ANI-1: A data set of 20M off-equilibrium DFT calculations for organic molecules

Justin S. Smith, Olexandr Isayev, Adrián E. RoitbergarXiv 2017 · DOI · arXiv

Abstract

A data set of more than 20 million density function theory calculated energies for non-equilibrium conformations from 57k small molecules. The molecules included contain from one to eight C, N, and O atoms. The non-equilibrium conformations were generated using a normal mode sampling scheme.

Introduction: Shaking Up Quantum Chemistry

In theoretical chemistry, a persistent "speed vs. accuracy" trade-off exists. Density Functional Theory (DFT) is the standard for providing accurate electronic properties, but its computational cost scales cubically ($O(N^3)$) with the number of electrons. This scaling makes simulating large proteins or long-term materials behavior computationally prohibitive.

Machine Learning (ML) offers a potential solution: training neural networks to "learn" the potential energy surface of molecules to achieve DFT-level accuracy at the speed of classical force fields. However, ML models require vast amounts of data. Previous benchmarks, such as QM9, focused on "equilibrium" structures—the lowest-energy states. Since real molecules are dynamic, building useful ML potentials requires data on non-equilibrium conformations. The ANI-1 dataset addresses this need by providing over 20 million high-fidelity calculations for off-equilibrium geometries.

Methods: The Generation Pipeline

The ANI-1 dataset was generated using a systematic five-step pipeline designed to sample the conformational space of small organic molecules.

The ANI-1 data generation pipeline, from molecule selection to final DFT energy calculation.
  1. Selection: 57,462 organic molecules containing C, N, and O with up to 8 heavy atoms were selected from the GDB-11 database.
  2. Initial 3D Modeling: RDKit generated 3D conformations, which were pre-optimized using the MMFF94 classical force field.
  3. DFT Optimization: The $\omega$B97x functional was used to find the "true" equilibrium state for each molecule.
  4. Normal Mode Sampling (NMS): To move beyond equilibrium structures, the authors calculated vibrational modes and displaced atoms along these paths at various simulated temperatures. This "shaking" process generated a wide range of off-equilibrium shapes.
  5. Final Calculation: Single-point DFT calculations were performed on approximately 20 million perturbed shapes to determine their exact total electronic energies.

Datasets and Scale

The resulting ANI-1 dataset is significantly larger and more diverse than its predecessors:

  • ANI-1: Comprises 20,057,374 "low energy" conformations (within 275 kcal/mol of the minimum) and 2,630,435 "high energy" conformations.
  • Comparison: ANI-1 is roughly 100x larger than benchmarks like QM9 and covers a much broader conformational space.
Distribution of total energies and populations across different GDB subsets and energy ranges.

Key Insights: Transferability and Local Environments

A central thesis of this work is that the physics of large, complex systems can be learned from the local chemical environments of small molecules. By using Normal Mode Sampling, the dataset captures the "slopes" of the potential energy surface, which is critical for calculating forces in molecular dynamics simulations. Unlike datasets providing long trajectories for single molecules, ANI-1 provides a massive variety of chemical environments, enhancing the transferability of trained models to unseen molecules.

Key Results

  • Scale and Coverage: With over 20 million data points, ANI-1 represents a massive leap forward. Even a 1% subsample covers vast areas of atomic distance and angular space missed by equilibrium-only datasets.
  • Predictive Power: A neural network potential trained on this data achieved an RMSE of just 0.57 kcal/mol compared to DFT on a 10-heavy-atom test set.
  • Generalization: The model successfully predicted energies for molecules with up to 50 atoms, despite being trained only on molecules with 8 or fewer heavy atoms.

Limitations and Future Work

While ANI-1 is a foundational resource, it currently lacks elemental diversity, being limited to H, C, N, and O. Future iterations will need to include elements like Sulfur, Phosphorus, and halogens. Additionally, while the $\omega$B97x functional is highly accurate, future datasets could aim for even higher levels of theory, such as CCSD(T), as computational resources expand.

Summary updated

Abstract

A data set of more than 20 million density function theory calculated energies for non-equilibrium conformations from 57k small molecules. The molecules included contain from one to eight C, N, and O atoms. The non-equilibrium conformations were generated using a normal mode sampling scheme.
BibTeX
@misc{smith2017ani,
  author = {Justin S. Smith and Olexandr Isayev and Adrián E. Roitberg},
  title = {{ANI-1: A data set of 20M off-equilibrium DFT calculations for organic molecules}},
  year = {2017},
  eprint = {1708.04987},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1708.04987},
}
Figure