de Santiago et al.

BaalChIP: Bayesian analysis of allele-specific

transcription factor binding in cancer genomesInes de Santiago†, Wei Liu†, Martin O’Reilly, Ke Yuan, Chandra Sekhar Reddy Chilamakuri, Bruce A.J.

Ponder, Kerstin B. Meyerˆ and Florian Markowetz*ˆ

*Correspondence: flo-

rian.markowetz@cruk.cam.ac.uk

University of Cambridge, Cancer

Research UK Cambridge

Institute, Robinson Way,

CB2 0RE Cambridge, UK

Full list of author information is

available at the end of the article

†Equal contributorsˆJoint

corresponding authors

Abstract

Allele-specific measurements of transcription factor binding from ChIP-seq data are

key to dissecting the allelic effects of non-coding variants and their contribution to

phenotypic diversity. However, most methods to detect allelic imbalance assume

diploid genomes. This assumption severely limits their applicability to cancer samples

with frequent DNA copy number changes. Here we present a Bayesian statistical

approach called BaalChIP to correct for the effect of background allele frequency on

the observed ChIP-seq read counts. BaalChIP allows the joint analysis of multiple

ChIP-seq samples across a single variant and outperforms competing approaches in

simulations. Using 548 ENCODE ChIP-seq and 6 targeted FAIRE-seq samples we

show that BaalChIP effectively corrects allele-specific analysis for copy number

variation and increases the power to detect putative cis-acting regulatory variants in

cancer genomes.

Keywords: Allele-specific binding; ChIP-sequencing; FAIRE-sequencing; Bayesian

statistics; Cancer; Copy-number change; Allele frequency

Background1

Allele-specific measurements of transcription factor binding from ChIP-seq data2

have provided important insights into the allelic effects of non-coding variants and3

their contribution to phenotypic diversity [1, 2, 3, 4, 5]. Since the majority of disease4

risk-associated SNPs occur in non-coding DNA, many of them might disrupt cis-5

regulatory elements. Thus, a direct method to identify functional SNPs is to use the6

information obtained from allelic-specific binding (ASB) of transcription factors.7

de Santiago et al. Page 2 of 57

ChIP-seq data are commonly used to study allele-specific binding of proteins at8

heterozygous non-coding SNPs and to infer putative effects of such variants on gene9

regulation. Allele-specific mapping corrects for environmental sources of variation10

that alter gene regulation, since both alleles are assayed in the same cellular context.11

Existing approaches to infer allelic imbalance Previous studies have used ChIP-seq12

and RNA-seq to identify ASB and allele-specific expression (ASE). These studies13

have described methods to address technical and methodological biases such as the14

sequence context of a SNP [6], alignment biases to the reference allele [7, 8], and15

the issue of increasing detection power by combining multiple SNPs in the same16

gene [9] or across multiple ChIP-seq samples [10]. However, all of these approaches17

are designed to examine the allelic imbalances in diploid samples and do not address18

copy number differences between the two alleles, an ubiquitous feature of cancer19

genomes (Figure S1).20

Few papers in the literature have tried to address confounding effects arising21

from copy number changes. Some studies have analyzed tumor and normal sam-22

ples without making any adjustments to the applied methodology, which risks23

identifying false positives where the detected ASE is mainly due to copy-number24

alterations [11]. Others have removed all sites overlapping copy number vari-25

ants [7, 12, 13], or used the genomic DNA allelic ratios to correct for the observed26

allelic imbalances [14, 15] or to remove SNPs with allelic imbalances detected in27

the control input DNA [2] which is only feasible when the coverage at any assayed28

heterozygous site is high (> 20x if a binomial test is to be applied to reach adequate29

statistical power [9, 12]). Recently, Liu et al. [16] developed cisASE, a likelihood-30

based method for detecting allele-specific expression. cisASE has been shown to31

successfully correct for copy number alterations that hinder the identification of32

ASE in RNA-seq data. Table S1 presents a summary of the different strategies33

of allelic-specific mapping analyses to deal with regions of altered copy-number.34

de Santiago et al. Page 3 of 57

While the analysis of allele-specific expression and binding are conceptually similar,35

in practice methods tailored to either ASE or ASB are not easily exchangeable,36

because the data types and thus the biological and statistical assumptions are dif-37

ferent.38

Bayesian analysis of allele specific binding To address the wide range of biases39

that can affect the detection of ASB from ChIP-seq data we developed BaalChIP40

(Bayesian Analysis of Allelic imbalance from ChIP-seq data) (Figure 1). BaalChIP41

makes it possible to quantify the significance of the allelic imbalance from ChIP-42

seq data in cell lines with genomic aberrations, which is a major advance over43

previous approaches. BaalChIP combines several important features: First, we use44

several strategies to rigorously account for allelic mapping bias, including filtering of45

SNPs in problematic alignment regions as well as simulations to identify reads with46

increased risk of mapping bias. Second, we implement a beta-binomial model for47

allelic read counts. This model accounts for the fact that the observed variance in the48

data is larger than expected from a standard binomial model, a phenomenon known49

as overdispersion [9, 17, 18]. Third, we take advantage of the fact that multiple50

transcription factor ChIP-seq data may be available for the same SNP to improve51

ASB detection. Finally, we use a Bayesian framework to account for the influence of52

the background allele composition and the reference mapping bias on the observed53

ChIP-seq allelic read count.54

We applied BaalChIP to 548 ENCODE samples [19] obtained from a panel of 1455

cell lines, representing different tissues and karyotypes, including 8 cancer cell lines56

(HeLa-S3, A549, MCF-7, T-47D, K562, HepG2, SK-N-SH, HL-60) and 6 non-cancer57

cell lines (H1-hESC, GM12878, GM12891, GM12892, MCF10A-Er-Src, IMR90), as58

well as 6 FAIRE-seq samples obtained from breast cancer cell lines (MDA-MB-59

134 and T-47D). We demonstrate that copy-number changes can easily give rise60

to spurious signals of allelic imbalance. We are able to integrate the quantitative61

de Santiago et al. Page 4 of 57

information obtained from ChIP-seq data along with information of the background62

allele composition to accurately detect allele-specific binding and correct for arti-63

facts caused by allele-specific losses or gains at the structural genomic level. Because64

of this integrated modeling, we were able to detect a large number of ASB events65

from ChIP-seq data generating a resource for future functional studies.66

BaalChIP is implemented as an R [20] package and freely available from67

Bioconductor repository https://bioconductor.org/packages/release/bioc/68

html/BaalChIP.html [21].69

Results70

Overview of BaalChIP workflow71

In this study we aim to devise a method that allows to correct for copy number72

changes and other biases in the analysis of ASB from ChIP-seq or similar data. The73

BaalChIP workflow requires three different sets of input data (Figure 1): the SNP74

variants file, ChIP-seq datasets (e.g. sets of BAM and BED files of ChIP-seq data75

obtained for different proteins that may or may not be grouped into replicates),76

and a corresponding set of genomic DNA files for each individual sample. The SNP77

variants file and the BED files are used to select the sites for the analysis, the ChIP-78

seq BAM files are used to compute the allelic read counts and the gDNA files will79

allow to determine reference-allele frequency (RAF) values to correct for effects of80

the background allele composition. Alternatively, RAF values may be directly pre-81

computed from BAF (B allele frequency) scores (Figure 1a). It has been shown that82

removing duplicate reads can reduce technical sources of biases at ASB sites [22]83

and we suggest that BAM files are pre-processed to contain only uniquely aligned84

reads or flagged duplicated reads.85

BaalChIP Quality Control86

The BaalChIP workflow starts by first computing the allelic read counts for each87

assayed heterozygous SNP overlapping genomic intervals of interest (e.g. ChIP-seq88

de Santiago et al. Page 5 of 57

peaks). By default, only uniquely mapped reads (MAQ > 15), and sites with base89

quality > 10 are used (Figure 1b). In the next step several quality control (QC)90

filters are applied to consider technical biases that may contribute to the false91

identification of regulatory SNPs (Figure 1b).92

In a first round, BaalChIP excludes SNPs mapping to regions of known prob-93

lematic read alignment: blacklisted regions [23], non-unique regions with genomic94

mappability score of less than 1 (based on the UCSC mappability track, 50bp seg-95

ments) [22], and collapsed repeat regions [24]. The excluded regions contain genomic96

intervals that are frequently enriched for repetitive elements and frequently cause97

ambiguous read mapping and sequencing artifacts [25].98

The second QC filter performs read-alignment simulations to consider an addi-99

tional type of read mapping bias named the intrinsic bias [6, 17]. This bias occurs100

due to intrinsic characteristics of the genome that translate into different probabil-101

ities of read mapping. Even when reads differ only in one location, reads carrying102

one of the alleles may have a higher chance of matching multiple locations (i.e. have103

many repeats in the genome) and may therefore be mapped to an incorrect locus.104

This, in turn, results in the underestimation of read counts and may cause both105

false-positive and false-negative inferences of ASB.106

The third QC filter selects those SNPs that pass all filters in all replicated samples,107

provided that replicated samples exist. This step will mainly remove SNPs in regions108

where the ChIP-seq signal is not consistently detected across all replicates (for109

instance when coverage is zero in one of the replicates).110

The final QC filter consists of removing possible homozygous SNPs by removing111

any site where only one allele is observed [4, 26].112

BaalChIP detection of allele-specific binding (ASB) events113

The filtered SNPs and their allelic read counts are merged into a table with the114

total number of read counts in the reference (Ref) and alternative (Alt) alleles. No115

de Santiago et al. Page 6 of 57

data is entered (missing data, NA) if a SNP did not pass the previously applied QC116

step for that sample (Figure 1c).117

BaalChIP considers two additional biases that may lead to inaccurate estimates118

of ASB: the reference mapping (RM) and the reference allele frequency (RAF) bi-119

ases. The RM bias occurs because the reference genome only represents a single120

"reference" allele at each SNP position. Reads that carry the "non-reference al-121

lele" have an extra mismatch to the reference sequence. Previous work has shown122

that this creates a marked bias in favor of the alignment of reads that contain the123

reference genome and could therefore affect the accuracy of allele-specific binding124

estimates [6].125

The RAF bias occurs due to alterations in the background abundance of each126

allele (e.g. in regions of copy-number alterations) and the correction for this bias is127

one of the key features of BaalChIP. RAF values at each heterozygous variant are128

used in the model likelihood to correct of the observed ChIP-seq read counts relative129

to the amount of the reference allele. These are given as relative measures from 0 to130

1, where values between 0.5 and 1 denote an underlying bias to the reference allele,131

and a value between 0 and 0.5 to the alternative allele. RAF scores can be obtained132

from SNP array B-allele frequencies (BAF) by assigning the correct reference and133

alternative alleles to allele A and B generic labels (RAF is equal to BAF if the134

reference allele corresponds to the B allele; RAF is equal to 1-BAF if the reference135

allele corresponds to the A allele). Alternatively, if whole-genome DNA (gDNA)136

sequencing samples are given as an input, BaalChIP calculates the RAF values137

directly from the gDNA allelic read counts.138

Finally, BaalChIP uses a beta-binomial distribution to model read count data139

therefore accounting for extra variability (over-dispersion) in allelic counts, a phe-140

nomenon that is often observed in sequencing data [9, 17]. The output of BaalChIP141

de Santiago et al. Page 7 of 57

is a posterior distribution of the estimated allelic balance ratio in read counts ob-142

served after considering all sources of underlying biases (Figure 1c).143

Evaluation of BaalChIP performance144

In a controlled simulation study, we compare BaalChIP with the two major compet-145

ing methods to infer allelic imbalance: binomial test and iASeq [10]. The binomial146

test is the most frequently used method for the analysis of allele-specificity from147

ChIP-seq data [1, 2, 3, 4, 7]. In a recent study [15], biases caused by differences148

in copy number were accounted for by weighting the binomial null with the allelic149

ratios observed from the gDNA samples (the number of reads mapping to each150

allele in the genomic DNA). Therefore, to take these strategy into consideration,151

we set the null hypothesis on the probability of success to the estimated RAF bias,152

instead of 0.5. iASeq is another available method that has been shown to improve153

the detection of ASB [10]. The iASeq method uses a Bayesian framework to combine154

information from different experiments or replicates, and models the read counts155

with a beta-binomal distribution, instead of a simple binomial distribution, to ac-156

count for extra-binomial variation caused by technical variability. Although iASeq157

was not originally designed to overcome copy-number biases, we included it in our158

simulation study because of its offered improvement of the traditional binomial test.159

All three methods are tested on synthetic data sets with varying sequencing depth,160

number of TF binding to a SNP and copy number states. In addition, we also exam-161

ine the robustness of the methods against a wide range of true allelic balance ratios.162

The allelic imbalance calling performance is assessed by ROC (Receiver operating163

characteristic) curves. In the absence of copy number changes, i.e. RAF is 0.5, the164

performances of the three methods are similar (Figure 2a, d and g).165

In the presence of copy number changes, BaalChIP shows significant improvements166

in the identification of true ASB events. In Figure 2 we highlight results on data sets167

with RAF set to either 0.3 or 0.1, which represent modest and severe background168

de Santiago et al. Page 8 of 57

allelic imbalance due to copy number changes. Significant improvements can be169

found across varying sequencing depth and across the number of TFs considered.170

Particularly, in the data sets with modest copy number changes, the performance171

of BaalChIP is similar to its performance in the absence of copy number changes,172

whereas the performance of both binomial test and iASseq suffers badly. When RAF173

is 0.1, BaalChIP is still able to achieve relatively good performance. In addition,174

the performance of BaalChIP benefits greatly from aggregating binding data for175

multiple TFs at the SNP (Figure S3).176

Overall, the simulation results show that BaalChIP is robust and performs well177

across a wide range of variations in the data set. In particular, in regions with copy178

number changes BaalChIP shows significant improvements over the state-of-the-art179

baselines. Thus BaalChIP offers a robust analysis of ASB in samples obtained with180

frequent copy number changes.181

Case study 1: ENCODE data182

We applied BaalChIP to 548 samples from the ENCODE project [19]. In total 271183

ChIP-seq experiments were analyzed, assaying a total of 8 cancer and 6 non-cancer184

cell lines representing different tissues. The data contained either 2 or 3 replicates185

per experiment and 4 to 42 DNA-binding proteins per cell line (Table S3 and Figure186

S4 show a summary of the cell lines, tissues, and number of ChIP-seq experiments187

used in this case study).188

The initial number of genotyped heterozygous SNPs per cell line ranged from189

139K to 356K SNPs. We selected those SNPs that occurred within ChIP-seq peaks190

in the corresponding cell lines, which amounts to between 1.6% and 5.8% of all191

SNPs. To ensure a reliable set of heterozygous SNPs we applied the BaalChIP QC192

step with the default parameters and options. We removed an average of 30.4%193

of all SNPs that hit peaks (Figure S5b); 0.8% of the excluded sites were found in194

regions of problematic read alignments; 7.4% had biases in simulations (Figure S5a);195

de Santiago et al. Page 9 of 57

16.4% were not consistent between replicates and 5.8% had only 1 observable allele196

(Figure S5c). After BaalChIP QC the number of heterozygous SNPs considered for197

downstream analysis ranged from 1,636-12,097 SNPs.198

Allele-specific copy-number alterations change ChIP-seq read densities199

First we examined of how allele-specific copy-number alterations affect the allelic200

ratios observed from ChIP-seq data (Figure 3). The relative presence of each allele201

is measured by the B allele frequency (BAF) obtained from SNP arrays [27].202

BAF values lie in the interval [0, 1]. A BAF value of 0.5 indicates an equal presence203

of the two alleles and is the expected value for heterozygous sites (AB) in a diploid204

genome, while values of 0 and 1 indicate homozygous genotypes (AA and BB). In205

a normal, non-contaminated diploid sample, a BAF plot will have three bands, one206

centered at 0.5 for AB genotypes, and two bands at 0 and 1 for the AA and BB207

homozygous genotypes, respectively.208

Figure 3a presents the BAF plots for chromosome 1 of three cancer cell lines209

(MCF-7, K-562 and SK-N-SH) and one non-cancer cell line obtained from normal210

blood (GM12878). As expected, the BAF plot of GM12878 is characterized by a211

typical diploid pattern of BAF values distributed around 0, 0.5 and 1 corresponding212

to the diploid genotypes AA, BB and AB. The few small deviations from these values213

can be attributed to germ line-based copy-number aberrations [28, 29]. The BAF214

plot for the SK-N-SH cell line, a glioblastoma cancer cell line known to have a near215

diploid karyotyope [30], is also relatively stable. In contrast, the MCF-7 and K-562216

cancer cell lines show more complex BAF plots due to a variety of copy-number217

aberrations.218

To evaluate the effects of allele-specific copy-number alterations on the ChIP-seq219

read densities, we compared for all four cell lines the BAF scores with ChIP-seq220

allelic ratios obtained at heterozygous SNPs (hetSNPs) (Figure 3b). We report221

ChIP-seq allelic ratios (AR) as the proportion of total reads carrying the reference222

de Santiago et al. Page 10 of 57

allele. We observed a clear correlation between the BAF scores and the imbalance223

in the ChIP-seq allelic ratios in MCF-7 and K-562 cancer cell lines (Spearman224

ρ = 0.704 and 0.525, respectively), while this correlation is not observed in the225

near-diploid SK-N-SH (Spearman ρ = 0.001) and the normal GM12878 cell lines226

(Spearman ρ = 0.036).227

The same effect can be observed in all other cell lines (Figure S6a) and is partic-228

ularly striking in cancer cells with the highest proportion of extreme BAF scores229

(< 0.4 or > 0.6; Figure S6b). Taken together, these results demonstrate the need230

to adjust for the background genomic abundance of the alleles when attempting to231

identify putative cis-acting regulatory SNPs from ChIP-seq data. These data also232

support our assumption that for allelic changes that do not change TF affinity, the233

allele frequency in ChIP-seq data is proportional to the presence of these alleles in234

the genomic DNA.235

Allele-specific amplification events explain most of the imbalances observed in236

ChIP-seq data in cancer cell lines237

Having obtained a reliable set of heterozygous SNPs per cell line, we next applied the238

Bayesian framework implemented in BaalChIP to each of the 14 cell lines individ-239

ually. To assess the importance of adjusting for copy-number biases, we performed240

the analysis with and without adjusting for the RAF bias. In a diploid sample, the241

allelic ratios (AR) are expected to be distributed around the 0.5 average, assuming242

that only a small proportion of sites carry an imbalance towards one of the alleles.243

However, as shown in Figure 3 copy-number alterations can affect read densities244

which will have an effect on ARs, in particular in cancer cells. Figure S7a shows245

the observed AR values in ChIP-seq data compared to the values before and after246

RAF correction for all analyzed cell lines. We observed that after RAF copy-number247

correction the ’corrected allelic ratios’ become more evenly distributed around an248

de Santiago et al. Page 11 of 57

average of 0.5. This effect is particularly notable in data obtained from cancer cell249

lines (Figure S7a).250

Overall, we found 2,438 ASB sites across all cell lines (Table S4), with an average251

of 3.1% SNPs displaying allele-specific binding using our chosen cutoffs. Figure S7b252

demonstrates that the copy-number correction has a strong effect in cancer cell253

lines. In the most extreme cases the number of allele-specific imbalances is reduced254

by 4-fold. In normal cells, and in cancer cells with near diploid genomes, such as255

HL-60 and SK-N-SH [30, 31], the total number of identified ASB sites due to copy-256

number effects is much lower (Figure S7b). This is consistent with the fact that257

non-cancer cell lines carry far fewer copy number aberrations. Because ASB is more258

easily detected in regions of high sequence coverage due to increases in statistical259

power (Figure S8), we repeated the same analysis after selecting only SNPs with260

30x-40x read coverage to control for read depth biases, and demonstrate that the261

same effect is verified (Figure S7c). Importantly, the data shown here demonstrates262

that adjusting for RAF is able to remove artifacts that are caused by copy number263

alterations in cancer cell lines.264

We observed higher rates of ASB on chromosome X of female cell lines (GM12878,265

GM12892, IMR90, MCF10, MCF-7, SK-N-SH) than in autosomal chromosomes266

(Figure S9, χ2 test p-value < 2.2e-16 for chromosome X versus autosomal identified267

sites comparison for all 6 cell lines). These cases might be explained by the extent268

of X-inactivation. In normal tissue X-inactivation is random. However, in clonal cell269

lines the same X-chromosome will continue to be silenced and most X-linked genes270

are expressed in a mono-allelic fashion [32].271

ASB events are consistent within and between cell lines272

We evaluated the correlation of allelic ratios between pairs of biological replicates,273

between distinct proteins bound at the same site, and at identical ASB sites in274

different cell lines. The allelic ratios of all shared heterozygous SNPs are well corre-275

de Santiago et al. Page 12 of 57

lated across biological replicates (Figure S10). Secondly, different proteins binding276

to the same SNP also display concordant allelic ratios (Figure S11) across cell lines.277

Our observations are consistent with the concordance of allele-specific binding of278

different co-bound TFs described in previous studies [1, 2, 3, 10, 12, 33, 34]. The279

Spearman correlation coefficients of all pairwise comparisons (between replicates280

and within cells between different DNA-binding proteins at the same SNP) are281

plotted as a boxplot in Figure S12a. Positive Spearman correlation coefficients are282

observed in every case, and in the majority of the cases we observe correlation coeffi-283

cients > 0.8. Therefore, overall we find that BaalChIP generates highly reproducible284

results, across separate ChIP-seq experiments.285

In addition, we analysed if heterozygous sites shared by cell lines had allelic ratios286

that were skewed towards the same allele or not. Of the identified ASB sites, only287

a small proportion (149 out of 2,438; 6.1%) were shared between cell lines (Figure288

S12b). Of these ASB sites, 91% (136 out of 149; Table S5) show an agreement in the289

direction of the allelic bias. Discordant cases may be explained by environmental290

or epigenetic factors, by different genomic context in different cells [1, 2], or by low291

sequencing coverage of ChIP-seq data. The high proportion of ASB with the same292

allelic bias further supports the robustness of BaalChIP.293

Functional annotation of ASB sites294

We then examined the genomic distribution of the identified ASB SNPs. A large295

proportion of ASB SNPs overlap introns and intergenic regions, and a considerable296

proportion is found at promoter-proximal regions, mainly reflecting the binding297

patterns of the ChIPed proteins (Figure S13). A large proportion of ASB SNPs298

overlap previously predicted enhancer regions and histone modifications associated299

with active enhancers (H3K4me1, H3K27ac), with an average of 70.2% of ASB SNPs300

occurring within cell-type specific putative enhancer regions (Figure S14). However,301

this enrichment is not significant when compared to non-ASB SNPs (Table S6,302

de Santiago et al. Page 13 of 57

χ2 pvalue > 0.05), and it is possibly mainly reflecting the distribution of binding303

sites (ChIP-seq peak regions) from which the initial set of heterozygous SNPs was304

sampled.305

Finally, to examine the putative functional mechanisms of ASB SNPs, we assessed306

if they modulated transcription factor binding affinity. We used predictions from307

HaploReg to assess if the observed ASB SNPs alter canonical transcription factor308

binding motifs. We found that 88% of ASB SNPs (range 83% to 92%) are motif309

disrupting SNPs but this enrichment was not statistically significant, since 85% of all310

tested SNPs under ChIP peaks were also found to alter motif scores. No significant311

difference in the magnitude of the change in binding affinity was observed between312

motif disrupting ASB SNPs when compared to all tested motif-disrupting SNPs,313

using the Kolmogorov–Smirnov test (Figure S15) To determine which TF-motifs314

are most likely to be disrupted, we grouped SNPs according to the DNA-binding315

proteins as identified by ChIP-seq peaks, and identified the top motifs disrupted316

by ASB SNPs when compared to non-ASB SNPs. In the majority of cases, the top317

disrupted motifs match the DNA-binding proteins used to generate the ChIP-seq318

data (Table S7). Overall, these results provide good evidence that the allele-specific319

binding we identify represents a true biological phenomenon.320

Case study 2: FAIRE-seq data321

To demonstrate the generality of our approach, we applied BaalChIP to targeted322

FAIRE-sequencing data obtained from two breast-cancer cell lines, MDA-MB-134323

and T-47D. FAIRE stands for Formaldehyde-Assisted Isolation of Regulatory Ele-324

ments and is an effective method to identify DNA regions in the genome associated325

with open chromatin [35]. The method is based on the fact that the formaldehyde326

cross-linking is more efficient in nucleosome-bound DNA than it is in nucleosome-327

depleted regions of the genome. Thus, FAIRE-seq identifies regions of open chro-328

de Santiago et al. Page 14 of 57

matin. One advantage of FAIRE-seq over ChIP-seq is that the assayed chromatin329

is not limited to the location of specific DNA-associated proteins.330

We chose to focus on the fraction of the genome that has been previously associ-331

ated to breast-cancer risk. To do so, we selected 75 previously known breast-cancer332

risk regions (Table S8) [36] covering a total of 4.93Mb of the human genome. We333

performed targeted sequencing of 3 replicated FAIRE samples per cell-line and the334

correspondent genomic DNA (gDNA) controls. Targeted sequencing of the gDNA335

samples allowed us to determine with confidence the genotype of a high number of336

sites at the assayed breast cancer risk regions. We identified a total of 3208 and 1624337

heterozygous SNPs in MDA-MB-134 and T-47D cells, respectively. In this dataset,338

the sequenced gDNA samples were used for the RAF correction step, i.e. allelic339

ratios at each SNP position were calculated directly from gDNA samples and used340

for bias correction. We first applied the BaalChIP QC pipeline to eliminate biases.341

We noticed that none of the SNPs in the selected targeted regions overlapped re-342

gions of potential problematic alignments, and only a small proportion of sites were343

eliminated during the BaalChIP QC step (< 0.3 %).344

We observed high correlation between allelic ratios in the gDNA and FAIRE345

samples (Figure 4a), indicating that observed allele-specificity in FAIRE-seq samples346

is primarily due to copy-number alterations and must therefore be corrected for.347

Figure 4b shows the observed allelic ratios (AR) obtained in the FAIRE-seq samples348

compared to the values after correcting for gDNA. After correction the ARs become349

more evenly distributed around an average of 0.5. We found that approximately350

0.65% (MDA-MB-134) and 0.56% (T-47D) of the tested sites in the selected risk351

regions were allele-specific. These correspond to a total of 21 sites in MDA-MB-134352

and 9 sites in T-47D cell lines (Table 1).353

Out of the 21 SNPs identified in the MDA-MB-134 cell line, 11 cluster in the354

10q26.13 region. This cluster includes the rs2981579 SNP in the second intron of355

de Santiago et al. Page 15 of 57

the FGFR2 gene (Table 1), which is the SNP with the strongest association with356

breast cancer in genome-wide analysis [37] In the 10q26.13 region, we identify two357

breast cancer risk SNPs, rs2981579 and rs2981578 with a strong allelic imbalance358

towards their risk alleles, rs2981579-A and rs2981578-C, respectively.359

Comparing BaalChIP to competing methods360

We applied two competing methods, the binomial test and the iASeq [10] method,361

to the ENCODE and FAIRE-seq datasets.362

When applying the iASeq method to the ENCODE dataset, the existence of363

missing data created an error. Missing data occur for those samples that do not364

contain ChIP-seq peaks at the SNP in question, or if the a SNP did not pass the365

previously applied QC step for that sample. To overcome this limitation of iASeq,366

we replaced every missing data point by zero, with the caveat that this ad hoc367

approach may create unknown biases in our results.368

When applying the binomial test, for each cell line, we pooled read count data369

from different experiments and replicates. While this approach maximizes statistical370

power, it may mask heterogeneity in ASB obtained from different experiments.371

Previous studies have shown that at least 20x read coverage at a particular SNP372

position is necessary to reach adequate statistical power with the binomial test [9,373

12]. Therefore, we restricted our analysis to SNPs with a coverage of at least 20374

reads in the pooled data. We included two different ways of performing the binomial375

test, either by setting the probability of success to 0.5 or by weighting the binomial376

null with the RAF scores. In real data, unlike simulations, we are not able to access377

the true-positive and false-positive rates and compute ROC curves. For this reason,378

we mainly focused on testing whether BaalChIP is comparable to existing methods379

while improving the biased ASB detection in high copy-number regions (Figure 5a)380

or the problem of overdispersion in the data (Figure 5b).381

de Santiago et al. Page 16 of 57

First, for consistency of the analysis, we obtained the set of heterozygous SNPs to382

test using the BaalChIP QC step. We then compared BaalChIP to the binomial test383

method with (p=RAF) and without (p=0.5) copy-number correction based on how384

many of the identified ASB SNPs were in CNA regions. For the comparison to the385

iASeq method, we computed the estimated posterior probability obtained for SNPs386

different regions of altered copy-numbers. Figure 5a shows that the binomial test387

(without RAF correction) and the iASeq methods show a substantial bias towards388

the increased detection of ASB events in regions altered copy-number, as expected.389

Both BaalChIP and iASeq employ a beta-binomial distribution to model read390

count data, which accounts for extra-binomial variability (overdispersion). To check391

the effect of overdispersion in the ENCODE and FAIRE-seq data we considered only392

SNPs at regions of normal or close to normal copy numbers (RAF betwen 0.4-0.6)393

and computed the percentage of detected ASB sites in bins of increasing depth of394

coverage. The overdispersion is often more accentuated for sites of higher coverage395

(> 200 pooled total counts), which will create a particularly sensitive scenario for396

the detection of ASB sites using the binomial distribution. Figure 5b demonstrates397

clearly the effect of overdispersion in the FAIRE-seq data, particularly visible at398

sites of high coverage, which shows the advantage of BaalChIP and iASeq over the399

binomial model. This effect is less pronounced in the ENCODE data, where depth400

of coverage is overall lower (Figure S16).401

These results suggest that for higher read counts random fluctuations in allelic402

ratios should be modelled with a beta-binomial distribution, rather than a binomial403

distribution, to reduce the number of false-positive results.404

Discussion405

ASB analysis is an important method for identifying putative regulatory SNPs406

that might have an effect on transcription factor binding and gene expression and407

might be associated with disease phenotypes. Our method for Bayesian Analysis408

de Santiago et al. Page 17 of 57

of Allelic imbalances from ChIP-seq data, called BaalChIP, has been motivated by409

the need to address the issue of detecting allelic-specific imbalances from ChIP-seq410

data obtained specifically from cancer cell lines, which frequently carry copy-number411

alterations.412

While allele-specific copy number alterations can be associated with changes in413

transcript abundance [38] and are possibly implicated in cancer phenotype, they can414

confound the identification of allele-specific binding at potentially regulatory sites.415

Traditional methods to identify ASB have not been able to distinguish between416

effects that are caused by allele specific amplification of binding sites versus the417

effects caused by allele-specific binding by specific TFs. Our method now allows to418

distinguish these and will therefore help to better delineate the causal mechanisms419

of disease.420

BaalChIP is a general framework applicable to a wide range of assays. We have421

applied it to ENCODE ChIP-seq and our own FAIRE-seq samples, and demon-422

strated the utility of BaalChIP to identify ASB events from cancer genomes. Each423

of these chromatin assays has its own advantages for the detection of allele-specific424

binding. ChIP-seq precisely determines the location of specific DNA-associated pro-425

teins, while FAIRE-seq identifies broader regions of open chromatin which might be426

less informative in terms of pinpointing the functional regulatory elements of the427

genome. On the other hand, the ChIP assay is limited to the availability of high-428

quality antibodies and only one factor can be tested per experiment. It is still an429

open question which chromatin assay will be the most informative for understanding430

allele-specific binding differences.431

Cancer cell lines carry frequent copy number alterations. Therefore an important432

issue is the determination of the background biases in allele frequency observed in433

regions of altered copy-number. For the ENCODE dataset, we have demonstrated434

that the use of BAF scores obtained from microarrays could be used to correct for435

de Santiago et al. Page 18 of 57

these effects in the allelic ratios, however it is also noteworthy that the microarray436

technology used by the ENCODE study did not generate BAF scores comprehen-437

sively for all possible heterozygous sites, limiting the number of SNPs at which the438

allele-specificity of each could be assayed. Therefore, for the majority of the samples439

in the ENCODE case study, it is likely that the regulatory potential of true causal440

SNPs was not directly assayed. As demonstrated by the analysis of FAIRE-seq data,441

this issue can be overcome by including genomic DNA-sequencing as control sam-442

ples to detect allele-specific biases, provided that there is sufficient read coverage at443

each site.444

The identified ASB sites may form the basis for future functional analysis of445

the genome. Of particular interest, within the FGFR2 gene, rs2981578 has been446

previously suggested to be the key causal variant. Indeed this SNP displays the447

highest allelic imbalance in a breast cancer cell line (Table 1).448

BaalChIP has been designed for data from cell lines, not tumor samples [39].449

Normal contamination and the existence of different clonal subpopulations in tumor450

samples pose additional challenges and can distort the expected distribution of451

heterozygous variant allele fractions. In the future, these challenges can potentially452

be overcome by combining the ideas implemented in BaalChIP with probabilistic453

methods developed for dissecting genetic heterogeneity and cancer evolution [40].454

Conclusions455

In summary, BaalChIP is a rigorous probabilistic method to detect allelic imbalance456

by correcting for the effect of background allele frequency on the observed ChIP-seq457

read counts. BaalChIP implements stringent filtering and preprocessing steps and458

allows the joint analysis of multiple ChIP-seq samples across a single variant. In459

simulations, BaalChIP outperformed competing approaches, and in an application460

to 548 ENCODE ChIP-seq and 6 targeted FAIRE-seq samples BaalChIP effectively461

de Santiago et al. Page 19 of 57

corrected allele-specific analysis for copy number changes and increased the power462

to detect putative cis-acting regulatory variants in cancer genomes.463

Methods464

BaalChIP Quality Control and filtering465

The QC and filtering steps implemented in the BaalChIP R package are:466

1 Keep only reads with MAPQ > 15 and with base quality > 10467

2 Within each cell line only consider heterozygous SNPs overlapping transcrip-468

tion factor binding sites identified by ChIP-seq peak calling;469

3 Exclude sites susceptible to allelic mapping bias in regions of known problem-470

atic read alignment [22, 23, 24];471

4 Apply simulation-based filtering to exclude SNPs with intrinsic bias to one of472

the alleles [6, 17];473

5 Consider only SNPs that are represented in all replicated samples after ap-474

plying all previous filters;475

6 Exclude possibly homozygous SNPs where only one allele was observed after476

pooling ChIP-seq reads from all examined samples [4, 26].477

Generating allele counts per SNP478

For each SNP, the number of reads carrying the reference and alternative alleles479

are computed using the pileup function and the PileupParam constructor of the480

Rsamtools package [41]. For each BAM file, BaalChIP only considers heterozygous481

SNPs overlapping the genomic regions in the corresponding BED files (peaks). Two482

arguments of the PileupParam constructor can be manipulated by the user: the483

min_mapq which refers to the minimum "MAPQ" value for an alignment to be484

included in pileup (reads with mapping quality score lower than this threshold are485

ignored; default is 15); and the min_base_quality which refers to the minimum486

quality value for each nucleotide in an alignment (bases at a particular location487

with base quality lower than this threshold are ignored; default is 10).488

de Santiago et al. Page 20 of 57

Data sources for regions of problematic alignments489

BaalChIP considers three sets of regions known to be of problematic read align-490

ment, by default: (1) blacklisted regions downloaded from the UCSC Genome491

Browser (mappability track; hg19, wgEncodeDacMapabilityConsensusExcludable492

and wgEncodeDukeMapabilityRegionsExcludable tables), (2) non-unique regions493

selected from DUKE uniqueness mappability track of the UCSC genome browser494

(hg19, wgEncodeCrgMapabilityAlign50mer table), and (3) collapsed repeat regions495

downloaded from [24] at the 0.1 % threshold. Sets of filtering regions used in this496

filtering step are fully customized and additional sets can be added by the user as497

GenomicRanges objects [42].498

Simulations to identify SNPs with intrinsic biases499

For each heterozygous site, BaalChIP simulates every possible read overlapping the500

site in four possible combinations - reads carrying the reference allele (plus and501

minus strand), and reads carrying the alternative allele (plus and minus strand).502

The simulated reads are constructed based on published methodology [6] (scripts503

shared by Degner et al. upon request) without taking into consideration different504

qualities at each base in the read or different read depth of coverage. As described505

by Degner et al. these parameters were sufficient to predict the SNPs that show an506

inherent bias [6].507

Reads are then aligned to the reference genome and BAM files are generated508

using Bowtie version 1.1.1 [43] and Picard tool version 1.47. The pipeline used509

to generate and align simulated reads can be fully customized with other aligners510

(e.g. BWA) and it is available under the file name "run_simulations.sh" found511

in the folder "extra" of the BaalChIP R package. Simulated allelic read counts512

are computed using the pileup function and the PileupParam constructor of the513

Rsamtools package [41]. For each SNP, the correct number of reads that should map514

to the reference and non-reference alleles is known (corresponds to twice the read515

de Santiago et al. Page 21 of 57

length for each allele); sites with incorrect number of read alignments are discarded516

from the analysis.517

Estimating reference mapping bias518

The reference mapping bias is calculated as described in [4, 26]. First, reads are519

combined across all heterozygous sites that pass the previous QC steps. The ex-520

pected allelic ratio (REF/TOTAL) for each cell line is then calculated separately521

for each allele combination. A minimum of 200 sites is required for each category;522

if there were less, a global estimate was used for that category. For each SNP this523

computation results in an allelic ratio µ ∈ (0, 1). These allelic ratios are used as524

part of a prior distribution in the Bayesian model described in the next section.525

BaalChIP Bayesian model description526

BaalChIP is designed to infer allelic imbalance from read counts data while inte-527

grating copy number and technical bias information. We assume that we are given528

a set of N datasets covering a SNP position. We jointly infer allelic imbalance from529

signals in all datasets covering the same SNP.530

For the nth dataset let dn ∈ N denote the total number of reads covering a SNP

position and an ∈ N the number of reads reporting the reference allele. BaalChIP

models an with a beta-binomial distribution, i.e. the binomial distribution in which

the probability of success is integrated out given that it follows the beta distribution:

p(an | dn, α, β) = BetaBin(an | dn, α, β). (1)

The beta-binomial distribution controls for overdispersion, i.e. the fact that the

increased variance in NGS data cannot be captured in the standard binomial

model [18]. To gain a more intuitive interpretation of the parameters, we re-

parametrize α and β in terms of the precision of the beta-binomial distribution,

de Santiago et al. Page 22 of 57

Λ, and the mean probability that a reference read is observed, θ:

α = θΛ (2)

β = (1− θ)Λ (3)

Including reference mapping bias In the next level of our model, we place a beta

distribution over θ, which allows us to naturally include the technical reference

mapping bias:

p(θ | α0, β0) = Beta (θ | α0, β0) , (4)

where, as before, we re-parametrise the shape parameters α0 and β0 by the mean

and variance of the distribution. Denoting the variance by λ and using as the mean

the reference mapping bias µ ∈ (0, 1) as discussed above, we get:

α0 =

((1− µ)

λ− 1

µ

)µ2 (5)

β0 =

((1− µ)

λ− 1

µ

)(µ− µ2

)(6)

As a consequence of this re-parametrisation, the reference mapping bias has a very531

intuitive interpretation as the a priori mean probability that a read reporting the532

reference allele is observed.533

Including reference allele frequency Next, we reparameterize θ as a function of

allelic balance ratio η and reference allele frequency ρ:

θ ≡ f(η, ρ) =ηρ

ηρ+ (1− η)(1− ρ). (7)

de Santiago et al. Page 23 of 57

Intuitively, this parameterization implements the Bayes’ rule with the following

definitions:

θ = Prob( allele = Ref | binding = Yes ),

η = Prob( binding = Yes | allele = Ref ),

1− η ≈ Prob( binding = Yes | allele = Alt ),

ρ = Prob( allele = Ref ),

1− ρ = Prob( allele = Alt ).

The auxiliary condition on binding status in the definition of θ reflects the fact534

that we only analysed peaks and thus assume that observed ChIP-seq reads are535

obtained after TF binding. Eq. 7 then formulates a ChIP-seq experiment as a process536

to obtain the posterior probability of a TF binding to the reference allele. In this537

model, the allelic balance ratio acts as the likelihood for TF binding to the reference538

allele and the RAF as the prior.539

To better understand how this model works, we will discuss two illustrative sce-540

narios. Example 1: Assuming no copy number alterations, i.e. RAF ρ = 0.5, and541

no allelic imbalance, i.e. η = 0.5, then θ = 0.5×0.50.5×0.5+0.5×0.5 = 0.5, which shows that542

BaalChIP contains the assumption of previous approaches as a special case. Exam-543

ple 2: In the event of LOH on the alternative allele, RAF ρ = 1, then irrespective544

of the allelic balance ratio, all observed reads will report the reference allele, which545

is reflected by θ = η×1η×1+(1−η)×0 = 1.546

Posterior distribution of allelic imbalance To rigorously represent the uncertainty

in the data, we adopt a full Bayesian approach targeting the posterior distribution

of allelic balance ratio η. We use the fact that the change of variables in Eq. 7 and

de Santiago et al. Page 24 of 57

the prior on θ directly define a prior over η, namely

p(η|ρ, µ, λ) = p(f(η, ρ)|µ, λ)∂f(η, ρ)

∂η, (8)

and write the posterior of η as follows:

p(η | {an}N1 , {bn}N1 , ρ,Λ, µ, λ

)(9)

∝N∏n=1

p(an|dn, f(η, ρ),Λ) p(f(η, ρ)|µ, λ)∂f(η, ρ)

∂η,

where

∂f(η, ρ)

∂η=

ρ

ηρ+ (1− η)(1− ρ)− ηρ(2ρ− 1)

(ηρ+ (1− η)(1− ρ))2. (10)

Inference of allelic imbalance We use the Metropolis-Hastings algorithm with ran-547

dom walk proposal to approach this distribution. In practice, we fix the Λ = 1000548

and λ = 0.05, which in our experience results in robust performance across a wide549

range of simulated and real datasets. Allelic imbalance calling is based on the high-550

est posterior density (HPD) interval, which is constructed from the MCMC trace551

of η as the shortest interval containing 95% of sampled values. Allelic imbalance is552

called, if the HPD does not contain the value 0.5, which is a rigorous way to decide553

that the data cannot be explained well by balanced alleles.554

In silico validations555

To thoroughly test the performance of allelic imbalance calling performance of556

BaalChIP, read counts data are generated using a wide range of parameter set-557

tings. Specifically, we varied the number N of data sets from 1 to 45, ρ from 0.1 to558

0.9 in steps of size 0.1, d from 1 to 100 in steps of size 7, and 1000 values of η are559

de Santiago et al. Page 25 of 57

evenly sampled from 0.1 to 0.9. These settings result in a set of 6, 075, 000 SNPs.560

The detailed simulation protocol is as follows:561

For each n ∈ [1, ..., N ] and ρ ∈ [0.1, ..., 0.9] and η ∈ [0.1, ..., 0.9]:562

1. Draw the number of available reference alleles:

tn ∼ Binomial(dn, ρ)

2. Draw the read counts reporting reference allele:

if η ≤ 0.5 :

an ∼ Binomial(tn, 2η)

if η > 0.5 :

bn ∼ Binomial(dn − tn, 2(1− η))

an = dn − bn

The performance is measured by ROC curves. The true allelic imbalance is deter-563

mine if the true η is outside the interval [0.45, 0.55].564

Baseline methods565

We compared BaalChIP against two baseline approaches: binomial test and566

iASeq [10]. Two-sided binomial tests are performed with the R function binom.test.567

To correct for reference mapping bias, the null hypothesis on the probability of suc-568

cess is set to be the previously estimated reference mapping bias (RAF), instead of569

0.5. We pooled data from all ChIP-seq samples for each analyzed cell line to max-570

imize power. For the analysis of real data, binomial test p-values were corrected571

for multiple testing using the false discovery rate (FDR) threshold of 0.01. FDR572

was calculated with the p.adjust function in R. For the simulated data, the ROC573

curves for binomial test are constructed based on the distance between reference574

de Santiago et al. Page 26 of 57

mapping bias and the mean of 95% confidence intervals for the probability of suc-575

cess. All iASseq analyses are performed with the function iASeqmotif where the576

number of non-null motif is set to be a vector [1, 2, ..., 5], the maximun number of577

iterations is 300 and the tolerance level of error is 0.001. The ROC curves for iASseq578

are constructed with bestmotif$p.post which is the posterior probability for each579

SNP being allele-specific event.580

Cell culture581

MDA-MB-134 and T-47D human breast cancer cells were cultured in RPMI (Invit-582

rogen) supplemented with 10% FBS and antibiotics. All cells were maintained at583

37C, 5% CO2. All cell lines were from the CRUK Cambridge Institute biorepository584

collection. Cell lines were authenticated by short tandem repeat genotyping using585

the GenePrint 10 (Promega) system and confirmed to be mycoplasma free.586

FAIRE and gDNA purification587

FAIRE stands for ’Formaldehyde assisted isolation of regulatory elements’ and was588

performed as previously described [35] with minor adaptations. Briefly, cells were589

fixed for 10 mins in 1% formaldehyde in FCS-free medium, washed and frozen. Nu-590

clei from MDA-MB-134 (3× 107 cells/tube) and T-47D (1.5× 107 cells/tube) were591

isolated and sonicated using 300µl volumes in 1.5ml eppendorfs, using a Diagenode592

Bioruptor. Sonication was performed for 20 cycles of 30s on/off at the "high" set-593

ting. The sonicate was subjected to three consecutive phenol–chloroform–isoamyl594

alcohol (25:24:1) extractions and reverse cross-linked overnight. DNA was purified595

by ethanol precipitation and quantified by Quanti-iT. Genomic DNA was isolated596

using Qiagen DNeasy Blood and tissue Kit according to protocol.597

Agilent SureSelect amplification, library preparation and sequencing598

DNA fragments were prepared for sequencing using the recommended protocol avail-599

able for SureSelect-Illumina sequencing. 69 known breast cancer risk tagging SNPs600

de Santiago et al. Page 27 of 57

were retrieved from [36]. SNAP Proxy Search was used to find SNPs correlated601

with the 69 tagging SNPs (r2 > 0.6 for 59 of the risk loci and r2 > 0.8 for 8 of the602

risk loci, within a distance of 500bp) using 1000 Genomes pilot 1 data. Genomic603

intervals were defined by the left-most to the right-most SNP in each LD block with604

an additional 400bp of flanking regions. An additional set of 7 SNPs with 500bp605

of flanking sequences were added manually to include the CCND1 (rs75915166,606

rs494406), MAP3K1 (rs10461612, rs112497245, rs17432750) and TERT gene regions607

(rs2736107, rs7705526). In total, we targeted 69 non-overlapping regions compris-608

ing 4.93Mb using SureSelect method (Table S8). DNA obtained from SureSelect609

solution-based sequence capture was subjected to Illumina HiSeq paired-end se-610

quencing (Illumina). Paired-end Sequencing was performed according to manufac-611

turer’s protocols.612

Pre-processing ENCODE samples613

We used publicly available ENCODE ChIP-seq and genotype datasets for a to-614

tal of 548 samples representing 271 different experiments. We included 8 cancer615

and 6 non-cancer cell lines representing different tissues. Table S3 shows a sum-616

mary of all tissues, cell lines and number of experiments included in this study.617

The ChIP-seq data was downloaded as reads mapped to the hg19 genome (BAM618

files) and correspondent peak calling files (BED files). Accession numbers of all619

public ChIP-seq datasets used in this study are provided in Table S2. Duplicated620

reads were marked using Picard tool version 1.47. ChIP-seq peak files were merged621

between replicates using HOMER version 5.4 mergePeaks with the default op-622

tion -d given to only consider peak ranges that overlapped for all replicates. Het-623

erozygous SNPs and BAF tracks were retrieved from the UCSC Genome browser624

(hg19, wgEncodeHaibGenotype track, wgEncodeHaibGenotypeBalleleSnp2015-03-625

04.tsv and wgEncodeHaibGenotypeGtypeSnp2014-09-15.tsv files). The initial num-626

ber of genotyped SNPs in the ENCODE files is 1.2 million SNPs. We only considered627

de Santiago et al. Page 28 of 57

SNPs listed in dbSNP (version 137 [44]). Homozygous SNPs (i.e., BAF > 0.9 or628

< 0.1) and SNPs with missing BAF scores were removed from the BAF tracks.629

BAF scores were converted to Reference Allele Frequency (RAF) scores using the630

information of A and B alleles in the Manifest file for the Illumina Human1M-Duo631

BeadChip (v3.0) downloaded from http://support.illumina.com/array/array_632

kits/human1m-duo_dna_analysis_kit/downloads.html.633

Pre-processing FAIRE-seq samples634

For sequence data from all FAIRE-seq and control samples, sequences were aligned635

to the human reference genome (GRCh37) using BWA version 0.7.12 [45]. Du-636

plicates were removed using Picard Tool version 1.47 and overlapping reads were637

clipped using clipOverlap tool from bamUtil repository version 1.0.14 with default638

parameters. SNPs were identified from gDNA sampes using the Genome Analysis639

ToolKit 3.4-46 software [46] across all gDNA samples simultaneously. As per GATK640

Best Practices recommendations [47, 48], duplicated reads were removed and local641

realignment and base quality recalibration were employed prior to SNP calling.642

Called SNPs were filtered using a hard filtering criteria (QD < 2.0, FS > 60.0, MQ643

< 30.0, MQRankSum < -12.5, ReadPosRankSum < -8.0).644

Applying BaalChIP to ENCODE and FAIRE-seq samples645

To ensure a reliable set of heterozygous SNPs we applied the BaalChIP (version646

0.1.9) quality control step and considered only uniquely mapping reads with MAQ647

> 15 and base call quality > 10. For the ENCODE dataset we removed from the648

analysis sets of SNPs based on the default six QC filters implemented within the649

BaalChIP pipeline. The FAIRE-seq samples contained paired-end sequenced reads650

of 125 bp. Since longer and paired-end reads reduce uncertainty of read alignment we651

did not consider two of the QC filters that are more relevant for shorter read lengths652

(of less than 50bp): the unique mappability and intrinsic bias filters. BaalChIP653

de Santiago et al. Page 29 of 57

(version 0.1.9) ASB Bayesian analysis was performed with the default parameters654

and options [21].655

Consistency of allelic ratios656

Consistency of the allelic ratios observed at ASB SNPs was analysed between (i)657

between replicates: pairs of replicated samples, (ii) within cell lines: pairs of ChIP-658

seq datasets from different proteins (pooled replicate data) in the same cell line (iii)659

between cell lines: pairs of cells (pooled data for each cell line). Correlations were660

calculated with Spearman correlation and required at least 15 shared ASB sites.661

SNP annotations relative to genes662

To annotate ASB SNPs with respect to gene annotations - 5’ UTR, 3’ UTR, pro-663

moter, spliceSite, coding, intron or intergenic - we used the VariantAnnotaion (ver-664

sion 1.8.13) package in R. To obtain the list of known genes and coordinates,665

we used the UCSC genome browser known gene annotation obtained from the666

TxDb.Hsapiens.UCSC.hg19.knownGene (version 2.6.2) library in R [20].667

Overlap with predicted enhancers668

To determine if ASB SNPs were enriched in any of the putative enhancer regions,669

we calculated the overlap of intergenic ASB SNPs in putative enhancer regions.670

The significance of the observed overlap was determined by a χ2 test by compar-671

ing the fraction of ASB SNPs in putative enhancer regions with the fraction of672

non-ASB heterozygous SNPs in putative enhancer regions. Putative enhancer re-673

gions were retrieved for 6 cell lines: GM12878, H1hESC, HeLa, HepG2, K562, A549674

based on predicted weak and strong enhancer sites and/or H3K27ac and H3K4me1675

chromatin marks. Enhancer site predictions were retrieved from human segmen-676

tations previously generated based on ENCODE data using Segway (downloaded677

from wgEncodeAwgSegmentation UCSC genome browser track).678

de Santiago et al. Page 30 of 57

TF motif disruption analysis679

In order to annotate the potential regulatory effects of the tested SNPs on tran-680

scription factor binding site (TFBS) motifs, the publically available HaploReg681

v2 database (accessible at http://www.broadinstitute.org/mammals/haploreg/682

haploreg_v2.php) was used. Haploreg calculates allele-specific changes in the log-683

odds (LOD) scores for position weight matrices (PWMs) of a regulatory motif684

based on a library of PWMs constructed from TRANSFAC, JASPAR, and and685

protein-binding microarray (PBM) experiments [49]. The magnitude of the change686

in binding affinity was calculated as the absolute difference (DELTA) of LOD scores687

(DELTA = LOD(ref) - LOD(alt)). The Kolmogorov-Smirnov test was used to com-688

pare the distributions of the absolute DELTA scores of motif disrupting ASB SNPs689

and all tested motif disrupting SNPs. To determine which TFBS-motifs ASB SNPs690

were more likely to disrupt, SNPs were grouped according to DNA-binding proteins691

as identified by ChIP-seq peaks. For each group, a Fisher exact test was used to692

identify TFBS-motifs that were more significantly disrupted by ASB SNPs when693

compared to non-ASB SNPs.694

Availability of data and materials695

ENCODE datasets used in this study are publicly available in the curated Gene Expression Omnibus (GEO)696

database. Accession numbers of all public ChIP-seq datasets used in this study are provided in Table S2.697

FAIRE-seq datasets used in this study are included in the GEO database (GEO accession number, GSE85261)698

Competing interests699

The authors declare that they have no competing interests.700

Author’s contributions701

IS, WL and KY designed and developed the Baal-ChIP statistical framework. IS and WL implemented702

BaalChIP package in R. IS analyzed the data. WL preformed simulation studies. KM and MOR designed and703

preformed FAIRE-seq experiments. BAJP, FM and KBM supervised the analysis. IS, WL, KY, KBM and FM704

wrote the manuscript. FM and KBM are co-corresponding authors.705

Funding706

We would like to acknowledge the support of The University of Cambridge, Cancer Research UK and Hutchison707

Whampoa Limited. Parts of this work were funded by CRUK core grant C14303/A17197 and A19274 and the708

Breast Cancer Research Foundation.709

de Santiago et al. Page 31 of 57

Acknowledgments710

We thank Thomas Carroll, Gordon Brown and Jing Su (CRUK Cambridge Institute, University of Cambridge)711

for suggestions and advice on ChIP-seq data analysis.712

References713

1. McDaniell, R., Lee, B.-K., Song, L., Liu, Z., Boyle, A.P., Erdos, M.R., Scott, L.J., Morken, M.A., Kucera,714

K.S., Battenhouse, A., et al.: Heritable individual-specific and allele-specific chromatin signatures in715

humans. Science 328(5975), 235–239 (2010)716

2. Reddy, T.E., Gertz, J., Pauli, F., Kucera, K.S., Varley, K.E., Newberry, K.M., Marinov, G.K., Mortazavi,717

A., Williams, B.A., Song, L., et al.: Effects of sequence variation on differential allelic transcription factor718

occupancy and gene expression. Genome research 22(5), 860–869 (2012)719

3. Kasowski, M., Kyriazopoulou-Panagiotopoulou, S., Grubert, F., Zaugg, J.B., Kundaje, A., Liu, Y., Boyle,720

A.P., Zhang, Q.C., Zakharia, F., Spacek, D.V., et al.: Extensive variation in chromatin states across721

humans. Science 342(6159), 750–752 (2013)722

4. Kilpinen, H., Waszak, S.M., Gschwind, A.R., Raghav, S.K., Witwicki, R.M., Orioli, A., Migliavacca, E.,723

Wiederkehr, M., Gutierrez-Arcelus, M., Panousis, N.I., et al.: Coordinated effects of sequence variation on724

dna binding, chromatin structure, and transcription. Science 342(6159), 744–747 (2013)725

5. McVicker, G., van de Geijn, B., Degner, J.F., Cain, C.E., Banovich, N.E., Raj, A., Lewellen, N., Myrthil,726

M., Gilad, Y., Pritchard, J.K.: Identification of genetic variants that affect histone modifications in human727

cells. Science 342(6159), 747–749 (2013)728

6. Degner, J.F., Marioni, J.C., Pai, A.A., Pickrell, J.K., Nkadori, E., Gilad, Y., Pritchard, J.K.: Effect of729

read-mapping biases on detecting allele-specific expression from rna-sequencing data. Bioinformatics730

25(24), 3207–3212 (2009)731

7. Rozowsky, J., Abyzov, A., Wang, J., Alves, P., Raha, D., Harmanci, A., Leng, J., Bjornson, R., Kong, Y.,732

Kitabayashi, N., et al.: Alleleseq: analysis of allele-specific expression and binding in a network framework.733

Molecular systems biology 7(1), 522 (2011)734

8. Satya, R.V., Zavaljevski, N., Reifman, J.: A new strategy to reduce allelic bias in rna-seq readmapping.735

Nucleic acids research, 425 (2012)736

9. Skelly, D.A., Johansson, M., Madeoy, J., Wakefield, J., Akey, J.M.: A powerful and flexible statistical737

framework for testing hypotheses of allele-specific gene expression from rna-seq data. Genome research738

21(10), 1728–1737 (2011)739

10. Wei, Y., Li, X., Wang, Q.-f., Ji, H.: iaseq: integrative analysis of allele-specificity of protein-dna740

interactions in multiple chip-seq datasets. BMC genomics 13(1), 681 (2012)741

11. Mayba, O., Gilbert, H.N., Liu, J., Haverty, P.M., Jhunjhunwala, S., Jiang, Z., Watanabe, C., Zhang, Z.:742

Mbased: allele-specific expression detection in cancer tissues and cell lines. Genome Biol 15(8), 405 (2014)743

12. Li, G., Bahn, J.H., Lee, J.-H., Peng, G., Chen, Z., Nelson, S.F., Xiao, X.: Identification of allele-specific744

alternative mrna processing via transcriptome sequencing. Nucleic acids research, 280 (2012)745

13. Chen, J., Rozowsky, J., Galeev, T.R., Harmanci, A., Kitchen, R., Bedford, J., Abyzov, A., Kong, Y.,746

Regan, L., Gerstein, M.: A uniform survey of allele-specific binding and expression over747

1000-genomes-project individuals. Nature communications 7 (2016)748

14. Almlöf, J.C., Lundmark, P., Lundmark, A., Ge, B., Pastinen, T., Goodall, A.H., Cambien, F., Deloukas, P.,749

Ouwehand, W.H., Syvänen, A.-C., et al.: Single nucleotide polymorphisms with cis-regulatory effects on750

long non-coding transcripts in human primary monocytes (2014)751

15. Bailey, S.D., Virtanen, C., Haibe-Kains, B., Lupien, M.: Abc: a tool to identify snvs causing allele-specific752

de Santiago et al. Page 32 of 57

transcription factor binding from chip-seq experiments. Bioinformatics, 321 (2015)753

16. Liu, Z., Gui, T., Wang, Z., Li, H., Fu, Y., Dong, X., Li, Y.: cisase: a likelihood-based method for detecting754

putative cis-regulated allele-specific expression in rna sequencing data. Bioinformatics (Oxford, England)755

32, 3291–3297 (2016). doi:10.1093/bioinformatics/btw416756

17. Pickrell, J.K., Marioni, J.C., Pai, A.A., Degner, J.F., Engelhardt, B.E., Nkadori, E., Veyrieras, J.-B.,757

Stephens, M., Gilad, Y., Pritchard, J.K.: Understanding mechanisms underlying human gene expression758

variation with rna sequencing. Nature 464(7289), 768–772 (2010)759

18. Roth, A., Khattra, J., Yap, D., Wan, A., Laks, E., Biele, J., Ha, G., Aparicio, S., Bouchard-Côté, A., Shah,760

S.P.: Pyclone: statistical inference of clonal population structure in cancer. Nature methods 11(4),761

396–398 (2014)762

19. Consortium, E.P.: An integrated encyclopedia of dna elements in the human genome. Nature 489(7414),763

57–74 (2012)764

20. R Core Team: R: A Language and Environment for Statistical Computing. R Foundation for Statistical765

Computing, Vienna, Austria (2014). R Foundation for Statistical Computing. http://www.R-project.org/766

21. de Santiago, I., Wei, L., Yuang, K., Markowetz, F.: Baal-ChIP: Bayesian Analysis of Allele-specific Binding767

Events from ChIP-seq Data. (2015). R package version 0.1.9.768

https://github.com/InesdeSantiago/BaalChIP769

22. Castel, S.E., Levy-Moonshine, A., Mohammadi, P., Banks, E., Lappalainen, T.: Tools and best practices770

for data processing in allelic expression analysis. Genome biology 16(1), 1 (2015)771

23. Fujita, P.A., Rhead, B., Zweig, A.S., Hinrichs, A.S., Karolchik, D., Cline, M.S., Goldman, M., Barber,772

G.P., Clawson, H., Coelho, A., et al.: The ucsc genome browser database: update 2011. Nucleic acids773

research, 963 (2010)774

24. Pickrell, J.K., Gaffney, D.J., Gilad, Y., Pritchard, J.K.: False positive peaks in chip-seq and other775

sequencing-based functional assays caused by unannotated high copy number regions. Bioinformatics776

27(15), 2144–2146 (2011)777

25. Carroll, T.S., Liang, Z., Salama, R., Stark, R., de Santiago, I.: Impact of artifact removal on chip quality778

metrics in chip-seq and chip-exo data. Front Genet 5, 75 (2014)779

26. Lappalainen, T., Sammeth, M., Friedländer, M.R., Aoen, P., Monlong, J., Rivas, M.A., Gonzàlez-Porta,780

M., Kurbatova, N., Griebel, T., Ferreira, P.G., et al.: Transcriptome and genome sequencing uncovers781

functional variation in humans. Nature 501(7468), 506–511 (2013)782

27. Peiffer, D.A., Le, J.M., Steemers, F.J., Chang, W., Jenniges, T., Garcia, F., Haden, K., Li, J., Shaw, C.A.,783

Belmont, J., et al.: High-resolution genomic profiling of chromosomal aberrations using infinium784

whole-genome genotyping. Genome research 16(9), 1136–1148 (2006)785

28. Iafrate, A.J., Feuk, L., Rivera, M.N., Listewnik, M.L., Donahoe, P.K., Qi, Y., Scherer, S.W., Lee, C.:786

Detection of large-scale variation in the human genome. Nature genetics 36(9), 949–951 (2004)787

29. Sebat, J., Lakshmi, B., Troge, J., Alexander, J., Young, J., Lundin, P., Månér, S., Massa, H., Walker, M.,788

Chi, M., et al.: Large-scale copy number polymorphism in the human genome. Science 305(5683), 525–528789

(2004)790

30. Biedler, J.L., Helson, L., Spengler, B.A.: Morphology and growth, tumorigenicity, and cytogenetics of791

human neuroblastoma cells in continuous culture. Cancer research 33(11), 2643–2652 (1973)792

31. Liang, J.C., Ning, Y., Wang, R.-y., Padilla-Nash, H.M., Schröck, E., Soenksen, D., Nagarajan, L., Ried, T.:793

Spectral karyotypic study of the hl-60 cell line: detection of complex rearrangements involving794

chromosomes 5, 7, and 16 and delineation of critical region of deletion on 5q31. 1. Cancer genetics and795

de Santiago et al. Page 33 of 57

cytogenetics 113(2), 105–109 (1999)796

32. Gimelbrant, A., Hutchinson, J.N., Thompson, B.R., Chess, A.: Widespread monoallelic expression on797

human autosomes. Science 318(5853), 1136–1140 (2007)798

33. Tang, F., Barbacioru, C., Nordman, E., Bao, S., Lee, C., Wang, X., Tuch, B.B., Heard, E., Lao, K., Surani,799

M.A.: Deterministic and stochastic allele specific gene expression in single mouse blastomeres. PLoS One800

6(6), 21208 (2011)801

34. Ni, Y., Hall, A.W., Battenhouse, A., Iyer, V.R.: Simultaneous snp identification and assessment of802

allele-specific bias from chip-seq data. BMC genetics 13(1), 46 (2012)803

35. Giresi, P.G., Kim, J., McDaniell, R.M., Iyer, V.R., Lieb, J.D.: Faire (formaldehyde-assisted isolation of804

regulatory elements) isolates active regulatory elements from human chromatin. Genome research 17(6),805

877–885 (2007)806

36. Michailidou, K., Hall, P., Gonzalez-Neira, A., Ghoussaini, M., Dennis, J., Milne, R.L., Schmidt, M.K.,807

Chang-Claude, J., Bojesen, S.E., Bolla, M.K., et al.: Large-scale genotyping identifies 41 new loci808

associated with breast cancer risk. Nature genetics 45(4), 353–361 (2013)809

37. Turnbull, C., Ahmed, S., Morrison, J., Pernet, D., Renwick, A., Maranian, M., Seal, S., Ghoussaini, M.,810

Hines, S., Healey, C.S., et al.: Genome-wide association study identifies five new breast cancer811

susceptibility loci. Nature genetics 42(6), 504–507 (2010)812

38. Tuch, B.B., Laborde, R.R., Xu, X., Gu, J., Chung, C.B., Monighetti, C.K., Stanley, S.J., Olsen, K.D.,813

Kasperbauer, J.L., Moore, E.J., et al.: Tumor transcriptome sequencing reveals allelic expression814

imbalances associated with copy number alterations. PLoS One 5(2), 9317 (2010)815

39. Ross-Innes, C.S., Stark, R., Teschendorff, A.E., Holmes, K.A., Ali, H.R., Dunning, M.J., Brown, G.D.,816

Gojis, O., Ellis, I.O., Green, A.R., et al.: Differential oestrogen receptor binding is associated with clinical817

outcome in breast cancer. Nature 481(7381), 389–393 (2012)818

40. Beerenwinkel, N., Schwarz, R.F., Gerstung, M., Markowetz, F.: Cancer evolution: mathematical models819

and computational inference. Systematic biology 64, 1–25 (2015).820

41. Morgan, M., Pagès, H., Obenchain, V., Hayden, N.: Rsamtools: Binary alignment (BAM), FASTA, variant821

call (BCF), and tabix file import. R package version 1.18. 2822

42. Lawrence, M., Huber, W., Pages, H., Aboyoun, P., Carlson, M., Gentleman, R., Morgan, M.T., Carey,823

V.J.: Software for computing and annotating genomic ranges. PLoS Comput Biol 9(8), 1003118 (2013)824

43. Langmead, B., Trapnell, C., Pop, M., Salzberg, S.L., et al.: Ultrafast and memory-efficient alignment of825

short dna sequences to the human genome. Genome biol 10(3), 25 (2009)826

44. Sherry, S.T., Ward, M.-H., Kholodov, M., Baker, J., Phan, L., Smigielski, E.M., Sirotkin, K.: dbsnp: the827

ncbi database of genetic variation. Nucleic acids research 29(1), 308–311 (2001)828

45. Li, H., Durbin, R.: Fast and accurate short read alignment with burrows–wheeler transform. Bioinformatics829

25(14), 1754–1760 (2009)830

46. McKenna, A., Hanna, M., Banks, E., Sivachenko, A., Cibulskis, K., Kernytsky, A., Garimella, K., Altshuler,831

D., Gabriel, S., Daly, M., et al.: The genome analysis toolkit: a mapreduce framework for analyzing832

next-generation dna sequencing data. Genome research 20(9), 1297–1303 (2010)833

47. DePristo, M.A., Banks, E., Poplin, R., Garimella, K.V., Maguire, J.R., Hartl, C., Philippakis, A.A.,834

Del Angel, G., Rivas, M.A., Hanna, M., et al.: A framework for variation discovery and genotyping using835

next-generation dna sequencing data. Nature genetics 43(5), 491–498 (2011)836

48. Auwera, G.A., Carneiro, M.O., Hartl, C., Poplin, R., del Angel, G., Levy-Moonshine, A., Jordan, T., Shakir,837

K., Roazen, D., Thibault, J., et al.: From fastq data to high-confidence variant calls: the genome analysis838

de Santiago et al. Page 34 of 57

toolkit best practices pipeline. Current protocols in bioinformatics, 11–10 (2013)839

49. Ward, L.D., Kellis, M.: Haploreg: a resource for exploring chromatin states, conservation, and regulatory840

motif alterations within sets of genetically linked variants. Nucleic acids research 40(D1), 930–934 (2012)841

de Santiago et al. Page 35 of 57

Figures842

0

2

4

6

0.00 0.25 0.50 0.75 1.00

SNP ID RM Bias RAF

rs2070942 0.51 0.34

rs4951389 0.51 0.69

rs10153729 0.50 0.59

Peak Ref Alt Ref Alt Peak Ref Alt

1 0 1 1 0 0 NA NA

1 10 12 4 7 1 3 4

1 1 0 0 1 0 NA NA

Rep1 + Rep2

TF1

...

TF2

SNP ID Lower Upper Allele Bias

rs2070942 0.57 0.95 Yes

rs4951389 0.13 0.40 Yes

rs10153729 0.07 0.87 No

• (2a) Blacklist regions

• (2a) non-unique regions

• (2a) Collapsed repeats

• (2b) Intrinsic bias

• (2d) Present in all reps• (2c) Only one allele

Bayesian Model

RM Biasbeta distribution

RAF,Ref & Total reads

beta binomial

TFn

x

Prior Likelihood

Posterior Distribution

Het. SNPs

gDNA BAM filesOr

pre-computed RAF scores

1

2

(a) Input data

(b) BaalChIP QC SNP filtering

(c) BaalChIP ASB detection

3

4

5

(d) Output

TFn

ChIP-seqBAM and BED files

Reads & peaks VariantsAllele frequency(for correction)

MAPQ<15base_qual<10

Figure 1 Description of BaalChIP model frame work. (a) The basic inputs for Baal are theChIP-seq raw read counts in a standard BAM alignment format, a BED file with the genomicregions of interest (such as ChIP-seq peaks) and a set of heterozygous SNPs in a tab-delimitedtext file. Optionally, genomic DNA (gDNA) BAM files can be specified for RAF computation,alternatively, the user can specify the pre-computed RAF scores for each variant. (b) The firstmodule of BaalChIP consists of (1) computing allelic read counts for each heterozygous SNP inpeak regions and (2) a round of filters to exclude heterozygous SNPs that are susceptible togenerating artifactual allele-specific binding effects. (3) the Reference Mapping (RM) bias and theReference Allele Frequency (RAF) are computed internally and the output consists of a datamatrix where RM and RAF scores are included alongside the information of allele counts for eachheterozygous SNP. The column ’Peak’ contains binary data used to state the called peaks. (c)The second module of BaalChIP consists of calling ASB binding events. (4) BaalChIP uses abeta-binomial Bayesian model to consider RM and RAF bias when detecting the ASB events. (d)The output from BaalChIP is a posterior distribution for each SNP. A threshold to identify SNPswith allelic bias is specified by the user (default value is a 95% interval). (5) The output ofBaalChIP is a credible interval (’Lower’ and ’Upper’) calculated based on the posteriordistribution. This interval corresponds to the true allelic ratio in read counts (i.e. after correctingfor RM and RAF biases). An ASB event is called if the lower and upper limits of the interval areoutside the 0.4-0.6 interval.

de Santiago et al. Page 36 of 57

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(a) TF = 3, RAF = 0.5, Reads per TF = 1

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(b) TF = 3, RAF = 0.3, Reads per TF = 1

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(c) TF = 3, RAF = 0.1, Reads per TF = 1

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(d) TF = 5, RAF = 0.5, Reads per TF = 8

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(e) TF = 5, RAF = 0.3, Reads per TF = 8

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(f) TF = 5, RAF = 0.1, Reads per TF = 8

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(g) TF = 15, RAF = 0.5, Reads per TF = 15

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(h) TF = 15, RAF = 0.3, Reads per TF = 15

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(i) TF = 15, RAF = 0.1, Reads per TF = 15

Baal−ChIP Binomial Test iASeq

Figure 2 The ROC curve comparison between BaalChIP and other allele-specific SNP finding

methods: binomial test and iASeq, using a simulated data set. The BaalChIP result is shown by

solid red line. Binomial test and iASeq are shown in dashed blue and black lines. The number of

TFs able to bind at a given SNP and reads per TF are increasing from 3 to 15 and 1 to 15,

respectively. RAF is decreasing from 0.5 to 0.1.

de Santiago et al. Page 37 of 57

MCF7

0 125 250

00.

51

cor=0.704

0 0.5 1

0.2

0.5

0.8

chr1 position (M)

AR (ChIP−seq)

B A

llele

Fre

qR

AF

a

b

K562

0 125 250

00.

51

cor=0.525

0 0.5 1

0.2

0.5

0.8

chr1 position (M)

AR (ChIP−seq)

B A

llele

Fre

qR

AF

a

b

SKNSH

0 125 250

00.

51

cor=0.001

0 0.5 1

0.2

0.5

0.8

chr1 position (M)

AR (ChIP−seq)

B A

llele

Fre

qR

AF

a

b

GM12878

0 125 250

00.

51

cor=−0.036

0 0.5 1

0.2

0.5

0.8

chr1 position (M)

AR (ChIP−seq)

B A

llele

Fre

qR

AF

a

b

Figure 3 Example cancer and non-cancer cell lines SNP and ChIP-seq ENCODE data. (a) B allele

frequencies (BAF) for chromosome 1 for three cancer cell lines (MCF-7, K562, SK-N-SH) and one

non-cancer cell line (GM12878). Individual SNPs are colored according to genotype values:

homozygous AA or BB (blue) and heterozygous AB (orange). (b) Correlations between the BAF

values and the ChIP-seq allelic ratios (AR) of heterozygous SNPs. RAF corresponds to the BAF

value with respect to the reference allele (RAF is equal to BAF if the reference allele corresponds

to the B allele; RAF is equal to 1-BAF if the reference allele corresponds to the A allele). The

fitted linear model (blue line) and the Spearman correlation coefficient (cor) show the relationship

between BAF and ChIP-seq allelic ratios at heterozygous sites.

de Santiago et al. Page 38 of 57

0.0 0.2 0.4 0.6 0.8 1.0

0.2

0.4

0.6

0.8

1.0

AR (FAIRE−seq)

AR

(gD

NA

)

cor=0.422

MDAMB134

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0 cor=0.706

T47Da

b

Allelic Ratio (AR)

densi

ty

0

3

6

9

0 0.25 0.50 0.75 10 0.25 0.50 0.75 1

Corrected ARFAIRE-seq AR

MDAMB134 T47D

Figure 4 ASB detection from FAIRE targeted sequencing data. (a) Correlations between the

allelic ratios obtained from genomic DNA and FAIRE-seq data. (b) Density plots showing the

distribution of allelic ratios before (green) and after (orange) BaalChIP correction. The adjusted

AR values were estimated by BaalChIP model after taking into account the RAF scores computed

directly from the control genomic DNA samples.

de Santiago et al. Page 39 of 57

ENCODE FAIRE-seq

[0.4

7 -

0.5

3]

[0.4

5-0

.47

)(0

.53

- 0

.55

]

[0.4

-0.4

5)

(0.5

5 -

0.6

]

[0.3

5-0

.4)

(0.6

- 0

.65

]

[0.3

-0.3

5)

(0.6

5 -

0.7

]

[0-0

.3)

(0.7

- 1

]

[0.4

7 -

0.5

3]

[0.4

5-0

.47

)(0

.53

- 0

.55

]

[0.4

-0.4

5)

(0.5

5 -

0.6

]

[0.3

5-0

.4)

(0.6

- 0

.65

]

[0.3

-0.3

5)

(0.6

5 -

0.7

]

[0-0

.3)

(0.7

- 1

]

Relative allele frequency

ENCODE FAIRE-seq

Depth of Coverage

a

b>

60

0

> 1

.5e+

03

Figure 5 Comparison of BaalChIP with other available methods. Left y-axis corresponds to the

frequency of ASB events called by BaalChIP or the binomial tests (red, blue and purple lines).

Right y-axis corresponds to the mean of the posterial probability given by the iASeq method

(black line). Small numbers at the top of the plot show the total number of tested heterozygous

sites in each bin. (a) SNPs were grouped in bins of different RAF intervals. The RAF intervals are

increasing in terms of distance to the diploid (RAF=0.5) value. The bionomial test (without RAF

correction; purple line) and the iASeq methods (black line) are biased towards the detection of

ASB events in regions of altered copy numbers. (b) SNPs were grouped in bins of different depth

of coverage. SNPs in regions RAF < 0.4 or RAF > 0.6 were excluded from this analysis. When

applying the binomial test (purple and blue lines), the frequency of ASB detection increases for

higher covered sites, while the same is not true when applying BaalChIP or the iASeq methods.

The effect is particularly visible for the FAIRE-seq dataset.

Tables843

de Santiago et al. Page 40 of 57

AC

C ARead

cou

nts

SNP

CC

C

AAAAAA

A

C

SNP

ChIP-seq read out:

Aneuploid region:

AAAAAA

C

AC

TFRegulatory variant in a diploid region:

Allele 1:Allele 2:

Allele 2:

Allele 1(amplicon)

Figure S1 Example illustration of a ChIP-seq read out at a DNA-binding site when a true

regulatory difference exists between two alleles (A allele depicted in orange and C allele depicted

in blue), and when there is no regulatory difference between the two alleles. In a diploid sample,

the protein (yellow circle) binds preferentially to the A allele. Here the regulatory effect is

observed as an imbalance in the allelic ratios obtained from ChIP-seq read counts, with a higher

number of reads carrying the A allele. In the presence of allele-specific copy number aberrations

such as an amplicon affecting the A allele, the direct ChIP-seq readout may reflect the relative

presence of the alleles, rather than the regulatory effect of the single nucleotide variant.

Consequently, in the presence of copy number changes, ChIP-seq allelic ratios are not sufficient to

uncover true cis-regulatory effects.

de Santiago et al. Page 41 of 57

Table 1 Heterozygous SNPs identified as allele-specific and the correspondent cell-lines and alleles inthe FAIRE-seq dataset. CHROM, POS correspond to the coordinates in the hg19 human referencegenome.

ID CHROM POS REF ALT ASB allele Allelic Imbalance GeneName Location

rs2373062 2 218322137 G C T47D-G 0.76 DIRC3 intron

rs9682063 3 4737842 A G T47D-G 0.34 ITPR1 intron

rs9292122 5 56087910 G A T47D-G 0.81 intergenic

rs201314815 7 144115683 C T T47D-T 0.28 intergenic

rs201949580 7 144115698 T G T47D-G 0.30 intergenic

rs200312388 7 144115704 T G T47D-G 0.29 intergenic

rs200496470 8 128323987 A G T47D-G 0.15 CASC8 intron

rs111929748 11 69379287 G T T47D-T 0.32 intergenic

rs12939887 17 53162672 A G T47D-A 0.67 STXBP4 intron

rs72650670 1 114394681 G A MDAMB134-G 0.76 PTPN22 coding

rs116782319 5 158260122 A T MDAMB134-T 0.33 EBF1 intron

rs12670370 7 144118594 C A MDAMB134-C 0.79 intergenic

rs2392922 8 129166840 A G MDAMB134-A 0.79 intergenic

rs17793170 8 129205679 G A MDAMB134-A 0.33 intergenic

rs2666763 10 22101608 G A MDAMB134-A 0.29 DNAJC1 intron

rs2941733 10 22241656 T C MDAMB134-T 0.74 DNAJC1 intron

rs4752570 10 123337066 T C MDAMB134-C 0.32 FGFR2 intron

rs2912780 10 123337117 C T MDAMB134-C 0.69 FGFR2 intron

rs2912779 10 123337182 T C MDAMB134-T 0.72 FGFR2 intron

rs2981579 10 123337335 A G MDAMB134-A 0.71 FGFR2 intron

rs2912778 10 123338654 G A MDAMB134-G 0.73 FGFR2 intron

rs11200017 10 123339685 G A MDAMB134-A 0.26 FGFR2 intron

rs2981578 10 123340311 C T MDAMB134-C 0.88 FGFR2 intron

rs11599804 10 123340664 G A MDAMB134-A 0.16 FGFR2 intron

rs34032268 10 123341525 A C MDAMB134-C 0.19 FGFR2 intron

rs2936870 10 123348902 T C MDAMB134-T 0.69 FGFR2 intron

rs3135718 10 123353869 C T MDAMB134-C 0.67 FGFR2 intron

rs112536831 19 17424329 G A MDAMB134-G 0.75 DDA1 intron

rs113207944 19 17438923 C T MDAMB134-C 0.74 ANO8 intron

rs10416361 19 44294617 A G MDAMB134-A 0.75 intergenic

de Santiago et al. Page 42 of 57

�µ

⌘ ⇢

⇤ an dn

N

Figure S2 Graphical model of BaalChIP. The gray circles represent given variables and white

circles are unknown variables. The direction of arrows indicates statistical dependency in the

model. The observed reference allele read count is an, total read counts is dn. n is the index of

given factor (e.g. transcription factor) binding to the region of the SNP of interest. The

rectangular plane and the capital N indicates that there are N factors binding to the SNP. The

primary variable of interest is η representing allelic balance ratio. Λ is the precision parameter.

The reference allele frequency at the SNP is ρ. dn, η, ρ,Λ are the parameters of a beta-binomial

distribution that governs the uncertainty in an. Finally, the reference mapping bias µ and λ are

the mean and variance of a beta distribution which serves as a prior over η.

de Santiago et al. Page 43 of 57

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(a) RAF = 0.5 & Reads per TF = 15

TF = 1TF = 5TF = 10TF = 20TF = 30

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

False Positive Rate

True

Pos

itive

Rat

e

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

(b) RAF = 0.5 & TF = 1

reads = 3reads = 30reads = 250reads = 520reads = 1000

Figure S3 BaalChIP simulation performance with varying number of transcription factors (a)

and read depth (b).

de Santiago et al. Page 44 of 57

ATF

3B

CLA

F1

CB

X3

CE

BP

BC

EB

PD

CR

EB

1C

TC

FC

TC

FL

E2F

6E

GR

1E

LF1

ET

S1

FO

SL1

GA

BPA

GAT

A2

MA

XM

EF

2AN

R2F

2N

RS

FP

ML

PO

L2P

U1

RA

D21

SIN

3AK

20S

IX5

SP

1S

P2

SR

FS

TAT

5ATA

F1

TAF

7T

EA

D4

TH

AP

1T

RIM

28U

SF

1Y

Y1

ZB

TB

33Z

BT

B7A

BA

F15

5B

AF

170

BR

CA

1cJ

UN

CH

D2

ELK

1G

CN

5G

TF

2F1

HC

FC

1IN

I1IR

F3

JUN

DM

AF

KM

AZ

MX

I1N

RF

1P

300

PR

DM

1C

OR

ES

TR

FX

5S

MC

3S

PT

20S

TAT

3T

BP

US

F2

ZK

SC

AN

1Z

NF

143

BC

L3F

OS

L2F

OX

A2

TC

F12

FO

XM

1G

ATA

3H

DA

C2

FO

XA

1B

HLH

E40

HN

F4A

HN

F4G

MB

D4

MY

BL2

NF

ICR

XR

AZ

EB

1P

BX

3cF

OS

cMY

CAT

F2

BC

L11A

NA

NO

GP

OU

5F1

SP

4B

ATF

EB

F1

IRF

4M

EF

2CM

TA3

NFA

TC

1PA

X5

PO

U2F

2R

UN

X3

TC

F3

CH

D1

K562

HeLa

A549

MCF7

T47D

HepG2

SKNSH

HL60

MCF10

H1hESC

GM12878

GM12891

GM12892

IMR90

DNA−binding proteins

Figure S4 Data matrix summarizing the ENCODE samples used in this study for allele-specific

binding analyses. More detailed description including assession numbers can be found in Table S2

and Table S3.

de Santiago et al. Page 45 of 57

75

80

85

90

95

30 35 40 45 50read length

Rightcalls(%)

0.00

0.25

0.50

0.75

1.00

A549

GM12878

GM12891

GM12892

H1hESC

HeLa

HepG2

HL60

IMR90

K562

MCF10

MCF7

SKNSH

T47D

cell line

Filtered/Total(%)

keyblacklist

highcov

mappability

intbias

replicates_merged

Only1Allele

pass

pass

Only1Allele

replicates_merged

intbias

mappabilityhighcovblacklista

b

c

Figure S5 BaalChIP SNP filtering. (a) Percentage of SNPs with the correct number of aligned

reads based on simulations of reads of different read lengths (28 to 50 mer). The percentage of

correct calls increased with read length. (b) Proportion of SNPs that were filtered out in each

filtering step. (b) The average proportion of SNPs that were filtered out in each filtering step.

Average was calculated accross all 14 ENCODE cell lines.

de Santiago et al. Page 46 of 57

cancer, A549 cancer, HeLa cancer, HepG2 cancer, HL60

cancer, K562 cancer, MCF7 cancer, SKNSH cancer, T47D

non−cancer, GM12878 non−cancer, GM12891 non−cancer, GM12892 non−cancer, H1hESC

non−cancer, IMR90 non−cancer, MCF10

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.00 0.25 0.50 0.75 1.000.00 0.25 0.50 0.75 1.00AR (ChIP-seq)

RA

F

0.0

0.2

0.4

0.6

0.0 0.2 0.4 0.6Correlation ChIP−seq AR vs. RAF scores

Pro

port

ion

ofsi

tes

inC

NA

regi

ons

celltype

cancer

non−cancer

a

b

Figure S6 Correlation between the BAF values and the ChIP-seq allelic ratios (AR). (a) Scatter

plots showing the correlation between RAF and AR at each SNP, for all considered cell lines of the

ENCODE dataset. RAF corresponds to the BAF value with respect to the reference allele (RAF is

equal to BAF if the reference allele corresponds to the B allele; RAF is equal to 1-BAF if the

reference allele corresponds to the A allele). The blue line shows the fitted linear regression line.

(b) Positive relationship between the Spearman correlation coefficient (obtained from the

correlation of AR with RAF scores) and the proportion of sites in copy-number altered (CNA)

regions. Each dot in the scatter plot corresponds to a cell line. Sites in copy number changed

regions correspond to heterozygous SNPs with BAF score < 0.4 or BAF > 0.6.

de Santiago et al. Page 47 of 57

●●●

●

●●

●

●

●

●●

●

●

●

●

●

0.025

0.050

0.075

0.100

0.125

0.0 0.2 0.4 0.6

●●

●

●

●

●●

●

●

●●

●

●

●

●

●0.05

0.10

0.15

0.20

0.0 0.2 0.4 0.6

Without RAF correction

With RAF correction

Cancer cell line

Non-cancer cell line

Pro

port

ion

of

ASB

site

sProportion of testedsites in CNA regions

Allelic Ratio (AR)

dens

ity

ChIP-seq AR (not corrected)

Corrected AR

K562 HeLa A549

MCF7 T47D HepG2

SKNSH HL60 MCF10

H1hESC GM12878 GM12891

GM12892 IMR90

0

1

2

3

0

1

2

3

0

1

2

3

0

1

2

3

0

1

2

3

0

1

2

3

4

0

1

2

3

0

1

2

3

0

1

2

3

0

1

2

3

4

0

1

2

3

4

0

1

2

3

0

1

2

3

0

1

2

3

4

0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

b

C

a

Pro

port

ion

of

ASB

site

s

Figure S7 Effects of BaalChIP adjustment of the allelic ratios (AR) after correcting for the

background Reference Allele Frequency (RAF) bias. (a) Density plots showing the distribution of

allelic ratios before (green) and after (orange) BaalChIP correction. The ARs before correction

were calculated directly from the ChIP-seq data and refer to the number of reads in the reference

allele divided by the total number of reads. The corrected AR values were estimated by BaalChIP

model after taking into account the RAF bias. The adjustment of AR is more dramatic in cancer

cell lines. (b) Proportion of ASB sites detected with (orange) and without (green) RAF correction

as a function of the proportion of sites within copy-number altered (CNA) regions. Each dot refers

to a cell line. Sites in regions of copy number change refer to any SNP with a BAF score higher

than 0.6 or lower than 0.4. Cancer cell lines have a higher proportion of SNPs in CNA regions and

benefit more from BaalChIP correction. (c) Same as (b) but after selecting SNPs with 30X-40X

coverage.

de Santiago et al. Page 48 of 57

0.00

0.04

0.08

0.12

20 30 40 50 20 30 40 50

Average coverage at tested sites

Pro

port

ion

of

AS

B s

ites

With RAF correction Without RAF correction

cell type

cancernon-cancer

Figure S8 Proportion of identified ASB SNPs per cell line. Correlation between the proportion of

identified ASB SNPs and the average coverage at all assayed heterozygous SNPs. The general

correlation is expected due to the higher power to detect ASB when coverage is high.

de Santiago et al. Page 49 of 57

0.0

0.2

0.4

0.0

0.1

0.2

0.000.010.020.030.04

0.00

0.05

0.10

0.15

0.0

0.1

0.2

0.3

0.0

0.1

0.2

0.3

GM

12878G

M12892

IMR

90M

CF

10M

CF

7S

KN

SH

chr1

chr1

0ch

r11

chr1

2ch

r13

chr1

4ch

r15

chr1

6ch

r17

chr1

8ch

r19

chr2

chr2

0ch

r21

chr2

2ch

r3ch

r4ch

r5ch

r6ch

r7ch

r8ch

r9ch

rX

Pro

port

ion

ofA

SB

site

s

cancernon−cancer

Figure S9 Higher rates of ASB on chromosome X of female cell lines than in autosomal

chromosomes.

de Santiago et al. Page 50 of 57

A549, FOSL2 A549, FOXA2 A549, GABPA A549, J UND A549, NRSF A549, P300 A549, POL2 A549, SIX5 A549, TCF12 GM12878, ATF2 GM12878, BATF GM12878, BCL11A GM12878, BCL3 GM12878, CREB1

GM12878, EBF1 GM12878, EGR1 GM12878, ELF1 GM12878, FOXM1 GM12878, GABPA GM12878, IRF4 GM12878, MEF2A GM12878, MTA3 GM12878, NFIC GM12878, NRSF GM12878, PAX5 GM12878, PBX3 GM12878, POL2 GM12878, POU2F2

GM12878, PU1 GM12878, RAD21 GM12878, RUNX3 GM12878, SIX5 GM12878, SP1 GM12878, SRF GM12878, STAT5A GM12878, TAF1 GM12878, TCF12 GM12878, TCF3 GM12878, YY1 GM12891, POL2 GM12891, POU2F2 GM12891, PU1

GM12891, TAF1 GM12891, YY1 GM12892, POL2 GM12892, YY1 H1hESC, CREB1 H1hESC, CTCF H1hESC, E2F6 H1hESC, EGR1 H1hESC, MAX H1hESC, NRSF H1hESC, POL2 H1hESC, RAD21 H1hESC, SIN3AK20 H1hESC, SP1

H1hESC, SRF H1hESC, TAF1H1hESC, TCF12 H1hESC, TEAD4 H1hESC, USF1 H1hESC, YY1 HeLa, BRCA1 HeLa, CEBPB HeLa, CHD2 HeLa, cJ UN HeLa, GTF2F1 HeLa, HCFC1 HeLa, J UND HeLa, MAFK

HeLa, MAX HeLa, MXI1 HeLa, P300 HeLa, POL2 HeLa, PRDM1 HeLa, RAD21 HeLa, RFX5 HeLa, SMC3 HeLa, STAT3 HeLa, TBP HeLa, USF2 HepG2, CEBPB HepG2, CREB1 HepG2, CTCF

HepG2, ELF1 HepG2, FOSL2 HepG2, FOXA1 HepG2, FOXA2 HepG2, GABPA HepG2, HNF4A HepG2, HNF4G HepG2, J UND HepG2, MAX HepG2, NR2F2 HepG2, P300 HepG2, POL2 HepG2, RAD21 HepG2, RXRA

HepG2, SRF HepG2, TAF1 HepG2, USF1 HepG2, YY1 HL60, GABPA HL60, NRSF HL60, POL2 HL60, PU1 IMR90, CEBPB IMR90, CTCF IMR90, MAFK IMR90, MXI1 IMR90, POL2 IMR90, RAD21

K562, ATF3 K562, CEBPB K562, CEBPD K562, CREB1 K562, CTCF K562, E2F6 K562, EGR1 K562, ELF1 K562, ETS1 K562, FOSL1 K562, GABPA K562, GATA2 K562, MAX K562, MEF2A

K562, NR2F2 K562, NRSF K562, PML K562, POL2 K562, PU1 K562, RAD21 K562, SP1 K562, SRF K562, STAT5A K562, TAF1 K562, TEAD4 K562, TRIM28 K562, USF1 K562, ZBTB7A

MCF10, cFOS MCF10, cMYC MCF10, POL2 MCF10, STAT3 MCF7, CEBPB MCF7, CTCF MCF7, EGR1 MCF7, ELF1 MCF7, FOSL2 MCF7, FOXM1 MCF7, GABPA MCF7, GATA3 MCF7, HDAC2 MCF7, J UND

MCF7, MAX MCF7, NR2F2 MCF7, NRSF MCF7, P300 MCF7, PML MCF7, RAD21 MCF7, SIN3AK20 MCF7, SRF MCF7, TAF1 MCF7, TCF12 MCF7, TEAD4 SKNSH, ELF1 SKNSH, FOSL2 SKNSH, FOXM1

SKNSH, GABPA SKNSH, GATA3 SKNSH, J UND SKNSH, MAX SKNSH, NFIC SKNSH, NRSF SKNSH, P300 SKNSH, PBX3 SKNSH, POL2 SKNSH, RXRA SKNSH, TCF12 SKNSH, TEAD4 SKNSH, USF1 SKNSH, YY1

SKNSH, ZBTB33 T47D, CTCF T47D, FOXA1 T47D, GATA3 T47D, P300

0

0.5

1

0 0.5 1Allelic Ratio (Replicate 1)

Alle

lic R

ati

o (

Replic

ate

2)

0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1

0

0.5

1

0

0.5

0

0.5

1

0

0.5

1

0

0.5

1

1

0

0.5

1

0

0.5

1

0

0.5

0

0.5

1

0

0.5

1

0

0.5

1

1

0

0.5

1

0

0.5

1

Figure S10 Consistency of ASB across biological replicates. Each dot in the scatter plot

represents the allelic ratio at a single SNP obtained for two replicated samples from the same

ChIP-seq experiment.

de Santiago et al. Page 51 of 57

K562 HeLa A549 MCF7

T47D HepG2 SKNSH HL60

MCF10 H1hESC GM12878 GM12891

GM12892 IMR90

0.0

0.5

1.0

0.0

0.5

1.0

0.0

0.5

1.0

0.0

0.5

1.0

0.0 0.5 1.00.0 0.5 1.0Allelic Ratio (Protein X)

Alle

lic R

atio

(P

rote

in Y

)

Figure S11 Consistency of ASB across pairs of different assayed proteins (Protein "X" and "Y"),

within the same cell line. Each dot in the scatter plot represents the allelic ratio at a single SNP

obtained for distinct ChIP-seq experiments in the same cell line (i.e distinct ChIPed proteins

bound at the same site).

de Santiago et al. Page 52 of 57

0.4

0.6

0.8

Rep sameCell

Spearman

correlationcoefficient

(124)

(18)

(5)(1) (1)

0

50

100

2 3 4 5 6Number of cell lines

Num

berofshared

ASBsites

a b

Figure S12 (a) Distribution of Spearman correlation coefficients obtained for the correlation of

allelic ratios obtained for pairs of biological replicates (Rep), and between distinct proteins bound

at the same site (sameCell). The allelic ratios are highly correlated in both scenarios. (b) Number

of shared ASB SNPs across cell lines. Numbers in parenthesis show the number of ASB SNPs

shared between 2 to 6 cell lines. 149 ASB SNPs (124+18+5+1+1) were concomitantly detected

in two or more cell lines.

de Santiago et al. Page 53 of 57

0

10

20

30

40

50 ASB

AssayedPe

rcenta

ge o

f O

verl

ap

coding

5' UTR

intergenicintro

n

promoter

3' UTR

Figure S13 Percentage of ASB SNPs overlapping gene annotations (5’ UTR, 3’ UTR, coding,

intergenic, intron or promoter). The largest proportion of ASB SNPs was found in introns and

intergenic regions. A similar distribution pattern is observed when considering all assayed

heterozygous SNPs.

de Santiago et al. Page 54 of 57

0

25

50

75K27ac

k4me1 E

K27ac WE E

K27ac WE E

K27ac WE E

K27ac WE E

K27ac WE

Cell−specific chromatin features

Percentageofoverlap

A549 GM12878 H1hESC HeLa HepG2 K562

*

0

25

50

75

A549

GM12878

H1hESC

HeLa

HepG2

K562

Cell−specific chromatin features (merged ranges)

Percentageofoverlap

observed expectedb

a

Figure S14 Percentage of ASB SNPs overlapping putative enhancer regions from the matched

cell line. (a) Putative enhancer regions were retrieved from publicly available ENCODE datasets

on H3K27ac and H3K4me1 modified regions, and previously predicted weak enhancers (WE) and

strong enhancer (E) sites obtained from Segway chromatin state annotations of the ENCODE

data. (b) Comparison of the observed and expected number of ASB SNPs that map to enhancer

regions. A significant difference (*) was only detected in HeLa cells.

de Santiago et al. Page 55 of 57

SKNSH T47D

IMR90 K562 MCF10 MCF7

H1hESC HeLa HepG2 HL60

A549 GM12878 GM12891 GM12892

ASB Assayed ASB Assayed

ASB Assayed ASB Assayed

0

5

10

15

20

25

0

5

10

15

20

25

0

5

10

15

20

25

0

5

10

15

20

25

abs(LOD(alt)−LOD(ref))

group

ASB

Assayed

p=0.63

p=0.81

p=0.56

p=0.85

p=0.84

p=0.99

p=0.97

p=0.63

p=0.94

p=0.89

p=0.81

p=0.6

p=0.98

p=0.03

SKNSH T47D

IMR90 K562 MCF10 MCF7

H1hESC HeLa HepG2 HL60

A549 GM12878 GM12891 GM12892

ASB Assayed ASB Assayed

ASB Assayed ASB Assayed

0

5

10

15

20

25

0

5

10

15

20

25

0

5

10

15

20

25

0

5

10

15

20

25

abs(LOD(alt)−LOD(ref))

group

ASB

Assayed

a

b

Figure S15 Distribution of absolute DELTA scores (deltaLOD) of motif-disrupting SNPs with

(a) deltaLOD > 0 and (b) deltaLOD > 5. The DELTA score corresponds to the change in the

PWM score between SNP alleles (DELTA = LOD(ref) - LOD(alt)). P-values correspond to the

Kolmogorov-Smirnov test when comparing the distribution of DELTA scores between

motif-disrupting ASB SNPs with all motif-disrupting SNPs.

de Santiago et al. Page 56 of 57

ENCODE (ChIP−seq)

10100

1000

K562

HeLa

A549

MCF7

T47D

HepG2

SKNSH

HL60

MCF10

H1hESC

GM12878

GM12891

GM12892

IMR90

Depth

FAIRE−seq gDNA

10100

1000

MDA134

T47D

MDA134

T47D

Depth

a b

Figure S16 Distribution of depth of coverage at tested heterozygous SNPs in (a) ENCODE

ChIP-seq datasets (pooled data across all ChIP-seq samples in each cell line) and (b) targeted

sequencing FAIRE-seq datasets (pooled data for three replicates).

de Santiago et al. Page 57 of 57

Additional Files844

Additional file 1 — supplementary_tables.xlsx845

Supplementary tables S1 - S8.846