MAINMAST-MELD-MDFF: Denovo Structure-Determination with Data-Guided Molecular Dynamics

Abstract

The past three decades of hybrid modeling has seen molecular dynamics (MD) simulations leverage NMR, X-ray crystallography and cryo EM data for the determination of medium- to high-resolution structures atomic structures. However, all these approaches, including our popular Molecular Dynamics Flexible Fitting (MDFF), and its various extensions work under the conventional molecular replacement paradigm, whereby any initial search model is morphed to satiate the data-imposed constraints. As a natural consequence, quality of the determined model remains heavily biased by choices of the initial model. Here, we deliver a novel modeling pipeline that iteratively combines minimum spanning tree-based backbone tracing tool (MAINMAST), Bayesian-likelihood based protein-folding methodology (MELD), and a resolution exchange-based fitting protocol (ReMDFF). Starting from only sequence information, the algorithm places C-alpha atoms into the density, fits a random coil to this C-alpha trace, generates protein secondary structures on-the-fly, and exhaustively samples the backbone and side chain geometries to deliver a refined model. Overcoming limitations of traditional approaches, the need for an initial model or homology information is completely subsided, and de novo modeling is now made feasible even at low resolutions. Available on cloud computing resources, the method is equally applicable for determining globular and transmembrane protein structures, as demonstrated for ubiquitin, TRPV1 and Magnesium-channel CorA (all in the 3-5 Å resolution range), and the de novo atomic structure of a lipoprotein flp3.