An efficient Bayesian kinetic lumping algorithm to identify metastable conformational states via Gibbs sampling.

Area of Science:

• Computational Biology
• Biophysics
• Statistical Mechanics

Background:

• Markov State Models (MSMs) predict long-timescale dynamics from molecular dynamics simulations.
• MSMs often contain numerous microstates, complicating mechanistic interpretation.
• Coarse-graining MSMs into macrostates aids understanding by lumping kinetically similar states.

Purpose of the Study:

• To introduce a probabilistic Gibbs lumping algorithm for kinetic lumping of MSMs.
• To enable Bayesian inference for assigning probabilities to kinetic lumpings.
• To facilitate the search for optimal macrostate models with high probability and metastability.

Main Methods:

• Developed a probabilistic Gibbs lumping algorithm using Bayesian inference.
• Modeled transitions between macrostates using Poisson processes to reflect timescale separation.
• Employed a Gibbs sampling algorithm to efficiently search for the optimal kinetic lumping.

Main Results:

• The Gibbs lumping algorithm assigns probabilities to kinetic lumpings.
• The method consistently produced lumped macrostate models with higher probability and metastability compared to six other algorithms.
• Demonstrated effectiveness on a 2D potential, alanine dipeptide, and a WW protein domain.

Conclusions:

• The Gibbs lumping algorithm provides a robust method for coarse-graining MSMs.
• This approach enhances the interpretability of complex biomolecular dynamics.
• The algorithm shows significant promise for investigating conformational changes in macromolecules.