Predição de regiões codificadoras de proteínas em rna circulares e transcriptoma em montagem de novo

Abstract

The prediction of protein-coding regions (CDSs) in circular RNAs (circRNAs) and de novo assembled transcriptomes represents significant open challenges in bioinformatics due to the non-canonical nature of translation in circRNAs and the heterogeneity of transcripts generated in assemblies. This thesis proposes computational methods that integrate complementary approaches of kernel methods and probabilistic models to overcome these limitations. We developed four main contributions: (i) circTIS, a tool based on a Support Vector Machine (SVM) with the weighted degree kernel (WDK) for predicting translation initiation sites (TIS) in circRNAs, achieving a precision of 93.12% and sensitivity of 86.03%, outperforming established tools; (ii) cirCodAn, an annotator employing generalized hidden Markov models (GHMMs) adapted to molecular cyclicity, achieving F1-scores of up to 77.06% in identifying CDSs in H. sapiens circRNAs; (iii) strkernels, a high-performance Python package with efficient implementations of string kernels, validated in tasks such as antimicrobial peptide classification (AUC-ROC of 0.99); and (iv) CodAnSVM, a hybrid method combining GHMMs and SVMs for annotating de novo assembled transcriptomes, demonstrating a 5 percentual points F1-score improvement over traditional methods for complete transcripts and satisfactory results for partial transcripts. The results show that integrating generative (GHMMs) and discriminative (SVMs) models enables precise analyses in challenging biological scenarios, such as predicting non-canonical initiation codons and annotating partial transcripts. The developed tools, available as open-source software, address methodological gaps in the study of coding circRNAs and the functional analysis of transcriptomes.