Density-based cluster algorithms for the identification of core sets.

Area of Science:

• Computational Chemistry
• Biophysics
• Data Science

Background:

• Markov state models (MSMs) are crucial for analyzing complex molecular dynamics.
• The core-set approach discretizes conformational space into metastable regions, but requires prior identification of these regions.
• Existing methods for identifying core sets can be computationally intensive and limited in scope.

Purpose of the Study:

• To investigate the use of density-based clustering algorithms for identifying core sets in molecular dynamics.
• To compare the performance of different density-based clustering algorithms (CNN, DBSCAN, Jarvis-Patrick) in constructing core-set models.
• To extend the applicability of the core-set method to systems exhibiting marginal metastability.

Main Methods:

• Application of three density-based clustering algorithms: CNN, DBSCAN, and Jarvis-Patrick.
• Construction and comparison of core-set models based on each clustering algorithm.
• Hierarchical density-based clustering with monitoring of the metric matrix structure.
• Testing the approach on molecular dynamics simulations of a flexible peptide.

Main Results:

• Core-set models derived from CNN and DBSCAN clustering demonstrated good convergence.
• Core-set models based on Jarvis-Patrick clustering were found to be unreliable.
• Converged core-set models achieved a significant reduction in the number of states compared to conventional MSMs.
• The density-based clustering approach successfully extended the core-set method to systems with marginal metastability.

Conclusions:

• Density-based clustering, particularly CNN and DBSCAN, provides an effective method for identifying core sets in molecular dynamics.
• This approach enables the construction of more efficient and broadly applicable core-set Markov state models.
• The method allows for high-resolution models capable of distinguishing subtle conformational differences, crucial for biological systems.