🧬 AlphaDynamics — Protein Torsion Dynamics from Sequence

A tiny (~123K param) neural propagator that predicts the Ramachandran density of a peptide's backbone (φ, ψ) angles from sequence alone.

On the canonical 4AA benchmark: 2.39× lower JSD than Microsoft Timewarp at 3000× fewer parameters. Cross-validated against Top8000 PDB statistics.

📦 pip install alphadynamics · 🐙 GitHub · 🤗 Model card

This demo is CPU-only — limited to 50 residues / 16 trajectories / 1000 steps. For longer peptides or larger ensembles, install locally.
1 16
50 1000
Try a known peptide:
Peptide sequence Trajectories (more = smoother density) Steps per trajectory

What this tool does

Predicts an ensemble of (φ, ψ) torsion-angle trajectories for any peptide sequence (4–100 residues recommended, capped at 50 on this demo). Useful for:

  • Quick conformational triage before launching expensive MD simulations
  • Comparing sequence variants / mutants side by side
  • Estimating α-helix / β-sheet / PPII basin populations
  • Teaching biochemistry with live, interactive Ramachandran plots
  • AI-for-biology baselines and benchmarks

Honest limits

  • Density only, not kinetics (no transition rates / dwell times)
  • Backbone only, no side-chain rotamers (χ angles)
  • Monomer only, no multimer / aggregation
  • Best for 4–100 residue peptides; reliability degrades outside

How to read the Ramachandran plot

  • α-R region (φ ≈ -60, ψ ≈ -45) → right-handed alpha-helix
  • β-sheet region (φ ≈ -120, ψ ≈ 120) → extended / beta-strand conformations
  • PPII region (φ ≈ -60, ψ ≈ 140) → polyproline-II extended (very common in short peptides in solution)
  • α-L region (φ ≈ 60, ψ ≈ 50) → left-handed helix, sterically forbidden for almost all amino acids; should be close to 0% if the model honors physics

Architecture (one paragraph)

A residue's (φ, ψ) is treated as a phase pair on a torus. An MLP emits per-residue oscillator parameters from sequence + position + current angles. A phase-flow ODE integrates 64 coupled phase oscillators with RK4. The result is decoded into a mixture of axis-independent von Mises distributions, sampled, and rolled out autoregressively.

This is the protein-dynamics application of a multi-year line of work on phase oscillators (REZON hardware, phase-entanglement-rc, theta-gamma neural coupling).

Created by Krzysztof Gwozdz 🇵🇱 · Apache 2.0 · Cite · GitHub Issues for feedback