SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials

Area of Science:

• Computational Chemistry
• Drug Discovery
• Materials Science

Background:

• Machine learning potentials (MLPs) are crucial for molecular simulations.
• Development of MLPs is limited by the availability of high-quality training datasets.
• Simulating drug-like small molecules interacting with proteins requires specialized data.

Purpose of the Study:

• Introduce the SPICE dataset, a novel quantum chemistry dataset.
• Facilitate the training of MLPs for simulating drug-protein interactions.
• Provide a valuable resource for creating transferable potential functions.

Main Methods:

• Compiled over 1.1 million conformations of small molecules, dimers, dipeptides, and solvated amino acids.
• Included 15 elements, charged/uncharged molecules, and diverse covalent/non-covalent interactions.
• Calculated forces and energies at the ωB97M-D3(BJ)/def2-TZVPPD level of theory, plus multipole moments and bond orders.

Main Results:

• Trained MLPs on the SPICE dataset achieved chemical accuracy.
• Demonstrated MLP performance across a wide chemical space.
• Validated the dataset's utility for developing accurate simulation potentials.

Conclusions:

• The SPICE dataset is a significant resource for advancing MLPs in molecular simulations.
• Enables more accurate and efficient modeling of drug-like molecules and protein interactions.
• Facilitates the creation of ready-to-use, transferable potential functions for computational chemistry.