Molecular Properties · maomlab
Molecule3D
A dataset of roughly four million DFT-derived 3D molecular structures and quantum properties for geometric learning.
What Molecule3D is
Molecule3D is a molecular property and structure dataset containing ground-state three-dimensional geometries derived from Density Functional Theory (DFT) calculations for approximately 4 million molecules. It was introduced to benchmark algorithms designed to predict 3D molecular coordinates and spatial arrangements directly from 2D molecular graphs.
Spatial structure dictates chemical reactivity, electronic behaviour, and biological activity. By linking chemical identifiers to 3D structures and calculated quantum mechanical descriptors, Molecule3D allows computational researchers to train conformer-aware representations and geometric deep learning models at scale.
How it was made
The dataset curates raw quantum chemical calculations by standardising molecular representations using RDKit and MolVS. Unparseable SMILES strings are filtered out, and each compound is paired with its corresponding Structure-Data File (SDF) containing 3D atomic coordinates.
Each record includes structural notations, PubChem compound identifiers (CID), and calculated physical properties such as Cartesian dipole moments (x, y, and z), highest occupied molecular orbital (HOMO) energy, lowest unoccupied molecular orbital (LUMO) energy, and self-consistent field (SCF) energy. The dataset is provided in both random and scaffold split configurations divided into training, testing, and validation partitions.
Key facts
- Original creators
- Zhao Xu, Youzhi Luo, Xuan Zhang, Xinyi Xu, Yaochen Xie, Meng Liu, Kaleb Dickerson, Cheng Deng, Maho Nakata, and Shuiwang Ji
- Available splits
- Random split and scaffold split (training, validation, test)
- Calculated properties
- HOMO, LUMO, SCF energy, and Cartesian dipole components (x, y, z)
- File format
- Parquet
- Primary modalities
- Tabular and text (SMILES, SDF strings)
What you can use it for
- Predicting 3D conformers from 2D graphs
- Benchmarking machine learning architectures that infer equilibrium 3D atomic coordinates from SMILES strings or molecular connectivity graphs.
- Quantum mechanical property prediction
- Training regression models, including gradient-boosted trees and neural networks, to estimate electronic properties such as HOMO, LUMO, and SCF energy.
- Geometric deep learning pre-training
- Pre-training 3D equivariant and invariant neural networks on millions of DFT-level geometries for downstream chemical and biological tasks.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from maomlab/Molecule3D at the moment you asked. Nothing is cached or stored — every row above came from that request.
Strengths and limits
Good at
- Provides DFT-level ground-state geometries and quantum properties for nearly four million molecules.
- Supplies both random and scaffold-based evaluation splits to assess out-of-distribution generalisation.
- Pre-processed with MolVS and RDKit to sanitise and standardise chemical representations.
- Delivers complete 3D structure information via SDF data alongside tabular quantum mechanical targets.
Watch out for
- Confined to DFT-optimised ground-state conformations rather than multi-conformer conformational ensembles or excited states.
- Small numbers of compounds with invalid SMILES parseable by RDKit were removed during curation.
Questions
Can I look inside Molecule3D without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is Molecule3D under?
Its authors declare gpl-3.0. Read the terms on the original page before reusing the data.
What electronic and physical properties are recorded for each molecule?
Each molecule includes computed Cartesian dipole moments across three axes (dipole x, y, and z), HOMO energy, LUMO energy, SCF energy, and an overall target variable Y.
How are the training, validation, and test sets structured?
The dataset provides two pre-defined splitting strategies: a random split and a scaffold split.
What 3D coordinate representations are provided?
Three-dimensional atomic configurations are provided directly as SDF formatted text alongside standardised SMILES strings and PubChem CIDs.
How to cite it
Xu, Zhao, Youzhi Luo, Xuan Zhang, Xinyi Xu, Yaochen Xie, Meng Liu, Kaleb Dickerson, Cheng Deng, Maho Nakata, and Shuiwang Ji. "Molecule3D: A Benchmark for Predicting 3D Geometries from Molecular Graphs." arXiv preprint arXiv:2110.01717 (2021).
Written from Molecule3D's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.