Cross-Scale Motion Pretraining for Data-Efficient RNA Dynamics Modeling

Learning surrogate models of RNA dynamics is data-intensive because each new sequence or condition requires computationally expensive atomistic simulations to generate training trajectories. We ask whether low-cost macroscopic trajectories can provide a useful initialization for learning molecular dynamics, despite the vast disparities in spatial and temporal scales. We do not assume that the underlying physics is scale invariant. Rather, we hypothesize that, after nondimensionalization, particle systems across scales exhibit transferable dynamical primitives, including local interaction patterns, correlated displacements, and constraint propagation. We pretrain a particle-based trajectory model on simulated elastic-body trajectories and then fine-tune the entire model on atomistic trajectories from four RNA systems. Compared with the same architecture initialized randomly, the pretrained model starts RNA fine-tuning with a lower validation loss and reaches comparable loss levels with fewer RNA-domain updates. Under limited RNA supervision, the pretrained model converges to a better solution, consistently achieving lower final validation loss across random seeds.