Job: Non-Uniform Refinement

Non-uniform refinement.

Description

Apply non-uniform refinement to achieve higher resolution and map quality, especially for membrane proteins. Non-uniform refinement iteratively accounts for regions of a structure that have disordered or flexible density causing local loss of resolution. Accounting for these regions can significantly improve resolution in other regions as well as overall map quality by impacting the alignment of particles and reducing the tendency for refinement algorithms to over-fit disordered regions.

The non-uniform refinement algorithm is described in complete detail, including detailed explanations of the important parameters, in this Nature Methods publication: https://www.nature.com/articles/s41592-020-00990-8

The Non-uniform Refinement (New) job is nearly identical to the Homogeneous Refinement (New) job type. Only a few new parameters are introduced, and all other parameters and inputs have the same meaning and usage.

The Non-uniform Refinement (New) job type has specifically two new important features that can drastically improve results on smaller proteins and membrane proteins. One of those features is the non-uniform adaptive regularization described in the paper noted above (this was also present in the Legacy Non-uniform Refinement job, though in a much slower implementation). The second is a new, automatically adaptive latent variable marginalization method. This new method performs highly efficient pose and shift marginalization over particle alignments during reconstruction steps, with a sampling strategy that automatically tunes to each particle's signal quality and noise level. Adaptive Marginalization can improve 3D structure quality on smaller proteins, as well as sometimes limit the influence of outlier images. This improvement is separate from non-uniform adaptive regularization, and using both together can provide even further improves results. Adaptive marginalization is fully GPU accelerated and only increases refinement runtime by a small factor, often hidden by disk IO and other operations.

Non-uniform adaptive regularization in this job type is also fully GPU accelerated. This imposes a limitation on refinement box size depending on GPU memory availability. Currently, 3 * box_size^3 * 4 bytes are needed. Therefore, a GPU with 12GB of RAM may be able to run a volume up to 1024^3.

CryoSPARC v4.4 includes performance optimizations that significantly accelerate the non-uniform cross validation regularization stage of Non-uniform Refinement. The regularization itself is 7x-10x faster, and can yield approximately 2x faster total runtime for Non-uniform Refinement jobs compared to v4.3.1.

The improved performance does requires more GPU memory than the previous implementation, and therefore a parameter Low-memory mode is available that when set, causes the job to revert to the previous implementation. If you encounter errors indicating that GPU memory is insufficient, turn on this mode. This is likely to occur with box sizes above 600 on 11GB GPUs, 700 for 16GB GPUs, and 882 for 32GB GPUs.

Input

• Initial model

• Particles

• Mask (optional)

Common Parameters

• Adaptive Marginalization: On by default, described above.

• Non-uniform Refine Enable: On by default, enabling non-uniform regularization.

• Non-uniform filter order: Order of the butterworth filter used during regularization, described in the paper.

• Non-uniform AWF: The adaptive window factor, also described in the paper. Trade off between fast transitions between regions (AWF should be lower) and more accurate local cross-validation test (AWF should be higher). The default of 3 is good, can try as low as 1.5.

Output

• Refined 3D maps

• Half-maps

• Mask used in refinement

• Mask used in FSC calculation

• Gold-standard FSC curve

• Plots, including orientation distribution

Common Next Steps

• Download and inspect map