![Analysis of Vegfa exon 6 muscle-dependent inclusion. A subset of the summary page produced by AVISPA is shown. (a) Feature enrichment analysis: the values of the features listed on the left are computed for Vegfa exon 6 and compared against matching feature values in a set of labeled exons. The four sets of exons compared against here are alternative exons ('AS', third column from the left), constitutive exons ('Const', third column from the right), exons differentially included in muscle ('Muscle Inc', second column from the right), and differentially excluded in muscle ('Muscle Exc', right most column). Relative enrichment or depletion of features is indicated using the heat map on the right. Only features with significantly low (blue) and high (red) values are shown here. The genomic region of each feature is indicated by the second from left column using the notation and colors in the top figure. (b) Stacked bar chart (left) of the normalized feature effect (NFE, y-axis) on splicing prediction. Only the top motifs are shown. Motif regions are annotated using the color scheme depicted below. Mapping of the motifs onto the UCSC genome browser is shown on the right. Tracks combining all motifs used by the code model (red), the unbiased motif search [5] (grey scaled), and conservation (blue) are added at the bottom.](https://www.researchgate.net/publication/258042106/figure/fig1/AS:338844378517547@1457798203384/Analysis-of-Vegfa-exon-6-muscle-dependent-inclusion-A-subset-of-the-summary-page_Q320.jpg)
SOF T W A R E Open Access
AVISPA: a web tool for the prediction and
analysis of alternative splicing
Yoseph Barash
1,2,4,5*
, Jorge Vaquero-Garcia
1,2
, Juan González-Vallinas
1,3†
, Hui Yuan Xiong
4†
, Weijun Gao
4
,
Leo J Lee
4
and Brendan J Frey
4,5
Abstract
Transcriptome complexity and its relation to numerous diseases underpins the need to predict in silico splice
variants and the regulatory elements that affect them. Building upon our recently described splicin g code, we
developed AVISPA, a Galaxy-based web tool for splicing prediction and analysis. Given an exon and its proximal
sequence, the tool predicts whether the exon is alternatively spliced, displays tissue-dependent splicing patterns,
and whether it has associated regulatory elements. We assess AVISPA's accuracy on an independent dataset of
tissue-dependent exons, and illustrate how the tool can be applied to analyze a gene of interest. AVISPA is
available at http://avispa.biociphers.org.
Alternative splicing (AS) is estimated to affect tran-
scripts from over 95% of human multi-exon genes [1,2],
with the most common class of AS involving cassette
exons. Thousands of alternative cassette exons have
been found to be differentially spliced between mamma-
lian tissues, with tissues such as the brain displaying the
most complex patterns [1,2]. These observations and the
association of many splicing defects with diseases [3]
motivated the recent derivation of a splicing code. The
code, comprising a model with a set of rules that can
predict splicing outcomes given genomic sequence and
cellular context [4,5], used over 1,000 regulatory fea-
tures. Trained using inclusion measurements for 3,700
cassette exons across 27 mouse tissues , the code’s model
was shown to predict differential AS in four tissue
groups: the central nervous system (CNS), muscle, di-
gestive, and embryo versus adult tissues.
The derivation of a predictive splicing code served as
proof-of-concept and enabled insights into RNA biogen-
esis [5,6], but was limited in scope. Specifically, it was
only applied to a subset of alternative exons in specific
studies. However, given the importance of splicing in the
study of gene regulation, development and disease, it
became important to translate the splicing code models
into a tool that would be accessible for researchers
in a wide range of fields. Here, we present AVISPA
(Advanced Visualization of Splicing Prediction and Ana-
lysis), a web tool that enables both prediction and spli-
cing analysis of alternative and tissue-dependent exons
in any gene of interest. Given an exon, the tool predicts
whether it is alternative and whether its inclusion is ex-
pected to change in different tissues. It reports whether
the exon is known to be alternative based on an internal
transcripts database, and perform s in silico splicing ana-
lysis, identifying putative regu latory elements and map-
ping those as tracks in the genome browser.
AVISPA’s pipeline is illustrated in Figure 1. Users sub-
mit a query by specifying the sequence or genomic coor-
dinates of either a single exon, or a triplet of exons that
includes the immediate up- and downstream exons of
the query exon. In the pre-processing step, the query is
matched against an internal database of exon triplets
mined from known transcripts and mapped to the refer-
ence genome. The result of the pre-processing is re-
ported in the AVISPA’s output and indicates existing
evidence for whether the exon is alternatively spliced
based on, for example, alignments of cDNA and E ST
data. After the query has been successfully matched,
RNA features are extracted from the query exon and
flanking regions [5]. At the first prediction stage, the
* Correspondence: yosephb@upenn.edu
†
Equal contributors
1
Department of Genetics, University of Pennsylvania, Philadelphia, PA 19104,
USA
2
Department of Computer and Information Science, University of
Pennsylvania, Philadelphia, PA 19104, USA
Full list of author information is available at the end of the article
© 2013 Barash et al.; 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.
Barash et al. Genome Biology 2013, 14:R114
http://genomebiology.com/2013/14/10/R114
extracted features are used to predict whether the query
exon is alternatively or constitutively spliced. If the
query is predicted to be an alternative cassette exon, a
second prediction step assesses whether the exon is dif-
ferentially included in specific tissues.
The new web tool offers marked improvements over
available software. First , it offers 'genome-wide' tissue-
dependent splicing predictions, where any exon can be
submitted as a query. By contrast, the original work only
allowed analysis on a previously mined set of approxi-
mately 12,000 cassette exons, while other tools focus on
quantifying experimental data or general splice site and
motif analysis [7-9]. Second, AVISPA offers a new in
silico analysis of regulatory features and the mapping of
putative regulatory sequence motifs in the genome. As
part of this analysis, motifs found to be robustly in-
cluded in the B ayesian ensemble of models and present
in the query are removed in silico to determine their ef-
fect on splicing prediction. The relative effect of the se
feature removals is reported as a bar chart of the nor-
malized feature effect (NFE). The putative regulatory
motifs are also mapped to the genome using the UCSC
genome browser, where they can be combined with
other tracks, such as known single nucleotide polymor-
phisms and binding measurements of known splicing
factors [10]. Additionally, the enrichment of the query’s
features is compared to refer ence groups such as alter-
native or constitutively spliced exons in AVISPA’s data-
base. Feature enrichment is reported using a standard
heat map ranging from blue, for relatively low values , to
red for relatively high values. For example, a relatively
strong 3’ splice site will appear red, indicating a high
score, while a weak splice site will be marked blue.
The new tool also includes several other improve-
ments. First, the prediction technique is now based on a
Bayesian neural network, which provides improved pre-
diction accuracy compared to a battery of other methods
[11]. Second, the original dataset of 3,700 cassette exons
has been expanded to approximately 30,000 exons using
data from 33 experiments in 11 mouse tissues [12].
Third, AVISPA uses an extended set of features that in-
clude computationally predicted nucleosome occupancy
[13] together with primary sequence motifs implicated
in general splicing regulation.
Assessing splicing prediction accuracy
The new two-stage prediction paradigm, combined with
the expanded dataset, yields a significant improvement
in detecting alternative cassette exons (Figure 2a). For
example, using only tissue-d ependent splicing predictors
achieves an area under the curve (AUC) of 64% for dis-
tinguishing between alternative and constitu tive exons,
compared to 86% by the first stage classifier. The im-
proved accuracy of 94% AUC achieved for detecting
tissue-dependent exons is to be expected, as many
regulatory features and higher intronic conservation
are associated with such exons. Notably, AVISPA’s se-
quence-based predictions offer a significant improve-
ment compared to a similar classifier that directly uses
normalized exon expression measurements from 33 ex-
periments [12]. The latter achieves an overall lower ac-
curacy of 71% AUC, with a significantly 2.5-fold lower
sensitivity (54% versus 21%) for high-confidence events
at a false positive rate of 2%. These results illustr ate the
usefulness of the new tool, which generalizes over ex-
perimental conditions and is not limited by technical
factors such as microarray noise or read coverage. We
note that these accuracy estimates can be considered as
lower bounds, as some of the events labeled as constitu-
tive in our database may be alternative.
Motif Effect
Motif 1
Motif 2
Motif 3
Motif Effect
Feat 1
Feat 2
Feat 4
Feat 1
Feat 3
Feat 4
PreA I1(5') A PostAI1(3') I2(5') I2(3')
Matched query
Alternative splicing
Constitutive
Alternative
Muscle
Digestive
Embryo
CNS
CNS-dependent splicing
Pr[AS]
Motif effect
Motif map
Motif 1
Motif 2
Motif 3
Feat 2
Feat 1
Feat 3
Feat 3
Feat 4
Feat 3
Feat 2
Low
values
High
values
Pr[Ts]
PreA A PostA
A
or
(1) Query submission
Cassette DB
Transcript DB
Genome
Unmatched Unmatched
Unmatched
Matched Matched Matched
(2) Query Matching
(4) Splicing regulatory analysis
(3) Splicing prediction
Feature enrichment
Reg F[AS] F[Con] F[Inc] F[Exc]
Figure 1 AVISPA’s analysis pipeline. The analysis is composed of the following steps. (1) Query submission: users submit a query composed of
either a single exon of interest or an exon triplet that also specifies the up- and downstream exons. (2) Query matching: the submitted query is
first matched against internal databases (DB) of known transcripts and alternative exons. If no match is found the query is searched against the
reference genome. If the query cannot be matched (red cross) an error is reported. (3) Splicing prediction: a successfully matched query
(light blue rectangle) is scored as an alternative cassette exon, followed by scoring for differential splicing in four tissue groups. (4) Splicing
analysis: if the query’s predictions pass a user-defined significance threshold a splicing analysis is performed. Analysis includes feature enrichment,
effect of in silico motif removal on splicing predictions, and mapping putative regulatory motifs to the genome. A visual summary of both predic-
tions and splicing analysis is produced (right).
Barash et al. Genome Biology 2013, 14:R114 Page 2 of 8
http://genomebiology.com/2013/14/10/R114
The new tool also achieves significant improvement in
detecting tissue-depend ent exons (Figure 2b). The over-
all accuracy in discriminating between tissue-dependent
and non-tissue-dependent exons is 89% AUC, but varies
considerably between tissues and between differential in-
clusion and exclusion in the same tissue type. For ex-
ample, the highest accuracy was achieved for detecting
increased inclusion of exons in CNS (94% AUC) and
muscle tissues (91% AUC), while the lowest accuracy was
for detecting increased exclusion in CNS (85% AUC) and
increased inclusion in embryonic tissues (82% AUC).
In order to test AVISPA on an independent dataset,
we computed predictions for a set of cassette exons re-
cently shown to be regulated by the Muscleblind-like
proteins Mbnl1/2 in mouse brain, muscle, and heart
[14]. Figure 2c shows AVISPA easily distinguished these
exons from constitutive exons (97% AUC), similar to its
performance in detecting tissue-dependent alternative
exons in the original test set. In discriminating the
Mbnl1/2-regulated exons from non-CNS- and non-
muscle-dependent exons, AVISPA achieves an AUC of
93% and 94%, respectively, while in silico removal of
Mbnl1/2 caused, on average, an almost two-fold larger
effect for Mbnl1/2-regulated exons compared to the ef-
fect for non-muscle- and non-heart-dependent exons.
The improved accuracy in detecting Mbnl1/2-regulated
exons compared to the detection of tissue-dependent
exons in the original test data is likely due to a lower
false detection rate from the RNA-Seq and CLIP-Seq ex-
periments in [14].
Finally, we also tested whether the regulatory features
added in the web tool were useful for splicing prediction.
As expected, many of the sequence motifs implicated in
general splicing regulation were included in the code,
especially for differentiating between alternative and
constitutive exons. By contrast, the relation between nu-
cleosome occupancy and alternative splicing is less well
understood, and has garnered much research attention
[15,16]. We found that the model selected features
representing nucleosome occupancy around the alterna-
tive exon, but training the model without these features
resulted in similar prediction accuracy (data not shown).
This result indicates that other features in our model,
such as di- and tri-nucleotide frequencies, already cap-
tured the 'predictive power' of computationally derived
nucleosome position features.
Vegfa in silico splicing analysis
Previous work demonstrated how the splicing code
model could be used to identify new regulatory ele-
ments, detect novel tissue-dependent splicing e vents,
and study the evolution of splicing across vertebrates
[6]. Here, we illustrate how the new tool can be used to
analyze a well-studied gene of major interest. We
applied AVISPA to the vascular endothelial growth fac-
tor A (Vegfa) gene. Vegfa has a complex and highly
Alternative exons
Pr[Ts] − 64%
33 exon arrays − 71%
Pr[AS] − 86%
Pr[AS] Tissue dep. − 91%
Rand
0.2
0.4
0.6
0.8
1
Sensitivity
0
Pr[AS] Mbnl dep. − 97%
CNS Mbnl dep. − 94%
Muscle Mbnl dep. − 95%
Rand
Rand
Tissue-dependent exons
0 0.2 0.4 0.6 0.8 1
1−Specificity
0 0.2 0.4 0.6 0.8 1
1−Specificity
0 0.2 0.4 0.6 0.8 1
1−Specificity
Mbnl1/2-dependent exons
abc
CNS.Inc − 94%
CNS.Exc − 85%
Muscle.Inc − 91%
Muscle.Exc − 86%
EM.Inc − 82%
EM.Exc − 89%
Digestive.Inc − 87%
Digestive.Exc − 88%
All Tiss. Dep. − 89%
Figure 2 Prediction accuracy. (a) Differentiating alternative (n = 11,773) from constitutive (n = 9,638) exons. Detecting which exons are
alternative (green) is significantly improved compared to a classifier that uses exon expression measurements from 33 experiments (cyan), and
compared to the original classifier trained to detect only tissue-dependent cassette exons (red). Detection of exons that exhibit tissue-dependent
splicing changes (blue, n = 659) is much more accurate. Numbers within each legend represent the area under the curve (AUC) (b) Identifying
tissue-dependent splicing. Detecting tissue-dependent splicing changes (n = 865) from a random set of non-tissue-dependent exons (n = 4,000)
achieves an overall accuracy of 89% AUC (black). Accuracy varies considerably between tissues and for detecting increased inclusion (solid line) or
exclusion (dashed) in a tissue (c) Detection accuracy for an independent set of Mbnl1/2-dependent exons [14] (n = 461). Differentiating between
Mbnl1/2-dependent exons and constitutive exons achieves 97% AUC. Accuracy in detecting Mbnl1/2-dependent exons from a random set of
non-tissue-dependent exons (n = 2,000) is approximately 94% AUC for both brain (blue) and muscle (red).
Barash et al. Genome Biology 2013, 14:R114 Page 3 of 8
http://genomebiology.com/2013/14/10/R114
conserved pattern of alternative splicing that changes
across tissues and developmental stages [17,18]. Its role
in angiogenesis, which is controlled in part by alternative
splicing, has made it an attractive target of several
anticancer therapies. Accordingly, there is considerable
interest in identifying the factors that regulate the spli-
cing of Vegfa transcripts [18,19]. Analyzing all Vegfa
exon triplets revealed that only exons 6 and 7 were pre-
dicted to be cassette exons, with a score corresponding
to a false positive rate of 0.009 and 0.017, respectively.
For comparison, other exons scores corresponded to a
false positive rate of 0.22 or higher (data not shown).
These prediction s are in line with annotated transcripts,
many of which skip exon 6, one that skips exon 7
(ENSMUST00000113519), and several that skip both.
Exons 6 and 7 were also both predicted, with a fal se
positive rate of less than 0.025, to exhibit differential
splicing in all four major tissue groups modeled. While
confidence in differential splicing was high, the predic-
tions were not conclusive as to whether a relative in-
crease or decrease of exon inclusion would occur in the
tissues. These results reflect the conserved and complex
splicing pattern of Vegfa, with RT-PCR experiments
showing exon 6 to have a complex bi-phasic increase of
inclusion in developing mouse and chicken heart [18].
Prediction of other splice variations of Vegfa, such as the
3’ splice site variation in exon 8, are currently not sup-
ported by the tool.
Figure 3 shows the regulatory feature analysis for
differential inclusion of Vegfa exon 6 in muscle. The
enrichment analysis in Figure 3a highlights that
the alternative exon is depleted of non-tissue-specific
exonic splicing enhancers and is highly enriched with
exonic splic ing silencers. Other highlighted features are
enriched secondary structure-free regions in the up-
stream intron, a distant first AG nucleotide upstream
and a particularly short preceding exon 5. The preceding
exon, for example, is 32 bp long, and the enrichment
analysis indicates that only 0.127% of the tool’s reference
set of alternative exons has a shorter preceding exon.
The most dominant effect of in silico motif removal
(Figure 3b) is for CU-rich elements known to bind Ptb1/
b
a
Figure 3 Analysis of Vegfa exon 6 muscle-dependent inclusion. A subset of the summary page produced by AVISPA is shown. (a) Feature
enrichment analysis: the values of the features listed on the left are computed for Vegfa exon 6 and compared against matching feature values in
a set of labeled exons. The four sets of exons compared against here are alternative exons ('AS', third column from the left), constitutive exons
('Const', third column from the right), exons differentially included in muscle ('Muscle Inc', second column from the right), and differentially
excluded in muscle ('Muscle Exc', right most column). Relative enrichment or depletion of features is indicated using the heat map on the right.
Only features with significantly low (blue) and high (red) values are shown here. The genomic region of each feature is indicated by the second
from left column using the notation and colors in the top figure. (b) Stacked bar chart (left) of the normalized feature effect (NFE, y-axis) on
splicing prediction. Only the top motifs are shown. Motif regions are annotated using the color scheme depicted below. Mapping of the motifs
onto the UCSC genome browser is shown on the right. Tracks combining all motifs used by the code model (red), the unbiased motif search [5]
(grey scaled), and conservation (blue) are added at the bottom.
Barash et al. Genome Biology 2013, 14:R114 Page 4 of 8
http://genomebiology.com/2013/14/10/R114
2, followed by an ACUAAY motif known to bind Quak-
ing (Qk). These splice factors have not been previously
reported to regulate Vegfa, but a recent study estimates
39% of regulated exons during myogenesis are under the
control of one or both of these splicing factors [20]. A
smaller effect on splicing prediction in muscle is associ-
ated with intronic motifs known to bind Cugbp1/2 and
Muscleblind-like protein (Mbnl1/2). Both Cugbp1/2 and
Mbnl1/2 have been shown to play an important role in
regulating splicing in developing hearts. Overexpressing
Cugbp1 or knockdown of Mbnl1 in the adult mouse
heart did not alter exon 6 inclusion levels significantly
[18], but recent results point to possible compensatory
effects between Mbnl1 and Mbnl2 [14]. Othe r elements
implicated in Vegfa splicing regulation include the short
YCAY motifs known to bind Nova proteins [21] and a
UGCAUG motif, known to bind the brain- and muscle-
specific splicing factor Fox-1 (A2bp 1) and its paralog
Fox-2 (Rbm9) [22]. While the Fox-1/2 binding site is
highly conserved, it resides over 1 kb downstream of
exon 6 and Fox-1/2 have not been pre viously reported to
regulate Vegfa. However, recent results indicate that
Fox-2 knockdown in mice clearly alters Vegfa splicing
pattern during heart development (Xiang-Dong Fu, per-
sonal communication). Smaller effects associated with
non-tissue-specific regulation include G-rich elements,
known to bind hnRNP-F/H, and U-rich elements that
are known to bind hnRNP-C and Tiar/Tia1 [23]. Not-
ably, Tia1 was previously reported to regulate Vegfa iso-
form expression [24]. Overall, our exploratory analysis of
Vegfa splicing is consistent with previou s results and
offers new insights into mechanism of Vegfa regulation
that are supported by recent experiments.
In summary, we presented a new tool, AVISPA, for in
silico prediction and analysis of alternative splicing. The
tool is not limit ed by technical constraints such as se-
quencing depth, and its predictions for alternatively
spliced exons generalize over unmeasured conditions.
Beyond the splicing outcom e, it offers researchers the
ability to identify putative regulatory elements and map
those to the genome. These capabilities were re cently
used in an independent study to identify TIA1 as a regu-
lator of an alternative exon coding miR-412 [25]. Here,
we used a recent genome-wide study to demon strate the
tool’s accuracy for predicting muscl e, heart, and brain
regulated exons and performed detailed in silico splicing
analysis for the vascular endothelial growth factor A.
Several important elements remain as on-going and
future enhancements of the tool. These include predic-
tions for species other than mouse, predictions for
additional forms of alternative splicing (for example,
alternative 3’ and 5
’ splice sites), and higher resolution of
tissue spe cificity. Currently, AVISPA’s predictions reflect
confidence in alternative splicing or in relative, tissue-
dependent, inclusion changes. Thus, users may infer an
exon is likely to be alternative or to be differentially in-
cluded in brain versus other tissues, but predictions for
absolute inclusion levels (for example, 20% inclusion in
brain, 40% inclusion in liver) are currently not sup-
ported. The tool has some technical limitations as well.
Users can only submit a single cassette exon as a query,
due to the computational burden involved in processing
a query. Queries must be based on annotated exons,
cannot contain exons shorter than 10 bases long, and
non-canonical splicing by the minor spliceosome is
not supported. Nonetheless, the ability to perform
splicing prediction irrespective of experimental limita-
tions , coupled with the new regulatory element s analysis,
should serve researchers studying gene regulation, RNA
biogenesis, and development. Moreover, AVISPA is built
as a flexible platform that can be repeatedly updated as
more data and improved models be come available. The
new computational analysis offered by AVISPA should
facilitate the discovery of novel splicing variant s, regula-
tory elements, and genomic variations affecting pheno-
typic variability or disease.
Materials and methods
Query matching against sequence database
The web-tool’s internal database includes three compo-
nents. The first is a database of 11,773 cassette exons that
we previously mined from sequence libraries [5]. The sec-
ond is a set of 9,638 exon triplets derived from Refseq [26]
and other sequence libraries as described in [5], where
every three constitutive exons in a transcript define a trip-
let. These triplets were also scanned against exon expres-
sion measurements in 11 mouse tissues [12] and triplets
suspected to contain an alternative cassette exon were re-
moved. A query’s sequence is matched against the two
transcript databases using BLAT with parameters set to
tileSize = 8, minMatch = 2, minIdentity = 88. The third
database component is the mouse assembly mm10 from
the UCSC Genome Browser [27]. Matching a query to the
reference genome is executed only if no match in the two
transcript-based databases is found, and only when gen-
omic coordinates for all three exons are specified.
Extended regulatory feature set
We extended the set of putative regulatory features to
include the occurrences of 350 new binding motifs in
the seven regions around a cassette exon as defined in
[5]. The motifs correspond to general splicing related
RNA binding proteins (RBPs), SR and SR-related pro-
teins (SC35, SRp20, 9G8, ASF/SF2, SRp30c, SRp38,
SRp40, SRp55, SRp75, Tra2α/β), and hnRNP proteins
(hnRNPA1, hnRNPA2/B1, hnRNPF/H, hnRNPG).
We also added features encoding computationally pre-
dicted nucleosome occupancy around the alternative
Barash et al. Genome Biology 2013, 14:R114 Page 5 of 8
http://genomebiology.com/2013/14/10/R114
exon [13]. Features were defined as the average and
maximal occupancy scores in the first 100 nucleotides in
each intron and the first or last 50 nucleotides of the
alternative exon.
Extended training set for tissue-specific alternative
splicing
A total of 33 data tracks for normalized expression mea-
surements using Affymetrix exon arrays were down-
loaded from the UCSC Genome Browser. The tracks are
composed of measurements in 11 mouse tissues (brain,
embryo, heart, kidney, liver, lung, muscle, ovary, spleen,
testis, thymus) with three replicates for each tissue [12].
The expression of each exon and the relative inclusion
of a putative cassette exon compared to its flanking
exons were used as input features to train an ensemble
of Bayesian neural networks [11]. The networks used
these input features to identify differential inclusion and
exclusion of alternative exons in the four tissue groups
previously identified (CNS, muscle, digestive, embryo).
Training was based on a subset of 3,770 cassette exons
for which three probabilities for increased inclusion
(q
inc
), increased exclusion (q
exc
) and no change (q
nc
)in
each of the four tissue groups was previously computed
[5]. This training step allowed the calibration of differen-
tial splicing estimation obtained from the new set of 33
experiments to the estimates used to train the original
splicing model [5]. The model ensemble was then used
to estimate differential splic ing (q
inc
,q
exc
,q
nc
) for the
remaining exons. The differential splicing estimates for
the original set of 3,770 exons were averaged between
the two datasets and care was taken to make sure pre-
dictions were based on non-overlapping training sets.
Predicting alternative cassette exons using expression
data and a single stage tissue-specific classifier
The33expressiondatatracksdescribedabovewerealso
used to train a Bayesian deep neural network classifier [11],
denoted '33 exon arrays'inFigure2a.Anyexontriplets
from the set of 11,773 cassette exons and 9,638 putative
constitutive exons that had missing data were removed,
maintaining a total of 8,986 for training and test purposes.
The prediction of alternative exons using a single stage
tissue classifier, denoted Pr[Ts] in Figure 2a, used a max
function over the chance of differential splicing (1 - p
nc
)
in each tissue.
Training a splicing code model for alternative exons and
for tissue-dependent splicing
For the purpose of inferring a regulatory model, we used
a Bayesian neural network that worked better for this
task than support vector machines, boosted decision
trees , and other leading machine learning techniques
[11]. To discriminate between alternative and constitutive
exons the network was set to have 10 hidden units and a
sparsity prior of 0.9 for connections between features and
hidden units. For predicting tissue-dependent splicing the
network was set to have 20 units and a sparsity prior of
0.95. Varying the sparsity prior between 0.85 and 0.95 and
adding up to 10 more hidden units did not have a signifi-
cant effect on the results (data not shown). An ensemble
of 5,000 models generated by Markov chain Monte Carlo
simulations was used to estimate differential splicing
(q
inc
,q
exc
,q
nc
) as was previously described [11].
Scoring tissue-dependent splicing
Under the new framework the probability that any given
triplet of exons contain a tissue-dependent cassette exon
can be expressed as:
PO
t
¼ ch r
e
Þ¼PASr
e
ÞPO
t
¼ ch r
e
; ASÞ;jðjðjð
where P(O
t
= ch|r
e
) denotes the probability to observe a
change in the exon’s inclusion level in tissue t given the
exon’s feature vector r
e
, P(AS|r
e
) is the probability the
exon is alternative, and P(O
t
= ch|r
e
, AS) is the probabil-
ity of observin g differential splicing given that the exon
is alternative. The first term on the right is computed by
the first stage predictor, while the second term is com-
puted by the second stage predictor.
ROC performance evaluation
Receiver operating characteristic (ROC) performance
was evaluated using repeated five-fold cross-validation
and care was taken to make sure predictions were based
on non-redundant training sets, as was previously de-
scribed [5]. Evaluation of discriminating between alter-
native and constitutive exons was based on a set of
11,773 cassette exons and 9,638 putative constitutive
exons derived from EST/cDNA sequences [5]. In order
to assess the accuracy of detecting cassette exons that
exhibit a tissue-dependent splicing pattern ( for example,
differential inclusion in muscle) we compared the scores
of such exons to those of a random set of exon triplets
that do not exhibit this splicing pattern. The random set
was selected using the following procedure. First, we
used the 33 genome-wide exon expression measure-
ments described above to quantify the inclusion level of
all exon triplets from all Refseq transcripts. Next, we dis-
carded triplets with missing data and required the rela-
tive expression of the upstream and downstream exons
to be no more than 1.5-fold apart in all experiments. In
order to avoid probe sets with little signal, we required
the up- and downstream exons to have a normalized ab-
solute value of at least 0.1 in at least 15 experiments.
Additionally, we required in at least three experiments
of the tissue group of interest (for e xample, digestive)
that the up- and downstream exons are not in the
Barash et al. Genome Biology 2013, 14:R114 Page 6 of 8
http://genomebiology.com/2013/14/10/R114
bottom 20 percentile. Finally, the relative expression of
each middle exon compared to its flanking exons was used
to estimate the chance it is differentially included in each
tissue group [28]. Any triplet that had a P-value of 0.7 or
higher was deemed non-tissue-dependent and a set of ap-
proximately 2,000 exons was then selected for each tissue
as a non-tissue-dependent exon set. Exons were selected
randomly from the respective genes and then randomly
from the relative order within the gene. We then verified
that these are not biased in terms of relative location
within the gene or gene length compared to a random
sample of triplets from the genome (data not shown).
While small variations in the parameters of the above
process did not have a notable effect on the results, we did
detect an apparent selection bias in this procedure. Specif-
ically, using expression measurements to select exons
based on high confidence in non-tissue-dependent spli-
cing may favor constitutive exons. Notably, the 'true' labels
of any given exon as alternative or constitutive is unavail-
able. However, since our prediction algorithm has proved
accurate in distinguishing alternative from constitutive
exons (Figure 2a), we applied it to the set of 2,000 non-
tissue-dependent exons selected for each tissue group.
Compared to a random set of 1,000 exon triplets, these
exons were biased towards constitutive exon scores
(Additional file 1). To correct for this apparent bias we
subsampled 1,000 exons for each tissue group so that their
scores as alternative match those in the random set
(Additional file 1, green and red lines). This corrected set
of a total of 4,000 predictions was then used for subsequent
analysis (Figure 2b,c). We note that without this correction
the initial set of non-tissue-dependent exons results in im-
proved performance compared to that shown in Figure 2.
In silico feature removal and normalized feature effect
In order to evaluate the relative effect of a putative regu-
latory sequence motif (for example, the occurrence of a
[U]GCAUG motif, known to bind Fox1/2, upstream of
the alternative exon), the feature is first set to zero. The
splicing predictions with the mutated feature, denoted
p
inc
Δf
; p
exc
Δf
; are then computed with the total effect on
differential splicing defined a s FE
f
¼jp
inc
p
inc
Δ
jþjp
exc
p
exc
Δ
j. This definition aims to capture the effect of features
that not only change the confidence in a splicing change
p
nc
; p
nc
Δf
; but also change the relative confidence in
either differential inclusion or exclusion. Finally, the nor-
malized feature effect (NFE) is defined as:
NFE
f
¼
FE
f
∑
j∈J
FE
j
where J is the set of robust features. By itself, the NFE has
no statistical significance measure associated with it. The
NFE serves mainly as a quantitative tool to guide re-
searchers interested in knowing which of the identified
regulatory features have a higher effect on the model’spre-
diction confidence.
Additional file
Additional file 1: Figure S1. Correcting constitutive exons selection
bias in non-tissue-dependent exons. Exon scores for being alternative
versus constitutive (x-axis) are plotted as a cumulative distribution func-
tion (CDF, y-axis). The initial set of selected non-tissue-dependent exons
(blue) was biased towards constitutive exons compared to a random
sample of 1,000 exon triplets from the genome (red). Subsampling the
original set of 2,000 exons per tissue to fit the score distribution of a
random set gave a good fit (green). Both green and red line plots are
accumulated over all exons in all tissues as no significant difference was
observed between the different tissues.
Abbreviations
AS: Alternative splicing; AUC: Area under the curve; AVISPA: Advanced
visualization of splicing prediction and analysis; CNS: Central nervous system;
EST: Expressed sequence tag; NFE: Normalized feature effect.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
YB and BJF conceived of the project. YB developed the combined prediction
framework and in silico feature analysis. YB, JVG and WG developed the
analysis pipeline with input from all authors. YB, JVG, WG and LJL created
the sequence databases. JGV, WG and JVG developed the web tool. HYX, YB
and BJF developed the prediction algorithms. YB and JVG performed the
data analysis. YB wrote the paper with input from BJF. All authors read and
approved the final manuscript.
Acknowledgements
We thank Ben Blencowe for his support and advice throughout this project. We
thank Kristen Lynch for helpful feedback and discussions; Xiang-Dong Fu for
sharing experimental results; members of the Barash, Lynch, Blencowe and Frey
labs for providing helpful comments on the manuscript and suggestions for the
web tool. Funding to JGV through FPI grant from the Spanish Ministry of Sci-
ence; HYX, WG and LJL were funded from NSERC Steacie and CIHR grants to
BJF. While at the University of Toronto, YB was funded from an OGI Spark grant
and a Genome Canada grant to BJF and others.
Author details
1
Department of Genetics, University of Pennsylvania, Philadelphia, PA 19104,
USA.
2
Department of Computer and Information Science, University of
Pennsylvania, Philadelphia, PA 19104, USA.
3
Universitat Pompeu Fabra,
Barcelona 08003, Spain.
4
Department of Electrical and Computer
Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada.
5
Banting
and Best Department of Medical Research, University of Toronto, Toronto,
ON M5G 1L6, Canada.
Received: 9 July 2013 Accepted: 11 October 2013
Published: 24 October 2013
References
1. 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, 40:1413–1415.
2. Wang ET, Sandberg R, Luo S, Khrebtukova I, Zhang L, Mayr C, Kingsmore SF,
Schroth GP, Burge CB: Alternative isoform regulation in human tissue
transcriptomes. Nature 2008, 456:470–476.
3. Wang ET, Cooper AT: Splicing in disease: disruption of the splicing code
and the decoding machinery. Nature 2007, 8:749–761.
4. Wang Z, Burge CB: Splicing regulation: from a parts list of regulatory
elements to an integrated splicing code. RNA 2008, 14:802–813.
Barash et al. Genome Biology 2013, 14:R114 Page 7 of 8
http://genomebiology.com/2013/14/10/R114
5. Barash Y, Calarco JA, Gao W, Pan Q, Wang X, Shai O, Blencowe BJ, Frey BJ:
Deciphering the splicing code. Nature 2010, 465:53–59.
6. Barbosa-Morais NL, Irimia M, Pan Q, Xiong HY, Gueroussov S, Lee LJ,
Slobodeniuc V, Kutter C, Watt S, Çolak R, Kim T, Misquitta-Ali CM, Wilson
MD, Kim PM, Odom DT, Frey BJ, Blencowe BJ: The evolutionary landscape
of alternative splicing in vertebrate species. Science 2012, 338:1587–1593.
7. Yeo G, Burge CB: Maximum entropy modeling of short sequence motifs
with applications to RNA splicing signals. J Comput Biol 2004, 11:377–394.
8. Dogan RI, Getoor L, Wilbur WJ, Mount SM: SplicePort–an interactive splice-
site analysis tool. Nucleic Acids Res 2007, 35:W285–W291.
9. Cartegni L, Wang J, Zhu Z, Zhang MQ, Krainer AR: ESEfinder: a web
resource to identify exonic splicing enhancers. Nucleic Acids Res 2003,
31:3568–3571.
10. Yeo GW, Coufal NG, Liang TY, Peng GE, Fu X-D, Gage FH: An RNA code for
the FOX2 splicing regulator revealed by mapping RNA-protein interac-
tions in stem cells. Nat Struct Mol Biol 2009, 16:130–137.
11. Xiong HY, Barash Y, Frey BJ: Bayesian prediction of tissue-regulated
splicing using RNA sequence and cellular context. Bioinformatics 2011,
27:2554–2562.
12. Pohl AA, Sugnet CW, Clark TA, Smith K, Fujita PA, Cline MS: Affy exon
tissues: exon levels in normal tissues in human, mouse and rat.
Bioinformatics 2009, 25:2442–2443.
13. Xi L, Fondufe-Mittendorf Y, Xia L, Flatow J, Widom J, Wang JP: Predicting
nucleosome positioning using a duration Hidden Markov Model.
BMC Bioinformatics 2010, 11:346.
14. Wang ET, Cody NAL, Jog S, Biancolella M, Wang TT, Treacy DJ, Luo S,
Schroth GP, Housman DE, Reddy S, Lécuyer E, Burge CB: Transcriptome-
wide regulation of pre-mRNA splicing and mRNA localization by muscle-
blind proteins. Cell 2012, 150:710–724.
15. Schwartz S, Meshorer E, Ast G: Chromatin organization marks exon-intron
structure. Nat Struct Mol Biol 2009, 16:990
–995.
16. Tilgner H, Nikolaou C, Althammer S, Sammeth M, Beato M, Valcárcel J,
Guigó R: Nucleosome positioning as a determinant of exon recognition.
Nat Struct Mol Biol 2009, 16:996–1001.
17. Harper SJ, Bates DO: VEGF-A splicing: the key to anti-angiogenic thera-
peutics? Nat Rev Cancer 2008, 8:880–887.
18. Kalsotra A, Xiao X, Ward AJ, Castle JC, Johnson JM, Burge CB, Cooper TA:
A postnatal switch of CELF and MBNL proteins reprograms alternative
splicing in the developing heart. Proc Natl Acad Sci USA 2008,
105:20333–20338.
19. Nowak DG, Amin EM, Rennel ES, Hoareau-Aveilla C, Gammons M, Damo-
doran G, Hagiwara M, Harper SJ, Woolard J, Ladomery MR, Bates DO:
Regulation of Vascular Endothelial Growth Factor (VEGF) Splicing from
Pro-angiogenic to Anti-angiogenic Isoforms: a Novel Therapeutic
Strategy for Angiogenesis. J Biol Chem 2009, 285:5532–5540.
20. Hall MP, Nagel RJ, Fagg WS, Shiue L, Cline MS, Perriman RJ, Donohue JP,
Ares M: Quaking and PTB control overlapping splicing regulatory
networks during muscle cell differentiation. RNA 2013, 19:627–638.
21. Ule J, Stefani G, Mele A, Ruggiu M, Wang X, Taneri B, Gaasterland T,
Blencowe BJ, Darnell RB: An RNA map predicting Nova-dependent
splicing regulation. Nature 2006, 444:580–586.
22. Kawamoto S: Neuron-specific alternative splicing of nonmuscle myosin II
heavy chain-B pre-mRNA requires a cis-acting intron sequence. J Biol
Chem 1996, 271:17613–17616.
23. Aznarez I, Barash Y, Shai O, He D, Zielenski J, Tsui LC, Parkinson J, Frey BJ,
Rommens JM, Blencowe BJ: A systematic analysis of intronic sequences
downstream of 5’ splice sites reveals a widespread role for U-rich motifs
and TIA1/TIAL1 proteins in alternative splicing regulation. Genome Res
2008, 18:1247 –1258.
24. Chen M, Manley JL: Mechanisms of alternative splicing regulation:
insights from molecular and genomics approaches. Nat Rev Mol Cell Biol
2009, 10:741–754.
25. Melamed Z, Levy A, Ashwal-Fluss R, Lev-Maor G, Mekahel K, Atias N, Gilad S,
Sharan R, Levy C, Kadener S, Ast G: Alternative Splicing Regulates Biogen-
esis of miRNAs Located across Exon-Intron Junctions. Mol Cell 2013,
50:869–
881.
26. Pruitt KD, Tatusova T, Brown GR, Maglott DR: NCBI Reference Sequences
(RefSeq): current status, new features and genome annotation policy.
Nucleic Acids Res 2012, 40:D130–D135.
27. Dreszer TR, Karolchik D, Zweig AS, Hinrichs AS, Raney BJ, Kuhn RM, Meyer
LR, Wong M, Sloan CA, Rosenbloom KR: The UCSC Genome Browser
database: extensions and updates 2011. Nucleic Acids Res 2012,
40:D918–D923.
28. Ben-Dor A, Friedman N, Yakhini Z: Scoring Genes for Relevance. Agilent; 2000.
doi:10.1186/gb-2013-14-10-r114
Cite this article as: Barash et al.: AVISPA: a web tool for the prediction
and analysis of alternative splicing. Genome Biology 2013 14:R114.
Submit your next manuscript to BioMed Central
and take full advantage of:
• Convenient online submission
• Thorough peer review
• No space constraints or color figure charges
• Immediate publication on acceptance
• Inclusion in PubMed, CAS, Scopus and Google Scholar
• Research which is freely available for redistribution
Submit your manuscript at
www.biomedcentral.com/submit
Barash et al. Genome Biology 2013, 14:R114 Page 8 of 8
http://genomebiology.com/2013/14/10/R114
