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

Ali Punjani; Haowei Zhang; David J. Fleet

. CC-BY-NC-ND 4.0 International license author/funder. It is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2019.12.15.877092 doi: bioRxiv preprint Figure 5 : Small membrane protein: Results of uniform and non-uniform refinement on a dataset of 42,740 particle images of a C3 symmetric 90kDa membrane protein with no soluble domains [6] . A: FSC curves computed using the same mask show numerical improvement from 3.9Å to 3.6Å, indicating improved global average resolution. B: Histograms of change in particle alignments between uniform and non-uniform refinement. Optimal regularization in non-uniform refinement improves the ability to align particles over iterations, causing these changes. D: Density map detail for transmembrane α-helices. Two (left, center) are peripheral, near the micelle. One (right) is central within the protein. Both maps are filtered using the corresponding FSC curve, sharpened with the same B-factor of −180Å 2 , no local filtering or sharpening is used, and thresholds are set to keep enclosed volume constant. Density differences are thus due to algorithmic rather than visualization differences. Non-uniform (purple) map has clear density for backbone and side-chains while for peripheral α-helices, helical pitch is barely resolved in uniform (grey) map density. The central α-helix (right) improves less dramatically. dataset, local regions of the map improve in quality more than the numerical resolution improvement might suggest. Figure 5C shows densities for three transmembrane α-helices from the protein, with two (left, center) being closer to the periphery of the protein near the micelle, and one (right) being closer to the center. The peripheral α-helices show significant improvement in the level of detail resolved in the density map, allowing tracing of backbone and side-chains in the non-uniform refinement result. By comparison, model building would be difficult with the corresponding uniform refinement result. The improvement of the central α-helix is less dramatic, and was chosen to indicate that some parts of the uniform refinement map are resolved at its nominal 3.9Å resolution. Figure 5B highlights the large changes in particle alignment pose and shift between uniform and non-uniform refinement that also occur with this dataset.