Non-uniform refinement: Adaptive regularization improves single particle cryo-EM reconstruction

Ali Punjani; Haowei Zhang; David J. Fleet

We introduce non-uniform refinement, a new algorithm for achieving higher resolution and higher quality reconstructions of protein structures from single particle cryo-EM data. Non-uniform refinement uses cross-validation regularization as a principled tool for parameter estimation in noisy inverse problems like cryo-EM reconstruction. In this way, we are able to incorporate into the algorithm, and exploit, the domain knowledge that many types of proteins contain regions with varying levels of disorder, flexibility, or occupancy. These variations lead uniform refinement algorithms to under-or over-regularize regions of a 3D density map during refinement. Instead, non-uniform refinement automatically identifies and treats local regions optimally, without introducing bias into "gold standard" estimation of resolution with the FSC and without any specific knowledge about any particular protein.