Abstract:
Species evolve by adapting to variable thermal conditions. The differences in the thermal
stability of hyperthermophilic, thermophilic and mesophilic proteins arise partly due to their
structural variations. The goal of this thesis is to identify structural features responsible
for these variations using machines learning techniques that use features derived from the
residue interaction graphs (RIG) and the amino acid sequences of proteins. For the RIG
model, we studied the features linked to thermal stability which capture different notions of
centrality, connection strength, weighted clustering coefficient and such. We evaluated them
against a few features that were hitherto not studied in the context of thermal classification
and demonstrated that the new features can significantly improve classification accuracy.
We further improved the performance by using a histogram of centrality values as a feature
vector instead of using a single statistic such as mean that has been the trend so far among
researchers. We discovered that the histograms corresponding to edge betweenness centrality,
current flow closeness centrality and 2-hop degree centrality lead to the best classification
accuracy among the network-based features. We also investigated the state-of-the-art features
based on amino acid sequences and proposed a new one using the amino acid tri-mers of
a protein. For empirical evaluation, we investigated a set of 842 hyperthermophilic, 533
thermophilic and 2248 mesophilic proteins and compared our proposed features with the
state-of-the-art features using commonly known classifiers such as SVM, ANN and random
forest. We obtained an overall accuracy greater than 90% which is significantly better than
what has been reported so far.
