To scrutinize how a protein folds at atomic resolution, we performed 200 molecular dynamics simulations (each of 50 ns) of the miniprotein Trp-cage on the computational grid. Within the trajectories, 58 folding and 31 unfolding events were identified and subjected to extensive comparison and classification. Based on an analogy with biological sequences, the folding and unfolding trajectories (arrays of sequential snapshots of structures) were aligned by dynamic programming allowing gaps. A phylogenetic tree derived from the alignments revealed four distinct groups of the trajectories, characterized by the Trp side-chain motions and the main-chain motions. It was found that only one group attained the native structure and that the other three led to pseudonative structures having the correct main-chain trace but different nonnative Trp side-chain rotamers, indicating that those four folded structures were each attained through a unique folding pathway.grid computing ͉ simulation ͉ trajectory alignment ͉ Trp-cage P rotein molecules rapidly fold into a unique, native structure despite the vast size of conformational space in the unfolded state (1). To resolve the paradox between the large space and the fast process, the folding pathway (2) and folding funnel (3) models have been proposed. The pathway model emphasizes narrow pathways in the conformational space, whereas the funnel theory focuses on quick folding on a funnel-like potential surface. A number of experimental results, particularly for small proteins, have been interpreted in terms of either the pathway or funnel model (4-11). However, these views do not necessarily present a comprehensive picture explaining in detail how a specific protein folds into the native structure. Recently, folding simulation at atomic resolution has become a realistic possibility to study the folding process of small proteins (12-18) in spite of the extremely large computational burden. An atomically detailed picture of the folding process could explain how a specific protein folds and complement the generic theories.The difficulty in simulating folding lies not only in the relatively short simulated time scale but also in two other issues. The first is the representation of the highly stochastic nature of protein motion. A single observation of the folding event cannot be a representative of various trajectories in the folding process. An ensemble comprising a diversity of folding trajectories is indispensable to develop a rigorous understanding. The second problem is how to analyze the complicated folding trajectories in a high-dimensional space. Projection of the motions onto two-to three-dimensional space has been used for small peptides (19,20), but it does not have high enough resolution to depict detailed dynamical features in a larger system. We need a more sophisticated technique to analyze large sets of trajectory data.In this study we calculated an ensemble of folding trajectories of a small protein, a 20-residue miniprotein Trp-cage [TC5b; NLYIQ WLKDG GPSSG RPPPS; P...