Protein dynamics is fundamentally a trajectory prediction problem, but molecular dynamics (MD) simulation remains expensive and static structure predictors do not model time-ordered motion. We present VelocityFM, a short-horizon protein trajectory predictor that applies rectified flow matching in velocity space over residue frames and torsions. The model combines six Invariant Point Attention (IPA) blocks with a two-layer per-residue temporal self-attention encoder, and is trained on 710 ATLAS proteins comprising 2090 filtered replicate trajectories. At the primary 128-frame rollout horizon, VelocityFM achieves a median TM-score of 0.929 on 72 held-out proteins, with 100% of proteins remaining above TM> 0.7 and 100% clash-free generation. Backbone geometry also remains strong, with a median Ramachandran favoured rate of 91.09%, while dynamics calibration is conservative with median RMSF ratio 0.697. These results show that velocity-space geometric learning can generalise short-horizon trajectory prediction to unseen proteins while preserving fold structure and geometric validity within its intended operating regime.
Jayathilake, L. et al. · CC-BY 4.0