MSBack: Multiscale Backmapping of Highly Coarse-Grained Proteins Using Constrained Diffusion

Abstract

Coarse-grained (CG) molecular dynamics is a powerful tool for simulating the collective behavior of biomolecules. However, the structural information lost during coarse-graining prevents the CG configurations from being more widely useful (e.g., for ligand binding). Regenerating the lost all-atom coordinates, or backmapping, is an unmet challenge for protein CG at resolutions lower than one coarse-grain site or bead per amino acid residue. This low resolution is computationally necessary to simulate many protein complexes including viruses like SARS-CoV-2 and HIV-1. We propose MSBack, a method to backmap highly CG proteins using a diffusion model for the all-atom coordinates constrained to fit the CG coordinates. This diffusion process works by perturbing a known all-atom structure and does not require retraining. We show that this stochastically generates a distribution of α-carbon traces that match the CG coordinates. By combining this with physics-based methods for smaller-length backmapping, we fully backmap a mature HIV-1 capsid bound with the small molecule inositol hexakisphosphate at 1 Å resolution.