abpark/transferable-cg

Dataset Overview

The data repository contains four trajectory datasets plus metadata files for train/test splits:

transferable-cg/
├── MFM/
├── MFM_100K/
├── MFM_test/
├── SM/
├── FM/
├── training_cath_1000.npy
└── test_cath_50.npy

Dataset Descriptions

Dataset Construction

All simulations used structures from 1000 CATH domains sourced from the mdCATH dataset, with 20 initial configurations per domain sampled from 320 K replica trajectories. Atomistic simulations were run with OpenMM 8.1.1 using the AMBER ff99SBdisp force field, TIP3P solvent (9 Å buffer, 150 mM salt), the OVRVO integrator at 300 K (2 fs timestep, γ = 1 ps⁻¹).

• —`FM`: Each configuration ran 100 ps equilibration + 3 ns production without hydrogen bond constraints; frames were saved every 10 ps (stored in {CATH_ID}_traj_all.hdf5), with every 60 ps frame selected for training (indexed by stride_60ps_indices.npy).
• —`SM`: Each configuration ran 5 ns equilibration + 20 ns production with hydrogen bond constraints, saving frames every 10 ps; 1,000 frames were randomly subsampled per domain for training (indexed by subset_1000_cg.npy).
• —`MFM`: Backbone Cα/C/N atoms were held fixed (zero mass) without H-bond constraints; simulations terminated once per-bead mean force standard error fell below 1 k_BT.
• —`MFM_100K`: Extended MFM with 80 additional structures per domain sampled from 348–450 K mdCATH trajectories (mean forces estimated at 300 K).
• —`MFM_test`: Same mean force protocol applied to 50 held-out CATH domains (< 40% sequence similarity to training), with structures drawn from all five mdCATH temperatures

Important Notes on Mean Force Datasets

⚠️ For `MFM`, `MFM_100K`, and `MFM_test` datasets: - Forces are recorded only for backbone atoms (Cα, C, N) of each residue and represent the estimated mean force on the Cα, C, and N for each structure - Forces for all other atoms are set to zero and should be ignored - Only positions/forces for Cα, C, and N atoms contain valid data

File Descriptions

File Structure

Each dataset folder contains subfolders organized by CATH domain ID:

{Dataset}/
└── {CATH_ID}/
    ├── {CATH_ID}_traj_all.hdf5
    └── {CATH_ID}.pdb
    └── subset_*.npy        # Only in SM/ and FM/

Per-Domain Files

HDF5 File Contents

Datasets

Each HDF5 file contains the following trajectory data:

Attributes (Metadata)

Each HDF5 file includes the following attributes:

For details on usage see docs/tutorials/training_dataset.md at https://github.com/rotskoff-group/transferable-cg.