High-resolution prediction of protein structures from their amino acid sequences and refinement of low-resolution protein structure models to produce more accurate structures are long-standing challenges in computational biology. With the explosion of genomic information and continued increase in experimentally solved structures, it is possible to produce homology models for many proteins whose structure is not experimentally solved. NMR offers a powerful alternative to structure determination especially when crystallographic methods are arduous or fail. However, NMR structures produced from limited experimental data and homology models from distant sequences are inherently low-resolution.

We present an energy-based rebuilding-and-refinement method that consistently generates models of atomic-level accuracy starting from low-resolution NMR models or homology models.

A stringent test and concomitant application of predicting structures to high resolution is to solve the x-ray crystallographic phase problem by molecular replacement. Estimating crystallographic phases by molecular replacement requires models of high accuracy (less than 1.5 ? backbone RMSD to native structure). We have shown that models produced by our refinement methodology have the accuracy and are, consequently, significantly better candidates for molecular replacement when compared to the best template from PDB or an NMR structure of the same protein.