Chromatin topology and distal elements underlie divergent cell-type-specific regulation of 9p21 locus cell cycle genes Authors: Elena Torlai Triglia, Tyler E. Miller, Neva C. Durand, Kyung Lock Kim, Salvador Casaní-Galdón, Joshua D’Antonio, Nauman Javed, Alessandro Paz Hernandez, Ayush Raman, Anna J. Cruz, Robbyn Issner, Abhishek Choudhary, François Aguet, Julia Verga, Michelle Chan, Varshini Ramanathan, Carman Man-Chung Li, Vivian Hecht, Anders S. Hansen, Martin J. Aryee, Charles B. Epstein, David R. Kelley, Fadi J. Najm, Noam Shoresh, Kristin Ardlie, David G. Hendrickson, J. Graham Ruby, Bradley E. Bernstein, Elizabeth Gaskell Published: 2026 ABSTRACT SummaryThe 9p21 locus encodes critical cell cycle regulators p16INK4A, p14ARF, and p15INK4B and contains a high density of SNPs associated with aging-related diseases. Progress toward understanding 9p21 regulatory mechanisms has been constrained by complex transcript structures and a shortage of cell models that retain physiologic locus control. We innovated tools to deconvolve the expression and regulation of 9p21 transcripts in single cells. We find that 9p21 transcripts exhibit highly divergent expression across tissues and cell types, with p15INK4B predominating. Despite being frequently conflated under the term CDKN2A, p14ARF and p16INK4A are independently expressed across tissues, cell types, aging, and in a fibroblast model of replicative senescence. We annotate a network of cis-regulatory elements (CREs), whose activities and interactions underpin these complex regulatory dynamics. Our systematic characterization of 9p21 transcripts and CREs provides mechanistic insights and the tools for future studies of this vital locus. Introduction The 9p21 locus encodes critical cell cycle regulators and tumor suppressors p16INK4A, p14ARF, and p15INK4B and a long noncoding RNA, ANRIL (Figure 1A). Deletion or silencing of 9p21 genes occurs in the majority of human cancers,1 and the locus harbors a high density of single-nucleotide polymorphisms (SNPs) associated with common diseases—including coronary artery disease (CAD), type 2 diabetes, germline cancer risk, and overall lifespan.2,3,4,5,6,7 Further, p16INK4A is an established marker of replicative senescence (RS) and aging and has garnered attention as a potential target for senolytic therapies.8,9,10 Despite the clear importance of 9p21 for human health, the mechanisms controlling expression of 9p21 transcripts are poorly understood, hampering efforts to both interpret disease risk variants and cancer-associated genome alterations and identify the cell states in which these genetic variants manifest.Figure 1Tissue-specific expression of 9p21 transcripts(A) Left: approximate genomic organization of the 9p21 transcripts—p16INK4A and p14ARF (CDKN2A), p15INK4B (CDKN2B), and the long noncoding RNA ANRIL (CDKN2B-AS1)—and expected results from various RNA-seq strategies. Right: reported functions of each protein encoded by the 9p21 cell cycle regulators.(B) Heatmap shows average transcript per million (TPM) per tissue site; RNA-seq data are from GTEx v.8 (n ≥ 70 donors per tissue). Tissues are ordered left to right by a p14ARF-to-p16INK4A ratio. Rows 1–2: isoform-level expression estimates for p16INK4A and p14ARF. Rows 3–6: gene-level expression estimates of CDKN2A, p15INK4B(CDKN2B), ANRIL (CDKN2B-AS1), and MTAP.(C) Scatterplots show the correlation (Pearson R) between TPM values for p16INK4A, p14ARF, or the sum of the two, with CDKN2A from two GTEx tissues at the extremes of the` p14ARF-to-p16INK4A ratio in (B). Each dot is a single donor sample.(D and E) UpSet plots of tissues from GTEx (D) and FANTOM (E) expressing one or more of the 9p21 transcripts. Threshold is ≥1 TPM using the average across donors for a given tissue in each dataset.(F) Violin plots show log expression (natural logarithm of TPM +1) of 9p21 transcripts across GTEx tissues grouped by donor age for 4 representative tissues. Each dot is a single donor sample. The white dot represents the median value. Adipose - visceral (omentum): n [20–30] = 49, n (60–70] = 169. Esophagus-gastroesophageal junction: n [20–30] = 36, n (60–70] = 97. Thyroid: n [20–30] = 53, n (60–70] = 219. Whole blood: n [20–30] = 73, n (60–70] = 253. ∗∗∗Bonferroni adjusted p value (adj p) < 0.001, ∗∗adj p < 0.01, and ∗adj p < 0.05. Tissue-specific expression of 9p21 transcripts (A) Left: approximate genomic organization of the 9p21 transcripts—p16INK4A and p14ARF (CDKN2A), p15INK4B (CDKN2B), and the long noncoding RNA ANRIL (CDKN2B-AS1)—and expected results from various RNA-seq strategies. Right: reported functions of each protein encoded by the 9p21 cell cycle regulators. (B) Heatmap shows average transcript per million (TPM) per tissue site; RNA-seq data are from GTEx v.8 (n ≥ 70 donors per tissue). Tissues are ordered left to right by a p14ARF-to-p16INK4A ratio. Rows 1–2: isoform-level expression estimates for p16INK4A and p14ARF. Rows 3–6: gene-level expression estimates of CDKN2A, p15INK4B(CDKN2B), ANRIL (CDKN2B-AS1), and MTAP. (C) Scatterplots show the correlation (Pearson R) between TPM values for p16INK4A, p14ARF, or the sum of the two, with CDKN2A from two GTEx tissues at the extremes of the` p14ARF-to-p16INK4A ratio in (B). Each dot is a single donor sample. (D and E) UpSet plots of tissues from GTEx (D) and FANTOM (E) expressing one or more of the 9p21 transcripts. Threshold is ≥1 TPM using the average across donors for a given tissue in each dataset. (F) Violin plots show log expression (natural logarithm of TPM +1) of 9p21 transcripts across GTEx tissues grouped by donor age for 4 representative tissues. Each dot is a single donor sample. The white dot represents the median value. Adipose - visceral (omentum): n [20–30] = 49, n (60–70] = 169. Esophagus-gastroesophageal junction: n [20–30] = 36, n (60–70] = 97. Thyroid: n [20–30] = 53, n (60–70] = 219. Whole blood: n [20–30] = 73, n (60–70] = 253. ∗∗∗Bonferroni adjusted p value (adj p) < 0.001, ∗∗adj p < 0.01, and ∗adj p < 0.05. The structure of the 9p21 transcripts is complex. p16INK4A and p14ARF (p19ARF in mice) differ in their first exon only, are annotated as a single gene (CDKN2A), and are largely indistinguishable in 3′ biased RNA sequencing (RNA-seq) data and single-cell RNA-seq (scRNA-seq) data (Figures 1A, S1A, and S1D–S1F).11 Consequently, despite PCR evidence of their divergent expression in bulk samples,12,13,14,15,16 p16INK4A and p14ARF are generally quantified jointly as CDKN2A in RNA-seq and scRNA-seq gene expression studies, confounding interpretation of expression dynamics across cell states and aging.10,11,17,18,19,20,21 Despite this overlap in their coding sequences, a frameshift in the distinct first exon results in the p16INK4A and p14ARF transcripts encoding entirely distinct protein products, with distinct biochemical and cellular functions.1,22,23 In contrast, the p15INK4B (CDKN2B) and p16INK4A proteins share ∼85% of their amino acid sequence and both function as cyclin-dependent kinase inhibitors. Adding to the complexity, the 9p21 locus also encodes the noncoding RNA ANRIL (also known as CDKN2B-AS1), which overlaps the p15INK4B transcript, and two unrelated genes: methylthioadenosine phosphorylase (MTAP) and doublesex- and Mab-3-related transcription factor A1 (DMRTA1). Several studies have painstakingly dissected cis-regulatory mechanisms of the 9p21 transcripts. Chromatin state mapping and reporter assays in cancer and primary cells identified enhancer-like sequences across 9p21.2,24,25,26 Studies of locus topology found that the cell cycle genes and MTAP are organized within a topologically associated domain (TAD) and identified multiple physical interactions between candidate cis-regulatory element (CRE) regions and promoters within the TAD, as well as to the neighboring TAD encoding the interferon (IFN) gene cluster.2,16,24 More recently, CRISPR perturbation studies in cancer cell lines have identified an element within the p14ARF coding region that represses p16INK4A (B-cell acute lymphoblastic leukemia, B-ALL cells)26 and a network of four CREs that, when deleted, reduce expression of all cell cycle genes— p16INK4A, p14ARF, and p15INK4B plus ANRIL24,25 (HeLa cells). Expression of all 9p21 genes is also controlled by methylation of histone H3 at lysine 27 (H3K27me3) as directed by the polycomb repressive complex proteins.24,25,27,28,29 Finally, ANRIL expression has been implicated in both repression and activation at this locus.24,25,27,30,31 Although these and other studies have expanded our understanding of 9p21 gene regulatory mechanisms, the complexities of the CDKN2A transcript structure, complex locus organization, and low transcript expression levels have confounded efforts to understand the cell-type-specific regulation and causal roles of the respective transcripts. For example, studies of cell-type-specific regulation and dynamics, as well as noncoding CRISPR screens with Perturb-seq, are hindered by the shared 3′ mRNA sequence between the CDKN2A transcripts p16INK4A and p14ARF. Furthermore, the 9p21 cell cycle transcripts are expressed at extremely low levels in non-immortalized cells in the absence of RS. This has necessitated screening CREs in cancer-derived or otherwise immortalized cell lines that have either lost portions of the locus or, conversely, express the genes at supraphysiologic levels due to loss or inhibition of downstream pathway components (e.g., Todd et al.32 and Collado et al.33). As such, we lack an understanding of the expression dynamics of 9p21 transcripts across tissues and cell types and their joint or independent regulation in a non-immortalized setting, including in vivo contexts where cells are subject to normal cell cycle control mechanisms and microenvironmental cues. To address these striking knowledge gaps, we integrated epigenomic and gene expression maps, high-resolution locus topology maps, CRISPRa Perturb-seq, and computational modeling to investigate gene regulatory mechanisms at 9p21. We show that p15INK4B is the predominant transcript across cell types and tissues. By deconvolving the CDKN2A transcripts (p16INK4A and p14ARF) in expression profiles, we find strong evidence that they are independently regulated in vivo and in RS. In immune cells, p14ARF appears to be the dominant CDKN2A transcript, while p16INK4A is often undetectable in vivo, with the exception of tissue fibroblasts. These data caution against using CDKN2A as a proxy for either gene. Indeed, we find that locus expression switches from p14ARF to p16INK4A and p15INK4B as cultured fibroblasts move from a cycling to a senescent state. In support of this striking independence among 9p21 transcripts, CRISPRa perturbations revealed that each promoter specifically regulates the coincident transcript but not others, with the exception of a bidirectional promoter shared by p14ARF and ANRIL. CRISPRa Perturb-seq, together with Region Capture Micro-C (RCMC), further nominated a minimum of 9 distal CREs that regulate one or more 9p21 transcripts. 8 of these distal CREs have highly conserved sequences and relatively few SNPs and are likely driven by an AP-1 transcription factor (TF) program. In contrast, one element forms a CTCF-CTCF loop suggestive of a structural role and specifically regulates p16INK4A and p15INK4B expression. We find that the CREs themselves do not change in accessibility as cells accumulate cell divisions. Instead, we find a general decompaction of locus topology specific to the cell cycle gene region of the 9p21 TAD, most likely enabling access and activation of p16INK4A and p15INK4B. Taken together, these data establish a detailed map of 9p21 gene regulation and CREs across human tissues and cell types and provide the tools to propel future studies of this vital locus. Results Discordant expression of CDKN2A (p16INK4A and p14ARF), CDKN2B (p15INK4B), ANRIL, and MTAP across tissues We first sought to understand the tissue expression dynamics and potential co-regulation of the 9p21 genes. While genes proximal along the linear chromosome are commonly co-expressed,34,35,36,37 complex loci, including functionally related genes, can exhibit finely-tuned and distinctly regulated expression controlled by networks of cell-type/state-specific CREs.38 Previous reports on the mRNA expression relationships between the 9p21 cell cycle genes—particularly at single-cell resolution—have been contradictory, often relying on CDKN2A as a proxy for p16INK4A expression.10,17,18,19,20,21,25,39 We systematically mapped expression of 9p21 transcripts across human tissues and individuals using the GTEx resource (n = 17,382 samples, from 948 donors and 54 tissue types)40 (Figures 1B–1D and 1F). We deconvolved and quantified the expression of p14ARF (ENST00000579755) and p16INK4A (ENST00000304494) by leveraging the exceptionally deep full-length mRNA-seq data (GTEx v.8). We find that the ratio of the CDKN2A transcripts, p14ARF or p16INK4A, varies across tissues and donors (Figures 1B, 1C, S1B, and S1C). CDKN2A expression is therefore an inaccurate proxy for either p14ARF or p16INK4A in most tissue samples (Figures 1B, 1C, S1B, and S1C). No 9p21 transcript pairs correlate across all tissues (Figures 1B–1D and S2). However, interesting tissue-specific patterns emerge. For example, p15INK4B and ANRIL are both highly expressed specifically in gut tissues, though ANRIL is not detected in many other tissues where p15INK4B is highly expressed (Figures 1B and S2). We replicated these findings in a second cohort, leveraging 5′ enriched CAGE data from FANTOM (Cap Analysis of Gene Expression) (n = 188 samples from 101 locations, 1–4 samples per tissue type)41 (Figures 1E; STAR Methods). In this dataset, we also note that several tissues express p15INK4B alongside p14ARF (Figure 1E). Thus, despite their proximity on the linear chromosome, coincident exons, and related cell cycle control functions, the 9p21 cell cycle transcripts have divergent expression patterns across tissues. p16INK4A is reported to increase in several tissues during aging in rodents and humans.8,42,43,44,45,46 p19ARF, the murine homolog of p14ARF, also increases with age in some tissues.43 We comprehensively examined the dynamics of 9p21 transcripts across human tissues with age in the GTEx cohort (age 20–70) (Figures 1F and S3). We find that p16INK4A expression increases from the youngest to the oldest donors across most tissues we examined. Whole blood is a notable exception: expression of p14ARF and p15INK4B differs significantly between the youngest and oldest GTEx donors, but p16INK4A does not. These data reinforce that tissue context should be considered when using p16INK4A as a biomarker of aging. Non-uniform clearance of senescent cells across tissues may also confound interpretation. In some tissue contexts, alternative 9p21 transcripts may be more informative. Deconvolution of CDKN2A to p16INK4A and p14ARF in single-cell gene expression data We next sought to determine 9p21 transcript expression at cell-type resolution. The unique portions of p16INK4A and p14ARF (exon1α or exon1β, respectively) are not detectable in gene expression data generated from commonly used single-cell technologies, which have a 3′ bias due to the final steps of the library preparation (Figures 1A and S1D–S1F). We therefore took advantage of 10× 5′ scRNA-seq data, in which reads represent the transcript-specific first exon. Typical RNA-seq analysis estimates gene-level expression abundances, which aggregate mapped reads to the gene annotation CDKN2A without differentiating p14ARF and p16INK4A. We therefore comprehensively reviewed human transcript annotations for p14ARF and p16INK4A and selected the major isoform for each gene, using GTEx data and literature evidence. We then designed a custom alignment pipeline that separately quantifies p14ARF and p16INK4A (Figure 2A; STAR Methods).Figure 2Cell-type-specific expression of 9p21 transcripts(A) Schematic of 9p21 cell cycle genes and their isoforms (GENCODE v.34). Major p14ARF and p16INK4A isoforms (CDKN2A-214 or ENST00000579755.1 and CDKN2A-201 or ENST00000304494.9, respectively) were used to design a custom gene transfer format (GTF) to annotate p16INK4A and p14ARF from 5′ single-cell gene expression profiles. A subset of isoforms of ANRIL and the protein-coding isoforms of CDKN2B are shown.(B) Heatmap shows percentages (%) of cells annotated in each cell type from human tissues expressing the 9p21 cell cycle genes p16INK4A, p14ARF, or p15INK4B or ANRIL. 10× 5′ data from He et al.47 analyzed with the custom GTF from (A) (cutoff for expression: unique molecular identifier [UMI] > 0). Dotted-line boxes highlight (1) p15INK4B and ANRIL expression in gut epithelial cells, (2) p14ARF expression in immune cells, and (3) tissue fibroblasts expressing p16INK4A, p14ARF, and p15INK4B.(C–E) Barplots show the percentage of each cell type expressing the 9p21 genes or ANRIL. 10× 5′ data analyzed with the custom GTF from (B). Expression cutoff UMI > 0. Error bars: standard error of the mean (SEM) between donor percentages. Number of donors are shown above the plot.(C) Adult gut tissues from Elmentaite et al.48 Top: all cell types; bottom: immune cells only.(D) Heart tissues from Koenig et al.49 Top: all cell types; bottom: immune cells only.(E) Immune cells. Top: 10× Genomics 5′ v.2 peripheral blood mononuclear cell (PBMC) example data from the manufacturer. Bottom: 10× 5′ data from CD4+ T cells from Hashimoto et al.50 Cell-type-specific expression of 9p21 transcripts (A) Schematic of 9p21 cell cycle genes and their isoforms (GENCODE v.34). Major p14ARF and p16INK4A isoforms (CDKN2A-214 or ENST00000579755.1 and CDKN2A-201 or ENST00000304494.9, respectively) were used to design a custom gene transfer format (GTF) to annotate p16INK4A and p14ARF from 5′ single-cell gene expression profiles. A subset of isoforms of ANRIL and the protein-coding isoforms of CDKN2B are shown. (B) Heatmap shows percentages (%) of cells annotated in each cell type from human tissues expressing the 9p21 cell cycle genes p16INK4A, p14ARF, or p15INK4B or ANRIL. 10× 5′ data from He et al.47 analyzed with the custom GTF from (A) (cutoff for expression: unique molecular identifier [UMI] > 0). Dotted-line boxes highlight (1) p15INK4B and ANRIL expression in gut epithelial cells, (2) p14ARF expression in immune cells, and (3) tissue fibroblasts expressing p16INK4A, p14ARF, and p15INK4B. (C–E) Barplots show the percentage of each cell type expressing the 9p21 genes or ANRIL. 10× 5′ data analyzed with the custom GTF from (B). Expression cutoff UMI > 0. Error bars: standard error of the mean (SEM) between donor percentages. Number of donors are shown above the plot. (C) Adult gut tissues from Elmentaite et al.48 Top: all cell types; bottom: immune cells only. (D) Heart tissues from Koenig et al.49 Top: all cell types; bottom: immune cells only. (E) Immune cells. Top: 10× Genomics 5′ v.2 peripheral blood mononuclear cell (PBMC) example data from the manufacturer. Bottom: 10× 5′ data from CD4+ T cells from Hashimoto et al.50 We applied our pipeline to analyze a 5′ human whole-body atlas from a single individual (84,363 cells, 15 organs).47 Clear patterns of divergent expression between p14ARF and p16INK4A emerge in the single-cell data beyond observations in bulk. Notably, p14ARF—but not p16INK4A—expression is enriched in the B and T cell lineages (Figures 2B and S4). Conversely, tissue fibroblasts account for a substantial majority of p16INK4A-expressing cells across tissues. We also observe p16INK4A expression in melanocytes, as expected.51 Consistent with the bulk tissue observations above, p15INK4B and ANRIL are both expressed (1.35%–32.4%) in epithelial cell types of the small intestine (tuft cells, absorptive cells, and enterocytes). These cells do not typically express p14ARF or p16INK4A. We confirmed these patterns of expression in four additional 10× 5′ scRNA-seq multi-donor tissue studies, including (1) multiple anatomical regions of the human gut (n = 5 adult donors, ∼82,000 single cells)48 (Figures 2C and S4), (2) left ventricular cardiac tissue (n = 1 healthy donor and n = 4 donors with cardiomyopathy, ∼35,000 single cells)49 (Figures 2D and S5), and (3) peripheral blood mononuclear cells (PBMCs) (n = 1 donor, ∼7,000 cells, and n = 2 donors, ∼9,000 cells)50 (Figures 2E and S5). Remarkably, we also find that the vast majority of CDKN2A expression in both tissue-resident and peripheral immune cells corresponds to p14ARF, with minor contributions from p16INK4A (Figures 2C–2E). This has implications for the interpretation of CDKN2A eQTLs from PBMCs.40,52,53 5′ scRNA-seq methods have been applied in limited settings. We therefore sought to distinguish p16INK4A and p14ARF in 3′ scRNA-seq libraries by adapting a PCR-based enrichment strategy based on our previous work54 (Figures 3A, S6, and S7). We amplified exon1α (p16INK4A) or exon1β (p14ARF), together with the cell barcode and unique molecular identifier (UMI), from the full-length cDNA created in 3′ scRNA-seq protocols (known as whole-transcriptome amplification [WTA]). We use the cell barcode to associate exon1α or exon1β UMIs with each cell, enabling sensitive detection and accurate quantification of p16INK4A and p14ARF (Figures 3A, S6, and S7). This enrichment method can be applied to existing scRNA-seq libraries where cDNA has been stored at the WTA stage. We applied the enrichment protocol to a published 3′ scRNA-seq library from mammary tissue (Figure 3B).55 p16INK4A and p14ARF were indistinguishable in the data prior to enrichment, despite evidence of CDKN2A expression in 5% of cells (Figure 3C). After applying our exon1α/1β enrichment protocol, we detected p16INK4A and p14ARF in 27% and 32% of cells, respectively. We confirmed the specificity of p16INK4A/p14ARF enrichment by comparing the percentage of cells expressing p16INK4A and p14ARF with cells expressing CDKN2A from the original scRNA-seq library (Figures 3D and S8). Consistent with observations from the 5′ data above (Figures 2B–2E), p14ARF expression was predominant in immune cells. In contrast, p16INK4A expression was predominant in specific classes of fibroblast and basal/myoepithelial (BAP) cells (Figure 3D).Figure 3Enrichment of p16INK4A and p14ARF from 3′ single-cell libraries reveals cell-type-specific expression differences(A) Overview schematic of the protocol for targeted enrichment of p16INK4A and p14ARF (exon1α/exon1β) from scRNA-seq libraries; protocol was adapted from our previous work.54 CB, cell barcode; UMI, unique molecular identifier.(B) Uniform manifold approximation and projection (UMAP) shows 3′ 10× data from selected breast tissue samples from Gray et al. n = 10 donors from which we performed targeted exon1α/exon1β enrichment. Cluster annotations are as in Gray et al.55. F1-3, fibroblasts; BAP1-3, basal/myoepithelial cells; AP, alveolar progenitor cells; BL, basal-luminal alveolar cells; HSa, hormone sensing alpha cells; HSb, hormone sensing beta cells; LE, lymphatic endothelial cells; M, myeloid cells; NK, NK cells; P, plasma cells; PE, pericytes/smooth muscle cells; T, T cells; and VE, vascular endothelial cells.(C) Barplot shows (bars 1–3): the percent (%) of cells from the samples in (B) expressing p16INK4A (estimated from reads mapping to exon1α), p14ARF (estimated from reads mapping to exon1β), or CDKN2A from the published data without enrichment, as well as (bars 4 and 5) p16INK4Aand p14ARF after exon1α/exon1β enrichment using the protocol from (A). Error bars are standard error between the mean percentage of cells per sample/donor, SEM.(D) Dot plot shows cell types from (B) ordered left to right by p14ARF to p16INK4A ratio. p16INK4A and p14ARF expression determined by the exon1α/exon1β enrichment protocol from (A), normalized by the total number of counts per cell. CDKN2A, MTAP, ANRIL, and p15INK4B (CDKN2B) expression from Gray et al.55 Enrichment of p16INK4A and p14ARF from 3′ single-cell libraries reveals cell-type-specific expression differences (A) Overview schematic of the protocol for targeted enrichment of p16INK4A and p14ARF (exon1α/exon1β) from scRNA-seq libraries; protocol was adapted from our previous work.54 CB, cell barcode; UMI, unique molecular identifier. (B) Uniform manifold approximation and projection (UMAP) shows 3′ 10× data from selected breast tissue samples from Gray et al. n = 10 donors from which we performed targeted exon1α/exon1β enrichment. Cluster annotations are as in Gray et al.55. F1-3, fibroblasts; BAP1-3, basal/myoepithelial cells; AP, alveolar progenitor cells; BL, basal-luminal alveolar cells; HSa, hormone sensing alpha cells; HSb, hormone sensing beta cells; LE, lymphatic endothelial cells; M, myeloid cells; NK, NK cells; P, plasma cells; PE, pericytes/smooth muscle cells; T, T cells; and VE, vascular endothelial cells. (C) Barplot shows (bars 1–3): the percent (%) of cells from the samples in (B) expressing p16INK4A (estimated from reads mapping to exon1α), p14ARF (estimated from reads mapping to exon1β), or CDKN2A from the published data without enrichment, as well as (bars 4 and 5) p16INK4Aand p14ARF after exon1α/exon1β enrichment using the protocol from (A). Error bars are standard error between the mean percentage of cells per sample/donor, SEM. (D) Dot plot shows cell types from (B) ordered left to right by p14ARF to p16INK4A ratio. p16INK4A and p14ARF expression determined by the exon1α/exon1β enrichment protocol from (A), normalized by the total number of counts per cell. CDKN2A, MTAP, ANRIL, and p15INK4B (CDKN2B) expression from Gray et al.55 In summary, we provide computational and experimental protocols for deconvolving p14ARF and p16INK4A expression in scRNA-seq experiments. We find that p14ARF and p16INK4A expression is divergent across tissues and cell types and caution against using CDKN2A as a proxy for either p16INK4A or p14ARF. Moreover, the 9p21 cell cycle genes are not co-expressed at bulk or single-cell resolution but rather show exquisitely cell-type-specific expression patterns. A switch from p14ARF during cycling to p16INK4A and p15INK4B at senescence We next sought to understand 9p21 transcript dynamics and their relationship to cellular programs. Because the inherent sparsity of single-cell data and the limited number of cells in each state in complex tissues renders it difficult to calculate robust associations, we turned to an in vitro model. The non-immortalized lung fibroblast line WI-38 robustly upregulates CDKN2A as cells approach replicative senescence (RS).17 We deployed our targeted amplification method to enrich and quantify p14ARF and p16INK4A transcripts from 3′ scRNA-seq libraries from Chan et al.17 for WI-38 cells at low (population doubling level [PDL]20) and high (PDL50) passage (Figure 4A and 4B). We validated that these cells had reached RS by applying the SenMayo and senescence-associated secretory phenotype (SASP) scores56 (Figures 4B, 4C, 4G, and S9). We next quantified p14ARF and p16INK4A expression in each cell (Figure 4D). Decline of p14ARF expression has been detected previously at the bulk level using RT-PCR/ddPCR in both replicative and oncogene-induced senescence (OIS).12,13,14,15,16 We confirmed that at the single-cell level, p14ARF expression is higher in the cells with lower replicative age and decreases with replicative age. In contrast, p16INK4A and p15INK4B expression followed the opposite trend (Figures 4D and 4E). Gene set enrichment analysis confirmed that genes correlated with p16INK4A are enriched for senescence-related, TGF-β, and extracellular matrix programs, typical of senescent cells in culture (Figure 4F).56 Genes correlated with p15INK4B share enrichment with TGF-β and extracellular matrix programs but are not enriched for other typical senescence programs. Conversely, genes correlated with p14ARF are enriched for cell-cycle programs and DNA replication.Figure 4A switch from p14ARF to p16INK4A and p15INK4B across fibroblast replicative age(A) Experimental design for the WI-38 fibroblast replicative senescence (RS) experiment from Chan et al.17(B) UMAP projection shows WI-38 cells from the RS time course from Chan et al.17 Increasing population doubling levels (PDLs) are colored from early (red) to late (blue) time points as in (A).(C) The UMAP projection from (B) colored by Senmayo score.(D) The same UMAP projection from (B), colored by expression of each of the 9p21 cell cycle genes and ANRIL. CDKN2A, p15INK4B, and ANRIL expression determined from the scRNA-seq data from Chan et al.17 p16INK4A and p14ARF expression determined by exon1α/exon1β enrichment protocol, normalised as in Figure 3D. Maximum value capped at 99th percentile.(E) Barplot shows the ratio of p14ARF to p16INK4A, as determined by the exon1α/exon1β enrichment protocol, across the RS time course.(F) Enriched ontologies among the top 200 genes most correlated (by PearsonR) with p14ARF, p16INK4A, or p15INK4B from EnrichR.57(G) Heatmap shows expression of marker genes and expression signatures from cells in the main cluster in (B)–(D). Cells arranged on pseudotime from cycling to senescent states as in Chan et al.17(H) SenMayo, SASP, and scores from genes most correlated with p14ARF, p16INK4A (from targeted enrichment), or p15INK4B (from original expression profiles) in cells as in (G). A switch from p14ARF to p16INK4A and p15INK4B across fibroblast replicative age (A) Experimental design for the WI-38 fibroblast replicative senescence (RS) experiment from Chan et al.17 (B) UMAP projection shows WI-38 cells from the RS time course from Chan et al.17 Increasing population doubling levels (PDLs) are colored from early (red) to late (blue) time points as in (A). (C) The UMAP projection from (B) colored by Senmayo score. (D) The same UMAP projection from (B), colored by expression of each of the 9p21 cell cycle genes and ANRIL. CDKN2A, p15INK4B, and ANRIL expression determined from the scRNA-seq data from Chan et al.17 p16INK4A and p14ARF expression determined by exon1α/exon1β enrichment protocol, normalised as in Figure 3D. Maximum value capped at 99th percentile. (E) Barplot shows the ratio of p14ARF to p16INK4A, as determined by the exon1α/exon1β enrichment protocol, across the RS time course. (F) Enriched ontologies among the top 200 genes most correlated (by PearsonR) with p14ARF, p16INK4A, or p15INK4B from EnrichR.57 (G) Heatmap shows expression of marker genes and expression signatures from cells in the main cluster in (B)–(D). Cells arranged on pseudotime from cycling to senescent states as in Chan et al.17 (H) SenMayo, SASP, and scores from genes most correlated with p14ARF, p16INK4A (from targeted enrichment), or p15INK4B (from original expression profiles) in cells as in (G). We next aligned these cells along our previously modeled pseudotime axis with proliferating cells as the trajectory “root”17 (Figure S9). While p16INK4A expression dramatically increases as cells reach RS, p14ARF expression is relatively uniform while cells are cycling and dramatically decreases at RS (Figures 4D, 4E, 4G, 4H, S9B, and S9C). p15INK4B expression also increases across the time course. In concordance with findings in vivo, we detect very little ANRIL in these fibroblasts (Figure 4D). Confirming the switch from cycling to senescent cells across pseudotime, we observe a concomitant increase in SenMayo and SASP,56 CDKN1A (p21), and a decrease in cell cycle markers. Hence, as WI-38 cells transition from a cycling to a senescent state, there is a distinct switch from p14ARF to p16INK4A and p15INK4B (Figures 4D–4H). This switch is also evident when viewing the data by PDL (Figure S9C). This switch is not apparent when measuring CDKN2A alone. These data strongly suggest that the 9p21 cell cycle genes are not coordinately regulated. Instead, these data are consistent with distinct promoters and cell-state-specific CREs, including for the CDKN2A transcripts p14ARF and/or p16INK4A. Our data also confirm at the single-cell level that in a fibroblast model, RS is associated with an increase in p15INK4B and p16INK4A and a dramatic reduction in p14ARF. p16INK4A, p14ARF, p15INK4B, ANRIL, and MTAP are encompassed in a single ∼500 kb TAD To explore the cis-regulatory mechanisms driving expression of the 9p21 genes, we next annotated topological and epigenomic features of the locus. Regulatory interactions between genes and CREs primarily occur between elements in the same TAD and may be distinguished as sub-TAD-level 3D contacts.58,59,60,61 We therefore utilized RCMC62 to map high-resolution 3D contacts across a 2.3 Mb region encompassing 9p21 and the adjacent IFN gene cluster (Figures 5A and 5B). We acquired these contact maps and also profiled, with chromatin immunoprecipitation (ChIP)-seq, the insulator protein CTCF and H3K27ac (a marker of active CREs) in lower-replicative-age (PDL21-33) and higher-replicative-age (PDL48-54) WI-38 (Figures 5B, S10, and S11). We determined contacts to 200 bp resolution, a 5-fold improvement over existing Hi-C datasets.63,64Figure 5Genomic and epigenomic landscape of 9p21(A) Region Capture Micro-C (RCMC) data of chr9:20,300,000–22,600,000 from WI-38 fibroblasts. Data were combined across all PDLs (Table S2). Contacts are shown at a 5 kb resolution.(B) RCMC data of chr9:21,750,000–22,500,000, zoomed in from (A). Contacts are shown at a 1 kb resolution. The remaining tracks from top to bottom: (1) eQTLs for CDKN2A, CDKN2B, ANRIL, and MTAP from selected GTEx v.10 tissues. Dot size represents posterior inclusion probability (PIP); (2 and 3) CTCF ChIP-seq, (4 and 5) H3K27ac ChIP-seq, and (6 and 7) ATAC-seq from WI-38 across replicative age (tracks are counts per million [CPM] normalized) (ATAC-seq data from Chan et al.,17 PDL20 and PDL50 WI-38, H3K27ac, and CTCF ChIP-seq were generated for this study from PDL33 and PDL54 WI-38); (8) candidate cis-regulatory elements (cCREs) from ENCODE v.365; (9) regions selected for CRISPRa screening; and (10) gene track, modified by annotating the CDKN2A isoform ENST00000304494.9 as p16INK4A and the ENST00000579755.1 isoform as p14ARF. All tracks were visualized in IGV.66,67 Putative TAD boundary insulator sites marked by strong CTCF ChIP-seq peaks are annotated in red boxes.(C) Summary of the regions selected for CRISPRa screening. TSS, transcription start site. Genomic and epigenomic landscape of 9p21 (A) Region Capture Micro-C (RCMC) data of chr9:20,300,000–22,600,000 from WI-38 fibroblasts. Data were combined across all PDLs (Table S2). Contacts are shown at a 5 kb resolution. (B) RCMC data of chr9:21,750,000–22,500,000, zoomed in from (A). Contacts are shown at a 1 kb resolution. The remaining tracks from top to bottom: (1) eQTLs for CDKN2A, CDKN2B, ANRIL, and MTAP from selected GTEx v.10 tissues. Dot size represents posterior inclusion probability (PIP); (2 and 3) CTCF ChIP-seq, (4 and 5) H3K27ac ChIP-seq, and (6 and 7) ATAC-seq from WI-38 across replicative age (tracks are counts per million [CPM] normalized) (ATAC-seq data from Chan et al.,17 PDL20 and PDL50 WI-38, H3K27ac, and CTCF ChIP-seq were generated for this study from PDL33 and PDL54 WI-38); (8) candidate cis-regulatory elements (cCREs) from ENCODE v.365; (9) regions selected for CRISPRa screening; and (10) gene track, modified by annotating the CDKN2A isoform ENST00000304494.9 as p16INK4A and the ENST00000579755.1 isoform as p14ARF. All tracks were visualized in IGV.66,67 Putative TAD boundary insulator sites marked by strong CTCF ChIP-seq peaks are annotated in red boxes. (C) Summary of the regions selected for CRISPRa screening. TSS, transcription start site. Consistent with prior lower-resolution studies in IMR-90 and HeLa cells,24,68 we annotate a TAD that harbors the 9p21 cell cycle regulators and MTAP. We observe multiple strong CTCF peaks at sites corresponding to the TAD boundaries in the RCMC data (Figure 5B). In both lower- and higher-replicative-age WI-38, the 9p21 TAD that includes p16INK4A, p14ARF, p15INK4B, ANRIL, and MTAP is distinct from neighboring TADs harboring the IFN cluster or DMRTA1 (Figure 5A). The higher resolution maps also reveal sub-TAD regions and loops (Figures 5A and 5B). MTAP and the CDKN2A/B cluster (p16INK4A, p14ARF, p15INK4B, and ANRIL) are partitioned into separate sub-regions of the TAD, suggesting they may be accessed by different CREs. Consistent with this hypothesis, bulk-tissue expression quantitative trait loci (eQTLs) for MTAP and the CDKN2A/B genes partition into distinct regions of the TAD (Figure 5B). Hence, the primary topological structure encompassing the cell cycle gene cluster in WI-38 is ∼500 kb in size and likely insulated from MTAP, the IFN cluster, and other genetic elements in the chromosomal region. Functional assessment of 9p21 promoters and distal CREs Previous studies of 9p21 cis regulation have either focused on promoters or annotated regulatory regions in cancer cell lines that do not undergo RS.14,24,26,69 Here, we sought to identify CREs for the respective 9p21 transcripts in the WI-38 model of RS using a CRISPRa screen, which we reasoned would be more sensitive for detecting lowly expressed transcripts. We utilized assay for transposase-accessible chromatin using sequencing (ATAC-seq) data for the same WI-38 RS time course as above17 (Figure 5B) to nominate 109 putatively “active” CREs, the majority of which have an H3K27ac signal (Figures 5B and S11). We supplemented this set with candidate CREs annotated in other cell types,65,70 yielding a final set of 173 potential CREs in the 9p21 locus (Figure 5C; STAR Methods). We cloned a library of 724 gRNAs against these target sites into a CROP-seq vector that enables gRNA and transcriptional readout by scRNA-seq71,72 (STAR Methods). We transfected the library pool at a high MOI into WI-38 cells stably expressing Cas9 fused to transactivator VP64 (CRISPRa73) and overexpressing the MS2-P65-HSF1 activator helper complex. We harvested cells 7 days after transfection, and read out transcriptional changes by scRNA-seq (Figure 6A) and a modified version of our p16INK4A/p14ARF enrichment that also captures the gRNA. We ultimately obtained three sets of measurements for 89,732 cells after quality control (QC), at a final MOI of 7.05: whole-transcriptome gene expression (used to determine p15INK4B and ANRIL), expression of p14ARF and p16INK4A from the enrichment protocol, and gRNA counts (Figure 6A). We deconvolved the impact of each gRNA on 9p21 gene expression using SCEPTRE74 (Figures 6B and S12; Table S4). The screen therefore assessed the effect of each gRNA on four target transcripts, for a total of 2,896 gRNA-gene pairs tested.Figure 6CRISPRa functionally links 9p21 CREs to target genes(A) Schematic of the CRISPRa screening workflow.(B) CRISPRa screen results overview, contextualized with CTCF and H3K27ac tracks. From top to bottom: (1) CTCF ChIP-seq track and (2) H3K27ac ChIP-seq, both from WI-38 at PDL54 (as in Figure 5B) (CPM normalized). (3–6) CRISPRa screen results from SCEPTRE: each track represents the −log10(p) for each gRNA-gene pair with a significance of p < 0.05, plotted at the gRNA binding site. Results for each target gene are shown as separate tracks. Expression values for CDKN2B and ANRIL are derived from scRNA-seq libraries and for p16INK4A and p14ARF from the exon1α/exon1β enrichment protocol. (7) Regions targeted by CRISPRa. (8) gRNAs validated from arrayed single gRNA transfection experiments and RT-qPCR. (9) Gene track, modified by annotating the CDKN2A isoform ENST00000304494.9 as p16INK4A and ENST00000579755.1 as p14ARF. All tracks were visualized with IGV.66,67(C) Zoomed-in views on the gRNAs that were taken forward into validation experiments, corresponding to 3 promoter regions and 9 distal CREs (E1–E9). Each section represents a 2–4 kb zoomed-in region of the data in (B).(D) RT-qPCR results for validation experiments. Each dot represents the average of 4 replicates: two separate time-series experiments, with cells collected on days 4 and 8 post-transfection. Data were normalized relative to expression of GAPDH and non-targeting control gRNAs. Color of the dot: average fold change. Size of the dot: combined Z score measuring the strength and significance of the effect of the guide on each transcript.(E) Dot plot shows ChromBPNet sequence attribution scores and −log10(q) of motif annotation strength for selected candidate TF family binding site (TFBS) within the promoter and CRE regions from the ChromBPNet PDL50 model. Top 4 rows: transcript(s) regulated by each CRE.(F) Zoomed-in views of E1 and E9. From top to bottom: CTCF and H3K27ac tracks from PDL54, ChromBPNet attribution score from model trained on PDL50 ATAC-seq, FOS ChIP-seq data from IMR-90 (ENCODE: ENCSR124AIG and ENCFF906BPJ), common SNPs (dbSNP155 common SNP track), and PhastCons100 conservation score. CRISPRa functionally links 9p21 CREs to target genes (A) Schematic of the CRISPRa screening workflow. (B) CRISPRa screen results overview, contextualized with CTCF and H3K27ac tracks. From top to bottom: (1) CTCF ChIP-seq track and (2) H3K27ac ChIP-seq, both from WI-38 at PDL54 (as in Figure 5B) (CPM normalized). (3–6) CRISPRa screen results from SCEPTRE: each track represents the −log10(p) for each gRNA-gene pair with a significance of p < 0.05, plotted at the gRNA binding site. Results for each target gene are shown as separate tracks. Expression values for CDKN2B and ANRIL are derived from scRNA-seq libraries and for p16INK4A and p14ARF from the exon1α/exon1β enrichment protocol. (7) Regions targeted by CRISPRa. (8) gRNAs validated from arrayed single gRNA transfection experiments and RT-qPCR. (9) Gene track, modified by annotating the CDKN2A isoform ENST00000304494.9 as p16INK4A and ENST00000579755.1 as p14ARF. All tracks were visualized with IGV.66,67 (C) Zoomed-in views on the gRNAs that were taken forward into validation experiments, corresponding to 3 promoter regions and 9 distal CREs (E1–E9). Each section represents a 2–4 kb zoomed-in region of the data in (B). (D) RT-qPCR results for validation experiments. Each dot represents the average of 4 replicates: two separate time-series experiments, with cells collected on days 4 and 8 post-transfection. Data were normalized relative to expression of GAPDH and non-targeting control gRNAs. Color of the dot: average fold change. Size of the dot: combined Z score measuring the strength and significance of the effect of the guide on each transcript. (E) Dot plot shows ChromBPNet sequence attribution scores and −log10(q) of motif annotation strength for selected candidate TF family binding site (TFBS) within the promoter and CRE regions from the ChromBPNet PDL50 model. Top 4 rows: transcript(s) regulated by each CRE. (F) Zoomed-in views of E1 and E9. From top to bottom: CTCF and H3K27ac tracks from PDL54, ChromBPNet attribution score from model trained on PDL50 ATAC-seq, FOS ChIP-seq data from IMR-90 (ENCODE: ENCSR124AIG and ENCFF906BPJ), common SNPs (dbSNP155 common SNP track), and PhastCons100 conservation score. The CRISPRa screen nominated 342 gRNA-gene pairs with significance of p < 0.05, including predicted gene promoters and distal CREs (Figures 6B and S12; Table S4). We selected 18 of the top-scoring guides across target transcripts for follow-up, added 2 gRNAs to more densely tile one CRE (later annotated as E7), and performed single gRNA transfections followed by RT-qPCR (n = 80 gRNA-gene pairs tested by RT-qPCR, of which 30 were significant in the SCEPTRE analysis, p ≤ 0.05) (Tables S6 and S10; STAR Methods). To assess concordance between the high-MOI and single-transfection experiments, we compared the effects on gene expression between the significant hits from the SCEPTRE analysis of the single-cell screen data and the RT-qPCR. 23 out of 30 total gRNA-gene pairs nominated by SCEPTRE for which we have matched RT-qPCR data have the same direction of effect in both the high-MOI screen and the single gRNA validation experiments, of which 18 were statistically significant in both assays (p ≤ 0.05 SCEPTRE, Z ≤ −1 RT-qPCR) (Table S10). The single gRNA transfections confirmed three promoter regions, with corresponding gRNAs increasing expression of p16INK4A, p14ARF/ANRIL, or p15INK4B (fold change>1, Z ≤ −1) (Figures 6C and 6D; Tables S6 and S10). In addition, these experiments confirmed 9 distal CREs as potential regulators of one or more 9p21 transcripts (annotated as E1–E9). These CREs were located 9 to >200 kb from the annotated promoter of their putative target genes and coincide with peaks of accessibility and/or H3K27ac in WI-38 (Figures 6C and S13). Given the close proximity and overlapping coding sequences of the 9p21 cell cycle genes, we next considered the inter-relationships among their promoters. Targeting CRISPRa to each of the 3 promoter regions robustly induced expression of the corresponding transcript(s) between ∼5- and ∼50-fold (Figures 6C and 6D; Table S6). Two of three gRNAs targeting the promoter between p14ARF and ANRIL induced both transcripts, consistent with prior reports that these genes share a common bidirectional promoter.75,76 Although p16INK4A and p14ARF are both encoded by the CDKN2A gene, their transcripts exhibit entirely distinct cell-type and tissue specificities at single-cell resolution (Figures 2, 3, and 4). This independence is further supported by the perturbations: CRISPRa targeting of the p16INK4A promoter upregulates its transcript but has essentially no impact on p14ARF (Figures 6C and 6D; Table S6). Conversely, targeting of the p14ARF promoter upregulated p14ARF and ANRIL but had no consistent effect on p16INK4A. Although ANRIL has been reported to repress p16INK4A and p15INK4B in other contexts,13,27 CRISPRa induction of ANRIL did not downregulate p16INK4A or p15INK4B expression in our assay (Table S6). CRISPRa targeting of the p15INK4B promoter increased its transcript levels by >50-fold. Yet, despite its location within an intron of ANRIL and its proximity to other cell cycle genes, this strong promoter perturbation did not increase expression of the other 9p21 transcripts (Figures 6C and 6D). These data build on our findings above that the 9p21 transcripts show distinct cell-type-specific patterns of expression in tissues, despite their proximity. In the fibroblast model, each transcript depends on the activity of its own promoter and is largely unaffected by other promoters and transcripts in the cluster. They raise the question of what additional factors, potentially including distal CREs and/or chromatin structures, underlie their differential activities. The 9p21 cell cycle genes are regulated by distal CREs and chromatin organization E1–E9 all reside within the predominant ∼500 kb sub-TAD that harbors the cell cycle genes and ANRIL (Figures 5A and 5B). 4 of the 9 distal CREs regulate more than one transcript (Figures 6C and 6D). The CREs that increase p16INK4A expression have no impact on p14ARF, or vice versa, further supporting their independence. Furthermore, we find that 8 of the 9 CREs increase p15INK4B expression (∼1.4- to 6-fold), consistent with our observation that p15INK4B is the predominant transcript in tissues. Unexpectedly, CREs that increase ANRIL expression do not affect p14ARF, and vice versa, suggesting a degree of independence even between these transcripts sharing a common promoter region. We next sought insight into the regulatory sequences and TFs that underpin the regulation of these CREs and their target transcripts at RS. We trained ChromBPNet on WI-38 ATAC data from PDL20 or PDL50.17,77 ChromBPNet accurately predicted accessibility (Pearson R in peaks: 0.732 for PDL20 and 0.699 for PDL50), including at E1–E9 and the transcript promoters (Figure S13C). Multiple sequences in all CREs and promoters were predicted to have an impact on chromatin accessibility, as quantified by the “attribution score” (Figures 6E, 6F, and S13C; Table S7). Detailed analysis of binding motifs in high-scoring sequences nominated TFs with potential roles in validated CREs, including a strong AP-1-driven program across multiple regions, consistent with the expression and increased motif accessibility of AP-1 TFs during RS17,78,79 (Figure S13A). Nominated TFs were mostly overlapping between PDL20 and PDL50 predictions (Table S7). Supporting these predictions, 7 of the 8 CREs with high-scoring AP-1 motifs display FOS binding in IMR-90 human fibroblasts (Figures 6F and S13C). Intriguingly, E1—the closest CRE to the p16INK4A promoter, a CRE for p16INK4A and p15INK4B, and the least conserved of the CREs in this study—was associated with a distinct set of features, including predicted accessibility-driving motifs for KLF factors, a strong CTCF ChIP-seq signal, and no evidence for FOS binding in IMR-90 (Figures 6E, 6F, S13C, and S14; Table S7). To evaluate the interactions between the 9p21 gene promoters and E1–E9, we next calculated loops across our combined WI-38 RCMC data (Figure 7A; Table S2; STAR Methods). We detected multiple looping interactions within the annotated CDKN2A/B cluster sub-TAD (Figures 5A and 7A). We also noted a loop from the CTCF-bound site in E1 to a strong CTCF-bound site in an MTAP intron. We also calculated “virtual 4C” profiles from the RCMC data, using each promoter and CRE as anchors (Figures 7A and S15). Consistent with the CRISPRa data, the virtual 4C profiles for both the p16INK4A promoter and E1 revealed an increased contact frequency compared to their local surroundings. Also consistent with the CRISPRa data, the p15INK4B promoter makes the most distal contacts with E2, E5, E6, and E9. Notably, E4 contacts the p15INK4B promoter and activates p15INK4B expression in our high-MOI screen but is not validated by single gRNA transfection and RT-qPCR. Unexpectedly, given their observed independence, the p15INK4B and p14ARF/ANRIL promoters also form a striking interaction (Figures 7A and S15). Finally, we find that E1—and the CDKN2A/B cluster promoters—contact a strong CTCF-bound site in the MTAP intron. We hypothesize that E1 potentially functions as an enhancer for p16INK4A and p15INK4B (Figures 6B–6D, 7A, and S15) and/or a structural element creating a boundary between the MTAP and CDKN2A/B sub-TADs (Figures 5B and 7A). Examination of the CTCF motifs in E1 reveals divergent motifs typical of structural CTCF-CTCF loop anchors (Figure 7A). Indeed, previous work identified a CTCF-bound region overlapping E1 in IMR-90 fibroblasts and implicated CTCF in control of p15INK4B and p16INK4A expression.16 Taken together, these results suggest that 9p21 is structured by CTCF-CTCF loops (at least partially anchored by E1 and an intronic CTCF-bound insulator site in an MTAP intron), enhancer-promoter loops, and a p15INK4B-p14ARF/ANRIL promoter-promoter loop.Figure 7Characterization of 9p21 distal CREs(A) 3D genome interactions across 9p21 chr9:21,800,000–22,225,000. From top to bottom: (1) loop plot representing regions of high interactions, identified from RCMC data combined from all WI-38 time points in our RS time course. (2) Arrows represent the presence and orientation of CTCF motifs at the major CTCF peaks. (3–6) ChIP-seq for CTCF and H3K27ac from WI-38 PDL33 or PDL54; tracks are CPM normalized. (7–10) Virtual 4C visualizations of RCMC data from four viewpoints (E1 or p16INK4A, p14ARF/ANRIL, and p15INK4B promoters) at 500 bp resolution. Anchor regions are indicated by the anchor symbol. (11) Track representing the location of validated CREs and promoter regions. (12) Gene track, modified by annotating the CDKN2A isoform ENST00000304494.9 as p16INK4A and the ENST00000579755.1 isoform as p14ARF. The orange-highlighted box marks a CTCF-bound region in MTAP that has strong interactions with E1 (gray-highlighted box). Blue highlight boxes mark the p16INK4A, p14ARF/ANRIL, and p15INK4B promoter regions.(B) Change in 3D genome interactions with replicative age. RCMC data are shown at a 1 kb resolution as a subtraction plot, comparing PDL48 and PDL21 WI-38 fibroblasts. Each RCMC dataset was normalized to total reads in chr9.(C) Fine-mapped GWAS variants (UKBB) for phenotypes associated with common diseases or the diseases hemoglobin A1c, type 2 diabetes, apolipoprotein B, coronary artery disease, pulse pressure, lymphocyte count, albumin/globulin ratio, glaucoma, and IGF1 (insulin-like growth factor 1). Bottom: track representing the location of lead germline SNPs from cancer GWASs, grouped in two sets: one including breast and colorectal cancer, glioma, and leukemia and the other including lung cancer and melanoma. Dot size represents −log10(p). eQTLs are the same as those in Figure 5. Dot size represents posterior inclusion probability (PIP). Linkage disequilibrium (LD) is from UKBB80; LD block annotation was determined using R2 > 0.5.81(D) Heatmap shows the 9p21 WI-38 CREs and promoters annotated in this study as open (blue) or closed (gray) in published single-nucleus ATAC-seq data from human tissues82 that overlap with a validated 9p21 CRE or promoter region and are labeled by the region they overlap with. There are no regions that overlap E2, but there are 2 regions overlapping each of the promoters and E6, shown as separate rows. Transcripts regulated by each CRE are annotated to the right. Bottom 2 annotation rows represent cell-type groups and the organ where each cell type is mostly identified. Characterization of 9p21 distal CREs (A) 3D genome interactions across 9p21 chr9:21,800,000–22,225,000. From top to bottom: (1) loop plot representing regions of high interactions, identified from RCMC data combined from all WI-38 time points in our RS time course. (2) Arrows represent the presence and orientation of CTCF motifs at the major CTCF peaks. (3–6) ChIP-seq for CTCF and H3K27ac from WI-38 PDL33 or PDL54; tracks are CPM normalized. (7–10) Virtual 4C visualizations of RCMC data from four viewpoints (E1 or p16INK4A, p14ARF/ANRIL, and p15INK4B promoters) at 500 bp resolution. Anchor regions are indicated by the anchor symbol. (11) Track representing the location of validated CREs and promoter regions. (12) Gene track, modified by annotating the CDKN2A isoform ENST00000304494.9 as p16INK4A and the ENST00000579755.1 isoform as p14ARF. The orange-highlighted box marks a CTCF-bound region in MTAP that has strong interactions with E1 (gray-highlighted box). Blue highlight boxes mark the p16INK4A, p14ARF/ANRIL, and p15INK4B promoter regions. (B) Change in 3D genome interactions with replicative age. RCMC data are shown at a 1 kb resolution as a subtraction plot, comparing PDL48 and PDL21 WI-38 fibroblasts. Each RCMC dataset was normalized to total reads in chr9. (C) Fine-mapped GWAS variants (UKBB) for phenotypes associated with common diseases or the diseases hemoglobin A1c, type 2 diabetes, apolipoprotein B, coronary artery disease, pulse pressure, lymphocyte count, albumin/globulin ratio, glaucoma, and IGF1 (insulin-like growth factor 1). Bottom: track representing the location of lead germline SNPs from cancer GWASs, grouped in two sets: one including breast and colorectal cancer, glioma, and leukemia and the other including lung cancer and melanoma. Dot size represents −log10(p). eQTLs are the same as those in Figure 5. Dot size represents posterior inclusion probability (PIP). Linkage disequilibrium (LD) is from UKBB80; LD block annotation was determined using R2 > 0.5.81 (D) Heatmap shows the 9p21 WI-38 CREs and promoters annotated in this study as open (blue) or closed (gray) in published single-nucleus ATAC-seq data from human tissues82 that overlap with a validated 9p21 CRE or promoter region and are labeled by the region they overlap with. There are no regions that overlap E2, but there are 2 regions overlapping each of the promoters and E6, shown as separate rows. Transcripts regulated by each CRE are annotated to the right. Bottom 2 annotation rows represent cell-type groups and the organ where each cell type is mostly identified. We next investigated whether these topological interactions shift with replicative age. Virtual 4C profiles from PDL21 and PDL48 WI-38 suggest a modest reduction in contact frequencies with increased replicative age, particularly at the CTCF-CTCF loop anchored by E1 and the p15INK4B-p14ARF promoter-promoter loop (Figure S15). Consistent with this, a broader examination of the overall locus topology suggested a larger-scale shift in 9p21 locus organization (Figures 7B, S16, and S17). We found that contact frequency was reduced in PDL48 across the CDKN2A/B cluster portion of the TAD, consistent with prior reports of decompaction at RS63,64 and reduction in contact frequency at 9p21 in OIS.16 This is in stark contrast to the DMRTA1 and IFN cluster regions, which have increased internal contact frequencies and appear to become more compact (Figures 7B and S17) and a genome-wide increase in CTCF binding at RS in WI-38 (Figure S13B). We speculate that this overall 9p21 locus decompaction accumulates over cell divisions, potentially in complement with changes to chromatin repressor and insulator activities.28,29,64,83 This may contribute to the switch from the expression of p14ARF to a p16INK4A- and p15INK4B-driven RS program. 9p21 CREs span major disease-associated LD blocks and are cell-type specific We next asked if E1–E9 overlapped any common SNPs or eQTLs. We collated fine-mapped variants associated with common diseases computed from the UK Biobank (UKBB; 337,000 British-ancestry individuals)80,84 and lead germline SNPs from cancer genome-wide association studies (GWASs). We also annotated linkage disequilibrium (LD) blocks using the UKBB80,81 (Figure 7C; STAR Methods). E1–E9 span the major LD blocks that harbor most of the common disease and germline cancer risk haplotypes in the 9p21 locus.2,3,4,5,6 Despite this, the specific elements—with the exception of E5—are largely devoid of GWAS SNPs and eQTLs (Figures 7C and S14). This lack of signal may be due to the high sequence conservation and depletion of common variants within the elements. The adjacent associations within the locus may reflect sequences that impact locus conformation or other CREs whose functions are limited to highly specific in vivo contexts. In this vein, we evaluated the chromatin accessibility of the 9p21 promoters and CREs in vivo using an atlas derived from single-nucleus ATAC-seq for human tissues82 (Figure 7D). The atlas aggregates genome-wide accessibility profiles for 111 adult cell types. The 9p21 promoters are broadly accessible, with regions in the p15INK4B promoter displaying accessibility across most profiled cell types. These data are consistent with our observations that p15INK4B is the predominant transcript in bulk and single-cell RNA expression data for tissues (Figures 1B and 2B). In contrast to the promoters, accessibility of the CREs is cell-type specific (Figure 7D), the sole exception being the CTCF-bound structural element, E1, which is more broadly accessible across cell types (Figure 7D). Notably, E1 also overlaps the LD block containing the germline cancer-associated variants, consistent with a role in cell cycle control (Figures 6F and S14). E5 is also notable for containing a CAD risk SNP, and its accessibility appears largely specific to GABAergic neurons (Figure 7D). Taken together, our results suggest that the 9p21 locus integrates a complex network of context-specific CREs with structural elements and topological changes that may accumulate with cell divisions to activate specific transcripts and downstream cell cycle control pathways in an exquisitely context-dependent manner. Discussion Despite their essential functions in cell cycle regulation and associations with polygenic common diseases and cancers, the study of the 9p21 transcripts in humans has been hindered by a lack of tools to parse the complex transcript structure and its regulation.10,18,85 Here, we present molecular and computational tools to overcome technical challenges and deconvolve transcript-specific patterns and mechanisms of gene regulation at 9p21 at single-cell resolution. By deconvolving the two protein-coding genes encoded by CDKN2A, p16INK4A and p14ARF, we show that these transcripts are entirely independent and distinctly regulated across cell types in vivo and in a fibroblast model of RS. Prior studies have noted discordant expression of these transcripts in bulk data from RS, OIS, and cancer.12,13,14,15,16 We extend these findings by demonstrating these dynamics in single cells, tissues and across a dense time course modeling RS. Further, we show that p16INK4A and p14ARF are under the tight control of distinct promoters and CREs, which likely underpin their distinct expression patterns across cell types and states. This is seemingly in contrast to the simpler situation in rodents, where the two genes are largely co-regulated.43,86,87,88 While others in the field have called out the need to separate these genes and their protein products in molecular assays,10 it has not been possible to distinguish between p16INK4A and p14ARF in typical 3′ single-cell expression profiles.10,11,17,18,19,20,21 Our work here further highlights that quantifying CDKN2A as a proxy for either p16INK4A or p14ARF likely masks important biology. It also provides molecular and computational tools that should enable other investigators to deconvolve these cell cycle regulators in other contexts. Clearance of senescent cells, with focus on p16INK4A-expressing cells, has garnered strong interest as a potential therapeutic strategy to extend human healthspan.8,89,90 Our study highlights that p16INK4A does increase with age but in a tissue-specific manner. These data caution that a more complete understanding of cell-type-specific patterns of p16INK4A, among other senescence biomarkers, in tissues over time is needed to underpin any potential senolytic approaches. Our enrichment protocol can now be readily applied to WTA from both existing and new studies to detect p16INK4A-expressing cells. More broadly, we find that the 9p21 cell cycle genes can be activated by 3 independent promoters and at least 9 CREs (E1–E9). These distal CREs are between 9 and 220 kb from their putative target promoters and show a strong preference for subsets of 9p21 transcripts, consistent with their independent cell-type and temporal expression patterns. Confirming their regulatory contributions by loss-of-function approaches remains a critical future goal. A deep learning model indicates that accessibility of distal CREs is underpinned mostly by AP-1 motifs, with a notable exception in the CTCF-bound element, E1. Dissecting the exact motifs and TFs that drive the potential functions of E1 will require detailed follow-up experiments. Two of the CREs validated in this study also regulate 9p21 transcripts in HeLa cells, though with very different transcript specificity, highlighting the need to study different cellular contexts24 (Figure S18). 9p21 promoters and distal CREs are accessible in human cells in vivo, with strong cell-type specificity. The distal CREs overlap at least 4 of the LD blocks in 9p21 containing disease-associated haplotypes but are depleted of disease-associated SNPs (with the exception of E5). This reinforces the importance of functional experiments to annotate conserved CREs that are not amenable to genetic QTL analyses. Finally, our topological analyses suggest that the cell cycle gene region of 9p21 is structured by CTCF-CTCF, enhancer-promoter and promoter-promoter interactions. Chromosomal contact frequencies, including these interactions, are reduced across the 9p21 TAD at senescence. We speculate that this decompaction, potentially driven by loss of polycomb binding28,29 and/or altered CTCF insulator activity,64,83 impacts the regulation of cell cycle gene promoters and downstream senescence programs. Such conformational changes may also contribute to the switch from a cycling-associated p14ARF program to a p16INK4A- and p15INK4B-driven program during RS. Overall, this study contributes detailed insight into the independent cell-type-specific expression and distal regulation of the 9p21 locus cell cycle regulators—p14ARF, p16INK4A, and p15INK4B—and ANRIL and provides new tools for studying these critical transcripts in single cells and tissues and across cell-state dynamics. Limitations of the study While high-MOI screening approaches increase the number of perturbations that can be tested, they can introduce false positives and noise into the data due to the potentially confounding effects of multiple gRNAs in the same cell. In this study, we took a conservative approach by utilizing a high-MOI screening approach to nominate gRNA-gene interactions, then drew conclusions from follow-up single transfections. Future studies will be needed to validate other gRNA-gene pairs that score in the high-MOI screen but were not explored further here. Our current data are also underpowered to determine the combinatorial effects of gRNA or CRE pairs. Future studies will also be needed to understand how CREs integrate other aging-associated changes in the epigenome—including binding of polycomb repressive complex repressors, chromatin states, and DNA methylation—to regulate 9p21 genes and shape cellular phenotypes during aging. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Elizabeth Gaskell (egaskell@broadinstitute.org). Materials availability Any materials generated specifically for this study are available from the Broad Institute Epigenomics Program or the Bernstein Lab at DFCI on request. Data and code availability •Data generated for this study are available at GEO: GSE309515.•Accession numbers for the published data utilized for this study are GEO: GSE175533 and GSE159929, E-MTAB-9543, GEO: GSE183852, hum0229.v1, and syn26560310•All code is available on Github (https://github.com/broadinstitute/CDKN2A-gene-regulation) and on Zenodo (10.5281/zenodo.20411705). •Data generated for this study are available at GEO: GSE309515. Data generated for this study are available at GEO: GSE309515. •Accession numbers for the published data utilized for this study are GEO: GSE175533 and GSE159929, E-MTAB-9543, GEO: GSE183852, hum0229.v1, and syn26560310 Accession numbers for the published data utilized for this study are GEO: GSE175533 and GSE159929, E-MTAB-9543, GEO: GSE183852, hum0229.v1, and syn26560310 •All code is available on Github (https://github.com/broadinstitute/CDKN2A-gene-regulation) and on Zenodo (10.5281/zenodo.20411705). All code is available on Github (https://github.com/broadinstitute/CDKN2A-gene-regulation) and on Zenodo (10.5281/zenodo.20411705). Acknowledgments We thank Joan Brugge for donating the scRNA-seq libraries for the breast tissue analysis. We thank the Broad Genetic Perturbation Platform and Broad Clinical Labs (BCL) for assistance with the design and creation of the CRISPRa library and DNA sequencing. We thank members of the Bernstein lab, the Broad Epigenomics Program, and the Broad-Calico alliance for helpful discussions. We thank Eleonora Adami from Brush & Pipette for assistance with the graphical abstract. This manuscript is the result of a collaborative effort between the Broad Institute and Calico Life Sciences LLC and was supported by funding from Calico Life Sciences LLC. E.G., B.E.B., and A.S.H. acknowledge support from the 10.13039/501100024571Novo Nordisk Foundation Center for Genomic Mechanisms of Disease (NNF21SA0072102). This work was also supported by funds from the Gene Regulation Observatory at the Broad Institute. Author contributions T.E.M., N.J., J.V., K.L.K., R.I., C.B.E., M.C., D.G.H., and J.D. performed all experiments. E.T.T., N.C.D., S.C.-G., A.P.H., N.S., V.R., M.J.A., A.C., A.R., V.H., and D.R.K. performed all data analysis. F.A. and K.A. contributed specifically to the human genetics analysis. A.J.C. and F.J.N. assisted with the CRISPRa validation experiments. V.R. and A.S.H. assisted with the RCMC analysis. C.M.-C.L. provided the mammary tissue scRNA-seq library. E.G., B.E.B., and J.G.R. designed and supervised the study. B.E.B., E.G., E.T.T., T.E.M., and J.G.R. wrote the paper. Declaration of interests M.C., D.G.H., J.G.R., and D.R.K. are employees of Calico Life Sciences LLC. F.A. was an employee of the Broad Institute when he contributed to the study and is now an employee of Predicta Biosciences. Calico Life Sciences LLC and the Broad Institute participated in interpreting the data and reviewing and approving the publication. B.E.B. declares outside interests in HiFiBio, Arsenal Biosciences, Cell Signaling Technologies, Sesame Therapeutics, and nChroma Bio. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesH3K27acActive Motifpart 39133, lot 31814008, RRID: AB_2561016CTCFDiagenodepart C15410210, lot A2359-00234P, RRID: AB_2753160Bacterial and virus strainsOne Shot STBL3 chemically competent E. coliInvitrogenC737303Chemicals, peptides, and recombinant proteinsKAPA hot start ready mixRoche07958935001Power SYBR™ Green PCR Master MixApplied bio systems43-676-59TRIzol™ Reagentinvitrogen15596026Maxima H Minus Reverse TranscriptaseThermo ScientificEP0751SYBR master mixApplied Biosystems43-444-63NP-40 SubstituteAbcamab142227disuccinimidyl glutarateThermoFisher20593methanol-free formaldehydeThermoFisher28906PBSThermoFisher10010023Tris buffer pH 7.5ThermoFisher15567027BSA, Molecular Biology GradeNEBB9000ST4 PNK (polynucleotide kinase)NEBM0201Protease Inhibitor CocktailSigma-Aldrich5056489001MNaseNEBM0247EGTAbioWORLD40520008ATPNEBP0756DTTSigma-Aldrich10197777001DNA Polymerase I Klenow FragmentNEBM0210dNTPsInvitrogen10297018Biotin-14-dATP, Biotin-11-dCTPJena BioscienceNU-835-BIO14, NU-809-BIOXEDTAInvitrogen15575020T4 DNA LigaseNEBM0202Exonuclease IIINEBM020620% SDS (Sodium dodecyl sulfate)InvitrogenAM9822Proteinase KNEBP8107RNaseAInvitrogenEN0531AgaroseVWR97062Dynabeads MyOne Streptavidin T1Invitrogen65601Tween 20Sigma-AldrichP8074NEB Multiplex Oligos for Illumina Primer Set 1NEBE7335Q5 High-Fidelity DNA PolymeraseNEBM0491Capture Custom PanelTwist Bioscience101001Standard Hybridization MixTwist Bioscience104178Universal BlockersTwist Bioscience100578Streptavidin Binding BeadsTwist Bioscience100983Equinox Library Amplification MixTwist Bioscience104178DNA Purification BeadsTwist Bioscience100983Critical commercial assaysDynabeads™ kilobaseBINDER™ KitThermo fisher60101Plasmid Miniprep KitQiagen27104Zymo DNA Clean & Concentrator kitZymo ResearchD4034Zymo Gel Purification kitZymo ResearchD4008Qubit 1× dsDNA High Sensitivity AssayInvitrogenQ33231Quick Blunting KitNEBE1201NEBNext Ultra II kitNEBE7645Deposited dataGTEx mRNA v8 TPM estimatesMelé et al.40Isoform-level expression estimates from “GTEx_Analysis_2017-06-05_v8_RSEMv1.3.0_transcript_tpm.gct.gz”,Gene-level expression estimates from “GTEx_Analysis_2017-06-05_v8_RNASeQCv1.1.9_gene_tpm.gct.gz”From (https://gtexportal.org/home/downloads/adult-gtex/bulk_tissue_expression)FANTOM CAGE TPM values hg38FANTOM Consortium41TPM values for each CAGE peak (“hg38_fair+new_CAGE_peaks_phase1and2_tpm.osc.txt”) from https://fantom.gsc.riken.jp/5/datafiles/reprocessed/hg38_v9/extra/CAGE_peaks_expression/.“Human tissue” samples list from (“00_human.tissue.hCAGE.hg38.assay_sdrf.txt”) from https://fantom.gsc.riken.jp/5/datafiles/reprocessed/hg38_v9/basic/human.tissue.hCAGE/Adult human atlas 5′ scRNA-seqHe et al.47SRA: SRP292721Intestinal tract atlas 5′ scRNA-seqElmentaite et al.48ArrayExpress: E-MTAB-9543Cardiac tissue 5′ scRNA-seqKoenig et al.49SRA: SRP33642810x Genomics example data–https://www.10xgenomics.com/resources/datasets/10k-human-pbmcs-5-v2-0-chromium-x-with-intronic-reads-2-standardT cells in supercentenarians 5′ scRNA-seqHashimoto et al.50NBDC Human Database https://humandbs.biosciencedbc.jp/en/hum0229-v1Breast atlas 3′ scRNA-seqGray et al.55https://www.synapse.org/Synapse:syn26560310WI-38 RS 3′ scRNA dataChan et al.17GEO: GSE175533WI-38 RS ATAC-seq dataChan et al.17GEO: GSE175533WI-38 RS mRNA-seq data – TPM tableChan et al.17(https://github.com/dghendrickson/hayflick/blob/main/data/RNAseq/TPM_tables/RS_batch_tpms.txtsci-ATAC-seq dataZhang et al.82https://decoder-genetics.wustl.edu/catlasv1/humanenhancer/data/cCRE_by_cell_type/ENCODE cCREs–https://hgdownload.soe.ucsc.edu/gbdb/hg38/encode3/ccre/encodeCcreCombined.bbENCODE HeLa ChIP-seq tracks–H3K27ac accession ENCFF850RSK/ENCSR000AOC and H3K4me1 accession ENCFF860GGZ/ENCSR000APWENCODE IMR-90 ChIP-seq tracks–FOS accession: ENCSR124AIG, ENCFF906BPJConservation scores (phastCons100way)–https://hgdownload.soe.ucsc.edu/goldenPath/hg38/phastCons100way/hg38.phastCons100way.bwCommon SNPs–http://hgdownload.soe.ucsc.edu/gbdb/hg38/snp/dbSnp155Common.bbFine-mapped variants associated with common diseases from UKBB–https://www.finucanelab.org/dataLead SNPs from cancer germline GWAS studies–Downloaded from the EBI GWAS catalog https://www.ebi.ac.uk/gwas on 11/03/2024 (e113)eQTLs for CDKN2A, CDKN2B, CDKN2B-AS1 (ANRIL) and MTAP–GTEx V10Linkage disequilibrium–UKBBExample datasets for 10x 3′ and 5′–(3′ data) https://www.10xgenomics.com/datasets/4-k-pbm-cs-from-a-healthy-donor-2-standard-1-2-0 and https://www.10xgenomics.com/datasets/frozen-pbm-cs-donor-a-1-standard-1-1-0); 5′ data: https://www.10xgenomics.com/resources/datasets/10k-human-pbmcs-5-v2-0-chromium-x-with-intronic-reads-2-standard and https://www.10xgenomics.com/resources/datasets/5k-human-ovarian-tumor-isolated-with-chromium-nuclei-isolation-kit-3-1-standardAll data generated for this study–GEO: GSE309515All code generated for this study–https://github.com/broadinstitute/CDKN2A-gene-regulation; 10.5281/zenodo.20411705Experimental models: Cell linesWI-38ATCCCCL-75OligonucleotidesHybridization MixTwist Bioscience104178Custom CDKN2A PanelsTwist Bioscience101001Universal BlockersTwist Bioscience100578Recombinant DNALenti dCAS-VP64_blastKonermann et al.74addgene, catalog # 61425CROP_sgRNA-MS2Alda-Catalinas et al.71addgene, catalog #153457LentiMPH v2Joung et al.91addgene, catalog #89308CROP-seq-optiHill et al.92addgene, catalog #106280pCMV_VSVgStewart et al.93addgene, catalog #8454psPAX2psPAX2 was a gift to addgene from Didier Tronoaddgene, catalog #12260Software and algorithmsComplexHeatmap 2.22.0Gu et al.94https://bioconductor.org/packages/release/bioc/html/ComplexHeatmap.htmlUpSetR v 1.4.0Conway et al.95https://github.com/hms-dbmi/UpSetRSTAR solo, v 2.7.10b_alpha_230301 (through Cumulus)Dobin et al. (STAR)96; Kaminow et al. (STARsolo)97; and Li et al. (Cumulus)98https://github.com/alexdobin/STAR/blob/master/docs/STARsolo.mdCumulus https://cumulus.readthedocs.io/scanpy 1.9.1Wolf et al.99https://scanpy.readthedocs.io/pandas 1.4.3Pandas on Zenodo: https://doi.org/10.5281/zenodo.3509134; McKinney100https://pandas.pydata.org/EnrichRChen et al.,101; Kuleshov et al.57; and Xie et al.102https://maayanlab.cloud/Enrichr/FastQC v0.11.9Andrews S (2010), Babraham Institute, Bioinformaticshttps://www.bioinformatics.babraham.ac.uk/projects/fastqc/multiQC v1.9Ewels et al.103https://github.com/multiqc/multiqcSkewerJiang et al.104https://github.com/relipmoc/skewerbowtie2 v2.4.1Langmead & Salzberg105; Langmead et al.106https://bowtie-bio.sourceforge.net/bowtie2/SAMBLASTERFaust & Hall107https://github.com/GregoryFaust/samblasterdeepTools2Ramírez et al.108https://deeptools.readthedocs.io/macs2 v2.2.7.1Zhang et al.109; Feng et al.110https://pypi.org/project/MACS2/ataqv tool v1.2.1Orchard et al.111https://github.com/ParkerLab/ataqvDiffBindRoss-Innes et al.112https://bioconductor.org/packages/release/bioc/html/DiffBind.htmlSCEPTREBarry et al.74https://katsevich-lab.github.io/sceptre/Juicer v2Durand et al.113Juicer/juicer_tools v2.20.00, incl. HiCCUPS; https://github.com/aidenlab/juicerhicstraw/strawDurand et al.113https://github.com/aidenlab/strawplotgardner v1.6.1, v1.12.0Kramer et al.114https://bioconductor.org/packages/release/bioc/html/plotgardener.htmlChromBPNet v0.1.7Pampari et al.77https://github.com/kundajelab/chrombpnetDeepLIFTShrikumar et al.115https://github.com/kundajelab/deepliftTOMTOMGupta et al.116https://meme-suite.org/meme/tools/tomtomSalmon v1.10.1Patro et al.117https://combine-lab.github.io/salmon/tximportSoneson et al.118https://bioconductor.org/packages/release/bioc/html/tximport.htmlEnsDb.Hsapiens.v86Rainer, Bioconductor119https://bioconductor.org/packages/EnsDb.Hsapiens.v86SeuratHao et al.120; Hao et al.121; Stuart and Butler et al.122; Butler et al.123; and Satija and Farrell124https://satijalab.org/seurat/SeuratDisk (v0.0.0.9015)Hoffman, GitHubhttps://github.com/mojaveazure/seurat-diskSAMtools (v1.11)Li et al.125; Danecek et al.126https://www.htslib.org/BEDTools (v2.29.0)Quinlan & Hall127https://bedtools.readthedocs.io/HOMER (v4.11.1)Heinz et al.128http://homer.ucsd.edu/homer/DESeq2Love et al.129https://bioconductor.org/packages/DESeq2UCSC and UCSC tools (bigWigToBedGraph, bedGraphToBigWig, LiftOver, BLAT)Casper et al.130; Kent et al.131; Hinrichs et al.(Liftover)132; and Kent et al. (bigwig tools)133https://genome.ucsc.edu/util.htmlENCODE ATAC-seq/ChIP-seq uniform processing pipelineHitz et al.134https://github.com/ENCODE-DCCCTCF (Bioconductor)Dozmorov et al.135https://bioconductor.org/packages/CTCFJASPAR 2022/2024 PWMsCastro-Mondragon et al.136; Rauluseviciute et al.137https://jaspar.elixir.no/ggplot2Wickham138https://ggplot2.tidyverse.org/matplotlib (v3.6.3)Hunter139https://matplotlib.org/seaborn (v0.12.2)Waskom140https://seaborn.pydata.org/NumPyHarris et al.141https://numpy.org/IGV/igv.jsRobinson et al.(IGV)66Robinson et al.(igv.js)67https://igv.org/RR Core Teamhttps://www.r-project.org/Python, stats packageVirtanen et al.(SciPy)142; Seabold & Perktold, (statsmodels)143SciPy v1.10.0 (scipy.stats) statsmodels v0.13.5Python (v3)Python Software Foundationhttps://www.python.org/Original code (this study)This paperhttps://github.com/broadinstitute/CDKN2A-gene-regulation;Zenodo: https://doi.org/10.5281/zenodo.20411705Other10 cm Tissue culture platesCorning3530036 well tissue culture platesCorning353046EMEMATCC30–2003Fetal Bovine serumPeak serumFBS-1penicillin/streptomycinGibco15140–163Trypsin-EDTAGibco25200114Opti-MEMThermoFischer31985062TransIT-LT1 (Mirus) transfection reagentMirusBioMIR2304 REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesH3K27acActive Motifpart 39133, lot 31814008, RRID: AB_2561016CTCFDiagenodepart C15410210, lot A2359-00234P, RRID: AB_2753160Bacterial and virus strainsOne Shot STBL3 chemically competent E. coliInvitrogenC737303Chemicals, peptides, and recombinant proteinsKAPA hot start ready mixRoche07958935001Power SYBR™ Green PCR Master MixApplied bio systems43-676-59TRIzol™ Reagentinvitrogen15596026Maxima H Minus Reverse TranscriptaseThermo ScientificEP0751SYBR master mixApplied Biosystems43-444-63NP-40 SubstituteAbcamab142227disuccinimidyl glutarateThermoFisher20593methanol-free formaldehydeThermoFisher28906PBSThermoFisher10010023Tris buffer pH 7.5ThermoFisher15567027BSA, Molecular Biology GradeNEBB9000ST4 PNK (polynucleotide kinase)NEBM0201Protease Inhibitor CocktailSigma-Aldrich5056489001MNaseNEBM0247EGTAbioWORLD40520008ATPNEBP0756DTTSigma-Aldrich10197777001DNA Polymerase I Klenow FragmentNEBM0210dNTPsInvitrogen10297018Biotin-14-dATP, Biotin-11-dCTPJena BioscienceNU-835-BIO14, NU-809-BIOXEDTAInvitrogen15575020T4 DNA LigaseNEBM0202Exonuclease IIINEBM020620% SDS (Sodium dodecyl sulfate)InvitrogenAM9822Proteinase KNEBP8107RNaseAInvitrogenEN0531AgaroseVWR97062Dynabeads MyOne Streptavidin T1Invitrogen65601Tween 20Sigma-AldrichP8074NEB Multiplex Oligos for Illumina Primer Set 1NEBE7335Q5 High-Fidelity DNA PolymeraseNEBM0491Capture Custom PanelTwist Bioscience101001Standard Hybridization MixTwist Bioscience104178Universal BlockersTwist Bioscience100578Streptavidin Binding BeadsTwist Bioscience100983Equinox Library Amplification MixTwist Bioscience104178DNA Purification BeadsTwist Bioscience100983Critical commercial assaysDynabeads™ kilobaseBINDER™ KitThermo fisher60101Plasmid Miniprep KitQiagen27104Zymo DNA Clean & Concentrator kitZymo ResearchD4034Zymo Gel Purification kitZymo ResearchD4008Qubit 1× dsDNA High Sensitivity AssayInvitrogenQ33231Quick Blunting KitNEBE1201NEBNext Ultra II kitNEBE7645Deposited dataGTEx mRNA v8 TPM estimatesMelé et al.40Isoform-level expression estimates from “GTEx_Analysis_2017-06-05_v8_RSEMv1.3.0_transcript_tpm.gct.gz”,Gene-level expression estimates from “GTEx_Analysis_2017-06-05_v8_RNASeQCv1.1.9_gene_tpm.gct.gz”From (https://gtexportal.org/home/downloads/adult-gtex/bulk_tissue_expression)FANTOM CAGE TPM values hg38FANTOM Consortium41TPM values for each CAGE peak (“hg38_fair+new_CAGE_peaks_phase1and2_tpm.osc.txt”) from https://fantom.gsc.riken.jp/5/datafiles/reprocessed/hg38_v9/extra/CAGE_peaks_expression/.“Human tissue” samples list from (“00_human.tissue.hCAGE.hg38.assay_sdrf.txt”) from https://fantom.gsc.riken.jp/5/datafiles/reprocessed/hg38_v9/basic/human.tissue.hCAGE/Adult human atlas 5′ scRNA-seqHe et al.47SRA: SRP292721Intestinal tract atlas 5′ scRNA-seqElmentaite et al.48ArrayExpress: E-MTAB-9543Cardiac tissue 5′ scRNA-seqKoenig et al.49SRA: SRP33642810x Genomics example data–https://www.10xgenomics.com/resources/datasets/10k-human-pbmcs-5-v2-0-chromium-x-with-intronic-reads-2-standardT cells in supercentenarians 5′ scRNA-seqHashimoto et al.50NBDC Human Database https://humandbs.biosciencedbc.jp/en/hum0229-v1Breast atlas 3′ scRNA-seqGray et al.55https://www.synapse.org/Synapse:syn26560310WI-38 RS 3′ scRNA dataChan et al.17GEO: GSE175533WI-38 RS ATAC-seq dataChan et al.17GEO: GSE175533WI-38 RS mRNA-seq data – TPM tableChan et al.17(https://github.com/dghendrickson/hayflick/blob/main/data/RNAseq/TPM_tables/RS_batch_tpms.txtsci-ATAC-seq dataZhang et al.82https://decoder-genetics.wustl.edu/catlasv1/humanenhancer/data/cCRE_by_cell_type/ENCODE cCREs–https://hgdownload.soe.ucsc.edu/gbdb/hg38/encode3/ccre/encodeCcreCombined.bbENCODE HeLa ChIP-seq tracks–H3K27ac accession ENCFF850RSK/ENCSR000AOC and H3K4me1 accession ENCFF860GGZ/ENCSR000APWENCODE IMR-90 ChIP-seq tracks–FOS accession: ENCSR124AIG, ENCFF906BPJConservation scores (phastCons100way)–https://hgdownload.soe.ucsc.edu/goldenPath/hg38/phastCons100way/hg38.phastCons100way.bwCommon SNPs–http://hgdownload.soe.ucsc.edu/gbdb/hg38/snp/dbSnp155Common.bbFine-mapped variants associated with common diseases from UKBB–https://www.finucanelab.org/dataLead SNPs from cancer germline GWAS studies–Downloaded from the EBI GWAS catalog https://www.ebi.ac.uk/gwas on 11/03/2024 (e113)eQTLs for CDKN2A, CDKN2B, CDKN2B-AS1 (ANRIL) and MTAP–GTEx V10Linkage disequilibrium–UKBBExample datasets for 10x 3′ and 5′–(3′ data) https://www.10xgenomics.com/datasets/4-k-pbm-cs-from-a-healthy-donor-2-standard-1-2-0 and https://www.10xgenomics.com/datasets/frozen-pbm-cs-donor-a-1-standard-1-1-0); 5′ data: https://www.10xgenomics.com/resources/datasets/10k-human-pbmcs-5-v2-0-chromium-x-with-intronic-reads-2-standard and https://www.10xgenomics.com/resources/datasets/5k-human-ovarian-tumor-isolated-with-chromium-nuclei-isolation-kit-3-1-standardAll data generated for this study–GEO: GSE309515All code generated for this study–https://github.com/broadinstitute/CDKN2A-gene-regulation; 10.5281/zenodo.20411705Experimental models: Cell linesWI-38ATCCCCL-75OligonucleotidesHybridization MixTwist Bioscience104178Custom CDKN2A PanelsTwist Bioscience101001Universal BlockersTwist Bioscience100578Recombinant DNALenti dCAS-VP64_blastKonermann et al.74addgene, catalog # 61425CROP_sgRNA-MS2Alda-Catalinas et al.71addgene, catalog #153457LentiMPH v2Joung et al.91addgene, catalog #89308CROP-seq-optiHill et al.92addgene, catalog #106280pCMV_VSVgStewart et al.93addgene, catalog #8454psPAX2psPAX2 was a gift to addgene from Didier Tronoaddgene, catalog #12260Software and algorithmsComplexHeatmap 2.22.0Gu et al.94https://bioconductor.org/packages/release/bioc/html/ComplexHeatmap.htmlUpSetR v 1.4.0Conway et al.95https://github.com/hms-dbmi/UpSetRSTAR solo, v 2.7.10b_alpha_230301 (through Cumulus)Dobin et al. (STAR)96; Kaminow et al. (STARsolo)97; and Li et al. (Cumulus)98https://github.com/alexdobin/STAR/blob/master/docs/STARsolo.mdCumulus https://cumulus.readthedocs.io/scanpy 1.9.1Wolf et al.99https://scanpy.readthedocs.io/pandas 1.4.3Pandas on Zenodo: https://doi.org/10.5281/zenodo.3509134; McKinney100https://pandas.pydata.org/EnrichRChen et al.,101; Kuleshov et al.57; and Xie et al.102https://maayanlab.cloud/Enrichr/FastQC v0.11.9Andrews S (2010), Babraham Institute, Bioinformaticshttps://www.bioinformatics.babraham.ac.uk/projects/fastqc/multiQC v1.9Ewels et al.103https://github.com/multiqc/multiqcSkewerJiang et al.104https://github.com/relipmoc/skewerbowtie2 v2.4.1Langmead & Salzberg105; Langmead et al.106https://bowtie-bio.sourceforge.net/bowtie2/SAMBLASTERFaust & Hall107https://github.com/GregoryFaust/samblasterdeepTools2Ramírez et al.108https://deeptools.readthedocs.io/macs2 v2.2.7.1Zhang et al.109; Feng et al.110https://pypi.org/project/MACS2/ataqv tool v1.2.1Orchard et al.111https://github.com/ParkerLab/ataqvDiffBindRoss-Innes et al.112https://bioconductor.org/packages/release/bioc/html/DiffBind.htmlSCEPTREBarry et al.74https://katsevich-lab.github.io/sceptre/Juicer v2Durand et al.113Juicer/juicer_tools v2.20.00, incl. HiCCUPS; https://github.com/aidenlab/juicerhicstraw/strawDurand et al.113https://github.com/aidenlab/strawplotgardner v1.6.1, v1.12.0Kramer et al.114https://bioconductor.org/packages/release/bioc/html/plotgardener.htmlChromBPNet v0.1.7Pampari et al.77https://github.com/kundajelab/chrombpnetDeepLIFTShrikumar et al.115https://github.com/kundajelab/deepliftTOMTOMGupta et al.116https://meme-suite.org/meme/tools/tomtomSalmon v1.10.1Patro et al.117https://combine-lab.github.io/salmon/tximportSoneson et al.118https://bioconductor.org/packages/release/bioc/html/tximport.htmlEnsDb.Hsapiens.v86Rainer, Bioconductor119https://bioconductor.org/packages/EnsDb.Hsapiens.v86SeuratHao et al.120; Hao et al.121; Stuart and Butler et al.122; Butler et al.123; and Satija and Farrell124https://satijalab.org/seurat/SeuratDisk (v0.0.0.9015)Hoffman, GitHubhttps://github.com/mojaveazure/seurat-diskSAMtools (v1.11)Li et al.125; Danecek et al.126https://www.htslib.org/BEDTools (v2.29.0)Quinlan & Hall127https://bedtools.readthedocs.io/HOMER (v4.11.1)Heinz et al.128http://homer.ucsd.edu/homer/DESeq2Love et al.129https://bioconductor.org/packages/DESeq2UCSC and UCSC tools (bigWigToBedGraph, bedGraphToBigWig, LiftOver, BLAT)Casper et al.130; Kent et al.131; Hinrichs et al.(Liftover)132; and Kent et al. (bigwig tools)133https://genome.ucsc.edu/util.htmlENCODE ATAC-seq/ChIP-seq uniform processing pipelineHitz et al.134https://github.com/ENCODE-DCCCTCF (Bioconductor)Dozmorov et al.135https://bioconductor.org/packages/CTCFJASPAR 2022/2024 PWMsCastro-Mondragon et al.136; Rauluseviciute et al.137https://jaspar.elixir.no/ggplot2Wickham138https://ggplot2.tidyverse.org/matplotlib (v3.6.3)Hunter139https://matplotlib.org/seaborn (v0.12.2)Waskom140https://seaborn.pydata.org/NumPyHarris et al.141https://numpy.org/IGV/igv.jsRobinson et al.(IGV)66Robinson et al.(igv.js)67https://igv.org/RR Core Teamhttps://www.r-project.org/Python, stats packageVirtanen et al.(SciPy)142; Seabold & Perktold, (statsmodels)143SciPy v1.10.0 (scipy.stats) statsmodels v0.13.5Python (v3)Python Software Foundationhttps://www.python.org/Original code (this study)This paperhttps://github.com/broadinstitute/CDKN2A-gene-regulation;Zenodo: https://doi.org/10.5281/zenodo.20411705Other10 cm Tissue culture platesCorning3530036 well tissue culture platesCorning353046EMEMATCC30–2003Fetal Bovine serumPeak serumFBS-1penicillin/streptomycinGibco15140–163Trypsin-EDTAGibco25200114Opti-MEMThermoFischer31985062TransIT-LT1 (Mirus) transfection reagentMirusBioMIR2304 Experimental model and study participant details WI-38 cell culture WI-38 cells were obtained from ATCC (CCL-75) at PDL 1. Cells were cultured as in Chan et al.17 Cells were grown in Eagle’s minimum essential medium (EMEM, ATCC, 30–2003), supplemented with 10% fetal bovine serum (Peak serum FBS-1), and 1% penicillin/streptomycin (Gibco 15140-163), and maintained in an incubator set to 37°C, 5% CO2, and 20% O2. Cells were seeded and maintained on 10 cm cell culture plates (Corning, 353003). The cells were split when the plates reached ∼70–80% confluence. Cells were passaged by washing the cell monolayer with PBS, followed by incubating the cells with Trypsin-EDTA (Gibco 25200114) for 2 min. Media was added to neutralize the Trypsin-EDTA and the cell suspension was collected into conical tubes. The cell suspension was spun at 200 x g for 5 min and the supernatant was aspirated. The resulting cell pellet was resuspended in media and cells were reseeded onto 10cm culture plates. PDL was tracked by the approximate doubling of cell confluence. Method details Transcript enrichment from single-cell libraries From Seq-Well single-cell libraries p14ARF or p16INK4A mRNAs were enriched from the WTA step of Seq-Well scRNA library preparation using a two PCR reaction chain, using a strategy similar to our previously reported methods for DNA mutation detection.54,144,145 Transcript specific primers are used for targeted biotinylation of our transcripts of interest while retaining the UMI and CB that the transcripts have from the Seq-Well protocol (Table S1). To perform PCR 1, a mix of 3 μM of targeted primers were added to 3 μM CA-SMART reverse primer to create a primer mix. 2.5 μL of primer mix was then added 12.5 μL of KAPA hot start ready mix (Roche 07958935001) to create a reaction master mix. As a template for this reaction, WTA libraries were pooled and diluted to obtain a 100 ng of cDNA in 10 μL of water. To this template 15 μL of master mix was added and run through a PCR reaction consisting of an initial 95 C denaturation step for 5 min, followed by 10 cycles of 98 C for 20 s, 65 C for 15 s, and 72 C for 3 min. The reaction ends with a final extension for 72 C for 5 min. Following this, a 1X AMPure bead clean up is performed to remove primer dimer noise from the reaction. The transcripts of interest are then separated via biotin pull down using the Dynabeads kilobaseBINDER Kit (Thermo, 60101) according to the manufacturer’s instructions. These enriched transcripts were then eluted into 23 μL of water. A second PCR reaction was then performed to add Illumina adapters P7 and P5, index barcodes i5 and i7 to identify the libraries (Table S1), and sequencing primer binding sites to the fragments. For this reaction, 23 μL of eluted streptavidin beads are added to 2 μL of 5 μM primer mix with sample barcode and 25 μL of KAPA hot start ready mix (Roche 07958935001) (Table S1). The PCR reaction is run with the following parameters: 95 C denaturation for 2 min followed by 10 cycles of 95 C for 20 s, 65 C for 20 s, and 72 C for 1 min and 20 s. This is then finished with 72 C for 5 min. Following this reaction, a 0.9X AMPure clean up was performed and the final product was eluted in 1X TE. Libraries were pooled (Table S9) and sequenced using a 75-cycle kit on the Illumina NextSeq 500 with custom read 1 (CR1P) and custom i5 primers (SW-Ci5P, Table S1), 20 cycles for read 1 (Cell barcode, CB + Unique Molecular Identifiers UMI), 56 cycles for read 2 (transcript sequence) and 2 × 8-bp library barcodes. From 10X single-cell libraries Akin to our previously reported methods for DNA mutation detection from single cells,54,144,145 a parallel 2 PCR enrichment strategy was used for 10x Genomics single-cell libraries. This does not require biotinylated primers and pull down as there are unique bidirectional primers for both reactions for 10x WTA. See primer sequences (Table S1) for primers for first and 2nd PCR reactions. Libraries were sequenced using a 75-cycle kit on the Illumina NextSeq 500 with custom read 1 (10x CR1P) and custom i5 primers (10x-Ci5P), 28 cycles for read 1 (CB + UMI) for v3 (26 cycles for v2), 48 cycles for read 2 (transcript sequence) for v3 (50 cycles for v2) and 2 × 8-bp library barcodes. CRISPRa screen gRNAs were designed against 173 regions (see Guide selection below and Table S3) and screened in WI-38 in a pooled CRISPRa screen. Changes in gene expression were measured with Seq-Well single-cell expression profiles and our p14ARF and p16INK4A enrichment protocol (see above). Previous studies of 9p21 cis-regulation have either focused on promoters, or annotated regulatory regions in cancer cell lines that do not enter replicative senescence.14,24,26 These cell lines typically express high levels of 9p21 transcripts independent of senescence and are therefore amenable to knock-down approaches. Conversely, in WI-38 cells expression of p16INK4A and p15INK4B prior to senescence is low, rendering knock down studies challenging, and motivating the choice of CRISPRa for this study. We chose to use a 3-vector system using CROP-sgRNA-MS2 (addgene 153457)71 for guide delivery, lenti dCAS-VP64_Blast (addgene 61425)73 for dCAS9 activator delivery, and lentiMPH v2 (addgene 89308)91 for activator helper complex delivery. WI-38 cells constitutively expressing both dCAS-VP64_Blast and lentiMPH v2 were created using lentivirus transduction and selection by Blasticidin (15 μg/mL) and Hygromycin (400 μg/mL), respectively. These selected cells were then transduced with the gRNA library in CROP-sgRNA-MS2 vector and selected using Puromycin (2 μg/mL). Guide selection All guide sequences in this study can be found in Table S3. Promoter guides were selected using CRISPick and utilized for positive control pilot experiments prior to the large-scale screen (data not shown). The full CRISPR library was designed and produced by the Broad Genetic Perturbation Platform. The library includes 724 guide RNAs targeting 173 regions (see section Selection of regions for CRISPRa screen experiment below for region selection, ∼3 gRNAs per region) and 83 safe harbor gRNAs as negative controls. Vectors and virus production For the large-scale pooled library screen, virus production was performed by the Broad Genetic Perturbation Platform. For validation experiments, annealed oligos were cloned into CROP-seq-opti (addgene 10628092) and CROP-seq-MS2 (Addgene 153457) backbones using BsmBI restriction enzyme for vector linearization and T7 ligase for ligation. Plasmids were transformed into One Shot STBL3 chemically competent E. coli (Invitrogen, C737303). After 1 h of shaking at 225 rpm and 37°C, bacterial cultures were plated on ampicillin agar. Single colonies were picked after overnight incubation at 37°C, grown in LB and shaken overnight at 225 rpm and 37°C. Plasmid DNA was isolated using a Plasmid Miniprep Kit (Qiagen) and concentrations were measured using a Nanodrop spectrophotometer and a Qubit fluorometer. HEK293FT cells (1 × 106 per well) were seeded in 6-well plates with 2 mL DMEM +10% FBS, 24 h before transfection. Transfection mixes included 250 μL Opti-MEM, 0.25 μg pCMV_VSVg (Addgene, 845493), 1.25 μg psPAX2 (Addgene, 12260), 1 μg guide carrying CROP-seq-opti or CROP-seq-MS2 vector, and 7.5 μL TransIT-LT1 (Mirus) transfection reagent. After a 20-min incubation, the solution was added dropwise to the 6-well and incubated at 37°C, with fresh media being added after 6–8 h. After 36 h, virus was collected, filtered, and snap frozen. Thawed viral supernatant was added to WI-38 (ATCC CCL-75) cells at 800 μL per 3 × 10ˆ4 cells with 8 μg/ml polybrene and cultured for 4 or 8 days. Cells were harvested and RNA extracted using trizol and analyzed by qRT-PCR (see below). Single-cell RNA sequencing with Seq-Well Single-cell RNA sequencing using Seq-Well was perfomed as described previously.146,147 Cells are prepared in a single cell suspension in cell media and loaded dropwise into a 90,000 well microarray pre-loaded with RNA capture beads. Single cells settle into wells with Poly-T RNA capture beads and the array is sealed with a semipermeable membrane. These sealed arrays are then submerged in a cell Lysis buffer which passes freely through the membrane. Cells are lysed and polyadenylated RNAs hybridize onto the RNA capture beads. The arrays are then unsealed and beads are washed out of the array. A reverse transcription reaction is then performed on the beads, generating cDNA attached to the capture beads. A second strand synthesis reaction is then run to generate a strand complementary to the cDNA. Following this whole transcriptome amplification (WTA) is performed to obtain a sequenceable quantity of cDNA. The product underwent a combination of tagmentation and PCR to generate dual-indexed sequencing libraries. Libraries were sequenced using a 75-cycle kit on the Illumina NextSeq 500 with custom read 1 (CR1P) and custom i5 primers (SW-Ci5P)(Table S1), 20 cycles for read 1 (CB + UMI), 56 cycles for read 2 (transcript sequence) and 2 × 8-bp library barcodes. We obtained data for 89,732 single cells after QC (Data processing and analysis). RT-qPCR validation of CRISPRa screen CRISPRa Perturb-seq results for 20 gRNAs that scored in our SCEPTRE analysis were validated with repeated arrayed CRISPRa experiments and qRT-PCR (Tables S5, S6 and S10). These 20 gRNAs included 6 guides targeting the 3 promoter regions, and 14 guides targeting 9 of the distal CRES (defined as E1-E9). We also introduced 2 additional gRNAs for E7 that were not included in our original screen (gRNAs 700 and 702 in Table S5). Induction of p14ARF or p16INK4A via promoter-targeting gRNAs induces cell death or senescence over time, confounding the interpretation of smaller effect size CRE-targeting gRNAs. Each gRNA transfection was therefore performed 4 times; two cultures were harvested at day 4 and two cultures harvested at day 8 post transfection. RNA was then isolated from cells treated with CRISPRa directed by a single guide via the following method. Cell pellets were lysed in Trizol and BCP was added to the suspension. This mixture was then spun down at 10,000 x g to allow for phase separation. The upper phase containing the RNA was then recovered and isopropanol and glycogen (50 μg/mL) were added to cause the RNA to precipitate. This suspension was left overnight at −20 C to allow precipitation. The RNA was then pelleted, supernatant was removed, and the RNA was resuspended in 50 μL of water. cDNA was generated from this RNA via reverse transcription using Maxima H Minus Reverse Transcriptase (Thermo Scientific, EP0751) according to manufacturer’s instructions. The cDNA was then analyzed using SYBR master mix (Applied Biosystems, 43-676-59) on a 7500 Fast Real Time system. Primers for p14ARF, p15INK4B, p16INK4A, ANRIL (Table S1) and GAPDH were used to measure their respective expression levels via qRT-PCR in technical triplicate. The expression levels of all transcripts were normalized to GAPDH and safe harbor guide controls (see “Data processing and analysis”). Total RNA-seq Total RNA was extracted and sequenced by Azenta (South Plainfield, NJ, USA) following their rRNA depletion protocol. As described in the Azenta protocol, RNA was extracted from frozen cell pellets using Trizol (Thermo Fisher Scientific) (per manufacturer’s instructions). RNA was quantified with Qubit 2.0 Fluorometer (ThermoFisher) and integrity checked with 4200 TapeStation (Agilent Technologies). Libraries were then prepared by using QIAGEN FastSelect rRNA HMR Kit. Library preparation uses "NEBNext Ultra II RNA Library Preparation Kit for Illumina," per manufacturer’s recommendations. In brief, RNAs are fragmented for 15 min at 94°C. First strand and second strand cDNA are ynthesized. cDNA fragments are end repaired and adenylated at 3′ends. Universal adapters are ligated to cDNA fragments, followed by index addition and library enrichment with limited cycle PCR. Sequencing libraries were validated using Agilent Tapestation 4200 and quantified with Qubit 2.0 Fluorometer and qPCR (KAPA Biosystems). The sequencing libraries were multiplexed and clustered on a lane of a flowcell. The flowcell was loaded on an Illumina HiSeq (4000/ equivalent) per manufacturer’s instructions. Libraries were sequenced as 2 × 150 Paired End (PE). Image analysis and base calling were conducted with HiSeq Control Software (HCS). Raw sequence data (.bcl files) generated from Illumina HiSeq were then converted into fastqs and de-multiplexed using bcl2fastq (v2.20, Illumina). One mismatch was allowed for index sequence identification. H3K27ac ChIP-seq H3K27ac ChIP experiments were performed to ENCODE standards. Samples from WI-38 at PDLs 53.5 and 54 were pooled for ChIP to create sufficient material for bulk ChIP-seq, and are referred to in the manuscript as PDL54 (Table S2). The antibody to H3K27ac used was a polyclonal from Active Motif: part number is 39133, lot number 31814008. The histone ChIP-seq experimental protocol is available at the ENCODE portal: https://www.encodeproject.org/documents/be2a0f12-af38-430c8f2d-57953baab5f5/@@download/attachment/Epigenomics_Alternative_Mag_Bead_ChIP_Protocol_v1.1_exp.pdf. Additional assay details are available at https://www.encodeproject.org/chip-seq/histone/. All commercial histone antibodies were validated by at least two independent experiments. Histone mark antibody standards are available at the ENCODE portal: https://www.encodeproject.org/documents/4bb40778-387a-47c4-ab24-cebe64ead5ae/@@download/attachment/ENCODE_Approved_Oct_2016_Histone_and_Chromatin_associated_Proteins_Antibody_Characterization_Guidelines.pdf. The quality of the histone ChIP-seq data was determined using the read depth, number of uniquely mapping reads over the total number of reads (i.e., non-redundant fraction, NRF), and two PCR bottlenecking coefficients (PBC1 and PBC2). CTCF ChIP-seq CTCF ChIP experiments were performed to ENCODE standards. Material from PDLs 53.5 and 54 was pooled as for the H3K27ac ChIP (Table S2). The antibody to CTCF used was a polyclonal antibody from Diagenode: part number is C15410210, lot number is A2359-00234P. All CTCF ChIP-seq protocols are available at the ENCODE portal: 1) https://www.encodeproject.org/documents/20ebf60b-4009-4a57-a540-8fd93407eccc/@@download/attachment/Epigenomics_CR_ChIP_Protocol_v1.0.pdf, 2) https://www.encodeproject.org/documents/6ecd8240-a351-479b-9de6-f09ca3702ac3/@@download/attachment/ChIP-seq_Protocol_v011014.pdf, 3) https://www.encodeproject.org/documents/a59e54bc-ec64-4401-8cf6-b60161e1eae9/@@download/attachment/EN-TEx%20ChIP-seq%20Protocol%20-%20Myers%20Lab.pdf, and 4) https://www.encodeproject.org/documents/f2aa60f2-90a6-4e4b-863a-c6831be371a2/@@download/attachment/ChIP-Seq%20Biorupter%20Pico%20TruSeq%20protocol%20for%20Syapse-c5bdc444fe0511e69d6a06346f39f379.pdf. Additional ChIP-seq protocol details are available at https://www.encodeproject.org/chip-seq/transcription_factor/. The quality of the ChIP-seq data was determined using the read depth, NRF, two PCR bottlenecking coefficients (PBC1 and PBC2), replicate concordance (i.e., IDR), and FRiP. Region capture micro-C Crosslinking Region capture micro-C (RCMC) was performed as described in Kim et al.62,148 Briefly, trypsinized cells were doubly crosslinked to fix protein-protein and protein-DNA interactions using 3 mM disuccinimidyl glutarate (ThermoFisher, 20593) and 1% methanol-free formaldehyde (ThermoFisher, 28906) in 1× PBS (ThermoFisher, 10010023), respectively. The crosslinking reaction was gently mixed at room temperature for 35 min, after which formaldehyde was added to a final concentration of 1%. The double crosslinking reaction was mixed at room temperature for an additional 10 min before quenching with Tris buffer pH 7.5 (ThermoFisher, 15567027) at a final concentration of 0.375 M. Crosslinked cells were washed twice with 1× PBS containing 100 μg/mL BSA (NEB, B9000), recounted to quantify any sample loss during fixation and then partitioned into 1M cell aliquots that were pelleted and snap-frozen in liquid nitrogen for storage at −80°C. MNase digestion Cell membranes were solubilized to extract intact nuclei by resuspending crosslinked 0.5–1 M cell pellets in Micro-C Buffer #1 (MB#1; 50 mM NaCl, 10 mM Tris-HCl pH 7.5, 5 mM MgCl2, 1 M CaCl2, 0.2% NP-40 Substitute (Abcam, ab142227), 1× Protease Inhibitor Cocktail (Sigma-Aldrich, 5056489001)) at 500 μL for 20 min on ice. Following an MB#1 wash, samples were fully resuspended in 500 μL MB#1. MNase titrations were then performed using 5% of the resuspended volume. The titration involved MNase digestion of 1–2% of the resuspended samples using varying concentrations of diluted MNase, followed by crosslink reversal, DNA purification, and D1000 HS TapeStation analysis to visualize and quantify the size distribution of mono- and di-nucleosome fragments. Based on the titration results, the optimal amount of MNase (NEB, M0247) was added. This digestion reaction was mixed at 37°C for 20 min on a thermomixer before being quenched with 20 mM EGTA (bioWORLD, 40520008) and heat inactivated at 65°C for 10 min. Digested nuclei were washed twice with ice-cold Micro-C Buffer #2 (50 mM NaCl, 10 mM Tris-HCl pH 7.5, 10 mM MgCl2 and 100 μg/mL BSA). End repair and labeling To generate blunt ends on digested DNA fragments before proximity ligation and add biotinylated nucleotides, a series of enzymatic processing steps were performed. First, to catalyze the addition of 5′-phosphate groups and the removal of 3′-phosphate groups, digested samples generated from ∼1 M cell inputs were incubated in end-repair reactions (25 U T4 Polynucleotide Kinase (NEB, M0201), 50 mM NaCl, 10 mM Tris-HCl pH 7.5, 10 mM MgCl2, 100 μg/mL BSA, 2 mM ATP (NEB, P0756) and 5 mM DTT (Sigma-Aldrich, 10197777001), in water) at 37°C for 15 min while mixing. To create 5′ fragment overhangs for end-blunting and labeling, 25 U DNA Polymerase I Klenow Fragment (NEB, M0210) was added to the reaction and incubated at 37°C for 15 min while mixing. Next, a mixture of dNTPs in end-labeling buffer (66 μM each of dTTP (Invitrogen, 10297018), dGTP (Invitrogen, 10297018), biotin-dATP (Jena Bioscience, NU-835-BIO14), biotin-dCTP (Jena Bioscience, NU-809-BIOX) and 100 μg/mL BSA in 1× T4 DNA Ligase Buffer) was added to the reaction. This reaction was incubated at room temperature for 45 min with interval mixing before being quenched by 30 mM EDTA (Invitrogen, #15575020) and heat inactivated at 65°C for 20 min. Finally, end-blunted and biotin-labeled nuclei were washed once with Micro-C Buffer #3 (50 mM Tris-HCl pH 7.5, 10 mM MgCl2 and 100 μg/mL BSA). Proximity ligation and removal of unligated biotin Proximity ligation was performed by incubating labeled chromatin in a ligation reaction (10,000 U T4 DNA Ligase (NEB, M0202), 1× T4 DNA Ligase Buffer, 100 μg/mL BSA) at 25°C for overnight with gentle mixing. To remove biotinylated dNTPs from all unligated fragment ends, samples were digested by 1,000 U Exonuclease III (NEB, M0206) in reaction buffer (1× NEBuffer #1 and 100 μg/mL BSA in water) at 37°C for 15 min with interval mixing. DNA purification and size selection To prepare ligated DNA for library generation, DNA was reverse crosslinked and proteins and RNA were digested by adding 100 μL ProK Buffer (20 mM Tris-HCl pH7.5, 100 mM NaCl, 1 mM EDTA, 1 mM EGTA, 0.5% Tx-100, 0.1% SDS (Invitrogen, AM9822), 24 U Proteinase K (NEB, P8107), and 100 μg/mL RNaseA (Invitrogen, EN0531), in water) to the samples and incubating at 65°C overnight. DNA was extracted using the Zymo DNA Clean & Concentrator kit (Zymo Research, D4034) according to the kit manual. Dinucleosome-sized DNA fragments (250–350 bp) were isolated by extraction from a 1% agarose gel (VWR, 97062). Gel extracts were purified using the Zymo Gel Purification kit (Zymo Research, D4008), and samples were quantified by Qubit 1× dsDNA High Sensitivity Assay (Invitrogen, Q33231). Sample ends were polished and blunted again using the Quick Blunting Kit (NEB, E1201) at 25°C for 45 min, followed by reaction inactivation at 65°C for 10 min. Ligated DNA contact fragments were isolated by pulling down biotin-bound fragments using Dynabeads MyOne Streptavidin T1 (Invitrogen, 65601). DNA samples were bound to beads in a Binding and Wash Buffer (1 M NaCl, 5 mM Tris-HCl pH 7.5, 500 μM EDTA, 0.1% Tween 20 (Sigma-Aldrich, P8074) in water) at room temperature for at least 30 min with mixing. After two washes with the Binding and Wash Buffer, the bead-bound samples were washed once with 10 mM Tris-HCl pH 7.5 before library preparation. Library preparation Illumina library preparation was performed using the NEBNext Ultra II kit (NEB, E7645) and NEB Multiplex Oligos for Illumina Primer Set 1 (NEB no. E7335) to end-repair, A-tail, and adaptor ligate the bead-bound samples. All steps were performed as directed by the manual, except that incubations included interval shaking (1 min on, 3 min off) at 1,000 rpm. Sample washes were performed using Binding and Wash Buffer and Tris Buffer (10 mM Tris-HCl pH 7.5 in water). After adaptor ligation, samples were amplified by PCR for 7–8 cycles and then purified using AMPure XP beads. Capture of target loci and sequencing 80-mer probes were designed to tile end-to-end without overlap across the CDKN2A capture loci (∼2.3Mb) through Twist Bioscience. Probes with high predicted likelihoods of off-target pull-down (for example, such as those in high-repeat regions) were masked and removed from the probe tiling, and probe coverage was double-checked to ensure the inclusion of key genomic features (e.g., all promoters and CTCF/H3K4me3 peaks in the locus) before finalization. Probe panels were synthesized and purchased as Custom Target Enrichment Panels from Twist Bioscience. Capture enrichment was performed in accordance with Twist Bioscience’s Target Enrichment Standard Hybridization v1 Protocol. Briefly, pooled sample libraries were dried and mixed with Hybridization Mix (Twist Bioscience, 104178), Custom CDKN2A Panels (Twist Bioscience, 101001) and Universal Blockers (Twist Bioscience, 100578). The library pool was hybridized to the biotinylated probe panel overnight, after which streptavidin beads (Twist Bioscience, 100983) were used to pull down probes with hybridized ligated fragments and then washed (Twist Bioscience, 104178) to remove unbound fragments. Bead-bound libraries were amplified by PCR using the Equinox Library Amplification Mix (Twist Bioscience, 104178) for 9 cycles and purified using AMPure XP beads. Pooled libraries were sequenced by paired-end 2 × 38 cycle sequencing kits with Illumina NextSeq500/550 system by the Broad Institute of MIT and Harvard’s Walk-Up Sequencing services. Quantification and statistical analysis SNP, eQTL, LD, and conservation score analysis Fine mapped variants associated with common diseases were collated from UK biobank (UKBB) fine mapped variant catalog from the Finucane lab (https://www.finucanelab.org/data). Lead SNPs from cancer germline GWAS studies were collated from the EBI GWAS catalog on 11/03/2024 (e113, https://www.ebi.ac.uk/gwas) and overlapped into a single track. Fine mapped eQTL were collated from GTEx V10.40 We selected the following tissues (GTEx abbreviation).-CDKN2A eQTLs: Brain - Anterior cingulate cortex (BRNACC), Brain - spinal cord (BRNSPC), Cultured fibroblasts (FIBRBLS), Whole blood (WHLBLD);-CDKN2B eQTLs: Brain - cortex (BRNCTXA), Esophagus - mucosa (ESPMCS), Muscle - Skeletal (MSCLSK), Skin - sun exposed (SKINS), Whole Blood (WHLBLD), Colon - sigmoid (CLNSGM).-ANRIL (CDKN2B-AS1) eQTLs: Testis (TESTIS).-MTAP eQTLs: Aorta (ARTAORT), Tibial Artery (ARTTBL), Hippocampus (BRNHPP), Breast mammary tissue (BREAST), Left ventricle (HRTLV), Tibial Nerve (NERVET), Not sun-exposed skin (suprapubic) (SKINNS), Thyroid (THYROID), Whole Blood (WHLBLD). -CDKN2A eQTLs: Brain - Anterior cingulate cortex (BRNACC), Brain - spinal cord (BRNSPC), Cultured fibroblasts (FIBRBLS), Whole blood (WHLBLD); CDKN2A eQTLs: Brain - Anterior cingulate cortex (BRNACC), Brain - spinal cord (BRNSPC), Cultured fibroblasts (FIBRBLS), Whole blood (WHLBLD); -CDKN2B eQTLs: Brain - cortex (BRNCTXA), Esophagus - mucosa (ESPMCS), Muscle - Skeletal (MSCLSK), Skin - sun exposed (SKINS), Whole Blood (WHLBLD), Colon - sigmoid (CLNSGM). CDKN2B eQTLs: Brain - cortex (BRNCTXA), Esophagus - mucosa (ESPMCS), Muscle - Skeletal (MSCLSK), Skin - sun exposed (SKINS), Whole Blood (WHLBLD), Colon - sigmoid (CLNSGM). -ANRIL (CDKN2B-AS1) eQTLs: Testis (TESTIS). ANRIL (CDKN2B-AS1) eQTLs: Testis (TESTIS). -MTAP eQTLs: Aorta (ARTAORT), Tibial Artery (ARTTBL), Hippocampus (BRNHPP), Breast mammary tissue (BREAST), Left ventricle (HRTLV), Tibial Nerve (NERVET), Not sun-exposed skin (suprapubic) (SKINNS), Thyroid (THYROID), Whole Blood (WHLBLD). MTAP eQTLs: Aorta (ARTAORT), Tibial Artery (ARTTBL), Hippocampus (BRNHPP), Breast mammary tissue (BREAST), Left ventricle (HRTLV), Tibial Nerve (NERVET), Not sun-exposed skin (suprapubic) (SKINNS), Thyroid (THYROID), Whole Blood (WHLBLD). eQTL were computed for each tissue independently by the GTEx V10 pipeline, and overlapped these into tracks for each transcript. We calculated LD from the UKBB (337K British-ancestry individuals), separating blocks when Rˆ2 > 0.580,81. Code to generate individual eQTL and GWAS tracks is available at https://github.com/broadinstitute/CDKN2A-gene-regulation. Additional code for generating figures is available at https://github.com/broadinstitute/pyqtl. For the analysis in Figure 6F and S14: Conservation scores correspond to phastCons100way score (0–1, clear to dark green, from https://hgdownload.soe.ucsc.edu/goldenPath/hg38/phastCons100way/hg38.phastCons100way.bw).149 Common SNPs were downloaded from Ensembl GRCh38150 (http://hgdownload.soe.ucsc.edu/gbdb/hg38/snp/dbSnp155Common.bb). Analysis of 9p21 genes in bulk RNA-seq (GTEx) and CAGE (FANTOM) Expression analysis of p14ARF and p16INK4A from isoform estimates is feasible in GTEx V8 dataset40,53,151 due to the exceptionally deep sequencing of mRNA-seq libraries, which allows detection of exon 1α and exon 1β. For GTEx V8, sequencing generated 76bp paired-end reads and an eight-base index barcode read, and was run with a coverage goal of 50M reads (the median achieved was ∼82M total reads)(https://www.gtexportal.org/home/methods#AboutSamples). Expression of p14ARF and p16INK4A from isoform estimates is feasible from the FANTOM5 consortium CAGE data (Cap analysis of gene expression) as the method focuses on transcript 5′ ends and thus captures exon 1α and exon 1β.41 Data preparation for GTEx: Bulk mRNA analysis We extracted isoform and gene expression estimates for 9p21 transcripts from 17,382 GTEx samples (across 948 donors and 54 tissue types). Specifically, we extracted isoform expression estimates in TPMs for ENST00000579755.1 and ENST00000304494.9 (for CDKN2A-p14 and CDKN2A-p16 respectively) from “GTEx_Analysis_2017-06-05_v8_RSEMv1.3.0_transcript_tpm.gct.gz” for all (https://gtexportal.org/home/downloads/adult-gtex/bulk_tissue_expression). We further extracted gene TPMs of CDKN2A, CDKN2B-AS1, CDKN2B, DMRTA1, and MTAP from “GTEx_Analysis_2017-06-05_v8_RNASeQCv1.1.9_gene_tpm.gct.gz” (https://gtexportal.org/home/downloads/adult-gtex/bulk_tissue_expression). We merged the tables for these 7 transcripts of interest, and used GTEx demographics data to assign age-groups to all donors, as discussed below. Data preparation FANTOM5: Bulk CAGE analysis We downloaded the hg38 table listing TPM values for each CAGE peak (“hg38_fair+new_CAGE_peaks_phase1and2_tpm.osc.txt”) from https://fantom.gsc.riken.jp/5/datafiles/reprocessed/hg38_v9/extra/CAGE_peaks_expression/. We extracted from TPM values for “Human tissue” samples (“00_human.tissue.hCAGE.hg38.assay_sdrf.txt”) from https://fantom.gsc.riken.jp/5/datafiles/reprocessed/hg38_v9/basic/human.tissue.hCAGE/. We excluded samples annotated as “Newborn”, “Fetal” and control samples (i.e., “Clontech Human Universal Reference Total RNA”, “SABiosciences XpressRef Human Universal Total RNA” and “Universal RNA - Human Normal Tissues Biochain”). If > 1 technical replicate was available we averaged TPM values across technical replicates. Further, we also averaged across donors when available. This allows us to obtain one value of TPM per transcript per tissue ID/location (n = 101), matching the analysis of GTEx samples in Figure 1D. The expression of the following CAGE peaks was used to estimate the expression of the 9p21 genes: hg19::chr9:21974820..21974906,-;hg_100253.1 was used to estimate p16INK4A, hg19::chr9:21994344..21994440,-;hg_100262.1 was used to estimate p14ARF, hg19::chr9:21994775..21994834,+;hg_96303.1 was used to estimate ANRIL (CDKN2B-AS1), hg19::chr9:22009297..22009324,-;hg_100272.1 was used to estimate p15 INK4B (CDKN2B). Data plotting (1)GTEx data heatmap (1)GTEx data heatmap GTEx data heatmap We calculated the mean values across donors per tissue for 54 tissues, and plotted these mean values as a heatmap (Figure 1B). We calculated the ratio of p14ARF/p16INK4A across the plot by removing samples where p14ARF and p16INK4A are both 0, and computing the ratio p16INK4A/p14ARF for remaining 15,348 out of 17,382 samples (Figure S1B). We plotted the data using ComplexHeatmap 2.22.0152 in R 4.4.2.(2)GTEx violin plots (2)GTEx violin plots GTEx violin plots We obtained log expression values by taking natural log of 1+TPM counts. Within each tissue, we plotted the log expression values for the genes of interest (CDKN2A, CDKN2A-p14, CDKN2A-p16, CDKN2B-AS1, CDKN2B, and MTAP) grouped by age groups (Figure S3). Age groups are defined as intervals of the form (X, Y], with X being excluded and Y included, e.g., X < age ≤ Y. The only exception is the [20, 30] age group, which includes all donors whose age is between 20 and 30 (both extremes included). Difference between expression values from individuals in the 20–30 years range compared to the 60–70 age range was calculated in python143 using the ttest_ind function (with equal_var = False, alternative = ‘two-sided’) from the scipy.stats package (v1.10.0).142 P-values from comparisons of all genes in 11 tissues of interest (Adipose - Visceral (Omentum), Artery - Aorta, Small Intestine - Terminal Ileum, Esophagus - Gastroesophageal Junction, Heart - Atrial Appendage, Lung, Prostate, Spleen, Thyroid, Whole Blood, Adrenal Gland) were corrected using Bonferroni correction (method = “Bonferroni”) via the multipletests function in statsmodels.stats.multitest (v0.13.5). Violin/scatterplots were generated using seaborn (v0.12.2).140 Number of data point in each group is listed in figure legends.(3)GTEx and FANTOM Upset plots (3)GTEx and FANTOM Upset plots GTEx and FANTOM Upset plots We set a threshold of ≥1 TPM to classify a gene as expressed in a given tissue using the average across donors for a given tissue in each dataset. To plot the data we used UpSetR (v 1.4.0).95 Gene track and ENCODE cCRE track To obtain a gene track which matched our definition of p14ARF and p16INK4A as ENST00000579755.1 and ENST00000304494.9, we modified the refGene.txt.gz annotation track (obtained from https://hgdownload.soe.ucsc.edu/goldenPath/hg38/database/). In particular, CDKN2A annotations were substituted by ENST00000579755.1 and ENST00000304494.9 (from Gencode v34) via custom code. Moreover, we simplified the track by displaying only the NR_003529 gene annotation for ANRIL, only NM_004936 for CDKN2B (p15INK4B), and removing the still poorly-characterised CDKN2A-DT. The ENCODE v365 cCRE track for hg38 was downloaded from the UCSC genome browser at: https://hgdownload.soe.ucsc.edu/gbdb/hg38/encode3/ccre/encodeCcreCombined.bb and modified to show the “ccre” ENCODE classification as label. Single-cell expression data analysis Custom GTFs To distinguish p14ARF and p16INK4A from 5′ enriched single-cell gene expression profiles, we generated custom versions of GTF (Gene Transfer Format) files containing the annotation of gene and transcript structures. We started with the GTF file for reference chromosomes only downloaded from the Gencode Release 34 (GRCh38.p13) Website (https://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_34/gencode.v34.annotation.gtf.gz). We then removed from the CDKN2A gene annotation all isoforms except: ENST00000304494.9 (canonical p16 transcript, shown in ENSEMBL as having a Transcript Support Level = 1). ENST00000579755.1 (canonical p14 transcript, TSL = 1). We then modified their entries so these transcripts would not be considered isoforms but individual genes, modifying both the gene_id entry (from gene_id “ENSG00000147889.17″ to gene_id “ENSG00000147889.17_p16″ and gene_id “ENSG00000147889.17_p14″, respectively) and the gene_name entry (from gene_name “CDKN2A” to gene_name “CDKN2A_p16″ and gene_name “CDKN2A_p14″, respectively). All other fields were maintained the same. We used the new GTF file, in combination with the Genome sequence, primary assembly only, fasta file downloaded from the same Gencode Release 34 Website https://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_34/GRCh38.primary_assembly.genome.fa.gz to generate a custom STAR solo reference, using the wrapper provided through Cumulus98 in Terra (https://app.terra.bio/).96,97,98 STAR version 2.7.10a_alpha_220818 was used for this step. This reference was used to remap published 5′ scRNA-seq datasets, as described below. To test whether it was possible to distinguish reads mapping to the transcript specific exons in 10x 3′ data (without targeted amplification), i.e., how many reads map to exon1α (p16INK4A) or exon 1β (p14ARF) versus the common exons, i.e., exons 2–3, we also generated a further modified version of the GTF above. ENSE00001833804.1 (exon1α, p16) and ENSE00002698483.1 (exon 1β, p14) were considered as two independent genes, names CDKN2A_p16_ex1 and CDKN2A_p14_ex1, respectively. One additional gene (CDKN2A_p14_p16) was created combining the annotations for exons 2 and 3 for p14 and p16, i.e., ENSE00003496053.1, ENSE00003529527.1, ENSE00003513347.1, ENSE00003565933.1. We used this version of the GTF to generate another custom STAR reference (following the same steps described above), which was used only to map the breast tissue single-cell expression profiles in Figure 3 (and obtain the estimates in Figure 3C). For all other quantifications of p14/p16 from 3′ libraries (10x or Seq-Well), which used target amplification, we generated an additional custom GTF tailored to the primer positioning and read structure. In this dataset, p14 reads were expected to fully align within exon 1β due to the location of the p14ARF primer, while p16INK4A reads spanned the splice junction at exon 1α, requiring inclusion of exons 1α and 2 to enable accurate alignment and quantification. Because the reads were 50 bp in length and included the primer sequence, they did not extend beyond exon 1 (p14ARF) or exon 1/2 (p16INK4A), though exon 3 is still included as part of the p16INK4A gene annotation. The target amplification GTF contained only gene, transcript, and exon features, with simplified gene names (“p14” and “p16”) and restricted coordinates (p14ARF defined from chr9: 21,994,139–21,994,624). All custom GTFs are provided at https://github.com/broadinstitute/CDKN2A-gene-regulation. Analysis of published single-cell gene expression profiles Single-cell 5′ profiles across 15 organs To identify how 9p21 transcripts are expressed across different cell types and tissues, we remapped single-cell expression profiles obtained with 10x 5′ assays from He et al.47 and available in Gene Expression Omnibus (GEO) through the accession code: GSE159929. Fastqs of the cDNA libraries were downloaded from SRA (accession code: SRP292721). We used STAR solo, version 2.7.10b_alpha_230301 (through the wrapper in Cumulus, specifying assay “tenX_5p”) to map the fastqs to the custom GTF. The mapped matrices were then aggregated. We filtered the aggregated matrix to include only the 84,363 cells deemed as high quality in the original publication, identified as the cells whose barcode was included in the metadata file Annotation_AHCA_alltissues_meta.data_84363_cell.txt downloaded from Github (Oct 2nd, 2023) from the repository https://github.com/bei-lab/scRNA-AHCA/blob/master/Cell_barcode_and_corresponding_cell_types_of_AHCA/. The metadata file included also cell annotations (“Cell_type_in_merged_data”) and tSNE coordinates (“tSNE_1”, “tSNE_2”). We generated a table listing the cell type annotation, organ of origin, and raw remapped expression of CDKN2A_p14, CDKN2A_p16, CDKN2B, and CDKN2B-AS1 for each cell. The cell type annotations from He et al., 2020 (43 categories) were merged in higher-level categories when the difference was only due to the expression of specific markers (e.g., cells labeled as “B Cell CD79A”, “B Cell MS4A1″ and “B Cell VPREB3″ were annotated as “B Cell”). Doing so, we obtained the 24 “broader” annotations shown in Figure S4B. We defined cells as expressing the 9p21 transcripts if they displayed at least one read mapped to that gene. We generated a table with the percentage of cells in each cell type and organ that expressed the 9p21 transcripts (by dividing the number of expressing cells by the total number of cells in each cell type for each organ). We removed from the table values from cell types accounting for less than 0.25% of the total number of cells in each organ. The filter allowed us to exclude inflated percentages of expressing cells due to the low total cells. Heatmaps were generated in R version 4.4.2 using ComplexHeatmap version 2.22.0. Code to reproduce these plots is available on https://github.com/broadinstitute/CDKN2A-gene-regulation. Single-cell 5′ profiles across gut To validate cell type/tissue specific patterns of expression for 9p21 transcripts, we remapped publicly available 5′ single-cell gene expression profiles from gut tissues from Elmentaite et al.48 We downloaded published fastqs (from 10x 5′ gene expression v2) from ArrayExpress (https://www.ebi.ac.uk/arrayexpress) at accession number E-MTAB-9543 (“end bias”: “5 prime tag”). We used STAR solo, version 2.7.10b_alpha_230301 (through the wrapper in Cumulus, specifying assay “tenX_5p”) to map the fastqs to the custom GTF. The mapped matrices were then aggregated. We extracted the raw count values for the 9p21 transcripts, and focused only on the cells that passed QC in the published analysis, identified as cells present in the normalised processed h5ad file associated with the publication, downloaded from the Gut cell atlas website at https://cellgeni.cog.sanger.ac.uk/gutcellatlas/Full_obj_log_counts_soupx_v2.h5ad. From this h5ad file, we also extracted the cell metadata to allow us to focus on cells coming from adult tissues (‘Age_group' classified as “Adult”, or “Adult_MLN”), resulting in 82,228 high quality cell profiles from 5 different donors (sampled at different gut tissue locations). We have calculated the number of cells expressing the different 9p21 transcripts per cell type, using the cell types annotation from the published object. We defined cells as expressing the genes of interest if they displayed at least one read mapped to that gene. We computed the number of expressing cells for each cell type and donor (noting that each donor was sampled at different gut locations). In the top plot in Figure 2C, ‘Immune cells’ was a new group created aggregating the annotations of ‘B cells', ‘T cells', ‘Plasma cells', ‘Myeloid' (from the ‘category' column in the published.obs file). Original annotations were then showed separately in the bottom plot. Bar plot showing mean values and SEM across donors were generated in python using matplotlib version 3.6.3. Code to reproduce these plots is available on https://github.com/broadinstitute/CDKN2A-gene-regulation. Single-cell 5′ profiles across heart We also reprocessed publicly available 5′ datasets from heart tissues from Koenig et al.49 Published fastq files from single-cell gene expression profiles from healthy donors and individuals with dilated cardiomyopathy, were obtained from SRA under accession number SRP336428. We restricted our analysis on 5 of the 7 single-cell libraries, as the fastq files for two samples (TWCM-H6-lib1 and TWCM-H7-lib1) appeared truncated. Raw fastqs were remapped with STAR solo, version 2.7.10b_alpha_230301 (through the wrapper in Cumulus, with assay “tenX_5p_pe”) to the custom GTF. We extracted the raw count values for the 9p21 transcripts, and focused only on the cells that passed QC in the published analysis, identified as cells present in the normalised processed file (“GSE183852_DCM_Cells.Robj.gz”), downloaded from GEO (GSE183852) and converted to a h5ad file using SeuratDisk v0.0.0.9015 (https://github.com/mojaveazure/seurat-disk, June 2023).120,121,122,123,124 As before, we have calculated the number of cells expressing the different 9p21 transcripts per cell type and donor. We defined cells as expressing the genes of interest if they displayed at least one read mapped to that gene. Bar plot showing mean values and SEM across donors were generated in python using matplotlib version 3.6.3.139 For ease of comparison across datasets, some of original annotations were refined (“Endothelium” cells were relabeled as “Endothelial”, “Neurons”: “Neuronal”), and other annotations were grouped. In particular.-“Fibroblasts”, “Smooth_Muscle” and “Pericytes” were grouped into “Mesenchymal” cells;-“B_Cells”, “T_Cells”, “NK_Cells”, “Macrophages”, “Monocytes” into “Immune” cells. -“Fibroblasts”, “Smooth_Muscle” and “Pericytes” were grouped into “Mesenchymal” cells; “Fibroblasts”, “Smooth_Muscle” and “Pericytes” were grouped into “Mesenchymal” cells; -“B_Cells”, “T_Cells”, “NK_Cells”, “Macrophages”, “Monocytes” into “Immune” cells. “B_Cells”, “T_Cells”, “NK_Cells”, “Macrophages”, “Monocytes” into “Immune” cells. In the Immune-focused bar plot, “Macrophages” and “Monocytes” were grouped into “Myeloid”. We removed from the plot values arising from cell types with <5 cells. The only value impacted was the percentage of B cells expressing p15 in HDCM5: one cell out of 4 B cells displayed p15 expression, resulting in an inflated percentage of p15-expressing cells. Code to reproduce these plots is available on https://github.com/broadinstitute/CDKN2A-gene-regulation. Single-cell 5′ profiles across immune cells We also reprocessed publicly available 5′ data from immune cells from Hashimoto et al.50 and from the Resource section of the website of 10x Genomics Fastq files related to 5′ gene expression profiling of CD4 T cells from 2 supercentenarians were obtained, upon approval, from the NBDC Human Database https://humandbs.biosciencedbc.jp/en/hum0229-v1 and remapped with STAR solo, version 2.7.10b_alpha_230301 (through the wrapper in Cumulus, with assay “tenX_5p_pe”) to the custom GTF. We restricted our analysis to the cells in the processed files in the original publication, identified as the ones presents in the processed “barcodes.tsv.gz” files in the archives files (“04.SC1.tar” and “05.SC2.tar”) downloaded from http://gerg.gsc.riken.jp/SC2018. As cell types annotations were not available, labels were reassigned via Leiden clustering (with resolution = 0.01 for SC1 and 0.05 for SC2) using the expression of genes: CD3D, CD4 and CD8A for T cells, HBA1 for Erythrocytes, KLRF1 for NK cells (some of which were shown in the original publication) as markers. As before, we have then calculated the number of cells expressing the different 9p21 transcripts per cell type and donor. We defined cells as expressing the genes of interest if they displayed at least one read mapped to that gene. Bar plot showing mean values and SEM across the two donors were generated in python using matplotlib version 3.6.3. Fastq files related to 5′ gene expression profiling of peripheral blood mononuclear cells (PBMCs) from a human healthy donor were downloaded from the 10x website at the link below: (https://www.10xgenomics.com/resources/datasets/10k-human-pbmcs-5-v2-0-chromium-x-with-intronic-reads-2-standard and remapped with STAR solo, version 2.7.10b_alpha_230301 (through the wrapper in Cumulus, with assay “tenX_5p”) to the custom GTF. We restricted the analysis to the barcodes in the filtered file “10k_PBMC_5pv2_nextgem_Chromium_X_intron_10k_PBMC_5pv2_nextgem_Chromium_X_intron_count_sample_feature_bc_matrix.h5” downloaded from the same link. We further removed cells that displayed expression in less than 200 genes, less than 1,000 or more than 20,000 counts, and ≥30% of counts mapping to mitochondrial genes: 6,823 cells passed these filters and were included in the analysis. As cell type annotations were unavailable, we assigned cells to cell types through Leiden clustering (resolution = 0.099) and exploration of marker genes. We manually annotated the clusters as T cells, NK cells, B cells and Myeloid. As before, we have then calculated the number of cells expressing the different 9p21 transcripts per cell type. We defined cells as expressing the genes of interest if they displayed at least one read mapped to that gene. Bar plot showing expressing cells per cell type was generated with matplotlib version 3.6.3. Single-cell 3′ profiles across breast tissue To assess how many reads from the original Gray et al. breast tissue dataset55 were mapping to exon1α (p16INK4A), exon 1β (p14ARF), or exon2&3, we downloaded fastq files from the Synapse repository (accession: syn26560310 in Oct 2022), https://www.synapse.org/Synapse:syn26560310, subfolder “single_cell_RNAseq_level_1”. We used STAR solo, version 2.7.10b_alpha_230301 (through the wrapper in Cumulus, specifying assay “tenX_v2”) to map the fastqs to the second custom GTF, generated as described above. For the dialup reads, we used the dialup gtf described above and mapped the fastqs from the dialup via STAR solo, version 2.7.10a_alpha_220818, with the flags: --outSAMtype BAM SortedByCoordinate --quantMode TranscriptomeSAM GeneCounts --alignSoftClipAtReferenceEnds Yes --outSAMstrandField intronMotif --outFilterMismatchNoverLmax 0.1 --chimMainSegmentMultNmax 1 --soloType CB_UMI_Simple --soloCBwhitelist 737K-august-2016.txt --soloCBmatchWLtype 1MM_multi_Nbase_pseudocounts --soloUMIfiltering MultiGeneUMI_CR --soloUMIdedup 1MM_CR --clipAdapterType CellRanger4 --outFilterScoreMin 30 --outSAMattributes CR UR CY UY CB UB --outFilterMatchNminOverLread 0.33 --outSAMattrRGline ID:rg1 SM:sm1 --chimJunctionOverhangMin 15 --alignIntronMax 1000000 --chimOutType Junctions WithinBAM SoftClip --outFilterMultimapNmax 20 --limitSjdbInsertNsj 1200000 --alignSJoverhangMin 8 --outFilterType BySJout --outFilterScoreMinOverLread 0.33 --outFilterIntronMotifs None --chimSegmentMin 15 --alignIntronMin 20 --alignSJDBoverhangMin 1 --alignMatesGapMax 999999 --outFilterMismatchNmax 999 --soloCBstart 1 --soloCBlen 16 --soloUMIstart 17 --soloUMIlen 10. Reads mapping to exon 1α and exon 1β were extracted from the mapped counts files and quantified as UMI counts per cell barcode. We restricted the analysis to cells passing QC in the original publication, i.e., the cell barcodes present in the processed file “Figures 1B and 2A_allCells_SingleCellExperiment.rds”, downloaded from Synapse (“single_cell_RNAseq_level_4″), where we also obtained cell type annotations and UMAP coordinates. For the dial-up comparison, we further restricted the analysis to the samples for which we obtained dial-up data, namely WT-A, WT-C, WT-D, BRCA1-C, BRCA1-D, BRCA1-H, BRCA2-A, BRCA2-B, BRCA2-D, and BRCA2-E (10 samples). We calculated the percentage of cells expressing CDKN2A-p16, CDKN2A-p14, or CDKN2A-p14 + p16 in the original dataset and compared this to the percentage of cells expressing p16 and p14 in the dial-up dataset. Pooled expression percentages were calculated across all cells, with the standard error of the mean computed across per-sample percentages to account for inter-sample variability. To examine cell-type-specific expression patterns, we generated dot plots comparing p16 and p14 dial-up expression against standard gene expression measurements for key 9p21 locus genes MTAP, CDKN2A, CDKN2B-AS1, and CDKN2B. For the dial-up analysis, we calculated mean expression ratios and the percentage of expressing cells for p14ARF and p16INK4A transcripts across different cell subtypes, with cell types ordered by their p14ARF/p16INK4A expression ratio in descending order. This same ordering was applied to visualize expression of canonical 9p21 genes. In both dot plots, point size represents the percentage of cells expressing each gene within a given cell subtype, while color intensity indicates the average expression level. Code to reproduce these plots is available at https://github.com/broadinstitute/CDKN2A-gene-regulation. WI-38 replicative senescence timecourse data We reanalyzed data from the time-course experiment of replicative senescence in WI-38 fibroblasts from17 using scanpy 1.9.1.99 Processed h5ad file (“GSE175533_sceasy_hay.h5ad.gz”) was downloaded from GEO, at accession GSE175533. We excluded from our analysis the expression profiles from immortalised hTERT cells (i.e., the cells classified as “tert” in the “exp” metadata column), obtaining a matrix of n = 8,953 cells from PDL25, 29, 33, 37, 46, 50 (1042–2105 cells per PDL). We combined the published expression values with the CDKN2A-p14ARF and p16INK4A values obtained from the dial-up experiments as -described above (Transcript enrichment from single cell libraries). Reads were sequenced with NextSeq and mapped via STAR version 2.7.10a with the flags --outSAMtype BAM SortedByCoordinate --quantMode TranscriptomeSAM GeneCounts --alignSoftClipAtReferenceEnds Yes --outSAMstrandField intronMotif --outFilterMismatchNoverLmax 0.1 --chimMainSegmentMultNmax 1 --soloType CB_UMI_Simple --soloCBwhitelist reference/737K-august-2016.txt --soloCBmatchWLtype 1MM_multi_Nbase_pseudocounts --soloUMIfiltering MultiGeneUMI_CR --soloUMIdedup 1MM_CR --clipAdapterType CellRanger4 --outFilterScoreMin 30 --outSAMattributes CR UR CY UY CB UB --outFilterMatchNminOverLread 0.33 --outSAMattrRGline ID:rg1 SM:sm1 --chimJunctionOverhangMin 15 --alignIntronMax 1000000 --chimOutType Junctions WithinBAM SoftClip --outFilterMultimapNmax 20 --limitSjdbInsertNsj 1200000 --alignSJoverhangMin 8 --outFilterType BySJout --outFilterScoreMinOverLread 0.33 --outFilterIntronMotifs None --chimSegmentMin 15 --alignIntronMin 20 --alignSJDBoverhangMin 1 --alignMatesGapMax 999999 --outFilterMismatchNmax 999 --soloCBstart 1 --soloCBlen 16 --soloUMIstart 17 --soloUMIlen 12. Using published PCA values, we recalculated nearest neighbors (sc.pp.neigbors) and UMAP coordinates to obtain the visualization in Figure 4. To obtain the list of genes most correlated with p14ARF and p16 INK4A, we generated a matrix with the per-cell expression values of highly variable genes (i.e., the 2000 genes identified as ‘vst.variable' genes in the ‘var’ section of the published object), p14ARF and p16INK4A (raw values, and values normalized by total counts). p15INK4B (CDKN2B) is included among the highly variable genes, while ANRIL (CDKN2B-AS1) is not, due to its low detection in this dataset. We then calculated Pearson correlation using pandas (1.4.3) corr(method='pearson’) function.100 The genes most correlated with normalised p14ARF and p16INK4A were their raw version, and CDKN2A. We picked the top 200 genes most correlated with the normalized expression values (excluding p14ARF/16INK4A themselves (raw/normalized), and CDKN2A) and generated signatures. Similarly, we extracted the top200 genes correlated to p15INK4B (excluding CDKN2B itself). Enriched ontologies were obtained using EnrichR.57,101,102 Ontologies shown are from BioPlanet 2019, sorted by p-value ranking. Gene signatures associated with Senescence (SenMayo) and SASP signatures were obtained from,56 Supplemental Data1 table. The SenMayo signature includes all 125 genes from the publication, while SASP genes are 83/125, the ones classified as “Secreted” in the “State” column in the published table. Signature scores were obtained using the “score_genes” function in scanpy. Pseudotime values as in17 were obtained from the authors, for n = 7,489 cells in the primary cluster, as in the original analysis. Heatmap was generated in R 4.4.2, using ComplexHeatmap (2.22.0152). Visual comparisons of read locations from 5′ and 3′ single-cell RNA-seq profiles Examples of single-cell expression profiles from 10x Genomics 3′ and 5′ datasets is shown in Figure S1, to show the absence and presence of reads mapping to exon1α and exon 1β. Datasets shown are: (3′ data) https://www.10xgenomics.com/datasets/4-k-pbm-cs-from-a-healthy-donor-2-standard-1-2-0 and https://www.10xgenomics.com/datasets/frozen-pbm-cs-donor-a-1-standard-1-1-0); 5′ data: https://www.10xgenomics.com/resources/datasets/10k-human-pbmcs-5-v2-0-chromium-x-with-intronic-reads-2-standard and https://www.10xgenomics.com/resources/datasets/5k-human-ovarian-tumor-isolated-with-chromium-nuclei-isolation-kit-3-1-standard. Bam files were downloaded from the 10x website and visualized with IGV. ATAC-seq data processing and analysis ATAC-seq raw reads from Chan et al. were obtained from GEO (GSE175533). The read quality was assessed using FastQC153 and multiQC v1.9.103 The Nextera PE adapters were removed using the Skewer104 tool. The trimmed reads were aligned using bowtie2 v2.4.1105,106 and PCR duplicates were removed using the SAMBLASTER tool.107 Properly paired reads (fragments) with mapping quality MAPQ ≥30 (using parameters within samtools view command -f 2 -q 30) were retained for further processing. The samtools v1.11125,126 was used to sort and index the bam files. The bamCoverage function in deepTools2108 was used to adjust reads due to the offset of Tn5 (+4 on the forward strand, −5 on the reverse strand) and visualization of the samples in bigwig format. Peak calling was performed using macs2 v2.2.7.1109) using the following parameters for the paired-end reads: -f BAMPE -g hs --keep-dup all --cutoff-analysis -p 0.01. The quality control analysis of the ATAC-seq dataset was performed using the ataqv tool v1.2.1.111 To obtain normalised “Count per million” tracks, bigwigs were converted to bedGraphs using bigWigToBedGraph from UCSC tools132,133 was then multiplied by a “scale factor” calculated as “1 million/number of unique aligned fragments (where each fragment is a read pair; number of unique aligned fragments were between 56 and 64 million) to obtain “Counts per million” tracks, and then converted back with bedGraphToBigWig. ChIP-seq data processing and analysis The CTCF and H3K27ac ChIP-seq datasets were processed via the ENCODE uniform pipeline154 Libraries produced from the same ChIP (technical replicates) were merged. Peak calling was performed on CTCF tracks (still within the pipeline) with HOMER version 4.11.1,128 with -style argument “factor”. pos2bed was then used to convert the result to a BED file. Only peaks with positive values for start and end coordinates are kept. The filtered BED was then sorted and any overlapping peaks merged. “Count per million” normalised ChIP-seq tracks were generated from bigwigs, as described above for ATAC-seq tracks. Number of unique aligned fragments were between 48 and 56 million for H3K27ac, and 48 and 66 millions for CTCF. For the analysis of CTCF occupancy in Figure S13B, we first selected a common set of CTCF peaks between conditions. To do so, we first intersected the two peak files for PDL33 (one for each biological replicate) using the bedtools (v2.29.0127) intersect function, with a reciprocal (-r) minimum overlap of 50% required (-f 0.5), to get a set of peaks in both PDL33 replicates. This file was then intersected (at the same conditions) with the peak file from PDL54. Using this set of common peaks, we then used the bedtools map function to calculate the sum of CPM signal overlapping each CTCF peak, genome-wide and in the 9p21 locus (defined as the chr9:21,750,000–22,500,000 interval). The signal in PDL33 (replicate 1) vs. PDL54 was then plotted in R (v4.4.2). CTCF motif identification CTCF motif track for hg38 was generated in R (v4.4.2) following the vignette (version 25/12/2022) in https://dozmorovlab.github.io/CTCF/articles/CTCF.html.135 In brief, we included CTCF sites identified using the all three CTCF position weight matrices (MA0139.1, MA1929.1, MA1930.1), generated from “hg38.JASPAR2022_CORE_vertebrates_non_redundant_v2”, using as p-value threshold 10−4. The GRanges object obtained was filtered to retain only standard chromosomes, sorted, and converted to a BED file. Total RNA-seq data processing Raw fastq reads from total RNA-seq of WI-38 over different PDLs were mapped with Salmon117 v1.10.1 to the hg38 Salmon Transcriptome index (“salmon_sa_index”) downloaded from http://refgenomes.databio.org/v3/genomes/splash/2230c535660fb4774114bfa966a62f823fdb6d21acf138d4. Transcript-level TPM abundances were then aggregated to gene-level abundances in R using the package tximport,118 using as Transcript to Gene mapping the “tx_id” and “gene_id” columns from the EnsDb.Hsapiens.v86 package.119 Mapping from Ensembl Gene IDs to Gene Symbols was performed using a table downloaded from Ensembl BioMart (GRCh38.p13) with ”Gene stable ID” (Ensembl Gene IDs) and ”Gene name” (Gene Symbols), downloaded 3/3/2023. The Senmayo “score” for total RNA-seq data in Figure S10B is the sum of the TPM values of all the Gene Symbols in the list obtained from,56 as described above. Selection of regions for CRISPRa screen experiment The processed shifted bam and macs2 peak calls were used to compute a consensus peak matrix across the dataset using the DiffBind R package.112 DiffBind uses DESeq2129 for the normalization (median of ratios) and differential accessibility analysis, where the time point was used as the covariate in the model. PDL50 was compared to each of the other time points in the dataset and differentially significant regions were called using an FDR/q-value threshold of 0.05 and absolute log2 fold change threshold of log2(1.5). We selected 109 regions either had open chromatin (peaks) or an increase/decrease in the size of the accessible chromatin in the 9p21 TAD. We additionally included a further 66 regions that had no ATAC-seq or H3K27ac peak in WI-38, but were annotated as a ENCODE cCRE in another cell type, and/or overlapped a fine-mapped GWAS variant; these were selected via manual inspection, resulting in a gross total of 175 regions. Two of these had no suitable PAM site for CRISPR targeting and were removed from future analysis, for a total of 173 regions used in the screen experiment. We also ensured that previously validated CREs from HeLa cells were included.24 Bed file listing the 173 screen regions is browsable/accessible via the IGV session at https://github.com/broadinstitute/CDKN2A-gene-regulation. CRISPRa screen data processing and analysis The screen sequencing data processed using custom R scripts, leveraging the R package SCEPTRE for statistical analysis of CRISPR perturbations in single cells.74 All the code is available at https://github.com/broadinstitute/CDKN2A-gene-regulation. First, an RDS object integrating both mRNA and CROP-seq alignment data is created, using standard Seurat (v5) and dplyr functions for filtering, subsetting, and aligning CRISPR guide RNA (gRNA) information with gene expression data. SCEPTRE was then run using the current modular high-MOI workflow (import_data → set_analysis_parameters → assign_grnas → run_calibration_check → run_power_check → run_discovery_analysis). Individual data processing steps were as follows: mRNA Processing Single-cell RNA-seq data from Seq-Well output was mapped via STAR as described in the file screen_alignment.sh in the above GitHub and imported and converted into a Seurat object. Cells were filtered based on the following quality control thresholds: 3 cells per feature, 200 features per cell, 500 ≤ nGenes ≤6000, nUMI ≥750, mitochondrial RNA content <15%. The resulting Seurat object, containing filtered gene-by-cell data, was saved for downstream integration with CRISPR gRNA data. CROP-seq Processing CROP-seq data containing guide RNA (gRNA) information per cell was read in and filtered for high-quality cells retained from the mRNA processing. Guide-RNA assignments and dial-up gene perturbation data were extracted; these matrices were generated per array, concatenated, and combined with the single-cell gene expression data into a unified gene-by-cell matrix. The per-cell guide information and CROP-seq UMI counts were stored as metadata for downstream SCEPTRE analysis. Statistical Analysis with SCEPTRE SCEPTRE is a scalable R package for statistically rigorous analysis of high-MOI single-cell CRISPR screens. The gene (response) and gRNA count matrices were extracted from the Seurat object. Cells with >50 gRNA UMIs were removed. Data were imported with import_data under a high-MOI model, treating each guide individually (grna_integration_strategy = “singleton”), with safe-harbor guides designated as non-targeting controls. In addition to the cell-level covariates computed automatically by SCEPTRE (gene and gRNA library sizes and complexity), we supplied log-transformed CROP-seq UMI counts, the number of assigned guides per cell, and batch (array) as extra covariates. gRNAs were assigned to cells with the mixture model. Assay calibration was assessed with run_calibration_check on the non-targeting/negative-control set, statistical power was confirmed on positive-control pairs with run_power_check, and candidate gRNA–gene associations were tested with run_discovery_analysis. Discovery (candidate) p-values were adjusted by the Benjamini–Hochberg method; positive controls were retained at their nominal p-values for visualization. Results were exported as per-gene bedGraph tracks (signed −log10 adjusted p-value, with sign taken from the log2 fold change, placed in 10-bp windows centered on each guide cut site) for downstream visualization, together with the full SCEPTRE output directory. RT-qPCR statistical analysis We selected 20 of the top scoring guides for follow up based on both their SCEPTRE predicted fold change and p-value. We also ensured a spread of gRNAs that impacted each (or multiple) of the 9p21 transcripts. We note that we identified many more significant gRNA-gene pairs that should be followed up in future studies. The impact of each of these 20 gRNAs, and a minimum of 2 safe harbor control gRNAs, on 9p21 transcript levels was measured across 4 independent transfections (with cells collected at either day 4 and day 8, as described above). For each of the 4 transfections, qRT-PCR was performed for each 9p21 transcript (i.e., p16INK4A, p14ARF, P15INK4B, ANRIL) and GAPDH in technical triplicate. The Ct values were averaged across the three technical replicates for each transcript. Ct values above Ct35 were considered ND (not detected). To calculate the impact of each gRNA on each 9p21 transcript, the Ct values were normalized by the average Ct value of the corresponding GAPDH control, in the presence of that gRNA (ΔCtgp). We then followed the following steps to calculate a Z score and a fold change.ΔΔCtgp=[ΔCtgp−mean(ΔCtgSH)]Zgp=ΔΔCtgpstd(ΔCtgSH)Foldchange=2−ΔΔCtgpWhere p is the gRNA/perturbation, g is the gene measured (i.e., p16INK4A, p14ARF, P15INK4B, ANRIL), SH is the safe harbor control gRNA, and Z is the Z score. The Z-scores were combined across the 4 transfection experiments using the following formula, where n is the number of biological replicates (day 4 and day 8 cells were considered replicates). We only considered a Z score significant if there are 3 or more replicates where a given transcript is expressed above the Ct35 threshold.combinedZ=∑Zgpn The average fold change and Z score were visualized with a dotplot generated in R using ggplot2.138 The size of the dot represents the strength and significance of the effect of the guide on each transcript. Z-scores of less than or equal to −1 are represented as a dot size of 1, Z-scores of less than or equal to −2 are represented as a dot size 2. Original data are shown in Table S6. Region Capture Micro-C data processing and analysis RCMC paired-end reads (2x 38 bp, generated as described above) fastqs were verified for quality control using FastQC (v0.11.9). Paired end reads were aligned to the UCSC hg38 genome using Juicer113 (default setting, version 2.0) and filtered to include only reads from the capture locus. The map resolution of 200bp, defined as the smallest locus size such that 80% of loci have at least 1,000 contacts, was calculated via the calculate_map_resolution.sh script from Juicer. Deduplicated paired reads were converted to contact matrices (.hic format file) using ‘juicer_tools’ with ‘pre’ at base pair delimited resolutions of 2.5 Mb, 1 Mb, 500 kb, 250 kb, 100 kb, 50 kb, 25 kb, 10 kb, 5 kb, 2 kb, 1 kb, 500 bp, 200 bp, 100 bp, 50 bp, 20 bp, and 10 bp. All contact matrices used for further analysis were VC-normalized with Juicer. RCMC contact maps were visualized alongside genomic annotations and ChIP-seq datasets using the ‘plotgardener’ package (version 1.6.1114) on R version 4.3.1. 9p21 promoter region and distal CRE definition To define the start-end boundaries of the 9p21 promoters and distal CRE, we have used the regions originally targeted by gRNAs (described in the section “Selection of regions for CRISPRa screen experiment”, and shown in Figure 5B). To define the p16 promoter region, we merged 2 “originally targeted” regions that were only 2bp apart and overlapped the first exon of p16. E6 was merged from 4 “originally targeted” regions which. The “original regions” are also visible in Figure 6C. The final coordinates of the promoter/CREs are listed in Table S5. Chromatin contact analysis All contact matrices from each replicate and time point (see Table S2) were combined after deduplication to create a merged.hic file. To call loops, the hiccups command from juicer_tools was used with default parameters and resolutions (1kb, 5kb). Subtraction plots of RCMC data (in Figures 7B and S17) were generated in R 4.4.2 using plotgardener v1.12.0.114 Hic files from PDL21 (2 replicates combined) and PDL48 were loaded at the resolution indicated in each panel. Counts within each dataframe were divided by the total number of counts in chr9 in that dataframe, and multiplied by 10ˆ6 to obtain normalized values. Normalised values for PDL21 were then subtracted to the normalized values for PDL48. Windows plotted were: chr9:20,300,000–22,600,000 and 21,750,000–22,500,000. Virtual 4C analysis Virtual 4C profiles were generated directly from the merged.hic file using a custom Python script based on numpy141 and the hicstraw library. For each genomic element within the 9p21 locus on chromosome 9, the element midpoint was used as a viewpoint. Observed Hi-C contacts with no normalization were extracted at 500bp resolution between the viewpoint bin and a contiguous genomic window spanning the MTAP gene through the most distal downstream element. Contact frequencies were aggregated by genomic bin and visualized as virtual 4C profiles, with the viewpoint and annotated regulatory elements overlaid. The source code is available at https://github.com/broadinstitute/CDKN2A-gene-regulation. Sci-ATAC-seq data To assess the accessibility across human tissues of our validated elements (promoters and enhancers), we queried data from the cis Elements Atlas for Human tissues from82 (accessible at https://decoder-genetics.wustl.edu/catlasv1/catlas_humanenhancer/#!/). We downloaded the location of peaks (bed format, cCREs.bed.gz), the cell type list (celltypes.txt.gz) and the matrix (matrix.mtx.gz) of accessibility by cell type from https://decoder-genetics.wustl.edu/catlasv1/humanenhancer/data/cCRE_by_cell_type/. We constrained the analysis to regions that overlapped with our elements (excluding regions where the overlap was less than 10% of our element (via bedtools intersect -f 0.1). All the elements in our validated list overlapped with at least one element identified in Zhang, et al.82 , with the exception of E2. The promoter regions and E6 overlapped with 2 elements each. E1 overlaps with Zhang et al.'s82 element 1081266 (chr9: 21965473–21965873), p16 promoter with elements 1081270–1081271 (chr9: 21974569–21974969 and chr9: 21974972–21975372), p14/ANRIL promoter with 1081279–1081280 (chr9: 21994127–21994527 and chr9: 21994543–21994943), p15 promoter with 1081284–1081285 (chr9: 22008572–22008972 and chr9: 22009192-2200959), E3 with 1081325 (chr9: 22097662–22098062), E4 with 1081336 (chr9: 22111805–22112205), E5 with 1081339 (chr9: 22114537–22114937), E6 with 1081340–1081341 (chr9: 22117621–22118021 and chr9: 22118178–22118578), E7 with 1081361 (chr9: 22148135–22148535), E8 with 1081374 (chr9: 22170176–22170576) and E9 with 1081388 (chr9: 22211645–22212045). We also constrained the analysis to only the 111 adult cell types in82 (i.e., we excluded from our analysis cell types with “Fetal” in their label). Each adult cell type is annotated (bottom annotation) by their “group”, and “main organ” of origin. The “group” is defined as the dendrogram label from Figure 1E in the original publication82) was downloaded from Mendeley Data (https://doi.org/10.17632/yv4fzv6cnm.1). Some of the 30 anatomical sites from the original publication were then grouped/renamed as follows: ‘Frontal Cortex' became ‘Brain'; ‘Sigmoid Colon', ‘Transverse Colon', ‘Esophagus GE Junction’, ‘Esophagus Mucosa', ‘Esophagus Muscularis', ‘Small Intestine', ‘Stomach' became ‘GI'; ‘Tibial Artery', ‘Aorta', ‘RA Appendage', ‘Left Atrium', ‘Left Ventricle', ‘Right Atrium', ‘Right Ventricle' became ‘Cardiovasc'; ‘Ovary', ‘Uterus', ‘Vagina’ became ‘Female Rep'; ‘Tibial Nerve' became ‘Nerve'; ‘Pancreatic Islet' and ‘Body of Pancreas' became ‘Pancreas'; ‘Skin Suprapublic' and ‘Skin Lower Leg' became ‘Skin'. For each cell type, we then nominated a “main organ” if there was one tissue that accounted for >70% of cells in that cell type. If no organ satisfied this criterion, the label for “organ” was assigned as “Multiple”. Cell type labels in the table from Mendeley Data were very similar but not identical to the labels in Figure 1E in the original publication and in the cell type list: small differences were amended manually. The accessibility across the adult cell types of the cis elements from82 that overlap our regions is displayed in the heatmap in Figure 7C. For each region we also displayed (right annotation in the heatmap) the effect that CRISPRa at that element has on 9p21 transcripts (manually binarized from results in Figure 6D). The heatmap in Figure 7C was generated using ComplexHeatmap 2.22.0152 in R 4.4.2. Attribution scores and TF predictions ChromBPNet (v0.1.7)77 was trained on PDL20 and PDL50 ATAC data (replicates merged for each PDL) in the default “pipeline” mode. Chromosomes 8 and 20 were used for model validation; chromosomes 1, 3 and 9 were used for model testing and remaining chromosomes were used for model training. PDL20 and PDL50 ChromBPNet models had a Pearson correlation coefficient (r) of 0.73 and 0.699 in the counts head, respectively. DeepLIFT155 was used to extract the model attribution scores in the E1-9 enhancers and promoters of the 9p21 locus. The resulting positive attribution scores were matched to JASPAR2024 vertebrate database136,137 using TOMTOM116 to annotate putative Transcription Factors (TF) that bind to the selected TF binding sites (TFBS). The candidate TF for each TFBS was selected by combining the mean attribution score of the TFBS and the resulting q-value of the TOMTOM search (Table S6). Expression of candidate TFs in WI-38 undergoing replicative senescence was assessed using the mRNA-seq analysis in.17 The table with TPM values was downloaded from (https://github.com/dghendrickson/hayflick/blob/main/data/RNAseq/TPM_tables/RS_batch_tpms.txt). Expression in 3 replicates from PDL20 and PDL50 is plotted with ComplexHeatmap 2.22.0152 in R 4.4.2. Comparison with published CREs in 9p21 To compare the CREs identified in this study with the CREs previously identified in human cells2 and characterized in HeLa cells,24 we obtained the list of 33 enhancer regions in 9p21 from Harismendy et al.,2 and converted their coordinates from NCBI36 (hg18) to hg38 using UCSC LiftOver132 with minimum ration of bases that must remap of 0.95. These regions were then manually matched with E1-E24 in Farooq et al., using as guidance their alignment to H3K4me1 and H3K27ac signal in HeLa (Encode tracks: H3K27ac accession ENCFF850RSK/ENCSR000AOC and H3K4me1 accession ENCFF860GGZ/ENCSR000APW, Figure S18). The coordinates of E17 and E18 (from Farooq et al.24) were manually added, as we did not find a matching region among the enhancers in Harismendy et al.2 The location of E17 was identified by aligning the sequences of the guide RNAs used for its editing (extracted from Table S2 in Farooq et al;24) to hg38 with Blat (https://genome.ucsc.edu/cgi-bin/hgBlat131). The approximate location of E18 was manually reconstructed by alignment with the H3K4me1 signal in HeLa, to represent the data shown in Farooq et al.24