VelocityFM

Protein function is driven by motion: enzymes flex through catalytic cycles, receptors toggle between active and inactive states, and binding sites open and close on timescales that a single static structure cannot describe. Molecular dynamics (MD) simulation is the standard tool for capturing this time-ordered behaviour, but it is computationally expensive, while fold-prediction models such as AlphaFold output one structure with no notion of how it evolves over time. VelocityFM reframes protein dynamics explicitly as a trajectory-prediction problem and learns to generate short MD-like rollouts directly.

Developed by Lahiru Jayathilake and Chithranjali Rupika Wijesinghe at the University of Colombo School of Computing with Ruvan Weerasinghe of the Informatics Institute of Technology, and posted to bioRxiv in June 2026, VelocityFM applies rectified flow matching in velocity space over residue frames and backbone torsions. Rather than denoising a fixed set of structures jointly, it learns the per-residue velocity field that carries a protein from one frame to the next, producing time-ordered motion within a short operating horizon.

VelocityFM sits alongside generative dynamics models such as AlphaFlow, which fine-tunes a folding network with flow matching to sample conformational ensembles, and ProAR, which generates trajectories autoregressively. It is a distinct model: its contribution is learning geometry directly in velocity space over frames and torsions, and demonstrating that this representation generalises short-horizon trajectory prediction to proteins never seen during training.

#Key Features

• Velocity-space flow matching: Uses rectified flow matching over residue-frame and torsion velocities, learning how a structure moves rather than only what conformations it visits, which yields time-ordered short-horizon trajectories.
• Geometry-aware architecture: Couples six Invariant Point Attention (IPA) blocks with a two-layer per-residue temporal self-attention encoder, combining SE(3)-equivariant spatial reasoning with explicit temporal context.
• Zero-shot generalisation: Applies to unseen proteins without retraining, reaching a median TM-score of 0.929 across 72 held-out proteins at the 128-frame rollout horizon.
• Geometrically valid output: Generates 100% clash-free structures with a median Ramachandran-favoured rate of 91.09%, indicating that predicted motion preserves physically realistic backbone geometry.
• Honest dynamics calibration: Reports a conservative median RMSF ratio of 0.697, transparently noting that the model under-predicts the full amplitude of fluctuations within its short-horizon regime.

#Technical Details

VelocityFM is a generative geometric model trained on MD trajectories from the ATLAS dataset, using 710 proteins comprising 2090 filtered replicate trajectories. The network represents protein state as residue frames plus backbone torsions and learns a velocity field under a rectified flow matching objective; six IPA blocks provide equivariant spatial attention over residues, while a two-layer per-residue temporal self-attention encoder supplies time context across the rollout. At its primary 128-frame rollout horizon the model achieves a median TM-score of 0.929 on 72 held-out proteins, with 100% of test proteins remaining above TM > 0.7 and 100% clash-free generation. Backbone quality stays strong (median Ramachandran-favoured rate 91.09%), and dynamics calibration is deliberately conservative, with a median RMSF ratio of 0.697 that the authors flag as under-estimating true fluctuation amplitude. The reported scope is explicitly short-horizon trajectory prediction rather than long-timescale equilibrium sampling.

#Applications

VelocityFM is aimed at computational structural biologists who need fast, time-ordered glimpses of protein motion without running full MD. Because it generalises zero-shot to unseen proteins, it can supply short trajectory rollouts for proteins lacking prior simulation data, generate physically plausible starting points or conformational priors for downstream MD, or screen many proteins for short-horizon flexibility patterns at a fraction of simulation cost. Its strong fold preservation and clash-free output make the generated frames directly usable in structure-based analysis pipelines, while its conservative fluctuation calibration signals where full MD remains necessary.

#Impact

VelocityFM contributes to the rapidly growing effort to replace or accelerate physics-based MD with learned generative models, and its framing of dynamics as velocity-space flow matching over frames and torsions is a distinctive alternative to ensemble-sampling and autoregressive approaches. The demonstration that this representation generalises short-horizon prediction to unseen proteins while maintaining fold integrity and geometric validity is its central result. As a recent preprint with no public code or weights release identified at the time of writing, its reported metrics await independent reproduction, and its honestly stated limitations—conservative RMSF amplitude and a short rollout horizon—mark it as a step toward, rather than a replacement for, full molecular dynamics.