Machine Learning Quantum Mechanical/Molecular Mechanical Potentials: Evaluating Transferability in Dihydrofolate...

Area of Science:

• Computational chemistry
• Biochemistry
• Machine learning

Background:

• Integrating machine learning potentials (MLPs) with quantum mechanical/molecular mechanical (QM/MM) simulations offers a powerful method for studying enzymatic catalysis.
• Generating training data for MLPs via QM/MM simulations is time-consuming and system-specific, hindering practical applications.

Purpose of the Study:

• To evaluate the transferability of a pretrained ΔMLP model across different enzyme mutations and environments.
• To assess the need for retraining MLPs for new enzyme systems or variants.

Main Methods:

• Utilized a QM/MM-based ML architecture to test a pretrained ΔMLP model.
• Evaluated transferability across single point substitutions, homologous enzymes, and aqueous environments.
• Compared free energy profiles with and without MLP retraining.

Main Results:

• The ΔMLP model accurately predicted the effects of enzyme mutations on electrostatic interactions and free energy profiles without retraining.
• Transferability was robust across various enzyme mutations and homologous enzymes.
• Significant limitations in transferability were observed when transitioning to water-rich molecular mechanics environments.

Conclusions:

• The QM/MM-based ML architecture demonstrates robustness for diverse enzyme systems.
• Pretrained MLPs can reduce the computational burden of studying enzymatic catalysis.
• Further research is needed to improve MLP transferability, especially for enzyme-to-solvent transitions.