A Quantum Chemistry Dataset of 120 Kilo Organic Molecules with 20 Million Conformations
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "A Quantum Chemistry Dataset of 120 Kilo Organic Molecules with 20 Million Conformations".
Title and authors: Mira: We're talking about the paper "A Quantum Chemistry Dataset of one hundred twenty Kilo Organic Molecules with twenty Million Conformations" here, which is introducing the Quantum Open Organic Molecular database <ref:2410.19316#pg1,the Quantum Open Organic Molecular>. It highlights how essential a large scale dataset is for AI in computational chemistry research right now.
Kai: The authors, Siqi Liu and Zhijian Zhou, are addressing the lack of publicly available data that has both high molecular diversity and high conformer counts. They're basically saying existing sets are too small or too limited in what elements they cover.
Lev: The process of generating twenty million conformers by systematically rotating bonds in thirty degree increments sounds like it’s going to be a very long and resource-intensive process to get that much data right.
The paper's summary: Mira: To summarize, the Quantum Open Organic Molecular database is an open source dataset of one hundred twenty thousand organic molecules and twenty million conformers that covers ten different elements <ref:2410.19316#pg1,the Quantum Open Organic Molecular>. It’s designed for professional research in organic molecular sciences.
Kai: The authors explain that this database addresses the problem of missing large scale quantum chemistry data, which is what existing public datasets often fail to provide because they are constrained by limited elemental diversity or small sample sizes with low heavy atom counts.
Lev: They use high precision Density Functional Theory, DFT, to get those property labels of molecules which is a standard way to get the data you need for training models.
Kai: Beyond just the properties, they also provide potential energy and forces for each conformation. This allows researchers to build force fields from this data to predict molecular behavior in simulations.
Mira: And they also improved the way they generate conformations by including nearby unstable conformations for each stable one, which helps in training models to get physically plausible structures for drug design.
Lev: That extra detail about unstable conformations is interesting because it directly addresses the quality of the molecular descriptions you feed into a model, and I think that makes a big difference when you're training something robust.
Kai: So they’re giving us a whole landscape of potential energy surfaces for these molecules, which is super useful for learning how molecules behave dynamically.
The paper's improvements: Mira: The authors point out several key ways this dataset improves things. First, they provide high precision quantum mechanical data that lets deep learning models directly predict those quantum properties instead of having to solve the many-body Schrödinger equation numerically.
Lev: That’s what we want—training models on data that bypasses the need for massive, slow numerical solvers, which is a huge computational win for running simulations.
Kai: Second, they provide potential energy and forces, which means you can build force fields from this data to do molecular dynamics simulations and other studies predicting molecular behavior.
Lev: That ability to generate force field parameters from this data is significant because it extends the applicability of those mathematical models to a broader range of chemical systems.
Mira: They also improved conformation generation by including nearby unstable conformations for each stable one, which helps in training models to get physically plausible structures for drug design.
Kai: So they’re giving us more than just static structures; we’re getting the full energy landscape for these molecules, which is super useful for learning how molecules behave dynamically.
Conclusion: Mira: To wrap up on this paper, the Quantum Open Organic Molecular database provides an extensive molecular landscape with one hundred twenty thousand molecules and twenty million conformers across ten different elements <ref:2410.19316#pg1,the Quantum Open Organic Molecular>.
Kai: It gives researchers a huge resource to explore structure-property relationships because it covers a much wider range of chemical space than previous datasets, especially in terms of heavy atom counts exceeding forty.
Lev: From my perspective on hardware, the main challenge is making sure that the methods used to generate and process this data don't become a bottleneck when we try to actually run these models on real hardware.
Mira: And they show that benchmark models like GemNet perform quite well, showing good generalization capabilities even when testing them on subset B, which has those representative drug molecules.
Kai: It’s a solid foundation for training specific models, with subsets A and B allowing researchers to test things like neural network potential regression tasks or machine learning force field tasks.
Lev: But they did have to make choices about how they split the data for interpolation versus extrapolation, which is a necessary compromise when you’re setting up these benchmarks.
Mira: The QO2Mol database stands as an open source resource that aims to advance molecular modeling by providing this comprehensive and diverse data set. It opens up new avenues for building more accurate predictive models in computational chemistry, which is what we were hoping for when we looked at this work.
Kai: We'll have to keep an eye on how researchers start using this data to push the limits of what these models can actually predict before we see if it translates into practical applications.
Lev: We’ll have to keep an eye on how researchers start using this data to push the limits of what these models can actually predict before we see if it translates into practical applications.
Weiqi Liu, Xi Ai, Zhijian Zhou, Chao Qu, Junyi An, Zhipeng Zhou, Yuan Cheng, Yinghui Xu, Fenglei Cao*, ªlan Qi*
Fudan University · INFTech · ª Shanghai Academy of Artificial Intelligence for Science
physics.chem-ph
Submitted: 2024-10-25
Updated: 2026-11-18
Code: https://github.com/saiscn/QO2Mol
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The gist: Artificial intelligence is revolutionizing computational chemistry by introducing the Quantum Open Organic Molecular (QO2Mol) database…
Key concepts
- QO2Mol Database
- This is a massive open-source dataset containing 120,000 organic molecules and 20 million different three-dimensional shapes (conformers) for each molecule. It includes data on ten elements and heavy atom counts over 40, derived from real chemical compounds to support advanced molecular research.
- B3LYP/def2-SVP Level
- This is a high-precision quantum mechanical calculation method used to determine the energy and forces of molecules. It allows researchers to directly predict complex quantum properties without needing to solve difficult mathematical equations, making it ideal for training accurate machine learning models.
- Data Segmentation (Subsets A, B, C)
- The data is divided into three distinct groups: Subset A has the most conformers and is used for training models on potential energy tasks. Subset B contains representative drug molecules used to test if the trained models can accurately predict properties for new, unseen molecules. Subset C serves as a validation set.
- Conformation Generation
- This process takes a basic molecular structure and optimizes it into its lowest energy states. The authors generate 20 million conformers by systematically rotating flexible bonds at room temperature to map out the molecule's entire potential energy landscape, ensuring the data accurately reflects physical reality.
Terminology
Summary
Artificial intelligence is revolutionizing computational chemistry by introducing the Quantum Open Organic Molecular (QO2Mol) database, a large-scale quantum chemistry dataset comprising 120,000 organic molecules and approximately 20 million conformers to support professional research
Dataset Scope and Quality
The QO2Mol database is designed for professional and transformative research in organic molecular sciences under an opensource license It comprises 120,000 organic molecules and approximately 20 million conformers, encompassing 10 different elements (C, H, O, N, S, P, F, Cl, Br, I), with heavy atom counts exceeding 40 These molecules are derived from fragments of compounds in ChEMBL The database employs high-precision B3LYP/def2-SVP quantum mechanical level, each conformation was meticulously computed for quantum mechanical properties, including potential energy and forces
Data Generation Methodology
The rigorous process of data selection, processing, and preparation in QO2Mol involves several key steps
** Molecule Fragmentation: The authors first derive a set of source compounds from ChEMBL Performing quantum mechanical calculations directly on these compounds is quite challenging due to the large size of these molecules They employed a Compound Fragmentation Process, dividing the source compounds into smaller fragments containing fewer heavy atoms A total of 120,000 fragmented molecules were selected based on three rules These rules included having top 90% occurrence frequency over the database, being labeled as important phosphate groups by their chemistry expert, and encompassing 10 different elements (C, H, O, N, S, P, F, Cl, Br, I) The selection criteria did not impose restrictions on the number of heavy atoms They ensured there was no fragment duplication during the generation procedure by utilizing InChIKey and canonical SMILES identifiers **
** Conformation Generation: For each selected fragment molecule, an initial 3D structure is generated using the RDKit package Then each initial structure is optimized to a local minimum at the B3LYP/def2-SVP precision level Structure optimization involved four convergence criteria to ensure the resulting structures are reasonable Conformation search is performed on optimized structures obtained in the previous step At room temperature, rotation is the most influential factor in constructing potential energy surfaces They perform rotational search in 30-degree increments each step on all rotatable bonds of each molecule By systematically rotating the flexible bonds of molecule to specific degrees, a series of new structures are generated Following this procedure, they obtained a total of 20 million conformers for the 120,000 molecules in their database **
Impact on Molecular Modeling
The QO2Mol dataset is intended to address the lack of publicly available large-scale quantum chemistry data Its extensive coverage of molecular structures and diverse elemental composition enables comprehensive studies of structure-property relationships, enhancing the accuracy and applicability of machine learning models in predicting molecular behaviors The dataset is crucial for advancing research in three specific areas
** Potential Energy Surface: High-precision QM data is required for training reliable models that predict QM properties without solving the many-body Schrödinger equation numerically The dataset supports methods that directly predict the QM properties instead of solving the many-body Schrodinger equation numerically **
** Force Field Parameters: Force fields are mathematical models essential for molecular dynamics simulations and other computational studies that predict molecular behavior The success of such optimizations extends the applicability of the force fields to a broader range of chemical and biological systems **
** Conformation Generation: High-quality data ensures accurate molecular descriptions and physically plausible conformations, which is crucial for applications like drug design It enables the training of robust models that can generalize across diverse molecular structures **
Data Segmentation for Learning Tasks
To support various learning tasks in this field, the data is divided into three subsets with different distribution patterns
** Subset A: This main subset encompasses the most extensive conformation data, containing 20 million conformers from more than 110,000 molecules Unlike previous datasets that only sample equilibrium conformations at local minima, subset A consists of equilibrium conformations at local minima and near-equilibrium conformations additionally sampled around local minima This subset can be used for various learning tasks, such as neural network potential (NNP) regression tasks, machine learning force field (MLFF) tasks, or denoising-like pretraining tasks **
** Subset B: This subset includes 2.4 million conformers generated from approximately 1,400 molecules This subset consists of carefully selected representative drug molecules, based on domain expert annotations, with a large number of heavy atoms ranging from 30 to 34 This subset facilitates multiple tasks, such as testing the model’s extrapolative and generalization capabilities and assessing its performance in real drug design workflows **
** Subset C: The third part includes molecules that are non-analogous to those in subsets A and B Subset C can be used for potential-related tasks either as a supplementary data source combined with the training set or as a validation set Researchers have the flexibility to combine them in various ways **
Benchmark Model Performance
The potential energy prediction task is a key benchmark task where models are evaluated on the QO2Mol dataset The authors evaluate the interpolation performance of potential prediction task over a series of benchmark models on subset A Subsequently, they undertook a more challenging task of employing these trained models to predict potential energies on the subset B in order to evaluate the extrapolation capability of benchmark models The results are presented in Table 2 GemNet stands out with the lowest MAE on test set A and relatively high generalization capability on test set B, indicating exceptional performance with a moderate number of parameters Equiformer and eSCN demonstrate good performance with lower MAE, balancing parameter count and accuracy effectively
Conclusion
In this paper, they present the QO2Mol database, a open-source large-scale data resource designed for organic molecular researchs This database stands out for its distinctive composition, as it comprises 120,000 organic molecules meticulously curated from real compound libraries With representation from 10 different elements and heavy atom counts exceeding 40, the QO2Mol database offers an extensive and diverse molecular landscape for research exploration By utilizing these data, researchers can simulate and predict molecular behaviors across different chemical environments, thereby advancing organic chemistry models and theories Future research endeavors could involve leveraging the diverse and extensive molecular data within the QO2Mol database to refine and optimize machine learning applications in the field of computational chemistry New algorithms could be developed to better predict the chemical properties of complex molecules, or existing models could be improved to enhance their generalization capabilities and accuracy Over time, they hope to see the QO2Mol database continually expand, incorporating new data to support a broader range of scientific research needs
A Key Information
All the documentation for their datasets, along with usage demo scripts via Python, are provided at https://github.com/saiscn/QO2Mol The license for this work uses CC BY-NC-SA 4. p. 0 p. 00 They utilize persistent cloud storage servers to provide accessing and downloading of the dataset The B Data File Format comprises several chunk files, each containing a list of molecular data objects Table S3 summarizes the presence of ring structures in the molecules The experiment details state that all experiment was conducted on A100 GPU cluster
B Data File Format
Table S2 provides supplementary thermochemical properties at local minima The QO2Mol database comprises several chunk files, each containing a list of molecular data objects
C Chemical Space
Relative to the QM9 database, which is limited to the elements C, H, O, N, and F, QO2Mol dataset encompasses a broader range of elements commonly found in organic molecules These include C, H, O, N, S, P, F, Cl, Br, and I QO2Mol dataset comprises a significantly larger number of molecules that contain the element F
D Experiment Details
They conducted all experiment on A100 GPU cluster For the interpolation task, they employ a 72%/18%/10% split for training, validation and testing on subset A For the extrapolation task, they use the entire subset B The cutoff radius is set to 5.0 angstrom for all models The training process was conducted using the AdamW optimizer with a cosine annealing learning rate scheduler
E Reference List
The references include Mayr et al., 2023, Chen et al. p. 0, 2023, and Agüero-Chapin et al. p. 0, 2022 Stokes et al. p. 0 p. 0, 2020, Zeng et al. p. 0 p. 0, 2024 Zura˙ nski et al. p. 0 p. 0, 2023, Wang et al. p. 0 p. 0, 2023, Pereira and Trofymchuk, 2023 Lin et al. p. 0 p. 0, 2023, Ding ´ et al. p. 0 p. 0, 2024 Yang et al. p. 0 p. 0, 2020, Cheng et al. p. 0 p. 0, 2021, Dai et al. p. 0 p. 0, 2021, Bu et al. p. 0 p. 0, 2023
F Table S3
Table S3 summarizes the presence of ring structures in the molecules QO2Mol QM9 Ring Size 3 3304 54489 Ring property Aromatic (5) 28264
G Experiment Details
The experiment details state that all experiment was conducted on A100 GPU cluster For the interpolation task, they employ a 72%/18%/10% split for training, validation and testing on subset A For the extrapolation task, they use the entire subset B The cutoff radius is set to 5.0 angstrom for all models The training process was conducted using the AdamW optimizer with a cosine annealing learning rate scheduler
H A Key Information
All the documentation for their datasets, along with usage demo scripts via Python, are provided at https://github.com/saiscn/QO2Mol The license for this work uses CC BY-NC-SA 4.0 They utilize persistent cloud storage servers to provide accessing and downloading of the dataset The B Data File Format comprises several chunk files, each containing a list of molecular data objects Table S3 summarizes the presence of ring structures in the molecules The experiment details state that all experiment was conducted on A100 GPU cluster
I Table S3
Table S3 summarizes the presence of ring structures in the molecules QO2Mol QM9 Ring Size 3 3304 54489 Ring property Aromatic (5) 28264
J Conclusion
In this paper, they present the QO2Mol database, a open-source large-scale data resource designed for organic molecular researchs This database stands out for its distinctive composition, as it comprises 120,000 organic molecules meticulously curated from real compound libraries With representation from 10 different elements and heavy atom counts exceeding 40, the QO2Mol database offers an extensive and diverse molecular landscape for research exploration By utilizing these data, researchers can simulate and predict molecular behaviors across different chemical environments, thereby advancing organic chemistry models and theories Future research endeavors could involve leveraging the diverse and extensive molecular data within the QO2Mol database to refine and optimize machine learning applications in the field of computational chemistry New algorithms could be developed to better predict the chemical properties of complex molecules, or existing models could be improved to enhance their generalization capabilities and accuracy Over time, they hope to see the QO2Mol database continually expand, incorporating new data to support a broader range of scientific research needs
K References
The references include Mayr et al., 2016, Chen et al. p. 0, 2023, Agüero-Chapin et al. p. 0, 2022 Stokes et al., 2020, Zeng et al., 2024 Zura˙ nski et al., 2023, Wang et al., 2023, Pereira and Trofymchuk, 2023 Lin et al., 2023, Ding ´ et al., 2024 Yang et al., 2020, Cheng et al., 2021, Dai et al., 2021, Bu et al., 2023
L Table S3
Table S3 summarizes the presence of ring structures in the molecules QO2Mol QM9 Ring Size 3 3304 54489 Ring property Aromatic (5) 28264
M Experiment Details
The experiment details state that all experiment was conducted on A100 GPU cluster For the interpolation task, they employ a 72%/18%/10% split for training, validation and testing on subset A For the extrapolation task, they use the entire subset B The cutoff radius is set to 5.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved Potential Energy Surfaces: The QO2Mol database provides
high-precision QM data
usingB3LYP/def2-SVP quantum mechanical level,
which allows deep learning mechanism todirectly predict the QM properties instead of solving the many-body Schrodinger equation numerically.
This enables neural networks to be trained on reliable data for predicting transition states and vibrational frequencies. -
Improved Force Field Parameters: The dataset supports the creation of
Force fields
by providing potential energy and forces, which are essential formolecular dynamics simulations and other computational studies that predict molecular behavior.
This allows models to generate force field parameters that extend applicability to a broader range of chemical systems. -
Improved Conformation Generation: By including
nearby unstable conformations for each stable conformation,
the database enhances the representation of theoverall molecular potential energy landscape.
This capability supports training robust models for generating physically plausible conformations, which is crucial for applications like drug design and improving predictive accuracy. -
Enhanced Molecular Diversity Modeling: The QO2Mol dataset covers a
wider variety of 10 elements
(C, H, O, N, S, P, F, Cl, Br, I) and heavy atom counts exceeding 40 compared to previous datasets. This extensive elemental coverage enables amore comprehensive exploration of the chemical space,
allowing models to better generalize across diverse molecular structures. -
Specialized Task Training: The data is segmented into subsets A (for NNP regression and MLFF), subset B (for testing extrapolation capabilities on representative drug molecules), and subset C (as a validation set). This segmentation allows researchers to train specific models for distinct tasks, such as
neural network potential (NNP) regression tasks
ormachine learning force field (MLFF) tasks.
-
Robust Model Benchmarking: Using the QO2Mol dataset, benchmark models like GemNet show
exceptional performance with a moderate number of parameters
in subset A and demonstrate high generalization capability on subset B, indicating superior predictive power for potential energy prediction.
Sources
Related papers
- A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions
- A Machine Learning Model for the Chemistry of a Solvated Electron
- Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set
- Charge Transfer with a Spin. II: A Framework for Diabatization which Localizes Charge and Spin
- Diabatic Seam Space Sampling for Hydrogen Tunneling Systems with Nuclear-Electronic Orbital Theory
- Augmented Roothaan-Hall Hessian Applied to Spin-Restricted Open-Shell Density-Functional Theory