On metastable conformational analysis of non-equilibrium biomolecular time series

We present a recently developed clustering method and specify it for the problem of identification of metastable conformations in nonequilibrium biomolecular time series. The approach is based on variational minimization of some novel regularized clustering functional. In context of conformational analysis, it allows one to combine the features of standard geometrical clustering techniques (like the Kmeans algorithm), dimension reduction methods (like principle component analysis), and dynamical machine learning approaches like hidden Markov models (HMMs). In contrast to the HMM-based approaches, no a priori assumptions about Markovianity of the underlying process and regarding probability distribution of the observed data are needed. The application of the computational framework is exemplified by means of conformational analysis of some penta-peptide torsion angle time series from a molecular dynamics simulation. Comparison of different versions of the presented algorithm is performed w.r.t. the metastability and geometrical resolution of the resulting conformations