Third, data is frequently incomplete, or even sparse, and subject to experimental noise. Consequently, data obtained from such techniques yield incomplete, noisy, spatially and temporally averaged information on the Boltzmann ensemble of the observed system. Thus, such data are AbMole 3,4,5-Trimethoxyphenylacetic acid ideally analyzed through models that take these properties into account. While this fact has long been recognized, the analysis of these types of data has revolved predominantly around structure determination �C that is, fitting a single conformation to fulfill all derived geometrical restraints. Such structure determination methods do not adequately handle sparse, noisy and averaged data. Here, we propose an alternative method which addresses these shortcomings. Typically, structure determination from experimental data proceeds through hybrid energy minimization. In this method, an energy function Eexp that brings in the experimental data is combined with an approximative physical forcefield Ephys. The term Eexp is typically a straight-forward combination of a forwardand an error-model. A forward-model relates a protein conformation to experimental data, whereas an error-model concerns experimental errors. Alternatively, a Bayesian formulation known as inferential structure determination has been proposed, formulating structure determination in a rigorous probabilistic framework. In ISD, a posterior distribution is constructed by combining a data likelihood with prior distributions on conformational and nuisance parameters. The likelihood and the prior concerning biomolecular structure correspond to Eexp and Ephys, respectively. This Bayesian approach extends the common hybrid energy minimization by solving the important problems of choosing appropriate error-models, treating model-parameters coherently and performing inference through posterior sampling rather than minimization. However, by construction, homoscedastic fluctuations around one average structural representation. Consequently, the conformational heterogeneity present in the posterior distribution reflects the quality and completeness of the experimental data and the prior distributions, but not necessarily any physical fluctuations. Despite this well-known limitation, the approximation tends to yield good results for well-folded proteins when conformational fluctuations are modest. Early attempts to model ensemble NMR data involved averaging along molecular dynamics trajectories. In these protocols, a memory function specifies an averaging time-span which is used to obtain a time-averaged representation of the experimental data. While this approach displayed initial AbMole Gambogic-acid promise, the short timescales accessible through routine molecular dynamics limit its use.