PMC

The application of SIDESPLITTER during the refinement process. Flow diagrams illustrating (A) pseudo-independent half-set projection matching refinement, (B) the application of FSC weighting as currently common in refinement, and (C) the application of SIDESPLITTER during such a refinement. Sources of information transfer between the pseudo-independent sides of the refinement are indicated with magenta arrows.

Refinement of synthetic data of known local resolution demonstrates that over-fitting occurs during independent half-set refinement, and that SIDESPLITTER refinement reduces over-fitting. Panels indicate the ground-truth (the known input structure), the output of idealised refinement against ground-truth, the output of standard independent refinement, and the output of refinement with SIDESPLITTER. Panels inset show density peaks in low local resolution. Over-refinement manifests as over-emphasis upon these peak regions. All volumes are shown as surfaces. The FSC curves between half-sets within soft spherical masks isolating part of each segment are inset, coloured according to the rainbow from red to purple (0.0125 to 0.35 voxels per cycle respectively). The corresponding masks are shown above, within the overall mask used during refinement (in grey).

Refinement of experimental datasets with substantial regions of known low local resolution / SNR shows that SIDESPLITTER refinement suppresses features within regions of low local resolution, and improves the quality of the final density. Comparison between the results of standard and SIDESPLITTER refinement for (A) EMD-9849, (B) EMD-20806 and (C) EMD-4038. Volumes are shown as transparent surfaces. Regions of high local resolution are inset in each case, with the corresponding structure fitted into the densities in question, demonstrating that signal is retained to high resolution in regions with a high local signal to noise ratio. Representative improvements in the SIDESPLITTER map are highlighted with arrowheads. In each case the two reconstructions are compared at an identical contour. For the overview images on the left, the contour level was chosen as that sufficient to exclude all background noise, whereas for the inset images on the right the contour level was chosen at the point at which the main chains became clearly interpretable.

Refinement of synthetic data generated from the 20S proteasome structure (PDB entry 6bdf). Three different refinements were made using standard RELION, using SIDESPLITTER or incorporating deepEMhancer.