ARTICLE
Received 18 Aug 2014 | Accepted 18 Nov 2014 | Published 9 Jan 2015
DOI: 10.1038/ncomms6901
OPEN
Combining gene mutation with gene
expression data improves outcome prediction
in myelodysplastic syndromes
Moritz Gerstung1,*, Andrea Pellagatti2,*, Luca Malcovati3,4, Aristoteles Giagounidis5, Matteo G. Della Porta3,6,
Martin Jädersten7, Hamid Dolatshad2, Amit Verma8, Nicholas C.P. Cross9, Paresh Vyas10, Sally Killick11,
Eva Hellström-Lindberg7, Mario Cazzola3,4, Elli Papaemmanuil1, Peter J. Campbell1 & Jacqueline Boultwood2
Cancer is a genetic disease, but two patients rarely have identical genotypes. Similarly,
patients differ in their clinicopathological parameters, but how genotypic and phenotypic
heterogeneity are interconnected is not well understood. Here we build statistical models to
disentangle the effect of 12 recurrently mutated genes and 4 cytogenetic alterations on gene
expression, diagnostic clinical variables and outcome in 124 patients with myelodysplastic
syndromes. Overall, one or more genetic lesions correlate with expression levels of B20% of
all genes, explaining 20–65% of observed expression variability. Differential expression
patterns vary between mutations and reflect the underlying biology, such as aberrant
polycomb repression for ASXL1 and EZH2 mutations or perturbed gene dosage for
copy-number changes. In predicting survival, genomic, transcriptomic and diagnostic clinical
variables all have utility, with the largest contribution from the transcriptome. Similar
observations are made on the TCGA acute myeloid leukaemia cohort, confirming the general
trends reported here.
1 Wellcome
Trust Sanger Institute, Hinxton CB10 1SA, UK. 2 LLR Molecular Haematology Unit, NDCLS, RDM, University of Oxford, Oxford OX3 9DU, UK.
of Hematology Oncology, Fondazione IRCCS Policlinico San Matteo, 27100 Pavia, Italy. 4 Departments of Molecular Medicine and Internal
Medicine and Medical Therapy, University of Pavia, 27100 Pavia, Italy. 5 Department of Hematology, Oncology, and Palliative Care, Marienhospital
Düsseldorf, 40479 Düsseldorf, Germany. 6 Department of Internal Medicine and Medical Therapy, University of Pavia, 27100 Pavia, Italy. 7 Division of
Hematology, Department of Medicine, Karolinska Institutet, SE-171 76 Stockholm, Sweden. 8 Albert Einstein College of Medicine, Bronx, New York 10461,
USA. 9 National Genetics Reference Laboratory, Salisbury NHS Foundation Trust, Salisbury SP2 8BJ, UK. 10 MRC Molecular Haematology Unit, Weatherall
Institute of Molecular Medicine, University of Oxford, Oxford OX3 9DS, UK. 11 Department of Haematology, Royal Bournemouth Hospital, Bournemouth BH7
7DW, UK. * These authors contributed equally to this work. Correspondence and requests for materials should be addressed to P.J.C. (email:
[email protected]) or to J.B. (email: [email protected]).
3 Department
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
1
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
I
t is well recognized that patients diagnosed with the same
cancer often have different constellations of pathological
parameters, frequently associated with distinct clinical and
prognostic features. In the last decade, molecular testing has
increasingly penetrated clinical practice, largely through screening
for driver mutations. Gene expression profiling has uncovered
many systematic differences between cancer and normal cells and
has enabled the definition of new, clinically relevant disease
subtypes. From a diagnostic standpoint, such expression profiles
are used as ‘biomarkers’, with no intent to address causality: it
suffices to correlate with the clinical feature of interest. Genomic
information occupies a fundamentally different niche in that the
screen assays causal, or driver, genes1,2. The vehicle by which
driver mutations cause cancer, however, is transcription acting
through an intricate cellular signalling circuitry that links the
genomic variants to the clinical phenotype of a cancer.
The myelodysplastic syndromes (MDS) represent a heterogeneous group of chronic blood cancers. MDS is characterized by
ineffective haematopoiesis resulting in peripheral cytopenias, and
patients typically have a hypercellular bone marrow3,4. About
30–40% of patients evolve to acute myeloid leukaemia (AML)
over months to years after diagnosis. The molecular pathogenesis
of MDS is increasingly well understood, charting the sets of
commonly mutated genes, chromosomal aberrations and gene
expression changes5–7. The most commonly mutated genes in
MDS are regulators of RNA splicing and epigenetic modifiers, but
signal transduction pathways and transcription factors are also
frequent targets6,8–11. In addition, MDS is characterized by
frequent cytogenetic aberrations such as deletions on the long
arms on chromosomes 5, 7 and 20, as well as more complex
karyotypes.
Many of the consequences of genetic and cytogenetic
alterations will affect gene expression by means of aberrant
transcription, epigenetic regulation, cell signalling and gene
dosage effects. We and others have identified many deregulated
genes and gene pathways in MDS using gene expression
profiling12–17. A fundamental limitation of these studies is the
unknown genetic background of the samples. As the recurrence
of mutations is typically lower than 10%, expression changes in
such small and unknown subgroups could not be reliably
mapped.
We recently published a large mutation screen of 111 cancer
genes in 738 MDS patients presenting a comprehensive map of
the mutational landscape of myelodysplasia11. Here we link these
genomic data with gene expression microarray data for 159 MDS
cases and 17 normal samples, which extend an existing cohort of
116 MDS cases previously published without mutation data12.
We deconvolute the expression of genes into contributions
stemming from each genetic and cytogenetic mutation, which
provides deep insights into how driver mutations interfere with
the transcriptomic state. We model the influence of mutations
and expression changes on diagnostic clinical variables as well as
survival and find that the transcriptome appears to be the most
powerful predictor of outcome.
Results
Mutation patterns correlate with global gene expression.
Variation in gene expression across patients with a particular type
of cancer may result from a number of factors. For this study, we
were primarily interested in isolating and defining the aggregate
effects of driver mutations on the transcriptome, noting that age,
sex, germline genetic background and other host factors may also
contribute as nuisance factors. We combined data from genomic
profiling of 111 relevant cancer genes11 with microarray gene
expression data from CD34 þ bone marrow cells of 159 MDS
2
patients and 17 normal individuals in total. Combined expression
and mutation data were available for 124/159 MDS patients
(Table 1; Supplementary Data 1). Outcome data were also
available and are released with the gene expression data (GEO
accession GSE58831). In addition, we release all code associated
with implementing the detailed statistical analyses that follow
(Supplementary Data 2), in order that this study can be replicated
on this data set and extended to other tumour types.
To obtain an overview of the main patterns of expression
changes, we first performed a principal components analysis
(PCA; Fig. 1a). This multivariate statistical technique collapses
multidimensional correlated data such as gene expression of more
than 20,000 genes into a smaller set of mutually uncorrelated
variables, ordered such that the first few principal components
explain the greatest amount of variation in the data. The PCA was
computed on all 176 cases with expression data to maximize the
stability of the components. In our data, the first two principal
components (PCs), respectively, account for 14.4% and 7.8% of
the total variability in gene expression; the first 20 PCs
cumulatively explain 67% of the variance (Supplementary
Fig. 1). The expression changes associated with PC1 are
dominated by genes related to haematopoietic differentiation
genes; for example, the stem cell factors KIT, CD34 and also FLT3
have positive values in PC1 while the monocyte-specific antigens
CD14 and CD163 as well as members of the a and b globin gene
clusters have negative values. The second component PC2 had
low levels of multiple chemokines and high levels of eosinophiland neutrophil-related genes as well as the haematopoietic
transcription factor KLF1. Notably, the observed principal
components did not lead to clearly separated groups of patients,
but rather to a continuum of expression changes (Fig. 1a).
Strikingly, overlaying the status of 12 recurrent (Z5 patients)
genetic and 4 cytogenetic alterations on to the first two principal
components demonstrated that driver mutations are correlated
with general gene expression profiles (Fig. 1a). For example,
patients with mutations in the RNA splicing factor SF3B1 tend to
have low scores on the first principal component, whereas
patients with mutations in two other splicing factors, SRSF2 and
ZRSR2, have high scores. Similarly, STAG2 mutations coincide
with extremely high values in the first principal component.
These general associations, however, do not necessarily imply
causation—STAG2 mutations tend to co-occur with SRSF2
mutations and more frequently in refractory cytopenia with
multilineage dysplasia and refractory anaemia with excess
blasts18, either of which may explain the correlation with a
given PC.
A linear model to deconvolute gene expression and mutations.
To explore predictors of gene expression in a multivariate
framework, we developed a linear modelling approach that
measures the association of expression levels on a gene-by-gene
basis with a number of potential predictors, including driver
mutations and nuisance variables (Fig. 1b). The normal samples
were included to identify changes common to all MDS samples.
Somatically acquired mutations and cytogenetic lesions were
encoded as being present/absent. The model assumes that each
mutation is associated with a certain set of expression changes
and that the expression pattern in cases with a complex genotype
comprising multiple alterations is the sum of the changes induced
by each mutation. We chose a linear model due to its interpretability and established statistical methods, enabling us to test
which transcripts are deregulated in the presence of specific
alterations, after correcting for other confounding variables such
as the nuisance factors and coexisting driver mutations. The
additivity assumption ignores potential interactions between
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
genetic lesions that may arise from the cellular signalling circuitry. However, inferring these interactions systematically from
the data would require many more cases, as the number of
gene:gene interaction pairs is quadratic and yields combinatorially many terms for higher order interactions. Hence, this
assumption appears necessary for statistically robust inference in
a data set of this size.
The transcriptome of MDS is globally perturbed by genetic
and cytogenetic driver mutations, with expression levels of
4,072/21,382 (19%) genes significantly associated with at least one
driver mutation, after correction for multiple hypothesis testing
(FDR-adjusted moderated F-statistico0.05; Supplementary
Data 3). For these genes, genomic alterations accounted for at
least R2 ¼ 20.3% of the observed inter-patient gene expression
variability; the strongest association of R2 ¼ 65% between
mutations and expression changes was observed for the iron
transporter ABCB7 (Fig. 1c). The observed variability can be
largely explained by the presence of SF3B1 mutations and del(5q)
leading to strong downregulation of ABCB7 mRNA (Fig. 1d).
Inherited missense mutations in ABCB7 have been linked to
hereditary X-linked sideroblastic anaemia and ataxia (OMIM
#301310)19, and the mechanistic role of ABCB7 in refractory
anaemia with ring sideroblasts (RARS) has been demonstrated20.
Similarly, somatic SF3B1 mutations were found to correlate
strongly with the presence of ring sideroblasts21, and our data
suggest an interesting three-way association across SF3B1
mutation, ABCB7 downregulation and the occurrence of ring
sideroblasts.
Our linear model also allows us to take a mutation-centric
view, specifying the set of gene expression changes that correlate
with a given driver alteration. The number of target genes whose
expression is differentially affected varies widely across the
different driver variants. For example, del(5q) is independently
correlated with expression levels of 741 genes; there were 605
target genes for patients with SF3B1 mutations, whereas driver
mutations in TET2 and DNMT3A altered expression of only 25
and 11 genes, respectively (Fig. 1e; Supplementary Data 3). These
striking differences cannot be explained by variable statistical
power, as all four of these genetic lesions are among the six most
common driver mutations in MDS11. Compared with normal
samples, 502 genes change in expression in MDS samples without
being attributable to a distinct driver mutation. Finally, the
expression of 58 genes correlated with sex, with the most
significant effect occurring in the XIST gene responsible for X
chromosome dosage compensation.
The genomic landscape of myeloid malignancies is characterized by a secondary structure with striking patterns of cooccurrence or mutual exclusivity with other driver genes11,22–24.
Table 1 | Clinical characteristics of the cohort.
Samples
MDS with expression data
MDS with sequencing data
Normals
159
124/159
17
Demographics
Gender
Age
102 male; 57 female
67 years median; 19–87 years range; 5 missing
MDS classification
RA
RARS
RARS-T
RCMD
RCMD-RS
RAEB
5qCMML
MDS-AML
missing
13
14
6
27
22
56
6
7
7
1
Blood and bone marrow counts at diagnosis
Haemoglobin
Platelets
Absolute neutrophil count
Bone marrow blasts
Ring sideroblasts
Serum ferritin
M:E ratio
9.5 g dl 1 median; 4.5–14.6 g dl 1 range; 15 missing
154 109 l 1 median; 10–45,000 109 l 1 range; 8 missing
1.8 109 l 1 median; 0.08–920 109 l 1 range; 20 missing
4% median; 0–63% range; 17 missing
0.5% median; 0–94% range; 31 missing
543 ng ml 1 median; 8–11,300 ng ml 1 range; 64 missing
2.03 median; 0.33–9 range; 70 missing
Outcome and follow-up
Follow up time
Outcome
AML transformation
1,040 days median; 0–3,141 days range; 36 missing
82 alive; 41 died; 36 missing
14 positive; 101 negative; 44 missing
Mutations
Number of recurrent cytogenetic lesions per patient
Number of recurrent point mutations and indels per patient
Variant allele fraction of point mutations and indels
0
81
0
31
Min
0.04
1
33
1
42
25%
0.27
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
2
3
2
30
50%
0.38
3
0
3
19
75%
0.45
Missing
7
4
2
Max
0.97
3
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
Principal components of gene expression
RUNX1
n=8
n=33
ASXL1
n=16
Expression Y = Genotype X × Effects β
DNMT3A
n=21
141 Samples
n=13
Hemoglobin expression
U2AF1
TP53
n=8
EZH2
n=6
n=7
STAG2
n=8
141 Samples
19 Variables
Decompose
expression of
each gene into
contributions
from variables
12 Driver
genes
4 Cytogenetic
Gender
Age
21,382 Genes
PC1
SRSF2
19 Variables
PC2
n=37
TET2
21,382 genes
Chemokine expression
SF3B1
Test which
Genes respond
to genotype
Normal
Variables affect
expression
del(5q)
n=5
tri(8)
n=20
del(12p)
n=7
Differentially expressed
genes by mutations
n=6
10
Observed
Random
Normal
n=6
Gender
n=17
Age
n=53
Density
8
del(20q)
Missing
n=60
Wild type
Mutant
6
4
4,072 Genes q < 0.05
2
ABCB7
Normal
0
Female
0.0
Age>median
0.2
0.4
0.6
6
56
83
58
72
60
14
21
67
11
M
*
*
*
*
*
*
*
*
* *
*
*
*
*
* *
* * * *
*
*
*
*
*
*
*
Correlated
1000
100
10
1
0.1
0.01
0.001
Odds ratio
2
50
1
33
6
18
3
9
12
6
11
15
9
10
SF
Model coefficients
6
–1
5
7
8
0
1
9
Predicted ABCB7 expression
ut
Ex atio
pr ns
es
si
on
1
74
5
60
25
0
3B
TE 1
SR T2
S
A F2
D SX
N L1
M
R T3A
U
N
U X1
2A
F
TP 1
EZ 53
ST H2
A
ZR G
SR2
JA 2
de K2
l(5
q)
de tri(8
l( )
de 12
l( p)
N 20q
or )
G ma
en l
de
r
Ag
e
Differentially expressed genes
100
del(20q)
del(12p)
tri(8)
del(5q)
JAK2
TP53
DNMT3A
TET2
STAG2
RUNX1
EZH2
ASXL1
SRSF2
U2AF1
ZRSR2
SF3B1
Overlap of expression targets
1
0
–1
500
200
7
Significant coefficients
of splice factors
U2AF1
SF3B1
ZRSR2
SRSF2
591
Exclusive
Not sig.
Q < 0.1
* P < 0.05
SF
ZR 3B
S 1
U R2
2
SR AF
1
ASSF
2
EZXL1
R H
U 2
ST NX
AG 1
D TE 2
N T
M 2
T3
TP A
JA 53
de K2
l(5
q
de tri )
l (8
de (12 )
l(2 p)
0q
)
Log FC
600
300
8
Co-occurrence of mutations
Significant mutation-expression
coefficients (q<0.05)
400
R2 = 0.65
9
6
Explained variance per gene R2
700
10
SF
3B
1
de
l(5
q)
JAK2
n=5
Observed ABCB7 expression
ZRSR2
Figure 1 | Patterns of mutation and differential expression. (a) Scatter plot of the first two principal components of gene expression data of 159 MDS and
17 normal samples overlaid with mutation status of the 12 most frequent point mutations, 4 cytogenetic alterations, as well as normal status, gender
and age above or below median (67 years). (b) Schematic linear decomposition of expression data by driver mutations and demographic variables.
(c) Distribution of the variance explained by genetic and cytogenetic alterations across genes (moderated F-test; FDRo0.05; n ¼ 141). (d) Scatter plot of
expression predictions for the ABCB7 gene versus observed expression values. The inset shows the model coefficients indicating the predicted magnitudes
of expression changes when a given alteration is present. (e) Statistically significant mutation expression interaction terms (moderated t-test; FDRo0.05;
n ¼ 141) for each alteration and demographic variable. The associated logarithmic expression fold change is indicated by colour. (f) Heatmap of observed
pairwise mutation patterns (odds ratio; upper triangle) and overlap of differentially expressed genes associated with each alteration (lower triangle).
Green/blue colours denote preferential co-mutation/high overlap, while pink/red colours indicate mutual exclusivity. (g) Venn diagram of differentially
expressed genes associated with spliceosome mutations.
The most frequently invoked explanation for a pair of genes
exhibiting mutual exclusivity is functional redundancy of the two
genes, potentially because they act in the same pathway25,26. We
therefore compared, for each pair of driver mutations, the extent
to which their sets of target genes overlapped (Fig. 1f). Gene pairs
that tend to be mutated together also show greater degrees of
overlap in downstream transcriptional changes than expected by
chance, and vice versa for abnormalities that tend to be mutually
exclusive. Four genes involved in RNA splicing are mutated in
MDS: SF3B1, SRSF2, U2AF1 and ZRSR2. Despite these genes
showing highly significant patterns of mutually exclusive
mutation11,23,24, we see a remarkably small overlap in their
consequences on gene expression (Fig. 1g). Thus, although
mutations in different genes within the same pathway may exhibit
similar oncogenic potential, this does not necessarily imply that
their transcriptional consequences will be equivalent. Indeed,
the lack of overlap between the transcriptional consequences of
4
mutations in different splicing factors may explain why they have
such profoundly different clinical phenotypes, with SF3B1
mutations driving a benign MDS associated with ring
sideroblasts and SRSF2 mutations driving a more aggressive
myelomonocytic disease. These data further suggest that the
reason mutations in splicing factors tend to be mutually exclusive
may not be because they are functionally redundant. It is possible,
for example, that more than one mutated splicing gene results in
a functional disadvantage.
Genetic mechanisms of mutation-induced expression changes.
The nature of differentially expressed genes associated with a
given driver abnormality may provide insight into its mode of
action. For driver point mutations, the number of differentially
expressed genes per chromosome broadly follows the gene
density on the autosomes (Fig. 2a). In contrast, for cytogenetic
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
Expression of driver genes
SF3B1
TET2
8.0
7.5
7.0
6.5
6.0
5.5
5.0
16(2)
X(19)
12
20(27)
14(3)
15(3)
19(2)
16(2)
17(3) 18(2)
11
11
10
9
10
9
8
or
m
11 al
6
8 wt
m
ut
or
m
11 al
7
7 wt
m
ut
N
N
17
17
Expression
N
or
m
10 al
3
21 wt
m
ut
17
Expression
N
or
m
11 al
6
6 wt
m
ut
17
ZRSR2
*
JAK2
**
*
9
8
7
6
5
4
9
8
7
6
5
Upregulated
Active promoter
Weak promoter
Poised promoter
Strong enhancer
Weak enhancer
Insulator
Txn transition
Txn elongation
Weak Txn
Repressed
Heterochrom/lo
Repetitive/CNV
5
0
–5
–10
–15
Downregulated
H3K27me3 TSS signal
Chromatin state:
10
H3K27me3 and expression of target genet genes
in normal CD34+ cells
***
**
***
***
20
15
Average target expression
15
10
5
2
1
0.5
0.2
an
3B
T 1
SR ET
S 2
A F
D SX 2
N L
M 1
R T3A
U
N
U X1
2A
F
TP 1
EZ 53
ST H
2
ZRAG
SR2
JA 2
D K
el 2
(5
q
D tr )
el i(8
(1 )
D 2q)
el
N (20
o q
G rm )
en al
de
r
Ag
e
SF
do
m
EZ
H
2
AS
XL
1
0.1
–20
R
Megabases
Expression
STAG2
***
***
8
ENCODE chromatin states of target genes in K562 cells
17
17
EZH2
*
10(4)
11(8)
N
or
m
10 al
8
16 w
m t
ut
m
91 al
33 wt
m
ut
N
or
m
87 al
37 wt
m
ut
Expression
12(4)
Expression
10(3)
N
or
10(2)
4(2) 1(3)
5(2)
7(3)
8(3)
9(3)
or
m
11 al
9
w
5 t
m
ut
7(4)
Y(15)
*
6.5
6.0
5.5
5.0
4.5
4.0
N
6(3)
15(2)
14(2)
13(3)
12(3)
1(2)
6(2)
9.0
8.5
8.0
7.5
7.0
6.5
6.0
17
16(6)
5(2)
3(4)
8(36)
4.8
4.6
4.4
4.2
4.0
3.8
3.6
3.4
17
5(3)
6(2)
7(4)
del(20q)
22(2)
20(2)
18(2)
17(2)
10 11
Gender
6
TP53
*
or
m
11 al
9
5 wt
m
ut
4(4)
8 9
8.0
7.5
7.0
6.5
6.0
5.5
7
U2AF1
*
N
2(8)
3(2)
7
RUNX1
**
**
***
Expression
X(4)
6
12(2)
4(9)
del(12p)
1(3)
16
15
14
13
12
8
5
17
12(32)
11(27)
10(26)
9(24)
6(73)
7(20) 8(20)
17(2)
DNMT3A
*
17
5
Expression
5(159)
1(5)
19
3
9
12.5
12.0
11.5
11.0
10.5
10.0
N
or
m
11 al
6
w
8 t
m
ut
17(33)
16(23)
15(15)
14(19)
2(4)
20
ASXL1
*
17
1(42) X(30)
22(21)
19(34)
X 22
1
2
Expression
2(41)
3(36)
4(16)
9.5
4
tri(8)
del(5q)
10.0
# Genes
17
4(21)
16(31)
16(11)
5(12)
4(20)
15(22)
15(16)
5(17)
14(23)
6(24 )
14(13)
13(15)
6(38)
13(8)
7(21 )
12(37)
7(22)
12(13)
8(8)
8(19)
11(26)
11(18)
9(11) 10(16)
9(16) 10(31)
3(39)
10.5
Expression
2(25)
3(12)
11.0
N
or
m
11 al
1
13 wt
m
ut
17(44)
11.5
17
2(52)
X(20)
22(8)
20(12)
19(11)
17(19)
1(25)
22(15)
19(22)
Expression
X(24)
1(67)
STAG2
Expression
12.0
SF3B1
SRSF2
Expression
**
N
or
m
11 al
6
8 wt
m
ut
Chromosomal distribution of target genes
wt
wt
10
5
Any EZH2 ASXL1
Figure 2 | Genetic mechanisms of differential expression. (a) Pie charts illustrating the chromosomal locations of differentially expressed genes.
(b) Expression boxplots of mutated genes in normal, wild-type MDS and MDS samples in which the indicated gene is found to be mutated. The lines
denote the expression predicted by the linear model and account for multivariate effects. Stars indicate significance, adjusted to the number of shown
comparisons. (***Po10 3, **Po10 2, *Po10 1; FDR-adjusted moderated t-test; n ¼ 141). (c) ENCODE chromosome states in K562 myeloid leukaemia
cells for differentially expressed genes for each driver alteration. The y-axis denotes the sum of the lengths of all differentially expressed genes in
Megabases. (d) Left: NIH roadmap H3K27me3 ChIP-seq enrichment at transcription start sites of genes differentially expressed in EZH2 and ASXL1 mutant
MDS. Right: observed expression changes associated with EZH2 and ASXL1 compared with 1,000 randomly selected genes.
abnormalities, the largest share of expression changes occurs at
the deleted or amplified genomic locus as a result of the altered
gene dosage. We also find, however, an appreciable set of
secondary, trans effects on other chromosomes. Reassuringly,
sex-specific effects predominantly localized to the X and Y
chromosomes.
The expression of the mutated gene itself also carries clues
about the pathogenicity and whether alternative mechanisms
such as epigenetic silencing or deregulation are acting in those
cases where the driver is not mutated. We observed lower
expression levels of SF3B1, SRSF2, TP53 and STAG2 in mutated
MDS cases compared with wild-type cases, consistent with
previous observations27 (Fig. 2b). The effect was most striking for
STAG2; a reduced expression of the STAG2 protein in mutant
cases has been reported in myeloid cell lines, but restricted to the
chromatin-associated protein fraction18. Conversely, RUNX1 was
more strongly expressed when mutated, which may indicate
either that RUNX1 mutations more frequently occur in cell types
with high RUNX1 activity, or that RUNX1 mutations perturb the
autocatalytic RUNX1 signalling network28.
The ENCODE consortium has provided a rich molecular
annotation of the genome in multiple cell types including K562
chronic myeloid leukaemia cells and the Gm12787 lymphoblastoid haematopoietic progenitor cell line. We explored the
distribution of genomic states among differentially regulated
genes, using ENCODE data on K562 cells29 (Fig. 2c).
Downregulation typically occurs in transcriptionally active
genes, whereas upregulation can also affect genes found to be
silenced and heterochromatic. This latter was especially
pronounced for the set of transcriptional changes associated
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
5
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
with ASXL1 and EZH2 mutations. The histone-lysine
N-methyltransferase EZH2 is part of the Polycomb Repressive
complex 2 (PRC2) and trimethylates histone H3 lysine 27
(H3K27) to silence chromatin; this process is facilitated by ASXL1
(ref. 30). We find that genes deregulated by EZH2 and ASXL1
mutations are only weakly expressed in normal cells and show an
enrichment of repressive H3K27me3 signal in data of normal
CD34 þ cells obtained from the NIH epigenome roadmap
consortium31 (Fig. 2d). Thus, driver mutations in EZH2 and
ASXL1 lead to a derepression of certain polycomb group target
loci leading to an increased expression in mutated cases. The loss
of PRC2 repression in ASXL1 mutant cell lines has been
functionally demonstrated in cancer cell lines32, as well as in
human CD34 þ cells following ASXL1 knockdown33, and it is
compelling to observe this mechanism in primary patient
samples.
measure the aggregated contribution of genetic, cytogenetic and
transcriptomic variables.
We exemplify these analyses with the use of genomic and
transcriptomic variables to predict the fraction of ring sideroblasts in the marrow (Fig. 3a). The two strongest predictors
were the presence of SF3B1 mutations and principal component 1
from the gene expression data. Including additional factors
further improves the cross-validated predictive accuracy R2,
before starting to overfit, leading to a decline in R2. The optimal
model had a predictive accuracy of R2 ¼ 55% indicating a good
agreement between predictions and observations (Fig. 3b). As
expected, SF3B1 mutations were positively correlated with the
number of ring sideroblasts and, together with the smaller
negative contributions of TET2, STAG2 and SRSF2 mutations,
accounted for 31% of the observed variance (Fig. 3c). A further
25% of the variance was explained by the expression data.
Similarly, we could model 35% of the variability in bone
marrow blast counts using PC1, PC2 and SF3B1 mutations, with
additional small contributions from mutations in ASXL1,
DNMT3A and U2AF1 (Fig. 3c). The largest contribution to the
explained variance stemmed from the expression data, in total
accounting for 32% of the explained variance, compared with
3% attributed to genetic variables. The residual variance of
100%—R2 ¼ 65% indicates that there are fluctuations in the
abundance of blast counts that are caused by factors, which are
either not included in our data set, such as somatic mutations in
genes that have not been sequenced, germline genetics or
epigenetics, or of technical nature.
In general, our ability to predict haematological variables from
genomic and transcriptomic data varied greatly. We were able to
model 18% of variance in haemoglobin levels and 19% of variance
Prediction of blood counts by expression and mutations. Blood
counts and bone marrow features are important diagnostic and
prognostic variables. In MDS, they are a phenotype, in the sense
that they imperfectly reflect the underlying biology of the disease,
and hence the genome and transcriptome state. To explore these
inter-relationships, we used generalized linear models to quantify
the association of common genetic and cytogenetic alterations, as
well as the first 20 principal components of the transcriptome,
with blood counts, bone marrow blast fractions, myeloid:erythroid ratio, ring sideroblast counts, serum ferritin, sex and age.
These models enable us to identify the most important genetic
and transcriptomic variables for predicting each variable and
Ringed sideroblasts
Optimal model
0.3
0.2
0.1
1
3B
SF
0.95
0.9
0.8
0.5
0.2
0.1
0.05
3B
PC1
P 1
PCC7
TE 11
P T2
ST C1
A 2
tri G2
(
PC19)
PC14
PC17
1
P 8
P C6
de C1
SRl(1 0
SF2)
C 2
B
PCL
PC9
PC4
P
al C25
t(1 0
7
dePC q)
l(2 16
AS 0q
XL )
TP 1
C 53
U
BC X1
O
PC R
JA 19
E K
ZR ZH2
SR 2
PT P 2
P C
R N1 3
U 1
U NX
2A 1
ID F1
chH2
r(
P 3)
PCC8
PH 15
d
F
D el 6
N (7
M q
T )
P 3A
de C1
l(5 3
tri q)
(
P 8)
de C2
l(Y
)
0.0
Coefficient
Explained variance R 2
Model coefficient-β
0.5
0.4
Absolute coefficients
−1 0 1
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
Observed
ringed sideroblasts
Explained variance R 2
0.6
R2 = 0.55
SF
0.01 0.05
1
1 Model coefficients
−1
3
5
9
9
1
6
3
6
1
0.95
9
* Significant
Genetics
Cytog.
Expr.
2
2
3
10 9
1 12
8
15
1
16
12
10 24 21 23
3
16 18
1 20 10
4 26 26 6 14
14 25
20 25 28
10 10 7
1 19
24 21 2
11
1
0.5 0.8
Variance components
LASSO rank
8
24 23
8
6 25
18
6
16
4
17 4
6 24
21
5 15
15
22
16
10
1
13 10
17 14 10
23
10
20 26 4 21
7
8
5 10 6
12 18 3
4
2 19
22
4
7
10 3
6
26 25 20
2
2
4 10
4
1
14
20 8 10
8 16 16 15 11
2
18 9
8 23 10 18
25
6
6
Genetics
Cytogenetics
*
4
9
4
*
*
*
1 14 17 4 16
26 8
13 4
13 8
14 7 10
10
1
6 13
12 3
6
10
7
4
3 10
2
2
16 4
6 12 11 6 16
20 2 17
8 16
14
2
12
7
11 11
7
5 13
7
*
7
*
8
SF
3B
T 1
SRET
2
S
A
D SXF2
N L
M 1
T
R 3A
U
N
U X1
2A
F
TP 1
EZ 53
H
I 2
STDH
2
ZRAG
SR2
2
C
BC BL
O
JA R
C K2
U
X
P
PT H 1
PNF6
R 11
ea
D rr
D el( (3)
el 5
(7 q)
/7
q)
D Tr
Ab el(i(8)
1
n( 2
17 )
q
D Tri( )
el 19
(2 )
0
D q)
el
(Y
PC )
PC1
PC2
PC3
PC4
PC5
PC6
PC7
P 8
PCC9
PC10
PC11
PC12
PC13
PC14
PC15
PC16
PC17
PC18
PC19
20
Abs. neutrophil count
Age
Bone marrow blasts
Cytopenia
Gender
Haemoglobin
M:E ratio
Platelets
Ringed sideroblasts
Serum ferritin
0.2
Predicted ringed sideroblasts
Model complexity
0
1 0.0
0.8
Explained variance
Expression
Figure 3 | Prediction of blood and bone marrow counts. (a) Variance explained by selected driver genes, cytogenetic lesions and first 20 transcriptome
principal components (red line±1 s.d.; fivefold cross validation) ordered by their occurrence in a LASSO penalized model. The optimal model maximizes
the explained variance R2. The right axis indicates the effect of each standardized covariate in the optimal model. (b) Scatter plot of predicted and observed
amounts of ringed sideroblasts on a double logit axis. The inset shows the model coefficients indicating the magnitude of each fold change of driver
alterations or a unit fold change in the expression components. (c) Heatmap of optimal model coefficients for eight blood and bone marrow counts plus
gender and age. LASSO-selected coefficients are coloured. The numbers on each tile denote the order in which variables are included indicating their
relative importance. Bold fonts are used for highly significant coefficients in which the explained variance is one s.d. below the maximum. The right bar plot
shows the estimated distribution of variance explained by genetic, cytogenetic and transcriptomic variables. Stars (*) denote models where R2 is greater
than zero by a margin of more than one s.d.
6
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
Prognostic power of expression, mutations and clinical data.
An important clinical aspiration is to accurately ascertain a
patient’s prognosis. With a high degree of interdependency
among genetic, cytogenetic, transcriptomic and haematological
variables, it is a challenge to derive the best combination of
predictors to calculate a patient’s risk. Currently in clinical
practice, the IPSS score is used for prognostication in MDS34.
Here, we used multivariate Cox proportional hazards models to
predict leukaemia-free survival. We were especially interested in
the extent of prognostic information contained by different
classes of predictor variables, such as genetic or transcriptomic
data, as this may provide which data types are best suited for
developing novel prediction schemes.
We therefore grouped the available data into five classes of
variables: gene mutations, gene expression, cytogenetics, diagnostic blood counts and demographic variables (gender and age).
To estimate the predictive accuracy of a given class of predictor
variables in an unbiased way, we used fivefold cross-validation
using four-fifths of the patients to train a multivariate Cox
proportional hazards model. We then apply this to predict
outcome on the remaining fifth of the patients, using Harrel’s C
statistic to measure the concordance between the predicted risk
and the observed outcome. A value of C ¼ 50% is equivalent to
random guessing and a value of 100% indicates that the risk ranks
the survival times of all patients correctly (Fig. 4a).
0.8
AML-free survival
Harrel's C
0.4
0.2
0.6
0.4
0.2
C = 0.68
Al
l
SS
Risk
Low
Intermediate
High
0.6
0.4
0.2
C = 0.76
0.0
0
20
100
40 60
Months
All
0.8
80
0.8
0.4
0.2
C = 0.69
0
Risk
Low
Intermediate
High
20
40 60
Months
Genetics (1%)
Cytogenetics (0%)
0.2
80
100
Survival risk contributions
Blood (36%)
Demographics (18%)
0.4
Risk
Low
Intermediate
High
0.6
100
0.6
C=0.76
C = 0.76
0.0
40 60 80
Months
20
Blood
1.0
0.0
1.0
AML-free survival
0.8
0
IP
G
C ene
yt
og tics
e
Ex ne
p ti
Bl res cs
o
D od sion
em c
og oun
ra ts
ph
ic
s
Expression
1.0
AML-free survival
0.8
0.0
0.0
Validation on TCGA AML data. To investigate whether the
observed findings were specific to our cohort of MDS samples
and to demonstrate the applicability of our methods to other
cancer types, we downloaded RNA-seq, mutations, cytogenetic
and clinical data from The Cancer Genome Atlas (TCGA) AML
Genetics Risk
Low
Intermediate
High
1.0
0.6
First, we separately evaluated the prognostic power of each data
type. The accuracy of genetics alone was C ¼ 68% (Fig. 4b), that
of blood counts and bone marrow features was C ¼ 69% (Fig. 4c),
somewhat inferior to that obtained by using expression data
(based on the first 20 PCs) of C ¼ 76% (Fig. 4d). Although we had
only a small data set, which generally reduces the cross-validated
model accuracy as test and training data splits may differ
substantially and also leads to an uncertainty of these estimates in
the order of a few percent, it is notable that our models achieved a
value greater than the current standard, the IPSS score (C ¼ 64%).
It therefore seems that these data categories individually possess
reasonable potential for prognostication that warrants future
investigation and validation in larger cohorts.
Combination of the variable classes resulted in a modest
increase of the predictive accuracy to C ¼ 76%, similar to the
value obtained by expression data alone (Fig. 4e). Decomposing
the contributions to the risk prediction by each category showed
that the relative contribution of cytogenetics and genetics did not
add measurably to the risk estimates (Fig. 4f). This behaviour was
confirmed by using random survival forests as a complementary
modelling approach (Supplementary Fig. 2). It therefore appears
as if the prognostic information present in genetics and
cytogenetics is, at least at the resolution possible with this cohort,
mostly contained in expression and blood count data, and does
not add independent prognostic information. It may also be that
the global transcriptomic changes capture biological variability
caused by genomic alterations or other factors, which we have not
assessed in our gene screen.
AML-free survival
in platelet counts, with broadly equal contributions from gene
mutation and gene expression data (Fig. 3c). No significant
association was found between our data and absolute neutrophil
counts, myeloid:erythroid ratio or serum ferritin levels, likely
because of the relatively high number of missing cases for the
latter two variables that limit our ability to train predictive models
(Table 1).
0
20
40 60
Months
80
100
Expression (46%)
Figure 4 | Prediction of survival. (a) Barplot of Harrel’s C-statistic for multivariate Cox proportional hazards models using genetic, cytogenetic,
expression, blood and bone marrow counts, demographics, or all variables as well as IPSS. Values are fivefold cross-validated (grey points) with
the bar showing the average across five splits; error bars denote the s.d. of the mean. (b–e) Kaplan–Meier curves for risk terciles of multivariate
survival model for different categories of predictive variables. (f) Distribution of risk contributions in the survival model using all covariates.
The relative contribution of genetics and cytogenetics is small, despite their marginal predictive potential.
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
7
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
cohort27 (Supplementary Data 4 and Supplementary Data 5).
Combined data were available for 173 patients and contained
mutation data for 30 recurrent (nX5) lesions. The
correspondence of gene expression and recurrent alterations
was very similar to our observations in MDS. The first
principal component of AML gene expression showed a
remarkable association with NPM1 and FLT3 mutations in
samples with high values of PC1; mutations in TP53 and RUNX1,
deletions of 5/5q and 7q, and trisomy 8 coincide with low
values of PC1. The recurrent balanced translocations t(15;17) and
t(8;21) occur predominantly in samples with low levels of PC2
(Fig. 5a).
These descriptive observations translate to 5,420/18,214 ¼ 30%
genes significantly associated with recurrent mutations and
explained variances R2 ranging from 21 to 69% (Fig. 5b). The
top 50 genes with the highest R2 contained many genes of the
HOX family, which are important regulators of development
and haematopoietic differentiation (Supplementary Data 4). The
number of associated expression changes per lesion ranged from
42 (KIT) to 3,157 for t(15;17), with a general trend of a high
number of differentially expressed genes in AML cases with
balanced translocations (Fig. 5c).
Prognostic blood counts, demographics, genetics, cytogenetics
and gene expression were found to posses prognostic power
(Fig. 5d). Although the uncertainty does not allow us to derive
definitive conclusions, we observed that the overall accuracy of
models fit to genetics and cytogenetics data appeared surprisingly
low (C ¼ 58% and C ¼ 54%, respectively), close to that of the
established cytogenetic risk classification (C ¼ 58%). The prognostic value of expression data was higher (C ¼ 62%), but by far
the most influential factor for survival in this cohort was patient
demographics (age, gender, C ¼ 67%).
Combing all data types in one prognostic model did not
significantly increase the concordance (C ¼ 68%) with the most
Differentially expressed
genes by mutations
DNMT3A
n=43
CEBPA
n=13
FAM5C
n=5
FLT3_ITD
n=36
FLT3_TKD
n=12
IDH1
n=16
10
Observed
Random
8
PC1
IDH2
n=17
KIT
n=7
KRAS
n=7
NPM1
n=48
NRAS
n=12
PHF6
Density
PC2
Gene expression principal components
n=5
6
5,420 genes q < 0.05
4
2
0
RUNX1
n=16
TET2
n=15
TP53
n=14
U2AF1
n=7
del(7q)
n=19
+8
+21
n=18
n=8
SMC1A
n=7
SMC3
WT1
n=10
n=7
STAG2
n=5
Complex
n=22
–5/del(5q)
n=14
t(15;17)
n=15
t(8;21)
n=7
inv(16)/t(16;16)
n=7
0.0
0.2
0.4
0.6
Explained variance per gene R2
Prognostic accuracy of
individual data types
0.70
0.60
C
Log FC
Expression (0.48± 0.2)
1
0
−1
2,500
0
8
29
1,
84
3
Translocations (0± 0.01)
CNA (0.01± 0.01)
Genetics (0.02± 0.05)
Blood (0.06± 0.05)
5
63
21
500
28
62 2
1,000
89
1
1,500
0
7 8
72
29
42 5
30
3
1,
20
20
8
18 3
96 9
11
6
90
6
18
56 9
57
0
83
61
7
24
91 5
30
24 2
8 8
92
28
89 0
2,000
61
14
3,000
Variance components (log hazard) in combined model
15
7
Significant mutation-expression
coefficients (q<0.05)
Missing
Wildtype
Mutant
3,
Age > median
n=85
C
D EB
N P
M A
FA T3A
FL M
FL T3 5C
T3 _IT
_T D
K
ID D
H
ID 1
H
2
K
KR IT
N AS
P
N M1
R
P AS
PT HF
P 6
R N1
A
R D21
U 1
SM NX
C 1
S 1A
STMC
AG3
TE 2
T T2
U P53
2A
F
C W 1
o
-5 mpT1
/d le
el x
de (5q
l(7 )
q)
+
t(1 + 8
in
2
5 1
v(
16 t ;17
)/t (8; )
(1 21
6 )
G ;16
en )
de
Ag r
e
Differentially expressed genes
Gender (male)
n=93
yt
0.50
B
og loo
en d
G etic
en s
+
G Cy
en t
e
S tic
G
en Ex td. s
+C pr risk
y es
D t+b sio
em lo n
og +e
ra xp
ph
ic
s
Al
l
RAD21
n=5
Concordance (5x CV)
PTPN11
n=8
Demographics (0.3± 0.02)
Figure 5 | TCGA AML cohort. (a) Principal components 1 and 2 of 173 AML cases with RNA-seq data, overlaid by mutation status of 30 recurrent genomic
lesions (green: point mutations and small indels; blue: copy-number alterations; orange: balanced translocations) as well as gender (male highlighted in
brown) and age (above median of 57.5 years in brown). (b) Distribution of gene expression variance explained by genetic and cytogenetic factors
(moderated F-test; FDRo0.05; n ¼ 173). (c) Number of differentially expressed genes per genetic lesion (moderated t-test; FDRo0.05; n ¼ 173).
(d) Prognostic accuracy of different data types and their combinations measured by a fivefold cross-validated C-statistic. Error bars denote the s.d.
of the mean. (e) Variance decomposition of the log-hazard predictions in a model comprising all data types. Values denoted by ± are the mean
prediction errors in each group.
8
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
Variance explained by each category
0–30%
0-65%
Genotype
R 2=0–55%
Expression
Blood
Survival
40%
C=76%
50%
<5%
MDS
AML
pr
e
te Ge ssio
d n n
ex om
pr ic
es s
si
on
0.45
ic
ed
ic
Ex
pr
e
te Ge ssio
d n n
ex om
pr ic
es s
si
on
0.3
0.60
ed
0.5
Concordance
0.75
0.7
Ex
Concordance
0.9
Pr
Discussion
We have performed a comprehensive analysis to study the
relationships between mutations in 12 genes frequently mutated
in MDS, common cytogenetic aberrations, gene expression
profiles from bone marrow CD34 þ cells, diagnostic clinical
variables as well as outcome in 124 MDS patients. Our
analysis allowed for systematically testing for associations across
driver alterations, expression changes, clinical variables and
outcome.
Some of the findings revealed by our systematic analysis
confirm known mechanisms, while other findings are more
surprising. First of all, it was somewhat unexpected that the
number of differentially expressed genes in MDS varied so
extensively among driver alterations (11 for DNMT3A to 741 in
del(5q)). The ability to detect transcription changes attributable
to a driver mutation may depend on the number of cases with
that alteration, although we did not observe a general trend in our
data. The second largest number of differentially expressed genes
was, for example, associated with STAG2 mutations, which were
observed in only eight cases. Analysing larger cohorts will refine
these estimates and should generally increase the number of
significant expression changes as more subtle differences can be
detected.
When analysing observational data, it is generally difficult to
discern the underlying causality. Here, that means that the
associated differences in expression patterns between different
mutant genotypes could, in principle, also indicate that mutations
occur preferentially in certain cell types. Ultimately, the given
data do not allow us to rule out this possibility. For many
examples discussed here, however, there exists additional
evidence from functional experiments illustrating the mechanistic
link between mutations and transcription changes, such as the
association between SF3B1 mutations, ABCB7 downregulation
and a sideroblastic phenotype, or the loss of PRC repression at
ASXL1 and EZH2 target genes.
Overall, our results can be broadly summarized as in Fig. 6a:
there is a considerable influence of the genotype on expression
data accounting for up to 65% of the observed expression
variability in some genes, which is remarkable given the technical
noise found in gene expression data. Both genotype and
expression data have a similar predictive contribution for
modelling blood counts. Genetics, clinical diagnostic variables
and demographics were all found to contain information for
predicting survival. When combining all available data types in a
multivariable survival model, prognostic accuracy is greater than
that based on individual data sets. This indicates that the accuracy
of prognostic models for MDS can benefit from incorporating
multiple data types, including gene expression data. The main
contributions to predict risk were attributed to expression and
blood counts; the contribution of genetics and cytogenetics was
smaller.
Very similar phenomena were observed when analysing the
TCGA AML cohort, where a comparable number and strength of
associations between gene expression and mutations were found.
When analysing outcome, we again observed that expression data
appeared at least as powerful in predicting survival as mutation
data, which is surprising given the established prognostic
importance of genetics and cytogenetics in AML.
0-30%
Pr
influential data components being expression and demographics,
confirming our findings in MDS. Taken together these observations support the notion that a substantial proportion of genetic
variability translates into distinct transcriptional states, which
appear at least as informative for predicting outcome as the
genetic data alone.
Figure 6 | Summary of findings. (a) Flow chart indicating explained
variance in MDS for different associations from genotype to survival. Values
on each arrow are the relative contribution to the explained variance
stemming from the variable category on the foot of the arrow in
multivariate model predicting the variable at the arrowhead using all
incoming variables. We observe a tendency to find stronger associations
between successive categories than between distal ones. (b,c) Prognostic
value of observed and predicted expression data and genomics in MDS and
AML. Shown are results from 100 test/training splits fitting a model to 4/5
of the data and evaluating the predictions on the remainder.
Although it is generally difficult to disentangle the effect of
multicollinear covariates, our findings are compatible with the
notion of a hierarchy in which the driver mutations in the
genome dictate intermediate phenotypes, such as gene expression
and blood counts, which ultimately determine outcome (Fig. 6a).
Hence, the genome indirectly controls survival, and this
information is largely contained in the intermediate phenotypes,
making these influential proxies for outcome. Conversely,
understanding the interrelationship of genotype and phenotypes
may help decipher the prognostic information that is genetically
encoded. If we, in hindsight, compute the predicted values of the
expression principal components, based on the linear expression
models defined earlier (using only genomic variables), we observe
an improved accuracy in outcome predictions compared with
pure genomic data (Fig. 6b,c). It thus appears as if mapping the
genotypes to expression data accentuates the genetic data such
that the prognostic information becomes easier to extract. The
observed expression data, however, still seems to be slightly
superior in its predictive performance, likely as it comprises the
effects of many more somatic and germline variants than the 17
mutations that were analysed here.
We would expect that more comprehensive sequencing data
and a better understanding of the underlying genetic and
epigenetic mechanisms will allow us to explain a large fraction
of the residual phenotypic variability and ultimately also to more
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
9
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
accurately predict survival. In this analysis we have not utilized
that mutations are sometimes present in only a subset of cells,
which may explain part of the observed phenotypic variability.
The reason for this was that individual estimates of the variant
allele frequencies of point mutations can be unreliable and were
not available for cytogenetic lesions. With whole-genome or
single-cell sequencing data one will be able to more precisely
reconstruct the clonal architecture of bone marrow cells;
accounting for subclonal heterogeneity may then increase the
fraction of explained phenotypic variance even further.
Transcriptomic and phenotypic heterogeneity in MDS has
been recognized for many years, but it has been unclear what the
genetic roots of this inter-patient variability are. Here we have
systematically decomposed the relation between genetic and
cytogenetic alterations, gene expression, blood and bone marrow
counts and survival. As we move towards integrating genomic
and transcriptional screens into the clinical management of
patients with cancer, these interconnected streams of data
will require careful modelling to ensure optimal predictive
performance.
Methods
Samples. The study was approved by the ethics committees (Oxford C00.196,
Bournemouth 9991/03/E, Duisburg 2283/03, Stockholm 410/03, Pavia 26264/2002)
and informed patient consent was obtained. A total of 159 MDS samples and 17
healthy controls were studied; no samples were excluded in the statistical analysis.
Gene expression data of 43 bone marrow samples from MDS patients without
published expression data were obtained using the protocol described in ref. 12. In
brief, CD34 þ cells were enriched from mononuclear cells using CD34 MicroBeads
(Miltenyi Biotec, Bergisch Gladbach, Germany). RNA was extracted using TRIZOL
(Invitrogen, Paisley, UK). Fifty nanograms of RNA was subsequently amplified and
biotin labelled. A total of 10 mg of labelled cRNA was hybridized to Affymetrix
GeneChip Human Genome U133 Plus 2.0 arrays (Affymetrix, Santa Clara, CA,
USA), which were scanned on an Affymetrix GeneChip Scanner 3000 (ref. 12).
Statistical computations. All calculations were performed using R version 3.0.1
(ref. 35). A detailed report of all analysis steps can be found in Supplementary
Data 2 and 4.
Modelling gene expression. Probe intensity values were normalized using the
gcrma bioconductor package. Normalized probe intensities were then average for
each gene in the case of multiple probes. Principal components were computed
using the prcomp() function.
Linear expression models were fit with the limma bioconductor package36.
Here the expression of gene k in patient i, Yik is modelled by the following
equation
X
Yik ¼
Xij bjk þ b0k þ e
ð1Þ
j¼1
Xij is the mutation matrix for patient i and mutation j, with entries Xij ¼ 1
indicating that patient i has an oncogenic mutation j and 0 otherwise. Similarly, for
j being gender Xij ¼ 1 denotes female sex; for j being age Xij takes integer values.
The coefficients bjk measure the expression change in gene k induced by the
presence of a mutation j. The entry b0k denotes the baseline expression level of
gene k. The symbol p denotes the number of covariates.
Using this model, finding significant effects amounts to testing whether the
coefficients bjk are different from zero. The limma package uses a moderated
t-statistic for these tests with a prior variance shared across genes. Similarly, one
can test for an overall association of any variable with a given gene using a
moderated F-test. We used the Benjamini-Hochberg correction for multiple
testing37. The accuracy of the applied tests and correction schemes was verified
using a permutation approach in which each covariate was randomly permuted
thereby breaking all correlations between genotype and expression.
Modelling blood counts. Blood and bone marrow counts were modelled by
regularized generalized linear models using the glmnet R package38,39. The
approach is to model the value of the blood count k, Zik in patient i by the
generalized linear function
!
p
p X
X
Zik ¼ f
subject to :
Xij bjk þ b0k þ e;
ð2Þ
bjk olk
j¼1
j¼1
The function f is a transformation depending on the range of Z.k, denoting the k-th
column of the matrix Z. For real Z.k it is the identity, for positive Z.k the logarithm,
10
and for Z.k in [0,1] or dichotomous Z.k it is the logit transform. The optimal penalty
lk was chosen to be the value maximizing the fivefold cross-validated generalized
coefficient of determination R2k, which is defined as
R2k ¼ 1 ðLðb:k ; 0Þ=Lðb:k ; lk ÞÞp=2
ð3Þ
where L(b.k;lk) denotes the likelihood function.
Survival models. Survival models were fitted using the Cox proportional hazards
model40. Here the hazard is modelled by the function
!
p
X
ð4Þ
Xij bj
li ðtÞ ¼ l0 ðtÞ exp j¼1
The parameters bj quantify the changes in the hazard rate imposed by mutation j.
To increase the stability we used a shared prior on the variances of the risk
coefficients.
We evaluated the prognostic accuracy of survival models using Harrel’s C
statistic, as implemented in the Hmisc R package41. This statistic measures the
fraction of pairs of patients with concordant risk predictions, exp(–Spj ¼ 1 Xij bj),
and outcome similarly to the area under the receiver operating characteristic curve.
To reduce the bias on the estimated risk, we used a fivefold cross-validating
scheme, in which the data is split into five parts of approximately same size.
Repeating five times, one part of the data was left out for training the model and
the C statistic was evaluated on the set aside test partition. From the five resulting
estimates of C we then report the average.
Suppose the p covariates can be partitioned into g groups, such as genetics,
cytogenetics, transcriptomics, and so on. The risk (log hazard) ri of patient i can be
decomposed into contributions from each group,
ri ¼ p
X
Xij bj ¼ XX
g
j¼1
Xij bj ¼: X
rig
ð5Þ
g
j eg
In general the risk components rig are correlated and the variance of the risk r.
(taken across patients) cannot be decomposed into positive variance components.
However, we may write
XX
X
Var½r: ¼
Cov½r:g ; r:h ¼:
sg ;
ð6Þ
g
h
g
where sg ¼ Sh Cov[r.g , r.h] is a generalized variance component of r, which reduces
to sg ¼ Var[r.g] if the covariance matrix Cov[r.g , r.h] is diagonal. The interpretation
of sg is that it measures the residual variance component stemming from group g
plus correlations from other groups.
Random survival forests were used as an alternative approach for predicting
outcome and measuring variable importance42. These are implemented in the
randomForestSRC R package.
References
1. Stratton, M. R. Exploring the genomes of cancer cells: progress and promise.
Science 331, 1553–1558 (2011).
2. Vogelstein, B. et al. Cancer genome landscapes. Science 339, 1546–1558 (2013).
3. Corey, S. J. et al. Myelodysplastic syndromes: the complexity of stem-cell
diseases. Nat. Rev. Cancer 7, 118–129 (2007).
4. Greenberg, P. L. The multifaceted nature of myelodysplastic syndromes:
clinical, molecular, and biological prognostic features. J. Natl Compr. Cancer
Network 11, 877–885 (2013).
5. Graubert, T. & Walter, M. J. Genetics of myelodysplastic syndromes: new
insights. Hematology Am. Soc. Educ. Program 2011, 543–549 (2011).
6. Lindsley, R. C. & Ebert, B. L. Molecular pathophysiology of myelodysplastic
syndromes. Ann. Rev. Pathol. 8, 21–47 (2013).
7. Raza, A. & Galili, N. The genetic basis of phenotypic heterogeneity in
myelodysplastic syndromes. Nat. Rev. Cancer 12, 849–859.
8. Bejar, R., Levine, R. & Ebert, B. L. Unraveling the molecular pathophysiology of
myelodysplastic syndromes. J. Clin. Onc. 29, 504–515 (2011).
9. Shih, A. H. & Levine, R. L. Molecular biology of myelodysplastic syndromes.
Seminars Oncol. 38, 613–620 (2011).
10. Kulasekararaj, A. G., Mohamedali, A. M. & Mufti, G. J. Recent advances in
understanding the molecular pathogenesis of myelodysplastic syndromes.
Br. J. Haematol. 162, 587–605 (2013).
11. Papaemmanuil, E. et al. Clinical and biological implications of driver mutations
in myelodysplastic syndromes. Blood 122, 3616–3627 (2013).
12. Pellagatti, A. et al. Deregulated gene expression pathways in
myelodysplastic syndrome hematopoietic stem cells. Leukemia 24, 756–764
(2010).
13. Mills, K. I. et al. Microarray-based classifiers and prognosis models identify
subgroups with distinct clinical outcomes and high risk of AML transformation
of myelodysplastic syndrome. Blood 114, 1063–1072 (2009).
14. Sridhar, K., Ross, D. T., Tibshirani, R., Butte, A. J. & Greenberg, P. L.
Relationship of differential gene expression profiles in CD34 þ myelodysplastic
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms6901
syndrome marrow cells to disease subtype and progression. Blood 114,
4847–4858 (2009).
15. Chen, G. et al. Distinctive gene expression profiles of CD34 cells from patients
with myelodysplastic syndrome characterized by specific chromosomal
abnormalities. Blood 104, 4210–4218 (2004).
16. Theilgaard-Mönch, K. et al. Gene expression profiling in MDS and AML:
potential and future avenues. Leukemia 25, 909–920 (2011).
17. Pellagatti, A. et al. Gene expression profiles of CD34 þ cells in myelodysplastic
syndromes: involvement of interferon-stimulated genes and correlation to FAB
subtype and karyotype. Blood 108, 337–345 (2006).
18. Kon, A. et al. Recurrent mutations in multiple components of the cohesin
complex in myeloid neoplasms. Nat. Genet. 45, 1232–1237 (2013).
19. Allikmets, R. et al. Mutation of a putative mitochondrial iron transporter gene
(ABC7) in X-linked sideroblastic anemia and ataxia (XLSA/A). Hum. Mol.
Genet. 8, 743–749 (1999).
20. Nikpour, M. et al. The transporter ABCB7 is a mediator of the phenotype of
acquired refractory anemia with ring sideroblasts. Leukemia 27, 889–896
(2013).
21. Malcovati, L. et al. Clinical significance of SF3B1 mutations in myelodysplastic
syndromes and myelodysplastic/myeloproliferative neoplasms. Blood 118,
6239–6246 (2011).
22. Makishima, H. et al. Mutations in the spliceosome machinery, a novel and
ubiquitous pathway in leukemogenesis. Blood 119, 3203–3210 (2012).
23. Damm, F. et al. Mutations affecting mRNA splicing define distinct clinical
phenotypes and correlate with patient outcome in myelodysplastic syndromes.
Blood 119, 3211–3218 (2012).
24. Haferlach, T. et al. Landscape of genetic lesions in 944 patients with
myelodysplastic syndromes. Leukemia 28, 241–247 (2014).
25. Vogelstein, B. & Kinzler, K. W. Cancer genes and the pathways they control.
Nat. Med. 10, 789–799 (2004).
26. McLendon, R. et al. Comprehensive genomic characterization defines human
glioblastoma genes and core pathways. Nature 455, 1061–1068 (2008).
27. Cancer Genome Atlas Research Network. Genomic and epigenomic landscapes
of adult de novo acute myeloid leukemia. N. Engl J. Med. 368, 2059–2074
(2013).
28. Pimanda, J. E. et al. The SCL transcriptional network and BMP signaling
pathway interact to regulate RUNX1 activity. Proc. Natl Acad. Sci. USA 104,
840–845 (2007).
29. Ernst, J. et al. Mapping and analysis of chromatin state dynamics in nine
human cell types. Nature 473, 43–49 (2011).
30. Beisel, C. & Paro, R. Silencing chromatin: comparing modes and mechanisms.
Nat. Rev. Genet. 12, 123–135 (2011).
31. Bernstein, B. E. et al. The NIH Roadmap Epigenomics Mapping Consortium.
Nat. Biotechnol. 28, 1045–1048 (2010).
32. Abdel-Wahab, O. et al. ASXL1 mutations promote myeloid transformation
through loss of PRC2-mediated gene repression. Cancer Cell 22, 180–193
(2012).
33. Davies, C. et al. Silencing of ASXL1 impairs the granulomonocytic lineage
potential of human CD34 þ progenitor cells. Br. J. Haematol. 160, 842–850
(2013).
34. Greenberg, P. L. et al. International scoring system for evaluating prognosis in
myelodysplastic syndromes. Blood 89, 2079–2088 (1997).
35. R Foundation for Statistical Computing. R: A Language and Environment for
Statistical Computing http://www.r-project.org (2012).
36. Smyth, G. K. Linear models and empirical Bayes methods for assessing
differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol.
3, Article3 (2004).
37. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: A Practical
and Powerful Approach to Multiple Testing. J. Royal Stat. Soc. Series B
(Methodological) 57, 289–300 (1995).
38. Simon, N., Friedman, J. H., Hastie, T. & Tibshirani, R. Regularization paths for
Cox’s proportional hazards model via coordinate descent. J. Stat. Softw. 39,
1–13 (2011).
39. Friedman, J., Hastie, T. & Tibshirani, R. Regularization paths for generalized
linear models via coordinate descent. J. Stat. Softw. 33, 1–22 (2010).
40. Cox, D. R. Regression models and life-tables. J. Royal Stat. Soc. Series B
(Methodological) 34, 187–220 (1972).
41. Harrell, Jr F. E., Lee, K. L. & Mark, D. B. Multivariable prognostic models:
issues in developing models, evaluating assumptions and adequacy, and
measuring and reducing errors. Stat. Med. 15, 361–387 (1996).
42. Ishwaran, H., Kogalur, U. B., Blackstone, E. H. & Lauer, M. S. Random survival
forests. Ann. Appl. Stat. 2, 841–860 (2008).
Acknowledgements
This work was supported by a Specialized Center of Research grant from the Leukemia
Lymphoma Society (L.L.S.), the Kay Kendall Leukaemia Fund and the Wellcome Trust to
P.J.C. (grant reference 077012/Z/05/Z). P.J.C. is personally funded through a Wellcome
Trust Senior Clinical Research Fellowship (grant reference WT088340MA). This work
was supported by Leukaemia & Lymphoma Research of the United Kingdom. A.P., P.V.
and J.B. acknowledge support by the National Institute for Health Research (NIHR)
Oxford Biomedical Research Centre Programme. E.P. is a European Hematology
Association fellow. Studies performed at the Department of Hematology Oncology,
Fondazione IRCCS Policlinico San Matteo, and Department of Molecular Medicine,
University of Pavia, Pavia, Italy, were supported by grants from AIRC (‘Special Program
Molecular Clinical Oncology 5 1000’, project #1005) and Fondazione Cariplo.
Author contributions
M.G. conceived and performed all bioinformatics analyses, created all figures and wrote
the paper. A.P. initiated the study, prepared expression arrays and helped writing the
paper. A.G., M.G.D.P. and M.J. provided samples. L.M., H.D., A.V., N.C.P.C., P.V., S.K.,
E.H.-L. and M.C. provided samples and comments on the manuscript. E.P. initiated the
study, discussed results and assisted writing the paper. P.J.C. and J.B. initiated and
oversaw the analysis and wrote the paper.
Additional information
Accession codes: Gene expression array data for 176 MDS and normal samples have
been deposited in the Gene Expression Omnibus under the accession code GSE58831.
Supplementary Information accompanies this paper at http://www.nature.com/
naturecommunications
Competing financial interests: P.J.C. is a founder, stock holder and consultant for 14M
Genomics Ltd. S.K. received honorariums for education, advisory board and attending
congresses over the last 3 years from Celgene and Novartis. The remaining authors
declare no competing financial interests.
Reprints and permission information is available online at http://npg.nature.com/
reprintsandpermissions/
How to cite this article: Gerstung, M. et al. Combining gene mutation with gene
expression data improves outcome prediction in myelodysplastic syndromes. Nat.
Commun. 6:5901 doi: 10.1038/ncomms6901 (2015).
This work is licensed under a Creative Commons Attribution 4.0
International License. The images or other third party material in this
article are included in the article’s Creative Commons license, unless indicated otherwise
in the credit line; if the material is not included under the Creative Commons license,
users will need to obtain permission from the license holder to reproduce the material.
To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/
NATURE COMMUNICATIONS | 6:5901 | DOI: 10.1038/ncomms6901 | www.nature.com/naturecommunications
& 2015 Macmillan Publishers Limited. All rights reserved.
11