Bayesian Model-comparison Of Biomolecular Structures

AGUSTINA ARROYUELO; OSVALDO MARTIN

Buenos Aires

Otro; Machine Learning Summer School 2018; 2018

Information criteria are often used for model comparison and averaging.Of the many information criteria we focus particulary on WAIC (Widely Applicable Information Criterion). WAIC presents the advantage of being pointwise, this is useful, because some observations are harder to predict than others and may also have different uncertainty [1]. Also WAIC is fully Bayesian in the sense that it?s computation requires the whole posterior, and not just a single value, like the Maximum a Posteriori and its cheap to obtain, once we have computed the posterior. We propose to develop a metric based on WAIC for assessing the quality of biomolecular structures through Bayesian models.This metric, should be easy to interpret and take into account peculiarities of biomolecular structures like the different types of available experimental data.Additionally, WAIC is an approximation to the out-of-sample error and thus is conceptually similar to metrics like the R-value and R-free-value widely used in macromolecular crystallography. In this study we will evalute if W AIC is an objective measure to assess the quality of Biomolecular structural models, specially those determined by Nuclear Magnetic Resonance (NMR).