Search for a command to run...
Abstract Single particle cryo-EM is a powerful method for studying proteins and other biological macromolecules. Many of these molecules comprise regions with varying structural properties including disorder, flexibility, and partial occupancy. These traits make computational 3D reconstruction from 2D images challenging. Detergent micelles and lipid nanodiscs, used to keep membrane proteins in solution, are common examples of locally disordered structures that can negatively affect existing iterative refinement algorithms which assume rigidity (or spatial uniformity). We introduce a cross-validation approach to derive non-uniform refinement , an algorithm that automatically regularizes 3D density maps during iterative refinement to account for spatial variability, yielding dramatically improved resolution and 3D map quality. We find that in common iterative refinement methods, regularization using spatially uniform filtering operations can simultaneously over- and under-regularize local regions of a 3D map. In contrast, non-uniform refinement removes noise in disordered regions while retaining signal useful for aligning particle images. Our results include state-of-the-art resolution 3D reconstructions of multiple membrane proteins with molecular weight as low as 90kDa. These results demonstrate that higher resolutions and improved 3D density map quality can be achieved even for small membrane proteins, an important use case for single particle cryo-EM, both in structural biology and drug discovery. Non-uniform refinement is implemented in the cryoSPARC software package and has already been used successfully in several notable structural studies.