Structural Knowledge Is What Matters in Protein-Ligand Binding Affinity Prediction

Binding affinity prediction is about estimating the degree to which a drug binds to a protein. Predicting the binding affinity between a drug and a protein in a computational process helps researchers filter huge libraries of compounds before performing expensive biochemical lab experiments. Currently, there is interest in predicting binding affinity through computational pattern recognition or machine learning methods instead of the classical physics-inspired methods, which are computationally intractable except for tiny chemical compounds. In the last five years, several machine learning-based methods have been presented, whose experimental validations have achieved increasing Pearson coefficients while trained and tested in the PDBBind 2016 and CASF 2016 databases, respectively. These methods have an important diversity of architectures that provide different properties. The aim of this paper is to discern which binary properties (existence or absence) of these methods make them return higher Pearson coefficients. Basically, the properties introduced are related to the level of structural knowledge, the presence of 3D information, and the introduction of the relationship between the drug and the protein in the input of the model. The t-test confirms that the important binary properties for having a high Pearson coefficient are the protein (or part of the protein) being represented and introduced into the computational model as a graph, the pocket and the drug-protein interaction being part of the input, and incorporating the distance between atoms and the type of chemical bonds into the model.