3D Protein Structure
Biologists have not had much success in the direct computer prediction of protein tertiary structure from raw sequence data (known as ab initio prediction). A 100 aa protein has 3 to the 200th power possible backbone configurations - many orders of magnitude beyond the capacity of even the fastest computers to evaluate. Yet a protein reliably folds into its true native 3D structure in solution in a few milliseconds. This is known as Levinthal's paradox.
On the other hand, there are probably only a few hundred basic protein sub-structures, but we don't yet have this vocabulary or the ability to recognize variants on a theme.
Algorithms have been developed that attempt to minimize the overall free energy of a protein molecule using either Monte-Carlo methods or Neural Net software. These applications run on mainframes or supercomputers - not readily accessible to the average biologist. Structural predictions from these programs are still unreliable.
The best approach to predicting 3D protein structure is to compare an unknown protein with similar sequences that have known structures. This is known as threading. However, threading only works for proteins with about 25% sequence similarity to a protein with known structure. The current state of the art in threading software requires many days of computing on a dedicated workstation in the hands of an expert. However, some websites offer quick approximations of threading computations. Overall, threading tecniques will improve as more 3-D structures are described.
Here are a few 3D structure prediction Web servers
The UCLA-DOE Protein Fold Recognition Server
Finally, even if a tertiary structure is predicted for a protein, techniques have not yet been developed for inferring the functional properties implied by that structure.
Using Computers for Molecular Biology
Stuart M. Brown, Ph.D., RCR, NYU Medical Center Comments to: browns02@mcrcr.med.nyu.edu