Meld Folds Nonthreadable Proteins

Protein structure prediction is important because not all protein structures are easily determined from experiment and being able to accurately and reliably predict structure from sequence allows for understanding structure-function relationships and provides structures for structure-based drug design. The state of the art in protein structure prediction is to use bioinformatics-based methods such as threading. This method is suitable for approximately 85% of the human proteome, but poses a problem for the remaining 15% of proteins that remain inaccessible, or “nonthreadable”. Bioinformatics algorithms fail to predict nonthreadable proteins because the required fragment structures are missing from the databases that threading relies upon to build an accurate protein model. When databases lack the necessary information to solve the problem, methods that rely on databases have no way to come up with a correct solution. An alternative approach is needed to solve nonthreadable protein structures. A physics-based (and database-free) method using atomistic molecular dynamics simulations should accurately predict these structures. Additionally, physics-based methods are advantageous because, in addition to providing native state protein structures, they provide dynamics and folding pathways; the simulations also obey Boltzmann's law and provide populations, thus giving access to folding free energies, which can be used to further refine our physics-based approach. Physics-based atomistic simulations have been limited by the computational expense of folding proteins, but MELD greatly accelerates folding simulations by using replica exchange molecular dynamics that incorporates native contact predictions in a probabilistic manner. We show that MELD successfully folds a number of nonthreadable proteins.