Development of a machine-learning model to predict Gibbs free energy of binding for protein-ligand complexes

The possibility of using the atomic coordinates of protein-ligand complexes to assess binding affinity has a beneficial impact in the early stages of drug development and design. From the computational view, the creation of reliable scoring functions is still an open problem in the simulation of biological systems, and the development of a new generation machine-learning model is an active research field. In this work, we propose a novel scoring function to predict Gibbs free energy of binding (ΔG) based on the crystallographic structure of complexes involving a protein and an active ligand. We made use of the energy terms available the AutoDock Vina scoring function and trained a novel function using the machine learning methods available in the program SAnDReS. We used a training set composed exclusively of high-resolution crystallographic structures for which the ΔG data was available. We describe here the methodology to develop a machine-learning model to predict binding affinity using the program SAnDReS. Statistical analysis of our machine-learning model indicated a superior performance when compared to the MolDock, Plants, AutoDock 4, and AutoDock Vina scoring functions. We expect that this new machine-learning model could improve drug design and development through the application of a reliable scoring function in the analysis virtual screening simulations. [Display omitted] •Development of a machine-learning model to predict free energy of binding for protein-ligand complexes;•The use of a dataset composed of 48 high-resolution crystallographic structures to be used to build a new scoring function;•Improved predictive power of the machine learning model to predict ΔG, when compared with classical scoring functions.