Computational structural biology has emerged as a peta-scale computing challenge that is already in great demand by biologists worldwide. Applications beyond the peta-scale are already running on grid computers: for example, in just the last year, in just one structural genomics project, the low-resolution form of a de novo protein structure prediction code known as Rosetta, exceeded 400 million CPU-days of run-time spanning a million of nodes. However, algorithmic advances are needed for proper computational scaling, and application specialization is required for the high-throughput biological user community.
The accurate prediction of protein structure at high resolution is perhaps the longest standing grand challenge in molecular biology. The problem of ab initio protein structure prediction has always had two components: first determining a sufficiently accurate yet efficiently computable potential surface and second a means to search this extremely rugged, high dimensional space efficiently. Recent breakthroughs applying brute force computation (Bradley et al, Science 2005) demonstrate that existing potential models can extend to high resolution structure determination. Confirming evidence of sufficiency also appears in the inverse problem of designing sequences that faithfully obtain their designed structure: critically, atomically accurate has been demonstrated for completely novel protein architectures.
Thus the age when computation can compete with and exceed experimental methods for true high-resolution structure determination is now plausibly in sight for at least single domain sized protein structures. Moreover, computational models offer the ability to go beyond static measurement and look at many other facets of biological modeling and simulation inaccessible to experimentation. The challenge is to muster sufficient algorithmic performance improvements to scale this to the application end user needs and to improve the optimizations methods towards consistency in converging to solutions.