Post-GWAS functional analysis identifies CUX1 as a regulator of p16INK4a and cellular senescence

Accumulation of senescent cells with age is an important driver of aging and age-related diseases. However, the mechanisms and signaling pathways that regulate senescence remain elusive. In this report, we performed post-genome-wide association studies (GWAS) functional studies on the CDKN2A/B locus, a locus known to be associated with multiple age-related diseases and overall human lifespan. We demonstrate that transcription factor CUX1 (Cut-Like Homeobox 1) specifically binds to an atherosclerosis-associated functional single-nucleotide polymorphism (fSNP) (rs1537371) within the locus and regulates the CDKN2A/B-encoded proteins p14ARF, p15INK4b and p16INK4a and the antisense noncoding RNA in the CDK4 (INK4) locus (ANRIL) in endothelial cells (ECs). Endothelial CUX1 expression correlates with telomeric length and is induced by both DNA-damaging agents and oxidative stress. Moreover, induction of CUX1 expression triggers both replicative and stress-induced senescence via activation of p16INK4a expression. Thus, our studies identify CUX1 as a regulator of p16INK4a-dependent endothelial senescence and a potential therapeutic target for atherosclerosis and other age-related diseases.

S ince its discovery by three independent GWAS 1-3 , the CDKN2A/B locus on chromosome 9p21 has been noted as being strongly associated with the susceptibility of coronary artery disease 4 . This locus has also been associated with a multitude of additional cardiovascular conditions, including myocardial infarction and carotid artery plaque formation, in addition to other diseases such as type 2 diabetes and various forms of cancer 5,6 . However, GWAS cannot on their own specify the causative fSNPs associated with these diseases. Of note, all these diseases are recognized as age-related pathologies in that their incidence markedly increases with age [7][8][9] . Consistent with this observation, a genetic association of this locus with frailty and overall lifespan has also been recently revealed [10][11][12] . Since this single region associates with multiple age-related diseases, each with its own distinct pathophysiological mechanism, these data imply that this locus may contribute to the progression of these age-related diseases by modulating some aspect of aging biology as a common unifying mechanism 13 .
Aging is a continuous process of gradual functional decline 7 . Increasing evidence has implicated that the accumulation of senescent cells with age is a molecular driver of this functional decline, and also an important contributor to age-related diseases 9,14-20 . Cellular senescence is defined as irreversible cell cycle arrest often accompanied by an enlarged and flattened cellular morphology. In association with this arrest, senescent cells also secrete multiple proinflammatory molecules such as the cytokines IL-6 and IL-1β and the cell adhesion molecule ICAM1, collectively known as the senescence-associated secretory phenotype (SASP). These factors can induce both low-grade chronic inflammation and endothelial remodeling 21 . Based on the initiating trigger, cellular senescence can be classified as either replicative or stress induced. Although both types of senescence are mediated through the p53/p21 and/ or p16 INK4a /retinoblastoma protein (RB) pathways, preference for one pathway over the other depends on cell type, species and the stimuli 14,17,20,22 .
The CDKN2A/B locus, spanning a 200-kb region, harbors three well-characterized tumor suppressor genes-p14 ARF , p15 INK4b , p16 INK4a -and ANRIL. Among these genes, p16 INK4a has been implicated in the p16 INK4a /RB pathway that leads to cellular senescence in a variety of cell types [23][24][25][26] . In addition, p16 INK4a expression is also used as one of the common markers for cellular senescence 27 . Other senescence markers include senescence-associated β-galactosidase (SA-β-gal) staining, γ-H2AX, telomere length and expression of SASP genes 27 . However, relatively little is known about how p16 INK4a expression is regulated in response to various stimuli that trigger senescence. a   rs10811640  rs10811660  rs10965248  rs113646886  rs1333047  rs1537371  rs2811709  rs3217977  rs568447  rs944799  rs3218007  rs77284052  rs10965215  rs944801  rs6475604  rs9632885  rs7857345  rs1537374  rs1333046  rs10965247 A   (2-4) three epigenetic tracks for H3K27ac, H3K4me1 and H3K4me3, known as transcriptional factor-binding sites; (5) DNase I hypersensitivity sites (DNase I HS) in human astrocytes; (6) predicted regulatory elements including promoters (red) and enhancers (gray); (7) annotated genes including p14 ARF , p16 INK4a , p15 INK4b and ANRIL. c, Zoomed-in view of the 4-kb genomic region around fSNP rs1537371, showing the same tracks as above plus the negative result from CUX1 ChIP-seq assay in three human cell lines, GM12878, K562 and MCF-7. d,e, Demonstration of fSNP rs1537371 by EMSA (d) and luciferase reporter assay (e). A, risk allele; C, nonrisk allele; T, very rare allele; RLA, relative luciferase activity. Data for EMSA represent n = 3 biologically independent experiments; data for luciferase reporter assays represent n = 6 biologically independent samples. f, Sequences showing mutations around rs1537371 in three independent CRISPR-cas9 clones (nos. 2, 19 and 56), together with wild-type sequence. CON, wild-type control. g, qPCR showing decreased expression of p16 INK4a , one of the potential risk genes in the three mutants. Data for qPCR analysis represent n = 3 biologically independent samples, each performed in duplicate. h, Dot plot of fSNP rs1537371 and p16 INK4a  In the present work, by coupling of regulatory element sequencing (Reel-seq) with flanking restriction-enhanced DNA pulldown-mass spectrometry (FREP-MS) and allele-imbalanced DNA pulldown-Western blot (AIDP-Wb), three techniques recently developed in our laboratory 28,29 , we discovered that CUX1, a transcription factor known to play roles in cell migration, proliferation, differentiation, DNA damage repair and tumorigenesis 30 , regulates both replicative and stress-induced senescence in human arterial ECs by modulation of p16 INK4a expression in response to telomere shortening, DNA damage and oxidative stress. We demonstrate that this regulation occurs via the specific binding of CUX1 to an atherosclerosis-associated fSNP, rs1537371, on the CDKN2A/B locus. Our findings reveal a new role of CUX1 in regulation of cellular senescence and provide new insights into how genetic variants in the CDKN2A/B locus can modulate susceptibility to atherosclerosis and other age-related complications.

Identification of a fSNP rs1537371 within the CDKN2A/B locus.
Previously we demonstrated the feasibility of using Reel-seq to identify fSNPs associated with susceptibility to breast cancer 29 . By applying Reel-seq to the disease-associated CDKN2A/B locus, 24 candidate fSNPs were identified from 193 SNPs revealed by GWAS in linkage disequilibrium (LD), with R 2 > 0.8 ( Fig. 1a and Supplementary Data 1). The locations of these 24 candidate fSNPs are listed in the SNP track in Fig. 1b, indicting their overlap with transcription factor-binding sites, DNase I hypersensitivity sites or predicted promoters and enhancers 31 . Twenty-two of these candidate fSNPs were demonstrated as probable fSNPs using allele-imbalanced electrophoresis mobility shift assay (EMSA) as a criterion (Fig. 1a). Among these 22 fSNPs, SNP rs1537371 is in LD (R 2 > 0.95) with a lead SNP, rs4977574, that is strongly associated with atherosclerosis in a European population 5 . To verify that rs1537371 is a fSNP, we first performed in silico analysis on this SNP using the Encyclopedia of DNA Elements (ENCODE) database. We found that rs1537371 is located in the H3K27ac (histone H3K27 acetylation), H3K4me1 (histone H3K4 methylation group 1) and DNase I hypersensitivity-enriched sites in both human umbilical vein endothelial cells (HUVECs) and human astrocytes (Fig. 1c). To further demonstrate that rs1537371 is functional, we repeated our EMSA analysis using nuclear extract (NE) isolated from primary ECs. This analysis revealed an alleleimbalanced gel shift pattern with the risk allele A noted to shift more than the nonrisk allele C (Fig. 1d). We next performed a luciferase reporter assay with both risk and nonrisk alleles in parallel. As noted with our EMSA results, allele-imbalanced luciferase reporter activity was also detected, with risk allele A having higher luciferase activity than nonrisk allele C (Fig. 1e). Finally, we performed CRISPR-cas9, a gene editing system with a synthetic guide RNA that targets the rs1537371 sequence in a human cell line. Three independent CRISPR-edited cell lines were generated (Fig. 1f). As indicated, clone 2 has a heterozygous deletion of 12 nucleotides, clone 56 has a 34-nucleotide deletion on one allele and a five-nucleotide deletion on the other while clone 19 carries a transversion from risk allele A to nonrisk allele C, which generates a homozygous C allele in this clone. Nevertheless, all these clones showed a significant reduction in p16 INK4a expression (Fig. 1g). In particular, downregulation of p16 INK4a in clone 19, carrying a C/C genotype compared to the C/A genotype of the parental control, suggests that risk allele A is a gainof-function allele. To complement this CRISPR-cas9 analysis, we also examined p16 INK4a expression in peripheral blood mononuclear cells (PBMCs) of 26 unrelated individuals of varying genotype. We found that the nine individuals with homozygous risk allele A/A expressed a significantly higher level of p16 INK4a than the eight individuals carrying homozygous nonrisk allele C/C (P = 0.047, n = 17) as measured by quantitative PCR (qPCR; Fig. 1h). Together, these results support rs1537371 as a fSNP with allele-imbalanced activity.
Identification of CUX1 specifically binding to rs1537371. To determine the protein(s) that specifically binds to rs1537371, we applied FREP-MS and identified six proteins binding it (Supplementary Table 1 and Supplementary Data 2). Among these six proteins we identified CUX1 as the top hit in this analysis. Of note, CUX1 is a member of the homeodomain family of DNAbinding proteins that have been reported to be involved in the regulation of cell proliferation, and to act as a tumor suppressor 30,32,33 .
To validate that CUX1 specifically binds to rs1537371, we first performed chromatin immunoprecipitation sequencing (ChIPseq) using an anti-CUX1 antibody in ECs. We first demonstrated the specificity of this anti-CUX1 antibody by showing a significant enrichment of the rs1537371-containing DNA fragment pulled down through use of this antibody versus an anti-IgG antibody (Fig.  2a, left). Using this anti-CUX1 antibody, we observed a significant decrease in the binding of CUX1 to the rs1537371-containing DNA fragment in CUX1 short hairpin RNA (shRNA) knockdown ECs compared to scrambled shRNA control ECs (Fig. 2a, left). In contrast, when we performed the same ChIP assay on two randomly chosen genomic regions we did not observe any obvious difference in CUX1 binding (Fig. 2a, right). To confirm the allele-imbalanced binding of CUX1 to risk allele A versus nonrisk allele C, we performed Sanger sequencing since the ECs we used carry the A/C heterozygous genotype. After sequencing rs1537371-containing DNA fragments from both inputs and ChIP samples, we observed 55% enrichment of the A allele versus the C allele from the ChIP samples (24A/15C) compared to the inputs (20A/19C), with P < 0.01 (n = 3) (Fig. 2b). These data demonstrate an endogenous, alleleimbalanced binding of CUX1 to the fSNP rs1537371. Next, we performed AIDP-Wb, a new DNA pulldown assay recently developed in our laboratory, for specific detection of allele-imbalanced binding of a known protein to a given fSNP 29 . Using this technique, we were able to confirm the allele-imbalanced binding of CUX1 to the fSNP rs1537371, again with risk allele A binding more CUX1 than nonrisk allele C (Fig. 2c). The noted allele-imbalanced binding of CUX1 in this AIDP-Wb assay further validates the supposition that rs1537371 is indeed a bona fide fSNP. We next performed a luciferase reporter assay with a construct containing risk allele A used in Fig. 1e. We performed this assay in the setting of either CUX1 knockdown by shRNA or CUX1 overexpression by lentiviral expression vector pLVX. As expected, increased or decreased luciferase activity was observed when CUX1 expression was either positively or negatively modulated (Fig. 2d, left in both panels). This alteration in luciferase activity was not observed when we performed the same assay using a control reporter construct containing an irrelevant SNP sequence (Fig. 2d, right in both panels). These data suggest that CUX1 can bind to fSNP rs1537371 and regulate transcriptional activity. In addition, we also performed an online search and identified the core binding motif of CUX1 as ATC[A/C]AT 34 . This motif showing an increase in replicative senescence from p5 ECs (left) to p10 ECs (middle), and a reduction in replicative senescence in CUX1 shRNA knockdown p10 ECs (right) compared to scrambled p10 ECs (middle). Right, quantitative plots are shown for both β-gal + cells (%) in SA-β-gal staining (top) and γ-H2AX foci/cells (%) with γ-H2AX staining (bottom). Data for SA-β-gal and γ-H2AX staining represent n = 3 biologically independent experiments. b, Immunoblots and qPCR showing increased expression of CUX1 and p16 INK4a in p10 ECs compared to p5 ECs. Data for immunoblot analysis represent n = 3 biologically independent experiments; data for qPCR analysis represent n = 3 biologically independent samples, each performed in duplicate. c, PCR-based analysis showing significant decrease in telomeric length from p5 to p10 ECs. Data for PCR analysis represent n = 4 biologically independent samples. d, Immunoblots and qPCR showing that shRNA knockdown of CUX1 resulted in decreased expression of p16 INK4a in p10 human ECs. Data for immunoblot analysis represent n = 3 biologically independent experiments; data for qPCR analysis represent n = 3 biologically independent samples, each performed in duplicate. e, qPCR showing significant downregulation of SASP genes, IL-6, IL-1β and ICAM1 in CUX1 shRNA knockdown p10 ECs. Data for qPCR analysis represent n = 3 biologically independent samples, each performed in duplicate. f,g, Decrease in both BrdU incorporation (f) and percentage of S/G2/M cell numbers (g) in p10 ECs (middle) versus p5 ECs (left) indicated an increase in replicative senescence. Knockdown of CUX1 in p10 ECs (right) resulted in recovery from both decreased BrdU incorporation and reduced percentage of S/G2/M cell numbers, indicating a blockage in cellular senescence in CUX1 shRNA knockdown p10 ECs. Data for BrdU incorporation represent n = 12 biologically independent samples; data for cell cycle analysis represent n = 3 biologically independent samples. P values were calculated using two-tailed Student's t-test, and all data are presented as mean ± s.e.).
is highly similar to the sequence around rs1537371, AT[G/T]AAA. These data thus support our premise that rs1537371 is a fSNP and that CUX1 can bind to rs1537371 in an allele-imbalanced manner. p14 ARF , p15 INK4b , p16 INK4a and ANRIL are regulated by CUX1x. The 200-kb region of the CDKN2A/B locus contains three tumor suppressor genes-p14 ARF , p15 INK4b and p16 INK4a -as well as the long noncoding RNA, ANRIL. To determine whether these four genes are regulated by CUX1, we performed shRNA knockdown of CUX1 in primary ECs. Using a shRNA lentivirus that carries a sequence targeting the CUX1 gene, we were able to generate polyclonal pools of ECs with reduced expression of CUX1 as detected by both qPCR and immunoblot analysis (Fig. 2e). Of note, human ECs appeared to express only full-length CUX1 with an apparent   35,36 . Nevertheless, as a result of CUX1 downregulation, a significant decrease in the expression of p14 ARF , p15 INK4b , p16 INK4a and ANRIL was evidenced at the messenger RNA level (Fig. 2f) and, in addition, decreased expression of coding proteins p14 ARF , p15 INK4b and p16 INK4a was detected by immunoblot analysis (Fig. 2g). These results thus demonstrate that CUX1 can modulate the expression of p14 ARF , p15 INK4b , p16 INK4a and ANRIL. To confirm these results, we also performed the same knockdown in ECs using a small interfering RNA (siRNA) that targets a CUX1 sequence different from that employed in CUX1 shRNA knockdown. This siRNA approach also resulted in a significant reduction in the expression of all four genes within the CDKN2A/B locus (Extended Data Fig. 1a). These data, together with data generated by various independent strategies (for example, CRISPR-cas9 gene editing, ChIP, AIDP-Wb, CUX1dependent luciferase reporter assay and in silico analysis), support the view that CUX1, as a transcription factor, regulates the expression of risk genes p14 ARF , p15 INK4b , p16 INK4a and ANRIL by binding to the fSNP rs1537371.

CUX1 regulates replicative senescence via p16 INK4a .
Because p16 INK4a has been implicated in senescence 9,14,37 , we reasoned that CUX1-mediated regulation of p16 INK4a expression might have important implications in endothelial senescence. To test this hypothesis, we first performed both SA-β-gal and γ-H2AX staining of primary human ECs collected at either early passage (p5) or late passage (p10). This analysis revealed the expected increase in both SA-β-gal + cells (Fig. 3a, top) and γ-H2AX foci (Fig. 3a, bottom) in p10 ECs (Fig. 3a, middle) versus p5 ECs (Fig. 3a, left). We next sought to determine correlation between the expression of CUX1 and p16 INK4a . As a result, passage-dependent increase in CUX1 expression and a concomitant increase in that of p16 INK4a from p5 to p10 ECs were detected at both mRNA and protein levels (Fig. 3b).
In addition, using a PCR-based assay, we also detected decreased telomere length in p10 versus p5 ECs (Fig. 3c), suggesting a negative correlation between the expression level of CUX1 and telomere length. Together, these data support a potential role of CUX1 in regulation of replicative senescence. To demonstrate that CUX1 is responsible for replicative senescence in ECs, we infected p10 ECs with a lentiviral shRNA targeting CUX1. Forty-eight hours after infection, downregulation of CUX1 and decreased expression of p16 INK4a were confirmed by both immunoblot and qPCR analysis (Fig. 3d). In these CUX1 shRNA knockdown p10 ECs, a significant decrease in both SA-β-gal and γ-H2AX staining was detected (Fig. 3a, right) in comparison with the scrambled control (Fig. 3a, middle), suggesting a reversal of cellular senescence in these cells. In addition, we also detected a significant downregulation of various SASP genes, including IL-6, IL-1β and ICAM1, in CUX1 shRNA knockdown p10 ECs (Fig. 3e). These data demonstrate that CUX1 can modulate replicative senescence. Similar findings were also observed when CUX1 expression was modulated independently by siRNA knockdown in ECs (Extended Data Fig. 1b). Moreover, because cell cycle arrest is an essential feature of cellular senescence 38 , we also performed a 5-bromo-2'-deoxyuridine (BrdU) incorporation assay and cell cycle analysis. As shown in Fig. 3f,g, a significant decrease in both BrdU incorporation and the percentage of cell numbers in the S/ G2/M phase was observed in p10 ECs (right) versus p5 ECs (left), suggesting potential cell cycle arrest. Consistent with the recovered SA-β-gal and γ-H2AX staining in CUX1 shRNA knockdown p10 ECs (Fig. 3a), reduction in both BrdU incorporation and percentage of cell numbers in the S/G2/M phase can also be reversed by CUX1 shRNA knockdown in p10 ECs (Fig. 3f,g, right).
In addition, to determine whether the regulation of replicative senescence by CUX1 is specific for ECs, we performed CUX1 shRNA knockdown in human primary vascular smooth muscle cells (VSMCs). Consistent with the results we observed in ECs, downregulation of CUX1 in VSMCs also resulted in a significant decrease in the expression of p14 ARF , p15 INK4b , p16 INK4a and ANRIL (Extended Data Fig. 2a). This was accompanied by a significant decrease in both SA-β-gal and γ-H2AX staining (Extended Data Fig. 2b), as well as reduced expression of SASP genes (Extended Data Fig. 2c). These data indicate that the role of CUX1 in regulation of replicative senescence is unlikely to be EC specific: it could also occur in other cell types relevant to atherogenesis, such as VSMCs.

Activation of cellular senescence by CUX1 requires p16 INK4a .
Among the four genes regulated by CUX1, p16 INK4a is known as a regulator of cellular senescence 9,14 . To determine whether p16 INK4a is the downstream mediator of CUX1 responsible for regulation of cellular senescence, we first overexpressed CUX1 in human ECs using the above-mentioned lentiviral expression vector pLVX-CUX1.  demonstrated by a decrease in both SA-β-gal and γ-H2AX staining (Fig. 4b, middle and right) as well as an increase in both BrdU incorporation (Fig. 4d, middle and right) and the percentage of cell numbers in the S/G2/M phase (Fig. 4e, middle and right). Together, these data suggest that CUX1 induces senescence in a p16 INK4aindependent manner. However, the increased expression level of SASP genes IL-6, IL-1β and ICAM1 in CUX1-overexpressed ECs remained unchanged following p16 INK4a knockdown (Fig. 4c, right).  These data are consistent with a previous publication showing that p16 INK4a can induce cellular senescence without the associated inflammatory secretory phenotypes 39 .
To further demonstrate that p16 INK4a is the downstream mediator of CUX1-induced cellular senescence, we also overexpressed p16 INK4a in CUX1 shRNA knockdown ECs. As expected, knockdown of CUX1 resulted in decreased expression of p16 INK4a in human ECs (Extended Data Fig. 3a, middle). Under this condition, reduced cellular senescence was observed as demonstrated by a decrease in both SA-β-gal and γ-H2AX staining (Extended Data Fig. 3b, middle). Consistent with this observation, when we overexpressed p16 INK4a using lentiviral expression vector p156RRL in CUX1 shRNA knockdown ECs (Extended Data Fig. 3a, right), we detected restoration of both SA-β-gal and γ-H2AX staining (Extended Data Fig. 3b, right), as well as in the expression of SASP genes IL-6, IL-1β and ICAM1 (Extended Data Fig. 3c, right). Of note, restoration of the expression of these SASP factors in p16 INK4a -overexpressed ECs was unexpected since, as mentioned above, p16 INK4a is not believed to be a SASPinducing factor 39 .
Since CUX1 also regulates the expression of p14 ARF , p15 INK4b and ANRIL (Fig. 2f,g), we next sought to determine whether these gene products play a role in regulation of endothelial senescence. We therefore performed siRNA knockdown of p14 ARF , p15 INK4b and ANRIL in CUX1-overexpressed human ECs (Extended Data Fig.  4a,b). However, we noted no obvious change in endothelial senescence following knockdown of p14 ARF , p15 INK4b or ANRIL (Extended Data Fig. 4c-e).
CUX1 regulates stress-induced senescence via p16 INK4a . Stressinduced premature senescence is another type of cellular senescence that can be triggered by various stimuli including DNA damage, oxidative stress, oncogene activation and metabolic dysregulation [40][41][42] . To determine whether CUX1 also plays a role in stress-induced premature senescence, we first investigated protein expression of CUX1 and p16 INK4a in ECs treated for 0, 4, 24 and 48 h with bleomycin (0.5 µg ml -1 ), a genotoxic drug known to induce senescence by the introduction of double-stranded DNA breaks 43 . We observed upregulation of both CUX1 and p16 INK4a expression at all time points, and a restoration of p16 INK4a expression to the level of untreated ECs with CUX1 knockdown by shRNA (Extended Data Fig. 5a). We also observed a corresponding induction of γ-H2AX staining (Extended Data Fig. 5b). Based on these observations, we performed a detailed analysis using human ECs treated with bleomycin (0.5 µg ml -1 ) for 24 h. Both immunoblot and qPCR analysis identified a significant increase in the expression of both CUX1 and p16 INK4a in bleomycin-treated ECs (Fig. 5a, left and middle).
Bleomycin-treated cells also demonstrated an increase in cellular senescence, as evident by increased staining for both SA-β-gal and γ-H2AX (Fig. 5b, left and middle), increased IL-6 and IL-1β expression (Fig. 5c, left and middle) and reduced BrdU incorporation (Fig. 5d, left and middle) and percentage of cells in S/G2/M phase (Fig. 5e, left and middle). To further demonstrate that CUX1 is required for the induction of this type of premature senescence, we downregulated CUX1 by shRNA in human ECs before bleomycin exposure (Fig. 5a, middle and right). As expected, bleomycininduced p16 INK4a expression was reduced in CUX1 knockdown ECs (Fig. 5a, middle and right). We also noted inhibition of bleomycininduced senescence, with a reduction in both SA-β-gal and γ-H2AX staining (Fig. 5b, middle and right), decreased expression of various SASP genes (Fig. 5c, middle and right), increased BrdU incorporation (Fig. 5d, middle and right) and restoration of cell cycle parameters (Fig. 5e, middle and right). Together, these data demonstrate that CUX1 modulates DNA damage-induced premature senescence, presumably via modulation of p16 INK4a expression in ECs.
To extend our findings on the role of CUX1 in regulation of stress-induced premature senescence, we repeated the above assays using exogenous H 2 O 2 to mimic conditions of oxidative stress 44 Fig. 6a, right) and, as a consequence, this further resulted in a reduction of H 2 O 2 -elicited senescent response (Extended Data Fig. 6b,c, right).

RNA-seq analysis of genes regulated by CUX1.
To further characterize how CUX1 regulates cellular senescence, we performed RNA sequencing (RNA-seq) analysis with total RNA isolated from human ECs at p10 treated with both a scrambled siRNA and a siRNA targeting human CUX1. We estimated that, at p10, approximately 20% of control ECs were senescent and that there was a reduction of ~70% in senescent cells following siRNA targeting of CUX1 (Fig. 3a). As a result, we identified 471 differentially expressed genes (DEGs) with fold change (FC) between the CUX1 siRNA-treated sample and the scrambled siRNA-treated control >1.5 and adjusted P < 0.05 (Supplementary Data 3, column D, E and F are scrambled siRNAtreated controls and Column G, H and I are CUX1 siRNA-treated samples). Among these 471 DEGs, we found that 228 genes were upregulated and 243 downregulated in CUX1 siRNA knockdown ECs, with the top 20 upregulated and downregulated DEGs listed in Extended Data Fig. 7a. While we found that CUX1, as well as p14 ARF , p16 INK4a and p15 INK4b , were significantly downregulated (Extended Data Fig. 7b), we detected no obvious alterations in other classical senescence markers, including p53 and HGBM1, nor did we observe classical SASP factors such as IL-6 among these 471 DEGs. This may relate to the relatively modest percentage of overall senescent cells in our cell-based model. Interestingly, we did identify downregulation of dimethylarginine dimethylaminohydrolase-1, an enzyme involved in asymmetric dimethylarginine (ADMA) degradation (Extended Data Fig. 7a). Of note, ADMA is an endothelial nitric oxide synthase (eNOS) inhibitor, and eNOS activity regulates endothelial cell senescence 45,46 . Similarly, we noted upregulation of DNA methyltransferase 3b (Extended Data Fig. 7a), which was previously reported to activate senescence markers including p16 INK4a and p21 CIP1/WAF1 by decreasing the methylation of CpG islands 47 . In addition, gene set enrichment analysis (GSEA) of the 471 DEGs identified 29 upregulated and five downregulated pathways enriched in CUX1 siRNA knockdown human ECs (nominal P < 0.025; Extended Data Fig. 7c). These pathways include cell cycle regulation, cancer cell growth, cell differentiation and apoptosis.

Elevated CUX1 and p16 INK4a expression in patients.
Previous studies have established that depletion of p16 INK4a -positive senescent cells by genetic manipulation or pharmacological strategies can delay the onset of age-related diseases, as well as extend longevity 9,14,15,48 .
Based on this association we hypothesized that, as a direct regulator of p16 INK4a , CUX1 expression might be upregulated in patients with age-related diseases such as atherosclerosis. To test this, human atherosclerotic plaques were obtained from patients undergoing carotid endarterectomy. Total RNA was isolated from both plaques and zones of normal appearance, and expression of CUX1 measured by qPCR. A significant induction in CUX1 expression was observed in plaque zones compared to normal zones (P = 0.036) (Fig. 6a). Consistent with this induction of CUX1, as well as a previous publication showing p16 INK4a is abundantly expressed in atherosclerotic lesions 49 , we also detected a significant increase in p16 INK4a expression with P = 0.011 in the plaque zones (Fig. 6b). Even with these limited numbers of samples, both Spearman correlation and trend analysis identified a significant correlation between the expression levels of CUX1 and p16 INK4a with P = 0.012 and 0.005, respectively (Fig. 6c,d). However, due to the limited number of samples, no significant association could be identified between the expression level of p16 INK4a and fSNP rs1537371 genotypes. To confirm these results, we also performed immunocytochemical staining using antibodies specifically against CUX1 and p16 INK4a . Only CUX1 or p16 INK4a staining colocalized with DAPI staining was used to calculate for fluorescence intensity. A similar induction of both CUX1 (green) and p16 INK4a (red) was observed by comparison of plaque zones to normal zones (Fig. 6e), and was quantitatively evaluated using the nonparametric Mann-Whitney test for pairwise comparisons with P = 0.0025 for CUX1 (n = 8) and P = 0.0006 for p16 INK4a (n = 8) (Fig.  6f). In addition, we also checked the expression of SASP genes IL-6, IL-1β and ICAM1. All three inflammatory markers showed a trend towards elevation in plaque zones, although with limited sample size none reached statistical significance (Fig. 6g). These data, together with our other findings, suggest that stress-induced upregulation of CUX1 may promote atherosclerosis by induction of cellular senescence through modulation of p16 INK4a expression.
To determine whether p53 also plays a role in the induction of cellular senescence in ECs, we first checked its expression in p5 and p10 ECs, as well as in ECs treated with both bleomycin and H 2 O 2 . We detected an elevated level of p53 expression in p10 ECs versus p5 ECs (Fig. 7a), as well as in bleomycin-and H 2 O 2 -treated ECs (Fig.  5a, left and middle and Extended Data Fig. 6a, left and middle). While these data are consistent with previous reports 51,52 , unexpectedly increased expression of p53 could be detected only at the protein level by immunoblot, but not at the mRNA level by qPCR analysis (Figs. 7a and 5a and Extended Data Fig. 6a). This suggests that regulation of p53 expression under these conditions is posttranscriptional. To check whether CUX1 regulates p53 expression in human ECs, we performed shRNA knockdown of CUX1 in p10 ECs. No significant change in p53 expression was observed in these ECs (Fig. 7b). We also did not detect any change in p53 expression when CUX1 was reduced by shRNA knockdown in ECs treated with either bleomycin (Fig. 5a, middle and right) or H 2 O 2 (Extended Data Fig. 6a, middle and right), suggesting that CUX1 regulates cellular senescence independently of p53, at least in ECs.
In addition, we also checked the possibility that p53 regulates CUX1 expression in human ECs which, in turn, regulates cellular senescence. As a result, no alteration in CUX1 expression was detected in p53 siRNA knockdown ECs (Fig. 7c,d). Also, no change in cellular senescence was observed between scrambled control ECs and p53 siRNA knockdown ECs, in terms of either SA-βgal (Fig. 7e) or γ-H2AX staining (Fig. 7f).

Discussion
In this report, we demonstrate that CUX1 regulates both replicative and stress-induced senescence in ECs by activation of the expression of p16 INK4a , a known regulator of senescence induction 9,14,15 . This activation depends on the allele-imbalanced binding of CUX1 to fSNP rs1537371, with the risk allele A binding more CUX1 than the nonrisk allele C. Increased binding of CUX1 to rs1537371 induced a higher level of p16 INK4a expression, which resulted in an increase in cellular senescence. Since accumulation of senescent cells has been detected in atherosclerotic plaques in patients with atherosclerosis 53,54 , we believe that our findings provide a potential pathophysiological mechanism that explains the contribution of fSNP rs1537371 to atherosclerotic risk (Fig. 8a). Consistent with the genetic data described above, we detected increased expression of CUX1 in ECs in a passage-dependent fashion, and following treatment with DNA-damaging agents or oxidative stress. We further demonstrate that induction of CUX1 by all the factors above can induce cellular senescence by activation of p16 INK4a , regardless of the genotype of fSNP rs1537371. We believe that these findings reveal a mechanism implicating CUX1 as an inducer of cellular senescence and, hence, as a potential driver of age-related diseases such as atherosclerosis (Fig. 8b).
In addition, our data suggest that CUX1-induced DNA damage, as detected by γ-H2AX, is p16 INK4a dependent. In particular, we observed that knockdown of p16 INK4a in CUX1-overexpressed ECs rescued the DNA damage phenotype (Figs. 4 and 5) while overexpression of p16 INK4a in CUX1 knockdown restored γ-H2AX staining (Extended Data Fig. 3). These data support the hypothesis that the effect of CUX1 on the DNA damage response is probably mediated through p16 INK4a . Unfortunately, precisely how p16 INK4a affects DNA repair is not well understood 55 . Of note, it was recently reported that upregulation of p16 INK4a decreases both nucleotide and deoxyribonucleotide synthesis and induces oncogene-induced senescence, probably by blocking the repair of DNA damage 56 .
However, we cannot exclude the possibility that CUX1 directly affects DNA repair, as previously reported 57,58 .
Collectively, in this report, we demonstrate how to potentially utilize post-GWAS functional studies to obtain important biological insights. In particular, our data reveal a new mechanism underlying the contribution of CUX1/rs1537371 in the pathogenesis of, and/or susceptibility to, aging and age-related diseases. However, to fully understand the role played by CUX1 in the regulation of cellular senescence and senescence-induced, age-related diseases, a comprehensive in vivo functional study with a CUX1 null mouse model would be warranted.

Primers and antibodies.
All primers used in this study were purchased from IDT and are listed in Supplementary Table 2. All antibodies used are listed in Supplementary Table 3, with corresponding supplier information.
Isolation of atherosclerotic plaques. Atherosclerotic plaques were obtained from patients undergoing carotid endarterectomy at the Department of Surgery at UPMC Presbyterian Hpospital. The use of human materials was approved by the University of Pittsburgh (Institutional Review Board no. STUDY18100138), and written informed consent was obtained from all individuals before operative procedures.

Reel-seq.
To identify fSNPs at the CDKN2A/B locus, a Reel-seq library containing both alleles of the 193 SNP sequences was built with the construct sequence shown in Supplementary Table 2. The library was amplified and regenerated by primers, sequecing and G3 with Accuprime Taq polymerase (Invitrogen). For screening, ~10 µg of NE isolated from PBMC cells (buffer for control) was mixed with ~50 ng of library DNA using the binding buffer provided with the LightShift Chemiluminescent EMSA Kit (Thermo Fisher Scientific), and subsequently incubated at room temperature (RT) for 2 h. The reaction was performed in triplicate, with three buffer-treated controls and three NE-treated samples. All samples were resolved on a 6% TBE native gel for gel shifting. After completion of electrophoresis, unshifted bands from each of the controls and samples were cut and isolated. The isolated library DNA was next amplified by PCR using sequecning and G3 primers, and regenerated libraries were used for the next round  of gel shifting. In total, seven rounds were performed. After screening, standard Illumina amplicon sequencing was performed with the PCR product from rounds 1, 4 and 7 (refs. 28,59 ).

FREP-MS.
FREP-MS assay was performed as previously described 28 . In brief, ~10 µg of FREP construct DNA (Supplementary Table 2 Luciferase reporter assay. Luciferase reporter assays were performed in 293T cells using the pGL3-Promoter vector (Promega, catalog no. E1761). Next, 31-bp, SNPcentered fragments were cloned into Sac I and Xho I sites in the pGL3-Promoter vector. For control, an irrelevant 31-bp DNA fragment was cloned into the same vector. The same amounts of both SNP and control construct were transfected into 293T cells by FuGENE HD transfection reagent (Promega), together with the same amount of control vector, pRL-TK, which provides constitutive expression of Relnlla luciferase (Promega). Luciferase reporter activity was measured by normalization of firefly luciferase reporter activity to Renilla luciferase activity using the Dual-Glo Luciferase Reporter Assay System (Promega). All experiments were performed according to the manufacturer's protocol. Data represent n = 6 independent biological replicates.
CRISPR-Cas9 genome editing. CRISPR-Cas9 was performed using the LentiCRISPR v.2 vector system (Addgene). Lentiviruses were infected into the human microglia cell line HMC3. Single, puromycin-resistant clones were selected using limited-dilution cloning in 96-well plates. Genomic DNA was isolated from each clone, and DNA fragments crossing fSNP rs1537371 were amplified and sequenced. Cells positive for mutations, except for homozygous mutations, were subcloned and the same DNA fragments were cloned into pGEM-T Easy vector (Promega) for sequencing of both alleles. For control, we used polyclonal cells targeted by the same CRISPR-cas9 targeting vector containing a guide RNA sequence irrelevant to rs1537371.
ChIP assay. ChIP was performed as described previously 60 . Briefly, scrambled control shRNA-infected human ECs and CUX1 shRNA knockdown ECs were crosslinked with 1% formaldehyde for 10 min. Sonication was carried out at 30% amplitude, with 20 s on and 50 s off for 5 min, followed by overnight incubation of 10 μg of anti-CUX1 antibody coupled to Dynabeads Protein A/G (Thermo Fisher Scientific, catalog nos. 10001D and 10003D) with sonicated samples at 4 °C. DNA pulled down by antibody, and input DNA, were purified with the Qiagen PCR purification kit after reversal of the crosslink. Purified DNAs were used for qPCR analysis of the DNA fragment containing fSNP rs1537371 with primers ChIP-F and ChIP-R (Supplementary Table 2). ChIP results were measured by normalization of ChIP DNA to input DNA. For antibody isotype control, a rabbit anti-IgG antibody was used; two randomly selected DNA regions were used as negative controls. For sequencing, PCR products from input and ChIP DNA in the scrambled shRNA control were cloned into pGEM-T Easy vector (Promega) and sequenced by Sanger sequencing. Data represent a combination of n = 3 independent samples.
AIDP-Wb analysis. AIDP-Wb was performed as previously described 29 . In brief, a 31-bp biotinylated SNP sequence centered with either the risk or nonrisk allele was generated by annealing two biotinylated primers (IDT). Approximately 1 µg of DNA was then attached to 40 µl of Dynabeads M-280 Streptavidin. DNA beads were mixed with ~100 µg of NE isolated from ECs at RT for 1 h, with rotation. After washing off unbound proteins, DNA-bound proteins were eluted using sample buffer and resolved on an SDS-PAGE gel for immunoblot analysis using an antibody directed against CUX1. For an internal loading control, the same blot was probed using an antibody directed against PARP-1. Data represent n = 3 independent biological replicates. Immunoblot analysis. Whole-cell lysates were prepared using RIPA buffer (Sigma). Cytosolic proteins and nuclear proteins were isolated with NE-PER Nuclear and Cytoplasmic Extraction Reagents (Thermo Scientific) according to the manufacturer's instructions. Proteins were resolved on SDS-PAGE gels and transferred to polyvinylidene difluoride membranes. Proteins were detected with gene-specific antibodies. All antibodies were purchased and used as listed in Supplementary Table 3. For a loading control, α-tubulin was used. Data represent n = 3 independent biological replicates.
Telomere length quantification. Genomic DNA was extracted from human ECs. The Absolute Human Telomere Length Quantification qPCR Assay Kit (ScienCell, catalog no. 8918) was used to measure telomere length. Data represent a combination of n = 3 independent samples.
BrdU proliferation assay. Human EC proliferation was determined by BrdU incorporation using the BrdU Cell Proliferation Assay Kit (Cell Signaling, catalog no. 6813). Briefly, human ECs were incubated with lentivirus(es) for 48 h. Cells were then subcultured (10,000 cells per well) in 96-well plates for 24 h with or without 0.5 μg ml -1 bleomycin, then 1× BrdU was added to the culture medium for DNA labeling. The labeling medium was removed after 2 h, then cells were fixed and DNA was denatured by the addition of 100 μl of fixing/denaturing solution for 30 min. The incorporated BrdU was then detected by a mouse anti-BrdU monoclonal antibody and measured using an anti-mouse IgG, horseradish peroxidase-linked antibody following the manufacturers' instructions. Data for the BrdU proliferation assay represent n = 12 independent biological samples.
Flow cytometry. Flow cytometry analysis was performed as previously reported 61 .
In brief, human ECs cultured in 12-well plates were infected with lentivirus(es) for 48 h. Cells were then incubated for 24 h in fresh medium with or without 0.5 μg ml -1 bleomycin and collected by trypsin. After washing in FACS buffer (PBS containing 5% bovine serum albumin (BSA)), cells were permeabilized with the FoxP3 permeabilization kit (eBioscience). Cell cycle analysis was performed with anti-Ki67 antibody and propidium iodide (eBioscience) staining. Data were acquired and analyzed using a Fortessa Flow Cytometer (Becton Dickinson) and FlowJo software (Tree Star). Data for flow cytometry analysis represent n = 3 independent biological samples. Overexpression of CUX1 and p16 INK4a . For overexpression of human CUX1 p200, human CUX1 cDNA from pXJ42-p200 CUX1 (Addgene) was cloned into the lentiviral expression vector pLVX puro using Xho I and Xba I cutting sites (Takara Bio) and confirmed by sequencing. p16 INK4a was overexpressed using lentiviral expression vector p156RRL (Addgene). Lentiviruses were generated by transfection of 293T cells and used to infect human ECs.

RNA-seq.
Total RNA was extracted from scrambled siRNA knockdown and CUX1 siRNA knockdown ECs at p10 using the RNeasy Mini kit (Qiagen, catalog no. 74104). All samples were quantified and assayed to confirm minimum RNA integrity number of at least 9.3 using an Agilent Bioanalyzer (High Sensitivity DNA Kit, catalog no. 5067-4626). Next, 1 µg of total RNA per sample underwent mRNA capture and was then fragmented at 94 °C for 6 min. Sequencing libraries were prepared according to the manufacturer's protocol using ten cycles of final amplification (KAPA mRNA HyperPrep Kit, catalog no. KK8580 and KAPA UDI Adapter Kit, catalog no. KK8727). Next-generation sequencing was performed on an Illumina NextSeq500 (75-bp paired end) to a targeted depth of ~20 million reads per sample.
RNA-seq data analysis. Paired-end sequencing reads (75 bp) were mapped using STAR to human genome assembly 38 (Hg38) with gencode v.38 annotation 62 . Genes with at least three sample counts >20 raw reads were then analyzed using the R Limma Voom library 63 RNA-seq quantitation pipeline. We defined linear contrast as: difference between CUX1 siRNA-treated samples versus scrambled siRNA-treated samples, normalized raw counts using the default Voom meanvariance transformation and finally calculated log 2 FC and adjusted P values. DEGs were defined as genes with Benjamini-Hochberg-corrected P < 0.05 and fold change >1.5 in the linear contrast between CUX1 siRNA-treated samples versus scrambled siRNA-treated samples. Differentially expressed genes were analyzed using Gene Set Enrichment Analysis 4.1 (GSEA) 64 . GSEA preranked analysis was performed with default parameters, including 1,000 permutations, minimum size 15, maximum size 500, normalization mean div and human gene symbol platform. GSEA input genes were ranked by log 2 FC, and their enrichment scores were calculated using the MSigDB7.0 C2 collection of expert curated gene sets. We defined significant gene sets as those with GSEA nominal P < 0.05 in either enriched or depleted C2 pathways.
Immunocytochemical staining for p16 INK4a and CUX1. Human atherosclerotic plaques were obtained from patients undergoing carotid endarterectomy. The part of the carotid artery showing hard, calcified tissue was used as a plaque zone while the part far from the calcified zone was used as the normal-appearing zone. Both plaque and normal-appearing zones were separated and fixed in 4% buffered formalin for 2 h and stored in 30% sucrose solution containing 0.05% sodium azide overnight. Sections of 10-µm thickness were permeabilized with 0.1% triton X-100 for 4 h and blocked overnight in PBS containing 2% BSA in 96-well cell culture plates. Sections were incubated for a further 24 h with primary antibodies against p16INK4a p16 (Invitrogen, catalog no. MA5-17142, 1:500 dilution) and CUX1 (Proteintech, catalog no. 11733-1-AP, 1:100 dilution). After washing with PBS, sections were incubated for 1 h at RT with fluorochrome-conjugated secondary antibodies (Alexa Fluor 488 goat anti-mouse and Alexa Fluor 647 goat anti-rabbit). Tissue sections were stained and mounted with VECTASHIELD DAPI, and images were taken using confocal laser microscopy and analyzed with imageJ. The data represent two independent experiments, with n = 8 plaque zones and n = 8 eight normal-appearing zones.
Datasets for comparison and visualization of fSNPs. The University of California, Santa Cruz genome browser was used to visualize data and create genomic view snapshots for regulatory regions of CDKN2A/B 65 .
Enhancer and promoter prediction. Enhancer and promoter prediction were performed using the GeneHancer database 66 .
DNase I hotspot. We used the track of DNase I Hypersensitivity on Human Astrocytes-spinal cord from ENCODE.
Histone marker. The layered H3K4Me1 and layered H3K27Ac tracks show where modification of histone proteins is suggestive of enhancer and, to a lesser extent, other regulatory activity. The layered H3K4Me3 track shows a histone mark associated with promoters 31 . We used the tracks of markers H3K27ac, H3K4me1 and H3K4me3 on HUVEC.
Transcription factor ChIP-seq data. This track shows DNA regions where transcription factors, and proteins responsible for modulation of gene transcription, bind-as assayed by ChIP-with antibodies specific to the transcription factor, followed by sequencing of the precipitated DNA (ChIP-seq). We used the tracks of CUX1 on GM12878, K562 and MCF-7.
Statistics and reproducibility. For normally distributed data, all data are presented as mean ± s.e. P values were calculated using two-tailed Student's t-test. Nonnormally distributed data relating to quantification of p16 INK4a expression in Fig. 1h and CUX1 and p16 INK4a immunocytochemical staining in Fig. 6f are presented as median ± interquartile range, and P values were calculated with the nonparametric Mann-Whitney test for pairwise comparison. All data presented in this work are reproducible. No statistical method was used to predetermine sample size. No data were excluded from the analyses. The experiments were not randomized, and no blinding was applied to allocation during experiments and outcome assessment except for phenotype-genotype analysis in Fig. 1h.
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
Reel-seq data, FREP-MS data and differentially expressed genes identified by RNA-seq are provided in Supplementary Data 1, 2 and 3. RNA-seq sequencing data have been deposited in GEO with accession code GSE186528. All other data are available from the corresponding author upon reasonable request.