Identification of Target Genes of the p16INK4A-pRB-E2F Pathway*□S

Received for publication, May 12, 2003, and in revised form, July 21, 2003Published, JBC Papers in Press, August 15, 2003, DOI 10.1074/jbc.M304930200

Richard Vernell‡, Kristian Helin‡§¶, and Heiko Muller‡§

From the ‡Department of Experimental Oncology, European Institute of Oncology, 20141 Milan, Italyand §The FIRC Institute of Molecular Oncology, 20139 Milan, Italy

Deregulation of the retinoblastoma protein (pRB)pathway is a hallmark of human cancer. The core mem-bers of this pathway include the tumor suppressorprotein, pRB, which through binding to a number ofcellular proteins, most notably members of the E2F tran-scription factor family, regulates progression throughthe cell division cycle. With the aim of identifying tran-scriptional changes provoked by deregulation of thepRB pathway, we have used cell lines that conditionallyexpress a constitutively active phosphorylation site mu-tant of pRB (pRB�CDK) or p16INK4A (p16). The expres-sion of pRB�CDK and p16 resulted in significant repres-sion and activation of a large number of genes asmeasured by high density oligonucleotide array analy-sis. Transcriptional changes were found in genes thatare essential for DNA replication and cell proliferation.In agreement with previous results, we found a highdegree of overlap between genes regulated by p16 andpRB. Data we have obtained previously for E2F familymembers showed that 74 of the genes repressed by pRBand p16 were induced by the E2Fs and 23 genes thatwere induced by pRB and p16 were repressed by theE2Fs. Thus, we have identified 97 genes as physiologicaltargets of the pRB pathway, and the further character-ization of these genes should provide insights into howthis pathway controls proliferation. We show that Gibbssampling detects enrichment of several sequence motifs,including E2F consensus binding sites, in the upstreamregions of these genes and use this enrichment in an insilico filtering process to refine microarray derivedgene lists.

Data generated in the last decade have pointed to a centralrole for the retinoblastoma protein (pRB) pathway in regulat-ing the progression through the G1 phase of the mammaliancell cycle (1). The core members of this pathway include, inaddition to pRB (and its family members p107 and p130), theD-type cyclins that, in association with CDK4 and CDK6, pro-mote proliferation of the cell cycle, in part through the phos-phorylation of the pRB family members, and the INK4 family ofcyclin-dependent kinase inhibitors that specifically bind and

inhibit the activity of CDK4 and CDK6. The high frequency bywhich alterations have been identified in the pRB pathway inhuman cancer taken together with the understanding of thecentral role of its core members in regulating cell proliferationhave led several researchers to suggest that the deregulation ofthe pRB pathway is an obligatory event in human cancer (e.g.see Ref. 2).

The biochemical mechanism by which pRB is restricting cellproliferation is widely believed to involve protein-protein inter-actions. Several hundred proteins have been reported to bind topRB (3), however, the relevance of most of these interactions ispoorly understood. The most studied and the best understoodtargets for pRB are members of the E2F transcription factorfamily (4, 5). The E2F transcription factors are essential for theproper transcriptional regulation of a number of genes, whosegene products control the progression through the cell cycle.These genes include CCNE1 (cyclin E1), CCNA2 (cyclin A2),and CDC25A, which are all essential for the entry into the Sphase of the cell cycle, and genes that are involved in theregulation of DNA replication, such as CDC6, DHFR, and TK1(thymidine kinase) (6, 7). When E2Fs bind to the promoters ofthese genes in complexes with pRB or its family members, geneexpression is repressed. The repression is effectively relievedby CDK1-mediated phosphorylation of the pRB family mem-bers, resulting in transcription of the E2F-regulated genes andsubsequent progression through the cell cycle.

Because of the central role of the E2Fs in regulating theprogression through the cell division cycle, we and others haveused gene expression profiling and chromosomal immunopre-cipitation assays to identify novel E2F target genes (5, 8–13).The data obtained from these expression profiles have shownthat the E2Fs, in addition to having a role in regulating theexpression of genes involved in cell cycle progression and DNAreplication, also regulate the expression of genes controllingdifferentiation, DNA repair, and apoptosis. The aim of thecurrent study was to identify transcriptional changes provokedby the reintroduction of a functional pRB pathway in humantumors. By performing such a study we expected to identifygenes whose deregulation, as a result of a non-functional pRBpathway, could contribute to transformation. Moreover, we alsoexpected that the gene expression would give valuable thera-peutic biomarkers, which could be useful when compoundsaimed at restoring a functional pRB pathway in tumor cellswill be tested. To perform these studies, we chose the widelyused human osteosarcoma cell line U-2 OS as an experimentalmodel because these cells do not express p16. Moreover, be-cause of the high level of CDK activity in U-2 OS cells, themajority of pRB is in its hyperphosphorylated state. Reintro-duction of p16 and a non-phosphorylatable version of pRB

* This work was supported by grants from Associazione Italiana perla Ricerca sul Cancro (AIRC), Fondazione Italiana per la Ricerca sulCancro (FIRC), the Italian Health Ministry, the Human Science Fron-tiers Science Program, and the European Unions Fifth FrameworkProgram. The costs of publication of this article were defrayed in part bythe payment of page charges. This article must therefore be herebymarked “advertisement” in accordance with 18 U.S.C. Section 1734solely to indicate this fact.

□S The on-line version of this article (available at http://www.jbc.org)contains Appendix 1 and table.

¶ To whom correspondence should be addressed: Dept. of Experimen-tal Oncology, European Institute of Oncology, Via Ripamonti 435,20141 Milan, Italy. Tel.: 39-0257489831; Fax: 39-0257489851; E-mail:khelin@ieo.it.

1 The abbreviations used are: CDK, cyclin-dependent kinase; HA,hemagglutinin; MES, 2-(N-morpholino)ethanesulfonic acid; PWM, po-sition weight matrix; TGF-�, transforming growth factor-�.

THE JOURNAL OF BIOLOGICAL CHEMISTRY Vol. 278, No. 46, Issue of November 14, pp. 46124–46137, 2003© 2003 by The American Society for Biochemistry and Molecular Biology, Inc. Printed in U.S.A.

This paper is available on line at http://www.jbc.org46124

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

results in cell cycle arrest in these cells, which can be abrogatedby re-expression of the E2Fs (14–18). Here, we report theidentification of 74 genes, whose expression is repressed by p16and pRB, and induced by E2F. Gibbs sampling performed inthe upstream regions of these genes detects the enrichment ofseveral sequence motifs, including E2F consensus bindingsites. The enriched sequence motifs were used in a search ofroughly 10,000 human upstream regulatory regions to identifynovel E2F target gene candidates and to refine the microarraydata.

EXPERIMENTAL PROCEDURES

Cell Culture and Cell Cycle Analysis—U-2 OS cells expressing tetra-cycline responsive p16 or constitutively active pRb (pRb�CDK) alleleswere described previously (17, 18). Cells were cultured in Dulbecco’smodified Eagle’s medium supplemented with 10% bovine calf serum(South American), 2 �g/ml puromycin. 200 �g/ml G418, and 2 �g/mltetracycline. U-2 OS cells expressing HA-tagged ER-E2F1 (5) werecultured in Dulbecco’s modified Eagle’s medium containing 10% bovinecalf serum and 2.5 �g/ml puromycin. Nearly confluent cultures of U-2OS clones were trypsinized and plated at 5 � 106 cells per 15-cm plateon the day before induction. Induction of E2F activity was accomplishedby addition of 4-hydroxytamoxifen to a final concentration of 300 nM for4 h. Exogenous p16 or pRb�CDK proteins were induced by rinsing cellstwice with phosphate-buffered saline followed by growing in tetracy-cline-free media. After the first hour of culture the media was changedagain. RNA was isolated at 12 h for pRb�CDK, or 16 h for p16. For cellcycle analysis, cells were pulse labeled for 30 min in 20 �M bromode-oxyuridine, harvested by trypsinzation, fixed in 1% paraformaldehydefor 5 min, and permeabilized in 70% methanol. DNA was denatured in2 N HCl for 20 min and cells were neutralized in 0.1 M sodium borate,pH 8.5, for 2 min. Bromodeoxyuridine was detected using fluoresceinisothiocyanate-labeled anti-bromodeoxyuridine monoclonal antibody(BD Biosciences), and DNA was stained in 0.5 ml of propidium iodide(10 �g/ml).

RNA Preparation—RNA for analysis by microarray and quantitativereal-time reverse transcriptase-PCR was isolated using the RNeasy kit(Qiagen) and integrity was determined by formaldehyde-agarose gelelectrophoresis. For use in quantitative PCR, the protocol includedon-column treatment with DNase I according to the manufacturer’sprotocol.

Immunostaining—Cells were cultured in the presence or absence oftetracycline for 16 h. Expression of HA-tagged pRb�CDK protein wasdetected with anti-HA (12CA5) antibody. Exogenous p16 protein wasdetected with anti-p16 (Santa Cruz sc-467).

Quantitative Reverse Transcriptase-PCR—Changes in gene expres-sion were confirmed using SYBR Green quantitative reverse tran-scriptase-PCR following the protocols from Applied Biosystems. Prim-ers used can be found in Supplemental Materials.

High Density Oligonucleotide Microarrays—We used HG-U95 A, B,C, D, and E oligonucleotide microarrays (Affymetrix) containing 62,907probe sets. Targets for hybridization to the microarrays were preparedas described (19, 20), except that hybridization was performed in 1 �MES buffer (0.1 M MES, pH 6.7, 1 M NaCl, 0.01% Triton X-100) andchips were washed in 0.1 � MES buffer. Target concentration was 30 �gof fragmented cRNA in 200 �l of hybridization solution. Images werescanned with a HewlettPackard GeneArray scanner and the imageswere analyzed using Affymetrix’s Microarray suite 5.0 software. Alltargets were hybridized to two different sets of HG-U95 microarrays(chips Av2 through E). Chip replicas were analyzed as described in Ref.5. The probe sets that were identified as regulated by the analysis ofchip replicas were inserted into GeneSpringTM 5.1 software as gene listsfor visualization, annotation, and clustering purposes. All data obtainedin the array experiments will be made available.2

Generation and Use of the Promoter-Exon 1 Data Base—We used theUniversity of California Santa Cruz genome browser (assembly re-leased 28th of June 2002)3 to retrieve genomic DNA sequence for 16,031human Reference Sequence accession numbers (NM_001234) availableas of 02/11/02 including 1000 bp of upstream sequence as reported bythe genome browser. Then, we blasted the exon 1 sequence as reportedby the genome browser to 100 bp of the 5�-end of the cDNA as reportedby the reference sequence entry. Only if this blast reported a hit with 99

or 100% identity in the correct orientation (plus/plus strand), did weconsider the promoter-exon 1 sequence for further analysis. This pro-cedure reduced the number of entries in the promoter-exon 1 data basefrom 16,031 to 9,652 entries. All position weight matrix searches usingthis data base were performed exactly as described by Ref. 21.

Position Weight Matrix Similarity (PWM Similarity)—The test forPWM similarity was performed in two different ways. 1) Using50,000,000 random 25-mer oligonucleotide sequences, each sequencewas presented to each PWM at the stringency indicated in Table II. APWM score was calculated as described in Ref. 21. If the PWM score forthat sequence was equal or higher than the stringency indicated inTable II, the oligonucleotide sequence was scored as a hit. In this test,similar matrices have the tendency to recognize similar sequences as ahit. The degree of similarity was quantified using the binomial distri-bution where the number of successes is the number of hits that twoPWMs have in common, the number of trials is the number of hits ofone PWM, and the probability of success is the number of hits of thesecond PWM divided by 50,000,000. The cumulative p value was thencalculated. PWMs were judged similar if this value fell between 0.99999and 1 (i.e. the number of successes lies far out in the tail of thedistribution). The relatively high number of random oligonucleotidesequences was necessary because half of the PWMs would recognizeonly 1 hit in roughly 10,000 sequences. 2) To visualize the similarity ofPWMs, the same random sequences were presented to the PWMs at astringency that would recognize 5% of the sequences as a hit. By doingso, each sequence was assigned a pattern of 0s and 1s, where 0 standsfor no hit with a given PWM and 1 stands for a hit with a given PWM.These patterns were analyzed by cluster analysis using the GeneSpring5.1. software (Silicon Genetics).

Scoring Scheme for in Silico E2F Target Gene Filter—Based on thetest for PWM similarity, each PWM was assigned to a PWM family: 1)E2F site (M00024, M00050, M00180, motif 5, motif “Kel et al.” (21), andmotif “Farnham”); 2) GC-rich (motif 7, motif 20, motif 33, motif 63, andmotif 71); 3) CCAAT box (motif 4); and 4) CCAAT-like (motif 45). Ifwithin a PWM family more than one PWM recognizes a binding site ina promoter, only the PWM with the highest enrichment factor is con-sidered. The enrichment factors of the PWMs that recognize a bindingsite in a promoter are multiplied to give the final score. Multiplicationof enrichment factors is justified because the probability to find two ormore motifs in the same promoter is calculated by multiplying theprobabilities of finding each motif in the promoter.

RESULTS

Previously, we have described the identification of severalhundred genes whose expression changes significantly uponactivation of E2F1, E2F2, or E2F3 (5). In total, between 7 and10% of all genes tested in this screen showed changes in ex-pression level. Computer-assisted motif searches in the pro-moters of these genes did not show a significant enrichment forthe consensus E2F DNA binding site (TTTSSCGC).4 There areseveral possibilities that could account for this observation.Although the more obvious explanation is that many of thegenes are false positives, and several of them are not directtargets of the E2Fs, we do not believe that this is the case,because Northern blotting or quantitative PCR confirmed themicroarray results for more than 95% of the genes tested (Ref.5, and data not shown). Moreover, our current understandingof the E2F binding site consensus and more generally themechanism of E2F mediated regulation of transcription is lim-ited. Several E2F-regulated genes have been described (e.g.CCNE1, CCNA2, and MYBL2) that do not contain perfect con-sensus E2F DNA binding sites (22–24). Likewise, a number ofgenes identified in an E2F4-specific chromatin immunoprecipi-tation assay do not contain a consensus E2F binding site (11).Furthermore, evidence is accumulating that E2F-mediatedtranscription control is exerted in tight collaboration with othertranscription factors such as RYBP, YY1, Max/Mga, and Smads(25–27). Considering these observations, we sought to identifythe functionally relevant E2F target genes by comparing theE2F-induced changes in gene expression to the corresponding

2 array.ifom-firc.it.3 genome.ucsc.edu. 4 H. Muller, unpublished data.

Gene Expression Profiling of the pRB Pathway 46125

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

changes induced by pRB and/or p16. Because endogenous E2Factivity is repressed by the induction of pRB or p16 in U-2 OScells, genes that in this setting show a pattern of regulationopposite to the one observed for induction of E2F activity arelikely to be the relevant E2F target genes that are deregulatedin cancer cells upon loss of p16 or pRB function.

We took advantage of two U-2 OS-derived cell lines thatoverexpress either pRb�CDK (18) or p16 (17) in a tetracycline-regulated manner. Fig. 1 shows that both cell lines uniformlyexpress high levels of p16 or pRb�CDK as tested by immuno-staining after removal of tetracycline (Fig. 1, A and B). Thisobservation was confirmed by Western blot analysis showingthat both pRb�CDK and p16 reach high levels of expression20 h after removal of tetracycline from the culture medium

(Fig. 1, E and F). Both proteins were undetectable at time 0 hand reached near maximum levels of expression 16 h afterinduction. In the p16 cell line, endogenous pRB began to accu-mulate in its hypophosphorylated form 12 h after induction ofp16 expression, whereas the hyperphosphorylated forms beganto diminish at this time (Fig. 1E). Northern blot analysis usinga PAI1 specific probe, a gene previously identified as an E2Frepressed and pRb-induced gene (5, 28), revealed significantaccumulation of PAI1 mRNA in both cell lines 12 h after in-duction of pRb�CDK and 16 h after induction of p16 expression(Fig. 1G). The accumulation of PAI1 mRNA upon induction ofpRb�CDK and p16 expression provides an example of a genewhose expression level changes in the opposite direction asobserved upon induction of E2F activity. The induction of

FIG. 1. Functional characterizationof U-2 OS cells expressing p16 orpRB�CDK in a tetracyclin-regulatedmanner. U-2 OS osteosarcoma cells sta-bly transfected with plasmids encodingtetracycline-repressed wild type p16 orHA-pRb�CDK were tested for uniform ex-pression and to determine the appropri-ate time points for RNA extraction. A,immunofluorescence using an antibodyspecific for p16 was used to determinethat most cells express the endogenousgene upon removal of tetracycline. Stain-ing with 4,6-diamidino-2-phenylindole(DAPI) was used to control for the numberof cells. B, immunofluorescence using anantibody against the HA tag present onHA-pRb�CDK demonstrated that mostcells express the exogenous gene. C andD, cell cycle analysis by propidium iodide-bromodeoxyuridine fluorescence-acti-vated cell sorter analysis. No cell cycleblock can be observed at the time pointsutilized for microarray analysis. E, West-ern immunoblot analysis demonstratesthat the endogenous pRB changes fromthe hyper- to the faster migrating, hypo-phosphorylated form (arrow) upon re-moval of tetracycline from the medium. F,Western immunoblot analysis demon-strating robust expression of HA-pRb�CDK (indicated by arrow) upon re-moval of tetracycline. G, expression of theE2F-repressed PAI1 gene was monitoredupon induction of p16 or HA-pRb�CDKby Northern blot analysis.

Gene Expression Profiling of the pRB Pathway46126

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

pRb�CDK and p16 expression is capable of arresting the cellcycle (17, 18). Cell cycle arrest in turn can regulate the expres-sion of hundreds of genes (e.g. see Ref. 29). We used the induc-tion of PAI1 expression and the repression of cyclin E1 expres-sion as parameters to identify the earliest time point wheresignificant changes in the expression can be observed in theabsence of detectable cell cycle arrest (Fig. 1, C and D). Thetime points chosen for further analysis were 12 h afterthe induction of pRb�CDK and 16 h after the induction of p16.The choice of early time points should ensure maximum spec-ificity in terms of regulation of endogenous E2F activity in theabsence of indirect interference because of cell cycle arrest.

We used oligonucleotide gene expression microarrays toidentify genes whose expression changes significantly in thetwo cell lines after induction of pRb�CDK and p16. Total RNAwas isolated 0 (control) and 12 h after induction of pRb�CDKexpression and 0 (control) and 16 h after p16 expression. Eachtarget cRNA was hybridized to two microarrays to facilitatestatistical analysis of the results. The data were analyzed usingMicroarray suite version 5.0 (Affymetrix) and GeneSpringTM

(Silicon Genetics) software. We also re-analyzed the microarraydata gathered for the induction of E2F1, -2, and -3 activityderived from the HU6800FL � Hu35k chipset (Affymetrix)that had been analyzed previously using Microarray suite ver-sion 4. During this procedure, around 5% of the genes consid-ered previously were discarded as insignificant but the totalnumber of genes found regulated was nearly doubled. Thesedifferences are entirely because of improvements in the statis-tical algorithm used by Microarray suite version 5.0, not todifferences in the analysis settings.

To compare the data that were derived from two differentchipsets, we restricted the following analyses to probe sets thatwere present on both chipsets. In total, we analyzed 42,089probe sets (24.600 Unigenes) in the E2F induction experimentsand 62,907 probe sets (40.915 Unigenes) in the pRb�CDK andp16 induction experiments. 23,527 probe sets (16.897 Uni-genes) were found to be comparable among each other and wereused for further analysis.

Activation of the E2Fs (E2F1, E2F2, and E2F3) or inducedexpression of pRB or p16 resulted in significant changes in theexpression of a number of genes. For each of the activated orinduced proteins, we organized the up-regulated and the re-pressed genes into a total of 10 lists of regulated genes, 5 listsof up-regulated genes, and 5 lists of down-regulated genes. Weasked how to combine the lists in a way that generates a highdegree of specificity without excessive loss of potential candi-date genes. We used published information about the identityof E2F target genes (Refs. 5 and 8–13, and references therein)and assembled a list of genes that had been studied in detail(bona fide target genes) or that had been found regulatedconsistently in different high throughput screens. This listcontains 47 genes, 38 of which were present on both chip setswe analyzed (Fig. 2A). We reasoned that a meaningful combi-nation of lists of regulated genes should maximize both thetotal number of known target genes as well as the proportion ofknown target genes relative to the total number of genes in thecombined list. Because the activity of E2Fs is repressed byinduction of pRB and/or p16, we focused our attention on com-binations of lists with opposing effects on gene expression of theE2Fs on the one hand and pRB/p16 on the other hand. In total,we tested 62 combinations of lists (using the Boolean ANDoperator, see Appendix 1, Supplemental Materials).

Fig. 2B shows the results of combining lists of E2F up-regulated genes with pRB and p16 down-regulated genes. Twodistinct classes of lists are immediately evident. Single listsand combined lists composed only of E2F induction data con-

tain many genes and a low percentage of known E2F targets(type I). Combined lists containing at least one list of E2Fup-regulated genes and one list of pRB or p16 down-regulatedgenes contain significantly fewer genes but a much larger per-centage of known E2F target genes (type II). The combinationof the E2F1, E2F2, and E2F3 derived lists (or any duplicatecombination thereof) does not lead to a significant increase inthe percentage of known E2F target genes in the resultingcombined list (see Fig. 2B). On the other hand, the p16 derivedsingle list of down-regulated genes contains the highest per-centage of known E2F target genes. Interestingly, both thepercentage and the total number of known E2F target genesare higher for the single list of p16 down-regulated genes (7.6%,28 genes) than in the triplicate combination of E2F up-regu-lated genes (1.4%, 5 genes), despite the similar total number ofgenes in these lists. This observation strongly suggests thatinduction of E2F activity alone is not a sufficient stimulus forthe induced expression of some bona fide E2F target genes likeCCNA2, DHFR, TS, and others in our system. The reason maybe that we are using proliferating U-2 OS cells rather thanquiescent cells employed in other studies (8, 10). Furthermore,U-2 OS cells seem to express low levels of BRG1, a necessarycomponent for pRB-mediated repression of E2F target genes(30). At least some of the E2F target genes may thereforealready be expressed at levels that cannot be increased furtherby induction of E2F activity, whereas they may be repressed byinduction of pRB or p16, depending on the availability of co-repressors. Nevertheless, the combination of lists with E2Fup-regulated genes and p16/pRB down-regulated genes leads toa significant increase in specificity of the resulting lists. Notethat the number of genes identified in the combined lists is onaverage nearly 100 times larger than the number of genes thatwould be expected by chance, whereas the number of genesexpected by chance is around 0.5 or less (Fig. 2C). We choselists with more then 20% of known E2F target genes for furtheranalysis. The lists satisfying this criterion are all composed ofboth the p16 and the pRB data and one or two E2F derived lists(see Fig. 2B, encircled).

We also compared p16 and pRB up-regulated genes to E2Fdown-regulated genes. In this case, an apparent increase inspecificity as seen in Fig. 2B was not observed (data notshown). Nevertheless, the number of genes found up-regulatedwith p16 and pRB and down-regulated with at least one E2Fwas significantly higher than would be expected by chance (Fig.2C). It is worth noting that none of the genes defined by thisfilter was a known E2F target gene as defined in Fig. 2A. Wewill call genes that are up-regulated by E2Fs and down-regu-lated by p16 and pRB cluster 1 genes, whereas the genes withopposite behavior will be called cluster 2 genes in thefollowing sections.

Fig. 3A shows a gene tree (generated with GeneSpring soft-ware) that contains the cluster 1 and cluster 2 genes. Cluster 1genes are more abundant than cluster 2 genes (55 Unigenesversus 22 Unigenes, see also Table I). None of the cluster 2genes is part of our list of known E2F target genes. Cluster 1genes display a slightly more homogeneous behavior than clus-ter 2 genes. Fig. 3B shows the -fold change values for cluster 1and cluster 2 genes in the five experiments. The -fold changesare somewhat more pronounced in the E2F experiments but nostriking differences can be observed between cluster 1 andcluster 2 genes in their overall behavior (except for the oppositedirection of change). We used quantitative PCR to confirm themicroarray-derived gene expression profiles for 25 genes thatare evenly distributed over cluster 1 and cluster 2. In all cases,the profiles observed in the microarray experiments were re-produced by this approach (Fig. 3, C and D). All cluster 1 and

Gene Expression Profiling of the pRB Pathway 46127

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

cluster 2 genes are listed in Table I. Table I contains threegenes whose expression levels did not change significantly inthe pRB and p16 microarray experiments. These genes areEGR1, PAI1, and CTGF. For these genes, we have previouslyverified by Northern blotting that they are cluster 1 (EGR1) orcluster 2 (CTGF, PAI1) genes (5). EGR1, PAI1, and CTGF aretherefore examples of false negatives. Less stringent analysisof the microarray data would eliminate these false negatives atthe cost of including potential false positives. We observed thatless stringent gene lists exhibit poorer performance in pro-moter motif searches and therefore preferred to work with highstringency gene lists even though we are aware that we will bemissing a number of true pRB pathway target genes.

We were wondering whether E2F binding sites and/or othersequence motifs are present in the promoter regions of cluster1 and cluster 2 genes. Sequence motif search algorithms can beused to identify motifs that are significantly overrepresented ina set of co-regulated genes. Some of these motifs may be over-represented because they are specifically required for coordi-nated expression of this set of genes, whereas others are gen-eral constituents of a promoter. To distinguish between thesepossibilities, we generated a data base of 9652 human promot-er-exon 1 sequences (see “Experimental Procedures” for de-

tails). We reasoned that specific motifs would be enriched in aset of co-regulated genes, whereas general motifs will not(Fig. 4A).

Transcription factor binding sites are generally described asPWMs. Kel et al. (21) recently described a set of experimentallyproven E2F binding sites that in a modified position weightmatrix search is capable of selecting E2F target genes in silico.The PWM is calculated from a multiple alignment of bindingsites of equal length. Given the PWM and applying the calcu-lation as described (21), any sequence motif of the same lengthas the position weight matrix can be assigned a score. Thehigher the score the higher the homology of the sequence motifunder study to the sequence motifs used to build the PWM. Byfixing a certain score value cutoff (stringency), this approachyields yes/no answers. This way a given sequence motif ispredicted to be homologous or not homologous to the sequencesused to build the PWM.

The method is associated with overprediction (false posi-tives) and with under prediction (false negatives) errors whosecontribution to the total error is dependent on the stringency ofthe search (Fig. 4, B–D, upper panels). For example, if we usedan E2F binding site matrix at a low stringency of the search, wewould identify promoters as containing E2F sites even though

FIG. 2. Identification of target genes of the p16-pRB-E2F pathway by a combination of gene expression lists obtained by highdensity oligonucleotide microarrays. A, list of known E2F target genes. B, number of genes and content of known E2F target genes in everyBoolean combination of gene lists. A gene was considered part of a combination of gene lists when it was found regulated in all constituting lists.Diamond, single gene lists; square, combination of 2 lists; triangle, combination of 3 lists; turquoise cross, combination of 4 lists; asterisk,combination of 5 lists. C, single and combined gene lists chosen for further analysis. Expected number of genes in combined gene lists is calculatedas the product of the fraction of regulated genes in each single experiment (e.g. 966/16898 for E2F1up) multiplied by the total number of Unigenestested (16898).

Gene Expression Profiling of the pRB Pathway46128

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

they do not contain them, whereas at a high stringency of thesearch we would miss promoters that contain E2F bindingsites. Because the optimal stringency (the one that minimizesthe error sum) depends on the nature of the multiple alignmentused to generate the PWM and not on the calculation itself, wetested multiple stringencies and show the associated error as aseparate curve for every PWM tested (Fig. 4, B–D, upper pan-els). The under prediction error was estimated by generatingthe position weight matrix using all sites of the multiple align-ment except for one and then calculating the stringency atwhich the missing site does not get predicted as being homol-ogous anymore. This procedure is repeated for all sequencesthat are part of the multiple alignment. Then, the fraction ofnon-predicted sequences is plotted for every stringency. Forexample, we used the E2F binding sites identified in promotersthat had been shown to bind E2F via chromatin IP as reported(11), generated a position weight matrix using all sites exceptfor one, and calculated the score for the E2F site that had beenleft out. This procedure is repeated for all the sites. By the end

of the procedure, every site has been assigned a score thatvaries slightly from one motif to the next. At low stringencies,the scores of all sites will be above that stringency and there-fore all sites will be recognized as being an E2F site, whereas athigher stringencies some scores will be below that stringencyand therefore these motifs will not be recognized as E2F sites.The fraction of these false negatives relative to all the sitesconstitutes the under prediction error at that stringency.

The data base we used to do motif searches contains 400 bpof sequence per gene around the transcriptional start site (seebelow). Therefore, the sequence motifs to be recognized areembedded in 400 bp of the promoter-exon 1 sequence. Accord-ingly, the overprediction error was estimated using the PWMgenerated from all sequences of the multiple alignment andcalculating the highest stringency at which each of 400 randomsequence motifs of the same length as the PWM gets predictedas being homologous. Then, for every stringency, the fraction ofpredicted sequences is plotted. The total prediction error isrepresented by the sum of the under prediction and the over-

FIG. 3. Gene expression profiles of pRB-p16-repressed and E2F-induced genes (cluster 1) and pRB-p16-induced and E2F-repressedgenes (cluster 2 genes). A, gene tree showing behavior of cluster 1 and cluster 2 genes as measured by microarrays. Red, up-regulation; green,down-regulation; yellow, control state of expression. B, -fold expression changes of cluster 1 and cluster 2 genes as measured by microarrays. C,gene tree showing behavior of selected cluster 1 and cluster 2 genes as measured by quantitative PCR. Red, up-regulation; green, down-regulation;yellow, control state of expression. The names of the tested genes are shown in Supplemental Materials. D, -fold expression changes of cluster 1and cluster 2 genes as measured by quantitative PCR.

Gene Expression Profiling of the pRB Pathway 46129

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

TABLE IpRB pathway target genes

The listed genes were found regulated by pRB and p16 and one of the E2Fs (E2F1, E2F2, or E2F3). Cluster 1 genes are up-regulated by E2F anddown-regulated by pRB and p16. Cluster 2 genes are down-regulated by E2F and up-regulated by pRB and p16.

PS_U95 GenBankTM accession UID Common name Cluster Known target Function

886_at M60527 709 DCK 1 Yes Deoxycytidine kinase54538_at A1343459 1634 CDC25A 1 Yes Cell cycle1491_at M31166 2050 PTX3 1 Inflammation1467_at U12535 2132 EPS8 1 Signal transduction34314_at X59543 2934 RRM1 1 Yes DNA replication1516_g_at L37374 4756 FEN1 1 Yes DNA replication/repair54081_at AI476789 6877 FLJ10483 155017_at AI912772 7988 141060_at M74093 9700 CCNE1 1 Yes Cell cycle63032_at AI653152 9914 FST 1 Signaling32682_at U09087 11355 TMPO 1 Yes Hormone/signaling56493_r_at A1634855 11355 TMPO 1 Yes Hormone/signaling48549_at AL043137 17834 DONSON 165531_at AA948676 19525 FLJ22794 144037_at AL079372 23044 MGC16386 11941_at U33761 23348 SKP2 1 Yes Cell cycle/proteolysis43549_at AA218886 24596 PIR51 1 DNA recombination/repair51130_at AA195220 26516 FLJ10604 1 Yes DNA replication/repair/Rothmund-Thomson syndrome55130_f_at AI206129 31442 RECQL4 151232_at AI684508 34045 CDCA4 11055_g_at M87339 35120 RFC4 1 Yes DNA replication1544_at U39817 36820 BLM 1 DNA replication/repair/Bloom syndrome58354_at A1656807 46677 PRO2000 162155_r_at AA974408 46677 PRO2000 139980_at AB000449 48269 VRK1 1 Serine/threonine kinase56811_at N62196 48480 1809_at U57094 50477 RAB27A 1 GTPase, vesicle transport1801_at U76638 54089 BARD1 1 Post-transcriptional processing of histone36913_at U75679 75257 SLBP 1 Yes mRNA36922_at X59618 75319 RRM2 1 Yes DNA replication947_at D55716 77152 MCM7 1 Yes DNA replication37305_at U61145 77256 EZH2 1 Yes Transcription/chromatin37218_at D64110 77311 BTG3 1 Cell growth1884_s_at M15796 78996 PCNA 1 Yes DNA replication/repair32589_at U20979 79018 CHAF1A 1 Chromatin remodeling37552_at U33632 79351 KCNK1 1 Potassium channel1670_at L23959 79353 TFDP1 1 Cell cycle/transcription38065_at X62534 80684 HMG2 1 Yes DNA packaging37646_at D26018 82502 POLD3 1 DNA replication47114_at AA256061 86617 161011_at A1203206 86617 147148_r_at AI377789 88959 C20orf154 140732_at D83243 89385 NPAT 161417_at AA425325 98427 158749_at T58129 109437 HUNK 1 Signaling38992_at X64229 110713 DEK 1 Oncogene/binds RNA and DNA50771_at AA652869 111496 139269_at L07541 115474 RFC3 1 Yes DNA replication31853_at AF080227 151461 EED 1 Transcription/chromatin63855_r_at AI859865 154443 MCM4 1 Yes DNA replication981_at X74794 154443 MCM4 1 Yes DNA replication40589_at U40572 172278 SNTB2 1 Signaling2031_s_at U03106 179665 CDKN1A 1 Cell cycle45243_at AI818209 180638 FLJ13081 142537_at AA281239 183085 CNOT6L 148996_at AA461343 184164 LOC115106 138847_at D79997 184339 MELK 1 Signaling56111_at AA905058 192843 FKSG14 1 Leucine zipper protein Also called SUZ12.40957_at D63881 197803 JJAZ1 1 transcription/chromatin32791_at L19183 199695 MAC30 157161_at N72576 226414 1 Hs.36167453663_at AI082535 231444 153839_at AI356843 249184 TCF19 1 Yes Transcription/cell cycle45284_at AF169958 260622 HSPC121 1 Butyrate-induced transcript, Rac-signaling pathway92111_g_at A1984229 260622 HSPC121 1 Butyrate-induced transcript, Rac-signaling pathway61866_at N63787 265592 148490_at AF063506 272027 FBXO5 1 Hs.38235633125_at AL043470 279841 FLJ10335 1 Also called Emil/early mitotic inhibitor/Mitosis44010_at AA977204 284137 FLJ12888 133145_at X99226 284153 FANCA 1 DNA repair/Fanconi anemia65720_at AA115295 284208 DKFZP434N161 154964_at AL118653 284270 157585_at AI202201 293836 132696_at X59841 294101 PBX3 1 Transcription789_at X52541 326035 EGR1 1 Yes Transcription44722_at AI224533 351623 DKFZp762L0311 1 Hs.349084

Gene Expression Profiling of the pRB Pathway46130

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

prediction errors (Fig. 4, B–D, upper panels).Using the procedure described above, we characterized

PWMs for the experimentally verified E2F binding sites de-scribed (21), as well as for a number of multiple alignments ofmotifs that have been identified by Gibbs sampling performedon the promoters of known E2F target genes (see below). Thetotal error curve associated with each multiple alignment de-scribes the predictive properties of the corresponding PWM.For the described motifs (21), the minimal total error is foundat stringencies around 0.85 (Fig. 4B, upper panel). At thisstringency, E2F sites that are present in the promoter of a geneof interest can be identified with the highest confidence. If E2Fsites are specifically enriched in the promoters of cluster 1and/or cluster 2 genes, this enrichment should be detectable ata stringency of 0.85. On the other hand, if only E2F sites arespecifically enriched in cluster 1 and/or cluster 2 genes, then formost other PWMs with predictive properties similar to the E2Fsite matrix such enrichment should be absent. Fig. 4, C and D,upper panels, shows two motifs with predictive properties sim-ilar to the one observed for the E2F matrix that was used asspecificity control.

We used our promoter-exon 1 data base to test for an enrich-ment of E2F binding sites and other motifs in our sets of genes.Because E2F sites are clustered around the transcriptionalstart site (21), 200 bp of sequence upstream and downstream ofthe transcription start site were included for every gene (nointron 1 sequences were included, i.e. only exon 1 was usedwhen exon 1 was shorter than 200 bp). Enrichment of E2Fbinding sites in a given set of genes (cluster) is measured bycomparing the number of genes containing the E2F site in ourcluster of interest to the expected number of genes containingan E2F site in a randomly selected set of genes of equal size. Ifthe difference between these numbers is larger than threestandard deviations, we consider the difference significant. Inthis scenario, particular care needs to be applied to the deter-mination of the expected number of genes containing a bindingsite in a random set of genes, and the standard deviationassociated with that number. To determine these two meas-ures, for every PWM search, 50 randomly selected sets of genesof the same size as cluster 1 (i.e. 50 times 27 randomly selectedpromoters) and cluster 2 (i.e. 50 times 8 randomly selectedpromoters) were generated. This way, for each random set ofgenes we obtained a number of genes predicted to contain a

binding site at any given stringency. The 50 resulting numbersat any given stringency are used to calculate the mean � S.D.at that stringency value (Fig. 4, B–D, middle panels).

Because it is not obvious that an enrichment of E2F bindingsites in our clusters 1 and 2, should it be detected, is biologi-cally relevant rather than an artifact, we included a set ofgenes that had been selected via chromatin immunoprecipita-tion using E2F1- and E2F4-specific antibodies (E2F-ChIPgenes) as a positive control (9, 11). In these genes, E2F bindingsites are present in the promoters with very high probability.As a further control, we included the promoters of the knownE2F target genes as depicted in Fig. 2A.

Next, we tested cluster 1, E2F-ChIP genes, and known E2Ftargets for their content of E2F binding sites that are predictedto be homologous to the experimentally proven E2F bindingsites used in the study by Kel et al. (21) (see Table II). Fig. 4B,upper panel, shows the errors associated with that search. Ascan be seen from the middle panel in Fig. 4B, in the range ofstringency values from 0.84 to 0.9 that have the best predictiveproperties (minimal total error sum), all three sets of genescontain significantly more genes predicted to contain an E2Fsite than could be expected by chance. Fig. 4B, lower panel,shows an increase in the enrichment factor with increasingstringency of the search. Fig. 4, C and D, shows the situationobtained for searches using a position weight matrix with sim-ilar predictive properties but describing a general motif. In thiscase, the curves observed for cluster 1 genes, E2F-ChIP genes,and known target genes fluctuate around the mean (Fig. 4, Cand D, middle panel) and no enrichment can be seen withincreasing stringency of the search (Fig. 4, C and D,lower panel).

These results suggest (i) that the enrichment for E2F bindingsites present in the E2F-ChIP set of genes is detectable usingour data base of promoters in conjunction with the PWM searchdescribed by Kel et al. (22) and (ii) that the cluster 1 genes showa similar if not even a more pronounced enrichment for E2Fbinding sites as compared with known E2F target genes. Forcluster 2 genes, no such enrichment was detectable (data notshown). The reason for the lack of enrichment is unclear at themoment, but may be purely technical because our promoterdata base contained only 10 promoter sequences for the 23genes belonging to cluster 2. With this low number of promot-ers, the standard deviation associated with the expected num-

TABLE I—continued

43055_at AI472084 356392 1 SKP232683_at U18271 138356_at M19481 165887_at AI830948 4994 TOB2 2 Signaling Hs.40897763396_at AL043006 17667 SH3BP4 2 Signaling50139_at AA747315 29068 256727_r_at N20929 34955 2 Hs.28768646074_at W51832 56123 2 Hs.34180636638_at X78947 75511 CTGF 2 Signaling64933_at AI458432 75611 LOC154952 2 Hs.115659-VIK/transcription33113_at U65093 82071 CITED2 2 Transcription672_at J03764 82085 PAI1 2 Proteolysis38418_at X59798 82932 CCND1 2 Cell cycle65880_at AA192516 83883 TMEPAI 255107_at AI916306 87125 EHD3 2 EH-containing protein48896_at AI082244 93764 CPA4 2 Carboxypeptidase/protein turnover48776_at AA843531 100002 DNLC2A 2 Motility/signaling48820_at AI972901 107882 DIP13B 2 Signaling57650_at AI075744 133207 PPFIBP1 2 Intracellular transport65917_at AW015575 180952 DCTN4 2 Hs.328865/signaling57027_at AI990405 194691 RAI3 265585_at AA527515 334860 MGC16279 2 Hs.38013760896_at AI634852 367688 258858_s_at AW004635 374612 2 Hs.18102233352_at X57985 263664_at AA186367 2

Gene Expression Profiling of the pRB Pathway 46131

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

ber of hits is too large to permit reliable detection of enrich-ment. Therefore, the following analysis was restricted tocluster 1 genes only.

Next, we asked whether similar enrichment could be de-tected for motifs other than E2F binding sites. To this end, wetested 280 publicly available PWMs (TRANSFAC) on cluster 1genes, E2F-ChIP genes, and known E2F target genes. Theresults of some of these searches are shown in Table II. Asexpected, the matrices M00024, M00050, and M00180 that aredescribing the E2F binding site show significant enrichment.No significant enrichment was observed for Sp1 (M00008), YY1(M00069), MYCMAX (M00123), p53 (M00272), the TATA ele-ment (M00216), and others (data not shown). In the case of theTATA element, the tendency was clearly toward exclusion fromcluster 1, E2F-ChIP, and known E2F target genes, which iscompatible with the lack of TATA elements in most known E2Ftarget genes.

The fact that E2F binding sites are significantly overrepre-sented in three independently defined sets of E2F target genes(known targets, E2F-ChIP targets, and cluster 1 genes) sug-gests that it might be possible to identify sequence motifs thatmay cooperate with E2F binding sites in the regulation of E2Ftarget genes. We turned to Gibbs sampling to identify suchmotifs. In particular, we used the AlignACE program (31, 32) toperform Gibbs sampling on cluster 1, E2F-ChIP, and knownE2F target genes. We obtained more that 120 different motifs(data not shown) that are overrepresented in these sets ofgenes as measured by the background model implemented byAlignACE. To understand whether any of these motifs arebiologically relevant for the regulation of E2F target genes, weneeded to distinguish the general motifs (e.g. Sp1 site andTATA box) from the specific ones (e.g. E2F site). Therefore, wetested each of the motifs for enrichment in E2F target genes asdescribed in Fig. 4. We call this test the motif specificity test.

FIG. 4. Strategy for the identification of sequence motifs that are enriched in the promoters of E2F target genes. A, known sequencemotifs or motifs identified by motif search algorithms can be significantly more abundant in the promoters of co-regulated genes, because they arespecifically needed for co-regulation of these genes or because they are generally required by the transcription machinery. To distinguish specificmotifs from general motifs, we measure the enrichment of the motif in sets of co-regulated genes as compared with randomly chosen sets of genesof equal size. B, upper panel, under prediction error (blue line), overprediction error (violet line), total prediction error (yellow line) for E2F bindingsite model described by Kel et al. (21) (see also Table II). x axis, stringency parameter; y axis, fraction of motifs predicted to fit the model. Middlepanel, E2F binding site model detects enrichment of this site in cluster 1 genes (brown line), E2F-ChIP genes (yellow line), and known E2F targetgenes (red line). The prediction error of the model (violet line) is identical to the total prediction error shown in the upper panel. The dotted blueline (mean � 3 S.D.) shows the significance cutoff for the smallest group of co-regulated genes (cluster 1 genes). For E2F-ChIP genes and knowntarget genes the standard deviations are slightly smaller (not shown). The black line shows the mean fraction of promoter sequences predicted tocontain the E2F site. The mean � S.D. (error bars) are calculated from 50 randomly selected sets of genes of the same size as cluster 1 genes. Lowerpanel, enrichment factors of E2F site for E2F-ChIP genes (blue bar), cluster 1 genes (red bar), and known target genes (white bar). The enrichmentfactors are calculated by dividing the observed fraction of genes predicted to contain an E2F site (brown, yellow, and red lines in middle panel) bythe expected mean fraction of genes predicted to contain an E2F site (black line in middle panel). C, like panel B for motif 9 that was identifiedby Gibbs sampling, whose position weight matrix has prediction properties similar to the properties of the E2F site matrix but is not specificallyenriched in sets of E2F target genes. D, like in panel C for motif 11 (another motif identified by Gibbs sampling).

Gene Expression Profiling of the pRB Pathway46132

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

TABLE IIMotifs enriched in E2F target gene promoters

Gene Expression Profiling of the pRB Pathway 46133

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

Whereas AlignACE measures overrepresentation of motifsbased on statistical arguments considering only the set of genesthat has been used as input, the motif specificity test measuresthe specificity of a motif to a set of genes of interest as com-pared with all the promoters in the data base. Most motifsfailed the specificity test. However, a significant number ofmotifs did indeed show specific enrichment in the three inde-pendent sets of E2F target genes mentioned above (Table II,motifs 4, 5, 7, 20, 33, 45, 63, and 71). These observationssuggest that the data base of promoters is of good quality andthat the sets of target E2F target genes are sufficiently refinedto allow for the identification of sequence motifs that are co-enriched together with E2F binding sites.

We were wondering whether the enrichment of sequencemotifs other than E2F sites in E2F target genes could be usedto generate an in silico filter for E2F target genes. Such a filtercould be used to enhance the specificity of microarray-derivedgene lists for known E2F target genes or to identify E2F targetgenes in other tissues, including genes that may not be ex-pressed in U-2 OS cells. To build such a filter, we first charac-terized all the motifs that had shown significant enrichment inE2F target genes in more detail (Table II). In particular, weidentified the stringency for each motif that represents a goodcompromise between maximum enrichment factor and mini-mum of total prediction error. As can be seen in Fig. 4B, lowerpanel, the maximum enrichment factor for E2F motifs is foundin the region of high search stringency, a region that is associ-ated with many false negative but few false positive predic-tions, and not in the region of minimal total prediction error.Because for most genes in the promoter data base we do notknow whether a binding site prediction is correct or not, deter-mination of the true prediction error of a given PWM is impos-sible. Therefore, we preferred high stringency searches. Forevery motif, we determined the stringencies at which 5 to 25%of the known E2F target genes are recognized as containingthat motif. Within that region we fixed the stringency that isassociated with the best enrichment factor for further analysis(see Table II). Table II also shows the enrichment factor ob-served at that stringency as well as the number of genes thathave been identified as containing that motif.

Next, we needed to devise a scoring scheme that wouldcombine the individual predictions for each motif in a way thatis most predictive for known E2F target genes. If the motifswere truly independent, just multiplying the enrichment fac-tors for each motif found would represent such a scheme. How-ever, some of the motifs are by definition not independent. Forexample, the motifs M00024, M00050, M00180, Kel et al., andFarnham are all E2F motifs. If a promoter contained an E2Fbinding site that is recognized by all these position weightmatrices, the resulting score would be largely overestimatingthe significance of that finding. This situation can be correctedby grouping these five matrices together. Then, if any one of thematrices of that group identifies a promoter as containing anE2F binding site, we use only one enrichment factor (the big-gest one) in the scoring scheme.

The situation with the remaining motifs is far less obvious,because we do not even know whether the motifs are similar ornot. However, if two matrices are similar to one another, theyshare a large subset of sequences that are recognized by bothmatrices at the stringency fixed in Table II. This subset wasexplicitly identified for 50,000,000 random 25-mer sequences.Then, we asked whether the number of sequences that arerecognized by both matrices is significantly larger than therandom expectation. This approach identified four PWM fam-ilies: E2F site (M00024, M00050, M00180, motif 5, motif Kel etal., motif Farnham), GC-rich (motif 7, motif 20, motif 33, motif

63, and motif 71), CCAAT box (motif 4), and CCAAT-like (motif45) (see “Experimental Procedures” for details of the calcula-tion). These families can be visualized in a cluster tree (Fig 5A).

Every motif shown in Table II is enriched in E2F targetgenes and it is reasonable to assume that the presence of morethan one of these motifs makes it more likely for a gene to beregulated by E2F. We devised a scoring scheme for the qualityof the in silico prediction based on the product of enrichmentfactors present in the promoter (see “Experimental Procedures”for details). By applying the in silico filter to the data base of9652 promoters we assigned a score to every promoter in thedata base. Most promoters have a score of one (i.e. no motifenriched in E2F target genes is present) but some promotershave significantly higher scores (data not shown). To estimatethe error rate of the filter, once again we used known E2Ftarget genes as defined in Fig. 2A as positive controls. Byincreasing the cutoff on the score (i.e. increasing the stringencyof the filter) we obtained lists of genes that contain ever lessfalse positives and ever more false negatives. Fig. 5B shows theprediction error of the in silico filter as a function of stringency.When the stringency is set between 1 and 7, the composition ofthe total error changes from being mainly false-positive errorsto being mainly false-negative errors. From stringency 7 to 14,there is a plateau with few changes in the error rate. Fromstringency 14 onward the false negative rate slowly growstoward 100% with a concomitant increase in the fraction ofknown E2F target genes among the hits. This fraction reaches18% at a stringency of 20.

We tested different stringencies of the in silico filter bycombining the in silico derived gene lists with the microarray-derived gene lists and asked whether the fraction of knownE2F target genes in the combined gene lists would grow. Theresults of these combinations using stringency 10 for the insilico filter are shown in Fig. 5, C and D. At this stringency, thepresence of at least two different cis-acting sequence elementsis required in the promoter because no single motif has anenrichment factor that is as big as 10. As can be seen from Fig.5, C and D, in all cases, in silico filtering of the microarray-derived gene lists leads to a significant enrichment of knownE2F target genes in the resulting combined gene lists. Theenrichment is much more pronounced for E2F down-regulatedgenes and for p16/pRB up-regulated genes. The reason for thisobservation is unclear at the moment, but it may reflect a biasin the set of known E2F target genes toward cluster 1 genes(up-regulated by E2F, down-regulated by pRB and p16). It isworth noting that the genes shown in Fig. 5E have been chosenbecause they have not yet been reported as E2F target genes.Many (but not all) of the known target genes have been pre-dicted by the in silico filter as well (e.g. MCM5, MCM6, MCM7,DHFR, RRM1, RRM2, RBL1, RFC3, and others). These genesare implicitly shown in Fig. 5, C and D. Some of the novel E2Ftarget genes (e.g. BLM and BTG3) reported here have beenpredicted by the in silico filter and have also been verified byquantitative PCR (Fig. 3) without being shown in Fig. 5E.

DISCUSSION

We report here the identification of several novel E2F targetgenes by combining the results of five independent microarrayscreens (induction of transcriptional activity of E2F1, E2F2,E2F3, induction of expression of p16, and pRb�CDK) and showthat the E2F binding site along with other motifs is enriched inthe upstream regulatory regions of these genes. A computa-tional approach to the identification of E2F target genes hasbeen described previously (21). This work employed the E2Fsite to search for novel targets. We show that this E2F motif isenriched 4.4-fold in our set of E2F target genes (Table II). Asimilar enrichment factor was observed for sets of E2F target

Gene Expression Profiling of the pRB Pathway46134

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

genes that were identified by chromatin immunoprecipitationor in bona fide E2F targets. Therefore, a search based on theE2F motif only is likely to be associated with a high rate oferror. In this report, we show that other motifs are co-enrichedin the promoters of E2F target genes. It is, therefore, possibleto build a more selective in silico filter. Genes identified by thisin silico filter were indeed regulated in our microarray screens(Fig. 5E).

To our knowledge, this is the first report that identifies abiologically meaningful motif (the E2F site) in the promoters ofmammalian genes that are co-regulated in microarray screens.The methodology works well in yeast (31, 32). In the mamma-lian system, several difficulties need to be overcome that aremainly related to the quality of the promoter data base and thelevel of noise in microarray screens. The promoter data baseused in this report is based on sequence annotations performedat the University of California Santa Cruz (Golden Path) andan additional cleaning step that ensures that the 5�-end of thereference sequence cDNA coincides with the exon 1 sequencereported in the genome browser. The noise level of the microar-ray data was lowered by cross-comparing five independentmicroarray screens. With these data on hand, we were able todistinguish general motifs found in most promoters from themotifs that are specifically enriched in E2F target genes using

the E2F binding site as a positive control. The enriched motifswere grouped into four families, GC-box, CCAAT-box, CCAAT-like box, and E2F binding sites. As opposed to the motif spec-ificity search performed by the AlignACE-ScanACE package(31), our approach takes advantage of biological data ratherthat statistical arguments to distinguish general from specificmotifs. Thus, the utility of our approach is 2-fold. Motifs thatare specifically enriched in groups of co-regulated genes mayreveal the involvement of a transcription factor binding to thatmotif in the regulation of these genes. On the other hand,enrichment of a set of motifs in the promoters of co-regulatedgenes can be used as an in silico filter to identify novel mem-bers of the co-regulated cluster. This is useful, for instance,when the tissue of interest is of low abundance or hard towork with.

Previously, we reported that E2F target genes cover a widerange of functions, namely that they are involved in the regu-lation of apoptosis, cell cycle, differentiation, and development.Independent studies have confirmed the relevance of E2F inthe regulation of apoptosis via transcription of APAF1 andseveral caspases (33, 34). However, several published reportshinted toward underrepresentation of cell cycle and checkpointregulators in our gene lists (8–11, 35). The reason for thesediscrepancies is unclear but may be related to the fact that we

FIG. 5. Identification of E2F target genes using in silico filtering of microarray data. A, similarity of PWMs visualized by clusteranalysis (see “Experimental Procedures” for details). B, error profile of the in silico filter. C, plot of the number of genes versus the fraction of knownE2F target genes in the indicated gene lists. Every gene list is shown before and after application of the in silico filter. D, fraction of known E2Ftarget genes in each gene list before (blue bar) and after (brown bar) application of the in silico filter. E, microarray data of some novel E2F-targetgenes that have been predicted in silico and have been found regulated in at least one microarray experiment.

Gene Expression Profiling of the pRB Pathway 46135

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

studied proliferating U-2 OS as opposed to quiescent cells usedby other groups. Furthermore, U-2 OS cells express low levelsof BRG1 whose activity contributes to pRB-mediated repres-sion of transcription (30, 36). In this system, many E2F targetgenes may be expressed at levels that cannot be increasedsignificantly by induction of E2F activity. Moreover, the use ofasynchronous cells may render signals from genes that aresensitive to E2F-mediated regulation only at specific points inthe cell cycle less robust.

We observed that induction of pRB�CDK was capable ofreverting the changes induced by activation of E2F (5). Inparticular, induction of pRB�CDK expression led to decreasedlevels of CCNE1 and EGR1 mRNAs, two genes that are in-duced by activation of E2F, while the levels PAI1 and CTGFmRNAs (whose expression levels are down-regulated by E2Factivity) were induced. These results suggest that the pRB-E2Fpathway-specific genes can be identified by cross-comparison ofgene lists that are specific for activators of the pathway(E2F1–3) with gene lists derived from suppressors of the path-way (pRB, p16). Importantly, cross-comparison of gene listsreduces the level of noise but enhances the significance of weaksignals. Consequently, regulators of the cell cycle, DNA repli-cation, DNA repair, and checkpoints constitute a major fractionof pRB-E2F pathway-specific genes defined here, in agreementwith other reports (8–11, 35).

Differentiation and development are complex processes thatare governed by a multitude of pathways. Mutual cross-talkbetween a number of these pathways has been well docu-mented (20, 37–48). The genes reported here provide startingpoints to investigate mechanistic links between some of thesepathways and the pRB-E2F pathway. Most signal transductionpathways are composed of a ligand, a receptor, one or moresignal transducers, a transcription factor, and target genes.Some of the target genes have feedback functions to monitorthe activity of the pathway, whereas others perform some bio-logical function, often by influencing the activity of other path-ways at one or more levels mentioned above (49). The nature ofthe pRB-E2F target genes identified here illustrates thisnotion.

For example, TGFBR2 and FST are genes involved in TGF-�signaling. The TGF-� superfamily of ligands comprises TGF-�,bone morphogenetic proteins, and activins, which are involvedin the regulation of various biological processes and induce G1

arrest in many cell types (50). Cross-talk between E2F, TGF-�,and c-Myc pathways has been reported recently (27, 51), andthe regulation of TGFBR2 and FST may add a further level ofcomplexity. EZH2, EED, and SUZ12 (also called JJAZ1) arethree genes of the polycomb group and were found in a commoncomplex (52). The EZH2-EED complex methylates histone H3at lysine 27. EZH2 contains a SET domain, which is a charac-teristic domain of most histone lysine methyltransferases andit is likely that EZH2 is the catalytic subunit of the EZH2-EEDcomplex. EZH2 has recently been shown to promote the pro-gression of prostate cancer (53). PBX3 is a homeobox gene ofthe TALE family with extensive homology to PBX1, a humanproto-oncogene (54). Milech et al. (55) reported that a splicevariant of PBX3 whose product cannot interact neither withMEIS proteins nor with PREP1 is preferentially expressed inleukemic cells. The SIL gene has been reported to be necessaryfor mouse embryonic axial development and left-right specifi-cation (56). SIL is involved in translocations in T-cell leukemiasand a transactivation domain-deficient variant can promoteT-cell malignancies in transgenic mice (57). The BTG3 andTOB2 genes are members of a family of seven proteins withpoorly defined function (58). It appears that all family memberscan inhibit proliferation. It has been suggested that members

of this family are induced by genotoxic stress and by changes inthe redox potential (59, 60). MAD4 is a member of the MADfamily of proteins that coordinate proliferation with differenti-ation by repressing E-box dependent transcription (61, 62).Thymopoietin (TMPO) and oncostatin M receptor (OSMR) areinvolved in hematopoietic cytokine signaling. Oncostatin M,the ligand of OSMR, has been reported to induce down-regula-tion of cyclin D1 and cyclin D2 expression via signal transduc-ers and activators of transcription 3 (63). TMPO is a hormonewith pleiotropic actions on prothymocytes, and mature T cells(64, 65). FLI1 is an ETS family transcription factor whose DNAbinding domain is an essential part of the EWS-FLI1 fusionprotein found in Ewing’s sarcoma (66, 67). This fusion proteindown-regulates expression of TGFBR2, induces a p53-depend-ent growth arrest, and induces expression of ID2 (68–70). Tosummarize, the pRB-E2F pathway-regulated genes reportedhere cover all types of molecules that constitute signal trans-duction pathways, namely ligands (e.g. FST and TMPO), recep-tors (e.g. TGFBR2 and OSMR), signal transducers (e.g.CDC25A, HSPC121), transcription factors including feedbackregulators (TFDP1, PBX3, and MAD4), and numerous execut-ers of biological functions like DNA replication and repair(MCM4 and FANCA). We believe that more detailed studies onthe genes reported here may help to understand how the pRB-E2F pathway contributes to the integration of proliferation,differentiation, and apoptosis during development.

Acknowledgments—We thank Elena Colli and Elena Prosperini forexcellent technical assistance. We thank Liang Zhu, Jiri Bartek, andJiri Lukas for the generous gift of reagents.

REFERENCES

1. Sherr, C. J. (1996) Science 274, 1672–16772. Hanahan, D., and Weinberg, R. A. (2000) Cell 100, 57–703. Morris, E. J., and Dyson, N. J. (2001) Adv. Cancer Res. 82, 1–544. Dyson, N. (1998) Genes Dev. 12, 2245–22625. Muller, H., Bracken, A. P., Vernell, R., Moroni, M. C., Christians, F., Grassilli,

E., Prosperini, E., Vigo, E., Oliner, J. D., and Helin, K. (2001) Genes Dev. 15,267–285

6. Helin, K. (1998) Curr. Opin. Genet. Dev. 8, 28–357. Trimarchi, J. M., and Lees, J. A. (2002) Nat. Rev. Mol. Cell Biol. 3, 11–208. Ishida, S., Huang, E., Zuzan, H., Spang, R., Leone, G., West, M., and Nevins,

J. R. (2001) Mol. Cell. Biol. 21, 4684–46999. Ren, B., Cam, H., Takahashi, Y., Volkert, T., Terragni, J., Young, R. A., and

Dynlacht, B. D. (2002) Genes Dev. 16, 245–25610. Polager, S., Kalma, Y., Berkovich, E., and Ginsberg, D. (2002) Oncogene 21,

437–44611. Weinmann, A. S., Yan, P. S., Oberley, M. J., Huang, T. H., and Farnham, P. J.

(2002) Genes Dev. 16, 235–24412. Markey, M. P., Angus, S. P., Strobeck, M. W., Williams, S. L., Gunawardena,

R. W., Aronow, B. J., and Knudsen, E. S. (2002) Cancer Res. 62, 6587–659713. Wells, J., Yan, P. S., Cechvala, M., Huang, T., and Farnham, P. J. (2003)

Oncogene 22, 1445–146014. Zhu, L., van den Heuvel, S., Helin, K., Fattaey, A., Ewen, M., Livingston, D.,

Dyson, N., and Harlow, E. (1993) Genes Dev. 7, 1111–112515. Lukas, J., Parry, D., Aagaard, L., Mann, D. J., Bartkova, J., Strauss, M.,

Peters, G., and Bartek, J. (1995) Nature 375, 503–50616. Lukas, J., Petersen, B. O., Holm, K., Bartek, J., and Helin, K. (1996) Mol. Cell.

Biol. 16, 1047–105717. Jiang, H., Chou, H. S., and Zhu, L. (1998) Mol. Cell Biol. 18, 5284–529018. Lukas, J., Sorensen, C. S., Lukas, C., Santoni-Rugiu, E., and Bartek, J. (1999)

Oncogene 18, 3930–393519. Lockhart, D. J., Dong, H., Byrne, M. C., Follettie, M. T., Gallo, M. V., Chee,

M. S., Mittmann, M., Wang, C., Kobayashi, M., Horton, H., and Brown, E. L.(1996) Nat. Biotechnol. 14, 1675–1680

20. Fambrough, D., McClure, K., Kazlauskas, A., and Lander, E. S. (1999) Cell 97,727–741

21. Kel, A. E., Kel-Margoulis, O. V., Farnham, P. J., Bartley, S. M., Wingender, E.,and Zhang, M. Q. (2001) J. Mol. Biol. 309, 99–120

22. Geng, Y., Eaton, E. N., Picon, M., Roberts, J. M., Lundberg, A. S., Gifford, A.,Sardet, C., and Weinberg, R. A. (1996) Oncogene 12, 1173–1180

23. Schulze, A., Zerfass, K., Spitkovsky, D., Middendorp, S., Berges, J., Helin, K.,Jansen-Durr, P., and Henglein, B. (1995) Proc. Natl. Acad. Sci. U. S. A. 92,11264–11268

24. Lam, E. W., and Watson, R. J. (1993) EMBO J. 12, 2705–271325. Schlisio, S., Halperin, T., Vidal, M., and Nevins, J. R. (2002) EMBO J. 21,

5775–578626. Ogawa, H., Ishiguro, K., Gaubatz, S., Livingston, D. M., and Nakatani, Y.

(2002) Science 296, 1132–113627. Chen, C. R., Kang, Y., Siegel, P. M., and Massague, J. (2002) Cell 110, 19–3228. Koziczak, M., Krek, W., and Nagamine, Y. (2000) Mol. Cell. Biol. 20,

2014–2022

Gene Expression Profiling of the pRB Pathway46136

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

29. Cho, R. J., Huang, M., Campbell, M. J., Dong, H., Steinmetz, L., Sapinoso, L.,Hampton, G., Elledge, S. J., Davis, R. W., and Lockhart, D. J. (2001) Nat.Genet. 27, 48–54

30. Zhang, H. S., Postigo, A. A., and Dean, D. C. (1999) Cell 97, 53–6131. Hughes, J. D., Estep, P. W., Tavazoie, S., and Church, G. M. (2000) J. Mol.

Biol. 296, 1205–121432. Roth, F. P., Hughes, J. D., Estep, P. W., and Church, G. M. (1998) Nat.

Biotechnol. 16, 939–94533. Moroni, M. C., Hickman, E. S., Denchi, E. L., Caprara, G., Colli, E., Cecconi, F.,

Muller, H., and Helin, K. (2001) Nat. Cell Biol. 3, 552–55834. Nahle, Z., Polakoff, J., Davuluri, R. V., McCurrach, M. E., Jacobson, M. D.,

Narita, M., Zhang, M. Q., Lazebnik, Y., Bar-Sagi, D., and Lowe, S. W. (2002)Nat. Cell Biol. 4, 859–864

35. Stanelle, J., Stiewe, T., Theseling, C. C., Peter, M., and Putzer, B. M. (2002)Nucleic Acids Res. 30, 1859–1867

36. Wong, A. K., Shanahan, F., Chen, Y., Lian, L., Ha, P., Hendricks, K., Ghaffari,S., Iliev, D., Penn, B., Woodland, A. M., Smith, R., Salada, G., Carillo, A.,Laity, K., Gupte, J., Swedlund, B., Tavtigian, S. V., Teng, D. H., and Lees,E. (2000) Cancer Res. 60, 6171–6177

37. Sherr, C. J., and McCormick, F. (2002) Cancer Cell 2, 103–11238. Crabtree, G. R., and Olson, E. N. (2002) Cell 109, (suppl.) S67–S7939. O’Shea, J. J., Gadina, M., and Schreiber, R. D. (2002) Cell 109, (suppl.)

S121–S13140. Rathmell, J. C., and Thompson, C. B. (2002) Cell 109, (suppl.) S97–S10741. van de Wetering, M., de Lau, W., and Clevers, H. (2002) Cell 109, (suppl.)

S13–S1942. Classon, M., and Harlow, E. (2002) Nat. Rev. Cancer 2, 910–91743. D’Andrea, A. D., and Grompe, M. (2003) Nat. Rev. Cancer 3, 23–3444. Downward, J. (2003) Nat. Rev. Cancer 3, 11–2245. Ferrara, N. (2002) Nat. Rev. Cancer 2, 795–80346. Karin, M., Cao, Y., Greten, F. R., and Li, Z. W. (2002) Nat. Rev. Cancer 2,

301–31047. Ruiz i Altaba, A., Sanchez, P., and Dahmane, N. (2002) Nat. Rev. Cancer 2,

361–37248. Vivanco, I., and Sawyers, C. L. (2002) Nat. Rev. Cancer 2, 489–50149. Freeman, M. (2000) Nature 408, 313–31950. Attisano, L., and Wrana, J. L. (2002) Science 296, 1646–164751. Feng, X. H., Liang, Y. Y., Liang, M., Zhai, W., and Lin, X. (2002) Mol. Cell 9,

133–14352. Cao, R., Wang, L., Wang, H., Xia, L., Erdjument-Bromage, H., Tempst, P.,

Jones, R. S., and Zhang, Y. (2002) Science 298, 1039–1043

53. Varambally, S., Dhanasekaran, S. M., Zhou, M., Barrette, T. R., Kumar-Sinha,C., Sanda, M. G., Ghosh, D., Pienta, K. J., Sewalt, R. G., Otte, A. P., Rubin,M. A., and Chinnaiyan, A. M. (2002) Nature 419, 624–629

54. Monica, K., Galili, N., Nourse, J., Saltman, D., and Cleary, M. L. (1991) Mol.Cell. Biol. 11, 6149–6157

55. Milech, N., Kees, U. R., and Watt, P. M. (2001) Genes Chromosomes Cancer 32,275–280

56. Izraeli, S., Lowe, L. A., Bertness, V. L., Good, D. J., Dorward, D. W., Kirsch,I. R., and Kuehn, M. R. (1999) Nature 399, 691–694

57. Aplan, P. D., Jones, C. A., Chervinsky, D. S., Zhao, X., Ellsworth, M., Wu, C.,McGuire, E. A., and Gross, K. W. (1997) EMBO J. 16, 2408–2419

58. Matsuda, S., Rouault, J., Magaud, J., and Berthet, C. (2001) FEBS Lett. 497,67–72

59. Seo, M. S., Lee, M. S., and Lim, I. K. (1999) Gene (Amst.) 240, 165–17360. Cortes, U., Moyret-Lalle, C., Falette, N., Duriez, C., Ghissassi, F. E., Barnas,

C., Morel, A. P., Hainaut, P., Magaud, J. P., and Puisieux, A. (2000) Mol.Carcinog. 27, 57–64

61. Hurlin, P. J., Queva, C., Koskinen, P. J., Steingrimsson, E., Ayer, D. E.,Copeland, N. G., Jenkins, N. A., and Eisenman, R. N. (1995) EMBO J. 14,5646–5659

62. McArthur, G. A., Laherty, C. D., Queva, C., Hurlin, P. J., Loo, L., James, L.,Grandori, C., Gallant, P., Shiio, Y., Hokanson, W. C., Bush, A. C., Cheng,P. F., Lawrence, Q. A., Pulverer, B., Koskinen, P. J., Foley, K. P., Ayer,D. E., and Eisenman, R. N. (1998) Cold Spring Harbor Symp. Quant. Biol.63, 423–433

63. Matsui, T., Kinoshita, T., Hirano, T., Yokota, T., and Miyajima, A. (2002)J. Biol. Chem. 277, 36167–36173

64. Basch, R. S., and Goldstein, G. (1975) Cell. Immunol. 20, 218–22865. Schlesinger, D. H., Goldstein, G., Scheid, M. P., and Boyse, E. A. (1975) Cell 5,

367–37066. Delattre, O., Zucman, J., Plougastel, B., Desmaze, C., Melot, T., Peter, M.,

Kovar, H., Joubert, I., de Jong, P., Rouleau, G., Aurias, A., and Thomas, G.(1992) Nature 359, 162–165

67. May, W. A., Lessnick, S. L., Braun, B. S., Klemsz, M., Lewis, B. C., Lunsford,L. B., Hromas, R., and Denny, C. T. (1993) Mol. Cell. Biol. 13, 7393–7398

68. Hahm, K. B., Cho, K., Lee, C., Im, Y. H., Chang, J., Choi, S. G., Sorensen, P. H.,Thiele, C. J., and Kim, S. J. (1999) Nat. Genet. 23, 222–227

69. Lessnick, S. L., Dacwag, C. S., and Golub, T. R. (2002) Cancer Cell 1, 393–40170. Nishimori, H., Sasaki, Y., Yoshida, K., Irifune, H., Zembutsu, H., Tanaka, T.,

Aoyama, T., Hosaka, T., Kawaguchi, S., Wada, T., Hata, J., Toguchida, J.,Nakamura, Y., and Tokino, T. (2002) Oncogene 21, 8302–8309

Gene Expression Profiling of the pRB Pathway 46137

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from

Richard Vernell, Kristian Helin and Heiko Müller-pRB-E2F PathwayINK4AIdentification of Target Genes of the p16

doi: 10.1074/jbc.M304930200 originally published online August 15, 20032003, 278:46124-46137.J. Biol. Chem. 

  10.1074/jbc.M304930200Access the most updated version of this article at doi:

 Alerts:

  When a correction for this article is posted• 

When this article is cited• 

to choose from all of JBC's e-mail alertsClick here

Supplemental material:

  http://www.jbc.org/content/suppl/2003/09/10/M304930200.DC1

  http://www.jbc.org/content/278/46/46124.full.html#ref-list-1

This article cites 70 references, 21 of which can be accessed free at

by guest on Decem

ber 21, 2020http://w

ww

.jbc.org/D

ownloaded from