Topology based machine learning models for drug design

Abstract

Binding affinity prediction from protein-ligand complex is a problem of interest as it is a key step in drug design. A good model for binding affinity prediction can help to lower time needed and cost of drug design. The binding affinity problem is unlike traditional machine learning tasks. Each protein-ligand complex consists of varying number and types of elements. For machine learning model to work, each input data must be of the same shape. It is also a difficult task to extract geometric features of protein-ligand complexes as well as the chemical interactions between the biomolecules. The paper explores the use of topological methods to featurize protein-ligand complexes to capture the geometric features and chemical interactions of the biomolecules before using machine learning techniques for the binding affinity prediction task. In this report, we followed two papers, (Meng, 2020) and (Zixuan Cang, 2018) closely and made changes to the final machine learning models. We compared our proposed models with some of the recent works and showed that our proposed models managed to outperform some of them.