Bioinformatics, 35(3), 2019, 389–397

doi: 10.1093/bioinformatics/bty624
Advance Access Publication Date: 13 July 2018
Original Paper

Sequence analysis

DMCM: a Data-adaptive Mutation Clustering

Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020

Method to identify cancer-related
mutation clusters
Xinguo Lu1,*, Xin Qian1, Xing Li1, Qiumai Miao1 and Shaoliang Peng1,2,*
1
College of Computer Science and Electronic Engineering, Hunan University, Changsha 410082, China and 2School
of Computer Science, National University of Defense Technology, Changsha 410073, China
*To whom correspondence should be addressed.
Associate Editor: Bonnie Berger
Received on March 7, 2018; revised on June 12, 2018; editorial decision on July 6, 2018; accepted on July 12, 2018

Abstract
Motivation: Functional somatic mutations within coding amino acid sequences confer growth
advantage in pathogenic process. Most existing methods for identifying cancer-related mutations
focus on the single amino acid or the entire gene level. However, gain-of-function mutations often
cluster in specific protein regions instead of existing independently in the amino acid sequences.
Some approaches for identifying mutation clusters with mutation density on amino acid chain
have been proposed recently. But their performance in identification of mutation clusters remains
to be improved.
Results: Here we present a Data-adaptive Mutation Clustering Method (DMCM), in which kernel
density estimate (KDE) with a data-adaptive bandwidth is applied to estimate the mutation density,
to find variable clusters with different lengths on amino acid sequences. We apply this approach in
the mutation data of 571 genes in over twenty cancer types from The Cancer Genome Atlas
(TCGA). We compare the DMCM with M2C, OncodriveCLUST and Pfam Domain and find that
DMCM tends to identify more significant clusters. The cross-validation analysis shows DMCM is ro-
bust and cluster cancer type enrichment analysis shows that specific cancer types are enriched for
specific mutation clusters.
Availability and implementation: DMCM is written in Python and analysis methods of DMCM are
written in R. They are all released online, available through https://github.com/XinguoLu/DMCM.
Contact: hnluxinguo@126.com or pengshaoliang@nudt.edu.cn
Supplementary information: Supplementary data are available at Bioinformatics online.

1 Introduction
Cancer genomes possess a large number of aberrations including som-
atic mutations and copy number variations (Lu et al., 2017).
Genomic aberration associated with cancers is a complicated patho-
logical phenomenon in carcinogenic process. Much of our previous
understanding of cancer genetics is grounded on the principle that
cancer arises from a clone that has accumulated the requisite
somatically-acquired genetic aberrations, leading to malignant trans-
formation (Stratton, 2011). Progression of such a transformed clone
to disseminated disease has previously been thought to be a linear
process driven by serially acquired new mutations, including base pair
substitutions, small insertions and deletions of bases, chromosomal
rearrangements and gains and losses in gene copy number. However,
recent insights have emerged from comprehensive genomic character-
ization using next-generation sequencing (NGS) technologies, which
has allowed for the sequencing of the coding portion of the genome
through hybrid-capture whole-exome sequencing (WES) or nearly all
base pairs in a tumor-normal pair by whole-genome sequencing
(WGS), revealing unanticipated complexities in the patterns of somat-
ic mutations in cancers (Chin et al., 2011; Meyerson et al., 2010).

C The Author(s) 2018. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
V 389
390
Cancers are clonal proliferations that arise owing to mutations (i.e.
driver mutations) that confer selective growth advantage on cells, but
most of somatic mutations in a patient have not been subject to selec-
tion (i.e. passenger mutations) (Watson et al., 2013). That is to say,
driver mutations in cancer-related genes alter the structure of protein
after coding and lead to the cancers. However, the characteristics of
driver and passenger mutations in cancer genomes are not currently
well distinguished. Thus, how to identify driver mutations accurately
and understand the exact consequences of such mutations still
remains a formidable challenge.
Pathogenesis of cancer with somatic mutations is constantly
undergoing researches with the development of various high
throughput analysis techniques. Meanwhile, NGS technologies and
large genomic projects like The Cancer Genome Atlas (TCGA) and
International Cancer Genome Consortium (ICGC) have generated a
large quantity of somatic mutation data on different functions in re-
cent years. Many computational methods based on statistical theory
have been proposed to identify the driver mutations and understand
the functional impact of these mutations. MuSiC (Dees et al., 2012)
identified mutational significance in cancer genomes and separated
the significant events which are likely drivers for disease from the
passenger mutations by using a variety of statistical methods.
OncodriveFM (Gonzalez-Perez and Lopez-Bigas, 2012) identified
lowly recurrently mutated driver events by applying FM bias to
three datasets of tumor somatic variants. ActiveDriver (Reimand
and Bader, 2014) provided a systematic analysis of somatic muta-
tions in phosphorylation signaling to predict novel cancer drivers.
These analytical approaches are used for gathering more mutations
and detecting the cancer-related drivers. But the hypothesis on the
uniform background noise distribution is unsatisfied for the real
characteristic of somatic mutations in cancer. MutSig (Lawrence
et al., 2013) detected cancer-related genes and ranked them based
on somatic mutation events. SomInaClust (Eynden et al., 2015)
identified driver genes based on their mutation pattern across tumor
samples and then classified them into oncogenes (OGs) and tumor
suppressor genes (TSGs) respectively. Module-Network (Lu et al.,
2018) distinguished the subtype-specific drivers from normal genes
by applying a module-network analysis. But these methods identify
cancer drivers by focusing on the level of entire gene. In addition,
many methods are proposed to identify cancer mutations based on a
single amino acid level (Lee et al., 2014; Lu et al., 2014; Network,
2014). However, mutations that associated with cancers do not
occur within the coding sequence at random positions. Cancer
mutations usually behave as clusters or hotspot regions of non-
synonymous mutations (e.g. missense mutations) in a cancer-related
gene. More importantly, the cause of altering the functions of
cancer-related genes is probably the interaction of different muta-
tion clusters (e.g. PIK3CA). Hence, the mutation cluster identifica-
tion has become the research focus in the pathogenesis of cancer.
Recently, a multiscale mutation clustering algorithm (M2C)
(Poole et al., 2017), based on kernel density estimate (KDE), was
proposed to identify variable length mutation clusters through the
amino acid chain. This approach identifies mutation cluster by using
28 different fixed bandwidths ranging from 2 to 450 (amino acid
units) in KDE and finally merged the 28 kernel density models gen-
erated by these bandwidths. Yet as we know that the result of kernel
density estimate at a fixed bandwidth is usually out of line with the
real result. We expect a kernel density model with a data-adaptive
bandwidth to satisfy the diversification of mutation frequency at
every point of the amino acid sequence.
To this end, we proposed an Data-adaptive Mutation Clustering
Method (DMCM) based on kernel density estimate with a
X.Lu et al.
data-adaptive bandwidth to identify cancer-related mutation clus-
ters. And we applied DMCM on over 500 mutated genes in 23
tumor types from TCGA and 1309 mutation clusters were identi-
fied. To validate the robustness of DMCM, we split the mutation
data into two partitions and used each partition to generate a new
KDE model for each gene, we then compared the resulting clusters
from each partition. These processes indicated that DMCM is ro-
bust. To validate the robustness of clusters, we calculated scores for
each cluster found by DMCM and found abundant robust clusters
Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020
with high scores. In order to test whether the clusters are enriched
for specific cancer types, we performed cluster cancer type enrich-
ment analysis and calculated P-value for each cluster by Fisher’s
Exact Test, and the result showed that the percent of clusters with
P-value < 0.05 is 100% which revealed that the clusters identified
by DMCM were significantly enriched for specific cancer types. To
give a clear description of this association, we listed several well-
known clusters in the cancer literature recovered by DMCM
and several novel clusters found by DMCM. Finally, DMCM is
applied to identify driver genes. The results are compared with
OncodriveCLUST and M2C. The comparison indicates that DMCM
makes an effective complements to other driver-detection methods.
2 Materials and methods
2.1 Benchmark dataset
In this literature, the benchmark dataset consists of a simulation
dataset and a real pan-cancer dataset. The simulation data, contain-
ing 100 data samples, is generated by conjoining five standard nor-
mal distributions Gi, i ¼ 1; 2; . . . ; 5. In these standard normal
distributions, li ; i ¼ 1; 2; . . . ; 5, is set to 0, 1, 1.5, 2, 3 and
ri ; i ¼ 1; 2; . . . ; 5, is set to 1, 2, 3, 4, 5 separately. Two percent of
samples are randomly selected to generate as the background noise
distribution. In this simulation data, 10 data points within the simu-
lated sequence are selected to construct the background noise.
The pan-cancer gene list is composed by 291 high-confidence
driver genes and 144 candidate driver genes (Tamborero et al.,
2013a). This pre-computed and curated gene list is a consensus
gene list generated by five methods (MuSiC, OncodriveFM,
OncodriveCLUST, ActiveDriver and MutSig) for considering three
mutation signals of positive selection (recurrence, functional impact
and clustering) (Sabarinathan et al., 2017; Tamborero et al., 2013a).
2.2 Constructing the synonymous mutations model for
background noise
In previous studies, the mutation noise on amino acid sequences is
assumed to be uniform distribution at every point (Stehr et al.,
2011; Ye et al., 2010). However, it is contentious that functional
mutations within the coding sequences (missense mutations, inser-
tion mutations or deletion mutations) and ordinary mutations (non-
sense mutations or silent mutations) occur at random positions on
amino acid chain (Poole et al., 2017; Tamborero et al., 2013 b). The
researches have shown that synonymous mutations including
coding-silent mutations are unsatisfied with a uniform distribution
but are instead accumulated at some regions in amino acid sequen-
ces (references more). Therefore, we constructed the background
noise model using the distribution of coding-silent mutations.
2.3 Kernel density estimate approach using a
data-adaptive bandwidth
Kernel density estimate (KDE) is applied to detect driver patterns
in previous mutation cluster identification methods. KDE is a
DMCM: Data-adaptive Mutation Clustering Method 391

nonparametric method for estimating the probability density. Given mutations in every amino acid position of a chain, and C ¼ 1.059 is
an n-sample x1 ; x2 ; . . . ; xn of independent random variables with the a constant.
common density f. Parzen-Rosenblatt kernel estimate of density f We assumed MSEðxÞ ¼ biasðf^ðxÞÞ2 þ varf^ðxÞ, then a data-
was defined as Equation (1): adaptive bandwidth showed as Equation (5) was obtained by differ-
x  x  entiating to MSE and setting the derivative equal to 0,
1 Xn
i
f^ðxÞ ¼ d K (1) Ð !15
nh i¼1 h
f ðxÞ KðtÞ2 dt 1
hðxÞ ¼ n5 (5)
l22 f 00 ðxÞ2
where f^ðxÞ is estimated probability density for x, n is the number of

Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020

samples which represents the length of amino acid sequence in our
However, the f(x) is the real distribution which is not known before.
mutation clustering algorithm, h is the bandwidth of KDE, d is the
Here we replace f(x) by f^ðxÞ showed as Equation (1) as an estimate.
number of data dimension and KðÞ is kernel function. d is set to 1 as 00
We then calculate the f^ ðxÞ by applying the gaussian kernel function
the somatic mutation events are located on the 1D amino acid
to Equations (1) and (5),
sequence.
To generate a KDE for one somatic mutation dataset, the 00 1 Xn
1 t2 t2
f^ ðxÞ ¼ 3 pffiffiffiffiffiffi ðt2 e 2  e 2 Þ (6)
kernel function KðÞ and the bandwidth h must be preset as input- nh i¼1 2p
ting. The effects of KðÞ and h are demonstrated in Supplementary
Fig. S1. The different kernel functions (Gaussian kernel function, where t ¼ xX
h . And then we obtain the data-adaptive bandwidth
i

Epanechnikov kernel function, Triangular kernel function or according to Equation (1), (5) and (6),
Biweight kernel function) are chosen for estimating the mutation 1

^ 9 1 f^ðxÞ5
frequency and the results shows that these estimators produce the hðxÞ ¼ h5 c5  25 (7)
Pn
almost same density curves (See Supplementary Fig. S1A). So the 1 2
i¼1 nh t KðtÞ  f^ðxÞ
following KDE will be conducted without consideration of kernel
function selection. However, the bandwidth is set to 0.8, 1, 2, 4, 8 where h ¼ hopt is the optimal fixed bandwidth calculated by
Ð 00
Equation (4), and c ¼ f^ ðxÞ2 dx ¼ 38 p2 h5 .
1
respectively with Gaussian kernel function and the result shows
that KDE is very sensitive to bandwidth selection. Tremendously Finally, the data-adaptive kernel density estimate function is:
divergent density curves are produced by their corresponding !
bandwidths (See Supplementary Fig. S1B). Hence, this sensitivity 1 X n
x  xi
f^ðxÞ ¼ K (8)
of the final density estimate to the selected bandwidth makes ^
nhðxÞ ^
hðxÞ
i¼1
this parameter an important area of focus for this work. Fitting
^
Where hðxÞ is the data-adaptive bandwidth calculated by
curve with variable bandwidth for different hotspots on amino
acid sequence is critical to explore cancer-related mutation Equation (7).
clusters.
To this end, we propose a data-adaptive KDE algorithm, 2.4 Generating the clusters
described in Algorithm 1 in Supplementary Files, to optimize the Clusters are referred as amino acid positions pairs (xi, xj) where xi is
bandwidth at every specific x on amino acid sequence. Firstly, the the start position of cluster and xj is the end position of the cluster.
optimal fixed bandwidth for a specific amino acid sequence is DMCM identifies cancer-related mutation clusters for a given data-
obtained by minimizing the Mean Integral Square Error (MISE). set by four steps.
MISE is defined with equation (2). Step 1: Background noise detection.
Ð Ð We identify the background noise distribution by applying an
MISEðf^ðxÞÞ ¼ ½biasðf^ðxÞÞ2 dx þ varf^ðxÞdx initial weight to coding-silent mutations. The initial weight of noise
ð 2 Ð
h4 KðtÞ2 dt distribution, denoted by Ui as Equation (9), is defined as the percent-
¼ l2 ðKÞ2 f 00 ðxÞ dx þ (2) age of coding-silent mutations at ith amino acid unit.
4 nh
 
1 XL
þO h4 þ M0 i
nh Ui ¼ ; Mtotal ¼ mutationi ; ð0  i  LÞ (9)
Mtotal i¼0
where biasðf^ðxÞÞ ¼ Ef^ðxÞ  f ðxÞ is the difference between the
where M0i is the number of coding-silent mutations at ith amino acid
expected f^ðxÞ and observed probability density f(x), varf^ðxÞ is the
Ð unit for a given gene g, L is the length of amino acid sequence of g.
variance for f^ðxÞ, and l2 ¼ t2 KðtÞdt, O is the little o notation. So
This background noise is filtered for generating the clusters.
the Asymptotics Mean Integrated Square Error (AMISE), defined
Step 2: Identify initial seed Gaussian.
with Equation (3), consists of the first two leading items of MISE.
^
We calculate the data-adaptive bandwidth, hðxÞ. An initial set of
ð 2 Ð
h4 KðtÞ2 dt n seed Gaussians is then identified, where n is the number of local
AMISEðf^ðxÞÞ ¼ l2 ðKÞ2 f 00 ðxÞ dx þ (3)
4 nh maxima in the density estimate. An Gaussian is defined by 2 varia-
bles, its mean and its standard deviation denoted by l and r
After taking one dimension gaussian function as kernel function,
respectively.
we obtained the optimal fixed bandwidth showed as Equation (4)
by differentiating AMISE and setting the derivative equal to 0, 1 pffi
xl2
Gðx; l; rÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi e 2r (10)
1 1
2pr2
hopt ¼ Cr2 n5 (4)
The parameters, l and r, are initially estimated from kernel
P
where hopt is the optimal fixed bandwidth, r ¼ 1n ni¼1 ðxi  x Þ, n is density estimate, f(x). The mean of each Gaussian and the standard
the length of amino acid chain to a gene, xi is the number of muta- deviation of each gaussian, l and r, are initialized as the location of
tions on specific amino acid position in a gene, x is the mean of the local maxima and the distance between the two adjacent local
392
minima around a given maxima of the KDE result respectively. We
define the amino acid location of the ith local maximum as (lmai,
f ðlmai Þ). This maximum will be bordered by two local minima,
(lmii, f ðlmii Þ) and (lmiiþ1 ; f ðlmiiþ1 Þ) to the left and right, respective-
ly. In the edge cases where there is no local minima before or after
ai, the values (0, f(0)) and (L, f(L)) are used as local minima depend-
ing upon the case, where L is the length of the gene in amino acids.
Then, the initial parameters, li and ri, of Gi are given by:
lmiiþ1  lmii
li ¼ lmai ri ¼ (11)
2
Step 3: Generate initial mutation clusters.
The parameters of means and standard deviations are optimized
with an Expectation Maximization (EM) algorithm. The EM
algorithm consists of 3 steps: E step (calculate the expectation of
l and r), M step (calculate the maximum likelihood estimate of
Gaussian model by these two parameters from E step) and iteration
step (repeat E step and M step until convergence of these two
parameters). And this process determine the boundary of mutation
clusters by which a set of mutation clusters is identified.
Step 4: Remove insufficient mutation.
Clusters which contains less than a mutations are removed to en-
sure that clusters have sufficient mutations for further statistic ana-
lysis. The selection of a depends on the richness of the mutation
dataset. The parameter a is set to 15 which means clusters that con-
tain less than 15 mutations are omitted for further analysis.
The details of Data-adaptive Mutation Clustering Method
(DMCM) is shown with Algorithm 2 in Supplementary Files.
2.5 Performance evaluation metrics
We apply the goodness-of-fit as one of the performance evaluation
metrics to validate if the density curve estimated by DMCM is high-
ly fitting to the real density curve from the data. According to
Nonparametric Goodness-of-fit Test (Chwialkowski et al., 2016),
we define the goodness-of-fit, denoted EAf^ðxÞ , of DMCM as the bias
between estimated density curve and real density curve from the
data, which is given by:
1X n
EAf^ðxÞ ¼ jf^ðxi Þ  f ðxi Þj (12)
n i¼1
where n is the number of data samples, f^ðxÞ and f(x) is the estimated
density and the real density at data point x respectively. The bias
indicates the goodness-of-fit of estimated density curve to the real
density curve. High bias represents poor accurate of estimated dens-
ity result.
To compare the superiority of DMCM clusters over other exist-
ing methods’ clusters, we define a ’cover’ as > 50% overlap, which
means the region of a cluster overlaps at least 50% of another clus-
ter’s region. Then, the cover rate of DMCM clusters overlapping
clusters identified by the reference method is defined as below:
Numcovered
Cover Ratio ¼ (13)
Numall
where Numcovered is the number of covered clusters detected by the
reference method, and Numall is the number of all clusters of the ref-
erence method.
In addition, we validate if specific cancer types are enriched for
specific mutation clusters by using Fisher’s Exact Test (Fisher, 1922)
to calculate the P-value. Our hypothesis is that H0: one specific clus-
ter is independent from a specific cancer type; H1: one specific
X.Lu et al.
Table 1. Table of variables in Fisher’s Exact Test
Cluster A Other clusters
Cancer type T N1 N2
Other cancer types N3 N4
cluster is significantly associated with a specific cancer type.
Relevant statistics are given in Table 1.
Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020
According to Fisher’s Exact Test, We calculate the P-value by:
   
N1 þ N3 N2 þ N4

N N2
P  value ¼  1  (14)
N1 þ N2 þ N3 þ N4
N1 þ N2
H1 is true if the calculated P-value < 0.05.
3 Experimental results and discussion
We conduct experiments on simulation data and pan-cancer data re-
spectively. For simulation dataset, the data-adaptive bandwidth is calcu-
lated to generate the KDE curves, which are compared with the results
generated by the non-data-adaptive method. For real pan-cancer data-
set, mutation clusters are identified by DMCM and then compared with
M2C (Poole et al., 2017), OncodriveCLUST (Tamborero et al., 2013 b)
and Pfam (Finn et al., 2010). Also, the cross-validation analysis and the
cluster cancer type enrichment analysis is performed.
3.1 Experiments on simulation data
3.1.1 Kernel density estimated by DMCM
Kernel density curve for simulation data is generated by our pro-
posed data-adaptive method of DMCM. We compare the result
with the curve estimated by the non-data-adaptive method. For non-
data-adaptive method, the optimal fixed bandwidth is calculated
(h ¼ 2.1) according to the Equations (2) and (4). We apply gaussian
function in the kernel density estimate and the density curve is
obtained as shown with the red dotted line in Figure 1A where the
green line is the real density curve for simulation data. For DMCM,
we firstly calculate the optimal fixed bandwidth, hopt, by minimizing
the MISE. MSE is defined as a transformation of MISE, then the
data-adaptive bandwidth is calculated using the Equation (5). The
data-adaptive density curve is shown with the red dashed line in
Figure 1B in which the green line represents the real density result.
From Figure 1A and B, the density curve generated by DMCM is
more fitting to the real density curve than by the non-data-adaptive
method. In other words, density curve generated by DMCM with a
data-adaptive bandwidth presents the detailed features better and is
closer to the actual density. Comparison of fixed bandwidth with
data-adaptive bandwidth is shown in Figure 1C from which we no-
tice that the fixed bandwidth is over-averaged. We also compare the
goodness of fit between the data-adaptive density result and the
non-data-adaptive density result by calculating the bias with
Equation (12). The result shows that, compared with non-data-
adaptive density f^ðxÞ, DMCM improves the goodness-of-fit by
73.7% (EAf^ðxÞ ¼ 0:00078 and EAf^ðxÞ ¼ 0:00296).
3.1.2 Clusters identified by DMCM
To further validate the superiority of DMCM, two sets of clusters
are identified by KDE with fixed bandwidth and DMCM respective-
ly (Fig. 2). Density curves are generated by KDE with fixed band-
width and DMCM and we present these density results with density
DMCM: Data-adaptive Mutation Clustering Method
Fig. 1. Comparison of KDE result generated with the optimal fixed bandwidth
(non-data-adaptive result) and data-adaptive bandwidth (DMCM result). A)
Comparison between the real density and the non-data-adaptive density. B)
Comparison between the real density and the DMCM density. C) Comparison
of optimal non-data-adaptive bandwidth (dotted) and the data-adaptive band-
width (dashed). D) Comparison of two bias curves
spectrums in Figure 2B and C respectively (Fig. 2B is the density re-
sult of KDE with fixed bandwidth and Figure 2C is the density result
of DMCM). This process result in 3 and 6 effective clusters for KDE
with fixed bandwidth and DMCM (clusters identified by KDE with
fixed bandwidth and DMCM are shown in Fig. 2D and E) respect-
ively, clusters with light grey are removed since the amount of data
samples in these clusters are too few to increase the statistic power.
The result shows that DMCM focuses on finding narrower clusters
and DMCM identifies more clusters than KDE with fixed band-
width since the data-adaptive bandwidth suits better the change of
data characteristics than the fixed bandwidth as described previous-
ly. We also calculate the cover ratio of clusters identified by DMCM
and non-data-adaptive method. The result presents that DMCM
clusters cover 100% of clusters identified by non-data-adaptive
method. On the other hand, clusters detected by non-data-adaptive
method cover only 66.7% (regions of two clusters of DMCM are
overlapped below 50% by non-data-adaptive clusters).
3.2 Experiments on pan-cancer mutation data
3.2.1 Pan-cancer data pre-processing
The raw pan-cancer gene list, including 435 genes, is obtained as
mentioned above by combining the results of five methods (MuSiC,
OncodriveFM, OncodriveCLUST, ActiveDriver and MutSig) for
exploiting all the signals of positive selections. Specifically, each
method considers a different signal of positive selections and five po-
tential candidate driver lists are identified. These potential candidate
driver lists are then merged into a consensus driver gene list. After
associating these consensus driver genes with TCGA somatic muta-
tion data of 23 cancer types, we finally obtain the raw somatic mu-
tation dataset which can be found in the Supplementary Data.
The cancer types we consider are: ACC, BLCA, BRCA, CESC,
COAD, GBM, HNSC, KICH, KIRC, KIRP, LAML, LGG, LIHC,
LUAD, LUSC, OV, PRAD, READ, SKCM, STAD, THCA, UCEC
and UCS. Full name of each cancer type can be found in T1 in
Supplementary Tables.
393
Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020
Fig. 2. Comparison of clusters identified by DMCM and KDE with optimal
fixed bandwidth. A) represents the histogram of the raw data distribution. B)
and C) shows two spectrums of mutation density calculated by KDE with
fixed bandwidth and DMCM respectively. D) and E) displays clusters identi-
fied by KDE with fixed bandwidth and DMCM respectively
3.2.2 Mutation clusters identified by DMCM
DMCM identified 1309 clusters for 392 genes out of the selected
435 genes. The remaining 43 genes without any clusters have all
their mutations referred as constructing background noise model by
DMCM and are omitted from further analysis. The following results
indicate that our method finds variable length of regions of proteins
which are enriched for mutations. Clusters span a wide range of
lengths: from 1 to 600 amino acid units, and the number of muta-
tions in each cluster ranges from 15 to 338 mutations. Figure 3A
shows the comparison of two density curves generated by DMCM
and non-data-adaptive method with an optimal fixed bandwidth of
4 respectively. Then, two sets of mutation clusters are identified by
using these two density curves as shown in Figure 3B and C. We find
that the density curve generated by DMCM are more precise than
by non-data-adaptive method. Clusters identified by non-data-
adaptive method are generally longer than clusters from DMCM.
We compare the clusters found by DMCM to multiscale mutation
clustering (M2C) algorithm (Poole et al., 2017), method based on
Density Based Spatial Clustering of Applications with Noise
(OncodriveCLUST) (Tamborero et al., 2013 b) and protein domains
from Pfam (Finn et al., 2010). M2C uses a multiscale clustering method
based on KDE with 23 fixed bandwidths to create 23 kernel densities
and finally merges these densities to identify mutation clusters. Our
method find more clusters based on KDE with data-adaptive bandwidth
than M2C (M2C found 1255 clusters in 393 genes). OncodriveCLUST’s
approach also uses a kernel smoother to create a mutation density, but
using only one predefined fixed bandwidth. Despite finding fewer clus-
ters (OncodriveCLUST found 5185 clusters in 514 genes), we find that
clusters identified by DMCM tend to be larger and have more muta-
tions. These statistics are represented in Figure 4A.
DMCM clusters cover a total of 71% of M2C clusters which cover
52% more than OncodriveCLUST clusters (OncodriveCLUST clusters
only cover 18% of M2C clusters). In addition, DMCM clusters cover a
total of 51% of OncodriveCLUST clusters which is 6% more than
M2C clusters (M2C clusters cover 48% of OncodriveCLUST clusters).
For DMCM clusters, clusters from M2C and OncodriveCLUST cover
50% and 10% of DMCM clusters separately. Finally, we note that
DMCM have a total of 34% of clusters located within or overlapping
with protein domains while M2C and OncodriveCLUST have 31%
and 22% of clusters overlap protein domains from Pfam respectively.
These statistics are summarized in Figure 4B and C.
394 X.Lu et al.

Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020

Fig. 3. DMCM Illustration on PTEN upon over twenty cancer types. A) Results of kernel density estimate with a self-adaptive bandwidth (in bold line) and a fixed
bandwidth (in ordinary line) were represented to make a comparison. B) The distribution of all the PTEN mutation events on amino acid sequences; mutation
clusters were identified by DMCM and represented in different colors with the amino acid interval on the right side. C) Mutation data of PTEN across all of pan-
cancer dataset and clusters identified by non-data-adaptive method (Color version of this figure is available at Bioinformatics online.)

Fig. 4. Cluster statistics comparing DMCM clusters with M2C clusters, OncodriveCLUST clusters and Pfam Domains. A) Cluster length histogram. B) Coverage
with competing methods and Pfam histogram

3.2.3 Cross-validation analysis of DMCM
We validate the robustness of DMCM by splitting our dataset
into two equally sized partitions (data from one partition is
regarded as the training data and data from the other partition is
regarded as the validation data) and running the algorithm separ-
ately on each partition. A data-adaptive KDE model is then gener-
ated for each gene from two partitions and every KDE model is
used to generate a mutation cluster set. First, We calculate the
log-likelihoods for KDE models from each partition and consider
if the log-likelihoods are highly correlated by calculating the
Spearman Correlation Coefficient. The result showed that the log-
likelihoods are significantly correlated (Spearman Correlation
Coefficient  0.99 and P-value  0). These results are plotted in
Figure 5 and indicate that the statistical model underlying
DMCM is robust.
We also calculate the cover ratio of clusters between the two par-
titions with Equation (13). We define ’covered’ to mean that for two
clusters in the same gene from different partitions, one of the clus-
ters overlaps the other by at least 50%, which is illustrated before.
Our validation analysis shows that on average DMCM robustness is
about 40%, meaning about 40% of clusters are covered. However,
we further note that smaller denser clusters are more highly con-
served and overlap by a greater percentage between partitions than
large sparse clusters.
DMCM: Data-adaptive Mutation Clustering Method
Fig. 5. Cross-validation of DMCM: each circle shows the log-likelihood of the
DMCM model trained from partition 1 to generate the data from partition 2
for a single gene (red circles). The opposite analysis, using partition 2’s
DMCM model to generate data from partition 1, is also shown (purple x’s)
(Color version of this figure is available at Bioinformatics online.)
In addition, We calculate a score under the background noise distri-
bution for each mutation cluster found by DMCM. This score, donated
as cluster scorec of a cluster c, is defined by the log value of the ratio of
the emission probability of the mutations in the cluster with the emis-
sion probability of the same mutations based upon the null hypothesis
of the background noise distribution across the cluster defined below:
! !
N1 þ N3 N2 þ N4
 tasis in Uterine Corpus Endometrial Carcinoma (FDR < 1%) (Ding
N1 N2
P  value ¼ ! (15)
N1 þ N2 þ N3 þ N4
N1 þ N2
M is the set of all n pan-cancer mutations in cluster c, Gðx; lc ; rc Þ
is the normalized Gaussian distribution with mean lc and standard
deviation rc representing the unweighted component of the mixture
model corresponding to cluster c. Finally P ¼ L1 is the emission
probability of single mutation by the background noise distribution
over the gene containing the clusters of length L (in amino acids). The
cluster score provides a good indication of how robust a cluster is like-
ly to be (See Supplementary Table T2, Supplementary Fig. S3).
Higher scores indicate increased robustness.
3.2.4 Cluster cancer type enrichment analysis
We calculate an enrichment P-value by using Fisher’s Exact Test to de-
termine if specific cancer types are enriched for specific mutation clus-
ters. A contingency table for each cluster cancer type pair is created
across the pan-cancer set of samples based upon the two Boolean varia-
bles: (i) Is the sample inside the cluster and (ii) Does the sample belong
to the cancer type being analyzed. These results are then tested for sig-
nificance together using the Benjamini-Hochberg method (Benjamini
and Hochberg, 1995) with results at FDR (False Discovery Rates) of 1,
5, 10 and 25% (See Supplementary Tables T3).
The result generated by cluster cancer type enrichment analysis
shows that about 57.9% of clusters (P < 0.01) and the percent of
clusters with P < 0.05 identified by DMCM is 100%. These results
confirmed that specific cancer types are enriched for specific muta-
tion clusters. The association between mutations of TP53 and 14
cancer types is illustrated (See Supplementary Fig. S3). The cluster
located in the region of (256, 260) is significantly related to BRCA,
LIHC, LUAD, LUSC, READ and UCEC which indicates this muta-
tion cluster is identified as a potential driver cluster for these cancer
395
types. Here, we highlight several well-known and novel clusters
detected by DMCM:
BRAF V600 mutation is a well-known gene mutation event which
is proved to be associated with many kinds of cancer types. DMCM
identified a amino acid cluster from 600-601 in BRAF within the can-
cer types of Melanoma, Rectum Adenocarcinoma (READ), Thyroid
Carcinoma (THCA), Lung Adenocarcinoma (LUAD), Breast Invasive
Carcinoma (BRCA), Stomach adenocarcinoma (STAD) and Ovarian
serous cystadenocarcinoma (OV), etc. Detection of cluster 600-601
Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020
mutation in BRAF is also recommended in cases of DNA mismatch
repair protein deficiency by the guide of the National Comprehensive
Cancer Network (NCCN) in the Lynch syndrome detection strategy.
In addition, BRAF is the most common gene mutation in thyroid pap-
illary carcinoma which is associated with an increase in mortality in
patients (FDR < 1%). More importantly, experiments of the BRAF
kinase inhibitor, named vemurafenib, in two phase of clinical trails
proved the inhibitor response rate of melanoma patients carrying
BRAF mutations of cluster 600-601 is over 50% and the survival rate
of 6 months is over 84%. Therefore, the detection of cluster 600-601
mutation in BRAF has become a necessary test for the diagnosis and
treatment of many kinds of tumors.
CTNNB1 has two mutation clusters from 25-45 and 322-354
which are associated with global changes of gene expression in Liver
Hepatocellular Carcinoma (LIHC) at a false discovery rate of 1% and
Adrenocortical Carcinoma (ACC) at a false discovery rate of 5%. pre-
viously, increased level of CTNNB1 expression caused by the mutation
cluster of amino acid positions 25-45 have been associated with metas-
et al., 2015). However, to our knowledge this somatic mutation region
have not been extensively studied. We note that clusters 25-45
and 322-354 have a much lower global gene expression association
P-values (about 21 and 6 orders of magnitude smaller, respectively) in
Liver Hepatocellular Carcinoma which perhaps reveals the specific
function to the Liver Hepatocellular Carcinoma. Thus, this gene may
be a candidate for further study of its role in Liver Hepatocellular
Carcinoma and probably the other tumor types mentioned above.
GPRIN2 has a mutation cluster in Adrenocortical Carcinoma
(ACC) and Skin Cutaneous Melanoma (SKCM): amino acid positions
39-126 with 23 somatic mutations. This cluster has a much lower glo-
bal gene expression association P-value (about 13 orders of magni-
tude smaller) with Adrenocortical Carcinoma which may reveal the
specific function to this disease (FDR < 1%). We hope our results are
helpful to the research of GPRIN2’s function.
3.2.5 Driver gene prediction by DMCM
To distinguish the cancer driver genes from passenger genes, we
calculate the gene clustering score and sort the gene list with their
clustering scores. The gene clustering score is generated as in
OncodriveCLUST and the P-value list is acquired. A comprehensive
driver gene list is obtained (See T4 in Supplementary Tables). We
compared the identified driver genes with which are derived by
OncodriveCLUST and M2C respectively on CGC gene list. The com-
parison is demonstrated in Figure 6. Some previous predicted driver
genes, such as JUN, CDH11 and EGFR in SKCM, are identified by
DMCM while they are not predicted by OncodriveCLUST and M2C.
The results show that potential driver genes are acquired by DMCM
and DMCM makes an effective complement to other methods.
3.2.6 Computational complexity analysis of DMCM
The experiments are run on a desktop workstation (Windows 7
operating system), the computer contains a Dual-Core AMD
396
Fig. 6. Selected top-ranking genes from DMCM analysis on two cancer types
(Only genes annotated in the Cancer Gene Census are depicted). Summary of
the results obtained by three methods (DMCM, OncodriveCLUST and M2C)
aimed to find driver genes for the two analyzed cancer types
Opteron, 24 GB of memory capacity and 500 GB of hard disk cap-
acity. We compare the time consumption of DMCM with M2C by
using the speed of convergence to evaluate the computational time
of these two algorithms since both of two algorithms are iterative
convergence algorithm. By now, the EM algorithm is one of the best
solutions to conquer convergence problems. The mutation cluster
generation process of these two algorithms is done by the EM algo-
rithm to optimize the cluster boundary. Hence, our time consump-
tion analysis focuses on comparing the speed of EM convergence.
Gene PTEN containing 550 mutations across 21 cancer types is used
to perform this comparison. For DMCM, the AMISE converges
after 40 iterations while it takes 60 iterations for M2C (See
Supplementary Fig. S4A). Moreover, M2C stores all the mutation
cluster lists generated from each fixed bandwidth (totally 28 band-
widths) in memory and finally merges 28 cluster lists to generate a
final list. Hence, there is potential memory overhead with the
increased size of dataset. In contrast, DMCM is more memory effi-
cient as it stores only one mutation cluster list generated by the data-
adaptive bandwidth. The comparison of memory usage in gene
PTEN is shown in Supplementary Figure S4B which shows that
M2C costs 5.5GB of memory space while DMCM only costs 4GB.
Taken together, DMCM is more efficient than M2C in both of com-
putational time consumption and memory usage.
4 Conclusion
Many previous methods identifying driver mutations focus on the
entire gene or the single amino acid level which may not be appro-
priate owing to the fact that mutations that associated with cancers
do not occur uniformly in a gene or at random positions within the
coding amino acid sequence. We have proposed a Data-adaptive
Mutation Clustering Method (DMCM) which combines these two
approaches by creating a data-adaptive bandwidth searching for
variable-length regions of interest within individual genes. DMCM
represents a data driven approach towards systematically identifying
X.Lu et al.
mutation clusters in genes. Both the cross-validation analysis and
cluster scores indicate the robustness of DMCM. To reveal the sig-
nificant relationship between specific mutation clusters and specific
cancer types, a cluster cancer type enrichment analysis is provided
by which we find that mutation clusters detected by DMCM are sig-
nificantly cancer-related. However, it is likely that many of our one-
dimensional clusters, if mapped onto the 3D structure of a protein,
would be more worthy in cancer studies. A future direction is to
carry out this mapping and determine more realistic structural muta-
Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020
tion clusters and we suspect that such an approach would further in-
crease statistical power.
Acknowledgements
The authors wish to thank all the anonymous reviewers whose constructive
comments will be very helpful to strengthen the presentation of this study.
Funding
This work was supported by the Natural Science Foundation of Hunan
Province, China (Grant No. 2018JJ2053).
Conflict of Interest: none declared.
References
Benjamini,Y. and Hochberg,Y. (1995) Controlling the false discovery rate: a
practical and powerful approach to multiple testing. J. R. Stat. Soc., 57,
289–300.
Chin,L. et al. (2011) Making sense of cancer genomic data. Genes Dev., 25,
534–555.
Chwialkowski,K. et al. (2016) A kernel test of goodness of fit. In: 33rd
International Conference on Machine Learning, ICML 2016, vol. 6, New
York City, NY, pp. 3854–3867.
Dees,N.D. et al. (2012) Music: identifying mutational significance in cancer
genomes. Genome Res., 22, 1589–1598.
Ding,J. et al. (2015) Systematic analysis of somatic mutations impacting gene
expression in 12 tumour types. Nat. Commun., 6, 8554.
Eynden,J.V.D. et al. (2015) Sominaclust: detection of cancer genes based
on somatic mutation patterns of inactivation and clustering. BMC
Bioinformatics, 16, 125.
Finn,R.D. et al. (2010) The pfam protein families database. Nucleic Acids
Res., 40, 290–301.
Fisher,R.A. (1922) On the interpretation of v2 from contingency tables, and
the calculation of p. J. R. Stat. Soc., 85, 87–94.
Gonzalez-Perez,A. and Lopez-Bigas,N. (2012) Functional impact bias reveals
cancer drivers. Nucleic Acids Res., 40, e169.
Lawrence,M.S. et al. (2013) Mutational heterogeneity in cancer and the search
for new cancer-associated genes. Nature, 499, 214–218.
Lee,C.S. et al. (2014) Recurrent point mutations in the kinetochore gene
knstrn in cutaneous squamous cell carcinoma. Nat. Genet., 46, 1060–1062.
Lu,X. et al. (2014) A co-expression modules based gene selection for cancer
recognition. J. Theor. Biol., 362, 75–82.
Lu,X. et al. (2017) Driver pattern identification over the gene co-expression of
drug response in ovarian cancer by integrating high throughput genomics
data. Sci. Rep., 7, 16188.
Lu,X. et al. (2018) The integrative method based on the module-network for
identifying driver genes in cancer subtypes. Molecules, 23, 183.
Meyerson,M. et al. (2010) Advances in understanding cancer genomes
through second-generation sequencing. Nat. Rev. Genet., 11, 685.
Network,T.C.G.A. (2014) Comprehensive molecular characterization of uro-
thelial bladder carcinoma. Nature, 507, 315.
Poole,W. et al. (2017) Multiscale mutation clustering algorithm identifies
pan-cancer mutational clusters associated with pathway-level changes in
gene expression. Plos Computat. Biol., 13, e1005347.
DMCM: Data-adaptive Mutation Clustering Method
Reimand,J. and Bader,G.D. (2014) Systematic analysis of somatic mutations
in phosphorylation signaling predicts novel cancer drivers. Mol. Syst. Biol.,
9, 637.
Sabarinathan,R. et al. (2017) The whole-genome panorama of cancer drivers.
bioRxiv, 190330. doi: https://doi.org/10.1101/190330.
Stehr,H. et al. (2011) The structural impact of cancer-associated missense
mutations in oncogenes and tumor suppressors. Mol. Cancer, 10, 54.
Stratton,M.R. (2011) Exploring the genomes of cancer cells: progress and
promise. Science, 331, 1553–1558.
397
Tamborero,D. et al. (2013a) Comprehensive identification of mutational can-
cer driver genes across 12 tumor types. Sci. Rep., 3, 2650.
Tamborero,D. et al. (2013b) Oncodriveclust: exploiting the positional clustering of
somatic mutations to identify cancer genes. Bioinformatics, 29, 2238.
Watson,I.R. et al. (2013) Emerging patterns of somatic mutations in cancer.
Nat. Rev. Genet., 14, 703–718.
Ye,J. et al. (2010) Statistical method on nonrandom clustering with ap-
plication to somatic mutations in cancer. BMC Bioinformatics,
11, 11.
Downloaded from https://academic.oup.com/bioinformatics/article/35/3/389/5053323 by University of Bologna user on 24 September 2020