Hi,

The adaptive window factor for non uniform refinement, if I understand it right, is a simple scaling factor - multiply the global resolution by the AWF and you get the local box size to use for the local res calculation (as in blocres).

The default for this parameter in “legacy” non uniform was 6, now it is 3 (with values as low as 1.5 suggested). Is this the same parameter as used for “legacy” non uniform? And what FSC threshold is being used (this is no longer exposed in the interface)?

Cheers

Oli

Hi @olibclarke,

The “New” non-uniform refinement uses a reparameterized/different calculation and definition of AWF than the “Legacy” implementation. The “New” version is exactly as is described in the Nature Methods paper: https://www.nature.com/articles/s41592-020-00990-8

As such, we no longer need local resolution/blocres/FSC thresholds and the only two parameters of the algorithm are AWF (default 3) and filter order (default 8). There is no finite local box size any more - all local estimation/filtering operations are done with an effective box size of the entire volume. There is also no FSC threshold as the new parameterization only relies on the definition of local filtering to define the local regularization parameters $\theta$ from the paper, rather than any explicit local resolution method that would have an FSC threshold.

The new parameterization and implementation is (in our hands) far more robust than the old one, much faster, and cleanly handles molecules of various sizes and SNR without changing params.

In fact in the paper (supp materials) we also have a test of different AWF choices and filter order choices on STRA6 and see that the results are quite insensitive to these.

Auto adaptive marginalization is independent of the above^ and also enabled by default in the “New” implementation - though note that this was not used in the paper, it’s a very recent addition. If you turn this off, things can run a fair bit faster (especially if there’s symmetry) and results will still be equal or slightly better than the “Legacy” implementation (and same as what’s in the paper). But if you leave it on, as you noticed, the results can be even better in many cases with smaller molecules or very poor SNR. We have found some nice improvements on GPCRs for example.

You can also turn off non-uniform regularization independently in the new job type - this leaves just fast homogeneous refinement + adaptive marginalization (and CTF refinement etc if you like). We haven’t played with this too much since we’ve never see non-uniform regularization make a result worse. It just costs a couple minutes more per iteration.

Hope this helps!

3 Likes

Ah I see! That is very helpful thank you Ali. Now I will read the paper more carefully!