Brain-derived exosomes as means to examine molecular mechanisms that affect cellular homeostasis. An informatics approach.

Postgraduate Thesis uoadl:3413077 18 Read counter

Unit:
Κατεύθυνση Βιοπληροφορική-Επιστήμη Βιοϊατρικών Δεδομένων
Πληροφορική
Deposit date:
2024-07-31
Year:
2024
Author:
Peza Alexandra
Supervisors info:
Δρ. Κωνσταντίνος Βεκρέλλης, Ερευνητής Α' Ίδρυμα Ιατροβιολογικών Ερευνών Ακαδημίας Αθηνών (ΙΙΒΕΑΑ)
Δρ. Αναστάσιος Δελής, Μεταδιδακτορικός Ερευνητής, Ίδρυμα Ιατροβιολογικών Ερευνών Ακαδημίας Αθηνών (ΙΙΒΕΑΑ)
Δρ. Θεόδωρος Δαλαμάγκας, Κύριος Ερευνητής, Ινστιτούτο Πληροφοριακών Συστημάτων Ερευνητικού Κέντρου «ΑΘΗΝΑ» (ΕΚ «Αθηνά»)
Original Title:
Brain-derived exosomes as means to examine molecular mechanisms that affect cellular homeostasis. An informatics approach.
Languages:
English
Translated title:
Brain-derived exosomes as means to examine molecular mechanisms that affect cellular homeostasis. An informatics approach.
Summary:
This thesis explores the role of brain-derived exosomes in understanding molecular mechanisms affecting cellular homeostasis, employing advanced bioinformatics and machine learning techniques. Exosomes, small extracellular vesicles, are integral in cellular communication within the central nervous system (CNS), influencing processes such as apoptosis, cell proliferation, and inflammatory responses. This study focuses on the internalization and endocytic trafficking of exosomes in microglia and astrocytes, using images generated from confocal microscopy and processed with Imaris software to extract relevant statistical information.
We utilized Topological Data Analysis (TDA) and various machine learning classifiers to analyze the dataset. Initially, the Mapper technique, a TDA approach, was applied to identify subcategories within predefined treatment groups. However, due to the high correlation and lack of sufficient data, this approach could not effectively distinguish the existing categories. Subsequently, we employed supervised machine learning methods like multiclass classification to build a predictive model capable of classifying images into four treatment classes. Hyper-parameter optimization using GridSearch and cross-validation was performed to enhance model accuracy.
The study also incorporated ensemble learning methods, specifically bagging and stacking, to improve model robustness and performance. Despite these efforts, the best accuracy achieved was 0.64%, indicating a need for further refinement. Recursive Feature Elimination (RFE) was used to identify and remove less significant features, which did not significantly improve the results. However, focusing on cytochalasin-treated and non-treated cells, reflecting actin-dependent internalization, led to a notable accuracy improvement, achieving 0.91% with k-nearest neighbors (k-NN) and 0.93% with bagging.
This research highlights the complexities of exosome internalization mechanisms and underscores the potential of combining topological and machine learning approaches to enhance our understanding of cellular processes. Future work will aim to refine these methodologies further, with a focus on expanding the dataset and exploring additional machine learning techniques to improve predictive accuracy and uncover deeper insights into the role of exosomes in cellular homeostasis.
Main subject category:
Science
Keywords:
exosomes, topological data analysis, machine learning, multi class classification, ensemble methods, bagging, combining binary classifiers, stacking
Index:
No
Number of index pages:
0
Contains images:
Yes
Number of references:
3
Number of pages:
74
File:
File access is restricted until 2025-07-31.

PEZA_ALEXANDRA_7115152100035_master_thesis_final.pdf
4 MB
File access is restricted until 2025-07-31.