Structural refinment by DeepEMhancer map


If using deepEMhancer, is it possible to use the density map modified by deepEMhancer for structural refinement in phenix, refmac, etc.?
If that is not appropriate, should I use the deepEMhancer map for manually building an atomic model in coot and then use the density map obtained by the model-free sharpening tools available in cryosparc, phenix, etc. for structural refinement?

Thank you very much,

I don’t see why you could not use deepemhancer maps for manual building, or real-space refinement. It is just another locally filtered map. In practice though, deep maps are most useful for illustration purposes (and may help with manual building), while real space refinement tends to give very similar results regardless of the way you sharpen/filter your map… I would be worried if a map massively changes results of model refinement.


I would argue that deepEMhancer is doing something fundamentally different from sharpening/filtering techniques that are rooted in gold-standard segregated half-maps. In every seminar I’ve heard from the developers, they’ve only ever endorsed the tool as a modeling aid; ;and perhaps rightly so, given the lack of validation that it’s free of hallucination. The recent trend of presenting/depositing deepEMhanced maps–or worse, composite maps thereof–as primary data is worrying to me.


I see how this can turn into an epic battle.

Half-maps and masks should be deposited along with whatever map you feel will be most useful for others (to at least reproduce graphics in your paper).

On a side note: I don’t see a fundamental difference between our “manual” building and real-space refinement. It’s the same thing (unlike in X-ray refinement where map is changing with your model).


I think this is because deepEMhancer map is processed by a very tight mask generated by the predictive model.
In general, we believe that tight masks can produce artificial densities.
According to your conversasion, the deepEMhancer map facilitates modeling, but it seems that it should be used only as a reference.

Thank you,

And by the way, I would like to ask, what kind of filtering and sharpening tools do you use?

Thank you,

Specifically, it is a non-linear filter that approximates Fourier amplitude correction by LocScale (which otherwise requires a high-quality model with meaningful ADPs).

Why would this be any different than using model phases to improve the experimental map in xray? Agree with peter.cherepanov, as long as one deposits the raw half maps etc.

Why would this be any different than using model phases to improve the experimental map in xray?

In crystallography context, better phases improve the entire map, even for unmodeled or less-well modeled elements. Here, the learned filter that approximates the correct amplitude scaling operates in real space, and is only trained using protein examples. Consequently in my experience it tends to enforce a “protein-like” appearance on protein density (generally a good thing), but weakens non-protein densities, and has no guarantee of improving the entire map or having the same kinds of effects in all map regions.

Yes but if one models incorrectly in X-ray and then uses model phases to calculate a new map it has even more bias than that from sharpened EM map as the phases dominate the resultant maps. Bad amplitudes and good phases can still produce a decent X-ray map. Bad phases and perfect amplitudes will produce a garbage map. But point well taken regarding the inherent training bias and I agree. Maybe a combined local filter map and model map subtracted local filter map for modeling non protein densities? Will give it a try.

This is my favorite use of deepEMhanced maps! (other than making pretty figures).

Specifically, I load the atomic model, the experimental map (the one fresh out of 3D refinement, whichever job I consider “final”) and the corresponding deepEMhanced map into ISOLDE. I then set up ISOLDE to only use the experimental map for MDFF (so only this map pulls on the atoms), and I use the deepEMhanced map only as a visual guide. Very often, it suggests side chain conformations that end up being happily stable in the combined pull of the MD force field and experimental map (which, often, is too weak to show side chain density, and without the deepEMhanced map I would have a difficult time finding a stable conformation for this side chain).