bojuntum/CGPeptides

CG-BGs: Coarse-Grained Boltzmann Generators

Paper: arXiv:2602.10637 | Code: github.com/tummfm/cg-bg

CG-BGs are flows trained on MD trajectories that generate coarse-grained molecular conformations, combined with a PMF model for importance reweighting to recover exact equilibrium statistics.

Molecules

Directory	System	Solvent
MB/	Mueller-Brown 2D toy potential	—
Ac-Ala-NHMe/	Alanine dipeptide (Ala2)	explicit + implicit
Ac-Ala3-NHMe/	Alanine tripeptide (Ala3)	explicit + implicit
Ac-Ala6-NHMe/	Alanine hexapeptide (Ala6)	explicit + implicit

Directory Structure

<molecule>/
  implicit/
    data.npz              # implicit-solvent baseline trajectory
    *.pdb                 # reference structure for the all-atom structure
  explicit/
    <cg_mapping>/
      <pmf_type>/
        <flow_type>/
          data.npz
          energy_params.pkl
          flow_params.pkl
          samples_and_weights.npz
      *.pdb               # reference structure for this CG mapping

CG mappings (<cg_mapping>)

For MB the <cg_mapping> level is omitted and the tree starts directly at pmf_type/flow_type/.

Name	Description	Nodes (Ala2 / Ala3 / Ala6)
heavy_atom	all non-hydrogen atoms	10 / 20 / —
core_beta	backbone + Cβ atoms	6 / 12 / 24

PMF types (<pmf_type>)

Name	Description
pmf_ub	PMF model trained on unbiased MD data
pmf_b	PMF model trained on WT-MetaD ($\gamma=9$) data

pmf_ub exists only for Ac-Ala-NHMe/explicit/heavy_atom/.

Flow types (<flow_type>)

Name	Description
flow_ub	Boltzmann generator trained on unbiased MD data
flow_b	Boltzmann generator trained on WT-MetaD ($\gamma=1.5$) data

Files

data.npz — training trajectory

MB:

Array	Shape	Description
R	(N, 2)	positions

Alanine explicit:

Array	Shape	Description
R	(N, n_nodes, 3)	CG positions (Å)
F	(N, n_nodes, 3)	CG forces
box	(N, 3, 3)	simulation box vectors
species	(N, n_nodes)	atom type indices
mask	(N, n_nodes)	valid-node mask
U	(N,)	potential energies

Alanine implicit:

Array	Shape	Description
R	(N, n_nodes, 3)	positions
F	(N, n_nodes, 3)	forces
box	(N, 3, 3)	box vectors
species	(N, n_nodes)	atom type indices
mask	(N, n_nodes)	valid-node mask
id	(N,)	frame identifier
r0	(N,)	umbrella restraint center
window	(N,)	umbrella window index
subset	(N,)	data subset label

energy_params.pkl — PMF model parameters

Pickled JAX parameter pytree for the trained PMF/energy model. Loaded by Stage 3 to evaluate energies and compute importance weights.

flow_params.pkl — normalizing flow parameters

Pickled JAX parameter pytree for the trained Boltzmann generator. Produced by Stage 1; loaded by Stage 2 for sampling.

samples_and_weights.npz — generated samples with importance weights

Output of Stages 2–3. Same spatial arrays as data.npz plus:

Array	Description
logp	log proposal probability from the flow
U	potential energy from the PMF model
logw	log importance weight (logw = -U/kT - logp)

Usage with cg-bg

Each leaf directory corresponds to one experiment config in configs/experiment/. The mapping is:

Experiment	Path
mb_ub	MB/pmf_b/flow_ub/
mb_b	MB/pmf_b/flow_b/
ala2_ha_ub	Ac-Ala-NHMe/explicit/heavy_atom/pmf_b/flow_ub/
ala2_ha_b	Ac-Ala-NHMe/explicit/heavy_atom/pmf_b/flow_b/
ala2_cb_ub	Ac-Ala-NHMe/explicit/core_beta/pmf_b/flow_ub/
ala2_cb_b	Ac-Ala-NHMe/explicit/core_beta/pmf_b/flow_b/
ala3_ha_ub	Ac-Ala3-NHMe/explicit/heavy_atom/pmf_b/flow_ub/
ala3_cb_ub	Ac-Ala3-NHMe/explicit/core_beta/pmf_b/flow_ub/
ala6_cb_ub	Ac-Ala6-NHMe/explicit/core_beta/pmf_b/flow_ub/

Files are fetched on demand via the ${hf:repo_id,path} Hydra resolver; no manual download is required. Set HF_REPO_ID in .env to point at a private fork.