Transfer Learning from Markov Models Leads to Efficient Sampling of Related Systems

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Biophysics

Background:

• Metadynamics simulations require efficient sampling of slow molecular motions.
• Transfer learning offers potential for accelerating simulations of related systems.

Purpose of the Study:

• To develop and benchmark a transfer learning methodology for enhanced sampling of protein mutants.
• To enable efficient thermodynamic comparisons between wild-type and mutant proteins.

Main Methods:

• Extension of time-structure-based independent component analysis (tICA) with transfer learning.
• Utilizing sequence mapping to transfer slow modes and structural information.
• Reweighting simulations using the multistate Bennett acceptance ratio.

Main Results:

• Successful benchmarking on alanine dipeptide dynamics across various force fields.
• Demonstrated application to the FIP35 WW domain, capturing mutation effects.

Conclusions:

• The proposed transfer learning method significantly enhances sampling efficiency for protein mutants.
• This approach facilitates accurate thermodynamic characterization and comparison of protein variants.