Toward Learned Chemical Perception of Force Field Typing Rules

Area of Science:

• Computational chemistry
• Molecular modeling
• Cheminformatics

Background:

• Molecular mechanics force fields are crucial for computational simulations but rely heavily on human expertise for parameterization.
• Defining atom types, which represent unique chemical environments, is a bottleneck due to its subjective and non-scalable nature.
• Current human-defined atom types can lack statistical rigor, leading to fitting issues and difficulties in extending to new chemical spaces.

Purpose of the Study:

• To develop automated methods for discovering chemical perception, specifically atom and fragment types, for molecular mechanics force fields.
• To replace subjective human input with statistically rigorous, data-driven approaches for force field development.
• To enhance the scalability and consistency of force field generation for diverse molecular systems.

Main Methods:

• Introduction of SMARTY for automated atom type discovery based on chemical perception.
• Introduction of SMIRKY for automated discovery of fragment types (nonbonded, bonds, angles, torsions).
• Utilizing Monte Carlo optimization with move sets in atom or fragment type space, driven by reference data.

Main Results:

• SMARTY successfully automates the rediscovery of human-defined atom types in existing small molecule force fields.
• SMIRKY automates the discovery of fragment types, demonstrating its capability in force field parameterization.
• The methods were validated using diverse molecular datasets, including a subset of the DrugBank database.

Conclusions:

• SMARTY and SMIRKY offer a robust, automated solution for chemical perception in force field development.
• These automated approaches address limitations of human-defined types, improving statistical justification and scalability.
• The developed methods pave the way for more efficient and consistent generation of force fields for various molecular systems.