Assessment of Embedding Schemes in a Hybrid Machine Learning/Classical Potentials (ML/MM) Approach.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning in Science

Background:

• Machine learning (ML) methods accurately predict molecular properties in vacuo but struggle with large system simulations.
• ML/MM (Machine Learning/Molecular Mechanics) methods combine ML with classical force fields for simulating complex systems, analogous to QM/MM.
• Accurate coupling description between ML and MM regions is crucial for ML/MM methods, with electrostatics being key.

Purpose of the Study:

• To evaluate mechanical embedding approaches for ML/MM coupling, focusing on atomic partial charges.
• To compare different atomic charge schemes and assess the impact of polarization corrections.
• To benchmark against QM(DFT)/MM electrostatic embedding calculations for accuracy.

Main Methods:

• Utilized ML/MM framework with in vacuo-derived atomic partial charges.
• Investigated various atomic charge schemes, including minimal basis iterative stockholder (MBIS).
• Incorporated a polarization correction using atomic polarizabilities.
• Benchmarked against QM(DFT)/MM electrostatic embedding on ~80k organic molecules from ANI-1x/ANI-2x databases solvated in water.

Main Results:

• Minimal basis iterative stockholder (MBIS) atomic charges showed the best agreement with reference coupling energies.
• A simple polarization correction significantly enhanced the accuracy of the mechanical embedding approaches.
• ML/MM methods with appropriate charge schemes and polarization corrections can effectively model inter-region interactions.

Conclusions:

• Mechanical embedding, particularly using MBIS charges and a polarization correction, is a viable and accurate strategy for ML/MM coupling.
• This approach provides a computationally efficient alternative to more expensive electrostatic embedding methods for large-scale simulations.
• The findings pave the way for more accurate and scalable molecular simulations using hybrid ML/MM techniques.