RES E AR C H Open Access
Lessons learned from additional research
analyses of unsolved clinical exome cases
Mohammad K. Eldomery
1,18†
, Zeynep Coban-Akdemir
1†
, Tamar Harel
1†
,JillA.Rosenfeld
1
, Tomasz Gambin
1,2
,
Asbjørg Stray-Pedersen
3
, Sébastien Küry
4
,SandraMercier
4,5
,DavorLessel
6
,JonasDenecke
7
, Wojciech Wiszniewski
1,8
,
Samantha Penney
1
, Pengfei Liu
1,9
,WeiminBi
1,9
, Seema R. Lalani
1,8
, Christian P. Schaaf
1,8,10
, Michael F. Wangler
1,8
,
Carlos A. Bacino
1,8
, Richard Alan Lewis
1,10
, Lorraine Potocki
1,8
, Brett H. Graham
1,8
,JohnW.Belmont
1,8
,
Fernando Scaglia
1,8
,JordanS.Orange
11,12
, Shalini N. Jhangiani
13
,TheodoreChiang
13
, Harsha Doddapaneni
13
,
Jianhong Hu
13
, Donna M. Muzny
13
, Fan Xia
1,9
, Arthur L. Beaudet
1,9
,EricBoerwinkle
13,14
, Christine M. Eng
1,9
,
Sharon E. Plon
1,8,11,15
,V.ReidSutton
1,8
, Richard A. Gibbs
1,13,16
, Jennifer E. Posey
1
, Yaping Yang
1,9
and
James R. Lupski
1,8,11,13,17*
Abstract
Background: Given the rarity of most single-gene Mendelian disorders, concerted efforts of data exchange
between clinical and scientific communities are critical to optimize molecular diagnosis and novel disease
gene discovery.
Methods: W e de signed and implemented prot ocols f or the study of cases fo r which a plausible molecul ar
diagnosis was not achieved in a clinical gen omics diag nostic labo ratory (i.e. unsolved clinical exomes). Such
cases were recruited to a research laboratory for further analyses, in order to potentially: (1) accelerate novel
disease gene discovery; (2) increase the molecular diagnostic yield of whole exome sequencing (WES); and
(3) gain insight into the genetic me chanisms of disease. Pilot proj ect d ata inclu ded 74 families, consisting
mostly of parent–offspring trios. Analyses performed on a research basis employed both WES from additional
family members and complementary bioin formatics approaches and protocols.
Results: Analysis of all possible m odes of Mendelian inheritance, focusing on both s ingle n ucleotid e variants
(SNV) and copy number variant (CNV) alleles, yielde d a likely contributory variant in 36% (27/74 ) of cases. If
one includes cand idate genes with variants identi fied within a single f amily, a potential contributory variant
was identified in a total of ~51% (38/74) of cases enrolled i n t his pi lot study. The molecular diagno sis was
achieved in 30/63 trios (47.6%). Besides this, the analysis workflow yielde d evidence for pathogenic var iants in
disease-associated genes in 4/6 single ton c ases (66.6%), 1/1 multiplex family involving three af fecte d sibli ngs,
and 3/4 (75%) quartet families. Both the analytical pipeline and the collaborative efforts between t he
diagnostic and research laboratories provided insights that allowed recent disease gene discoveries (PURA,
TANGO2, EMC1, GNB5, ATAD3A,andMIPEP) and increased the number of novel genes, defined in this study
as genes identified in more than one family (DHX30 and EBF3).
(Continued on next page)
* Correspondence: jlupski@bcm.edu
†
Equal contributors
1
Department of Molecular and Human Genetics, Baylor College of Medicine,
Houston, TX 77030, USA
8
Texas Children’s Hospital, Houston, TX 77030, USA
Full list of author information is available at the end of the article
© The Author(s). 2017 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and
reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to
the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver
(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
Eldomery et al. Genome Medicine (2017) 9:26
DOI 10.1186/s13073-017-0412-6
(Continued from previous page)
Conclusion: An efficient genomi cs pi peline in which clinical sequencing in a d iagnos tic lab oratory is followed
by the detailed reanalysis of unsolved cases in a research environment, supplemented w ith WES data from
additional family members, and subject to adjuvant bioinformatics analyses including relaxed variant filtering
parameters in informatics pipelines, can enhance the molecular diagn ostic yield and provide mechani stic
insights into Mendelian disorders. Implementing these approaches requires co llaborative cli nical mol ecular
diagnostic and research efforts.
Background
Applications to clinical practice of whole exome sequen-
cing (WES) and whole genome sequencing (WGS)
technologies and the computational interpretation of
rare variants in genome data have been revolutionary,
allowing conclusions to diagnostic odysseys and enabling
molecular diagnoses for thousands of patients [1–7].
Moreover, such genome-wide assays have enabled insights
into multi-locus contributions to disease [8]. Recent re-
ports document an initial ~25–30% rate of molecular
diagnosis in known disea se genes for patients referred
for exome sequencing and interpretation [3, 5, 9–12].
The remaining un diagnosed individuals may represent:
(1) limitations in concluding a molecular diagnosis
using the current experimental and analytical methods
of clinical genomics practice; or (2) our limited under-
standing of the genetics of human disease. Collabor-
ation between the clinical, clinical molecular diagnostic,
and research communities may optimize discovery of
disease genes, considering the rarity of specific genetic
disorders [13–17].
We performed a pilot study for systematic transfer of
molecularly “unsolved” exomes from the clinical envir-
onment to a research setting, in order to potentially fuel
human genetic disease gene disc overy. The WES data
from 74 probands for whom clinical singleton WES did
not reveal a secure molecular diagnosis were augmented
with WES from addit ional family members, where avail-
able. Additional bioinformatics filters, database re-
sources, and interpretive analyses were implemented,
leveraging systematic studies emerging from the research
laboratory. A likely disease contributory gene and poten-
tial molecular diagnosis (i.e. known disease gene or
novel gene identified in more than one family) was iden-
tified in 36% of the probands, and a candidate gene find-
ing (i.e. identified in a single family) was identified in
15% of patients. This experience and resulting findings
offer the opportunity to systematically compare different
but complementary approaches to optimize molecular
diagnostic yield. Several novel gene discoveries (PURA,
TANGO2, EMC1, GNB5, ATAD3A, MIPEP)[18–23]
were facilitated by this collaborative and systematic clin-
ical/research laboratory approach, and additional novel
disease genes were found in multiple families (DHX30,
EBF3), together highlighting different genetic contribu-
tions to pathogenicity [24–26].
Methods
Recruitment of non-diagnostic clinical exome cases into
research
Probands, whos e DNA had been analyzed at Baylor
Genetics (BG) laboratory for clinical diagnostic WES,
and for whom a molecular diagnosis (defined as a patho-
genic or likely pathogenic variant according to American
College of Medical Genetics and Genomics [ACMG]
guidelines) was not achieved at the time of initial report-
ing, were classified as “unsolved ” and enrolled into the
study [3, 27]. For this pilot study, a total of 74 unsolved
clinical WES cases, analyzed by the clini cal genomics
laboratory between April 2012 and April 2014, were
enrolled in research between July 2013 and March 2015.
Enrollment proceeded in serial order with no specific
inclusion or exclusion criteria, other than the inability to
achieve a molecular diagnosis in the clinical genomics
laboratory and parents consenting to research analyses.
Depending on the clinical situation and individual avail-
ability, additional family members (e.g. parents or
affected siblings) were also enrolled.
Our pilot study consisted of 74 cases including 63
trios, four quartets, one multiplex family involving three
affected siblings, and six singleton cases for which par-
ental samples were unavailable. Prior diagnostic work-up
was variable from case to case based on the referring
physician’s differential diagnosis (e.g. single gene testing,
enzyme assays, array comparative-genome hybridization)
and included a proband-only WES in all cases. Mito-
chondrial DNA sequencing was performed for all ca ses
undergoing clinical WES through December 2014. De-
tailed clinical phenotype data were collected and entered
into PhenoDB [14, 28], after con tacting the families/
patients to obtain informed consent, and did not influ-
ence the choice of cases. However, we retrospectively
analyzed the phenotypic features of this cohort and
found that developmental delay/intellectual di sability
(DD/ID) wa s the most prevalent phenotype in our pilot
study, consistent with the nature of cases referred for
clinical WES [3, 5]. Our phenotypic analysis revealed
59 cases with syndromic DD/ID and one with non-
Eldomery et al. Genome Medicine (2017) 9:26 Page 2 of 15
syndromic DD/ID. Additionally, 14 other phenotypes
encountered in the 74 cases of the pilot study were
metabolic, gastrointestinal, and mitochondrial abnormal-
ities (Additional file 1: Table S1).
Whole exome sequencing and annotation
Exome capture was performed with Nimblegen reagents
using the Baylor College of Medicine (BCM) Human
Genome Sequencing Center (HGSC) custom-designed
capture reagent VCRome 2.1 for both clinical and re-
search laboratory exomes. This capture reagent contains
more tha n 196K targets and 42 Mbp of genomic regions
and includes predicted coding exons from Vega , CCDS,
and RefSeq. Samples are multiplexed (six-plex format)
for both capture and sequencing and full-length blocking
oligos were employed for hybridization to enhance on-
target specificity [3, 5, 29]. Clinical WES targets the cod-
ing exons of ~20,000 genes with 130X average depth of
coverage and greate r than 95% of the targeted bases
having >20 reads [3, 5, 29, 30]. Research WES (for add-
itional fa mily members) had an average depth of cover-
age of 95X, with >92% of the targeted bases having >20
reads. The raw sequence data were post-processed using
the Mercury pipeline [31]. First, the raw sequencing data
(bcl files) were converted to fastq files using Casava.
Then, the Burrows-Wheeler Alignment (BWA) tool was
utilized to map short reads to the human genome refer-
ence sequence (GRCh37). Finally, the recalibration and
variant calling were performed using GATK [30] and the
Atlas2 suite, respectively [32]. The Mercury pipeline is
available in the cloud via DNANexus (http://blog.dna-
nexus.com/2013-10-22-run-mercury-variant-calling-pipe-
line/). For research cases, exome variant analyses were
then independently performed in the Baylor Hopkins
Center for Mendelian Genomics (BHCMG) [15] und er a
Research Protocol approved by the Institutional Review
Board (IRB) for Human Subje cts Research at BCM; this
protocol enables bi-directional transfer of samples and
data between the clinical and research laboratories.
Identification of de novo single nucleotide variants (SNVs)
in BHCMG
De novo variants were identified by an in-house developed
software called DNM (de novo mutation)-Finder (https://
github.com/BCM-Lupskilab/DNM-Finder), available upon
request. Parental variants were subtracted in silico from
the proband’s variants in vcf f iles , w hile incorporating
read numbe r information extracted from BA M files.
Filtering was then implemented using the following
criteria: (1) an a lternative variant read count greater
than 5 in the proband; (2) ratio of alternative variant
read count to reference variant read count greater
than 30% in the proband; (3) reference variant read
count greater than 10 in both parents; and (4) ratio
of alternative variant read count to reference variant
read count less than 5% in both parents.
SNV prioritization and filtering workflow
We designed a stepwise analysis workflow. This workflow
consists of four major steps of analysis that examined: (1)
recessive homozygous predicted loss-of-function (LOF) var-
iants (stop-gain, frameshift indels, and splice site variants)
and/or missense variants; (2) compound heterozygous LOF
and/or missense variants; (3) heterozygous LOF variants,
which facilitated the detection of potential truncating de
novo mutations in proband-only WES cases; and (4) de
novo variants and potential parental mosaic variants using
trio-WES (Fig. 1a). Variants were prioritized for further
study based on their minor allele frequencies (MAF; <0.5%)
and the output from several prediction software algorithms
that identified cross-species conservation or variant effects
on protein function (Fig. 1a) [33]. Subsequently, each vari-
ant was further filtered based on additional gene and
variant-level information from publicly available databases:
Online Mendelian Inheritance in Man (OMIM: http://
www.omim.org); PubMed; the Human Gene Mutation
Database (HGMD: http://www.hgmd.cf.ac.uk); and ClinV ar
(https://www.ncbi.nlm.nih. gov/clinvar/). This sequential
analysis eliminates the bias towards a particular mode of in-
heritance, mutation type, and the prior knowledge of the
role of a specific gene in human disease. Additionally, this
approach also integrates the fundamentals of WES analysis
based on previously described SNV workflows and is com-
patible with ACMG guidelines for SNV filtering and classi-
fication of variant pathogenicity [3, 9, 27]. Analysis of
interactomes, gene families, protein–protein interactions,
co-expression data, and available literature further priori-
tized candidate variants (Fig. 1b). The patterns of segrega-
tion of the remaining candidate variants were analyzed by
Sanger sequencing in available family members. Finally,
the most promising candidate genes/variants were used to
interrogate local databases and GeneMatcher [13, 14] to
find additional affected cases with damaging variants in
the same gene and with similar clinical phenotypes. When
the BHCMG identified suspected causative variants in ei-
ther n ewly discovered or recently published known
disease genes , it wa s communicated to the clinical ex-
ome laboratory. Sanger confirmation of the suspe cted
pathogenic variant was performed for all of the identified
variants (confirmation rate o f 100%) under the aus-
pices of the C AP/CLIA-certified diagnostic laboratory
and a formal updated clinical report wa s issued. Simi-
larly, identified variants both in novel genes and in re-
cent r eports from the clinical exome laboratory were
reported back to BHCMG, creating a bidire c tional
channel of intercommunication between research and
clinical laboratories.
Eldomery et al. Genome Medicine (2017) 9:26 Page 3 of 15
We also designed a parallel computational analysis
of “bulk data” in order to accelerate our WES data
analyses by scanning the BHCMG database for rare
homozygous/heterozygous stop-gain variants culled from
WES data of ~5000 research participants including our
pilot study cases. In this analysis, we targeted the rare
(MAF <0.5%) homozygous stop-gain variants from ~5000
research participants (including the 74 families) for
each ge ne. Genes were then sorted according to the
number of homozy gous ra re stop-gain variants exist-
ing in our database . Second, the genes were analyzed
further according to SN V prioritization and filtering
workflow as described above (Fig. 1). This approach
was designed to accelerate the discovery of novel
disease genes that exhibit phenotypic consequences
through a loss-of-function mechanism.
Detection of CNVs
CNV detection from WES data has been employed by dif-
ferent clinical laboratories to improve the molecular diag-
nostic rate [34, 35]. We applied several computational
algorithms (CoNVex, Sanger Centre [ftp://ftp.sanger.ac.uk/
pub/users/pv1/CoNVex/Docs/CoNVex.pdf], CoNIFER, and
XHMM) to WES data to identify potential disease associ-
ated CNVs; these tools detect a clinically relevant intra-
genic CNV when at least three contiguous exons are
deleted [36, 37]. Therefore, in addition to these algo-
rithms, we developed an in-house pipeline HMZDelFinder
Fig. 1 Analysis of WES data. a SNVs were filtered and prioritized according to specific criteria, including mode of inheritance, mutation type,
variant frequency, conservation, and predictions of pathogenicity. b Candidate genes were further prioritized by data mining, taking into
consideration gene function, expression, and networks. In addition, other cohorts were interrogated for additional families with variants in the
same candidate gene. MutationMapper (http://www.cbioportal.org/mutation_mapper.jsp), ARIC Atherosclerosis Risk in Communities Study, AR-
Hom autosomal recessive-homozygous, BHCMG Baylor-Hopkins Center for Mendelian Genomics, BG Baylor Genetics laboratories, CNV copy number
variation, Comp compound, db database, ExAC Exome Aggregation Consortium, Het heterozygous, HGMD Human Gene Mutation Database, MAF minor
allele frequency; SNV single nucleotide variant, XLR-Hem X-linked recessive-hemizygous
Eldomery et al. Genome Medicine (2017) 9:26 Page 4 of 15
(https://github.com/BCM-Lupskilab/HMZDelFinder) [38]
to detect potential homozygous and hemizygous small in-
tragenic deletions from WES data, including single exon
“dropout alleles” that may be less robustly identified by
current software.
Results
The findings were sorted into three major categories: (1)
known disease genes; (2) novel genes: genes with pre-
dicted pathogenic variants in two or more cases with
similar phenotypes; and (3) potential candidate genes:
genes with variants in a single case (Table 1). Predicted
pathogenic variants in known or novel disease genes
were identified in 27/74 (36%) families studied. In
addition, predicted pathogenic variants in candidate
genes were identified in 11/74 (15%) of cases, yielding an
overall potential solved rate of 38/74 (51%) of cases. Our
findings are represented by six “lessons learned” clusters:
(1) de novo changes in known genes; (2) de novo
changes in novel and candidate genes; (3) potential
mosaicism in parents; (4) biallelic or hemizygous vari-
ants in genes known to convey recessive disease traits;
(5) biallelic variants in novel and candidate recessive
disease genes; and (6) blended phenotypes resulting
from dual molecular diagnoses (Fig. 2a, b). Examples for
each of these scenarios are presented below.
De novo changes in known genes
Analysis of WES data from sets of parents and offspring
initially yielded ~50–200 putative de novo variants per
trio. Further prioritization, based upon read coverage,
MAF and mutation type (non-synonymous, stop-gain,
frameshift indels, and splicing variants), reduced the
number of potential pathogenic de novo variants per
family to ~0–5 variants per proband (Additional file 2:
Figure S1). We detected de novo variants in five recently
published genes (Table 1, Additional file 3: Table S2):
ZBTB20 associated with Primrose syndrome [39] (MIM
259050); NR2F1 causing the Bosch-Boonstra-Schaaf
optic atrophy syndrome [40] (MIM 615722); DDX3X
associated with X-linked intellectual disability [41]
(MIM 300958); CACNA1A implicated in non-fluctuating
ataxia [42]; and NALCN associated with congenital con-
tractures of the limbs and face, hypotonia, and develop-
mental delay [43] (MIM 616266).
De novo changes in novel and candidate genes
We further identified de novo variants in five novel and
ten candidate genes (Table 1, Additional file 3: Table S2).
A de novo SNV in DHX30 c.2344C > T (p.Arg782Trp)
was found in a patient with microcephaly, developmen-
tal delay/intellectual disability (DD/ID), mild cerebral
volume loss, hypotonia, seizures, short sta ture, failure to
thrive, and generalized hirsutism. Through data ex-
change [3, 5] with the clinical exome laboratory (Baylor
Genetics) and GeneMatcher [13, 14], we found three
additional unrelated participants with overlapping fea-
tures, each of whom harbored a de novo SNV in
DHX30. Interestingly, two of these part icipants shared
the same de novo SNV c.2344C > T, occurring at a CpG
dinucleotide (Fig. 3a) [44]. DEAH (Asp-Glu-Ala-His)
box helicase 30 (DHX30, MIM 616423) is a member of
the RNA helicase family including DHX and DDX pro-
teins involved in DNA transcription, splicing and trans -
lation [45–47], and is highly expressed in the brain
during neurogenesis [45]. Homozygous knockout of
Dhx30 in mice is leth al [45]. DDX3X, another member
of this family, has been recognized recently to play a
major role in DD/ID [41]. Additionally, DHX37 has been
proposed as a candidate gene for brain malformations
[48], highlighting the importance of RNA helicases in
development of the central nervous system.
GSPT2 and EBF3 each harbored one de novo variant
in the genomes from two separate probands (Table 1,
Additional file 3: Table S2). Each of these genes
is located within CNV intervals identified previously
Table 1 Molecular diagnoses in 74 cases are represented as three major categories: known genes, novel genes and candidate genes
Inheritance Known genes Novel genes Candidate genes
De novo CACNA1A, DDX3X(X2)
a
, NALCN(X2),
NR2F1
a
, ZBTB20
ATAD3A, DHX30, EBF3, EMC1, PURA
a
CDK20 + HIVEP1, DNAH7, GSPT2 GUCY2C,
MICALL2 + SLC30A7, MPP4, SYN3, SYTL2
Autosomal/X-linked Recessive ABCA4, DDX3X, FBXL4
a
, NAA10,
SLC13A5
a
(X2), TRAPPC11, ZNF335
a
GNB5, MIPEP, TANGO2
a
ACOT1
a
, NRXN3, USP19
Other NA NA
UPD SLC1A4
a
Mosaic PIK3CD
Dual molecular diagnosis PMPCA + KCND3
a
POLRIC + SCNIB
a
a
Cases independently solved by the clinical exome laboratory re-analysis
Three major categories of identified genes include 13 candidate genes identified in 11 families and 26 known or novel genes in 27 families. Note that some
families had more than one molecular diagnosis (indicated by PMPCA + KCND3, POLR1C + SCN1B, CDK20 + HIVEP1, MICALL2 + SLC30A7) and some genes were
identified in more than one family (indicated by “(X2)” in the table)
NA non-applicable
Eldomery et al. Genome Medicine (2017) 9:26 Page 5 of 15