macrophage fraction did not show significant variation (Supplementary Fig. 9n), there was a marked increase in MHC-II^high macrophages in the combination groups, indicating a rise in M1-like macrophage abundance, consistent with prior studies^51,54. Importantly, total leukocytes, dendritic cells, monocytes and MHC-II^high macrophages but not total macrophages demonstrated inverse correlation with individual tumor weight, whereas neutrophils showed a positive correlation (Fig. 7l and Supplementary Fig. 9p, q), suggesting a distinct role for these myeloid cells in driving the phenotypic changes. Despite limited overall lymphocyte infiltration, the A-196 and Emavusertib combination exhibited significantly higher levels of total T cells compared to Emavusertib alone (Supplementary Fig. 9o), with total T cell fractions also negatively correlating with tumor weights (Supplementary Fig. 9r). To further confirm immune activation, we performed a cytokine array analysis on tumor lysates. The A-196 plus ADU-S100 combination demonstrated enhanced secretion of canonical cGAS–STING downstream targets, including CCL5, TNF-α, CXCL1, and CCL2, indicating activation of NF-κB and type I interferon signaling (Supplementary Fig. 9s). In the A-196 plus Emavusertib cohort, we observed a more pronounced suppression of M2-like macrophage-associated factors, including Amphiregulin, CCL17, CCL12 and CCL11, alongside an upregulation of M1-associated cytokines such as IL-6, TNF-α, and IL-23. This supports the macrophage subtype transition observed in flow cytometry (Fig. 7k).

Discussion

Effective treatment of TNBC remains a significant challenge, in part because of the limited number of therapeutic targets encoded within the euchromatin. The potential for targeting vulnerabilities embedded within the heterochromatin has been insufficiently explored, largely due to the limited availability of tractable heterochromatin regulators. In this study, we identified the heterochromatin mark H4K20me3 as the most heterogeneous histone modification in TNBC. H4K20me3 marks a subset of heterochromatic regions that display greater epigenetic and transcriptional activity than H3K9me3, a heterochromatin mark often deemed functionally redundant. In basal TNBC, reducing H4K20me3 through inhibition of the histone methyltransferases KMT5B and KMT5C induced a shift toward a mesenchymal-like cell state, creating vulnerabilities that included enhanced sensitivity to ER stress and a more immune-active tumor microenvironment. Exploiting these vulnerabilities by combining KMT5B and KMT5C inhibition with an ER stress inducer or immune modulators synergistically suppressed TNBC growth (Fig. 7m).

H3K9me3 and H4K20me3 are well-established heterochromatin marks with overlapping and often redundant roles in the formation and maintenance of constitutive heterochromatin¹⁰. Both modifications can recruit HP1 proteins and interact with 53BP1 at sites of DNA damage⁵⁵. However, our comparative analysis of H3K9me3 and H4K20me3 across three TNBC cell lines revealed notable distinctions. Globally, H3K9me3-enriched regions generally exhibited higher read counts, consistent with constitutive heterochromatin, whereas H4K20me3-enriched regions showed moderate read counts, consistent with facultative or more permissive chromatin states. Although the two marks were positive correlated, we identified specific genomic regions uniquely enriched for H4K20me3. Previous studies have suggested that H3K9me3 is required for KMT5B and KMT5C to deposit H4K20me3⁵⁶. Our findings, however, suggest the existence of H3K9me3-independent pathways for H4K20me3 deposition. Notably, genomic regions selectively enriched for H4K20me3 displayed increased chromatin accessibility and transcriptional activity in both TNBC cell lines and patient samples. These findings imply that H4K20me3 alone may be insufficient to establish stable constitutive heterochromatin, possibly due to reduced recruitment of HP1 proteins⁵⁷ or its inability to drive heterochromatin-specific liquid-liquid phase separation⁵⁸. Intriguingly, a previous study identified a subset H4K20me3-marked regions co-localizing with active chromatin marks H3K4me3 and H3K36me3 in embryonic stem cells, creating a bivalent-like chromatin states with decreased RNAPII pausing⁵⁹. Furthermore, whereas H4K20me3 depletion uniquely sensitized cells to ER stress, depletion of H3K9me3 via KMT1A&1B knockdown modestly attenuated the ER stress response. One plausible explanation is that H3K9me3 is crucial for maintaining higher-order 3D genome organization⁶⁰; thus, its disruption may preferentially impair long-range enhancer–promoter interactions at select UPR loci, thereby limiting efficient transcriptional activation independently of H4K20me3. The clinical relevance of H4K20me3 and H3K9me3 in breast cancer remains underexplored. Prior studies using single-marker staining reported reduced levels of both marks in breast tumors relative to normal mammary glands, consistent with our observations^61,62. Our multicolor immunofluorescence analyses revealed distinct patterns, with H4K20me3 showing a more heterogeneous distribution in breast cancer, particularly in early-stage TNBC. Together with our mechanistic findings, these data suggest that cells with low H4K20me3 may represent a subgroup of UPR-high mesenchymal cancer cells. This pronounced heterogeneity suggests that H4K20me3 may regulate specific genomic regions that control cellular plasticity. Our observations are consistent with previous reports in pancreatic⁶³ and colon cancers⁶⁴ that described role for H4K20me3 in regulating epithelial-mesenchymal plasticity and stem-like properties. Intriguingly, although A-196 increases chromatin accessibility at mesenchymal lineage loci in SUM149 cells, basal lineage loci remain largely unaffected. This indicates that downregulation of basal lineage genes is unlikely to result from direct chromatin closure and may instead involve alternative transcriptional mechanisms, such as recruitment of transcriptional repressors. For example, ZEB1 can suppress epithelial programs without inducing chromatin closure in breast and colorectal cancer^65,66. Further studies will be needed to define these mechanisms.

H4K20me3 formation is catalyzed by KMT5B and KMT5C, although the functional distinctions between these two enzymes remain poorly characterized. Our cell line-based omics analyses indicate that KMT5C, but not KMT5B, is significantly upregulated in mesenchymal TNBC and strongly positively correlates with H4K20me3 levels across TNBC cell lines. However, experimental data demonstrate that knockdown of either enzyme reduces H4K20me3 and is accompanied by an increase in H4K20me1. This apparent discrepancy may reflect the distinct regulatory properties of KMT5B and KMT5C. KMT5B expression is relatively stable and appears essential for maintaining basal H4K20me3 levels, whereas KMT5C expression is more variable and more tightly regulated, potentially contributing to lineage plasticity. In SUM149 cells, the increase in H4K20me3 signal following A-196 treatment, specifically at H3K9me3-enriched regions, indicates that the inhibitory effect of KMT5B and KMT5C inhibition on H4K20me3 is not global but rather contingent on local H3K9me3 enrichment. This suggests that KMT5B and KMT5C blockade may trigger a negative-feedback or compensatory mechanism that enhances H4K20me3 deposition in H3K9me3-rich regions through alternative, less characterized H4K20 methyltransferases, such as SMYD3 and SMYD5^67,68.

In this study, we utilized multi-omic profiling to examine how mesenchymal lineage features are regulated in relation to cellular sensing of unfolded protein burden. Although we did not detect significant differences in overall protein synthesis, transcriptional activity, or mutational burden as indirect indicators of misfolded protein load, these metrics may not fully capture the complexity of protein misfolding and aggregation. Further research employing protein-folding reporters, such as the Thioflavin T Binding Assay for amyloid fibrils⁶⁹, will be needed to provide deeper insights. A previous study reported enhanced ER stress sensitivity associated with EMT, largely due to increased extracellular matrix secretion⁷⁰. In contrast, our study primarily investigated chromatin-level mechanisms underlying this sensitivity. Our data revealed enhanced chromatin accessibility at UPR gene loci in mesenchymal cells. Notably, the mesenchymal TNBC master regulator PRRX1 bound UPR gene loci, and its overexpression amplified the UPR response. This effect may occur through multiple non-mutually exclusive mechanisms. PRRX1 may serve as a direct pioneer factor or recruit other epigenetic activators to the promoters of UPR genes. This possibility is supported by our observation that PRRX1 overexpression increased H3K27ac signal at PRRX1 binding sites. These findings are consistent with previous studies highlighting the role of mesenchymal transcription factors in altering the epigenetic landscape. For example, ZEB1 has been shown to interact with SETD1B to promote H3K4me3 deposition, and facilitate chromatin opening⁷¹. Additionally, PRRX1 may directly interact with UPR transcription factors, thereby enhancing their activity through a lineage-independent, non-canonical manner. Although our study primarily examined the interaction between PRRX1 and the UPR, other molecular characteristics associated with mesenchymal-like states may also play a role. For instance, previous research has indicated that the loss of cytokeratin can intensify UPR activation through the PERK/eIF2α axis⁷².

Our scRNA-seq analysis of mouse syngeneic 4T1 and EMT6 tumors demonstrated that A-196 treatment induces substantial changes in both cancer cells and the tumor microenvironment. While we focused on the relationship between the unfolded protein response and mesenchymal features, it remains unclear whether and how the tumor microenvironment contributes to the differential regulation of UPR. One potential explanation involves pro-inflammatory cytokines produced by nonmalignant cells, such as infiltrating macrophages. For instance, increased secretion of CXCL1 and CXCL2 may specifically activate neutrophils, thereby enhancing neutrophil UPR, which may be transmitted to neighboring cancer cells^73,74. In addition, another highly induced cytokine, TNFα, may amplify UPR in a non-cell-autonomous manner. TNFα has been shown to induce endoplasmic reticulum stress through a reactive oxygen species (ROS)-dependent mechanism, and the resulting UPR acts as a protective feedback mechanism that mitigates further ROS accumulation⁷⁵. Microenvironmental mechanical stress, such as that imposed by the extracellular matrix, as well as nutrient deprivation and hypoxia within the tumor core—conditions absent under standard in vitro culture conditions—may also trigger UPR^76,77. In this context, the pro-inflammatory tumor microenvironment may represent a downstream consequence, rather than a direct cause, of elevated UPR in cancer cells, as UPR is known to activate pro-inflammatory pathways^78,79. Our in vitro findings of globally derepressed repetitive elements following A-196 treatment suggest an alternative mechanism: transcription of retroelements may activate the cGAS-STING pathway^80,81, thereby enhancing innate immune responses, including the interferon signaling observed in our RNA-seq analysis. Notably, recent studies have highlighted the essential role of STING in ER stress induction in breast cancer⁴⁹, further supporting this hypothesis. The strong synergism of combining A-196 with the STING agonist ADUS100 in EMT6 syngeneic models underscores the potential therapeutic value of this signaling axis in boosting anti-tumor immunity.

Our study identified a role of histone H4K20me3 modification in TNBC, although certain limitations remain. For instance, the relationship between H4K20me3, mesenchymal lineage, and the ER stress response requires further validation at single-cell resolution across large-scale clinical specimens through spatial characterization. In addition, although short-term systemic administration of A-196 in vivo showed no signs of toxicity, its long-term effects require comprehensive evaluation to establish safety and support potential clinical translation. Despite these limitations, our study provides a deeper understanding of how H4K20me3 shapes heterochromatin in TNBC. We uncovered stress-response vulnerabilities linked to genetic information embedded within H4K20me3-regulated heterochromatin and identified therapeutic strategies that may benefit patients with TNBC.

Methods

Ethics statement

All animal studies were conducted in compliance with the relevant ethical guidelines and approved by the Dana-Farber Cancer Institute Animal Care and Use Committee (IACUC) protocol #11-023. The tissue microarray (TMA) used in this study in Fig. 2 was purchased from Novus Biologicals (No. NBP2-30212). The tissues were collected and commercialized by The SUPERBIOCHIPS Company (South Korea), which has confirmed that comprehensive informed consent, including authorization for the use of hospital information and biological materials for research purposes, was obtained from all patients prior to surgery. Patient identifiers were removed to protect privacy and confidentiality. Although the U.S. Code of Federal Regulations exempts the collection of pathological specimens from Institutional Review Board (IRB) approval requirements under certain circumstances, The SUPERBIOCHIPS Company maintains an External Review Board composed of pathologists from major hospitals and legal experts. This board oversees the ethical and legal aspects of tissue collection and use, ensuring that all procedures protect the rights and welfare of human subjects and safeguard patient privacy and confidentiality.

Cell line and model construction

SUM149, SUM185, SUM1315, and SUM229 cells were obtained from Dr. Stephen Ethier (University of Michigan). SUM149 cell line was cultured in DMEM/F12 supplemented with 5% fetal bovine serum (FBS), 10 mmol/L HEPES pH 7.4, 1 μg/mL hydrocortisone, 5 μg/mL insulin. SUM185, SUM1315 and SUM229 was cultured in 1:1 mix of DMEM/F12 10% FBS with complete HMEC medium. MDA-MB-231, MDA-MB-468, HCC1806, Hs578T, and EMT6 cell lines were obtained from ATCC. HCC3153 cells were provided by Adi Gazdar/Joe Gray (LBNL), while 4T1 were obtained from the Karmanos Cancer Institute. MDA-MB-231 and MDA-MB-468 cells were cultured in DMEM supplemented with 10% FBS. HCC1806 and HCC3153 cells was cultured in RPMI-1640 medium plus 10% FBS, 10 mmol/L HEPES pH 7.4 and 1 mM sodium pyruvate. Hs578T cells were cultured in DMEM supplemented with 10% FBS and 0.01 mg/ml human insulin. 4T1 cells was cultured DMEM supplied with 10% FBS and 2 mM L-Glutamine. EMT6 cells cultured in Waymouth’s Medium plus 10% FBS. Dox-inducible human shKMT5B, mouse shKmt5b&Kmt5c and H4 mutant expressing derivatives of cell lines were generated by infection with lentivirus. Plasmids of doxycycline-inducible shKMT5B (V3SH11252-225844350) plasmids were purchased from Horizon Discovery. Histone H4 wild-type (VB211109-1220ybt), H4K5A (VB211109-1221bvs), H4K8A (VB211109-1222ydh), H4K12A (VB211109-1223vpw), H4K16A VB211109-1224jcg) and H4K20A (VB211109-1225brb) lentiviral vectors were custom ordered from Vector Builder. These plasmids were conjugated with EGFP:TA2:Puro resistant sequence. For the 4T1 and EMT6 shLacZ models, two lentiviral plasmids (VB250908-1654mjc and VB250908-1653dnu) carrying distinct shLacZ sequences and conferring puromycin or blasticidin resistance, respectively, were co-transfected into cells. For the shKmt5b&Kmt5c models, cells were co-transfected with a shKmt5c plasmid conferring blasticidin resistance (VB250908-1652sch) and a shKmt5b plasmid conferring puromycin resistance (VB250908-1649rsv). Stable cell populations were selected using both antibiotics. All the above-mentioned lentiviral vectors used in this study were synthesized by VectorBuilder. Sequences of shRNAs are listed in Supplementary Table 3. For cell model generation, 12 μg of the corresponding lentiviral vector, 8 μg of psPAX2 (Addgene 12260) and 4 μg of pMD2.G (Addgene 12259) were mixed with 60 μl of Lipofectamine 3000 in 1 ml OptiMem. In total, 1 × 10⁷ 293FT cells were seeded and kept upright while adding the transfection mix. Flask contents were pipetted to homogenize, and cells were allowed to attach overnight at standard culture conditions. Media was then changed with 12 ml of pre-warmed fresh media. After a 48 h incubation, viral supernatant was collected, and fresh media added for two more collections after 12 h and 24 h. Viral supernatants were kept at 4 °C, combined, and spun at 500×g for 10 min at 4 °C. Supernatants were then filtered through a 0.45-μm PVDF syringe filter. Supernatants were then added to the targeted cells in 6-well plates with twice of infection in the presence of 10 μg/ml Polybrene transfection reagent. After 2 days, cells were grown in fresh media and selection was then started with 2 μg/ml puromycin or 6 μg/ml blasticidin. For H4 models, selected cells were further FACS sorted for top10% GFP+ cells to enrich for subpopulation with high ectopic copies. HCC3153 doxycycline-inducible PRRX1 overexpression model¹⁹, SUM149 paclitaxel-resistant EpCAM-high and EpCAM-negative models⁴², and SUM149 Fucci model⁴⁵ were previously described.

Reagent and compounds

A-196 (HY-100201), paclitaxel (HY-B0015), Palbociclib (HY-50767A), ADUS100 (HY-12885B) and Emavusertib (HY-135317) were purchased from MedChem Express. Tunicamycin (S7894), HA15 (S8299) and Tanespimycin (17-AAG, S1141) were obtained from SelleckChem. Doxorubicin was purchased from Sigma-Aldrich (D1515). InVivoMAb anti-mouse PD-1 (BE0146) and InVivoMAb rat IgG2a isotype control (BE0089) were purchased from BioXcell. For siRNA knockdown, oligos were purchased from Horizon Discovery (siKMT5B: L-013366-01-0005, siKMT5C L-018622-02-0005, siKMT1A: L-009604-00-0005, siKMT1B: L-008512-00-0005, siZEB1: L-006564-01-0005, siPRRX1: L-012402-00-0005, Non-targeting control: D-001810-10-20). For each target, four siRNAs were pooled (SMARTpool) and used for transfection. Sequences of each siRNA were included in Supplementary Table 3. Reverse transfection was performed on cell suspensions in a six well plate using 0.0625 nmol of siRNA per well with 5 µl Lipofectamine RNAiMAX Reagent (Life Technology 13778075). For experiments with ER stress inducer treatment, cells were first pre-treated with or without 1 μM A-196 for 5 days and then stimulated with indicated doses of tunicamycin, 17-AAG or HA15 for 12 h for immunoblot analysis, or 12–14 days for colony formation assay.

Mouse tumorigenesis assays

All animal experiments were performed following protocol #11–023 approved by DFCI IACUC. Mice were housed 5 mice/cage with ad libitum access to food and water in 20 °C ambient temperature, 40–50% humidity, and 12-h light/12-h dark cycle. Exponentially growing cells were resuspended in ice cold PBS and 50% Matrigel (BD Biosciences) and orthotopically injected into the inguinal mammary fat pads in a 50 μl volume of 6-week-old female NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) mice for SUM149 xenograft or female Balb/c mice (Jackson Laboratory, RRID:IMSR_JAX:000646) for 4T1 and EMT6 tumorigenesis assay. Bilateral mammary fat pad injection was performed except for experiment in Fig. 4b, where only one mammary gland was injected. Mice were randomized after all mice developed palpable tumors to ensure no pre-treatment tumor volume differences. Compounds were administered as below: A-196 (25 mg/kg daily oral gavage), tunicamycin (0.25 mg/kg, daily intraperitoneal injection), 17-AAG (12.5 mg/kg, 5 days/week intraperitoneal injection), ADUS100 (2 mg/kg intratumor injection twice a week), Emavusertib (50 mg/kg daily oral gavage), anti-PD-1 or IgG isotype (10 mg/kg intraperitoneal injection twice a week) and paclitaxel (10 mg/kg intraperitoneal injection twice a week). All compounds were dissolved in 90% corn oil + 10% DMSO except ADUS100 (in PBS) and anti-PD-1 or IgG isotype control (in InVivoPure pH 7.0 or 6.5 Dilution Buffer, BioXcell). For doxycycline-inducible models, mice were maintained under control or doxycycline-containing diets (625 ppm) after cell implantation. Animals were euthanized, and tumors were harvested when tumors in the any group reached maximum allowed size or reached the humane end point. Tumor volumes were calculated using the formula (length ×  width²)/2. The maximum tumor size permitted by the DFCI IACUC under this protocol was 2 cm³, and this limit was not exceeded at any point during the study. CO₂ euthanasia was conducted using a flow meter at a rate of 30-70% (average 30% of chamber volume of CO₂ delivery per minute) in a chamber as per IACUC protocol #11-023.

Immunoblot analyses

Cells were lysed using RIPA buffer supplemented with Halt protease and phosphatase inhibitor Cocktail (Life Technologies) and subjected to sonication with a cup horn at 70% amplitude for 5 min, cycling 20 s on and 10 s off. For all the A-196-related experiments, cells were pre-treated with 1 µM A-196 for 5 days followed by other perturbations. For nucleosome preparation, extraction kit from Active Motif (53504) was used following the manufacturer's instructions. Lysates mixed with loading buffer, were separated on 4%-12% Bis-Tris gels or 10% Tricine gels (Life Technologies), and transferred to PVDF membranes at 80 V for 2 h at 4 °C. Membranes were blocked with 5% milk in PBS-T (0.1% Tween 20 in PBS) for 1 h at room temperature and incubated with primary antibodies (1:2000 for tubulin antibody, 1:500 for ATF4 antibody in mouse tumors and 1:1000 for all other antibodies) in 5% milk PBS-T overnight at 4 °C. Following three washes of 5 min each, membranes were incubated with secondary antibodies (1:2000 dilution), washed twice for 5 min, and then for 15 min. Protein bands were visualized using Clarity or Clarity Max ECL western blotting substrates (Bio-Rad) and detected with the Bio-Rad ChemiDoc Imaging System, with exposure times ranging from 1 s to 10 min. Quantification of immunoblot intensity was performed using ImageJ. Detailed antibody information can be found in Supplementary Data 6.

Immunofluorescence staining and analyses

For human tissue microarray, a commercially available FFPE slide was purchased from Novus (NBP2-30212). Tissue sections from FFPE samples were deparaffinized, followed by antigen retrieval using Target Retrieval Solution (pH 6 or pH 9, Agilent) for 40 min with a steamer. After blocking with 5% BSA in TBST with 0.1% tritonX-100 for 1 h, slides were incubated overnight at 4 °C with primary antibodies. For H4K20me3 and H3K9me3 co-staining, primary antibodies against H3K9me3 (ab8898, 1:100) and H4K20me3 (ab9053, 1:100) were first pre-conjugated with Alexa Fluor 488 (Abcam 236553) and Alexa Fluor 555 (ab269820, 1:100). For co-staining of EpCAM (ab2929, 1:100) and phosphor-histone H3 (Ser10) (pHH3, ab5176, 1:100), antibodies were mixed and directly used. Primary antibodies were diluted in 5% BSA with 0.1% tritonX-100 in TBST. Secondary antibodies were used for co-staining of EpCAM and pHH3. Goat anti-mouse Alexa Fluor 488 (Invitrogen, A-11001, 1:200) and goat anti-rabbit Alexa Fluor 555 (Invitrogen, A-21428, 1:200), were applied for 2 h at room temperature in 5% BSA in TBST. Slides were then mounted using mounting medium with DAPI. Images were captured with a Nikon ECLIPSE Ti2-E fluorescence microscope. For TMA quantification, whole slides were scanned each protein intensity in epithelial cell nuclei were quantified by QuPath software. For EpCAM and pHH3 co-staining, two representative images of each tumor were taken and pHH3 positive cells were quantified using QuPath software. Both analyses were conducted in a blinded manner. For staining on cell lines, SUM149 cells were pre-treated with or without 1 µM A-196 for 5 days and seeded on a coverslip with 30-50% confluency. Cells were then stimulated with 1 µM tunicamycin for 12 h and fixed with 4% formaldehyde and blocked with 5% BSA solution plus 0.1% tritonX-100. Primary antibody against CHOP (Cell Signaling Technology, 2895, 1:100) and H4K20me3 (Abcam, ab9053, 1:100) was applied to co-stain the cells overnight. Secondary antibodies including goat anti-mouse Alexa Fluor 488 (Invitrogen, A-11001, 1:200) and goat anti-rabbit Alexa Fluor 555 (Invitrogen, A-21428, 1:200) were used following primary antibody incubation. Coverslips were mounted and images were taken using Nikon ECLIPSE Ti2-E fluorescence microscope. Signal intensity of each target protein in individual cells were quantified by QuPath software with at least 40 cells being scored from 3 different regions of each treatment.

Cell growth assay

For colony growth assay, cells were plated in 6-well plates at 5,000 cells per well in triplicates. For all the A-196-related experiments, cells were pre-treated with 1µM A-196 for 5 days followed by the indicated perturbations. For siRNA knockdown experiments, siRNA transfection was refreshed every 6 days. After 12 to 14 days of growth, cells were fixed in ice-cold methanol (Fisher Scientific) for 10 min, followed by staining with 0.5% crystal violet at room temperature for 15 min. Images were captured after washing the wells three times with ddH₂O. For quantification, crystal violet was dissolved in 300 μl of 10% SDS per well. The absorbance at OD595 of 100 μl of de-stained solution from each well was measured using a microplate reader with three technical replicates. For proliferation assay, 3000 cells were seeded into each well of a 96 well plate under desired treatment. Cell numbers were determined by FluoReporter™ Blue Fluorometric dsDNA Quantitation Kit (Thermo Fisher Scientific).

Flow cytometry

Tumor tissues were smashed and digested for 60 min using digestion media (2% w/v collagenase IV, 2% w/v hyaluronidase and 2% w/v BSA in DMEM/F12) at 37 °C in shaker for 1 h. Cell suspensions were filtered through a 500 µm mesh, washed with PBS and frozen in 10% DMSO/FBS at -80 °C. For flow cytometry analysis, digested cells were first stained with Live/dead cell staining kit (Life Technology, L34957) and fixed using eBioscience™ Intracellular Fixation & Permeabilization Buffer (Thermo Fisher, 88-8824-00). Cells were then stained by two antibody panels (1:200) against lymphocyte and myeloid cell markers separately. The lymphocyte panel is a mixture of the following antibodies: CD45-FITC (eBioscience, 11-0451-82), MHC-II-eFluor450 (eBioscience, 48-5321-82), CD3e-PE (BioLegend, 100206), CD49b-APC-Cy7 (BioLegend, 108920), CD19-BV650 (BioLegend, 115541), NK1.1-APC-Cy7(BioLegend, 108724), CD4-BV605 (BioLegend, 100547), CD8a-BV711 (BioLegend, 100747, RRID:AB_11219594), γδTCR-PerCP-Cy5.5 (BioLegend, 118118), CD69-PE/Cyanine7 (BioLegend, 104512), CD25-Alexa Fluor 700 (BioLegend, 102024) and FoxP3-APC (eBioscience, 17-5773-82). The myeloid marker panel includes CD11b-BV711 (BioLegend, 101241), CD11c-BV605 (BioLegend, 117333), F4-80-APC (BioLegend, 123116), Ly6C-PerCP-Cy5.5 (BioLegend, 128012), Ly6G-APC-Cy7 (BioLegend, 127624), CD103-PE/Cyanine7 (BioLegend, 121426), PDCA-1-PE (eBioscience, 12-3172-81), CD80-BV650 (BioLegend, 104731), CD86-Alexa Fluor 700 (BioLegend, 105024) as well as the CD45-FITC and MHC-II-eFluor450 listed above. Flow cytometry were analyzed using LSRFortessa™ High-Parameter Flow Cytometer (BD) after channel compensation. Gating strategy for each cell subtype are shown in Supplementary Fig. 5a, b. For in vitro SUM149 Fucci model experiment, cells were treated with or without 1 µM A-196 for 5 days, and then treated with 0, 100 and 500 nM of tunicamycin for 48 h. Cells together with flow-ups in the supernatants were collected and subjected to flow cytometry analysis using mCherry and GFP channel. Flow Jo was used to quantify cell fractions within each phase. Cell gating strategy is shown in Supplementary Fig. 5j. Detailed antibody information can be found in Supplementary Data 6.

Cytokine array

Snap-frozen tumor tissues were mechanically homogenized in PBS containing 1% Triton-X and protease and phosphatase inhibitors, followed by one freeze-thaw cycle. After centrifugation, the supernatant was collected. The protein concentration of the supernatant from each individual tumor was measured using the Pierce™ BCA Protein Assay Reagent (Thermo Fisher). A total of 200 µg of protein evenly pooled from each sample within the same group was combined for cytokine quantification. Cytokine levels were determined using either Proteome Profiler Mouse XL Cytokine Array (R&D Systems, ARY028) or Proteome Profiler Human XL Cytokine Array Kit (R&D Systems, ARY022B) analyzed on a Chemidoc MP (Bio-Rad) device, according to the manufacturer’s instructions. Intensities of each cytokine dots were quantified using ImageJ, average intensity from two technical replicates were then normalized to the positive control dots from the membrane.

qRT-PCR

Different cell models were seeded into six-well plate with 50,000 cells per well using biological triplicates. After the respective treatments, RNAs were extracted using Qiagen RNeasy Kit, and cDNAs were synthesized using PrimeScript RT Master Mix with 1000–200 ng RNA as input (Takara Bio, #RR036). qRT-PCR reactions were performed with SybrGreen Supermix (Bio-Rad, #1726275) and data were recorded using CFX Manager software (Bio-Rad Version 3.1), and the ΔΔCt method was used to analyze relative mRNA fold changes and GAPDH levels were measured as the internal control. Primers for qRT-PCR are listed in Supplementary Table 4.

PRISM screen

Cell lines

The current PRISM cell set consists of 931 cell lines representing more than 45 lineages including both adherent and suspension/hematopoietic cell lines. These cell lines largely overlap with and reflect the diversity of the Cancer Cell Line Encyclopedia (CCLE) cell lines (see https://portals.broadinstitute.org/ccle). Adherent cell lines were cultured in phenol red-free RPMI supplemented with 10% FBS, while suspension lines were maintained in phenol red-free RPMI with 20% FBS. Parental cell lines were stably transduced with a lentiviral vector carrying a unique 24-nucleotide DNA barcode, followed by selection with blasticidin. After selection, the barcoded lines were expanded and underwent quality control, including tests for mycoplasma contamination, SNP-based cell line authentication, and barcode sequence verification. Successfully validated lines were grouped into pools of 20–25 cell lines based on similar doubling times and cryopreserved in assay-ready vials.

PRISM screening

Test compound A-196 was dispensed into 384-well plates using an 8-point, threefold serial dilution format, with each concentration tested in triplicate. Thawed cell line pools were then seeded into these assay-ready plates. Adherent cell pools were plated at a density of 1250 cells per well, while suspension and mixed adherent/suspension pools were plated at 2000 cells per well. Following a 5-day incubation period, cells were lysed, and lysates from replicate wells were pooled before proceeding with barcode amplification and detection.

Barcode amplification and detection

Each cell line’s unique barcode is embedded in the 3′ untranslated region (UTR) of the blasticidin resistance gene, allowing it to be transcribed as part of the mRNA. Total mRNA was isolated using magnetic beads that bind to polyA tails. The captured mRNA was then reverse transcribed into cDNA, followed by PCR amplification of the region containing the unique PRISM barcode. Amplified products were hybridized to Luminex beads conjugated with probes specific to each barcode sequence. Detection was performed using a Luminex scanner, and barcode abundance was quantified as median fluorescent intensity (MFI).

Data processing

• I.Each detection well included ten spike-in control barcodes present at defined, increasing abundances. For each plate, a reference profile was generated by computing the median log₂(MFI) for each spike-in barcode across all negative control wells.
• II.For each well, a monotonic smoothing p-spline was fitted to align the spike-in control levels with the reference profile. The fitted spline was then used to transform the log₂(MFI) values of each cell barcode, enabling well-to-well comparisons by correcting for amplification and detection variability.
• III.
  Next, the separability between negative and positive control treatments was assessed. In particular, we calculated the error rate of the optimum simple threshold classifier between the control samples for each cell line and plate combination. Error rate is a measure of overlap of the two control sets and was defined as

  $$\mathrm{Error}=\frac{FP+FN}{n}$$ 1

  where FP is false positives, FN is false negatives, and n is the total number of controls. A threshold was set between the distributions of positive and negative control log₂(MFI) values (with everything below the threshold said to be positive and above said to be negative) such that this value is minimized. Additionally, we also calculated the dynamic range of each cell line. Dynamic range was defined as

  $$DR={\mathrm{\mu }}_{-}-{\mathrm{\mu }}_{+}$$ 2

  where μ + /− stood for the median of the normalized logMFI values in positive/negative control samples.
  Cell lines with an error rate greater than 0.05 or a dynamic range below 1.74 were excluded from downstream analysis. To ensure reproducibility, cell lines with fewer than two passing replicates were also removed. Viability was then calculated by normalizing each well’s signal to the median of the negative control wells on the same plate. Log-fold-change viabilities were computed as

  $${\mathrm{Log}}_{viability}=lo{\mathrm{g}}_2\left(X\right)-lo{\mathrm{g}}_2\left({\mathrm{\mu }}_{-}\right)$$ 3

  where log₂(x) is the corrected log₂(MFI) value in the treatment and log₂(μ − ) is the median corrected log₂(MFI) in the negative control wells in the same plate.
• IV.Log-viability scores were corrected for batch effects coming from pools and culture conditions using the ComBat algorithm⁸².
• V.We fit a robust four-parameter logistic curve to the response of each cell line to the compound:

$$\mathrm{f}\left(\mathrm{x}\right)=b+\frac{a-b}{1+e^{s\log \left(\frac{x}{EC50}\right)}}$$ 4

With the following restrictions:

1. Upper asymptote of the curve be between 0.99 and 1.01

2. The lower asymptote of the curve be between 0 and 1.01

3. No enforce decreasing curves

4. Initializing the curve fitting algorithm to guess an upper asymptote of 1 and a lower asymptote of 0.5

5. When the standard curve fit fails, we report the robust fits provided by the dr4pl R-package.

and computed AUC values for each dose-response curve and IC₅₀ values for curves that dropped below 50% viability.

Finally, the replicates were collapsed to a treatment-level profile by computing the median log-viability score for each cell line.

After data processing, we calculated univariate associations between the PRISM sensitivity profiles (each dose, log₂(AUC), and log₂(IC₅₀)) and used log₂(AUC) for the downstream analysis.

scRNA-seq data generation and analyses

Samples from tumor tissues were prepared as described for flow cytometry. Three tumors collected from three different animals were used as biological replicates per group. Cells were resuspended in 0.04% UltraPure BSA (Sigma-Aldrich) in PBS and immediately processed for library preps with DFCI Translational Immunogenomics Lab (TIGL). Gene expression library preparation was conducted using 10x Genomics ChromiumTM instrument (10x Genomics) according to the manufacturer’s recommendations using Chromium Next GEM Single Cell 5’ HT Kit v2 (10x Genomics). Quality controls for amplified cDNA libraries and final sequencing libraries were performed using Bioanalyzer High Sensitivity DNA Kit (Agilent). Equimolar ratios of libraries were sequenced on an Illumina NovaSeq 6000 (Illumina, RRID:SCR_016387) targeting 2000 reads per cell using 150 bp read pairs per library at the Dana-Farber Cancer Institute Molecular Biology Core Facilities. Raw data was processed using CellRanger (v7.0.1, RRID:SCR_023221) with bcltofastq function to obtain fastq files for each sample. Files were further processed using Cell Ranger Count 7.0.1 to obtain counts after aligning with mouse mm10 genome assembly. H5 files were then processed in R using Seurat package⁸³ (v 5.0.2). Low quality reads were removed according to following criteria: nCount_RNA > 1000, nFeature_RNA < 1000, percentage of mitochondrial reads <20, percentage of ribosomal reads <40 and log10 genes per UMI > 0.8. Normalization was performed using LogNormalize followed by SCTransform regressing ribosomal percentage, mitochondrial percentage, nCount_RNA and nFeature_RNA. Clustering and subclustering was performed using with the top 30 principal components. To merge 4T1 and EMT6 models, “HarmonyIntegration” was used. Next, cluster were manually annotated using gene.module.scores with specific marker genes. Cell subtype was annotated using established marker genes and confirmed with “Cell Types” search functions in Enrichr⁸⁴ using the top marker genes from the “FindMarker” function of Seurat. Within each subset cells were again annotated using selected marker genes and gene.module.scores and scores were plotted according to each genotype. UMAP visualization was performed using ‘scCustomize’ package⁸⁵ (v2.1.2). Neighborhood analysis for myeloid population was performed using the “miloR” package⁸⁵ (v1.10.0). Differentially expressing genes in each population was derived using “FindMarker” function using the cutoff of FDR < 0.05 and expression in at least 20% of detected cells. Module scores of specific gene signatures in clinical sample analysis were calculated using “UCell” package⁸⁶ (v 2.6.2).

RNA-seq data generation and analyses

For bulk RNA-seq analysis of tumor samples (3–4 tumors collected from different animals were used as biological replicates per group) and cells (n = 2 biological replicates) after desired treatments, total RNA was isolated using the Qiagen RNeasy Mini Kit. Library preparation and sequencing were conducted by Azenta RNA-seq services, yielding 15-30 million reads per sample. RNA-seq data were processed using the VIPER pipeline⁸⁷. Reads were mapped to the hg19 human genome using STAR (RRID:SCR_004463). Genes with zero counts across all samples were filtered out, and the remaining counts were normalized using edgeR⁸⁸ (RRID:SCR_012802) with a log2 TMM transformation to counts per million (log2(TMM-CPM + 1)). PCA plots were generated with edgeR, while heatmaps were visualized using the ‘ComplexHeatmap’ package⁸⁹ (RRID:SCR_017270). Differential expression analysis was performed with DESeq2⁹⁰ based on the experimental design. Pathway enrichment scores of each sample were computed using the ‘gsva’ package⁹¹ (RRID:SCR_021058). Pathway enrichment of a given list of differentially expressing genes was analyzed using Enrichr website (https://maayanlab.cloud/Enrichr/) using the Hallmark gene set for enrichment score, odd ratio and FDR values. For inference of transcriptional regulators, LISA⁴⁶ was used (http://lisa.cistrome.org/). For lineage signatures, pre-ranked analytic approach based on A-196-induced fold changes was used via Gene Set Enrichment Analysis (GSEA)⁹² software (v.4.3.2) with 1000 times of permutation test. Signatures used for enrichment analysis are listed in Supplementary Data 2.

ChIP-seq library generation

For histone modification, ATF4 and E2F4 ChIP-seq, cells grown in 150 mm tissue culture plates were fixed for 10 min at 37 °C in 1% formaldehyde (Electron Microscopy Sciences, 15714) with fixation buffer (50 mM HEPES-KOH pH 7.5, 100 mM NaCl, 1 mM EDTA pH 8, and 0.5 mM EGTA pH 8). Crosslinking was quenched with 0.125 M glycine for 5 min on a rocker. Cells were washed twice with ice-cold PBS and scraped into 10 mL PBS. After spinning at 200 g for 5 min, the supernatant was removed, and the pellet was flash frozen. The frozen pellet was resuspended in 1 mL of Cell Lysis Buffer (50 mM HEPES pH 8, 140 mM NaCl, 1 mM EDTA pH 8, 10% glycerol, 0.5% NP-40, 0.25% Triton X-100, and protease/phosphatase inhibitors) and incubated for 10 min at 4 °C with gentle rotation. Nuclei were pelleted by centrifugation at 1700× g for 5 min at 4 °C and washed twice with 1 mL Wash Buffer (10 mM Tris-HCl pH 8, 200 mM NaCl, 1 mM EDTA pH 8, 0.5 mM EGTA pH 8, and protease/phosphatase inhibitors). Nuclei were then resuspended in 1 mL Shearing Buffer (10 mM Tris-HCl pH 7.4, 1 mM EDTA pH 8, 0.1% SDS, 1% Triton X-100, 0.1% Sodium Deoxycholate, 0.25% N-Lauroylsarcosine, 1 mM DTT, and protease/phosphatase inhibitors) and sonicated in a Covaris E220 focused-ultrasonicator (peak intensity 140, 5% duty cycle, 200 cycles per burst) for 10–12 min using 1 mL AFA fiber tubes. Sonication was performed for each ChIP with 5 × 10⁶ cells per tube. Lysates were centrifuged at 10,000× g for 15 min, and the cleared supernatants were transferred to fresh tubes. NaCl was added to a final concentration of 150 mM. Samples were pre-cleared by incubating with 40 µL/ml of Dynabeads Protein G (Life Technologies, 10003D) for 1 h at 4 °C with rotation. After magnetic separation, the appropriate primary antibody (5 µg per sample for all the antibodies) was added for overnight immunoprecipitation at 4 °C with gentle rotation. Chromatin bound to the antibody was precipitated by adding pre-washed Dynabeads Protein G for 2 h at 4 °C with rotation. Beads were sequentially washed with low salt wash buffer (20 mM Tris-HCl pH 8, 150 mM NaCl, 10 mM EDTA, and 1% SDS), high salt wash buffer (50 mM Tris-HCl pH 8, 10 mM EDTA, and 1% SDS), LiCl wash buffer (50 mM Tris-HCl pH 8, 10 mM EDTA, and 1% SDS), and twice with 1× TE buffer. Each wash was done for 5 min at 4 °C with rotation. DNA was eluted from the beads with 300 µL of 100 mM sodium bicarbonate and 1% SDS for 30 min at room temperature with constant shaking. Crosslinking was reversed overnight at 65 °C, followed by RNA and protein digestion: 0.2 mg/ml RNase A for 30 min at 37 °C, then 0.2 mg/ml Proteinase K for 1 h at 55 °C. DNA was extracted by phenol/chloroform/isoamyl alcohol (Calbiochem, #516726), centrifuged at 20,000× g for 5 min at room temperature. DNA in the upper layer was precipitated by adding 1 volume isopropanol, 1/3 volume 2 M sodium perchlorate, and 5 µL glycogen. After centrifugation at 20,000× g for 10 min, the pelleted DNA was washed twice with 70% ethanol and resuspended in 20 µL low TE buffer. ChIP-seq libraries were prepared using the ThruPLEX DNA-seq Kit (Rubicon Genomics, R400427) according to the manufacturer’s instructions. Library fragments were size selected by PAGE within a 150–700 kb range. Sequencing was performed on a NextSeq6000 (Illumina) with 150 base pair paired-end reads.

ATAC-seq library generation

SUM149 cells were first pre-treated with or without 1 μM A-196 for 5 days and then subjected for 12 h of 1 μM tunicamycin stimulation. Cells were then viability frozen in full culture medium with 10% DMSO. Fifty thousand cells were then resuspended in 1 ml of cold ATAC-seq resuspension buffer (RSB) (10 mM Tris-HCl (pH 7.4), 10 mM NaCl and 3 mM MgCl2 in water). Cells were centrifuged at maximum speed for 10 min in a prechilled (4 °C) fixed-angle centrifuge. After centrifugation, supernatant was carefully aspirated. Cell pellets were then resuspended in 50 μl of ATAC-seq RSB containing 0.1% NP40, 0.1% Tween-20 and 0.01% digitonin by pipetting up and down three times and incubated on ice for 3 min. After lysis, 1 ml of ATAC-seq RSB containing only 0.1% Tween-20 was added and the tubes were inverted to mix. Nuclei were then centrifuged for 5 min at max speed in a prechilled fixed-angle centrifuge. Supernatant was removed and nuclei were resuspended in 50 μl of transposition mix (25 μl 2× TD buffer, 2.5 μl transposase (100 nM final), 16.5 μl PBS, 0.5 μl 1% digitonin, 0.5 μl 10% Tween20 and 5 μl water46) by pipetting up and down six times. Transposition reactions were incubated at 37 °C for 30 min in a thermomixer with shaking at 1000 r.p.m. Reactions were cleaned up with Qiagen MinElute columns. Libraries were amplified as previously described⁹³. In all, 35-bp paired-end reads were sequenced on a NextSeq500 instrument (Illumina).

ChIP-seq and ATAC-seq data analyses

ChIP-seq and ATAC-seq data processing were based on chips pipeline⁹⁴. Reads were aligned to hg19 genome using BWA-MEM aligner⁹⁵. Peak calling was performed using MACS2 v2.1.2 with FDR < 0.01 as the cutoff⁹⁶ with H4K20me3 and H3K9me3 peaks called using the broad peak mode. Seqplots⁹⁷ was used to make the ChIP-Seq heatmaps and intensity plots using Bigwig files generated from the MACS2 treat_pileup track, which represents the scaled fragment pileup (read coverage) across the genome. Genomic track visualization were conducted using WashU Epigenome Browser⁹⁸. For motif enrichment analysis, fasta sequence files were extracted based on Bed files using BEDtools⁹⁹ and then input into MEME Suite AME module¹⁰⁰ using HOCOMOCO full v11 database¹⁰¹ for calculation. E value below 10^-5 is considered as a significant enrichment. Motif enrichment was visualized using “circlize” package. For UPR target motif search, MEME Suite FIMO module was used. For global signal intensity comparison for H4 modifications and H3K9me3, signal intensities at each 1 kb bin of genome were calculated using the Cobra software, where multiBamSummary function was used from deepTools 3.5.0¹⁰² and differential marked regions were derived by DESeq2 package for differential bins in H4 WT and mutant models (Padj <0.05), H3K9me3 versus H4K20me3 baseline intensity (Padj <0.1) and by LIMMA¹⁰³ (v3.58.1) for differential bins in A-196-treated H4K20me3 and H4K20me1 bins with Padj <0.05. For ATAC-seq differentially peak analysis, all ATAC peaks were first merged, signal intensities were quantified on the concatenated sets and differential peaks were calculated using DEseq2 using cutoff of adjust P value < 0.05. Heatmap of differential ATAC sites were generated using deepTools 3.5.0¹⁰². For integration of RNA-seq and ChIP-seq/ATAC-seq, BETA⁴⁴ (RRID:SCR_005396) was performed as previously described⁴⁴. Briefly, differentially expressed (DE) genes were first computed using DESeq2 (RRID:SCR_015687)⁹⁰ and used as one of the inputs. BETA basic modules were used to compute the statistical associations between DE genes and DB peaks using 100 kb as the range to link gene TSS to each peak. P values were derived using one-tailed Kolmogorov–Smirnov test for upregulated and downregulated genes, respectively. For genomic feature enrichment analysis, ChIPseeker¹⁰⁴ package (v1.46.1) was used taking −/+ 3 kb from TSS as the promoter region. For genomic region-associated gene inference, the BETA minus function was used. Gene density of a given region was calculated by dividing the number of associated genes to the total length of the regions. To project signal intensity to a given genomic region, Bigwig files were used and normalized read coverage were quantified using multiBigwigSummary function from deepTools 3.5. Particularly, the genome-wide RepeatMasker coordinates of hg19 assembly was downloaded from UCSC genome browser. ATAC-seq signal was projected to the corresponding bed file and differentially accessible repetitive elements (FDR < 0.05 and |log2FC | >1) were derived from DESeq2.

Statistics and reproducibility

All quantitative data are presented as the mean values  ±  standard deviation with indicated independent experiments in the corresponding legends. All box plots span the upper quartile (upper limit), median (center) and lower quartile (lower limit). Whiskers extend a maximum of 1.5× IQR. No statistical methods were used to predetermine the sample size for the experiments. Statistic test was conducted using GraphPad PRISM v10.3.1 or embedded test in “ggpubr” package in R v4.3.1. Non-parametric tests were used, should data for matched comparisons did not pass the test. Non-paired, two-sided t test or Mann–Whitney U test was used for single comparisons. One-way ANOVA corrected for multiple comparisons or Kruskal–Wallis test was used for multi-group comparisons. All tests were performed with a 95% confidence interval. P values are indicated for each experiment. In all in vivo experiments, mice were randomized to treatment groups; otherwise, experiments were not randomized. Image quantification in Figs. 2b, c and  7g were analyzed blindly. The investigators were not blinded to allocation during other experiments and outcome assessment. Sequencing data failed at QC were excluded from the analyses. For computational analyses performed on fixed datasets, the analyses were conducted a single time because application of the same computational pipeline to the same fixed dataset would be expected to produce identical results. Nevertheless, all analyses incorporated the appropriate biological replicates available within each dataset, and the corresponding statistical methods are described in the manuscript. For animal experiments, statistical analyses were performed using data from individual tumors collected from different mice, considered an independent biological replicates. The animal experiments were not repeated as independent cohorts due to resource constraints and the animal number limitations specified in the approved IACUC protocol. For bulk RNA-seq, ATAC-seq, and ChIP-seq experiments, all analyses were performed using n = 2 independent biological replicates per condition. Unlike conventional low-dimensional in vitro assays, in which statistical comparisons are typically based on measurements of a single endpoint, genomic sequencing experiments generate high-dimensional datasets comprising measurements across thousands of genes or genomic regions simultaneously. Differential gene expression and chromatin enrichment analyses were therefore performed using established computational frameworks specifically developed for these data types, which model sequencing counts across all measured features while incorporating biological replication to estimate dispersion and statistical significance. For all reported analyses, gene expression levels and chromatin signal intensities were quantified using data from both biological replicates, thereby reducing the impact of technical variation and providing representative measurements for downstream statistical analyses. For scRNA-seq analyses, the number of biological samples is indicated in the corresponding figure legends. Mouse tumor scRNA-seq experiments were performed using three independent tumors per group. The analyses shown in Supplementary Fig. 6b, c were based on a previously published public scRNA-seq dataset comprising one sample containing 2239 cells for the indicated cell line. As these analyses were performed on an existing publicly available dataset, no additional biological replicates were generated in the present study. The results are presented as exploratory analyses of this published dataset and are interpreted accordingly. Figure 7m was created in BioRender.com. All figures were assembled with Affinity Designer 2.0. Statistic test in R uses a precision floating-point format, which has a lower limit of approximately 2.2 × 10⁻¹⁶, and the smallest allowable reported P value in GraphPad PRISM is 1 × 10⁻¹⁵. When P values fall below these thresholds, the tools report a range (that is, P  <  2.2 × 10⁻¹⁶ in R and P  <  1 × 10⁻¹⁵ in GraphPad) rather than attempt to report a less precise or unreliable value.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization, Z.L. and K.P.; methodology, Z.L.; formal analysis, Z.L., A.T., P.F., T.B., E.R.J., R.L., and M.M.G.; investigation, Z.L., A.T., J.N., M.S., X-Y.H., M.G.S., M.M.R., and J.A.R; resource: L.E.S; writing—original draft, Z.L. and K.P.; writing—review and editing, all authors; supervision, H.W.L. and K.P.; funding acquisition, H.W.L., K.P., and Z.L. All authors helped to design the study and write the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

K.P. discloses support for the research of this work from the National Cancer Institute R35 CA197623 and the Dana-Farber Cancer Institute Innovation Research Fund. K.P. and H.W.L. disclose support for the research of this work from the National Cancer Institute P01CA250959. Z.L. discloses support for the research of this work from the National Cancer Institute K99 CA297044, from the Charles A. King Trust Postdoctoral Research.

Data availability

All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplemental Information. All raw and processed data newly generated from this study including bulk RNA-seq, ATAC-seq, ChIP-seq and scRNA-seq are deposited to GEO database under accession number GSE289193. All the genomic data were aligned to human reference genome GRCh37/hg19 [https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.13/] or mouse reference genome GRCm38/mm10 [https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001635.20/]. For public dataset analysis, processed and normalized bulk RNA-seq, histone mass spectrometry and H3K27ac ChIP-seq datasets in 34 TNBC cell lines; RNA-seq from 15 TNBC PDX models; PRRX1 ChIP-seq from Hs578T, MDA-MB-157 and MDA-MB436 cell lines; H3K27ac ChIP-seq in HCC3153 PRRX1 overexpression models were described in our prior studies¹⁹ and are available in the GEO database under accession code GSE202776. RNA-seq from SUM149 paclitaxel-resistant derivative EpCAM-high and EpCAM-negative cells⁴² used in this study are available in the GEO database under accession code GSE163397. H4K20me3 and H3K9me3 ChIP-seq from Hs578T cell line¹⁶ was reprocessed and the raw data are available in the GEO database under accession code GSE143653. ATAC-seq of Hs578T cells used in this study are available in the GEO database under accession code GSE223182¹⁰⁵. For integrative analysis, ATAC-seq and H3K27ac ChIP-seq data of MDA-MB-231 cells used in this study are available in the GEO database under accession code GSE72141¹⁰⁶. H3K27me3 ChIP-seq data from MDA-MB-231 cells used in this study are available in the GEO database under accession code GSE73702¹⁰⁷. ATAC-seq data from MDA-MB-468 used in this study are available in the GEO database under accession code GSE157082¹⁰⁸. H3K27ac ChIP-seq data of MDA-MB-468 used in this study are available in the GEO database under accession code GSE69107¹⁰⁹]. H3K27me3 ChIP-seq data of MDA-MB-468 used in this study are available in the GEO database under accession code GSE189215. The mRNA expression data and the clinical data of TCGA²⁸ and METABRIC²⁷ are available from the TCGA data portal [https://portal.gdc.cancer.gov] and Synapse under accession code Syn1688369, respectively. For TCGA, RNA-seq reads were reprocessed using Salmon (v0.14.1)87 and log2(transcripts per million + 1) values were used. For genes with multiple probes in METABRIC, probes with the highest IQR were selected to represent the gene. Patients were assigned a TNBC status based on the downloaded clinical annotation files with “Negative” annotation for ER, PR, and HER2. RNA-seq data from 101 tumors from CPTAC cohort were obtained from Krug et al.⁴³. And the matched histone mass spectrometry data was downloaded from Panorama: [https://panoramaweb.org/LINCS/GCP/project-begin.view] (CPTAC plate). Tumor mutations burden was calculated using SNV data obtained from FireBrowse (TCGA) [http://firebrowse.org/] or DepMap (CCLE) [https://sites.broadinstitute.org/ccle/]. Mutations subtypes were classified into truncated (nonsense, frame-shift deletion, frame-shift insertion, splice-site) and non-truncated mutations (missense, in-frame deletion, in-frame insertion, nonstop). TMB was calculated as 2× Truncating mutation numbers + non-truncating mutation numbers. scRNA-seq data of 19 TNBC tumors were downloaded and reprocessed from two prior works, data are available in the GEO database under accession code GSE161892⁴¹ and [https://lambrechtslab.sites.vib.be/en/single-cell]⁴⁰. The remaining data are available within the Article, Supplementary Information or Source Data file. Source data are provided with this paper.

Competing interests

K.P. serves on the Scientific Advisory Board of Ideaya Biosciences, is an adviser to Curie.Bio, holds equity in Antares Therapeutics and stock options in Ideaya Biosciences, receives funding from Novartis, and received honoraria from AstraZeneca and Novartis, and payment from ELI Lilly for Scorpion Biosciences purchase in the past 24 months. T.M.B. is a shareholder of Merck & Co., Novo Nordisk A/S, Eli Lilly & Co., AbbVie Inc., and Johnson & Johnson. L.E.S. is current employee of AstraZeneca. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76773-0.