Single-cell epigenetic profiling reveals a tumor-intrinsic interferon response program in ccRCC tied to poor prognosis and BAP1 loss

Abstract

Transcriptional programs in renal cell carcinoma (RCC) have been linked to tumor heterogeneity and clinical outcomes, but analogous efforts to define chromatin programs shaping disease biology have been limited. Here, we generated single-cell ATAC-seq profiles from patients with RCC and integrated them with three previously published datasets to identify chromatin programs in tumor cells. We identified an interferon response program enriched in BAP1-mutant tumors, and, in bulk ATAC-seq cohorts with linked clinical data, this program was associated with poor prognosis. Mechanistic analyses in isogenic models suggested that BAP1 loss induces a tumor-intrinsic interferon response, with dysregulated endogenous retroviruses as a potential upstream trigger. We further characterized the BAP1 mutation–associated tumor microenvironment across single-cell, bulk, and multiplex immunofluorescence data, identifying features of both inflammation and immune evasion. Together, our findings nominate tumor-intrinsic interferon signaling as a candidate driver of BAP1-associated aggressiveness in RCC and highlight immune evasion pathways as potential therapeutic targets.

Single-cell chromatin profiling reveals how BAP1 loss drives interferon signaling and aggressiveness in ccRCC.

INTRODUCTION

Renal cell carcinoma (RCC) is a cancer partly defined by recurrent mutations in epigenetic regulators. Deficiency in these epigenetic regulators has been associated with distinct evolutionary paths, influencing core disease biology, prognosis, and therapeutic outcomes (1–5). For example, mutations in BAP1 define a particularly aggressive subset of disease characterized by high tumor grade, sarcomatoid differentiation, and tumor necrosis (6–10). Conversely, the near mutually exclusive PBRM1-deficient RCC tends to be of lower grade and may serve as a predictive biomarker of response to anti-angiogenic therapy and immune checkpoint blockade (ICB) in select clinical contexts (5, 11–16).

While these associations underscore the importance of epigenetic regulation in RCC, the heterogeneity of alterations raises the question of whether tumors converge on shared epigenetic programs with functional or clinical relevance. Notably, recurrent transcriptional programs have already been described in RCC. For example, Bi et al. (17) used single-cell RNA sequencing (scRNA-seq) to define two tumor-intrinsic programs: one enriched for angiogenesis and kidney differentiation and the other marked by metabolic plasticity and immune evasion; the former was predictive of ICB response. In bulk RNA sequencing (RNA-seq) data, Motzer et al. (18) identified seven transcriptional programs, including subsets associated with therapeutic benefit from anti-angiogenic therapies or ICB.

Analogous efforts to define clinically and biologically relevant epigenetic programs have been more limited. Yu et al. (19) profiled 19 clear cell RCC tumors with single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) and identified tumor chromatin states, but these were largely sample specific and they did not evaluate their association to clinical outcomes. In bulk ATAC-seq data, Fukagawa et al. (20) observed three epigenetic subtypes, including an immune-enriched group associated with shorter disease-free survival (DFS). However, it remains unclear to what extent these programs reflect tumor cells versus nontumor components of the tumor microenvironment (TME). Defining such programs would clarify how epigenetic dysregulation shapes RCC outcomes and highlight shared pathways that could serve as biomarkers or therapeutic targets.

Here, we generated scATAC-seq data from patients with RCC and integrated it with three published datasets (19, 21, 22) to define recurrent chromatin programs in tumor cells. We identified an interferon (IFN) response program enriched in BAP1-mutant tumors, and, in bulk sequenced patient cohorts, this program was associated with poor prognosis. Mechanistic analyses suggested that BAP1 loss induces tumor-intrinsic IFN signaling and that dysregulation of endogenous retroviruses (ERVs) may act as an upstream trigger. These findings nominate persistent tumor-intrinsic IFN signaling as a candidate driver of BAP1-associated aggressiveness.

RESULTS

Single-cell chromatin accessibility atlas of RCC

To identify shared epigenetic programs in RCC tumor cells, we generated scATAC-seq data from 16 tumor biopsies (13 patients) and combined it with three previously published studies (19, 21, 22), yielding an integrated dataset of 177,845 cells from 61 biopsies across 58 patients (Fig. 1A).

Fig. 1. Single-cell chromatin accessibility atlas of RCC.

(A) Overview of the study design. (B) Comutation plot of the clinical and genomic characteristics of the biopsies included in the study. Each row is a clinical variable or gene, and each column corresponds to an individual biopsy. NA, not applicable. (C) Heatmap of a subset of chromosome arm-level copy number estimates, split by tumor and nontumor cells, downsampled to 10,000 cells each for visualization. Histology color mapping in (B) legend. (D) Heatmap of scaled average gene activity scores across broad nontumor cell types, downsampled to 100 cells per group for visualization. The top 100 differentially accessible genes (DAGs) per cell type are shown. Canonical marker genes in top 100 sets are annotated. (E) Uniform manifold approximation and projection (UMAP) visualization of all cells included in the study. Broad cell types are annotated; cells shown in light gray were excluded from downstream analyses. T_reg, regulatory T cell; NK, natural killer; cDC, conventional dendritic cells; MoDC, monocyte-derived dendritic cells.

Our internal scATAC-seq dataset was derived from frozen tissue, with four patients also profiled using multiome single-cell RNA and ATAC sequencing (scRNA-ATAC-seq). Among published datasets, Wu et al. (21) profiled frozen samples, whereas Long et al. (22) and Yu et al. (19) profiled fresh tissue. Most specimens across internal and external cohorts were derived from the kidney, with a minority from bone, lymph node, and visceral metastases. Two specimens had non–clear-cell subtypes, papillary and TFE3 translocation (identified by a TFE3 fluorescence in situ hybridization test), while all other specimens were of clear cell histology. The combined cohort contained a mixture of disease stages and grades. Genomically, the cohort had a high frequency of somatic mutations in VHL, PBRM1, KDM5C, BAP1, and SETD2, recapitulating known genomic characteristics of RCC (Fig. 1B and table S1) (23).

To detect shared cell types and states across samples, we corrected low-dimensional embeddings for sample-specific batch effects. Tumor cells were identified primarily through copy number inference; as expected, most tumor cells from clear cell tumors had chromosome 3p loss alongside other arm level gains and losses, distinguishing them from nonmalignant cells (Fig. 1C). Nonmalignant cell types were annotated through iterative reclustering, evaluation of differentially accessible genes (DAGs), and canonical marker gene accessibility (Fig. 1, D and E, and Materials and Methods). Broad cell types were captured consistently across cohorts and sample preservation methods, and quality control (QC) metrics were comparable (fig. S1, A and B, and table S1). Together with the variation in genotype, disease stage, and patient characteristics, this dataset provides the foundation for discovery analyses of generalizable chromatin programs in tumor cells.

Histologic subtype–driven epigenetic heterogeneity of RCC tumor cells

To detect epigenetic programs shared across patient tumors, we first performed dimension reduction and unsupervised clustering within all tumor cells. RCC histologic subtype emerged as a major source of variation (Fig. 2, A and B, and fig. S2A), consistent with prior reports (20, 24). One cluster, RADIL-high, was specific to the single papillary RCC sample and showed increased accessibility at RADIL, PGAP3, and HOXB7. Another cluster, TRIM63-high, was enriched in the translocation RCC sample and had increased accessibility at TRIM63, TIMP2, and PRCD (Fig. 2B and table S2). Subtype-specific accessibility patterns could not be generalized further given the limited number of papillary and translocation samples.

Fig. 2. Shared epigenetic programs in ccRCC tumor cells are functionally and clinically relevant.

(A) UMAP of putative tumor cells from all samples, colored by malignant cell state; selected clusters are annotated with representative genes showing high accessibility. (B) Heatmap of scaled average gene activity scores across tumor states, downsampled to 100 cells each for visualization. The top eight DAGs per state are shown. (C) UMAP of ccRCC tumor cells highlighting four shared clusters (C0 to C3) derived from subclustering the ccRCC-balanced state. (D) Heatmap of differentially accessible peaks (DAPs) across tumor states (C0 to C3). Each column represents the scaled average accessibility of cells from a given biopsy assigned to that tumor state. Columns are grouped by tumor state and cohort. (E) Bar plot of selected pathways enriched in state-specific peak sets, identified with GREAT [binomial test with false discovery rate (FDR) correction]. HIF-1α, hypoxia-inducible factor–1α. (F) Heatmap of the top 10 transcription factor (TF) motif enrichments across tumor state peak sets, showing −log₁₀ q values from a hypergeometric test with FDR correction. Values were capped at −log₁₀ q = 10 for visualization. (G) TF footprinting for HNF4A (top), FOS (middle), and RELA (bottom) across tumor states. (H) Heatmap of Cliff’s delta effect sizes for associations between tumor programs (C0 to C3) and epigenetic modifier mutations; asterisks denote significance by two-sided Wilcoxon test with FDR correction (**q < 0.01). (I) Boxplot of per-biopsy median C3 program scores stratified by BAP1 mutation status (two-sided Wilcoxon test with FDR correction). (J) Forest plot of hazard ratios (HR) for C1 to C3 programs in multivariable Cox models of overall survival (OS) and DFS across bulk ATAC-seq cohorts (*P < 0.05; n.s., not significant). (K) Kaplan-Meier curves for DFS in early-stage patients, comparing top quartile (Q4) versus Q1 to Q3 C3 program proportion. P values from log-rank test (P_log-rank) and univariable Cox model [P_Cox(UVA)] are shown.

Among ccRCC samples, several clusters were highly sample specific (e.g., PDGFRA-high, CADM1-high, and NCAM1-high; fig. S2A). In contrast, a group of clusters contained cells from multiple tumors, which we collectively termed “ccRCC-balanced” (Fig. 2, A and B; fig. S2A; and table S2).

Four shared chromatin programs in clear cell RCC tumor cells

Because ccRCC represented most of tumors in our dataset, we next focused on this subtype to resolve epigenetic heterogeneity beyond subtype-specific differences. Subclustering of the ccRCC-balanced subset (n = 56 tumors; 88,739 tumor cells) revealed four epigenetic programs shared across patients, cohorts, and disease stages, which we termed C0, C1, C2, and C3 (Fig. 2C and fig. S2B). We further identified 1027 regions with significantly different accessibility across these states. There were no peaks enriched for accessibility in C0, suggesting that this state may be comparatively epigenetically quiescent (Fig. 2D and table S2).

To define the biological properties of each state, we performed pathway enrichment analysis on state-specific peaks using GREAT (25). Peaks uniquely accessible in C1 were enriched for solute transport pathways, including glucose, bile salts, organic acids, and ions, consistent with kidney-associated transport and differentiation processes. C2 was enriched for angiogenesis-related pathways, including vascular disease, the hypoxia-inducible factor–1α transcription factor (TF) network, and vascular endothelial growth factor signaling regulation. Last, C3 showed increased accessibility at loci linked to IFN signaling and inflammation, including type I/II IFN signaling, nephritis, and cytokine production (Fig. 2E, fig. S2C, and table S2).

To identify TFs potentially regulating these tumor states, we tested for motif enrichment within the defining peaks of each state. Each state showed distinct TF motif enrichments. C1 peaks were enriched for motifs of TFs related to kidney function, including HNF4A, a key regulator of proximal tubule development (26). Because C1 was characterized by kidney differentiation pathways and HNF4A motif enrichment, we considered whether this program might reflect contaminating nonmalignant kidney cells. However, tumor cells in C1 exhibited chromosome 3p loss at similar levels to the other tumor states, indicating that this program represents malignant rather than contaminating normal cells (fig. S2D). C2 peaks were enriched for FOS and JUN motifs. Last, C3 peaks were enriched for nuclear factor κB (NF-κB) family motifs (e.g., REL, RELA, and NFKB2), which are closely linked to innate immunity and type I IFN signaling (Fig. 2F and table S2). TF footprinting analysis supported these findings, where accessible regions containing these motifs showed greater accessibility in tumor cells from the corresponding state (Fig. 2G). Together, these analyses defined four recurrent epigenetic programs in ccRCC tumor cells: C0, an epigenetically quiescent state; C1, characterized by kidney differentiation-associated features; C2, marked by angiogenesis-related pathways and AP-1 activity; and C3, enriched for IFN signaling and NF-κB activity.

Chromatin programs linked to genotype and prognosis in ccRCC

Given that epigenetic modifier genes are recurrently mutated in ccRCC, we next tested whether these alterations were associated with the shared chromatin programs. Biopsies with putative loss-of-function (pLOF) BAP1 mutations showed increased accessibility in the IFN response–high C3 program (Fig. 2, H and I), with minimal confounding effects from disease stage and the presence of sarcomatoid features (fig. S2, E and F). In contrast, biopsies with mutations in PBRM1, SETD2, or KDM5C were not enriched for accessibility in any of the shared programs (Fig. 2H and table S2).

Next, we asked whether the chromatin programs were associated with clinical outcomes in ccRCC. We projected program signatures into two clinically-annotated bulk ATAC-seq cohorts [Fukagawa et al. (20) and The Cancer Genome Atlas Kidney Renal Clear Cell Carcinoma (TCGA KIRC) (27)] and evaluated associations with overall survival (OS) and DFS. In multivariable Cox proportional hazards models controlling for age, sex, stage, tumor purity, cohort, and sequencing depth, the IFN response–high C3 program was significantly associated with shorter DFS (Fig. 2J and fig. S2G). No other programs were associated with DFS or OS (fig. S2G). In Kaplan-Meier analysis restricted to early-stage tumors, patients in the highest quartile of C3 scores also had shorter DFS compared to all others. Because C3 was enriched in BAP1-mutant tumors, we further tested whether its prognostic association was independent of BAP1 status. Adjustment for BAP1 status yielded a similar effect (hazard ratio of 2.74 versus 2.83 unadjusted), although significance was lost (P = 0.109). These findings suggest that C3 may underlie features of the aggressive phenotype observed in BAP1-mutant tumors (fig. S2H). Together, these results identify the IFN-high C3 program as an epigenetic state enriched in BAP1-mutant tumors and associated with adverse outcomes.

BAP1 loss induces a tumor-intrinsic IFN response in ccRCC

The enrichment of BAP1 mutation in the IFN-high C3 program suggests that BAP1 loss may be one determinant of this chromatin state. To test this, we next performed a supervised comparison of BAP1-mutant and wild-type (WT) tumor cells, asking whether an IFN-high signature is recapitulated when stratifying by BAP1 mutation status.

Consistent with prior reports, patients with BAP1-mutated tumors were enriched for advanced stage disease (fig. S3A) (7, 28). To minimize the confounding effect of stage, all subsequent analyses were restricted to advanced-stage samples (n = 20 biopsies). Differential peak accessibility analysis identified 1272 regions with increased accessibility and 51 regions with decreased accessibility in BAP1-mutant tumor cells relative to WT (Fig. 3A and table S3). Because this restriction reduced sample size (n = 20 biopsies), we repeated the analysis including all tumors (n = 53 biopsies) with stage as a covariate. Results from this model were highly concordant with the stage-restricted analysis [Pearson correlation coefficient (r) = 0.93 for log₂ fold changes; fig. S3B and table S3], and accessibility at BAP1-enriched peaks continued to separate tumors by BAP1 status irrespective of stage (fig. S3C). Together, these results support that the observed differences reflect true BAP1-associated biology rather than a sample size artifact. In contrast to a prior study (21), we did not observe globally reduced accessibility in BAP1-mutant tumors; instead, that effect appeared to be cohort specific (fig. S3D).

Fig. 3. BAP1 loss induces a tumor-intrinsic IFN response in ccRCC.

(A) Heatmap of DAPs between BAP1-mutated and WT tumor cells. Each column represents the scaled average accessibility of cells from a given biopsy; columns are grouped by BAP1 mutation status, stage, and cohort. The number of significant peaks is not proportional to the heatmap display. (B) Dot plot of pathways enriched in peaks with increased accessibility in BAP1 pLOF tumor cells, identified with GREAT (binomial test with FDR correction). Displayed pathways are the top 15 Gene Ontology Biological Process (GO BP) terms. (C) Motif enrichment in BAP1 pLOF mutation-associated peaks. Significant motifs were identified by hypergeometric test with FDR correction (n.s., not significant). Motifs are ordered by q value (left to right). Labels for select motifs are displayed. (D) Schematic of isogenic BAP1 model generation and molecular profiling (31). (E) Volcano plot of pathway enrichment from GSEA of RNA-seq data comparing BAP1 KO and WT tumor cells across Hallmark gene sets and the tumor type I IFN–inducible signature (32). Pathways with q < 0.01 (GSEA P value with FDR correction) are annotated. NES, normalized enrichment score. (F) Volcano plot of pathway enrichment from overrepresentation analysis of differentially expressed proteins (DEPs) between BAP1 KO and WT tumor cells across Hallmark gene sets and the tumor type I IFN–inducible signature (32). Pathways with q < 0.01 (hypergeometric test with FDR correction) are annotated. (G) Heatmap of DEPs between BAP1 KO and WT tumor cells that overlap the tumor type I IFN–inducible signature of Bi et al. (32). Values represent scaled protein abundances (z-scores).

Peaks with increased accessibility in BAP1-mutant tumor cells were enriched for type I and II IFN signaling pathways, among others (Fig. 3B). In contrast, peaks with decreased accessibility were not significantly enriched for any annotated pathways. Regions with increased accessibility in BAP1-mutant tumor cells were also enriched for motifs of NF-κB family TFs, as well as FOS and JUN motifs (Fig. 3C and table S3). Peaks with decreased accessibility were not significantly enriched for any TF motifs. These findings are concordant with prior bulk transcriptomic and proteomic studies linking BAP1 mutation to IFN and inflammatory signaling (18, 29, 30) and extend these to demonstrate that this association is evident within tumor cells at the chromatin level.

To test whether BAP1 loss is sufficient to induce an IFN-high state, we analyzed an isogenic CRISPR-Cas9 BAP1 knockout (KO) model in the 786-O ccRCC cell line (Fig. 3D) (31). In bulk RNA-seq data, pathway enrichment analysis revealed strong induction of IFN-related pathways, including Hallmark IFN-α and IFN-γ responses, as well as a type I IFN–inducible gene set previously described in ccRCC tumor cells (Fig. 3E and fig. S3E) (32). Similarly, proteomic analysis confirmed enrichment of IFN response pathways, with the same signatures as the top hits (Fig. 3, F and G; fig. S3F; and table S3). Together, these results demonstrate that BAP1 loss induces a tumor-intrinsic IFN response in ccRCC, detectable in both patient tumors and an isogenic model.

ERV activity associates with IFN signaling and BAP1 mutation in ccRCC

Given that BAP1 loss induces a tumor-intrinsic IFN program, we next asked what upstream mechanisms might trigger this phenotype. One candidate is viral mimicry, supported by a prior report of elevated ERV expression in BAP1-mutant ccRCC tumors (33).

To explore this, we first developed a chromatin-based measure of IFN signaling in tumor cells (Fig. 4A). Using multiome scRNA-ATAC-seq samples, we identified genomic regions whose accessibility correlated with RNA-based measures of type I and II IFN signaling in tumor cells and used these peak sets as IFN1 and IFNG signatures (table S4). Accessibility at these regions correlated with RNA-defined IFN activity (Fig. 4B), and the peak sets were enriched for IFN-related pathways and binding motifs of IFN-related TFs such as interferon regulatory factors (IRFs), signal transducers and activators of transcription (STATs), and NF-κB family members (Fig. 4, C and D; fig. S4, A to C; and table S4).

Fig. 4. ERV activity is associated with IFN signaling and BAP1 mutation in ccRCC.

(A) Overview of methodology for identifying IFN-associated ERVs in tumor cells. (B) Scatter plot of RNA- versus ATAC-derived IFN1 signaling in tumor cells from multiome scRNA-ATAC-seq samples. Each point is a cell, colored by sample. Pearson correlation coefficient (r) and regression line are shown. (C) Dot plot of pathways enriched in the IFN1-associated peak set, identified with GREAT (binomial test, FDR correction). Displayed pathways include the top five GO BP terms and top five MSigDB pathways. (D) Motif enrichment in the IFN1-associated peak set. Significant motifs identified by hypergeometric test with FDR correction (n.s., not significant). Motifs ordered by q value (left to right); labels for select motifs are shown. (E) Scatter plot of mixed-effects model coefficients for ERVs, comparing associations with IFN1 (x axis) and IFNG (y axis) signaling. Only ERVs with positive coefficients are shown. Labeled ERVs significantly associated with both (FDR q < 0.05, coefficient > 0). (F) Volcano plot of Cliff’s delta effect sizes for associations between BAP1 mutation status and accessibility of IFN-associated ERVs in advanced-stage ccRCC tumor cells. Significant ERVs (one-sided Wilcoxon test; alternative: pLOF > WT) in red. (G) Boxplot of median accessibility of ERV3-16A3_LTR in BAP1 pLOF tumor cells, BAP1 WT tumor cells, and immune cells. q values for BAP1 pLOF versus WT (one-sided Wilcoxon test with FDR correction; alternative: pLOF > WT) and nominal P values are shown; two-sided P values are shown for immune cell comparisons. (H) Boxplot of normalized accessibility of ERV3-16A3_LTR in BAP1 pLOF versus WT tumors across bulk ATAC-seq cohorts. One-sided Wilcoxon test P values (alternative: pLOF > WT, P_wilcox); multivariable model P values (P_mv) are shown. (I) Boxplot of normalized expression of ERV3-16A3_LTR in BAP1 pLOF versus WT tumors in TCGA KIRC cohort. One-sided Wilcoxon test P values (alternative: pLOF > WT, P_wilcox); multivariable model P values (P_mv) are shown.

We then quantified ERV accessibility in our scATAC-seq dataset using a single-cell transposable element (TE) profiling approach (34). To validate this quantification, we first asked whether TE accessibility could reproduce expected biological signals. As reported previously (34), TE profiles alone distinguished broad cell types, and, in tumor cells, we observed enriched activity of HERVE-int, a tumor-specific ERV previously identified at the RNA level in ccRCC (fig. S4, D and E) (35).

With per-cell measures of IFN signaling and ERV accessibility, we next tested whether ERVs were linked to IFN signaling in ccRCC tumor cells. Nineteen ERVs were significantly associated with higher chromatin-based IFN scores after controlling for technical and biological covariates (Fig. 4E and table S4). This set included HERVE-int, consistent with its known immunogenic role in ccRCC (35, 36).

Last, we examined whether IFN-associated ERVs were linked to BAP1 mutation status. Among the 19 ERVs identified, ERV3-16A3_LTR was most strongly associated with BAP1 status, showing greater accessibility in BAP1-mutant tumor cells at nominal significance (Fig. 4, F and G, and table S4). In bulk ATAC-seq, ERV3-16A3_LTR accessibility was also higher in BAP1-mutant tumors (P = 0.046), with a trending association after adjustment for clinical and technical covariates (P = 0.076), likely reflecting the limited number of BAP1-mutant samples in this dataset (n = 5; Fig. 4H and fig. S4F). Bulk RNA-seq from TCGA KIRC likewise showed significantly higher ERV3-16A3_LTR expression in BAP1-mutant tumors in both unadjusted and adjusted analyses, with the association strongest in early-stage disease (Fig. 4I and fig. S4, G to I). In a second bulk RNA-seq dataset, the JAVELIN Renal 101 cohort composed of advanced-stage tumors, we did not observe an association between BAP1 status and ERV3-16A3_LTR expression, suggesting that this link may be attenuated in advanced disease (fig. S4J). Last, we evaluated whether BAP1 loss was sufficient to induce ERV3-16A3_LTR expression. In an isogenic BAP1 model, results were heterogeneous relative to WT, indicating that BAP1 loss alone may be insufficient to consistently drive its activation (fig. S4K). Together, these analyses identify ERV3-16A3_LTR as a BAP1-associated ERV linked to IFN signaling in ccRCC tumors.

BAP1 loss shapes an inflamed yet immune evasive TME

Thus far, we have focused on tumor-intrinsic phenotypes of BAP1 loss. To understand their broader significance, we next asked whether these changes influence the TME and the ways that tumor cells engage with infiltrating immune cells.

Prior studies have linked BAP1 mutation to an inflamed TME (8, 18, 29, 30, 37). To explore whether tumor-intrinsic IFN signaling might contribute to immune infiltration, we evaluated accessibility at immune-recruiting loci in BAP1-mutant tumors. Tumor cells from BAP1-mutant lesions showed increased accessibility at loci encoding chemokines and cytokines that recruit CD8⁺ T cells, natural killer cells, and myeloid populations (e.g., CXCL10, CXCL11, CCL5, and CSF1; Fig. 5A, fig. S5A, and table S5).

Fig. 5. BAP1 loss shapes an inflamed yet immune evasive TME.

(A) Coverage plots at CXCL10, CXCL11, CCL5, and CSF1 comparing chromatin accessibility in BAP1 pLOF versus WT ccRCC tumor cells. Downsampled to ≤1000 cells per biopsy for visualization and restricted to tumor states C0 to C3 and advanced-stage patients. (B) Boxplots of broad immune cell-type proportions in BAP1 pLOF versus WT tumors, restricted to advanced-stage patients. Each point represents a per-biopsy proportion. The monocyte comparison was significant (two-sided Wilcoxon test with FDR correction); other comparisons were not. (C) Boxplots of CD8⁺ cell density, CD8⁺PD1⁺ cell density, and the CD8⁺PD1⁺/CD8⁺ ratio in tumors with BAP1 pLOF versus WT. q values are from one-sided Wilcoxon tests (alternative: pLOF > WT) with FDR correction. Color denotes ICB exposure at biopsy; shapes indicate lymph node versus non–lymph node samples. (D) Representative mIF images [one region of interest (ROI) each] from two BAP1 pLOF samples (lung, lymph node biopsies) and two WT samples (lung, lymph node biopsies), all ICB exposed. First row, CD8; second row, PD1. PAX8 and DAPI are shown in both. (E) Coverage plots at HLA-E, HLA-G, LGALS9, and VSIR comparing chromatin accessibility in BAP1 pLOF versus WT ccRCC tumor cells. Downsampled to ≤1000 cells per biopsy for visualization and restricted to tumor states C0 to C3 and advanced-stage patients.

We next compared immune cell proportions between BAP1-mutant and WT tumors in the scATAC-seq dataset. Monocyte proportions were significantly higher in BAP1-mutant tumors, while CD8⁺ T cells and tumor-associated macrophages (TAMs) showed directionally similar but nonsignificant increases (Fig. 5B). These patterns persisted when excluding sarcomatoid cases and stratifying by cohort and showed similar directionality in bulk RNA-seq deconvolution (TCGA KIRC and JAVELIN Renal 101) (fig. S5, B to J).

In multiplex immunofluorescence (mIF) data, BAP1-mutant tumors (n = 3) showed a suggestive trend toward higher overall CD8⁺ density and nominally significant increases in PD1⁺ CD8⁺ T cell density and the CD8⁺PD1⁺/CD8⁺ ratio across intratumoral (IT) and tumor-stroma interface (TSI) regions (Fig. 5C and table S5). These findings were consistent when restricting to ICB-exposed tumors and when including a non-ccRCC BAP1-mutant case (fig. S5, K and L, and table S5). Representative regions of interest (ROIs) were selected to illustrate staining patterns, and, in both lymph node and lung metastases, BAP1-mutant tumors showed stronger IT CD8⁺ and PD1⁺ signals compared to WT (Fig. 5D).

Despite this inflamed milieu, BAP1-mutant tumors are aggressive and do not exhibit improved responses to immunotherapy, prompting us to examine whether immune evasion programs are differentially engaged in BAP1-mutant tumors. Tumor cells from BAP1-mutant lesions showed increased chromatin accessibility at multiple immune checkpoint and evasion loci, including HLA-G, HLA-E, LGALS9, and VSIR (Fig. 5E, fig. S5M, and table S5). Together with the preceding observations, these findings are consistent with a model in which IFN signaling in BAP1-mutant tumors not only recruits and activates immune cells but also induces counterregulatory pathways that enable immune evasion, contributing to an inflamed yet aggressive TME.

DISCUSSION

While recurrent transcriptional programs in ccRCC have been linked to tumor heterogeneity and therapeutic outcomes (17, 18), analogous efforts in the epigenetic space have been limited by small sample size, lack of clinical outcome-linked profiling, and difficulty distinguishing tumor from microenvironmental signals (19, 20). In this study, we integrated newly generated and previously published scATAC-seq data (19, 21, 22) with clinical outcome-linked bulk ATAC-seq cohorts (20, 27) to define recurrent epigenetic programs in tumor cells. This framework uncovered shared chromatin programs across patients that were associated with both disease variation and clinical outcomes in ccRCC, providing additional insight into the role of epigenetic dysregulation in this cancer.

Applying this approach, we identified four recurrent chromatin programs in ccRCC tumor cells: a kidney differentiation program (C1), an angiogenesis-related program (C2), an IFN response program (C3), and an epigenetically quiescent state (C0). Among these, the IFN-high C3 program emerged as clinically relevant, being enriched in tumors with aggressive features, such as BAP1 mutation, and associated with shorter DFS. This observation is thematically consistent with prior studies linking inflammatory programs to poor prognosis in RCC in other data modalities (20, 21, 38–40). Our work extends these findings by demonstrating that an IFN response program is embedded in tumor cell chromatin, recurred across patients, and carries prognostic relevance.

Through supervised analysis of BAP1-mutant tumors and validation in an isogenic model, we show that BAP1 loss induces a tumor-intrinsic IFN response. Previous studies have associated BAP1 mutation with an inflamed molecular phenotype in RCC (8, 18, 29, 30, 37), but it was unclear whether the observed signal reflected malignant cells or immune infiltration. Our results indicate that tumor cells themselves contribute to this IFN signal. Because IFN signaling in tumor cells has been linked to immune evasion and resistance to immunotherapy in other cancers (41, 42), this program may also provide a mechanistic basis for the aggressive behavior of BAP1-mutant RCC.

ERVs are a well-established trigger of IFN signaling through viral mimicry (43–45), and prior studies have reported elevated immunogenic ERV expression in BAP1-mutant tumors (33). However, those observations were made in bulk RNA-seq, and subsequent work raised the possibility that the signal could be immune cell derived (46). Here, we show that multiple ERVs are linked to IFN signaling within ccRCC tumor cells, including ERV3-16A3_LTR, which we found to be associated with BAP1 mutation across single-cell and bulk chromatin and transcriptomic data. Although results from an isogenic BAP1 model were heterogeneous, suggesting that BAP1 loss may be permissive but not solely sufficient for its activation, these findings implicate ERV dysregulation as a candidate upstream driver of the IFN phenotype observed in BAP1-mutant tumors.

Extending beyond tumor-intrinsic chromatin programs, our analyses indicate that BAP1-mutant tumors exist in a paradoxically inflamed yet immune evasive microenvironment. We observed increased accessibility at chemokine and cytokine loci such as CXCL10, CXCL11, CCL5, and CSF1, consistent with increased cytotoxic and myeloid infiltration observed across scATAC-seq, bulk RNA-seq, and mIF datasets and with prior reports linking BAP1 loss to an inflamed TME (8, 18, 29, 30, 37). These findings suggest that tumor-intrinsic IFN signaling may directly contribute to immune recruitment. At the same time, BAP1-mutant tumors displayed greater accessibility at multiple immune checkpoint and evasion loci, including HLA-G and HLA-E, which have been implicated in immune escape (47–49), as well as LGALS9 and VSIR, which contribute to T cell suppression (50, 51). Together, these results support a model in which persistent tumor-intrinsic IFN signaling promotes immune infiltration while simultaneously engaging counterregulatory pathways, helping to explain why BAP1-mutant tumors are both inflamed and clinically aggressive.

Our study has several limitations. Integration of multiple scATAC-seq datasets, processed under different conditions (fresh versus frozen), introduces technical variability that can influence results. In addition, both the single-cell and bulk ATAC-seq cohorts remain modest in size, which constrains power to identify rarer tumor states, subtle genotype-phenotype associations, and robust survival correlations. These sample size constraints are particularly relevant for analyses of BAP1-mutant tumors, where the number of mutated samples in scATAC-seq and mIF datasets was limited, reducing statistical confidence in linking specific ERVs or immune microenvironment features to genotype. In addition, while our QC strategy was designed to minimize technical noise, any filtering carries the risk of excluding rare or low-quality cells that may reflect true biological heterogeneity. We attempted to mitigate this by confirming the persistence of broad cell types, recurrent tumor states, and orthogonal validation across independent systems but cannot rule out subtle loss of diversity. More broadly, chromatin accessibility is an indirect measure of regulatory activity and does not capture transcriptional or protein-level consequences, limiting the interpretability of the findings. Furthermore, functional validation of the BAP1-IFN axis was limited to a single isogenic BAP1 model, restricting generalizability. In addition, while our chromatin analyses suggest that BAP1-mutant tumor cells may directly promote immune recruitment through chemokine accessibility, we cannot exclude the possibility that this reflects secondary influences from infiltrating immune cells, hypoxia signaling (52), or other microenvironmental factors. Last, although BAP1 is an H2AK119ub deubiquitinase that counteracts Polycomb-mediated repression and might be expected to decrease accessibility when mutated (53–55), we instead observed more regions with increased rather than decreased accessibility in BAP1-mutant tumors, a paradox that underscores the complexity of its regulatory role.

To mitigate these challenges, we implemented stringent per-sample QC to remove doublets and artifacts and harmonized embeddings with sample ID as the covariate before downstream analyses. Where possible in single-cell analyses, measures were summarized at the per-biopsy level or modeled with random effects to avoid single samples or cohorts driving associations. In bulk analyses, cohort effects were explicitly modeled and, where necessary, addressed with batch correction methods. We also confirmed reproducibility by assessing replicate samples from the same biopsy, which showed high concordance of tumor cell accessibility profiles (fig. S5N). Across analyses, we prioritized cross-cohort reproducibility, visualized biopsy-level behavior in supplementary plots, and validated key observations in orthogonal systems including bulk ATAC-seq, bulk RNA-seq, mIF, and an isogenic BAP1 model. These safeguards cannot eliminate all sources of bias, but they substantially reduce the risk that conclusions reflect artifacts of a particular dataset or processing choice.

Future work should further dissect the role of tumor-intrinsic IFN signaling in BAP1-mutant tumors, particularly how it shapes the TME in in vivo models. Orthogonal profiling at the RNA and protein level will be important to confirm whether immune evasion pathways identified at the chromatin level are functionally active and to test whether these mechanisms represent targetable vulnerabilities.

In summary, our study defines recurrent malignant epigenetic programs in ccRCC, including an IFN response program that is linked to BAP1 genotype and worse patient prognosis. We explore potential mechanisms of activation, including direct effects of BAP1 loss and ERV dysregulation, and characterize the BAP1 mutation-associated TME using multiple data modalities. Together, these findings suggest that persistent tumor-intrinsic IFN signaling contributes to the aggressiveness of BAP1-mutant tumors and highlight immune evasion pathways as potential therapeutic targets in this molecular subset.

MATERIALS AND METHODS

Internal patient cohort scATAC-seq and multiome scRNA-ATAC-seq data generation

Human tissue samples were collected after written informed consent obtained under the Institutional Review Board–approved Dana-Farber Cancer Institute research protocol no. 15-349. Nuclei isolation was performed as previously described (56), using low-retention microcentrifuge tubes (Fisher Scientific, Hampton, NH, USA) throughout to minimize nuclei loss. The cohort included both optimal cutting temperature (OCT) compound-embedded and snap-frozen tumor tissues. For OCT-embedded samples, an additional step involved separating the tissue from OCT by carefully excising it with tweezers and scalpels. Tissues were manually dissociated by finely chopping with spring scissors for 10 min, homogenized in TST solution, filtered through either a 35-μm fluorescence activated cell sorting (FACS) tube filter for scATAC-seq or a 30-μm magnetic-activated cell sorting (MACS) SmartStrainer for multiome scRNA-ATAC-seq (Miltenyi Biotec, Germany), and pelleted by centrifugation at 500g for 10 min at 4°C. The nuclei pellet was resuspended in a lysis buffer to permeabilize the nuclei before pelleting by centrifugation for 10 min at 500g at 4°C. The final nuclei pellet was resuspended in 100 μl of 10x Genomics Diluted Nuclei Buffer, and trypan blue–stained nuclei were counted by eye using INCYTO C-Chip Neubauer Improved Disposable Hemacytometers (VWR International Ltd., Radnor, PA, USA).

Approximately 16,000 to 25,000 nuclei per sample were loaded per channel of the Chromium Next GEM Chip for processing on the 10x Chromium Controller (10x Genomics, Pleasanton, CA, USA). For scATAC-seq, transposition and library construction were carried out as per the manufacturer’s instructions (Chromium Next GEM Single Cell ATAC User Guide). For multiome scRNA-ATAC-seq, both transposition and cDNA generation, followed by library construction, were performed according to the Chromium Next GEM Single Cell Multiome ATAC + Gene Expression User Guide (Rev F). Libraries from both methods were normalized and pooled for sequencing on two NovaSeq SP-100 flow cells (Illumina Inc., San Diego, CA, USA).

scATAC-seq data availability

BAM files for the internal RCC cohort were deposited to dbGaP under accession phs002065.v2.p1. For the Wu et al. (21) cohort, FASTQs were obtained through the Cancer Data Service with dbGaP approval (phs001287.v16.p6). For the Long et al. (22) and Yu et al. (19) cohorts, FASTQs were retrieved from the Sequence Read Archive (PRJNA768891 and PRJNA855923).

scATAC-seq FASTQ processing and fragment file generation

For all scATAC-seq and multiome scRNA-ATAC-seq samples (internal and external cohorts), FASTQ files were processed using Cell Ranger ATAC v2.0.0 or Cell Ranger ARC v2.0.0, respectively, for read filtering, alignment, and cell calling. Reads were aligned to the GRCh38 human reference genome (10x reference build: refdata-cellranger-arc-GRCh38-2020-A-2.0.0). The resulting fragment files and fragment index files were used as inputs for downstream analysis with Signac v1.10.0 (57).

Quality control of scATAC-seq data

To ensure the reliability of scATAC-seq data, we implemented an extensive QC process at the individual sample level. First, we generated a filtered fragment file to only include fragments from cell-associated barcodes as determined by Cell Ranger’s cell calling algorithm. Using this fragment file, we identified accessible peaks with MACS2 v2.2.9.1. Count matrices for each sample were constructed by quantifying insertion events within sample-level peak sets for each cell. Doublets were identified with scDblFinder v1.16.0 (58) and removed.

For all remaining cells, we evaluated multiple QC metrics computed in Signac, including the fraction of reads in peaks (FRiP), the fraction overlapping ENCODE blacklist regions (blacklist fraction), the nucleosome signal (mononucleosome to nucleosome-free fragment ratio), the transcription start site (TSS) enrichment score, and the total number of fragments overlapping peaks. To define thresholds, we adopted a per-sample statistical approach on the basis of the median and median absolute deviation (MAD). Cells were excluded if they contained fewer than 3000 fragments overlapping peaks (or <−3 MAD for internal samples) or more than the median of +3 MAD. Likewise, cells with FRiP below the median of −3 MAD, blacklist fraction or nucleosome signal above the median of +3 MAD, or TSS enrichment below the median of −3 MAD were excluded.

During downstream integration and annotation, we further excluded clusters that appeared to represent contaminants, ambiguous cell identities, or artifacts driven by technical metrics. For instance, a small cluster within the T cell compartment showed unexpectedly high accessibility of myeloid markers, consistent with ambient DNA contamination, and was removed. Last, because FRiP was observed to drive some clustering results, we applied an additional sample-level cutoff, excluding cells below the median of −1.25 MAD for FRiP.

scATAC-seq data integration

All QC-passing cells from internal and external cohorts were combined into a single dataset using a shared peak set. This peak set was generated by performing broad cell-type–level peak calling with MACS2 on the internal scATAC-seq samples. Individual count matrices were then constructed and these merged into a single matrix.

Normalization and dimensionality reduction were carried out in Signac using latent semantic indexing (LSI; RunTFIDF and RunSVD). Batch effects were corrected using Harmony (59) with sample ID as the covariate. The corrected embeddings were used to construct a nearest-neighbor graph with Seurat v4.3.0, and clustering was performed with the smart local moving (SLM) algorithm.

To facilitate cell-type annotation, we calculated gene activity scores in Signac and copy number profiles with CopyscAT v0.30 (60). On the basis of these annotations, broad cell types were defined and peaks were recalled within each broad type using MACS2. For all downstream analyses, we used this broad cell-type–specific peak set. Count matrices were reconstructed, renormalized, reduced with LSI, and batch corrected with Harmony, yielding the integrated dataset used for subsequent analyses.

Cell-type annotation

Within the integrated dataset, we used the 2nd through 30th Harmony-adjusted LSI components to perform clustering with Seurat (SLM algorithm, resolution of 0.1). Clusters were annotated into broad lineages (lymphoid, myeloid, stromal, and tumor) on the basis of differential gene activity score analysis, inspection of insertion events at known marker genes, and copy number estimates derived from CopyscAT using default parameters.

To achieve more granular annotation, we iteratively subsetted cells within each lineage and cell type, renormalized and performed dimensionality reduction using LSI, and reclustered across a range of resolutions and LSI component numbers. These finer clusters were assigned to specific cell types or states on the basis of a combination of differential gene activity score analysis, TF motif enrichment, and copy number profiles.

Mutation and clinical annotation of scATAC-seq cohorts

Mutation and clinical data were assembled for all scATAC-seq cohorts, combining internal and published datasets. For the internal cohort, mutational profiling was performed using whole-exome sequencing (WES) or panel sequencing via OncoPanel where additional tissue was available. Briefly, exome capture was performed using the Illumina’s Rapid Capture Exome Kit; reads were aligned to the hg19 reference using BWA v0.5.9 (61); and variants were called, filtered, and annotated using the Broad Institute’s CGA WES Characterization pipeline (https://app.terra.bio/#workflows/getzlab/CGA_WES_Characterization_Pipeline_v0.1_Dec2018/). This pipeline incorporates multiple tools, including MuTect (62), ContEst (63), Strelka (64), Orientation Bias Filter (65), DeTiN (66), AllelicCapSeg (67), MAFPoNFilter (68), RealignmentFilter, ABSOLUTE (69), GATK (70), PicardTools (71, 72), Variant Effect Predictor (73), and Oncotator (74).

Mutational data for the Wu et al. (21) cohort were obtained from the Genomic Data Commons (GDC) (75) as mutation annotation format (MAF) files. When possible, we selected the MAF corresponding to the same sample ID as the scATAC-seq specimen; otherwise, we used a MAF from another specimen from the same patient. For the Yu et al. (19) and Long et al. (22) cohorts, mutation data were extracted from the supplemental tables of the original publications.

Across all scATAC-seq cohorts, pLOF mutations were defined as frameshift insertions or deletions, nonsense mutations (stop gained), nonstop mutations (loss of a stop codon), start-lost events, splice-site mutations (splice donor or acceptor), or other truncating alterations. Mutational status was determined at the bulk tumor level and then mapped back to the corresponding scATAC-seq samples. As such, when we refer to “BAP1-mutant tumor cells” or similar phrases in the text, this indicates tumor cells derived from a patient tumor carrying a BAP1 pLOF mutation as determined by bulk sequencing. We did not perform mutation calling at the single-cell level. Clinical data for the Wu et al. (21) and Yu et al. (19) cohorts were obtained from their respective supplemental tables, while clinical data for the Long et al. (22) cohort were compiled from both the supplementary materials and additional information was provided by the study authors.

Differential gene accessibility between nontumor broad cell types

DAGs across broad nontumor cell types were identified using Seurat’s FindAllMarkers function with the logistic regression test, adjusting for sequencing depth (latent.vars = nCount_ATAC). Cell identities were downsampled to a maximum of 1000 cells per type (max.cells.per.ident = 1000). Genes were required to be detected in at least 10% of cells in either of the two populations (min.pct = 0.1) and to meet a minimum log₂ fold-change threshold of 0.1. Statistical significance was defined at Bonferroni-adjusted P < 0.05.

Identification of RCC tumor states

To identify RCC tumor states, we subsetted the integrated dataset to RCC tumor cells. Within this subset, data were renormalized, and dimensionality reduction was performed using LSI. The resulting LSI components were batch corrected with Harmony, using sample ID as the covariate. Harmony-adjusted components 2 to 50 were used to construct a nearest-neighbor graph, and clustering was performed with Seurat (SLM algorithm, resolution of 0.1).

Annotation of tumor states was performed iteratively. Initial clusters were evaluated using differential gene accessibility analyses, and clusters with similar accessibility profiles were annotated as the same state. After final states were defined, differential gene accessibility analysis was repeated to identify marker genes for each state (FindMarkers, logistic regression test adjusting for sequencing depth). The number of cells was capped at 1000 per state (max.cells.per.ident = 1000). Genes were required to be accessible in at least 20% of cells in a given state (min.pct = 0.2), with a minimum log₂ fold-change threshold of 0.1. Statistical significance was defined as Bonferroni-adjusted P < 0.05.

Identification and characterization of shared ccRCC tumor states

To focus on epigenetic programs in ccRCC tumor cells, we subsetted the tumor cells to those with ccRCC histology and included in the ccRCC-balanced annotation. Within this subset, data were renormalized, and dimensionality reduction was performed using LSI. The resulting LSI components were batch corrected with Harmony (sample ID as covariate), and Harmony-adjusted components 2 to 20 were used to construct a nearest-neighbor graph. Clustering was performed with Seurat (SLM algorithm, resolution of 0.2).

We advanced four shared tumor states (C0 to C3) for downstream analyses. Highly sample-specific clusters were excluded. Differential peak accessibility between states was assessed with Seurat FindMarkers using the logistic regression test, adjusting for sequencing depth. The number of cells per state was capped at 1000 (max.cells.per.ident = 1000). Peaks were required to be accessible in at least 10% of cells within a state (min.pct = 0.1) and to meet a minimum log₂ fold-change threshold of 0.1. Statistical significance was defined at Bonferroni-adjusted P < 0.05.

Significant peak sets defined from ccRCC tumor states were converted to hg19 genomic coordinates using the rtracklayer liftOver function. Pathway enrichment was then performed with GREAT v3.0.0 using default parameters (25). Enriched terms were required to meet the following criteria: region fold enrichment of >2, at least two associated genes, and q < 0.05 [binomial test, false discovery rate (FDR) adjusted].

TF motif enrichment was assessed using the Signac FindMotifs function with default parameters, which applies a hypergeometric test with FDR correction. Background peaks were defined as those accessible in at least 10% of all cells across the comparison groups. To control for nucleotide composition, background peaks were further restricted to match the GC content distribution of the test peak set using the Signac MatchRegionStats function with default parameters. TF footprinting was performed with the Signac Footprint function using default parameters.

Association of epigenetic modifier mutations with ccRCC tumor programs

To assess whether ccRCC tumor programs were associated with specific epigenetic modifier mutations, we performed separate analyses for BAP1, PBRM1, SETD2, and KDM5C. For each gene, ccRCC tumor cells were subset to those with either a pLOF mutation in that gene or WT status. Program signature scores (calculated as described in the following section) were summarized as per-biopsy medians, and these values were used as the unit of comparison. For each mutation-program pair, per-biopsy median scores were compared between mutant and WT groups using two-sided Wilcoxon rank sum tests, with P values adjusted across the 12 comparisons using the FDR method. Effect sizes were estimated with Cliff’s delta (R package effsize v0.8.1).

scATAC-seq peak set scoring

Cells were scored for accessibility of significant peak sets using a background-subtracted approach. For each peak set, the average normalized counts were calculated per cell by summing the normalized counts across all peaks in the set and dividing by the total number of peaks. A GC-matched reference peak set was then constructed from peaks open in at least 10% of cells (Signac AccessiblePeaks function), restricted to match the GC content distribution of the test set (Signac MatchRegionStats function). Scores for the reference set were calculated in the same manner, and the final peak set score for each cell was defined as the difference between the test and reference values.

Bulk ATAC-seq data acquisition and processing

For the Fukagawa et al. (20) cohort, raw FASTQ files for ccRCC samples were obtained from the European Genome–phenome Archive (EGAD00001009868). Reads were processed using the ENCODE-DCC ATAC-seq pipeline (v2.2.2; https://github.com/ENCODE-DCC/atac-seq-pipeline) with the hg38 reference genome (gs://encode-pipeline-genome-data/genome_tsv/v4/hg38.terra.tsv). Libraries were set as paired-end and adapters automatically detected, and all other pipeline parameters were left at default settings. Mutation and clinical annotations were obtained from the supplementary data accompanying the study, and tumor purity estimates were provided directly by the authors.

For the TCGA KIRC cohort, aligned ATAC-seq BAM files were downloaded from the GDC. Reads were converted to FASTQ format and processed with the ENCODE-DCC ATAC-seq pipeline using identical parameters as for the Fukagawa et al. (20) cohort. Clinical annotations were downloaded from cBioPortal (www.cbioportal.org/study/summary?id=kirc_tcga_pan_can_atlas_2018) (23, 76–78), and somatic mutation calls were obtained from the same resource. Tumor purity estimates were obtained from the supplementary data of Aran et al. (79). Mutations were classified using the same pLOF definition as in the scATAC-seq cohorts, described in the “Mutation and clinical annotation of scATAC-seq cohorts” section.

Bulk ATAC-seq signature scoring

To project ccRCC tumor programs (C1 to C3) into the bulk ATAC-seq cohorts, samples were quantified over predefined peak regions using bedtools coverage v2.28.0. A consensus reference peak set was generated for the Fukagawa et al. (20) cohort by merging sample-level ENCODE-DCC ATAC-seq pipeline-filtered MACS2 calls, retaining peaks present in ≥3 samples, 20 to 10,000 base pairs in width, and located on standard chromosomes. Tumor-state peak sets (C1 to C3) defined from scATAC-seq were appended to this reference, and the combined set was used for quantification of both Fukagawa et al. and TCGA KIRC cohort samples.

Signature scores for each tumor-state peak set were then calculated using the same background-subtracted approach as in the scATAC-seq analysis. For each set, the average counts per million (CPM)-normalized coverage was computed across peaks and compared to a GC-matched background of 10,000 regions (generated with Signac MatchRegionStats function, excluding peaks within 1 kb of the test set). The final score for each sample was defined as the difference between the test and background values.

Bulk ATAC-seq tumor program survival analysis

To mitigate cohort-specific differences in score ranges, tumor-state signature scores were converted to relative proportions per sample [e.g., C1/(C1 + C2 + C3)]. All continuous covariates were scaled before modeling. Associations between tumor-state signatures and OS and DFS were evaluated using Cox proportional hazards models (coxph function, survival R package v3.8-3). Models included sex, age, stage (early versus late), tumor purity, sequencing depth (total fragments in peaks), and cohort as covariates. Kaplan-Meier survival curves were generated (survfit function, survival R package), and statistical significance was evaluated using a log-rank test and univariable Cox proportional hazards model.

Peak-level characterization of BAP1 mutation-associated epigenetic effects in tumor cells (scATAC-seq)

Within the ccRCC-balanced tumor cell subset, we compared chromatin accessibility between BAP1 pLOF mutant and WT tumor cells, restricting the analysis to advanced disease samples. Differential peak accessibility was assessed using Seurat FindMarkers with logistic regression, adjusting for sequencing depth and stage. Peaks were required to be accessible in at least 10% of cells in either group (min.pct = 0.1) and to meet a minimum log₂ fold-change threshold of 0.1. Significance was defined as Bonferroni-adjusted P < 0.05, and peaks with log₂ fold-change of >1.5 (in either direction) were retained for downstream analyses. Pathway enrichment and TF motif enrichment were performed as described above in the “Identification and characterization of shared ccRCC tumor states” section.

RNA-seq data acquisition and analysis of isogenic BAP1 model

RNA-seq FASTQ files for two BAP1 KO monoclonal 786-O cell lines and one isogenic WT control were obtained from the study of Chen et al. (31). FASTQs were aligned to the hg19 reference genome using STAR (80) v2.7.0f (parameters: --outFilterMultimapNmax 20 --outFilterMismatchNmax 999 --outFilterMismatchNoverReadLmax 0.04 --alignIntronMin 20 --alignIntronMax 1250000 --alignMatesGapMax 1250000). Sequencing primers were removed with cutadapt (81), and gene expression was quantified using RNA-Seq by Expectation-Maximization (RSEM) (82) v1.3.1.

Differential expression analysis was performed with DESeq2 v1.42.1. Genes were then ranked by log₂ fold change (KO versus WT) and used as input to fgsea v1.35.0. Enrichment was tested against the Molecular Signatures Database (MSigDB) Hallmark gene sets (category H) and a tumor-cell IFN-stimulated gene (ISG) signature defined by Bi et al. (32). Specifically, this ISG signature included genes that the study found to be significantly induced by type I IFN in tumor cells (β > 0, FDR q value of <0.01) and annotated in their “putative ISG” category.

Proteomic data acquisition and analysis of isogenic BAP1 model

Proteomic data were obtained from the PRIDE database (accession PXD012288) and analyzed following the differential expression procedures described by Chen et al. (31). Differentially expressed proteins (DEPs) in both directions were identified and analyzed for pathway enrichment using the clusterProfiler v4.10.1 enricher function, with the background universe defined as all proteins detected in the experiment. Pathway enrichment was assessed against the same pathway collections used for RNA-seq analyses described in the “RNA-seq data acquisition and analysis of isogenic BAP1 model” section.

Nontargeting CRISPR-Cas9 control in 786-O cells

Because CRISPR-Cas9 editing itself can transiently induce an IFN response (83–85), we performed a nontargeting CRISPR-Cas9 control in the 786-O cell line. 786-O human RCC cells were a generous gift from S. Viswanathan laboratory (Dana-Farber Cancer Institute) and were maintained in RPMI 1640 medium (Gibco, 11775-093) supplemented with 10% fetal bovine serum (Gibco, A5670701) and 1% penicillin-streptomycin (Gibco,15140-122). Cells were cultured in a humidified incubator at 37°C with 5% CO₂ and passaged every 3 to 4 days.

For CRISPR-Cas9–mediated editing, 786-O cells were transfected with pSpCase9 BB-2A–green fluorescent protein (GFP) PX458 (GeneScript, no. SC1823) using Lipofectamine Reagent p3000 (Invitrogen, L3000001) in CTS Opti-MEM (Gibco, A41248). At 48 hours posttransfection, GFP-positive cells were isolated by FACS. We acknowledge the Hematologic Neoplasia Flow Cytometry Core at Dana-Farber Cancer Institute for technical support and instrumentation used in this study. Cells were subsequently seeding for harvesting at days 2, 6, 10, and 12.

Total RNA was isolated using the RNeasy Mini Kit (QIAGEN, 74106), and cDNA was synthesized using a High-Capacity cDNA RT kit (Applied Biosystems, 436814), according to the manufacturer’s protocol. Quantitative polymerase chain reaction was performed using the PowerUp SYBR Master Mix (Applied Biosystems, A25742) on a 7500 Real-Time Fast System (Applied Biosystems). Primers were purchased from Integrated DNA Technologies. Primers for target genes are listed in Table 1. Relative gene expression was calculated using the ΔΔCT method, normalizing to ꞵ-actin as the housekeeping control.

Table 1. Primers used in quantitative polymerase chain reaction.

ACTB, beta-actin; IRF1, interferon regulatory factors 1; ISG15, IFN-stimulated gene 15; MX1, MX dynamin-like GTPase 1; OAS1, 2′-5′-oligoadenylate synthetase 1; STAT1, signal transducer and activator of transcription 1.

Primer name	Forward primer, 5′-3′	Reverse primer, 5′-3′
ACTB	CTCTTCCAGCCTTCCTTCCT	AGCACTGTGTTGGCGTACAG
IRF1	ACCCTGGCTAGAGATGCAGA	TGCTTTGTATCGGCCTGTGT
ISG15	GGTGGACAAATGCGACGAAC	TCGAAGGTCAGCCAGAACAG
MX1	CAGCTCAGGGGCTTTGGAAT	CCTTGGAATGGTGGCTGGAT
OAS1	ACAGGAAACTTGGGTGGTGG	GGTCTCATCGTCTGCACTGT
STAT1	GAGTTGATTTCTGTGTCTGAAG	ATGAGAAGGAAAACTGTCGC

Multiome scRNA-ATAC-seq RNA quality control

Ambient RNA was removed with CellBender remove-background (v0.2.0; 150 epochs, 25,000 droplets, “full” model), and multiplets were detected with Scrublet (v0.2.1, expected doublet rate of 0.1) and excluded. Cells with ≤200 detected genes or ≥5% mitochondrial transcripts were also excluded. Counts were normalized using Seurat’s log normalization.

Deriving and characterizing IFN-associated peak sets

To identify IFN-associated peak sets, we leveraged four multiome scRNA-ATAC-seq samples, retaining putative tumor cells that passed QC for both modalities. IFN signaling was quantified from RNA using CytoSig (86), with raw counts exported from Seurat via DropletUtils v1.23.0 function write10xCounts. To identify chromatin regions associated with IFN activity, we tested 80,768 peaks (accessible in >5% of tumor cells) using mixed-effects models (lme4 v1.1-32). For each peak, accessibility was modeled as a function of IFN1 or IFNG CytoSig score and sequencing depth, with sample ID as a random effect. Peaks were considered IFN-associated if the IFN coefficient was positive and the nominal P value of <0.05. Pathway enrichment and TF motif enrichment were performed as described in the “Identification and characterization of shared ccRCC tumor states” section.

TE quantification and processing in scATAC-seq

Position-sorted BAM files from the Cell Ranger ATAC and ARC pipelines were resorted by read name using Picard’s SortSam. TE accessibility was quantified using an adapted version of scTE (https://github.com/sabrinacamp2/scTE), modified to include a maximum fragment length filter for paired-end alignments. TE count matrices were generated with the following command:

scTEATAC -CB CB -UMI False -p 6 -i ${bam} --expect-cells 20000 -x ../hg38.te.atac.idx -o ${sample_id}.

To retain only high-quality cells, TE counts were subset to barcodes that passed peak-based QC in the integrated scATAC dataset. The resulting cell x TE matrix was combined with cell-type metadata and converted into a Seurat object (minimum of 10 cells per TE, minimum of 200 TEs per cell). Counts were log normalized, scaled, and reduced by principal components analysis.

Identification of IFN-associated ERVs

We restricted the analysis to ccRCC tumor cells assigned to the shared epigenetic states (C0 to C3) with quantified TE accessibility. Each cell was scored for IFN1 and IFNG signaling using the derived IFN-associated peak sets and the background-subtracted scoring method described in the “scATAC-seq peak set scoring” section (here, reference peaks required only ≥5% accessibility).

TE features to test were limited to ERVs by filtering for features containing “HERV” or “ERV” and requiring nonzero accessibility in more than 10 cells, resulting in 58 ERVs. For each ERV and each IFN measure (IFN1 and IFNG), we fit mixed-effects models with the lme4 package lmer function, using IFN score as the outcome and ERV accessibility, sequencing depth, and sample ID as predictors (sample ID as a random effect). ERVs were considered IFN associated if they had a positive coefficient for both IFN1 and IFNG and a FDR q value of <0.05.

Association of IFN-associated ERVs with BAP1 mutation status (scATAC-seq)

We tested whether IFN-associated ERVs were enriched in BAP1-mutant tumors. Analyses were restricted to late-stage ccRCC tumor cells (states C0 to C3) and to biopsies classified as BAP1 pLOF or BAP1 WT. For each ERV feature, accessibility was aggregated to the per-biopsy median and compared between BAP1 groups using a one-sided Wilcoxon rank sum test (alternative: pLOF > WT). Effect sizes were estimated with Cliff’s delta and 95% confidence intervals. P values were adjusted across ERVs using the FDR method; significance was defined as FDR q < 0.05.

Association of ERV3-16A3_LTR accessibility and BAP1 mutation status (bulk ATAC-seq)

Aligned BAM files from the ENCODE-DCC ATAC-seq pipeline were quantified over TE features. TE count matrices were generated with the following command using the adapted version of scTE (https://github.com/sabrinacamp2/scTE): scTEATAC -CB False -UMI False -p 1 -i ${bam} -x hg38.te.atac.idx --min_counts 0 -o ${sample_id}. Counts were log₂ CPM normalized, and batch effects across cohorts were corrected with ComBat (sva v3.50.0) (87), specifying cohort as the batch variable and adjusting for BAP1 status, age, sex, stage group, and tumor purity.

The relationship between ERV3-16A3_LTR accessibility and BAP1 mutation status was assessed using both one-sided Wilcoxon rank sum tests (alternative: pLOF > WT) and multivariable linear models. Multivariable models included BAP1 status, age, sex, stage (early versus late), tumor purity, sequencing depth, and cohort as covariates. Analyses were performed on the combined Fukagawa et al. and TCGA KIRC cohorts.

Bulk RNA-seq data acquisition and processing

For the TCGA KIRC cohort, hg38-aligned RNA-seq BAM files were downloaded from the GDC. Clinical annotations and somatic mutation calls were obtained from cBioPortal (www.cbioportal.org/study/summary?id=kirc_tcga_pan_can_atlas_2018) (23, 76–78). Tumor purity estimates are obtained from the supplementary data of Aran et al. (79). Mutations were classified using the pLOF definition as in the scATAC-seq cohorts.

For the JAVELIN Renal 101 cohort, paired-end RNA-seq data were obtained from published results (88, 89) and realigned to the hg19 reference genome using STAR v2.7.0f (parameters: --outFilterMultimapNmax 20 --outFilterMismatchNmax 999 --outFilterMismatchNoverReadLmax 0.04 --alignIntronMin 20 --alignIntronMax 1250000 --alignMatesGapMax 1250000). WES data for JAVELIN Renal 101 were obtained from the same publications and processed following the same pipeline as for internal cohort WES data in the “Mutation and clinical annotation of scATAC-seq cohorts” section. Mutations were classified using the same pLOF definition as in the scATAC-seq cohorts, described in the “Mutation and clinical annotation of scATAC-seq cohorts” section. Clinical data for JAVELIN Renal 101 were obtained from the published studies.

Association of ERV3-16A3_LTR expression and BAP1 mutation status (bulk RNA-seq)

We assessed ERV3-16A3_LTR expression in two patient cohorts (TCGA KIRC and JAVELIN Renal 101) and in the isogenic BAP1 model of Chen et al. (31). Bulk RNA-seq BAM files were quantified for TE expression using the adapted version of scTE (https://github.com/sabrinacamp2/scTE). The following command was used to generate TE count matrices: scTE -i ${bam} -o ${sample_id} -x /scTE/${genome}.exclusive.idx -CB False -UMI False --hdf5 False. The genome index matched the alignment reference for each dataset (hg38 for TCGA KIRC, hg19 for JAVELIN Renal 101, and hg19 for isogenic BAP1 model). Counts were log₂ CPM normalized, with library size defined as the total counts per sample. Associations between ERV3-16A3_LTR expression and BAP1 mutation status were evaluated using both one-sided Wilcoxon rank sum tests (alternative: pLOF > WT) and multivariable linear models. In TCGA KIRC, the multivariable model included BAP1 status, age, sex, clinical stage, history of neoadjuvant treatment, tumor purity, and library size. In JAVELIN Renal 101, models included BAP1 status, age, sex, tumor location, tumor purity, and library size.

Differential gene accessibility analysis between BAP1 mutated and WT tumor cells (scATAC-seq)

DAGs were identified by comparing BAP1 pLOF and WT tumor cells within advanced-stage ccRCC, restricted to cells in the ccRCC-balanced tumor states (C0 to C3). Analyses were performed using the Seurat FindAllMarkers function with logistic regression, adjusting for sequencing depth. Genes were required to be detected in at least 5% of cells in either group (min.pct = 0.05) and to meet a log₂ fold-change threshold of 0.05. Significance was defined as Bonferroni-adjusted P < 0.05.

Immune attractant and immune checkpoint and evasion genes

Following differential gene accessibility analysis between BAP1 pLOF and WT tumor cells, we specifically examined predefined gene panels for immune cell attractants and immune checkpoint/evasion genes. The immune cell attractant gene set included CXCL9, CXCL10, CXCL11, CXCL12, CCL5, CCL2, CCL4, CCL20, CCL17, CCL22, and CSF1. The checkpoint/evasion gene set included CD274, PDCD1LG2, C10orf54 (VSIR/VISTA), LGALS9, MIF, VSIG4, HLA-E, NECTIN2, PVR, CD276, NT5E, ENTPD1, VTCN1, HLA-G, and IDO1.

mIF analysis

Preprocessed mIF data for ccRCC were obtained from the Dana-Farber ImmunoProfile platform via OncDRS (90, 91). For cases annotated only as “RCC” without a specified subtype, pathology reports were manually reviewed to confirm ccRCC. BAP1 mutation status was derived from OncoPanel sequencing of matched patients, with pLOF mutations as defined in the scATAC-seq cohorts. For the BAP1-mutant tumors analyzed by mIF, mutation profiling was performed on a primary tumor sample rather than the metastatic biopsy used for mIF. The variant allele fraction was consistent with expectations for a clonal event (0.5× tumor purity), supporting the assumption that the mutation was also present in the metastatic biopsy analyzed by mIF. One non-ccRCC case harbored a BAP1 structural variant and had OncoPanel and mIF performed on the same specimen; this case was only included in a supplementary analysis.

ICB exposure at the time of biopsy was manually annotated by chart review. Six immune features provided by the ImmunoProfile pipeline (90, 91) were analyzed: CD8⁺ cell density in IT and TSI regions, CD8⁺PD1⁺ cell density in IT and TSI, and the ratios of CD8⁺PD1⁺/CD8⁺ cells in each region. Outliers were excluded if any of the four density features exceeded five times the interquartile range beyond Q1 or Q3. Associations between immune features and BAP1 mutation status were tested using one-sided Wilcoxon rank sum tests (alternative: pLOF > WT), with P values adjusted for multiple testing using the FDR method.

Representative ROIs for two BAP1-mutant and two BAP1-WT tumors were chosen for visualization. We selected matched biopsy site samples (one lung and one lymph node per genotype) and restricted selection to ROIs annotated as IT only. All selected samples were ICB exposed. ROIs were visualized using QuPath v0.6.0, with consistent colors assigned to 4′,6-diamidino-2-phenylindole (DAPI), paired box 8 (PAX8), cluster of differentiation 8 (CD8), and programmed cell death protein 1 (PD1). Minimum and maximum intensity values were adjusted manually on a per-sample basis.

Bulk RNA-seq gene quantification, cell-type deconvolution, and association with BAP1 mutation status

For the TCGA KIRC cohort, RNA-seq expression data were downloaded from the UCSC Xena Browser (92) as log₂[transcripts per million (TPM) + 1] values. These were converted back to TPM before analysis [TPM = 2^log₂^(TPM+1) − 1]. Ensembl identifiers were mapped to gene symbols using the GENCODE v36 annotation.

For the JAVELIN Renal 101 cohort, RNA-seq data were aligned with STAR as described in the “Bulk RNA-seq data acquisition and processing” section. For CIBERSORTx analysis, aligned reads were further processed to obtain gene-level TPM estimates: Sequencing primers were removed with cutadapt, and gene expression was quantified using RSEM.

Immune cell proportions were estimated using CIBERSORTx Fractions (93). Briefly, we used the scRNA-seq reference described by Bi et al. (17) as the refsample, provided by study authors. Nonlogged TPM matrices for TCGA KIRC and JAVELIN Renal 101 were supplied as mixture files, with S-mode batch correction, no quantile normalization, and one permutation.

Analyses were restricted to tumors with BAP1 pLOF mutations or WT status. Associations between immune cell fractions and BAP1 mutation status were tested using multivariable linear models. In TCGA KIRC, models included BAP1 mutation status, age, sex, clinical stage, and history of neoadjuvant treatment. In JAVELIN Renal 101, models included BAP1 mutation status, age, sex, and tumor location. Immune subsets tested were CD8⁺ T cells, TAMs, and monocytes.

Within-biopsy reproducibility analysis

To assess intrabiopsy consistency of chromatin accessibility, we restricted the analysis to ccRCC-balanced tumor cells from biopsies with at least two sequenced libraries. For each library, accessibility was aggregated to a sample-level profile by averaging the normalized ATAC signal across all cells. Peaks were retained if accessible in 2 to 98% of samples. Pairwise Pearson correlations were then computed between the resulting sample-level profiles (fig. S5N).

Statistical analysis

Statistical tests are provided in the figure legends and in the corresponding Methods subsections for each analysis.

Supplementary Materials

The PDF file includes:

Figs. S1 to S5

Legends for tables S1 to S5

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S5