HRDetect is a predictor of BRCA1 and BRCA2 deficiency based
on mutational-signatures
Helen Davies
1,*
, Dominik Glodzik
1,*
, Sandro Morganella
1
, Lucy R. Yates
1,2
, Johan Staaf
3
,
Xueqing Zou
1
, Manasa Ramakrishna
1,4
, Sancha Martin
1
, Sandrine Boyault
5
, Anieta M.
Sieuwerts
6
, Peter T. Simpson
7
, Tari A. King
8
, Keiran Raine
1
, Jorunn E. Eyfjord
9
, Gu Kong
10
,
Åke Borg
3
, Ewan Birney
11
, Hendrik G. Stunnenberg
12
, Marc J. van de Vijver
13
, Anne-Lise
Børresen-Dale
14,15
, John W.M. Martens
6
, Paul N. Span
16
, Sunil R Lakhani
7,17
, Anne
Vincent-Salomon
18
, Christos Sotiriou
19
, Andrew Tutt
20,21
, Alastair M. Thompson
22
, Steven
Van Laere
23,24
, Andrea L. Richardson
25,26
, Alain Viari
27,28
, Peter J Campbell
1
, Michael R.
Stratton
1
, and Serena Nik-Zainal
1,29
1
Wellcome Trust Sanger Institute, Hinxton, Cambridge CB10 1SA, UK
2
Guys and St Thomas’
NHS Trust, London, UK
3
Division of Oncology and Pathology, Department of Clinical Sciences
Lund, Lund University, Lund, SE-223 81, Sweden
4
Oncology, Innovative Medicines and Early
Development Biotech Unit, AstraZeneca, Hodgkin Building, Chesterford Research Park, Little
Chesterford, Cambridge CB10 1XL, UK
5
Centre Léon Bérard, Translational Research Lab
Department, 28, rue Laënnec, 69373 Lyon Cedex 08, France
6
Department of Medical Oncology,
Erasmus MC Cancer Institute and Cancer Genomics Netherlands, Erasmus University Medical
Center, Rotterdam 3015CN, The Netherlands
7
The University of Queensland: UQ Centre for
Clinical Research and School of Medicine, Brisbane, Queensland 4029, Australia
8
Memorial
Sloan Kettering Cancer Center, 1275 York Ave, New York, NY 10065, United States
9
Cancer
Research Laboratory, Faculty of Medicine, University of Iceland, 101 Reykjavik, Iceland
10
Department of Pathology, College of Medicine, Hanyang University, Seoul, 133-791, South
Corresponding author : Serena Nik-Zainal (snz@sanger.ac.uk).
*
shared first authorship
Equal contributions: HD and DG contributed equally towards the development of the principles and the analyses in this manuscript.
Competing Financial Interests
H. Davies, D. Glodzik and S. Nik-Zainal are inventors on a patent application encompassing the code and intellectual principle on this
algorithm. A. Tutt has been in receipt of payments from the Institute of Cancer Research Rewards to inventors scheme associated with
the invention of PARP inhibitors as therapy for
BRCA1
and
BRCA2
mutation associated cancers.
Author Contributions
H.D., D.G., S.N.-Z. drove the development of the intellectual concepts, performed analyses and wrote the manuscript.
S.M., J.S., X.Z. and M.R. contributed towards data curation and performed analyses.
L.R.Y., S.B., A.M.S., P.T.S., T.A.K., J.E.E., P.N.S., S.R.L., A.V.-S., C.S., A.T., A.M.T., S.V.L., A.L. contributed new samples and/or to
experimental design of the study.
S.M. was the scientific project coordinator.
K.R. provided bioinformatics support.
P.J.C. provided infrastructure at Wellcome Trust Sanger Institute.
G.K., A.B., E.B., H.G.S., M.JvdV., A.-L.B.-D., J.W.M.M., A.M.T., A.L.R., A.V. and M.R.S. originally conceived the concept of the
Breast Cancer Consortium that generated the data resource which has been utilised for these analyses, contributed samples old and
new and contributed comments towards the manuscript.
Data access
Breast cancer whole genome sequence BAM files and CEL files from Affymetrix SNP6 arrays are available from European Genome-
phenome Archive (EGA), see Accession codes sections for details of accession codes.
Europe PMC Funders Group
Author Manuscript
Nat Med. Author manuscript; available in PMC 2018 March 02.
Published in final edited form as:
Nat Med
. 2017 April ; 23(4): 517–525. doi:10.1038/nm.4292.
Europe PMC Funders Author Manuscripts Europe PMC Funders Author Manuscripts
Korea
11
European Molecular Biology Laboratory, European Bioinformatics Institute, Wellcome
Trust Genome Campus,Hinxton, Cambridgeshire, CB10 1SD
12
Department of Molecular Biology,
Faculties of Science and Medicine, Radboud University, 6525GA, Nijmegen, Netherlands
13
Department of Pathology, Academic Medical Center, Meibergdreef 9, 1105 AZ Amsterdam, The
Netherlands
14
Department of Cancer Genetics, Institute for Cancer Research, Oslo University
Hospital The Norwegian Radium Hospital Oslo 0310, Norway
15
K.G. Jebsen Centre for Breast
Cancer Research, Institute for Clinical Medicine, University of Oslo, Oslo 0310, Norway
16
Department of Radiation Oncology, and department of Laboratory Medicine, Radboud
university medical center, Nijmegen 6525GA, The Netherlands
17
Pathology Queensland, The
Royal Brisbane and Women’s Hospital, Brisbane, Queensland 4029, Australia
18
Institut Curie,
Department of Pathology and INSERM U934, 26 rue d'Ulm, 75248 Paris Cedex 05, France
19
Breast Cancer Translational Research Laboratory, Université Libre de Bruxelles, Institut Jules
Bordet, Bd de Waterloo 121, B-1000 Brussels, Belgium
20
Breast Cancer Now Research Unit,
King’s College, London, UK
21
Breast Cancer Now Toby Robin’s Research Centre, Institute of
Cancer Research, London, UK
22
Department of Breast Surgical Oncology, University of Texas
MD Anderson Cancer Center, 1400 Pressler Street, Houston, Texas 77030, USA
23
Translational
Cancer Research Unit, Center for Oncological Research, Faculty of Medicine and Health
Sciences, University of Antwerp, Antwerp, Belgium
24
HistoGeneX NV, Wilrijk, Belgium
25
Department of Pathology, Brigham and Women's Hospital, Boston, MA 02115 USA
26
Dana-
Farber Cancer Institute, Boston, MA 02215 USA
27
Equipe Erable, INRIA Grenoble-Rhône-Alpes,
655, Avenue de l'Europe, 38330 Montbonnot-Saint Martin, France
28
Synergie Lyon Cancer,
Centre Léon Bérard, 28 rue Laënnec, Lyon Cedex 08, France
29
East Anglian Medical Genetics
Service, Cambridge University Hospitals NHS Foundation Trust, Cambridge CB2 9NB, UK
Abstract
Approximately 1-5% of breast cancers are attributed to inherited mutations in
BRCA1
or
BRCA2
and are selectively sensitive to poly (ADP-ribose) polymerase (PARP) inhibitors. Germline and/or
somatic mutations in
BRCA1
/
BRCA2
in other cancer types also confer selective sensitivity to
PARP inhibitors. Thus, assays to detect
BRCA1
/
BRCA2
deficient tumours have been sought.
Recently, somatic substitution, insertion/deletion and rearrangement patterns or
mutational
signatures
were associated with
BRCA1
/
BRCA2
dysfunction. We used a supervised lasso logistic
regression model to identify six critically distinguishing mutational signatures predictive of
BRCA1
/
BRCA2
deficiency. A weighted model called HRDetect was developed to accurately
detect
BRCA1
/
BRCA2
deficient samples. HRDetect identifies
BRCA1
/
BRCA2
deficient tumours
with 98.7% sensitivity (AUC 0.98). Application of this model in a cohort of 560 breast cancer
patients with 22 known germline
BRCA1
/
BRCA2
mutation carriers, allowed us to identify an
additional 22 somatic
BRCA1
/
BRCA2
null tumours and 47 tumours with functional
BRCA1
/
BRCA2
-deficiency where no mutation was detected. We validated HRDetect on independent
cohorts of breast, ovarian and pancreatic cancers, and demonstrate efficacy on alternative
sequencing strategies. Integrating all classes of mutational signatures thus reveals a larger
proportion of breast cancer patients (of up to 22%) than hitherto appreciated (~1-5%) that could
have selective therapeutic sensitivity to PARP-inhibition.
Davies et al. Page 2
Nat Med
. Author manuscript; available in PMC 2018 March 02.
Europe PMC Funders Author Manuscripts Europe PMC Funders Author Manuscripts
Introduction
A small fraction of breast cancers (~1-5%)1–3 are attributed to familial mutations in the
BRCA1
and
BRCA2
cancer susceptibility genes. Heterozygous germline mutations in
BRCA1
and
BRCA2
confer elevated lifetime risks of breast, ovarian and other cancers4,5.
BRCA1 and BRCA2 proteins have multiple distinct roles in maintaining genome integrity,
particularly, through Homologous Recombination (HR)-mediated double strand break
(DSB) repair6. These classical tumour suppressor genes usually lose the wild-type allele
during tumorigenesis to become fully inactivated7.
BRCA1
/
BRCA2
null tumours are thus
deficient in HR and selectively sensitive to compounds that increase the demand on HR8.
Poly (ADP-ribose) polymerase (PARP) inhibitors are an example of therapeutic compounds
that cause replication fork stalling and collapse leading to increased DSBs9. The inability to
perform HR-dependent DSB repair ultimately leads to selective tumour cell death10,11.
Preclinical studies and Phase I/II breast and ovarian clinical trials12,13 have shown PARP-
inhibitor efficacy in familial
BRCA1
and
BRCA2
patients. However, PARP-inhibition has
applications beyond that of germline mutated tumours14. Effective PARP-inhibition
maintenance therapy has been demonstrated in high grade serous ovarian cancer with
germline or somatic
BRCA1
/
BRCA2
mutations15. Thus, extensive efforts have been put
into identification of molecular features of tumours that are
BRCA1
/
BRCA2
deficient,
referred to historically as “BRCAness”, whether inactivated through germline, somatic or
secondary means, including promoter hypermethylation or inactivation of a related gene in
the HR pathway.
Gene-specific sequencing strategies including sequencing all known HR genes, Multiplex
Ligation-dependent Probe Amplification (MLPA)16, promoter hypermethylation assays17,
transcriptional metagene signatures18–20, copy number-based methods (e.g. HRD
(Homologous Recombination Deficiency) index and genomic “scars”)21–23 and functional
assays of HR competence24 have been developed to detect
BRCA1
/
BRCA2
deficiency.
However, these indices have had limited predictive success. A recent review suggests that a
good predictor of the biological status of a HR-deficient tumour is essential, as the cohort of
tumours that demonstrate BRCAness and that could be selectively sensitive to PARP-
inhibitors is likely not limited to the small proportion of familial breast and ovarian cancers,
but extends to a larger fraction of sporadic breast and ovarian cancers as well as other cancer
types25.
Recent advances in sequencing technology26 have significantly reduced sequencing costs,
permitting whole genome sequencing (WGS) for the detection of all somatic mutations
including base substitutions, insertions/deletions (indels), rearrangements and copy number
aberrations in human cancers. Deep analysis reveals patterns of mutations, or somatic
mutational signatures, which are the physiological readout of the DNA damage and DNA
repair processes that have occurred through tumorigenesis27–31. These patterns are
indicators of past and on-going exposures, whether of environmental insults such as
ultraviolet radiation, or of endogenous biochemical degradation and deficiencies of DNA
repair pathways like HR.
Davies et al. Page 3
Nat Med
. Author manuscript; available in PMC 2018 March 02.
Europe PMC Funders Author Manuscripts Europe PMC Funders Author Manuscripts
We reason that mutational signatures which report
BRCA1
/
BRCA2
deficiency in germline
mutated tumours could be used as a predictor of other tumours that also have this deficiency.
Previously, base substitution Signature 3 was shown to distinguish germline
BRCA1/
BRCA2
null from sporadic cancers in a small subset of breast cancers29,30 and
subsequently extended to pancreatic32,33, ovarian34 and stomach cancer35. However,
selecting a cut-off to discriminate
BRCA1
/
BRCA2
-deficient from -proficient cancers is not
straightforward when using this signature alone. Recent characterisation of a large cohort of
WGS breast cancers27,28 has provided new insights. A defect in a single gene such as
BRCA1
/
BRCA2
does not result in a single signature – it gives rise to at least five mutational
signatures of all classes, including base substitutions, indels and rearrangements27,28.
Unlike most biomarkers, these multiple mutational signatures are the direct consequence of
abrogation of DSB repair pathways. Thus, in the current analysis, we exploit this
observation to quantitatively define genomic features of
BRCA1
/
BRCA2
deficiency and
present a WGS-based predictor with remarkable preformance for detection of HR-deficient
tumours.
Results
Quantitatively defining features of “BRCA”ness
24 women carrying inherited predisposition mutations in
BRCA1
(5) and
BRCA2
(19) were
recruited into a breast cancer genome sequencing study involving 560 patients27. Loss of
the wild-type allele predicted to result in complete inactivation of the relevant protein was
observed in 22 of the 24 breast cancers.
These 22 tumours had a distinguishing genomic profile: overrepresentation of base-
substitution Signatures 3 or 8, an excess of large deletions (>3bp) with microhomology at
the junction of the deletion, Rearrangement Signature 5, and copy number profiles
associated with widespread loss of heterozygosity (Figure 1). Additionally, BRCA1 null
tumours also had an excess of Rearrangement Signature 3 (characterized by short <10kb)
tandem duplications) mainly, and a lesser contribution of Rearrangement Signature 1
(typified by long >100kb tandem duplications)27.
The 22
BRCA1
/
BRCA2
null tumours were used in a first training set to quantitatively define
features of
BRCA1
/
BRCA2
deficiency. They were contrasted to a cohort of 235 sporadic
breast cancers with quiescent genomic profiles, distinct from
BRCA1
/
BRCA2
null cancers.
Somatic variants of all classes of mutation had been previously called. Twelve base
substitution, two indel and six rearrangement mutational signatures were previously
extracted and HRD copy number indices were obtained (Supplementary Table 1). A
supervised learning lasso logistic regression model was applied to counts of mutational
signatures and to HRD indices that were log-transformed and normalized to permit
comparability between genomic parameters (Supplementary Table 2).
An iterative ten-fold nested cross-validation strategy was adopted where 90% of samples
were used for model parameter selection and the weights for each parameter were tested on
Davies et al. Page 4
Nat Med
. Author manuscript; available in PMC 2018 March 02.
Europe PMC Funders Author Manuscripts Europe PMC Funders Author Manuscripts
the remaining 10% of samples. This was performed to ensure that parameters identified as
putative predictors of
BRCA1
/
BRCA2
deficiency were robust and generalizable.
Five distinguishing parameters with different individual weights were found to convey the
greatest difference between
BRCA1
/
BRCA2
-deficient cancers and sporadic breast cancers:
microhomology-mediated indels, the HRD index, base substitution Signature 3,
Rearrangement Signature 3 and Rearrangement Signature 5(Supplementary Table 2).
Identification of additional BRCA1 and BRCA2 null tumors
The selected parameters were applied across the cohort of 560 breast cancers to test the
performance of our model in predicting
BRCA1
/
BRCA2
deficiency, and to detect other
cancers with similar characteristics to germline
BRCA1
/
BRCA2
null tumours Figure 2 for
workflow, Supplementary Table 2). The resulting distribution of probabilities of
BRCA1
/
BRCA2
deficiency was a strikingly steep sigmoidal curve with clear distinction between
patients predicted to have high or low probabilities of
BRCA1
/
BRCA2
deficiency. Apart
from the 22 positive controls from the training set, 90/538 additional tumor samples were
identified as having a probability of
BRCA1
/
BRCA2
deficiency exceeding 70%, bringing
the total proportion of patients predicted to have a high level of
BRCA1
/
BRCA2
deficiency
to 20%.
This result prompted us to look for additional
BRCA1/BRCA2
mutations (germline and
somatic) in the cohort of 560 patients. 33 were found to carry pathogenic germline variants
in
BRCA1
/
BRCA2
with corresponding somatic inactivation of the alternative allele. This
more than doubles the number of women harbouring familial cancer predisposition alleles
than originally recruited into the study, carrying significant clinical genetic counselling
implications and potential for active surveillance and/or treatment choices, for affected
patients and their families.
22 patients had early, clonal, somatic
BRCA1
/
BRCA2
mutations (eight) or promoter DNA
hypermethylation of
BRCA1
(fourteen) with inactivation of the alternative allele. The
remaining tumours with probability of BRCA1/BRCA2 deficiency exceeding 70% did not
demonstrate biallelic inactivation of
BRCA1
/
BRCA2,
although DNA methylation data were
not available for a subset of patients.
6 BRCA1-null samples had probabilities of 0.006-0.64 and were missed because the
algorithm had been trained on a small cohort of five
BRCA1
tumours out of the total 22 in
the training set, suggesting that algorithm retraining on a larger and more balanced cohort
was prudent.
HRDetect: Predictor of BRCA1/BRCA2 deficiency in cancer
Given that additional patients were identified as null for
BRCA1
/
BRCA2
, we performed
another iteration of the lasso logistic regression model on a larger, better-powered training
set comprising 77 samples (22 known germline, 33 new germline diagnoses, 22 somatic)
(Figure 2).
Davies et al. Page 5
Nat Med
. Author manuscript; available in PMC 2018 March 02.
Europe PMC Funders Author Manuscripts Europe PMC Funders Author Manuscripts