dc.description.abstract | As more and more genomes are sequences it becomes ever more needed to have access to fast
methods of analyzing them. One such analysis is finding regions of the genome that contain
descriptions of proteins we are familiar with. Pfam is a database containing these descriptions,
and tools like HMMER can match known proteins/protein substructures onto sequences. This
process is slow, and this thesis explores how to speed it up by creating machine-learning models
that filter out the sequences less likely to contain known proteins.
This is done by analyzing large datasets containing hundreds of species sampled by different
means and by employing different target values. Through this valuable insight such as high
GC content negatively impacting model performance and that machine learning models such
as convolutional neural networks can describe specific Pfam patterns close to perfectly, with
accuracies of 99.7% on balanced datasets.
The end result of the thesis is that a filtering mechanism should be possible to create, but it
would require significantly more work to get it working to a degree where it both included the
necessary. Ensuring that enough Pfam entries are represented would be a good starting step.
One of the models this thesis employs is the Tsetlin machine, hoping its unique structure would
make it well suited to such data. The results did, however, not show this to be the case. While
the Tsetlin machine might still be well suited for similar data, some changes would have to be
made to how they were employed in this thesis. | |