Machine Learning (ML) is a field of artificial intelligence focused on the design and
implementation of algorithms that enable creation of models for clustering, classification,
prediction, ranking and similar inference tasks based on information contained in data.
Many ML algorithms have been successfully utilized in a variety of applications. The
problem addressed in this thesis is from the field of bioinformatics and deals with the
recognition of polyadenylation (poly(A)) sites in the genomic sequence of the plant
Arabidopsis thaliana. During the RNA processing, a tail consisting of a number of
consecutive adenine (A) nucleotides is added to the terminal nucleotide of the 3’-
untranslated region (3’UTR) of the primary RNA. The process in which these A
nucleotides are added is called polyadenylation. The location in the genomic DNA
sequence that corresponds to the start of terminal A nucleotides (i.e. to the end of 3’UTR)
is known as a poly(A) site. Recognition of the poly(A) sites in DNA sequence is
important for better gene annotation and understanding of gene regulation. In this study,
we built an artificial neural network (ANN) for the recognition of poly(A) sites in the
Arabidopsis thaliana genome. Our study demonstrates that this model achieves improved
accuracy compared to the existing predictive models for this purpose. The key factor
contributing to the enhanced predictive performance of our ANN model is a
distinguishing set of features used in creation of the model. These features include a
number of physico-chemical characteristics of relevance, such as dinucleotide
thermodynamic characteristics, electron-ion interaction potential, etc., but also many of the statistical properties of the DNA sequences from the region surrounding poly(A) site,
such as nucleotide and polynucleotide properties, common motifs, etc. Our ANN model
was compared in performance with several other ML models, as well as with the PAC
tool that is specifically developed for poly(A) site recognition in Arabidopsis thaliana
and rice. The comparison analysis shows that our model performs better compared to the
others available, and achieves on average 93% accuracy.
| Date of Award | Sep 2011 |
|---|
| Original language | English (US) |
|---|
| Awarding Institution | - Computer, Electrical and Mathematical Sciences and Engineering
|
|---|
| Supervisor | Vladimir Bajic (Supervisor) |
|---|