We model loop regions of proteins based on a fragment assembly method. The fragments that
comprise candidates of the local structure of a protein loop are collected from a structure database,
and all loop conformations possible from a smooth assembly of the fragments are generated. For
each of the fragment-assembled conformations, a Monte Carlo simulation in the conformational
subspace that satises the loop closure constraint is performed to minimize the root-mean-square
deviation of the backbone dihedral angles from the fragment angles. The side-chains are then
built using a rotamer library, and the backbone and the side-chain conformations are optimized
locally with the AMBER 96 force eld without solvation terms to remove steric clashes. The
resulting conformations are then ranked using the DFIRE potential. A test prediction for eight
protein loops with sizes ranging from 8 to 12 residues is presented to show the feasibility of our
method. Tests with further optimization using Monte Carlo with minimization show that extensive
conformational optimization leading to deviation from the original fragment-assembled structures
tend to deteriorate the prediction accuracy, suggesting that the utilization of fragment information
is superior to purely energy-based methods.