Protein Structure Prediction with MELD x MD

Predicting protein structures from their amino acid sequences is a major challenge, especially with atomistic molecular dynamics (MD) simulations. The two major limitations to MD are poor sampling due to high computational cost and inaccurate force fields. Bioinformatics-based methods are an alternative approach to structure prediction that rely on databases of structures or structural fragments, as in threading. Threading has been very successful in predicting protein structures, but about 15% of proteins are not amenable to threading, the so-called nonthreadables. The nonthreadables provide a test set for MELD x MD, which does not rely on bioinformatics databases. We show that MELD x MD accurately predicts 20/41 nonthreadables and give high Boltzmann populations for successful predictions. We also discuss MELD x MD structure prediction on a larger set of proteins, particularly how the method is limited by sampling and force fields, but also how MELD x MD can overcome these limitations.