Assessment of uncertainty in protein backbone NMR assignments using Bayesian model averaging

Nuclear magnetic resonance (NMR) spectroscopy is a key experimental technique for determination of protein structure, dynamics, and interactions. However, the data it provides are noisy and incomplete, and there is no generally accepted measure of uncertainty in the resulting interpretations. This paper proposes a statistical modeling framework for an essential step in the NMR process, backbone resonance assignment. Our approach views NMR data as random observations, and employs Bayesian model averaging to quantify uncertainty in terms of standard errors. Results with experimental data demonstrate that the approach is able to quantify the ambiguity in spectral information as well as the ambiguity in assignment. The proposed framework is complementary to existing automated methods. It can benefit NMR spectroscopy by identifying regions of assigned sequence that are certain enough for use in further analysis (e.g. structure determination and dynamics), and by annotating deposited assignments with the amount of information provided by the particular protocols utilized.