mdRNAcat: A Molecular Dynamics RNA Catalog of REST2 Conformational Ensembles

This dataset comprises force-labelled, all-atom MD trajectories from 48 RNA systems. Simulations were performed using GROMACS and parameterized using the classical force field amber99sb*-ildn with parmbsc0 corrections and DESRES RNA parameters and a Replica Exchange with Solute Tempering 2 (REST2) scheme with 32 replica per system from 300 to 500 K.

Analysis and Machine-Readable Metadata

A git repo with analysis scripts is available here: https://github.com/SDLAML/mdRNAcat.

The dataset includes a croissant.json file containing machine-readable metadata in the MLCommons Croissant format.

The metadata describes the dataset, its distribution, structure, and provenance. The accompanying GitHub repository contains in-depth analysis of the dataset, including both simulation metadata and coverage of sequence and structural space.

Dataset Size

Total RNA systems	48
Systems with 200 ns simulation time	45
Systems with 300 ns simulation time	3
Replicas	32
State trajectories	1,536
Total sampled time	316.8 µs
Total sampling at 300 K	9.9 µs
Average atoms per system	1,955
Average residues per system	61
Total file size	699 GB

Getting Started with an Example

To load the dataset simulation trajectories, simply install the HF cli tool huggingface_hub, optionally also hf-xet, and use it to download single RNA system files:

# Install the Hugging Face client and HDF5 reader
pip install -U huggingface_hub hf-xet h5py

# Download a single RNA system from the dataset
hf download sdlaml/mdrnacat \
    6qn3.h5 6qn3_replica.h5 \
    --repo-type dataset \
    --local-dir ./mdrnacat

The RNA coordinates and forces stored in the <pdbid>.h5 and <pdbid>_replica.h5 files are accessible using the following code snippet.

import h5py
import numpy as np


def get_state(root, target_temperature=300.0):
    """Find a REST2 state by its effective temperature.

    The group names are discovered rather than assuming
    a particular string representation of temperature.
    """
    states = []

    for name, group in root.items():
        if not isinstance(group, h5py.Group):
            continue
        if "0" not in group:
            continue

        traj = group["0"]
        if "temperature" in traj.attrs:
            states.append(traj)

    if not states:
        raise ValueError("No REST2 trajectory groups found")

    return min(
        states,
        key=lambda g: abs(
            float(g.attrs["temperature"]) - target_temperature
        ),
    )


# -------------------------------------------------------
# 1. Load the densely sampled reference-state trajectory
# -------------------------------------------------------

pdb_id = "6qn3"

with h5py.File(f"mdrnacat/{pdb_id}.h5", "r") as h5:
    root = h5[pdb_id]
    traj = get_state(root, 300.0)

    # Read metadata without loading the trajectory.
    n_frames, n_atoms, _ = traj["coords"].shape
    stride_ps = float(traj.attrs["frameStridePs"])

    print(f"System: {pdb_id}")
    print(f"Atoms: {n_atoms}")
    print(f"Frames: {n_frames}")
    print(f"Frame interval: {stride_ps} ps")
    print(f"Temperature: {traj.attrs['temperature']} K")

    # Load only the first 10 frames.
    # Shapes: (10, n_atoms, 3)
    coords = traj["coords"][:10]
    forces = traj["forces"][:10]

    # Obtain the per-atom chemical element and residue.
    elements = root["element"].asstr()[:]
    residues = root["resname"].asstr()[:]

    print("Coordinates:", coords.shape)
    print("Forces:", forces.shape)
    print("First atoms:", elements[:5])
    print("First residues:", residues[:5])

    # Units:
    # coords: Å
    # forces: kcal / (mol Å)


# -------------------------------------------------------
# 2. Load reference-Hamiltonian forces from REST2 states
# -------------------------------------------------------

with h5py.File(f"mdrnacat/{pdb_id}_replica.h5", "r") as h5:
    root = h5[pdb_id]

    # Select the highest effective-temperature state.
    traj = get_state(root, 500.0)

    coords = traj["coords"][:10]

    # Forces from the scaled REST2 Hamiltonian.
    forces_rest2 = traj["forces"][:10]

    # Forces reevaluated under the unscaled Hamiltonian.
    forces_reference = traj["scaledForces"][:10]

    temperature = float(traj.attrs["temperature"])

    print(f"\nEffective temperature: {temperature} K")
    print("Coordinates:", coords.shape)
    print("REST2 forces:", forces_rest2.shape)
    print("Reference forces:", forces_reference.shape)

    # Inspect the effective-temperature ladder.
    temperatures = root["repex"]["temps"][:]
    print("REST2 temperature ladder:", temperatures)

Data Structure and HDF5 Schema

mdRNAcat is distributed in HDF5 format, following a structure similar to mdCATH to facilitate interoperability with existing molecular machine-learning workflows.

HDF5 Hierarchy

Each RNA system is stored in two HDF5 files:

• <pdbid>.h5: Densely sampled 300 K reference-state trajectory (1 ps resolution).
• <pdbid>_replica.h5: All 32 REST2 thermodynamic states, with additional annotations and replica-exchange metadata (100 ps resolution).

The following diagram illustrates the organization using PDB entry 6qn3 as an example:

6qn3.h5
└── 6qn3/
    ├── chain, element, resid, resname, z
    ├── pdb, top
    ├── [root attributes]
    │   ├── numProteinAtoms, numResidues
    │   ├── numReplicas
    │   └── rest2TempMin, rest2TempMax
    └── <T = 300 K>/
        └── 0/
            ├── coords             [F, N, 3]
            ├── forces             [F, N, 3]
            ├── box                [3, 3]
            └── [trajectory attributes]
                ├── temperature
                ├── numFrames
                └── frameStridePs

6qn3_replica.h5
└── 6qn3/
    ├── chain, element, resid, resname, z
    ├── pdb, top
    ├── [root attributes]
    ├── repex/
    │   ├── terms
    │   ├── energy, energyTime, energyNSamples
    │   ├── replicaIndex, replicaIndexTime
    │   ├── nAttempts, nExchanges
    │   └── temps
    ├── <T = 300 K>/
    │   └── 0/
    │       ├── coords             [F, N, 3]
    │       ├── forces             [F, N, 3]
    │       ├── scaledForces       [F, N, 3]
    │       ├── gyrationRadius     [F]
    │       ├── rmsd               [F]
    │       ├── rmsf               [R]
    │       ├── secondaryStructure [F, R]
    │       ├── sugarPucker        [F, R]
    │       ├── box                [3, 3]
    │       └── [trajectory attributes]
    ├── <T = ...>/
    │   └── 0/
    │       └── ...
    └── <T = 500 K>/
        └── 0/
            └── [same trajectory schema]

Here, N denotes atoms, R residues, and F saved frames. The <T> labels denote REST2 effective temperatures, not physical thermostat temperatures.

The actual temperature-group names can be discovered programmatically, as demonstrated below.

File Types and Contents

Each RNA system is identified by its PDB ID and has two associated HDF5 files:

File	Contents	Temporal resolution
<pdbid>.h5	Reference-state (300 K) coordinates and instantaneous forces	1 ps
<pdbid>_replica.h5	All 32 REST2 thermodynamic states, forces, structural annotations, and replica-exchange metadata	100 ps

For example, 6qn3.h5 contains the densely sampled reference-state trajectory, while 6qn3_replica.h5 contains configurations from all REST2 states.

All files use the PDB identifier as their root group. Within that group, trajectory data are organized by REST2 effective temperature (<T>/0/), while replica-exchange records are stored under repex/.

Root-level Metadata

The following datasets and attributes are stored at <pdbid>/.

Field	Shape	Type	Description
chain	N	String	Chain identifier
element	N	String	Chemical element
resid	N	Integer	Residue number
resname	N	String	Residue name
z	N	Integer	Atomic number
pdb	1	String	RNA reference PDB
top	1	String	GROMACS topology

Root-group attributes

Attribute	Type	Description
numProteinAtoms	Integer	Number of RNA atoms (N)
numResidues	Integer	Number of residues (R)
numReplicas	Integer	Number of REST2 states (M)
rest2TempMin	Float	Minimum REST2 effective temperature (K)
rest2TempMax	Float	Maximum REST2 effective temperature (K)

The numProteinAtoms attribute name is retained for compatibility with the mdCATH-style schema; here it refers to RNA atoms.

Trajectory Data

Each effective-temperature group contains a trajectory subgroup at <pdbid>/<T>/0/.

Field	Shape	Type	Unit	Description
coords	F × N × 3	Float	Å	Atomic coordinates
forces	F × N × 3	Float	kcal/mol/Å	Instantaneous atomic forces
scaledForces*	F × N × 3	Float	kcal/mol/Å	Forces recomputed using the unscaled (λ = 1) reference Hamiltonian
gyrationRadius*	F	Double	nm	Radius of gyration
rmsd*	F	Float	nm	RMSD relative to the first frame
rmsf*	R	Float	nm	Per-residue RMSF
secondaryStructure*	F × R	String	—	Base-pairing dot-bracket annotations
sugarPucker*	F × R	Float	degrees	Ribose pseudorotation phase
box	3 × 3	Float	nm	Simulation unit cell

Fields marked with an asterisk are included in the all-state _replica.h5 files.

Trajectory-group attributes

Attribute	Type	Unit	Description
temperature	Float	K	REST2 effective temperature
numFrames	Integer	—	Number of saved frames (F)
frameStridePs	Float	ps	Time between saved frames

Replica-exchange Metadata

The _replica.h5 files additionally contain a <pdbid>/repex/ group documenting the REST2 exchange process.

Field	Shape	Type	Unit	Description
terms	K	String	—	Interaction-energy term names
energy	M × S × K	Float	kJ/mol	Per-state interaction energies
energyTime	M × S	Float	ps	Energy sampling times
energyNSamples	M	Integer	—	Valid energy samples per state
replicaIndex	X × M	Integer	—	Replica occupancy for each thermodynamic state
replicaIndexTime	X	Float	ps	Exchange-record times
nAttempts	P	Float	—	Exchange attempts per log segment
nExchanges	P × (M − 1)	Float	—	Accepted exchanges between adjacent states
temps	M	Float	K	REST2 effective-temperature ladder

Dimension notation

• N: Number of RNA atoms
• R: Number of RNA residues
• F: Number of saved trajectory frames
• M: Number of REST2 thermodynamic states
• K: Number of interaction-energy terms
• S: Number of energy-sampling positions
• X: Number of replica-exchange records
• P: Number of simulation-log segments

Important Considerations

REST2 thermodynamic states: The 300–500 K values describe effective temperatures corresponding to Hamiltonian scaling. All simulations use a physical thermostat temperature of 300 K. Higher effective-temperature states sample modified Hamiltonians.

Reference-Hamiltonian forces: The scaledForces arrays contain forces reevaluated under the unscaled reference Hamiltonian for configurations obtained from other REST2 states. These configurations can be useful for force-based machine learning, but their sampling distributions do not automatically correspond to the reference-state Boltzmann distribution. Appropriate statistical treatment may be necessary for thermodynamic applications.

Units: mdRNAcat intentionally follows different unit conventions for different observables. Coordinates and forces follow mdCATH conventions (Å and kcal/mol/Å), while structural observables and replica-exchange quantities retain their specified simulation-tool conventions. Check the units in the tables above before combining fields.

For further information on data generation, simulation parameters, and analysis methods, refer to the accompanying mdRNAcat manuscript.

License and Citation

The mdRNAcat molecular simulation dataset is released under the Creative Commons Attribution 4.0 International License.

The accompanying software and analysis utilities are distributed separately under the MIT License.

If you use mdRNAcat in your research, please cite the associated mdRNAcat publication. Citation information is provided in CITATION.cff and the BibTeX entry below [WIP].

Please cite the dataset's specific version when reproducibility depends on the released configurations or force data.