University of Thessaly (UTH)
Information mining in molecular biology with emphasis on language processing and machine learning techniques
Abstract
dc:descriptionMicroRNAs (miRNAs) are small non-coding RNAs that function as key post-transcriptional regulators of gene expression, modulating diverse pathways through complex gene regulatory networks. The biological consequences of miRNA activity depend not only on their biogenesis and target recognition but also on a variety of other molecular and cellular factors. Genetic variants in miRNAs and their binding sites can alter these interactions, leading to changes in gene regulation that, together with many other elements, contribute to disease pathogenesis and cancer susceptibility. As a result, detailed study of miRNA–gene associations remains vital for molecular medicine and translational genomics. This thesis presents the development and evaluation of computational pipelines for the automated identification and characterization of miRNA–gene interactions from biomedical literature and experimental data. Advanced natural language processing (NLP) methodologies - including transformer-based and deep learning models—were used for large-scale screening, annotation, and extraction of candidate associations. The pipeline combined careful entity normalization, dependency parsing, and contextual filtering, followed by systematic evaluation against manually curated gold-standard datasets. To improve precision and the capture of semantically meaningful relationships, large language models (LLMs) were incorporated as post-processing filters, using prompt design and in-context examples to refine predictions and exclude erroneous extractions. In parallel, this thesis introduces microT-CNN: a convolutional neural network designed for miRNA target prediction at the transcriptome scale. Trained on diverse high-throughput experimental datasets - such as CLIP-seq, chimeric miRNA–target fragments, and tissue-matched perturbation experiments—microT-CNN identifies both canonical and non-canonical miRNA binding within 3’ UTR and coding regions. The model also captures virus-encoded miRNA–host gene interactions, and leverages features such as sequence conservation, regional accessibility, and binding structure. This flexible architecture enables microT-CNN to advance the characterization of miRNA targeting beyond traditional models. Extensions of these information extraction methods enabled the systematic identification of gene–variant–cancer associations, using entity-primed LLM pipelines applied to a large corpus of PubMed abstracts. The resulting platform supports consistent annotation and extraction of complex triplet relationships among genes, variants, and cancer types—facilitating research into molecular mechanisms, disease associations, and clinical outcomes. Collectively, the datasets and computational tools produced in this thesis provide a framework for translational research, hypothesis generation, and integration with biomedical databases. While these approaches offer notable advances, ongoing efforts to incorporate new experimental and literature resources, improve entity recognition, and adapt to other biomolecular association tasks will be essential for further progress. It is anticipated that the contributions of this work will support continued development in molecular network discovery and precision medicine. During the course of this thesis, the candidate contributed to five scientific studies spanning machine learning, natural language processing, deep learning, and large-scale data mining in computational biology. The research focused on both the automated extraction and curation of molecular biomarker associations - using advanced NLP pipelines and large language models - and the development of neural networks tailored for transcriptome-scale miRNA target prediction. Central topics included microRNAs, genetic variants, genes, and cancer types. Special emphasis was placed on the creation, benchmarking, and application of robust information extraction and prediction algorithms to advance translational genomics and biomedical knowledge discovery. These works have been published in high-impact, international journals and have collectively received 418 citations to date (Google Scholar, 04/2025). The following publications are organized by their principal research topic, with a summary of the candidate’s main contributions and the broader impact of each study.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zacharopolou, Elissavet
- Ζαχαροπούλου, Ελισσάβετ
Subjects
dc:subject × 19- Επεξεργασία φυσικής γλώσσας
- Μηχανική μάθηση
- Μεγάλα γλωσσικά μοντέλα
- Εξόρυξη πληροφορίας
- ΜικροRNA
- Γονιδιακή ρύθμιση
- Natural language processing
- Machine learning
- Large language models
- Information extraction
- MicroRNAs
- Gene regulation
- Cancer Sensitively
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Βιοπληροφορική
- Computer and Information Sciences
- Natural Sciences
- Bioinformatics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58973