The limits of homology modeling

The computational prediction of three-dimensional structures of protein sequences may be performed using a wide variety of techniques, such as homology modeling or threading. In threading, the correct fold is searched for by evaluating the energy of the intended sequence when it is "forced" to adopt each of the known folding patterns. In homology modeling, one looks for a high-similarity protein sequence with experimentally-determined 3D structure, and mutates it in silico until the desired sequence is obtained. Many different programs and web-servers are now available for these tasks, differing among themselves in the forcefields used, alignment algorithms, etc. Performance is usually quite good when templates with similarity >40% are used.

Recently, two small proteins with very high homology (>95%) but widely differing structure have been designed and studied. Starting from a pair of proteins with < 20 % identity and different 3D structures, the authors gradually mutated one sequence into the other, and ended up generating two sequences differing only in one amino acid, but with different folds. Attempts to unravel the precise mechanisms governing the selection of one fold over the other have however been inconclusive, because current molecular dynamics protocols and force fields are not accurate enough to measure the small energy differences involved.