Design of Carbon Nanotube Inhibitors for Main Proteinase of SARS-CoV-2: A Combined Deep Learning and Molecular Dynamics Simulation Study

The rapid development of machine learning (ML) and deep learning (DL) methods provides new opportunities for innovative drug discovery. While these techniques are widely used in docking organic molecules (drugs) with protein, an evaluation of the performance of ML and DL in treating nanostructures remains lacking. This situation becomes the main hindrance for the development of nanostructure-based medicine and medical materials. In this study, we have compared the performance of a recently developed DL model, named DeepRMSD + Vina, with the traditional molecular dynamic (MD) simulations in treating the docking problem of a carbon nanotube, the most representative nanomaterial, and the main proteinase (M^pro) of SARS-CoV-2 as the representative case. Our results indicate that most of the DL-generated structures are in good agreement with the structures optimized by MD. However, minor discrepancies were indeed observed where structural alterations happened for the flexible loops near the binding pocket of M^pro, which were not addressed in the DL-generated structures. Further analyses demonstrated that the DL-generated binding pose is a metastable conformation at a local minimum on the energy surface. The transition from such a local minimum structure to a global minimum structure has to overcome an energy barrier that is accompanied by a flip of the flexible loops. In brief, we demonstrate that the DL method has considerably high efficacy in treating nanobiosystems, with potential implications for the design of nanomedicine materials. This study also shed new light on the future development direction of the DL method for enhanced docking accuracy.