Live model · runs in your browser

Peptide torsion dynamics
from sequence alone

An open-source, sequence-only neural propagator for backbone (φ, ψ) ensembles of short peptides. Type a sequence, get a Ramachandran density and basin populations in seconds — no MD, no install.

2.39×
closer to MD than Timewarp
(mean JSD, 4-residue benchmark)
~78K
parameters · ≈3000× smaller
than the Cartesian baseline
4–20 aa
peptide length range
backbone torsions only
Trajectories8
More trajectories → smoother density (slower).
Steps per trajectory500
Longer rollout → more samples on the torus.

Ramachandran density

Enter a sequence and hit Generate — the first run wakes the model (~30–60 s), then it's a few seconds.

What this is — and what it isn't

  • Is: the density of backbone (φ, ψ) torsion ensembles for short peptides — where the chain spends time on the Ramachandran torus.
  • Isn't: not 3D de-novo folding, not side-chain rotamers, not kinetics/rates. It captures where, not when.
  • Benchmark honesty: the 2.39× figure is mean JSD on 4-residue peptides (0.196 vs Timewarp 0.468). On longer peptides the gap narrows.
  • Open source, Apache-2.0 — pip install alphadynamics. The button above calls the public model on Hugging Face; nothing is stored.