---
license: mit
tags:
- chemistry
- biology
---


This repository contains the dataset used in the paper: **"Scalable Machine Learning Force Fields for Macromolecular Systems Through Long-Range Aware Message Passing"** [[link](https://arxiv.org/abs/2601.03774)].

### Computational Details
All data in this repository were generated using Density Functional Theory (DFT) calculations to ensure the physical accuracy for machine learning training.

* Functional: The ωB97X-D3 range-separated hybrid functional was employed, which includes empirical dispersion corrections to accurately capture long-range van der Waals interactions.

* Basis Set: The def2-SVP basis set was used for all atomic species, providing a robust balance between computational efficiency and electronic structure accuracy.

* System Scale: The dataset covers a wide range of molecular sizes, featuring atomic systems with up to 1,200 atoms, making it uniquely suited for developing scalable models for macromolecular systems.

* Data Samples: 677,753.

### Dataset Components
The dataset consists of three primary compressed archives, each catering to different aspects of macromolecular force field modeling:

1. deshaw_protein.tar.gz

  This subset contains protein structures and conformational data extracted from the DE Shaw Research molecular dynamics trajectories [1]. It provides high-fidelity biological samples essential for evaluating the model's ability to generalize across complex protein folding and fluctuation landscapes.
  * Average molecule size: 1065
  * Samples: 42,763

2. di_molecule_interaction.tar.gz

  This dataset focuses on non-bonded interactions. It was constructed by systematically increasing the distance between two distinct molecules (dimers). It is specifically designed to benchmark long-range interaction modeling, capturing how energy and forces decay as a function of intermolecular separation.
  * Average molecule size: 79
  * Samples: 504,990

3. md_traj.tar.gz

  This file contains various molecular dynamics (MD) trajectories generated in-house. It includes a diverse set of chemical systems (e.g., NaCl in water) used to train and validate the model's stability and accuracy in simulating temporal evolution and thermodynamic properties.
  * Average molecule size: 525
  * Samples: 130,000

### How-to Load
The data is stored in **LMDB** format. Using the `md_traj` data as an example, you can load the data as follows:

```python
import lmdb
import pickle

# Open the LMDB environment
env = lmdb.open("md_traj/train/NaCl/data_0.lmdb", readonly=True, lock=False, subdir=False)

with env.begin() as txn:
    length = txn.stat()['entries']
    data_list = []
    for idx in range(length):
        byte_data = txn.get(f"{idx}".encode())
        if byte_data:
            data_list.append(pickle.loads(byte_data))

# See the keys stored in data
if data_list:
    print(data_list[0].keys())

```

### Data Structure

The loaded objects contain the following keys:

* `forces`: Atomic forces with shape \[N,3\], in kcal/mol/Angstrom.
* `cluster_ids`: IDs for each atom within the prebuilt clusters.
* `order`: The frame index in the MD simulation.
* `pos`: Atomic positions with shape \[N,3\], in units of Angstrom
* `cluster_centers`: Geometric centers of the prebuilt clusters.
* `energy`: Total molecular energy in kcal/mol/Angstrom.
* `atomic_numbers`: List of atomic numbers with shape \[N\].

For the **Di-Molecule** dataset, the following additional properties are included:

* `mol_a`: The PubChem ID of molecule A.
* `mol_b`: The PubChem ID of molecule B.
* `distance`: The horizontal distance between the right-most atom of A and the left-most atom of B, in units of Angstrom.

For further details, please refer to the [original paper](https://arxiv.org/abs/2601.03774) and the official repository: [IQuestLab/UBio-MolFM](https://github.com/IQuestLab/UBio-MolFM).

[1]: Lindorff-Larsen, et al. "How fast-folding proteins fold." Science 334.6055 (2011): 517-520.