Aprimorando a classificação de mirtrons com aumento de dados e modelos transformers

Abstract

Among the classes of non-coding RNA (ncRNA) sequences, microRNAs and mirtrons are the most extensively studied in the literature, primarily due to their importance in understanding the regulatory mechanisms of these small ncRNAs. From a computational perspective, particularly regarding mirtron classification, a key challenge is the data imbalance problem, which can hinder the generalization ability of machine learning models and contribute, for instance, to model overfitting. This issue compromises the accuracy and reliability of analyses, hindering progress in the identification and characterization of these ncRNA classes. To mitigate the challenges of data imbalance in small sample datasets, this study explores two main strategies. The first involves the development of GENNUS: GENerative Approaches for NUcleotide Sequences, a framework of generative techniques aimed at synthesizing mirtron data. GENNUS comprises two approaches based on Generative Adversarial Networks (GAN) and three approaches based on the Synthetic Minority Over-sampling Technique (SMOTE). The results demonstrate that these approaches are capable of generating synthetic data while preserving the intrinsic patterns and inherent diversity of real mirtron sequences. Moreover, the use of synthetic data eliminates the need for extensive feature engineering, making the modeling process more efficient and accessible. The second strategy explores deep learning techniques, specifically Transformer-based architectures, for the classification of mirtrons and canonical miRNAs. These models leverage the Transformer’s ability to capture long-range dependencies and complex relationships within nucleotide sequences. The results indicate that Transformer-based classifiers outperform stateof-the-art methods, yielding substantial improvements in both generalization and classification accuracy. This approach highlights the potential of deep learning models to enhance ncRNA analysis, paving the way for significant advancements in the field of bioinformatics. This work emphasizes the transformative impact of synthetic data generation and the use of advanced deep learning architectures in addressing the limitations imposed by data imbalance in genomics. The proposed methodologies offer more effective and scalable solutions for ncRNA classification, contributing to a deeper understanding of genetic regulation and enabling new discoveries in computational molecular biology. Additionally, this work also presents contributions in the classification of long non-coding RNAs (lncRNA), further expanding and demonstrating the applicability of machine learning techniques across different contexts of genomic analysis.