% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{Sitani:989346,
author = {Sitani, Divya},
othercontributors = {Zimmer-Bensch, Geraldine Marion and Carloni, Paolo},
title = {{M}achine learning methods for prediction of
protein-protein interaction hot spot residues},
school = {RWTH Aachen University},
type = {Dissertation},
address = {Aachen},
publisher = {RWTH Aachen University},
reportid = {RWTH-2024-06740},
pages = {1 Online-Ressource : Illustrationen},
year = {2024},
note = {Veröffentlicht auf dem Publikationsserver der RWTH Aachen
University; Dissertation, RWTH Aachen University, 2024},
abstract = {Protein–protein interactions (PPIs) form a vast and
intricate network of reactions important for the regulation
and execution of most biological processes [Rao+14]. PPIs
occur when two proteins make direct physical contact via
their surface residues and form an interface, which is a
non-uniform surface on a protein-protein complex [GS10].
Even though a protein interface may occupy a large area,
only a small subset of its buried residues plays a crucial
role in the binding free energy of the complex [BT98;
Jan95]. These energetically key residues are known as hot
spots. The experimental method to identify them is Alanine
Scanning Mutagenesis (ASM) where systematically each
interface residue is mutated to Alanine and the consequent
change in binding energy ΔΔGbinding between the wild type
and the mutant complex is measured. If (ΔΔGbinding) is
larger than a certain threshold, typically 2 kcal/mol, the
interface residue is defined as a hot spot or else it is
considered a null spot [MFR07; CW89; BT98]. The so-called
hot spot residues are often enriched in disease-associated
mutations [Ten+09]. These mutations often cause disrupted or
erroneous protein interactions, resulting in phenotypic
changes that might cause a disease. Moreover, with the
discovery of hot spots in protein-protein interfaces, it has
become possible to target a broader range of PPIs with small
molecule drugs. The identification of hot spots has helped
researchers to identify molecules that interact at these
sites, thus interfering with PPIs and the downstream
pathways they mediate [Pet+16a; Pet+16b; Sco+16]. Therefore,
predicting hot spots is crucial to understand the effect of
disease-associated mutations on PPIs and for drug discovery
[Mur+17]. As mentioned before, experimentally hot spots can
be found out by using ASM, but it is quite costly and
tedious and this has led to the use of computational methods
to predict hot spot residues. Previous computational
approaches included molecular dynamics and knowledge-based
methods [GNS02; KB02; MK99; HMK02; GF08; Bre+09]. However,
such approaches were time-consuming and hence limited in the
number of hot spots predicted. This led to an increased use
of machine learning (ML) based methods for hot spot
prediction in recent years [DPM07; Den+13; CKL09a; CKL09b;
Ass+10]. Such ML approaches capitalize on the availability
of experimental datasets containing protein-protein complex
structures and ASM-derived hotspot data. However, as it
often happens with biological data repositories, such
hotspot datasets often contain noise [Mor+17; KC21]. If
machine learning (ML) algorithms are trained and predictions
are made on this "noisy" data, the results will not be
accurate [GG19]. The earlier ML-based approaches for hot
spot prediction did not take this issue into account. In
this thesis, I describe the basic concepts and recent
advances of machine learning applications in finding the
protein–protein interaction hot spots. To reduce the
effects of noise in hot spot prediction, I have proposed the
method RBHS (Robust Principal Component Analysis-(RPCA)
based Prediction of Protein-Protein Interaction Hot Spots)
in this thesis [Sit+21]. I use RPCA [Can+11] followed by
feature selection using Extreme Gradient Boosting (XGBoost)
[CG16] on the data matrix containing protein sequence and
structure-based features calculated on the interface
residues. I trained several popular machine learning
classifiers on the benchmark dataset HB-34 [LLD18] and
evaluated the performance of my proposed method on the
independent test set BID-18 [LLD18]. After extensive
computational experimentation and comparison with the
existing state-of-the-art approaches to predict hot spots, I
was able to show that my method is quite efficient in
identifying hot spot residues crucial for protein-protein
interactions. Finally, I discuss the challenges and future
directions in the prediction of hot spots in this thesis.},
cin = {164620 / 160000},
ddc = {570},
cid = {$I:(DE-82)164620_20181217$ / $I:(DE-82)160000_20140620$},
typ = {PUB:(DE-HGF)11},
doi = {10.18154/RWTH-2024-06740},
url = {https://publications.rwth-aachen.de/record/989346},
}
Gast ::