Spectral Diffusion for Protein Dynamics

Generative models present a promising alternative to expensive molecular dynamics for computationally querying protein dynamics, yet many existing approaches treat ensembles as unordered snapshots rather than temporally coherent trajectories. We present a new physics-informed representation using Fourier transforms as an inductive bias for the multiscale temporal nature of protein dynamics. Diffusion in the spectral domain allows for disentangling of dynamics into slow conformational modes and fast atomic jitter, enabling rapid, improved prediction of dynamics across a range of temperatures. This is facilitated by direct denoising of structure and temperature conditioned spectral volumes where the low frequencies encode per-residue flexibility. Trained on the mdCATH dataset, we evaluate our model, DynaMode, on a held-out test set achieving an RMSF pearson $r$ of $0.844$. However, we suffer from significantly more steric clashes than standard molecular dynamics, suggesting more explicit structural reasoning is necessary for state of the art dynamics emulation.