Melhoria da quantidade e qualidade de predições do Genppi via uso do Random Forest em Common Lisp para identificar similaridade proteica.

Abstract

The context of this study lies in the growing need for efficient computational methods to predict PPIs, given the importance of these interactions for understanding cellular biology and developing new drugs. Experimental methods are costly and time-consuming, making computational approaches a promising alternative. The original version of Genppi already represented an advancement in this field, but its capabilities were limited by its dependence on high protein sequence similarity. This work presents a significant advancement in the development of the Genppi software, a bioinformatics tool dedicated to predicting protein-protein interaction (PPI) networks in bacterial genomes. Version 1.5 of Genppi incorporates the Random Forest algorithm, implemented in Common Lisp, with the goal of overcoming the limitations of the previous version and expanding the accuracy and scope of protein interaction analysis, especially in scenarios with low sequence similarity. The research addresses both the implementation of machine learning and the optimization of network generation through the RIS algorithm (Reduced Interaction Sampling), developed to handle the computational complexity inherent in analyzing complete genomes.

The methodology employed was multifaceted. Initially, the Random Forest algorithm was integrated into Genppi, enabling efficient protein similarity classification. Subsequently, to optimize software performance, the RIS algorithm was implemented to reduce the computational complexity associated with generating interaction networks from conserved phylogenetic profiles with a large number of genes. Evaluating the effectiveness of these implementations involved generating 180 interaction networks in different configurations, varying the operating mode (Features, Machine Learning), the Top_N parameter (defining the number of most relevant nodes), and the number of files analyzed. The generated networks were subjected to rigorous analysis using various topological metrics (Degree, Betweenness Centrality, Bridging Centrality) and statistical tests (Kolmogorov-Smirnov, Kendall Tau correlation, Fleiss’ Kappa) to assess the stability, accuracy, and structural integrity of the networks.

The results demonstrated that the integration of Random Forest in Machine Learning mode significantly expanded Genppi’s ability to predict protein interactions. Applying the model to the genome of the bacterium Buchnera Aphidicola revealed an overlap of up to 62% with interactions documented in the STRING database, validating the accuracy and completeness of the predictions. The Features mode, which utilizes predefined biophysical characteristics, also presented consistent results, maintaining high structural stability even with the application of the RIS algorithm. Analysis of the RIS algorithm revealed that, although it introduces some randomness, the identification of critical nodes remains robust for Top_N values greater than or equal to 100, indicating that reducing the number of edges does not significantly compromise the network’s integrity. The Weighted Global Mean (MGP) and other indicators confirmed that the algorithm preserves the global structure of the networks, even with reduced computational complexity.

In conclusion, this work demonstrates that the combination of the Random Forest algorithm with the RIS algorithm represents a significant advancement in the prediction of protein interaction networks. The integration of Random Forest allowed Genppi to overcome the limitation of dependence on high sequence similarity, while the RIS algorithm optimized software performance, enabling efficient analysis of complex genomes. Validation of the results with the STRING database and rigorous statistical analysis confirm the accuracy and robustness of the predictions generated by Genppi 1.5. Public availability of the data and code promotes transparency and reproducibility of the research, encouraging collaboration and advancement in bioinformatics. This study contributes to the development of more effective tools for understanding molecular processes in bacterial organisms, with potential applications in biomedicine, biotechnology, and other scientific fields. Genppi, enhanced with these new features, solidifies its position as a robust and scalable solution for predicting protein-protein interactions based on machine learning and biophysical characteristics of genomes.