README.md

---
license: cc-by-4.0
language:
- en
pretty_name: OpenFF AshGC v1.0
size_categories:
- 100K<n<1M
---

# Dataset Card for OpenFF AshGC v1.0 data

<!-- Provide a quick summary of the dataset. -->
This dataset includes the training, validation, and testing data for the AshGC v1.0 charge model,
a graph convolutional neural network designed to assign partial atomic charges of semi-empirical
quality to molecules for usage with Open Force Field (OpenFF) force fields.

AshGC is designed to efficiently produce conformer-independent charges of semi-empirical quality
at linear computational cost for molecules ranging from small molecules to macromolecules.

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

The OpenFF AshGC v1.0 dataset contains molecular graphs (as mapped SMILES) and
corresponding AM1-BCC ELF10 (Electrostatically Least-interacting Functional groups, 10-conformer)
partial atomic charges for training and evaluating the AshGC neural network charge model.

For training and validation purposes, datasets also include conformers and computed geometries.


The dataset comprises 271,204 molecules in the initial training set,
39,909 molecules in the fine-tuning set (small molecules < 200 Da),
61,593 molecules in the validation set,
and 45,596 molecules in the test set.
Molecules span diverse chemical space including drug-like molecules,
peptides up to 7 residues (heptamers),
and molecules with post-translational modifications.

The set of elements covered includes H, C, N, O, F, P, S, Cl, Br, I.
Molecules range from 1 to 82 heavy atoms and -6 to 8 molecular charge.

- **Curated by:** Open Force Field Initiative, Open Molecular Software Foundation
- **Funded by:** Open Force Field Initiative
- **Shared by:** Open Force Field Initiative
- **Language(s):** Not applicable (molecular data)
- **License:** CC BY 4.0


### Dataset Sources [optional]

<!-- Provide the basic links for the dataset. -->

- **Repository:** https://github.com/openforcefield/ashgc-v1.0-fit
- **Paper:** https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-597h9


## Uses

<!-- Address questions around how the dataset is intended to be used. -->


### Direct Use

<!-- This section describes suitable use cases for the dataset. -->

This dataset was used for:

1. **Training graph neural networks** to predict partial atomic charges for molecular dynamics simulations
2. **Benchmarking charge assignment methods** against AM1-BCC ELF10 reference charges
3. **Benchmarking differences between AM1-BCC backends** between AmberTools and OpenEye (the reference)



### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->

This dataset should not be used for:

- Molecules containing elements outside the supported set: H, C, N, O, F, P, S, Cl, Br, I
- Systems requiring highly accurate quantum mechanical charge distributions (the AM1-BCC charges are semi-empirical quality)
- Combined with datasets of primarily AmberTools charges, as there are systematic differences between OpenEye and AmberTools. (Note, a subset of the data here does contain AmberTools charges and may be used)
- Combined with datasets of single-conformer AM1-BCC charges, as the charges here are averaged over multiple conformers

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

The schema is:

- `mapped_smiles`: mapped SMILES
- `conformers`: all conformer geometries in Angstrom, flattened into a list
- `n_conformers`: the number of conformers in `conformers`
- `esp_lengths`: a list of the number of ESP grid points per conformer
- `am1bcc_charges`: AM1-BCC ELF10 charges
- `am1bcc_esps`: computed ESPs using AM1-BCC ELF10 charges
- `am1bcc_dipoles`: computed dipoles using AM1-BCC dipoles
- `esp_grid_inverse_distances`: the inverse distances from each atom to each ESP grid point, for ease of ESP computation

Some subsets contain AmberTools charges as well.
In these cases, the AmberTools charges are marked either by folder name (starting with `ambertools_`)
or by column name (ending with `_ambertools`).
Otherwise, charges should be assumed to originate from OpenEye's `oequacpac`.

## Dataset Creation

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->

### Curation Rationale

The AshGC dataset was created to train and benchmark the AshGC v1.0 charge model.
Please see the [preprint](https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-597h9)
for more information.

### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->

A brief summary is provided for more, but please see the paper for more information.

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

**Initial molecular graphs were collected from:**
- ChEMBL eps 78 and ZINC eps 78 ([Bleizziffer et al. 2018](https://doi.org/10.1021/acs.jcim.7b00663))
- Enamine Discovery Diversity Sets (DDS-10 and DDS-50, accessed April 2023)
- NCI Open Database (Release 4, 2012)
- PDB Ligand Expo (accessed April 2023)
- ChEMBL v33 (accessed October 2023)
- OpenFF Industry Benchmark Season 1 v1.1
- [SPICE dataset v1.1.0](https://doi.org/10.1038/s41597-022-01882-6)

**Data processing pipeline:**

1. **Molecular filtering:**
   - Most datasets filtered to 200-400 Da molecular weight
   - Elements limited to H, C, N, O, F, P, S, Cl, Br, I
   - Up to 10 rotatable bonds
   - ChEMBL eps 78 and ZINC eps 78: 250-350 Da, up to 7 rotatable bonds
   - Small-chembl: 1-199 Da (no rotatable bond filter)

2. **Protomer enumeration:**
   - Up to 2 protomers per molecule using OpenEye oequacpac (OpenEye version 2022.2.1)

3. **Conformer generation:**
   - 1000 conformers per molecule using OpenEye OMEGA
   - RMSD cutoff: 0.05 Å
   - Lowest 2% by electrostatic energy selected
   - Up to 10 diverse conformers chosen via ELF10 method

4. **Charge generation:**
   - Single-conformer AM1-BCC charges computed for each conformer
   - Charges averaged across all conformers to produce final ELF10 charges
   - Generated using both OpenEye Toolkit v2022.2.1 and AmberTools v22.0 (via OpenFF Toolkit v0.11.1)

5. **Diversity selection:**
   - Molecules labeled using hashed atom pair fingerprints (RDKit v2022.03.5, maxLength=2, 2048 bits)
   - Molecules with rare atom environments selected
   - Top 4 molecules per underrepresented environment added

6. **Train/validation/test splits:**
   - Morgan fingerprints (radius 3, 2048 bits) computed
   - Training set selected by MaxMin algorithm (Tanimoto distance)
   - Test set filtered to < 0.7 Tanimoto similarity with training/validation

**Peptide datasets:**

Two specialized peptide datasets were generated to extend coverage to larger biomolecules using RDKit v2022.03.5:

- **Tetrapeptides:** 1-4 amino acid chains from 20 natural amino acids (combinations with replacement)
- **PTMs (heptamers):** "G-V-triplet-V-G" structure with post-translational modifications (lipidation, phosphorylation, glycosylation, nitrosylation)


#### Who are the source data producers?

<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->

We drew molecular graphs from the data sources below, although charges were generated as above.

**Primary data sources:**
- ChEMBL database (EMBL-EBI)
- ZINC database and ChEMBL subset via [Bleiziffer et al 2018](https://doi.org/10.1021/acs.jcim.7b00663)
- Enamine Ltd. (commercial compound library)
- National Cancer Institute (NCI) Open Database
- Protein Data Bank (PDB) Ligand Expo
- [SPICE dataset](https://doi.org/10.1038/s41597-022-01882-6)
- OpenFF Industry Benchmark (Open Force Field Initiative)



## Citation [optional]

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

[More Information Needed]

**APA:**

[More Information Needed]


## More Information

**Related resources:**

- **Training repository:** https://github.com/openforcefield/ashgc-v1.0-fit
- **Benchmarking repository:** https://github.com/openforcefield/ash-sage-rc2
- **OpenFF Toolkit:** https://github.com/openforcefield/openff-toolkit
- **NAGL library:** https://github.com/openforcefield/openff-nagl
- **Sage 2.3.0 force field:** Available through OpenFF Toolkit

## Dataset Card Authors

- Lily Wang (Open Force Field Initiative)

## Dataset Card Contact

- **Primary contact:** info@openforcefield.org