Investigação sobre a capacidade de predição de afinidade de ligação entre moléculas em sistemas hospedeiro-hóspede por meio de métodos de aprendizado de máquina.

Abstract

The insertion of in silico experiments in the scientific context in recent decades has allowed the consolidation of interdisciplinary areas such as bioinformatics, computational biology, computational chemistry, among others, which seek to describe, understand and predict natural events through mathematical equations and computational methods. In this context, it is frequent that researchers are interested in predicting interaction measures between molecules, mainly to enable the rational study of drugs. Performing screenings of potential drugs computationally aims to reduce time to discover new drugs and reduce the high number of laboratory tests that make the whole process more expensive. Researchers usually perform computational molecular screenings through docking techniques, which define degrees of freedom for molecular representations within a simulation grid. The goal of the process is to evolve an optimization in this space that aims to find the geometric configuration of a possible bond between molecules and calculate metrics relating to this interaction state. To this end, the literature already presents several proposals for the formulation of objective functions for the search, sometimes based on mathematical models from the perspective of classical physics, sometimes based on models based on quantum theory. Recently, as an alternative, predictive models based on data and adjusted by computational machine learning methods have been proposed. Surprisingly, some of these methods have shown better results than the physical models, with lower prediction time once trained. Therefore, Machine Learning (ML) techniques are an integral part of rational drug design and discovery. Cyclodextrins (CDs) are nano-cages (nanohorns) used to enhance the delivery of insoluble or toxic drugs to the body. Due to the chemical similarity between CDs and proteins, ML approaches can vastly benefit studies in the field by identifying promising carriers for a given molecule of interest. In the present work, the performance of three well-known ML methods in the literature - Support Vector Regression (ε-SVR), Gaussian Process Regression (GPR), and eXtreme Gradient Boosting (XGB) - are evaluated to predict the binding affinity of cyclodextrin and ligands of interest in a host-ligand system (host-guest). We have tuned the hyperparameters of the proposed ML methods in a Random Search strategy. The results show the consistency of the methodology used by presenting controlled average error results. The best prediction performance was obtained by a GPR model optimized in random search, fitting the data well (R2 = 0.803) with low prediction errors (RMSE = 1.811kJ/mol and MAE = 1.201kJ/mol).