Development of molecular docking methods using Artificial Intelligence techniques

VILLARREAL MA; QUIROGA, RODRIGO

Congreso; LI Reunión Anual de la SAB; 2023

Molecular docking is a computational method for predicting the energy, position andorientation of ligand binding to a protein, and is a key tool in structure-based drug design.In essence, the method involves finding the global minimum of a mathematical functionrepresenting the affinity energy of a ligand at a given position on the protein surface. Thisfunction, called the scoring function, is developed on the basis of different approaches.These include fully empirical, force field-based and artificial intelligence(AI)-basedmethods. Whatever approach is used to describe the scoring function, there are alwaysparameters that need to be optimized for a successful application.To parameterize a scoring function, the correlation between predicted and experimentallymeasured binding values is usually optimized using a large set of protein-ligandcomplexes for which both their structure and binding energy are known. However thisstrategy may prove to be suboptimal since the global minimum of the scoring function isnot guaranteed to match the experimental structure of the protein-ligand complex. Thisrequirement is key for molecular docking applications.In this work we develop a self-consistent methodology to parameterize scoring functionsusing AI techniques. By applying automatic derivatives, mini-batch annealing and self-supervised learning we show that our method is able to optimize scoring functions formolecular docking of up to 500 parameters in a fast and efficient way. Throughcomputational experiments we compare the advantages of this new methodology andalso evaluate the effect of using different loss functions and regularization terms. Ourmethod is written in the Julia programming language, developed as open source and iseasily adaptable to a wide variety of scoring functions.