Avaliação do potencial preditivo de assinaturas estruturais na identificação de interações proteína-peptídeo

Abstract

Protein-peptide interactions are essential for the proper functioning of the human body, playing crucial roles in various biological processes. In silico studies based on artificial intelligence have become fundamental for the development of peptide-based therapies. However, predicting these interactions remains a challenge in structural bioinformatics, primarily due to the intrinsic flexibility of peptides and the difficulty in developing more accurate metrics for ranking the expected positions of ligands in molecular docking. In this study, we evaluated the predictive potential of structural signatures in identifying protein-peptide interactions using machine learning models. The dissertation is structured into four main parts: I. Evaluation of three docking and molecular modeling tools for protein-peptide complexes: HPEPDOCK, HDOCK, and AlphaFold Multimer; II. Comparison of peptide modeling in holo and apo forms using AlphaFold Multimer; III. Development of machine learning models for predicting real and fake protein-peptide complexes using structural signatures of the protein-peptide interface; and IV. Development of machine learning models for predicting protein-peptide interactions using concatenated structural signatures from protein interface and peptide. The main results indicate that the three docking tools evaluated (HPEPDOCK, HDOCK, and AlphaFold Multimer) showed similar performances in complex modeling, effectively distinguishing real complexes from false ones. HPEPDOCK stood out with the highest median DockQ score (0.935) when comparing predicted complexes with experimental ones. In the comparison between peptides modeled in holo and apo forms, the results were similar, with an average RMSD of 0.214. Machine learning models for classifying real and fake complexes were developed, achieving an F1-score of 0.741 for the best model, using Neural Networks. Additionally, for protein-peptide interaction prediction, the best model developed with the Gradient Boosting algorithm achieved an F1-score of 0.930. In summary, structural signatures proved to be valuable tools for training protein-peptide interaction prediction models. The developed models can be used to facilitate virtual peptide screening, contributing to the development of new peptide-based therapies and drugs.