AlphaDynamics

Compact sequence-only neural propagator for protein torsion dynamics.

2.39× lower JSD than Microsoft Timewarp · 3000× fewer parameters · 64 phase oscillators

A tiny (~123K parameter) neural propagator that, given only a protein sequence, predicts an ensemble of torsion-angle (φ, ψ) trajectories matching the marginal Ramachandran density of long-timescale molecular dynamics simulations.

On the canonical 4AA benchmark from microsoft/timewarp it produces densities that are 2.39× closer to ground-truth MD than Microsoft Research's Timewarp model (396M parameters), at roughly 3000× fewer parameters.

Quickstart

Try in your browser (no install)

👉 huggingface.co/spaces/krissss0/alphadynamics

Type a peptide sequence, get an interactive Ramachandran plot in 30 seconds.

Use locally

The simplest path — just type alphadynamics and answer the prompts:

pip install alphadynamics
alphadynamics
# Sequence (1-letter AA, e.g. AAAY): KLVFFAE
# How many independent trajectories? [16]: <Enter>
# How many timesteps per trajectory? [2500]: <Enter>
# Device (cuda/cpu/auto) [auto]: <Enter>
# Output file [...]: <Enter>
# Save Ramachandran plot? [b=both, p=PNG, h=HTML, n=no] [b]: <Enter>

Or as a one-liner for power users:

alphadynamics predict --sequence AAAY --n-ensemble 16 --rollout-steps 2500 --plot --plot-html -o traj.npz

Or from Python:

from alphadynamics import predict_torsion_ensemble

traj = predict_torsion_ensemble(
    "KLVFFAE",          # amyloid-β fragment
    n_ensemble=16,
    rollout_steps=2500,
    seed=42,
)
print(traj.shape)        # (16, 2500, 7, 2)  — ensemble × time × residues × [φ,ψ]

Headline result

Canonical Ramachandran Jensen-Shannon divergence (val-only ground truth, 36 bins, no smoothing) on the canonical 4AA test set, averaged over three peptides AAAY, AACE, AAEW:

Model	Params	Mean JSD	4AA wins	Notes
Microsoft Timewarp	396 M	0.468	0 / 3	published baseline
AlphaDynamics	123 K	0.196	3 / 3	2.39× lower, 3000× smaller

Cross-validated against Top8000 PDB statistics (Richardson Lab, Duke). Forbidden α-L region: 0.7–1.0% — at the MolProbity error level (0.5%) for real high-resolution PDB structures, meaning the model honors atomic steric exclusion.

How it works

A residue's torsion state (φ, ψ) is treated as a phase pair on a torus. Conditioned on the amino-acid identity, position, and current angles, an MLP emits per-residue oscillator parameters: an intrinsic frequency, a coupling matrix, and an anchor phase. A phase-flow ODE then integrates the joint state of 64 coupled oscillators with classical RK4 over a fixed horizon t_max=4.0 (8 substeps). The integrated phase state is decoded into a mixture of axis-independent von Mises distributions per residue, from which the next torsion frame is sampled. Rolled out autoregressively, this defines a transferable sequence-only propagator over the torsion torus.

The architecture is the protein-dynamics application of a multi-year line of work in phase-oscillator computation: REZON hardware (80 physical PCB oscillators), phase-entanglement-rc (formal phase computing), and phase-cnot-neuroscience (theta-gamma coupling validated on rat and human LFP recordings).

Files

ad_transfer_v2_clean_best.pt (310 KB) — the main propagator, ~78K params
ad_init_full_1477_best.pt (181 KB) — von Mises mixture prior, ~45K params

Honest caveats

Density only, not kinetics. Captures where the peptide spends time on the Ramachandran torus, not when transitions happen.
Backbone only. No side-chain rotamers, no Cartesian xyz, no docking.
Trained on 4–98 residues. Reliability degrades outside this range.
Monomer only. No multimer / aggregation modeling.

Best for

Quick conformational triage before launching expensive MD
Comparing peptide sequence variants / mutants
Estimating α/β/PPII basin populations
Sanity-checking before large compute commitment
AI-for-bio baselines and benchmarks
Biochemistry / biophysics teaching with interactive Ramachandran plots

Citation

@software{gwozdz2026alphadynamics,
  author  = {Gwozdz, Krzysztof},
  title   = {AlphaDynamics: Compact sequence-only neural propagator
             for protein torsion dynamics},
  year    = {2026},
  url     = {https://github.com/krisss0mecom/AlphaDynamics},
  license = {Apache-2.0},
  doi     = {10.5281/zenodo.19788564}
}

Author

Krzysztof Gwozdz — independent researcher, Poland 🇵🇱 krisss0gwo@gmail.com

This is a free contribution to the protein-dynamics community.

License

Apache 2.0. Free for any use including commercial.

krissss0
/

alphadynamics