In this masters thesis, we tackled the problem of classification of Tobamovirus
genomic sequences. Tobamovirus is a plant RNA-virus that mutates quickly
and can cause significant agronomical damage. The goal of our thesis was
to improve the speed and ease of use of an existing classification system
while also improving the classification accuracy. We managed to develop
several containers and processes for training of deep genomic models, using
a high-performance computer cluster. We thoroughly evaluated two models,
DNABERT-2 and DNABERT-S. Despite good results during training,
DNABERT-2 did not gain a good generalization for the classification of Tobamovirus
genome sequences. Performance on the test data was low, the
most successful DNABERT-2 model achieved a ROC AUC value of 0.58. As
such, it is not suitable for further research. DNABERT-S, which was developed
for species classification, achieved better results. The most successful
DNABERT-S model achieved a ROC AUC value of 0.77. Although the performance
of DNABERT-S is acceptable as a start, further development and
research are needed to obtain significant improvements. The DNABERT-S
model generated unstable and unpredictable results, which prevents us from
declaring it a good replacement for the current classification system.
|
