Multistep generative backmapping of coarse-grained structures

Data-driven backmapping from coarse-grained (CG) to fine-grained (FG) representations remains challenging for complex biomolecular systems such as proteins, where methods often suffer from limited accuracy, training instability, and compromised physical realism. We present a novel multistep generative framework that enables stepwise refinement from CG beads to full FG detail by integrating conditional Variational Autoencoders with graph-based neural networks. The probabilistic formulation of multistep backmapping is outlined, and numerical experiments on proteins with diverse structures and ultra-coarse representations demonstrate that multistep schemes substantially enhance reconstruction accuracy while improving computational efficiency during training compared to single-step alternatives.