Source-linked AI summary
SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials
Peter Eastman, Pavan Kumar Behara, David L. Dotson, Raimondas Galvelis, John E. Herr, Josh T. Horton, Yuezhi Mao, John D. Chodera, Benjamin P. Pritchard, Yuanqing Wang, Gianni De Fabritiis, Thomas E. Markland
TL;DR
The paper targets accurate, general-purpose potentials for simulating drug-like molecules interacting with proteins. It constructs SPICE from complementary subsets and trains proof-of-concept models, achieving low energy errors across the dataset while remaining scoped to conformational changes.
Problem
Accurate potentials for drug-like molecules interacting with proteins should generalize across molecules rather than be specialized to one system.
Source: Dataset Requirements · p. 2Method
SPICE combines subsets covering covalent and noncovalent interactions and trains machine learning potentials on the complete dataset as a proof of concept.
Source: Content of the SPICE Dataset · p. 5, Training a Potential Energy Model · p. 8Results
4.663 kJ/mol mean absolute error and 2.912 kJ/mol median absolute error were obtained across the complete dataset, with most conformations below 4.184 kJ/mol error.
Source: Analysis of the trained models · p. 11, Molecules with large errors · p. 15Takeaways & Limitations
The results demonstrate that SPICE can support high-quality potential functions covering a wide range of chemical space.
Source: Training a Potential Energy Model · p. 8, Molecules with large errors · p. 15The initial dataset excludes very high-energy conformations involving covalent bond formation or breaking and therefore targets conformational changes rather than chemical reactions.
Source: Dataset Requirements · p. 3
Abstract
from arXiv · showhide
Machine learning potentials are an important tool for molecular simulation, but their development is held back by a shortage of high quality datasets to train them on. We describe the SPICE dataset, a new quantum chemistry dataset for training potentials relevant to simulating drug-like small molecules interacting with proteins. It contains over 1.1 million conformations for a diverse set of small molecules, dimers, dipeptides, and solvated amino acids. It includes 15 elements, charged and uncharged molecules, and a wide range of covalent and non-covalent interactions. It provides both forces and energies calculated at the ωB97M-D3(BJ)/def2-TZVPPD level of theory, along with other useful quantities such as multipole moments and bond orders. We train a set of machine learning potentials on it and demonstrate that they can achieve chemical accuracy across a broad region of chemical space. It can serve as a valuable resource for the creation of transferable, ready to use potential functions for use in molecular simulations.
Dataset Requirements
SPICE is designed as a general-purpose potential-training dataset covering chemically relevant elements, charges, interactions, and conformations. Its requirements emphasize forces, accurate quantum chemistry, useful auxiliary properties, and continued versioned growth.
- SPICE targets general-purpose potentials for drug-like molecules interacting with proteins, rather than models specialized to one system.Source: Dataset Requirements · p. 2
- The dataset must cover diverse elements, chemical groups, charged and polar molecules, and both covalent and non-covalent interactions.Source: Dataset Requirements · p. 2
- It must sample higher-energy as well as low-energy conformations because molecular dynamics and conformational searches explore more than energy-minimized structures.Source: Dataset Requirements · pp. 2–3
- The initial scope excludes very high-energy conformations involving covalent-bond formation or breaking, focusing instead on conformational changes.Source: Dataset Requirements · p. 3
- Forces provide 3N components for an N-atom molecule, compared with one energy, substantially increasing the information available for training.Source: Dataset Requirements · p. 3
- Training data should use the most accurate practical theory and include inexpensive auxiliary quantities such as multipole moments and partial charges.Source: Dataset Requirements · p. 3
- Existing datasets commonly lack forces, charged molecules, broad elemental coverage, non-equilibrium conformations, or complete intramolecular interactions.Source: Dataset Requirements · pp. 4–5
Methods
SPICE combines complementary subsets covering covalent and non-covalent chemistry, diverse drug-like molecules, peptides, solvated amino acids, ions, and conformational variation. Energies and forces are computed at a practical high-accuracy DFT level, alongside multiple molecular properties.
Content of the SPICE Dataset
- SPICE combines subsets designed to broaden sampling across interaction types and chemical motifs.Source: Content of the SPICE Dataset · p. 5
- The dipeptide subset covers covalent interactions from the 20 natural amino acids and common protonation variants.Source: Content of the SPICE Dataset · p. 5
- The dataset samples 50 conformations per molecule, balancing low- and high-energy structures generated through molecular dynamics and partial minimization.Source: Content of the SPICE Dataset · p. 5
- Solvated amino-acid systems contain 20 nearby TIP3P-FB waters to sample protein–water and water–water interactions.Source: Content of the SPICE Dataset · p. 6
- The PubChem subset begins from approximately 1.5 million records and filters for single, non-radical molecules with 3–50 atoms and supported elements and force-field parameters.Source: Content of the SPICE Dataset · p. 6
- The non-covalent subset contributes 345,682 dimer conformations spanning a wide range of interactions and complementing primarily covalent subsets.Source: Content of the SPICE Dataset · p. 6
Quantum Chemistry Calculations
- All conformations receive DFT forces and energies at the ωB97M-D3(BJ)/def2-TZVPPD level, selected as the most accurate practical option within the computational budget.Source: Quantum Chemistry Calculations · pp. 7–8
- SPICE also stores dipole and quadrupole moments, MBIS atomic multipoles and charges, and Wiberg and Mayer bond orders.Source: Quantum Chemistry Calculations · p. 8
Data Records
SPICE is distributed as a single HDF5 data file organized by molecule or cluster, with fields describing the available properties.
- The dataset is deposited in Zenodo as SPICE-1.1.2.hdf5 and organized into one top-level group per molecule or cluster.Source: Data Records · p. 8
- Each molecule or cluster group contains data fields for the different available types of information, with detailed descriptions in the Zenodo record.Source: Data Records · p. 8
Technical Validation
The proof-of-concept models use an equivariant transformer ensemble trained on filtered SPICE data with energy and force losses. Training converges, generalizes closely from training to test conformations, and achieves mostly chemically accurate energy predictions, while errors remain associated with charges, molecular size, and unusual chemistries.
Model and training setup
- Five Equivariant Transformer models with identical hyperparameters differ only by random seed and receive minimal hyperparameter tuning.Each model uses six layers, eight attention heads, embedding width 128, a 1 nm cutoff, and 64 radial basis functions.Source: Training a Potential Energy Model · p. 9
- The model represents each unique element–formal-charge combination as one of 28 atom types, whose frequencies range from over 12 million neutral carbons to 750 positively charged phosphorus atoms.Formal-charge-based atom types provide the model's charge representation, but remain valid only when bonding patterns do not change.Source: Training a Potential Energy Model · p. 9, Table 2 · p. 9
Data preparation
- Approximately 2% of highly strained conformations are discarded, leaving 1,109,212 samples, with 55,461 randomly selected for testing.Samples are filtered when any force component exceeds 1 hartree/bohr.Source: Filtering of strained molecules · p. 10
Training behavior
- All five models largely plateau during training, suggesting convergence and limited benefit from further training.The loss combines weighted L2 energy and force errors, with weights of 1 (kJ/mol)^-2 and 1 (kJ/mol/Å)^-2, respectively.Source: Optimization of model parameters · p. 10, Figure 2 · p. 10
Validation results
- The test-loss mean is 264 versus 247 for training, corresponding to an RMS prediction error factor of 1.03 and indicating little overfitting.Across models, training loss ranges from 179 to 305 and test loss from 195 to 318.Source: Optimization of model parameters · pp. 10–11, Figure 3 · p. 11
Error analysis
- The best model has dataset-wide energy MAE 4.663 kJ/mol and median absolute error 2.912 kJ/mol, while errors tend to increase roughly linearly with molecular size.The lower median indicates that a small fraction of large-error samples raises the mean; unusual chemistries, including phosphinic acids, are also prominent among high-error molecules.Source: Analysis of the trained models · pp. 11–12, Molecules with large errors · p. 13
Competing interests
The authors disclose advisory-board memberships, equity interests, and extensive funding relationships involving scientific, pharmaceutical, and technology organizations.
- JDC serves on the Scientific Advisory Boards of four companies and holds equity interests in two of them.Source: Competing interests · p. 16