16 posters,  50 authors,  15 institutions
ePostersLive® by SciGen® Technologies S.A. All rights reserved.

P07
DeepExon: Deep Learning Model for Recognition of Acceptor Splice Site

Rate

No votes yet

During the RNA processing, the intron regions are spliced
out, and the remaining exon regions are connected to form
mature RNA.
• To correctly annotate genes contained in a DNA sequence,
one also has to correctly annotate boundaries between
exons and introns in case a gene contains introns.
• These boundaries are demarcated by the splice sites (SS),
namely, donors and acceptors.
• Donors and acceptor sites are characterized in almost all
cases by canonical GT and AG dinucleotides in DNA.
Motivation
• Computational recognition of SS is attractive as laboratory
experiments are not required, but it suffers from a large
number of false positive predictions.
• Therefore, deriving accurate models to predict real SS is
challenging as most of the GT/AG dinucleotides present
in a DNA sequence do not have roles of SS. The current
state-of-the-art methods to computationally solve the
problem of large number of false positive predictions
remain ineffective.

During the RNA processing, the intron regions are spliced
out, and the remaining exon regions are connected to form
mature RNA.
• To correctly annotate genes contained in a DNA sequence,
one also has to correctly annotate boundaries between
exons and introns in case a gene contains introns.
• These boundaries are demarcated by the splice sites (SS),
namely, donors and acceptors.
• Donors and acceptor sites are characterized in almost all
cases by canonical GT and AG dinucleotides in DNA.
Motivation
• Computational recognition of SS is attractive as laboratory
experiments are not required, but it suffers from a large
number of false positive predictions.
• Therefore, deriving accurate models to predict real SS is
challenging as most of the GT/AG dinucleotides present
in a DNA sequence do not have roles of SS. The current
state-of-the-art methods to computationally solve the
problem of large number of false positive predictions
remain ineffective.

The Convolutional layer is the core building block of a
CNN; it consists of a set of K filters (kernels), where each
filter has a width and a height to automatically capture a
special pattern in the sequence.
• Then, we applied a Max pooling layer which identifies
regions of our input with a high response to a particular
filter aiming to obtain a local invariance and also to reduce
the size of the input volume.
• The last layer is fully connected NN followed by a softmax
classifier, which computes the final output prediction
scores. For generalization, Dropout layer is added to avoid
overfitting.

As our proposed model "DeepExon" structure allows our network
to learn richer features as signal progresses through the
layers, it significantly improves the accuracy of acceptor SS prediction
over the previous state-of-the-art models, achieving 99%
for both sensitivity and specificity for recognizing acceptor SS
in DNA sequence in the testing set. The model is implemented
to run on GPU.

Enter Poster ID (e.gGoNextPreviousCurrent