Exploring the quality of protein structural models from a Bayesian perspective

JORGE VILA; AGUSTINA ARROYUELO; OSVALDO MARTIN

Congreso; 1st Congress of Women in Bioinformatics and Data Science Latin America; 2020

Weexplore how ideas and practices common in Bayesian modeling can be applied in the contextof Nuclear Magnetic Resonance ( protein structure validation We fit a Bayesian hierarchicallinear model to experimental and theoretical 13 C α Chemical S hifts 1 The latter ones arecomputed from 3 D protein models through the Che Shift 2 software 2 Based on the premisethat evaluation of the Bayesian model´s fit may reveal aspects of the quality of a 3 D proteinstructure, we propose such evaluation in terms of the Expected Log point wise PredictiveDensity ( estimated by Pareto smooth importance sampling leave one out cross validationPSIS LOO CV), LOO for short LOO offers an accurate, reliable and fast estimate of the out ofsample prediction accuracy (prediction accuracy computed from data not used to train themodel) without requiring to re fit the model 3 Exact leave one out cross validation requires tore fit a model n times, with n being the size of the data set (i e the number of 13 C α ChemicalShifts Finally, we present visualizations that can help interpret these comparisons