Cross-validation tests for cryo-electron microscopy using an independent set of images

Abstract

ABSTRACT: In addition to the chemical composition, information about the three-dimensional structure of a biomolecule is vital for understanding its biological function. For many years, resolv- ing structures of biomolecules was exclusive of X-ray crystallography and nuclear magnetic resonance (NMR) techniques. However, due to technological and software improvements, cryo-electron microscopy (cryo-EM) has emerged as an alternative for resolving complexes that were infeasible for crystallization or too large for NMR. Currently, cryo-EM is able to provide near-atomic resolution and close-to-native structures . Moreover, it enables extracting dynamical information, such as free-energy landscapes, from thermal states in the micrographs. The “resolution revolution” in cryo-EM has provoked an avalanche of reported cryo-EM maps. Recent statistics show an exponentially-growing number of reported maps spatially resolved by cryo-EM with their mean resolution decreasing from ∼ 10 Å (in 2013) to 4 Å (for 2018). The resolution revolution brings with it the need of creating robust and reliable methodolo- gies to validate the increasingly large number of maps. Some advances have been done along these lines: the tilt-pair analysis , the gold-standard procedure and the high-frequency randomization have shown to be reliable validation tools. However, it has recently been shown that these methods remain sensitive to overfitting (treating noise as true signal) and subjective criteria. In this work, I will present a novel methodology for validating cryo-EM maps. The method is based on cross-validation criteria where the reconstructed maps are compared against a set of experimental images (raw data) not used in the reconstruction procedure. Such comparison is carried out by calculating the probability that an image is the projection of a given map. The information from these probabilities led us to propose two validation criteria, which are tested over three well-behaved systems and two systems that present overfitting. The results prove that our methodology is able to identify overfitted maps.