Nutmeg and SPICE: Models and Data for Biomolecular Machine Learning.

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Machine learning potentials (MLPs) require large datasets of quantum chemistry calculations for training.
• Existing datasets may lack sufficient sampling of chemical space or detailed data on noncovalent interactions.
• Accurate modeling of charged and polar molecules presents a significant challenge for MLPs.

Purpose of the Study:

• To introduce version 2 of the SPICE dataset, expanding chemical space sampling and noncovalent interaction data.
• To train and evaluate new MLPs, named Nutmeg, based on the TensorNet architecture using the enhanced dataset.
• To develop and assess a novel mechanism for improving MLP performance on charged and polar molecules.

Main Methods:

• Expansion of the SPICE dataset with increased chemical space sampling and noncovalent interaction data.
• Training of Nutmeg potential energy functions utilizing the TensorNet architecture.
• Implementation of a novel mechanism involving precomputed partial charges to guide MLPs for charged and polar molecules.

Main Results:

• Nutmeg models demonstrate high accuracy in reproducing energy differences between molecular conformations, even for highly charged or large molecules.
• The trained models produce stable molecular dynamics trajectories.
• The computational speed of the Nutmeg models is suitable for routine simulations of small molecules.

Conclusions:

• The enhanced SPICE dataset and the developed Nutmeg models represent a significant advancement in machine learning for quantum chemistry.
• The novel charge injection mechanism effectively improves the performance of MLPs on challenging molecular systems.
• These findings pave the way for more efficient and accurate molecular simulations in computational chemistry.