Genomas bacterianos antifúngicos: um pipeline baseado empython para extração de características e classificação por aprendizado de máquina

Abstract

The evolution of agriculture has brought with it the challenge of managing pests that threaten food productivity and quality. Initially, the use of pesticides was an effective solution to combat insects, rodents, fungi, and bacteria. However, the indiscriminate use of these products has led to serious environmental and public health problems, such as water and soil contamination, the accumulation of toxic residues in food, and the development of pesticide-resistant pests. This scenario has driven the search for more sustainable alternatives that are less harmful to the environment and human health. Within this context, the use of microorganisms that establish beneficial associations with plants has emerged as a promising strategy. Bacteria belonging to the genera Pseudomonas and Bacillus are known for their ability to synthesize antifungal metabolites, contributing to the biological control of plant diseases in agricultural crops. The primary objective of this research is to develop an automated pipeline using the Python programming language to extract genomic features from bacterial genomes and classify these bacteria based on their antifungal activity, employing Machine Learning algorithms. The central tool used in this study is MathFeature, which was designed to extract both biological and mathematical information from genomic sequences. This study aims to analyze the features extracted by MathFeature to train a machine learning model capable of performing binary classification of bacteria (with antifungal properties or not). The research also includes a review of computational tools and biological databases, such as GenBank and MIBiG, in addition to discussing previous studies that explore bacterial pathogen suppression capabilities and genomic feature extraction methods combined with Machine Learning algorithms. In the initial stages of this research, tests were conducted using classical Machine Learning algorithms (SVM, Logistic Regression, Random Forest, and Extra Trees) to validate the proposed approach. However, the results indicated significant challenges, such as the difficulty of adapting the models to the dataset and the imbalance between bacterial classes with and without antifungal activity. To overcome these limitations, a second phase was conducted, incorporating new strategies, including additional algorithms (Gradient Boosting, AdaBoost, Multi-Layer Perceptron, and LightGBM), class balancing techniques (SMOTE), and advanced feature selection methods (ANOVA, RFE, and Mutual Information). These approaches have proven to be promising in improving the model’s predictive capability, highlighting the need for continuous refinement of the proposed pipeline.