Bambu: desenvolvimento de uma ferramenta para QSAR baseada em aprendizado de máquina

Abstract

BioAssays Model Builder - BAMBU is a tool developed to assist researchers in research related to the development of new drugs. The identification of molecules with pharmacological potential is traditionally performed through discoveries of natural compounds, for which the drug discovery approach is used, within it the High Throughput Screening (HTS). Another approach used to develop these tools is machine learning (ML), a sub-area of artificial intelligence, which aims to develop algorithms capable of solving problems for which they were not explicitly programmed. It can be divided into two fundamental areas, supervised learning that aims to learn from the data provided and unsupervised learning that learns from sample patterns. One of the methodologies used in supervised ML in the context of drug discovery is the Quantitative Structure-Activity Relationship (QSAR), a quantitative study to see interactions between organic molecules and chemical structures in a three-dimensional way that aims to study the ligand. The bamboo tool joins the databases from HTS studies with the QSAR methodology that employs the use of supervised ML and uses ML algorithms to distinguish molecules with pharmacological potential from those that do not and can be used to assist in the search for new approaches therapeutics for various diseases including cancer and neurodegenerative diseases. To compose the tool we used models based on decision trees, neural networks and linear regression. As a balancing strategy, undersampling, oversampling, tomek links and SMOTE approaches are used. To assess the functioning of the tool, ranking metrics are used such as precision, recall, f1-score and accuracy. The use of models allows the tool to be able to separate active from inactive molecules and this is proven and achieves the objective of the work when we observe the precision data, which is one of the metrics used to validate that tells us if the model is managing to separate active molecules from inactive molecules. We also observed that models based on decision trees and decision tree ensembles are the ones with the most satisfactory results.