Generating Global Protein Dynamics with Structure-From-FRET

Abstract

The remarkable specificity and diversity of protein function are consequences of native conformational fluctuations, and more specifically transitions between conformational substates. Experimentally, these dynamics are increasingly resolvable (with high temporal detail) using single molecule Forster Resonance Energy Transfer (smFRET) methods, which can capture rare structural extensions not discernible with bulk methods. Still, even with multiple fluorophore-pairs, dye-separation trajectories lack the atomistic detail of crystallographic or NMR-based methods, and, at present, merging smFRET data with high-resolution structural data remains problematic. Moreover, many functionally-relevant transitions proceed within timescales far beyond those accessible to current molecular dynamics approaches. We address this common tradeoff between detail and sampling by posing the following inference problem: given a single dynamic distance constraint, what is the most likely protein structure? For a solution, our Kalman-filter-inspired approach, Structure-from-FRET, generates an atomistic protein trajectory most likely responsible for the observed smFRET distance constraint: a statistical inference inverse problem. As a general analysis and visualization tool, we apply the method to the model hinge protein adenylate kinase. Millisecond-resolution FRET efficiencies indicate approximate Lid/Core distance traces, which constrain inference of global protein geometry. Resultant trajectories span extended temporal (seconds) and conformational (tens of Angstroms) scales, without imposed potentials or perturbations between reference states. Here, the inter-residue distances constrain a state-space estimation with backbone atom coordinates as variables. Experimental (constraint) uncertainty is explicitly considered and each update step maximizes spatial information from previous ‘frames’. We compare generated trajectories against those from interpolative or elastic network approaches, but also with known PDB structures and long run MD simulations. Results suggest Structure-from-FRET is reliable for modeling novel biocatalysts, inhibitor targets, or other proteins for which multiple structures or atomistic simulations are unavailable.