Available via license: CC BY 2.0
Content may be subject to copyright.
BioMed Central
Page 1 of 13
(page number not for citation purposes)
BMC Bioinformatics
Open Access
Methodology article
Differential splicing using whole-transcript microarrays
Mark D Robinson*
1,2,3
and Terence P Speed
3
Address:
1
Department of Medical Biology, The University of Melbourne, Parkville, Victoria 3010, Australia,
2
Cancer Research Program, Garvan
Institute of Medical Research, Darlinghurst, NSW 2010, Australia and
3
Bioinformatics Division, Walter and Eliza Hall Institute of Medical
Research, Parkville, Victoria 3050, Australia
Email: Mark D Robinson* - mrobinson@wehi.edu.au; Terence P Speed - terry@wehi.edu.au
* Corresponding author
Abstract
Background: The latest generation of Affymetrix microarrays are designed to interrogate
expression over the entire length of every locus, thus giving the opportunity to study alternative
splicing genome-wide. The Exon 1.0 ST (sense target) platform, with versions for Human, Mouse
and Rat, is designed primarily to probe every known or predicted exon. The smaller Gene 1.0 ST
array is designed as an expression microarray but still interrogates expression with probes along
the full length of each well-characterized transcript. We explore the possibility of using the Gene
1.0 ST platform to identify differential splicing events.
Results: We propose a strategy to score differential splicing by using the auxiliary information
from fitting the statistical model, RMA (robust multichip analysis). RMA partitions the probe-level
data into probe effects and expression levels, operating robustly so that if a small number of probes
behave differently than the rest, they are downweighted in the fitting step. We argue that adjacent
poorly fitting probes for a given sample can be evidence of differential splicing and have designed a
statistic to search for this behaviour. Using a public tissue panel dataset, we show many examples
of tissue-specific alternative splicing. Furthermore, we show that evidence for putative alternative
splicing has a strong correspondence between the Gene 1.0 ST and Exon 1.0 ST platforms.
Conclusion: We propose a new approach, FIRMAGene, to search for differentially spliced genes
using the Gene 1.0 ST platform. Such an analysis complements the search for differential
expression. We validate the method by illustrating several known examples and we note some of
the challenges in interpreting the probe-level data.
Software implementing our methods is freely available as an R package.
Background
Alternative splicing
Alternative splicing is the ubiquitous phenomenon where
the same genetic locus can transcribe multiple messenger
RNAs (mRNAs), by splicing out different subsets of
intronic regions from a common pre-mRNA product.
Splice variants of a gene can be functionally distinct and
generate considerable proteomic diversity. Despite early
estimates of near 50% [1], it is now thought that greater
than 90% of all human genes exhibit alternative splicing
[2,3], accounting for much of the complexity of metazoan
organisms. Alternative splicing is known to be prominent
in many important physiological processes, such as cell
differentiation, apoptosis and development, and is espe-
Published: 22 May 2009
BMC Bioinformatics 2009, 10:156 doi:10.1186/1471-2105-10-156
Received: 7 December 2008
Accepted: 22 May 2009
This article is available from: http://www.biomedcentral.com/1471-2105/10/156
© 2009 Robinson and Speed; licensee BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0
),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 2 of 13
(page number not for citation purposes)
cially prevalent in the nervous system [4-6]. Mis-regula-
tion or mutations that affect the splicing mechanism can
result in disease, including cancer [7]. It is no surprise
then that alternative splice variants have been observed in
a tissue-specific or cancer-specific manner.
Until recently, predicting alternative splice events usually
involved the comparison of expressed sequence tags
(ESTs) across several libraries. For example, algorithms
that compare EST abundance across human tissues
deduced many tissue-specific isoforms [8,9]. Recently,
DNA microarrays have been successfully utilized to
explore alternative splicing, finding many genes with
known and putative tissue-specific isoforms [1,10,11].
In this study, we propose a statistical method of scoring
differential splicing for the Gene 1.0 ST (hereafter referred
to as Gene) array data, which is the latest generation of
Affymetrix genome-wide expression profiling chips. Note
that the aim of this work is not to suggest the Gene plat-
form as a replacement for the Exon 1.0 ST (referred to as
Exon) array. The considerations of cost, probe coverage
and protocol (e.g. amount of RNA needed) will ultimately
guide this decision for experimenters. We expect the Gene
platform will be used regularly for expression profiling
studies and here, we describe the potential to identify dif-
ferential splicing at no additional experimental cost. We
are simply providing a additional data analysis-based ave-
nue of interrogation. Here, our motivation is to outline
the possibilities and limitations of using the Gene array
for the detection of differential splicing, not to rigour-
ously compare and contrast the platforms.
Both the Gene and Exon arrays interrogate well-anno-
tated exonic content. Perhaps not surprising given the two
platforms share a large number of probes, we have discov-
ered that many of the patterns observed in Exon data are
also observed in Gene data. In addition, we note some of
the challenges and ambiguities of analyzing whole tran-
script microarray data in the context of alternative splic-
ing.
We have shown previously that Gene has similar perform-
ance to Exon and a previous generation of Affymetrix
chips, in various respects in the context of gene expression
[12]. There is certainly value in having an expression plat-
form that can additionally deduce alternative splice
forms. Exon does this [12,13,19]. We show here that
Gene has potential to do so as well, if we are willing to
interrogate only well-annotated content and have reduced
coverage for some transcripts.
Differential splicing
It is worth noting at the outset that microarrays, in gen-
eral, will not be able to detect alternative splicing, per se.
For example, if an exon is spliced out of all tissues or sam-
ples in the study, there is no ability to detect it as alterna-
tive splicing. So, the focus of the methodology presented
here and other related methods is on detecting differential
splicing, or more generally, the differential expression of
alternative isoforms.
Affymetrix array design
Figure 1 shows a UCSC browser view [14] of the locations
of Exon probes and probesets and Gene probes for a sin-
gle human gene, SLC25A3 (solute carrier family 25, mem-
ber 3). As is standard with Affymetrix design, all probes
are 25 base pairs, however, on the newer generation of
chips, there is no mismatch probe for every perfect match
probe. The Exon probesets, one for each probe selection
region (PSR), are shown in black for well-annotated exons
and 2 shades of grey depending on the original prediction
evidence. PSRs are defined by Affymetrix according to
whether a particular region may act as an independent
unit, based on several levels of annotation projected to the
genome. The array design for Exon aims to have 4 probes
per PSR whereas the Gene array has approximately 25
probes per transcript cluster [12].
UCSC browser view of solute carrier family 25, member 3 (SLC25A3)Figure 1
UCSC browser view of solute carrier family 25, member 3 (SLC25A3). Custom tracks have been added for the loca-
tions of the 25-mer probes for the Affymetrix Gene, Exon and HG-U133 human expression arrays, relative to the locations
of exons for RefSeq or Ensembl gene builds. The Exon probesets are shown in black and grey in the lowermost track. Several
probes are common to both the Gene and Exon platforms.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 3 of 13
(page number not for citation purposes)
The Gene platform shares a large number of its probes
(approximately 65%) with the Exon array, but also
includes a significant number of probes unique to the
platform. In terms of differential splice detection, the cov-
erage by either platform is locus-specific. The ability to
detect differential isoform usage will depend not only on
the number of probes covering the region, but the nature
of the splicing, the degree of differential usage and the per-
formance of the probes near to the event. This could also
mean there is a bias in the ability to detect differential
splicing with Gene through genes having fewer rather
than more exons. In general, genes with fewer than 5 or 6
exons will have more probes per exon on the Gene array.
We have not studied this possible bias in any detail.
Instead, we focus on determining differential splicing pre-
dictions based on the available data with the current Gene
design.
RMA decomposition
After background adjustment and normalization, one of
the commonly used methods for summarizing probe-
level Affymetrix data into expression levels is robust mul-
tichip analysis (RMA) [15]. The approach accounts for rel-
ative probe-specific effects according to the following
model:
where Y
ij
are the log
2
background adjusted and normal-
ized intensities for probe j from sample i,
α
i
are the chip
effects (i = 1, ..., N) and
β
j
are the probe effects (j = 1, ..., J),
given N samples and J probes and
ε
ij
are the errors. For
simplicity, a subscript for gene is suppressed here since all
models are fit to genes one by one. The constraint
is imposed to make the probe effects relative
and identifiable. Figure 2A illustrates probe-level data for
a gene that is strongly differentially expressed between
heart and brain across a full mixture of RNA samples (red
– 100% heart, green – 100% brain, blue – mixture). The
most striking observation of the probe-level data is the
parallelism across all samples, largely due to the
sequence-specific probe intensity effects. That is, because
this gene is differentially expressed between brain and
heart, each individual probe shows a relative change in
abundance, even though the range of intensity for each
probe may be quite different. RMA models this behaviour
by estimating probe-specific effects (Figure 2B), leaving
the relevant sample-specific features (chip effects, Figure
2C) for downstream analysis of expression. The residuals
(Figure 2D), which are the differences between the
observed intensities and that explained by the model, are
random and centred around 0. The models are fitted
robustly using iteratively reweighted least squares [16] so
that individual observations do not have undue influence
in the estimation of
α
i
and
β
j
.
Next, we show that alternative splicing can be highlighted
by focusing on the residuals. Take for example WNK1
(lysine deficient protein kinase 1), a gene known to
express a kidney-specific isoform having a 5' region
spliced out [8]. Figure 3 shows the probe-level data for
WNK1, as well as the residuals after fitting the RMA
model. Figure 3A illustrates quite clearly that several
probes near the 5' region of the gene for WNK1 are
expressed at noticeably lower levels in the kidney samples
than in the remaining samples, as would be expected.
Since the RMA model is fitted robustly, the 5' probes for
the kidney samples, which depart from the parallelism we
saw previously, are downweighted, and so have a rela-
tively small influence on the overall estimation of chip
and probe effects. However, for the determination of dif-
ferential splicing, these observations in the 5' region of the
gene are very much of interest. Figure 3B highlights a
sequence of residuals that appear very different from the
rest of the gene. We return to this observation in the next
Section. Figure 3C shows the genomic context of the
probe-level data and the known Ensembl trancripts for
WNK1. The sequence of residuals showing the persistently
low values suggest kidney-specific expression of the short
transcript ENST00000340908, in agreement with the pre-
viously published result [8].
Related Work
To the best of our knowledge, this paper is the first
attempt at using the Gene platform to investigate alterna-
tive splicing. Several methods have been proposed for the
differential splicing analysis of Exon data, including the
Splicing Index (SI) [10], pattern-based correlation (PAC)
[17], microarray analysis of different splicing (MADS)
[18] and finding isoforms using the robust multichip
average algorithm (FIRMA) [19]. The SI forms a score that
represents the difference between the gene-level summary
(as fitted by an RMA-like algorithm) and an exon-level
summary, requiring two estimation steps. Effectively, the
method estimates probe effects twice independently, one
with all probes for the gene and another with only the
probes for a probeset. We do not see SI as a feasible
approach with Gene data, since even if we were to create
probesets that represent exons, often very few probes will
available and it will be difficult to get reliable estimates.
The fewer probes per probeset will have a similar effect on
applying FIRMA directly to Gene data. FIRMA fits the
standard RMA model (as above) to all probes for a given
gene and summarizes probeset-wise departures from the
model through the residuals. With very few probes, the
probeset summaries of residuals may not be precise. PAC,
Y
ij i j ij
=++
abe
(1)
b
j
j
J
=
=
∑
0
1
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 4 of 13
(page number not for citation purposes)
on the other hand is an all-sample approach that scores
each probeset on whether it correlates with the rest of a
gene, over all samples. Simulation studies for a modest
number of chips (e.g. 20) show that FIRMA and SI gener-
ally outperform PAC [19]. MADS is a new approach for
Exon data that combines several steps, including probe
selection and compensation for sequence-specific cross-
hybridization effects. Though it has not been applied to
the Gene platform, it appears that since calculations are
done at the probe level and combined together to make
inferences about probesets, it may be possible to adapt the
method.
Differential splicing
Scoring persistence of residuals
The method presented in this paper differs from previous
approaches in that we focus on identifying genes with pos-
sible alternative splice forms, instead of highlighting
exons or probesets. This has a subtle statistical advantage
in that the multiple testing penalty is considerably
smaller.
As mentioned above, the residuals from the RMA model
hold the key to finding differential splicing events. Instead
of focusing on individual exons (and the organization of
RMA decomposition of probe-level Affymetrix dataFigure 2
RMA decomposition of probe-level Affymetrix data. Panel A shows the background-adjusted and normalized probe-
level data for PRRX1, from the Affymetrix mixture dataset (see Methods). The probes are displayed in the order which they
map to the human genome (not to scale), and lines join all probe intensities of the same sample. PRRX1 is expressed signifi-
cantly higher in heart tissue compared to brain. Three replicates of pure heart tissue are shown as red lines; green lines repre-
sent pure brain tissue replicates and the blue lines represent a mixture of 75% brain tissue and 25% heart tissue. Panel B shows
the estimated relative probe effects. Panel C shows the chip effects (i.e. summarized expression levels) and Panel D shows
residuals, using the same colour scheme.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 5 of 13
(page number not for citation purposes)
annotation that that requires), we score a persistent devi-
ation from zero of adjacent residuals. The residuals are
defined as:
where and are estimated using robust fits of Equa-
tion 1. In order to normalize across genes, we calculate
standardized residuals = r
ij
/MAD{r
uv
, u = 1, ..., N; v = 1,
..., J} where MAD(.) is defined as 1.4826 times the median
absolute deviation from 0 over all residuals for that gene
and all samples. In the absence of alternative splicing,
standardized residuals will have approximately unit vari-
ance. FIRMA [19] takes advantage of the Exon array
design, where each PSR has 4 probes and residuals can be
summarized at the probeset level. If a particular PSR is dif-
ferentially spliced, then it is expected that most if not all
probes for the PSR would have a large-in-magnitude resid-
ual (i.e. not fit well by the RMA model). For the Gene
array, we are not guaranteed 4 probes per exon and,
depending on the probes designed for a particular tran-
script, may have very little power to detect single exons
that are differentially spliced. Since the performance of
the summary will be related to the number of observa-
tions used to calculate it, we consider an alternative pro-
cedure. We take the approach of finding a persistence of
residuals that are away from zero and in the same direc-
rY
ij ij
=− +().
(2)
ˆ
a
i
ˆ
b
j
%
r
ij
Differential splicing of WNK1Figure 3
Differential splicing of WNK1. Panels A and B show the normalized data and residuals, respectively, of WNK1 for the
Affymetrix tissue dataset (see Methods). The three replicates for human kidney tissue are shown as blue lines, and the remain-
ing 10 tissues (30 samples) are shown with black lines. Panel C shows the set of exonic regions joined together in a gene model
(green) and the three known Ensembl transcripts (blue). The blue lines linking Panels B and C illustrate the correspondence
between probes and exons.
ˆ
a
i
ˆ
b
j
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 6 of 13
(page number not for citation purposes)
tion, thus entirely avoiding the non-uniformity of probes
per exon. We only require that the probes are put in
genomic order for the calculations below. Several adjacent
probes, interrogating exon regions that are adjacent in an
mRNA product, that show the same departure from the
model are evidence of potential differential splicing. For
example, Figure 3 illustrates that probes 2–8 of WNK1 all
have strongly negative residuals. Such an observation is
unlikely to occur by chance.
One possibility to highlight such persistence of residuals
is an extreme value of the absolute values of all partial
sums of adjacent residuals. A natural statistic, inspired by
the monitoring of nuclear material unaccounted for
(MUF) [20], is the maximum absolute partial sum:
over the J(J + 1)/2 possible consecutive sums of J probes.
This calculation is repeated separately for each gene, giv-
ing a score for each gene and each sample. That is, this
approach can be applied in the absence of replicates.
However, if replicates are available, we recommend
precomputing a probe-wise summarized residual
and use this in place of in
Equation 3, where i(k) is the index of replicate k.
The MUF statistic is very flexible. An extreme MUF statistic
can result from a single probe if it is extreme enough. But,
it can also highlight a subtle change that persists across
any number of probes if the score is deemed to be
extreme. Notice that the denominator of the partial sums
is the square root of the number of data points. This
ensures the variances (of the sum) are constant, thus
putting all the partial sums on an even footing.
As the number of probes increases, there are more partial
sums to consider, making the distribution of maximum
order statistics more likely to take on more extreme val-
ues. To alleviate this, we repeatedly sample J probes from
the empirical distribution of all standardized residuals
and calculated the MUF score, giving a null distribution of
MUF scores for J probes. A false discovery rate can be cal-
culated for the discoveries above a given quantile of the
null distributions.
We call this approach FIRMAGene, since it is only a small
modification to FIRMA [19], in terms of operating on
residuals from an RMA fit, but is applied to the Affymetrix
Gene 1.0 ST platform and scores differential splicing at the
gene level instead of the probeset level.
Results
Validation of using the Gene platform for splicing
We first validate the approach of using the Gene platform
for differential splice detection by comparing the residuals
for a gene known to express a vastly different isoform in
human brain [21], using the publicly available data of the
same tissue RNA hybridized to both the Gene and Exon
platforms. Figure 4 shows residuals plots for MBP (myelin
basic protein) and highlights a very distinct pattern in the
brain samples. This pattern is observed almost identically
from the 36 probes represented on the Gene array or 72
probes from the Exon array. The exact splicing mecha-
nism is not as apparent as in the previous example
(WNK1, Figure 3), but it is straightforward to put the
probe-level data in the context of known genome annota-
tion. For a genome-wide comparison, we matched the
probes from the Gene array to the Affymetrix-defined
probesets of the Exon array, allowing us to run FIRMA on
the Gene platform. Note that we are not advocating the
use of FIRMA for the Gene platform, although we do
highlight that it can be done and allows us to make the
comparison. See Methods for further description of proce-
dures used to construct the annotation. FIRMA scores are
calculated for Gene and Exon data, generating a table of
scores by probeset and sample, one for each platform.
Note that the summaries for the two platforms are often
from different numbers of probes and therefore have dif-
ferent precision. Table 1 give a cross-tabulation of the
numbers of probes for both platforms amongst matched
probesets. The Exon array most often has 4 probes per
probeset, whereas the Gene platform most often has 1 or
2 probes. In some cases (e.g. genes with few exons), Gene
will have more than 4 probes. Taking the average of
FIRMA scores over the 3 brain replicates, Figure 5 illus-
trates convincing genome-wide evidence that extreme
residuals observed on the Exon array are also observed on
the Gene array (correlation r = 0.53 over more than
230,000 Exon probesets). This is especially promising
considering the majority of summarized sets of residuals
will be centred close to 0. Shown in Figure 5 are summa-
ries for the brain replicates from each platform, since
brain tissue is expected to exhibit more alternative splic-
ing than most other tissues.
Next, we were interested to determine whether Gene data
is able to detect a significant proportion of the differential
splicing events that FIRMA detects on Exon data. The tis-
sue panel dataset, where the same source of RNA is
hybridized to both platforms, is an ideal test set for this
comparison. We applied FIRMA to the Exon data and FIR-
MAGene to the Gene data. We compared the top 100
probeset-tissue scores from FIRMA to the corresponding
gene-tissue scores from FIRMAGene, as shown in Figure 6.
The vast majority of the MUF statistics are large in magni-
tude, suggesting that Gene platform is quite capable of
detecting similar differential splicing events. In fact, 86 of
V
r
ij
js
e
es
i
J
seJ
()
max=
=
∑
−+
≤≤≤1
1
%
(3)
rnr
kj t ik j
k
n
k
⋅
=
=
()
∑
1
1
/
()
%
%
r
ij
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 7 of 13
(page number not for citation purposes)
the 100 corresponding gene-tissue scores have MUF statis-
tics more extreme than the 95th percentile of their null
distribution. One the other hand, because FIRMA gives
(sub-)exon-level and FIRMAGene gives gene-level statis-
tics, there may be some cases where the scores do not cor-
respond. For example, 4 of the 14 MUF statistics that are
not extreme have no Gene probes represented in the
region where the Exon probes are. Furthermore, since the
MUF score is an extreme value statistic, there may be set of
probes within the gene that are more extreme in the oppo-
site direction, as shown in 8 of the 14 non-extreme MUF
scores. Overall, this analysis suggests that the Gene plat-
form will be quite promising for the analysis of differen-
tial splicing.
Tissue panel dataset
The publicly available 11 tissue panel dataset, where the
same human tissue RNA was run on both the Gene and
Exon platforms in a single laboratory by Affymetrix, pro-
vides an ideal testing ground for the methodology and for
illustrating of some of the features of whole-transcript
microarray data. Although there are many individual
examples in the literature, there is no readily available
positive control set of tissue-specific alternative splice
events that can be used for benchmarking. However,
tables of EST-based predictions exist. A rigorous compari-
son of EST predictions and microarray analysis of alterna-
tive splicing events is beyond the scope of this study.
Instead, we calculate scores genome-wide (using Gene
data) across the 11 tissues and show that many of the top
ranking scores have been observed previously to either
have tissue-specific variants or tissue-specific expression
patterns.
Figure 7 shows the genome-wide scores, stratified by the
number of probes for each gene. The plot shows only the
genes that have between 10 and 70 probes (nearly 95% of
the genes on the array). Because genes with more probes
have more partial sums to consider, the maximum gently
increases with the number of probes per gene. The two
examples shown earlier, WNK1 for kidney tissue and MBP
for brain tissue, have high scores, as highlighted. Table 2
shows the top scoring gene-tissue combinations. Of the
top 20 gene-tissues scores, many of them have previous
evidence of tissue-specific behaviour. Plots of the normal-
ized data and residuals can be found in Additional file 1,
in addition to a list of publications corroborating the tis-
sue-specific evidence.
Some tissues have considerably more differential splice
detections. For example, of the top 1000 gene-tissue
scores (see Additional file 2), the top three tissues are tes-
tis (295), brain (258) and liver (116). This corresponds
with previous EST studies where brain, liver and testis
have the highest percentage of alternatively spliced genes
[22].
Conclusion and Discussion
We have proposed a novel scoring method called FIR-
MAGene based on decomposing probe-level microarray
Normalized probe-level data and RMA residuals for MBPFigure 4
Normalized probe-level data and RMA residuals for MBP. Panels A and B show the residuals for Gene and Exon for
RMA fits, respectively. There are 36 probes for Gene and 72 probes for Exon. Both panels show 33 lines, one for each
hybridization (11 tissues with 3 biological replicates each). The brain and muscle replicates are shown blue and red lines,
respectively.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 8 of 13
(page number not for citation purposes)
data with a linear model. The major motivation for this
work is to provide an extra investigation, in addition to
differential expression analysis, thus giving researchers
added value from their collected data. The design of the
latest generation of Affymetrix expression array facilitates
this. Using a public tissue panel dataset, we show the
method highlights many previously known and poten-
tially new differential tissue-specific splice events and
shows strong correspondence with the Exon array over the
same RNA samples. The strategy we propose can be
applied directly to the Affymetrix Human, Mouse and Rat
Gene 1.0 ST platforms, or any other whole-transcript plat-
form that exhibits probe-specific effects. Although we
have not investigated thoroughly, FIRMAGene may be
A comparison of FIRMA scores for Gene and Exon platformsFigure 5
A comparison of FIRMA scores for Gene and Exon platforms. Each point in the scatter plot represents an Exon
probeset that has been matched to probes on the Gene array. The X-axis gives the averaged (over brain replicates) FIRMA
score for Exon data. The Y-axis gives the average FIRMA score for the corresponding Gene samples.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 9 of 13
(page number not for citation purposes)
useful for the Exon array. It comes at some additional
computational cost, since there are more probes (and
therefore even more partial sums), but it may be better
able to highlight smaller, but persistent, changes in adja-
cent probesets. The procedure can operate in a single sam-
ple mode or can make use of replicates. Technical or
biological replicates can be used, although significant
detections from the latter will give more generalizable
results. One subtle difference that FIRMAGene takes
advantage of, is the fact that scoring by gene instead of by
exon results in a much smaller penalty for multiple test-
ing.
The Gene platform will be used in various profiling stud-
ies and this work simply provides an additional analysis
that will be of interest. The approach is not without limi-
tations. The Gene platform only covers well-annotated
exons, whereas the Exon platform covers a considerable
amount of additional content, based on either EST evi-
dence or computational predictions. However, having no
features representing predicted exonic content has some
advantages. For example, in the analysis of Exon data, it is
not always clear whether to include all probesets (for well-
characterized and predicted exons together) in the RMA
model fit. The MADS approach uses a computational
probe selection for this [18]. In many cases, the probes for
content with weak evidence are not used for the primary
analysis [19]. Since the well-characterized exonic content
on the Exon array only represents approximately 20% of
all features, the selection of probesets to include may have
a large impact on the differential splice detections. In
addition, for short genes, the Gene platform will generally
have more probes than the Exon array, giving potentially
higher power to detect new variants.
Since we are scoring a gene over all partial sums of probes,
the MUF score is very flexible. It simultaneously searches
for extreme residuals over any number of adjacent probes,
including a single probe if it is extreme enough. There are
variations of the MUF score that may be worth pursuing
for a more refined mapping of differential splice events.
For example, it is generally unreasonable that all residuals
for a single sample will be non-zero. It may suffice to con-
sider only partial sums of length less than
LJ/2O, for exam-
ple. Another variation would be to target specific patterns.
For example, SLC25A3 (solute carrier family 25, member
3) has a very distinct mutually exclusive differential splic-
ing pattern (see Additional file 1). If this or other distinc-
tive patterns were of particular interest, the scoring of
adjacent residuals could be tailored towards it.
It is difficult to know in what experimental circumstances
the Gene platform and a procedure such as FIRMAGene
will be most successful. We have shown FIRMAGene can
be useful in a panel of tissues, where in general the major-
ity of samples exhibit the same probe-level pattern and
only a small number of samples differ. We expect the pro-
cedure will be useful even in a balanced two group com-
parison, where differential isoform usage would still
present as a persistent departure from the linear model.
However, there may be limitations in the robust fitting for
probe effects in cases where the probe intensities are split
into two distinct groups. One possible option would be to
use existing Gene data (e.g. from a public source), in order
to stabilize the probe effect determination. We have not
investigated this thoroughly. As mentioned above, micro-
arrays are only able to detect differential splicing, so in
order to detect such events, there needs to be enough var-
iation amongst the samples for a pattern to stick out.
Depending on the strength of the difference and the
number of probes represented on the array for the alterna-
tive spicing event (which can vary from gene to gene), a
large sample size may be required.
Identifying departures through residuals from the RMA
model will not always be perfect. Some departures from
the RMA linear model may not be alternative splicing at
all. In some cases, large residuals may be a result of cross-
hybridizing probes, or through probes that have a differ-
ent range of intensity, or are induced through, for exam-
ple, an exon that is not expressed in any of the samples in
combination with strong differential expression. It may be
possible to compensate for cross-hybridization, as dem-
onstrated recently (see [23,24]). With relevance to studies
involving human populations, it has been recently shown
that single nucleotide polymorphisms can significantly
affect probe-level Exon data [25]. In addition, a resource
has now been created to track Exon probes that may be
affected [26]. Individual probe performance aside, we
argue that most of the detected examples are biologically
meaningful and these problems are not isolated to FIR-
MAGene and represent the challenging nature of design-
ing methodology that operates over a range of probe
behaviours. Other procedures, such as SI, MADS or PAC if
they were to be adapted to the Gene platform, would need
to effectively deal with these same challenges.
There are a number of other issues that we are aware of,
but are beyond the scope of this investigation. For exam-
Table 1: Number of probes per matched probeset. After taking a
subset of common probesets, the numbers of probes for each
matched Exon probe selection region are given.
No. of Gene Probes
No. of Exon Probes12345
168020000
226544191 6 1 1
3279427881155 11 7
4 87092 83621 24219 12301 4019
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 10 of 13
(page number not for citation purposes)
FIRMAGene scores for the top 100 FIRMA scores for the Affymetrix tissue panel datasetFigure 6
FIRMAGene scores for the top 100 FIRMA scores for the Affymetrix tissue panel dataset. The X-axis gives the
FIRMA score (calculated on the Affymetrix Exon tissue panel dataset) for the top 100 probesets with 4 probes. The Y-axis
gives the signed MUF scores (calculated on the Affymetrix Gene tissue panel dataset) for the corresponding genes. Circles
which are filled in correspond to MUF scores that are more extreme than the 95
th
percentile of the permutation-based null
distribution.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 11 of 13
(page number not for citation purposes)
ple, in some cases, the probes for a gene are overlapping.
This may induce a correlation between residuals of neigh-
bouring probes. The current model assumes independ-
ence for all probes and makes no compensation for this.
As evidenced by the top ranked genes, our current scoring
scheme seems reasonable and does highlight interesting
cases. Despite the limitations mentioned above, this
research highlights an additional avenue of investigation
beyond differential expression that is freely available at a
minimal additional computational cost.
Methods
Datasets
The mixture dataset used for illustration of RMA (Figure
2) and the tissue panel datasets (Gene and Exon) were
run by Affymetrix and made publicly available (see http:/
/www.affymetrix.com). Briefly, the mixture dataset com-
prises 33 total samples, 3 technical replicates each of 11
separate mixtures. The tissue panel datasets use the same
RNA on both the Gene and Exon platforms. Again, there
are 33 total samples representing 3 biological replicates of
each of the following human tissues: brain, thyroid,
breast, pancreas, prostate, heart, skeletal muscle, kidney,
testis, spleen and liver.
FIRMAGene scoring for the Affymetrix tissue panel datasetFigure 7
FIRMAGene scoring for the Affymetrix tissue panel dataset. Each point in the plot gives a FIRMAGene score for a tis-
sue-gene combination. The X-axis gives the number of probes and Y-axis gives the raw MUF score. Jitter is added in the X
dimension. The simulated null distribution is shown as blue boxplots. The scores for gene-tissue combinations shown in Figures
3 and 4 are highlighted.
Table 2: Top scoring tissue-specific differential splicing
candidates.
ID Sample Score* Symbol
7922737 Testis 24.76 C1orf14
8086077 Brain 21.84 CLASP2
8086842 Brain 20.99 MAP4
7957746 SkMus 19.76 SLC25A3
7957746 Heart 19.41 SLC25A3
8165653 Heart 18.72 --
8166876 Testis 18.29 DDX3X
8064191 Brain 18.08 TPD52L2
8007188 Brain 18.01 CNP
7922627 Kidney 18.01 NPHS2
8170215 Liver 17.93 F9
8100458 Testis 17.84 PDCL2
7962194 Testis 17.78 LOC440093
7940971 Testis 17.56 KCNK4
8176419 Testis 17.18 TSPY2
8155203 Brain 16.97 CLTA
8170390 Brain 16.84 --
8023889 Brain 16.79 MBP
8176544 Testis 16.64 TSPY1
8024194 Testis 16.32 GPX4
The Affymetrix Human Gene 1.0 ST identifiers and gene symbols are
given for the top 20 tissue-gene combinations.
*See Methods.
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 12 of 13
(page number not for citation purposes)
Data processing
All data processing has been performed in the open source
statistical package R[27] and the methods implemented in
this paper are available from the authors as an R package,
operating on objects created using the aroma.affymetrix
package [28]. Chip definition files (CDFs) have been cre-
ated for both arrays, based on library files and annotation
made available from Affymetrix, using the Bioconductor
[29] affxparser package. To facilitate alternative splicing
analysis, probe collections are organized in a gene-centric
fashion, so that probes from all known isoforms for a
gene can be analyzed by a single framework (i.e. fit with
the RMA model). For the Exon platform, we are used core
probesets only. For the Gene array, the coordinates of the
probes are matched to the Exon probeset coordinates, so
that summaries for the same regions can be compared.
Some probes for the Gene array, however, fall outside the
region of Exon's PSR. These are still kept within the Gene
probe collection, but not used for the comparison.
Running FIRMAGene consists of the following steps: 1) fit
the RMA probe-level model robustly for each gene, 2)
standardize the residuals by dividing by the gene-wise
MAD and summarize over residuals if replicates are used,
3) calculate the maximum MUF score for each sample, 4)
given the number of probes for a gene, sample a large
number of vectors of residuals (from the empirical distri-
bution of all residuals) of same length, calculate the MUF
score on each one to generate the null distribution, 5) at a
given cutoff, calculate the false discovery rate. An example
R script for running these steps on the tissue dataset is pro-
vided in Additional file 3.
The score represented in Table 2 compares the tissue-gene
score to the mean and standard deviation of the permuta-
tion-based null distribution (subtract mean, divide by
standard deviation).
Authors' contributions
MDR conceived the original idea, analyzed the data,
implemented the software and wrote the paper. TPS
refined the statistical analysis and directed the project.
Additional material
Acknowledgements
The authors wish to thank Henrik Bengtsson for various discussions
regarding implementation using aroma.affymetrix data structures. In addi-
tion, we wish to thank Anna Tsykin, Mark Cowley and Paul Spellman for
helpful comments on an earlier version of the manuscript and Elizabeth Pur-
dom for discussions regarding FIRMA.
References
1. Johnson JM, Castle J, Garrett-Engele P, Kan Z, Loerch PM, Armour
CD, Santos R, Schadt EE, Stoughton R, Shoemaker DD: Genome-
wide survey of human alternative pre-mRNA splicing with
exon junction microarrays. Science 2003, 302(5653):2141-2144.
2. Pan Q, Shai O, Lee LJ, Frey BJ, Blencowe BJ: Deep surveying of
alternative splicing complexity in the human transcriptome
by high-throughput sequencing. Nat Genet 2008,
40(12):1413-1415.
3. Wang ET, Sandberg R, Luo S, Khrebtukova I, Zhang L, Mayr C,
Kingsmore SF, Schroth GP, Burge CB: Alternative isoform regu-
lation in human tissue transcriptomes. Nature 2008,
456:470-476.
4. Schwerk C, Schulze-Osthoff K: Regulation of apoptosis by alter-
native pre-mRNA splicing. Mol Cell 2005, 19:1-13.
5. Singh N, Preiser P, Renia L, Balu B, Barnwell J, Blair P, Jarra W, Voza
T, Landau I, Adams JH: Conservation and developmental con-
trol of alternative splicing in maebl among malaria parasites.
J Mol Biol 2004, 343(3):589-599.
6. Grabowski PJ, Black DL: Alternative RNA splicing in the nerv-
ous system. Prog Neurobiol 2001, 65(3):289-308.
7. Venables JP: Aberrant and alternative splicing in cancer. Cancer
Res 2004, 64(21):7647-7654.
8. Xu Q, Modrek B, Lee C: Genome-wide detection of tissue-spe-
cific alternative splicing in the human transcriptome. Nucleic
Acids Res 2002, 30(17):3754-3766.
9. Gupta S, Zink D, Korn B, Vingron M, Haas SA: Strengths and
weaknesses of EST-based prediction of tissue-specific alter-
native splicing. BMC Genomics 2004, 5:72.
10. Clark TA, Schweitzer AC, Chen TX, Staples MK, Lu G, Wang H, Wil-
liams A, Blume JE: Discovery of tissue-specific exons using com-
prehensive human exon microarrays. Genome Biol 2007,
8(4):R64.
11. Gardina PJ, Clark TA, Shimada B, Staples MK, Yang Q, Veitch J, Sch-
weitzer A, Awad T, Sugnet C, Dee S, Davies C, Williams A, Turpaz Y:
Alternative splicing and differential gene expression in colon
cancer detected by a whole genome exon array. BMC Genom-
ics 2006, 7:325.
12. Robinson MD, Speed TP: A comparison of Affymetrix gene
expression arrays. BMC Bioinformatics 2007, 8:449.
13. Bemmo A, Benovoy D, Kwan T, Gaffney DJ, Jensen RV, Majewski J:
Gene expression and isoform variation analysis using
Affymetrix Exon Arrays. BMC Genomics 2008, 9:529.
Additional file 1
Plots and corroborating evidence for the top 20 gene-tissue scores.
Probe-level data and residuals for the top 20 gene-tissue scores, from
applying FIRMAGene to the Affymetrix tissue panel dataset. Additionally,
links to various corroborating evidence of tissue-specific splicing or expres-
sion.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-
2105-10-156-S1.pdf]
Additional file 2
Top 1000 Gene-tissue scores for the tissue panel dataset. Table giving
the probeset identifier, tissue sample, FIRMAGene score and gene symbol,
after applying FIRMAGene to the Affymetrix tissue panel dataset.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-
2105-10-156-S2.zip]
Additional file 3
Example R script for FIRMAGene (R). Source code example to run FIR-
MAGene on the Affymetrix tissue panel dataset.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-
2105-10-156-S3.zip]
Publish with Bio Med Central and every
scientist can read your work free of charge
"BioMed Central will be the most significant development for
disseminating the results of biomedical research in our lifetime."
Sir Paul Nurse, Cancer Research UK
Your research papers will be:
available free of charge to the entire biomedical community
peer reviewed and published immediately upon acceptance
cited in PubMed and archived on PubMed Central
yours — you keep the copyright
Submit your manuscript here:
http://www.biomedcentral.com/info/publishing_adv.asp
BioMedcentral
BMC Bioinformatics 2009, 10:156 http://www.biomedcentral.com/1471-2105/10/156
Page 13 of 13
(page number not for citation purposes)
14. Kent WJ, Sugnet CW, Furey TS, Roskin KM, Pringle TH, Zahler AM,
Haussler D: The human genome browser at UCSC. Genome
Res 2002, 12(6):996-1006.
15. Irizarry RA, Bolstad BM, Collin F, Cope LM, Hobbs B, Speed TP:
Summaries of Affymetrix GeneChip probe level data. Nucleic
Acids Res 2003, 31(4):e15.
16. Marazzi A: Algorithms, Routines and S Functions for Robust Statistics
Pacific Grove, California: Wadsworth & Brooks/Cole; 1993.
17. French PJ, Peeters J, Horsman S, Duijm E, Siccama I, Bent MJ van den,
Luider TM, Kros JM, Spek P van der, Smitt PAS: Identification of dif-
ferentially regulated splice variants and novel exons in glial
brain tumors using exon expression arrays. Cancer Res 2007,
67(12):5635-5642.
18. Xing Y, Stoilov P, Kapur K, Han A, Jiang H, Shen S, Black DL, Wong
WH: MADS: a new and improved method for analysis of dif-
ferential alternative splicing by exon-tiling microarrays. RNA
2008, 14(8):1470-1479.
19. Purdom E, Simpson KM, Robinson MD, Conboy JG, Lapuk AV, Speed
TP: FIRMA: a method for detection of alternative splicing
from exon array data. Bioinformatics 2008, 24(15):1707-1714.
20. Speed TP, Culpin D: The Role of Statistics in Nuclear Materials
Accounting: Issues and Problems. Journal of the Royal Statistical
Society Series A 1986, 149:281-313.
21. de Ferra F, Engh H, Hudson L, Kamholz J, Puckett C, Molineaux S, Laz-
zarini RA: Alternative splicing accounts for the four forms of
myelin basic protein. Cell 1985, 43(3 Pt 2):721-727.
22. Yeo G, Holste D, Kreiman G, Burge CB: Variation in alternative
splicing across human tissues. Genome Biol 2004, 5(10):R74.
23. Huang JC, Morris QD, Hughes TR, Frey BJ: GenXHC: a probabil-
istic generative model for cross-hybridization compensation
in high-density genome-wide microarray data. Bioinformatics
2005, 21(Suppl 1):i222-i231.
24. Kapur K, Jiang H, Xing Y, Wong WH: Cross-Hybridization Mod-
eling on Affymetrix Exon Arrays. Bioinformatics 2008.
25. Zhang W, Duan S, Bleibel W, Wisel S, Huang R, Wu X, He L, Clark
T, Chen T, Schweitzer A, Blume J, Dolan M, Cox N: Identification
of common genetic variants that account for transcript iso-
form variation between human populations. Hum Genet 2008,
125:81-93.
26. Duan S, Zhang W, Bleibel WK, Cox NJ, Dolan ME: SNPinProbe
1.0: A database for filtering out probes in the Affymetrix
GeneChip(R) Human Exon 1.0 ST array potentially affected
by SNPs. Bioinformation 2008, 2(10):469-470.
27. R Development Core Team: R: A Language and Environment for Statis-
tical Computing 2008 [http://www.R-project.org
]. R Foundation for
Statistical Computing, Vienna, Austria ISBN 3-900051-07-0
28. Bengtsson H, Simpson K, Bullard J, Hansen K: aroma.affyme
trix: An R framework for analyzing small to large Affyme-
trix data sets in bounded memory. Tech Report 745, Department
of Statistics, University of California, Berkeley 2008.
29. Gentleman RC, Carey VJ, Bates DM, Bolstad B, Dettling M, Dudoit S,
Ellis B, Gautier L, Ge Y, Gentry J, Hornik K, Hothorn T, Huber W,
Iacus S, Irizarry R, Leisch F, Li C, Maechler M, Rossini AJ, Sawitzki G,
Smith C, Smyth G, Tierney L, Yang JYH, Zhang J: Bioconductor:
open software development for computational biology and
bioinformatics. Genome Biol 2004, 5(10):R80.
