The paper presented the approach of Perron cluster analysis (PCCA) as developed up to now by the author, SCHUTTE, and their co-workers. The key idea of this approach is to directly identify almost invariant sets of the Hamiltonian system as metastable conformations together with their life spans and transition patterns. This leads to the numerical solution of a Perron cluster eigenvalue problem for some Markov operator. Discretization of that operator via hybrid Monte Carlo methods generates transition matrices for nearly uncoupled Markov chains. The discretization of the Markov operator requires careful consideration to avoid the curse of dimension. Three variants have been developed: 1/Temperature embedding via Boltzmann distribution; 2/PCCA combined with Kohonen's self-organized neural networks; 3/Successive PCCA of dihedrals.