A carregar...
Publicação
Predicting non-coding RNA function using Artificial Intelligence
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.44 MB | Adobe PDF |
Resumo(s)
Non-coding RNAs (ncRNAs) represent the majority of human gene products and are involved in various
important biological processes, being considered relevant disease biomarkers and therapeutic agents.
However, there are few functional annotation databases dedicated to ncRNAs and information about
these biomolecules remains sparsely distributed, mostly in the form of scientific research articles. It is
then of pivotal importance to aggregate and summarize the existing information.
Natural Language Processing methods applied to text mining enable automatic information extraction
and summarization from textual data. These techniques can be used to generate collections of annotated
sentences expressing relations between entities, called relational corpora.
In this work, a text mining pipeline was implemented to generate a ncRNA-phenotype relational
corpus (ncoRP) using Distant Supervision Relation Extraction (DSRE), consisting of 21,608 annotated
articles, 2,835 unique ncRNAs, 1,118 unique phenotypes and 35,295 unique relations, with a precision of
0.761 and F1-score of 0.593, calculated through human validation. DSRE methods require a set of predocumented relations to function, as such, a high-fidelity ncRNA-phenotype relation dataset, consisting
of 214,300 unique relations, was created by the aggregation of five ncRNA-disease functional annotation
databases. Then, both ncoRP and the relation dataset represent important contributions towards solving
the problem with the sparseness of information about ncRNAs.
Large Language Models (LLMs) are an emerging type of language model, showing great capabilities
in general task-solving through text generation, without the requirement of fine-tuning with large datasets.
This benefit shows promise for applications in Relation Extraction (RE), when compared to data-intensive
state-of-the-art deep learning methods. In this work, a LLM RE methodology is proposed and evaluated,
achieving an F1-score of 0.978 by combining the RE task with a preceding sentence filtering task and
applying prompting principles such as in-context learning and Chain-of-Thought self-explanation.
Descrição
Tese de mestrado, Bioinformática e Biologia Computacional, 2024, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
RNAs não codificantes Extração de Relações Prospeção de Texto Supervisão à Distância Grandes Modelos de Linguagem Teses de mestrado - 2024