Gkac 1143

Published online 8 December 2022 Nucleic Acids Research, 2022, Vol. 50, No.
21 12131–12148
https://doi.org/10.1093/nar/gkac1143
Cis-regulatory mutations associate with

transcriptional and post-transcriptional deregulation
of gene regulatory programs in cancers
Jaime A. Castro-Mondragon 1 , Miriam Ragle Aure 2,3 , Ole Christian Lingjærde 2,4,5
,
Anita Langerød2 , John W.M. Martens 6 , Anne-Lise Børresen-Dale 2 ,
Vessela N. Kristensen 2,3 and Anthony Mathelier 1,2,3,*
1
Centre for Molecular Medicine Norway (NCMM), Nordic EMBL Partnership, University of Oslo, 0318 Oslo, Norway,
Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

2
Department of Cancer Genetics, Institute for Cancer Research, Oslo University Hospital Radiumhospitalet, 0310
Oslo, Norway, 3 Department of Medical Genetics, Institute of Clinical Medicine, University of Oslo and Oslo University
Hospital, Oslo, Norway, 4 Centre for Bioinformatics, Department of Informatics, University of Oslo, Gaustadalléen 23
B, N-0373 Oslo, Norway, 5 KG Jebsen Centre for B-cell malignancies, Institute for Clinical Medicine, University of Oslo,
Ullernchausseen 70, N-0372 Oslo, Norway and 6 Erasmus MC Cancer Institute and Cancer Genomics Netherlands,
University Medical Center Rotterdam, Department of Medical Oncology, 3015GD Rotterdam, The Netherlands
Received May 18, 2022; Revised November 03, 2022; Editorial Decision November 11, 2022; Accepted November 17, 2022
ABSTRACT our method predicts cis-regulatory mutations related

to the dysregulation of key gene regulatory networks
Most cancer alterations occur in the noncoding por-
in cancer patients.
tion of the human genome, where regulatory regions
control gene expression. The discovery of noncod-
ing mutations altering the cells’ regulatory programs INTRODUCTION
has been limited to few examples with high recur- Dysregulation of the gene expression regulatory programs
rence or high functional impact. Here, we show that in a cell is a hallmark of cancer. The often observed aber-
transcription factor binding sites (TFBSs) have simi- rant gene expression in cancer can be triggered by dereg-
lar mutation loads to those in protein-coding exons. ulation at any regulatory level (transcriptional and post-
By combining cancer somatic mutations in TFBSs transcriptional) (1,2). While the majority of studies have fo-
and expression data for protein-coding and miRNA cused on the mutations lying within protein-coding regions,
most alterations occur in the noncoding portion of the hu-
genes, we evaluate the combined effects of tran-
man genome, where cis-regulatory elements reside and act
scriptional and post-transcriptional alterations on as genetic switches to ensure that gene expression occurs
the regulatory programs in cancers. The analysis of at correct times and intensities in the correct cells and tis-
seven TCGA cohorts culminates with the identifica- sues (3). Molecular alterations in these regions can mod-
tion of protein-coding and miRNA genes linked to ulate the entire regulatory network of the cells, conferring
mutations at TFBSs that are associated with a cas- oncogenic traits associated with clinical and histopathologi-
cading trans-effect deregulation on the cells’ regu- cal features in cancer (3). So far, identification of noncoding
latory programs. Our analyses of cis-regulatory mu- cancer driver events at cis-regulatory regions has been lim-
tations associated with miRNAs recurrently predict ited to few examples with high recurrence or high functional
12 mature miRNAs (derived from 7 precursors) as- impact (3–7). Based on mutation recurrence along the hu-
sociated with the deregulation of their target gene man genome, the Pan-Cancer Analysis of Whole Genomes
(PCAWG) consortium reported that patients harbor an av-
networks. The predictions are enriched for cancer-
erage of ∼4.6 driver events in their tumors. The PCAWG
associated protein-coding and miRNA genes and consortium estimated that driver point mutations in non-
highlight cis-regulatory mutations associated with coding regions (∼1.2 per patient) were less frequent than
the dysregulation of key pathways associated with driver point mutations in protein-coding genes (∼2.6 per
carcinogenesis. By combining transcriptional and patient) (8). Large-scale discovery of noncoding drivers has
post-transcriptional regulation of gene expression, been hindered by their low level of recurrence, the varying
* To whom correspondence should be addressed. Tel: +47 228 40 561; Email: [email protected]

C The Author(s) 2022. Published by Oxford University Press on behalf of Nucleic Acids Research.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which
permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
12132 Nucleic Acids Research, 2022, Vol. 50, No. 21
target size of functional elements, technical shortcomings, Despite active research on post-transcriptional regula-
and their composite effect with small individual effect size tion and the identification of miRNAs and their targets (35),
on multiple regulatory regions, e.g. slightly altering, but not the understanding of miRNA transcriptional regulation is
obliterating, protein-DNA interactions (4,8). Furthermore, currently limited (30). One obstacle was the lack of precise
while high-impact driver mutations are typically found and identification of pri-miRNA TSSs. The FANTOM5 consor-
reported, medium-impact putative passenger mutations can tium recently took advantage of the cap analysis of gene ex-
have an aggregated effect on tumorigenesis, beyond the al- pression (CAGE) technology to identify pri-miRNA TSSs
ready annotated driver events (9). genome-wide from different cell types and tissues in human
Gene expression is mainly regulated at the transcrip- and mouse (36). Given their short size and the fact that
tional level by the binding of transcription factors (TFs) to they are not recurrently mutated (8), we hypothesize that
promoters (cis-regulatory regions surrounding genes’ tran- the driver potential of miRNAs in cancer could be triggered
scription start sites, TSSs) and enhancers (cis-regulatory re- by cis-regulatory mutations that alter their expression with
gions distal to genes) at TF binding sites (TFBSs) (10,11). a downstream cascading effect on the gene regulatory pro-
Most of the studies that predict noncoding driver muta- grams of the cancer cells.

tions in cis-regulatory regions rely on the identification of The increasing data accumulation of high-quality direct
mutational hotspots, which are regions with higher muta- TF-DNA interactions (37,38), pri-miRNA TSS locations
tion frequencies than expected by chance (8,12–18). Other (36), somatic cancer mutations and cancer cell expression
studies explore somatic mutations with a potential effect on data (39) provides an unprecedented opportunity to ana-
TF-DNA interactions (19–22), based on DNA sequence in- lyze alterations of gene regulatory programs in cancer by
formation alone, and confirm the potential impact of the combining transcriptional and post-transcriptional levels
predicted mutations on gene expression by in vitro experi- of gene expression regulation. The PCAWG consortium
ments. It has also been attempted to directly combine so- stated that the community is facing a ‘paucity’ in the dis-
matic mutation data with gene expression information to covery of noncoding cancer drivers that could be improved
evaluate the impact of the mutations in cancer samples. For by analyzing larger sample datasets (8). We hypothesize that
instance, causal cis-regulatory variations in breast cancers focusing on regulatory variants within TFBSs associated
have been identified by differential allele-specific expression with protein-coding and miRNA genes combined with gene
of genes between cancer and normal cells (23,24). Muta- expression data has the potential to pinpoint cis-regulatory
tions close to the TSSs of genes were shown to exert an in- variants linked to the dysregulation of key gene regulatory
cis effect on the expression of the corresponding genes (25). networks in cancer patients.
A tool that can be used to associate mutations with changes To this end, we adapted the framework of the xseq tool
in expression in gene networks is xseq (26). The tool was de- to predict cis-regulatory somatic mutations associated with
veloped to predict mutations in protein-coding exons with the dysregulation of gene networks by considering both
trans-effect (26) and it was adapted to consider noncod- protein-coding and miRNA genes. We predict genes asso-
ing mutations associated with protein-coding genes in B cell ciated with cis-regulatory mutations with cascading trans-
lymphomas (27). This methodology specifically assesses the effects on the gene regulatory program alteration across
trans-associations between mutations and gene network ex- seven cancer patient cohorts from The Cancer Genome
pression alteration in cancer samples through either exonic Atlas (TCGA) (39). This analysis reveals 12 mature miR-
or cis-regulatory mutations linked to protein-coding genes NAs recurrently associated with cis-regulatory somatic mu-
(26,27). tations in different cohorts. Functional enrichment analy-
At the post-transcriptional level, one way to further con- ses of the dysregulated networks downstream of the pre-
trol gene expression is through miRNAs acting as ‘buffers’ dicted protein-coding and miRNA genes confirm that path-
to induce translational repression and mRNA degradation ways known to be associated with carcinogenesis are re-
(28,29). miRNA biogenesis generally occurs in mammals currently disrupted. We conclude that the interpretation of
in three steps: transcription of a primary transcript (pri- noncoding mutations can be improved by focusing on TF-
miRNA) that can be several kilobases long, cleavage of DNA interactions with the combined analysis of both tran-
the pri-miRNA into a precursor (pre-miRNA) of ∼70bp, scriptional and post-transcriptional regulation of gene ex-
and cleavage of the precursor to produce mature miRNAs pression to revert the paucity in the discovery of cancer-
of ∼22 bp (29,30). The mature miRNA sequence is then associated noncoding events.
loaded in the RNA-induced silencing complex to specifi-
cally target mRNAs for repression through base-pair com-
MATERIALS AND METHODS
plementarity at the 3’UTR of mRNA targets. A miRNA se-
quence is predicted to target tens to thousands of mRNAs All analyses were performed using the hg19 human genome
(31). The miRNA-mediated regulation of mRNA transla- assembly. When data were obtained from another human
tion is not an on/off system but rather an interplay be- genome assembly, coordinates were converted to the hg19
tween miRNA-binding site specificity, and miRNA and assembly using the liftOver tool provided by the UCSC
mRNA abundance (28,32). Therefore, even small changes in Genome Browser (40,41).
miRNA abundance may affect the expression of several di-
rect targets but also other mRNAs through a cascading ef-
Cancer patient data
fect, potentially leading to dysregulation patterns observed
in cancer. This observation, amongst others, suggests that We considered TCGA (39) cohort samples for which trios
miRNAs can act as cancer drivers (33,34). of (i) whole genome somatic mutations, (ii) RNA-seq, and
Nucleic Acids Research, 2022, Vol. 50, No. 21 12133
(iii) small RNA-seq data were available with at least 30 pa- Mutation rate analysis
tients per cohort. Data were downloaded from the Inter-
For each sample, we calculated the mutation rates by divid-
national Cancer Genome Consortium (ICGC) portal (42)
ing the number of mutated nucleotides within a set of re-
through the icgc-get client (Additional file 5). Altogether, we
gions (TFBSs, exons, and flanking regions) by the number
collected data for 349 samples from seven TCGA patient co-
of nucleotides covered by the given set of regions. TFBS ge-
horts (35–89 donors per cohort; Additional file 1): BRCA-
nomic positions were obtained from UniBind (38) (see be-
US (breast invasive carcinoma), HNSC-US (head and neck
low). Protein-coding exon coordinates were retrieved from
squamous cell carcinoma), LIHC-US (liver hepatocellular
RefSeq Curated (51) (Additional file 5). Flanking regions
carcinoma), LUAD-US (lung adenocarcinoma), LUSC-US
were computed by (i) extending TFBS or exonic regions by
(lung squamous cell carcinoma), STAD-US (stomach ade-
100, 500 and 1000 nucleotides on both sides using the flank
nocarcinoma), and UCEC-US (uterine corpus endometrial
bedtools subcommand and (ii) removing regions overlap-
carcinoma).
ping TFBSs and exonic regions using the subtract bedtools
We retrieved data from 256 samples collected by the
subcommand. Sets of regions were independently merged
ICGC Breast Cancer Working group (43,44) for which

using the merge subcommand of the bedtools (52).
trios of whole genome somatic mutations, RNA-seq, and
Random expectations for mutation rates were computed
miRNA microarray data were available (Additional file
using 150 random sets of somatic mutations and applying
4). miRNA expression was measured using the Human
the mutation rate computation described above. The ran-
miRNA Microarray Slide (Release 19.0) with Design ID
dom sets of mutations were generated by shuffling the orig-
046064 (Agilent Technologies, Santa Clara, CA, USA; see
inal coordinates within the same chromosomes using the
(43) for details).
shuffle subcommand of the bedtools with the -chrom option.
Somatic single nucleotide variants (SNVs) and small in-
miRNAs
sertions and deletions (indels) called by MuSE (45) were
Genomic coordinates of human pre-miRNAs were re-
retrieved from the ICGC portal for TCGA samples. For
trieved from miRBase v20 (53) and used to predict miRNA
ICGC samples, we retrieved SNVs and indels called by the
TSSs from CAGE data by the FANTOM5 consortium
tools CaVEMan (46) and Pindel (47), respectively, used in
(36). When miRNA names in the miRNA-related files (ex-
the original study (43).
pression, survival, cancer-associated miRNAs) used in this
study were mapped to older versions of miRBase (start-
RNA-seq and small RNA-seq normalization ing from version v10), we updated the names according
to the miRBase version (v22) using the miRBaseConverter
Both RNA-seq and small RNA-seq raw counts were filtered
R/Bioconductor package (54).
to remove all genes with 0 reads in >50% of the samples
for a given cohort. For each cohort, both matrices (RNA-
seq and small RNA-seq) of raw counts were normalized Transcription factor binding sites
to counts per million (cpm) using the cpm function from
the R package edgeR (48) and the cpm values were scaled TFBSs were retrieved from the UniBind database (2019 ver-
by log2 conversion. To avoid zeros, we added a pseudo- sion) at https://unibind.uio.no (38) (Additional file 5). The
count of 1. Note that small RNA-seq reads were mapped to TFBSs correspond to high confidence direct TF-DNA in-
pre-miRNA coordinates by TCGA, providing information teractions with both experimental (through ChIP-seq) and
about pre-miRNA expression and not mature miRNAs. computational (through position weight matrices (PWMs)
The normalized microarray miRNA expression matrix from JASPAR (55)) evidence (37,38). Indeed, these TFBSs
for ICGC samples was retrieved from the original study were predicted with high PWM scores and proximity to
where normalization was performed using the 90th per- ChIP-seq peak summits and were derived from 1983 ChIP-
centile methodology (43). We used the normalized RNA- seq experiments for 231 TFs across 315 cell types and tissues
seq matrix provided by the ICGC Breast Cancer Working (38).
Group (43).
TFBS-gene association
Copy number alteration computation
We used the cis-regulatory element-gene associations from
We downloaded copy number alteration (CNA) values pre- the GeneHancer database (v4.9), derived from eight sources
dicted using the GISTIC2 tool (49) for TCGA samples to associate TFBSs to genes (Additional file 5; Supplemen-
through the Firebrowse database at http://firebrowse.org tary Figure S8) (56). TFBSs overlapping a cis-regulatory el-
(Additional file 5). ICGC CNA estimates were computed ement annotated in GeneHancer were associated with the
using ASCAT (v2.1.1) (50) and converted into GISTIC for- corresponding gene in GeneHancer. TFBSs not overlap-
mat with -2 for homozygous loss (nMinor + nMajor = 0), –1 ping annotated elements were associated with the closest
for hemizygous loss (nMinor + nMajor = 1), 0 for normal TSS (for a protein-coding or a miRNA gene). We consid-
(nMinor + nMajor = 2), 1 for three copies (nMinor + nMa- ered TSSs associated with protein-coding genes from Ref-
jor = 3), and 2 for more than three copies (nMinor + nMa- Seq Curated (51) and TSSs associated with miRNAs by
jor > 3). The CNA values assigned to the protein-coding FANTOM5 (36). With this approach, about half of the TF-
genes were used in the xseq analysis to remove cis-effects BSs were associated with protein-coding or miRNA genes
of CNAs on the gene expression dysregulation assessment using GeneHancer associations and the other half with the
(26). closest TSS.
TFBS mutations xseq analyses

Somatic mutations were intersected with TFBS locations The likely associations between mutations and dysregula-
using the intersect subcommand of bedtools v2.25.0 (52). tion of protein-coding gene or miRNA target networks were
All mutations in TFBSs associated with miRNAs were con- calculated with xseq (26). This method requires the follow-
sidered for the xseq analysis (see below). For mutations in ing as input: a gene expression matrix of samples (RNA-
TFBSs associated with protein-coding genes, we followed seq matrix), a binary sample-gene mutation matrix (a ma-
the approach previously used by Mathelier et al. for the xseq trix indicating that a particular gene in a given sample is as-
analysis (27). Specifically, we restricted the analysis to mu- sociated with a mutation), and a weighted network of con-
tations associated with genes potentially dysregulated in the nected genes. Taking advantage of the gene expression in-
corresponding samples. Following (27), genes were consid- formation, the method identifies genes in the sample-gene
ered as potentially dysregulated in a given sample in cohort matrix whose biological partners (from the biological net-
C if its expression value v satisfied v < ␮-1␴ or v > ␮+1␴ work) have expressions that deviate from neutral. This is
(i.e. z-value< -1 or z-value > 1) where ␮ and ␴ correspond computed by decomposing the expression distribution of

to the mean and standard deviation of the expression values each connected gene into three components (or regulatory
of the gene in C. status, namely, downregulation, neutral, and upregulation).
Enrichment of both upregulated and downregulated genes
Loss-of-function mutations within a set of biological partners is evaluated in individual
samples and then across a cohort using a Bayesian hierar-
Following Ding et al. (26) for protein-coding exonic regions, chical network. xseq outputs posterior probabilities associ-
we considered only LoF mutations that are either (i) non- ated with: (i) a sample-specific gene regulatory status (GRS,
sense mutations (disruptive in-frame deletion, disruptive in- the probability of a given gene being dysregulated in a sam-
frame insertion, stop gained, start lost, stop lost, and stop ple) for each gene connected to the gene associated with a
retained variant), (ii) frameshift mutations (frameshift vari- mutation in a given sample, (ii) a sample-specific dysregu-
ant, initiator codon variant) or (iii) splice-site mutations lation probability (SSD, the probability that a mutation in
(splice region variant, splice donor variant, splice acceptor a given gene in a given sample is associated with dysregula-
variant). The analysis was performed using somatic mutation of the gene’s network) and (iii) a dysregulation across
tion data obtained from whole exon sequencing in the same the cohort probability (DAC, the probability that mutations
TCGA samples as for the other analyses. linked to a gene are associated with the dysregulation of its
network across patients) (Supplementary Figure S9). In a
Protein-coding gene networks first step, we removed lowly expressed genes in a cohort fol-
Protein-coding gene networks were retrieved from (26) and lowing the approach described by Ding et al. (26). Briefly,
were composed of 898,032 interactions. Briefly, the net- xseq considers the 90th percentile of expression for each
works were constructed by combining gene associations gene and decomposes the distribution of these values into
from STRING v9.1 functional protein association (57), two Gaussian distributions corresponding to low and high
KEGG pathway datasets (58), WikiPathway (59) and Bio- expression values. We considered for further analysis the
Cyc (60) as integrated into the IntPath database (61), and genes for which their 90th percentile of expression values
TF-target links from ENCODE (62) (see (26) for more de- lie within the high expression distribution with a posterior
tails). We updated the weights of the connections when- probability ≥0.8 (see Ding et al. (26) for details). Next, xseq
ever possible using the methods provided in xseq, following was used to compute all the posterior probabilities to pre-
the methodology described in (26). Specifically, the original dict genes and cis-regulatory mutations in the cancer patient
weight between a given gene g and a biological partner gene cohorts.
p was updated to 1 if p was found differentially expressed
(Benjamini–Hochberg adjusted P-value ≤ 0.05) in samples Selection of predicted genes
where g is mutated in the same cohort (see Materials and
We considered cohort-specific FDR computation to predict
methods in (26) for details). If there existed such genes p,
miRNAs and protein-coding genes. Specifically, we gener-
then only these genes were kept connected to g. Original
ated a set of 100 random controls for each cohort where
weights were kept otherwise.
the original network and the gene-sample association ta-
bles were shuffled; the RNA-seq matrix was not shuffled.
miRNA–target networks
For the biological networks, we kept the original num-
miRNAs were associated with potential target protein- ber of edges, but both the target genes and their connec-
coding genes using predictions from TargetScan v7.2 (31). tion weights were shuffled. xseq was applied to each ran-
From the list of targets for each miRNA, we filtered out dom control independently and the results of the 100 con-
the targets with less than two predicted binding sites for trols were aggregated to compute the threshold t on DAC
the given miRNA to reduce false positives (63,64). miRNA- that corresponds to an FDR of 0.05. If the corresponding
target weights were computed as t score / 100 where t score threshold on the DAC cohort-specific posterior probabil-
corresponds to the targetScan context++ score percentiles ity was <0.5, we chose 0.5 as the threshold. We consid-
from TargetScan. We updated the weights of the connec- ered genes with DAC above this threshold and SSD ≥0.5
tions whenever possible following the same strategy as for in at least two samples as potential cancer-associated
protein-coding genes (see above). genes.
Dysregulation heatmaps Results accessibility

The dysregulated networks for predicted protein-coding The analysis with all the scripts and parameters can
and miRNA genes are visualized as heatmaps where be found through the following link: https://bitbucket.
columns correspond to mutated samples and rows to con- org/CBGR/workspace/projects/DYS. We provide (i) the
nected genes. Heatmaps were constructed with connected source code for the analysis at https://bitbucket.org/CBGR/
genes dysregulated (GRS ≥ 0.5) in at least one sample dysmir manuscript/src/master/ and (ii) a pipeline for users
with SSD ≥0.5. These genes are referred to as dysregulated to run similar analyses with their data at https://bitbucket.
genes. org/CBGR/dysmir pipeline/src/master/.
Aggregated and sample-specific networks RESULTS

To evaluate whether the protein-coding genes predicted by Transcription factor binding sites harbor a similar mutational

cis-regulatory mutations are connected in the filtered net- load as protein-coding exons
works (see Protein coding gene networks section), we built
an aggregate network using all the predicted protein-coding We study the occurrence of somatic mutations from whole
genes within a cohort. We counted the number of clusters genome sequencing of 349 samples from seven cancer pa-
using the R packages igraph (65) and ggnetwork (66). Sim- tient cohorts (35–89 samples per cohort) covering seven dis-
ilarly, we built sample-specific networks and counted the tinct cancer types from TCGA (39) (Additional files 1–2).
number of clusters in each sample, only considering the pre- Specifically, we select samples where trios of somatic muta-
dicted genes with DAC ≥t (with t being the threshold on the tions, RNA-seq, and small RNA-seq data are available. In
DAC that corresponds to an FDR of 0.05, see above) and aggregate, we examine 11 434 931 somatic single nucleotide
SSD ≥0.5. variants and small insertions and deletions (from 2832 to 1
014 969 per sample; Additional file 2; Supplementary Fig-
ure S1).
Functional enrichment analysis To identify cancer-associated cis-regulatory mutations,
we consider a set of TFBSs predicted as direct TF-DNA in-
Given a list of dysregulated genes, functional en- teractions in the human genome and stored in the UniBind
richment analyses were performed using the R database (38). These TFBS predictions are supported by
package enrichR (67) for the following databases: both experimental (based on ChIP-seq) and computational
KEGG 2021 Human, WikiPathways 2019 Human, evidence (based on position weight matrices) of direct TF–
GO Biological Process 2021 and Panther 2016. DNA interactions (see Materials and Methods and refer-
ences (37,38) for details). We first assess whether this set of
TFBSs represents regions of functional interest similar to
Enrichment for cancer-associated genes and TFs the coding portion of the human genome commonly stud-
Given a set of genes, we assessed their enrichment for ied to predict cancer-associated mutations. These TFBSs
cancer-associated genes or TFs using hypergeometric tests cover ∼2.2% (68 071 257 nt) of the human genome, close
using the stats::phyper function in R. The list of cancer to the exonic coverage of protein-coding genes (∼2.6%; 81
protein-coding genes considered was constructed by con- 416 464 nt). Focusing on the somatic mutations, we observe
sidering genes that appear in at least two of the following that 1–2% of the mutations in each sample lie within these
databases: Network Cancer Gene (68), inToGen (69) and TFBSs (median of 277 mutations per sample; Additional
Cancer Gene Census (70). Cancer miRNA genes were re- file 2; Supplementary Figure S2). Mutation rates in TFBSs
trieved from miRCancer (71) with data from 1 May 2019. vary between cancer cohorts but are similar to the mutation
TF genes were retrieved from the human transcription fac- rates observed in exons (two-tailed Wilcoxon tests P-values
tor database (11). between 0.13 and 0.96; Figure 1 and Supplementary Fig-
ures S3 and S4). TFBSs are less mutated than their flanking
regions (Figure 1). Note that regions of 1 kb surrounding
TFBSs harbor mutation rates similar to what is expected
Survival analysis
by chance (two-tailed Wilcoxon tests P-values between 0.56
To test whether miRNA expression was associated with sur- and 0.95; Supplementary Figures S3 and S4). While exons
vival, we used the METABRIC breast cancer cohort (72) exhibit mutation rates similar to those observed within TF-
with miRNA microarray expression (73) available for 1282 BSs (Supplementary Figure S5), their flanking regions show
tumors. Expression values were downloaded from the Eu- a smaller increase in mutation rates than the increase de-
ropean Genome-Phenome Archive, www.ebi.ac.uk/ega, ac- tected in the vicinity of TFBSs (Supplementary Figures S6
cession number EGAS00000000122. Follow-up data were and S7).
available from Curtis et al. (72). Kaplan–Meier survival Taken together, these results confirm that the noncoding
analyses and log-rank tests were performed using the R mutation frequencies in the studied set of TFBSs follow a
package survival with tumors separated into ‘high’ or ‘low’ similar pattern to what is observed in protein-coding exons.
miRNA expression groups depending on expression val- It provides a posteriori confirmation that the set of TFBSs
ues above or below the median. P-values were adjusted we consider is likely composed of functional regions in the
for multiple testing according to the Benjamini–Hochberg human genome and can be used to highlight cis-regulatory
method. mutations of functional interest in cancer genomes.
BRCA−US LIHC−US UCEC−US HNSC−US
−4
−5
Log10 mutation rate
−6
BS
BS
BS
BS
BS
nic
ing
ing
ng
n
STAD−US LUAD−US LUSC−US
Exo
i
Exo
TF
_TF
TF
TF
TF
cod
cod
d
_co
m_
m_
nk_
nk_
nk_
nk_
ank
ank
ndo
ndo
Fla
Fla
Fla
Fla
_Fl
−4
_Fl
Ra
nt_
nt_
Ra

nt_
nt_
0nt
0nt
500
100
100
500
100
100
−5
−6
BS
nic
ng
ng
nic
ing
ng
g
BS
BS
BS
BS
BS
BS
BS
BS
BS
BS
nic
ng
ing
g
BS
BS
BS
BS
n
din
Exo
din
Exo
din
Exo
i
odi
odi
odi
Exo
Exo
Exo
TF
_TF
TF
TF
TF
cod
TF
_TF
TF
TF
TF
cod
TF
_TF
TF
TF
TF
cod
_co
_co
_co
m_
m_
m_
k_c
k_c
k_c
m_
nk_
nk_
m_
nk_
nk_
m_
nk_
nk_
nk_
nk_
nk_
ank
ank
ank
ank
ank
ank
ndo
ndo
ndo
ndo
ndo
ndo
lan
lan
n
Fla
Fla
Fla
Fla
Fla
Fla
Fla
Fla
Fla
Fla
_Fl
_Fl
_Fl
_Fl
_Fl
_Fl
F
Ra
Ra
Ra
nt_
nt_
nt_
nt_
nt_
nt_
Ra
Ra
Ra
nt_
nt_
nt_
nt_
nt_
nt_
0nt
0nt
0nt
0nt
0nt
0nt
500
100
500
100
500
100
100
500
100
500
100
500
100
100
100
100
100
100
Figure 1. Comparison of mutation rates in TFBSs and exons versus their flanking regions and random mutation rates. Each panel corresponds to a specific
cancer cohort (see title boxes) and each point corresponds to a sample. On each panel, the two central boxplots (shadowed) represent mutation rates in
TFBS and exonic regions, the remaining box plots correspond to mutation rates in increasing-size flanking regions (100, 500 and 1000 nt) and mutation
rates expected by chance (150 randomly distributed sets of mutations in the genome; Material and Methods).
Cis-regulatory and loss-of-function mutations complemen- hort (Supplementary Figure S9). Genes with low expres-
tarily alter protein-coding gene networks sion in a given cohort were filtered out; the distribution of
the 90th percentile of expression for genes was decomposed
We then seek to predict the cis-regulatory mutations that
into two Gaussian distributions corresponding to low and
lie in these TFBSs and that lead to cascading effects on
high expression values and only genes lying in the high ex-
gene network deregulation, a hallmark of carcinogenesis.
pression distribution were retained (Materials and Meth-
We first focus on the mutations in TFBSs linked to protein-
ods). Furthermore, gene expression is corrected for copy
coding genes and compare their effect on gene regulation
number alterations (amplifications and deletions detected
to that of mutations altering the function of the protein-
by GISTIC2 (49)) to compensate for copy number-related
coding genes. We consider a protein-coding gene to be mu-
cis-effects on expression (Material and Methods). LoF mu-
tated through either a loss-of-function (LoF) somatic mu-
tations and mutations that overlap TFBSs are analyzed
tation in one of its exons as in (26) or a somatic mutation
independently. Finally, we consider predictions that sat-
overlapping a TFBS associated with the gene. TFBSs are
isfy a false discovery rate (FDR) <0.05, computed empiri-
linked to protein-coding or miRNA genes based on cis-
cally for each cohort using random controls (Materials and
regulatory element-to-gene associations from GeneHancer
Methods).
(56) or distances to TSSs (Materials and Methods; Supple-
Out of the 7275 unique protein-coding genes linked to so-
mentary Figure S8). We estimate the potential trans-effect
matic mutations in the seven TCGA cohorts, 237 are associ-
of the mutations on expression disruption in protein-coding
ated with the deregulation of transcriptional networks in at
gene networks using the xseq tool, following approaches
least one cohort. Of these, 21 harbor LoF mutations (TP53
implemented in previous studies (26,27). Specifically, the
and RPL22 are predicted with LoF mutations in three and
method uses a hierarchical bayesian approach to associate
two cohorts, respectively; Figure 2A) and 219 are linked
mutations with expression dysregulation in biological net-
to cis-regulatory mutations associated to transcriptional
works associated with the mutated protein-coding genes.
deregulation (24 genes are found in more than one cohort;
In a nutshell, it assesses the posterior probability of the
Figure 2A, Supplementary Figures S10 and S11, and Ad-
likely association between observing mutations in a set of
ditional File 3). Three genes are linked to dysregulated net-
patients and observed deviations from neutral expression
works in association with both LoF and cis-regulatory mu-
in these samples for protein-coding genes in the same net-
tations but in different patients and/or cohorts: ACVR2A,
work. The likely trans-associations between mutations and
ARID1A and GATA3 (Figure 2A). These three genes are
gene network deregulation are first assessed in a sample-
already known cancer drivers that we predict to be im-
specific manner and then across samples from the same co-

Figure 2. Pan-cancer predicted protein-coding genes. (A) Predictions are obtained applying the xseq tool when considering the effect on gene deregulation of
protein-coding genes mutated through either LoF (red triangles) or cis-regulatory (TFBS; green triangles) mutations, independently. Genes, with mutations
predicted to affect gene regulation in at least two cohorts are depicted here. Enrichment for cancer-associated genes (red stars) and TFs (blue stars) are
evaluated using hypergeometric tests (p-values provided in the legend; Material and Methods). (B) Samples where genes are predicted with cis-regulatory
mutations are considered for each cohort and assessed for the presence of LoF mutations in the same genes for the same cohort (TFBS & Exon) or no LoF
mutation in the corresponding gene (TFBS only).
pacted by alternative mutational mechanisms (LoF or cis- cading trans-effect in gene network dysregulation but the
regulatory mutations). The remaining genes are either as- method cannot identify the specific main driver event or the
sociated with LoF or cis-regulatory mutations across co- combination of cis-regulatory mutations. When considering
horts (TP53, RPL22 with LoF mutations; e.g. PIK3C3 and all the predicted genes per cohort, we detect a similar pat-
CHRM3 with cis-regulatory mutations; Figure 2). tern with subnetworks of interconnected genes with a max-
From the combined list of 237 predicted protein-coding imum of 12 subgraphs containing at least two nodes per co-
genes (Additional File 3), 81 are already annotated as hort (mean = 3; median = 4.13; Figure 3B and Supplemen-
cancer-associated genes (P-value = 9.3e–17; hypergeomet- tary Figure S13). Altogether, these interconnected subnet-
ric test) and 29 as TFs (P-value = 0.025; Supplementary works suggest that the predicted genes are likely involved in
Figures S10 and S11). We observe 28 genes to be predicted similar biological pathways with altered expression associ-
in at least two cohorts. These 28 genes are enriched for al- ated with cis-regulatory somatic mutations.
ready known cancer-associated genes (P-value = 1.4e–6; hy-
pergeometric test) but not for TFs (P-value = 0.21; hyper-
Deregulation of transcriptional activity and cancer pathways
geometric test) (Figure 2A).
are trans-effect signatures of the predicted cis-regulatory and
The genes predicted through cis-regulatory mutations
loss-of-function mutations
rarely contained LoF mutation in the same tumors (Fig-
ure 2B and Supplementary Figure S12). We interpret this To shed light on the functional role of the somatic muta-
to mean that LoF and cis-regulatory mutations are possibly tions predicted to be associated with a cascading effect, we
complementary mechanisms that alter the gene regulatory perform enrichment analyses on the altered gene expres-
programs of cancer cells. We observe that multiple genes can sion profiles. One advantage of xseq is its capacity to high-
be predicted through cis-regulatory mutations in the same light the specific genes in the biological networks that are
sample. Furthermore, these genes tend to be interconnected dysregulated in the samples harboring the somatic muta-
in the dysregulated genes’ networks (Figure 3A). All these tions considered (Material and Methods) (26). These genes
genes are predicted through mutations associated with cas- are consistently found to be either up- or down-regulated

Figure 3. Predicted protein-coding genes linked to cis-regulatory mutations are connected in biological networks. (A) Subgraph detected among the pre-
dicted protein-coding genes in the Uterine Corpus Endometrial Carcinoma (UCEC-US) cohort. The number of samples in which each gene is predicted
is shown within the nodes. TF names are highlighted with an orange background. (B) Heatmap showing the number of subgraphs (y-axis) found among
the predicted protein-coding genes linked to cis-mutations in the TCGA cohorts (x-axis). The number of nodes within a subgraph is indicated in each cell.
Genes not connected to any other predicted gene are not shown.
in the samples with predicted disrupted expression (see the Combining transcriptional and post-transcriptional regula-
blue and red colors in the upper and lower clusters in Fig- tion highlights pan-cancer miRNAs associated with gene ex-
ure 4A). These results highlight sets of genes up- or down- pression alteration in tumors
regulated across samples where cancer-associated genes are
The analysis of mutations linked to protein-coding genes
predicted.
presented above demonstrates that our methodology pin-
We assess the biological relevance of the networks pre-
points cis-regulatory mutations likely associated with car-
dicted to be dysregulated in association with either LoF or
cinogenesis. We hypothesize that our method could high-
cis-regulatory mutations linked to the protein-coding genes.
light cis-regulatory mutations linked to miRNAs with
Functional enrichment analysis is performed using path-
downstream cascading effects on the gene regulatory pro-
ways from KEGG (58), WikiPathways (59) and Panther
grams of the cells because miRNAs are involved in post-
(74), and gene ontology biological processes (GO BP (75))
transcriptional regulation of gene expression. This novel ap-
with the EnrichR tool (67). The dysregulated genes in the
proach of functional analysis of mutations aims to combine
networks are enriched for transcriptional activity (‘regula-
transcriptional (through mutations in TFBSs) and post-
tion of transcription, DNA-templated’ from GO BP; Sup-
transcriptional (through regulatory networks of miRNA–
plementary Figure S14). Combined with the enrichment
targets) regulation to predict miRNAs associated with a
of TFs in the complete list of predicted cancer-associated
trans-effect on gene expression alteration through somatic
genes, this result emphasizes that the alteration of tran-
mutations in cis-regulatory elements.
scriptional regulation is likely a common feature of can-
Specifically, we adapt the xseq framework to infer cis-
cer cells throughout cancer types. Pathways already known
regulatory somatic mutations linked to miRNAs and as-
to be associated with carcinogenesis (e.g. ‘Pathways in can-
sociated with a cascading effect on miRNA target net-
cer’, ‘JAK-STAT signaling’, ‘PI3K-Akt signaling’, ‘p53 sig-
works dysregulation. Similar to the analysis of protein-
naling pathway’, ‘Focal adhesion’ and ‘Apoptosis’; Figures
coding genes, we estimate the posterior probability of the
4B, C and Supplementary Figures S14–S17) are at the top
likely association between the presence of mutations in TF-
of the enriched terms. The enrichment for cancer pathways
BSs linked to a miRNA with observed deviations from neu-
confirms that our approach identifies somatic exonic and
tral expression of the miRNA’s target genes. We consider
cis-regulatory mutations associated with potential protein-
miRNAs from miRBase (53) and their corresponding TSSs,
coding cancer-associated genes with cascading effect on reg-
which were identified using CAGE (Materials and Meth-
ulatory alteration of key cancer-related pathways. Our re-
ods) (36). To assess the cascading effect of mutations linked
sults suggest that alteration of gene network expression
to miRNAs on their targets’ expression, we examined the
could be achieved through cis-regulatory mutations asso-
protein-coding genes predicted by TargetScan (31) to be tar-
ciated with different genes in different patients but involved
gets of each miRNA. We limited the set of miRNA–target
in the same pathways.

B C
Figure 4. Dysregulated protein-coding gene networks and functional enrichment analysis. (A) Dysregulated gene network in samples where FUS
is predicted through cis-regulatory mutations in breast cancer (BRCA-US) (rows: dysregulated genes associated with FUS; columns: samples with
FUS-associated cis-regulatory mutations). The color scale represents the gene regulatory status posterior probability (red: up-regulation; blue: down-
regulation––posterior probability * (–1)). The top horizontal bar shows the sample-specific dysregulation posterior probability computed by xseq for the
samples harboring a cis-regulatory mutation in the FUS gene. The horizontal bar below shows the gene expression z-value of FUS (Materials and Meth-
ods). (B) KEGG 2021 most enriched terms computed from all the dysregulated genes associated with the predicted protein-coding genes (A is one example
for FUS) by xseq with LoF mutations and (C) cis-regulatory mutations in TCGA cohorts (columns). Terms (rows) are ordered by their mean rank across
all cohorts. Significance is provided as –log10 (P-value).
genes pairs to those where at least two target sites for the 5C), arguing for a potential link between viral infections
miRNA are predicted to reduce false positive predictions and cancer initiation/progression, as previously suggested
(63,64) (Materials and Methods). Note that we separately (89,90), via miRNAs.
analyze miRNAs from both arms (5p and 3p) for each pre- Altogether, this study provides the first foray into the
miRNA sufficiently expressed in a TCGA cohort (Materials analysis of a combined effect of coherent transcriptional
and Methods). and post-transcriptional dysregulation downstream of so-
Applying this analysis to the seven TCGA cohorts, we matic cis-regulatory mutations associated with miRNAs in
predict 68 mature miRNAs, derived from 47 pre-miRNAs, cancer cells. It highlights a core set of miRNAs associated
as associated with mutations in TFBSs and deregulation of with cis-regulatory mutations that are linked to a cascading
expression for their target genes (Figure 5A and Supple- alteration of gene regulatory networks involved in cancer
mentary Figure S18; Additional File 3). From these 68 miR- onset and progression.
NAs, 54 are already annotated as cancer-associated miR-
NAs in the miRCancer database (71) (P-value = 5e–23;
Complementary analysis of an independent breast cancer co-
hypergeometric test), which is derived from text-mining of

hort supports dysregulation of specific pathways
the scientific literature in PubMed (76). Moreover, miR-
Cancer provides information about the cancer types that Further, we aim to validate the recurrence of the predic-
are associated with miRNAs in the literature; ∼27% cancer- tions for breast cancer obtained from the 92 samples of the
associated miRNAs we predict are supported by the litera- BRCA-US cohort from TCGA in a complementary cohort.
ture to be involved in the same cancer type as the cohort We apply the same methodology with the same parameters
from which they were identified (P-value = 3.97e–14; hy- to the ICGC breast cancer cohort (43), which is composed
pergeometric test). of 256 breast cancer samples with the same trio of data types
Among these, we identify a core set of 12 mature miR- available (WGS, RNA-seq, and miRNA expression - from
NAs (derived from 7 pre-miRNAs) that are identified in at microarrays; Additional file 4).
least four out of the seven cohorts (Figure 5A and Sup- Similar to the BRCA-US analysis on protein-coding
plementary Figure S18): hsa-miR-20a-3p, hsa-miR-92a-1- genes, our analysis of the ICGC cohort predicts known can-
5p (predicted in all seven cohorts), hsa-miR-18a-5p (six co- cer drivers identified by associating LoF or cis-regulatory
horts), hsa-miR-20a-5p, hsa-miR-18a-3p, hsa-miR-17-5p, mutations with dysregulation of their respective gene net-
hsa-miR-17-3p, hsa-miR-155-5p (five cohorts), hsa-miR- works. Breast cancers can be categorized into estrogen re-
155-3p, hsa-miR-708-3p, hsa-miR-708-5p and hsa-miR- ceptor positive (ER+) and negative (ER–), each subtype
205-5p (four cohorts). All these miRNAs are derived from harboring a distinctive signature of gene expression with
precursors of already established oncomiRs or tumor sup- prognostic and predictive impact. We explore how the dis-
pressor miRNAs, or are known to be involved in immune tribution of ER status in patients from the two cohorts
response or inflammation (77–88). Note that hsa-miR-17- can impact the predictions of cancer-associated genes. The
3p, hsa-miR-17-5p, hsa-miR-18a-5p, hsa-miR-18a-3p, hsa- BRCA-US cohort is composed of approximately the same
miR-20a-3p, hsa-miR-20a-5p and hsa-miR-92a-1-5p are number of ER+ and ER– patients while the ICGC cohort
part of a single miRNA cluster on chromosome 13 and this is composed of 72% of ER+ patients. Given the size of
polycistronic cluster (known as miR-17-92) is well known the ICGC cohort (256 samples), it is possible to perform
to be composed of oncomiRs involved in proliferation and two additional analyses on ER+ (184 samples) and ER–
tumor angiogenesis as well as reducing apoptosis of cancer samples (72 samples) independently. The analysis of cis-
cells (77). regulatory mutations associated with protein-coding genes
When visualizing the dysregulated networks of miRNA reveals two predictions specifically common to BRCA-US,
targets in samples harboring cis-regulatory alterations as- ICGC, and ICGC ER+ cohorts (IL12RB1 and TOP1), one
sociated with the predicted cancer-associated miRNAs, we specifically common to BRCA-US and ICGC (B4GALT3),
detect subsets of the networks as up- or down-regulated one specifically common to BRCA-US and ICGC ER+
across patients from the same cohort (Figure 5B). The func- (CTSS), and three common to ICGC and ICGC ER–
tional pathways are similar to those detected with protein- (MEF2A, RB1 and RGS1) (Supplementary Figure S20).
coding gene networks (Figure 4B-C and 5C). Note that the Out of these seven genes, four are known cancer-associated
miRNA-target networks observed with altered expression genes (B4GALT3, CTSS, RB1 and TOP1). Despite this
for a given miRNA may vary between cohorts for the same small intersection, the functional enrichment analyses of
miRNA because some targets are specifically expressed or the dysregulated genes associated with all predicted genes
altered in a subset of tissues or cell types (Supplementary are similar in the cohorts (Supplementary Figure S21), sug-
Figure S19). gesting that although the predictions vary among cohorts
Similar to our previous observations with protein-coding with different etiology, the dysregulated pathways are likely
genes, miRNA targets with altered expression downstream the same. Furthermore, we detect enrichment of similar key
of cis-regulatory mutations are enriched for transcrip- cancer pathways when considering the dysregulated genes
tional activity terms and in biological pathways associated associated with the predicted cancer-associated genes (Sup-
with carcinogenesis (Figure 5C). Furthermore, these net- plementary Figure S21).
works are recurrently found when considering disrupted To confirm whether common pathways are deregulated
target genes in each cohort independently (Supplemen- despite the prediction of different genes, we construct the
tary Figures S14–S17). We discover several virus infection- network of all the predicted genes when considering pa-
related terms enriched across the cohorts (Figures 4B-C and tients from BRCA-US, ICGC, ICGC ER+ and ICGC ER–.
B C
Figure 5. Overview of miRNA driver predictions and their dysregulated target networks. (A) Pre-miRNAs with mature miRNAs predicted as potential
drivers by xseq. Cell colors indicate the posterior probability computed over the corresponding cohort. Red stars indicate that the miRNA is annotated
as a cancer-associated miRNA in miRCancer (71). Blue stars indicate that the miRNA was reported as a cancer-associated miRNA in the specific cancer
type where it is predicted by xseq, according to miRCancer annotation. (B) Dysregulated network of target genes for miRNA hsa-mir-20a-5p predicted in
liver hepatocellular carcinoma (LIHC-US) (rows: dysregulated targets; columns: samples with cis-regulatory mutations associated with hsa-mir-20a-5p).
The top color scale represents the gene regulatory status posterior probability (red: up-regulation; blue: down-regulation - posterior probability * (-1)). The
horizontal bar below shows the miRNA expression z-value (Materials and Methods). (C) KEGG 2021 most enriched terms (rows) for all the dysregulated
genes associated with the identified miRNA drivers across TCGA cohorts (columns). Terms are ordered by their mean rank across all cohorts. Significance
is provided as –log10 (P-value).
Genes are linked in the network if they are known biological ple cis-regulatory mutation information with gene expres-
partners in the original network (Figure 6). The constructed sion data from the same samples to highlight direct evidence
network comprises 87 genes, which are all connected in a of the regulatory impact of the mutations. By integrating
single dense network, where the top three (hub) genes with whole-genome somatic mutations, RNA-seq, small RNA-
the largest in-degree are JUN, RB1 and TP53. This obser- seq, and copy number aberrations (CNA) data with gene
vation highlights that the predicted genes across the cohorts regulatory networks, we perform pan-cancer predictions of
are likely involved in similar biological pathways, which is protein-coding and miRNA genes associated with somatic
supported by the functional enrichment results above. It cis-regulatory mutations in patients from seven distinct can-
suggests that the same pathways tend to be dysregulated cer types. Our study provides a large-scale foray into pre-
through mutations associated with different genes. dicting cancer-associated protein-coding and miRNA genes
We predict one miRNA (hsa-mir-378a-3p) associated by combining both transcriptional and post-transcriptional
with cis-regulatory mutations in the ICGC cohort when information. Our results provide new insights into the po-
considering all samples (Supplementary Figure S22). We tential impacts and causes of the alterations of gene regula-
do not predict any driver miRNAs associated with cis- tory programs observed in cancer cells along with the cas-

regulatory mutations when examining specifically the cading effects on key biological pathways.
ER+ samples. However, we identify hsa-mir-17-3p, hsa- We specifically focus on somatic mutations that reside
mir-17–5p, hsa-mir-18a-5p, hsa-mir-20a-5p, hsa-mir-21– within a high-quality dataset of TFBSs that represent di-
5p, hsa-mir-155–5p, hsa-mir-590-5p, and hsa-mir-629-3p rect TF-DNA interactions, which cover ∼2% of the human
when considering ER- samples. Out of these eight miR- genome, with both experimental and computational evi-
NAs, two are predicted in the BRCA-US cohort (hsa-mir- dence (38). We acknowledge that this set of TFBSs might
17-3p and hsa-mir-18a-5p; Supplementary Figure S22) and represent a limited subset of all potential TFBSs in the
five are recurrently found in at least 5 out of the 7 TCGA human genome as it was derived from experiments avail-
cohorts (Figure 5 and Supplementary Figure S22). As ex- able for a reduced number of TFs and cell types/tissues
pected, these results confirm that the cohort clinicopatho- (231 TFs out of the ∼1600 human TFs reported (11) and
logical composition impacts the predictions as it can impact 315 cell types and tissues). Moreover, some TFBSs might
the landscape gene expression distributions across samples. not be relevant or functional in the cell type of origin as-
Nevertheless, the complementary analyses of the BRCA- sociated with the cancer types studied here. Nevertheless,
US and ICGC breast cancer cohorts exhibit hsa-mir-17-3p we provide evidence that the regions considered are likely
and hsa-mir-18-5p as recurrently predicted breast cancer- enriched for functional genomic elements since they har-
associated miRNAs linked to cis-regulatory mutations and bor mutation rates similar to what is observed in exonic
dysregulation of their target gene networks. Functional en- regions (Figure 1). This observation is complementary to
richment analysis confirms that the dysregulated miRNA other studies that showed similar mutation rates in pro-
target gene networks are enriched for genes involved in tran- moters and enhancers compared to protein-coding exons
scriptional regulation and cancer-relevant pathways such as (12,91) and a negative selection for cancer mutations at TF-
the p38 MAPK signaling, ErbB signaling, and DNA dam- BSs (92). The reduced mutation rates in exons and the lim-
age response (Supplementary Figure S23). ited increase in surrounding regions can be attributed to in-
Finally, we evaluate the clinical potential of the pre- creased mismatch repair and nucleotide excision repair in
dicted breast cancer miRNAs for breast cancer survival exons as previously shown (93,94). The decreased mutation
estimation. For this purpose, we consider a third cohort, rates when considering TFBSs are in line with our previ-
METABRIC (72), which is composed of 1282 samples. We ous observation in B-cell lymphomas (27). Nevertheless, it
compute Kaplan–Meier survival curves and log-rank tests is somewhat in disagreement with previous studies showing
using miRNA expression from the METABRIC cohort for that nucleotide excision repair is impaired by the binding of
the miRNAs predicted as drivers in the BRCA-US and TFs to DNA (95,96). We hypothesize that the differences
ICGC cohorts (for 26 of the predicted miRNAs in breast observed could be partially explained by the fact that (i) our
cancer). Examining both overall survival and breast cancer- mutation rate analysis considered TFBSs predicted from
specific survival values, we observe log-rank test P-values several cell lines and tissues instead of focusing on TFs and
<0.05 for hsa-mir-29a-3p, hsa-mir-20a-5p, and hsa-mir- TFBSs specific to the considered cell types or conditions
20a-3p (Figure 7 and Supplementary Figures S24 and S25). (such as UV-exposure in melanoma) and (ii) we do not fil-
Note that hsa-mir-20a-5p and hsa-mir-20a-3p are recur- ter TFBSs based on open chromatin data from matched cell
rently predicted in at least five out of the seven ICGC co- types.
horts. Taken together, these results reinforce a posteriori the Contrary to previous studies assessing the impact of mu-
biomarker potential of some miRNAs we predicted as their tations on TF-DNA binding affinity or the enrichment for
level of expression could be used for prognosis. mutations in cis-regulatory regions (97–100), we particu-
larly evaluate the impact of cis-regulatory mutations on
expression alteration in gene networks. As such, our ap-
DISCUSSION
proach does not quantify the direct impact of individual
In this study, we explore how cis-regulatory somatic mu- mutations on the obliteration of TF–DNA interactions but
tations at TFBSs can be used to predict genes with a cas- uses RNA information as the ultimate readout. Although
cading trans-effect on gene regulatory network dysregula- other features can be used to highlight variants of interest,
tion. Contrary to most methods that predict cancer-driving it has previously been shown that machine learning meth-
events based on the recurrence of mutations, we seek to cou- ods used to assess the effect of mutations on TF binding
Aggregated network − Nodes: 87 − Subgraphs: 1
POLR1A
PDE4D
ADCY7
RGS2
ADCY9
FCGR3A HGS
PTGER4 RGS1
PSMD3
SNRPE RGS1
PLCB3
LAMB1
ITGA2 CBLB FCER1G GRB2
FUS
ERBB4 CXCR4
ITGA2 SYK PSMD9 HEXIM1
IL1B

CXCR4 PLA2G4A
IL12RB1 CDC23
ICAM1 PTK2B DHX9
SRSF1 CDC23
EIF4B ZFPM1
RPLP2 CALM2
MEF2A MEF2A JUN HNRNPA2B1
SSR3
CDH1 GATA3
MAPK10 PPP1CA
RPL26
PPP2R5A MSH6
KAT5 KNTC1
NCOA3 UBE2I CREB1 CSNK2B
HLA−DRB5 CLSPN
RB1 KIF2C ASF1B
SNW1 CREBBP TP53 JUNB CCNE1
HLA−DRB5
RB1 MCM3 MELK
ETF1 NCOR2 HSP90AA1 ATF3 ASPM
MEF2D
MED18 CENPF MAPRE1
ETS1 SMAD4 NUP98 PRIM1
POLA2 ZWINT
ACD
CREB3L4 POLA2
SENP3 MDM4
FES
FBXO31
BAMBI
ZNF143
ICGC (all) ICGC ER+ ICGC ER− BRCA−US
Indegree
10 20 30
Figure 6. Predicted genes in breast cancer cohorts are connected in the biological network. Network representing the predicted protein-coding genes in
ICGC (all samples), ICGC ER+, ER– and BRCA-US cohorts. The names of the genes predicted in two or more cohorts are displayed several times with
different colors.
affinity poorly predict the effects on expression as reported The analysis of protein-coding genes predicts 28 genes in
by massively parallel assays (101). A previous method sys- at least two (out of the seven) TCGA cohorts analyzed, with
tematically assessed the potential impact of somatic muta- many already known cancer drivers (Figure 2A). We ob-
tions in genomic tiles near genes’ TSSs on gene expression serve that the protein-coding genes predicted through the
(25). Here, we consider mutations lying within a specific set analysis of cis-regulatory mutations generally do not con-
of pre-defined TFBSs without restrictions on distances to tain mutations in exonic regions for the same patients (Fig-
TSSs and evaluate the trans-association of the mutations ure 2B and Supplementary Figure S12). This observation
with genes’ network deregulation. Our approach is some- suggests complementary mechanisms acting upon gene ex-
what similar to a genome-wide association study frame- pression dysregulation with cascading effects on regulatory
work focused on TFBSs to reduce the search space. More- network disruption. We hypothesize that either the final
over, our strategy is not directly assessing the effect on TF- product of a gene may be altered due to LoF mutations or
DNA interactions, i.e. the gain/loss of TFBSs, but rather fo- the expression of the gene is altered through cis-regulatory
cuses on the association with gene expression deregulation. mutations, which, in both cases, alter the activity of biolog-
Although we focused on somatic mutations and small in- ical networks.
dels at cis-regulatory elements, we acknowledge that CNAs Given that miRNAs cover a small portion of the hu-
such as duplications or deletions are likely to contribute to man genome, they harbor a small number of somatic muta-
gene expression alteration as well. Nevertheless, our analy- tions (8), limiting the possibility to affect gene expression.
ses considered CNAs to ensure that the predicted deregula- The potential mechanism that we propose here is the alter-
tions were not confounded with CNAs. Further work and ation of their regulatory elements. Our study highlights cis-
a complementary computational framework will be neces- regulatory mutations linked to miRNAs that are associated
sary to bring together single nucleotide variants, small in- with dysregulation of expression of the miRNA targets. In
dels, CNAs, and structural variations and assess their com- our pan-cancer analysis, we discover a core set of 12 mature
bined impact on gene expression deregulation in cancer. miRNAs associated with the dysregulation of key pathways

Figure 7. Survival curve analysis for some predicted miRNA drivers. Kaplan–Meier survival curves were obtained using the METABRIC cohort for the
most significant driver miRNAs identified in the breast cancer cohorts. Samples were separated into two groups according to the level of miRNA expression
(above/below the median). Log-rank test p-values are indicated. OS: overall survival. BCSS: breast cancer-specific survival.
involved in carcinogenesis. This core set of miRNAs repre- network lead to the same phenotype), which may have orig-
sents a common feature for gene expression dysregulation inated because the dysregulated genes are connected in the
associated with cancer onset or progression. We note that biological network (Figure 6). Moreover, as originally de-
several of these miRNAs are established oncomiRs, which scribed in Ding et al. (26), the xseq probabilistic framework
promote carcinogenesis. The Kaplan–Meier plots in Figure highlights the specific samples where mutations are associ-
7 for hsa-mir-29a-3p, hsa-mir20a-3p, hsa-mir-20a-5p, and ated with an impact on gene expression (Figure 4A). This
hsa-mir-145-3p show that higher expression correlates with dichotomy can, in principle, be used to stratify samples and
poorer survival rates, which would indicate that these miR- mutations but, in this study, is limited by the number of sam-
NAs act as oncomiRs in breast cancer, possibly targeting ples considered.
tumor suppressor genes or pathways. We apply our methodology to two cohorts of breast can-
The analysis of the dysregulated networks of the pre- cer samples (BRCA-US and ICGC). Given the large num-
dicted cancer-associated genes (protein-coding and miR- ber of samples in ICGC (n = 256), we perform three analy-
NAs) shows that many genes are dysregulated in a few sam- ses separately by considering (i) all samples, (ii) ER+ sam-
ples but rarely across all the mutated samples (Figure 5B). ples and (iii) ER– samples. Predictions vary depending on
However, the functional enrichment analysis of the dys- the samples’ histopathology. This is particularly important
regulated genes shows consistency across cohorts and the for methods relying on gene expression, which is influenced
analyzed types of mutations (LoF and cis-regulatory) for by the clinical composition of the cohorts. We acknowl-
both protein-coding and miRNA genes, even when there is edge that methodological differences between the BRCA-
a small intersection among the predicted genes in cohorts US and ICGC cohorts (e.g. different somatic mutation call-
of the same cancer type (Supplementary Figures S20 and ing algorithms, RNA-seq versus microarrays, and normal-
S21). Altogether, these observations suggest a phenotypic ization of RNA-seq raw counts) can provide additional ex-
heterogeneity (i.e. alterations of different parts of the same planations for the variation in predictions, which is the case
with the BRCA-US and ICGC cohorts that were indepen- Our study represents, to our knowledge, the first large-scale
dently normalized. Although only a few of the predicted analysis of cis-regulatory mutations that are linked to gene
protein-coding genes are predicted in both the ICGC and expression alteration in key cancer-associated pathways.
the BRCA-US cohorts (Supplementary Figure S20), the Our results suggest that this process can be achieved flexibly
functional enrichment analysis of the dysregulated gene net- because although we observe different genes in different pa-
works is consistent (Supplementary Figure S21). This ob- tients, all are associated with deregulation of the same path-
servation suggests common dysregulated pathways that act ways. Combining transcriptional and post-transcriptional
as attractors and that could originate from (non-recurrent) information, we identify a core set of 12 miRNAs linked
distinct cancer-associated events. It underlines the impor- to altered cancer pathways across cancer types. These pan-
tance of addressing cancer as a disease with perturbations cancer results provide new insights into the impact and po-
manifested at the gene network level. Our miRNA analyses tential causes of miRNA-mediated gene expression dysreg-
target gene expression alteration recurrently altered across ulation. This work extends our capacity to address the dis-
the BRCA-US and ICGC ER– breast cancer cohorts and covery gap of cancer-associated event identification through
highlight two miRNAs (hsa-mir-17-3p and hsa-mir-18-5p) the analysis of noncoding mutations and miRNA genes.

associated with cis-regulatory mutations.
Despite the multiple lines of evidence for the prediction of DATA AVAILABILITY
cancer-associated genes in this study, we acknowledge that
the predictions can provide false positives and false nega- The analysis with all the scripts and parameters
tives due to multiple reasons such as: (i) a limited number can be found through the following link: https:
of TFs with high-quality TFBSs; (ii) TFBS-target gene as- //bitbucket.org/CBGR/workspace/projects/DYS. We
sociations obtained by a naive approach combining infor- provide (i) the source code for the analysis and (ii)
mation from an integrative database (56) and association a pipeline for users to run similar analysis with their
to the closest TSS (Supplementary Figure S8)––we hypoth- own data. The repositories can be accessed with
esize that many of these associations may be irrelevant or the following links: for the dysmiR pipeline: https:
incorrect and many others are missing; (iii) a diversity of tu- //bitbucket.org/CBGR/dysmir pipeline, for the manuscript:
mor purity within the considered samples, despite the orig- https://bitbucket.org/CBGR/dysmir manuscript.
inal threshold of 80% used by TCGA; (iv) a limited number
of WGS datasets (tens of samples) within each cohort, com- SUPPLEMENTARY DATA
pared to the number of samples with WXS (hundreds) used
Supplementary Data are available at NAR Online.
in other studies; (v) prior networks that might be incom-
plete or with incorrect associations. Importantly, one of the
main limitations of this project is the low number of tumor ACKNOWLEDGEMENTS
samples with both WGS and RNA-seq data; this limitation As research parasites (102), we thank the TCGA and ICGC
not only biases the community research toward the study consortia and other researchers for making their data pub-
of exonic regions but also limits the statistical power of the licly available. We thank Jiarui Ding for his help in us-
methods assessing the impact of cis-regulatory mutations ing xseq for this study; Marcel Smid for providing raw
on gene network expression alteration. RNA-seq data for the ICGC cohort; Georgios Magklaras,
Altogether, we argue that our capacity to predict cancer- Georgios Marselis, Harold Gutch, and Torfinn Nome for
associated cis-regulation mutations will increase as more their IT support; Ingrid Kjelsvik and Elisa Bjørgo for ad-
high-quality TFBSs for more TFs and improved methods ministrative support; Rafael Riudavets-Puig, Roza Berhanu
to associate TFBSs with their target genes become avail- Lemma, Sebastian Waszak, Marieke Kuijjer, Yuvia A.
able. In addition, focusing on cis-regulatory regions specif- Perez-Rico, and Alexandra Gade for helpful comments on
ically open or active in cancer samples would inform where the manuscript; and the Mathelier, Kuijjer, and Kristensen’s
somatic mutations are likely effective. We expect that with groups for insightful discussions throughout the execution
more WGS, RNA-seq and other genomics datasets derived of this project.
from cancer samples available, the community will revert Author contributions: A.M. was responsible for the project
the paucity in the detection of noncoding cancer-associated conception and supervision. J.A.C.M. was responsible for
events (8). the analysis design and execution, and for its implementa-
tion. J.A.C.M. and M.R.A. undertook bioinformatic anal-
CONCLUSION ysis. J.A.C.M., M.R.A. and A.M. wrote the manuscript
with input from all co-authors. O.C.L. contributed with
By integrating whole-genome somatic mutations, RNA- CNA values for the ICGC cohort. J.W.M.M. contributed
seq, and small RNA-seq data with gene regulatory networks with clinical data. A.L. managed samples and clinical data.
across seven cancer types, we identify cis-regulatory mu- A.M., J.A.C.M., M.R.A., A.L.B.D. and V.K. contributed
tations associated with the dysregulation of gene regula- to the data analysis and scientific input. All authors read
tory networks through specific protein-coding and miRNA and approved the final manuscript.
genes. The enrichment for known cancer-associated genes
and the functional enrichment analysis reinforce a posteri-
FUNDING
ori the predicted protein-coding and miRNA genes as being
involved in biological pathway alteration affecting cancer Norwegian Research Council [187615]; Helse Sør-Øst; Uni-
development through exonic and cis-regulatory alterations. versity of Oslo through the Centre for Molecular Medicine
Norway (NCMM) (to A.M. and J.A.C.M.); Norwegian Re- Interpretation Working Group, Gallinger,S. and Stein,L.PCAWG
search Council [288404 to J.A.C.M. and Mathelier group]; Consortium (2020) Combined burden and functional impact tests
for cancer driver discovery using driverpower. Nat. Commun., 11,
Norwegian Cancer Society [197884 to Mathelier group]; 734.
M.R.A. was a postdoctoral fellow of the South Eastern 19. Kalender Atak,Z., Imrichova,H., Svetlichnyy,D., Hulselmans,G.,
Norway Health Authority [2014021 to A.L.B.D.]; a re- Christiaens,V., Reumers,J., Ceulemans,H. and Aerts,S. (2017)
search fellow of the Norwegian Cancer Society [711164 to Identification of cis-regulatory mutations generating de novo edges
V.N.K.]. Funding for open access charge: Norwegian Re- in personalized cancer gene regulatory networks. Genome Med., 9,
80.
search Council. 20. Fu,Y., Liu,Z., Lou,S., Bedford,J., Mu,X.J., Yip,K.Y., Khurana,E.
Conflict of interest statement. None declared. and Gerstein,M. (2014) FunSeq2: a framework for prioritizing
noncoding regulatory variants in cancer. Genome Biol., 15, 480.
21. Ritchie,G.R.S., Dunham,I., Zeggini,E. and Flicek,P. (2014)
REFERENCES Functional annotation of noncoding sequence variants. Nat.
1. Bradner,J.E., Hnisz,D. and Young,R.A. (2017) Transcriptional Methods, 11, 294–296.
addiction in cancer. Cell, 168, 629–643. 22. Boyle,A.P., Hong,E.L., Hariharan,M., Cheng,Y., Schaub,M.A.,

2. Melo,S.A. and Esteller,M. (2011) Dysregulation of microRNAs in Kasowski,M., Karczewski,K.J., Park,J., Hitz,B.C., Weng,S. et al.
cancer: playing with fire. FEBS Lett., 585, 2087–2099. (2012) Annotation of functional variation in personal genomes
3. Khurana,E., Fu,Y., Chakravarty,D., Demichelis,F., Rubin,M.A. using RegulomeDB. Genome Res., 22, 1790–1797.
and Gerstein,M. (2016) Role of non-coding sequence variants in 23. Przytycki,P.F. and Singh,M. (2020) Differential allele-specific
cancer. Nat. Rev. Genet., 17, 93–108. expression uncovers breast cancer genes dysregulated by cis
4. Cuykendall,T.N., Rubin,M.A. and Khurana,E. (2017) Non-coding noncoding mutations. Cell Syst., 10, 193–203.
genetic variation in cancer. Curr. Opin. Syst. Biol., 1, 9–15. 24. Cheng,Z., Vermeulen,M., Rollins-Green,M., DeVeale,B. and
5. Mathelier,A., Shi,W. and Wasserman,W.W. (2015) Identification of Babak,T. (2021) Cis-regulatory mutations with driver hallmarksin
altered cis-regulatory elements in human disease. Trends Genet., 31, major cancers. iScience, 24, 102144.
67–76. 25. Fredriksson,N.J., Ny,L., Nilsson,J.A. and Larsson,E. (2014)
6. Patel,M.B. and Wang,J. (2018) The Identification and Interpretation Systematic analysis of noncoding somatic mutations and gene
of cis-Regulatory Noncoding Mutations in Cancer. High expression alterations across 14 tumor types. Nat. Genet., 46,
Throughput, 8, 1. 1258–1263.
7. Gan,K.A., Carrasco Pro,S., Sewell,J.A. and Fuxman Bass,J.I. (2018) 26. Ding,J., McConechy,M.K., Horlings,H.M., Ha,G., Chun Chan,F.,
Identification of single nucleotide non-coding driver mutations in Funnell,T., Mullaly,S.C., Reimand,J., Bashashati,A., Bader,G.D.
cancer. Front. Genet., 9, 16. et al. (2015) Systematic analysis of somatic mutations impacting
8. Rheinbay,E., Nielsen,M.M., Abascal,F., Wala,J.A., Shapira,O., gene expression in 12 tumour types. Nat. Commun., 6, 8554.
Tiao,G., Hornshøj,H., Hess,J.M., Juul,R.I., Lin,Z. et al. (2020) 27. Mathelier,A., Lefebvre,C., Zhang,A.W., Arenillas,D.J., Ding,J.,
Analyses of non-coding somatic drivers in 2,658 cancer whole Wasserman,W.W. and Shah,S.P. (2015) Cis-regulatory somatic
genomes. Nature, 578, 102–111. mutations and gene-expression alteration in B-cell lymphomas.
9. Kumar,S., Warrell,J., Li,S., McGillivray,P.D., Meyerson,W., Genome Biol., 16, 84.
Salichos,L., Harmanci,A., Martinez-Fundichely,A., Chan,C.W.Y., 28. Bartel,D.P. and Chen,C.-Z. (2004) Micromanagers of gene
Nielsen,M.M. et al. (2020) Passenger mutations in more than 2,500 expression: the potentially widespread influence of metazoan
cancer genomes: overall molecular functional impact and microRNAs. Nat. Rev. Genet., 5, 396–400.
consequences. Cell, 180, 915–927. 29. Bartel,D.P. (2018) Metazoan MicroRNAs. Cell, 173, 20–51.
10. Wasserman,W.W. and Sandelin,A. (2004) Applied bioinformatics for 30. Ha,M. and Kim,V.N. (2014) Regulation of microRNA biogenesis.
the identification of regulatory elements. Nat. Rev. Genet., 5, Nat. Rev. Mol. Cell Biol., 15, 509–524.
276–287. 31. Agarwal,V., Bell,G.W., Nam,J.-W. and Bartel,D.P. (2015) Predicting
11. Lambert,S.A., Jolma,A., Campitelli,L.F., Das,P.K., Yin,Y., effective microRNA target sites in mammalian mRNAs. Elife, 4,
Albu,M., Chen,X., Taipale,J., Hughes,T.R. and Weirauch,M.T. e05005.
(2018) The human transcription factors. Cell, 172, 650–665. 32. Brancati,G. and Großhans,H. (2018) An interplay of miRNA
12. Weinhold,N., Jacobsen,A., Schultz,N., Sander,C. and Lee,W. (2014) abundance and target site architecture determines miRNA activity
Genome-wide analysis of noncoding regulatory mutations in cancer. and specificity. Nucleic Acids Res., 46, 3259–3269.
Nat. Genet., 46, 1160–1165. 33. Lin,S. and Gregory,R.I. (2015) MicroRNA biogenesis pathways in
13. Fujimoto,A., Furuta,M., Totoki,Y., Tsunoda,T., Kato,M., cancer. Nat. Rev. Cancer, 15, 321–333.
Shiraishi,Y., Tanaka,H., Taniguchi,H., Kawakami,Y., Ueno,M. 34. Peng,Y. and Croce,C.M. (2016) The role of micrornas in human
et al. (2016) Whole-genome mutational landscape and cancer. Signal Transduct. Targeted Ther., 1, 15004.
characterization of noncoding and structural mutations in liver 35. Kern,F., Backes,C., Hirsch,P., Fehlmann,T., Hart,M., Meese,E. and
cancer. Nat. Genet., 48, 500–509. Keller,A. (2019) What’s the target: understanding two decades of in
14. Lochovsky,L., Zhang,J., Fu,Y., Khurana,E. and Gerstein,M. (2015) silico microRNA-target prediction. Brief. Bioinform., 21, 1999–2010.
LARVA: an integrative framework for large-scale analysis of 36. de Rie,D., Abugessaisa,I., Alam,T., Arner,E., Arner,P., Ashoor,H.,
recurrent variants in noncoding annotations. Nucleic Acids Res., 43, Åström,G., Babina,M., Bertin,N., Burroughs,A.M. et al. (2017) An
8123–8134. integrated expression atlas of miRNAs and their promoters in
15. Lochovsky,L., Zhang,J. and Gerstein,M. (2018) MOAT: efficient human and mouse. Nat. Biotech., 35, 872–878.
detection of highly mutated regions with the mutations 37. Puig,R.R., Boddie,P., Khan,A., Castro-Mondragon,J.A. and
overburdening annotations tool. Bioinformatics, 34, 1031–1033. Mathelier,A. (2021) UniBind: maps of high-confidence direct
16. Rheinbay,E., Parasuraman,P., Grimsby,J., Tiao,G., Engreitz,J.M., TF-DNA interactions across nine species. BMC Genomics, 22, 482.
Kim,J., Lawrence,M.S., Taylor-Weiner,A., Rodriguez-Cuevas,S., 38. Gheorghe,M., Sandve,G.S., Khan,A., Cheneby,J., Ballester,B. and
Rosenberg,M. et al. (2017) Recurrent and functional regulatory Mathelier,A. (2019) A map of direct TF–DNA interactions in the
mutations in breast cancer. Nature, 547, 55–60. human genome. Nucleic Acids Res., 47, e21.
17. Feigin,M.E., Garvin,T., Bailey,P., Waddell,N., Chang,D.K., 39. The Cancer Genome Atlas Research Network, Weinstein,J.N.,
Kelley,D.R., Shuai,S., Gallinger,S., McPherson,J.D., Collisson,E.A., Mills,G.B., Mills Shaw,K.R., Ozenberger,B.A.,
Grimmond,S.M. et al. (2017) Recurrent noncoding regulatory Ellrott,K., Shmulevich,I., Sander,C. and Stuart,J.M. (2013) The
mutations in pancreatic ductal adenocarcinoma. Nat. Genet., 49, cancer genome atlas pan-cancer analysis project. Nat. Genet., 45,
825–833. 1113–1120.
18. Shuai,S., PCAWG Drivers and Functional Interpretation Working 40. Hinrichs,A.S., Karolchik,D., Baertsch,R., Barber,G.P., Bejerano,G.,
Group and PCAWG ConsortiumPCAWG Drivers and Functional Clawson,H., Diekhans,M., Furey,T.S., Harte,R.A., Hsu,F. et al.
(2006) The UCSC genome browser database: update 2006. Nucleic 58. Kanehisa,M. and Goto,S. (2000) KEGG: kyoto encyclopedia of
Acids Res., 34, D590–D598. genes and genomes. Nucleic Acids Res., 28, 27–30.
41. Karolchik,D. and James Kent,W. (2003) The UCSC genome 59. Pico,A.R., Kelder,T., van Iersel,M.P., Hanspers,K., Conklin,B.R.
browser. Curr. Protoc. Bioinformatics, and Evelo,C. (2008) WikiPathways: pathway editing for the people.
https://doi.org/10.1002/0471250953.bi0104s00. PLoS Biol., 6, e184.
42. Zhang,J., Baran,J., Cros,A., Guberman,J.M., Haider,S., Hsu,J., 60. Karp,P.D., Ouzounis,C.A., Moore-Kochlacs,C., Goldovsky,L.,
Liang,Y., Rivkin,E., Wang,J., Whitty,B. et al. (2011) International Kaipa,P., Ahrén,D., Tsoka,S., Darzentas,N., Kunin,V. and
cancer genome consortium data Portal––a one-stop shop for cancer López-Bigas,N. (2005) Expansion of the biocyc collection of
genomics data. Database, 2011, bar026. pathway/genome databases to 160 genomes. Nucleic Acids Res., 33,
43. Nik-Zainal,S., Davies,H., Staaf,J., Ramakrishna,M., Glodzik,D., 6083–6089.
Zou,X., Martincorena,I., Alexandrov,L.B., Martin,S., Wedge,D.C. 61. Zhou,H., Jin,J., Zhang,H., Yi,B., Wozniak,M. and Wong,L. (2012)
et al. (2016) Landscape of somatic mutations in 560 breast cancer IntPath–an integrated pathway gene relationship database for model
whole-genome sequences. Nature, 534, 47–54. organisms and important pathogens. BMC Syst. Biol., 6(Suppl. 2),
44. Smid,M., Rodrı́guez-González,F.G., Sieuwerts,A.M., Salgado,R., S2.
Smissen,Prager-Vander, Vlugt-Daane,W.J.C., van der,M., van 62. Gerstein,M.B., Kundaje,A., Hariharan,M., Landt,S.G., Yan,K.-K.,
Galen,A., Nik-Zainal,S., Staaf,J. et al. (2016) Breast cancer genome Cheng,C., Mu,X.J., Khurana,E., Rozowsky,J., Alexander,R. et al.

and transcriptome integration implicates specific mutational (2012) Architecture of the human regulatory network derived from
signatures with immune cell infiltration. Nat. Commun., 7, 12910. ENCODE data. Nature, 489, 91–100.
45. Fan,Y., Xi,L., Hughes,D.S.T., Zhang,J., Zhang,J., Futreal,P.A., 63. Krek,A., Grün,D., Poy,M.N., Wolf,R., Rosenberg,L., Epstein,E.J.,
Wheeler,D.A. and Wang,W. (2016) MuSE: accounting for tumor MacMenamin,P., da Piedade,I., Gunsalus,K.C., Stoffel,M. et al.
heterogeneity using a sample-specific error model improves (2005) Combinatorial microRNA target predictions. Nat. Genet.,
sensitivity and specificity in mutation calling from sequencing data. 37, 495–500.
Genome Biol., 17, 178. 64. Seitz,H. (2017) Issues in current microRNA target identification
46. Jones,D., Raine,K.M., Davies,H., Tarpey,P.S., Butler,A.P., methods. RNA Biol., 14, 831–834.
Teague,J.W., Nik-Zainal,S. and Campbell,P.J. (2016) 65. Csardi,G. and Nepusz,T. (2006) The igraph software package for
cgpCaVEManWrapper: simple execution of caveman in order to complex network research. Interjournal, complex systems, 1695, 1–9.
detect somatic single nucleotide variants in NGS data. Curr. Protoc. 66. Tyner,S., Briatte,F. and Hofmann,H. (2017) Network visualization
Bioinformatics, 56, 15.10.1–15.10.18. with ggplot2. The R Journal, 9, 27.
47. Raine,K.M., Hinton,J., Butler,A.P., Teague,J.W., Davies,H., 67. Kuleshov,M.V., Jones,M.R., Rouillard,A.D., Fernandez,N.F.,
Tarpey,P., Nik-Zainal,S. and Campbell,P.J. (2015) cgpPindel: Duan,Q., Wang,Z., Koplev,S., Jenkins,S.L., Jagodnik,K.M.,
identifying somatically acquired insertion and deletion events from Lachmann,A. et al. (2016) Enrichr: a comprehensive gene set
paired end sequencing. Curr. Protoc. Bioinformatics, 52, enrichment analysis web server 2016 update. Nucleic Acids Res., 44,
15.7.1–15.7.12. W90–W97.
48. Robinson,M.D., McCarthy,D.J. and Smyth,G.K. (2010) edgeR: a 68. Repana,D., Nulsen,J., Dressler,L., Bortolomeazzi,M., Venkata,S.K.,
bioconductor package for differential expression analysis of digital Tourna,A., Yakovleva,A., Palmieri,T. and Ciccarelli,F.D. (2019) The
gene expression data. Bioinformatics, 26, 139–140. network of cancer genes (NCG): a comprehensive catalogue of
49. Mermel,C.H., Schumacher,S.E., Hill,B., Meyerson,M.L., known and candidate cancer genes from cancer sequencing screens.
Beroukhim,R. and Getz,G. (2011) GISTIC2.0 facilitates sensitive Genome Biol., 20, 1.
and confident localization of the targets of focal somatic 69. Gundem,G., Perez-Llamas,C., Jene-Sanz,A., Kedzierska,A.,
copy-number alteration in human cancers. Genome Biol., 12, R41. Islam,A., Deu-Pons,J., Furney,S.J. and Lopez-Bigas,N. (2010)
50. Van Loo,P., Nordgard,S.H., Lingjærde,O.C., Russnes,H.G., IntOGen: integration and data mining of multidimensional
Rye,I.H., Sun,W., Weigman,V.J., Marynen,P., Zetterberg,A., oncogenomic data. Nat. Methods, 7, 92–93.
Naume,B. et al. (2010) Allele-specific copy number analysis of 70. Sondka,Z., Bamford,S., Cole,C.G., Ward,S.A., Dunham,I. and
tumors. Proc. Natl. Acad. Sci. U.S.A., 107, 16910–16915. Forbes,S.A. (2018) The COSMIC cancer gene census: describing
51. O’Leary,N.A., Wright,M.W., Brister,J.R., Ciufo,S., Haddad,D., genetic dysfunction across all human cancers. Nat. Rev. Cancer, 18,
McVeigh,R., Rajput,B., Robbertse,B., Smith-White,B., 696–705.
Ako-Adjei,D. et al. (2016) Reference sequence (RefSeq) database at 71. Xie,B., Ding,Q., Han,H. and Wu,D. (2013) miRCancer: a
NCBI: current status, taxonomic expansion, and functional microRNA–cancer association database constructed by text mining
annotation. Nucleic Acids Res., 44, D733–D745. on literature. Bioinformatics, 29, 638–644.
52. Quinlan,A.R. and Hall,I.M. (2010) BEDTools: a flexible suite of 72. Curtis,C., Shah,S.P., Chin,S.-F., Turashvili,G., Rueda,O.M.,
utilities for comparing genomic features. Bioinformatics, 26, Dunning,M.J., Speed,D., Lynch,A.G., Samarajiwa,S., Yuan,Y. et al.
841–842. (2012) The genomic and transcriptomic architecture of 2,000 breast
53. Kozomara,A. and Griffiths-Jones,S. (2014) miRBase: annotating tumours reveals novel subgroups. Nature, 486, 346–352.
high confidence microRNAs using deep sequencing data. Nucleic 73. Dvinge,H., Git,A., Gräf,S., Salmon-Divon,M., Curtis,C.,
Acids Res., 42, D68–D73. Sottoriva,A., Zhao,Y., Hirst,M., Armisen,J., Miska,E.A. et al.
54. Xu,T., Su,N., Liu,L., Zhang,J., Wang,H., Zhang,W., Gui,J., Yu,K., (2013) The shaping and functional consequences of the microRNA
Li,J. and Le,T.D. (2018) miRBaseConverter: an R/Bioconductor landscape in breast cancer. Nature, 497, 378–382.
package for converting and retrieving miRNA name, accession, 74. Thomas,P.D., Campbell,M.J., Kejariwal,A., Mi,H., Karlak,B.,
sequence and family information in different versions of miRBase. Daverman,R., Diemer,K., Muruganujan,A. and Narechania,A.
BMC Bioinformatics, 19, 514. (2003) PANTHER: a library of protein families and subfamilies
55. Khan,A., Fornes,O., Stigliani,A., Gheorghe,M., indexed by function. Genome Res., 13, 2129–2141.
Castro-Mondragon,J.A., van der Lee,R., Bessy,A., Chèneby,J., 75. Ashburner,M., Ball,C.A., Blake,J.A., Botstein,D., Butler,H.,
Kulkarni,S.R., Tan,G. et al. (2018) JASPAR 2018: update of the Cherry,J.M., Davis,A.P., Dolinski,K., Dwight,S.S., Eppig,J.T. et al.
open-access database of transcription factor binding profiles and its (2000) Gene ontology: tool for the unification of biology. The gene
web framework. Nucleic Acids Res., 46, D1284. ontology consortium. Nat. Genet., 25, 25–29.
56. Fishilevich,S., Nudel,R., Rappaport,N., Hadar,R., Plaschkes,I., Iny 76. Sayers,E.W., Barrett,T., Benson,D.A., Bolton,E., Bryant,S.H.,
Stein,T., Rosen,N., Kohn,A., Twik,M., Safran,M. et al. (2017) Canese,K., Chetvernin,V., Church,D.M., DiCuccio,M., Federhen,S.
GeneHancer: genome-wide integration of enhancers and target et al. (2011) Database resources of the national center for
genes in genecards. Database, 2017, bax028. biotechnology information. Nucleic Acids Res., 39, D38–D51.
57. Franceschini,A., Szklarczyk,D., Frankild,S., Kuhn,M., 77. Mendell,J.T. (2008) miRiad roles for the miR-17-92 cluster in
Simonovic,M., Roth,A., Lin,J., Minguez,P., Bork,P., von Mering,C. development and disease. Cell, 133, 217–222.
et al. (2013) STRING v9.1: protein-protein interaction networks, 78. Teng,G. and Papavasiliou,F.N. (2009) Shhh! Silencing by
with increased coverage and integration. Nucleic Acids Res., 41, microRNA-155. Philos. Trans. R. Soc. Lond. B Biol. Sci., 364,
D808–D815. 631–637.
79. Nadiminty,N., Tummala,R., Lou,W., Zhu,Y., Shi,X.-B., Zou,J.X., 91. Lee,W., Jiang,Z., Liu,J., Haverty,P.M., Guan,Y., Stinson,J., Yue,P.,
Chen,H., Zhang,J., Chen,X., Luo,J. et al. (2012) MicroRNA let-7c is Zhang,Y., Pant,K.P., Bhatt,D. et al. (2010) The mutation spectrum
downregulated in prostate cancer and suppresses prostate cancer revealed by paired genome sequences from a lung cancer patient.
growth. PLoS One, 7, e32832. Nature, 465, 473–477.
80. Tehler,D., Høyland-Kroghsbo,N.M. and Lund,A.H. (2011) The 92. Vorontsov,I.E., Khimulya,G., Lukianova,E.N., Nikolaeva,D.D.,
miR-10 microRNA precursor family. RNA Biol., 8, 728–734. Eliseeva,I.A., Kulakovskiy,I.V. and Makeev,V.J. (2016) Negative
81. Ke,K. and Lou,T. (2017) MicroRNA-10a suppresses breast cancer selection maintains transcription factor binding motifs in human
progression via PI3K/Akt/mTOR pathway. Oncol. Lett., 14, cancer. BMC Genomics, 17, 395.
5994–6000. 93. Martincorena,I., Raine,K.M., Gerstung,M., Dawson,K.J.,
82. Mu,N., Gu,J., Huang,T., Zhang,C., Shu,Z., Li,M., Hao,Q., Li,W., Haase,K., Van Loo,P., Davies,H., Stratton,M.R. and Campbell,P.J.
Zhang,W., Zhao,J. et al. (2016) A novel (2017) Universal patterns of selection in cancer and somatic tissues.
NF-␬B/YY1/microRNA-10a regulatory circuit in fibroblast-like Cell, 171, 1029–1041.
synoviocytes regulates inflammation in rheumatoid arthritis. Sci. 94. Frigola,J., Sabarinathan,R., Mularoni,L., Muiños,F.,
Rep., 6, 20059. Gonzalez-Perez,A. and López-Bigas,N. (2017) Reduced mutation
83. Hirschberger,S., Hinske,L.C. and Kreth,S. (2018) MiRNAs: rate in exons due to differential mismatch repair. Nat. Genet., 49,
dynamic regulators of immune cell functions in inflammation and 1684–1692.

cancer. Cancer Lett., 431, 11–21. 95. Sabarinathan,R., Mularoni,L., Deu-Pons,J., Gonzalez-Perez,A. and
84. Kim,E.-A., Kim,S.-W., Nam,J., Sung,E.-G., Song,I.-H., Kim,J.-Y., López-Bigas,N. (2016) Nucleotide excision repair is impaired by
Kwon,T.K. and Lee,T.-J. (2016) Inhibition of c-FLIPL expression by binding of transcription factors to DNA. Nature, 532, 264–267.
miRNA-708 increases the sensitivity of renal cancer cells to 96. Kaiser,V.B., Taylor,M.S. and Semple,C.A. (2016) Mutational biases
anti-cancer drugs. Oncotarget, 7, 31832–31846. drive elevated rates of substitution at regulatory sites across cancer
85. Lee,J.-W., Guan,W., Han,S., Hong,D.-K., Kim,L.-S. and Kim,H. types. PLoS Genet., 12, e1006207.
(2018) MicroRNA-708-3p mediates metastasis and chemoresistance 97. Nishizaki,S.S., Ng,N., Dong,S., Porter,R.S., Morterud,C.,
through inhibition of epithelial-to-mesenchymal transition in breast Williams,C., Asman,C., Switzenberg,J.A. and Boyle,A.P. (2020)
cancer. Cancer Sci., 109, 1404–1413. Predicting the effects of SNPs on transcription factor binding
86. Li,J., Li,L., Li,Z., Gong,G., Chen,P., Liu,H., Wang,J., Liu,Y. and affinity. Bioinformatics, 36, 364–372.
Wu,X. (2015) The role of miR-205 in the VEGF-mediated 98. Fornes,O., Gheorghe,M., Richmond,P.A., Arenillas,D.J.,
promotion of human ovarian cancer cell invasion. Gynecol. Oncol., Wasserman,W.W. and Mathelier,A. (2018) MANTA2, update of the
137, 125–133. mongo database for the analysis of transcription factor binding site
87. Tuo,H., Wang,Y., Wang,L., Yao,B., Li,Q., Wang,C., Liu,Z., Han,S., alterations. Sci. Data, 5, 180141.
Yin,G., Tu,K. et al. (2017) MiR-324-3p promotes tumor growth 99. Gonzalez-Perez,A., Sabarinathan,R. and Lopez-Bigas,N. (2019)
through targeting DACT1 and activation of Wnt/␤-catenin pathway Local determinants of the mutational landscape of the human
in hepatocellular carcinoma. Oncotarget, 8, 65687–65698. genome. Cell, 177, 101–114.
88. Lu,J., Lu,S., Li,J., Yu,Q., Liu,L. and Li,Q. (2018) MiR-629-5p 100. Carrasco Pro,S., Bulekova,K., Gregor,B., Labadorf,A. and Fuxman
promotes colorectal cancer progression through targetting CXXC Bass,J.I. (2021) Prediction of genome-wide effects of single
finger protein 4. Biosci. Rep., 38, BSR20180613. nucleotide variants on transcription factor binding. Sci. Rep., 10,
89. Karimzadeh,M., Arlidge,C., Rostami,A., Lupien,M., Bratman,S.V. 17632.
and Hoffman,M.M. (2020) Viral integration transforms chromatin 101. Kircher,M., Xiong,C., Martin,B., Schubach,M., Inoue,F.,
to drive oncogenesis. bioRxiv doi: Bell,R.J.A., Costello,J.F., Shendure,J. and Ahituv,N. (2019)
https://doi.org/10.1101/2020.02.12.942755, 14 May 2021, preprint: Saturation mutagenesis of twenty disease-associated regulatory
not peer reviewed. elements at single base-pair resolution. Nat. Commun., 10, 3583.
90. Zapatka,M., Borozan,I., Brewer,D.S., Iskar,M., Grundhoff,A., 102. Longo,D.L. and Drazen,J.M. (2016) Data sharing. N. Engl. J. Med.,
Alawi,M., Desai,N., Sültmann,H., Moch,H., Pathogens,PCAWG 374, 276–277.
et al. (2020) The landscape of viral associations in human cancers.
Nat. Genet., 52, 320–330.

Gkac 1143

Uploaded by

Copyright:

Available Formats

Gkac 1143

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gkac 1143

Uploaded by

Copyright:

Available Formats

Published online 8 December 2022 Nucleic Acids Research, 2022, Vol. 50, No.

Cis-regulatory mutations associate with

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

ABSTRACT our method predicts cis-regulatory mutations related

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

TFBS mutations xseq analyses

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Dysregulation heatmaps Results accessibility

Aggregated and sample-specific networks RESULTS

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

BRCA−US LIHC−US UCEC−US HNSC−US

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Aggregated network − Nodes: 87 − Subgraphs: 1

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

ICGC (all) ICGC ER+ ICGC ER− BRCA−US

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

Downloaded from https://academic.oup.com/nar/article/50/21/12131/6882141 by guest on 05 April 2023

You might also like