DynaProt

Checkpoints for "Learning residue level protein dynamics with multiscale Gaussians" (Mihir Bafna, Bowen Jing, Bonnie Berger — CSAIL, MIT).

Folder	Model	Params	Output
dynaprot-M	DynaProt-M (marginal)	955K	per-residue 3×3 covariances — anisotropic flexibility, and RMSF as sqrt(Tr(Σᵢ))
dynaprot-J	DynaProt-J (joint)	1.9M	N×N scalar residue–residue coupling matrix

Both were trained on the ATLAS MD dataset (3 × 100 ns per protein) using the AlphaFlow train/val/test split, with no large-scale PDB pretraining.

Usage

pip install "dynaprot[hub] @ git+https://github.com/MihirBafna/dynaprot"
dynaprot predict my_protein.pdb -o out/ --num-samples 250

Weights download automatically on first use. In Python:

from dynaprot.inference import DynaProtPredictor

predictor  = DynaProtPredictor.from_pretrained(device="cuda:0")
prediction = predictor.predict("my_protein.pdb", chain_id="A")

prediction.rmsf                  # (N,)     per-residue flexibility, Å
prediction.marginal_covariances  # (N,3,3)  anisotropic Gaussian blobs, global frame
prediction.correlation           # (N,N)    residue-residue coupling
prediction.joint_covariance      # (3N,3N)  composed joint covariance
prediction.sample_ensemble(250)  # (250,N,3)

Citation

@inproceedings{bafna2026dynaprot,
  title     = {Learning residue level protein dynamics with multiscale Gaussians},
  author    = {Bafna, Mihir and Jing, Bowen and Berger, Bonnie},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}