Molecule3D

Molecule3D is a comprehensive dataset containing ground-state geometries derived from Density Functional Theory (DFT) calculations for approximately 4 million molecules. This is a mirror of the Official Github repo where the dataset was uploaded in 2021.

Preprocseeing

[Update on 2025.08.16 - version 1.1.1] We removed invalid SMILES strings which could not be parsed by RDKit.

• Random split

1. train : removed 60 smiles strings from 2339788 strings
2. test : removed 35 smiles strings from 779930 strings
3. validation : removed 26 smiles strings from 779929 strings

• Scaffold split

1. train : removed 46 smiles strings from 2339788 strings
2. test : removed 71 smiles strings from 779930 strings
3. validation : removed 4 smiles strings from 779929 strings

We also updated the README.md file.

We utilized the raw data uploaded on Github and performed several preprocessing:

1. Sanitize the molecules using RDKit and MolVS (standardize SMILES format)
2. Combine the SMILES strings, SDF data, and 3D molecular properties for each molecule.
3. Split the dataset using random split and scaffold split (train, test, validation)

If you would like to try these processes with the original dataset, please follow the instructions in the Preprocessing Script file located in our Molecule3D repository.

Quickstart Usage

Load a dataset in python

Each subset can be loaded into python using the Huggingface datasets library. First, from the command line install the datasets library

$ pip install datasets

then, from within python load the datasets library

>>> import datasets

and load one of the Molecule3D datasets, e.g.,

>>> Molecule3D = datasets.load_dataset('maomlab/Molecule3D', name = 'Molecule3D_random_split')   # can put 'Molecule3D_scaffold_split' for the name as well
README.md: 100%  4.95k/4.95k [00:00<00:00, 559kB/s]
Generating train split: 100%  2339788/2339788 [00:34<00:00, 85817.85 examples/s]
Generating test split: 100%  779930/779930 [00:15<00:00, 96660.33 examples/s]
Generating validation split:  100% 779929/779929 [00:09<00:00, 79064.99 examples/s]

and inspecting the dataset

>>> Molecule3D
DatasetDict({
train: Dataset({
    features: ['index', 'SMILES', 'sdf', 'cid', 'dipole x', 'dipole y', 'dipole z', 'homo', 'lumo', 'Y', 'scf energy'],
    num_rows: 2339788
})
test: Dataset({
    features: ['index', 'SMILES', 'sdf', 'cid', 'dipole x', 'dipole y', 'dipole z', 'homo', 'lumo', 'Y', 'scf energy'],
    num_rows: 779930
})
validation: Dataset({
    features: ['index', 'SMILES', 'sdf', 'cid', 'dipole x', 'dipole y', 'dipole z', 'homo', 'lumo', 'Y', 'scf energy'],
    num_rows: 779929
})

})

Use a dataset to train a model

One way to use the dataset is through the MolFlux package developed by Exscientia. First, from the command line, install MolFlux library with catboost and rdkit support

pip install 'molflux[catboost,rdkit]'

then load, featurize, split, fit, and evaluate the catboost model

import json
from datasets import load_dataset
from molflux.datasets import featurise_dataset
from molflux.features import load_from_dicts as load_representations_from_dicts
from molflux.splits import load_from_dict as load_split_from_dict
from molflux.modelzoo import load_from_dict as load_model_from_dict
from molflux.metrics import load_suite

split_dataset = load_dataset('maomlab/Molecule3D', name = 'Molecule3D_random_split')  # can put 'Molecule3D_scaffold_split' for the name as well

split_featurised_dataset = featurise_dataset(
  split_dataset,
  column = "SMILES",
  representations = load_representations_from_dicts([{"name": "morgan"}, {"name": "maccs_rdkit"}]))

model = load_model_from_dict({
    "name": "cat_boost_regressor",
    "config": {
        "x_features": ['SMILES::morgan', 'SMILES::maccs_rdkit'],
        "y_features": ['Y']}})

model.train(split_featurised_dataset["train"])
preds = model.predict(split_featurised_dataset["test"])

regression_suite = load_suite("regression")

scores = regression_suite.compute(
    references=split_featurised_dataset["test"]['Y'],
    predictions=preds["cat_boost_regressor::Y"])    

Citation

@misc{https://doi.org/10.48550/arxiv.2110.01717, doi = {10.48550/ARXIV.2110.01717}, url = {https://arxiv.org/abs/2110.01717}, author = {Xu, Zhao and Luo, Youzhi and Zhang, Xuan and Xu, Xinyi and Xie, Yaochen and Liu, Meng and Dickerson, Kaleb and Deng, Cheng and Nakata, Maho and Ji, Shuiwang}, keywords = {Machine Learning (cs.LG), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences}, title = {Molecule3D: A Benchmark for Predicting 3D Geometries from Molecular Graphs}, publisher = {arXiv}, year = {2021}, copyright = {arXiv.org perpetual, non-exclusive license} }