Chip-On-Chip Significance Analysis Reveals Large-Scale Binding and Regulation by Human Transcription Factor Oncogenes

ChIP-on-chip significance analysis reveals large-scale binding and regulation by human transcription factor oncogenes

Adam A. Margolina,b,c,1, Teresa Palomerod,e, Pavel Sumazinb, Andrea Califanoa,b,d,2,3, Adolfo A. Ferrandod,e,f,2,3, and Gustavo Stolovitzkyb,c,2,3

aDepartment of Biomedical Informatics, bJoint Centers for Systems Biology, dInstitute for Cancer Genetics, eDepartment of Pathology, and fDepartment of Pediatrics, Columbia University, New York, NY 10032; and cFunctional Genomics and Systems Biology Group, IBM T.J. Watson Research Center, Yorktown Heights, NY 10598

Edited by Barry H. Honig, Columbia University, New York, NY, and approved November 1, 2008 (received for review July 9, 2008) ChIP-on-chip has emerged as a powerful tool to dissect the complex tions. Although several studies have experimentally validated novel network of regulatory interactions between transcription factors target collections produced at a given statistical threshold (8–12), and their targets. However, most ChIP-on-chip analysis methods these studies likely miss a large number of true binding events, use conservative approaches aimed at minimizing false-positive obscuring the full complexity of transcriptional processes. transcription factor targets. We present a model with improved Using an empirically determined model of the distribution of sensitivity in detecting binding events from ChIP-on-chip data. Its intensity ratios for non-IP-enriched probes in ChIP2 experiments, application to human T cells, followed by extensive biochemical we developed an analytical method called ChIP2 Significance validation, reveals that 3 oncogenic transcription factors, NOTCH1, Analysis (CSA). When applied to ChIP2 data from the NOTCH1, MYC, and HES1, bind to several thousand target gene promoters, MYC, and HES1 protooncogenes in human T cell acute lympho- up to an order of magnitude increase over conventional analysis blastic leukemia (T-ALL) cells, CSA increased the number of methods. Gene expression proﬁling upon NOTCH1 inhibition detected binding sites by up to an order of magnitude compared shows broad-scale functional regulation across the entire range of with other routinely used methods. Both binding site analysis and predicted target genes, establishing a closer link between occu- biochemical validation demonstrate quantitative agreement with pancy and regulation. Finally, the increased sensitivity reveals a CSA-predicted false-positive rates. Analysis of gene expression combinatorial regulatory program in which MYC cobinds to virtu- signatures indicates functional regulation by NOTCH1 across the ally all NOTCH1-bound promoters. Overall, these results suggest an entire range of predicted targets. Finally, the increased sensitivity unappreciated complexity of transcriptional regulatory networks reveals that virtually all NOTCH1-bound promoters are also bound and highlight the fundamental importance of genome-scale anal- by MYC. Overall, these results highlight the power of the proposed ysis to represent transcriptional programs. analysis framework for the identification of transcriptional networks and provide an improved and fundamentally different pic- regulatory networks ͉ T cell lymphoblastic leukemia ͉ ture of the transcriptional programs controlled by NOTCH1, transcriptional regulation ͉ systems biology HES1, and MYC in T-ALL.

he dysregulated activity of oncogenic transcription factors Results T(TFs) contributes to neoplastic transformation by promoting Probe Statistics Are Accurately Modeled by CSA. T-ALL is a malig- aberrant expression of target genes involved in regulating cell nant tumor characterized by the aberrant activation of oncogenic homeostasis. Therefore, characterization of the regulatory net- TFs (13). We recently demonstrated that constitutive activation of works controlled by these TFs is a critical objective in understanding NOTCH1 signaling due to mutations in the NOTCH1 gene acti- the molecular mechanisms of cell transformation. ChIP-on-chip vates a transcriptional network that controls leukemic cell growth (ChIP2) (1) has emerged as a promising technology in the dissection (11, 14–16). These studies also demonstrated a fundamental role of transcriptional networks by providing high-resolution maps of for HES1 and MYC as transcriptional mediators of NOTCH1 genome-wide TF–chromatin interactions. signals (15, 17). To characterize the structure of the oncogenic ChIP2 uses microarray technology to measure the relative abun- transcriptional network driven by activated NOTCH1 in T cell dance of genomic fragments derived from an immunoprecipitate transformation, we sought to identify the direct transcriptional (IP) sample, which is enriched in fragments bound by an immuno- precipitated protein (usually a TF), and a whole-cell extract (WCE) sample, containing fragments derived from a total chromatin Author contributions: A.A.M., P.S., A.C., A.A.F., and G.S. designed research; A.A.M., T.P., and P.S. performed research; T.P. and A.A.F. contributed new reagents/analytic tools; preparation (input control) or an immunoprecipitation with a A.A.M. and P.S. analyzed data; and A.A.M., T.P., P.S., A.C., A.A.F., and G.S. wrote the paper. nonspecific control antibody (2). The 2 samples may either be The authors declare no conﬂict of interest. hybridized to different arrays or labeled with different dyes and This article is a PNAS Direct Submission. hybridized to the same array. Correct interpretation of ChIP2 data Freely available online through the PNAS open access option. depends critically on an accurate statistical model to compute the Data deposition: The microarray data have been deposited in the Gene Expression Omnibus probability that a given IP/WCE ratio is produced by a binding (GEO) Database, www.ncbi.nlm.nih.gov/geo (accession no. GSE12868). ChIP2 data is at event rather than experimental noise. http://wiki.c2b2.columbia.edu/califanolab/PNASAM2009/. 2 Recently, several elegant ChIP analysis methods have been 1Present address: The Broad Institute of MIT and Harvard, 7 Cambridge Center, Cambridge, proposed to tackle problems such as integrating measurements MA 02142. from adjacent probes (3–6) or inferring binding site locations at 2A.C., A.A.F., and G.S. contributed equally to this work. subprobe resolution (7). However, the lower-level problem of 3To whom correspondence may be addressed. E-mail: [email protected], califano@ developing an accurate error model to define meaningful statistical c2b2.columbia.edu, or [email protected]. thresholds has received comparably little attention [see SI and Fig. This article contains supporting information online at www.pnas.org/cgi/content/full/ 1]. Thus, ChIP2 data analysis methods often use highly conservative 0806445106/DCSupplemental. approaches aimed at minimizing the rate of false-positive predic- © 2008 by The National Academy of Sciences of the USA

244–249 ͉ PNAS ͉ January 6, 2009 ͉ vol. 106 ͉ no. 1 www.pnas.org͞cgi͞doi͞10.1073͞pnas.0806445106 Downloaded by guest on September 27, 2021 accurate description of the individual and combinatorial regulatory programs controlled by these TFs. We first generated an empirical model of the distribution of IP/WCE intensity ratios for probes associated with unbound fragments (see Materials and Methods), and we used it to assign a P value to each probe in the analysis of ChIP2 assays representing replicate experiments for NOTCH1, MYC, and HES1. ChIP2 assays for these TFs were performed in HPB-ALL cells, a well-characterized T-ALL cell line with high expression levels of activated NOTCH1, MYC, and HES1. For NOTCH1, ChIP2 assays were also performed in CUTLL1 cells, another NOTCH1-dependent T-ALL cell line. The magnitude versus amplitude plots (Fig. 2A) of the intensity- dependent distributions of probe-ratio values showed marked dif- ferences for the four experiments. In each case CSA accurately modeled the left tail of the probe ratio probability distribution, where the contribution from bound probes is expected to be Fig. 1. Modeling errors of methods that use whole-dataset statistics for minimal (Fig. 2 A and B). We note that if bound-probe ratios are either normalization or significance detection. Blue bars represent a histo- well separated from the experimental noise, the P value distribution 2 gram of log2 IP/WCE probe ratio values from a MYC ChIP experiment. The for all probes should be uniform between zero and one (unbound histogram displays distinct, overlapping distributions for bound and unbound probes) with a single peak near zero (bound probes). Importantly, probes. The dotted red curve shows the log2 ratio values after mean centering, CSA accurately captured these statistical properties (see SI). a common normalization technique that, for this experiment, adjusts the mean of the null distribution to be negative to compensate for the large 2 number of high-ratio values for the bound probes. The green curve represents Improved ChIP Sensitivity by CSA. CSA then incorporates the probe a Gaussian fitted to the overall distribution, demonstrating that analysis significance model with an analytical method that integrates the methods that fit a global error model to these data will significantly overes- statistics for replicate experiments and probes with nearby genomic timate the variance of the null distribution and will incur a high false-negative locations (to account for ChIP2 fragmentation lengths, see Materials rate, as shown by the black arrow, which represents 2 standard deviations and Methods). We used CSA to compute the false discovery rate from the mean of the green curve. (FDR) associated with the most significant 500-bp region on each of the 16,697 promoters represented on the array. Analysis of NOTCH1, MYC, and HES1 promoter occupancy in T-ALL targets of NOTCH1, HES1, and MYC. We hypothesized that the showed a larger than anticipated number of candidate target genes development of an accurate statistical model would result in for these TFs. Specifically, using CSA at a conservative FDR of improved sensitivity in the identification of TF targets and a more 0.05, the number of promoters on the array bound by the TFs in this GENETICS

Fig. 2. CSA determination of ChIP2 target genes. (A) Magnitude (M) versus amplitude (A) plots with confidence intervals inferred by CSA. The x axis represents the amplitude, calculated as the average log2 intensity of the IP and WCE channels. The y axis represents the magnitude, calculated as the log2 ratio of IP/WCE. The black line represents the intensity-dependent mean of the inferred null distribution, and the colored lines represent confidence intervals corresponding to P values of 0.1, 0.01, and 0.001 probability. Because we are only interested in positive-valued probes, confidence intervals are computed based on a 1-tail test, and statistically significant probes lie above the upper confidence interval lines. For ease of visualization we plot the lower confidence interval lines as 1 minus the corresponding P value. As shown, for all 3 TFs a large number of probes are significantly enriched in the IP channel, and MYC displays substantially more enrichment. (B) Graphic representation of the inferred distribution of P(M͉A ϭ 11). The blue curve represents the empirical conditional distribution of M computed at the particular value of A ϭ 11. The dotted black line represents the inferred mean of the null distribution, and the dotted red line represents the inferred null distribution. (C) Magnitude versus amplitude plots with colors representing Ϫlog10 P values of the CSA- inferred null distribution. As expected, the model reveals an intensity-dependent mean and variance of the null distribution, with increased variance at low- intensity levels, as well as sometimes for extremely high-intensity levels due to saturation effects.

Margolin et al. PNAS ͉ January 6, 2009 ͉ vol. 106 ͉ no. 1 ͉ 245 Downloaded by guest on September 27, 2021 Table 1. Number of predicted target genes for various methods TF Cell line SAEM ChIPOTle Chipper CSA M Ͼ 1

MYC HPB-ALL 127 5,130 5,684 8,016 8,534 HES1 HPB-ALL 187 2,899 1,228 3,074 1,470 NOTCH1 CUTLL1 647 3,014 1,386 3,154 907 NOTCH1 HPB-ALL 410 2,641 1,361 2,471 841

For each of the 4 ChIP2 experiments, we compared the number of targets inferred by CSA and by 3 other algorithms. The SAEM column refers to genes that were output as inferred targets using the standard Agilent software. Inferred targets for ChIPOTle, Chipper, and CSA are reported using an FDR cutoff of 5%. The last column lists the number of probes that were at least 2-fold enriched in the IP channel for both replicates, and provides a simple heuristic of the overall enrichment in the experiment.

study are: MYC (8,016; 48.0%), NOTCH1 in CUTLL1 (3,154; 18.9%), HES1 (3,074; 18.4%), and NOTCH1 in HPB-ALL (2,471; 14.8%) (Table 1). Although the numbers reported above are far larger than the number of predicted targets commonly reported in ChIP2 analysis studies, we also compared against predictions from several pub- Fig. 3. The percentages of identiﬁed sequences containing a binding site for lished analysis methods with available software. One class of MYC are plotted as a function of the total number of rank ordered sequences methods relies heavily on analyzing the shape of ratio values from using a threshold that yields a 30% false-positive rate. (Inset) For bins of 100 multiple probes with nearby genomic proximity (3–7). These meth- genes ranked by CSA, we computed MYC binding site enrichment P values ods are generally not applicable to the relatively sparse arrays used relative to a background of unbound promoter fragments (solid blue curve). in this study and produced very few predicted targets. As an initial The x axis represents the center of each bin. For each bin we also approximated Ϫ Ϫ benchmark, we compared against the Single Array Error Model the expected percent of bound genes as (FDRr*r FDRl*l)/(r l), where FDRr (SAEM) (1, 18), the standard method packaged with the Agilent and FDRl represent the CSA-inferred FDRs for the genes at the right and left analysis software. This method models the intensity-dependent edges of the bin, respectively, and r and l represent their ranks. The dotted red curve displays this quantity, which is in excellent agreement with the se- variance of probe ratio values, but then computes significance based quence-based enrichment P values. on whole-dataset statistics. CSA predicted approximately an order of magnitude more bound promoters than SAEM (Table 1). Two published methods, ChIPOTle (19) and Chipper (8), compute methods. The highest-scoring Ϸ2,000 sites identified by Chipper significance using only probes with low ratio values, and these performed better than those identified by CSA; however, CSA- methods indeed predicted more targets than SAEM. However, identified fragments beyond this ranking were significantly more both methods use normalization techniques that, to varying extents, enriched in MYC binding sites. Comparing nonoverlapping rely on whole-dataset statistics, and as a result CSA predicted more fragments in the top 5,000 promoters inferred by CSA versus targets than both. This was most apparent for MYC, which con- each of the other 3 methods demonstrated a statistically signif- tained the largest number of high ratio probes. For the other TFs, icant enrichment of MYC binding sites in CSA-inferred frag- CSA predicted roughly the same number of targets as ChIPOTle ments (P Ͻ 10Ϫ10, based on the hypergeometric distribution, for and twice as many targets as Chipper (Table 1). each comparison). To compare the predicted false-positive rate by CSA with the Accuracy of CSA Predictions Is Supported by Binding Site Enrichment significance of MYC binding site enrichment, we binned the Analysis. As a first test of the broad TF binding predictions fragments based on their CSA rankings (100/bin) and assessed generated by CSA, we evaluated the enrichment of MYC binding whether the MYC/M00322 motif could be successfully used to sites, using the TRANSFAC (20) position-specific scoring matrix distinguish the fragments in each bin from those in the negative set, M00322, in the promoters of target genes identified by CSA and S(Ϫ). The classification P value based on binding site enrichment was other analysis methods. The DNA-binding component of in excellent agreement with the CSA-inferred false-positive rate of NOTCH1 transcriptional complexes, CSL, is not represented in each bin, suggesting significant enrichment of MYC sites in the TRANSFAC or JASPAR (21), and the only HES1-associated Ϸ matrix was found to be of low quality and a poor predictor of HES1 promoters of 7,000 genes, corresponding to the range of high- binding, independent of the algorithm used. For each analysis confidence targets predicted by CSA (Fig. 3 Inset). Beyond this method, promoters were ranked by their P values computed from threshold, both quantities degraded very rapidly, and for ranks Ϸ the MYC ChIP2 experiment, and MYC/M00322 matching sites greater than 8,000 the CSA-inferred false-positive rate reached were identified in the 600-bp fixed-length window centered on the 100% and, correspondingly, fragments showed no ability to be most significant probe in the highest-scoring promoter region. The classified by MYC binding sites. match threshold was set so that a negative set, S(Ϫ),of3,000 Overall, these results suggest that CSA has increased sensitivity fragments showing the least amount of MYC binding would in identifying a larger number of binding events and that meaningful produce a false-positive rate of 30%. Details on the procedure are statistical cutoffs can be determined from data. given in the SI. Analysis of the cumulative proportion MYC/M00322- Experimental Validation of CSA TF Binding Predictions. To further test matching fragments as a function of their ChIP2 ranking by the the accuracy of CSA-based TF target predictions, we performed corresponding method showed that fragments inferred by all independent chromatin immunoprecipitation (ChIP) experiments methods were enriched in MYC/M00322 sites and that site for each of the 4 ChIP2 conditions and tested the IP enrichment of enrichment was correlated with the ChIP2 ranking (Fig. 3). specific promoters by quantitative PCR (qPCR). We first analyzed However, fragments identified by CSA were more likely to 8 predicted NOTCH1 targets in HPB-ALL cells, randomly sampled contain MYC binding sites than those identified by the other at an FDR Յ20%. Seven of these 8 predicted fragments were

246 ͉ www.pnas.org͞cgi͞doi͞10.1073͞pnas.0806445106 Margolin et al. Downloaded by guest on September 27, 2021 validated as bound by NOTCH1, and only the least significant false-positives beyond the top 2,000 targets and, correspondingly, fragment failed validation (Table 2). their likelihoods to be expressed and regulated by NOTCH1 We tested an additional 12 targets for HES1 and MYC in decreased. However, even for genes with ChIP2 ranks between HPB-ALL and for NOTCH1 in CUTLL1, sampling predicted 4,000 and 5,000, there was significant enrichment for both the targets uniformly at an FDR of 20% (i.e., 20% expected false- percent of expressed genes (59.4%; P Ͻ 10Ϫ33) and the expression positives) (Table 2). In this analysis, 26 of 36 targets (72.2%) were change upon NOTCH1 inhibition (P Ͻ 10Ϫ15). These results positive by ChIP/qPCR, and 9 (25%) were negative. The remaining demonstrate that, in contrast with previous analysis based on a gene (the second least significant for MYC) could not be amplified limited number of targets (17), NOTCH1 directly contributes to the by qPCR. Nonvalidated/false-positive targets were, in general, at transcriptional activity of thousands of genes. the end of the ranked lists (Table 2). The only outlier was the first-ranked fragment for HES1 (KIAA1407 gene promoter). To Interaction of NOTCH1 and MYC Regulatory Networks. NOTCH1 and obtain experimental evidence on the robustness of our validation MYC operate as highly interrelated regulators of cell growth, assay, we randomly selected 10 genomic regions not identified as proliferation, and survival during T cell development and transfor- bound by MYC and 10 not identified as bound by HES1. Nine mation. In a recent study (11), we compared the regulatory selected regions were within promoters, and 11 were in intergenic networks controlled by NOTCH1 and MYC by using the AR- regions. As expected, none of these 20 regions showed evidence of ACNE reverse engineering algorithm (23, 24) to predict 58 and 61 binding by MYC or HES1 when tested by ChIP/qPCR. targets of NOTCH1 and MYC, respectively, and observed a For all experiments, numerous validated genes had CSA ranks in significant overlap of 12 genes between the 2 lists (P Ͻ 10Ϫ51). We the thousands. The lowest-ranking validated genes before encoun- went on to characterize a feed-forward loop in which NOTCH1 tering a false-positive were as follows: 2,223 for NOTCH1 in directly regulates MYC, and both TFs regulate a common set of CUTLL1; 2,958 for NOTCH1 in HPB-ALL; 4,901 for MYC; and targets promoting leukemic cell growth. Based on these findings, we although the top-ranking gene for HES1 failed validation, the sought to further investigate the relationship between the genes following 7, down to rank 3,247, were positive. Notably, many of the bound by MYC and by NOTCH1 using the much larger list of validated targets showed subtle ChIP2 signals. For example, 2 targets inferred by CSA. Strikingly, the analysis predicted that MYC C6orf82, a validated HES1 target, had ChIP binding ratios in bound to 1,668 of the 1,804 (92.5%; P Ͻ 10Ϫ11) genes that were replicate experiments of 1.37 and 1.68 for the most significant probe bound by NOTCH1, using a ChIP2 FDR threshold of 0.01. In in its promoter, and there was no enrichment (ratios of 0.81 and agreement with the fundamental role of NOTCH1 in controlling 1.15) for its adjacent probe. However, upon ChIP/qPCR validation, leukemia cell growth (11), the NOTCH1-bound genes were highly this region showed binding ratios of 2.69 and 4.65. ChIP/qPCR enriched in gene ontology (GO) (25) categories related to cellular SI Ϫ results are available in the . growth and metabolism, such as cellular metabolism (P Ͻ 10 41), Overall, 33 of the 44 genes (75%) selected from those with an Ϫ Ϫ RNA metabolism (P Ͻ 10 24), and protein biosynthesis (P Ͻ 10 9). FDR of 20% by CSA were validated by ChIP/qPCR. These The complete output of the GO enrichment analysis is given in biochemical validation results support our computationally derived the SI. conclusions regarding the broad range of binding for all tested TFs and demonstrate the power of CSA for reducing the false-negative Discussion GENETICS rate in ChIP2 experiments. We have shown that the choice of a realistic statistical model can dramatically affect the result of ChIP2 data analysis and its biolog- NOTCH1 Regulates Direct Target Genes Predicted by CSA. To test whether CSA-predicted NOTCH1-bound genes are also function- ical interpretation and proposed the CSA algorithm to assign ally regulated by this TF, we treated a panel of 10 T-ALL cell lines meaningful statistical significance scores used to predict a more with Compound E, a ␥-secretase inhibitor that blocks an essential complete range of TF–target interactions. The method of assessing proteolytic cleavage step required for release of the intracellular probe statistical significance relies on minimal assumptions: that the domains of NOTCH1 from the membrane and their translocation null distribution is symmetric and that bound fragments do not significantly affect the left tail of the null hypothesis statistics. As a to the nucleus (22). Genome-wide expression profiles of cells 2 treated for 72 h with Compound E (100 nM) or vehicle only result, it should generalize well to ChIP experiments performed (DMSO) were measured using microarrays, and expression changes using other platforms and cellular conditions. We used an inde- were compared with NOTCH1 promoter occupancy identified by pendence model for replicate experiments and adjacent probes in CSA analysis of the ChIP2 data. Overall, 11,606 genes were the null hypothesis. Although this assumption is valid for relatively represented on both the ChIP2 and the expression arrays. For each sparse arrays, denser arrays may introduce correlation for unbound gene we computed: (i) the ChIP2 FDR based on the highest-scoring nearby probes that are within the DNA fragmentation length, leading to unrealistically low FDR values if independence is as- 500-bp region in its promoter; (ii)thelog2 expression ratio of the control versus treatment, averaged over the 10 cell lines and sumed. We therefore recommend caution that the independence duplicate experiments; and (iii) the number of microarray experi- assumption applies when analyzing denser arrays, in which case the ments in which the gene was expressed (not called absent by CSA method may be further improved by incorporating existing, MAS5), considering both Compound E-treated and DMSO- more sophisticated models for the integration of data from nearby treated samples (because the group of expressed genes is essentially probes (3–7). However, for the arrays used in this study, which contain the same for both treatments, considering expressed genes using an average probe spacing of more than 200 nucleotides, we show that only one subgroup does not substantially change the results). our results are in quantitatively good agreement with biochemical Predicted NOTCH1-bound genes were more likely to be ex- validation assays and that no correction seems to be required. pressed than genes not identified as bound by NOTCH1 and The analysis of ChIP2 data from 3 oncogenic TFs reveals that showed clear down-regulation upon NOTCH1 inhibition (Fig. 4). CSA identifies far more bound gene promoters than standard The 2,000 most confident NOTCH1 targets (FDR Ͻ 0.058) were analyses. Specifically, CSA predicts that each studied TF binds to expressed in 83.3% of experiments, whereas the 6,000 least confi- several thousand target genes, with MYC binding to roughly half of dent NOTCH1 targets were expressed in 38.8% of experiments the assayed promoters, providing additional insight into the ex- (P Ͻ 10Ϫ100). The top 2,000 targets also showed coordinated treme pluripotency of this protooncogene (26). These predictions down-regulation upon NOTCH1 inhibition that was subtle in might still be an underestimate, because only the proximal pro- magnitude (mean ϭ 12.2%) but extremely significant moter regions (Ϫ0.8kbtoϩ0.2 kb, relative to transcription start (P Ͻ 10Ϫ100). The ChIP2 analysis predicted a rapid increase in site) are represented on the arrays used in this study.

Margolin et al. PNAS ͉ January 6, 2009 ͉ vol. 106 ͉ no. 1 ͉ 247 Downloaded by guest on September 27, 2021 Table 2. Validation of predicted targets at 20% FDR Gene CSA FDR Validated Rank

NOTCH1 in HPB-ALL FLJ13798 2.29EϪ12 Yes 35 RAB18 2.03EϪ05 Yes 674 PORIMIN 0.0015 Yes 1301 ZMAT2 0.0035 Yes 1497 PSENEN 0.0068 Yes 1675 LMAN2 0.0379 Yes 2316 XKR9 0.1054 Yes 2958 THPO 0.125 No 3119 MYC PRKACB 1.96EϪ13 Yes 337 LOH12CR1 8.44EϪ11 Yes 1131 HS3ST3B1 3.62EϪ09 Yes 1885 POLR2I 6.11EϪ08 Yes 2639 TXLNB 6.07EϪ07 Yes 3393 PARP1 4.80EϪ06 Yes 4147 KIAA1984 2.93EϪ05 Yes 4901 ZNF233 1.91EϪ04 No 5655 KIF5B 0.0014 Yes 6409 Fig. 4. Regulation of NOTCH1 target genes as a function of ChIP2 rank. Genes HIST1H2AK 0.0083 Yes 7163 are ranked according to their ChIP2 FDR, plotted in green, as inferred by CSA CPE 0.0417 N/A 7917 (FDRs are displayed with a maximum value of 1). The blue curve displays the PEX16 0.126 No 8671 median log2 expression ratio of vehicle control compared with Compound E HES1 treatment across bins of 250 genes, with the 95% confidence interval plotted KIAA1407 2.96EϪ07 No 216 in red. Positive values indicate down-regulation upon NOTCH1 inhibition. The MGC3121 0.0001 Yes 649 x axis represents the center of each bin. The heat map above the plot displays PRKDC 0.0015 Yes 1082 the average percent of experiments in which the genes in the corresponding GTF3C2 0.0057 Yes 1515 bin are expressed. Expression change upon NOTCH1 inhibition and the per- 2 FAM20B 0.0126 Yes 1948 cent of expressed genes are both highly correlated with the ChIP ranking, and CHRM5 0.0239 Yes 2381 they remain significantly enriched for more than 5,000 predicted targets. BTBD9 0.0393 Yes 2814 C6orf82 0.0577 Yes 3247 WDSUB1 0.0815 No 3680 albeit weakly, regulated upon NOTCH1 inhibition. These results DACH2 0.1074 No 4113 are highly consistent with a previous study performed in yeast (27) NARG1L 0.1405 Yes 4546 that also observed correspondence of ChIP2 results with both CHORDC1 0.1775 No 4979 binding site enrichment and expression changes for a large number NOTCH1 in CUTLL1 of genes. RNF139 5.41EϪ10 Yes 171 GO enrichment analysis shows that NOTCH1 subtly regulates a ELP3 2.76EϪ07 Yes 513 large number of genes involved in the cellular growth machinery. Ϫ BAT2 5.22E 06 Yes 855 These results add an additional layer of regulation to the effects of DDX5 7.45EϪ05 Yes 1197 ZNF436 4.05EϪ04 Yes 1539 NOTCH1 signaling in promoting cell growth, with important ETFDH 1.44EϪ03 Yes 1881 implications for understanding the role of NOTCH1 signaling in DCP1A 4.41EϪ03 Yes 2223 development and transformation. Thus, in addition to the estab- MRPL48 0.0109 No 2565 lished role of NOTCH1 in promoting growth through its interaction MYB 0.0303 No 2907 with MYC (17) and the PI3K-AKT (15) signaling pathway, KCTD16 0.0599 Yes 3249 NOTCH1 also has a direct effect in promoting cell growth. This ADAR 0.1022 Yes 3591 irreversibly couples the developmental programs involved in stem BXDC5 0.1549 No 3933 cell homeostasis and lineage commitment activated upon NOTCH1 We used ChIP/qPCR to test 12 genes for NOTCH1 in CUTLL1 and for HES1 and activation with the metabolic pathways needed for the expansion of MYC in HPB-ALL (1 gene for MYC could not be amplified), and 8 genes for stem cells and T cell progenitors. NOTCH1 in HPB-ALL. The table columns are as follows: Gene, the gene name; Finally, the availability of a more complete repertoire of bound CSA FDR, the FDR computed by CSA—that is, the percentage of genes with promoters allows us to truly assess the extent of a TF’s regulatory ranks lower than the current gene that are expected not to be bound by the TF; Validated, whether the gene was positively validated by ChIP/qPCR; Rank, program and the combinatorial overlap between independent the rank of the gene when all genes are sorted by their CSA-inferred statistical programs. Our analysis shows that 92.5% of the promoters bound significance. by NOTCH1 are also bound by MYC. Indeed, it appears that NOTCH1 coregulates a specific subset of the MYC regulatory program. Although this was previously hinted at by the similarity of CSA predictions were validated by 3 independent tests. ChIP/ the regulatory programs inferred for the 2 TFs by expression qPCR experiments are in excellent correspondence with CSA- analysis (17), the true extent of this overlap can only be grasped inferred FDRs, especially considering that ChIP/qPCR itself has a after resolving a more complete map of NOTCH1 and MYC 10–20% false-negative rate (9–12). Computational validation by targets. While contributing to our understanding of transcriptional sequence analysis further indicates that CSA-inferred FDRs are in regulation at the genome-scale, our findings suggest an even greater agreement with MYC binding site enrichments. Finally, gene than expected complexity of transcriptional networks. expression analysis after NOTCH1 inhibition both provides further support for the CSA predictions and creates a stronger than Materials and Methods expected association between bound and regulated genes. We CSA Algorithm. The CSA algorithm takes as input probe intensity measurements found that NOTCH1 binds to a large number of promoters after background subtraction and correction for factors such as spatial position on (Ͼ2,000) and that the set of corresponding genes is consistently, the array and print tip variability. In this study we used the standard Agilent proce-

248 ͉ www.pnas.org͞cgi͞doi͞10.1073͞pnas.0806445106 Margolin et al. Downloaded by guest on September 27, 2021 ⌫M Ϫ͚M j ⌫M dure to obtain background-corrected probe intensity values, and we determined the be evaluated as CDF( jϭ1log(pi)), where CDF is the gamma cumulative distri- statistical signiﬁcance of binding regions by using the procedure described below. bution function with mean 1 and M degrees of freedom.

Single-Probe Significance Analysis. For each probe, the statistical significance of Combining Regions. Because of sonication, the signal derived from a binding a measurement is inferred by computing the conditional probability of the event may be detected by multiple probes in close genomic proximity to the binding site. To compute a combined statistic representing the probability of a magnitude (M) given the amplitude (A), P (M͉A), where M ϭ log2(IP/WCE) and null binding event within the region spanned by multiple probes, we adapt a com- A ϭ [log2(IP) ϩ log2(WCE)]/2, under the null hypothesis (i.e., no enrichment in the monly used strategy (19) of using a fixed-size sliding window and integrating the IP compared with the WCE channel). Here, IP and WCE represent, respectively, the values of probes falling within this window. Based on published measurements probe intensity measurements for the IP and WCE channels. The dependency of of fragmentation lengths (7), in this work we used a 500-bp window and a step M on A is illustrated in Fig. 2A. size of 150 bp. Assuming that measurements from adjacent probes are indepen- The method begins by estimating the joint probability distribution, P(M,A), dent in the null hypothesis, Fisher’s method can again be applied to integrate the using a bivariate Gaussian kernel density estimator (28). The kernel width of the values from nearby probes. That is, let Wrepresent the set of probes falling within estimator is calculated using the AMISE criterion (29). Conditioning on A yields a given 500-bp window. The integrated probability for this region is then calcu- the conditional distribution P(M͉A) ϭ P(M,A)/P(A), where P(A) is calculated using lated as a univariate Gaussian kernel. For a particular average intensity value, A0, the conditional mean of the null distribution is inferred as ϭ ⌫M*͉W͉ͩ Ϫ͸ ͸ M ͑ j͒ͪ pregion CDF i␧W jϭ1 log pi ␮ ͉ ϭ ͑ ͉ ϭ ͒ ˆ M A0 argmax P M A A0 . M To compute the probability that any region within a gene’s promoter is bound, The conditional null distribution given A ϭ A0 is inferred by projecting P(M͉A ϭ we consider the most significant window, controlling for multiple tests using ␮ Ͻ ␮ ͉ A0) across ˆ M͉A0 for M ˆ M͉A0. This procedure is used to calculate Pnull(M A) for an Bonferroni correction based on the number of probes in the promoter. This evenly spaced grid of A and M values, excluding the 1% of probes with the lowest correction is not exact, because the number of tests (i.e., the number of windows A values (which are assigned a P value of 1). In this work we used step sizes of 0.05 containing unique subsets of probes) is likely greater than the number of probes and 0.01 for the A and M values, respectively. The complete conditional null in a promoter, causing underestimation of the significance, and the tests are not independent (i.e., the same probe may fall within multiple windows), causing distribution, Pnull(M͉A), is computed using 2-dimensional linear interpolation. For each probe, statistical significance is assessed using a 1-tailed test with reference overestimation of the significance. However, because the number of probes in to this distribution. Because the distribution is empirical, there is a limit to the each promoter (and therefore the number of probes within each window) is inferable minimum P value, which depends on the number of arrayed probes. For relatively small, especially for the arrays used in this study, we expect this simpli- fication to have little impact on the calculated statistics. For very dense arrays, a the arrays used in this study, we set the minimum P value to 10Ϫ5, which is roughly more sophisticated multiple-test correction procedure, such as those described in 1 divided by the number of probes on the array. We stress the importance of using (31), may yield more accurate results. an empirical distribution because we have observed that the empirical data generally display significantly non-Gaussian tails. FDR Calculation. After computing a corrected P value for each gene representing the probability that the most significant region on the gene’s promoter is bound, Combining Replicates. We use Fisher’s method (30) to compute an aggregate P we control for multiple tests across genes and compute a false discovery rate using value for each probe based on measurements from replicate experiments, under j the Benjamini–Hochberg procedure (32). Let pk represent the corrected P value the null hypothesis that the probe is unbound in all replicates. Let pi denote the 2 computed for gene k, let rk represent the rank of gene k sorted by the ChIP P th th P value computed for the i probe in the j replicate experiment. Assuming that values, and let G represent the total number of genes on the array; then, the false replicates are independent in the null hypothesis, a test statistic for evaluating the GENETICS discovery rate for gene k is computed as FDRk ϭ G*pk/rk . probability of a joint observation of P values across experiments is the product of ៮ ϭ⌸M j ៮ the individual P values, si jϭ1pi, where si is the test statistic and M is the number ACKNOWLEDGMENTS. A.A.M. was supported by an IBM PhD fellowship. This of replicate experiments. If modeled correctly, P values under the null hypothesis work was supported by National Cancer Institute Grants R01CA109755 (to A.C.) should be uniformly distributed (See SI). It is useful to log-transform this equation and R01CA120196 (to A.A.F.), National Institute of Allergy and Infectious Diseases Grant R01AI066116, the Alex Lemonade Stand Foundation (T.P.), The Cancer such that we evaluate Ϫlog(s៮ ) ϭϪ͚M log(pj). Because the logarithm of a uni- i jϭ1 i Research Institute, the WOLF Foundation, the National Centers for Biomedical form distribution is exponentially distributed with mean 1, this equation is a sum Computing National Institutes of Health Roadmap Initiative (U54CA121852), and of exponentially distributed random variables, which is a gamma-distributed the Leukemia and Lymphoma Society (Grants 1287-08 and 6237-08). A.A.F. is a random variable with mean 1 and M degrees of freedom. Thus, significance can Leukemia and Lymphoma Scholar.

1. Ren B, et al. (2000) Genome-wide location and function of DNA binding proteins. 16. Weng AP, et al. (2004) Activating mutations of NOTCH1 in human T cell acute Science 290:2306–2309. lymphoblastic leukemia. Science 306:269–271. 2. Buck MJ, Lieb JD (2004) ChIP-chip: Considerations for the design, analysis, and appli- 17. Palomero T, et al. (2006) NOTCH1 directly regulates c-MYC and activates a feed- cation of genome-wide chromatin immunoprecipitation experiments. Genomics forward-loop transcriptional network promoting leukemic cell growth. Proc Natl Acad 83:349–360. Sci USA 103:18261–18266. 3. Glynn EF, et al. (2004) Genome-wide mapping of the cohesin complex in the yeast 18. Hughes TR, et al. (2000) Functional discovery via a compendium of expression profiles. Saccharomyces cerevisiae. PLoS Biol 2:E259. Cell 102:109–126. 4. Kim TH, et al. (2005) A high-resolution map of active promoters in the human genome. 19. Buck MJ, Nobel AB, Lieb JD (2005) ChIPOTle: a user-friendly tool for the analysis of Nature 436:876–880. ChIP-chip data. Genome Biol 6:R97. 5. Li W, Meyer CA, Liu XS (2005) A hidden Markov model for analyzing ChIP-chip experiments on genome tiling arrays and its application to p53 binding sequences. 20. Matys V, et al. (2003) TRANSFAC: Transcriptional regulation, from patterns to profiles. Bioinformatics 21(Suppl 1):i274–i282. Nucleic Acids Res 31:374–378. 6. Zheng M, Barrera LO, Ren B, Wu YN (2007) ChIP-chip: Data, model, and analysis. 21. Sandelin A, et al. (2004) JASPAR: An open-access database for eukaryotic transcription Biometrics 63:787–796. factor binding profiles. Nucleic Acids Res 32:D91–D94. 7. Qi Y, et al. (2006) High-resolution computational models of genome binding events. 22. Miele L (2006) Notch signaling. Clin Cancer Res 12:1074–1079. Nat Biotechnol 24:963–970. 23. Basso K, et al. (2005) Reverse engineering of regulatory networks in human B cells. Nat 8. Gibbons FD, Proft M, Struhl K, Roth FP (2005) Chipper: Discovering transcription-factor Genet 37:382–390. targets from chromatin immunoprecipitation microarrays using variance stabilization. 24. Margolin AA, et al. (2006) ARACNE: An algorithm for the reconstruction of gene regula- Genome Biol 6:R96. tory networks in a mammalian cellular context BMC Bioinformatics 7(Suppl 1):S7. 9. Lee TI, et al. (2002) Transcriptional regulatory networks in Saccharomyces cerevisiae. 25. Ashburner M, et al. (2000) Gene ontology: Tool for the unification of biology. The Gene Science 298:799–804. Ontology Consortium. Nat Genet 25:25–29. 10. Polo JM, et al. (2007) Transcriptional signature with differential expression of BCL6 26. Pelengaris S, Khan M (2003) The many faces of c-MYC. Arch Biochem Biophys 416:129–136. target genes accurately identifies BCL6-dependent diffuse large B cell lymphomas. 27. Tanay A (2006) Extensive low-affinity transcriptional interactions in the yeast genome. Proc Natl Acad Sci USA 104:3207–3212. Genome Res 16:962–972. 11. Palomero T, et al. (2006) Transcriptional regulatory networks downstream of TAL1/SCL 28. Beirlant J, Dudewicz E, Gyorfi L, van der Meulen E (1997) Nonparametric entropy in T-cell acute lymphoblastic leukemia. Blood 108:986–992. estimation: An overview. Int J Math Stat Sci 6:17–39. 12. Odom DT, et al. (2004) Control of pancreas and liver gene expression by HNF transcription factors. Science 303:1378–1381. 29. Sheather SJ, Jones MC (1991) A reliable data-based bandwidth selection method for 13. Ferrando AA, et al. (2002) Gene expression signatures define novel oncogenic path- kernel density-estimation. J R Stat Soc Ser B 53:683–690. ways in T cell acute lymphoblastic leukemia. Cancer Cell 1:75–87. 30. Mosteller F, Fisher RA (1948) Questions and answers. Am Stat 2:30–31. 14. Palomero T, et al. (2006) Activating mutations in NOTCH1 in acute myeloid leukemia 31. Keles S, van der Laan MJ, Dudoit S, Cawley SE (2006) Multiple testing methods for and lineage switch leukemias. Leukemia 20:1963–1966. ChIP-Chip high density oligonucleotide array data. J Comput Biol 13:579–613. 15. Palomero T, et al. (2007) Mutational loss of PTEN induces resistance to NOTCH1 32. Benjamini Y, Hochberg Y (1995) Controlling the false discovery rate - a practical and inhibition in T-cell leukemia. Nat Med 13:1203–1210. powerful approach to multiple testing. J R Stat Soc Ser B 57:289–300.

Margolin et al. PNAS ͉ January 6, 2009 ͉ vol. 106 ͉ no. 1 ͉ 249 Downloaded by guest on September 27, 2021